From pubconference at gmail.com Tue Feb 1 00:35:24 2022 From: pubconference at gmail.com (Pub Conference) Date: Tue, 1 Feb 2022 00:35:24 -0500 Subject: Connectionists: Neural Computing and Applications (NCAA) - Call for Topical Collectionon (Deadline March 31, 2022) Message-ID: Neural Computing and Applications Topical Collection on Interpretation of Deep Learning: Prediction, Representation, Modeling and Utilization https://www.springer.com/journal/521/updates/19187658 Aims, Scope and Objective While Big Data offers the great potential for revolutionizing all aspects of our society, harvesting of valuable knowledge from Big Data is an extremely challenging task. The large scale and rapidly growing information hidden in the unprecedented volumes of non-traditional data requires the development of decision-making algorithms. Recent successes in machine learning, particularly deep learning, has led to breakthroughs in real-world applications such as autonomous driving, healthcare, cybersecurity, speech and image recognition, personalized news feeds, and financial markets. While these models may provide the state-of-the-art and impressive prediction accuracies, they usually offer little insight into the inner workings of the model and how a decision is made. The decision-makers cannot obtain human-intelligible explanations for the decisions of models, which impede the applications in mission-critical areas. This situation is even severely worse in complex data analytics. It is, therefore, imperative to develop explainable computation intelligent learning models with excellent predictive accuracy to provide safe, reliable, and scientific basis for determination. Numerous recent works have presented various endeavors on this issue but left many important questions unresolved. The first challenging problem is how to construct self-explanatory models or how to improve the explicit understanding and explainability of a model without the loss of accuracy. In addition, high dimensional or ultra-high dimensional data are common in large and complex data analytics. In these cases, the construction of interpretable model becomes quite difficult and complex. Further, how to evaluate and quantify the explainability of a model is lack of consistent and clear description. Moreover, auditable, repeatable, and reliable process of the computational models is crucial to decision-makers. For example, decision-makers need explicit explanation and analysis of the intermediate features produced in a model, thus the interpretation of intermediate processes is requisite. Subsequently, the problem of efficient optimization exists in explainable computational intelligent models. These raise many essential issues on how to develop explainable data analytics in computational intelligence. This Topical Collection aims to bring together original research articles and review articles that will present the latest theoretical and technical advancements of machine and deep learning models. We hope that this Topical Collection will: 1) improve the understanding and explainability of machine learning and deep neural networks; 2) enhance the mathematical foundation of deep neural networks; and 3) increase the computational efficiency and stability of the machine and deep learning training process with new algorithms that will scale. Potential topics include but are not limited to the following: - Interpretability of deep learning models - Quantifying or visualizing the interpretability of deep neural networks - Neural networks, fuzzy logic, and evolutionary based interpretable control systems - Supervised, unsupervised, and reinforcement learning - Extracting understanding from large-scale and heterogeneous data - Dimensionality reduction of large scale and complex data and sparse modeling - Stability improvement of deep neural network optimization - Optimization methods for deep learning - Privacy preserving machine learning (e.g., federated machine learning, learning over encrypted data) - Novel deep learning approaches in the applications of image/signal processing, business intelligence, games, healthcare, bioinformatics, and security Guest Editors Nian Zhang (Lead Guest Editor), University of the District of Columbia, Washington, DC, USA, nzhang at udc.edu Jian Wang, China University of Petroleum (East China), Qingdao, China, wangjiannl at upc.edu.cn Leszek Rutkowski, Czestochowa University of Technology, Poland, leszek.rutkowski at pcz.pl Important Dates Deadline for Submissions: March 31, 2022 First Review Decision: May 31, 2022 Revisions Due: June 30, 2022 Deadline for 2nd Review: July 31, 2022 Final Decisions: August 31, 2022 Final Manuscript: September 30, 2022 Peer Review Process All the papers will go through peer review, and will be reviewed by at least three reviewers. A thorough check will be completed, and the guest editors will check any significant similarity between the manuscript under consideration and any published paper or submitted manuscripts of which they are aware. In such case, the article will be directly rejected without proceeding further. Guest editors will make all reasonable effort to receive the reviewer?s comments and recommendation on time. The submitted papers must provide original research that has not been published nor currently under review by other venues. Previously published conference papers should be clearly identified by the authors at the submission stage and an explanation should be provided about how such papers have been extended to be considered for this special issue (with at least 30% difference from the original works). Submission Guidelines Paper submissions for the special issue should strictly follow the submission format and guidelines ( https://www.springer.com/journal/521/submission-guidelines). Each manuscript should not exceed 16 pages in length (inclusive of figures and tables). Manuscripts must be submitted to the journal online system at https://www.editorialmanager.com/ncaa/default.aspx. Authors should select ?TC: Interpretation of Deep Learning? during the submission step ?Additional Information?. -------------- next part -------------- An HTML attachment was scrubbed... URL: From donatello.conte at univ-tours.fr Tue Feb 1 03:15:00 2022 From: donatello.conte at univ-tours.fr (Donatello Conte) Date: Tue, 1 Feb 2022 09:15:00 +0100 Subject: Connectionists: =?iso-8859-1?q?CFP_-_IEEE_Transactions_on_Emergin?= =?iso-8859-1?q?g_Topics_in_Computing__Special_Issue_on_=22Emerging?= =?iso-8859-1?q?_Trends_and_Advances_in_Graph-based_Methods_and_App?= =?iso-8859-1?q?lications=22?= Message-ID: <008a01d81743$cef378e0$6cda6aa0$@univ-tours.fr> ******************************************************** CFP - Apologies for multiple posting ******************************************************** Dear Colleague, we would like to invite you to submit a paper to the Special Section on ?Emerging Trends and Advances in Graph-based Methods and Applications? to be published in IEEE Transactions on Emerging Topics in Computing (ISSN 21686750, IF 7.691) ******************************************************** TOPICS OF INTEREST ******************************************************** - Advances in computing and learning on graphs - Graph Neural Networks (GNNs) - Graph data fusion methods and graph embedding techniques - Efficient, parallel, and distributed processing frameworks for big graphs - Novel dynamic, spatial, and temporal graphs for recognition and learning - Emerging graph-based methods in computer vision - Interactivity, explainability, and trust in graph-based learning methods - Applications of GNNs - Human behavior and scene understanding using graphs - Benchmarks for GNNs - Graph signal processing - Application of graph data processing in biology, healthcare, transportation, natural language processing, social networks, etc. For further details please visit the special issue website: https://www.computer.org/digital-library/journals/ec/graph-based-methods ******************************************************** IMPORTANT DATES ******************************************************** Submission deadline: 30 June 2022 First review notification: 30 November 2022 Revision submission: 15 February 2023 Second review notification: 30 April 2023 Journal publication: first half of 2023 Should you have any further enquiry please do not hesitate to contact us. Best regards, Donatello Conte, Universit? de Tours, Laboratoire d?Informatique Fondamentale et Appliqu?e de Tours, France Alessandro D?Amelio, Universit? degli Studi di Milano, Dipartimento di Informatica, Italy Raffaella Lanzarotti, Universit? degli Studi di Milano, Dipartimento di Informatica, Italy Jianyi Lin, Universit? Cattolica del Sacro Cuore, Dipartimento di Scienze Statistiche, Italy Jean-Yves Ramel, Universit? de Tours, Laboratoire d?Informatique Fondamentale et Appliqu?e de Tours, France (Guest Editors) -------------- next part -------------- An HTML attachment was scrubbed... URL: From Tom.Verguts at UGent.be Tue Feb 1 02:15:57 2022 From: Tom.Verguts at UGent.be (Tom Verguts) Date: Tue, 1 Feb 2022 07:15:57 +0000 Subject: Connectionists: new book Message-ID: <73CB88CB-269A-43E4-AEF5-C242366DB25A@ugent.be> Dear colleagues This book may be useful for your teaching of modeling cognitive processes: https://mitpress.mit.edu/books/introduction-modeling-cognitive-processes best wishes, Tom -------------- next part -------------- An HTML attachment was scrubbed... URL: From m.plumbley at surrey.ac.uk Tue Feb 1 04:09:57 2022 From: m.plumbley at surrey.ac.uk (Mark Plumbley) Date: Tue, 1 Feb 2022 09:09:57 +0000 Subject: Connectionists: JOB: Head of Department of Computer Science, University of Surrey Message-ID: Dear All, (with apologies for crossposting) This opportunity for a new Head of Department may be of interest to some people on this list. I would particularly like to encourage applications from women, disabled and Black, Asian and minority ethnic candidates, since these groups are currently underrepresented at this level in our area. Please forward to anyone who you think may be interested. Many thanks, Mark --- Head of Department of Computer Science, University of Surrey Location: Guildford, UK Closing date: 22 February 2022 The University of Surrey seeks applications from exceptional leaders for the post of Head of Department of Computer Science. This is an exciting opportunity to lead a growing Department, developing and implementing its future strategy, further enhancing its research profile, and ensuring students at all levels receive an outstanding experience. The new Head will bring experience, drive and aspirational leadership to support the members of the Department to achieve the highest standards in research and teaching. The Department of Computer Science has a long-standing reputation for its vibrant and supportive teaching and research environment. We have a growing complement of around 30 academic staff, and a current profile of 400+ undergraduate students, 100+ Masters students, and 40+ PhD students. Our Computer Science BSc and Computing and Information Technology BSc programmes have been running successfully for many years and continue to attract strong students. The Department offers Information Security MSc and Data Science MSc programmes with healthy student numbers. The Department is home to Surrey Centre for Cyber Security, which has world-leading research expertise in applied cryptography, trusted computing, secure systems, privacy and authentication, secure communications, blockchain and distributed ledger technologies, and security verification. The Distributed and Networked Systems group is internationally recognised for its work in consensus protocols, distributed trust and coordination, fault-tolerance, edge and cloud computing, networks in space, web tracking and privacy, and online harms. The Nature Inspired Computing and Engineering (NICE) group specialises in machine learning, trustworthy AI, optimisation, systems biology, and image processing, and is a key partner in the new Surrey Institute for People-Centred AI. As part of the School of Computer Science and Electronic Engineering, we also have strong links with other Research Centres in the School, including the Centre for Vision Speech and Signal Processing, the Institute for Communication Systems, and the Surrey Space Centre. The successful candidate will bring broad-based experience of visible and inclusive senior leadership, as well as effective management of people and resources. They will demonstrate outstanding achievements in teaching and research, with international recognition in their academic field. The University and the Department specifically are committed to building a culturally diverse organisation, and the Department was awarded a Bronze Athena SWAN award. Applications are strongly encouraged from female and minority candidates. Applicants are welcome to contact Prof Mark Plumbley, Head of School of Computer Science and Electronic Engineering (m.plumbley at surrey.ac.uk) for further information or to discuss the post. For more information and to apply, visit: https://jobs.surrey.ac.uk/083821 -- Prof Mark D Plumbley Head of School of Computer Science and Electronic Engineering Email: Head-of-School-CSEE at surrey.ac.uk Professor of Signal Processing University of Surrey, Guildford, Surrey, GU2 7XH, UK Email: m.plumbley at surrey.ac.uk PA: Kelly Green Email: k.d.green at surrey.ac.uk -------------- next part -------------- An HTML attachment was scrubbed... URL: From barak at pearlmutter.net Tue Feb 1 06:17:12 2022 From: barak at pearlmutter.net (Barak A. Pearlmutter) Date: Tue, 1 Feb 2022 11:17:12 +0000 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: References: <27D911A3-9C51-48A6-8034-7FF3A3E89BBB@princeton.edu> <2f1d9928-543f-f4a0-feab-5a5a0cc1d4d7@rubic.rutgers.edu> <307D9939-4F3A-40FF-A19F-3CEABEAE315C@supsi.ch> <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> Message-ID: J?rgen, It's fantastic that you're helping expose people to some important bits of scientific literature. But... > Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions If you actually read Minsk and Papert's "Perceptrons" book, this is not a misconception it encourages. It defines a "k-th order perceptron" as a linear threshold unit preceded by an arbitrary set of fixed nonlinearities with fan-in k. (A linear threshold unit with binary inputs would, in this terminology, be a 1st-order perceptron.) All their theorems are for k>1. For instance, they prove that a k-th order perceptron cannot do (k+1)-bit parity, which in the special case of k=1 simplifies to the trivial observation that a simple linear threshold unit cannot do xor. [image: perceptrons-book-cover-1.jpg] [image: perceptron-diagram-1.jpg] This is why you're not supposed to directly cite things you have not actually read: it's too easy to misconstrue them based on inaccurate summaries transmitted over a series of biased noisy compressive channels. Cheers, --Barak. -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: perceptrons-book-cover-1.jpg Type: image/jpeg Size: 20796 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: perceptron-diagram-1.jpg Type: image/jpeg Size: 541745 bytes Desc: not available URL: From juergen at idsia.ch Tue Feb 1 06:33:17 2022 From: juergen at idsia.ch (Schmidhuber Juergen) Date: Tue, 1 Feb 2022 11:33:17 +0000 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: References: <27D911A3-9C51-48A6-8034-7FF3A3E89BBB@princeton.edu> <2f1d9928-543f-f4a0-feab-5a5a0cc1d4d7@rubic.rutgers.edu> <307D9939-4F3A-40FF-A19F-3CEABEAE315C@supsi.ch> <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> Message-ID: Thanks, Barak! Indeed, I should have said in the email msg that _others_ interpreted the book of Minsky & Papert [M69] in this way. My report explicitly mentions Terry [S20] who wrote in 2020: "The great expectations in the press (Fig. 3) were dashed by Minsky and Papert (7), who showed in their book Perceptrons that a perceptron can only represent categories that are linearly separable in weight space. Although at the end of their book Minsky and Papert considered the prospect of generalizing single- to multiple-layer perceptrons, one layer feeding into the next, they doubted there would ever be a way to train these more powerful multilayer perceptrons. Unfortunately, many took this doubt to be definitive, and the field was abandoned until a new generation of neural network researchers took a fresh look at the problem in the 1980s.? However, as mentioned above, the 1969 book [M69] addressed a "problem" of Gauss & Legendre's shallow learning (~1800) [DL1-2] that had already been solved 4 years prior by Ivakhnenko & Lapa's popular deep learning method [DEEP1-2][DL2] (and then also in 1967 by Amari's SGD for MLPs [GD1-2]). Deep learning research was not abandoned in the 1970s. It was alive and kicking, especially outside of the Anglosphere. [DEEP2][GD1-3][CNN1][DL1-2] See Sec. II and Sec. XIII of the report: https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html Cheers, J?rgen > On 1 Feb 2022, at 14:17, Barak A. Pearlmutter wrote: > > J?rgen, > > It's fantastic that you're helping expose people to some important bits of scientific literature. > > But... > > > Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions > > If you actually read Minsk and Papert's "Perceptrons" book, this is not a misconception it encourages. It defines a "k-th order perceptron" as a linear threshold unit preceded by an arbitrary set of fixed nonlinearities with fan-in k. (A linear threshold unit with binary inputs would, in this terminology, be a 1st-order perceptron.) All their theorems are for k>1. For instance, they prove that a k-th order perceptron cannot do (k+1)-bit parity, which in the special case of k=1 simplifies to the trivial observation that a simple linear threshold unit cannot do xor. > > This is why you're not supposed to directly cite things you have not actually read: it's too easy to misconstrue them based on inaccurate summaries transmitted over a series of biased noisy compressive channels. > > Cheers, > > --Barak. From pmitra3 at gmail.com Tue Feb 1 09:38:46 2022 From: pmitra3 at gmail.com (Prasenjit Mitra) Date: Tue, 1 Feb 2022 09:38:46 -0500 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: References: <27D911A3-9C51-48A6-8034-7FF3A3E89BBB@princeton.edu> <2f1d9928-543f-f4a0-feab-5a5a0cc1d4d7@rubic.rutgers.edu> <307D9939-4F3A-40FF-A19F-3CEABEAE315C@supsi.ch> <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> Message-ID: <03E01C81-9B5B-417B-990A-45AB9642046F@gmail.com> All this is fascinating discussion. I fear that this will be lost in the bowels of some mailing list archive. We will repeat our not knowing or misrepresenting what was done in the past and not knowing people?s opinions and perspectives of what led to what if we do not document this well. Is there somewhere we should put all this information Wiki-style? Maybe in a wikipedia page and a whole set of connected, hierarchically organized pages? Or sort of a history of ML and related topics Wikipedia or wiki site? Does something like this already exist and I am woefully naive? If so, please provide some pointers someone. The reason I say this is because it can and perhaps has been captured in books, but books are not deliberative in nature. That is, it will reflect the opinions (and biases) of an author but not the deliberations of a community. Despite edit wars, the history of the edits and reverses provides valuable information if done thoughtfully with the reasoning provided, which I expect academics to do. I am not wedded to a solution but I would love this to be captured and preserved and added and shaped and refined perhaps to come to some consensus or consensuses ? Best, Prasenjit (Mitra) Professor, Penn State > On Feb 1, 2022, at 6:17 AM, Barak A. Pearlmutter wrote: > > J?rgen, > > It's fantastic that you're helping expose people to some important bits of scientific literature. > > But... > > > Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions > > If you actually read Minsk and Papert's "Perceptrons" book, this is not a misconception it encourages. It defines a "k-th order perceptron" as a linear threshold unit preceded by an arbitrary set of fixed nonlinearities with fan-in k. (A linear threshold unit with binary inputs would, in this terminology, be a 1st-order perceptron.) All their theorems are for k>1. For instance, they prove that a k-th order perceptron cannot do (k+1)-bit parity, which in the special case of k=1 simplifies to the trivial observation that a simple linear threshold unit cannot do xor. > > This is why you're not supposed to directly cite things you have not actually read: it's too easy to misconstrue them based on inaccurate summaries transmitted over a series of biased noisy compressive channels. > > Cheers, > > --Barak. From juergen at idsia.ch Tue Feb 1 09:46:32 2022 From: juergen at idsia.ch (Schmidhuber Juergen) Date: Tue, 1 Feb 2022 14:46:32 +0000 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: References: <27D911A3-9C51-48A6-8034-7FF3A3E89BBB@princeton.edu> <2f1d9928-543f-f4a0-feab-5a5a0cc1d4d7@rubic.rutgers.edu> <307D9939-4F3A-40FF-A19F-3CEABEAE315C@supsi.ch> <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> Message-ID: <0736A8A3-3CF9-40E1-A7EE-E061FD52F0FD@supsi.ch> Yes, Barak, in hindsight we know well that both Rosenblatt [R58-62] and Minsky [M69] were wrong in several ways. But by now we have known that for many decades! Obviously, today?s surveys should reflect this. However, the previously mentioned self-serving DL surveys (2015-2021) [DL3, DL3a, S20] are still promulgating the same old misleading PDP narratives, although their authors know better. Alarmingly, revisionist surveys of the so-called "DL conspiracy" and the similar "ELM conspiracy? (mentioned in the previous msg) are still being cited! What?s wrong with our field? J?rgen References: https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html > On 1 Feb 2022, at 17:07, Barak A. Pearlmutter wrote: > > > "The great expectations in the press (Fig. 3) were dashed by Minsky and Papert (7), > > Just to be clear, below are some of those "great expectations in the press" that Minsky and Papert were referring to. The first I've included is actually by Rosenblatt. So, Minsky and Papert were wrong in their self-declared intuition that learning in multilayer systems wouldn't be possible. Maybe evidence was even available before they wrote that paragraph, although they were apparently not aware of it. But Rosenblatt was wrong too, in his predictions. And he wasn't pitching it as his intution: he said it was a done deal, a fait accompli! So looking back, both were wrong. But who was *more* wrong? > > --Barak. > > > From pfbaldi at ics.uci.edu Tue Feb 1 09:40:40 2022 From: pfbaldi at ics.uci.edu (Baldi,Pierre) Date: Tue, 1 Feb 2022 06:40:40 -0800 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: References: <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> Message-ID: <6f02b384-6b29-431f-1235-babfcd6d8b16@ics.uci.edu> M&P wrote: " The perceptron has shown itself worthy of study despite (and even because of!) its severe limitations, It has many features to attract attention: its linearity; its intriguing learning theorem; its clear paradigmatic simplicity as a kind of parallel computation.There is no reason to suppose that any of these virtues carry over to the many-layered version.***Nevertheless, we consider it to be an important research problem to elucidate (or reject) our intuitive judgment that the extension is sterile. Perhaps some powerful convergence theorem will be discovered, or some profound reason for the failure to produce an interesting "learning theorem" for the multilayered machine will be found." (pp231-232)* * * The claim that they killed NNs in the 1970s is indeed an exaggeration propagated by the PDP group for obvious reasons. There was plenty of NN research in the 1970s--Amari, Grossberg, etc. Not surprisingly, the same is true of the so-called "second neural network winter". **** * * On 2/1/2022 3:33 AM, Schmidhuber Juergen wrote: > Thanks, Barak! Indeed, I should have said in the email msg that _others_ interpreted the book of Minsky & Papert [M69] in this way. My report explicitly mentions Terry [S20] who wrote in 2020: > > "The great expectations in the press (Fig. 3) were dashed by Minsky and Papert (7), who showed in their book Perceptrons that a perceptron can only represent categories that are linearly separable in weight space. Although at the end of their book Minsky and Papert considered the prospect of generalizing single- to multiple-layer perceptrons, one layer feeding into the next, they doubted there would ever be a way to train these more powerful multilayer perceptrons. Unfortunately, many took this doubt to be definitive, and the field was abandoned until a new generation of neural network researchers took a fresh look at the problem in the 1980s.? > > However, as mentioned above, the 1969 book [M69] addressed a "problem" of Gauss & Legendre's shallow learning (~1800) [DL1-2] that had already been solved 4 years prior by Ivakhnenko & Lapa's popular deep learning method [DEEP1-2][DL2] (and then also in 1967 by Amari's SGD for MLPs [GD1-2]). Deep learning research was not abandoned in the 1970s. It was alive and kicking, especially outside of the Anglosphere. [DEEP2][GD1-3][CNN1][DL1-2] > > See Sec. II and Sec. XIII of the report:https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html > > Cheers, > J?rgen > > > > >> On 1 Feb 2022, at 14:17, Barak A. Pearlmutter wrote: >> >> J?rgen, >> >> It's fantastic that you're helping expose people to some important bits of scientific literature. >> >> But... >> >>> Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions >> If you actually read Minsk and Papert's "Perceptrons" book, this is not a misconception it encourages. It defines a "k-th order perceptron" as a linear threshold unit preceded by an arbitrary set of fixed nonlinearities with fan-in k. (A linear threshold unit with binary inputs would, in this terminology, be a 1st-order perceptron.) All their theorems are for k>1. For instance, they prove that a k-th order perceptron cannot do (k+1)-bit parity, which in the special case of k=1 simplifies to the trivial observation that a simple linear threshold unit cannot do xor. >> >> This is why you're not supposed to directly cite things you have not actually read: it's too easy to misconstrue them based on inaccurate summaries transmitted over a series of biased noisy compressive channels. >> >> Cheers, >> >> --Barak. > -- Pierre Baldi, Ph.D. Distinguished Professor, Department of Computer Science Director, Institute for Genomics and Bioinformatics Associate Director, Center for Machine Learning and Intelligent Systems University of California, Irvine Irvine, CA 92697-3435 (949) 824-5809 (949) 824-9813 [FAX] Assistant: Janet Kojko at uci.edu -------------- next part -------------- An HTML attachment was scrubbed... URL: From barak at pearlmutter.net Tue Feb 1 09:07:28 2022 From: barak at pearlmutter.net (Barak A. Pearlmutter) Date: Tue, 1 Feb 2022 14:07:28 +0000 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: References: <27D911A3-9C51-48A6-8034-7FF3A3E89BBB@princeton.edu> <2f1d9928-543f-f4a0-feab-5a5a0cc1d4d7@rubic.rutgers.edu> <307D9939-4F3A-40FF-A19F-3CEABEAE315C@supsi.ch> <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> Message-ID: > "The great expectations in the press (Fig. 3) were dashed by Minsky and Papert (7), Just to be clear, below are some of those "great expectations in the press" that Minsky and Papert were referring to. The first I've included is actually by Rosenblatt. So, Minsky and Papert were wrong in their self-declared intuition that learning in multilayer systems wouldn't be possible. Maybe evidence was even available before they wrote that paragraph, although they were apparently not aware of it. But Rosenblatt was wrong too, in his predictions. And he wasn't pitching it as his intution: he said it was a done deal, a fait accompli! So looking back, both were wrong. But who was *more* wrong? --Barak. [image: perceptron-cornell-research-trends-1958.jpg] [image: perceptron-rl-nyt-p25-8-Jul-1958.png] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: perceptron-cornell-research-trends-1958.jpg Type: image/jpeg Size: 301560 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: perceptron-rl-nyt-p25-8-Jul-1958.png Type: image/png Size: 714307 bytes Desc: not available URL: From pfbaldi at ics.uci.edu Tue Feb 1 13:22:29 2022 From: pfbaldi at ics.uci.edu (Baldi,Pierre) Date: Tue, 1 Feb 2022 10:22:29 -0800 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: <6f02b384-6b29-431f-1235-babfcd6d8b16@ics.uci.edu> References: <6f02b384-6b29-431f-1235-babfcd6d8b16@ics.uci.edu> Message-ID: <3c452866-ffc9-e144-7c6a-d03e8079a048@ics.uci.edu> In their Perceptrons book, Minsky and Papert wrote: " The perceptron has shown itself worthy of study despite (and even because of!) its severe limitations, It has many features to attract attention: its linearity; its intriguing learning theorem; its clear paradigmatic simplicity as a kind of parallel computation.There is no reason to suppose that any of these virtues carry over to the many-layered version.***Nevertheless, we consider it to be an important research problem to elucidate (or reject) our intuitive judgment that the extension is sterile. Perhaps some powerful convergence theorem will be discovered, or some profound reason for the failure to produce an interesting "learning theorem" for the multilayered machine will be found." (pp231-232)* * * The claim that they killed NNs in the 1970s is indeed an exaggeration propagated by the PDP group for obvious reasons. There was plenty of NN research in the 1970s--Amari, Grossberg, etc.? Not surprisingly, the same is true of the so-called "second neural network winter". **** On 2/1/2022 3:33 AM, Schmidhuber Juergen wrote: > Thanks, Barak! Indeed, I should have said in the email msg that _others_ interpreted the book of Minsky & Papert [M69] in this way. My report explicitly mentions Terry [S20] who wrote in 2020: > > "The great expectations in the press (Fig. 3) were dashed by Minsky and Papert (7), who showed in their book Perceptrons that a perceptron can only represent categories that are linearly separable in weight space. Although at the end of their book Minsky and Papert considered the prospect of generalizing single- to multiple-layer perceptrons, one layer feeding into the next, they doubted there would ever be a way to train these more powerful multilayer perceptrons. Unfortunately, many took this doubt to be definitive, and the field was abandoned until a new generation of neural network researchers took a fresh look at the problem in the 1980s.? > > However, as mentioned above, the 1969 book [M69] addressed a "problem" of Gauss & Legendre's shallow learning (~1800) [DL1-2] that had already been solved 4 years prior by Ivakhnenko & Lapa's popular deep learning method [DEEP1-2][DL2] (and then also in 1967 by Amari's SGD for MLPs [GD1-2]). Deep learning research was not abandoned in the 1970s. It was alive and kicking, especially outside of the Anglosphere. [DEEP2][GD1-3][CNN1][DL1-2] > > See Sec. II and Sec. XIII of the report:https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html > > Cheers, > J?rgen > > > > >> On 1 Feb 2022, at 14:17, Barak A. Pearlmutter wrote: >> >> J?rgen, >> >> It's fantastic that you're helping expose people to some important bits of scientific literature. >> >> But... >> >>> Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions >> If you actually read Minsk and Papert's "Perceptrons" book, this is not a misconception it encourages. It defines a "k-th order perceptron" as a linear threshold unit preceded by an arbitrary set of fixed nonlinearities with fan-in k. (A linear threshold unit with binary inputs would, in this terminology, be a 1st-order perceptron.) All their theorems are for k>1. For instance, they prove that a k-th order perceptron cannot do (k+1)-bit parity, which in the special case of k=1 simplifies to the trivial observation that a simple linear threshold unit cannot do xor. >> >> This is why you're not supposed to directly cite things you have not actually read: it's too easy to misconstrue them based on inaccurate summaries transmitted over a series of biased noisy compressive channels. >> >> Cheers, >> >> --Barak. -- Pierre Baldi, Ph.D. Distinguished Professor, Department of Computer Science Director, Institute for Genomics and Bioinformatics Associate Director, Center for Machine Learning and Intelligent Systems University of California, Irvine Irvine, CA 92697-3435 (949) 824-5809 (949) 824-9813 [FAX] Assistant: Janet Kojko at uci.edu -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Tue Feb 1 15:19:43 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Tue, 1 Feb 2022 20:19:43 +0000 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: <6f02b384-6b29-431f-1235-babfcd6d8b16@ics.uci.edu> References: <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> <6f02b384-6b29-431f-1235-babfcd6d8b16@ics.uci.edu> Message-ID: I had some communication with Marvin Minsky and Jerry Fodor years ago. Here are two quotes from them: 1. Marvin Minsky: "Don?t pay any attention to the critics. Don?t even ignore them.? -By Sam Goldwyn 2. Jerry Fodor: ?Arguing with Connectionists is like arguing with zombies; both are dead, but neither has noticed it.? Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Baldi,Pierre Sent: Tuesday, February 1, 2022 7:41 AM To: Schmidhuber Juergen ; connectionists at cs.cmu.edu Subject: Re: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. M&P wrote: " The perceptron has shown itself worthy of study despite (and even because of!) its severe limitations, It has many features to attract attention: its linearity; its intriguing learning theorem; its clear paradigmatic simplicity as a kind of parallel computation. There is no reason to suppose that any of these virtues carry over to the many-layered version. Nevertheless, we consider it to be an important research problem to elucidate (or reject) our intuitive judgment that the extension is sterile. Perhaps some powerful convergence theorem will be discovered, or some profound reason for the failure to produce an interesting "learning theorem" for the multilayered machine will be found." (pp231-232) The claim that they killed NNs in the 1970s is indeed an exaggeration propagated by the PDP group for obvious reasons. There was plenty of NN research in the 1970s--Amari, Grossberg, etc. Not surprisingly, the same is true of the so-called "second neural network winter". On 2/1/2022 3:33 AM, Schmidhuber Juergen wrote: Thanks, Barak! Indeed, I should have said in the email msg that _others_ interpreted the book of Minsky & Papert [M69] in this way. My report explicitly mentions Terry [S20] who wrote in 2020: "The great expectations in the press (Fig. 3) were dashed by Minsky and Papert (7), who showed in their book Perceptrons that a perceptron can only represent categories that are linearly separable in weight space. Although at the end of their book Minsky and Papert considered the prospect of generalizing single- to multiple-layer perceptrons, one layer feeding into the next, they doubted there would ever be a way to train these more powerful multilayer perceptrons. Unfortunately, many took this doubt to be definitive, and the field was abandoned until a new generation of neural network researchers took a fresh look at the problem in the 1980s.? However, as mentioned above, the 1969 book [M69] addressed a "problem" of Gauss & Legendre's shallow learning (~1800) [DL1-2] that had already been solved 4 years prior by Ivakhnenko & Lapa's popular deep learning method [DEEP1-2][DL2] (and then also in 1967 by Amari's SGD for MLPs [GD1-2]). Deep learning research was not abandoned in the 1970s. It was alive and kicking, especially outside of the Anglosphere. [DEEP2][GD1-3][CNN1][DL1-2] See Sec. II and Sec. XIII of the report: https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html Cheers, J?rgen On 1 Feb 2022, at 14:17, Barak A. Pearlmutter wrote: J?rgen, It's fantastic that you're helping expose people to some important bits of scientific literature. But... Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions If you actually read Minsk and Papert's "Perceptrons" book, this is not a misconception it encourages. It defines a "k-th order perceptron" as a linear threshold unit preceded by an arbitrary set of fixed nonlinearities with fan-in k. (A linear threshold unit with binary inputs would, in this terminology, be a 1st-order perceptron.) All their theorems are for k>1. For instance, they prove that a k-th order perceptron cannot do (k+1)-bit parity, which in the special case of k=1 simplifies to the trivial observation that a simple linear threshold unit cannot do xor. This is why you're not supposed to directly cite things you have not actually read: it's too easy to misconstrue them based on inaccurate summaries transmitted over a series of biased noisy compressive channels. Cheers, --Barak. -- Pierre Baldi, Ph.D. Distinguished Professor, Department of Computer Science Director, Institute for Genomics and Bioinformatics Associate Director, Center for Machine Learning and Intelligent Systems University of California, Irvine Irvine, CA 92697-3435 (949) 824-5809 (949) 824-9813 [FAX] Assistant: Janet Ko jko at uci.edu -------------- next part -------------- An HTML attachment was scrubbed... URL: From bisant at umbc.edu Tue Feb 1 13:27:34 2022 From: bisant at umbc.edu (David B) Date: Tue, 1 Feb 2022 13:27:34 -0500 Subject: Connectionists: FLAIRS-35 May 15-18, 2022, Jensen Beach, Florida, Paper Deadline Message-ID: FLAIRS-35 Special Track on Neural Networks and Data Mining The Florida Artificial Intelligence Research Symposium (FLAIRS) is an interdisciplinary conference which features double-blind reviewing, free tutorials, and a warm and sunny venue. Abstract Due Date: February 7, 2022 Submission Due Date: February 14, 2022 Conference: May 15-18, 2022 Jensen Beach, Florida Website: https://sites.google.com/view/flairs-35-nn-dm-track/home URL: https://www.flairs-35.info/call-for-papers Papers are being solicited for a special track on Neural Networks and Data Mining at the 35th International FLAIRS Conference (https://www.flairs-35.info/home). This special track will be devoted to neural networks and data mining with the aim of presenting new and important contributions in these areas. Papers and contributions are encouraged for any work related to neural networks, data mining, or the intersection thereof. Topics of interest may include (but are in no way limited to): applications such as Pattern Recognition, Control and Process Monitoring, Biomedical Applications, Robotics, Text Mining, Diagnostic Problems, Telecommunications, Power Systems, Signal Processing; Intelligence analysis, medical and health applications, text, video, and multi-media mining, E-commerce and web data, financial data analysis, cyber security, remote sensing, earth sciences, bioinformatics, and astronomy; algorithms such as new developments in Back Propagation, RBF, SVM, Deep Learning, Ensemble Methods, Kernel Approaches; hybrid approaches such as Neural Networks/Genetic Algorithms, Neural Network/Expert Systems, Causal Nets trained with Backpropagation, and Neural Network/Fuzzy Logic applications such as Intelligence analysis, medical and health applications, text, video, and multi-media mining, E-commerce and web data, financial data analysis, cyber security, remote sensing, earth sciences, bioinformatics, and astronomy; modeling algorithms such as hidden Markov models, decision trees, neural networks, statistical methods, or probabilistic methods; case studies in areas of application, or over different algorithms and approaches; graph modeling, pattern discovery, and anomaly detection; feature extraction and selection; post-processing techniques such as visualization, summarization, or trending; preprocessing and data reduction; and knowledge engineering or warehousing. Questions regarding the track should be addressed to: David Bisant at bisant at umbc.edu, Steven Gutstein at s.m.gutstein at gmail.com, or Bill Eberle at weberle at tntech.edu. From jose at rubic.rutgers.edu Tue Feb 1 16:48:32 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Tue, 1 Feb 2022 16:48:32 -0500 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: References: <2f1d9928-543f-f4a0-feab-5a5a0cc1d4d7@rubic.rutgers.edu> <307D9939-4F3A-40FF-A19F-3CEABEAE315C@supsi.ch> <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> Message-ID: <7f97db1f-13e1-ba48-8b02-f3a2c4769df9@rubic.rutgers.edu> Jeurgen:? Even some of us lowly psychologists know some math.?? And its not about the math.. its about the context (is this sounding like an echo?) Let me try again.. and I think your good intentions but misguided reconstruction of history is appearing to me, to be? perverse. You tip your hand when you talk about "rebranding".??? Also that the PDP books were a "conspiracy". But lets go point by point. (1) we already agreed that the Perceptron? was not linear regression--lets not go backwards. Closer to logistic regression.?? If you are talking about Widrow and Hoff, well it is the Delta Rule-- SSE kind of regression.?? But where did the Delta rule come from?? Lets look at Math.? So there is some nice papers by Gluck and Thompson (80s) showing how Pavlovian conditioning is exactly the Delta rule and even more relevant was shown to account for majority of classical (pavlovian) conditioning was the Rescorla-Wagner (1972) model-- \Delta V_A = [\alpha_A\beta_1](\lambda_1 - V_{AX}), which of course was Ivan Petrovich Pavlov discovery of classical conditioning (1880s). Why aren't you citing him??? What about John Brodeus Watson and Burris Fredrick Skinner???????? At least they were focused on learning albeit? *just* in biological systems.? But these were actual? natural world discoveries. (2) Function approximation.?? Ok Juergen, claims that everything?? is really?? just X, reminds me of the man with a Hammer to whom everything looks like a nail!????? To the point: its incidental.? Yes, Neural networks are function approximators, but that is incidental to the original more general context (PDP)? as a way to create "internal representations".?? The function approximation was a Bonus! (3) Branding.?? OMG.? So you seem to believe that everyone is cynical and will put their intellectual finger in the air to find out what to call what they are doing!?? Jeez, I hope this isn't true.? But the narrative you eschew is in fact something that Minsky would talk about (I remember this at lunch with him in the 90s at Thinking Machines), and he was quite clear that Perceptron was failing well? before the 1969 book (trying to do speech recognition with a perceptron--yikes), but in a piling on kind of way Perceptrons killed the perceptron, but it was the linearity focus (as BAP points out) and the lack of depth. (4) Group Method of Handling Data.?? Frankly, the only one I can find that branded GMHD as a NeuroNet (as they call it) was you. There is a 2017 reference, but they reference you again. (5) Its just names,? fashion and preference.. or no actual concepts matter.? Really? There was an french mathematician named Fourier in the 19th century who came up with an idea of periodic function decomposition into weighted trigonometric functions.. but he had no math.?? And Laplace Legendre and others said he had no math!? So they prevented him from publishing for FIFTEEN YEARS..?? 150 years later after Tukey invented the FFT, its the most common transform used and misused? in general. Concepts lead to math.. and that may lead to further formalism.. but don't mistake the math for the concept behind it.??? The context matters and you are confusing syntax for semantics! Cheers, Steve On 1/31/22 11:38 AM, Schmidhuber Juergen wrote: > Steve, do you really want to erase the very origins of shallow learning (Gauss & Legendre ~1800) and deep learning (DL, Ivakhnenko & Lapa 1965) from the field's history? Why? Because they did not use modern terminology such as "artificial neural nets (NNs)" and "learning internal representations"? Names change all the time like fashions; the only thing that counts is the math. Not only mathematicians but also psychologists like yourself will agree. > > Again: the linear regressor of Legendre & Gauss is formally identical to what was much later called a linear NN for function approximation (FA), minimizing mean squared error, still widely used today. No history of "shallow learning" (without adaptive hidden layers) is complete without this original shallow learner of 2 centuries ago. Many NN courses actually introduce simple NNs in this mathematically and historically correct way, then proceed to DL NNs with several adaptive hidden layers. > > And of course, no DL history is complete without the origins of functional DL in 1965 [DEEP1-2]. Back then, Ivakhnenko and Lapa published the first general, working DL algorithm for supervised deep feedforward multilayer perceptrons (MLPs) with arbitrarily many layers of neuron-like elements, using nonlinear activation functions (actually Kolmogorov-Gabor polynomials) that combine both additions (like in linear NNs) and multiplications (basically they had deep NNs with gates, including higher order gates). They incrementally trained and pruned their DL networks layer by layer to learn internal representations, using regression and a separate validation set (network depth > 7 by 1971). They had standard justifications of DL such as: "a multilayered structure is a computationally feasible way to implement multinomials of very high degree" [DEEP2] (that cannot be approximated by simple linear NNs). Of course, their DL was automated, and many people have used it up to the 2000s ! > - just follow the numerous citations. > > I don't get your comments about Ivakhnenko's DL and function approximation (FA). FA is for all kinds of functions, including your "cognitive or perceptual or motor functions." NNs are used as FAs all the time. Like other NNs, Ivakhnenko's nets can be used as FAs for your motor control problems. You boldly claim: "This was not in the intellectual space" of Ivakhnenko's method. But obviously it was. > > Interestingly, 2 years later, Amari (1967-68) [GD1-2] trained his deep MLPs through a different DL method, namely, stochastic gradient descent (1951-52)[STO51-52]. His paper also did not contain the "modern" expression "learning internal representations in NNs." But that's what it was about. Math and algorithms are immune to rebranding. > > You may not like the fact that neither the original shallow learning (Gauss & Legendre ~1800) nor the original working DL (Ivakhnenko & Lapa 1965; Amari 1967) were biologically inspired. They were motivated through math and problem solving. The NN rebranding came later. Proper scientific credit assignment does not care for changes in terminology. > > BTW, unfortunately, Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions. But actually he had much more interesting MLPs with a non-learning randomized first layer and an adaptive output layer. So Rosenblatt basically had what much later was rebranded as "Extreme Learning Machines (ELMs)." The revisionist narrative of ELMs (see this web site https://elmorigin.wixsite.com/originofelm) is a bit like the revisionist narrative of DL criticized by my report. Some ELM guys apparently thought they can get away with blatant improper credit assignment. After all, the criticized DL guys seemed to get away with it on an even grander scale. They called themselves the "DL conspiracy" [DLC]; the "ELM conspiracy" is similar. What an embarrassing lack of maturity of our field. > > Fortunately, more and more ML researchers are helping to set things straight. "In science, by definition, the facts will always win in the end. As long as the facts have not yet won it's not yet the end." [T21v1] > > References as always under https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html > > J?rgen > > >> On 27 Jan 2022, at 17:37, Stephen Jos? Hanson wrote: >> >> >> >> Juergen, I have read through GMHD paper and a 1971 Review paper by Ivakhnenko. These are papers about function approximation. The method proposes to use series of polynomial functions that are stacked in filtered sets. The filtered sets are chosen based on best fit, and from what I can tell are manually grown.. so this must of been a tedious and slow process (I assume could be automated). So are the GMHDs "deep", in that they are stacked 4 deep in figure 1 (8 deep in another). Interestingly, they are using (with obvious FA justification) polynomials of various degree. Has this much to do with neural networks? Yes, there were examples initiated by Rumelhart (and me: https://www.routledge.com/Backpropagation-Theory-Architectures-and-Applications/Chauvin-Rumelhart/p/book/9780805812596), based on poly-synaptic dendrite complexity, but not in the GMHD paper.. which was specifically about function approximation. Ivakhnenko, lists four reasons for the approach t! > hey took: mainly reducing data size and being more efficient with data that one had. No mention of "internal representations" >> So when Terry, talks about "internal representations" --does he mean function approximation? Not so much. That of course is part of this, but the actual focus is on cognitive or perceptual or motor functions. Representation in the brain. Hidden units (which could be polynomials) cluster and project and model the input features wrt to the function constraints conditioned by training data. This is more similar to model specification through function space search. And the original Rumelhart meaning of internal representation in PDP vol 1, was in the case of representation certain binary functions (XOR), but more generally about the need for "neurons" (inter-neurons) explicitly between input (sensory) and output (motor). Consider NETTALK, in which I did the first hierarchical clustering of the hidden units over the input features (letters). What appeared wasn't probably surprising.. but without model specification, the network (w.hidden units), learned VOWELS and ! > CONSONANT distinctions just from training (Hanson & Burr, 1990). This would be a clear example of "internal representations" in the sense of Rumelhart. This was not in the intellectual space of Ivakhnenko's Group Method of Handling Data. (some of this is discussed in more detail in some recent conversations with Terry Sejnowski and another one to appear shortly with Geoff Hinton (AIHUB.org look in Opinions). >> Now I suppose one could be cynical and opportunistic, and even conclude if you wanted to get more clicks, rather than title your article GROUP METHOD OF HANDLING DATA, you should at least consider: NEURAL NETWORKS FOR HANDLING DATA, even if you didn't think neural networks had anything to do with your algorithm, after all everyone else is! Might get it published in this time frame, or even read. This is not scholarship. These publications threads are related but not dependent. And although they diverge they could be informative if one were to try and develop polynomial inductive growth networks (see Falhman, 1989; Cascade correlation and Hanson 1990: Meiosis nets) to motor control in the brain. But that's not what happened. I think, like Gauss, you need to drop this specific claim as well. >> >> With best regards, >> >> Steve > > On 25 Jan 2022, at 20:03, Schmidhuber Juergen wrote: > > PS: Terry, you also wrote: "Our precious time is better spent moving the field forward.? However, it seems like in recent years much of your own precious time has gone to promulgating a revisionist history of deep learning (and writing the corresponding "amicus curiae" letters to award committees). For a recent example, your 2020 deep learning survey in PNAS [S20] claims that your 1985 Boltzmann machine [BM] was the first NN to learn internal representations. This paper [BM] neither cited the internal representations learnt by Ivakhnenko & Lapa's deep nets in 1965 [DEEP1-2] nor those learnt by Amari?s stochastic gradient descent for MLPs in 1967-1968 [GD1-2]. Nor did your recent survey [S20] attempt to correct this as good science should strive to do. On the other hand, it seems you celebrated your co-author's birthday in a special session while you were head of NeurIPS, instead of correcting these inaccuracies and celebrating the true pioneers of deep learning, such as ! > Ivakhnenko and Amari. Even your recent interview https://blog.paperspace.com/terry-sejnowski-boltzmann-machines/ claims: "Our goal was to try to take a network with multiple layers - an input layer, an output layer and layers in between ? and make it learn. It was generally thought, because of early work that was done in AI in the 60s, that no one would ever find such a learning algorithm because it was just too mathematically difficult.? You wrote this although you knew exactly that such learning algorithms were first created in the 1960s, and that they worked. You are a well-known scientist, head of NeurIPS, and chief editor of a major journal. You must correct this. We must all be better than this as scientists. We owe it to both the past, present, and future scientists as well as those we ultimately serve. > > The last paragraph of my report https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html quotes Elvis Presley: "Truth is like the sun. You can shut it out for a time, but it ain't goin' away.? I wonder how the future will reflect on the choices we make now. > > J?rgen > > >> On 3 Jan 2022, at 11:38, Schmidhuber Juergen wrote: >> >> Terry, please don't throw smoke candles like that! >> >> This is not about basic math such as Calculus (actually first published by Leibniz; later Newton was also credited for his unpublished work; Archimedes already had special cases thereof over 2000 years ago; the Indian Kerala school made essential contributions around 1400). In fact, my report addresses such smoke candles in Sec. XII: "Some claim that 'backpropagation' is just the chain rule of Leibniz (1676) & L'Hopital (1696).' No, it is the efficient way of applying the chain rule to big networks with differentiable nodes (there are also many inefficient ways of doing this). It was not published until 1970 [BP1]." >> >> You write: "All these threads will be sorted out by historians one hundred years from now." To answer that, let me just cut and paste the last sentence of my conclusions: "However, today's scientists won't have to wait for AI historians to establish proper credit assignment. It is easy enough to do the right thing right now." >> >> You write: "let us be good role models and mentors" to the new generation. Then please do what's right! Your recent survey [S20] does not help. It's mentioned in my report as follows: "ACM seems to be influenced by a misleading 'history of deep learning' propagated by LBH & co-authors, e.g., Sejnowski [S20] (see Sec. XIII). It goes more or less like this: 'In 1969, Minsky & Papert [M69] showed that shallow NNs without hidden layers are very limited and the field was abandoned until a new generation of neural network researchers took a fresh look at the problem in the 1980s [S20].' However, as mentioned above, the 1969 book [M69] addressed a 'problem' of Gauss & Legendre's shallow learning (~1800)[DL1-2] that had already been solved 4 years prior by Ivakhnenko & Lapa's popular deep learning method [DEEP1-2][DL2] (and then also by Amari's SGD for MLPs [GD1-2]). Minsky was apparently unaware of this and failed to correct it later [HIN](Sec. I).... deep learning research was a! > live and kicking also in the 1970s, especially outside of the Anglosphere." >> Just follow ACM's Code of Ethics and Professional Conduct [ACM18] which states: "Computing professionals should therefore credit the creators of ideas, inventions, work, and artifacts, and respect copyrights, patents, trade secrets, license agreements, and other methods of protecting authors' works." No need to wait for 100 years. >> >> J?rgen >> >> >> >> >> >>> On 2 Jan 2022, at 23:29, Terry Sejnowski wrote: >>> >>> We would be remiss not to acknowledge that backprop would not be possible without the calculus, >>> so Isaac newton should also have been given credit, at least as much credit as Gauss. >>> >>> All these threads will be sorted out by historians one hundred years from now. >>> Our precious time is better spent moving the field forward. There is much more to discover. >>> >>> A new generation with better computational and mathematical tools than we had back >>> in the last century have joined us, so let us be good role models and mentors to them. >>> >>> Terry > > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From jose at rubic.rutgers.edu Tue Feb 1 17:57:36 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Tue, 1 Feb 2022 17:57:36 -0500 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: References: <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> <6f02b384-6b29-431f-1235-babfcd6d8b16@ics.uci.edu> Message-ID: <762d59f0-dbe6-152d-bb7b-ba1b5dd38e62@rubic.rutgers.edu> On 2/1/22 3:19 PM, Asim Roy wrote: > > I had some communication with Marvin Minsky and Jerry Fodor years ago. > Here are two quotes from them: > > > 2. Jerry Fodor: ?Arguing with Connectionists is like arguing with > zombies; both are dead,?but ?neither has noticed it.? > I used to argue with Jerry at our Pizza lunchs, and miss him as he had great counter arguments? to almost anything....? but he would have hated the DL results with language...? GPT-xs --would have liked to know what his counters would have been.. but no, he didn't think all connectionists were zombies.. Sorry Asim. Steve > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > *From:* Connectionists > *On Behalf Of *Baldi,Pierre > *Sent:* Tuesday, February 1, 2022 7:41 AM > *To:* Schmidhuber Juergen ; connectionists at cs.cmu.edu > *Subject:* Re: Connectionists: Scientific Integrity, the 2021 Turing > Lecture, etc. > > M&P wrote: > > " The perceptron has shown itself worthy of study despite (and even > because of!) its severe limitations, It has many features to attract > attention: its linearity; its intriguing learning theorem; its clear > paradigmatic simplicity as a kind of parallel computation. There is no > reason to suppose that any of these virtues carry over to the > many-layered version.*Nevertheless, we consider it to be an important > research problem to elucidate (or reject) our intuitive judgment that > the extension is sterile. Perhaps some powerful convergence theorem > will be discovered, or some profound reason for the failure to produce > an interesting "learning theorem" for the multilayered machine will be > found." (pp231-232)* > > The claim that they killed NNs in the 1970s is indeed an exaggeration > propagated by the PDP group for obvious reasons. There was plenty of > NN research in the 1970s--Amari, Grossberg, etc.? Not surprisingly, > the same is true of the so-called "second neural network winter". > > On 2/1/2022 3:33 AM, Schmidhuber Juergen wrote: > > Thanks, Barak! Indeed, I should have said in the email msg that _others_ interpreted the book of Minsky & Papert [M69] in this way. My report explicitly mentions Terry [S20] who wrote in 2020: > > "The great expectations in the press (Fig. 3) were dashed by Minsky and Papert (7), who showed in their book Perceptrons that a perceptron can only represent categories that are linearly separable in weight space. Although at the end of their book Minsky and Papert considered the prospect of generalizing single- to multiple-layer perceptrons, one layer feeding into the next, they doubted there would ever be a way to train these more powerful multilayer perceptrons. Unfortunately, many took this doubt to be definitive, and the field was abandoned until a new generation of neural network researchers took a fresh look at the problem in the 1980s.? > > However, as mentioned above, the 1969 book [M69] addressed a "problem" of Gauss & Legendre's shallow learning (~1800) [DL1-2] that had already been solved 4 years prior by Ivakhnenko & Lapa's popular deep learning method [DEEP1-2][DL2] (and then also in 1967 by Amari's SGD for MLPs [GD1-2]). Deep learning research was not abandoned in the 1970s. It was alive and kicking, especially outside of the Anglosphere. [DEEP2][GD1-3][CNN1][DL1-2] > > See Sec. II and Sec. XIII of the report:https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html > > Cheers, > > J?rgen > > On 1 Feb 2022, at 14:17, Barak A. Pearlmutter wrote: > > J?rgen, > > It's fantastic that you're helping expose people to some important bits of scientific literature. > > But... > > Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions > > If you actually read Minsk and Papert's "Perceptrons" book, this is not a misconception it encourages. It defines a "k-th order perceptron" as a linear threshold unit preceded by an arbitrary set of fixed nonlinearities with fan-in k. (A linear threshold unit with binary inputs would, in this terminology, be a 1st-order perceptron.) All their theorems are for k>1. For instance, they prove that a k-th order perceptron cannot do (k+1)-bit parity, which in the special case of k=1 simplifies to the trivial observation that a simple linear threshold unit cannot do xor. > > > > This is why you're not supposed to directly cite things you have not actually read: it's too easy to misconstrue them based on inaccurate summaries transmitted over a series of biased noisy compressive channels. > > Cheers, > > --Barak. > > -- > Pierre Baldi, Ph.D. > Distinguished Professor, Department of Computer Science > Director, Institute for Genomics and Bioinformatics > Associate Director, Center for Machine Learning and Intelligent Systems > University of California, Irvine > Irvine, CA 92697-3435 > (949) 824-5809 > (949) 824-9813 [FAX] > Assistant: Janet Kojko at uci.edu -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From rloosemore at susaro.com Tue Feb 1 17:51:00 2022 From: rloosemore at susaro.com (Richard Loosemore) Date: Tue, 1 Feb 2022 17:51:00 -0500 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: References: Message-ID: <335F02E1-1B72-402C-AE9D-F24290922B6D@susaro.com> Sigh. So storm! Very teacup! Please, some of us were around in the 70s and 80s, and I gotta say that a certain amount of guff is being talked. I will confine myself to two points. 1) It was not ?PDP propaganda? that interest in neural nets was killed off by Minsky and Papert, it was already a widely acknowledged fact by the time I took a course on Neural Nets given by John G. Taylor (at Kings College London) in 1980/81. It doesn?t matter that M&P qualified their critique (everyone knew they said that) and it doesn?t matter that a few islands of NN did continue, what mattered was that there was HUGE excitement about the possibility of neurally inspired learning systems, in the wake of Rosenblatt, and then it died almost totally, with everyone citing M&P as the reason. Nobody gave a hoot that M&P themselves said that it shouldn?t be killed off). THAT?S WHAT SCIENCE DOES, IT BANDWAGONS AROUND MOST OF THE TIME. 2) Yes, you can trace the origins of various NN ideas back into the mists of Gaussian prehistory, and into the far corners of the scienceiverse, but scientists are and always have been self-serving, hypocritical bastards when it comes to attribution: they cram their references sections with whichever famous people?s papers happen to be the Soup du Jour, without EVER reading more than a tiny fraction of those papers. Entire oeuvres are quoted without anyone ever bothering to check if what is in them is diamond or dog poop, and nobody gives a damn if some little nobody writes something brilliant if they are not a Cool Kid. So, get over yourselves. Most of this debate is about self-preening, not giving credit where credit is due, and while there are some very valid points about the Gang of Four taking credit for what is actually a hundred people?s work, that message has gotten lost in a Boston Harbor so full of tea that there?s no water left in it anymore. Richard Loosemore Humble Nobody at Cornell University. From kaiolae at ifi.uio.no Wed Feb 2 03:53:35 2022 From: kaiolae at ifi.uio.no (Kai Olav Ellefsen) Date: Wed, 2 Feb 2022 08:53:35 +0000 Subject: Connectionists: PhD-position in adaptive robotics with learned behavior libraries at the University of Oslo Message-ID: <4806dcf18a8e4862ae95af047d4316d5@ifi.uio.no> At the University of Oslo, we are looking for a PhD candidate focusing on learning hierarchical libraries of behaviors for an assistive robot. The candidate would be part of the Robotics and Intelligent Systems (ROBIN) research group who has several related projects ongoing, that the PhD candidate could benefit from collaborating with. The full details are available here: https://www.jobbnorge.no/en/available-jobs/job/219688/phd-research-fellow-in-learned-robot-behavior-libraries -Kai Olav Ellefsen -------------- next part -------------- An HTML attachment was scrubbed... URL: From aihuborg at gmail.com Wed Feb 2 08:45:02 2022 From: aihuborg at gmail.com (AIhub) Date: Wed, 2 Feb 2022 13:45:02 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org ( https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI , NeurIPS , ICML , AIJ /IJCAI , ACM SIGAI , EurAI/AICOMM, CLAIRE and RoboCup . Twitter: @aihuborg -------------- next part -------------- An HTML attachment was scrubbed... URL: From coralie.gregoire at insa-lyon.fr Wed Feb 2 08:39:49 2022 From: coralie.gregoire at insa-lyon.fr (Coralie Gregoire) Date: Wed, 2 Feb 2022 14:39:49 +0100 (CET) Subject: Connectionists: [CFP EXTENDED DEADLINES] The ACM Web Conference 2022 - PhD Symposium - Web Developer and W3C Track Message-ID: <916242303.4598298.1643809189083.JavaMail.zimbra@insa-lyon.fr> ?[Apologies for the cross-posting, this call is sent to numerous lists you may have subscribed to] [CFP EXTENDED DEADLINES] The ACM Web Conference 2022 - PhD Symposium - Web Developer and W3C Track *Important dates* - NEW EXTENDED SUBMISSION DEADLINE: February 10, 2022 - Notification of acceptance: February 24, 2022 - Camera-ready version: March 10, 2022 We invite contributions to the PhD Symposium and Web Developer and W3C track to be held at The Web Conference 2022 (formerly known as WWW). The conference will take place online, hosted by Lyon, France, on April 25-29, 2022. ------------------------------------------------------------ Call for PhD Symposium Papers *Important dates* - NEW EXTENDED SUBMISSION DEADLINE: February 10, 2022 - Notification of acceptance: February 24, 2022 - Camera-ready version: March 10, 2022 - PhD Symposium: April 26, 2022 (to be confirmed) All submission deadlines are end-of-day in the Anywhere on Earth (AoE) time zone. *PhD Symposium chairs: (www2022-phd-symposium at easychair.org)* - Hala Skaf-Molli (University of Nantes, France) - Elena Demidova (University of Bonn, Germany) The PhD Symposium of The Web Conference 2022 welcomes submissions from PhD students on their ongoing research related to the main conference topics. These topics include: Semantics and Knowledge, Web Search, Web Systems and Infrastructure, Web Mining and Content Analysis, Economics, Monetization, and Online Markets, User Modeling and Personalization, Web and Society, Web of Things, Ubiquitous and Mobile Computing, Social Network Analysis and Graph Algorithms, Security, Privacy, Trust and Social Web (see also the Research Tracks). We are particularly interested in submissions aimed to enhance our understanding of the Web, provide intuitive access to Web information and knowledge, strengthen the positive impact of the Web on society, take advantage of the Web of Things and enhance security, privacy protection, and trust. The goal of the PhD Symposium is to provide a platform for PhD students to present and receive feedback on their ongoing research. Students at different stages of their research will have the opportunity to present and discuss their research questions, goals, methods, and results. The symposium aims to provide students guidance on various aspects of their research from established researchers and other PhD students working in research areas related to the World Wide Web. Finally, the symposium aims to enable PhD students to interact with other participants of The Web Conference and potential collaborators by stimulating the exchange of ideas and experiences. *Eligibility* The PhD Symposium is open to all PhD students. PhD students at the beginning stages of their doctoral work are particularly welcome when they have a well-defined problem statement and ideas about the solutions they would like to discuss. PhD students in a more advanced stage of their work are also welcome to share and discuss their research results and experiences. *Submission Guidelines* Submissions should be written based on the following structure, which focuses on the key methodological components required for a sound research synthesis: - Abstract: A self-sustained short description of the paper. - Introduction/Motivation: Provide a general introduction to the topic and indicate its importance/impact on Web research and real-world applications. - Problem: Describe the core problem of the PhD thesis. - State of the art: Briefly describe the most relevant related work. - Proposed approach: Briefly present the approach taken and motivate how this is novel regarding existing works. - Methodology: Sketch the methodology that is (or will be) adopted and, in particular, the approach to be taken for evaluating the results of the work. - Results: Describe the current status of the work and the most significant results that have been obtained so far. - Conclusions and future work: Conclude and specify the major items of future work. Submissions should be written in English and must be no longer than five (5) pages in length (according to the ACM format acmart.cls, using the ?sigconf? option). Submissions must be in PDF and must be made through the EasyChair system at https://easychair.org/conferences/?conf=thewebconf2022 (select the PhD Symposium track). Submissions must be single-author and be on the topic of the doctoral work. The supervisor?s name must be clearly marked (? supervised by ? ?) on the paper, under the author?s name. Submissions that do not comply with the formatting guidelines will be rejected without review. Selected papers will be published in the Companion Proceedings of The Web Conference 2022 and made available through the ACM Digital Library. *Review Process* All submissions will be reviewed by the members of the Program Committee of the PhD Symposium, who are experienced researchers in the relevant areas. Students of accepted submissions will have the opportunity to discuss their submissions in more detail and receive additional feedback from mentors. You can reach the PhD symposium chairs at www2022-phd-symposium at easychair.org ------------------------------------------------------------ Call for WebConf 2022 Developer and W3C Track *Important dates* - NEW EXTENDED SUBMISSION DEADLINE: February 10, 2022 - Notification to authors: March 3rd, 2022 - Camera ready: March 10th, 2022 All submission deadlines are end-of-day in the Anywhere on Earth (AoE) time zone. *Web Developer and W3C Track Chairs (www2022-web-dev-w3c at easychair.org)* - Dominique Hazael-Massieux (W3C) - Tom Steiner (Google LLC) The Web Conference 2022 Developer and W3C Track is part of The Web Conference 2022 in Lyon, France. Participation in the developers track will require registration of at least one author for the conference. The Web Conference 2022 Developer and W3C Track presents an opportunity to share the latest developments across the technical community, both in terms of technologies and in terms of tooling. We will be running as a regular track with live (preferred) or pre-recorded (as an alternative format) presentations during the conference, showcasing community expertise and progress. The event will take place according to the CET time zone (Paris time). We will do our best to accommodate speakers in as convenient as possible time slots for their local time zones. While we are open to any contributions that are relevant for the Web space, here are a few areas that we are particularly interested in: - Tools and methodologies to measure and reduce the environmental impact of the Web. - New usage patterns enabled by Progressive Web Apps and new browser APIs. - Work-arounds, data-based quantification and identification of Web compatibility issues. - Web tooling and developer experience, in particular towards reducing the complexity for newcomers: How can we get closer to the magic of hitting "view source" as a way to get people started? - Tools and frameworks that enable the convergence of real-time communication and streaming media. - Decentralized architectures for the Web, such as those emerging from projects and movements such as Solid, or "Web3". - Peer to peer architectures and protocols. - Identity management (DID, WebAuthN, Federated Credential Management). *Submission guidelines* Submissions can take several form and authors can choose one or multiple submission entry among the following choices: - Papers: papers are limited to 6 pages, including references. Submissions are NOT anonymous. It is the authors? responsibility to ensure that their submissions adhere strictly to the required format. In particular, the format cannot be modified with the objective of squeezing in more material. Papers will be published in the ACM The Web Conference Companion Proceedings archived by the ACM Digital Library, as open access, if the authors wish so. - Links to code repositories on GitHub (with sufficient description and documentation). - Links to recorded demos (as a complement to the above, ideally following established best practices as proposed by the W3C). - Any other resource reachable on the Web. Submissions will be handled via Easychair, at https://easychair.org/conferences/?conf=thewebconf2022, selecting the Web Developer and W3C track. *Formatting the submissions* Submissions must adhere to the ACM template and format published in the ACM guidelines at https://www.acm.org/publications/proceedings-template. Please remember to add Concepts and Keywords and use the template in traditional double-column format to prepare your submissions. For example, Word users may use the Word Interim template, and LaTeX users may use the sample-sigconf template. For Overleaf users, you may want to use https://www.overleaf.com/latex/templates/association-for-computing-machinery-acm-sig-proceedings-template/bmvfhcdnxfty. Submissions for review must be in PDF format. They must be self-contained and written in English. Submissions that do not follow these guidelines, or do not view or print properly, will be rejected without review. *Ethical use of data and informed consent* As a published ACM author, you and your co-authors are subject to all ACM Publications Policies, including ACM?s new Publications Policy on Research Involving Human Participants and Subjects. When appropriate, authors are encouraged to include a section on the ethical use of data and/or informed consent in their paper. Note that submitting your research for approval by the author(s)? institutional ethics review body (IRB) may not always be sufficient. Even if such research has been signed off by your IRB, the programme committee might raise additional concerns about the ethical implications of the work and include these concerns in its review. *Publication policy* Accepted papers will require a further revision in order to meet the requirements and page limits of the camera-ready format required by ACM. Instructions for the preparation of the camera-ready versions of the papers will be provided after acceptance. ============================================================ Contact us: contact at thewebconf.org - Facebook: https://www.facebook.com/TheWebConf - Twitter: https://twitter.com/TheWebConf - LinkedIn: https://www.linkedin.com/showcase/18819430/admin/ - Website: https://www2022.thewebconf.org/ ============================================== From marcello.pelillo at gmail.com Wed Feb 2 10:05:12 2022 From: marcello.pelillo at gmail.com (Marcello Pelillo) Date: Wed, 2 Feb 2022 16:05:12 +0100 Subject: Connectionists: Frontiers in Computer Vision - Call for contributions and Research Topics proposals Message-ID: The *Computer Vision* section of *Frontiers in Computer Science* welcomes original contributions in all relevant areas of computer vision, from both academia and industry https://www.frontiersin.org/journals/computer-science/sections/computer-vision Among the distinguishing features of Frontiers' open-access journals are fast publication time and an innovative collaborative peer-review process (see here for details). We also particularly welcome Research Topics proposals on cutting-edge themes: https://www.frontiersin.org/about/research-topics Here's a list of the most recent ones: - Novel Methods for Human Face and Body Perception and Generation - Segmentation and Classification: Theories, Algorithms, and Applications - Deep Learning with Few Labeled Training Data in Computer Vision and Image Analysis - Face Analysis in Computer Vision and Biometrics - Continual Unsupervised Learning in Computer Vision - Object Localization - Synchronization in Computer Vision - Sketching for Human Expressivity If you have any questions about the journal, feel free to contact me or the editorial office. Best regards -mp -- Marcello Pelillo, *FIEEE, FIAPR, FAAIA* Professor of Computer Science Ca' Foscari University of Venice, Italy IEEE SMC Distinguished Lecturer Specialty Chief Editor, *Computer Vision - Frontiers in Computer Science* -------------- next part -------------- An HTML attachment was scrubbed... URL: From dwang at cse.ohio-state.edu Wed Feb 2 10:54:07 2022 From: dwang at cse.ohio-state.edu (Wang, Deliang) Date: Wed, 2 Feb 2022 15:54:07 +0000 Subject: Connectionists: NEURAL NETWORKS, Feb. 2022 Message-ID: Neural Networks - Volume 146, February 2022 https://www.journals.elsevier.com/neural-networks Multiple-view flexible semi-supervised classification through consistent graph construction and label propagation Najmeh Ziraki, Fadi Dornaika, Alireza Bosaghzadeh Using top-down modulation to optimally balance shared versus separated task representations Pieter Verbeke, Tom Verguts Computational epidemiology study of homeostatic compensation during sensorimotor aging Niceto R. Luque, Francisco Naveros, Denis Sheynikhovich, Eduardo Ros, Angelo Arleo Predictive accuracy of CNN for cortical oscillatory activity in an acute rat model of parkinsonism Ali Abdul Nabi Ali, Mesbah Alam, Simon C. Klein, Nicolai Behmann, ... Kerstin Schwabe Multi-layer information fusion based on graph convolutional network for knowledge-driven herb recommendation Yun Yang, Yulong Rao, Minghao Yu, Yan Kang CSITime: Privacy-preserving human activity recognition using WiFi channel state information Santosh Kumar Yadav, Siva Sai, Akshay Gundewar, Heena Rathore, ... Mohit Mathur Imitation and mirror systems in robots through Deep Modality Blending Networks M. Yunus Seker, Alper Ahmetoglu, Yukie Nagai, Minoru Asada, ... Emre Ugur Multi-view Teacher-Student Network Yingjie Tian, Shiding Sun, Jingjing Tang Dilated projection correction network based on autoencoder for hyperspectral image super-resolution Xinya Wang, Jiayi Ma, Junjun Jiang, Xiao-Ping Zhang A hybridization of distributed policy and heuristic augmentation for improving federated learning approach Dawid Polap, Marcin Wozniak Feedforward neural networks initialization based on discriminant learning Kateryna Chumachenko, Alexandros Iosifidis, Moncef Gabbouj LOss-Based SensiTivity rEgulaRization: Towards deep sparse neural networks Enzo Tartaglione, Andrea Bragagnolo, Attilio Fiandrotti, Marco Grangetto Leveraging hierarchy in multimodal generative models for effective cross-modality inference Miguel Vasco, Hang Yin, Francisco S. Melo, Ana Paiva Efficient correntropy-based multi-view clustering with anchor graph embedding Ben Yang, Xuetao Zhang, Badong Chen, Feiping Nie, ... Zhixiong Nan Deep two-way matrix reordering for relational data analysis Chihiro Watanabe, Taiji Suzuki Deep semi-supervised learning via dynamic anchor graph embedding in latent space Enmei Tu, Zihao Wang, Jie Yang, Nikola Kasabov Functional connectivity inference from fMRI data using multivariate information measures Qiang Li An inertial neural network approach for robust time-of-arrival localization considering clock asynchronization Chentao Xu, Qingshan Liu Stability and dissipativity criteria for neural networks with time-varying delays via an augmented zero equality approach S.H. Lee, M.J. Park, D.H. Ji, O.M. Kwon Understanding and mitigating noise in trained deep neural networks Nadezhda Semenova, Laurent Larger, Daniel Brunner A second-order accelerated neurodynamic approach for distributed convex optimization Xinrui Jiang, Sitian Qin, Xiaoping Xue, Xinzhi Liu Fixed/Preassigned-time synchronization of quaternion-valued neural networks via pure power-law control Wanlu Wei, Juan Yu, Leimin Wang, Cheng Hu, Haijun Jiang ARCNN framework for multimodal infodemic detection Chahat Raj, Priyanka Meel Event-centric multi-modal fusion method for dense video captioning Zhi Chang, Dexin Zhao, Huilin Chen, Jingdan Li, Pengfei Liu Deep neural network enabled corrective source term approach to hybrid analysis and modeling Sindre Stenen Blakseth, Adil Rasheed, Trond Kvamsdal, Omer San NeuroLISP: High-level symbolic programming with attractor neural networks Gregory P. Davis, Garrett E. Katz, Rodolphe J. Gentili, James A. Reggia Transformers for modeling physical systems Nicholas Geneva, Nicholas Zabaras -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Wed Feb 2 12:51:16 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Wed, 2 Feb 2022 09:51:16 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf , also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > Stephen Hanson in conversation with Geoff Hinton > > In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. > > You can watch the discussion, and read the transcript, here: > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > About AIhub: > AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/ ). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI , NeurIPS , ICML , AIJ /IJCAI , ACM SIGAI , EurAI/AICOMM, CLAIRE and RoboCup . > Twitter: @aihuborg -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Wed Feb 2 21:52:30 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Wed, 2 Feb 2022 18:52:30 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Dear Geoff, Causality is often hard to establish, but I didn't start this thread; I merely responded to a false assertion of yours that was publicized at the top. More broadly, it's a shame for the field that you won't engage in the real issues at hand, even with a clear home-court advantage. Gary > > On Feb 2, 2022, at 12:52, Geoffrey Hinton wrote: > ? > You started this thread and it was a mistake for me to engage in arguing with you. I have said all I want to say. You have endless time for arguing and I don't. I find it more productive to spend time writing programs to see what works and what doesn't. You should try it sometime. > > Geoff > > > On Wed, Feb 2, 2022 at 3:25 PM Gary Marcus wrote: >> Dear Geoff, and interested others, >> >> What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: >> >> >> >> One could >> >> a. avert one?s eyes and deem the anomalous outputs irrelevant >> or >> b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. >> >> Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). >> >> So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. >> >> All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. >> >> With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) >> >> (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >> >> Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. >> >> Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? >> >> In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. >> >> Gary >> >> >>> On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: >>> ? >>> Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. >>> >>> A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. >> >>> But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? >>> >>> Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. >>> I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. >>> >>> Geoff >>> >>> >>> >>> >>> >>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, >>>> >>>> There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. >>>> >>>> The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. >>>> >>>> Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? >>>> >>>> I never said any such thing. >>>> >>>> What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. >>>> >>>> I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. >>>> >>>> I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: >>>> >>>> You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: >>>> >>>> Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. >>>> >>>> It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. >>>> >>>> Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. >>>> >>>> Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: >>>> >>>> The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. >>>> >>>> Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. >>>> >>>> Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). >>>> >>>> More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. >>>> >>>> As Herb Simon once observed, science does not have to be zero-sum. >>>> >>>> Sincerely, >>>> Gary Marcus >>>> Professor Emeritus >>>> New York University >>>> >>>>> On Feb 2, 2022, at 06:12, AIhub wrote: >>>>> ? >>>>> Stephen Hanson in conversation with Geoff Hinton >>>>> >>>>> In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. >>>>> >>>>> You can watch the discussion, and read the transcript, here: >>>>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>>>> >>>>> About AIhub: >>>>> AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. >>>>> Twitter: @aihuborg -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 319547 bytes Desc: not available URL: From gary.marcus at nyu.edu Wed Feb 2 15:25:34 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Wed, 2 Feb 2022 12:25:34 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387 ) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary > On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: > > ? > Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. > > A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. > But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? > > Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. > I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. > > Geoff > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, > > There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. > > The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? > > I never said any such thing. > > What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. > > I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. > > I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: > > You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: > > Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. > > It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. > > Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. > > Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf , also emphasizing issues of understanding and meaning: > > The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. > > Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. > > Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). > > More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. > > As Herb Simon once observed, science does not have to be zero-sum. > > Sincerely, > Gary Marcus > Professor Emeritus > New York University > >> On Feb 2, 2022, at 06:12, AIhub > wrote: >> >> ? >> Stephen Hanson in conversation with Geoff Hinton >> >> In the latest episode of this video series for AIhub.org , Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. >> >> You can watch the discussion, and read the transcript, here: >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> About AIhub: >> AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/ ). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI , NeurIPS , ICML , AIJ /IJCAI , ACM SIGAI , EurAI/AICOMM, CLAIRE and RoboCup . >> Twitter: @aihuborg -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 319547 bytes Desc: not available URL: From maanakg at gmail.com Wed Feb 2 18:14:24 2022 From: maanakg at gmail.com (Maanak Gupta) Date: Wed, 2 Feb 2022 17:14:24 -0600 Subject: Connectionists: Call for Papers: 27th ACM Symposium on Access Control Models and Technologies Message-ID: ACM SACMAT 2022 New York City, New York ----------------------------------------------- | Hybrid Conference (Online + In-person) | ----------------------------------------------- Call for Research Papers ============================================================== Papers offering novel research contributions are solicited for submission. Accepted papers will be presented at the symposium and published by the ACM in the symposium proceedings. In addition to the regular research track, this year SACMAT will again host the special track -- "Blue Sky/Vision Track". Researchers are invited to submit papers describing promising new ideas and challenges of interest to the community as well as access control needs emerging from other fields. We are particularly looking for potentially disruptive and new ideas which can shape the research agenda for the next 10 years. We also encourage submissions to the "Work-in-progress Track" to present ideas that may have not been completely developed and experimentally evaluated. Topics of Interest ============================================================== Submissions to the regular track covering any relevant area of access control are welcomed. Areas include, but are not limited to, the following: * Systems: * Operating systems * Cloud systems and their security * Distributed systems * Fog and Edge-computing systems * Cyber-physical and Embedded systems * Mobile systems * Autonomous systems (e.g., UAV security, autonomous vehicles, etc) * IoT systems (e.g., home-automation systems) * WWW * Design for resiliency * Designing systems with zero-trust architecture * Network: * Network systems (e.g., Software-defined network, Network function virtualization) * Corporate and Military-grade Networks * Wireless and Cellular Networks * Opportunistic Network (e.g., delay-tolerant network, P2P) * Overlay Network * Satellite Network * Privacy and Privacy-enhancing Technologies: * Mixers and Mixnets * Anonymous protocols (e.g., Tor) * Online social networks (OSN) * Anonymous communication and censorship resistance * Access control and identity management with privacy * Cryptographic tools for privacy * Data protection technologies * Attacks on Privacy and their defenses * Authentication: * Password-based Authentication * Biometric-based Authentication * Location-based Authentication * Identity management * Usable authentication * Mechanisms: * Blockchain Technologies * AI/ML Technologies * Cryptographic Technologies * Programming-language based Technologies * Hardware-security Technologies (e.g., Intel SGX, ARM TrustZone) * Economic models and game theory * Trust Management * Usable mechanisms * Data Security: * Big data * Databases and data management * Data leakage prevention * Data protection on untrusted infrastructure * Policies and Models: * Novel policy language design * New Access Control Models * Extension of policy languages * Extension of Models * Analysis of policy languages * Analysis of Models * Policy engineering and policy mining * Verification of policy languages * Efficient enforcement of policies * Usable access control policy New in ACM SACMAT 2022 ============================================================== We are moving ACM SACMAT 2022 to have two submission cycles. Authors submitting papers in the first submission cycle will have the opportunity to receive a major revision verdict in addition to the usual accept and reject verdicts. Authors can decide to prepare a revised version of the paper and submit it to the second submission cycle for consideration. Major revision papers will be reviewed by the program committee members based on the criteria set forward by them in the first submission cycle. Regular Track Paper Submission and Format ============================================================== Papers must be written in?English. Authors are required to use the ACM format for papers, using the two-column SIG Proceedings Template (the sigconf template for LaTex) available in the following link: https://www.acm.org/publications/authors/submissions The length of the paper in the proceedings format must not exceed?twelve?US letter pages formatted for 8.5" x 11" paper and be no more than 5MB in size. It is the responsibility of the authors to ensure that their submissions will print easily on simple default configurations. The submission must be anonymous, so information that might identify the authors - including author names, affiliations, acknowledgments, or obvious self-citations - must be excluded. It is the authors' responsibility to ensure that their anonymity is preserved when citing their work. Submissions should be made to the EasyChair conference management system by the paper submission deadline of: November 15th, 2021 (Submission Cycle 1) February 18th, 2022 (Submission Cycle 2) Submission Link: https://easychair.org/conferences/?conf=acmsacmat2022 All submissions must contain a significant original contribution. That is, submitted papers must not substantially overlap papers that have been published or that are simultaneously submitted to a journal, conference, or workshop. In particular, simultaneous submission of the same work is not allowed. Wherever appropriate, relevant related work, including that of the authors, must be cited. Submissions that are not accepted as full papers may be invited to appear as short papers. At least one author from each accepted paper must register for the conference before the camera-ready deadline. Blue Sky Track Paper Submission and Format ============================================================== All submissions to this track should be in the same format as for the regular track, but the length must not exceed ten US letter pages, and the submissions are not required to be anonymized (optional). Submissions to this track should be submitted to the EasyChair conference management system by the same deadline as for the regular track. Work-in-progress Track Paper Submission and Format ============================================================== Authors are invited to submit papers in the newly introduced work-in-progress track. This track is introduced for (junior) authors, ideally, Ph.D. and Master's students, to obtain early, constructive feedback on their work. Submissions in this track should follow the same format as for the regular track papers while limiting the total number of pages to six US letter pages. Paper submitted in this track should be anonymized and can be submitted to the EasyChair conference management system by the same deadline as for the regular track. Call for Lightning Talk ============================================================== Participants are invited to submit proposals for 5-minute lightning talks describing recently published results, work in progress, wild ideas, etc. Lightning talks are a new feature of SACMAT, introduced this year to partially replace the informal sharing of ideas at in-person meetings. Submissions are expected??by May 27, 2022. Notification of acceptance will be on June 3, 2022. Call for Posters ============================================================== SACMAT 2022 will include a poster session to promote discussion of ongoing projects among researchers in the field of access control and computer security. Posters can cover preliminary or exploratory work with interesting ideas, or research projects in the early stages with promising results in all aspects of access control and computer security. Authors interested in displaying a poster must submit a poster abstract in the same format as for the regular track, but the length must not exceed three US letter pages, and the submission should not be anonymized. The title should start with "Poster:". Accepted poster abstracts will be included in the conference proceedings. Submissions should be emailed to the poster chair by Apr 15th, 2022. The subject line should include "SACMAT 2022 Poster:" followed by the poster title. Call for Demos ============================================================== A demonstration proposal should clearly describe (1) the overall architecture of the system or technology to be demonstrated, and (2) one or more demonstration scenarios that describe how the audience, interacting with the demonstration system or the demonstrator, will gain an understanding of the underlying technology. Submissions will be evaluated based on the motivation of the work behind the use of the system or technology to be demonstrated and its novelty. The subject line should include "SACMAT 2022 Demo:" followed by the demo title. Demonstration proposals should be in the same format as for the regular track, but the length must not exceed four US letter pages, and the submission should not be anonymized. A two-page description of the demonstration will be included in the conference proceedings. Submissions should be emailed to the Demonstrations Chair by Apr 15th, 2022. Financial Conflict of Interest (COI) Disclosure: ============================================================== In the interests of transparency and to help readers form their own judgments of potential bias, ACM SACMAT requires authors and PC members to declare any competing financial and/or non-financial interests in relation to the work described. Definition ------------------------- For the purposes of this policy, competing interests are defined as financial and non-financial interests that could directly undermine, or be perceived to undermine the objectivity, integrity, and value of a publication, through a potential influence on the judgments and actions of authors with regard to objective data presentation, analysis, and interpretation. Financial competing interests include any of the following: Funding: Research support (including salaries, equipment, supplies, and other expenses) by organizations that may gain or lose financially through this publication. A specific role for the funding provider in the conceptualization, design, data collection, analysis, decision to publish, or preparation of the manuscript, should be disclosed. Employment: Recent (while engaged in the research project), present or anticipated employment by any organization that may gain or lose financially through this publication. Personal financial interests: Ownership or contractual interest in stocks or shares of companies that may gain or lose financially through publication; consultation fees or other forms of remuneration (including reimbursements for attending symposia) from organizations that may gain or lose financially; patents or patent applications (awarded or pending) filed by the authors or their institutions whose value may be affected by publication. For patents and patent applications, disclosure of the following information is requested: patent applicant (whether author or institution), name of the inventor(s), application number, the status of the application, specific aspect of manuscript covered in the patent application. It is difficult to specify a threshold at which a financial interest becomes significant, but note that many US universities require faculty members to disclose interests exceeding $10,000 or 5% equity in a company. Any such figure is necessarily arbitrary, so we offer as one possible practical alternative guideline: "Any undeclared competing financial interests that could embarrass you were they to become publicly known after your work was published." We do not consider diversified mutual funds or investment trusts to constitute a competing financial interest. Also, for employees in non-executive or leadership positions, we do not consider financial interest related to stocks or shares in their company to constitute a competing financial interest, as long as they are publishing under their company affiliation. Non-financial competing interests: Non-financial competing interests can take different forms, including personal or professional relations with organizations and individuals. We would encourage authors and PC members to declare any unpaid roles or relationships that might have a bearing on the publication process. Examples of non-financial competing interests include (but are not limited to): * Unpaid membership in a government or non-governmental organization * Unpaid membership in an advocacy or lobbying organization * Unpaid advisory position in a commercial organization * Writing or consulting for an educational company * Acting as an expert witness Conference Code of Conduct and Etiquette ============================================================== ACM SACMAT will follow the ACM Policy Against Harassment at ACM Activities. Please familiarize yourself with the ACM Policy Against Harassment (available at https://www.acm.org/special-interest-groups/volunteer-resources/officers-manual/ policy-against-discrimination-and-harassment) and guide to Reporting Unacceptable Behavior (available at https://www.acm.org/about-acm/reporting-unacceptable-behavior). AUTHORS TAKE NOTE ============================================================== The official publication date is the date the proceedings are made available in the ACM Digital Library. This date may be up to two weeks before the first day of your conference. The official publication date affects the deadline for any patent filings related to published work. (For those rare conferences whose proceedings are published in the ACM Digital Library after the conference is over, the official publication date remains the first day of the conference.) Important dates ============================================================== **Note that, these dates are currently only tentative and subject to change.** * Paper submission: November 15th, 2021 (Submission Cycle 1) February 18th, 2022 (Submission Cycle 2) * Rebuttal: December 16th - December 20th, 2021 (Submission Cycle 1) March 24th - March 28th, 2022 (Submission Cycle 2) * Notifications: January 14th, 2022 (Submission Cycle 1) April 8th, 2022 (Submission Cycle 2) * Systems demo and Poster submissions: April 15th, 2022 * Systems demo and Poster notifications: April 22nd, 2022 * Panel Proposal: March 18th, 2022 * Camera-ready paper submission: April 29th, 2022 * Conference date: June 8 - June 10, 2022 ReplyForward -------------- next part -------------- An HTML attachment was scrubbed... URL: From amir.kalfat at gmail.com Wed Feb 2 14:44:59 2022 From: amir.kalfat at gmail.com (Amir Aly) Date: Wed, 2 Feb 2022 19:44:59 +0000 Subject: Connectionists: CRNS Talk Series (2) - Live Talk by Dr. Rachid Alami - CNRS, France Message-ID: Dear All * Apologies for cross posting* The *Center for Robotics and Neural Systems* (CRNS) is pleased to announce the talk of *Dr. Rachid Alami* who is a research director and team leader at the *National Center for Scientific Research* (CNRS), France on Wednesday, February 9th from *11:00 am* to *12:30* *pm* (*London time*) over *Zoom*. >> *Events*: The CRNS talk series will cover a wide range of topics including social and cognitive robotics, computational neuroscience, computational linguistics, cognitive vision, machine learning and AI, and applications to autism. More details are available here: https://www.plymouth.ac.uk/research/robotics-neural-systems/whats-on >> *Link for the next event (No Registration is Required)*: Join Zoom Meeting https://plymouth.zoom.us/j/98559430652?pwd=aDBpNWtZd3FZKzNVQ0FMbndxWjdWdz09&from=addon >> *Title of the talk*: Models and Decisional issues for Human-Robot Joint Action *Abstract*: This talk will address some key decisional issues that are necessary for a cognitive and interactive robot which shares space and tasks with humans. We adopt a constructive approach based on the identification and the effective implementation of individual and collaborative skills. The system is comprehensive since it aims at dealing with a complete set of abilities articulated so that the robot controller is effectively able to conduct in a flexible and fluent manner a human-robot joint action seen as a collaborative problem solving and task achievement. These abilities include geometric reasoning and situation assessment based essentially on perspective-taking and affordances, management and exploitation of each agent (human and robot) knowledge in a separate cognitive model, human-aware task planning and interleaved execution of shared plans. We will also discuss the key issues linked to the pertinence and the acceptability by the human of the robot behaviour, and how this influence qualitatively the robot decisional, planning, control and communication processes. >> If you have any questions, please don't hesitate to contact me, Regards ---------------- *Dr. Amir Aly* Lecturer in Artificial Intelligence and Robotics Center for Robotics and Neural Systems (CRNS) School of Engineering, Computing, and Mathematics Room B332, Portland Square, Drake Circus, PL4 8AA University of Plymouth, UK -------------- next part -------------- An HTML attachment was scrubbed... URL: From g.goodhill at wustl.edu Wed Feb 2 17:15:37 2022 From: g.goodhill at wustl.edu (Goodhill, Geoffrey) Date: Wed, 2 Feb 2022 22:15:37 +0000 Subject: Connectionists: Postdoc position in the development of light field imaging technologies for recording neural activity in freely-behaving animals Message-ID: <40EE4DF5-69F4-47E7-AEAA-ECE13F6AE0B6@wustl.edu> The Goodhill lab at Washington University in St Louis is looking for a postdoc scientist for an NIH-funded project to develop novel microscopy methods for whole-brain calcium imaging in freely-behaving zebrafish. This is a collaboration with Oliver Cossairt at Northwestern University. The lab focuses on the computational principles underlying the development of neural circuits and behavior. The goal of this project is to construct light-field imaging techniques to record the activity of ~100,000 neurons in the larval zebrafish brain during unconstrained hunting behavior (for more details see https://reporter.nih.gov/search/q1hIGuT01Em96OpHwTdvDg/project-details/10300922). We are looking for someone with experience in building new optical imaging technologies to work with others in the lab from a diverse array of backgrounds including biology, mathematics, physics and engineering. Washington University in St Louis is ranked in the top 10 institutions globally for Neuroscience and Behavior, and offers an outstanding intellectual environment for neuroscience research. In 2023 the Goodhill lab will move along with over 100 other labs into Washington University's new start-of-the-art 600,000 sq ft Neuroscience Research Building (https://neuroscience.wustl.edu/research/neuroscience-building). For more information about St Louis see https://explorestlouis.com. To apply please send a detailed CV and cover letter explaining your interest to g.goodhill at wustl.edu. Review of applications will continue until the position is filled. Professor Geoffrey J Goodhill Departments of Developmental Biology, Neuroscience, Biomedical Engineering, and Electrical and Systems Engineering https://neuroscience.wustl.edu/people/geoffrey-goodhill-phd Email: g.goodhill at wustl.edu -------------- next part -------------- An HTML attachment was scrubbed... URL: From geoffrey.hinton at gmail.com Wed Feb 2 15:52:05 2022 From: geoffrey.hinton at gmail.com (Geoffrey Hinton) Date: Wed, 2 Feb 2022 15:52:05 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> Message-ID: You started this thread and it was a mistake for me to engage in arguing with you. I have said all I want to say. You have endless time for arguing and I don't. I find it more productive to spend time writing programs to see what works and what doesn't. You should try it sometime. Geoff On Wed, Feb 2, 2022 at 3:25 PM Gary Marcus wrote: > Dear Geoff, and interested others, > > What, for example, would you make of a system that often drew the > red-hatted hamster you requested, and perhaps a fifth of the time gave you > utter nonsense? Or say one that you trained to create birds but sometimes > output stuff like this: > > > One could > > a. avert one?s eyes and deem the anomalous outputs irrelevant > or > b. wonder if it might be possible that sometimes the system gets the right > answer for the wrong reasons (eg partial historical contingency), and > wonder whether another approach might be indicated. > > Benchmarks are harder than they look; most of the field has come to > recognize that. The Turing Test has turned out to be a lousy measure of > intelligence, easily gamed. It has turned out empirically that the Winograd > Schema Challenge did not measure common sense as well as Hector might have > thought. (As it happens, I am a minor coauthor of a very recent review on > this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no > way means machines now have common sense; many people from many different > perspectives recognize that (including, e.g., Yann LeCun, who generally > tends to be more aligned with you than with me). > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote > me; but what you said about me and machine translation remains your > invention, and it is inexcusable that you simply ignored my 2019 > clarification. On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains unsolved. > > All of the problems LLMs have with coherence, reliability, truthfulness, > misinformation, etc stand witness to that fact. (Their persistent inability > to filter out toxic and insulting remarks stems from the same.) I am hardly > the only person in the field to see that progress on any given benchmark > does not inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > With respect to embeddings: Embeddings are very good for natural language > *processing*; but NLP is not the same as NL*U* ? when it comes to > *understanding*, their worth is still an open question. Perhaps they will > turn out to be necessary; they clearly aren?t sufficient. In their extreme, > they might even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a specific set of > numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > Notably absent from your email is any kind of apology for misrepresenting > my position. It?s fine to say that ?many people thirty years ago once > thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. > I have consistently felt throughout our interactions that you have mistaken > me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me > for having made that error. I am still not he. > > Which maybe connects to the last point; if you read my work, you would see > thirty years of arguments *for* neural networks, just not in the way that > you want them to exist. I have ALWAYS argued that there is a role for them; > characterizing me as a person ?strongly opposed to neural networks? misses > the whole point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > In the last two decades or so you have insisted (for reasons you have > never fully clarified, so far as I know) on abandoning symbol-manipulation, > but the reverse is not the case: I have *never* called for dismissal of > neural networks, but rather for some hybrid between the two (as you > yourself contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded and broke down, > and where symbols could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > Gary > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: > > ? > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > > > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > Geoff > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > >> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural >> network community, >> >> There has been a lot of recent discussion on this list about framing and >> scientific integrity. Often the first step in restructuring narratives is >> to bully and dehumanize critics. The second is to misrepresent their >> position. People in positions of power are sometimes tempted to do this. >> >> The Hinton-Hanson interview that you just published is a real-time >> example of just that. It opens with a needless and largely content-free >> personal attack on a single scholar (me), with the explicit intention of >> discrediting that person. Worse, the only substantive thing it says is >> false. >> >> Hinton says ?In 2015 he [Marcus] made a prediction that computers >> wouldn?t be able to do machine translation.? >> >> I never said any such thing. >> >> What I predicted, rather, was that multilayer perceptrons, as they >> existed then, would not (on their own, absent other mechanisms) >> understand language. Seven years later, they still haven?t, except in >> the most superficial way. >> >> I made no comment whatsoever about machine translation, which I view as a >> separate problem, solvable to a certain degree by correspondance without >> semantics. >> >> I specifically tried to clarify Hinton?s confusion in 2019, but, >> disappointingly, he has continued to purvey misinformation despite that >> clarification. Here is what I wrote privately to him then, which should >> have put the matter to rest: >> >> You have taken a single out of context quote [from 2015] and >> misrepresented it. The quote, which you have prominently displayed at the >> bottom on your own web page, says: >> >> Hierarchies of features are less suited to challenges such as language, >> inference, and high-level planning. For example, as Noam Chomsky famously >> pointed out, language is filled with sentences you haven't seen >> before. Pure classifier systems don't know what to do with such sentences. >> The talent of feature detectors -- in identifying which member of some >> category something belongs to -- doesn't translate into understanding >> novel sentences, in which each sentence has its own unique meaning. >> >> It does not say "neural nets would not be able to deal with novel >> sentences"; it says that hierachies of features detectors (on their own, if >> you read the context of the essay) would have trouble understanding novel sentences. >> >> >> Google Translate does yet not understand the content of the sentences is >> translates. It cannot reliably answer questions about who did what to whom, >> or why, it cannot infer the order of the events in paragraphs, it can't >> determine the internal consistency of those events, and so forth. >> >> Since then, a number of scholars, such as the the computational linguist >> Emily Bender, have made similar points, and indeed current LLM difficulties >> with misinformation, incoherence and fabrication all follow from these >> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf >> , >> also emphasizing issues of understanding and meaning: >> >> The success of the large neural language models on many NLP tasks is >> exciting. However, we find that these successes sometimes lead to hype in >> which these models are being described as ?understanding? language or >> capturing ?meaning?. In this position paper, we argue that a system trained >> only on form has a priori no way to learn meaning. .. a clear understanding >> of the distinction between form and meaning will help guide the field >> towards better science around natural language understanding. >> >> Her later article with Gebru on language models ?stochastic parrots? is >> in some ways an extension of this point; machine translation requires >> mimicry, true understanding (which is what I was discussing in 2015) >> requires something deeper than that. >> >> Hinton?s intellectual error here is in equating machine translation with >> the deeper comprehension that robust natural language understanding will >> require; as Bender and Koller observed, the two appear not to be the same. >> (There is a longer discussion of the relation between language >> understanding and machine translation, and why the latter has turned out to >> be more approachable than the former, in my 2019 book with Ernest Davis). >> >> More broadly, Hinton?s ongoing dismissiveness of research from >> perspectives other than his own (e.g. linguistics) have done the field a >> disservice. >> >> As Herb Simon once observed, science does not have to be zero-sum. >> >> Sincerely, >> Gary Marcus >> Professor Emeritus >> New York University >> >> On Feb 2, 2022, at 06:12, AIhub wrote: >> >> ? >> Stephen Hanson in conversation with Geoff Hinton >> >> In the latest episode of this video series for AIhub.org >> , >> Stephen Hanson talks to Geoff Hinton about neural networks, >> backpropagation, overparameterization, digit recognition, voxel cells, >> syntax and semantics, Winograd sentences, and more. >> >> You can watch the discussion, and read the transcript, here: >> >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> >> About AIhub: >> AIhub is a non-profit dedicated to connecting the AI community to the >> public by providing free, high-quality information through AIhub.org >> >> (https://aihub.org/ >> ). >> We help researchers publish the latest AI news, summaries of their work, >> opinion pieces, tutorials and more. We are supported by many leading >> scientific organizations in AI, namely AAAI >> , >> NeurIPS >> , >> ICML >> , >> AIJ >> >> /IJCAI >> , >> ACM SIGAI >> , >> EurAI/AICOMM, CLAIRE >> >> and RoboCup >> >> . >> Twitter: @aihuborg >> >> -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 319547 bytes Desc: not available URL: From geoffrey.hinton at gmail.com Wed Feb 2 14:22:28 2022 From: geoffrey.hinton at gmail.com (Geoffrey Hinton) Date: Wed, 2 Feb 2022 14:22:28 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural > network community, > > There has been a lot of recent discussion on this list about framing and > scientific integrity. Often the first step in restructuring narratives is > to bully and dehumanize critics. The second is to misrepresent their > position. People in positions of power are sometimes tempted to do this. > > The Hinton-Hanson interview that you just published is a real-time example > of just that. It opens with a needless and largely content-free personal > attack on a single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive thing it says is > false. > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t > be able to do machine translation.? > > I never said any such thing. > > What I predicted, rather, was that multilayer perceptrons, as they existed > then, would not (on their own, absent other mechanisms) understand language. > Seven years later, they still haven?t, except in the most superficial way. > > > I made no comment whatsoever about machine translation, which I view as a > separate problem, solvable to a certain degree by correspondance without > semantics. > > I specifically tried to clarify Hinton?s confusion in 2019, but, > disappointingly, he has continued to purvey misinformation despite that > clarification. Here is what I wrote privately to him then, which should > have put the matter to rest: > > You have taken a single out of context quote [from 2015] and > misrepresented it. The quote, which you have prominently displayed at the > bottom on your own web page, says: > > Hierarchies of features are less suited to challenges such as language, > inference, and high-level planning. For example, as Noam Chomsky famously > pointed out, language is filled with sentences you haven't seen > before. Pure classifier systems don't know what to do with such sentences. > The talent of feature detectors -- in identifying which member of some > category something belongs to -- doesn't translate into understanding > novel sentences, in which each sentence has its own unique meaning. > > It does not say "neural nets would not be able to deal with novel > sentences"; it says that hierachies of features detectors (on their own, if > you read the context of the essay) would have trouble understanding novel sentences. > > > Google Translate does yet not understand the content of the sentences is > translates. It cannot reliably answer questions about who did what to whom, > or why, it cannot infer the order of the events in paragraphs, it can't > determine the internal consistency of those events, and so forth. > > Since then, a number of scholars, such as the the computational linguist > Emily Bender, have made similar points, and indeed current LLM difficulties > with misinformation, incoherence and fabrication all follow from these > concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter > with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, > also emphasizing issues of understanding and meaning: > > The success of the large neural language models on many NLP tasks is > exciting. However, we find that these successes sometimes lead to hype in > which these models are being described as ?understanding? language or > capturing ?meaning?. In this position paper, we argue that a system trained > only on form has a priori no way to learn meaning. .. a clear understanding > of the distinction between form and meaning will help guide the field > towards better science around natural language understanding. > > Her later article with Gebru on language models ?stochastic parrots? is in > some ways an extension of this point; machine translation requires mimicry, > true understanding (which is what I was discussing in 2015) requires > something deeper than that. > > Hinton?s intellectual error here is in equating machine translation with > the deeper comprehension that robust natural language understanding will > require; as Bender and Koller observed, the two appear not to be the same. > (There is a longer discussion of the relation between language > understanding and machine translation, and why the latter has turned out to > be more approachable than the former, in my 2019 book with Ernest Davis). > > More broadly, Hinton?s ongoing dismissiveness of research from > perspectives other than his own (e.g. linguistics) have done the field a > disservice. > > As Herb Simon once observed, science does not have to be zero-sum. > > Sincerely, > Gary Marcus > Professor Emeritus > New York University > > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > Stephen Hanson in conversation with Geoff Hinton > > In the latest episode of this video series for AIhub.org, Stephen Hanson > talks to Geoff Hinton about neural networks, backpropagation, > overparameterization, digit recognition, voxel cells, syntax and semantics, > Winograd sentences, and more. > > You can watch the discussion, and read the transcript, here: > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > About AIhub: > AIhub is a non-profit dedicated to connecting the AI community to the > public by providing free, high-quality information through AIhub.org ( > https://aihub.org/ > ). > We help researchers publish the latest AI news, summaries of their work, > opinion pieces, tutorials and more. We are supported by many leading > scientific organizations in AI, namely AAAI > , > NeurIPS > , > ICML > , > AIJ > > /IJCAI > , > ACM SIGAI > , > EurAI/AICOMM, CLAIRE > > and RoboCup > > . > Twitter: @aihuborg > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Thu Feb 3 02:52:21 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Thu, 3 Feb 2022 07:52:21 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> Message-ID: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: [cid:image001.png at 01D81897.B1E39500] One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub > wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 319547 bytes Desc: image001.png URL: From danko.nikolic at gmail.com Thu Feb 3 06:19:53 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Thu, 3 Feb 2022 12:19:53 +0100 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> Message-ID: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- Virus-free. www.avast.com <#DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > Without getting into the specific dispute between Gary and Geoff, I think > with approaches similar to GLOM, we are finally headed in the right > direction. There?s plenty of neurophysiological evidence for single-cell > abstractions and multisensory neurons in the brain, which one might claim > correspond to symbols. And I think we can finally reconcile the decades old > dispute between Symbolic AI and Connectionism. > > > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > GARY: I have *never* called for dismissal of neural networks, but rather > for some hybrid between the two (as you yourself contemplated in 1991); the > point of the 2001 book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols could complement > them. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > *From:* Connectionists *On > Behalf Of *Gary Marcus > *Sent:* Wednesday, February 2, 2022 1:26 PM > *To:* Geoffrey Hinton > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Geoff, and interested others, > > > > What, for example, would you make of a system that often drew the > red-hatted hamster you requested, and perhaps a fifth of the time gave you > utter nonsense? Or say one that you trained to create birds but sometimes > output stuff like this: > > > > > > One could > > > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > or > > b. wonder if it might be possible that sometimes the system gets the right > answer for the wrong reasons (eg partial historical contingency), and > wonder whether another approach might be indicated. > > > > Benchmarks are harder than they look; most of the field has come to > recognize that. The Turing Test has turned out to be a lousy measure of > intelligence, easily gamed. It has turned out empirically that the Winograd > Schema Challenge did not measure common sense as well as Hector might have > thought. (As it happens, I am a minor coauthor of a very recent review on > this very topic: https://arxiv.org/abs/2201.02387 > ) > But its conquest in no way means machines now have common sense; many > people from many different perspectives recognize that (including, e.g., > Yann LeCun, who generally tends to be more aligned with you than with me). > > > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote > me; but what you said about me and machine translation remains your > invention, and it is inexcusable that you simply ignored my 2019 > clarification. On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains unsolved. > > > > All of the problems LLMs have with coherence, reliability, truthfulness, > misinformation, etc stand witness to that fact. (Their persistent inability > to filter out toxic and insulting remarks stems from the same.) I am hardly > the only person in the field to see that progress on any given benchmark > does not inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > > > With respect to embeddings: Embeddings are very good for natural language > *processing*; but NLP is not the same as NL*U* ? when it comes to > *understanding*, their worth is still an open question. Perhaps they will > turn out to be necessary; they clearly aren?t sufficient. In their extreme, > they might even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a specific set of > numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > > > (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > > Notably absent from your email is any kind of apology for misrepresenting > my position. It?s fine to say that ?many people thirty years ago once > thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. > I have consistently felt throughout our interactions that you have mistaken > me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me > for having made that error. I am still not he. > > > > Which maybe connects to the last point; if you read my work, you would see > thirty years of arguments *for* neural networks, just not in the way that > you want them to exist. I have ALWAYS argued that there is a role for them; > characterizing me as a person ?strongly opposed to neural networks? misses > the whole point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > > > In the last two decades or so you have insisted (for reasons you have > never fully clarified, so far as I know) on abandoning symbol-manipulation, > but the reverse is not the case: I have *never* called for dismissal of > neural networks, but rather for some hybrid between the two (as you > yourself contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded and broke down, > and where symbols could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > > > Gary > > > > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: > > ? > > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > > > > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > > > Geoff > > > > > > > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural > network community, > > > > There has been a lot of recent discussion on this list about framing and > scientific integrity. Often the first step in restructuring narratives is > to bully and dehumanize critics. The second is to misrepresent their > position. People in positions of power are sometimes tempted to do this. > > > > The Hinton-Hanson interview that you just published is a real-time example > of just that. It opens with a needless and largely content-free personal > attack on a single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive thing it says is > false. > > > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t > be able to do machine translation.? > > > > I never said any such thing. > > > > What I predicted, rather, was that multilayer perceptrons, as they existed > then, would not (on their own, absent other mechanisms) *understand* language. > Seven years later, they still haven?t, except in the most superficial way. > > > > > I made no comment whatsoever about machine translation, which I view as a > separate problem, solvable to a certain degree by correspondance without > semantics. > > > > I specifically tried to clarify Hinton?s confusion in 2019, but, > disappointingly, he has continued to purvey misinformation despite that > clarification. Here is what I wrote privately to him then, which should > have put the matter to rest: > > > > You have taken a single out of context quote [from 2015] and > misrepresented it. The quote, which you have prominently displayed at the > bottom on your own web page, says: > > > > Hierarchies of features are less suited to challenges such as language, > inference, and high-level planning. For example, as Noam Chomsky famously > pointed out, language is filled with sentences you haven't seen > before. Pure classifier systems don't know what to do with such sentences. > The talent of feature detectors -- in identifying which member of some > category something belongs to -- doesn't translate into understanding > novel sentences, in which each sentence has its own unique meaning. > > > > It does *not* say "neural nets would not be able to deal with novel > sentences"; it says that hierachies of features detectors (on their own, if > you read the context of the essay) would have trouble *understanding *novel sentences. > > > > > Google Translate does yet not *understand* the content of the sentences > is translates. It cannot reliably answer questions about who did what to > whom, or why, it cannot infer the order of the events in paragraphs, it > can't determine the internal consistency of those events, and so forth. > > > > Since then, a number of scholars, such as the the computational linguist > Emily Bender, have made similar points, and indeed current LLM difficulties > with misinformation, incoherence and fabrication all follow from these > concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter > with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf > , > also emphasizing issues of understanding and meaning: > > > > *The success of the large neural language models on many NLP tasks is > exciting. However, we find that these successes sometimes lead to hype in > which these models are being described as ?understanding? language or > capturing ?meaning?. In this position paper, we argue that a system trained > only on form has a priori no way to learn meaning. .. a clear understanding > of the distinction between form and meaning will help guide the field > towards better science around natural language understanding. * > > > > Her later article with Gebru on language models ?stochastic parrots? is in > some ways an extension of this point; machine translation requires mimicry, > true understanding (which is what I was discussing in 2015) requires > something deeper than that. > > > > Hinton?s intellectual error here is in equating machine translation with > the deeper comprehension that robust natural language understanding will > require; as Bender and Koller observed, the two appear not to be the same. > (There is a longer discussion of the relation between language > understanding and machine translation, and why the latter has turned out to > be more approachable than the former, in my 2019 book with Ernest Davis). > > > > More broadly, Hinton?s ongoing dismissiveness of research from > perspectives other than his own (e.g. linguistics) have done the field a > disservice. > > > > As Herb Simon once observed, science does not have to be zero-sum. > > > > Sincerely, > > Gary Marcus > > Professor Emeritus > > New York University > > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > > Stephen Hanson in conversation with Geoff Hinton > > > > In the latest episode of this video series for AIhub.org > , > Stephen Hanson talks to Geoff Hinton about neural networks, > backpropagation, overparameterization, digit recognition, voxel cells, > syntax and semantics, Winograd sentences, and more. > > > > You can watch the discussion, and read the transcript, here: > > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > > > About AIhub: > > AIhub is a non-profit dedicated to connecting the AI community to the > public by providing free, high-quality information through AIhub.org > > (https://aihub.org/ > ). > We help researchers publish the latest AI news, summaries of their work, > opinion pieces, tutorials and more. We are supported by many leading > scientific organizations in AI, namely AAAI > , > NeurIPS > , > ICML > , > AIJ > > /IJCAI > , > ACM SIGAI > , > EurAI/AICOMM, CLAIRE > > and RoboCup > > . > > Twitter: @aihuborg > > Virus-free. www.avast.com <#DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 319547 bytes Desc: not available URL: From ioannakoroni at csd.auth.gr Thu Feb 3 07:09:12 2022 From: ioannakoroni at csd.auth.gr (Ioanna Koroni) Date: Thu, 3 Feb 2022 14:09:12 +0200 Subject: Connectionists: Asynchronous Web e-Courses on Computer Vision on offer. Free access to course material References: <008601d81847$04daea70$0e90bf50$@csd.auth.gr> Message-ID: <0a5801d818f6$dbc26040$934720c0$@csd.auth.gr> Dear Computer Vision, Machine Learning, Image Processing Engineers, Scientists and Enthusiasts, you are welcomed to register and attend the Web e-Course on Computer Vision consisting of the following two CVML Web e-Course Modules on offer (total 20 lectures): Computer Vision (12 Lectures), http://icarus.csd.auth.gr/computer-vision-web-module/ 1. Introduction to Computer Vision 2. Digital Images and Videos 3. Image acquisition 4. Camera geometry 5. Stereo and Multiview Imaging 6. Neural Semantic 3D World Modeling and Mapping 7. Structure from Motion 8. Simultaneous Localization and Mapping 9. Neural SLAM 10. 3D Object Localization 11. Object Pose Estimation 12. Computational Cinematography 2D Computer Vision/Image Analysis (8 Lectures), http://icarus.csd.auth.gr/2d-computer-vision-and-image-analysis-web-module/ 1. Introduction to 2D Computer Vision 2. Edge Detection 3. Region Segmentation 4. Image Features 5. Image Registration 6. Shape Description 7. Mathematical Morphology 8. Computational Geometry Around 50% of the lectures provide free access to the full lecture pdf! You can find sample Web e-Course Module material to make up your mind and/or can perform CVML Web e-Course registration in: http://icarus.csd.auth.gr/cvml-web-lecture-series/ For questions, please contact: Ioanna Koroni > More information on this Web e-course: This Web e-Course Computer Vision material that can cover a semester course, but you can master it in approximately 1 month. Course materials are at senior undergraduate/MSc level in a CS, CSE, EE or ECE or related Engineering or Science Department. Their structure, level and offer are completely different from what you can find in either Coursera or Udemy. CVML Web e-Course Module materials typically consist of: a) a lecture pdf/ppt, b) lecture self-assessment understanding questionnaire and lecture video, programming exercises, tutorial exercises (for several modules/lectures) and overall course module satisfaction questionnaire. Asynchronous tutor support will be provided in case of questions. Course materials have been very successfully used in many top conference keynote speeches/tutorials worldwide and in short courses, summer schools, semester courses delivered by AIIA Lab physically or on-line from 2018 onwards, attracting many hundreds of registrants. More information on other CVML Web e-course: Several other Web e-Course Modules are offered on Deep Learning, Computer Vision, Autonomous Systems, Signal/Image/Video Processing, Human-centered Computing, Social Media, Mathematical Foundations, CVML SW tools. See: http://icarus.csd.auth.gr/cvml-web-lecture-series/ You can combine CVML Web e-Course Modules to create CVML Web e-Courses (typically consisting of 16 lectures) of your own choice that cater your personal education needs. Each CVML Web e-Course you will create (typically 16 lectures) provides you material that can cover a semester course, but you can master it in approximately 1 month. Academic/Research/Industry offer and arrangements Special arrangements can be made to offer the material of these CVML Web e-Course Modules at University/Department/Company level: * by granting access to the material to University/research/industry lecturers to be used as an aid in their teaching, * by enabling class registration in CVML Web e-Courses * by delivering such live short courses physically or on-line by Prof. Ioannis Pitas * by combinations of the above. The CVML Web e-Course is organized by Prof. I. Pitas, IEEE and EURASIP fellow, Coordinator of International AI Doctoral Academy (AIDA), past Chair of the IEEE SPS Autonomous Systems Initiative, Director of the Artificial Intelligence and Information analysis Lab (AIIA Lab), Aristotle University of Thessaloniki, Greece, Coordinator of the European Horizon2020 R&D project Multidrone. He is ranked 249-top Computer Science and Electronics scientist internationally by Guide2research (2018). He has 34100+ citations to his work and h-index 87+. The informatics Department at AUTH ranked 106th internationally in the field of Computer Science for 2019 in the Leiden Ranking list Relevant links: 1. Prof. I. Pitas: https://scholar.google.gr/citations?user=lWmGADwAAAAJ &hl=el 2. International AI Doctoral Academy (AIDA): https://www.i-aida.org/ 3. Horizon2020 EU funded R&D project Aerial-Core: https://aerial-core.eu/ 4. Horizon2020 EU funded R&D project Multidrone: https://multidrone.eu/ 5. Horizon2020 EU funded R&D project AI4Media: https://ai4media.eu/ 6. AIIA Lab: https://aiia.csd.auth.gr/ Sincerely yours Prof. I. Pitas Director of the Artificial Intelligence and Information analysis Lab (AIIA Lab) Aristotle University of Thessaloniki, Greece Post scriptum: To stay current on CVML matters, you may want to register to the CVML email list, following instructions in https://lists.auth.gr/sympa/info/cvml -- This email has been checked for viruses by Avast antivirus software. https://www.avast.com/antivirus -------------- next part -------------- An HTML attachment was scrubbed... URL: From Ulrich.Bodenhofer at fh-hagenberg.at Thu Feb 3 07:05:10 2022 From: Ulrich.Bodenhofer at fh-hagenberg.at (Bodenhofer Ulrich) Date: Thu, 3 Feb 2022 12:05:10 +0000 Subject: Connectionists: 5 PhD Positions in Human-Centered Artificial Intelligence (Hagenberg and Linz, Austria) Message-ID: <83fd623328714ec68c73f57f759642a1@fhooembox1.fhooe.at> Job Announcement for 5 Fully-Funded Positions as PhD Student Researchers in Austria We are very happy to announce an exciting new research and training program for PhD students: Human-Centered Artificial Intelligence (HCAI). HCAI will start on April 1, 2022. The HCAI program will fund a total of five PhD student researchers for up to 4 years. Employment will be for 30 hours/week, according to the regulations of the funding body, i.e., the Austrian Science Fund (FWF). To fill these open positions, we are seeking highly motivated students with a background in Artificial Intelligence, Human-computer Interaction, Machine Learning, Information Visualization, Information Retrieval, and/or Recommender Systems, holding a Master's degree (or equivalent). The students will be employed either at the Johannes Kepler University (JKU) Linz or the University of Applied Sciences, Upper Austria (FH O?), Hagenberg Campus. Selected students will be given the unique opportunity to work on exciting and timely topics, at the intersection of basic and applied research, thanks to the close collaboration between the universities. For more information, please visit: https://bit.ly/3Gf6t2e From jose at rubic.rutgers.edu Thu Feb 3 08:10:25 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Thu, 3 Feb 2022 08:10:25 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <9b988b04-250b-56c8-c3ec-9509bba4bdc8@rubic.rutgers.edu> Since AIHUB was the one who posted the recent discussion between myself and Geoff, I suppose that could be the proximate causal event. But in looking over the transcript, I think you are overreacting to the single en passant comment in an hour about you.?? Hardly, a? "personal attack". Moreover, I've known you for over 30 years, and frankly it does sound like something you might have said );-) ? And I agree your positions in various books and articles have evolved over time.? ? But I and AIHUB are always happy to correct transcripts that have seriously miss-represented someone's claims. I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. Its important for those paying attention to this thread, to realize these are still very early times.? ? Many more shoes will drop in the next few years.? I for one don't believe one of those shoes will be Hybrid approaches to AI,? I've seen that movie before and it didn't end well. Best and hope you are doing well. Steve On 2/2/22 9:52 PM, Gary Marcus wrote: > Dear Geoff, > > Causality is often hard to establish, but I didn't start this thread; > I merely responded to a false assertion of yours that was publicized > at the top. > > More broadly, it's a shame for the field that you won't engage in the > real issues at hand, even with a clear home-court advantage. > > Gary > >> >> On Feb 2, 2022, at 12:52, Geoffrey Hinton >> wrote: >> >> ? >> You started this thread and it was a mistake for me to engage in >> arguing with you. I have said all I want to say.? You have endless >> time for arguing and I don't. I find it more productive to spend time >> writing programs to see what works and what doesn't. You should try >> it sometime. >> >> Geoff >> >> >> -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From krallinger.martin at gmail.com Thu Feb 3 08:07:38 2022 From: krallinger.martin at gmail.com (Martin Krallinger) Date: Thu, 3 Feb 2022 14:07:38 +0100 Subject: Connectionists: Text Mining research position at my group (Barcelona Supercomputing Center, Text Mining Unit) In-Reply-To: References: Message-ID: We are looking for a motivated Text Mining researcher at my group, please share with potential candidates or apply if you are interested in the position. Reference: 16_22_LS_TM_RE2 Job: Text Mining research position at Barcelona Supercomputing Center URL: https://www.bsc.es/join-us/job-opportunities/1622lstmre2 Context And Mission The Text Mining Unit of the BSC-CNS (TeMU-BSC) is funded through the Plan de Impulso de las Tecnolog?as del Lenguaje de la Agenda Digital, by the Secretary of State of Telecommunications and the Information. It is the first publicly funded text mining unit in Spain and has the aim to promote the development of natural language processing and machine translation resources for Spanish and other co-official languages in the area of biomedicine. We search for a technical researcher in bioinformatics, medical informatics, natural language processing, computational linguistics or language engineering with a strong background in Machine Learning (ML), who will be responsible of the integration of data and the implementation of search tools into the platform for natural language processing designed by the unit. Key Duties Design, implementation, and evaluation of text mining, NLP, deep learning and ML tools and models applied to the clinical and biomedical application domain. Coordination and organization of shared tasks and evaluation campaigns (like IberLEF, CLEF eHealth,..). Technical coordination and supervision of Gold Standard annotation projects to generate high-quality text corpora. Requirements Education: University degree in computer science, mathematics, statistics, data scientist, physics, bioinformatics, telecommunications, electrical engineering or equivalent. Essential Knowledge and Professional Experience: Experience with Deep Learning and statistical data mining frameworks: Keras, Tensorflow, PySpark, PyTorch, Spacy, etc. Experience with ML algorithms and techniques: LDA, Topic Modelling, LSTM, KNN, SVM, Decision Trees, Clustering, Word Embeddings, etc. Experience in the development or management of software resources/tools, Github + github projects. Experience with NLP components and platforms. Additional Knowledge and Professional Experience: Strong programming skills in at least one of the following languages: Java, Python, C++, Scala, R. Experience and skills related to bash, Docker, Kubernetes, Unity testing, Collab Competences: Good communication and presentation skills. Strong technical writing skills. Ability to work both independently and within a team. Used to work under pressure under strict deadlines -- ======================================= Martin Krallinger, Dr. Head of Biological Text Mining Unit Barcelona Supercomputing Center (BSC-CNS) ======================================= -------------- next part -------------- An HTML attachment was scrubbed... URL: From chaumann at cor-lab.Uni-Bielefeld.DE Thu Feb 3 09:17:06 2022 From: chaumann at cor-lab.Uni-Bielefeld.DE (Carola Haumann) Date: Thu, 3 Feb 2022 15:17:06 +0100 Subject: Connectionists: =?utf-8?q?Save_the_date=3A_Next-generation_AIoT_a?= =?utf-8?q?pplications_=E2=80=93_VEDLIoT-Open_is_coming_soon?= Message-ID: Open Call Announcement: VEDLIoT ? ?Very Efficient Deep Learning in IoT? Save the date: the call will launch on 1st of March 2022 and has a deadline on 8th of May 2022 at 23:59 CEST. ******************************************** VEDLIoT ? ?Very Efficient Deep Learning in IoT? is an EU H2020 ICT-56-2020 funded research project running for 36 months, driven by challenging use cases in key sectors like automotive, automation, and smart home. The main objective of VEDLIoT is to develop the next generation of connected IoT devices utilising distributed deep learning. You can become part of the project by applying for an Open Call project! We expect to fund around ten research experiments incorporating additional use cases in the project utilising the developed technologies. This open call for cascaded funding is foreseen to explore new opportunities by extending the application of the VEDLIoT platform to a more extensive set of new and relevant use cases. It is expected that open call projects leverage VEDLIoT technologies for their own AI-related IoT use case, thereby broadening the VEDLIoT use-case basis and help making the overall concept more robust. *VEDLIoT-Open in a nutshell* Project acronym: VEDLIoT Project full name: Very Efficient Deep Learning in IoT Project VEDLIoT, co-funded by the European Union?s Horizon 2020 Research and Innovation Programme under grant agreement No. 957197, foresees as an eligible activity the provision of financial support to third parties, as a means to achieve its own objectives. For this Open Call, the types of activities to that qualify for receiving financial support are the next generation of AIoT applications in areas such as wearables, transportation, agriculture, homes, health, energy, and manufacturing. Call identifier: VEDLIoT - Open Call title: Next-generation AIoT applications ? VEDLIoT-Open Publication date: 01/03/2022 Deadline: 08/05/2022 at 23:59h CEST. Indicative budget for the call: 840.000 ? Expected duration of participation: 9 to 12 months Indicative budget for each proposal: up to 120,000 ? (including 25 % indirect costs, at a funding/reimbursement rate of 70 %) Language in which proposal must be submitted: English All OC-application documents can be found here, starting from 1st of March 2022: https://vedliot.eu/use-cases/open-call/ Contact us: info at vedliot.eu vedliot-open-support at vedliot.[eu|ai|io] From bwyble at gmail.com Thu Feb 3 08:30:00 2022 From: bwyble at gmail.com (Brad Wyble) Date: Thu, 3 Feb 2022 08:30:00 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> Message-ID: With respect to moving the goalposts, I don't think that's actually what is happening in this case. Rather we're learning how to better map the relationship between doing and understanding. When we didn't have the machines that could translate, we assumed that a model which could translate would have all of these other desirable properties, such as the ability to generalize, and resilience to making bizarre errors. But now that such machines exist, we can see that one does not bring the other. It's akin to seeing goalposts at the top of the hill and when you get there you realize that the goalposts are actually on a different hill that is much farther away. These NLP algorithms are a huge advance, both in terms of what they help us to do, and also in helping us to perceive what "understanding" really means. -Brad On Thu, Feb 3, 2022 at 6:53 AM Danko Nikolic wrote: > G. Hinton wrote: "I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster wearing a red hat and > it draws such a picture, it understood the request." > > I would like to suggest why drawing a hamster with a red hat does not > necessarily imply understanding of the statement "hamster wearing a red > hat". > To understand that "hamster wearing a red hat" would mean inferring, in > newly emerging situations of this hamster, all the real-life > implications that the red hat brings to the little animal. > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > ...and so on. > > Countless many questions can be asked. One has understood "hamster wearing > a red hat" only if one can answer reasonably well many of such real-life > relevant questions. Similarly, a student has understood materias in a class > only if they can apply the materials in real-life situations (e.g., > applying Pythagora's theorem). If a student gives a correct answer to a > multiple choice question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > I also suggest that understanding also comes together with effective > learning: We store new information in such a way that we can recall it > later and use it effectively i.e., make good inferences in newly emerging > situations based on this knowledge. > > In short: Understanding makes us humans able to 1) learn with a few > examples and 2) apply the knowledge to a broad set of situations. > > No neural network today has such capabilities and we don't know how to > give them such capabilities. Neural networks need large amounts of > training examples that cover a large variety of situations and then > the networks can only deal with what the training examples have already > covered. Neural networks cannot extrapolate in that 'understanding' sense. > > I suggest that understanding truly extrapolates from a piece of knowledge. > It is not about satisfying a task such as translation between languages or > drawing hamsters with hats. It is how you got the capability to complete > the task: Did you only have a few examples that covered something different > but related and then you extrapolated from that knowledge? If yes, this is > going in the direction of understanding. Have you seen countless examples > and then interpolated among them? Then perhaps it is not understanding. > > So, for the case of drawing a hamster wearing a red hat, understanding > perhaps would have taken place if the following happened before that: > > 1) first, the network learned about hamsters (not many examples) > 2) after that the network learned about red hats (outside the context of > hamsters and without many examples) > 3) finally the network learned about drawing (outside of the context of > hats and hamsters, not many examples) > > After that, the network is asked to draw a hamster with a red hat. If it > does it successfully, maybe we have started cracking the problem of > understanding. > > Note also that this requires the network to learn sequentially without > exhibiting catastrophic forgetting of the previous knowledge, which is > possibly also a consequence of human learning by understanding. > > > Danko > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > --- A progress usually starts with an insight --- > > > > Virus-free. > www.avast.com > > <#m_-4683042230811366927_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > >> Without getting into the specific dispute between Gary and Geoff, I think >> with approaches similar to GLOM, we are finally headed in the right >> direction. There?s plenty of neurophysiological evidence for single-cell >> abstractions and multisensory neurons in the brain, which one might claim >> correspond to symbols. And I think we can finally reconcile the decades old >> dispute between Symbolic AI and Connectionism. >> >> >> >> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >> an effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> GARY: I have *never* called for dismissal of neural networks, but rather >> for some hybrid between the two (as you yourself contemplated in 1991); the >> point of the 2001 book was to characterize exactly where multilayer >> perceptrons succeeded and broke down, and where symbols could complement >> them. >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> *From:* Connectionists *On >> Behalf Of *Gary Marcus >> *Sent:* Wednesday, February 2, 2022 1:26 PM >> *To:* Geoffrey Hinton >> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Geoff, and interested others, >> >> >> >> What, for example, would you make of a system that often drew the >> red-hatted hamster you requested, and perhaps a fifth of the time gave you >> utter nonsense? Or say one that you trained to create birds but sometimes >> output stuff like this: >> >> >> >> >> >> One could >> >> >> >> a. avert one?s eyes and deem the anomalous outputs irrelevant >> >> or >> >> b. wonder if it might be possible that sometimes the system gets the >> right answer for the wrong reasons (eg partial historical contingency), and >> wonder whether another approach might be indicated. >> >> >> >> Benchmarks are harder than they look; most of the field has come to >> recognize that. The Turing Test has turned out to be a lousy measure of >> intelligence, easily gamed. It has turned out empirically that the Winograd >> Schema Challenge did not measure common sense as well as Hector might have >> thought. (As it happens, I am a minor coauthor of a very recent review on >> this very topic: https://arxiv.org/abs/2201.02387 >> ) >> But its conquest in no way means machines now have common sense; many >> people from many different perspectives recognize that (including, e.g., >> Yann LeCun, who generally tends to be more aligned with you than with me). >> >> >> >> So: on the goalpost of the Winograd schema, I was wrong, and you can >> quote me; but what you said about me and machine translation remains your >> invention, and it is inexcusable that you simply ignored my 2019 >> clarification. On the essential goal of trying to reach meaning and >> understanding, I remain unmoved; the problem remains unsolved. >> >> >> >> All of the problems LLMs have with coherence, reliability, truthfulness, >> misinformation, etc stand witness to that fact. (Their persistent inability >> to filter out toxic and insulting remarks stems from the same.) I am hardly >> the only person in the field to see that progress on any given benchmark >> does not inherently mean that the deep underlying problems have solved. >> You, yourself, in fact, have occasionally made that point. >> >> >> >> With respect to embeddings: Embeddings are very good for natural language >> *processing*; but NLP is not the same as NL*U* ? when it comes to >> *understanding*, their worth is still an open question. Perhaps they >> will turn out to be necessary; they clearly aren?t sufficient. In their >> extreme, they might even collapse into being symbols, in the sense of >> uniquely identifiable encodings, akin to the ASCII code, in which a >> specific set of numbers stands for a specific word or concept. (Wouldn?t >> that be ironic?) >> >> >> >> (Your GLOM, which as you know I praised publicly, is in many ways an >> effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> >> >> Notably absent from your email is any kind of apology for misrepresenting >> my position. It?s fine to say that ?many people thirty years ago once >> thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. >> I have consistently felt throughout our interactions that you have mistaken >> me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me >> for having made that error. I am still not he. >> >> >> >> Which maybe connects to the last point; if you read my work, you would >> see thirty years of arguments *for* neural networks, just not in the way >> that you want them to exist. I have ALWAYS argued that there is a role for >> them; characterizing me as a person ?strongly opposed to neural networks? >> misses the whole point of my 2001 book, which was subtitled ?Integrating >> Connectionism and Cognitive Science.? >> >> >> >> In the last two decades or so you have insisted (for reasons you have >> never fully clarified, so far as I know) on abandoning symbol-manipulation, >> but the reverse is not the case: I have *never* called for dismissal of >> neural networks, but rather for some hybrid between the two (as you >> yourself contemplated in 1991); the point of the 2001 book was to >> characterize exactly where multilayer perceptrons succeeded and broke down, >> and where symbols could complement them. It?s a rhetorical trick (which is >> what the previous thread was about) to pretend otherwise. >> >> >> >> Gary >> >> >> >> >> >> On Feb 2, 2022, at 11:22, Geoffrey Hinton >> wrote: >> >> ? >> >> Embeddings are just vectors of soft feature detectors and they are very >> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >> opposite. >> >> >> >> A few decades ago, everyone I knew then would have agreed that the >> ability to translate a sentence into many different languages was strong >> evidence that you understood it. >> >> >> >> But once neural networks could do that, their critics moved the >> goalposts. An exception is Hector Levesque who defined the goalposts more >> sharply by saying that the ability to get pronoun references correct in >> Winograd sentences is a crucial test. Neural nets are improving at that but >> still have some way to go. Will Gary agree that when they can get pronoun >> references correct in Winograd sentences they really do understand? Or does >> he want to reserve the right to weasel out of that too? >> >> >> >> Some people, like Gary, appear to be strongly opposed to neural networks >> because they do not fit their preconceived notions of how the mind should >> work. >> >> I believe that any reasonable person would admit that if you ask a neural >> net to draw a picture of a hamster wearing a red hat and it draws such a >> picture, it understood the request. >> >> >> >> Geoff >> >> >> >> >> >> >> >> >> >> >> >> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >> >> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural >> network community, >> >> >> >> There has been a lot of recent discussion on this list about framing and >> scientific integrity. Often the first step in restructuring narratives is >> to bully and dehumanize critics. The second is to misrepresent their >> position. People in positions of power are sometimes tempted to do this. >> >> >> >> The Hinton-Hanson interview that you just published is a real-time >> example of just that. It opens with a needless and largely content-free >> personal attack on a single scholar (me), with the explicit intention of >> discrediting that person. Worse, the only substantive thing it says is >> false. >> >> >> >> Hinton says ?In 2015 he [Marcus] made a prediction that computers >> wouldn?t be able to do machine translation.? >> >> >> >> I never said any such thing. >> >> >> >> What I predicted, rather, was that multilayer perceptrons, as they >> existed then, would not (on their own, absent other mechanisms) >> *understand* language. Seven years later, they still haven?t, except in >> the most superficial way. >> >> >> >> I made no comment whatsoever about machine translation, which I view as a >> separate problem, solvable to a certain degree by correspondance without >> semantics. >> >> >> >> I specifically tried to clarify Hinton?s confusion in 2019, but, >> disappointingly, he has continued to purvey misinformation despite that >> clarification. Here is what I wrote privately to him then, which should >> have put the matter to rest: >> >> >> >> You have taken a single out of context quote [from 2015] and >> misrepresented it. The quote, which you have prominently displayed at the >> bottom on your own web page, says: >> >> >> >> Hierarchies of features are less suited to challenges such as language, >> inference, and high-level planning. For example, as Noam Chomsky famously >> pointed out, language is filled with sentences you haven't seen >> before. Pure classifier systems don't know what to do with such sentences. >> The talent of feature detectors -- in identifying which member of some >> category something belongs to -- doesn't translate into understanding >> novel sentences, in which each sentence has its own unique meaning. >> >> >> >> It does *not* say "neural nets would not be able to deal with novel >> sentences"; it says that hierachies of features detectors (on their own, if >> you read the context of the essay) would have trouble *understanding *novel sentences. >> >> >> >> >> Google Translate does yet not *understand* the content of the sentences >> is translates. It cannot reliably answer questions about who did what to >> whom, or why, it cannot infer the order of the events in paragraphs, it >> can't determine the internal consistency of those events, and so forth. >> >> >> >> Since then, a number of scholars, such as the the computational linguist >> Emily Bender, have made similar points, and indeed current LLM difficulties >> with misinformation, incoherence and fabrication all follow from these >> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf >> , >> also emphasizing issues of understanding and meaning: >> >> >> >> *The success of the large neural language models on many NLP tasks is >> exciting. However, we find that these successes sometimes lead to hype in >> which these models are being described as ?understanding? language or >> capturing ?meaning?. In this position paper, we argue that a system trained >> only on form has a priori no way to learn meaning. .. a clear understanding >> of the distinction between form and meaning will help guide the field >> towards better science around natural language understanding. * >> >> >> >> Her later article with Gebru on language models ?stochastic parrots? is >> in some ways an extension of this point; machine translation requires >> mimicry, true understanding (which is what I was discussing in 2015) >> requires something deeper than that. >> >> >> >> Hinton?s intellectual error here is in equating machine translation with >> the deeper comprehension that robust natural language understanding will >> require; as Bender and Koller observed, the two appear not to be the same. >> (There is a longer discussion of the relation between language >> understanding and machine translation, and why the latter has turned out to >> be more approachable than the former, in my 2019 book with Ernest Davis). >> >> >> >> More broadly, Hinton?s ongoing dismissiveness of research from >> perspectives other than his own (e.g. linguistics) have done the field a >> disservice. >> >> >> >> As Herb Simon once observed, science does not have to be zero-sum. >> >> >> >> Sincerely, >> >> Gary Marcus >> >> Professor Emeritus >> >> New York University >> >> >> >> On Feb 2, 2022, at 06:12, AIhub wrote: >> >> ? >> >> Stephen Hanson in conversation with Geoff Hinton >> >> >> >> In the latest episode of this video series for AIhub.org >> , >> Stephen Hanson talks to Geoff Hinton about neural networks, >> backpropagation, overparameterization, digit recognition, voxel cells, >> syntax and semantics, Winograd sentences, and more. >> >> >> >> You can watch the discussion, and read the transcript, here: >> >> >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> >> >> >> About AIhub: >> >> AIhub is a non-profit dedicated to connecting the AI community to the >> public by providing free, high-quality information through AIhub.org >> >> (https://aihub.org/ >> ). >> We help researchers publish the latest AI news, summaries of their work, >> opinion pieces, tutorials and more. We are supported by many leading >> scientific organizations in AI, namely AAAI >> , >> NeurIPS >> , >> ICML >> , >> AIJ >> >> /IJCAI >> , >> ACM SIGAI >> , >> EurAI/AICOMM, CLAIRE >> >> and RoboCup >> >> . >> >> Twitter: @aihuborg >> >> > > Virus-free. > www.avast.com > > <#m_-4683042230811366927_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> > -- Brad Wyble Associate Professor Psychology Department Penn State University http://wyblelab.com -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 319547 bytes Desc: not available URL: From helma.torkamaan at uni-due.de Thu Feb 3 10:37:01 2022 From: helma.torkamaan at uni-due.de (Helma Torkamaan) Date: Thu, 3 Feb 2022 16:37:01 +0100 Subject: Connectionists: ACM UMAP 2022: Last Call for Papers Message-ID: <1fb09a05-1fae-16a2-46b3-bae8b0665f6d@uni-due.de> --- Please forward to anyone who might be interested --- --- Apologies for cross-posting --- -------------------------------------------------------- # **30th ACM Conference on User Modeling, Adaptation and Personalization (UMAP ?22)** Barcelona*, Spain, July 4?7, 2022 http://www.um.org/umap2022/ (*) Due to the ongoing COVID-19 pandemic, we are planning for a hybrid conference and will accommodate online presentations where needed. Submission Deadline: - Abstracts due: *February 10, 2022* (mandatory) - Full paper due: February 17, 2022 -------------------------------------------------------- **BACKGROUND AND SCOPE** ============================ **ACM UMAP** ? ***User Modeling, Adaptation and Personalization*** ? is the premier international conference for researchers and practitioners working on systems that adapt to individual users or to groups of users, and that collect, represent, and model user information. **ACM UMAP** is sponsored by ACM SIGCHI (https://sigchi.org) and SIGWEB (https://www.sigweb.org), and organized with User Modeling Inc. (https://um.org) as the core Steering Committee, extended with past years? chairs. The proceedings are published by the **ACM** and will be part of the ACM Digital Library (https://dl.acm.org). **ACM UMAP** covers a wide variety of research areas where personalization and adaptation may be applied. The main theme of **UMAP 2022** is ***?User control in personalized systems?***. Specifically, we welcome submissions related to user modeling, personalization, and adaptation in all areas of personalized systems, with an emphasis on how to balance adaptivity and user control. Below we present a short (but not prescriptive) list of topics of importance to the conference. ACM UMAP is co-located and collaborates with the ACM Hypertext conference (https://ht.acm.org/ht2022/). UMAP takes place one week after Hypertext, and uses the same submission dates and formats. We expect authors to submit research on personalized systems to UMAP and invite authors to submit their Web-related work without a focus on personalization to the Hypertext conference. The two conferences will organize one shared track on **personalized recommender systems** (same track chairs and PC, see the track description). -------------------------------------------------------- ?**IMPORTANT DATES** ============================ - Paper Abstracts:?? February 10, 2022 (mandatory) - Full paper:???????????? February 17, 2022 - Notification:?????????? April 11, 2022 - Conference:????????? July 4-July 7, 2022 **Note**: The submissions deadlines are at 11:59pm AoE time (Anywhere on Earth) -------------------------------------------------------- ?**CONFERENCE TOPICS** ?============================ We welcome submissions related to *user modeling, personalization, and adaptation in any area*. The topics listed below are not intended to limit possible contributions. **Detailed descriptions and the suggested topics for each track are reported in the online version of the CFP on the UMAP 2022 web site.** ### **Personalized Recommender Systems*** **Track Chairs: Osnat Mokryn (University of Haifa), Eva Zangerle (University of Innsbruck, Austria) and Markus Zanker (University of Bolzano, Italy, and University of Klagenfurt, Austria)** (*) This is a joint track between ACM UMAP and ACM Hypertext (same track chairs, overlapping PC). Authors planning to contribute to this track can submit to either conference, depending on their broader interest in either Hypertext or UMAP. Track chairs organize a special issue in the journal New Review of Hypermedia and Multimedia. This track aims to provide a forum for researchers and practitioners to discuss open challenges, latest solutions and novel research approaches in the field of recommender systems. In addition to mature research works addressing technical aspects pertaining to recommendations, we also particularly welcome research contributions that address questions related to the user perception and the business value of recommender systems. ### **Adaptive Hypermedia, Semantic, and Social Web** **Track Chairs: Alexandra I. Cristea (Durham University, UK) and Peter Brusilovsky (University of Pittsburgh, US)** This track aims to provide a forum to researchers to discuss open research problems, solid solutions, latest challenges, novel applications, and innovative research approaches in adaptive hypermedia, semantic and social web. We invite original submissions addressing all aspects of personalization, user models building, and personal experience in online social systems. ### **Intelligent User Interfaces** **Track chairs: Elisabeth Lex (Graz University of Technology, Austria) and Marko Tkalcic (University of Primorska, Slovenia)** This topic can be characterized by exploring how to make the interaction between computers and people smarter and more productive, which may leverage solutions from human-computer interaction, data mining, natural language processing, information visualization, and knowledge representation and reasoning. ### **Technology-Enhanced Adaptive Learning** **Track chairs: Judy Kay (University of Sydney, Australia) and Sharon Hsiao (Santa Clara University, US)** This track invites researchers, developers, and practitioners from various disciplines to present their innovative learning solutions, share acquired experience, and discuss their modeling challenges for personalized adaptive learning. ### **Fairness, Transparency, Accountability, and Privacy** **Track chairs: Bamshad Mobasher (DePaul University College of Computing and Digital Media, US) and Munindar P. Singh (NC State University, US)** Adaptive systems researchers and developers have a social responsibility to care about the impact of their technologies on individual people (users, providers, and other stakeholders) and on society. This track invites work that pertains to the science of building, maintaining, evaluating, and studying adaptive systems that are fair, transparent, respectful of users? privacy, and beneficial to society. ### **Personalization for Persuasive and Behavior Change Systems** **Track chairs: Julita Vassileva (University of Saskatchewan, Canada) and Panagiotis Germanakos (SAP SE, Germany)** This track invites original submissions addressing the areas of personalization and tailoring for persuasive technologies, including but not limited to personalization models, user models, computational personalization, design and evaluation methods, and personal experience designing personalized and adaptive behaviour change technologies. ### **Virtual Assistants and Personalized Human-robot Interaction** **Track chairs: Radhika Garg (Syracuse University, US) and Cristina Gena (University of Torino, Italy)** This track aims at investigating new models and techniques for the adaptation of synthetic companions (e.g., virtual assistants, chatbots, social robots) to the individual user. ### **Research Methods and Reproducibility** **Track chairs: Odd Erik Gundersen (Norwegian University of Science and Technology, Norway) and Dietmar Jannach (University of Klagenfurt, Austria)** This track accepts works on methodologies for the evaluation of personalized systems, benchmarks, measurement scales, with particular attention to reproducibility of results and of techniques. -------------------------------------------------------- ?**SUBMISSION AND REVIEW PROCESS** ?============================ Please consult the conference website for the submission link: http://www.um.org/umap2022/. The maximum length is **14 pages (excluding references) in the ACM new single-column format**. We encourage papers of any length up to 14 pages; reviewers will be asked to comment on whether the length is appropriate for the contribution. **Additional review criteria are available in the online version of the CFP on the UMAP 2022 web site.** Each accepted paper will be included in the conference proceedings and presented at the conference. UMAP uses a **double blind** review process. Authors must omit their names and affiliations from submissions, and avoid obvious identifying statements. For instance, citations to the authors' own prior work should be made in the third person. Failure to anonymize your submission results in the desk-rejection of your paper. -------------------------------------------------------- **ORGANIZERS** ============================ **General chairs** - Ludovico Boratto, University of Cagliari, Italy - Alejandro Bellog?n, Universidad Aut?noma de Madrid, Spain - Olga C. Santos, Spanish National University for Distance Education, Spain **Program Chairs** - Liliana Ardissono, University of Torino, Italy - Bart Knijnenburg, Clemson University, US -------------------------------------------------------- **RELATED EVENTS** ============================ Separate calls will be sent for Workshops, Doctoral Consortium, and Demo/Late-Breaking Results, as these have different deadlines and submission requirements. -- Helma Torkamaan Researcher University of Duisburg-Essen Interactive Systems Group Room LF 288, Forsthausweg 2, 47057 Duisburg, GERMANY helma.torkamaan at uni-due.de / +49 203 379-2276 -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Thu Feb 3 11:25:34 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Thu, 3 Feb 2022 08:25:34 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> Message-ID: <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." > > I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". > To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. > > What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) > What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) > What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) > What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) > > ...and so on. > > Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). > > I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. > > In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. > > No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. > > I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. > > So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: > > 1) first, the network learned about hamsters (not many examples) > 2) after that the network learned about red hats (outside the context of hamsters and without many examples) > 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) > > After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. > > Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. > > > Danko > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > --- A progress usually starts with an insight --- > > > Virus-free. www.avast.com > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: > Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. > > > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) > > GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > > > From: Connectionists > On Behalf Of Gary Marcus > Sent: Wednesday, February 2, 2022 1:26 PM > To: Geoffrey Hinton > > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > > > Dear Geoff, and interested others, > > > > What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: > > > > > > > > One could > > > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > or > > b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. > > > > Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387 ) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). > > > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. > > > > All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. > > > > With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > > > (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) > > > > Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. > > > > Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? > > > > In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. > > > > Gary > > > > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: > > ? > > Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. > > > > A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. > > > > > But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? > > > > Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. > > I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. > > > > Geoff > > > > > > > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, > > > > There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. > > > > The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. > > > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? > > > > I never said any such thing. > > > > What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. > > > > I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. > > > > I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: > > > > You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: > > > > Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. > > > > It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. > > > > Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. > > > > Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf , also emphasizing issues of understanding and meaning: > > > > The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. > > > > Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. > > > > Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). > > > > More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. > > > > As Herb Simon once observed, science does not have to be zero-sum. > > > > Sincerely, > > Gary Marcus > > Professor Emeritus > > New York University > > > > > On Feb 2, 2022, at 06:12, AIhub > wrote: > > ? > > Stephen Hanson in conversation with Geoff Hinton > > > > In the latest episode of this video series for AIhub.org , Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. > > > > You can watch the discussion, and read the transcript, here: > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > About AIhub: > > AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information throughAIhub.org (https://aihub.org/ ). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI , NeurIPS , ICML , AIJ /IJCAI , ACM SIGAI , EurAI/AICOMM, CLAIRE and RoboCup . > > Twitter: @aihuborg > > > Virus-free. www.avast.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From nagai.yukie at mail.u-tokyo.ac.jp Thu Feb 3 23:55:48 2022 From: nagai.yukie at mail.u-tokyo.ac.jp (nagai.yukie at mail.u-tokyo.ac.jp) Date: Fri, 4 Feb 2022 04:55:48 +0000 Subject: Connectionists: Assistant Professor / Postdoc Researcher in computational neuroscience and/or developmental robotics at the University of Tokyo Message-ID: Dear colleagues, The International Research Center for Neurointelligence (IRCN) at the University of Tokyo is looking for highly motivated full-time Assistant Professor / Postdoc researchers in the field of computational neuroscience and/or developmental robotics. The selected candidates will join Cognitive Developmental Robotics Laboratory head by Prof. Yukie Nagai and investigate the underlying mechanisms of human cognitive development and disorders by means of computational approaches. Please send the required documents if you are interested in. Assistant Professor / Postdoc Researcher at IRCN, the University of Tokyo https://ircn.jp/wp-content/uploads/2021/12/20211220_IRCN_Nagai_lab_2EN.pdf Cognitive Developmental Robotics Laboratory (Nagai Lab) http://developmental-robotics.jp ---------- 1. Job title / Number of positions Project Research Associate or Project Researcher (Postdoctoral Fellow) / One or two 2. Employment period Starting Date: Negotiable Contract duration: until March 31, 2022 3. Renewable The contract is renewable on a fiscal year basis (from April 1 to March 31; every year) according to research budget, research activity, and research achievements. Contract duration is until March 31, 2026. 4. Place of work International Research Center for Neurointelligence, The University of Tokyo 7-3-1 Hongo, Bunkyo-ku Tokyo, 113-0033 JAPAN 5. Description Successful applicants will work in the fields of computational neuroscience, cognitive developmental robotics, machine learning, and relevant topics. They will investigate the principle of human cognitive development and disorders by modeling computational neural networks inspired by the human brain and/or analyzing human cognitive behaviors. Please see the lab homepage (https://developmental-robotics.jp/en/) for more details. 6. Salary and benefits Salary: To be determined in accordance with the University of Tokyo Regulations Commuter allowance: JPY55,000 per month at maximum No retirement benefits or bonuses 7. Qualifications - PhD in engineering, computer science, cognitive science, neuroscience, or relevant fields - Communication skills in English - Programming skills - Experiences in computational neuroscience and/or cognitive developmental robotics are preferred 8. Application documents (1) Cover letter (A4 1 page) (2) Curriculum vita (3) Publication list (4) Research plan (A4 2-3 pages) (5) Name, affiliation, and email address of two references, one of which should be a previous employer or supervisor 9. Submission Please send the application documents (PDF) to yukie at ircn.jp 10. Application deadline / Selection process When the position is filled All applications will be screened, and only those qualified will be scheduled for an interview (on-site or online). 11. Inquiries Please contact Prof. Yukie Nagai at IRCN, the University of Tokyo yukie at ircn.jp ? Yukie Nagai, Ph.D. Project Professor, The University of Tokyo nagai.yukie at mail.u-tokyo.ac.jp | https://developmental-robotics.jp CREST Cognitive Mirroring: https://cognitive-mirroring.org CREST Cognitive Feeling: https://cognitive-feeling.jp From gary.marcus at nyu.edu Thu Feb 3 12:06:19 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Thu, 3 Feb 2022 09:06:19 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <9b988b04-250b-56c8-c3ec-9509bba4bdc8@rubic.rutgers.edu> References: <9b988b04-250b-56c8-c3ec-9509bba4bdc8@rubic.rutgers.edu> Message-ID: <21C84870-23E7-4F53-8466-E09CBFBB0221@nyu.edu> Dear Steve, Thanks for your gracious note, but the passage I objected to was not one line but several, (eg ?I wish people would stop taking [Marcus] so seriously?), and, critically, Hinton?s fabrication regarding machine translation was not an innocent mistake. People misrepresent me all the time, but in this case I had already specifically advised Hinton in private that what he said was misrepresenting me on this exact point. Moreover, the bullying campaign goes back several years. For example, moments before I gave a 2019 talk to an audience of computer scientists in Toronto, Hinton replied-all to the invite, in an effort to persuade the audience not to go. As it happens, his email didn?t get through to the list but I was cc?d (possibly inadvertently) and did receive it; the bullying intent was quite apparent. If AIHub believed in its charter, and wished to elevate the discourse, it would take down the entire passage and not support a persistent campaign to misrepresent and stifle a critical voice. By all means, if Hinton wants to write a reasoned critique of my work, publish it. But as one eminent computer scientist wrote to me yesterday, ?cronyism, collusion and now bullying and dehumanizing critics ? should not be rewarded by this community.? Thank you for allowing me to clarify. And yes, I will read the rest of the interview with interest, and of course I?d love to get your thoughts about how the ?hybrid movie? might end, but I?ll put that in a separate thread :) Cheers. Gary > On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson wrote: > > Since AIHUB was the one who posted the recent discussion between myself and Geoff, I suppose that could be the proximate causal event. > But in looking over the transcript, I think you are overreacting to the single en passant comment in an hour about you. Hardly, a "personal attack". > > Moreover, I've known you for over 30 years, and frankly it does sound like something you might have said );-) And I agree your positions in various books and articles have evolved over time. But I and AIHUB are always happy to correct transcripts that have seriously miss-represented someone's claims. > > I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. > Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. > > Best and hope you are doing well. > > Steve > > > > > > > > On 2/2/22 9:52 PM, Gary Marcus wrote: >> Dear Geoff, >> >> Causality is often hard to establish, but I didn't start this thread; I merely responded to a false assertion of yours that was publicized at the top. >> >> More broadly, it's a shame for the field that you won't engage in the real issues at hand, even with a clear home-court advantage. >> >> Gary >> >>> >>> On Feb 2, 2022, at 12:52, Geoffrey Hinton wrote: >>> >>> ? >>> You started this thread and it was a mistake for me to engage in arguing with you. I have said all I want to say. You have endless time for arguing and I don't. I find it more productive to spend time writing programs to see what works and what doesn't. You should try it sometime. >>> >>> Geoff >>> >>> >>> > -- > -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Thu Feb 3 15:35:49 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Thu, 3 Feb 2022 20:35:49 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. DANKO: What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Gary Marcus Sent: Thursday, February 3, 2022 9:26 AM To: Danko Nikolic Cc: Asim Roy ; Geoffrey Hinton ; AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub > wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Thu Feb 3 11:59:58 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Thu, 3 Feb 2022 08:59:58 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <9b988b04-250b-56c8-c3ec-9509bba4bdc8@rubic.rutgers.edu> References: <9b988b04-250b-56c8-c3ec-9509bba4bdc8@rubic.rutgers.edu> Message-ID: <8EA19260-6F0C-421D-8AF3-55E74C2FC3E9@nyu.edu> Steve, I?d love to hear you elaborate on this part, Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. I?d love your take on why you think the impetus towards hybrid models ended badly before, and why you think that the mistakes of the past can?t be corrected. Also it? would be really instructive to compare with deep learning, which lost steam for quite some time, but reemerged much stronger than ever before. Might not the same happen with hybrid models? I am cc?ing some folks (possibly not on this list) who have recently been sympathetic to hybrid models, in hopes of a rich discussion. (And, Geoff, still cc?d, I?d genuinely welcome your thoughts if you want to add them, despite our recent friction.) Cheers, Gary > On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson wrote: > > > I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. > Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. > > Best and hope you are doing well. > > Steve > -------------- next part -------------- An HTML attachment was scrubbed... URL: From d.bach at uni-bonn.de Thu Feb 3 12:17:00 2022 From: d.bach at uni-bonn.de (Dominik Bach) Date: Thu, 3 Feb 2022 17:17:00 +0000 Subject: Connectionists: Post doc positions in AI and Neuroscience at University of Bonn, Germany (topics: motion capture, virtual reality, learning theory, ML and software development) In-Reply-To: References: Message-ID: <57cc32fb-d99b-a3aa-498c-72b497e59d5b@uni-bonn.de> /Please circulate - apologies for cross-posting./ The Hertz Chair for Artificial Intelligence and Neuroscience at University of Bonn is looking to recruit postdoctoral fellows for an interdisciplinary neuroscience research program coordinated and supervised by Professor Dominik Bach. This program brings together researchers with expertise in cognitive(-computational) science, movement science, machine-learning, and software development. This provides an exciting opportunity for postdoctoral candidates to work at the cutting edge of human cognitive science and neuroscience research. Collaboration partners in this endeavour are based at Max-Planck-Institute for Biological Cybernetics in T?bingen (Germany), University of T?bingen (Germany), Max-Planck UCL Centre for Computational Psychiatry (UK) and Wellcome Centre for Human Neuroimaging (UK). The aim of the research is to understand the *cognitive neurobiology of human threat avoidance*, in terms of acute escape behaviour as well as medium- and long-term threat forecasting. Our research strongly builds on computational modelling of behaviour and neural systems, theories of artificial agents, machine-learning methods such as pose estimation and motion sequencing, and research automation by software design and by self-learning data analysis methods. Our team culture is collaborative, agile, and shaped by technical sophistication. We believe in open, reproducible, and sustainable precision science. We host a state-of-the-art virtual reality and motion capture lab, and have access to human neuroimaging facilities (3 T and 7 T MRI, OPM-MEG). The successful candidates will be based at the *University of Bonn, Campus Endenich*, in direct vicinity to natural and computer science departments and other interdisciplinary Hertz Chairs. The University of Bonn is an internationally leading research university, providing an intellectually stimulating environment. At University of Bonn, postdoctoral salaries start at around 55'000 ?/year depending on prior post-doctoral experience. The positions are available on or after 1 April 2022. An initial appointment for a two-year period will be made with potential for extension depending on successful performance of research and publications. University of Bonn is committed to diversity and encourages applications from underrepresented groups. Qualified postdoctoral applicants should submit a current CV including publication list, a personal statement describing their experience and interests, and contact information for three references to d.bach at uni-bonn.de. *Post doc positions are initially based in the following fields. We welcome enquries from candidates in related fields of cognitive-computational neuroscience.* * **Post doc Motion Capture* The goal is to understand human motor behaviour under acute, immediate threat. We investigate this in an immersive virtual reality (VR) environment, in which people can move to avoid various threats. The candidate will conduct full-body markerless and marker-based motion capture, pose estimation, recover kinematics, and structure the recorded movement trajectories with statistical and machine-learning models. Applicants should have (or be close to obtaining) a PhD in machine-learning, robotics, computer science, motor science, biomechanics, computational neuroscience, or a related area, by the agreed start date of the position. Experience with motion capture, pose estimation, inverse kinematics (in humans or robots), movement trajectory analysis and structuring/sequencing are essential. Strong background in contemporary machine-learning and applied statistics is essential, as are solid mathematical skills and good general IT and software development knowledge. Familiarity with virtual reality and/or human/animal defensive behaviour would be desirable. *Post doc VR* The goal is to develop a cognitive-computational understanding of human decision-making under acute, immediate threat. We investigate this in an immersive virtual reality (VR) environment, in which people can move to avoid various threats. The candidate's role will be to maintain and advance an existing Unity-based research platform, build specific suitable scenarios, conduct experimental studies with this setup, and analyse the data. Applicants should have (or be close to obtaining) a PhD in cognitive-computational (neuro)science, applied machine-learning, biomechanics, motor science, a quantitative field of psychology (e.g. decision-making, perception), or a related area by the agreed start date of the position. Experience with Unity and C# are essential, familiarity with R would be desirable. The successful candidate will have experience in programming is essential, solid knowledge of decision science, applied statistics and a good publication record. *Post doc learning theory* The goal is to understand the computational algorithms by which humans learn to predict and avoid threat. Experimentally, we investigate this using human fear conditioning and VR-based avoidance learning. The candidate will build and test computational learning models using existing experimental data, and design new experiments to disambiguate candidate models. They will maintain and advance software frameworks for model benchmarking and Bayesian experimental design optimisation, and model-based data analysis. Applicants should have (or be close to obtaining) a PhD in cognitive-computational (neuro)science, computer science, machine learning, mathematics, a quantitative field of psychology (e.g. decision-making, perception), or a related area by the agreed start date of the position. Experience with learning theory in biological or artificial agents is essential; familiarity with analysis of biological/psychological data would be desirable. The successful candidate will have solid knowledge of mathematical statistics and experience with modern software development techniques. *Post doc software development* Our goal is to develop open, reproducible and sustainable, precision methods in the field of human cognitive neuroscience. To this end, we develop and maintain several software frameworks for computational model benchmarking, model-based data analysis, Bayesian experimental design optimisation, and collaborative data bases. The candidate will build on contemporary methods theory and metrology to advance and integrate these tools into an autonomous, continuously integrating, and self-learning software ecosystem. Applicants should have (or be close to obtaining) a PhD in computer science, cognitive-computational (neuro)science, software engineering, machine learning, mathematics, a quantitative field of psychology (e.g. decision-making, perception), or a related area by the agreed start date of the position. Experience with modern software development techniques is essential. A good understanding of cognitive-computational modelling and data sharing practices would be desirable. The successful candidate will have solid knowledge of applied statistics and machine learning, and experience with managing biological/psychological data. -------------- next part -------------- An HTML attachment was scrubbed... URL: From tgd at oregonstate.edu Thu Feb 3 12:31:37 2022 From: tgd at oregonstate.edu (Dietterich, Thomas) Date: Thu, 3 Feb 2022 17:31:37 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub > wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From lorincz at inf.elte.hu Thu Feb 3 13:34:18 2022 From: lorincz at inf.elte.hu (Andras Lorincz) Date: Thu, 3 Feb 2022 18:34:18 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: I would say that Geoff's example still does not go beyond "Chinese room capabilities". I lack reassuring formulation/evidence that there is anything/something/nothing beyond "Chinese room" type feedforward intelligence. Andr?s ------------------------------------ Andras Lorincz http://nipg.inf.elte.hu/ Fellow of the European Association for Artificial Intelligence Department of Artificial Intelligence Faculty of Informatics Eotvos Lorand University Budapest, Hungary ________________________________ From: Connectionists on behalf of Gary Marcus Sent: Thursday, February 3, 2022 5:25 PM To: Danko Nikolic Cc: connectionists at mailman.srv.cs.cmu.edu ; AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub > wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From laurent.perrinet at univ-amu.fr Fri Feb 4 03:19:25 2022 From: laurent.perrinet at univ-amu.fr (PERRINET Laurent) Date: Fri, 4 Feb 2022 08:19:25 +0000 Subject: Connectionists: crowd-sourcing COSYNE post-review score sheet Message-ID: <36CC6CD1-C12C-4E5E-BA9C-37E6FE5999F8@univ-amu.fr> Dear community COSYNE is a great conference which plays a pivotal role in our field. Raw numbers we were given are * 881 submitted abstracts * 215 independent reviewers * 2639 reviews If you have submitted an abstract (or several) you have recently received your scores. I am not affiliated to COSYNE - yet I would like to contribute in some way and would like to ask one minute of your time to report the raw scores from your reviewers: https://forms.gle/p7eG1p6dJAkr4Cyg7 (Do one form per abstract.) For this crowd-sourcing effort to have a most positive impact, I will share the results and summarize in a few lines them in one week time (11/02). The more numerous your feedbacks, the higher the precision of results! Thanks in advance for your action, Laurent PS: if any similar initiative already exists, I'll be more than willing to receive feedback -- Laurent Perrinet - INT (UMR 7289) AMU/CNRS https://laurentperrinet.github.io/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Fri Feb 4 06:51:34 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Fri, 4 Feb 2022 12:51:34 +0100 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: I suppose everyone agrees that "the brain is a physical system", and that "There is no ?magic? inside the brain", and that '?understanding? is just part of ?learning.?' Also, we can agree that some sort of simulation takes place behind understanding. However, there still is a problem: Neural network's can't implement the needed simulations; they cannot achieve the same cognitive effect that human minds can (or animal minds can). We don't know a way of wiring a neural network such that it could perform the simulations (understandings) necessary to find the answers to real-life questions, such as the hamster with a hat problem. In other words, neural networks, as we know them today, cannot: 1) learn from a small number of examples (simulation or not) 2) apply the knowledge to a wide range of situations We, as scientists, do not understand understanding. Our technology's simulations (their depth of understanding) are no match for the simulations (depth of understanding) that the biological brain performs. I think that scientific integrity also covers acknowledging when we did not (yet) succeed in solving a certain problem. There is still significant work to be done. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- Virenfrei. www.avast.com <#DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> On Thu, Feb 3, 2022 at 9:35 PM Asim Roy wrote: > First of all, the brain is a physical system. There is no ?magic? inside > the brain that does the ?understanding? part. Take for example learning to > play tennis. You hit a few balls - some the right way and some wrong ? but > you fairly quickly learn to hit them right most of the time. So there is > obviously some simulation going on in the brain about hitting the ball in > different ways and ?learning? its consequences. What you are calling > ?understanding? is really these simulations about different scenarios. It?s > also very similar to augmentation used to train image recognition systems > where you rotate images, obscure parts and so on, so that you still can say > it?s a cat even though you see only the cat?s face or whiskers or a cat > flipped on its back. So, if the following questions relate to > ?understanding,? you can easily resolve this by simulating such scenarios > when ?teaching? the system. There?s nothing ?magical? about > ?understanding.? As I said, bear in mind that the brain, after all, is a > physical system and ?teaching? and ?understanding? is embodied in that > physical system, not outside it. So ?understanding? is just part of > ?learning,? nothing more. > > > > DANKO: > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Gary Marcus > *Sent:* Thursday, February 3, 2022 9:26 AM > *To:* Danko Nikolic > *Cc:* Asim Roy ; Geoffrey Hinton < > geoffrey.hinton at gmail.com>; AIhub ; > connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Danko, > > > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, > in which he said (paraphrasing from memory, because I don?t remember the > precise words) that the famous 200 Quoc Le unsupervised model [ > https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf > ] > had learned the concept of a ca. In reality the model had clustered > together some catlike images based on the image statistics that it had > extracted, but it was a long way from a full, counterfactual-supporting > concept of a cat, much as you describe below. > > > > I fully agree with you that the reason for even having a semantics is as > you put it, "to 1) learn with a few examples and 2) apply the knowledge to > a broad set of situations.? GPT-3 sometimes gives the appearance of having > done so, but it falls apart under close inspection, so the problem remains > unsolved. > > > > Gary > > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > > > G. Hinton wrote: "I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster wearing a red hat and > it draws such a picture, it understood the request." > > > > I would like to suggest why drawing a hamster with a red hat does not > necessarily imply understanding of the statement "hamster wearing a red > hat". > > To understand that "hamster wearing a red hat" would mean inferring, in > newly emerging situations of this hamster, all the real-life > implications that the red hat brings to the little animal. > > > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > ...and so on. > > > > Countless many questions can be asked. One has understood "hamster wearing > a red hat" only if one can answer reasonably well many of such real-life > relevant questions. Similarly, a student has understood materias in a class > only if they can apply the materials in real-life situations (e.g., > applying Pythagora's theorem). If a student gives a correct answer to a > multiple choice question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > > > I also suggest that understanding also comes together with effective > learning: We store new information in such a way that we can recall it > later and use it effectively i.e., make good inferences in newly emerging > situations based on this knowledge. > > > > In short: Understanding makes us humans able to 1) learn with a few > examples and 2) apply the knowledge to a broad set of situations. > > > > No neural network today has such capabilities and we don't know how to > give them such capabilities. Neural networks need large amounts of > training examples that cover a large variety of situations and then > the networks can only deal with what the training examples have already > covered. Neural networks cannot extrapolate in that 'understanding' sense. > > > > I suggest that understanding truly extrapolates from a piece of knowledge. > It is not about satisfying a task such as translation between languages or > drawing hamsters with hats. It is how you got the capability to complete > the task: Did you only have a few examples that covered something different > but related and then you extrapolated from that knowledge? If yes, this is > going in the direction of understanding. Have you seen countless examples > and then interpolated among them? Then perhaps it is not understanding. > > > > So, for the case of drawing a hamster wearing a red hat, understanding > perhaps would have taken place if the following happened before that: > > > > 1) first, the network learned about hamsters (not many examples) > > 2) after that the network learned about red hats (outside the context of > hamsters and without many examples) > > 3) finally the network learned about drawing (outside of the context of > hats and hamsters, not many examples) > > > > After that, the network is asked to draw a hamster with a red hat. If it > does it successfully, maybe we have started cracking the problem of > understanding. > > > > Note also that this requires the network to learn sequentially without > exhibiting catastrophic forgetting of the previous knowledge, which is > possibly also a consequence of human learning by understanding. > > > > > > Danko > > > > > > > > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > > > > Virus-free. www.avast.com > > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > > Without getting into the specific dispute between Gary and Geoff, I think > with approaches similar to GLOM, we are finally headed in the right > direction. There?s plenty of neurophysiological evidence for single-cell > abstractions and multisensory neurons in the brain, which one might claim > correspond to symbols. And I think we can finally reconcile the decades old > dispute between Symbolic AI and Connectionism. > > > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > GARY: I have *never* called for dismissal of neural networks, but rather > for some hybrid between the two (as you yourself contemplated in 1991); the > point of the 2001 book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols could complement > them. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Gary Marcus > *Sent:* Wednesday, February 2, 2022 1:26 PM > *To:* Geoffrey Hinton > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Geoff, and interested others, > > > > What, for example, would you make of a system that often drew the > red-hatted hamster you requested, and perhaps a fifth of the time gave you > utter nonsense? Or say one that you trained to create birds but sometimes > output stuff like this: > > > > > > > > One could > > > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > or > > b. wonder if it might be possible that sometimes the system gets the right > answer for the wrong reasons (eg partial historical contingency), and > wonder whether another approach might be indicated. > > > > Benchmarks are harder than they look; most of the field has come to > recognize that. The Turing Test has turned out to be a lousy measure of > intelligence, easily gamed. It has turned out empirically that the Winograd > Schema Challenge did not measure common sense as well as Hector might have > thought. (As it happens, I am a minor coauthor of a very recent review on > this very topic: https://arxiv.org/abs/2201.02387 > ) > But its conquest in no way means machines now have common sense; many > people from many different perspectives recognize that (including, e.g., > Yann LeCun, who generally tends to be more aligned with you than with me). > > > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote > me; but what you said about me and machine translation remains your > invention, and it is inexcusable that you simply ignored my 2019 > clarification. On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains unsolved. > > > > All of the problems LLMs have with coherence, reliability, truthfulness, > misinformation, etc stand witness to that fact. (Their persistent inability > to filter out toxic and insulting remarks stems from the same.) I am hardly > the only person in the field to see that progress on any given benchmark > does not inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > > > With respect to embeddings: Embeddings are very good for natural language > *processing*; but NLP is not the same as NL*U* ? when it comes to > *understanding*, their worth is still an open question. Perhaps they will > turn out to be necessary; they clearly aren?t sufficient. In their extreme, > they might even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a specific set of > numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > > > (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > > Notably absent from your email is any kind of apology for misrepresenting > my position. It?s fine to say that ?many people thirty years ago once > thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. > I have consistently felt throughout our interactions that you have mistaken > me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me > for having made that error. I am still not he. > > > > Which maybe connects to the last point; if you read my work, you would see > thirty years of arguments *for* neural networks, just not in the way that > you want them to exist. I have ALWAYS argued that there is a role for them; > characterizing me as a person ?strongly opposed to neural networks? misses > the whole point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > > > In the last two decades or so you have insisted (for reasons you have > never fully clarified, so far as I know) on abandoning symbol-manipulation, > but the reverse is not the case: I have *never* called for dismissal of > neural networks, but rather for some hybrid between the two (as you > yourself contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded and broke down, > and where symbols could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > > > Gary > > > > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: > > ? > > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > > > > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > > > Geoff > > > > > > > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural > network community, > > > > There has been a lot of recent discussion on this list about framing and > scientific integrity. Often the first step in restructuring narratives is > to bully and dehumanize critics. The second is to misrepresent their > position. People in positions of power are sometimes tempted to do this. > > > > The Hinton-Hanson interview that you just published is a real-time example > of just that. It opens with a needless and largely content-free personal > attack on a single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive thing it says is > false. > > > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t > be able to do machine translation.? > > > > I never said any such thing. > > > > What I predicted, rather, was that multilayer perceptrons, as they existed > then, would not (on their own, absent other mechanisms) *understand* language. > Seven years later, they still haven?t, except in the most superficial way. > > > > > I made no comment whatsoever about machine translation, which I view as a > separate problem, solvable to a certain degree by correspondance without > semantics. > > > > I specifically tried to clarify Hinton?s confusion in 2019, but, > disappointingly, he has continued to purvey misinformation despite that > clarification. Here is what I wrote privately to him then, which should > have put the matter to rest: > > > > You have taken a single out of context quote [from 2015] and > misrepresented it. The quote, which you have prominently displayed at the > bottom on your own web page, says: > > > > Hierarchies of features are less suited to challenges such as language, > inference, and high-level planning. For example, as Noam Chomsky famously > pointed out, language is filled with sentences you haven't seen > before. Pure classifier systems don't know what to do with such sentences. > The talent of feature detectors -- in identifying which member of some > category something belongs to -- doesn't translate into understanding > novel sentences, in which each sentence has its own unique meaning. > > > > It does *not* say "neural nets would not be able to deal with novel > sentences"; it says that hierachies of features detectors (on their own, if > you read the context of the essay) would have trouble *understanding *novel sentences. > > > > > Google Translate does yet not *understand* the content of the sentences > is translates. It cannot reliably answer questions about who did what to > whom, or why, it cannot infer the order of the events in paragraphs, it > can't determine the internal consistency of those events, and so forth. > > > > Since then, a number of scholars, such as the the computational linguist > Emily Bender, have made similar points, and indeed current LLM difficulties > with misinformation, incoherence and fabrication all follow from these > concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter > with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf > , > also emphasizing issues of understanding and meaning: > > > > *The success of the large neural language models on many NLP tasks is > exciting. However, we find that these successes sometimes lead to hype in > which these models are being described as ?understanding? language or > capturing ?meaning?. In this position paper, we argue that a system trained > only on form has a priori no way to learn meaning. .. a clear understanding > of the distinction between form and meaning will help guide the field > towards better science around natural language understanding. * > > > > Her later article with Gebru on language models ?stochastic parrots? is in > some ways an extension of this point; machine translation requires mimicry, > true understanding (which is what I was discussing in 2015) requires > something deeper than that. > > > > Hinton?s intellectual error here is in equating machine translation with > the deeper comprehension that robust natural language understanding will > require; as Bender and Koller observed, the two appear not to be the same. > (There is a longer discussion of the relation between language > understanding and machine translation, and why the latter has turned out to > be more approachable than the former, in my 2019 book with Ernest Davis). > > > > More broadly, Hinton?s ongoing dismissiveness of research from > perspectives other than his own (e.g. linguistics) have done the field a > disservice. > > > > As Herb Simon once observed, science does not have to be zero-sum. > > > > Sincerely, > > Gary Marcus > > Professor Emeritus > > New York University > > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > > Stephen Hanson in conversation with Geoff Hinton > > > > In the latest episode of this video series for AIhub.org > , > Stephen Hanson talks to Geoff Hinton about neural networks, > backpropagation, overparameterization, digit recognition, voxel cells, > syntax and semantics, Winograd sentences, and more. > > > > You can watch the discussion, and read the transcript, here: > > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > > > About AIhub: > > AIhub is a non-profit dedicated to connecting the AI community to the > public by providing free, high-quality information through AIhub.org > > (https://aihub.org/ > ). > We help researchers publish the latest AI news, summaries of their work, > opinion pieces, tutorials and more. We are supported by many leading > scientific organizations in AI, namely AAAI > , > NeurIPS > , > ICML > , > AIJ > > /IJCAI > , > ACM SIGAI > , > EurAI/AICOMM, CLAIRE > > and RoboCup > > . > > Twitter: @aihuborg > > > > > > > Virus-free. www.avast.com > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Fri Feb 4 07:00:45 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Fri, 4 Feb 2022 07:00:45 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: <6f8f9f7f-053d-a633-805d-bacd28967edf@rubic.rutgers.edu> Tom, understanding is a theorem? you mean it should be a theorem? and yes, if you are having brain surgery.. you hope your surgeon, "understands" what they are doing.. Steve On 2/3/22 12:31 PM, Dietterich, Thomas wrote: > > ?Understanding? is not a Boolean. It is a theorem that no system can > enumerate all of the consequences of a state of affairs in the world. > > For low-stakes application work, we can be satisfied by a system that > ?does the right thing?. If the system draws a good picture, that?s > sufficient. It ?understood? the request. > > But for higher-stakes applications---and for advancing the > science---we seek a causal account of how the components of a system > cause it to do the right thing. We are hoping that a small set of > mechanisms can produce broad coverage of intelligent behavior. This > gives us confidence that the system will respond correctly outside of > the narrow tasks on which we have tested it. > > --Tom > > Thomas G. Dietterich, Distinguished Professor Emeritus > > School of Electrical Engineering and Computer Science > > US Mail: 1148 Kelley Engineering Center > > Office: 2067 Kelley Engineering Center > > Oregon State Univ., Corvallis, OR 97331-5501 > > Voice: 541-737-5559; FAX: 541-737-1300 > > URL: http://web.engr.oregonstate.edu/~tgd/ > > > *From:* Connectionists > *On Behalf Of *Gary Marcus > *Sent:* Thursday, February 3, 2022 8:26 AM > *To:* Danko Nikolic > *Cc:* connectionists at mailman.srv.cs.cmu.edu; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with > Geoff Hinton > > [This email originated from outside of OSU. Use caution with links and > attachments.] > > Dear Danko, > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED > talk, in which he said (paraphrasing from memory, because I don?t > remember the precise words) that the famous 200 Quoc Le unsupervised > model > [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf > ] > had learned the concept of a ca. In reality the model had clustered > together some catlike images based on the image statistics that it had > extracted, but it was a long way from a full, > counterfactual-supporting concept of a cat, much as you describe below. > > I fully agree with you that the reason for even having a semantics is > as you put it, "to 1) learn with a few examples and 2) apply the > knowledge to a broad set of situations.? GPT-3 sometimes gives the > appearance of having done so, but it falls apart under close > inspection, so the problem remains unsolved. > > Gary > > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: > > G. Hinton wrote: "I believe that any reasonable person would admit > that if you ask a neural net to draw a picture of a hamster > wearing a red hat and it draws such a picture, it understood?the > request." > > I would like to suggest why drawing a?hamster with a red?hat does > not necessarily imply understanding of the statement "hamster > wearing a red hat". > > To understand that "hamster wearing a red hat" would mean > inferring, in newly?emerging situations of this hamster, all the > real-life implications?that the red hat brings to the little animal. > > What would happen to the hat if the hamster rolls on its back? > (Would the hat fall off?) > > What would happen to the red hat when the hamster enters its lair? > (Would the hat fall?off?) > > What would happen to that hamster when it goes foraging? (Would > the red hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? > (Would it be easier for predators to spot the hamster?) > > ...and so on. > > Countless many questions can be asked. One has understood "hamster > wearing a red hat" only if one can answer reasonably well many of > such real-life relevant questions. Similarly, a student > has?understood materias in a class only if they can apply the > materials in real-life situations (e.g., applying Pythagora's > theorem). If a student gives a correct answer to a multiple?choice > question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > I also suggest that understanding also comes together with > effective learning: We store new information in such a way that we > can recall it later and use it effectively? i.e., make good > inferences in newly emerging situations based on this knowledge. > > In short: Understanding makes us humans able to 1) learn with a > few examples and 2) apply the knowledge to a broad set of situations. > > No neural network?today has such capabilities and we don't know > how to give them such capabilities. Neural networks need large > amounts of training?examples that cover a large variety of > situations and then the?networks can only deal with what the > training examples have already covered. Neural networks cannot > extrapolate in that 'understanding' sense. > > I suggest that understanding truly extrapolates from a piece of > knowledge. It is not about satisfying a task such as translation > between languages or drawing hamsters with hats. It is how you got > the capability to complete the task: Did you only have a few > examples that covered something different but related and then you > extrapolated from that knowledge? If yes, this is going in the > direction of understanding. Have you seen countless examples and > then interpolated among them? Then perhaps it is not understanding. > > So, for the case of drawing a hamster wearing a red hat, > understanding perhaps would have taken place if the following > happened before that: > > 1) first, the network learned about hamsters (not many examples) > > 2) after that the network learned about red hats (outside the > context of hamsters and without many examples) > > 3) finally the network learned about drawing (outside of the > context of hats and hamsters, not many examples) > > After that, the network is asked to draw a hamster with a red hat. > If it does it successfully, maybe we have started cracking the > problem of understanding. > > Note also that this requires the network to learn sequentially > without exhibiting catastrophic forgetting of the previous > knowledge, which is possibly also a consequence of human learning > by understanding. > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > > --- A progress usually starts with an insight --- > > > > > > Virus-free. www.avast.com > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: > > Without getting into the specific dispute between Gary and > Geoff, I think with approaches similar to GLOM, we are finally > headed in the right direction. There?s plenty of > neurophysiological evidence for single-cell abstractions and > multisensory neurons in the brain, which one might claim > correspond to symbols. And I think we can finally reconcile > the decades old dispute between Symbolic AI and Connectionism. > > GARY: (Your GLOM, which as you know I praised publicly, is in > many ways an effort to wind up with encodings that effectively > serve as symbols in exactly that way, guaranteed to serve as > consistent representations of specific concepts.) > > GARY: I have /never/ called for dismissal of neural networks, > but rather for some hybrid between the two (as you yourself > contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded > and broke down, and where symbols could complement them. > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > *From:* Connectionists > > *On > Behalf Of *Gary Marcus > *Sent:* Wednesday, February 2, 2022 1:26 PM > *To:* Geoffrey Hinton > > *Cc:* AIhub >; > connectionists at mailman.srv.cs.cmu.edu > > *Subject:* Re: Connectionists: Stephen Hanson in conversation > with Geoff Hinton > > Dear Geoff, and interested others, > > What, for example, would you make of a system that often?drew > the red-hatted hamster you requested, and perhaps a fifth of > the time gave you utter nonsense?? Or say one that you trained > to create birds but sometimes output stuff like this: > > > > One could > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > or > > b. wonder if it might be possible that sometimes the system > gets the right answer for the wrong reasons (eg partial > historical contingency), and wonder whether another approach > might be indicated. > > Benchmarks are harder than they look; most of the field has > come to recognize that. The Turing Test has turned out to be a > lousy measure of intelligence, easily gamed. It has turned out > empirically that the Winograd Schema Challenge did not measure > common sense as well as Hector might have thought. (As it > happens, I am a minor coauthor of a very recent review on this > very topic: https://arxiv.org/abs/2201.02387 > ) > But its conquest in no way means machines now have common > sense; many people from many different perspectives recognize > that (including, e.g., Yann LeCun, who generally tends to be > more aligned with you than with me). > > So: on the goalpost of the Winograd schema, I was wrong, and > you can quote me; but what you said about me and machine > translation remains your invention, and it is inexcusable that > you simply ignored my 2019 clarification. On the essential > goal of trying to reach meaning and understanding, I remain > unmoved; the problem remains unsolved. > > All of the problems LLMs have with coherence, reliability, > truthfulness, misinformation, etc stand witness to that fact. > (Their persistent inability to filter out toxic and insulting > remarks stems from the same.) I am hardly the only person in > the field to see that progress on any given benchmark does not > inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > With respect to embeddings: Embeddings are very good for > natural language /processing/; but NLP is not the same as > NL/U/ ? when it comes to /understanding/, their worth is still > an open question. Perhaps they will turn out to be necessary; > they clearly aren?t sufficient. In their extreme, they might > even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a > specific set of numbers stands for a specific word or concept. > (Wouldn?t that be ironic?) > > (Your GLOM, which as you know I praised publicly, is in many > ways an effort to wind up with encodings that effectively > serve as symbols in exactly that way, guaranteed to serve as > consistent representations of specific concepts.) > > Notably absent from your email is any kind of apology for > misrepresenting my position. It?s fine to say that ?many > people thirty years ago once thought X? and another to say > ?Gary Marcus said X in 2015?, when I didn?t. I have > consistently felt throughout our interactions that you have > mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS > 2014) apologized to me for having made that error. I am still > not he. > > Which maybe connects to the last point; if you read my work, > you would see thirty years of arguments /for/?neural networks, > just not in the way that you want them to exist. I have ALWAYS > argued that there is a role for them; ?characterizing me as a > person ?strongly?opposed to neural networks? misses the whole > point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > In the last two decades or so you have insisted (for reasons > you have never fully clarified, so far as I know) on > abandoning symbol-manipulation, but the reverse is not the > case: I have /never/ called for dismissal of neural networks, > but rather for some hybrid between the two (as you yourself > contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded > and broke down, and where symbols could complement them. It?s > a rhetorical trick (which is what the previous thread was > about) to pretend otherwise. > > Gary > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > > wrote: > > ? > > Embeddings are just vectors of soft feature detectors and > they are very good for NLP.? The quote on my webpage from > Gary's 2015 chapter implies the opposite. > > A few decades ago, everyone I knew then would?have agreed > that the ability to translate a sentence into many > different languages was strong evidence that you > understood it. > > But once neural networks could do that, their critics > moved the goalposts. An exception is Hector Levesque who > defined the goalposts more sharply by saying that the > ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at > that but still have some way to go. Will Gary agree that > when they can get pronoun references?correct in Winograd > sentences they really?do understand? Or does he want to > reserve the right to weasel out of that too? > > Some people, like Gary, appear to be strongly?opposed to > neural networks because?they do not fit their preconceived > notions of how the mind should work. > > I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster > wearing a red hat and it draws such a picture, it > understood?the request. > > Geoff > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > > wrote: > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, > and the larger neural network community, > > There has been a lot of recent discussion on this list > about framing and scientific integrity. Often the > first step in restructuring narratives is to bully and > dehumanize critics. The second is to misrepresent > their position. People in positions of power are > sometimes tempted to do this. > > The Hinton-Hanson interview that you just published is > a real-time example of just that. It opens with a > needless and largely content-free personal attack on a > single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive > thing it says is false. > > Hinton says ?In 2015 he [Marcus] made a prediction > that computers wouldn?t be able to do machine > translation.? > > I never said any such thing. > > What I predicted, rather, was that multilayer > perceptrons, as they existed then, would not (on their > own, absent other mechanisms) /understand/?language. > Seven years later, they still haven?t, except in the > most superficial way. > > I made no comment whatsoever about machine > translation, which I view as a separate problem, > solvable to a certain degree by correspondance without > semantics. > > I specifically tried to clarify Hinton?s confusion in > 2019, but, disappointingly, he has continued to purvey > misinformation despite that clarification. Here is > what I wrote privately to him then, which should have > put the matter to rest: > > You have taken a single out of context quote [from > 2015] and misrepresented it. The quote, which you have > prominently displayed at the bottom on your own web > page, says: > > Hierarchies of features are less suited to challenges > such as language, inference, and high-level planning. > For example, as Noam Chomsky famously pointed out, > language is filled with sentences you haven't seen > before.?Pure classifier systems don't know what to do > with such sentences. The talent of feature detectors > -- in??identifying which member of some category > something belongs to -- doesn't translate into > understanding novel??sentences, in which each sentence > has its own unique meaning. > > It does /not/?say "neural nets would not be able to > deal with novel sentences"; it says that hierachies of > features detectors (on their own, if you read the > context of the essay) would have trouble > /understanding /novel?sentences. > > Google Translate does yet not /understand/?the content > of?the sentences is translates. It cannot reliably > answer questions about who did what to whom, or why, > it cannot infer the order of the events in paragraphs, > it can't determine the internal consistency of those > events, and so forth. > > Since then, a number of scholars, such as the the > computational linguist Emily Bender, have made similar > points, and indeed current LLM difficulties with > misinformation, incoherence and fabrication all follow > from these concerns. Quoting from Bender?s > prizewinning 2020 ACL article on the matter with > Alexander Koller, > https://aclanthology.org/2020.acl-main.463.pdf > , > also emphasizing issues of understanding and meaning: > > /The success of the large neural language models on > many NLP tasks is exciting. However, we find that > these successes sometimes lead to hype in which these > models are being described as ?understanding? language > or capturing ?meaning?. In this position paper, we > argue that a system trained only on form has a priori > no way to learn meaning. .. a clear understanding of > the distinction between form and meaning will help > guide the field towards better science around natural > language understanding. / > > Her later article with Gebru on language models > ?stochastic parrots? is in some ways an extension of > this point; machine translation requires mimicry, true > understanding (which is what I was discussing in 2015) > requires something deeper than that. > > Hinton?s intellectual error here is in equating > machine translation with the deeper comprehension that > robust natural language understanding will require; as > Bender and Koller observed, the two appear not to be > the same. (There is a longer discussion of the > relation between language understanding and machine > translation, and why the latter has turned out to be > more approachable than the former, in my 2019 book > with Ernest Davis). > > More broadly, Hinton?s ongoing dismissiveness of > research from perspectives other than his own (e.g. > linguistics) have done the field a disservice. > > As Herb Simon once observed, science does not have to > be zero-sum. > > Sincerely, > > Gary Marcus > > Professor Emeritus > > New York University > > On Feb 2, 2022, at 06:12, AIhub > > > wrote: > > ? > > Stephen Hanson in conversation with Geoff Hinton > > In the latest episode of this video series for > AIhub.org > , > Stephen Hanson talks to? Geoff Hinton?about neural > networks, backpropagation, overparameterization, > digit recognition, voxel cells, syntax and > semantics, Winograd sentences, and more. > > You can watch the discussion, and read the > transcript, here: > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > About AIhub: > > AIhub is a non-profit dedicated to connecting the > AI community to the public by providing free, > high-quality information through AIhub.org > > (https://aihub.org/ > ). > We help researchers publish the latest AI news, > summaries of their work, opinion pieces, tutorials > and more.? We are supported by many leading > scientific organizations in AI, namely AAAI > , > NeurIPS > , > ICML > , > AIJ > /IJCAI > , > ACM SIGAI > , > EurAI/AICOMM, CLAIRE > > and RoboCup > . > > Twitter: @aihuborg > > > > > > Virus-free. www.avast.com > > > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From barak at cs.nuim.ie Fri Feb 4 07:45:42 2022 From: barak at cs.nuim.ie (Barak A. Pearlmutter) Date: Fri, 4 Feb 2022 12:45:42 +0000 Subject: Connectionists: Postdoc and PhD Student Positions at Intersection of Neural Networks and Programming Language Theory Message-ID: To take a short break from castigating other scientists and reciting the Gospel of Materialism (F=ma implies we are mechanisms, amen), I'd like to mention some openings in our group. PhD student and postdoctoral openings on the MAIVV project https://cs.nuim.ie/research/pop/maivv.html at the interface of programming language theory and machine learning/neural networks. See https://cs.nuim.ie/research/pop/#Hiring for details and contact info. --Barak Pearlmutter, Dept of Computer Science, Maynooth University. From r.jolivet at ucl.ac.uk Fri Feb 4 08:15:19 2022 From: r.jolivet at ucl.ac.uk (Jolivet, Renaud) Date: Fri, 4 Feb 2022 13:15:19 +0000 Subject: Connectionists: {SPAM?} Re: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: <76ADFF55-D9A5-4B89-AFD1-E71A74B40D50@ucl.ac.uk> > However, there still is a problem: Neural network's can't implement the needed simulations; they cannot achieve the same cognitive effect that human minds can (or animal minds can). 80% of the cells in the human cortex are not neurons, but glial cells (incl. vasculature). Neuroscience first has to become less neurocentric. Cheers, Renaud From jose at rubic.rutgers.edu Fri Feb 4 07:17:27 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Fri, 4 Feb 2022 07:17:27 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <9b988b04-250b-56c8-c3ec-9509bba4bdc8@rubic.rutgers.edu> <8EA19260-6F0C-421D-8AF3-55E74C2FC3E9@nyu.edu> Message-ID: Geoff's position is pretty clear.?? He said in the conversation we had and in this thread,? "vectors of soft features", Some of my claim is in several of the conversations with Mike Jordan and Rich Sutton, but briefly, there are a number of very large costly efforts from the 1970s and 1980s, to create, deploy and curate symbol AI systems that were massive failures. Not counterfactuals,? but factuals that failed.?? The MCC comes to mind with Adm Bobby Inmann's? national US mandate to counter the Japanese so called"Fifth-generation AI systems"? as a massive failure of symbolic AI. -------------------- In 1982, Japan launched its Fifth Generation Computer Systems project (FGCS), designed to develop intelligent software that would run on novel computer hardware. As the first national, large-scale artificial intelligence (AI) research and development (R&D) project to be free from military influence and corporate profit motives, the FGCS was open, international, and oriented around public goods. On 2/3/22 6:34 PM, Francesca Rossi2 wrote: > Hi all. > > Thanks Gary for adding me to this thread. > > I also would be interested in knowing why Steve thinks that NS AI did not work in the past, and why this is an indication that it cannot work now or in the future. > > Thanks, > Francesca. > ------------------ > > Francesca Rossi > IBM Fellow and AI Ethics Global Leader > T.J. Watson Research Center, Yorktown Heights, USA > +1-617-3869639 > > ________________________________________ > From: Artur Garcez > Sent: Thursday, February 3, 2022 6:00 PM > To: Gary Marcus > Cc: Stephen Jos? Hanson; Geoffrey Hinton; AIhub; connectionists at mailman.srv.cs.cmu.edu; Luis Lamb; Josh Tenenbaum; Anima Anandkumar; Francesca Rossi2; Swarat Chaudhuri; Gadi Singer > Subject: [EXTERNAL] Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally ZjQcmQRYFpfptBannerStart > This Message Is From an External Sender > This message came from outside your organization. > ZjQcmQRYFpfptBannerEnd > > It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. > > Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally symbolic), there is a clear resurgence of neuro-symbolic approaches. > > Best wishes, > Artur > > > On Thu, Feb 3, 2022 at 5:00 PM Gary Marcus > wrote: > Steve, > > I?d love to hear you elaborate on this part, > > Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. > > > I?d love your take on why you think the impetus towards hybrid models ended badly before, and why you think that the mistakes of the past can?t be corrected. Also it? would be really instructive to compare with deep learning, which lost steam for quite some time, but reemerged much stronger than ever before. Might not the same happen with hybrid models? > > I am cc?ing some folks (possibly not on this list) who have recently been sympathetic to hybrid models, in hopes of a rich discussion. (And, Geoff, still cc?d, I?d genuinely welcome your thoughts if you want to add them, despite our recent friction.) > > Cheers, > Gary > > > On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson > wrote: > > > I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. > > Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. > > Best and hope you are doing well. > > Steve > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From jose at rubic.rutgers.edu Fri Feb 4 08:53:00 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Fri, 4 Feb 2022 08:53:00 -0500 Subject: Connectionists: {SPAM?} Re: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <76ADFF55-D9A5-4B89-AFD1-E71A74B40D50@ucl.ac.uk> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <76ADFF55-D9A5-4B89-AFD1-E71A74B40D50@ucl.ac.uk> Message-ID: <9b7da1fc-b2fe-73c0-bdee-c9d088023006@rubic.rutgers.edu> Good point. fortunately, there are some folks at Rutgers that has been working this very issue for? quite a while: **Konstantinos Michmizos and NIH just announced this new program this morning. New Funding Opportunity:***Neuro-Glia Computational Mechanisms Governing Complex Behaviors *_https://grants.nih.gov/grants/guide/notice-files/NOT-MH-22-090.html_ Cheers Steve On 2/4/22 8:15 AM, Jolivet, Renaud wrote: >> However, there still is a problem: Neural network's can't implement the needed simulations; they cannot achieve the same cognitive effect that human minds can (or animal minds can). > 80% of the cells in the human cortex are not neurons, but glial cells (incl. vasculature). > > Neuroscience first has to become less neurocentric. > > Cheers, > Renaud -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From krallinger.martin at gmail.com Fri Feb 4 08:45:30 2022 From: krallinger.martin at gmail.com (Martin Krallinger) Date: Fri, 4 Feb 2022 14:45:30 +0100 Subject: Connectionists: Text Mining research position at my group (Barcelona Supercomputing Center, Text Mining Unit) In-Reply-To: References: Message-ID: We are looking for a motivated Text Mining researcher at my group, please share with potential candidates or apply if you are interested in the position. Reference: 16_22_LS_TM_RE2 Job: Text Mining research position at Barcelona Supercomputing Center URL: https://www.bsc.es/join-us/job-opportunities/1622lstmre2 Context And Mission The Text Mining Unit of the BSC-CNS (TeMU-BSC) is funded through the Plan de Impulso de las Tecnolog?as del Lenguaje de la Agenda Digital, by the Secretary of State of Telecommunications and the Information. It is the first publicly funded text mining unit in Spain and has the aim to promote the development of natural language processing and machine translation resources for Spanish and other co-official languages in the area of biomedicine. We search for a technical researcher in bioinformatics, medical informatics, natural language processing, computational linguistics or language engineering with a strong background in Machine Learning (ML), who will be responsible of the integration of data and the implementation of search tools into the platform for natural language processing designed by the unit. Key Duties Design, implementation, and evaluation of text mining, NLP, deep learning and ML tools and models applied to the clinical and biomedical application domain. Coordination and organization of shared tasks and evaluation campaigns (like IberLEF, CLEF eHealth,..). Technical coordination and supervision of Gold Standard annotation projects to generate high-quality text corpora. Requirements Education: University degree in computer science, mathematics, statistics, data scientist, physics, bioinformatics, telecommunications, electrical engineering or equivalent. Essential Knowledge and Professional Experience: Experience with Deep Learning and statistical data mining frameworks: Keras, Tensorflow, PySpark, PyTorch, Spacy, etc. Experience with ML algorithms and techniques: LDA, Topic Modelling, LSTM, KNN, SVM, Decision Trees, Clustering, Word Embeddings, etc. Experience in the development or management of software resources/tools, Github + github projects. Experience with NLP components and platforms. Additional Knowledge and Professional Experience: Strong programming skills in at least one of the following languages: Java, Python, C++, Scala, R. Experience and skills related to bash, Docker, Kubernetes, Unity testing, Collab Competences: Good communication and presentation skills. Strong technical writing skills. Ability to work both independently and within a team. Used to work under pressure under strict deadlines -- ======================================= Martin Krallinger, Dr. Head of Biological Text Mining Unit Barcelona Supercomputing Center (BSC-CNS) ======================================= -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Fri Feb 4 10:04:36 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Fri, 4 Feb 2022 10:04:36 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <1A52CB03-212F-446F-95A5-EDE3A18C614A@nyu.edu> References: <1A52CB03-212F-446F-95A5-EDE3A18C614A@nyu.edu> Message-ID: <096679d9-37ca-0f01-2ba8-007c95e34d91@rubic.rutgers.edu> Well I don't like counterfactual arguments or ones that start with "It can't be done with neural networks.."--as this amounts to the old Rumelhart saw, of "proof by lack of imagination". I think my position and others (I can't speak for Geoff and won't) is more of a "purist" view that brains have computationally complete representational power to do what ever is required of human level mental processing.? AI symbol systems are remote descriptions of this level of processing.???? Looking at 1000s of brain scans, one begins to see a pattern of interacting large and smaller scale networks, probably related to Resting state and the Default Mode networks in some important competitive way.?? But what one doesn't find is modular structure (e.g. face area.. nope)? or evidence of "symbols" being processed.??? Research on Numbers is interesting in this regard, as number representation should provide some evidence of? discrete symbol processing as would? letters.?? But again the processing states from brain imaging? more generally appear to be distributed representations of some sort. One other direction has to do with prior rules that could be neurally coded and therefore provide an immediate bias in learning and thus dramatically reduce the number of examples required for? asymptotic learning.???? Some of this has been done with pre-training-- on let's say 1000s of videos that are relatively generic, prior to learning on a small set of videos related to a specific topic-- say two individuals playing a monopoly game.? In that case, no game-like videos were sampled in the pre-training, and the LSTM was trained to detect change point on 2 minutes of video, achieving a 97% match with human parsers.??? In these senses I have no problem with this type of hybrid training. Steve On 2/4/22 9:07 AM, Gary Marcus wrote: > ?The whole point of the neurosymbolic approach is to develop systems > that can accommodate both vectors and symbols, since neither on their > own seems adequate. > > If there are arguments against trying to do that, we would be interested. > >> On Feb 4, 2022, at 4:17 AM, Stephen Jos? Hanson >> wrote: >> >> ? >> >> Geoff's position is pretty clear.?? He said in the conversation we >> had and in this thread, "vectors of soft features", >> >> Some of my claim is in several of the conversations with Mike Jordan >> and Rich Sutton, but briefly,? there are a number of >> very large costly efforts from the 1970s and 1980s, to create, deploy >> and curate symbol AI systems that were massive failures.? Not >> counterfactuals,? but factuals that failed.?? The MCC comes to mind >> with Adm Bobby Inmann's? national US mandate to counter the Japanese >> so called"Fifth-generation AI systems"? as a massive failure of >> symbolic AI. >> >> -------------------- >> >> In 1982, Japan launched its Fifth Generation Computer Systems project >> (FGCS), designed to develop intelligent software that would run on >> novel computer hardware. As the first national, large-scale >> artificial intelligence (AI) research and development (R&D) project >> to be free from military influence and corporate profit motives, the >> FGCS was open, international, and oriented around public goods. >> >> On 2/3/22 6:34 PM, Francesca Rossi2 wrote: >>> Hi all. >>> >>> Thanks Gary for adding me to this thread. >>> >>> I also would be interested in knowing why Steve thinks that NS AI did not work in the past, and why this is an indication that it cannot work now or in the future. >>> >>> Thanks, >>> Francesca. >>> ------------------ >>> >>> Francesca Rossi >>> IBM Fellow and AI Ethics Global Leader >>> T.J. Watson Research Center, Yorktown Heights, USA >>> +1-617-3869639 >>> >>> ________________________________________ >>> From: Artur Garcez >>> Sent: Thursday, February 3, 2022 6:00 PM >>> To: Gary Marcus >>> Cc: Stephen Jos? Hanson; Geoffrey Hinton; AIhub;connectionists at mailman.srv.cs.cmu.edu; Luis Lamb; Josh Tenenbaum; Anima Anandkumar; Francesca Rossi2; Swarat Chaudhuri; Gadi Singer >>> Subject: [EXTERNAL] Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally ZjQcmQRYFpfptBannerStart >>> This Message Is From an External Sender >>> This message came from outside your organization. >>> ZjQcmQRYFpfptBannerEnd >>> >>> It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. >>> >>> Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally symbolic), there is a clear resurgence of neuro-symbolic approaches. >>> >>> Best wishes, >>> Artur >>> >>> >>> On Thu, Feb 3, 2022 at 5:00 PM Gary Marcus > wrote: >>> Steve, >>> >>> I?d love to hear you elaborate on this part, >>> >>> Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. >>> >>> >>> I?d love your take on why you think the impetus towards hybrid models ended badly before, and why you think that the mistakes of the past can?t be corrected. Also it? would be really instructive to compare with deep learning, which lost steam for quite some time, but reemerged much stronger than ever before. Might not the same happen with hybrid models? >>> >>> I am cc?ing some folks (possibly not on this list) who have recently been sympathetic to hybrid models, in hopes of a rich discussion. (And, Geoff, still cc?d, I?d genuinely welcome your thoughts if you want to add them, despite our recent friction.) >>> >>> Cheers, >>> Gary >>> >>> >>> On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson > wrote: >>> >>> >>> I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. >>> >>> Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. >>> >>> Best and hope you are doing well. >>> >>> Steve >>> >> -- >> -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From gary.marcus at nyu.edu Fri Feb 4 09:07:14 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Fri, 4 Feb 2022 06:07:14 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: <1A52CB03-212F-446F-95A5-EDE3A18C614A@nyu.edu> ?The whole point of the neurosymbolic approach is to develop systems that can accommodate both vectors and symbols, since neither on their own seems adequate. If there are arguments against trying to do that, we would be interested. > On Feb 4, 2022, at 4:17 AM, Stephen Jos? Hanson wrote: > ? > Geoff's position is pretty clear. He said in the conversation we had and in this thread, "vectors of soft features", > > Some of my claim is in several of the conversations with Mike Jordan and Rich Sutton, but briefly, there are a number of > very large costly efforts from the 1970s and 1980s, to create, deploy and curate symbol AI systems that were massive failures. Not counterfactuals, but factuals that failed. The MCC comes to mind with Adm Bobby Inmann's national US mandate to counter the Japanese so called"Fifth-generation AI systems" as a massive failure of symbolic AI. > > -------------------- > > In 1982, Japan launched its Fifth Generation Computer Systems project (FGCS), designed to develop intelligent software that would run on novel computer hardware. As the first national, large-scale artificial intelligence (AI) research and development (R&D) project to be free from military influence and corporate profit motives, the FGCS was open, international, and oriented around public goods. > > On 2/3/22 6:34 PM, Francesca Rossi2 wrote: >> Hi all. >> >> Thanks Gary for adding me to this thread. >> >> I also would be interested in knowing why Steve thinks that NS AI did not work in the past, and why this is an indication that it cannot work now or in the future. >> >> Thanks, >> Francesca. >> ------------------ >> >> Francesca Rossi >> IBM Fellow and AI Ethics Global Leader >> T.J. Watson Research Center, Yorktown Heights, USA >> +1-617-3869639 >> >> ________________________________________ >> From: Artur Garcez >> Sent: Thursday, February 3, 2022 6:00 PM >> To: Gary Marcus >> Cc: Stephen Jos? Hanson; Geoffrey Hinton; AIhub; connectionists at mailman.srv.cs.cmu.edu; Luis Lamb; Josh Tenenbaum; Anima Anandkumar; Francesca Rossi2; Swarat Chaudhuri; Gadi Singer >> Subject: [EXTERNAL] Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >> >> It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally ZjQcmQRYFpfptBannerStart >> This Message Is From an External Sender >> This message came from outside your organization. >> ZjQcmQRYFpfptBannerEnd >> >> It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. >> >> Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally symbolic), there is a clear resurgence of neuro-symbolic approaches. >> >> Best wishes, >> Artur >> >> >> On Thu, Feb 3, 2022 at 5:00 PM Gary Marcus > wrote: >> Steve, >> >> I?d love to hear you elaborate on this part, >> >> Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. >> >> >> I?d love your take on why you think the impetus towards hybrid models ended badly before, and why you think that the mistakes of the past can?t be corrected. Also it? would be really instructive to compare with deep learning, which lost steam for quite some time, but reemerged much stronger than ever before. Might not the same happen with hybrid models? >> >> I am cc?ing some folks (possibly not on this list) who have recently been sympathetic to hybrid models, in hopes of a rich discussion. (And, Geoff, still cc?d, I?d genuinely welcome your thoughts if you want to add them, despite our recent friction.) >> >> Cheers, >> Gary >> >> >> On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson > wrote: >> >> >> I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. >> >> Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. >> >> Best and hope you are doing well. >> >> Steve >> > -- > -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Fri Feb 4 13:58:20 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Fri, 4 Feb 2022 13:58:20 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <6f8f9f7f-053d-a633-805d-bacd28967edf@rubic.rutgers.edu> Message-ID: <5461fab3-a86f-a37b-ba3b-6cc1d9ac928d@rubic.rutgers.edu> I see, sort of a bound. so "understanding" could be a continuum..??? and further dependent on whom we are interacting with... On 2/4/22 1:37 PM, Dietterich, Thomas wrote: > > I mean that if you only say a system is ?understanding? X if it can > enumerate all of the consequences of X, then you have solved what is > known as the ?Ramification Problem?. And it is easy to show that this > is impossible. Hence, our criteria for saying that a system > ?understandings? must lie somewhere between ?doing the right thing in > this one situation? and knowing all of the consequences of its beliefs. > > --Tom > > Thomas G. Dietterich, Distinguished Professor Voice: 541-737-5559 > > School of Electrical Engineering????????????? FAX: 541-737-1300 > > ? and Computer Science??????????????????????? URL: > eecs.oregonstate.edu/~tgd > > US Mail: 1148 Kelley Engineering Center > > Office: 2067 Kelley Engineering Center > > Oregon State Univ., Corvallis, OR 97331-5501 > > *From:*Stephen Jos? Hanson > *Sent:* Friday, February 4, 2022 4:01 AM > *To:* Dietterich, Thomas ; Gary Marcus > ; Danko Nikolic > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with > Geoff Hinton > > [This email originated from outside of OSU. Use caution with links and > attachments.] > > Tom, understanding is a theorem? > > you mean it should be a theorem? > > and yes, if you are having brain surgery.. you hope your surgeon, > "understands" what they are doing.. > > Steve > > On 2/3/22 12:31 PM, Dietterich, Thomas wrote: > > ?Understanding? is not a Boolean. It is a theorem that no system > can enumerate all of the consequences of a state of affairs in the > world. > > For low-stakes application work, we can be satisfied by a system > that ?does the right thing?. If the system draws a good picture, > that?s sufficient. It ?understood? the request. > > But for higher-stakes applications---and for advancing the > science---we seek a causal account of how the components of a > system cause it to do the right thing. We are hoping that a small > set of mechanisms can produce broad coverage of intelligent > behavior. This gives us confidence that the system will respond > correctly outside of the narrow tasks on which we have tested it. > > --Tom > > Thomas G. Dietterich, Distinguished Professor Emeritus > > School of Electrical Engineering and Computer Science > > US Mail: 1148 Kelley Engineering Center > > Office: 2067 Kelley Engineering Center > > Oregon State Univ., Corvallis, OR 97331-5501 > > Voice: 541-737-5559; FAX: 541-737-1300 > > URL: http://web.engr.oregonstate.edu/~tgd/ > > > *From:* Connectionists > > *On Behalf > Of *Gary Marcus > *Sent:* Thursday, February 3, 2022 8:26 AM > *To:* Danko Nikolic > > *Cc:* connectionists at mailman.srv.cs.cmu.edu > ; AIhub > > *Subject:* Re: Connectionists: Stephen Hanson in conversation with > Geoff Hinton > > [This email originated from outside of OSU. Use caution with links > and attachments.] > > Dear Danko, > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 > TED talk, in which he said (paraphrasing from memory, because I > don?t remember the precise words) that the famous 200 Quoc Le > unsupervised model > [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf > ] > had learned the concept of a ca. In reality the model had > clustered together some catlike images based on the image > statistics that it had extracted, but it was a long way from a > full, counterfactual-supporting concept of a cat, much as you > describe below. > > I fully agree with you that the reason for even having a semantics > is as you put it, "to 1) learn with a few examples and 2) apply > the knowledge to a broad set of situations.? GPT-3 sometimes gives > the appearance of having done so, but it falls apart under close > inspection, so the problem remains unsolved. > > Gary > > > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic > > wrote: > > G. Hinton wrote: "I believe that any reasonable person would > admit that if you ask a neural net to draw a picture of a > hamster wearing a red hat and it draws such a picture, it > understood?the request." > > I would like to suggest why drawing a?hamster with a red?hat > does not necessarily imply understanding of the statement > "hamster wearing a red hat". > > To understand that "hamster wearing a red hat" would mean > inferring, in newly?emerging situations of this hamster, all > the real-life implications?that the red hat brings to the > little animal. > > What would happen to the hat if the hamster rolls on its back? > (Would the hat fall off?) > > What would happen to the red hat when the hamster enters its > lair? (Would the hat fall?off?) > > What would happen to that hamster when it goes foraging? > (Would the red hat have an influence on finding food?) > > What would happen in a situation of being chased by a > predator? (Would it be easier for predators to spot the hamster?) > > ...and so on. > > Countless many questions can be asked. One has understood > "hamster wearing a red hat" only if one can answer reasonably > well many of such real-life relevant questions. Similarly, a > student has?understood materias in a class only if they can > apply the materials in real-life situations (e.g., applying > Pythagora's theorem). If a student gives a correct answer to a > multiple?choice question, we don't know whether the student > understood the material or whether this was just rote learning > (often, it is rote learning). > > I also suggest that understanding also comes together with > effective learning: We store new information in such a way > that we can recall it later and use it effectively i.e., make > good inferences in newly emerging situations based on this > knowledge. > > In short: Understanding makes us humans able to 1) learn with > a few examples and 2) apply the knowledge to a broad set of > situations. > > No neural network?today has such capabilities and we don't > know how to give them such capabilities. Neural networks need > large amounts of training?examples that cover a large variety > of situations and then the?networks can only deal with what > the training examples have already covered. Neural networks > cannot extrapolate in that 'understanding' sense. > > I suggest that understanding truly extrapolates from a piece > of knowledge. It is not about satisfying a task such as > translation between languages or drawing hamsters with hats. > It is how you got the capability to complete the task: Did you > only have a few examples that covered something different but > related and then you extrapolated from that knowledge? If yes, > this is going in the direction of understanding. Have you seen > countless examples and then interpolated among them? Then > perhaps it is not understanding. > > So, for the case of drawing a hamster wearing a red hat, > understanding perhaps would have taken place if the following > happened before that: > > 1) first, the network learned about hamsters (not many examples) > > 2) after that the network learned about red hats (outside the > context of hamsters and without many examples) > > 3) finally the network learned about drawing (outside of the > context of hats and hamsters, not many examples) > > After that, the network is asked to draw a hamster with a red > hat. If it does it successfully, maybe we have started > cracking the problem of understanding. > > Note also that this requires the network to learn sequentially > without exhibiting catastrophic forgetting of the previous > knowledge, which is possibly also a consequence of human > learning by understanding. > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > > --- A progress usually starts with an insight --- > > > > > > Virus-free. www.avast.com > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: > > Without getting into the specific dispute between Gary and > Geoff, I think with approaches similar to GLOM, we are > finally headed in the right direction. There?s plenty of > neurophysiological evidence for single-cell abstractions > and multisensory neurons in the brain, which one might > claim correspond to symbols. And I think we can finally > reconcile the decades old dispute between Symbolic AI and > Connectionism. > > GARY: (Your GLOM, which as you know I praised publicly, is > in many ways an effort to wind up with encodings that > effectively serve as symbols in exactly that way, > guaranteed to serve as consistent representations of > specific concepts.) > > GARY: I have /never/ called for dismissal of neural > networks, but rather for some hybrid between the two (as > you yourself contemplated in 1991); the point of the 2001 > book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols > could complement them. > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > *From:* Connectionists > > > *On Behalf Of *Gary Marcus > *Sent:* Wednesday, February 2, 2022 1:26 PM > *To:* Geoffrey Hinton > > *Cc:* AIhub >; > connectionists at mailman.srv.cs.cmu.edu > > *Subject:* Re: Connectionists: Stephen Hanson in > conversation with Geoff Hinton > > Dear Geoff, and interested others, > > What, for example, would you make of a system that > often?drew the red-hatted hamster you requested, and > perhaps a fifth of the time gave you utter nonsense?? Or > say one that you trained to create birds but sometimes > output stuff like this: > > > > One could > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > or > > b. wonder if it might be possible that sometimes the > system gets the right answer for the wrong reasons (eg > partial historical contingency), and wonder whether > another approach might be indicated. > > Benchmarks are harder than they look; most of the field > has come to recognize that. The Turing Test has turned out > to be a lousy measure of intelligence, easily gamed. It > has turned out empirically that the Winograd Schema > Challenge did not measure common sense as well as Hector > might have thought. (As it happens, I am a minor coauthor > of a very recent review on this very topic: > https://arxiv.org/abs/2201.02387 > ) > But its conquest in no way means machines now have common > sense; many people from many different perspectives > recognize that (including, e.g., Yann LeCun, who generally > tends to be more aligned with you than with me). > > So: on the goalpost of the Winograd schema, I was wrong, > and you can quote me; but what you said about me and > machine translation remains your invention, and it is > inexcusable that you simply ignored my 2019 clarification. > On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains > unsolved. > > All of the problems LLMs have with coherence, reliability, > truthfulness, misinformation, etc stand witness to that > fact. (Their persistent inability to filter out toxic and > insulting remarks stems from the same.) I am hardly the > only person in the field to see that progress on any given > benchmark does not inherently mean that the deep > underlying problems have solved. You, yourself, in fact, > have occasionally made that point. > > With respect to embeddings: Embeddings are very good for > natural language /processing/; but NLP is not the same as > NL/U/ ? when it comes to /understanding/, their worth is > still an open question. Perhaps they will turn out to be > necessary; they clearly aren?t sufficient. In their > extreme, they might even collapse into being symbols, in > the sense of uniquely identifiable encodings, akin to the > ASCII code, in which a specific set of numbers stands for > a specific word or concept. (Wouldn?t that be ironic?) > > (Your GLOM, which as you know I praised publicly, is in > many ways an effort to wind up with encodings that > effectively serve as symbols in exactly that way, > guaranteed to serve as consistent representations of > specific concepts.) > > Notably absent from your email is any kind of apology for > misrepresenting my position. It?s fine to say that ?many > people thirty years ago once thought X? and another to say > ?Gary Marcus said X in 2015?, when I didn?t. I have > consistently felt throughout our interactions that you > have mistaken me for Zenon Pylyshyn; indeed, you once (at > NeurIPS 2014) apologized to me for having made that error. > I am still not he. > > Which maybe connects to the last point; if you read my > work, you would see thirty years of arguments /for/?neural > networks, just not in the way that you want them to exist. > I have ALWAYS argued that there is a role for them; > ?characterizing me as a person ?strongly?opposed to neural > networks? misses the whole point of my 2001 book, which > was subtitled ?Integrating Connectionism and Cognitive > Science.? > > In the last two decades or so you have insisted (for > reasons you have never fully clarified, so far as I know) > on abandoning symbol-manipulation, but the reverse is not > the case: I have /never/ called for dismissal of neural > networks, but rather for some hybrid between the two (as > you yourself contemplated in 1991); the point of the 2001 > book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols > could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > Gary > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > > wrote: > > ? > > Embeddings are just vectors of soft feature detectors > and they are very good for NLP.? The quote on my > webpage from Gary's 2015 chapter implies the opposite. > > A few decades ago, everyone I knew then would?have > agreed that the ability to translate a sentence into > many different languages was strong evidence that you > understood it. > > But once neural networks could do that, their critics > moved the goalposts. An exception is Hector Levesque > who defined the goalposts more sharply by saying that > the ability to get pronoun references correct in > Winograd sentences is a crucial test. Neural nets are > improving at that but still have some way to go. Will > Gary agree that when they can get pronoun > references?correct in Winograd sentences they > really?do understand? Or does he want to reserve the > right to weasel out of that too? > > Some people, like Gary, appear to be strongly?opposed > to neural networks because?they do not fit their > preconceived notions of how the mind should work. > > I believe that any reasonable person would admit that > if you ask a neural net to draw a picture of a hamster > wearing a red hat and it draws such a picture, it > understood?the request. > > Geoff > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > > wrote: > > Dear AI Hub, cc: Steven Hanson and Geoffrey > Hinton, and the larger neural network community, > > There has been a lot of recent discussion on this > list about framing and scientific integrity. Often > the first step in restructuring narratives is to > bully and dehumanize critics. The second is to > misrepresent their position. People in positions > of power are sometimes tempted to do this. > > The Hinton-Hanson interview that you just > published is a real-time example of just that. It > opens with a needless and largely content-free > personal attack on a single scholar (me), with the > explicit intention of discrediting that person. > Worse, the only substantive thing it says is false. > > Hinton says ?In 2015 he [Marcus] made a prediction > that computers wouldn?t be able to do machine > translation.? > > I never said any such thing. > > What I predicted, rather, was that multilayer > perceptrons, as they existed then, would not (on > their own, absent other mechanisms) > /understand/?language. Seven years later, they > still haven?t, except in the most superficial way. > > I made no comment whatsoever about machine > translation, which I view as a separate problem, > solvable to a certain degree by correspondance > without semantics. > > I specifically tried to clarify Hinton?s confusion > in 2019, but, disappointingly, he has continued to > purvey misinformation despite that clarification. > Here is what I wrote privately to him then, which > should have put the matter to rest: > > You have taken a single out of context quote [from > 2015] and misrepresented it. The quote, which you > have prominently displayed at the bottom on your > own web page, says: > > Hierarchies of features are less suited to > challenges such as language, inference, and > high-level planning. For example, as Noam Chomsky > famously pointed out, language is filled with > sentences you haven't seen before.?Pure classifier > systems don't know what to do with such sentences. > The talent of feature detectors -- in??identifying > which member of some category something belongs to > -- doesn't translate into understanding > novel??sentences, in which each sentence has its > own unique meaning. > > It does /not/?say "neural nets would not be able > to deal with novel sentences"; it says that > hierachies of features detectors (on their own, if > you read the context of the essay) would have > trouble /understanding /novel?sentences. > > Google Translate does yet not /understand/?the > content of?the sentences is translates. It cannot > reliably answer questions about who did what to > whom, or why, it cannot infer the order of the > events in paragraphs, it can't determine the > internal consistency of those events, and so forth. > > Since then, a number of scholars, such as the the > computational linguist Emily Bender, have made > similar points, and indeed current LLM > difficulties with misinformation, incoherence and > fabrication all follow from these concerns. > Quoting from Bender?s prizewinning 2020 ACL > article on the matter with Alexander Koller, > https://aclanthology.org/2020.acl-main.463.pdf > , > also emphasizing issues of understanding and meaning: > > /The success of the large neural language models > on many NLP tasks is exciting. However, we find > that these successes sometimes lead to hype in > which these models are being described as > ?understanding? language or capturing ?meaning?. > In this position paper, we argue that a system > trained only on form has a priori no way to learn > meaning. .. a clear understanding of the > distinction between form and meaning will help > guide the field towards better science around > natural language understanding. / > > Her later article with Gebru on language models > ?stochastic parrots? is in some ways an extension > of this point; machine translation requires > mimicry, true understanding (which is what I was > discussing in 2015) requires something deeper than > that. > > Hinton?s intellectual error here is in equating > machine translation with the deeper comprehension > that robust natural language understanding will > require; as Bender and Koller observed, the two > appear not to be the same. (There is a longer > discussion of the relation between language > understanding and machine translation, and why the > latter has turned out to be more approachable than > the former, in my 2019 book with Ernest Davis). > > More broadly, Hinton?s ongoing dismissiveness of > research from perspectives other than his own > (e.g. linguistics) have done the field a disservice. > > As Herb Simon once observed, science does not have > to be zero-sum. > > Sincerely, > > Gary Marcus > > Professor Emeritus > > New York University > > On Feb 2, 2022, at 06:12, AIhub > > wrote: > > ? > > Stephen Hanson in conversation with Geoff Hinton > > In the latest episode of this video series for > AIhub.org > , > Stephen Hanson talks to? Geoff Hinton?about > neural networks, backpropagation, > overparameterization, digit recognition, voxel > cells, syntax and semantics, Winograd > sentences, and more. > > You can watch the discussion, and read the > transcript, here: > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > About AIhub: > > AIhub is a non-profit dedicated to connecting > the AI community to the public by providing > free, high-quality information through > AIhub.org > > (https://aihub.org/ > ). > We help researchers publish the latest AI > news, summaries of their work, opinion pieces, > tutorials and more.? We are supported by many > leading scientific organizations in AI, namely > AAAI > , > NeurIPS > , > ICML > , > AIJ > /IJCAI > , > ACM SIGAI > , > EurAI/AICOMM, CLAIRE > > and RoboCup > . > > Twitter: @aihuborg > > > > > > Virus-free. www.avast.com > > > > -- > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 19957 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From achler at gmail.com Fri Feb 4 11:56:18 2022 From: achler at gmail.com (Tsvi Achler) Date: Fri, 4 Feb 2022 08:56:18 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Gary Marcus and Geoffrey Hinton, It is unfortunate but telling to see heated arguments between both of you, Through some of your past statements, I find both of you are some of the more open minded members of the community. But through your arguments with each other it is clear you are bound by, and buy into the awful nature of academic administration that was established in the middle ages where everyone is dug into their positions, maintaining resources and attending to politics of power. There are new ways to see and address the problems you discuss, but you, like most academics, are caught up in this system. >From your combined silence towards novelty and new ideas, neither of you seem interested in moving the field forward. Sincerely, -Tsvi Tsvi Achler MD/PhD The Science: https://www.youtube.com/watch?v=F-GBIZoZ1mI&list=PL4nMP8F3B7bg3cNWWwLG8BX-wER2PeB-3&index=1 The Academics: https://www.youtube.com/watch?v=dgVDYshmpao&list=PLM3bZImI0fj3rM3ZrzSYbfozkf8m4102j&index=1 Founder Optimizing Mind: https://www.youtube.com/watch?v=xYw0sg1OaAk On Thu, Feb 3, 2022 at 8:30 AM Gary Marcus wrote: > Dear Danko, > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, > in which he said (paraphrasing from memory, because I don?t remember the > precise words) that the famous 200 Quoc Le unsupervised model [ > https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] > had learned the concept of a ca. In reality the model had clustered > together some catlike images based on the image statistics that it had > extracted, but it was a long way from a full, counterfactual-supporting > concept of a cat, much as you describe below. > > I fully agree with you that the reason for even having a semantics is as > you put it, "to 1) learn with a few examples and 2) apply the knowledge to > a broad set of situations.? GPT-3 sometimes gives the appearance of having > done so, but it falls apart under close inspection, so the problem remains > unsolved. > > Gary > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > G. Hinton wrote: "I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster wearing a red hat and > it draws such a picture, it understood the request." > > I would like to suggest why drawing a hamster with a red hat does not > necessarily imply understanding of the statement "hamster wearing a red > hat". > To understand that "hamster wearing a red hat" would mean inferring, in > newly emerging situations of this hamster, all the real-life > implications that the red hat brings to the little animal. > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > ...and so on. > > Countless many questions can be asked. One has understood "hamster wearing > a red hat" only if one can answer reasonably well many of such real-life > relevant questions. Similarly, a student has understood materias in a class > only if they can apply the materials in real-life situations (e.g., > applying Pythagora's theorem). If a student gives a correct answer to a > multiple choice question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > I also suggest that understanding also comes together with effective > learning: We store new information in such a way that we can recall it > later and use it effectively i.e., make good inferences in newly emerging > situations based on this knowledge. > > In short: Understanding makes us humans able to 1) learn with a few > examples and 2) apply the knowledge to a broad set of situations. > > No neural network today has such capabilities and we don't know how to > give them such capabilities. Neural networks need large amounts of > training examples that cover a large variety of situations and then > the networks can only deal with what the training examples have already > covered. Neural networks cannot extrapolate in that 'understanding' sense. > > I suggest that understanding truly extrapolates from a piece of knowledge. > It is not about satisfying a task such as translation between languages or > drawing hamsters with hats. It is how you got the capability to complete > the task: Did you only have a few examples that covered something different > but related and then you extrapolated from that knowledge? If yes, this is > going in the direction of understanding. Have you seen countless examples > and then interpolated among them? Then perhaps it is not understanding. > > So, for the case of drawing a hamster wearing a red hat, understanding > perhaps would have taken place if the following happened before that: > > 1) first, the network learned about hamsters (not many examples) > 2) after that the network learned about red hats (outside the context of > hamsters and without many examples) > 3) finally the network learned about drawing (outside of the context of > hats and hamsters, not many examples) > > After that, the network is asked to draw a hamster with a red hat. If it > does it successfully, maybe we have started cracking the problem of > understanding. > > Note also that this requires the network to learn sequentially without > exhibiting catastrophic forgetting of the previous knowledge, which is > possibly also a consequence of human learning by understanding. > > > Danko > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > --- A progress usually starts with an insight --- > > > > Virus-free. > www.avast.com > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > >> Without getting into the specific dispute between Gary and Geoff, I think >> with approaches similar to GLOM, we are finally headed in the right >> direction. There?s plenty of neurophysiological evidence for single-cell >> abstractions and multisensory neurons in the brain, which one might claim >> correspond to symbols. And I think we can finally reconcile the decades old >> dispute between Symbolic AI and Connectionism. >> >> >> >> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >> an effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> GARY: I have *never* called for dismissal of neural networks, but rather >> for some hybrid between the two (as you yourself contemplated in 1991); the >> point of the 2001 book was to characterize exactly where multilayer >> perceptrons succeeded and broke down, and where symbols could complement >> them. >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> >> *From:* Connectionists *On >> Behalf Of *Gary Marcus >> *Sent:* Wednesday, February 2, 2022 1:26 PM >> *To:* Geoffrey Hinton >> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Geoff, and interested others, >> >> >> >> What, for example, would you make of a system that often drew the >> red-hatted hamster you requested, and perhaps a fifth of the time gave you >> utter nonsense? Or say one that you trained to create birds but sometimes >> output stuff like this: >> >> >> >> >> >> >> >> One could >> >> >> >> a. avert one?s eyes and deem the anomalous outputs irrelevant >> >> or >> >> b. wonder if it might be possible that sometimes the system gets the >> right answer for the wrong reasons (eg partial historical contingency), and >> wonder whether another approach might be indicated. >> >> >> >> Benchmarks are harder than they look; most of the field has come to >> recognize that. The Turing Test has turned out to be a lousy measure of >> intelligence, easily gamed. It has turned out empirically that the Winograd >> Schema Challenge did not measure common sense as well as Hector might have >> thought. (As it happens, I am a minor coauthor of a very recent review on >> this very topic: https://arxiv.org/abs/2201.02387 >> ) >> But its conquest in no way means machines now have common sense; many >> people from many different perspectives recognize that (including, e.g., >> Yann LeCun, who generally tends to be more aligned with you than with me). >> >> >> >> So: on the goalpost of the Winograd schema, I was wrong, and you can >> quote me; but what you said about me and machine translation remains your >> invention, and it is inexcusable that you simply ignored my 2019 >> clarification. On the essential goal of trying to reach meaning and >> understanding, I remain unmoved; the problem remains unsolved. >> >> >> >> All of the problems LLMs have with coherence, reliability, truthfulness, >> misinformation, etc stand witness to that fact. (Their persistent inability >> to filter out toxic and insulting remarks stems from the same.) I am hardly >> the only person in the field to see that progress on any given benchmark >> does not inherently mean that the deep underlying problems have solved. >> You, yourself, in fact, have occasionally made that point. >> >> >> >> With respect to embeddings: Embeddings are very good for natural language >> *processing*; but NLP is not the same as NL*U* ? when it comes to >> *understanding*, their worth is still an open question. Perhaps they >> will turn out to be necessary; they clearly aren?t sufficient. In their >> extreme, they might even collapse into being symbols, in the sense of >> uniquely identifiable encodings, akin to the ASCII code, in which a >> specific set of numbers stands for a specific word or concept. (Wouldn?t >> that be ironic?) >> >> >> >> (Your GLOM, which as you know I praised publicly, is in many ways an >> effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> >> >> Notably absent from your email is any kind of apology for misrepresenting >> my position. It?s fine to say that ?many people thirty years ago once >> thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. >> I have consistently felt throughout our interactions that you have mistaken >> me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me >> for having made that error. I am still not he. >> >> >> >> Which maybe connects to the last point; if you read my work, you would >> see thirty years of arguments *for* neural networks, just not in the way >> that you want them to exist. I have ALWAYS argued that there is a role for >> them; characterizing me as a person ?strongly opposed to neural networks? >> misses the whole point of my 2001 book, which was subtitled ?Integrating >> Connectionism and Cognitive Science.? >> >> >> >> In the last two decades or so you have insisted (for reasons you have >> never fully clarified, so far as I know) on abandoning symbol-manipulation, >> but the reverse is not the case: I have *never* called for dismissal of >> neural networks, but rather for some hybrid between the two (as you >> yourself contemplated in 1991); the point of the 2001 book was to >> characterize exactly where multilayer perceptrons succeeded and broke down, >> and where symbols could complement them. It?s a rhetorical trick (which is >> what the previous thread was about) to pretend otherwise. >> >> >> >> Gary >> >> >> >> >> >> On Feb 2, 2022, at 11:22, Geoffrey Hinton >> wrote: >> >> ? >> >> Embeddings are just vectors of soft feature detectors and they are very >> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >> opposite. >> >> >> >> A few decades ago, everyone I knew then would have agreed that the >> ability to translate a sentence into many different languages was strong >> evidence that you understood it. >> >> >> >> But once neural networks could do that, their critics moved the >> goalposts. An exception is Hector Levesque who defined the goalposts more >> sharply by saying that the ability to get pronoun references correct in >> Winograd sentences is a crucial test. Neural nets are improving at that but >> still have some way to go. Will Gary agree that when they can get pronoun >> references correct in Winograd sentences they really do understand? Or does >> he want to reserve the right to weasel out of that too? >> >> >> >> Some people, like Gary, appear to be strongly opposed to neural networks >> because they do not fit their preconceived notions of how the mind should >> work. >> >> I believe that any reasonable person would admit that if you ask a neural >> net to draw a picture of a hamster wearing a red hat and it draws such a >> picture, it understood the request. >> >> >> >> Geoff >> >> >> >> >> >> >> >> >> >> >> >> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >> >> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural >> network community, >> >> >> >> There has been a lot of recent discussion on this list about framing and >> scientific integrity. Often the first step in restructuring narratives is >> to bully and dehumanize critics. The second is to misrepresent their >> position. People in positions of power are sometimes tempted to do this. >> >> >> >> The Hinton-Hanson interview that you just published is a real-time >> example of just that. It opens with a needless and largely content-free >> personal attack on a single scholar (me), with the explicit intention of >> discrediting that person. Worse, the only substantive thing it says is >> false. >> >> >> >> Hinton says ?In 2015 he [Marcus] made a prediction that computers >> wouldn?t be able to do machine translation.? >> >> >> >> I never said any such thing. >> >> >> >> What I predicted, rather, was that multilayer perceptrons, as they >> existed then, would not (on their own, absent other mechanisms) >> *understand* language. Seven years later, they still haven?t, except in >> the most superficial way. >> >> >> >> I made no comment whatsoever about machine translation, which I view as a >> separate problem, solvable to a certain degree by correspondance without >> semantics. >> >> >> >> I specifically tried to clarify Hinton?s confusion in 2019, but, >> disappointingly, he has continued to purvey misinformation despite that >> clarification. Here is what I wrote privately to him then, which should >> have put the matter to rest: >> >> >> >> You have taken a single out of context quote [from 2015] and >> misrepresented it. The quote, which you have prominently displayed at the >> bottom on your own web page, says: >> >> >> >> Hierarchies of features are less suited to challenges such as language, >> inference, and high-level planning. For example, as Noam Chomsky famously >> pointed out, language is filled with sentences you haven't seen >> before. Pure classifier systems don't know what to do with such sentences. >> The talent of feature detectors -- in identifying which member of some >> category something belongs to -- doesn't translate into understanding >> novel sentences, in which each sentence has its own unique meaning. >> >> >> >> It does *not* say "neural nets would not be able to deal with novel >> sentences"; it says that hierachies of features detectors (on their own, if >> you read the context of the essay) would have trouble *understanding *novel sentences. >> >> >> >> >> Google Translate does yet not *understand* the content of the sentences >> is translates. It cannot reliably answer questions about who did what to >> whom, or why, it cannot infer the order of the events in paragraphs, it >> can't determine the internal consistency of those events, and so forth. >> >> >> >> Since then, a number of scholars, such as the the computational linguist >> Emily Bender, have made similar points, and indeed current LLM difficulties >> with misinformation, incoherence and fabrication all follow from these >> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf >> , >> also emphasizing issues of understanding and meaning: >> >> >> >> *The success of the large neural language models on many NLP tasks is >> exciting. However, we find that these successes sometimes lead to hype in >> which these models are being described as ?understanding? language or >> capturing ?meaning?. In this position paper, we argue that a system trained >> only on form has a priori no way to learn meaning. .. a clear understanding >> of the distinction between form and meaning will help guide the field >> towards better science around natural language understanding. * >> >> >> >> Her later article with Gebru on language models ?stochastic parrots? is >> in some ways an extension of this point; machine translation requires >> mimicry, true understanding (which is what I was discussing in 2015) >> requires something deeper than that. >> >> >> >> Hinton?s intellectual error here is in equating machine translation with >> the deeper comprehension that robust natural language understanding will >> require; as Bender and Koller observed, the two appear not to be the same. >> (There is a longer discussion of the relation between language >> understanding and machine translation, and why the latter has turned out to >> be more approachable than the former, in my 2019 book with Ernest Davis). >> >> >> >> More broadly, Hinton?s ongoing dismissiveness of research from >> perspectives other than his own (e.g. linguistics) have done the field a >> disservice. >> >> >> >> As Herb Simon once observed, science does not have to be zero-sum. >> >> >> >> Sincerely, >> >> Gary Marcus >> >> Professor Emeritus >> >> New York University >> >> >> >> On Feb 2, 2022, at 06:12, AIhub wrote: >> >> ? >> >> Stephen Hanson in conversation with Geoff Hinton >> >> >> >> In the latest episode of this video series for AIhub.org >> , >> Stephen Hanson talks to Geoff Hinton about neural networks, >> backpropagation, overparameterization, digit recognition, voxel cells, >> syntax and semantics, Winograd sentences, and more. >> >> >> >> You can watch the discussion, and read the transcript, here: >> >> >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> >> >> >> About AIhub: >> >> AIhub is a non-profit dedicated to connecting the AI community to the >> public by providing free, high-quality information through AIhub.org >> >> (https://aihub.org/ >> ). >> We help researchers publish the latest AI news, summaries of their work, >> opinion pieces, tutorials and more. We are supported by many leading >> scientific organizations in AI, namely AAAI >> , >> NeurIPS >> , >> ICML >> , >> AIJ >> >> /IJCAI >> , >> ACM SIGAI >> , >> EurAI/AICOMM, CLAIRE >> >> and RoboCup >> >> . >> >> Twitter: @aihuborg >> >> > > Virus-free. > www.avast.com > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary at eng.ucsd.edu Fri Feb 4 13:19:37 2022 From: gary at eng.ucsd.edu (gary@ucsd.edu) Date: Fri, 4 Feb 2022 10:19:37 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: This is an argument from lack of imagination, as Pat Churchland used to say. All you have to notice, is that your brain is a neural net work. What are the alternatives? On Fri, Feb 4, 2022 at 4:08 AM Danko Nikolic wrote: > > I suppose everyone agrees that "the brain is a physical system", > and that "There is no ?magic? inside the brain", > and that '?understanding? is just part of ?learning.?' > > Also, we can agree that some sort of simulation takes place behind > understanding. > > However, there still is a problem: Neural network's can't implement the > needed simulations; they cannot achieve the same cognitive effect that > human minds can (or animal minds can). > > We don't know a way of wiring a neural network such that it could perform > the simulations (understandings) necessary to find the answers to real-life > questions, such as the hamster with a hat problem. > > In other words, neural networks, as we know them today, cannot: > > 1) learn from a small number of examples (simulation or not) > 2) apply the knowledge to a wide range of situations > > > We, as scientists, do not understand understanding. Our technology's > simulations (their depth of understanding) are no match for the simulations > (depth of understanding) that the biological brain performs. > > I think that scientific integrity also covers acknowledging when we did > not (yet) succeed in solving a certain problem. There is still significant > work to be done. > > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > --- A progress usually starts with an insight --- > > > > Virenfrei. > www.avast.com > > <#m_-1469727422087267219_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> > > On Thu, Feb 3, 2022 at 9:35 PM Asim Roy wrote: > >> First of all, the brain is a physical system. There is no ?magic? inside >> the brain that does the ?understanding? part. Take for example learning to >> play tennis. You hit a few balls - some the right way and some wrong ? but >> you fairly quickly learn to hit them right most of the time. So there is >> obviously some simulation going on in the brain about hitting the ball in >> different ways and ?learning? its consequences. What you are calling >> ?understanding? is really these simulations about different scenarios. It?s >> also very similar to augmentation used to train image recognition systems >> where you rotate images, obscure parts and so on, so that you still can say >> it?s a cat even though you see only the cat?s face or whiskers or a cat >> flipped on its back. So, if the following questions relate to >> ?understanding,? you can easily resolve this by simulating such scenarios >> when ?teaching? the system. There?s nothing ?magical? about >> ?understanding.? As I said, bear in mind that the brain, after all, is a >> physical system and ?teaching? and ?understanding? is embodied in that >> physical system, not outside it. So ?understanding? is just part of >> ?learning,? nothing more. >> >> >> >> DANKO: >> >> What would happen to the hat if the hamster rolls on its back? (Would the >> hat fall off?) >> >> What would happen to the red hat when the hamster enters its lair? (Would >> the hat fall off?) >> >> What would happen to that hamster when it goes foraging? (Would the red >> hat have an influence on finding food?) >> >> What would happen in a situation of being chased by a predator? (Would it >> be easier for predators to spot the hamster?) >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> >> *From:* Gary Marcus >> *Sent:* Thursday, February 3, 2022 9:26 AM >> *To:* Danko Nikolic >> *Cc:* Asim Roy ; Geoffrey Hinton < >> geoffrey.hinton at gmail.com>; AIhub ; >> connectionists at mailman.srv.cs.cmu.edu >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Danko, >> >> >> >> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >> talk, in which he said (paraphrasing from memory, because I don?t remember >> the precise words) that the famous 200 Quoc Le unsupervised model [ >> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >> ] >> had learned the concept of a ca. In reality the model had clustered >> together some catlike images based on the image statistics that it had >> extracted, but it was a long way from a full, counterfactual-supporting >> concept of a cat, much as you describe below. >> >> >> >> I fully agree with you that the reason for even having a semantics is as >> you put it, "to 1) learn with a few examples and 2) apply the knowledge to >> a broad set of situations.? GPT-3 sometimes gives the appearance of having >> done so, but it falls apart under close inspection, so the problem remains >> unsolved. >> >> >> >> Gary >> >> >> >> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >> wrote: >> >> >> >> G. Hinton wrote: "I believe that any reasonable person would admit that >> if you ask a neural net to draw a picture of a hamster wearing a red hat >> and it draws such a picture, it understood the request." >> >> >> >> I would like to suggest why drawing a hamster with a red hat does not >> necessarily imply understanding of the statement "hamster wearing a red >> hat". >> >> To understand that "hamster wearing a red hat" would mean inferring, in >> newly emerging situations of this hamster, all the real-life >> implications that the red hat brings to the little animal. >> >> >> >> What would happen to the hat if the hamster rolls on its back? (Would the >> hat fall off?) >> >> What would happen to the red hat when the hamster enters its lair? (Would >> the hat fall off?) >> >> What would happen to that hamster when it goes foraging? (Would the red >> hat have an influence on finding food?) >> >> What would happen in a situation of being chased by a predator? (Would it >> be easier for predators to spot the hamster?) >> >> >> >> ...and so on. >> >> >> >> Countless many questions can be asked. One has understood "hamster >> wearing a red hat" only if one can answer reasonably well many of such >> real-life relevant questions. Similarly, a student has understood materias >> in a class only if they can apply the materials in real-life situations >> (e.g., applying Pythagora's theorem). If a student gives a correct answer >> to a multiple choice question, we don't know whether the student understood >> the material or whether this was just rote learning (often, it is rote >> learning). >> >> >> >> I also suggest that understanding also comes together with effective >> learning: We store new information in such a way that we can recall it >> later and use it effectively i.e., make good inferences in newly emerging >> situations based on this knowledge. >> >> >> >> In short: Understanding makes us humans able to 1) learn with a few >> examples and 2) apply the knowledge to a broad set of situations. >> >> >> >> No neural network today has such capabilities and we don't know how to >> give them such capabilities. Neural networks need large amounts of >> training examples that cover a large variety of situations and then >> the networks can only deal with what the training examples have already >> covered. Neural networks cannot extrapolate in that 'understanding' sense. >> >> >> >> I suggest that understanding truly extrapolates from a piece of >> knowledge. It is not about satisfying a task such as translation between >> languages or drawing hamsters with hats. It is how you got the capability >> to complete the task: Did you only have a few examples that covered >> something different but related and then you extrapolated from that >> knowledge? If yes, this is going in the direction of understanding. Have >> you seen countless examples and then interpolated among them? Then perhaps >> it is not understanding. >> >> >> >> So, for the case of drawing a hamster wearing a red hat, understanding >> perhaps would have taken place if the following happened before that: >> >> >> >> 1) first, the network learned about hamsters (not many examples) >> >> 2) after that the network learned about red hats (outside the context of >> hamsters and without many examples) >> >> 3) finally the network learned about drawing (outside of the context of >> hats and hamsters, not many examples) >> >> >> >> After that, the network is asked to draw a hamster with a red hat. If it >> does it successfully, maybe we have started cracking the problem of >> understanding. >> >> >> >> Note also that this requires the network to learn sequentially without >> exhibiting catastrophic forgetting of the previous knowledge, which is >> possibly also a consequence of human learning by understanding. >> >> >> >> >> >> Danko >> >> >> >> >> >> >> >> >> >> >> >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> >> https://www.linkedin.com/in/danko-nikolic/ >> >> >> --- A progress usually starts with an insight --- >> >> >> >> >> >> >> >> >> Virus-free. www.avast.com >> >> >> >> >> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >> >> Without getting into the specific dispute between Gary and Geoff, I think >> with approaches similar to GLOM, we are finally headed in the right >> direction. There?s plenty of neurophysiological evidence for single-cell >> abstractions and multisensory neurons in the brain, which one might claim >> correspond to symbols. And I think we can finally reconcile the decades old >> dispute between Symbolic AI and Connectionism. >> >> >> >> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >> an effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> GARY: I have *never* called for dismissal of neural networks, but rather >> for some hybrid between the two (as you yourself contemplated in 1991); the >> point of the 2001 book was to characterize exactly where multilayer >> perceptrons succeeded and broke down, and where symbols could complement >> them. >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> >> *From:* Connectionists *On >> Behalf Of *Gary Marcus >> *Sent:* Wednesday, February 2, 2022 1:26 PM >> *To:* Geoffrey Hinton >> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Geoff, and interested others, >> >> >> >> What, for example, would you make of a system that often drew the >> red-hatted hamster you requested, and perhaps a fifth of the time gave you >> utter nonsense? Or say one that you trained to create birds but sometimes >> output stuff like this: >> >> >> >> >> >> >> >> One could >> >> >> >> a. avert one?s eyes and deem the anomalous outputs irrelevant >> >> or >> >> b. wonder if it might be possible that sometimes the system gets the >> right answer for the wrong reasons (eg partial historical contingency), and >> wonder whether another approach might be indicated. >> >> >> >> Benchmarks are harder than they look; most of the field has come to >> recognize that. The Turing Test has turned out to be a lousy measure of >> intelligence, easily gamed. It has turned out empirically that the Winograd >> Schema Challenge did not measure common sense as well as Hector might have >> thought. (As it happens, I am a minor coauthor of a very recent review on >> this very topic: https://arxiv.org/abs/2201.02387 >> ) >> But its conquest in no way means machines now have common sense; many >> people from many different perspectives recognize that (including, e.g., >> Yann LeCun, who generally tends to be more aligned with you than with me). >> >> >> >> So: on the goalpost of the Winograd schema, I was wrong, and you can >> quote me; but what you said about me and machine translation remains your >> invention, and it is inexcusable that you simply ignored my 2019 >> clarification. On the essential goal of trying to reach meaning and >> understanding, I remain unmoved; the problem remains unsolved. >> >> >> >> All of the problems LLMs have with coherence, reliability, truthfulness, >> misinformation, etc stand witness to that fact. (Their persistent inability >> to filter out toxic and insulting remarks stems from the same.) I am hardly >> the only person in the field to see that progress on any given benchmark >> does not inherently mean that the deep underlying problems have solved. >> You, yourself, in fact, have occasionally made that point. >> >> >> >> With respect to embeddings: Embeddings are very good for natural language >> *processing*; but NLP is not the same as NL*U* ? when it comes to >> *understanding*, their worth is still an open question. Perhaps they >> will turn out to be necessary; they clearly aren?t sufficient. In their >> extreme, they might even collapse into being symbols, in the sense of >> uniquely identifiable encodings, akin to the ASCII code, in which a >> specific set of numbers stands for a specific word or concept. (Wouldn?t >> that be ironic?) >> >> >> >> (Your GLOM, which as you know I praised publicly, is in many ways an >> effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> >> >> Notably absent from your email is any kind of apology for misrepresenting >> my position. It?s fine to say that ?many people thirty years ago once >> thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. >> I have consistently felt throughout our interactions that you have mistaken >> me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me >> for having made that error. I am still not he. >> >> >> >> Which maybe connects to the last point; if you read my work, you would >> see thirty years of arguments *for* neural networks, just not in the way >> that you want them to exist. I have ALWAYS argued that there is a role for >> them; characterizing me as a person ?strongly opposed to neural networks? >> misses the whole point of my 2001 book, which was subtitled ?Integrating >> Connectionism and Cognitive Science.? >> >> >> >> In the last two decades or so you have insisted (for reasons you have >> never fully clarified, so far as I know) on abandoning symbol-manipulation, >> but the reverse is not the case: I have *never* called for dismissal of >> neural networks, but rather for some hybrid between the two (as you >> yourself contemplated in 1991); the point of the 2001 book was to >> characterize exactly where multilayer perceptrons succeeded and broke down, >> and where symbols could complement them. It?s a rhetorical trick (which is >> what the previous thread was about) to pretend otherwise. >> >> >> >> Gary >> >> >> >> >> >> On Feb 2, 2022, at 11:22, Geoffrey Hinton >> wrote: >> >> ? >> >> Embeddings are just vectors of soft feature detectors and they are very >> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >> opposite. >> >> >> >> A few decades ago, everyone I knew then would have agreed that the >> ability to translate a sentence into many different languages was strong >> evidence that you understood it. >> >> >> >> But once neural networks could do that, their critics moved the >> goalposts. An exception is Hector Levesque who defined the goalposts more >> sharply by saying that the ability to get pronoun references correct in >> Winograd sentences is a crucial test. Neural nets are improving at that but >> still have some way to go. Will Gary agree that when they can get pronoun >> references correct in Winograd sentences they really do understand? Or does >> he want to reserve the right to weasel out of that too? >> >> >> >> Some people, like Gary, appear to be strongly opposed to neural networks >> because they do not fit their preconceived notions of how the mind should >> work. >> >> I believe that any reasonable person would admit that if you ask a neural >> net to draw a picture of a hamster wearing a red hat and it draws such a >> picture, it understood the request. >> >> >> >> Geoff >> >> >> >> >> >> >> >> >> >> >> >> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >> >> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural >> network community, >> >> >> >> There has been a lot of recent discussion on this list about framing and >> scientific integrity. Often the first step in restructuring narratives is >> to bully and dehumanize critics. The second is to misrepresent their >> position. People in positions of power are sometimes tempted to do this. >> >> >> >> The Hinton-Hanson interview that you just published is a real-time >> example of just that. It opens with a needless and largely content-free >> personal attack on a single scholar (me), with the explicit intention of >> discrediting that person. Worse, the only substantive thing it says is >> false. >> >> >> >> Hinton says ?In 2015 he [Marcus] made a prediction that computers >> wouldn?t be able to do machine translation.? >> >> >> >> I never said any such thing. >> >> >> >> What I predicted, rather, was that multilayer perceptrons, as they >> existed then, would not (on their own, absent other mechanisms) >> *understand* language. Seven years later, they still haven?t, except in >> the most superficial way. >> >> >> >> I made no comment whatsoever about machine translation, which I view as a >> separate problem, solvable to a certain degree by correspondance without >> semantics. >> >> >> >> I specifically tried to clarify Hinton?s confusion in 2019, but, >> disappointingly, he has continued to purvey misinformation despite that >> clarification. Here is what I wrote privately to him then, which should >> have put the matter to rest: >> >> >> >> You have taken a single out of context quote [from 2015] and >> misrepresented it. The quote, which you have prominently displayed at the >> bottom on your own web page, says: >> >> >> >> Hierarchies of features are less suited to challenges such as language, >> inference, and high-level planning. For example, as Noam Chomsky famously >> pointed out, language is filled with sentences you haven't seen >> before. Pure classifier systems don't know what to do with such sentences. >> The talent of feature detectors -- in identifying which member of some >> category something belongs to -- doesn't translate into understanding >> novel sentences, in which each sentence has its own unique meaning. >> >> >> >> It does *not* say "neural nets would not be able to deal with novel >> sentences"; it says that hierachies of features detectors (on their own, if >> you read the context of the essay) would have trouble *understanding *novel sentences. >> >> >> >> >> Google Translate does yet not *understand* the content of the sentences >> is translates. It cannot reliably answer questions about who did what to >> whom, or why, it cannot infer the order of the events in paragraphs, it >> can't determine the internal consistency of those events, and so forth. >> >> >> >> Since then, a number of scholars, such as the the computational linguist >> Emily Bender, have made similar points, and indeed current LLM difficulties >> with misinformation, incoherence and fabrication all follow from these >> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf >> , >> also emphasizing issues of understanding and meaning: >> >> >> >> *The success of the large neural language models on many NLP tasks is >> exciting. However, we find that these successes sometimes lead to hype in >> which these models are being described as ?understanding? language or >> capturing ?meaning?. In this position paper, we argue that a system trained >> only on form has a priori no way to learn meaning. .. a clear understanding >> of the distinction between form and meaning will help guide the field >> towards better science around natural language understanding. * >> >> >> >> Her later article with Gebru on language models ?stochastic parrots? is >> in some ways an extension of this point; machine translation requires >> mimicry, true understanding (which is what I was discussing in 2015) >> requires something deeper than that. >> >> >> >> Hinton?s intellectual error here is in equating machine translation with >> the deeper comprehension that robust natural language understanding will >> require; as Bender and Koller observed, the two appear not to be the same. >> (There is a longer discussion of the relation between language >> understanding and machine translation, and why the latter has turned out to >> be more approachable than the former, in my 2019 book with Ernest Davis). >> >> >> >> More broadly, Hinton?s ongoing dismissiveness of research from >> perspectives other than his own (e.g. linguistics) have done the field a >> disservice. >> >> >> >> As Herb Simon once observed, science does not have to be zero-sum. >> >> >> >> Sincerely, >> >> Gary Marcus >> >> Professor Emeritus >> >> New York University >> >> >> >> On Feb 2, 2022, at 06:12, AIhub wrote: >> >> ? >> >> Stephen Hanson in conversation with Geoff Hinton >> >> >> >> In the latest episode of this video series for AIhub.org >> , >> Stephen Hanson talks to Geoff Hinton about neural networks, >> backpropagation, overparameterization, digit recognition, voxel cells, >> syntax and semantics, Winograd sentences, and more. >> >> >> >> You can watch the discussion, and read the transcript, here: >> >> >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> >> >> >> About AIhub: >> >> AIhub is a non-profit dedicated to connecting the AI community to the >> public by providing free, high-quality information through AIhub.org >> >> (https://aihub.org/ >> ). >> We help researchers publish the latest AI news, summaries of their work, >> opinion pieces, tutorials and more. We are supported by many leading >> scientific organizations in AI, namely AAAI >> , >> NeurIPS >> , >> ICML >> , >> AIJ >> >> /IJCAI >> , >> ACM SIGAI >> , >> EurAI/AICOMM, CLAIRE >> >> and RoboCup >> >> . >> >> Twitter: @aihuborg >> >> >> >> >> >> >> Virus-free. www.avast.com >> >> >> >> > -- Gary Cottrell 858-534-6640 FAX: 858-534-7029 Computer Science and Engineering 0404 IF USING FEDEX INCLUDE THE FOLLOWING LINE: CSE Building, Room 4130 University of California San Diego - 9500 Gilman Drive # 0404 La Jolla, Ca. 92093-0404 Email: gary at ucsd.edu Home page: http://www-cse.ucsd.edu/~gary/ Schedule: http://tinyurl.com/b7gxpwo Blind certainty - a close-mindedness that amounts to an imprisonment so total, that the prisoner doesn?t even know that he?s locked up. -David Foster Wallace Power to the people! ?Patti Smith Except when they?re delusional ?Gary Cottrell This song makes me nostalgic for a memory I don't have -- Tess Cottrell *Listen carefully,Neither the VedasNor the Qur'anWill teach you this:Put the bit in its mouth,The saddle on its back,Your foot in the stirrup,And ride your wild runaway mindAll the way to heaven.* -- Kabir -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Fri Feb 4 15:59:12 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Fri, 4 Feb 2022 15:59:12 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <9b988b04-250b-56c8-c3ec-9509bba4bdc8@rubic.rutgers.edu> <8EA19260-6F0C-421D-8AF3-55E74C2FC3E9@nyu.edu> Message-ID: Francesca, thanks for the clear distinctions between now and then.? I was also at MCC during this time as a corporate visitor from Bell Labs (Bellcore) at the time-- we ponied up our 1M$ and 4 lucky MTS got to visit MCC every month.. and see the magic.?? Oh and have great BBQ! I sat in on Doug's stuff mostly, partly because it was such a grand and enormous plan.? And although I acknowledge it might have been? missing some Graph modeling and heaven's to Betsy exclusively using Dolphins!.. And high school students (as if only they had common sense!).,,,at least on vu-graphs Doug said CYC would become conscious in 2005 or? 2010 or 2012 and knowledge would flow out of the wall like electricity. --hallelujah. OK Ok.. maybe a bit unfair.. despite the enormous investment and the lack of return,? Doug and many others thought analogy just in the way he was implementing it was-- learning.???? And CYC was about scale... we just have to wait till it swallowed enough knowledge. So lets say, If we had CYC now and it was actually working.. would you really want to hook up a Transformer (being feed wikipedia) to it?? Why?? Because Deep Learning is more a robust learning? In some ways the GPT-Xs are CYC... just not with human understandable knowledge structures.? Could a decoder be built to create human recognizable structures from a GPT... maybe.. but it would no doubt be an approximation and couldn't really be used to reconstruct the GPT that we decoded -- it would be lost.?? Its a trapdoor. Right now? we have learning systems that work because they are composed of very many layers with billions-trillions of connections, and we have no idea why they work at all.? None. (there is that 500 page book that just dropped on statistical mechanics of DL). Things are very early yet. And I am sure experiments in hybrids will appear in NIPS for years to come.. Let them bloom if they can.???? But I doubt it. Steve On 2/4/22 3:18 PM, Francesca Rossi2 wrote: > Hi Stephen. > > I was at MCC in 1987-88, so I am aware of that effort. > As you may know, MCC included many different projects. The most visible one in trying to achieve "general" AI was CYC (I was in another project, called LDL and led by Carlo Zaniolo, now UCLA), and in my view it did not succeed because it was trying to codify all human knowledge manually and with logic. The Internet was not used yet, and knowledge graphs were not there. > > Both MCC and FGCS were relying on the assumption that everything could be coded in logic, and not through learning from data (as suggested, for example, by Udi Shapiro and others). FGCS was also claiming that a specialized hardware was needed. What neuro-symbolic AI researchers are advocating now is a fruitful way to combine learning from data and symbols/logic-based reasoning, which was not what was done at that time. > > Francesca. > ----------------------- > > Francesca Rossi > IBM Fellow and AI Ethics Global Leader > T.J. Watson Research Center, Yorktown Heights, USA > +1-617-3869639 > > ________________________________________ > From: Stephen Jos? Hanson > Sent: Friday, February 4, 2022 7:17 AM > To: Francesca Rossi2; Artur Garcez; Gary Marcus > Cc: Geoffrey Hinton; AIhub; connectionists at mailman.srv.cs.cmu.edu; Luis Lamb; Josh Tenenbaum; Anima Anandkumar; Swarat Chaudhuri; Gadi Singer > Subject: [EXTERNAL] Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Geoff's position is pretty clear. He said in the conversation we had and in this thread, "vectors of soft features", Some of my claim is in several of the conversations with Mike Jordan and Rich Sutton, but briefly, there are a number of ? ? ZjQcmQRYFpfptBannerStart > This Message Is From an External Sender > This message came from outside your organization. > ZjQcmQRYFpfptBannerEnd > > Geoff's position is pretty clear. He said in the conversation we had and in this thread, "vectors of soft features", > > Some of my claim is in several of the conversations with Mike Jordan and Rich Sutton, but briefly, there are a number of > very large costly efforts from the 1970s and 1980s, to create, deploy and curate symbol AI systems that were massive failures. Not counterfactuals, but factuals that failed. The MCC comes to mind with Adm Bobby Inmann's national US mandate to counter the Japanese so called"Fifth-generation AI systems" as a massive failure of symbolic AI. > > -------------------- > > In 1982, Japan launched its Fifth Generation Computer Systems project (FGCS), designed to develop intelligent software that would run on novel computer hardware. As the first national, large-scale artificial intelligence (AI) research and development (R&D) project to be free from military influence and corporate profit motives, the FGCS was open, international, and oriented around public goods. > > On 2/3/22 6:34 PM, Francesca Rossi2 wrote: > > Hi all. > > > > Thanks Gary for adding me to this thread. > > > > I also would be interested in knowing why Steve thinks that NS AI did not work in the past, and why this is an indication that it cannot work now or in the future. > > > > Thanks, > > Francesca. > > ------------------ > > > > Francesca Rossi > > IBM Fellow and AI Ethics Global Leader > > T.J. Watson Research Center, Yorktown Heights, USA > > +1-617-3869639 > > > > ________________________________________ > > From: Artur Garcez > > Sent: Thursday, February 3, 2022 6:00 PM > > To: Gary Marcus > > Cc: Stephen Jos? Hanson; Geoffrey Hinton; AIhub; connectionists at mailman.srv.cs.cmu.edu; Luis Lamb; Josh Tenenbaum; Anima Anandkumar; Francesca Rossi2; Swarat Chaudhuri; Gadi Singer > > Subject: [EXTERNAL] Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > > > It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally ZjQcmQRYFpfptBannerStart > > This Message Is From an External Sender > > This message came from outside your organization. > > ZjQcmQRYFpfptBannerEnd > > > > It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. > > > > Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally symbolic), there is a clear resurgence of neuro-symbolic approaches. > > > > Best wishes, > > Artur > > > > > > On Thu, Feb 3, 2022 at 5:00 PM Gary Marcus > wrote: > > Steve, > > > > I?d love to hear you elaborate on this part, > > > > Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. > > > > > > I?d love your take on why you think the impetus towards hybrid models ended badly before, and why you think that the mistakes of the past can?t be corrected. Also it? would be really instructive to compare with deep learning, which lost steam for quite some time, but reemerged much stronger than ever before. Might not the same happen with hybrid models? > > > > I am cc?ing some folks (possibly not on this list) who have recently been sympathetic to hybrid models, in hopes of a rich discussion. (And, Geoff, still cc?d, I?d genuinely welcome your thoughts if you want to add them, despite our recent friction.) > > > > Cheers, > > Gary > > > > > > On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson > wrote: > > > > > > I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. > > > > Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. > > > > Best and hope you are doing well. > > > > Steve > > > > > > -- > [cid:part1.23688013.E2D7C5E3 at rubic.rutgers.edu] -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From tgd at oregonstate.edu Fri Feb 4 13:37:23 2022 From: tgd at oregonstate.edu (Dietterich, Thomas) Date: Fri, 4 Feb 2022 18:37:23 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <6f8f9f7f-053d-a633-805d-bacd28967edf@rubic.rutgers.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <6f8f9f7f-053d-a633-805d-bacd28967edf@rubic.rutgers.edu> Message-ID: I mean that if you only say a system is ?understanding? X if it can enumerate all of the consequences of X, then you have solved what is known as the ?Ramification Problem?. And it is easy to show that this is impossible. Hence, our criteria for saying that a system ?understandings? must lie somewhere between ?doing the right thing in this one situation? and knowing all of the consequences of its beliefs. --Tom Thomas G. Dietterich, Distinguished Professor Voice: 541-737-5559 School of Electrical Engineering FAX: 541-737-1300 and Computer Science URL: eecs.oregonstate.edu/~tgd US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 From: Stephen Jos? Hanson Sent: Friday, February 4, 2022 4:01 AM To: Dietterich, Thomas ; Gary Marcus ; Danko Nikolic Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Tom, understanding is a theorem? you mean it should be a theorem? and yes, if you are having brain surgery.. you hope your surgeon, "understands" what they are doing.. Steve On 2/3/22 12:31 PM, Dietterich, Thomas wrote: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub > wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com -- [cid:image001.png at 01D819B2.C34F3210] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 19957 bytes Desc: image001.png URL: From geoffrey.hinton at gmail.com Fri Feb 4 15:24:02 2022 From: geoffrey.hinton at gmail.com (Geoffrey Hinton) Date: Fri, 4 Feb 2022 15:24:02 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas wrote: > ?Understanding? is not a Boolean. It is a theorem that no system can > enumerate all of the consequences of a state of affairs in the world. > > > > For low-stakes application work, we can be satisfied by a system that > ?does the right thing?. If the system draws a good picture, that?s > sufficient. It ?understood? the request. > > > > But for higher-stakes applications---and for advancing the science---we > seek a causal account of how the components of a system cause it to do the > right thing. We are hoping that a small set of mechanisms can produce broad > coverage of intelligent behavior. This gives us confidence that the system > will respond correctly outside of the narrow tasks on which we have tested > it. > > > > --Tom > > > > Thomas G. Dietterich, Distinguished Professor Emeritus > > School of Electrical Engineering and Computer > Science > > US Mail: 1148 Kelley Engineering Center > > > > Office: 2067 Kelley Engineering Center > > Oregon State Univ., Corvallis, OR 97331-5501 > > Voice: 541-737-5559; FAX: 541-737-1300 > > URL: http://web.engr.oregonstate.edu/~tgd/ > > > > *From:* Connectionists *On > Behalf Of *Gary Marcus > *Sent:* Thursday, February 3, 2022 8:26 AM > *To:* Danko Nikolic > *Cc:* connectionists at mailman.srv.cs.cmu.edu; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > [This email originated from outside of OSU. Use caution with links and > attachments.] > > Dear Danko, > > > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, > in which he said (paraphrasing from memory, because I don?t remember the > precise words) that the famous 200 Quoc Le unsupervised model [ > https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf > ] > had learned the concept of a ca. In reality the model had clustered > together some catlike images based on the image statistics that it had > extracted, but it was a long way from a full, counterfactual-supporting > concept of a cat, much as you describe below. > > > > I fully agree with you that the reason for even having a semantics is as > you put it, "to 1) learn with a few examples and 2) apply the knowledge to > a broad set of situations.? GPT-3 sometimes gives the appearance of having > done so, but it falls apart under close inspection, so the problem remains > unsolved. > > > > Gary > > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > > > G. Hinton wrote: "I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster wearing a red hat and > it draws such a picture, it understood the request." > > > > I would like to suggest why drawing a hamster with a red hat does not > necessarily imply understanding of the statement "hamster wearing a red > hat". > > To understand that "hamster wearing a red hat" would mean inferring, in > newly emerging situations of this hamster, all the real-life > implications that the red hat brings to the little animal. > > > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > ...and so on. > > > > Countless many questions can be asked. One has understood "hamster wearing > a red hat" only if one can answer reasonably well many of such real-life > relevant questions. Similarly, a student has understood materias in a class > only if they can apply the materials in real-life situations (e.g., > applying Pythagora's theorem). If a student gives a correct answer to a > multiple choice question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > > > I also suggest that understanding also comes together with effective > learning: We store new information in such a way that we can recall it > later and use it effectively i.e., make good inferences in newly emerging > situations based on this knowledge. > > > > In short: Understanding makes us humans able to 1) learn with a few > examples and 2) apply the knowledge to a broad set of situations. > > > > No neural network today has such capabilities and we don't know how to > give them such capabilities. Neural networks need large amounts of > training examples that cover a large variety of situations and then > the networks can only deal with what the training examples have already > covered. Neural networks cannot extrapolate in that 'understanding' sense. > > > > I suggest that understanding truly extrapolates from a piece of knowledge. > It is not about satisfying a task such as translation between languages or > drawing hamsters with hats. It is how you got the capability to complete > the task: Did you only have a few examples that covered something different > but related and then you extrapolated from that knowledge? If yes, this is > going in the direction of understanding. Have you seen countless examples > and then interpolated among them? Then perhaps it is not understanding. > > > > So, for the case of drawing a hamster wearing a red hat, understanding > perhaps would have taken place if the following happened before that: > > > > 1) first, the network learned about hamsters (not many examples) > > 2) after that the network learned about red hats (outside the context of > hamsters and without many examples) > > 3) finally the network learned about drawing (outside of the context of > hats and hamsters, not many examples) > > > > After that, the network is asked to draw a hamster with a red hat. If it > does it successfully, maybe we have started cracking the problem of > understanding. > > > > Note also that this requires the network to learn sequentially without > exhibiting catastrophic forgetting of the previous knowledge, which is > possibly also a consequence of human learning by understanding. > > > > > > Danko > > > > > > > > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > > > > Virus-free. www.avast.com > > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > > Without getting into the specific dispute between Gary and Geoff, I think > with approaches similar to GLOM, we are finally headed in the right > direction. There?s plenty of neurophysiological evidence for single-cell > abstractions and multisensory neurons in the brain, which one might claim > correspond to symbols. And I think we can finally reconcile the decades old > dispute between Symbolic AI and Connectionism. > > > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > GARY: I have *never* called for dismissal of neural networks, but rather > for some hybrid between the two (as you yourself contemplated in 1991); the > point of the 2001 book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols could complement > them. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Gary Marcus > *Sent:* Wednesday, February 2, 2022 1:26 PM > *To:* Geoffrey Hinton > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Geoff, and interested others, > > > > What, for example, would you make of a system that often drew the > red-hatted hamster you requested, and perhaps a fifth of the time gave you > utter nonsense? Or say one that you trained to create birds but sometimes > output stuff like this: > > > > > > > > One could > > > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > or > > b. wonder if it might be possible that sometimes the system gets the right > answer for the wrong reasons (eg partial historical contingency), and > wonder whether another approach might be indicated. > > > > Benchmarks are harder than they look; most of the field has come to > recognize that. The Turing Test has turned out to be a lousy measure of > intelligence, easily gamed. It has turned out empirically that the Winograd > Schema Challenge did not measure common sense as well as Hector might have > thought. (As it happens, I am a minor coauthor of a very recent review on > this very topic: https://arxiv.org/abs/2201.02387 > ) > But its conquest in no way means machines now have common sense; many > people from many different perspectives recognize that (including, e.g., > Yann LeCun, who generally tends to be more aligned with you than with me). > > > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote > me; but what you said about me and machine translation remains your > invention, and it is inexcusable that you simply ignored my 2019 > clarification. On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains unsolved. > > > > All of the problems LLMs have with coherence, reliability, truthfulness, > misinformation, etc stand witness to that fact. (Their persistent inability > to filter out toxic and insulting remarks stems from the same.) I am hardly > the only person in the field to see that progress on any given benchmark > does not inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > > > With respect to embeddings: Embeddings are very good for natural language > *processing*; but NLP is not the same as NL*U* ? when it comes to > *understanding*, their worth is still an open question. Perhaps they will > turn out to be necessary; they clearly aren?t sufficient. In their extreme, > they might even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a specific set of > numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > > > (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > > Notably absent from your email is any kind of apology for misrepresenting > my position. It?s fine to say that ?many people thirty years ago once > thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. > I have consistently felt throughout our interactions that you have mistaken > me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me > for having made that error. I am still not he. > > > > Which maybe connects to the last point; if you read my work, you would see > thirty years of arguments *for* neural networks, just not in the way that > you want them to exist. I have ALWAYS argued that there is a role for them; > characterizing me as a person ?strongly opposed to neural networks? misses > the whole point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > > > In the last two decades or so you have insisted (for reasons you have > never fully clarified, so far as I know) on abandoning symbol-manipulation, > but the reverse is not the case: I have *never* called for dismissal of > neural networks, but rather for some hybrid between the two (as you > yourself contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded and broke down, > and where symbols could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > > > Gary > > > > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: > > ? > > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > > > > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > > > Geoff > > > > > > > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural > network community, > > > > There has been a lot of recent discussion on this list about framing and > scientific integrity. Often the first step in restructuring narratives is > to bully and dehumanize critics. The second is to misrepresent their > position. People in positions of power are sometimes tempted to do this. > > > > The Hinton-Hanson interview that you just published is a real-time example > of just that. It opens with a needless and largely content-free personal > attack on a single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive thing it says is > false. > > > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t > be able to do machine translation.? > > > > I never said any such thing. > > > > What I predicted, rather, was that multilayer perceptrons, as they existed > then, would not (on their own, absent other mechanisms) *understand* language. > Seven years later, they still haven?t, except in the most superficial way. > > > > > I made no comment whatsoever about machine translation, which I view as a > separate problem, solvable to a certain degree by correspondance without > semantics. > > > > I specifically tried to clarify Hinton?s confusion in 2019, but, > disappointingly, he has continued to purvey misinformation despite that > clarification. Here is what I wrote privately to him then, which should > have put the matter to rest: > > > > You have taken a single out of context quote [from 2015] and > misrepresented it. The quote, which you have prominently displayed at the > bottom on your own web page, says: > > > > Hierarchies of features are less suited to challenges such as language, > inference, and high-level planning. For example, as Noam Chomsky famously > pointed out, language is filled with sentences you haven't seen > before. Pure classifier systems don't know what to do with such sentences. > The talent of feature detectors -- in identifying which member of some > category something belongs to -- doesn't translate into understanding > novel sentences, in which each sentence has its own unique meaning. > > > > It does *not* say "neural nets would not be able to deal with novel > sentences"; it says that hierachies of features detectors (on their own, if > you read the context of the essay) would have trouble *understanding *novel sentences. > > > > > Google Translate does yet not *understand* the content of the sentences > is translates. It cannot reliably answer questions about who did what to > whom, or why, it cannot infer the order of the events in paragraphs, it > can't determine the internal consistency of those events, and so forth. > > > > Since then, a number of scholars, such as the the computational linguist > Emily Bender, have made similar points, and indeed current LLM difficulties > with misinformation, incoherence and fabrication all follow from these > concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter > with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf > , > also emphasizing issues of understanding and meaning: > > > > *The success of the large neural language models on many NLP tasks is > exciting. However, we find that these successes sometimes lead to hype in > which these models are being described as ?understanding? language or > capturing ?meaning?. In this position paper, we argue that a system trained > only on form has a priori no way to learn meaning. .. a clear understanding > of the distinction between form and meaning will help guide the field > towards better science around natural language understanding. * > > > > Her later article with Gebru on language models ?stochastic parrots? is in > some ways an extension of this point; machine translation requires mimicry, > true understanding (which is what I was discussing in 2015) requires > something deeper than that. > > > > Hinton?s intellectual error here is in equating machine translation with > the deeper comprehension that robust natural language understanding will > require; as Bender and Koller observed, the two appear not to be the same. > (There is a longer discussion of the relation between language > understanding and machine translation, and why the latter has turned out to > be more approachable than the former, in my 2019 book with Ernest Davis). > > > > More broadly, Hinton?s ongoing dismissiveness of research from > perspectives other than his own (e.g. linguistics) have done the field a > disservice. > > > > As Herb Simon once observed, science does not have to be zero-sum. > > > > Sincerely, > > Gary Marcus > > Professor Emeritus > > New York University > > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > > Stephen Hanson in conversation with Geoff Hinton > > > > In the latest episode of this video series for AIhub.org > , > Stephen Hanson talks to Geoff Hinton about neural networks, > backpropagation, overparameterization, digit recognition, voxel cells, > syntax and semantics, Winograd sentences, and more. > > > > You can watch the discussion, and read the transcript, here: > > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > > > About AIhub: > > AIhub is a non-profit dedicated to connecting the AI community to the > public by providing free, high-quality information through AIhub.org > > (https://aihub.org/ > ). > We help researchers publish the latest AI news, summaries of their work, > opinion pieces, tutorials and more. We are supported by many leading > scientific organizations in AI, namely AAAI > , > NeurIPS > , > ICML > , > AIJ > > /IJCAI > , > ACM SIGAI > , > EurAI/AICOMM, CLAIRE > > and RoboCup > > . > > Twitter: @aihuborg > > > > > > > Virus-free. www.avast.com > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Fri Feb 4 14:52:42 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Fri, 4 Feb 2022 11:52:42 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: <303504A2-453D-4DE8-8A34-C41693041954@nyu.edu> ?Steve, The phrase I always liked was ?poverty of the imagination arguments?; I share your disdain for them. But that?s why I think you should be careful of any retreat into biological plausibility. As even Jay McClelland has acknowledged, we do know that some humans some of the time manipulate symbols. So wetware-based symbols are not literally biologically impossible; the real question for cognitive neuroscience is about the scope and development of symbols. For engineering, the real question is, are they useful. Certainly for software engineering in general, they are indispensable. Beyond this, none of the available AI approaches map particularly neatly onto what we know about the brain, and none of what we know about the brain is understood well enough to solve AI. All the examples you point to, for instance, are actually controversial, not decisive. As you probably know, for example, Nancy Kanwisher has a different take on domain-specificity than you do (https://web.mit.edu/bcs/nklab/), with evidence of specialization early in life, and Jeff Bowers has argued that the grandmother cell hypothesis has been dismissed prematurely (https://jeffbowers.blogs.bristol.ac.uk/blog/grandmother-cells/); there?s also a long literature on the possible neural realization of rules, both in humans and other animals. I don?t know what the right answers are there, but nor do I think that neurosymbolic systems are beholden to them anymore than CNNs are bound to whether or not the brain performs back-propagation. Finally, as a reminder, ?Distributed? per se in not the right question; in some technical sense ASCII encodings are distributed, and about as symbolic as you can get. The proper question is really what you do with your encodings; the neurosymbolic approach is trying to broaden the available range of options. Gary > On Feb 4, 2022, at 07:04, Stephen Jos? Hanson wrote: > ? > Well I don't like counterfactual arguments or ones that start with "It can't be done with neural networks.."--as this amounts to the old Rumelhart saw, of "proof by lack of imagination". > > I think my position and others (I can't speak for Geoff and won't) is more of a "purist" view that brains have computationally complete representational power to do what ever is required of human level mental processing. AI symbol systems are remote descriptions of this level of processing. Looking at 1000s of brain scans, one begins to see a pattern of interacting large and smaller scale networks, probably related to Resting state and the Default Mode networks in some important competitive way. But what one doesn't find is modular structure (e.g. face area.. nope) or evidence of "symbols" being processed. Research on Numbers is interesting in this regard, as number representation should provide some evidence of discrete symbol processing as would letters. But again the processing states from brain imaging more generally appear to be distributed representations of some sort. > > One other direction has to do with prior rules that could be neurally coded and therefore provide an immediate bias in learning and thus dramatically reduce the number of examples required for asymptotic learning. Some of this has been done with pre-training-- on let's say 1000s of videos that are relatively generic, prior to learning on a small set of videos related to a specific topic-- say two individuals playing a monopoly game. In that case, no game-like videos were sampled in the pre-training, and the LSTM was trained to detect change point on 2 minutes of video, achieving a 97% match with human parsers. In these senses I have no problem with this type of hybrid training. > > Steve > > On 2/4/22 9:07 AM, Gary Marcus wrote: >> ?The whole point of the neurosymbolic approach is to develop systems that can accommodate both vectors and symbols, since neither on their own seems adequate. >> >> If there are arguments against trying to do that, we would be interested. >> >>> On Feb 4, 2022, at 4:17 AM, Stephen Jos? Hanson wrote: >>> ? >>> Geoff's position is pretty clear. He said in the conversation we had and in this thread, "vectors of soft features", >>> >>> Some of my claim is in several of the conversations with Mike Jordan and Rich Sutton, but briefly, there are a number of >>> very large costly efforts from the 1970s and 1980s, to create, deploy and curate symbol AI systems that were massive failures. Not counterfactuals, but factuals that failed. The MCC comes to mind with Adm Bobby Inmann's national US mandate to counter the Japanese so called"Fifth-generation AI systems" as a massive failure of symbolic AI. >>> >>> -------------------- >>> >>> In 1982, Japan launched its Fifth Generation Computer Systems project (FGCS), designed to develop intelligent software that would run on novel computer hardware. As the first national, large-scale artificial intelligence (AI) research and development (R&D) project to be free from military influence and corporate profit motives, the FGCS was open, international, and oriented around public goods. >>> >>> On 2/3/22 6:34 PM, Francesca Rossi2 wrote: >>>> Hi all. >>>> >>>> Thanks Gary for adding me to this thread. >>>> >>>> I also would be interested in knowing why Steve thinks that NS AI did not work in the past, and why this is an indication that it cannot work now or in the future. >>>> >>>> Thanks, >>>> Francesca. >>>> ------------------ >>>> >>>> Francesca Rossi >>>> IBM Fellow and AI Ethics Global Leader >>>> T.J. Watson Research Center, Yorktown Heights, USA >>>> +1-617-3869639 >>>> >>>> ________________________________________ >>>> From: Artur Garcez >>>> Sent: Thursday, February 3, 2022 6:00 PM >>>> To: Gary Marcus >>>> Cc: Stephen Jos? Hanson; Geoffrey Hinton; AIhub; connectionists at mailman.srv.cs.cmu.edu; Luis Lamb; Josh Tenenbaum; Anima Anandkumar; Francesca Rossi2; Swarat Chaudhuri; Gadi Singer >>>> Subject: [EXTERNAL] Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>> >>>> It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally ZjQcmQRYFpfptBannerStart >>>> This Message Is From an External Sender >>>> This message came from outside your organization. >>>> ZjQcmQRYFpfptBannerEnd >>>> >>>> It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. >>>> >>>> Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally symbolic), there is a clear resurgence of neuro-symbolic approaches. >>>> >>>> Best wishes, >>>> Artur >>>> >>>> >>>> On Thu, Feb 3, 2022 at 5:00 PM Gary Marcus > wrote: >>>> Steve, >>>> >>>> I?d love to hear you elaborate on this part, >>>> >>>> Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. >>>> >>>> >>>> I?d love your take on why you think the impetus towards hybrid models ended badly before, and why you think that the mistakes of the past can?t be corrected. Also it? would be really instructive to compare with deep learning, which lost steam for quite some time, but reemerged much stronger than ever before. Might not the same happen with hybrid models? >>>> >>>> I am cc?ing some folks (possibly not on this list) who have recently been sympathetic to hybrid models, in hopes of a rich discussion. (And, Geoff, still cc?d, I?d genuinely welcome your thoughts if you want to add them, despite our recent friction.) >>>> >>>> Cheers, >>>> Gary >>>> >>>> >>>> On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson > wrote: >>>> >>>> >>>> I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. >>>> >>>> Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. >>>> >>>> Best and hope you are doing well. >>>> >>>> Steve >>>> >>> -- >>> > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From minaiaa at gmail.com Sat Feb 5 01:37:46 2022 From: minaiaa at gmail.com (Ali Minai) Date: Sat, 5 Feb 2022 01:37:46 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Asim Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! Ali *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: > First of all, the brain is a physical system. There is no ?magic? inside > the brain that does the ?understanding? part. Take for example learning to > play tennis. You hit a few balls - some the right way and some wrong ? but > you fairly quickly learn to hit them right most of the time. So there is > obviously some simulation going on in the brain about hitting the ball in > different ways and ?learning? its consequences. What you are calling > ?understanding? is really these simulations about different scenarios. It?s > also very similar to augmentation used to train image recognition systems > where you rotate images, obscure parts and so on, so that you still can say > it?s a cat even though you see only the cat?s face or whiskers or a cat > flipped on its back. So, if the following questions relate to > ?understanding,? you can easily resolve this by simulating such scenarios > when ?teaching? the system. There?s nothing ?magical? about > ?understanding.? As I said, bear in mind that the brain, after all, is a > physical system and ?teaching? and ?understanding? is embodied in that > physical system, not outside it. So ?understanding? is just part of > ?learning,? nothing more. > > > > DANKO: > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Gary Marcus > *Sent:* Thursday, February 3, 2022 9:26 AM > *To:* Danko Nikolic > *Cc:* Asim Roy ; Geoffrey Hinton < > geoffrey.hinton at gmail.com>; AIhub ; > connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Danko, > > > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, > in which he said (paraphrasing from memory, because I don?t remember the > precise words) that the famous 200 Quoc Le unsupervised model [ > https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf > ] > had learned the concept of a ca. In reality the model had clustered > together some catlike images based on the image statistics that it had > extracted, but it was a long way from a full, counterfactual-supporting > concept of a cat, much as you describe below. > > > > I fully agree with you that the reason for even having a semantics is as > you put it, "to 1) learn with a few examples and 2) apply the knowledge to > a broad set of situations.? GPT-3 sometimes gives the appearance of having > done so, but it falls apart under close inspection, so the problem remains > unsolved. > > > > Gary > > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > > > G. Hinton wrote: "I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster wearing a red hat and > it draws such a picture, it understood the request." > > > > I would like to suggest why drawing a hamster with a red hat does not > necessarily imply understanding of the statement "hamster wearing a red > hat". > > To understand that "hamster wearing a red hat" would mean inferring, in > newly emerging situations of this hamster, all the real-life > implications that the red hat brings to the little animal. > > > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > ...and so on. > > > > Countless many questions can be asked. One has understood "hamster wearing > a red hat" only if one can answer reasonably well many of such real-life > relevant questions. Similarly, a student has understood materias in a class > only if they can apply the materials in real-life situations (e.g., > applying Pythagora's theorem). If a student gives a correct answer to a > multiple choice question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > > > I also suggest that understanding also comes together with effective > learning: We store new information in such a way that we can recall it > later and use it effectively i.e., make good inferences in newly emerging > situations based on this knowledge. > > > > In short: Understanding makes us humans able to 1) learn with a few > examples and 2) apply the knowledge to a broad set of situations. > > > > No neural network today has such capabilities and we don't know how to > give them such capabilities. Neural networks need large amounts of > training examples that cover a large variety of situations and then > the networks can only deal with what the training examples have already > covered. Neural networks cannot extrapolate in that 'understanding' sense. > > > > I suggest that understanding truly extrapolates from a piece of knowledge. > It is not about satisfying a task such as translation between languages or > drawing hamsters with hats. It is how you got the capability to complete > the task: Did you only have a few examples that covered something different > but related and then you extrapolated from that knowledge? If yes, this is > going in the direction of understanding. Have you seen countless examples > and then interpolated among them? Then perhaps it is not understanding. > > > > So, for the case of drawing a hamster wearing a red hat, understanding > perhaps would have taken place if the following happened before that: > > > > 1) first, the network learned about hamsters (not many examples) > > 2) after that the network learned about red hats (outside the context of > hamsters and without many examples) > > 3) finally the network learned about drawing (outside of the context of > hats and hamsters, not many examples) > > > > After that, the network is asked to draw a hamster with a red hat. If it > does it successfully, maybe we have started cracking the problem of > understanding. > > > > Note also that this requires the network to learn sequentially without > exhibiting catastrophic forgetting of the previous knowledge, which is > possibly also a consequence of human learning by understanding. > > > > > > Danko > > > > > > > > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > > > > Virus-free. www.avast.com > > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > > Without getting into the specific dispute between Gary and Geoff, I think > with approaches similar to GLOM, we are finally headed in the right > direction. There?s plenty of neurophysiological evidence for single-cell > abstractions and multisensory neurons in the brain, which one might claim > correspond to symbols. And I think we can finally reconcile the decades old > dispute between Symbolic AI and Connectionism. > > > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > GARY: I have *never* called for dismissal of neural networks, but rather > for some hybrid between the two (as you yourself contemplated in 1991); the > point of the 2001 book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols could complement > them. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Gary Marcus > *Sent:* Wednesday, February 2, 2022 1:26 PM > *To:* Geoffrey Hinton > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Geoff, and interested others, > > > > What, for example, would you make of a system that often drew the > red-hatted hamster you requested, and perhaps a fifth of the time gave you > utter nonsense? Or say one that you trained to create birds but sometimes > output stuff like this: > > > > > > > > One could > > > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > or > > b. wonder if it might be possible that sometimes the system gets the right > answer for the wrong reasons (eg partial historical contingency), and > wonder whether another approach might be indicated. > > > > Benchmarks are harder than they look; most of the field has come to > recognize that. The Turing Test has turned out to be a lousy measure of > intelligence, easily gamed. It has turned out empirically that the Winograd > Schema Challenge did not measure common sense as well as Hector might have > thought. (As it happens, I am a minor coauthor of a very recent review on > this very topic: https://arxiv.org/abs/2201.02387 > ) > But its conquest in no way means machines now have common sense; many > people from many different perspectives recognize that (including, e.g., > Yann LeCun, who generally tends to be more aligned with you than with me). > > > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote > me; but what you said about me and machine translation remains your > invention, and it is inexcusable that you simply ignored my 2019 > clarification. On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains unsolved. > > > > All of the problems LLMs have with coherence, reliability, truthfulness, > misinformation, etc stand witness to that fact. (Their persistent inability > to filter out toxic and insulting remarks stems from the same.) I am hardly > the only person in the field to see that progress on any given benchmark > does not inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > > > With respect to embeddings: Embeddings are very good for natural language > *processing*; but NLP is not the same as NL*U* ? when it comes to > *understanding*, their worth is still an open question. Perhaps they will > turn out to be necessary; they clearly aren?t sufficient. In their extreme, > they might even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a specific set of > numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > > > (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > > Notably absent from your email is any kind of apology for misrepresenting > my position. It?s fine to say that ?many people thirty years ago once > thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. > I have consistently felt throughout our interactions that you have mistaken > me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me > for having made that error. I am still not he. > > > > Which maybe connects to the last point; if you read my work, you would see > thirty years of arguments *for* neural networks, just not in the way that > you want them to exist. I have ALWAYS argued that there is a role for them; > characterizing me as a person ?strongly opposed to neural networks? misses > the whole point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > > > In the last two decades or so you have insisted (for reasons you have > never fully clarified, so far as I know) on abandoning symbol-manipulation, > but the reverse is not the case: I have *never* called for dismissal of > neural networks, but rather for some hybrid between the two (as you > yourself contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded and broke down, > and where symbols could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > > > Gary > > > > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: > > ? > > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > > > > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > > > Geoff > > > > > > > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural > network community, > > > > There has been a lot of recent discussion on this list about framing and > scientific integrity. Often the first step in restructuring narratives is > to bully and dehumanize critics. The second is to misrepresent their > position. People in positions of power are sometimes tempted to do this. > > > > The Hinton-Hanson interview that you just published is a real-time example > of just that. It opens with a needless and largely content-free personal > attack on a single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive thing it says is > false. > > > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t > be able to do machine translation.? > > > > I never said any such thing. > > > > What I predicted, rather, was that multilayer perceptrons, as they existed > then, would not (on their own, absent other mechanisms) *understand* language. > Seven years later, they still haven?t, except in the most superficial way. > > > > > I made no comment whatsoever about machine translation, which I view as a > separate problem, solvable to a certain degree by correspondance without > semantics. > > > > I specifically tried to clarify Hinton?s confusion in 2019, but, > disappointingly, he has continued to purvey misinformation despite that > clarification. Here is what I wrote privately to him then, which should > have put the matter to rest: > > > > You have taken a single out of context quote [from 2015] and > misrepresented it. The quote, which you have prominently displayed at the > bottom on your own web page, says: > > > > Hierarchies of features are less suited to challenges such as language, > inference, and high-level planning. For example, as Noam Chomsky famously > pointed out, language is filled with sentences you haven't seen > before. Pure classifier systems don't know what to do with such sentences. > The talent of feature detectors -- in identifying which member of some > category something belongs to -- doesn't translate into understanding > novel sentences, in which each sentence has its own unique meaning. > > > > It does *not* say "neural nets would not be able to deal with novel > sentences"; it says that hierachies of features detectors (on their own, if > you read the context of the essay) would have trouble *understanding *novel sentences. > > > > > Google Translate does yet not *understand* the content of the sentences > is translates. It cannot reliably answer questions about who did what to > whom, or why, it cannot infer the order of the events in paragraphs, it > can't determine the internal consistency of those events, and so forth. > > > > Since then, a number of scholars, such as the the computational linguist > Emily Bender, have made similar points, and indeed current LLM difficulties > with misinformation, incoherence and fabrication all follow from these > concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter > with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf > , > also emphasizing issues of understanding and meaning: > > > > *The success of the large neural language models on many NLP tasks is > exciting. However, we find that these successes sometimes lead to hype in > which these models are being described as ?understanding? language or > capturing ?meaning?. In this position paper, we argue that a system trained > only on form has a priori no way to learn meaning. .. a clear understanding > of the distinction between form and meaning will help guide the field > towards better science around natural language understanding. * > > > > Her later article with Gebru on language models ?stochastic parrots? is in > some ways an extension of this point; machine translation requires mimicry, > true understanding (which is what I was discussing in 2015) requires > something deeper than that. > > > > Hinton?s intellectual error here is in equating machine translation with > the deeper comprehension that robust natural language understanding will > require; as Bender and Koller observed, the two appear not to be the same. > (There is a longer discussion of the relation between language > understanding and machine translation, and why the latter has turned out to > be more approachable than the former, in my 2019 book with Ernest Davis). > > > > More broadly, Hinton?s ongoing dismissiveness of research from > perspectives other than his own (e.g. linguistics) have done the field a > disservice. > > > > As Herb Simon once observed, science does not have to be zero-sum. > > > > Sincerely, > > Gary Marcus > > Professor Emeritus > > New York University > > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > > Stephen Hanson in conversation with Geoff Hinton > > > > In the latest episode of this video series for AIhub.org > , > Stephen Hanson talks to Geoff Hinton about neural networks, > backpropagation, overparameterization, digit recognition, voxel cells, > syntax and semantics, Winograd sentences, and more. > > > > You can watch the discussion, and read the transcript, here: > > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > > > About AIhub: > > AIhub is a non-profit dedicated to connecting the AI community to the > public by providing free, high-quality information through AIhub.org > > (https://aihub.org/ > ). > We help researchers publish the latest AI news, summaries of their work, > opinion pieces, tutorials and more. We are supported by many leading > scientific organizations in AI, namely AAAI > , > NeurIPS > , > ICML > , > AIJ > > /IJCAI > , > ACM SIGAI > , > EurAI/AICOMM, CLAIRE > > and RoboCup > > . > > Twitter: @aihuborg > > > > > > > Virus-free. www.avast.com > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Sat Feb 5 05:11:51 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Sat, 5 Feb 2022 11:11:51 +0100 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Gary, you wrote: "What are the alternatives?" There is at least one alternative: the theory of practopoiesis which suggests that it is not the neural networks that "compute" the mental operations. It is instead the quick adaptations of neurons who are responsible for thinking and perceiving. The network only serves the function of bringing in the information and sending it out. The adaptations are suggested to do the central part of the cognition. So far, this is all hypothetical. If we develop these ideas into a working system, this would be an entirely new paradigm. It would be like the third paradigm: 1) manipulation of symbols 2) neural net 3) fast adaptations Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Fri, Feb 4, 2022 at 7:19 PM gary at ucsd.edu wrote: > This is an argument from lack of imagination, as Pat Churchland used to > say. All you have to notice, is that your brain is a neural net work. What > are the alternatives? > > On Fri, Feb 4, 2022 at 4:08 AM Danko Nikolic > wrote: > >> >> I suppose everyone agrees that "the brain is a physical system", >> and that "There is no ?magic? inside the brain", >> and that '?understanding? is just part of ?learning.?' >> >> Also, we can agree that some sort of simulation takes place behind >> understanding. >> >> However, there still is a problem: Neural network's can't implement the >> needed simulations; they cannot achieve the same cognitive effect that >> human minds can (or animal minds can). >> >> We don't know a way of wiring a neural network such that it could perform >> the simulations (understandings) necessary to find the answers to real-life >> questions, such as the hamster with a hat problem. >> >> In other words, neural networks, as we know them today, cannot: >> >> 1) learn from a small number of examples (simulation or not) >> 2) apply the knowledge to a wide range of situations >> >> >> We, as scientists, do not understand understanding. Our technology's >> simulations (their depth of understanding) are no match for the simulations >> (depth of understanding) that the biological brain performs. >> >> I think that scientific integrity also covers acknowledging when we did >> not (yet) succeed in solving a certain problem. There is still significant >> work to be done. >> >> >> Danko >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> https://www.linkedin.com/in/danko-nikolic/ >> --- A progress usually starts with an insight --- >> >> >> >> Virenfrei. >> www.avast.com >> >> <#m_-3229424020171779455_m_-1469727422087267219_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> >> >> On Thu, Feb 3, 2022 at 9:35 PM Asim Roy wrote: >> >>> First of all, the brain is a physical system. There is no ?magic? inside >>> the brain that does the ?understanding? part. Take for example learning to >>> play tennis. You hit a few balls - some the right way and some wrong ? but >>> you fairly quickly learn to hit them right most of the time. So there is >>> obviously some simulation going on in the brain about hitting the ball in >>> different ways and ?learning? its consequences. What you are calling >>> ?understanding? is really these simulations about different scenarios. It?s >>> also very similar to augmentation used to train image recognition systems >>> where you rotate images, obscure parts and so on, so that you still can say >>> it?s a cat even though you see only the cat?s face or whiskers or a cat >>> flipped on its back. So, if the following questions relate to >>> ?understanding,? you can easily resolve this by simulating such scenarios >>> when ?teaching? the system. There?s nothing ?magical? about >>> ?understanding.? As I said, bear in mind that the brain, after all, is a >>> physical system and ?teaching? and ?understanding? is embodied in that >>> physical system, not outside it. So ?understanding? is just part of >>> ?learning,? nothing more. >>> >>> >>> >>> DANKO: >>> >>> What would happen to the hat if the hamster rolls on its back? (Would >>> the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? >>> (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red >>> hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would >>> it be easier for predators to spot the hamster?) >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> *From:* Gary Marcus >>> *Sent:* Thursday, February 3, 2022 9:26 AM >>> *To:* Danko Nikolic >>> *Cc:* Asim Roy ; Geoffrey Hinton < >>> geoffrey.hinton at gmail.com>; AIhub ; >>> connectionists at mailman.srv.cs.cmu.edu >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Dear Danko, >>> >>> >>> >>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >>> talk, in which he said (paraphrasing from memory, because I don?t remember >>> the precise words) that the famous 200 Quoc Le unsupervised model [ >>> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >>> ] >>> had learned the concept of a ca. In reality the model had clustered >>> together some catlike images based on the image statistics that it had >>> extracted, but it was a long way from a full, counterfactual-supporting >>> concept of a cat, much as you describe below. >>> >>> >>> >>> I fully agree with you that the reason for even having a semantics is as >>> you put it, "to 1) learn with a few examples and 2) apply the knowledge to >>> a broad set of situations.? GPT-3 sometimes gives the appearance of having >>> done so, but it falls apart under close inspection, so the problem remains >>> unsolved. >>> >>> >>> >>> Gary >>> >>> >>> >>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >>> wrote: >>> >>> >>> >>> G. Hinton wrote: "I believe that any reasonable person would admit that >>> if you ask a neural net to draw a picture of a hamster wearing a red hat >>> and it draws such a picture, it understood the request." >>> >>> >>> >>> I would like to suggest why drawing a hamster with a red hat does not >>> necessarily imply understanding of the statement "hamster wearing a red >>> hat". >>> >>> To understand that "hamster wearing a red hat" would mean inferring, in >>> newly emerging situations of this hamster, all the real-life >>> implications that the red hat brings to the little animal. >>> >>> >>> >>> What would happen to the hat if the hamster rolls on its back? (Would >>> the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? >>> (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red >>> hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would >>> it be easier for predators to spot the hamster?) >>> >>> >>> >>> ...and so on. >>> >>> >>> >>> Countless many questions can be asked. One has understood "hamster >>> wearing a red hat" only if one can answer reasonably well many of such >>> real-life relevant questions. Similarly, a student has understood materias >>> in a class only if they can apply the materials in real-life situations >>> (e.g., applying Pythagora's theorem). If a student gives a correct answer >>> to a multiple choice question, we don't know whether the student understood >>> the material or whether this was just rote learning (often, it is rote >>> learning). >>> >>> >>> >>> I also suggest that understanding also comes together with effective >>> learning: We store new information in such a way that we can recall it >>> later and use it effectively i.e., make good inferences in newly emerging >>> situations based on this knowledge. >>> >>> >>> >>> In short: Understanding makes us humans able to 1) learn with a few >>> examples and 2) apply the knowledge to a broad set of situations. >>> >>> >>> >>> No neural network today has such capabilities and we don't know how to >>> give them such capabilities. Neural networks need large amounts of >>> training examples that cover a large variety of situations and then >>> the networks can only deal with what the training examples have already >>> covered. Neural networks cannot extrapolate in that 'understanding' sense. >>> >>> >>> >>> I suggest that understanding truly extrapolates from a piece of >>> knowledge. It is not about satisfying a task such as translation between >>> languages or drawing hamsters with hats. It is how you got the capability >>> to complete the task: Did you only have a few examples that covered >>> something different but related and then you extrapolated from that >>> knowledge? If yes, this is going in the direction of understanding. Have >>> you seen countless examples and then interpolated among them? Then perhaps >>> it is not understanding. >>> >>> >>> >>> So, for the case of drawing a hamster wearing a red hat, understanding >>> perhaps would have taken place if the following happened before that: >>> >>> >>> >>> 1) first, the network learned about hamsters (not many examples) >>> >>> 2) after that the network learned about red hats (outside the context of >>> hamsters and without many examples) >>> >>> 3) finally the network learned about drawing (outside of the context of >>> hats and hamsters, not many examples) >>> >>> >>> >>> After that, the network is asked to draw a hamster with a red hat. If it >>> does it successfully, maybe we have started cracking the problem of >>> understanding. >>> >>> >>> >>> Note also that this requires the network to learn sequentially without >>> exhibiting catastrophic forgetting of the previous knowledge, which is >>> possibly also a consequence of human learning by understanding. >>> >>> >>> >>> >>> >>> Danko >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> >>> --- A progress usually starts with an insight --- >>> >>> >>> >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> >>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>> >>> Without getting into the specific dispute between Gary and Geoff, I >>> think with approaches similar to GLOM, we are finally headed in the right >>> direction. There?s plenty of neurophysiological evidence for single-cell >>> abstractions and multisensory neurons in the brain, which one might claim >>> correspond to symbols. And I think we can finally reconcile the decades old >>> dispute between Symbolic AI and Connectionism. >>> >>> >>> >>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >>> an effort to wind up with encodings that effectively serve as symbols in >>> exactly that way, guaranteed to serve as consistent representations of >>> specific concepts.) >>> >>> GARY: I have *never* called for dismissal of neural networks, but >>> rather for some hybrid between the two (as you yourself contemplated in >>> 1991); the point of the 2001 book was to characterize exactly where >>> multilayer perceptrons succeeded and broke down, and where symbols could >>> complement them. >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> *From:* Connectionists *On >>> Behalf Of *Gary Marcus >>> *Sent:* Wednesday, February 2, 2022 1:26 PM >>> *To:* Geoffrey Hinton >>> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Dear Geoff, and interested others, >>> >>> >>> >>> What, for example, would you make of a system that often drew the >>> red-hatted hamster you requested, and perhaps a fifth of the time gave you >>> utter nonsense? Or say one that you trained to create birds but sometimes >>> output stuff like this: >>> >>> >>> >>> >>> >>> >>> >>> One could >>> >>> >>> >>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>> >>> or >>> >>> b. wonder if it might be possible that sometimes the system gets the >>> right answer for the wrong reasons (eg partial historical contingency), and >>> wonder whether another approach might be indicated. >>> >>> >>> >>> Benchmarks are harder than they look; most of the field has come to >>> recognize that. The Turing Test has turned out to be a lousy measure of >>> intelligence, easily gamed. It has turned out empirically that the Winograd >>> Schema Challenge did not measure common sense as well as Hector might have >>> thought. (As it happens, I am a minor coauthor of a very recent review on >>> this very topic: https://arxiv.org/abs/2201.02387 >>> ) >>> But its conquest in no way means machines now have common sense; many >>> people from many different perspectives recognize that (including, e.g., >>> Yann LeCun, who generally tends to be more aligned with you than with me). >>> >>> >>> >>> So: on the goalpost of the Winograd schema, I was wrong, and you can >>> quote me; but what you said about me and machine translation remains your >>> invention, and it is inexcusable that you simply ignored my 2019 >>> clarification. On the essential goal of trying to reach meaning and >>> understanding, I remain unmoved; the problem remains unsolved. >>> >>> >>> >>> All of the problems LLMs have with coherence, reliability, truthfulness, >>> misinformation, etc stand witness to that fact. (Their persistent inability >>> to filter out toxic and insulting remarks stems from the same.) I am hardly >>> the only person in the field to see that progress on any given benchmark >>> does not inherently mean that the deep underlying problems have solved. >>> You, yourself, in fact, have occasionally made that point. >>> >>> >>> >>> With respect to embeddings: Embeddings are very good for natural >>> language *processing*; but NLP is not the same as NL*U* ? when it comes >>> to *understanding*, their worth is still an open question. Perhaps they >>> will turn out to be necessary; they clearly aren?t sufficient. In their >>> extreme, they might even collapse into being symbols, in the sense of >>> uniquely identifiable encodings, akin to the ASCII code, in which a >>> specific set of numbers stands for a specific word or concept. (Wouldn?t >>> that be ironic?) >>> >>> >>> >>> (Your GLOM, which as you know I praised publicly, is in many ways an >>> effort to wind up with encodings that effectively serve as symbols in >>> exactly that way, guaranteed to serve as consistent representations of >>> specific concepts.) >>> >>> >>> >>> Notably absent from your email is any kind of apology for >>> misrepresenting my position. It?s fine to say that ?many people thirty >>> years ago once thought X? and another to say ?Gary Marcus said X in 2015?, >>> when I didn?t. I have consistently felt throughout our interactions that >>> you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) >>> apologized to me for having made that error. I am still not he. >>> >>> >>> >>> Which maybe connects to the last point; if you read my work, you would >>> see thirty years of arguments *for* neural networks, just not in the >>> way that you want them to exist. I have ALWAYS argued that there is a role >>> for them; characterizing me as a person ?strongly opposed to neural >>> networks? misses the whole point of my 2001 book, which was subtitled >>> ?Integrating Connectionism and Cognitive Science.? >>> >>> >>> >>> In the last two decades or so you have insisted (for reasons you have >>> never fully clarified, so far as I know) on abandoning symbol-manipulation, >>> but the reverse is not the case: I have *never* called for dismissal of >>> neural networks, but rather for some hybrid between the two (as you >>> yourself contemplated in 1991); the point of the 2001 book was to >>> characterize exactly where multilayer perceptrons succeeded and broke down, >>> and where symbols could complement them. It?s a rhetorical trick (which is >>> what the previous thread was about) to pretend otherwise. >>> >>> >>> >>> Gary >>> >>> >>> >>> >>> >>> On Feb 2, 2022, at 11:22, Geoffrey Hinton >>> wrote: >>> >>> ? >>> >>> Embeddings are just vectors of soft feature detectors and they are very >>> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >>> opposite. >>> >>> >>> >>> A few decades ago, everyone I knew then would have agreed that the >>> ability to translate a sentence into many different languages was strong >>> evidence that you understood it. >>> >>> >>> >>> But once neural networks could do that, their critics moved the >>> goalposts. An exception is Hector Levesque who defined the goalposts more >>> sharply by saying that the ability to get pronoun references correct in >>> Winograd sentences is a crucial test. Neural nets are improving at that but >>> still have some way to go. Will Gary agree that when they can get pronoun >>> references correct in Winograd sentences they really do understand? Or does >>> he want to reserve the right to weasel out of that too? >>> >>> >>> >>> Some people, like Gary, appear to be strongly opposed to neural networks >>> because they do not fit their preconceived notions of how the mind should >>> work. >>> >>> I believe that any reasonable person would admit that if you ask a >>> neural net to draw a picture of a hamster wearing a red hat and it draws >>> such a picture, it understood the request. >>> >>> >>> >>> Geoff >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>> >>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger >>> neural network community, >>> >>> >>> >>> There has been a lot of recent discussion on this list about framing and >>> scientific integrity. Often the first step in restructuring narratives is >>> to bully and dehumanize critics. The second is to misrepresent their >>> position. People in positions of power are sometimes tempted to do this. >>> >>> >>> >>> The Hinton-Hanson interview that you just published is a real-time >>> example of just that. It opens with a needless and largely content-free >>> personal attack on a single scholar (me), with the explicit intention of >>> discrediting that person. Worse, the only substantive thing it says is >>> false. >>> >>> >>> >>> Hinton says ?In 2015 he [Marcus] made a prediction that computers >>> wouldn?t be able to do machine translation.? >>> >>> >>> >>> I never said any such thing. >>> >>> >>> >>> What I predicted, rather, was that multilayer perceptrons, as they >>> existed then, would not (on their own, absent other mechanisms) >>> *understand* language. Seven years later, they still haven?t, except in >>> the most superficial way. >>> >>> >>> >>> I made no comment whatsoever about machine translation, which I view as >>> a separate problem, solvable to a certain degree by correspondance without >>> semantics. >>> >>> >>> >>> I specifically tried to clarify Hinton?s confusion in 2019, but, >>> disappointingly, he has continued to purvey misinformation despite that >>> clarification. Here is what I wrote privately to him then, which should >>> have put the matter to rest: >>> >>> >>> >>> You have taken a single out of context quote [from 2015] and >>> misrepresented it. The quote, which you have prominently displayed at the >>> bottom on your own web page, says: >>> >>> >>> >>> Hierarchies of features are less suited to challenges such as language, >>> inference, and high-level planning. For example, as Noam Chomsky famously >>> pointed out, language is filled with sentences you haven't seen >>> before. Pure classifier systems don't know what to do with such sentences. >>> The talent of feature detectors -- in identifying which member of some >>> category something belongs to -- doesn't translate into understanding >>> novel sentences, in which each sentence has its own unique meaning. >>> >>> >>> >>> It does *not* say "neural nets would not be able to deal with novel >>> sentences"; it says that hierachies of features detectors (on their own, if >>> you read the context of the essay) would have trouble *understanding *novel sentences. >>> >>> >>> >>> >>> Google Translate does yet not *understand* the content of the sentences >>> is translates. It cannot reliably answer questions about who did what to >>> whom, or why, it cannot infer the order of the events in paragraphs, it >>> can't determine the internal consistency of those events, and so forth. >>> >>> >>> >>> Since then, a number of scholars, such as the the computational linguist >>> Emily Bender, have made similar points, and indeed current LLM difficulties >>> with misinformation, incoherence and fabrication all follow from these >>> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >>> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf >>> , >>> also emphasizing issues of understanding and meaning: >>> >>> >>> >>> *The success of the large neural language models on many NLP tasks is >>> exciting. However, we find that these successes sometimes lead to hype in >>> which these models are being described as ?understanding? language or >>> capturing ?meaning?. In this position paper, we argue that a system trained >>> only on form has a priori no way to learn meaning. .. a clear understanding >>> of the distinction between form and meaning will help guide the field >>> towards better science around natural language understanding. * >>> >>> >>> >>> Her later article with Gebru on language models ?stochastic parrots? is >>> in some ways an extension of this point; machine translation requires >>> mimicry, true understanding (which is what I was discussing in 2015) >>> requires something deeper than that. >>> >>> >>> >>> Hinton?s intellectual error here is in equating machine translation with >>> the deeper comprehension that robust natural language understanding will >>> require; as Bender and Koller observed, the two appear not to be the same. >>> (There is a longer discussion of the relation between language >>> understanding and machine translation, and why the latter has turned out to >>> be more approachable than the former, in my 2019 book with Ernest Davis). >>> >>> >>> >>> More broadly, Hinton?s ongoing dismissiveness of research from >>> perspectives other than his own (e.g. linguistics) have done the field a >>> disservice. >>> >>> >>> >>> As Herb Simon once observed, science does not have to be zero-sum. >>> >>> >>> >>> Sincerely, >>> >>> Gary Marcus >>> >>> Professor Emeritus >>> >>> New York University >>> >>> >>> >>> On Feb 2, 2022, at 06:12, AIhub wrote: >>> >>> ? >>> >>> Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> In the latest episode of this video series for AIhub.org >>> , >>> Stephen Hanson talks to Geoff Hinton about neural networks, >>> backpropagation, overparameterization, digit recognition, voxel cells, >>> syntax and semantics, Winograd sentences, and more. >>> >>> >>> >>> You can watch the discussion, and read the transcript, here: >>> >>> >>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>> >>> >>> >>> >>> About AIhub: >>> >>> AIhub is a non-profit dedicated to connecting the AI community to the >>> public by providing free, high-quality information through AIhub.org >>> >>> (https://aihub.org/ >>> ). >>> We help researchers publish the latest AI news, summaries of their work, >>> opinion pieces, tutorials and more. We are supported by many leading >>> scientific organizations in AI, namely AAAI >>> , >>> NeurIPS >>> , >>> ICML >>> , >>> AIJ >>> >>> /IJCAI >>> , >>> ACM SIGAI >>> , >>> EurAI/AICOMM, CLAIRE >>> >>> and RoboCup >>> >>> . >>> >>> Twitter: @aihuborg >>> >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> >> -- > Gary Cottrell 858-534-6640 FAX: 858-534-7029 > Computer Science and Engineering 0404 > IF USING FEDEX INCLUDE THE FOLLOWING LINE: > CSE Building, Room 4130 > University of California San Diego - > 9500 Gilman Drive # 0404 > La Jolla, Ca. 92093-0404 > > Email: gary at ucsd.edu > Home page: http://www-cse.ucsd.edu/~gary/ > Schedule: http://tinyurl.com/b7gxpwo > > Blind certainty - a close-mindedness that amounts to an imprisonment so > total, that the prisoner doesn?t even know that he?s locked up. -David > Foster Wallace > > > Power to the people! ?Patti Smith > > Except when they?re delusional ?Gary Cottrell > > > This song makes me nostalgic for a memory I don't have -- Tess Cottrell > > > > > > > > > > > *Listen carefully,Neither the VedasNor the Qur'anWill teach you this:Put > the bit in its mouth,The saddle on its back,Your foot in the stirrup,And > ride your wild runaway mindAll the way to heaven.* > > -- Kabir > -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Sat Feb 5 11:26:29 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Sat, 5 Feb 2022 08:26:29 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Presumably we can all agree that the brain is neural network of some form. But which? Danko?s point was about functional limits on ?neural networks, as we know them today?. Neural networks as we know them today do not strongly resemble human brain in their structure, development, or functionality. We need to expand our search space. I fully concur with Danko?s concluding paragraph. > On Feb 4, 2022, at 13:46, gary at eng.ucsd.edu wrote: > > ? > This is an argument from lack of imagination, as Pat Churchland used to say. All you have to notice, is that your brain is a neural net work. What are the alternatives? > >> On Fri, Feb 4, 2022 at 4:08 AM Danko Nikolic wrote: >> >> I suppose everyone agrees that "the brain is a physical system", >> and that "There is no ?magic? inside the brain", >> and that '?understanding? is just part of ?learning.?' >> >> Also, we can agree that some sort of simulation takes place behind understanding. >> >> However, there still is a problem: Neural network's can't implement the needed simulations; they cannot achieve the same cognitive effect that human minds can (or animal minds can). >> >> We don't know a way of wiring a neural network such that it could perform the simulations (understandings) necessary to find the answers to real-life questions, such as the hamster with a hat problem. >> >> In other words, neural networks, as we know them today, cannot: >> >> 1) learn from a small number of examples (simulation or not) >> 2) apply the knowledge to a wide range of situations >> >> >> We, as scientists, do not understand understanding. Our technology's simulations (their depth of understanding) are no match for the simulations (depth of understanding) that the biological brain performs. >> >> I think that scientific integrity also covers acknowledging when we did not (yet) succeed in solving a certain problem. There is still significant work to be done. >> >> >> Danko >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> https://www.linkedin.com/in/danko-nikolic/ >> --- A progress usually starts with an insight --- >> >> >> Virenfrei. www.avast.com >> >>> On Thu, Feb 3, 2022 at 9:35 PM Asim Roy wrote: >>> First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. >>> >>> >>> >>> DANKO: >>> >>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> From: Gary Marcus >>> Sent: Thursday, February 3, 2022 9:26 AM >>> To: Danko Nikolic >>> Cc: Asim Roy ; Geoffrey Hinton ; AIhub ; connectionists at mailman.srv.cs.cmu.edu >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> Dear Danko, >>> >>> >>> >>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. >>> >>> >>> >>> I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. >>> >>> >>> >>> Gary >>> >>> >>> >>> >>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: >>> >>> >>> >>> G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." >>> >>> >>> >>> I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". >>> >>> To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. >>> >>> >>> >>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>> >>> >>> >>> ...and so on. >>> >>> >>> >>> Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). >>> >>> >>> >>> I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. >>> >>> >>> >>> In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. >>> >>> >>> >>> No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. >>> >>> >>> >>> I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. >>> >>> >>> >>> So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: >>> >>> >>> >>> 1) first, the network learned about hamsters (not many examples) >>> >>> 2) after that the network learned about red hats (outside the context of hamsters and without many examples) >>> >>> 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) >>> >>> >>> >>> After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. >>> >>> >>> >>> Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. >>> >>> >>> >>> >>> >>> Danko >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> --- A progress usually starts with an insight --- >>> >>> >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>> >>> Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. >>> >>> >>> >>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>> >>> GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> From: Connectionists On Behalf Of Gary Marcus >>> Sent: Wednesday, February 2, 2022 1:26 PM >>> To: Geoffrey Hinton >>> Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> Dear Geoff, and interested others, >>> >>> >>> >>> What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: >>> >>> >>> >>> >>> >>> >>> >>> One could >>> >>> >>> >>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>> >>> or >>> >>> b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. >>> >>> >>> >>> Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). >>> >>> >>> >>> So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. >>> >>> >>> >>> All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. >>> >>> >>> >>> With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) >>> >>> >>> >>> (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>> >>> >>> >>> Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. >>> >>> >>> >>> Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? >>> >>> >>> >>> In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. >>> >>> >>> >>> Gary >>> >>> >>> >>> >>> >>> On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: >>> >>> ? >>> >>> Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. >>> >>> >>> >>> A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. >>> >>> >>> >>> But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? >>> >>> >>> >>> Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. >>> >>> I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. >>> >>> >>> >>> Geoff >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>> >>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, >>> >>> >>> >>> There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. >>> >>> >>> >>> The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. >>> >>> >>> >>> Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? >>> >>> >>> >>> I never said any such thing. >>> >>> >>> >>> What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. >>> >>> >>> >>> I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. >>> >>> >>> >>> I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: >>> >>> >>> >>> You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: >>> >>> >>> >>> Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. >>> >>> >>> >>> It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. >>> >>> >>> >>> Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. >>> >>> >>> >>> Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: >>> >>> >>> >>> The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. >>> >>> >>> >>> Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. >>> >>> >>> >>> Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). >>> >>> >>> >>> More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. >>> >>> >>> >>> As Herb Simon once observed, science does not have to be zero-sum. >>> >>> >>> >>> Sincerely, >>> >>> Gary Marcus >>> >>> Professor Emeritus >>> >>> New York University >>> >>> >>> >>> On Feb 2, 2022, at 06:12, AIhub wrote: >>> >>> ? >>> >>> Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. >>> >>> >>> >>> You can watch the discussion, and read the transcript, here: >>> >>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>> >>> >>> >>> About AIhub: >>> >>> AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. >>> >>> Twitter: @aihuborg >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> > -- > Gary Cottrell 858-534-6640 FAX: 858-534-7029 > Computer Science and Engineering 0404 > IF USING FEDEX INCLUDE THE FOLLOWING LINE: > CSE Building, Room 4130 > University of California San Diego - > 9500 Gilman Drive # 0404 > La Jolla, Ca. 92093-0404 > > Email: gary at ucsd.edu > Home page: http://www-cse.ucsd.edu/~gary/ > Schedule: http://tinyurl.com/b7gxpwo > > Blind certainty - a close-mindedness that amounts to an imprisonment so total, that the prisoner doesn?t even know that he?s locked up. -David Foster Wallace > > Power to the people! ?Patti Smith > Except when they?re delusional ?Gary Cottrell > > This song makes me nostalgic for a memory I don't have -- Tess Cottrell > > Listen carefully, > Neither the Vedas > Nor the Qur'an > Will teach you this: > Put the bit in its mouth, > The saddle on its back, > Your foot in the stirrup, > And ride your wild runaway mind > All the way to heaven. > -- Kabir -------------- next part -------------- An HTML attachment was scrubbed... URL: From minaiaa at gmail.com Sat Feb 5 00:45:52 2022 From: minaiaa at gmail.com (Ali Minai) Date: Sat, 5 Feb 2022 00:45:52 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: I say that when a machine can correctly translate the meaning of a Persian ghazal by Rumi or Hafez - with all its symbolism, metaphor, historical allusions, etc, - into English, it has understood the language. Until then, it is just superficial mimicry. As of now, when I want to get a laugh on social media, I take a bit of Persian or Urdu (or Hindi or Arabic for that matter), have Google or Facebook translate it, and post it. Without fail, it is hilarious and "not even wrong". A distributional view of language without grounding in embodied experience, historical knowledge, cultural knowledge, etc., can only go so far. It is surprising how far it has gone, but nowhere near far enough. This does not mean that I am arguing for the impossibility of understanding by machines. Or that I am arguing against a neural approach - which I think is the right one. We just happen to have taken a very superficial view of how a system built from neurons should do things to realize a mind, and, thanks to our computational muscle, have taken that superficial vision to an extreme. There are notable successes - especially where the models actually follow the principles that the brain uses, however simplistically. CNN-based networks are a good example; they are a reasonable approximation to the early visual system. But until we bring a deep understanding of the biological organism - including evolution and development - to how we try to build minds from matter, we will remain mired in these superficial successes and impressive irrelevancies like learning to play chess and Go. One problem is that there are two mutually incompatible visions of AI. One is to build an actual intelligence - autonomous, self-motivated, understanding, creative. The other is to build "smart" tools for human purposes, to the point of including purposes that today require mental work, such as translation, video captioning, recommendation, summarization, driving etc., just as, until recently, arithmetic required mental work. In the near future, AI may well write legal briefs, prescribe medication, and teach courses fairly competently. But this is all the work of smart, obedient servants learning the craft, not autonomous, creative intelligences comparable to our own - or even to a bird at this point. Unfortunately, trying to do the latter has no immediate payoff, which is what today's AI is mostly about. As for academic research, most people seem more focused on getting that additional 0.8% in performance on a benchmark dataset by adding a few million parameters to a model that already has 200 million parameters, all so they can get a publication in ICLR or ACL. Very few people have the incentive or the inclination to focus on addressing fundamental conceptual issues that would get us out of our current local conceptual minimum in AI but solve no lucrative problem or meet any popular benchmark. Ali *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Thu, Feb 3, 2022 at 2:24 AM Geoffrey Hinton wrote: > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > Geoff > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > >> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural >> network community, >> >> There has been a lot of recent discussion on this list about framing and >> scientific integrity. Often the first step in restructuring narratives is >> to bully and dehumanize critics. The second is to misrepresent their >> position. People in positions of power are sometimes tempted to do this. >> >> The Hinton-Hanson interview that you just published is a real-time >> example of just that. It opens with a needless and largely content-free >> personal attack on a single scholar (me), with the explicit intention of >> discrediting that person. Worse, the only substantive thing it says is >> false. >> >> Hinton says ?In 2015 he [Marcus] made a prediction that computers >> wouldn?t be able to do machine translation.? >> >> I never said any such thing. >> >> What I predicted, rather, was that multilayer perceptrons, as they >> existed then, would not (on their own, absent other mechanisms) >> understand language. Seven years later, they still haven?t, except in >> the most superficial way. >> >> I made no comment whatsoever about machine translation, which I view as a >> separate problem, solvable to a certain degree by correspondance without >> semantics. >> >> I specifically tried to clarify Hinton?s confusion in 2019, but, >> disappointingly, he has continued to purvey misinformation despite that >> clarification. Here is what I wrote privately to him then, which should >> have put the matter to rest: >> >> You have taken a single out of context quote [from 2015] and >> misrepresented it. The quote, which you have prominently displayed at the >> bottom on your own web page, says: >> >> Hierarchies of features are less suited to challenges such as language, >> inference, and high-level planning. For example, as Noam Chomsky famously >> pointed out, language is filled with sentences you haven't seen >> before. Pure classifier systems don't know what to do with such sentences. >> The talent of feature detectors -- in identifying which member of some >> category something belongs to -- doesn't translate into understanding >> novel sentences, in which each sentence has its own unique meaning. >> >> It does not say "neural nets would not be able to deal with novel >> sentences"; it says that hierachies of features detectors (on their own, if >> you read the context of the essay) would have trouble understanding novel sentences. >> >> >> Google Translate does yet not understand the content of the sentences is >> translates. It cannot reliably answer questions about who did what to whom, >> or why, it cannot infer the order of the events in paragraphs, it can't >> determine the internal consistency of those events, and so forth. >> >> Since then, a number of scholars, such as the the computational linguist >> Emily Bender, have made similar points, and indeed current LLM difficulties >> with misinformation, incoherence and fabrication all follow from these >> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, >> also emphasizing issues of understanding and meaning: >> >> The success of the large neural language models on many NLP tasks is >> exciting. However, we find that these successes sometimes lead to hype in >> which these models are being described as ?understanding? language or >> capturing ?meaning?. In this position paper, we argue that a system trained >> only on form has a priori no way to learn meaning. .. a clear understanding >> of the distinction between form and meaning will help guide the field >> towards better science around natural language understanding. >> >> Her later article with Gebru on language models ?stochastic parrots? is >> in some ways an extension of this point; machine translation requires >> mimicry, true understanding (which is what I was discussing in 2015) >> requires something deeper than that. >> >> Hinton?s intellectual error here is in equating machine translation with >> the deeper comprehension that robust natural language understanding will >> require; as Bender and Koller observed, the two appear not to be the same. >> (There is a longer discussion of the relation between language >> understanding and machine translation, and why the latter has turned out to >> be more approachable than the former, in my 2019 book with Ernest Davis). >> >> More broadly, Hinton?s ongoing dismissiveness of research from >> perspectives other than his own (e.g. linguistics) have done the field a >> disservice. >> >> As Herb Simon once observed, science does not have to be zero-sum. >> >> Sincerely, >> Gary Marcus >> Professor Emeritus >> New York University >> >> On Feb 2, 2022, at 06:12, AIhub wrote: >> >> ? >> Stephen Hanson in conversation with Geoff Hinton >> >> In the latest episode of this video series for AIhub.org, Stephen Hanson >> talks to Geoff Hinton about neural networks, backpropagation, >> overparameterization, digit recognition, voxel cells, syntax and semantics, >> Winograd sentences, and more. >> >> You can watch the discussion, and read the transcript, here: >> >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> >> About AIhub: >> AIhub is a non-profit dedicated to connecting the AI community to the >> public by providing free, high-quality information through AIhub.org ( >> https://aihub.org/ >> ). >> We help researchers publish the latest AI news, summaries of their work, >> opinion pieces, tutorials and more. We are supported by many leading >> scientific organizations in AI, namely AAAI >> , >> NeurIPS >> , >> ICML >> , >> AIJ >> >> /IJCAI >> , >> ACM SIGAI >> , >> EurAI/AICOMM, CLAIRE >> >> and RoboCup >> >> . >> Twitter: @aihuborg >> >> -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Sat Feb 5 04:30:55 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sat, 5 Feb 2022 09:30:55 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: All, I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? https://www.youtube.com/watch?v=gn4nRCC9TwQ https://www.youtube.com/watch?v=8sO7VS3q8d0 Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Ali Minai Sent: Friday, February 4, 2022 11:38 PM To: Asim Roy Cc: Gary Marcus ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Asim Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! Ali Ali A. Minai, Ph.D. Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Fri, Feb 4, 2022 at 2:42 AM Asim Roy > wrote: First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. DANKO: What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Gary Marcus > Sent: Thursday, February 3, 2022 9:26 AM To: Danko Nikolic > Cc: Asim Roy >; Geoffrey Hinton >; AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub > wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From arbib at usc.edu Sat Feb 5 17:54:04 2022 From: arbib at usc.edu (Michael Arbib) Date: Sat, 5 Feb 2022 22:54:04 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: I think this is unhelpful, Gary. The question is not whether our brains can do certain things (not always very well) using neurons that, e.g., co-exist with glia, are cells with nuclei containing complex genetic detail, and have axons whose behavior can be well described by the Hodgkin-Huxley equations ? it is whether networks of a specific kind of ?something with only very superficial similarity to our neurons? can do certain tasks. As a complementary theme, the history of computers starts with implementing numerical computations in terms of zeros and ones, and then inventing ever more subtle ways to organize computation in terms of higher-level structures whose relations to 0s and 1s is of limited importance in the sense that we program now in terms of high-level constructs, not in machine language. Putting the two together, the issue is ?what are the high-level languages of brain operation?? If we wish to implement them in silicon their relation to biological neurons may be as secondary as the relation of the implementation to 0s and 1s. That?s why, in my own work, I explore networks of ?schemas? as well as networks of ?neurons? in seeking to understand the cognitive neuroscience of vision, action, language, etc. From: Connectionists On Behalf Of gary at ucsd.edu Sent: Friday, February 4, 2022 10:20 AM To: Danko Nikolic Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton This is an argument from lack of imagination, as Pat Churchland used to say. All you have to notice, is that your brain is a neural net work. What are the alternatives? -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Sat Feb 5 23:56:27 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sun, 6 Feb 2022 04:56:27 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <303504A2-453D-4DE8-8A34-C41693041954@nyu.edu> References: <303504A2-453D-4DE8-8A34-C41693041954@nyu.edu> Message-ID: There was another recent attempt to take down the grandmother cell idea: Frontiers | The Value of Failure in Science: The Story of Grandmother Cells in Neuroscience | Neuroscience (frontiersin.org) And here?s my commentary defending grandmother cells: Frontiers | Commentary: The Value of Failure in Science: The Story of Grandmother Cells in Neuroscience | Neuroscience (frontiersin.org) By the way, we have had vigorous private arguments over the years about grandmother cells and many were involved ? from Jay McClelland and Christof Koch to Walter Freeman and Bernard Baars. As far as I can tell, the brain uses abstractions at the single cell level that can be argued to be symbols. Short arguments are in the commentary, based on observations by neurophysiologists themselves. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Gary Marcus Sent: Friday, February 4, 2022 12:53 PM To: Stephen Jos? Hanson Cc: connectionists at mailman.srv.cs.cmu.edu; Stevan Harnad ; Francesca Rossi2 ; Artur Garcez ; Anima Anandkumar ; Luis Lamb ; Gadi Singer ; Josh Tenenbaum ; AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton ?Steve, The phrase I always liked was ?poverty of the imagination arguments?; I share your disdain for them. But that?s why I think you should be careful of any retreat into biological plausibility. As even Jay McClelland has acknowledged, we do know that some humans some of the time manipulate symbols. So wetware-based symbols are not literally biologically impossible; the real question for cognitive neuroscience is about the scope and development of symbols. For engineering, the real question is, are they useful. Certainly for software engineering in general, they are indispensable. Beyond this, none of the available AI approaches map particularly neatly onto what we know about the brain, and none of what we know about the brain is understood well enough to solve AI. All the examples you point to, for instance, are actually controversial, not decisive. As you probably know, for example, Nancy Kanwisher has a different take on domain-specificity than you do (https://web.mit.edu/bcs/nklab/), with evidence of specialization early in life, and Jeff Bowers has argued that the grandmother cell hypothesis has been dismissed prematurely (https://jeffbowers.blogs.bristol.ac.uk/blog/grandmother-cells/); there?s also a long literature on the possible neural realization of rules, both in humans and other animals. I don?t know what the right answers are there, but nor do I think that neurosymbolic systems are beholden to them anymore than CNNs are bound to whether or not the brain performs back-propagation. Finally, as a reminder, ?Distributed? per se in not the right question; in some technical sense ASCII encodings are distributed, and about as symbolic as you can get. The proper question is really what you do with your encodings; the neurosymbolic approach is trying to broaden the available range of options. Gary On Feb 4, 2022, at 07:04, Stephen Jos? Hanson > wrote: ? Well I don't like counterfactual arguments or ones that start with "It can't be done with neural networks.."--as this amounts to the old Rumelhart saw, of "proof by lack of imagination". I think my position and others (I can't speak for Geoff and won't) is more of a "purist" view that brains have computationally complete representational power to do what ever is required of human level mental processing. AI symbol systems are remote descriptions of this level of processing. Looking at 1000s of brain scans, one begins to see a pattern of interacting large and smaller scale networks, probably related to Resting state and the Default Mode networks in some important competitive way. But what one doesn't find is modular structure (e.g. face area.. nope) or evidence of "symbols" being processed. Research on Numbers is interesting in this regard, as number representation should provide some evidence of discrete symbol processing as would letters. But again the processing states from brain imaging more generally appear to be distributed representations of some sort. One other direction has to do with prior rules that could be neurally coded and therefore provide an immediate bias in learning and thus dramatically reduce the number of examples required for asymptotic learning. Some of this has been done with pre-training-- on let's say 1000s of videos that are relatively generic, prior to learning on a small set of videos related to a specific topic-- say two individuals playing a monopoly game. In that case, no game-like videos were sampled in the pre-training, and the LSTM was trained to detect change point on 2 minutes of video, achieving a 97% match with human parsers. In these senses I have no problem with this type of hybrid training. Steve On 2/4/22 9:07 AM, Gary Marcus wrote: ?The whole point of the neurosymbolic approach is to develop systems that can accommodate both vectors and symbols, since neither on their own seems adequate. If there are arguments against trying to do that, we would be interested. On Feb 4, 2022, at 4:17 AM, Stephen Jos? Hanson wrote: ? Geoff's position is pretty clear. He said in the conversation we had and in this thread, "vectors of soft features", Some of my claim is in several of the conversations with Mike Jordan and Rich Sutton, but briefly, there are a number of very large costly efforts from the 1970s and 1980s, to create, deploy and curate symbol AI systems that were massive failures. Not counterfactuals, but factuals that failed. The MCC comes to mind with Adm Bobby Inmann's national US mandate to counter the Japanese so called"Fifth-generation AI systems" as a massive failure of symbolic AI. -------------------- In 1982, Japan launched its Fifth Generation Computer Systems project (FGCS), designed to develop intelligent software that would run on novel computer hardware. As the first national, large-scale artificial intelligence (AI) research and development (R&D) project to be free from military influence and corporate profit motives, the FGCS was open, international, and oriented around public goods. On 2/3/22 6:34 PM, Francesca Rossi2 wrote: Hi all. Thanks Gary for adding me to this thread. I also would be interested in knowing why Steve thinks that NS AI did not work in the past, and why this is an indication that it cannot work now or in the future. Thanks, Francesca. ------------------ Francesca Rossi IBM Fellow and AI Ethics Global Leader T.J. Watson Research Center, Yorktown Heights, USA +1-617-3869639 ________________________________________ From: Artur Garcez Sent: Thursday, February 3, 2022 6:00 PM To: Gary Marcus Cc: Stephen Jos? Hanson; Geoffrey Hinton; AIhub; connectionists at mailman.srv.cs.cmu.edu; Luis Lamb; Josh Tenenbaum; Anima Anandkumar; Francesca Rossi2; Swarat Chaudhuri; Gadi Singer Subject: [EXTERNAL] Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally ZjQcmQRYFpfptBannerStart This Message Is From an External Sender This message came from outside your organization. ZjQcmQRYFpfptBannerEnd It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally symbolic), there is a clear resurgence of neuro-symbolic approaches. Best wishes, Artur On Thu, Feb 3, 2022 at 5:00 PM Gary Marcus > wrote: Steve, I?d love to hear you elaborate on this part, Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. I?d love your take on why you think the impetus towards hybrid models ended badly before, and why you think that the mistakes of the past can?t be corrected. Also it? would be really instructive to compare with deep learning, which lost steam for quite some time, but reemerged much stronger than ever before. Might not the same happen with hybrid models? I am cc?ing some folks (possibly not on this list) who have recently been sympathetic to hybrid models, in hopes of a rich discussion. (And, Geoff, still cc?d, I?d genuinely welcome your thoughts if you want to add them, despite our recent friction.) Cheers, Gary On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson > wrote: I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. Best and hope you are doing well. Steve -- -- [cid:image001.png at 01D81AD9.13EC7A80] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 19957 bytes Desc: image001.png URL: From gary.marcus at nyu.edu Fri Feb 4 17:21:46 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Fri, 4 Feb 2022 14:21:46 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <86E5466E-2316-4046-921E-5DCCCAD8E539@nyu.edu> Geoff and I may not be so far apart as either of us assumed. He seems willing to take on board symbols in the first sense that he describes below, centering around hierarchy and wholes in parts, which is at the heart of what I have been advocating for, and I am comfortable with his concerns about the limitations of purely arbitrary symbols. Neither of us sees much hope in pure rule-based approaches, and we both see the value of ML. I also agree that the recognition of a ?2? is unlikely to ground in any simple causal, symbolic description; that?s part of why I think we need the ?neuro-? in neurosymbolic. And, as Geoff notes below, we also need to be able to deal with compositions of wholes out of parts (and other wholes), and that?s part of why I expect to see symbols as part of the final picture. All that said, I don?t think that fans of symbols need be committed to symbols being purely arbitrary, and perhaps the thing that Geoff suspects might be a bone of contention need not be. The ASCII code, for example, is a symbolic code (in that each character has a unique encoding) that has a partial similarity space, with capital letters forming one cluster, lower cases letters forming another, and digits a third. Binary digits as symbolic representations for (eg) integers are also entirely non-arbitrary. Why restrict ourselves to purely arbitrary symbols? To Geoff?s point, it is interesting that humans are seemingly more facile with symbols that are non-arbitrary, as one can seen for example in the old literature on the Wason card selection task. I don?t think machines have to behave here like humans, but humans are an interesting proof of concept. If we can develop a bridge between neural networks and symbols that passes through stable yet semantically meaningful embeddings/symbols on the ways to structure, compositionality and operations over variables, perhaps we can both feel vindicated. Gary > On Feb 4, 2022, at 12:24, Geoffrey Hinton wrote: > > ? > I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. > > This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. > > >> On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas wrote: >> ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. >> >> >> >> For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. >> >> >> >> But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. >> >> >> >> --Tom >> >> >> >> Thomas G. Dietterich, Distinguished Professor Emeritus >> >> School of Electrical Engineering and Computer Science >> >> US Mail: 1148 Kelley Engineering Center >> >> >> >> Office: 2067 Kelley Engineering Center >> >> Oregon State Univ., Corvallis, OR 97331-5501 >> >> Voice: 541-737-5559; FAX: 541-737-1300 >> >> URL: http://web.engr.oregonstate.edu/~tgd/ >> >> >> >> From: Connectionists On Behalf Of Gary Marcus >> Sent: Thursday, February 3, 2022 8:26 AM >> To: Danko Nikolic >> Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub >> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >> >> >> >> [This email originated from outside of OSU. Use caution with links and attachments.] >> >> Dear Danko, >> >> >> >> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. >> >> >> >> I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. >> >> >> >> Gary >> >> >> >> >> On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: >> >> >> >> G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." >> >> >> >> I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". >> >> To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. >> >> >> >> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >> >> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >> >> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >> >> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >> >> >> >> ...and so on. >> >> >> >> Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). >> >> >> >> I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. >> >> >> >> In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. >> >> >> >> No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. >> >> >> >> I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. >> >> >> >> So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: >> >> >> >> 1) first, the network learned about hamsters (not many examples) >> >> 2) after that the network learned about red hats (outside the context of hamsters and without many examples) >> >> 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) >> >> >> >> After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. >> >> >> >> Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. >> >> >> >> >> >> Danko >> >> >> >> >> >> >> >> >> >> >> >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> https://www.linkedin.com/in/danko-nikolic/ >> >> --- A progress usually starts with an insight --- >> >> >> >> >> >> >> >> Virus-free. www.avast.com >> >> >> >> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >> >> Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. >> >> >> >> GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >> >> GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> From: Connectionists On Behalf Of Gary Marcus >> Sent: Wednesday, February 2, 2022 1:26 PM >> To: Geoffrey Hinton >> Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu >> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >> >> >> >> Dear Geoff, and interested others, >> >> >> >> What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: >> >> >> >> >> >> >> >> One could >> >> >> >> a. avert one?s eyes and deem the anomalous outputs irrelevant >> >> or >> >> b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. >> >> >> >> Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). >> >> >> >> So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. >> >> >> >> All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. >> >> >> >> With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) >> >> >> >> (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >> >> >> >> Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. >> >> >> >> Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? >> >> >> >> In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. >> >> >> >> Gary >> >> >> >> >> >> On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: >> >> ? >> >> Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. >> >> >> >> A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. >> >> >> >> But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? >> >> >> >> Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. >> >> I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. >> >> >> >> Geoff >> >> >> >> >> >> >> >> >> >> >> >> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >> >> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, >> >> >> >> There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. >> >> >> >> The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. >> >> >> >> Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? >> >> >> >> I never said any such thing. >> >> >> >> What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. >> >> >> >> I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. >> >> >> >> I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: >> >> >> >> You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: >> >> >> >> Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. >> >> >> >> It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. >> >> >> >> Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. >> >> >> >> Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: >> >> >> >> The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. >> >> >> >> Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. >> >> >> >> Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). >> >> >> >> More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. >> >> >> >> As Herb Simon once observed, science does not have to be zero-sum. >> >> >> >> Sincerely, >> >> Gary Marcus >> >> Professor Emeritus >> >> New York University >> >> >> >> On Feb 2, 2022, at 06:12, AIhub wrote: >> >> ? >> >> Stephen Hanson in conversation with Geoff Hinton >> >> >> >> In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. >> >> >> >> You can watch the discussion, and read the transcript, here: >> >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> >> >> About AIhub: >> >> AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. >> >> Twitter: @aihuborg >> >> >> >> >> >> Virus-free. www.avast.com >> >> -------------- next part -------------- An HTML attachment was scrubbed... URL: From david at irdta.eu Sat Feb 5 09:15:23 2022 From: david at irdta.eu (David Silva - IRDTA) Date: Sat, 5 Feb 2022 15:15:23 +0100 (CET) Subject: Connectionists: DeepLearn 2022 Summer: early registration February 23 Message-ID: <2133773186.2471935.1644070523476@webmail.strato.com> ****************************************************************** 6th INTERNATIONAL GRAN CANARIA SCHOOL ON DEEP LEARNING DeepLearn 2022 Summer Las Palmas de Gran Canaria, Spain July 25-29, 2022 https://irdta.eu/deeplearn/2022su/ ***************** Co-organized by: University of Las Palmas de Gran Canaria Institute for Research Development, Training and Advice ? IRDTA Brussels/London ****************************************************************** Early registration: February 23, 2022 ****************************************************************** SCOPE: DeepLearn 2022 Summer will be a research training event with a global scope aiming at updating participants on the most recent advances in the critical and fast developing area of deep learning. Previous events were held in Bilbao, Genova, Warsaw, Las Palmas de Gran Canaria, Bournemouth, and Guimar?es. Deep learning is a branch of artificial intelligence covering a spectrum of current frontier research and industrial innovation that provides more efficient algorithms to deal with large-scale data in a huge variety of environments: computer vision, neurosciences, speech recognition, language processing, human-computer interaction, drug discovery, biomedical informatics, image analysis, recommender systems, advertising, fraud detection, robotics, games, finance, biotechnology, physics experiments, biometrics, communications, climate sciences, etc. etc. Renowned academics and industry pioneers will lecture and share their views with the audience. Most deep learning subareas will be displayed, and main challenges identified through 24 four-hour and a half courses and 3 keynote lectures, which will tackle the most active and promising topics. The organizers are convinced that outstanding speakers will attract the brightest and most motivated students. Face to face interaction and networking will be main ingredients of the event. It will be also possible to fully participate in vivo remotely. An open session will give participants the opportunity to present their own work in progress in 5 minutes. Moreover, there will be two special sessions with industrial and recruitment profiles. ADDRESSED TO: Graduate students, postgraduate students and industry practitioners will be typical profiles of participants. However, there are no formal pre-requisites for attendance in terms of academic degrees, so people less or more advanced in their career will be welcome as well. Since there will be a variety of levels, specific knowledge background may be assumed for some of the courses. Overall, DeepLearn 2022 Summer is addressed to students, researchers and practitioners who want to keep themselves updated about recent developments and future trends. All will surely find it fruitful to listen to and discuss with major researchers, industry leaders and innovators. VENUE: DeepLearn 2022 Summer will take place in Las Palmas de Gran Canaria, on the Atlantic Ocean, with a mild climate throughout the year, sandy beaches and a renowned carnival. The venue will be: Instituci?n Ferial de Canarias Avenida de la Feria, 1 35012 Las Palmas de Gran Canaria https://www.infecar.es/index.php?option=com_k2&view=item&layout=item&id=360&Itemid=896 STRUCTURE: 3 courses will run in parallel during the whole event. Participants will be able to freely choose the courses they wish to attend as well as to move from one to another. Full live online participation will be possible. However, the organizers highlight the importance of face to face interaction and networking in this kind of research training event. KEYNOTE SPEAKERS: Wahid Bhimji (Lawrence Berkeley National Laboratory), Deep Learning on Supercomputers for Fundamental Science Joachim M. Buhmann (Swiss Federal Institute of Technology Zurich), Machine Learning -- A Paradigm Shift in Human Thought!? Kate Saenko (Boston University), Overcoming Dataset Bias in Deep Learning PROFESSORS AND COURSES: T?lay Adal? (University of Maryland Baltimore County), [intermediate] Data Fusion Using Matrix and Tensor Factorizations Pierre Baldi (University of California Irvine), [intermediate/advanced] Deep Learning: From Theory to Applications in the Natural Sciences Arindam Banerjee (University of Illinois Urbana-Champaign), [intermediate/advanced] Deep Generative and Dynamical Models Mikhail Belkin (University of California San Diego), [intermediate/advanced] Modern Machine Learning and Deep Learning through the Prism of Interpolation Dumitru Erhan (Google), [intermediate/advanced] Visual Self-supervised Learning and World Models Arthur Gretton (University College London), [intermediate/advanced] Probability Divergences and Generative Models Phillip Isola (Massachusetts Institute of Technology), [intermediate] Deep Generative Models Mohit Iyyer (University of Massachusetts Amherst), [intermediate/advanced] Natural Language Generation Irwin King (Chinese University of Hong Kong), [intermediate/advanced] Deep Learning on Graphs Vincent Lepetit (Paris Institute of Technology), [intermediate] Deep Learning and 3D Reasoning for 3D Scene Understanding Yan Liu (University of Southern California), [introductory/intermediate] Deep Learning for Time Series Dimitris N. Metaxas (Rutgers, The State University of New Jersey), [intermediate/advanced] Model-based, Explainable, Semisupervised and Unsupervised Machine Learning for Dynamic Analytics in Computer Vision and Medical Image Analysis Sean Meyn (University of Florida), [introductory/intermediate] Reinforcement Learning: Fundamentals, and Roadmaps for Successful Design Louis-Philippe Morency (Carnegie Mellon University), [intermediate/advanced] Multimodal Machine Learning Wojciech Samek (Fraunhofer Heinrich Hertz Institute), [introductory/intermediate] Explainable AI: Concepts, Methods and Applications Clara I. S?nchez (University of Amsterdam), [introductory/intermediate] Mechanisms for Trustworthy AI in Medical Image Analysis and Healthcare Bj?rn W. Schuller (Imperial College London), [introductory/intermediate] Deep Multimedia Processing Jonathon Shlens (Apple), [introductory/intermediate] An Introduction to Computer Vision and Convolution Neural Networks Johan Suykens (KU Leuven), [introductory/intermediate] Deep Learning, Neural Networks and Kernel Machines Csaba Szepesv?ri (University of Alberta), [intermediate/advanced] Tools and Techniques of Reinforcement Learning to Overcome Bellman's Curse of Dimensionality 1. Murat Tekalp (Ko? University), [intermediate/advanced] Deep Learning for Image/Video Restoration and Compression Alexandre Tkatchenko (University of Luxembourg), [introductory/intermediate] Machine Learning for Physics and Chemistry Li Xiong (Emory University), [introductory/intermediate] Differential Privacy and Certified Robustness for Deep Learning Ming Yuan (Columbia University), [intermediate/advanced] Low Rank Tensor Methods in High Dimensional Data Analysis OPEN SESSION: An open session will collect 5-minute voluntary presentations of work in progress by participants. They should submit a half-page abstract containing the title, authors, and summary of the research to david at irdta.eu by July 17, 2022. INDUSTRIAL SESSION: A session will be devoted to 10-minute demonstrations of practical applications of deep learning in industry. Companies interested in contributing are welcome to submit a 1-page abstract containing the program of the demonstration and the logistics needed. People in charge of the demonstration must register for the event. Expressions of interest have to be submitted to david at irdta.eu by July 17, 2022. EMPLOYER SESSION: Firms searching for personnel well skilled in deep learning will have a space reserved for one-to-one contacts. It is recommended to produce a 1-page .pdf leaflet with a brief description of the company and the profiles looked for to be circulated among the participants prior to the event. People in charge of the search must register for the event. Expressions of interest have to be submitted to david at irdta.eu by July 17, 2022. ORGANIZING COMMITTEE: Marisol Izquierdo (Las Palmas de Gran Canaria, local chair) Carlos Mart?n-Vide (Tarragona, program chair) Sara Morales (Brussels) David Silva (London, organization chair) REGISTRATION: It has to be done at https://irdta.eu/deeplearn/2022su/registration/ The selection of 8 courses requested in the registration template is only tentative and non-binding. For the sake of organization, it will be helpful to have an estimation of the respective demand for each course. During the event, participants will be free to attend the courses they wish. Since the capacity of the venue is limited, registration requests will be processed on a first come first served basis. The registration period will be closed and the on-line registration tool disabled when the capacity of the venue will have got exhausted. It is highly recommended to register prior to the event. FEES: Fees comprise access to all courses and lunches. There are several early registration deadlines. Fees depend on the registration deadline. The fees for on site and for online participation are the same. ACCOMMODATION: Accommodation suggestions will be available in due time at https://irdta.eu/deeplearn/2022su/accommodation/ CERTIFICATE: A certificate of successful participation in the event will be delivered indicating the number of hours of lectures. QUESTIONS AND FURTHER INFORMATION: david at irdta.eu ACKNOWLEDGMENTS: Cabildo de Gran Canaria Universidad de Las Palmas de Gran Canaria Universitat Rovira i Virgili Institute for Research Development, Training and Advice ? IRDTA, Brussels/London -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Sat Feb 5 21:58:09 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sun, 6 Feb 2022 02:58:09 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <537DF004-25CE-45A2-8155-D7E6018F4EE5@nyu.edu> References: <537DF004-25CE-45A2-8155-D7E6018F4EE5@nyu.edu> Message-ID: Gary, I don?t get much into the type of cognitive modeling you are talking about, but I would guess that the type problem can generally be handled by neural network models and tokens can be resolved with some memory-based system. But to the heart of the question, this is what so-called ?understanding? reduces to computation wise? Asim From: Gary Marcus Sent: Saturday, February 5, 2022 8:39 AM To: Asim Roy Cc: Ali Minai ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton There is no magic in understanding, just computation that has been realized in the wetware of humans and that eventually can be realized in machines. But understanding is not (just) learning. Understanding incorporates (or works in tandem with) learning - but also, critically, in tandem with inference, and the development and maintenance of cognitive models. Part of developing an understanding of cats in general is to learn long term-knowledge about their properties, both directly (e.g., through observation) and indirectly (eg through learning facts about animals in general that can be extended to cats), often through inference (if all animals have DNA, and a cat is an animal, it must also have DNA). The understanding of a particular cat also involves direct observation, but also inference (eg one might surmise that the reason that Fluffy is running about the room is that Fluffy suspects there is a mouse stirring somewhere nearby). But all of that, I would say, is subservient to the construction of cognitive models that can be routinely updated (e.g., Fluffy is currently in the living room, skittering about, perhaps looking for a mouse). In humans, those dynamic, relational models, which form part of an understanding, can support inference (if Fluffy is in the living room, we can infer that Fluffy is not outside, not lost, etc). Without such models - which I think represent a core part of understanding - AGI is an unlikely prospect. Current neural networks, as it happens, are better at acquiring long-term knowledge (cats have whiskers) than they are at dynamically updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic model that I am describing. To a modest degree they can approximate it on the basis of large samples of texts, but their ultimate incoherence stems from the fact that they do not have robust internal cognitive models that they can update on the fly. Without such cognitive models you can still capture some aspects of understanding (eg predicting that cats are likely to be furry), but things fall apart quickly; inference is never reliable, and coherence is fleeting. As a final note, one of the most foundational challenges in constructing adequate cognitive models of the world is to have a clear distinction between individuals and kinds; as I emphasized 20 years ago (in The Algebraic Mind), this has always been a weakness in neural networks, and I don?t think that the type-token problem has yet been solved. Gary On Feb 5, 2022, at 01:31, Asim Roy > wrote: ? All, I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? https://www.youtube.com/watch?v=gn4nRCC9TwQ https://www.youtube.com/watch?v=8sO7VS3q8d0 Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Ali Minai > Sent: Friday, February 4, 2022 11:38 PM To: Asim Roy > Cc: Gary Marcus >; Danko Nikolic >; Brad Wyble >; connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Asim Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! Ali Ali A. Minai, Ph.D. Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Fri, Feb 4, 2022 at 2:42 AM Asim Roy > wrote: First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. DANKO: What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Gary Marcus > Sent: Thursday, February 3, 2022 9:26 AM To: Danko Nikolic > Cc: Asim Roy >; Geoffrey Hinton >; AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub > wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary at eng.ucsd.edu Sat Feb 5 14:05:27 2022 From: gary at eng.ucsd.edu (gary@ucsd.edu) Date: Sat, 5 Feb 2022 11:05:27 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Please avert your gaze while I apply Ockham?s Razor? On Sat, Feb 5, 2022 at 2:12 AM Danko Nikolic wrote: > Gary, you wrote: "What are the alternatives?" > > There is at least one alternative: the theory of practopoiesis which > suggests that it is not the neural networks that "compute" the mental > operations. > It is instead the quick adaptations of neurons who are responsible for > thinking and perceiving. The network only serves the function of bringing > in the information and sending it out. > > The adaptations are suggested to do the central part of the cognition. > > So far, this is all hypothetical. If we develop these ideas into a working > system, this would be an entirely new paradigm. It would be like the third > paradigm: > 1) manipulation of symbols > 2) neural net > 3) fast adaptations > > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > --- A progress usually starts with an insight --- > > > On Fri, Feb 4, 2022 at 7:19 PM gary at ucsd.edu wrote: > >> This is an argument from lack of imagination, as Pat Churchland used to >> say. All you have to notice, is that your brain is a neural net work. What >> are the alternatives? >> >> On Fri, Feb 4, 2022 at 4:08 AM Danko Nikolic >> wrote: >> >>> >>> I suppose everyone agrees that "the brain is a physical system", >>> and that "There is no ?magic? inside the brain", >>> and that '?understanding? is just part of ?learning.?' >>> >>> Also, we can agree that some sort of simulation takes place behind >>> understanding. >>> >>> However, there still is a problem: Neural network's can't implement the >>> needed simulations; they cannot achieve the same cognitive effect that >>> human minds can (or animal minds can). >>> >>> We don't know a way of wiring a neural network such that it could >>> perform the simulations (understandings) necessary to find the answers to >>> real-life questions, such as the hamster with a hat problem. >>> >>> In other words, neural networks, as we know them today, cannot: >>> >>> 1) learn from a small number of examples (simulation or not) >>> 2) apply the knowledge to a wide range of situations >>> >>> >>> We, as scientists, do not understand understanding. Our technology's >>> simulations (their depth of understanding) are no match for the simulations >>> (depth of understanding) that the biological brain performs. >>> >>> I think that scientific integrity also covers acknowledging when we did >>> not (yet) succeed in solving a certain problem. There is still significant >>> work to be done. >>> >>> >>> Danko >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> --- A progress usually starts with an insight --- >>> >>> >>> >>> Virenfrei. >>> www.avast.com >>> >>> <#m_8423976727351221435_m_-3229424020171779455_m_-1469727422087267219_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> >>> >>> On Thu, Feb 3, 2022 at 9:35 PM Asim Roy wrote: >>> >>>> First of all, the brain is a physical system. There is no ?magic? >>>> inside the brain that does the ?understanding? part. Take for example >>>> learning to play tennis. You hit a few balls - some the right way and some >>>> wrong ? but you fairly quickly learn to hit them right most of the time. So >>>> there is obviously some simulation going on in the brain about hitting the >>>> ball in different ways and ?learning? its consequences. What you are >>>> calling ?understanding? is really these simulations about different >>>> scenarios. It?s also very similar to augmentation used to train image >>>> recognition systems where you rotate images, obscure parts and so on, so >>>> that you still can say it?s a cat even though you see only the cat?s face >>>> or whiskers or a cat flipped on its back. So, if the following questions >>>> relate to ?understanding,? you can easily resolve this by simulating such >>>> scenarios when ?teaching? the system. There?s nothing ?magical? about >>>> ?understanding.? As I said, bear in mind that the brain, after all, is a >>>> physical system and ?teaching? and ?understanding? is embodied in that >>>> physical system, not outside it. So ?understanding? is just part of >>>> ?learning,? nothing more. >>>> >>>> >>>> >>>> DANKO: >>>> >>>> What would happen to the hat if the hamster rolls on its back? (Would >>>> the hat fall off?) >>>> >>>> What would happen to the red hat when the hamster enters its lair? >>>> (Would the hat fall off?) >>>> >>>> What would happen to that hamster when it goes foraging? (Would the red >>>> hat have an influence on finding food?) >>>> >>>> What would happen in a situation of being chased by a predator? (Would >>>> it be easier for predators to spot the hamster?) >>>> >>>> >>>> >>>> Asim Roy >>>> >>>> Professor, Information Systems >>>> >>>> Arizona State University >>>> >>>> Lifeboat Foundation Bios: Professor Asim Roy >>>> >>>> >>>> Asim Roy | iSearch (asu.edu) >>>> >>>> >>>> >>>> >>>> >>>> >>>> *From:* Gary Marcus >>>> *Sent:* Thursday, February 3, 2022 9:26 AM >>>> *To:* Danko Nikolic >>>> *Cc:* Asim Roy ; Geoffrey Hinton < >>>> geoffrey.hinton at gmail.com>; AIhub ; >>>> connectionists at mailman.srv.cs.cmu.edu >>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>> Geoff Hinton >>>> >>>> >>>> >>>> Dear Danko, >>>> >>>> >>>> >>>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >>>> talk, in which he said (paraphrasing from memory, because I don?t remember >>>> the precise words) that the famous 200 Quoc Le unsupervised model [ >>>> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >>>> ] >>>> had learned the concept of a ca. In reality the model had clustered >>>> together some catlike images based on the image statistics that it had >>>> extracted, but it was a long way from a full, counterfactual-supporting >>>> concept of a cat, much as you describe below. >>>> >>>> >>>> >>>> I fully agree with you that the reason for even having a semantics is >>>> as you put it, "to 1) learn with a few examples and 2) apply the knowledge >>>> to a broad set of situations.? GPT-3 sometimes gives the appearance of >>>> having done so, but it falls apart under close inspection, so the problem >>>> remains unsolved. >>>> >>>> >>>> >>>> Gary >>>> >>>> >>>> >>>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >>>> wrote: >>>> >>>> >>>> >>>> G. Hinton wrote: "I believe that any reasonable person would admit that >>>> if you ask a neural net to draw a picture of a hamster wearing a red hat >>>> and it draws such a picture, it understood the request." >>>> >>>> >>>> >>>> I would like to suggest why drawing a hamster with a red hat does not >>>> necessarily imply understanding of the statement "hamster wearing a red >>>> hat". >>>> >>>> To understand that "hamster wearing a red hat" would mean inferring, in >>>> newly emerging situations of this hamster, all the real-life >>>> implications that the red hat brings to the little animal. >>>> >>>> >>>> >>>> What would happen to the hat if the hamster rolls on its back? (Would >>>> the hat fall off?) >>>> >>>> What would happen to the red hat when the hamster enters its lair? >>>> (Would the hat fall off?) >>>> >>>> What would happen to that hamster when it goes foraging? (Would the red >>>> hat have an influence on finding food?) >>>> >>>> What would happen in a situation of being chased by a predator? (Would >>>> it be easier for predators to spot the hamster?) >>>> >>>> >>>> >>>> ...and so on. >>>> >>>> >>>> >>>> Countless many questions can be asked. One has understood "hamster >>>> wearing a red hat" only if one can answer reasonably well many of such >>>> real-life relevant questions. Similarly, a student has understood materias >>>> in a class only if they can apply the materials in real-life situations >>>> (e.g., applying Pythagora's theorem). If a student gives a correct answer >>>> to a multiple choice question, we don't know whether the student understood >>>> the material or whether this was just rote learning (often, it is rote >>>> learning). >>>> >>>> >>>> >>>> I also suggest that understanding also comes together with effective >>>> learning: We store new information in such a way that we can recall it >>>> later and use it effectively i.e., make good inferences in newly emerging >>>> situations based on this knowledge. >>>> >>>> >>>> >>>> In short: Understanding makes us humans able to 1) learn with a few >>>> examples and 2) apply the knowledge to a broad set of situations. >>>> >>>> >>>> >>>> No neural network today has such capabilities and we don't know how to >>>> give them such capabilities. Neural networks need large amounts of >>>> training examples that cover a large variety of situations and then >>>> the networks can only deal with what the training examples have already >>>> covered. Neural networks cannot extrapolate in that 'understanding' sense. >>>> >>>> >>>> >>>> I suggest that understanding truly extrapolates from a piece of >>>> knowledge. It is not about satisfying a task such as translation between >>>> languages or drawing hamsters with hats. It is how you got the capability >>>> to complete the task: Did you only have a few examples that covered >>>> something different but related and then you extrapolated from that >>>> knowledge? If yes, this is going in the direction of understanding. Have >>>> you seen countless examples and then interpolated among them? Then perhaps >>>> it is not understanding. >>>> >>>> >>>> >>>> So, for the case of drawing a hamster wearing a red hat, understanding >>>> perhaps would have taken place if the following happened before that: >>>> >>>> >>>> >>>> 1) first, the network learned about hamsters (not many examples) >>>> >>>> 2) after that the network learned about red hats (outside the context >>>> of hamsters and without many examples) >>>> >>>> 3) finally the network learned about drawing (outside of the context of >>>> hats and hamsters, not many examples) >>>> >>>> >>>> >>>> After that, the network is asked to draw a hamster with a red hat. If >>>> it does it successfully, maybe we have started cracking the problem of >>>> understanding. >>>> >>>> >>>> >>>> Note also that this requires the network to learn sequentially without >>>> exhibiting catastrophic forgetting of the previous knowledge, which is >>>> possibly also a consequence of human learning by understanding. >>>> >>>> >>>> >>>> >>>> >>>> Danko >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> Dr. Danko Nikoli? >>>> www.danko-nikolic.com >>>> >>>> https://www.linkedin.com/in/danko-nikolic/ >>>> >>>> >>>> --- A progress usually starts with an insight --- >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> Virus-free. www.avast.com >>>> >>>> >>>> >>>> >>>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>>> >>>> Without getting into the specific dispute between Gary and Geoff, I >>>> think with approaches similar to GLOM, we are finally headed in the right >>>> direction. There?s plenty of neurophysiological evidence for single-cell >>>> abstractions and multisensory neurons in the brain, which one might claim >>>> correspond to symbols. And I think we can finally reconcile the decades old >>>> dispute between Symbolic AI and Connectionism. >>>> >>>> >>>> >>>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >>>> an effort to wind up with encodings that effectively serve as symbols in >>>> exactly that way, guaranteed to serve as consistent representations of >>>> specific concepts.) >>>> >>>> GARY: I have *never* called for dismissal of neural networks, but >>>> rather for some hybrid between the two (as you yourself contemplated in >>>> 1991); the point of the 2001 book was to characterize exactly where >>>> multilayer perceptrons succeeded and broke down, and where symbols could >>>> complement them. >>>> >>>> >>>> >>>> Asim Roy >>>> >>>> Professor, Information Systems >>>> >>>> Arizona State University >>>> >>>> Lifeboat Foundation Bios: Professor Asim Roy >>>> >>>> >>>> Asim Roy | iSearch (asu.edu) >>>> >>>> >>>> >>>> >>>> >>>> >>>> *From:* Connectionists *On >>>> Behalf Of *Gary Marcus >>>> *Sent:* Wednesday, February 2, 2022 1:26 PM >>>> *To:* Geoffrey Hinton >>>> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>> Geoff Hinton >>>> >>>> >>>> >>>> Dear Geoff, and interested others, >>>> >>>> >>>> >>>> What, for example, would you make of a system that often drew the >>>> red-hatted hamster you requested, and perhaps a fifth of the time gave you >>>> utter nonsense? Or say one that you trained to create birds but sometimes >>>> output stuff like this: >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> One could >>>> >>>> >>>> >>>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>>> >>>> or >>>> >>>> b. wonder if it might be possible that sometimes the system gets the >>>> right answer for the wrong reasons (eg partial historical contingency), and >>>> wonder whether another approach might be indicated. >>>> >>>> >>>> >>>> Benchmarks are harder than they look; most of the field has come to >>>> recognize that. The Turing Test has turned out to be a lousy measure of >>>> intelligence, easily gamed. It has turned out empirically that the Winograd >>>> Schema Challenge did not measure common sense as well as Hector might have >>>> thought. (As it happens, I am a minor coauthor of a very recent review on >>>> this very topic: https://arxiv.org/abs/2201.02387 >>>> ) >>>> But its conquest in no way means machines now have common sense; many >>>> people from many different perspectives recognize that (including, e.g., >>>> Yann LeCun, who generally tends to be more aligned with you than with me). >>>> >>>> >>>> >>>> So: on the goalpost of the Winograd schema, I was wrong, and you can >>>> quote me; but what you said about me and machine translation remains your >>>> invention, and it is inexcusable that you simply ignored my 2019 >>>> clarification. On the essential goal of trying to reach meaning and >>>> understanding, I remain unmoved; the problem remains unsolved. >>>> >>>> >>>> >>>> All of the problems LLMs have with coherence, reliability, >>>> truthfulness, misinformation, etc stand witness to that fact. (Their >>>> persistent inability to filter out toxic and insulting remarks stems from >>>> the same.) I am hardly the only person in the field to see that progress on >>>> any given benchmark does not inherently mean that the deep underlying >>>> problems have solved. You, yourself, in fact, have occasionally made that >>>> point. >>>> >>>> >>>> >>>> With respect to embeddings: Embeddings are very good for natural >>>> language *processing*; but NLP is not the same as NL*U* ? when it >>>> comes to *understanding*, their worth is still an open question. >>>> Perhaps they will turn out to be necessary; they clearly aren?t sufficient. >>>> In their extreme, they might even collapse into being symbols, in the sense >>>> of uniquely identifiable encodings, akin to the ASCII code, in which a >>>> specific set of numbers stands for a specific word or concept. (Wouldn?t >>>> that be ironic?) >>>> >>>> >>>> >>>> (Your GLOM, which as you know I praised publicly, is in many ways an >>>> effort to wind up with encodings that effectively serve as symbols in >>>> exactly that way, guaranteed to serve as consistent representations of >>>> specific concepts.) >>>> >>>> >>>> >>>> Notably absent from your email is any kind of apology for >>>> misrepresenting my position. It?s fine to say that ?many people thirty >>>> years ago once thought X? and another to say ?Gary Marcus said X in 2015?, >>>> when I didn?t. I have consistently felt throughout our interactions that >>>> you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) >>>> apologized to me for having made that error. I am still not he. >>>> >>>> >>>> >>>> Which maybe connects to the last point; if you read my work, you would >>>> see thirty years of arguments *for* neural networks, just not in the >>>> way that you want them to exist. I have ALWAYS argued that there is a role >>>> for them; characterizing me as a person ?strongly opposed to neural >>>> networks? misses the whole point of my 2001 book, which was subtitled >>>> ?Integrating Connectionism and Cognitive Science.? >>>> >>>> >>>> >>>> In the last two decades or so you have insisted (for reasons you have >>>> never fully clarified, so far as I know) on abandoning symbol-manipulation, >>>> but the reverse is not the case: I have *never* called for dismissal >>>> of neural networks, but rather for some hybrid between the two (as you >>>> yourself contemplated in 1991); the point of the 2001 book was to >>>> characterize exactly where multilayer perceptrons succeeded and broke down, >>>> and where symbols could complement them. It?s a rhetorical trick (which is >>>> what the previous thread was about) to pretend otherwise. >>>> >>>> >>>> >>>> Gary >>>> >>>> >>>> >>>> >>>> >>>> On Feb 2, 2022, at 11:22, Geoffrey Hinton >>>> wrote: >>>> >>>> ? >>>> >>>> Embeddings are just vectors of soft feature detectors and they are very >>>> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >>>> opposite. >>>> >>>> >>>> >>>> A few decades ago, everyone I knew then would have agreed that the >>>> ability to translate a sentence into many different languages was strong >>>> evidence that you understood it. >>>> >>>> >>>> >>>> But once neural networks could do that, their critics moved the >>>> goalposts. An exception is Hector Levesque who defined the goalposts more >>>> sharply by saying that the ability to get pronoun references correct in >>>> Winograd sentences is a crucial test. Neural nets are improving at that but >>>> still have some way to go. Will Gary agree that when they can get pronoun >>>> references correct in Winograd sentences they really do understand? Or does >>>> he want to reserve the right to weasel out of that too? >>>> >>>> >>>> >>>> Some people, like Gary, appear to be strongly opposed to neural >>>> networks because they do not fit their preconceived notions of how the mind >>>> should work. >>>> >>>> I believe that any reasonable person would admit that if you ask a >>>> neural net to draw a picture of a hamster wearing a red hat and it draws >>>> such a picture, it understood the request. >>>> >>>> >>>> >>>> Geoff >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>>> >>>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger >>>> neural network community, >>>> >>>> >>>> >>>> There has been a lot of recent discussion on this list about framing >>>> and scientific integrity. Often the first step in restructuring narratives >>>> is to bully and dehumanize critics. The second is to misrepresent their >>>> position. People in positions of power are sometimes tempted to do this. >>>> >>>> >>>> >>>> The Hinton-Hanson interview that you just published is a real-time >>>> example of just that. It opens with a needless and largely content-free >>>> personal attack on a single scholar (me), with the explicit intention of >>>> discrediting that person. Worse, the only substantive thing it says is >>>> false. >>>> >>>> >>>> >>>> Hinton says ?In 2015 he [Marcus] made a prediction that computers >>>> wouldn?t be able to do machine translation.? >>>> >>>> >>>> >>>> I never said any such thing. >>>> >>>> >>>> >>>> What I predicted, rather, was that multilayer perceptrons, as they >>>> existed then, would not (on their own, absent other mechanisms) >>>> *understand* language. Seven years later, they still haven?t, except >>>> in the most superficial way. >>>> >>>> >>>> >>>> I made no comment whatsoever about machine translation, which I view as >>>> a separate problem, solvable to a certain degree by correspondance without >>>> semantics. >>>> >>>> >>>> >>>> I specifically tried to clarify Hinton?s confusion in 2019, but, >>>> disappointingly, he has continued to purvey misinformation despite that >>>> clarification. Here is what I wrote privately to him then, which should >>>> have put the matter to rest: >>>> >>>> >>>> >>>> You have taken a single out of context quote [from 2015] and >>>> misrepresented it. The quote, which you have prominently displayed at the >>>> bottom on your own web page, says: >>>> >>>> >>>> >>>> Hierarchies of features are less suited to challenges such as language, >>>> inference, and high-level planning. For example, as Noam Chomsky famously >>>> pointed out, language is filled with sentences you haven't seen >>>> before. Pure classifier systems don't know what to do with such sentences. >>>> The talent of feature detectors -- in identifying which member of some >>>> category something belongs to -- doesn't translate into understanding >>>> novel sentences, in which each sentence has its own unique meaning. >>>> >>>> >>>> >>>> It does *not* say "neural nets would not be able to deal with novel >>>> sentences"; it says that hierachies of features detectors (on their own, if >>>> you read the context of the essay) would have trouble *understanding *novel sentences. >>>> >>>> >>>> >>>> >>>> Google Translate does yet not *understand* the content of the >>>> sentences is translates. It cannot reliably answer questions about who did >>>> what to whom, or why, it cannot infer the order of the events in >>>> paragraphs, it can't determine the internal consistency of those events, >>>> and so forth. >>>> >>>> >>>> >>>> Since then, a number of scholars, such as the the computational >>>> linguist Emily Bender, have made similar points, and indeed current LLM >>>> difficulties with misinformation, incoherence and fabrication all follow >>>> from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on >>>> the matter with Alexander Koller, >>>> https://aclanthology.org/2020.acl-main.463.pdf >>>> , >>>> also emphasizing issues of understanding and meaning: >>>> >>>> >>>> >>>> *The success of the large neural language models on many NLP tasks is >>>> exciting. However, we find that these successes sometimes lead to hype in >>>> which these models are being described as ?understanding? language or >>>> capturing ?meaning?. In this position paper, we argue that a system trained >>>> only on form has a priori no way to learn meaning. .. a clear understanding >>>> of the distinction between form and meaning will help guide the field >>>> towards better science around natural language understanding. * >>>> >>>> >>>> >>>> Her later article with Gebru on language models ?stochastic parrots? is >>>> in some ways an extension of this point; machine translation requires >>>> mimicry, true understanding (which is what I was discussing in 2015) >>>> requires something deeper than that. >>>> >>>> >>>> >>>> Hinton?s intellectual error here is in equating machine translation >>>> with the deeper comprehension that robust natural language understanding >>>> will require; as Bender and Koller observed, the two appear not to be the >>>> same. (There is a longer discussion of the relation between language >>>> understanding and machine translation, and why the latter has turned out to >>>> be more approachable than the former, in my 2019 book with Ernest Davis). >>>> >>>> >>>> >>>> More broadly, Hinton?s ongoing dismissiveness of research from >>>> perspectives other than his own (e.g. linguistics) have done the field a >>>> disservice. >>>> >>>> >>>> >>>> As Herb Simon once observed, science does not have to be zero-sum. >>>> >>>> >>>> >>>> Sincerely, >>>> >>>> Gary Marcus >>>> >>>> Professor Emeritus >>>> >>>> New York University >>>> >>>> >>>> >>>> On Feb 2, 2022, at 06:12, AIhub wrote: >>>> >>>> ? >>>> >>>> Stephen Hanson in conversation with Geoff Hinton >>>> >>>> >>>> >>>> In the latest episode of this video series for AIhub.org >>>> , >>>> Stephen Hanson talks to Geoff Hinton about neural networks, >>>> backpropagation, overparameterization, digit recognition, voxel cells, >>>> syntax and semantics, Winograd sentences, and more. >>>> >>>> >>>> >>>> You can watch the discussion, and read the transcript, here: >>>> >>>> >>>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>>> >>>> >>>> >>>> >>>> About AIhub: >>>> >>>> AIhub is a non-profit dedicated to connecting the AI community to the >>>> public by providing free, high-quality information through AIhub.org >>>> >>>> (https://aihub.org/ >>>> ). >>>> We help researchers publish the latest AI news, summaries of their work, >>>> opinion pieces, tutorials and more. We are supported by many leading >>>> scientific organizations in AI, namely AAAI >>>> , >>>> NeurIPS >>>> , >>>> ICML >>>> , >>>> AIJ >>>> >>>> /IJCAI >>>> , >>>> ACM SIGAI >>>> , >>>> EurAI/AICOMM, CLAIRE >>>> >>>> and RoboCup >>>> >>>> . >>>> >>>> Twitter: @aihuborg >>>> >>>> >>>> >>>> >>>> >>>> >>>> Virus-free. www.avast.com >>>> >>>> >>>> >>>> >>> -- >> Gary Cottrell 858-534-6640 FAX: 858-534-7029 >> Computer Science and Engineering 0404 >> IF USING FEDEX INCLUDE THE FOLLOWING LINE: >> CSE Building, Room 4130 >> University of California San Diego - >> 9500 Gilman Drive # 0404 >> >> La Jolla, Ca. 92093-0404 >> >> >> Email: gary at ucsd.edu >> Home page: http://www-cse.ucsd.edu/~gary/ >> Schedule: http://tinyurl.com/b7gxpwo >> >> >> Blind certainty - a close-mindedness that amounts to an imprisonment so >> total, that the prisoner doesn?t even know that he?s locked up. -David >> Foster Wallace >> >> >> Power to the people! ?Patti Smith >> >> Except when they?re delusional ?Gary Cottrell >> >> >> This song makes me nostalgic for a memory I don't have -- Tess Cottrell >> >> >> >> >> >> >> >> >> >> >> *Listen carefully,Neither the VedasNor the Qur'anWill teach you this:Put >> the bit in its mouth,The saddle on its back,Your foot in the stirrup,And >> ride your wild runaway mindAll the way to heaven.* >> >> -- Kabir >> > -- Gary Cottrell 858-534-6640 FAX: 858-534-7029 Computer Science and Engineering 0404 IF USING FEDEX INCLUDE THE FOLLOWING LINE: CSE Building, Room 4130 University of California San Diego - 9500 Gilman Drive # 0404 La Jolla, Ca. 92093-0404 Email: gary at ucsd.edu Home page: http://www-cse.ucsd.edu/~gary/ Schedule: http://tinyurl.com/b7gxpwo Blind certainty - a close-mindedness that amounts to an imprisonment so total, that the prisoner doesn?t even know that he?s locked up. -David Foster Wallace Power to the people! ?Patti Smith Except when they?re delusional ?Gary Cottrell This song makes me nostalgic for a memory I don't have -- Tess Cottrell *Listen carefully,Neither the VedasNor the Qur'anWill teach you this:Put the bit in its mouth,The saddle on its back,Your foot in the stirrup,And ride your wild runaway mindAll the way to heaven.* -- Kabir -------------- next part -------------- An HTML attachment was scrubbed... URL: From steve at bu.edu Sat Feb 5 12:53:20 2022 From: steve at bu.edu (Grossberg, Stephen) Date: Sat, 5 Feb 2022 17:53:20 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton: Some biological neural models of of causality, ambiguous visual percepts, and handwritten letters In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear All, The problems that have been raised below about by Geoffrey Hinton and others concerning causality, ambiguous visual percepts, and handwritten letters have been at least partially solved by specific neural models. Below I take the liberty of mentioning some of these models and where you can read more if you are interested: CAUSALITY AND ADAPTIVE RESONANCE THEORY Many of you may already know that Adaptive Resonance Theory, or ART, is currently the most advanced cognitive and neural theory that explains how humans learn to consciously attend, recognize, and predict objects and events in a changing world. ART does this by explaining how we learn to attend to the critical feature patterns that successfully predict what will happen next in a familiar environment, while suppressing irrelevant features. These critical feature patterns can be learned at multiple levels of brain organization, ranging from concrete to abstract. Critical features are the ones that we come to believe cause predicted outcomes. It is fair to ask: Why should anyone believe this ART proposal? There are several types of reasons: First, all the foundational hypotheses of ART have been supported by subsequent psychological and neurobiological experiments. Second, ART proposes principled and unifying explanations of hundreds of additional experiments. It has also made numerous predictions, many of which have been supported by subsequent psychological and/or neurobiological experiments. These successes may derive from the fact that ART explains how human brains solve the stability-plasticity dilemma; that is, how we can learn quickly without experiencing catastrophic forgetting. Third, there is a deeper reason to believe that ART explains fundamental properties of how our brains make our minds: In 1980, in the journal Psychological Review, I derived ART from a thought experiment whose hypotheses are facts that are familiar to us all because they are ubiquitous constraints on the evolution of our brains. The words mind and brain are not mentioned during this thought experiment. The thought experiment from which ART is derived emerges from an analysis of how any system can autonomously correct predictive errors in a changing world that is filled with unexpected events. ART and its variations are thus universal solutions of this fundamental problem. Perhaps that it why ART has been applied to help design autonomous adaptively intelligent systems for large-scale applications. I was particularly excited when I realized that ART also helps to explain how, where in our brains, and why evolution created conscious states of seeing, hearing, feeling, and knowing, and how these conscious states enable planning and action to realize valued goals. ART has also been extended to explain and predict a lot of data about cognitive-emotional interactions and affective neuroscience. When these cognitive and/or cognitive-emotional processes break down in specific ways, then behavioral symptoms of mental disorders emerge, including Alzheimer's disease, autism, Fragile X syndrome, medial temporal amnesia, schizophrenia, ADHD, PTSD, auditory and visual neglect and agnosia, and disorders of slow-wave sleep. Here are two articles that discuss how ART learns the critical feature patterns that predict future outcomes and thus that embody our understanding of their causes: Grossberg, S. (2017). Towards solving the Hard Problem of Consciousness: The varieties of brain resonances and the conscious experiences that they support. Neural Networks, 87, 38?95. https://www.sciencedirect.com/science/article/pii/S0893608016301800 Grossberg, S. (2021). A canonical laminar neocortical circuit whose bottom-up, horizontal, and top-down pathways control attention, learning, and prediction. Frontiers in Systems Neuroscience. Published online: 23 April 2021. https://www.frontiersin.org/articles/10.3389/fnsys.2021.650263/full I have also discussed these issues in my recent book: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind which just won the 2022 PROSE book award in Neuroscience from the Association of American Publishers: https://global.oup.com/academic/product/conscious-mind-resonant-brain-9780190070557?cc=us&lang=en& https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 Some of the following results are also summarized in the book: AMBIGUOUS VISUAL PERCEPTS AND BISTABILITY The following articles propose model explanations of ambiguous visual percepts that are more complicated than the ones that Geoffrey Hinton mentioned. They show how attention shifts can dramatically alter the way in which we consciously see ambiguous visual percepts of various kinds: Grossberg, S., and Swaminathan, G. (2004). A laminar cortical model for 3D perception of slanted and curved surfaces and of 2D images: development, attention and bistability. Vision Research, 44, 1147-1187. https://sites.bu.edu/steveg/files/2016/06/GroSwa2004VR.pdf Grossberg, S., Yazdanbakhsh, A., Cao, Y., and Swaminathan, G. (2008). How does binocular rivalry emerge from cortical mechanisms of 3-D vision? Vision Research, 48, 2232-2250. https://sites.bu.edu/steveg/files/2016/06/GroYazCaoSwaVR2008.pdf PRODUCTION AND PERCEPTION OF HANDWRITTEN LETTERS The following articles propose how children and adults learn and recognize cursive script, which includes the problem of producing and recognizing isolated cursive letters: Bullock, D., Grossberg, S., and Mannes, C. (1993). A neural network model for cursive script production. Biological Cybernetics, 70, 15-28. https://sites.bu.edu/steveg/files/2016/06/BulGroMan1993BiolCyb.pdf Grossberg, S. and Paine, R.W.(2000). A neural model of corticocerebellar interactions during attentive imitation and predictive learning of sequential handwriting movements. Neural Networks, 13, 999-1046. https://sites.bu.edu/steveg/files/2016/06/GroPai00.pdf Paine, R.W., Grossberg, S., and Van Gemmert, A.W.A. (2004). A quantitative evaluation of the AVITEWRITE model of handwriting learning . Human Movement Science, 23, 837-860. https://sites.bu.edu/steveg/files/2016/06/PaiGrovanGem2004HMS.pdf If you have any comments or questions about the above results, please feel free to send them to me. I will do my best to reply. Best, Steve Grossberg Stephen Grossberg http://en.wikipedia.org/wiki/Stephen_Grossberg http://scholar.google.com/citations?user=3BIV70wAAAAJ&hl=en https://youtu.be/9n5AnvFur7I https://www.youtube.com/watch?v=_hBye6JQCh4 https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 Wang Professor of Cognitive and Neural Systems Director, Center for Adaptive Systems Professor Emeritus of Mathematics & Statistics, Psychological & Brain Sciences, and Biomedical Engineering Boston University sites.bu.edu/steveg steve at bu.edu ________________________________ From: Connectionists on behalf of Geoffrey Hinton Sent: Friday, February 4, 2022 3:24 PM To: Dietterich, Thomas Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists > On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic > Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- [X] Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub > wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg [X] Virus-free. www.avast.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Sun Feb 6 05:27:39 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Sun, 6 Feb 2022 11:27:39 +0100 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Hi Gary, you said: "Please avert your gaze while I apply Ockham?s Razor?" I dare you to apply Ockham's razor. Practopoiesis is designed with the Ockham's razor in mind: To account for as many mental phenomena as possible by making as few assumptions as possible. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Sat, Feb 5, 2022 at 8:05 PM gary at ucsd.edu wrote: > Please avert your gaze while I apply Ockham?s Razor? > > On Sat, Feb 5, 2022 at 2:12 AM Danko Nikolic > wrote: > >> Gary, you wrote: "What are the alternatives?" >> >> There is at least one alternative: the theory of practopoiesis which >> suggests that it is not the neural networks that "compute" the mental >> operations. >> It is instead the quick adaptations of neurons who are responsible for >> thinking and perceiving. The network only serves the function of bringing >> in the information and sending it out. >> >> The adaptations are suggested to do the central part of the cognition. >> >> So far, this is all hypothetical. If we develop these ideas into a >> working system, this would be an entirely new paradigm. It would be like >> the third paradigm: >> 1) manipulation of symbols >> 2) neural net >> 3) fast adaptations >> >> >> Danko >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> >> https://www.linkedin.com/in/danko-nikolic/ >> >> --- A progress usually starts with an insight --- >> >> >> On Fri, Feb 4, 2022 at 7:19 PM gary at ucsd.edu wrote: >> >>> This is an argument from lack of imagination, as Pat Churchland used to >>> say. All you have to notice, is that your brain is a neural net work. What >>> are the alternatives? >>> >>> On Fri, Feb 4, 2022 at 4:08 AM Danko Nikolic >>> wrote: >>> >>>> >>>> I suppose everyone agrees that "the brain is a physical system", >>>> and that "There is no ?magic? inside the brain", >>>> and that '?understanding? is just part of ?learning.?' >>>> >>>> Also, we can agree that some sort of simulation takes place behind >>>> understanding. >>>> >>>> However, there still is a problem: Neural network's can't implement the >>>> needed simulations; they cannot achieve the same cognitive effect that >>>> human minds can (or animal minds can). >>>> >>>> We don't know a way of wiring a neural network such that it could >>>> perform the simulations (understandings) necessary to find the answers to >>>> real-life questions, such as the hamster with a hat problem. >>>> >>>> In other words, neural networks, as we know them today, cannot: >>>> >>>> 1) learn from a small number of examples (simulation or not) >>>> 2) apply the knowledge to a wide range of situations >>>> >>>> >>>> We, as scientists, do not understand understanding. Our technology's >>>> simulations (their depth of understanding) are no match for the simulations >>>> (depth of understanding) that the biological brain performs. >>>> >>>> I think that scientific integrity also covers acknowledging when we did >>>> not (yet) succeed in solving a certain problem. There is still significant >>>> work to be done. >>>> >>>> >>>> Danko >>>> >>>> Dr. Danko Nikoli? >>>> www.danko-nikolic.com >>>> >>>> https://www.linkedin.com/in/danko-nikolic/ >>>> >>>> --- A progress usually starts with an insight --- >>>> >>>> >>>> >>>> Virenfrei. >>>> www.avast.com >>>> >>>> <#m_-134745596574091214_m_8423976727351221435_m_-3229424020171779455_m_-1469727422087267219_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> >>>> >>>> On Thu, Feb 3, 2022 at 9:35 PM Asim Roy wrote: >>>> >>>>> First of all, the brain is a physical system. There is no ?magic? >>>>> inside the brain that does the ?understanding? part. Take for example >>>>> learning to play tennis. You hit a few balls - some the right way and some >>>>> wrong ? but you fairly quickly learn to hit them right most of the time. So >>>>> there is obviously some simulation going on in the brain about hitting the >>>>> ball in different ways and ?learning? its consequences. What you are >>>>> calling ?understanding? is really these simulations about different >>>>> scenarios. It?s also very similar to augmentation used to train image >>>>> recognition systems where you rotate images, obscure parts and so on, so >>>>> that you still can say it?s a cat even though you see only the cat?s face >>>>> or whiskers or a cat flipped on its back. So, if the following questions >>>>> relate to ?understanding,? you can easily resolve this by simulating such >>>>> scenarios when ?teaching? the system. There?s nothing ?magical? about >>>>> ?understanding.? As I said, bear in mind that the brain, after all, is a >>>>> physical system and ?teaching? and ?understanding? is embodied in that >>>>> physical system, not outside it. So ?understanding? is just part of >>>>> ?learning,? nothing more. >>>>> >>>>> >>>>> >>>>> DANKO: >>>>> >>>>> What would happen to the hat if the hamster rolls on its back? (Would >>>>> the hat fall off?) >>>>> >>>>> What would happen to the red hat when the hamster enters its lair? >>>>> (Would the hat fall off?) >>>>> >>>>> What would happen to that hamster when it goes foraging? (Would the >>>>> red hat have an influence on finding food?) >>>>> >>>>> What would happen in a situation of being chased by a predator? (Would >>>>> it be easier for predators to spot the hamster?) >>>>> >>>>> >>>>> >>>>> Asim Roy >>>>> >>>>> Professor, Information Systems >>>>> >>>>> Arizona State University >>>>> >>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>> >>>>> >>>>> Asim Roy | iSearch (asu.edu) >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> *From:* Gary Marcus >>>>> *Sent:* Thursday, February 3, 2022 9:26 AM >>>>> *To:* Danko Nikolic >>>>> *Cc:* Asim Roy ; Geoffrey Hinton < >>>>> geoffrey.hinton at gmail.com>; AIhub ; >>>>> connectionists at mailman.srv.cs.cmu.edu >>>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>>> Geoff Hinton >>>>> >>>>> >>>>> >>>>> Dear Danko, >>>>> >>>>> >>>>> >>>>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >>>>> talk, in which he said (paraphrasing from memory, because I don?t remember >>>>> the precise words) that the famous 200 Quoc Le unsupervised model [ >>>>> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >>>>> ] >>>>> had learned the concept of a ca. In reality the model had clustered >>>>> together some catlike images based on the image statistics that it had >>>>> extracted, but it was a long way from a full, counterfactual-supporting >>>>> concept of a cat, much as you describe below. >>>>> >>>>> >>>>> >>>>> I fully agree with you that the reason for even having a semantics is >>>>> as you put it, "to 1) learn with a few examples and 2) apply the knowledge >>>>> to a broad set of situations.? GPT-3 sometimes gives the appearance of >>>>> having done so, but it falls apart under close inspection, so the problem >>>>> remains unsolved. >>>>> >>>>> >>>>> >>>>> Gary >>>>> >>>>> >>>>> >>>>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >>>>> wrote: >>>>> >>>>> >>>>> >>>>> G. Hinton wrote: "I believe that any reasonable person would admit >>>>> that if you ask a neural net to draw a picture of a hamster wearing a red >>>>> hat and it draws such a picture, it understood the request." >>>>> >>>>> >>>>> >>>>> I would like to suggest why drawing a hamster with a red hat does not >>>>> necessarily imply understanding of the statement "hamster wearing a red >>>>> hat". >>>>> >>>>> To understand that "hamster wearing a red hat" would mean inferring, >>>>> in newly emerging situations of this hamster, all the real-life >>>>> implications that the red hat brings to the little animal. >>>>> >>>>> >>>>> >>>>> What would happen to the hat if the hamster rolls on its back? (Would >>>>> the hat fall off?) >>>>> >>>>> What would happen to the red hat when the hamster enters its lair? >>>>> (Would the hat fall off?) >>>>> >>>>> What would happen to that hamster when it goes foraging? (Would the >>>>> red hat have an influence on finding food?) >>>>> >>>>> What would happen in a situation of being chased by a predator? (Would >>>>> it be easier for predators to spot the hamster?) >>>>> >>>>> >>>>> >>>>> ...and so on. >>>>> >>>>> >>>>> >>>>> Countless many questions can be asked. One has understood "hamster >>>>> wearing a red hat" only if one can answer reasonably well many of such >>>>> real-life relevant questions. Similarly, a student has understood materias >>>>> in a class only if they can apply the materials in real-life situations >>>>> (e.g., applying Pythagora's theorem). If a student gives a correct answer >>>>> to a multiple choice question, we don't know whether the student understood >>>>> the material or whether this was just rote learning (often, it is rote >>>>> learning). >>>>> >>>>> >>>>> >>>>> I also suggest that understanding also comes together with effective >>>>> learning: We store new information in such a way that we can recall it >>>>> later and use it effectively i.e., make good inferences in newly emerging >>>>> situations based on this knowledge. >>>>> >>>>> >>>>> >>>>> In short: Understanding makes us humans able to 1) learn with a few >>>>> examples and 2) apply the knowledge to a broad set of situations. >>>>> >>>>> >>>>> >>>>> No neural network today has such capabilities and we don't know how to >>>>> give them such capabilities. Neural networks need large amounts of >>>>> training examples that cover a large variety of situations and then >>>>> the networks can only deal with what the training examples have already >>>>> covered. Neural networks cannot extrapolate in that 'understanding' sense. >>>>> >>>>> >>>>> >>>>> I suggest that understanding truly extrapolates from a piece of >>>>> knowledge. It is not about satisfying a task such as translation between >>>>> languages or drawing hamsters with hats. It is how you got the capability >>>>> to complete the task: Did you only have a few examples that covered >>>>> something different but related and then you extrapolated from that >>>>> knowledge? If yes, this is going in the direction of understanding. Have >>>>> you seen countless examples and then interpolated among them? Then perhaps >>>>> it is not understanding. >>>>> >>>>> >>>>> >>>>> So, for the case of drawing a hamster wearing a red hat, understanding >>>>> perhaps would have taken place if the following happened before that: >>>>> >>>>> >>>>> >>>>> 1) first, the network learned about hamsters (not many examples) >>>>> >>>>> 2) after that the network learned about red hats (outside the context >>>>> of hamsters and without many examples) >>>>> >>>>> 3) finally the network learned about drawing (outside of the context >>>>> of hats and hamsters, not many examples) >>>>> >>>>> >>>>> >>>>> After that, the network is asked to draw a hamster with a red hat. If >>>>> it does it successfully, maybe we have started cracking the problem of >>>>> understanding. >>>>> >>>>> >>>>> >>>>> Note also that this requires the network to learn sequentially without >>>>> exhibiting catastrophic forgetting of the previous knowledge, which is >>>>> possibly also a consequence of human learning by understanding. >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> Danko >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> Dr. Danko Nikoli? >>>>> www.danko-nikolic.com >>>>> >>>>> https://www.linkedin.com/in/danko-nikolic/ >>>>> >>>>> >>>>> --- A progress usually starts with an insight --- >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> Virus-free. www.avast.com >>>>> >>>>> >>>>> >>>>> >>>>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>>>> >>>>> Without getting into the specific dispute between Gary and Geoff, I >>>>> think with approaches similar to GLOM, we are finally headed in the right >>>>> direction. There?s plenty of neurophysiological evidence for single-cell >>>>> abstractions and multisensory neurons in the brain, which one might claim >>>>> correspond to symbols. And I think we can finally reconcile the decades old >>>>> dispute between Symbolic AI and Connectionism. >>>>> >>>>> >>>>> >>>>> GARY: (Your GLOM, which as you know I praised publicly, is in many >>>>> ways an effort to wind up with encodings that effectively serve as symbols >>>>> in exactly that way, guaranteed to serve as consistent representations of >>>>> specific concepts.) >>>>> >>>>> GARY: I have *never* called for dismissal of neural networks, but >>>>> rather for some hybrid between the two (as you yourself contemplated in >>>>> 1991); the point of the 2001 book was to characterize exactly where >>>>> multilayer perceptrons succeeded and broke down, and where symbols could >>>>> complement them. >>>>> >>>>> >>>>> >>>>> Asim Roy >>>>> >>>>> Professor, Information Systems >>>>> >>>>> Arizona State University >>>>> >>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>> >>>>> >>>>> Asim Roy | iSearch (asu.edu) >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> *From:* Connectionists >>>>> *On Behalf Of *Gary Marcus >>>>> *Sent:* Wednesday, February 2, 2022 1:26 PM >>>>> *To:* Geoffrey Hinton >>>>> *Cc:* AIhub ; >>>>> connectionists at mailman.srv.cs.cmu.edu >>>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>>> Geoff Hinton >>>>> >>>>> >>>>> >>>>> Dear Geoff, and interested others, >>>>> >>>>> >>>>> >>>>> What, for example, would you make of a system that often drew the >>>>> red-hatted hamster you requested, and perhaps a fifth of the time gave you >>>>> utter nonsense? Or say one that you trained to create birds but sometimes >>>>> output stuff like this: >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> One could >>>>> >>>>> >>>>> >>>>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>>>> >>>>> or >>>>> >>>>> b. wonder if it might be possible that sometimes the system gets the >>>>> right answer for the wrong reasons (eg partial historical contingency), and >>>>> wonder whether another approach might be indicated. >>>>> >>>>> >>>>> >>>>> Benchmarks are harder than they look; most of the field has come to >>>>> recognize that. The Turing Test has turned out to be a lousy measure of >>>>> intelligence, easily gamed. It has turned out empirically that the Winograd >>>>> Schema Challenge did not measure common sense as well as Hector might have >>>>> thought. (As it happens, I am a minor coauthor of a very recent review on >>>>> this very topic: https://arxiv.org/abs/2201.02387 >>>>> ) >>>>> But its conquest in no way means machines now have common sense; many >>>>> people from many different perspectives recognize that (including, e.g., >>>>> Yann LeCun, who generally tends to be more aligned with you than with me). >>>>> >>>>> >>>>> >>>>> So: on the goalpost of the Winograd schema, I was wrong, and you can >>>>> quote me; but what you said about me and machine translation remains your >>>>> invention, and it is inexcusable that you simply ignored my 2019 >>>>> clarification. On the essential goal of trying to reach meaning and >>>>> understanding, I remain unmoved; the problem remains unsolved. >>>>> >>>>> >>>>> >>>>> All of the problems LLMs have with coherence, reliability, >>>>> truthfulness, misinformation, etc stand witness to that fact. (Their >>>>> persistent inability to filter out toxic and insulting remarks stems from >>>>> the same.) I am hardly the only person in the field to see that progress on >>>>> any given benchmark does not inherently mean that the deep underlying >>>>> problems have solved. You, yourself, in fact, have occasionally made that >>>>> point. >>>>> >>>>> >>>>> >>>>> With respect to embeddings: Embeddings are very good for natural >>>>> language *processing*; but NLP is not the same as NL*U* ? when it >>>>> comes to *understanding*, their worth is still an open question. >>>>> Perhaps they will turn out to be necessary; they clearly aren?t sufficient. >>>>> In their extreme, they might even collapse into being symbols, in the sense >>>>> of uniquely identifiable encodings, akin to the ASCII code, in which a >>>>> specific set of numbers stands for a specific word or concept. (Wouldn?t >>>>> that be ironic?) >>>>> >>>>> >>>>> >>>>> (Your GLOM, which as you know I praised publicly, is in many ways an >>>>> effort to wind up with encodings that effectively serve as symbols in >>>>> exactly that way, guaranteed to serve as consistent representations of >>>>> specific concepts.) >>>>> >>>>> >>>>> >>>>> Notably absent from your email is any kind of apology for >>>>> misrepresenting my position. It?s fine to say that ?many people thirty >>>>> years ago once thought X? and another to say ?Gary Marcus said X in 2015?, >>>>> when I didn?t. I have consistently felt throughout our interactions that >>>>> you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) >>>>> apologized to me for having made that error. I am still not he. >>>>> >>>>> >>>>> >>>>> Which maybe connects to the last point; if you read my work, you would >>>>> see thirty years of arguments *for* neural networks, just not in the >>>>> way that you want them to exist. I have ALWAYS argued that there is a role >>>>> for them; characterizing me as a person ?strongly opposed to neural >>>>> networks? misses the whole point of my 2001 book, which was subtitled >>>>> ?Integrating Connectionism and Cognitive Science.? >>>>> >>>>> >>>>> >>>>> In the last two decades or so you have insisted (for reasons you have >>>>> never fully clarified, so far as I know) on abandoning symbol-manipulation, >>>>> but the reverse is not the case: I have *never* called for dismissal >>>>> of neural networks, but rather for some hybrid between the two (as you >>>>> yourself contemplated in 1991); the point of the 2001 book was to >>>>> characterize exactly where multilayer perceptrons succeeded and broke down, >>>>> and where symbols could complement them. It?s a rhetorical trick (which is >>>>> what the previous thread was about) to pretend otherwise. >>>>> >>>>> >>>>> >>>>> Gary >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> On Feb 2, 2022, at 11:22, Geoffrey Hinton >>>>> wrote: >>>>> >>>>> ? >>>>> >>>>> Embeddings are just vectors of soft feature detectors and they are >>>>> very good for NLP. The quote on my webpage from Gary's 2015 chapter >>>>> implies the opposite. >>>>> >>>>> >>>>> >>>>> A few decades ago, everyone I knew then would have agreed that the >>>>> ability to translate a sentence into many different languages was strong >>>>> evidence that you understood it. >>>>> >>>>> >>>>> >>>>> But once neural networks could do that, their critics moved the >>>>> goalposts. An exception is Hector Levesque who defined the goalposts more >>>>> sharply by saying that the ability to get pronoun references correct in >>>>> Winograd sentences is a crucial test. Neural nets are improving at that but >>>>> still have some way to go. Will Gary agree that when they can get pronoun >>>>> references correct in Winograd sentences they really do understand? Or does >>>>> he want to reserve the right to weasel out of that too? >>>>> >>>>> >>>>> >>>>> Some people, like Gary, appear to be strongly opposed to neural >>>>> networks because they do not fit their preconceived notions of how the mind >>>>> should work. >>>>> >>>>> I believe that any reasonable person would admit that if you ask a >>>>> neural net to draw a picture of a hamster wearing a red hat and it draws >>>>> such a picture, it understood the request. >>>>> >>>>> >>>>> >>>>> Geoff >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus >>>>> wrote: >>>>> >>>>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger >>>>> neural network community, >>>>> >>>>> >>>>> >>>>> There has been a lot of recent discussion on this list about framing >>>>> and scientific integrity. Often the first step in restructuring narratives >>>>> is to bully and dehumanize critics. The second is to misrepresent their >>>>> position. People in positions of power are sometimes tempted to do this. >>>>> >>>>> >>>>> >>>>> The Hinton-Hanson interview that you just published is a real-time >>>>> example of just that. It opens with a needless and largely content-free >>>>> personal attack on a single scholar (me), with the explicit intention of >>>>> discrediting that person. Worse, the only substantive thing it says is >>>>> false. >>>>> >>>>> >>>>> >>>>> Hinton says ?In 2015 he [Marcus] made a prediction that computers >>>>> wouldn?t be able to do machine translation.? >>>>> >>>>> >>>>> >>>>> I never said any such thing. >>>>> >>>>> >>>>> >>>>> What I predicted, rather, was that multilayer perceptrons, as they >>>>> existed then, would not (on their own, absent other mechanisms) >>>>> *understand* language. Seven years later, they still haven?t, except >>>>> in the most superficial way. >>>>> >>>>> >>>>> >>>>> I made no comment whatsoever about machine translation, which I view >>>>> as a separate problem, solvable to a certain degree by correspondance >>>>> without semantics. >>>>> >>>>> >>>>> >>>>> I specifically tried to clarify Hinton?s confusion in 2019, but, >>>>> disappointingly, he has continued to purvey misinformation despite that >>>>> clarification. Here is what I wrote privately to him then, which should >>>>> have put the matter to rest: >>>>> >>>>> >>>>> >>>>> You have taken a single out of context quote [from 2015] and >>>>> misrepresented it. The quote, which you have prominently displayed at the >>>>> bottom on your own web page, says: >>>>> >>>>> >>>>> >>>>> Hierarchies of features are less suited to challenges such as >>>>> language, inference, and high-level planning. For example, as Noam Chomsky >>>>> famously pointed out, language is filled with sentences you haven't seen >>>>> before. Pure classifier systems don't know what to do with such sentences. >>>>> The talent of feature detectors -- in identifying which member of some >>>>> category something belongs to -- doesn't translate into understanding >>>>> novel sentences, in which each sentence has its own unique meaning. >>>>> >>>>> >>>>> >>>>> It does *not* say "neural nets would not be able to deal with novel >>>>> sentences"; it says that hierachies of features detectors (on their own, if >>>>> you read the context of the essay) would have trouble *understanding *novel sentences. >>>>> >>>>> >>>>> >>>>> >>>>> Google Translate does yet not *understand* the content of the >>>>> sentences is translates. It cannot reliably answer questions about who did >>>>> what to whom, or why, it cannot infer the order of the events in >>>>> paragraphs, it can't determine the internal consistency of those events, >>>>> and so forth. >>>>> >>>>> >>>>> >>>>> Since then, a number of scholars, such as the the computational >>>>> linguist Emily Bender, have made similar points, and indeed current LLM >>>>> difficulties with misinformation, incoherence and fabrication all follow >>>>> from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on >>>>> the matter with Alexander Koller, >>>>> https://aclanthology.org/2020.acl-main.463.pdf >>>>> , >>>>> also emphasizing issues of understanding and meaning: >>>>> >>>>> >>>>> >>>>> *The success of the large neural language models on many NLP tasks is >>>>> exciting. However, we find that these successes sometimes lead to hype in >>>>> which these models are being described as ?understanding? language or >>>>> capturing ?meaning?. In this position paper, we argue that a system trained >>>>> only on form has a priori no way to learn meaning. .. a clear understanding >>>>> of the distinction between form and meaning will help guide the field >>>>> towards better science around natural language understanding. * >>>>> >>>>> >>>>> >>>>> Her later article with Gebru on language models ?stochastic parrots? >>>>> is in some ways an extension of this point; machine translation requires >>>>> mimicry, true understanding (which is what I was discussing in 2015) >>>>> requires something deeper than that. >>>>> >>>>> >>>>> >>>>> Hinton?s intellectual error here is in equating machine translation >>>>> with the deeper comprehension that robust natural language understanding >>>>> will require; as Bender and Koller observed, the two appear not to be the >>>>> same. (There is a longer discussion of the relation between language >>>>> understanding and machine translation, and why the latter has turned out to >>>>> be more approachable than the former, in my 2019 book with Ernest Davis). >>>>> >>>>> >>>>> >>>>> More broadly, Hinton?s ongoing dismissiveness of research from >>>>> perspectives other than his own (e.g. linguistics) have done the field a >>>>> disservice. >>>>> >>>>> >>>>> >>>>> As Herb Simon once observed, science does not have to be zero-sum. >>>>> >>>>> >>>>> >>>>> Sincerely, >>>>> >>>>> Gary Marcus >>>>> >>>>> Professor Emeritus >>>>> >>>>> New York University >>>>> >>>>> >>>>> >>>>> On Feb 2, 2022, at 06:12, AIhub wrote: >>>>> >>>>> ? >>>>> >>>>> Stephen Hanson in conversation with Geoff Hinton >>>>> >>>>> >>>>> >>>>> In the latest episode of this video series for AIhub.org >>>>> , >>>>> Stephen Hanson talks to Geoff Hinton about neural networks, >>>>> backpropagation, overparameterization, digit recognition, voxel cells, >>>>> syntax and semantics, Winograd sentences, and more. >>>>> >>>>> >>>>> >>>>> You can watch the discussion, and read the transcript, here: >>>>> >>>>> >>>>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>>>> >>>>> >>>>> >>>>> >>>>> About AIhub: >>>>> >>>>> AIhub is a non-profit dedicated to connecting the AI community to the >>>>> public by providing free, high-quality information through AIhub.org >>>>> >>>>> (https://aihub.org/ >>>>> ). >>>>> We help researchers publish the latest AI news, summaries of their work, >>>>> opinion pieces, tutorials and more. We are supported by many leading >>>>> scientific organizations in AI, namely AAAI >>>>> , >>>>> NeurIPS >>>>> , >>>>> ICML >>>>> , >>>>> AIJ >>>>> >>>>> /IJCAI >>>>> , >>>>> ACM SIGAI >>>>> , >>>>> EurAI/AICOMM, CLAIRE >>>>> >>>>> and RoboCup >>>>> >>>>> . >>>>> >>>>> Twitter: @aihuborg >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> Virus-free. www.avast.com >>>>> >>>>> >>>>> >>>>> >>>> -- >>> Gary Cottrell 858-534-6640 FAX: 858-534-7029 >>> Computer Science and Engineering 0404 >>> IF USING FEDEX INCLUDE THE FOLLOWING LINE: >>> CSE Building, Room 4130 >>> University of California San Diego - >>> 9500 Gilman Drive # 0404 >>> >>> La Jolla, Ca. 92093-0404 >>> >>> >>> Email: gary at ucsd.edu >>> Home page: http://www-cse.ucsd.edu/~gary/ >>> Schedule: http://tinyurl.com/b7gxpwo >>> >>> >>> Blind certainty - a close-mindedness that amounts to an imprisonment so >>> total, that the prisoner doesn?t even know that he?s locked up. -David >>> Foster Wallace >>> >>> >>> Power to the people! ?Patti Smith >>> >>> Except when they?re delusional ?Gary Cottrell >>> >>> >>> This song makes me nostalgic for a memory I don't have -- Tess Cottrell >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> *Listen carefully,Neither the VedasNor the Qur'anWill teach you this:Put >>> the bit in its mouth,The saddle on its back,Your foot in the stirrup,And >>> ride your wild runaway mindAll the way to heaven.* >>> >>> -- Kabir >>> >> -- > Gary Cottrell 858-534-6640 FAX: 858-534-7029 > Computer Science and Engineering 0404 > IF USING FEDEX INCLUDE THE FOLLOWING LINE: > CSE Building, Room 4130 > University of California San Diego - > 9500 Gilman Drive # 0404 > La Jolla, Ca. 92093-0404 > > Email: gary at ucsd.edu > Home page: http://www-cse.ucsd.edu/~gary/ > Schedule: http://tinyurl.com/b7gxpwo > > Blind certainty - a close-mindedness that amounts to an imprisonment so > total, that the prisoner doesn?t even know that he?s locked up. -David > Foster Wallace > > > Power to the people! ?Patti Smith > > Except when they?re delusional ?Gary Cottrell > > > This song makes me nostalgic for a memory I don't have -- Tess Cottrell > > > > > > > > > > > *Listen carefully,Neither the VedasNor the Qur'anWill teach you this:Put > the bit in its mouth,The saddle on its back,Your foot in the stirrup,And > ride your wild runaway mindAll the way to heaven.* > > -- Kabir > -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Sun Feb 6 10:12:04 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Sun, 6 Feb 2022 07:12:04 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <02FB3C99-FD80-49BE-BE7C-1969B58E513F@nyu.edu> Dear Stevan, You can define things as you like, but too strict a reading on 4 and 5 leads to weird places. Yes, a literal Turing Tape doesn?t care what my bits are, but I can build algorithms in higher-level languages, or even in low-level hardware, that are sensitive to the structure of the symbols that I am passing. For example, programmers often use signed representations for integers, in which the leftmost bit in the representation of an integer is used to represent whether the number is positive or negative and then further computation (eg comparison at the hardware or software level) is condition thereon [eg https://www.tutorialspoint.com/signed-binary-integers]. In this way, the representation of individual binary integers is far from arbitrary. I would not wish to conclude that they are therefore not symbols, nor that, eg., the comparison I would then between them is not computation or not algorithmic. (And of course the representations of the magnitude of the integer is perfectly orderly, the antithesis ofarbitrary) All of this can be mapped onto a Turing tape, but it doesn?t mean that the microprocessor opcodes or programming instructions that manipulate numbers with reference to the sign bit aren?t manipulating symbols or aren?t algorithmic. It means that we need ways of talking about symbols that are composable, potentially in meaningful ways. I take embeddings to be a very useful extension of that idea, of composed symbols, in which you wind up with a set of stable encodings in which the individual bits are not in fact arbitrary. You still wind up with a (useful) string of bits to represent something. (Binary or otherwise). But you can leverage those bits in interesting ways. Yes, you can do whatever you wind up with in a cumbersome fashion on a Turing Tape, but so what? That fact alone doesn?t tell us all that much about the nature of the computations we want to perform, since (resource and timing limits aside) there is nothing on the table or seriously envisioned that couldn?t be mapped onto the infinite space of things that could be so mapped. Gary > On Feb 6, 2022, at 05:48, Stevan Harnad wrote: > > ? > 0. It might help if we stop ?cognitizing? computation and symbols. > > 1. Computation is not a subset of AI. > > 2. AI (whether ?symbolic? AI or ?connectionist? AI) is an application of computation to cogsci. > > 3. Computation is the manipulation of symbols based on formal rules (algorithms). > > 4. Symbols are objects or states whose physical ?shape? is arbitrary in relation to what they can be used and interpreted as referring to. > > 5. An algorithm (executable physically as a Turing Machine) manipulates symbols based on their (arbitrary) shapes, not their interpretations (if any). > > 6. The algorithms of interest in computation are those that have at least one meaningful interpretation. > > 7. Examples of symbol shapes are numbers (1, 2, 3), words (one, two, three; onyx, tool, threnody), or any object or state that is used as a symbol by a Turing Machine that is executing an algorithm (symbol-manipulation rules). > > 8. Neither a sensorimotor feature of an object in the world, nor a sensorimotor feature-detector of a robot interacting with the world, is a symbol (except in the trivial sense that any arbitrary shape can be used as a symbol). > > 9. What sensorimotor features and sensorimotor feature-detectors (whether ?symbolic? or ?connectionist?) might be good for is connecting symbols inside symbol systems (e.g., robots) to the objects that they can be interpreted as referring to. > > 10. If you are interpreting ?symbol? in a wider sense than this formal, literal one, then you are closer to lit-crit than cogsci. > > Stevan Harnad > > > From: Asim Roy > Date: Saturday, February 5, 2022 at 11:59 PM > To: Gary Marcus , Stephen Jos? Hanson > Cc: connectionists at mailman.srv.cs.cmu.edu , Stevan Harnad , Francesca Rossi2 , Artur Garcez , Anima Anandkumar , Luis Lamb , Gadi Singer , Josh Tenenbaum , AIhub > Subject: RE: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > CAUTION: This e-mail originated outside the University of Southampton. > There was another recent attempt to take down the grandmother cell idea: > > Frontiers | The Value of Failure in Science: The Story of Grandmother Cells in Neuroscience | Neuroscience (frontiersin.org) > > And here?s my commentary defending grandmother cells: > > Frontiers | Commentary: The Value of Failure in Science: The Story of Grandmother Cells in Neuroscience | Neuroscience (frontiersin.org) > > By the way, we have had vigorous private arguments over the years about grandmother cells and many were involved ? from Jay McClelland and Christof Koch to Walter Freeman and Bernard Baars. As far as I can tell, the brain uses abstractions at the single cell level that can be argued to be symbols. Short arguments are in the commentary, based on observations by neurophysiologists themselves. > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > From: Connectionists On Behalf Of Gary Marcus > Sent: Friday, February 4, 2022 12:53 PM > To: Stephen Jos? Hanson > Cc: connectionists at mailman.srv.cs.cmu.edu; Stevan Harnad ; Francesca Rossi2 ; Artur Garcez ; Anima Anandkumar ; Luis Lamb ; Gadi Singer ; Josh Tenenbaum ; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > ?Steve, > > The phrase I always liked was ?poverty of the imagination arguments?; I share your disdain for them. But that?s why I think you should be careful of any retreat into biological plausibility. As even Jay McClelland has acknowledged, we do know that some humans some of the time manipulate symbols. So wetware-based symbols are not literally biologically impossible; the real question for cognitive neuroscience is about the scope and development of symbols. > > For engineering, the real question is, are they useful. Certainly for software engineering in general, they are indispensable. > > Beyond this, none of the available AI approaches map particularly neatly onto what we know about the brain, and none of what we know about the brain is understood well enough to solve AI. All the examples you point to, for instance, are actually controversial, not decisive. As you probably know, for example, Nancy Kanwisher has a different take on domain-specificity than you do (https://web.mit.edu/bcs/nklab/), with evidence of specialization early in life, and Jeff Bowers has argued that the grandmother cell hypothesis has been dismissed prematurely (https://jeffbowers.blogs.bristol.ac.uk/blog/grandmother-cells/); there?s also a long literature on the possible neural realization of rules, both in humans and other animals. > > I don?t know what the right answers are there, but nor do I think that neurosymbolic systems are beholden to them anymore than CNNs are bound to whether or not the brain performs back-propagation. > > Finally, as a reminder, ?Distributed? per se in not the right question; in some technical sense ASCII encodings are distributed, and about as symbolic as you can get. The proper question is really what you do with your encodings; the neurosymbolic approach is trying to broaden the available range of options. > > Gary > > On Feb 4, 2022, at 07:04, Stephen Jos? Hanson wrote: > > ? > Well I don't like counterfactual arguments or ones that start with "It can't be done with neural networks.."--as this amounts to the old Rumelhart saw, of "proof by lack of imagination". > > I think my position and others (I can't speak for Geoff and won't) is more of a "purist" view that brains have computationally complete representational power to do what ever is required of human level mental processing. AI symbol systems are remote descriptions of this level of processing. Looking at 1000s of brain scans, one begins to see a pattern of interacting large and smaller scale networks, probably related to Resting state and the Default Mode networks in some important competitive way. But what one doesn't find is modular structure (e.g. face area.. nope) or evidence of "symbols" being processed. Research on Numbers is interesting in this regard, as number representation should provide some evidence of discrete symbol processing as would letters. But again the processing states from brain imaging more generally appear to be distributed representations of some sort. > > One other direction has to do with prior rules that could be neurally coded and therefore provide an immediate bias in learning and thus dramatically reduce the number of examples required for asymptotic learning. Some of this has been done with pre-training-- on let's say 1000s of videos that are relatively generic, prior to learning on a small set of videos related to a specific topic-- say two individuals playing a monopoly game. In that case, no game-like videos were sampled in the pre-training, and the LSTM was trained to detect change point on 2 minutes of video, achieving a 97% match with human parsers. In these senses I have no problem with this type of hybrid training. > > Steve > > On 2/4/22 9:07 AM, Gary Marcus wrote: > ?The whole point of the neurosymbolic approach is to develop systems that can accommodate both vectors and symbols, since neither on their own seems adequate. > > If there are arguments against trying to do that, we would be interested. > > > > On Feb 4, 2022, at 4:17 AM, Stephen Jos? Hanson wrote: > > ? > Geoff's position is pretty clear. He said in the conversation we had and in this thread, "vectors of soft features", > > Some of my claim is in several of the conversations with Mike Jordan and Rich Sutton, but briefly, there are a number of > very large costly efforts from the 1970s and 1980s, to create, deploy and curate symbol AI systems that were massive failures. Not counterfactuals, but factuals that failed. The MCC comes to mind with Adm Bobby Inmann's national US mandate to counter the Japanese so called"Fifth-generation AI systems" as a massive failure of symbolic AI. > > -------------------- > > In 1982, Japan launched its Fifth Generation Computer Systems project (FGCS), designed to develop intelligent software that would run on novel computer hardware. As the first national, large-scale artificial intelligence (AI) research and development (R&D) project to be free from military influence and corporate profit motives, the FGCS was open, international, and oriented around public goods. > > On 2/3/22 6:34 PM, Francesca Rossi2 wrote: > Hi all. > > Thanks Gary for adding me to this thread. > > I also would be interested in knowing why Steve thinks that NS AI did not work in the past, and why this is an indication that it cannot work now or in the future. > > Thanks, > Francesca. > ------------------ > > Francesca Rossi > IBM Fellow and AI Ethics Global Leader > T.J. Watson Research Center, Yorktown Heights, USA > +1-617-3869639 > > ________________________________________ > From: Artur Garcez > Sent: Thursday, February 3, 2022 6:00 PM > To: Gary Marcus > Cc: Stephen Jos? Hanson; Geoffrey Hinton; AIhub; connectionists at mailman.srv.cs.cmu.edu; Luis Lamb; Josh Tenenbaum; Anima Anandkumar; Francesca Rossi2; Swarat Chaudhuri; Gadi Singer > Subject: [EXTERNAL] Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally ZjQcmQRYFpfptBannerStart > This Message Is From an External Sender > This message came from outside your organization. > ZjQcmQRYFpfptBannerEnd > > It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. > > Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally symbolic), there is a clear resurgence of neuro-symbolic approaches. > > Best wishes, > Artur > > > On Thu, Feb 3, 2022 at 5:00 PM Gary Marcus > wrote: > Steve, > > I?d love to hear you elaborate on this part, > > Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. > > > I?d love your take on why you think the impetus towards hybrid models ended badly before, and why you think that the mistakes of the past can?t be corrected. Also it? would be really instructive to compare with deep learning, which lost steam for quite some time, but reemerged much stronger than ever before. Might not the same happen with hybrid models? > > I am cc?ing some folks (possibly not on this list) who have recently been sympathetic to hybrid models, in hopes of a rich discussion. (And, Geoff, still cc?d, I?d genuinely welcome your thoughts if you want to add them, despite our recent friction.) > > Cheers, > Gary > > > On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson > wrote: > > > I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. > > Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. > > Best and hope you are doing well. > > Steve > > -- > > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 19969 bytes Desc: not available URL: From dayan at tue.mpg.de Sat Feb 5 16:46:17 2022 From: dayan at tue.mpg.de (Peter Dayan) Date: Sat, 5 Feb 2022 22:46:17 +0100 Subject: Connectionists: Cognitive & Computational Neuroscience postdoc In-Reply-To: <20211116211002.c4m2vcdxbyahtfth@tuebingen.mpg.de> References: <20211014105046.wtbpasjz3dbrdz6v@tuebingen.mpg.de> <20211116211002.c4m2vcdxbyahtfth@tuebingen.mpg.de> Message-ID: <20220205214617.wfon2plevuxo6fva@tuebingen.mpg.de> Cognitive and Computational Neuroscientist Postdoctoral position to understand the function of long range connectivity in human cognition We are seeking a new colleague to work on a collaborative project studying human cognition between the laboratories of Professor Linda Richards at Washington University in St. Louis and Professor Peter Dayan at Max Planck Institute for Biological Cybernetics in Tubingen. The position will be based at Washington University in St. Louis. We are studying how people with corpus callosum dysgenesis, a major change in their global brain wiring connectivity, are able to process and integrate information during tasks involving decision making, social intelligence and memory. We are seeking a creative scientist with a background in cognitive neuroscience and computational neuroscience to develop analysis tools and pipelines and new models of cognitive function that will help guide our experimental approaches. Applicants should submit their CV and a cover letter explaining their background and interest in the position and working in both laboratories by April 4th, 2022. Please see job posting for additional information. Linda J. Richards AO, FAA, FAHMS, PhD Edison Professor and Chair, Department of Neuroscience Director, McDonnell Center for Cellular & Molecular Neurobiology Washington University School of Medicine MSC 8108-12-8, 660 S. Euclid Avenue, St. Louis, MO 63110 Office Phone +1 314 362 4198; Cell Phone +1 314 532 7187 Linda.Richards at wustl.edu (My pronouns: she, her, hers) -------------- next part -------------- A non-text attachment was scrubbed... Name: smime.p7s Type: application/x-pkcs7-signature Size: 4936 bytes Desc: not available URL: From george at cs.ucy.ac.cy Sun Feb 6 07:29:50 2022 From: george at cs.ucy.ac.cy (George A. Papadopoulos) Date: Sun, 6 Feb 2022 14:29:50 +0200 Subject: Connectionists: 8th IEEE International Smart Cities Conference (ISC2 2022): First Call for Papers Message-ID: <4S4GUOF1-GJG8-5LJ-0OWX-HEX8GTOU2OHY@cs.ucy.ac.cy> *** First Call for Papers *** 8th IEEE International Smart Cities Conference (ISC2 2022) "Community Smartification and towards ZERO emission Smart Cities for a Green New Era" September 26-29, 2022, Aliathon Resort, Paphos, Cyprus https://attend.ieee.org/isc2-2022? The Call-for-Papers for the 8th IEEE International Smart Cities Conference (ISC2 2022) is OPEN! The IEEE ISC2 2022 Organizing Committee is glad to announce that it is currently accepting papers for the 8th edition of the IEEE International Smart Cities Conference.? This year, the IEEE ISC2 will be held in-person on September 26-29, 2022, in Paphos, Cyprus and the theme of the conference this year is ?Community Smartification and towards ZERO emission Smart Cities for a Green New Era.? The IEEE ISC2 is the flagship conference sponsored by the IEEE Smart Cities Technical Community, a coalition of eight IEEE technical societies and organizations. Learn more about IEEE Smart Cities here. Besides contributions addressing the conference theme, authors are welcome to submit their original research results in traditional topics across broad application and functional domains, within the context of smart urban infrastructure systems. The complete list of technical areas can be found here: https://attend.ieee.org/isc2-2022/call-for-papers/ . Important Dates ? Conference Paper Submission Deadline ? May 15, 2022 ? Conference Acceptance Notification ? July 15, 2022 ? Conference Camera-ready Deadline ? July 31, 2022 ? Conference Dates ? September 26-29, 2022 Paper Submission Guidelines Prospective authors are invited to submit high quality original (Full or Short) papers via the EasyChair submission site at https://easychair.org/conferences/?conf=isc22022 . Full papers should describe novel research contributions with evaluation results and are limited to seven (7) pages. Short papers, limited in length to four (4) pages, should be more visionary in nature and are meant to discuss new challenges and visions, highlight early research results, and explore novel research directions. All submitted papers must be unpublished and not considered elsewhere for publication, should be written in English and formatted according to IEEE Template ( https://www.ieee.org/conferences/publishing/templates.html ). Each submitted paper will pass through the standard IEEE peer-review process. If accepted and presented at the conference, it will appear in the conference proceedings and be submitted for inclusion in the IEEE Xplore Digital Library. The conference organizers are currently negotiating a number of special issues with high quality journals. More information will be available on the conference web site. The best papers will be awarded in the conference Best Paper Award and a Best Student Paper Award contests.? For more information, please visit the conference website: https://attend.ieee.org/isc2-2022 . Should you need further clarifications or have any inquiries, please do not hesitate to contact us at: isc22022 at easychair.org . If you want to become a Sponsor of our event, please send an email to: smartcities-info at ieee.org . Organizing Committee https://attend.ieee.org/isc2-2022/chairs/ Special Track Chairs Committee https://attend.ieee.org/isc2-2022/special-tracks-committee2/ IEEE Smart Cities https://smartcities.ieee.org -------------- next part -------------- An HTML attachment was scrubbed... URL: From george at cs.ucy.ac.cy Sat Feb 5 08:25:00 2022 From: george at cs.ucy.ac.cy (George A. Papadopoulos) Date: Sat, 5 Feb 2022 15:25:00 +0200 Subject: Connectionists: 38th IEEE International Conference on Software Maintenance and Evolution (ICSME 2022): Third Call for Contributions Message-ID: *** Third Call for Contributions *** 38th IEEE International Conference on Software Maintenance and Evolution (ICSME 2022) 3?7 October, 2022, 4* Atlantica Miramare Beach Hotel, Limassol, Cyprus https://cyprusconferences.org/icsme2022/ Goals and Scope The IEEE International Conference on Software Maintenance and Evolution (ICSME) is the premier forum for researchers and practitioners to present and discuss the most recent innovations, trends, experiences, and challenges in software maintenance and evolution. We invite high quality submissions describing significant, original, and unpublished results related to but not limited to any of the following software maintenance and evolution topics (in alphabetical order): ? Change and defect management ? Code cloning and provenance ? Concept and feature location ? Continuous integration/deployment ? Empirical studies of software maintenance and evolution ? Evolution of non-code artifacts ? Human aspects of software maintenance and evolution ? Maintenance and evolution of model-based methods ? Maintenance and evolution processes ? Maintenance and evolution of mobile apps ? Maintenance and evolution of service-oriented and cloud computing systems ? Maintenance versus release process ? Mining software repositories ? Productivity of software engineers during maintenance and evolution ? Release engineering ? Reverse engineering and re-engineering ? Run-time evolution and dynamic configuration ? Software and system comprehension ? Software migration and renovation ? Software quality assessment ? Software refactoring and restructuring ? Software testing theory and practice ? Source code analysis and manipulation ? Technical Debt ICSME welcomes innovative ideas that are timely, well presented, and evaluated. All submissions must position themselves within the existing literature, describe the relevance of the results to specific software engineering goals, and include a clear motivation and presentation of the work. ICSME invites contributions to a number of different tracks. These are listed below with more information, including submission instructions, available on the conference web site (submission dates are midnight, AoE, UTC-12). The submission link for all tracks is: https://easychair.org/conferences/?conf=icsme2022 . Please use this link and choose the appropriate track for your submission. Research Track ? Abstract Submission: March 25th, 2022. ? Paper Submission: April 1st, 2022. ? Notification: June 10th, 2022. ? Camera-Ready: July 1st, 2022. Please refer to https://cyprusconferences.org/icsme2022/call-for-research-track/ . Doctoral Symposium ? Paper Submission: July 8th, 2022. ? Notification: August 1st, 2022. ? Camera-Ready: August 19th, 2022. Please refer to https://cyprusconferences.org/icsme2022/call-for-doctoral-symposium/ . Journal First Track ? Paper Submission: June 17th, 2022. ? Notification: July 8th, 2022. Please refer to https://cyprusconferences.org/icsme2022/call-for-journal-first-track/ . Tool Demo Track ? Abstract Submission: June 13th, 2022. ? Paper Submission: June 20th, 2022. ? Notification: July 19th, 2022. ? Camera-Ready: July 26th, 2022. Please refer to https://cyprusconferences.org/icsme2022/tool-demo-track/ . Joint Artifact Evaluation Track and ROSE Festival ? Artifact Submission: August 26th, 2022. ? Notification: September 16th, 2022. Please refer to https://cyprusconferences.org/icsme2022/call-for-joint-artifact-evaluation-track-and-rose-festival-track/ . Industry Track ? Full/Short Papers Abstract Submission: April 22nd, 2022. ? Full/Short Paper Submission: April 29th, 2022. ? Full/Short Paper Notification: June 15th, 2022. ? Full/Short Camera-Ready: July 15th, 2022. ? Extended Abstract Submission: May 20th, 2022. ? Extended Abstract Notification: June 15th, 2022. ? Camera-Ready: July 15th, 2022. Please refer to https://cyprusconferences.org/icsme2022/call-for-industry-track/ . News Ideas and Emerging Results Track ? Abstract Submission: June 17th, 2022. ? Paper Submission: June 24th, 2022. ? Notification: July 18th, 2022. ? Camera-Ready: July 31st, 2022. Please refer to https://cyprusconferences.org/icsme2022/new-ideas-and-emerging-results/ . Registered Reports Track ? Initial Report Submission: June 3rd, 2022. ? Feedback from PC: July 8th, 2022. ? Authors Response: July 22nd, 2022. ? Notification for Stage 1: August 12th, 2022. ? Submission of Accepted Report: August 19th, 2022. Please refer to https://cyprusconferences.org/icsme2022/registered-reports-track/ . Organisation General Chairs ? Rainer Koschke, University of Bremen, Germany ? George A. Papadopoulos, University of Cyprus, Cyprus Program Chairs ? Paris Avgeriou, University of Groningen, The Netherlands ? Dave Binkley, Loyola University Maryland, USA Local Organising Chair and Industry Liaison ? Georgia M. Kapitsaki, University of Cyprus, Cyprus New Ideas and Emerging Results Track Chairs ? Eleni Constantinou, Eindhoven University of Technology, The Netherlands ? Christian Newman, Rochester Institute of Technology, USA Tool Demonstrations Chairs ? Sherlock Licorish, University of Otago, New Zealand ? Gilles Perrouin, University of Namur, Belgium Industry Track Chairs ? Andrea Capiluppi, University of Groningen, The Netherlands ? Shi Han, Microsoft Research Beijing, China Journal First Track Chairs ? Alexander Chatzigeorgiou, University of Macedonia, Greece ? Amjed Tahir, Massey University, New Zealand Doctoral Symposium ? Matthias Galster, University of Canterbury, New Zealand ? Mark Hills, East Carolina University, USA Joint Artifact Evaluation Track and ROSE Festival Chairs ? Maria Papoutsoglou, University of Cyprus, Cyprus ? Christoph Treude, University of Melbourne, Australia Most Influential Paper Awards Chairs ? Massimiliano Di Penta, University of Sannio, Italy ? Jonathan I. Maletic, Kent State University, USA Diversity and Inclusion ? Hadil?Abukwaik, ABB Corporate Research, Germany ? Sonia Haiduc, Florida State University, USA Registered Reports ? Maria Teresa Baldassarre, University of Bari, Italy ? Mike Papadakis, University of Luxembourg, Luxembourg -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Sun Feb 6 12:05:38 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Sun, 6 Feb 2022 12:05:38 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Asim, I try to be brief so that I can explain why many of us have missed, and will continue to miss, the boat. In some of my talks, I have a ppt slide "The brain is like blindmen and an elephant". Unfortunately, your "identify objects based on its parts" is a good traditional idea from pattern recognition that is still a blindman. Your idea does not explain many other problems without which we will never understand a biological brain. For example, your idea does not explain how the brain learns planning and discovery in a cluttered world. We must solve many million-dollar problems holistically. Please watch my YouTube video: Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar Problems Solved https://youtu.be/Dgx1dLCdSKY Best regards, -John On Sat, Feb 5, 2022 at 12:01 AM Asim Roy wrote: > I am responding to this part of Geoffrey Hinton?s note: > > > > *?I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. We can introspect on how we do it > and this may or may not give some insight into how we check our answer, but > the immediate sense that a handwritten 2 is a 2 is computed by a neural net > that is not functionally equivalent to any simple and easily explainable > procedure.?* > > > > The causal explanation is actually done quite simply, and we are doing it > currently. I can talk about this now because Arizona State University (ASU) > has filed a provisional patent application on the technology. The basic > idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable > Artificial Intelligence (darpa.mil) > ) and > illustrated in the figure below. The idea is to identify objects based on > its parts. So, the figure below says that it?s a cat because it has fur, > whiskers, and claws plus an unlabeled visual feature. I am not sure if > DARPA got anything close to this from its funding of various entities. What > this means is that you need a parts model. And we do that. In the case of > MNIST handwritten digits that Geoff mentions, we ?teach? this parts model > what the top part of a digit ?3? looks like, what the bottom part looks > like and so on. And we also teach connectivity between parts and the > composition of objects from parts. And we do that for all digits. And we > get a symbolic model sitting on top of a CNN model that provides the > explanation that Geoff is referring to as the causal explanation. This > ?teaching? is similar to the way you would teach a kid to recognize > different digits. > > > > An advantage of this parts model, in addition to being in an explainable > symbolic form, is robustness to adversarial attack. We recently tested on > the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast > gradient method to 27%, our XAI model maintained an accuracy of 90%, > probably higher. In general, it would be hard to make a school bus look > like an ostrich, with a few pixel changes, if you can identify the parts of > a school bus and an ostrich. > > > > A parts model that DARPA wanted provides both a symbolic explanation and > adversarial protection. The problem that Geoffrey is referring to is solved. > > > > I am doing a tutorial on this at IEEE World Congress on Computational > Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy > 18-23 July ). I am copying the organizers and want > to thank them for accepting the tutorial proposal. The only other > presentation I have done on this is at a Military Operations Research > Society (MORS) meeting last December. > > > > So, back to the future. Hybrid models might indeed save deep learning > models and let us deploy these models without concern. We might not even > need adversarial training of any kind. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > www.teuvonet.com > > > > > [image: Timeline Description automatically generated] > > > > *From:* Connectionists *On > Behalf Of *Geoffrey Hinton > *Sent:* Friday, February 4, 2022 1:24 PM > *To:* Dietterich, Thomas > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. We can introspect on how we do it > and this may or may not give some insight into how we check our answer, but > the immediate sense that a handwritten 2 is a 2 is computed by a neural net > that is not functionally equivalent to any simple and easily explainable > procedure. > > > > This does not mean that we should give up on trying to make artificial > neural nets work more like real ones. People can see a tilted square as > either an upright diamond or a tilted square and, so far as I know, a > convnet does not exhibit this type of alternate percept. People seem to > impose hierarchical structural descriptions on images and sound waves and > they clearly impose intrinsic coordinate frames on wholes and parts. If > this is what Gary means by symbolic then I don?t disagree that neural nets > should do symbol processing. However, there is a very different meaning of > "symbolic". A pure atomic symbol has no internal structure. The form of the > symbol itself tells you nothing about what it denotes. The only relevant > properties it has are that it's identical to other instances of the > same symbol and different from all other symbols. That's totally different > from a neural net that uses embedding vectors. Embedding vectors have a > rich internal structure that dictates how they interact with other > embedding vectors. What I really object to is the following approach: Start > with pure symbols and rules for how to manipulate structures made out of > pure symbols. These structures themselves can be denoted by symbols that > correspond to memory addresses where the bits in the address tell you > nothing about the content of the structure at that address. Then when the > rule-based approach doesn't work for dealing with the real world (e.g. > machine translation) try to use neural nets to convert the real world into > pure symbols and then carry on with the rule-based approach. That is like > using an electric motor to inject the gasoline into the same old gasoline > engine instead of just replacing the gasoline engine with an electric motor. > > > > > > On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: > > ?Understanding? is not a Boolean. It is a theorem that no system can > enumerate all of the consequences of a state of affairs in the world. > > > > For low-stakes application work, we can be satisfied by a system that > ?does the right thing?. If the system draws a good picture, that?s > sufficient. It ?understood? the request. > > > > But for higher-stakes applications---and for advancing the science---we > seek a causal account of how the components of a system cause it to do the > right thing. We are hoping that a small set of mechanisms can produce broad > coverage of intelligent behavior. This gives us confidence that the system > will respond correctly outside of the narrow tasks on which we have tested > it. > > > > --Tom > > > > Thomas G. Dietterich, Distinguished Professor Emeritus > > School of Electrical Engineering and Computer > Science > > US Mail: 1148 Kelley Engineering Center > > > > Office: 2067 Kelley Engineering Center > > Oregon State Univ., Corvallis, OR 97331-5501 > > Voice: 541-737-5559; FAX: 541-737-1300 > > URL: http://web.engr.oregonstate.edu/~tgd/ > > > > > *From:* Connectionists *On > Behalf Of *Gary Marcus > *Sent:* Thursday, February 3, 2022 8:26 AM > *To:* Danko Nikolic > *Cc:* connectionists at mailman.srv.cs.cmu.edu; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > [This email originated from outside of OSU. Use caution with links and > attachments.] > > Dear Danko, > > > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, > in which he said (paraphrasing from memory, because I don?t remember the > precise words) that the famous 200 Quoc Le unsupervised model [ > https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf > ] > had learned the concept of a ca. In reality the model had clustered > together some catlike images based on the image statistics that it had > extracted, but it was a long way from a full, counterfactual-supporting > concept of a cat, much as you describe below. > > > > I fully agree with you that the reason for even having a semantics is as > you put it, "to 1) learn with a few examples and 2) apply the knowledge to > a broad set of situations.? GPT-3 sometimes gives the appearance of having > done so, but it falls apart under close inspection, so the problem remains > unsolved. > > > > Gary > > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > > > G. Hinton wrote: "I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster wearing a red hat and > it draws such a picture, it understood the request." > > > > I would like to suggest why drawing a hamster with a red hat does not > necessarily imply understanding of the statement "hamster wearing a red > hat". > > To understand that "hamster wearing a red hat" would mean inferring, in > newly emerging situations of this hamster, all the real-life > implications that the red hat brings to the little animal. > > > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > ...and so on. > > > > Countless many questions can be asked. One has understood "hamster wearing > a red hat" only if one can answer reasonably well many of such real-life > relevant questions. Similarly, a student has understood materias in a class > only if they can apply the materials in real-life situations (e.g., > applying Pythagora's theorem). If a student gives a correct answer to a > multiple choice question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > > > I also suggest that understanding also comes together with effective > learning: We store new information in such a way that we can recall it > later and use it effectively i.e., make good inferences in newly emerging > situations based on this knowledge. > > > > In short: Understanding makes us humans able to 1) learn with a few > examples and 2) apply the knowledge to a broad set of situations. > > > > No neural network today has such capabilities and we don't know how to > give them such capabilities. Neural networks need large amounts of > training examples that cover a large variety of situations and then > the networks can only deal with what the training examples have already > covered. Neural networks cannot extrapolate in that 'understanding' sense. > > > > I suggest that understanding truly extrapolates from a piece of knowledge. > It is not about satisfying a task such as translation between languages or > drawing hamsters with hats. It is how you got the capability to complete > the task: Did you only have a few examples that covered something different > but related and then you extrapolated from that knowledge? If yes, this is > going in the direction of understanding. Have you seen countless examples > and then interpolated among them? Then perhaps it is not understanding. > > > > So, for the case of drawing a hamster wearing a red hat, understanding > perhaps would have taken place if the following happened before that: > > > > 1) first, the network learned about hamsters (not many examples) > > 2) after that the network learned about red hats (outside the context of > hamsters and without many examples) > > 3) finally the network learned about drawing (outside of the context of > hats and hamsters, not many examples) > > > > After that, the network is asked to draw a hamster with a red hat. If it > does it successfully, maybe we have started cracking the problem of > understanding. > > > > Note also that this requires the network to learn sequentially without > exhibiting catastrophic forgetting of the previous knowledge, which is > possibly also a consequence of human learning by understanding. > > > > > > Danko > > > > > > > > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > > > > Virus-free. *www.avast.com* > > > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy <*ASIM.ROY at asu.edu*> wrote: > > > Without getting into the specific dispute between Gary and Geoff, I think > with approaches similar to GLOM, we are finally headed in the right > direction. There?s plenty of neurophysiological evidence for single-cell > abstractions and multisensory neurons in the brain, which one might claim > correspond to symbols. And I think we can finally reconcile the decades old > dispute between Symbolic AI and Connectionism. > > > > > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > GARY: I have *never* called for dismissal of neural networks, but rather > for some hybrid between the two (as you yourself contemplated in 1991); the > point of the 2001 book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols could complement > them. > > > > > > Asim Roy > > > Professor, Information Systems > > > Arizona State University > > > *Lifeboat Foundation Bios: Professor Asim Roy* > > > *Asim Roy | iSearch (asu.edu)* > > > > > > > > > > > > > *From: Connectionists On > Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: > Geoffrey Hinton Cc: AIhub ; > connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen > Hanson in conversation with Geoff Hinton > * > > > > > Dear Geoff, and interested others, > > > > > > What, for example, would you make of a system that often drew the > red-hatted hamster you requested, and perhaps a fifth of the time gave you > utter nonsense? Or say one that you trained to create birds but sometimes > output stuff like this: > > > > > > > > > > > > One could > > > > > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > > or > > > b. wonder if it might be possible that sometimes the system gets the right > answer for the wrong reasons (eg partial historical contingency), and > wonder whether another approach might be indicated. > > > > > > Benchmarks are harder than they look; most of the field has come to > recognize that. The Turing Test has turned out to be a lousy measure of > intelligence, easily gamed. It has turned out empirically that the Winograd > Schema Challenge did not measure common sense as well as Hector might have > thought. (As it happens, I am a minor coauthor of a very recent review on > this very topic: *https://arxiv.org/abs/2201.02387*) But its conquest in > no way means machines now have common sense; many people from many > different perspectives recognize that (including, e.g., Yann LeCun, who > generally tends to be more aligned with you than with me). > > > > > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote > me; but what you said about me and machine translation remains your > invention, and it is inexcusable that you simply ignored my 2019 > clarification. On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains unsolved. > > > > > > All of the problems LLMs have with coherence, reliability, truthfulness, > misinformation, etc stand witness to that fact. (Their persistent inability > to filter out toxic and insulting remarks stems from the same.) I am hardly > the only person in the field to see that progress on any given benchmark > does not inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > > > > > With respect to embeddings: Embeddings are very good for natural language > *processing*; but NLP is not the same as NL*U* ? when it comes to > *understanding*, their worth is still an open question. Perhaps they will > turn out to be necessary; they clearly aren?t sufficient. In their extreme, > they might even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a specific set of > numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > > > > > (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > > > > Notably absent from your email is any kind of apology for misrepresenting > my position. It?s fine to say that ?many people thirty years ago once > thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. > I have consistently felt throughout our interactions that you have mistaken > me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me > for having made that error. I am still not he. > > > > > > Which maybe connects to the last point; if you read my work, you would see > thirty years of arguments *for* neural networks, just not in the way that > you want them to exist. I have ALWAYS argued that there is a role for them; > characterizing me as a person ?strongly opposed to neural networks? misses > the whole point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > > > > > In the last two decades or so you have insisted (for reasons you have > never fully clarified, so far as I know) on abandoning symbol-manipulation, > but the reverse is not the case: I have *never* called for dismissal of > neural networks, but rather for some hybrid between the two (as you > yourself contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded and broke down, > and where symbols could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > > > > > Gary > > > > > > > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton <*geoffrey.hinton at gmail.com*> > wrote: > > > ? > > > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > > > > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > > > > > > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > > > > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > > > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > > > > > Geoff > > > > > > > > > > > > > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus <*gary.marcus at nyu.edu*> wrote: > > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural > network community, > > > > > > There has been a lot of recent discussion on this list about framing and > scientific integrity. Often the first step in restructuring narratives is > to bully and dehumanize critics. The second is to misrepresent their > position. People in positions of power are sometimes tempted to do this. > > > > > > The Hinton-Hanson interview that you just published is a real-time example > of just that. It opens with a needless and largely content-free personal > attack on a single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive thing it says is > false. > > > > > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t > be able to do machine translation.? > > > > > > I never said any such thing. > > > > > > What I predicted, rather, was that multilayer perceptrons, as they existed > then, would not (on their own, absent other mechanisms) *understand* language. > Seven years later, they still haven?t, except in the most superficial way. > > > > > > > I made no comment whatsoever about machine translation, which I view as a > separate problem, solvable to a certain degree by correspondance without > semantics. > > > > > > I specifically tried to clarify Hinton?s confusion in 2019, but, > disappointingly, he has continued to purvey misinformation despite that > clarification. Here is what I wrote privately to him then, which should > have put the matter to rest: > > > > > > You have taken a single out of context quote [from 2015] and > misrepresented it. The quote, which you have prominently displayed at the > bottom on your own web page, says: > > > > > > Hierarchies of features are less suited to challenges such as language, > inference, and high-level planning. For example, as Noam Chomsky famously > pointed out, language is filled with sentences you haven't seen > before. Pure classifier systems don't know what to do with such sentences. > The talent of feature detectors -- in identifying which member of some > category something belongs to -- doesn't translate into understanding > novel sentences, in which each sentence has its own unique meaning. > > > > > > It does *not* say "neural nets would not be able to deal with novel > sentences"; it says that hierachies of features detectors (on their own, if > you read the context of the essay) would have trouble *understanding *novel sentences. > > > > > > > Google Translate does yet not *understand* the content of the sentences > is translates. It cannot reliably answer questions about who did what to > whom, or why, it cannot infer the order of the events in paragraphs, it > can't determine the internal consistency of those events, and so forth. > > > > > > Since then, a number of scholars, such as the the computational linguist > Emily Bender, have made similar points, and indeed current LLM difficulties > with misinformation, incoherence and fabrication all follow from these > concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter > with Alexander Koller, *https://aclanthology.org/2020.acl-main.463.pdf*, > also emphasizing issues of understanding and meaning: > > > > > > *The success of the large neural language models on many NLP tasks is > exciting. However, we find that these successes sometimes lead to hype in > which these models are being described as ?understanding? language or > capturing ?meaning?. In this position paper, we argue that a system trained > only on form has a priori no way to learn meaning. .. a clear understanding > of the distinction between form and meaning will help guide the field > towards better science around natural language understanding. > * > > > > > Her later article with Gebru on language models ?stochastic parrots? is in > some ways an extension of this point; machine translation requires mimicry, > true understanding (which is what I was discussing in 2015) requires > something deeper than that. > > > > > > Hinton?s intellectual error here is in equating machine translation with > the deeper comprehension that robust natural language understanding will > require; as Bender and Koller observed, the two appear not to be the same. > (There is a longer discussion of the relation between language > understanding and machine translation, and why the latter has turned out to > be more approachable than the former, in my 2019 book with Ernest Davis). > > > > > > More broadly, Hinton?s ongoing dismissiveness of research from > perspectives other than his own (e.g. linguistics) have done the field a > disservice. > > > > > > As Herb Simon once observed, science does not have to be zero-sum. > > > > > > Sincerely, > > > Gary Marcus > > > Professor Emeritus > > > New York University > > > > > > On Feb 2, 2022, at 06:12, AIhub <*aihuborg at gmail.com*> wrote: > > > ? > > > Stephen Hanson in conversation with Geoff Hinton > > > > > > In the latest episode of this video series for *AIhub.org*, Stephen > Hanson talks to Geoff Hinton about neural networks, backpropagation, > overparameterization, digit recognition, voxel cells, syntax and semantics, > Winograd sentences, and more. > > > > > > You can watch the discussion, and read the transcript, here: > > > > > *https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/* > > > > > > About AIhub: > > > AIhub is a non-profit dedicated to connecting the AI community to the > public by providing free, high-quality information through *AIhub.org* ( > *https://aihub.org/*). We help researchers publish the latest AI news, > summaries of their work, opinion pieces, tutorials and more. We are > supported by many leading scientific organizations in AI, namely *AAAI*, > *NeurIPS*, *ICML*, *AIJ*/*IJCAI*, *ACM SIGAI*, EurAI/AICOMM, *CLAIRE* and > *RoboCup*. > > > Twitter: @aihuborg > > > > > > > > > Virus-free. *www.avast.com* > > > > > > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: not available URL: From ASIM.ROY at asu.edu Sun Feb 6 15:15:42 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sun, 6 Feb 2022 20:15:42 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <981624DC-3FA6-440E-910F-055B1FB77C41@nyu.edu> References: <981624DC-3FA6-440E-910F-055B1FB77C41@nyu.edu> Message-ID: Dear Gary, I don?t disagree with you. I think it would help research if we define several small, bounded ?understanding? problems rather than this humongous one that we think the human brain handles. May be define ?understanding? in the context of a robot learning to walk. Or define ?understanding? for an artificial mouse that?s simulated. That way, we have some well-defined problems to resolve and they could serve as building blocks for larger problems. In essence, we need to solve simple problems first before we can handle the more complicated ones. Asim From: Gary Marcus Sent: Sunday, February 6, 2022 7:42 AM To: Asim Roy Cc: Ali Minai ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, Sorry for a long answer to your short but rich questions. * Yes, memory in my view has to be part of the answer to the type-token problem. Symbol systems encoded in memory allow a natural way to set up records, and something akin to that seems necessary. Pure multilayer perceptrons struggle with type-token distinctions precisely because they lack such records. On the positive side, I see more and more movement towards recordlike stores (eg w key-value stores in memory networks), and I think that is an important and necessary step, very familiar from the symbol-manipulating playbook, sometimes implemented in new ways. * But ultimately, handling the type-token distinction requires considerable inferential overhead beyond the memory representation of a record per se. How do you determine when to denote something (e.g. Felix) as an instance, and of which kinds (cat, animal etc), and how do you leverage that knowledge once you determine it? * In the limit we reason about types vs tokens in fairly subtle ways, eg in guessing whether a glass that we put down at party is likely to be ours. The reverse is also important: we need to be learn particular traits for individuals and not erroneously generalize them to the class; if my aunt Esther wins the lottery, one shouldn?t infer that all of my aunts or all of my relatives or adult females have won the lottery. so you need both representational machinery that can distinguish eg my cat from cats in general and reasoning machinery to decide at what level certain learned knowledge should inhere. (I had a whole chapter about this sort of thing in The Algebraic Mind if you are interested, and Mike Mozer had a book about types and tokens in neural networks in the mid 1990s). * Yes, part (though not all!) of what we do when we set up cognitive models in our heads is to track particular individuals and their properties. If you only had to correlate kinds (cats) and their properties (have fur) you could maybe get away with a multilayer perceptron, but once you need to track individuals, yes, you really need some kind of memory-based records. * As far as I can tell, Transformers can sometimes approximate some of this for a few sentences, but not over long stretches. As a small terminological aside; for me cognitive models ? cognitive modeling. Cognitive modeling is about building psychological or computational models of how people think, whereas what I mean by a cognitive model is a representation of eg the entities in some situation and the relations between those entities. To your closing question, none of us yet really knows how to build understanding into machines. A solid type-token distinction, both in terms of representation and reasoning, is critical for general intelligence, but hardly sufficient. Personally, I think some minimal prerequisites would be: * representations of space, time, causality, individuals, kinds, persons, places, objects, etc. * representations of abstractions that can hold over all entities in a class * compositionality (if we are talking about human-like understanding) * capacity to construct and update cognitive models on the fly * capacity to reason over entities in those models * ability to learn about new entities and their properties Much of my last book (Rebooting AI, w Ernie Davis) is about the above list. The section in the language chapter on a children?s story in which man has lost is wallet is an especially vivid worked example. Later chapters elaborate some of the challenges in representing space, time, and causality. Gary On Feb 5, 2022, at 18:58, Asim Roy > wrote: ? Gary, I don?t get much into the type of cognitive modeling you are talking about, but I would guess that the type problem can generally be handled by neural network models and tokens can be resolved with some memory-based system. But to the heart of the question, this is what so-called ?understanding? reduces to computation wise? Asim From: Gary Marcus > Sent: Saturday, February 5, 2022 8:39 AM To: Asim Roy > Cc: Ali Minai >; Danko Nikolic >; Brad Wyble >; connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton There is no magic in understanding, just computation that has been realized in the wetware of humans and that eventually can be realized in machines. But understanding is not (just) learning. Understanding incorporates (or works in tandem with) learning - but also, critically, in tandem with inference, and the development and maintenance of cognitive models. Part of developing an understanding of cats in general is to learn long term-knowledge about their properties, both directly (e.g., through observation) and indirectly (eg through learning facts about animals in general that can be extended to cats), often through inference (if all animals have DNA, and a cat is an animal, it must also have DNA). The understanding of a particular cat also involves direct observation, but also inference (eg one might surmise that the reason that Fluffy is running about the room is that Fluffy suspects there is a mouse stirring somewhere nearby). But all of that, I would say, is subservient to the construction of cognitive models that can be routinely updated (e.g., Fluffy is currently in the living room, skittering about, perhaps looking for a mouse). In humans, those dynamic, relational models, which form part of an understanding, can support inference (if Fluffy is in the living room, we can infer that Fluffy is not outside, not lost, etc). Without such models - which I think represent a core part of understanding - AGI is an unlikely prospect. Current neural networks, as it happens, are better at acquiring long-term knowledge (cats have whiskers) than they are at dynamically updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic model that I am describing. To a modest degree they can approximate it on the basis of large samples of texts, but their ultimate incoherence stems from the fact that they do not have robust internal cognitive models that they can update on the fly. Without such cognitive models you can still capture some aspects of understanding (eg predicting that cats are likely to be furry), but things fall apart quickly; inference is never reliable, and coherence is fleeting. As a final note, one of the most foundational challenges in constructing adequate cognitive models of the world is to have a clear distinction between individuals and kinds; as I emphasized 20 years ago (in The Algebraic Mind), this has always been a weakness in neural networks, and I don?t think that the type-token problem has yet been solved. Gary On Feb 5, 2022, at 01:31, Asim Roy > wrote: ? All, I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? https://www.youtube.com/watch?v=gn4nRCC9TwQ https://www.youtube.com/watch?v=8sO7VS3q8d0 Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Ali Minai > Sent: Friday, February 4, 2022 11:38 PM To: Asim Roy > Cc: Gary Marcus >; Danko Nikolic >; Brad Wyble >; connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Asim Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! Ali Ali A. Minai, Ph.D. Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Fri, Feb 4, 2022 at 2:42 AM Asim Roy > wrote: First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. DANKO: What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Gary Marcus > Sent: Thursday, February 3, 2022 9:26 AM To: Danko Nikolic > Cc: Asim Roy >; Geoffrey Hinton >; AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub > wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg [https://ipmcdn.avast.com/images/icons/icon-envelope-tick-round-orange-animated-no-repeat-v1.gif] Virus-free. www.avast.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary at eng.ucsd.edu Sun Feb 6 17:34:56 2022 From: gary at eng.ucsd.edu (gary@ucsd.edu) Date: Sun, 6 Feb 2022 14:34:56 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: "practopoietic theory challenges is the generally accepted idea that the dynamics of a neural network (with its excitatory and inhibitory mechanisms) is sufficient to implement a mind. Practopoiesis tells us that this is not enough." So you are making additional assumptions. Hence Ockham's razor applies... On Sun, Feb 6, 2022 at 2:27 AM Danko Nikolic wrote: > Hi Gary, > > you said: "Please avert your gaze while I apply Ockham?s Razor?" > > I dare you to apply Ockham's razor. Practopoiesis is designed with the > Ockham's razor in mind: To account for as many mental phenomena as possible > by making as few assumptions as possible. > > Danko > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > --- A progress usually starts with an insight --- > > > On Sat, Feb 5, 2022 at 8:05 PM gary at ucsd.edu wrote: > >> Please avert your gaze while I apply Ockham?s Razor? >> >> On Sat, Feb 5, 2022 at 2:12 AM Danko Nikolic >> wrote: >> >>> Gary, you wrote: "What are the alternatives?" >>> >>> There is at least one alternative: the theory of practopoiesis which >>> suggests that it is not the neural networks that "compute" the mental >>> operations. >>> It is instead the quick adaptations of neurons who are responsible for >>> thinking and perceiving. The network only serves the function of bringing >>> in the information and sending it out. >>> >>> The adaptations are suggested to do the central part of the cognition. >>> >>> So far, this is all hypothetical. If we develop these ideas into a >>> working system, this would be an entirely new paradigm. It would be like >>> the third paradigm: >>> 1) manipulation of symbols >>> 2) neural net >>> 3) fast adaptations >>> >>> >>> Danko >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> --- A progress usually starts with an insight --- >>> >>> >>> On Fri, Feb 4, 2022 at 7:19 PM gary at ucsd.edu wrote: >>> >>>> This is an argument from lack of imagination, as Pat Churchland used to >>>> say. All you have to notice, is that your brain is a neural net work. What >>>> are the alternatives? >>>> >>>> On Fri, Feb 4, 2022 at 4:08 AM Danko Nikolic >>>> wrote: >>>> >>>>> >>>>> I suppose everyone agrees that "the brain is a physical system", >>>>> and that "There is no ?magic? inside the brain", >>>>> and that '?understanding? is just part of ?learning.?' >>>>> >>>>> Also, we can agree that some sort of simulation takes place behind >>>>> understanding. >>>>> >>>>> However, there still is a problem: Neural network's can't implement >>>>> the needed simulations; they cannot achieve the same cognitive effect that >>>>> human minds can (or animal minds can). >>>>> >>>>> We don't know a way of wiring a neural network such that it could >>>>> perform the simulations (understandings) necessary to find the answers to >>>>> real-life questions, such as the hamster with a hat problem. >>>>> >>>>> In other words, neural networks, as we know them today, cannot: >>>>> >>>>> 1) learn from a small number of examples (simulation or not) >>>>> 2) apply the knowledge to a wide range of situations >>>>> >>>>> >>>>> We, as scientists, do not understand understanding. Our technology's >>>>> simulations (their depth of understanding) are no match for the simulations >>>>> (depth of understanding) that the biological brain performs. >>>>> >>>>> I think that scientific integrity also covers acknowledging when we >>>>> did not (yet) succeed in solving a certain problem. There is still >>>>> significant work to be done. >>>>> >>>>> >>>>> Danko >>>>> >>>>> Dr. Danko Nikoli? >>>>> www.danko-nikolic.com >>>>> >>>>> https://www.linkedin.com/in/danko-nikolic/ >>>>> >>>>> --- A progress usually starts with an insight --- >>>>> >>>>> >>>>> >>>>> Virenfrei. >>>>> www.avast.com >>>>> >>>>> <#m_3368963304466174950_m_-134745596574091214_m_8423976727351221435_m_-3229424020171779455_m_-1469727422087267219_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> >>>>> >>>>> On Thu, Feb 3, 2022 at 9:35 PM Asim Roy wrote: >>>>> >>>>>> First of all, the brain is a physical system. There is no ?magic? >>>>>> inside the brain that does the ?understanding? part. Take for example >>>>>> learning to play tennis. You hit a few balls - some the right way and some >>>>>> wrong ? but you fairly quickly learn to hit them right most of the time. So >>>>>> there is obviously some simulation going on in the brain about hitting the >>>>>> ball in different ways and ?learning? its consequences. What you are >>>>>> calling ?understanding? is really these simulations about different >>>>>> scenarios. It?s also very similar to augmentation used to train image >>>>>> recognition systems where you rotate images, obscure parts and so on, so >>>>>> that you still can say it?s a cat even though you see only the cat?s face >>>>>> or whiskers or a cat flipped on its back. So, if the following questions >>>>>> relate to ?understanding,? you can easily resolve this by simulating such >>>>>> scenarios when ?teaching? the system. There?s nothing ?magical? about >>>>>> ?understanding.? As I said, bear in mind that the brain, after all, is a >>>>>> physical system and ?teaching? and ?understanding? is embodied in that >>>>>> physical system, not outside it. So ?understanding? is just part of >>>>>> ?learning,? nothing more. >>>>>> >>>>>> >>>>>> >>>>>> DANKO: >>>>>> >>>>>> What would happen to the hat if the hamster rolls on its back? (Would >>>>>> the hat fall off?) >>>>>> >>>>>> What would happen to the red hat when the hamster enters its lair? >>>>>> (Would the hat fall off?) >>>>>> >>>>>> What would happen to that hamster when it goes foraging? (Would the >>>>>> red hat have an influence on finding food?) >>>>>> >>>>>> What would happen in a situation of being chased by a predator? >>>>>> (Would it be easier for predators to spot the hamster?) >>>>>> >>>>>> >>>>>> >>>>>> Asim Roy >>>>>> >>>>>> Professor, Information Systems >>>>>> >>>>>> Arizona State University >>>>>> >>>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>>> >>>>>> >>>>>> Asim Roy | iSearch (asu.edu) >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> *From:* Gary Marcus >>>>>> *Sent:* Thursday, February 3, 2022 9:26 AM >>>>>> *To:* Danko Nikolic >>>>>> *Cc:* Asim Roy ; Geoffrey Hinton < >>>>>> geoffrey.hinton at gmail.com>; AIhub ; >>>>>> connectionists at mailman.srv.cs.cmu.edu >>>>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>>>> Geoff Hinton >>>>>> >>>>>> >>>>>> >>>>>> Dear Danko, >>>>>> >>>>>> >>>>>> >>>>>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >>>>>> talk, in which he said (paraphrasing from memory, because I don?t remember >>>>>> the precise words) that the famous 200 Quoc Le unsupervised model [ >>>>>> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >>>>>> ] >>>>>> had learned the concept of a ca. In reality the model had clustered >>>>>> together some catlike images based on the image statistics that it had >>>>>> extracted, but it was a long way from a full, counterfactual-supporting >>>>>> concept of a cat, much as you describe below. >>>>>> >>>>>> >>>>>> >>>>>> I fully agree with you that the reason for even having a semantics is >>>>>> as you put it, "to 1) learn with a few examples and 2) apply the knowledge >>>>>> to a broad set of situations.? GPT-3 sometimes gives the appearance of >>>>>> having done so, but it falls apart under close inspection, so the problem >>>>>> remains unsolved. >>>>>> >>>>>> >>>>>> >>>>>> Gary >>>>>> >>>>>> >>>>>> >>>>>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >>>>>> wrote: >>>>>> >>>>>> >>>>>> >>>>>> G. Hinton wrote: "I believe that any reasonable person would admit >>>>>> that if you ask a neural net to draw a picture of a hamster wearing a red >>>>>> hat and it draws such a picture, it understood the request." >>>>>> >>>>>> >>>>>> >>>>>> I would like to suggest why drawing a hamster with a red hat does not >>>>>> necessarily imply understanding of the statement "hamster wearing a red >>>>>> hat". >>>>>> >>>>>> To understand that "hamster wearing a red hat" would mean inferring, >>>>>> in newly emerging situations of this hamster, all the real-life >>>>>> implications that the red hat brings to the little animal. >>>>>> >>>>>> >>>>>> >>>>>> What would happen to the hat if the hamster rolls on its back? (Would >>>>>> the hat fall off?) >>>>>> >>>>>> What would happen to the red hat when the hamster enters its lair? >>>>>> (Would the hat fall off?) >>>>>> >>>>>> What would happen to that hamster when it goes foraging? (Would the >>>>>> red hat have an influence on finding food?) >>>>>> >>>>>> What would happen in a situation of being chased by a predator? >>>>>> (Would it be easier for predators to spot the hamster?) >>>>>> >>>>>> >>>>>> >>>>>> ...and so on. >>>>>> >>>>>> >>>>>> >>>>>> Countless many questions can be asked. One has understood "hamster >>>>>> wearing a red hat" only if one can answer reasonably well many of such >>>>>> real-life relevant questions. Similarly, a student has understood materias >>>>>> in a class only if they can apply the materials in real-life situations >>>>>> (e.g., applying Pythagora's theorem). If a student gives a correct answer >>>>>> to a multiple choice question, we don't know whether the student understood >>>>>> the material or whether this was just rote learning (often, it is rote >>>>>> learning). >>>>>> >>>>>> >>>>>> >>>>>> I also suggest that understanding also comes together with effective >>>>>> learning: We store new information in such a way that we can recall it >>>>>> later and use it effectively i.e., make good inferences in newly emerging >>>>>> situations based on this knowledge. >>>>>> >>>>>> >>>>>> >>>>>> In short: Understanding makes us humans able to 1) learn with a few >>>>>> examples and 2) apply the knowledge to a broad set of situations. >>>>>> >>>>>> >>>>>> >>>>>> No neural network today has such capabilities and we don't know how >>>>>> to give them such capabilities. Neural networks need large amounts of >>>>>> training examples that cover a large variety of situations and then >>>>>> the networks can only deal with what the training examples have already >>>>>> covered. Neural networks cannot extrapolate in that 'understanding' sense. >>>>>> >>>>>> >>>>>> >>>>>> I suggest that understanding truly extrapolates from a piece of >>>>>> knowledge. It is not about satisfying a task such as translation between >>>>>> languages or drawing hamsters with hats. It is how you got the capability >>>>>> to complete the task: Did you only have a few examples that covered >>>>>> something different but related and then you extrapolated from that >>>>>> knowledge? If yes, this is going in the direction of understanding. Have >>>>>> you seen countless examples and then interpolated among them? Then perhaps >>>>>> it is not understanding. >>>>>> >>>>>> >>>>>> >>>>>> So, for the case of drawing a hamster wearing a red hat, >>>>>> understanding perhaps would have taken place if the following happened >>>>>> before that: >>>>>> >>>>>> >>>>>> >>>>>> 1) first, the network learned about hamsters (not many examples) >>>>>> >>>>>> 2) after that the network learned about red hats (outside the context >>>>>> of hamsters and without many examples) >>>>>> >>>>>> 3) finally the network learned about drawing (outside of the context >>>>>> of hats and hamsters, not many examples) >>>>>> >>>>>> >>>>>> >>>>>> After that, the network is asked to draw a hamster with a red hat. If >>>>>> it does it successfully, maybe we have started cracking the problem of >>>>>> understanding. >>>>>> >>>>>> >>>>>> >>>>>> Note also that this requires the network to learn sequentially >>>>>> without exhibiting catastrophic forgetting of the previous knowledge, which >>>>>> is possibly also a consequence of human learning by understanding. >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> Danko >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> Dr. Danko Nikoli? >>>>>> www.danko-nikolic.com >>>>>> >>>>>> https://www.linkedin.com/in/danko-nikolic/ >>>>>> >>>>>> >>>>>> --- A progress usually starts with an insight --- >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> Virus-free. www.avast.com >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>>>>> >>>>>> Without getting into the specific dispute between Gary and Geoff, I >>>>>> think with approaches similar to GLOM, we are finally headed in the right >>>>>> direction. There?s plenty of neurophysiological evidence for single-cell >>>>>> abstractions and multisensory neurons in the brain, which one might claim >>>>>> correspond to symbols. And I think we can finally reconcile the decades old >>>>>> dispute between Symbolic AI and Connectionism. >>>>>> >>>>>> >>>>>> >>>>>> GARY: (Your GLOM, which as you know I praised publicly, is in many >>>>>> ways an effort to wind up with encodings that effectively serve as symbols >>>>>> in exactly that way, guaranteed to serve as consistent representations of >>>>>> specific concepts.) >>>>>> >>>>>> GARY: I have *never* called for dismissal of neural networks, but >>>>>> rather for some hybrid between the two (as you yourself contemplated in >>>>>> 1991); the point of the 2001 book was to characterize exactly where >>>>>> multilayer perceptrons succeeded and broke down, and where symbols could >>>>>> complement them. >>>>>> >>>>>> >>>>>> >>>>>> Asim Roy >>>>>> >>>>>> Professor, Information Systems >>>>>> >>>>>> Arizona State University >>>>>> >>>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>>> >>>>>> >>>>>> Asim Roy | iSearch (asu.edu) >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> *From:* Connectionists >>>>>> *On Behalf Of *Gary Marcus >>>>>> *Sent:* Wednesday, February 2, 2022 1:26 PM >>>>>> *To:* Geoffrey Hinton >>>>>> *Cc:* AIhub ; >>>>>> connectionists at mailman.srv.cs.cmu.edu >>>>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>>>> Geoff Hinton >>>>>> >>>>>> >>>>>> >>>>>> Dear Geoff, and interested others, >>>>>> >>>>>> >>>>>> >>>>>> What, for example, would you make of a system that often drew the >>>>>> red-hatted hamster you requested, and perhaps a fifth of the time gave you >>>>>> utter nonsense? Or say one that you trained to create birds but sometimes >>>>>> output stuff like this: >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> One could >>>>>> >>>>>> >>>>>> >>>>>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>>>>> >>>>>> or >>>>>> >>>>>> b. wonder if it might be possible that sometimes the system gets the >>>>>> right answer for the wrong reasons (eg partial historical contingency), and >>>>>> wonder whether another approach might be indicated. >>>>>> >>>>>> >>>>>> >>>>>> Benchmarks are harder than they look; most of the field has come to >>>>>> recognize that. The Turing Test has turned out to be a lousy measure of >>>>>> intelligence, easily gamed. It has turned out empirically that the Winograd >>>>>> Schema Challenge did not measure common sense as well as Hector might have >>>>>> thought. (As it happens, I am a minor coauthor of a very recent review on >>>>>> this very topic: https://arxiv.org/abs/2201.02387 >>>>>> ) >>>>>> But its conquest in no way means machines now have common sense; many >>>>>> people from many different perspectives recognize that (including, e.g., >>>>>> Yann LeCun, who generally tends to be more aligned with you than with me). >>>>>> >>>>>> >>>>>> >>>>>> So: on the goalpost of the Winograd schema, I was wrong, and you can >>>>>> quote me; but what you said about me and machine translation remains your >>>>>> invention, and it is inexcusable that you simply ignored my 2019 >>>>>> clarification. On the essential goal of trying to reach meaning and >>>>>> understanding, I remain unmoved; the problem remains unsolved. >>>>>> >>>>>> >>>>>> >>>>>> All of the problems LLMs have with coherence, reliability, >>>>>> truthfulness, misinformation, etc stand witness to that fact. (Their >>>>>> persistent inability to filter out toxic and insulting remarks stems from >>>>>> the same.) I am hardly the only person in the field to see that progress on >>>>>> any given benchmark does not inherently mean that the deep underlying >>>>>> problems have solved. You, yourself, in fact, have occasionally made that >>>>>> point. >>>>>> >>>>>> >>>>>> >>>>>> With respect to embeddings: Embeddings are very good for natural >>>>>> language *processing*; but NLP is not the same as NL*U* ? when it >>>>>> comes to *understanding*, their worth is still an open question. >>>>>> Perhaps they will turn out to be necessary; they clearly aren?t sufficient. >>>>>> In their extreme, they might even collapse into being symbols, in the sense >>>>>> of uniquely identifiable encodings, akin to the ASCII code, in which a >>>>>> specific set of numbers stands for a specific word or concept. (Wouldn?t >>>>>> that be ironic?) >>>>>> >>>>>> >>>>>> >>>>>> (Your GLOM, which as you know I praised publicly, is in many ways an >>>>>> effort to wind up with encodings that effectively serve as symbols in >>>>>> exactly that way, guaranteed to serve as consistent representations of >>>>>> specific concepts.) >>>>>> >>>>>> >>>>>> >>>>>> Notably absent from your email is any kind of apology for >>>>>> misrepresenting my position. It?s fine to say that ?many people thirty >>>>>> years ago once thought X? and another to say ?Gary Marcus said X in 2015?, >>>>>> when I didn?t. I have consistently felt throughout our interactions that >>>>>> you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) >>>>>> apologized to me for having made that error. I am still not he. >>>>>> >>>>>> >>>>>> >>>>>> Which maybe connects to the last point; if you read my work, you >>>>>> would see thirty years of arguments *for* neural networks, just not >>>>>> in the way that you want them to exist. I have ALWAYS argued that there is >>>>>> a role for them; characterizing me as a person ?strongly opposed to neural >>>>>> networks? misses the whole point of my 2001 book, which was subtitled >>>>>> ?Integrating Connectionism and Cognitive Science.? >>>>>> >>>>>> >>>>>> >>>>>> In the last two decades or so you have insisted (for reasons you have >>>>>> never fully clarified, so far as I know) on abandoning symbol-manipulation, >>>>>> but the reverse is not the case: I have *never* called for dismissal >>>>>> of neural networks, but rather for some hybrid between the two (as you >>>>>> yourself contemplated in 1991); the point of the 2001 book was to >>>>>> characterize exactly where multilayer perceptrons succeeded and broke down, >>>>>> and where symbols could complement them. It?s a rhetorical trick (which is >>>>>> what the previous thread was about) to pretend otherwise. >>>>>> >>>>>> >>>>>> >>>>>> Gary >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> On Feb 2, 2022, at 11:22, Geoffrey Hinton >>>>>> wrote: >>>>>> >>>>>> ? >>>>>> >>>>>> Embeddings are just vectors of soft feature detectors and they are >>>>>> very good for NLP. The quote on my webpage from Gary's 2015 chapter >>>>>> implies the opposite. >>>>>> >>>>>> >>>>>> >>>>>> A few decades ago, everyone I knew then would have agreed that the >>>>>> ability to translate a sentence into many different languages was strong >>>>>> evidence that you understood it. >>>>>> >>>>>> >>>>>> >>>>>> But once neural networks could do that, their critics moved the >>>>>> goalposts. An exception is Hector Levesque who defined the goalposts more >>>>>> sharply by saying that the ability to get pronoun references correct in >>>>>> Winograd sentences is a crucial test. Neural nets are improving at that but >>>>>> still have some way to go. Will Gary agree that when they can get pronoun >>>>>> references correct in Winograd sentences they really do understand? Or does >>>>>> he want to reserve the right to weasel out of that too? >>>>>> >>>>>> >>>>>> >>>>>> Some people, like Gary, appear to be strongly opposed to neural >>>>>> networks because they do not fit their preconceived notions of how the mind >>>>>> should work. >>>>>> >>>>>> I believe that any reasonable person would admit that if you ask a >>>>>> neural net to draw a picture of a hamster wearing a red hat and it draws >>>>>> such a picture, it understood the request. >>>>>> >>>>>> >>>>>> >>>>>> Geoff >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus >>>>>> wrote: >>>>>> >>>>>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger >>>>>> neural network community, >>>>>> >>>>>> >>>>>> >>>>>> There has been a lot of recent discussion on this list about framing >>>>>> and scientific integrity. Often the first step in restructuring narratives >>>>>> is to bully and dehumanize critics. The second is to misrepresent their >>>>>> position. People in positions of power are sometimes tempted to do this. >>>>>> >>>>>> >>>>>> >>>>>> The Hinton-Hanson interview that you just published is a real-time >>>>>> example of just that. It opens with a needless and largely content-free >>>>>> personal attack on a single scholar (me), with the explicit intention of >>>>>> discrediting that person. Worse, the only substantive thing it says is >>>>>> false. >>>>>> >>>>>> >>>>>> >>>>>> Hinton says ?In 2015 he [Marcus] made a prediction that computers >>>>>> wouldn?t be able to do machine translation.? >>>>>> >>>>>> >>>>>> >>>>>> I never said any such thing. >>>>>> >>>>>> >>>>>> >>>>>> What I predicted, rather, was that multilayer perceptrons, as they >>>>>> existed then, would not (on their own, absent other mechanisms) >>>>>> *understand* language. Seven years later, they still haven?t, except >>>>>> in the most superficial way. >>>>>> >>>>>> >>>>>> >>>>>> I made no comment whatsoever about machine translation, which I view >>>>>> as a separate problem, solvable to a certain degree by correspondance >>>>>> without semantics. >>>>>> >>>>>> >>>>>> >>>>>> I specifically tried to clarify Hinton?s confusion in 2019, but, >>>>>> disappointingly, he has continued to purvey misinformation despite that >>>>>> clarification. Here is what I wrote privately to him then, which should >>>>>> have put the matter to rest: >>>>>> >>>>>> >>>>>> >>>>>> You have taken a single out of context quote [from 2015] and >>>>>> misrepresented it. The quote, which you have prominently displayed at the >>>>>> bottom on your own web page, says: >>>>>> >>>>>> >>>>>> >>>>>> Hierarchies of features are less suited to challenges such as >>>>>> language, inference, and high-level planning. For example, as Noam Chomsky >>>>>> famously pointed out, language is filled with sentences you haven't seen >>>>>> before. Pure classifier systems don't know what to do with such sentences. >>>>>> The talent of feature detectors -- in identifying which member of some >>>>>> category something belongs to -- doesn't translate into understanding >>>>>> novel sentences, in which each sentence has its own unique meaning. >>>>>> >>>>>> >>>>>> >>>>>> It does *not* say "neural nets would not be able to deal with novel >>>>>> sentences"; it says that hierachies of features detectors (on their own, if >>>>>> you read the context of the essay) would have trouble >>>>>> *understanding *novel sentences. >>>>>> >>>>>> >>>>>> >>>>>> Google Translate does yet not *understand* the content of the >>>>>> sentences is translates. It cannot reliably answer questions about who did >>>>>> what to whom, or why, it cannot infer the order of the events in >>>>>> paragraphs, it can't determine the internal consistency of those events, >>>>>> and so forth. >>>>>> >>>>>> >>>>>> >>>>>> Since then, a number of scholars, such as the the computational >>>>>> linguist Emily Bender, have made similar points, and indeed current LLM >>>>>> difficulties with misinformation, incoherence and fabrication all follow >>>>>> from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on >>>>>> the matter with Alexander Koller, >>>>>> https://aclanthology.org/2020.acl-main.463.pdf >>>>>> , >>>>>> also emphasizing issues of understanding and meaning: >>>>>> >>>>>> >>>>>> >>>>>> *The success of the large neural language models on many NLP tasks is >>>>>> exciting. However, we find that these successes sometimes lead to hype in >>>>>> which these models are being described as ?understanding? language or >>>>>> capturing ?meaning?. In this position paper, we argue that a system trained >>>>>> only on form has a priori no way to learn meaning. .. a clear understanding >>>>>> of the distinction between form and meaning will help guide the field >>>>>> towards better science around natural language understanding. * >>>>>> >>>>>> >>>>>> >>>>>> Her later article with Gebru on language models ?stochastic parrots? >>>>>> is in some ways an extension of this point; machine translation requires >>>>>> mimicry, true understanding (which is what I was discussing in 2015) >>>>>> requires something deeper than that. >>>>>> >>>>>> >>>>>> >>>>>> Hinton?s intellectual error here is in equating machine translation >>>>>> with the deeper comprehension that robust natural language understanding >>>>>> will require; as Bender and Koller observed, the two appear not to be the >>>>>> same. (There is a longer discussion of the relation between language >>>>>> understanding and machine translation, and why the latter has turned out to >>>>>> be more approachable than the former, in my 2019 book with Ernest Davis). >>>>>> >>>>>> >>>>>> >>>>>> More broadly, Hinton?s ongoing dismissiveness of research from >>>>>> perspectives other than his own (e.g. linguistics) have done the field a >>>>>> disservice. >>>>>> >>>>>> >>>>>> >>>>>> As Herb Simon once observed, science does not have to be zero-sum. >>>>>> >>>>>> >>>>>> >>>>>> Sincerely, >>>>>> >>>>>> Gary Marcus >>>>>> >>>>>> Professor Emeritus >>>>>> >>>>>> New York University >>>>>> >>>>>> >>>>>> >>>>>> On Feb 2, 2022, at 06:12, AIhub wrote: >>>>>> >>>>>> ? >>>>>> >>>>>> Stephen Hanson in conversation with Geoff Hinton >>>>>> >>>>>> >>>>>> >>>>>> In the latest episode of this video series for AIhub.org >>>>>> , >>>>>> Stephen Hanson talks to Geoff Hinton about neural networks, >>>>>> backpropagation, overparameterization, digit recognition, voxel cells, >>>>>> syntax and semantics, Winograd sentences, and more. >>>>>> >>>>>> >>>>>> >>>>>> You can watch the discussion, and read the transcript, here: >>>>>> >>>>>> >>>>>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> About AIhub: >>>>>> >>>>>> AIhub is a non-profit dedicated to connecting the AI community to the >>>>>> public by providing free, high-quality information through AIhub.org >>>>>> >>>>>> (https://aihub.org/ >>>>>> ). >>>>>> We help researchers publish the latest AI news, summaries of their work, >>>>>> opinion pieces, tutorials and more. We are supported by many leading >>>>>> scientific organizations in AI, namely AAAI >>>>>> , >>>>>> NeurIPS >>>>>> , >>>>>> ICML >>>>>> , >>>>>> AIJ >>>>>> >>>>>> /IJCAI >>>>>> , >>>>>> ACM SIGAI >>>>>> , >>>>>> EurAI/AICOMM, CLAIRE >>>>>> >>>>>> and RoboCup >>>>>> >>>>>> . >>>>>> >>>>>> Twitter: @aihuborg >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> Virus-free. www.avast.com >>>>>> >>>>>> >>>>>> >>>>>> >>>>> -- >>>> Gary Cottrell 858-534-6640 FAX: 858-534-7029 >>>> Computer Science and Engineering 0404 >>>> IF USING FEDEX INCLUDE THE FOLLOWING LINE: >>>> CSE Building, Room 4130 >>>> University of California San Diego >>>> - >>>> 9500 Gilman Drive # 0404 >>>> >>>> La Jolla, Ca. 92093-0404 >>>> >>>> >>>> Email: gary at ucsd.edu >>>> Home page: http://www-cse.ucsd.edu/~gary/ >>>> Schedule: http://tinyurl.com/b7gxpwo >>>> >>>> >>>> Blind certainty - a close-mindedness that amounts to an imprisonment so >>>> total, that the prisoner doesn?t even know that he?s locked up. -David >>>> Foster Wallace >>>> >>>> >>>> Power to the people! ?Patti Smith >>>> >>>> Except when they?re delusional ?Gary Cottrell >>>> >>>> >>>> This song makes me nostalgic for a memory I don't have -- Tess Cottrell >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> *Listen carefully,Neither the VedasNor the Qur'anWill teach you >>>> this:Put the bit in its mouth,The saddle on its back,Your foot in the >>>> stirrup,And ride your wild runaway mindAll the way to heaven.* >>>> >>>> -- Kabir >>>> >>> -- >> Gary Cottrell 858-534-6640 FAX: 858-534-7029 >> Computer Science and Engineering 0404 >> IF USING FEDEX INCLUDE THE FOLLOWING LINE: >> CSE Building, Room 4130 >> University of California San Diego - >> 9500 Gilman Drive # 0404 >> La Jolla, Ca. 92093-0404 >> >> Email: gary at ucsd.edu >> Home page: http://www-cse.ucsd.edu/~gary/ >> Schedule: http://tinyurl.com/b7gxpwo >> >> >> Blind certainty - a close-mindedness that amounts to an imprisonment so >> total, that the prisoner doesn?t even know that he?s locked up. -David >> Foster Wallace >> >> >> Power to the people! ?Patti Smith >> >> Except when they?re delusional ?Gary Cottrell >> >> >> This song makes me nostalgic for a memory I don't have -- Tess Cottrell >> >> >> >> >> >> >> >> >> >> >> *Listen carefully,Neither the VedasNor the Qur'anWill teach you this:Put >> the bit in its mouth,The saddle on its back,Your foot in the stirrup,And >> ride your wild runaway mindAll the way to heaven.* >> >> -- Kabir >> > -- Gary Cottrell 858-534-6640 FAX: 858-534-7029 Computer Science and Engineering 0404 IF USING FEDEX INCLUDE THE FOLLOWING LINE: CSE Building, Room 4130 University of California San Diego - 9500 Gilman Drive # 0404 La Jolla, Ca. 92093-0404 Email: gary at ucsd.edu Home page: http://www-cse.ucsd.edu/~gary/ Schedule: http://tinyurl.com/b7gxpwo Blind certainty - a close-mindedness that amounts to an imprisonment so total, that the prisoner doesn?t even know that he?s locked up. -David Foster Wallace Power to the people! ?Patti Smith Except when they?re delusional ?Gary Cottrell This song makes me nostalgic for a memory I don't have -- Tess Cottrell *Listen carefully,Neither the VedasNor the Qur'anWill teach you this:Put the bit in its mouth,The saddle on its back,Your foot in the stirrup,And ride your wild runaway mindAll the way to heaven.* -- Kabir -------------- next part -------------- An HTML attachment was scrubbed... URL: From juergen at idsia.ch Sun Feb 6 03:44:02 2022 From: juergen at idsia.ch (Schmidhuber Juergen) Date: Sun, 6 Feb 2022 08:44:02 +0000 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: <7f97db1f-13e1-ba48-8b02-f3a2c4769df9@rubic.rutgers.edu> References: <2f1d9928-543f-f4a0-feab-5a5a0cc1d4d7@rubic.rutgers.edu> <307D9939-4F3A-40FF-A19F-3CEABEAE315C@supsi.ch> <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> <7f97db1f-13e1-ba48-8b02-f3a2c4769df9@rubic.rutgers.edu> Message-ID: <6E8002C5-64E7-41DB-930F-B77BF78F600A@supsi.ch> Steve, it?s simple: the original ?shallow learning? (~1800) is much older than your relatively recent ?shallow learning? references (mostly from the 1900s). No need to mention all of them in this report, which is really about "deep learning? (see title) with adaptive hidden units, which started to work in the 1960s, first through layer by layer training (USSR, 1965), then through stochastic gradient descent (SGD) in relatively deep nets (Japan, 1967). The reverse mode of automatic differentiation (now called backpropagation) appeared 3 years later (Finland, 1970). No fancy talk about syntax vs semantics can justify a revisionist history of deep learning that does not mention these achievements. Cheers, J?rgen https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html On 2 Feb 2022, at 00:48, Stephen Jos? Hanson > wrote: Jeurgen: Even some of us lowly psychologists know some math. And its not about the math.. its about the context (is this sounding like an echo?) Let me try again.. and I think your good intentions but misguided reconstruction of history is appearing to me, to be perverse. You tip your hand when you talk about "rebranding". Also that the PDP books were a "conspiracy". But lets go point by point. (1) we already agreed that the Perceptron was not linear regression--lets not go backwards. Closer to logistic regression. If you are talking about Widrow and Hoff, well it is the Delta Rule-- SSE kind of regression. But where did the Delta rule come from? Lets look at Math. So there is some nice papers by Gluck and Thompson (80s) showing how Pavlovian conditioning is exactly the Delta rule and even more relevant was shown to account for majority of classical (pavlovian) conditioning was the Rescorla-Wagner (1972) model-- \Delta V_A = [\alpha_A\beta_1](\lambda_1 - V_{AX}), which of course was Ivan Petrovich Pavlov discovery of classical conditioning (1880s). Why aren't you citing him? What about John Brodeus Watson and Burris Fredrick Skinner? At least they were focused on learning albeit *just* in biological systems. But these were actual natural world discoveries. (2) Function approximation. Ok Juergen, claims that everything is really just X, reminds me of the man with a Hammer to whom everything looks like a nail! To the point: its incidental. Yes, Neural networks are function approximators, but that is incidental to the original more general context (PDP) as a way to create "internal representations". The function approximation was a Bonus! (3) Branding. OMG. So you seem to believe that everyone is cynical and will put their intellectual finger in the air to find out what to call what they are doing! Jeez, I hope this isn't true. But the narrative you eschew is in fact something that Minsky would talk about (I remember this at lunch with him in the 90s at Thinking Machines), and he was quite clear that Perceptron was failing well before the 1969 book (trying to do speech recognition with a perceptron--yikes), but in a piling on kind of way Perceptrons killed the perceptron, but it was the linearity focus (as BAP points out) and the lack of depth. (4) Group Method of Handling Data. Frankly, the only one I can find that branded GMHD as a NeuroNet (as they call it) was you. There is a 2017 reference, but they reference you again. (5) Its just names, fashion and preference.. or no actual concepts matter. Really? There was an french mathematician named Fourier in the 19th century who came up with an idea of periodic function decomposition into weighted trigonometric functions.. but he had no math. And Laplace Legendre and others said he had no math! So they prevented him from publishing for FIFTEEN YEARS.. 150 years later after Tukey invented the FFT, its the most common transform used and misused in general. Concepts lead to math.. and that may lead to further formalism.. but don't mistake the math for the concept behind it. The context matters and you are confusing syntax for semantics! Cheers, Steve On 1/31/22 11:38 AM, Schmidhuber Juergen wrote: Steve, do you really want to erase the very origins of shallow learning (Gauss & Legendre ~1800) and deep learning (DL, Ivakhnenko & Lapa 1965) from the field's history? Why? Because they did not use modern terminology such as "artificial neural nets (NNs)" and "learning internal representations"? Names change all the time like fashions; the only thing that counts is the math. Not only mathematicians but also psychologists like yourself will agree. Again: the linear regressor of Legendre & Gauss is formally identical to what was much later called a linear NN for function approximation (FA), minimizing mean squared error, still widely used today. No history of "shallow learning" (without adaptive hidden layers) is complete without this original shallow learner of 2 centuries ago. Many NN courses actually introduce simple NNs in this mathematically and historically correct way, then proceed to DL NNs with several adaptive hidden layers. And of course, no DL history is complete without the origins of functional DL in 1965 [DEEP1-2]. Back then, Ivakhnenko and Lapa published the first general, working DL algorithm for supervised deep feedforward multilayer perceptrons (MLPs) with arbitrarily many layers of neuron-like elements, using nonlinear activation functions (actually Kolmogorov-Gabor polynomials) that combine both additions (like in linear NNs) and multiplications (basically they had deep NNs with gates, including higher order gates). They incrementally trained and pruned their DL networks layer by layer to learn internal representations, using regression and a separate validation set (network depth > 7 by 1971). They had standard justifications of DL such as: "a multilayered structure is a computationally feasible way to implement multinomials of very high degree" [DEEP2] (that cannot be approximated by simple linear NNs). Of course, their DL was automated, and many people have used it up to the 2000s ! - just follow the numerous citations. I don't get your comments about Ivakhnenko's DL and function approximation (FA). FA is for all kinds of functions, including your "cognitive or perceptual or motor functions." NNs are used as FAs all the time. Like other NNs, Ivakhnenko's nets can be used as FAs for your motor control problems. You boldly claim: "This was not in the intellectual space" of Ivakhnenko's method. But obviously it was. Interestingly, 2 years later, Amari (1967-68) [GD1-2] trained his deep MLPs through a different DL method, namely, stochastic gradient descent (1951-52)[STO51-52]. His paper also did not contain the "modern" expression "learning internal representations in NNs." But that's what it was about. Math and algorithms are immune to rebranding. You may not like the fact that neither the original shallow learning (Gauss & Legendre ~1800) nor the original working DL (Ivakhnenko & Lapa 1965; Amari 1967) were biologically inspired. They were motivated through math and problem solving. The NN rebranding came later. Proper scientific credit assignment does not care for changes in terminology. BTW, unfortunately, Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions. But actually he had much more interesting MLPs with a non-learning randomized first layer and an adaptive output layer. So Rosenblatt basically had what much later was rebranded as "Extreme Learning Machines (ELMs)." The revisionist narrative of ELMs (see this web site https://elmorigin.wixsite.com/originofelm) is a bit like the revisionist narrative of DL criticized by my report. Some ELM guys apparently thought they can get away with blatant improper credit assignment. After all, the criticized DL guys seemed to get away with it on an even grander scale. They called themselves the "DL conspiracy" [DLC]; the "ELM conspiracy" is similar. What an embarrassing lack of maturity of our field. Fortunately, more and more ML researchers are helping to set things straight. "In science, by definition, the facts will always win in the end. As long as the facts have not yet won it's not yet the end." [T21v1] References as always under https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html J?rgen On 27 Jan 2022, at 17:37, Stephen Jos? Hanson wrote: Juergen, I have read through GMHD paper and a 1971 Review paper by Ivakhnenko. These are papers about function approximation. The method proposes to use series of polynomial functions that are stacked in filtered sets. The filtered sets are chosen based on best fit, and from what I can tell are manually grown.. so this must of been a tedious and slow process (I assume could be automated). So are the GMHDs "deep", in that they are stacked 4 deep in figure 1 (8 deep in another). Interestingly, they are using (with obvious FA justification) polynomials of various degree. Has this much to do with neural networks? Yes, there were examples initiated by Rumelhart (and me: https://www.routledge.com/Backpropagation-Theory-Architectures-and-Applications/Chauvin-Rumelhart/p/book/9780805812596), based on poly-synaptic dendrite complexity, but not in the GMHD paper.. which was specifically about function approximation. Ivakhnenko, lists four reasons for the approach t! hey took: mainly reducing data size and being more efficient with data that one had. No mention of "internal representations" So when Terry, talks about "internal representations" --does he mean function approximation? Not so much. That of course is part of this, but the actual focus is on cognitive or perceptual or motor functions. Representation in the brain. Hidden units (which could be polynomials) cluster and project and model the input features wrt to the function constraints conditioned by training data. This is more similar to model specification through function space search. And the original Rumelhart meaning of internal representation in PDP vol 1, was in the case of representation certain binary functions (XOR), but more generally about the need for "neurons" (inter-neurons) explicitly between input (sensory) and output (motor). Consider NETTALK, in which I did the first hierarchical clustering of the hidden units over the input features (letters). What appeared wasn't probably surprising.. but without model specification, the network (w.hidden units), learned VOWELS and ! CONSONANT distinctions just from training (Hanson & Burr, 1990). This would be a clear example of "internal representations" in the sense of Rumelhart. This was not in the intellectual space of Ivakhnenko's Group Method of Handling Data. (some of this is discussed in more detail in some recent conversations with Terry Sejnowski and another one to appear shortly with Geoff Hinton (AIHUB.org look in Opinions). Now I suppose one could be cynical and opportunistic, and even conclude if you wanted to get more clicks, rather than title your article GROUP METHOD OF HANDLING DATA, you should at least consider: NEURAL NETWORKS FOR HANDLING DATA, even if you didn't think neural networks had anything to do with your algorithm, after all everyone else is! Might get it published in this time frame, or even read. This is not scholarship. These publications threads are related but not dependent. And although they diverge they could be informative if one were to try and develop polynomial inductive growth networks (see Falhman, 1989; Cascade correlation and Hanson 1990: Meiosis nets) to motor control in the brain. But that's not what happened. I think, like Gauss, you need to drop this specific claim as well. With best regards, Steve On 25 Jan 2022, at 20:03, Schmidhuber Juergen wrote: PS: Terry, you also wrote: "Our precious time is better spent moving the field forward.? However, it seems like in recent years much of your own precious time has gone to promulgating a revisionist history of deep learning (and writing the corresponding "amicus curiae" letters to award committees). For a recent example, your 2020 deep learning survey in PNAS [S20] claims that your 1985 Boltzmann machine [BM] was the first NN to learn internal representations. This paper [BM] neither cited the internal representations learnt by Ivakhnenko & Lapa's deep nets in 1965 [DEEP1-2] nor those learnt by Amari?s stochastic gradient descent for MLPs in 1967-1968 [GD1-2]. Nor did your recent survey [S20] attempt to correct this as good science should strive to do. On the other hand, it seems you celebrated your co-author's birthday in a special session while you were head of NeurIPS, instead of correcting these inaccuracies and celebrating the true pioneers of deep learning, such as ! Ivakhnenko and Amari. Even your recent interview https://blog.paperspace.com/terry-sejnowski-boltzmann-machines/ claims: "Our goal was to try to take a network with multiple layers - an input layer, an output layer and layers in between ? and make it learn. It was generally thought, because of early work that was done in AI in the 60s, that no one would ever find such a learning algorithm because it was just too mathematically difficult.? You wrote this although you knew exactly that such learning algorithms were first created in the 1960s, and that they worked. You are a well-known scientist, head of NeurIPS, and chief editor of a major journal. You must correct this. We must all be better than this as scientists. We owe it to both the past, present, and future scientists as well as those we ultimately serve. The last paragraph of my report https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html quotes Elvis Presley: "Truth is like the sun. You can shut it out for a time, but it ain't goin' away.? I wonder how the future will reflect on the choices we make now. J?rgen On 3 Jan 2022, at 11:38, Schmidhuber Juergen wrote: Terry, please don't throw smoke candles like that! This is not about basic math such as Calculus (actually first published by Leibniz; later Newton was also credited for his unpublished work; Archimedes already had special cases thereof over 2000 years ago; the Indian Kerala school made essential contributions around 1400). In fact, my report addresses such smoke candles in Sec. XII: "Some claim that 'backpropagation' is just the chain rule of Leibniz (1676) & L'Hopital (1696).' No, it is the efficient way of applying the chain rule to big networks with differentiable nodes (there are also many inefficient ways of doing this). It was not published until 1970 [BP1]." You write: "All these threads will be sorted out by historians one hundred years from now." To answer that, let me just cut and paste the last sentence of my conclusions: "However, today's scientists won't have to wait for AI historians to establish proper credit assignment. It is easy enough to do the right thing right now." You write: "let us be good role models and mentors" to the new generation. Then please do what's right! Your recent survey [S20] does not help. It's mentioned in my report as follows: "ACM seems to be influenced by a misleading 'history of deep learning' propagated by LBH & co-authors, e.g., Sejnowski [S20] (see Sec. XIII). It goes more or less like this: 'In 1969, Minsky & Papert [M69] showed that shallow NNs without hidden layers are very limited and the field was abandoned until a new generation of neural network researchers took a fresh look at the problem in the 1980s [S20].' However, as mentioned above, the 1969 book [M69] addressed a 'problem' of Gauss & Legendre's shallow learning (~1800)[DL1-2] that had already been solved 4 years prior by Ivakhnenko & Lapa's popular deep learning method [DEEP1-2][DL2] (and then also by Amari's SGD for MLPs [GD1-2]). Minsky was apparently unaware of this and failed to correct it later [HIN](Sec. I).... deep learning research was a! live and kicking also in the 1970s, especially outside of the Anglosphere." Just follow ACM's Code of Ethics and Professional Conduct [ACM18] which states: "Computing professionals should therefore credit the creators of ideas, inventions, work, and artifacts, and respect copyrights, patents, trade secrets, license agreements, and other methods of protecting authors' works." No need to wait for 100 years. J?rgen On 2 Jan 2022, at 23:29, Terry Sejnowski wrote: We would be remiss not to acknowledge that backprop would not be possible without the calculus, so Isaac newton should also have been given credit, at least as much credit as Gauss. All these threads will be sorted out by historians one hundred years from now. Our precious time is better spent moving the field forward. There is much more to discover. A new generation with better computational and mathematical tools than we had back in the last century have joined us, so let us be good role models and mentors to them. Terry -- -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Fri Feb 4 22:36:07 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Fri, 4 Feb 2022 22:36:07 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <303504A2-453D-4DE8-8A34-C41693041954@nyu.edu> References: <303504A2-453D-4DE8-8A34-C41693041954@nyu.edu> Message-ID: <702d7ba1-5891-d55f-1f8a-a188c4190b44@rubic.rutgers.edu> Gary, Not retreating.??? Simply stating the obvious. Brains are where? symbols, as we talk about them, are.?? But what are they exactly in a brain? I think that is a question that starts with connections and layers..? not? some sort of specialized software,? this is not about software engineering but about the science of the thing. Utility is important in many domains. but appears to me to be a retreat that you are didactically hiding behind ("even Jay"? comeon.) I am aware of simulations using RNN to counter you claim that RNNs could not learn certain sequential behavior. (https://www.semanticscholar.org/paper/On-the-Emergence-of-Rules-in-Neural-Networks-Hanson-Negishi/4bca27b823c9724d910b4637fd489343233570f8). I just can't take modularity of the brain seriously anymore, as cognitive neuroscience continues to embrace generic networks (resting state) and distributed representations-- things are moving on (see --the Failure of Blobology-SJH).? Face areas?? why would thetr be Face areas (different from neural patches)?. what would they do in any case---store all the faces you've seen--unlikely, process faces into parts from whole??? What was the point of a face area in the first place.. more likely to be some sort of? WAVELET? that incidentally? encodes faces,? 1963 Cadillacs and greebles (https://psycnet.apa.org/record/2008-00548-008; https://pubmed.ncbi.nlm.nih.gov/19883493/) then some specific type/token face thing. I think this is more about hoping there is some common ground that doesn't really exist. Most are beginning to see that? deep learning is a fundamental step forward in AI and yes is very non-biologically plausible, except for the obvious parts that are in the mammalian brain--layers? (the cortex has 6) and connections. Its better to focus on why it works and how to make mathematical sense of it. Steve On 2/4/22 2:52 PM, Gary Marcus wrote: > ?Steve, > > The phrase I always liked was ?poverty of the imagination arguments?; > I share your disdain for them. But that?s why I think you should be > careful of any retreat into biological plausibility. As even Jay > McClelland has acknowledged, we do know that some humans some of the > time manipulate symbols. So wetware-based symbols are not literally > biologically impossible; the real question for cognitive neuroscience > is about the scope and development of symbols. > > For engineering, the real question is, are they useful. Certainly for > software engineering in general, they are indispensable. > > Beyond this, none of the available AI approaches map particularly > neatly onto what we know about the brain, and none of what we know > about the brain is understood well enough to solve AI. ?All the > examples you point to, for instance, are actually controversial, not > decisive. As you probably know, for example, Nancy Kanwisher has a > different take on domain-specificity than you do > (https://web.mit.edu/bcs/nklab/ ), > with evidence of specialization early in life, and Jeff Bowers has > argued that the grandmother cell hypothesis has been dismissed > prematurely > ?(https://jeffbowers.blogs.bristol.ac.uk/blog/grandmother-cells/ > ); > there?s also a long literature on the possible neural realization of > rules, both in humans and other animals. > > I don?t know what the right answers are there, but nor do I think that > neurosymbolic systems are beholden to them anymore than CNNs are bound > to whether or not the brain performs back-propagation. > > Finally, as a reminder, ??Distributed? per se in not the right > question; in some technical sense ASCII encodings are distributed, and > about as symbolic as you can get. The proper question is really what > you do with your encodings; the neurosymbolic approach is trying to > broaden the available range of options. > > Gary > >> On Feb 4, 2022, at 07:04, Stephen Jos? Hanson >> wrote: >> >> ? >> >> Well I don't like counterfactual arguments or ones that start with >> "It can't be done with neural networks.."--as this amounts to the old >> Rumelhart saw, of "proof by lack of imagination". >> >> I think my position and others (I can't speak for Geoff and won't) is >> more of a "purist" view that brains have computationally complete >> representational power to do what ever is required of human level >> mental processing.? AI symbol systems are remote descriptions of this >> level of processing. Looking at 1000s of brain scans, one begins to >> see a pattern of interacting large and smaller scale networks, >> probably related to Resting state and the Default Mode networks in >> some important competitive way.?? But what one doesn't find is >> modular structure (e.g. face area.. nope)? or evidence of "symbols"? >> being processed. Research on Numbers is interesting in this regard, >> as number representation should provide some evidence of discrete >> symbol processing as would? letters.?? But again the processing >> states from brain imaging? more generally appear to be distributed >> representations of some sort. >> >> One other direction has to do with prior rules that could be neurally >> coded and therefore provide an immediate bias in learning and thus >> dramatically reduce the number of examples required for? asymptotic >> learning.???? Some of this has been done with pre-training-- on let's >> say 1000s of videos that are relatively generic, prior to learning on >> a small set of videos related to a specific topic-- say two >> individuals playing a monopoly game.? In that case, no game-like >> videos were sampled in the pre-training, and the LSTM was trained to >> detect change point on 2 minutes of video, achieving a 97% match with >> human parsers.??? In these senses I have no problem with this type of >> hybrid training. >> >> Steve >> >> On 2/4/22 9:07 AM, Gary Marcus wrote: >>> ?The whole point of the neurosymbolic approach is to develop systems >>> that can accommodate both vectors and symbols, since neither on >>> their own seems adequate. >>> >>> If there are arguments against trying to do that, we would be >>> interested. >>> >>>> On Feb 4, 2022, at 4:17 AM, Stephen Jos? Hanson >>>> wrote: >>>> >>>> ? >>>> >>>> Geoff's position is pretty clear.?? He said in the conversation we >>>> had and in this thread,? "vectors of soft features", >>>> >>>> Some of my claim is in several of the conversations with Mike >>>> Jordan and Rich Sutton, but briefly,? there are a number of >>>> very large costly efforts from the 1970s and 1980s, to create, >>>> deploy and curate symbol AI systems that were massive failures.? >>>> Not counterfactuals,? but factuals that failed. The MCC comes to >>>> mind with Adm Bobby Inmann's national US mandate to counter the >>>> Japanese so called"Fifth-generation AI systems"? as a massive >>>> failure of symbolic AI. >>>> >>>> -------------------- >>>> >>>> In 1982, Japan launched its Fifth Generation Computer Systems >>>> project (FGCS), designed to develop intelligent software that would >>>> run on novel computer hardware. As the first national, large-scale >>>> artificial intelligence (AI) research and development (R&D) project >>>> to be free from military influence and corporate profit motives, >>>> the FGCS was open, international, and oriented around public goods. >>>> >>>> On 2/3/22 6:34 PM, Francesca Rossi2 wrote: >>>>> Hi all. >>>>> >>>>> Thanks Gary for adding me to this thread. >>>>> >>>>> I also would be interested in knowing why Steve thinks that NS AI did not work in the past, and why this is an indication that it cannot work now or in the future. >>>>> >>>>> Thanks, >>>>> Francesca. >>>>> ------------------ >>>>> >>>>> Francesca Rossi >>>>> IBM Fellow and AI Ethics Global Leader >>>>> T.J. Watson Research Center, Yorktown Heights, USA >>>>> +1-617-3869639 >>>>> >>>>> ________________________________________ >>>>> From: Artur Garcez >>>>> Sent: Thursday, February 3, 2022 6:00 PM >>>>> To: Gary Marcus >>>>> Cc: Stephen Jos? Hanson; Geoffrey Hinton; AIhub;connectionists at mailman.srv.cs.cmu.edu; Luis Lamb; Josh Tenenbaum; Anima Anandkumar; Francesca Rossi2; Swarat Chaudhuri; Gadi Singer >>>>> Subject: [EXTERNAL] Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>>> >>>>> It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally ZjQcmQRYFpfptBannerStart >>>>> This Message Is From an External Sender >>>>> This message came from outside your organization. >>>>> ZjQcmQRYFpfptBannerEnd >>>>> >>>>> It would be great to hear Geoff's account with historical reference to his 1990 edited special volume of the AI journal on connectionist symbol processing. >>>>> >>>>> Judging from recent reviewing for NeurIPS, ICLR, ICML but also KR, AAAI, IJCAI (traditionally symbolic), there is a clear resurgence of neuro-symbolic approaches. >>>>> >>>>> Best wishes, >>>>> Artur >>>>> >>>>> >>>>> On Thu, Feb 3, 2022 at 5:00 PM Gary Marcus > wrote: >>>>> Steve, >>>>> >>>>> I?d love to hear you elaborate on this part, >>>>> >>>>> Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. >>>>> >>>>> >>>>> I?d love your take on why you think the impetus towards hybrid models ended badly before, and why you think that the mistakes of the past can?t be corrected. Also it? would be really instructive to compare with deep learning, which lost steam for quite some time, but reemerged much stronger than ever before. Might not the same happen with hybrid models? >>>>> >>>>> I am cc?ing some folks (possibly not on this list) who have recently been sympathetic to hybrid models, in hopes of a rich discussion. (And, Geoff, still cc?d, I?d genuinely welcome your thoughts if you want to add them, despite our recent friction.) >>>>> >>>>> Cheers, >>>>> Gary >>>>> >>>>> >>>>> On Feb 3, 2022, at 5:10 AM, Stephen Jos? Hanson > wrote: >>>>> >>>>> >>>>> I would encourage you to read the whole transcript, as you will see the discussion does intersect with a number of issues you raised in an earlier post on what is learned/represented in DLs. >>>>> >>>>> Its important for those paying attention to this thread, to realize these are still very early times. Many more shoes will drop in the next few years. I for one don't believe one of those shoes will be Hybrid approaches to AI, I've seen that movie before and it didn't end well. >>>>> >>>>> Best and hope you are doing well. >>>>> >>>>> Steve >>>>> >>>> -- >>>> >> -- -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From pau at ini.uzh.ch Sat Feb 5 11:29:58 2022 From: pau at ini.uzh.ch (pau) Date: Sat, 05 Feb 2022 17:29:58 +0100 Subject: Connectionists: Please stop having conversations on this mailing list Message-ID: Dear connectionists, To the best of my understanding, the aim of this mailing list is to meet the needs of working professionals. I understand by this sharing information about events, publications, developments, etc. that can be useful for someone who works in this field. If someone wants to argue, discuss, praise or assign blame, agree or disagree, please do so in private correspondence or in non-work-specific social media channels. Best, P. From gary.marcus at nyu.edu Sat Feb 5 10:38:55 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Sat, 5 Feb 2022 07:38:55 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <537DF004-25CE-45A2-8155-D7E6018F4EE5@nyu.edu> There is no magic in understanding, just computation that has been realized in the wetware of humans and that eventually can be realized in machines. But understanding is not (just) learning. Understanding incorporates (or works in tandem with) learning - but also, critically, in tandem with inference, and the development and maintenance of cognitive models. Part of developing an understanding of cats in general is to learn long term-knowledge about their properties, both directly (e.g., through observation) and indirectly (eg through learning facts about animals in general that can be extended to cats), often through inference (if all animals have DNA, and a cat is an animal, it must also have DNA). The understanding of a particular cat also involves direct observation, but also inference (eg one might surmise that the reason that Fluffy is running about the room is that Fluffy suspects there is a mouse stirring somewhere nearby). But all of that, I would say, is subservient to the construction of cognitive models that can be routinely updated (e.g., Fluffy is currently in the living room, skittering about, perhaps looking for a mouse). In humans, those dynamic, relational models, which form part of an understanding, can support inference (if Fluffy is in the living room, we can infer that Fluffy is not outside, not lost, etc). Without such models - which I think represent a core part of understanding - AGI is an unlikely prospect. Current neural networks, as it happens, are better at acquiring long-term knowledge (cats have whiskers) than they are at dynamically updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic model that I am describing. To a modest degree they can approximate it on the basis of large samples of texts, but their ultimate incoherence stems from the fact that they do not have robust internal cognitive models that they can update on the fly. Without such cognitive models you can still capture some aspects of understanding (eg predicting that cats are likely to be furry), but things fall apart quickly; inference is never reliable, and coherence is fleeting. As a final note, one of the most foundational challenges in constructing adequate cognitive models of the world is to have a clear distinction between individuals and kinds; as I emphasized 20 years ago (in The Algebraic Mind), this has always been a weakness in neural networks, and I don?t think that the type-token problem has yet been solved. Gary > On Feb 5, 2022, at 01:31, Asim Roy wrote: > > ? > All, > > I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? > > https://www.youtube.com/watch?v=gn4nRCC9TwQ > https://www.youtube.com/watch?v=8sO7VS3q8d0 > > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > > > From: Ali Minai > Sent: Friday, February 4, 2022 11:38 PM > To: Asim Roy > Cc: Gary Marcus ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Asim > > Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. > > FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! > > Ali > > Ali A. Minai, Ph.D. > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > > On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: > First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. > > DANKO: > What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) > What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) > What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) > What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > From: Gary Marcus > Sent: Thursday, February 3, 2022 9:26 AM > To: Danko Nikolic > Cc: Asim Roy ; Geoffrey Hinton ; AIhub ; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Danko, > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. > > I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. > > Gary > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." > > I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". > To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. > > What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) > What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) > What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) > What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) > > ...and so on. > > Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). > > I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. > > In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. > > No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. > > I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. > > So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: > > 1) first, the network learned about hamsters (not many examples) > 2) after that the network learned about red hats (outside the context of hamsters and without many examples) > 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) > > After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. > > Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. > > > Danko > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > --- A progress usually starts with an insight --- > > > > Virus-free. www.avast.com > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) > GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > From: Connectionists On Behalf Of Gary Marcus > Sent: Wednesday, February 2, 2022 1:26 PM > To: Geoffrey Hinton > Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Geoff, and interested others, > > What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: > > > > One could > > a. avert one?s eyes and deem the anomalous outputs irrelevant > or > b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. > > Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. > > All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. > > With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) > > Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. > > Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? > > In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. > > Gary > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: > > ? > Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. > > A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. > > > But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? > > Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. > I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. > > Geoff > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, > > There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. > > The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? > > I never said any such thing. > > What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. > > I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. > > I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: > > You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: > > Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. > > It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. > > Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. > > Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: > > The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. > > Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. > > Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). > > More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. > > As Herb Simon once observed, science does not have to be zero-sum. > > Sincerely, > Gary Marcus > Professor Emeritus > New York University > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > Stephen Hanson in conversation with Geoff Hinton > > In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. > > You can watch the discussion, and read the transcript, here: > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > About AIhub: > AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. > Twitter: @aihuborg > > > Virus-free. www.avast.com > -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Sun Feb 6 22:08:07 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Mon, 7 Feb 2022 03:08:07 +0000 Subject: Connectionists: A New Society for Explainable AI + a possible inaugural conference in San Francisco in late July, early August Message-ID: Dear Colleagues, Some of us are thinking of forming a new Society for Explainable AI. If there?s enough interest, we can have our first conference in late July or early August in San Francisco. And a separate journal can follow that. You can email me if you have interest. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) -------------- next part -------------- An HTML attachment was scrubbed... URL: From minaiaa at gmail.com Sun Feb 6 16:18:55 2022 From: minaiaa at gmail.com (Ali Minai) Date: Sun, 6 Feb 2022 16:18:55 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <981624DC-3FA6-440E-910F-055B1FB77C41@nyu.edu> References: <981624DC-3FA6-440E-910F-055B1FB77C41@nyu.edu> Message-ID: Gary, That?s a very interesting and accurate list of capabilities that a general intelligent system must have and that our AI does not. Of course, the list is familiar to me from having read your book. However, I have a somewhat different take on this whole thing. All the things we discuss here ? symbols/no symbols, parts/wholes, supervised/unsupervised, token/type, etc., are useful categories and distinctions for our analysis of the problem, and are partly a result of the historical evolution of the field of AI in particular and of philosophy in general. The categories are not wrong in any way, of course, but they are posterior to the actual system ? good for describing and analyzing it, and for validating our versions of it (which is how you use them). I think they are less useful as prescriptions for how to build our AI systems. If intelligent systems did not already exist and we were building them from scratch (please ignore the impossibility of that), having a list of ?must haves? would be great. But intelligent systems already exist ? from humans to fish ? and they already have these capacities to a greater or lesser degree because of the physics of their biology. A cat?s intelligence does not care whether it has symbols or not, and nor does mine or yours. Whatever we describe as symbolic processing post-facto has already been done by brains for at least tens of millions of years. Instead of getting caught up in ?how to add symbols into our neural models?, we should be investigating how what we see as symbolic processing emerges from animal brains, and then replicate those brains to the degree necessary. If we can do that, symbolic processing will already be present. But it cannot be done piece by piece. It must take the integrity of the whole brain and the body it is part of, and its environment, into account. That?s why I think that a much better ? though a very long ? route to AI is to start by understanding how a fish brain makes the intelligence of a fish possible, and then boot up our knowledge across phylogenetic stages: Bottom up reverse engineering rather than top-down engineering. That?s the way Nature built up to human intelligence, and we will succeed only by reverse engineering it. Of course, we can do it much faster and with shortcuts because we are intelligent, purposive agents, but working top-down by building piecewise systems that satisfy a list of attributes will not get us there. Among other things, those pieces will be impossible to integrate into the kind of intelligence that can have those general models of the world that you rightly point to as being necessary. I think that one thing that has been a great boon to the AI enterprise has also been one of the greatest impediments to its complete success, and that is the ?computationalization? of intelligence. On the one hand, thinking of intelligence computationally allows us to describe it abstractly and in a principled, formal way. It also resonates with the fact that we are trying to implement intelligence through computational machines. But, on the flip side, this view of intelligence divorces it from its physics ? from the fact that real intelligence in animals emerges from the physics of the physical system. That system is not a collection of its capabilities; rather, those capabilities are immanent in it by virtue of its physics. When we try to build those capabilities computationally, i.e., through code, we are making the same error that the practitioners of old-style ?symbolic AI? made ? what I call the ?professors are smarter than Nature? error, i.e., the idea that we are going to enumerate (or describe) all the things that underlie intelligence and implement them one by one until we get complete intelligence. We will never be able to enumerate all those capabilities, and will never be able to get to that complete intelligence. The only difference between us and the ?symbolists? of yore is that we are replacing giant LISP and Prolog programs with giant neural networks. Otherwise, we are using our models exactly as they were trying to use their models, and we will fail just as they did unless we get back to biology and the real thing. I will say again that the way we do AI today is driven more by habit and the incentives of the academic and corporate marketplaces than by a deep, long-term view of AI as a great exploratory project in fundamental science. We are just building AI to drive our cars, translate our documents, write our reports, and do our shopping. What that will teach us about actual intelligence is just incidental. My apologies too for a long response. Ali *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Sun, Feb 6, 2022 at 9:42 AM Gary Marcus wrote: > Dear Asim, > > > Sorry for a long answer to your short but rich questions. > > - Yes, memory in my view has to be part of the answer to the > type-token problem. Symbol systems encoded in memory allow a natural way to > set up records, and something akin to that seems necessary. Pure multilayer > perceptrons struggle with type-token distinctions precisely because they > lack such records. On the positive side, I see more and more movement > towards recordlike stores (eg w key-value stores in memory networks), and I > think that is an important and necessary step, very familiar from the > symbol-manipulating playbook, sometimes implemented in new ways. > - But ultimately, handling the type-token distinction requires > considerable inferential overhead beyond the memory representation of a > record per se. How do you determine when to denote something (e.g. > Felix) as an instance, and of which kinds (cat, animal etc), and how do you > leverage that knowledge once you determine it? > - In the limit we reason about types vs tokens in fairly subtle ways, > eg in guessing whether a glass that we put down at party is likely to be > ours. The reverse is also important: we need to be learn particular > traits for individuals and not erroneously generalize them to the class; if > my aunt Esther wins the lottery, one shouldn?t infer that all of my > aunts or all of my relatives or adult females have won the lottery. so you > need both representational machinery that can distinguish eg my cat from > cats in general and reasoning machinery to decide at what level certain > learned knowledge should inhere. (I had a whole chapter about this sort of > thing in The Algebraic Mind if you are interested, and Mike Mozer had a > book about types and tokens in neural networks in the mid 1990s). > - Yes, part (though not all!) of what we do when we set up cognitive > models in our heads is to track particular individuals and their > properties. If you only had to correlate kinds (cats) and their properties > (have fur) you could maybe get away with a multilayer perceptron, but once > you need to track individuals, yes, you really need some kind of > memory-based records. > - As far as I can tell, Transformers can sometimes approximate some of > this for a few sentences, but not over long stretches. > > > As a small terminological aside; for me cognitive models ? cognitive > modeling. Cognitive modeling is about building psychological or > computational models of how people think, whereas what I mean by a cognitive > model is a representation of eg the entities in some situation and the > relations between those entities. > > > To your closing question, none of us yet really knows how to build > understanding into machines. A solid type-token distinction, both in > terms of representation and reasoning, is critical for general > intelligence, but hardly sufficient. Personally, I think some minimal > prerequisites would be: > > - representations of space, time, causality, individuals, kinds, > persons, places, objects, etc. > - representations of abstractions that can hold over all entities in a > class > - compositionality (if we are talking about human-like understanding) > - capacity to construct and update cognitive models on the fly > - capacity to reason over entities in those models > - ability to learn about new entities and their properties > > Much of my last book (*Rebooting AI*, w Ernie Davis) is about the above > list. The section in the language chapter on a children?s story in which > man has lost is wallet is an especially vivid worked example. Later > chapters elaborate some of the challenges in representing space, time, and > causality. > > > Gary > > > On Feb 5, 2022, at 18:58, Asim Roy wrote: > > ? > > Gary, > > > > I don?t get much into the type of cognitive modeling you are talking > about, but I would guess that the type problem can generally be handled by > neural network models and tokens can be resolved with some memory-based > system. But to the heart of the question, this is what so-called > ?understanding? reduces to computation wise? > > > > Asim > > > > *From:* Gary Marcus > *Sent:* Saturday, February 5, 2022 8:39 AM > *To:* Asim Roy > *Cc:* Ali Minai ; Danko Nikolic < > danko.nikolic at gmail.com>; Brad Wyble ; > connectionists at mailman.srv.cs.cmu.edu; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > There is no magic in understanding, just computation that has been > realized in the wetware of humans and that eventually can be realized in > machines. But understanding is not (just) learning. > > > > Understanding incorporates (or works in tandem with) learning - but also, > critically, in tandem with inference, *and the development and > maintenance of cognitive models*. Part of developing an understanding of > cats in general is to learn long term-knowledge about their properties, > both directly (e.g., through observation) and indirectly (eg through > learning facts about animals in general that can be extended to cats), > often through inference (if all animals have DNA, and a cat is an animal, > it must also have DNA). The understanding of a particular cat also > involves direct observation, but also inference (eg one might surmise > that the reason that Fluffy is running about the room is that Fluffy > suspects there is a mouse stirring somewhere nearby). *But all of that, I > would say, is subservient to the construction of cognitive models that can > be routinely updated *(e.g., Fluffy is currently in the living room, > skittering about, perhaps looking for a mouse). > > > > In humans, those dynamic, relational models, which form part of an > understanding, can support inference (if Fluffy is in the living room, we > can infer that Fluffy is not outside, not lost, etc). Without such models - > which I think represent a core part of understanding - AGI is an unlikely > prospect. > > > > Current neural networks, as it happens, are better at acquiring long-term > knowledge (cats have whiskers) than they are at dynamically updating > cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic > model that I am describing. To a modest degree they can approximate it on > the basis of large samples of texts, but their ultimate incoherence stems > from the fact that they do not have robust internal cognitive models that > they can update on the fly. > > > > Without such cognitive models you can still capture some aspects of > understanding (eg predicting that cats are likely to be furry), but things > fall apart quickly; inference is never reliable, and coherence is fleeting. > > > > As a final note, one of the most foundational challenges in constructing > adequate cognitive models of the world is to have a clear distinction > between individuals and kinds; as I emphasized 20 years ago (in The > Algebraic Mind), this has always been a weakness in neural networks, and I > don?t think that the type-token problem has yet been solved. > > > > Gary > > > > > > On Feb 5, 2022, at 01:31, Asim Roy wrote: > > ? > > All, > > > > I think the broader question was ?understanding.? Here are two Youtube > videos showing simple robots ?learning? to walk. They are purely physical > systems. Do they ?understand? anything ? such as the need to go around an > obstacle, jumping over an obstacle, walking up and down stairs and so on? > By the way, they ?learn? to do these things on their own, literally > unsupervised, very much like babies. The basic question is: what is > ?understanding? if not ?learning?? Is there some other mechanism (magic) at > play in our brain that helps us ?understand?? > > > > https://www.youtube.com/watch?v=gn4nRCC9TwQ > > > https://www.youtube.com/watch?v=8sO7VS3q8d0 > > > > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > > > > > *From:* Ali Minai > *Sent:* Friday, February 4, 2022 11:38 PM > *To:* Asim Roy > *Cc:* Gary Marcus ; Danko Nikolic < > danko.nikolic at gmail.com>; Brad Wyble ; > connectionists at mailman.srv.cs.cmu.edu; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Asim > > > > Of course there's nothing magical about understanding, and the mind has to > emerge from the physical system, but our AI models at this point are not > even close to realizing how that happens. We are, at best, simulating a > superficial approximation of a few parts of the real thing. A single, > integrated system where all the aspects of intelligence emerge from the > same deep, well-differentiated physical substrate is far beyond our > capacity. Paying more attention to neurobiology will be essential to get > there, but so will paying attention to development - both physical and > cognitive - and evolution. The configuration of priors by evolution is key > to understanding how real intelligence learns so quickly and from so > little. This is not an argument for using genetic algorithms to design our > systems, just for understanding the tricks evolution has used and > replicating them by design. Development is more feasible to do > computationally, but hardly any models have looked at it except in a > superficial sense. Nature creates basic intelligence not so much by > configuring functions by explicit training as by tweaking, modulating, > ramifying, and combining existing ones in a multi-scale self-organization > process. We then learn much more complicated things (like playing chess) by > exploiting that substrate, and using explicit instruction or learning by > practice. The fundamental lesson of complex systems is that complexity is > built in stages - each level exploiting the organization of the level below > it. We see it in evolution, development, societal evolution, the evolution > of technology, etc. Our approach in AI, in contrast, is to initialize a > giant, naive system and train it to do something really complicated - but > really specific - by training the hell out of it. Sure, now we do build > many systems on top of pre-trained models like GPT-3 and BERT, which is > better, but those models were again trained by the same none-to-all process > I decried above. Contrast that with how humans acquire language, and how > they integrate it into their *entire* perceptual, cognitive, and behavioral > repertoire, not focusing just on this or that task. The age of symbolic AI > may have passed, but the reductionistic mindset has not. We cannot build > minds by chopping it into separate verticals. > > > > FTR, I'd say that the emergence of models such as GLOM and Hawkins and > Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but > they are, I think, looking in the right direction. With a million miles to > go! > > > > Ali > > > > *Ali A. Minai, Ph.D.* > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > > 828 Rhodes Hall > > University of Cincinnati > Cincinnati, OH 45221-0030 > > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > > > > > > On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: > > First of all, the brain is a physical system. There is no ?magic? inside > the brain that does the ?understanding? part. Take for example learning to > play tennis. You hit a few balls - some the right way and some wrong ? but > you fairly quickly learn to hit them right most of the time. So there is > obviously some simulation going on in the brain about hitting the ball in > different ways and ?learning? its consequences. What you are calling > ?understanding? is really these simulations about different scenarios. It?s > also very similar to augmentation used to train image recognition systems > where you rotate images, obscure parts and so on, so that you still can say > it?s a cat even though you see only the cat?s face or whiskers or a cat > flipped on its back. So, if the following questions relate to > ?understanding,? you can easily resolve this by simulating such scenarios > when ?teaching? the system. There?s nothing ?magical? about > ?understanding.? As I said, bear in mind that the brain, after all, is a > physical system and ?teaching? and ?understanding? is embodied in that > physical system, not outside it. So ?understanding? is just part of > ?learning,? nothing more. > > > > DANKO: > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Gary Marcus > *Sent:* Thursday, February 3, 2022 9:26 AM > *To:* Danko Nikolic > *Cc:* Asim Roy ; Geoffrey Hinton < > geoffrey.hinton at gmail.com>; AIhub ; > connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Danko, > > > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, > in which he said (paraphrasing from memory, because I don?t remember the > precise words) that the famous 200 Quoc Le unsupervised model [ > https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf > ] > had learned the concept of a ca. In reality the model had clustered > together some catlike images based on the image statistics that it had > extracted, but it was a long way from a full, counterfactual-supporting > concept of a cat, much as you describe below. > > > > I fully agree with you that the reason for even having a semantics is as > you put it, "to 1) learn with a few examples and 2) apply the knowledge to > a broad set of situations.? GPT-3 sometimes gives the appearance of having > done so, but it falls apart under close inspection, so the problem remains > unsolved. > > > > Gary > > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > > > G. Hinton wrote: "I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster wearing a red hat and > it draws such a picture, it understood the request." > > > > I would like to suggest why drawing a hamster with a red hat does not > necessarily imply understanding of the statement "hamster wearing a red > hat". > > To understand that "hamster wearing a red hat" would mean inferring, in > newly emerging situations of this hamster, all the real-life > implications that the red hat brings to the little animal. > > > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > ...and so on. > > > > Countless many questions can be asked. One has understood "hamster wearing > a red hat" only if one can answer reasonably well many of such real-life > relevant questions. Similarly, a student has understood materias in a class > only if they can apply the materials in real-life situations (e.g., > applying Pythagora's theorem). If a student gives a correct answer to a > multiple choice question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > > > I also suggest that understanding also comes together with effective > learning: We store new information in such a way that we can recall it > later and use it effectively i.e., make good inferences in newly emerging > situations based on this knowledge. > > > > In short: Understanding makes us humans able to 1) learn with a few > examples and 2) apply the knowledge to a broad set of situations. > > > > No neural network today has such capabilities and we don't know how to > give them such capabilities. Neural networks need large amounts of > training examples that cover a large variety of situations and then > the networks can only deal with what the training examples have already > covered. Neural networks cannot extrapolate in that 'understanding' sense. > > > > I suggest that understanding truly extrapolates from a piece of knowledge. > It is not about satisfying a task such as translation between languages or > drawing hamsters with hats. It is how you got the capability to complete > the task: Did you only have a few examples that covered something different > but related and then you extrapolated from that knowledge? If yes, this is > going in the direction of understanding. Have you seen countless examples > and then interpolated among them? Then perhaps it is not understanding. > > > > So, for the case of drawing a hamster wearing a red hat, understanding > perhaps would have taken place if the following happened before that: > > > > 1) first, the network learned about hamsters (not many examples) > > 2) after that the network learned about red hats (outside the context of > hamsters and without many examples) > > 3) finally the network learned about drawing (outside of the context of > hats and hamsters, not many examples) > > > > After that, the network is asked to draw a hamster with a red hat. If it > does it successfully, maybe we have started cracking the problem of > understanding. > > > > Note also that this requires the network to learn sequentially without > exhibiting catastrophic forgetting of the previous knowledge, which is > possibly also a consequence of human learning by understanding. > > > > > > Danko > > > > > > > > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > > > > Virus-free. www.avast.com > > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > > Without getting into the specific dispute between Gary and Geoff, I think > with approaches similar to GLOM, we are finally headed in the right > direction. There?s plenty of neurophysiological evidence for single-cell > abstractions and multisensory neurons in the brain, which one might claim > correspond to symbols. And I think we can finally reconcile the decades old > dispute between Symbolic AI and Connectionism. > > > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > GARY: I have *never* called for dismissal of neural networks, but rather > for some hybrid between the two (as you yourself contemplated in 1991); the > point of the 2001 book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols could complement > them. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Gary Marcus > *Sent:* Wednesday, February 2, 2022 1:26 PM > *To:* Geoffrey Hinton > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Geoff, and interested others, > > > > What, for example, would you make of a system that often drew the > red-hatted hamster you requested, and perhaps a fifth of the time gave you > utter nonsense? Or say one that you trained to create birds but sometimes > output stuff like this: > > > > > > > > One could > > > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > or > > b. wonder if it might be possible that sometimes the system gets the right > answer for the wrong reasons (eg partial historical contingency), and > wonder whether another approach might be indicated. > > > > Benchmarks are harder than they look; most of the field has come to > recognize that. The Turing Test has turned out to be a lousy measure of > intelligence, easily gamed. It has turned out empirically that the Winograd > Schema Challenge did not measure common sense as well as Hector might have > thought. (As it happens, I am a minor coauthor of a very recent review on > this very topic: https://arxiv.org/abs/2201.02387 > ) > But its conquest in no way means machines now have common sense; many > people from many different perspectives recognize that (including, e.g., > Yann LeCun, who generally tends to be more aligned with you than with me). > > > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote > me; but what you said about me and machine translation remains your > invention, and it is inexcusable that you simply ignored my 2019 > clarification. On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains unsolved. > > > > All of the problems LLMs have with coherence, reliability, truthfulness, > misinformation, etc stand witness to that fact. (Their persistent inability > to filter out toxic and insulting remarks stems from the same.) I am hardly > the only person in the field to see that progress on any given benchmark > does not inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > > > With respect to embeddings: Embeddings are very good for natural language > *processing*; but NLP is not the same as NL*U* ? when it comes to > *understanding*, their worth is still an open question. Perhaps they will > turn out to be necessary; they clearly aren?t sufficient. In their extreme, > they might even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a specific set of > numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > > > (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > > Notably absent from your email is any kind of apology for misrepresenting > my position. It?s fine to say that ?many people thirty years ago once > thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. > I have consistently felt throughout our interactions that you have mistaken > me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me > for having made that error. I am still not he. > > > > Which maybe connects to the last point; if you read my work, you would see > thirty years of arguments *for* neural networks, just not in the way that > you want them to exist. I have ALWAYS argued that there is a role for them; > characterizing me as a person ?strongly opposed to neural networks? misses > the whole point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > > > In the last two decades or so you have insisted (for reasons you have > never fully clarified, so far as I know) on abandoning symbol-manipulation, > but the reverse is not the case: I have *never* called for dismissal of > neural networks, but rather for some hybrid between the two (as you > yourself contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded and broke down, > and where symbols could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > > > Gary > > > > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: > > ? > > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > > > > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > > > Geoff > > > > > > > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural > network community, > > > > There has been a lot of recent discussion on this list about framing and > scientific integrity. Often the first step in restructuring narratives is > to bully and dehumanize critics. The second is to misrepresent their > position. People in positions of power are sometimes tempted to do this. > > > > The Hinton-Hanson interview that you just published is a real-time example > of just that. It opens with a needless and largely content-free personal > attack on a single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive thing it says is > false. > > > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t > be able to do machine translation.? > > > > I never said any such thing. > > > > What I predicted, rather, was that multilayer perceptrons, as they existed > then, would not (on their own, absent other mechanisms) *understand* language. > Seven years later, they still haven?t, except in the most superficial way. > > > > > I made no comment whatsoever about machine translation, which I view as a > separate problem, solvable to a certain degree by correspondance without > semantics. > > > > I specifically tried to clarify Hinton?s confusion in 2019, but, > disappointingly, he has continued to purvey misinformation despite that > clarification. Here is what I wrote privately to him then, which should > have put the matter to rest: > > > > You have taken a single out of context quote [from 2015] and > misrepresented it. The quote, which you have prominently displayed at the > bottom on your own web page, says: > > > > Hierarchies of features are less suited to challenges such as language, > inference, and high-level planning. For example, as Noam Chomsky famously > pointed out, language is filled with sentences you haven't seen > before. Pure classifier systems don't know what to do with such sentences. > The talent of feature detectors -- in identifying which member of some > category something belongs to -- doesn't translate into understanding > novel sentences, in which each sentence has its own unique meaning. > > > > It does *not* say "neural nets would not be able to deal with novel > sentences"; it says that hierachies of features detectors (on their own, if > you read the context of the essay) would have trouble *understanding *novel sentences. > > > > > Google Translate does yet not *understand* the content of the sentences > is translates. It cannot reliably answer questions about who did what to > whom, or why, it cannot infer the order of the events in paragraphs, it > can't determine the internal consistency of those events, and so forth. > > > > Since then, a number of scholars, such as the the computational linguist > Emily Bender, have made similar points, and indeed current LLM difficulties > with misinformation, incoherence and fabrication all follow from these > concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter > with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf > , > also emphasizing issues of understanding and meaning: > > > > *The success of the large neural language models on many NLP tasks is > exciting. However, we find that these successes sometimes lead to hype in > which these models are being described as ?understanding? language or > capturing ?meaning?. In this position paper, we argue that a system trained > only on form has a priori no way to learn meaning. .. a clear understanding > of the distinction between form and meaning will help guide the field > towards better science around natural language understanding. * > > > > Her later article with Gebru on language models ?stochastic parrots? is in > some ways an extension of this point; machine translation requires mimicry, > true understanding (which is what I was discussing in 2015) requires > something deeper than that. > > > > Hinton?s intellectual error here is in equating machine translation with > the deeper comprehension that robust natural language understanding will > require; as Bender and Koller observed, the two appear not to be the same. > (There is a longer discussion of the relation between language > understanding and machine translation, and why the latter has turned out to > be more approachable than the former, in my 2019 book with Ernest Davis). > > > > More broadly, Hinton?s ongoing dismissiveness of research from > perspectives other than his own (e.g. linguistics) have done the field a > disservice. > > > > As Herb Simon once observed, science does not have to be zero-sum. > > > > Sincerely, > > Gary Marcus > > Professor Emeritus > > New York University > > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > > Stephen Hanson in conversation with Geoff Hinton > > > > In the latest episode of this video series for AIhub.org > , > Stephen Hanson talks to Geoff Hinton about neural networks, > backpropagation, overparameterization, digit recognition, voxel cells, > syntax and semantics, Winograd sentences, and more. > > > > You can watch the discussion, and read the transcript, here: > > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > > > About AIhub: > > AIhub is a non-profit dedicated to connecting the AI community to the > public by providing free, high-quality information through AIhub.org > > (https://aihub.org/ > ). > We help researchers publish the latest AI news, summaries of their work, > opinion pieces, tutorials and more. We are supported by many leading > scientific organizations in AI, namely AAAI > , > NeurIPS > , > ICML > , > AIJ > > /IJCAI > , > ACM SIGAI > , > EurAI/AICOMM, CLAIRE > > and RoboCup > > . > > Twitter: @aihuborg > > > > > > > Virus-free. www.avast.com > > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Sun Feb 6 22:38:07 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Sun, 6 Feb 2022 22:38:07 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Asim, The brain does not assume a single object in a cluttered science. Thus, a simple explanation like "recognition by parts" (but without object abstraction) should be invalid. It is like a chicken and egg problem. Both chicken and egg are absent. We must not assume egg is there or chicken is there. Best regards, -John On Sun, Feb 6, 2022 at 2:42 PM Asim Roy wrote: > Dear John, > > > > You are right and I admit I am not solving all of the problems. It?s just > in reference to this one problem that Geoffrey Hinton mentions that I think > can be resolved: > > *?I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. We can introspect on how we do it > and this may or may not give some insight into how we check our answer, but > the immediate sense that a handwritten 2 is a 2 is computed by a neural net > that is not functionally equivalent to any simple and easily explainable > procedure.?* > > Best, > > Asim > > > > *From:* Juyang Weng > *Sent:* Sunday, February 6, 2022 10:06 AM > *To:* Asim Roy > *Cc:* Geoffrey Hinton ; Dietterich, Thomas < > tgd at oregonstate.edu>; AIhub ; > connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; > Danko Nikolic ; Stephen Jos? Hanson < > jose at rubic.rutgers.edu>; Marek Reformat ; MARCO > GORI ; Alessandro Sperduti < > alessandro.sperduti at unipd.it>; Xiaodong Li ; > Hava Siegelmann ; Peter Tino < > P.Tino at cs.bham.ac.uk>; Bing Xue ; Ali Minai < > minaiaa at gmail.com>; Claudius Gros ; > Jean-Philippe Thivierge ; Tsvi Achler > ; Prof A Hussain > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Asim, > > I try to be brief so that I can explain why many of us have missed, and > will continue to miss, the boat. > In some of my talks, I have a ppt slide "The brain is like blindmen and an > elephant". > > Unfortunately, your "identify objects based on its parts" is a good > traditional idea from pattern recognition that is still a blindman. > > Your idea does not explain many other problems without which we will never > understand a biological brain. > > For example, your idea does not explain how the brain learns planning and > discovery in a cluttered world. > > We must solve many million-dollar problems holistically. Please watch my > YouTube video: > Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar > Problems Solved > https://youtu.be/Dgx1dLCdSKY > > > Best regards, > > -John > > > > On Sat, Feb 5, 2022 at 12:01 AM Asim Roy wrote: > > I am responding to this part of Geoffrey Hinton?s note: > > > > *?I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. We can introspect on how we do it > and this may or may not give some insight into how we check our answer, but > the immediate sense that a handwritten 2 is a 2 is computed by a neural net > that is not functionally equivalent to any simple and easily explainable > procedure.?* > > > > The causal explanation is actually done quite simply, and we are doing it > currently. I can talk about this now because Arizona State University (ASU) > has filed a provisional patent application on the technology. The basic > idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable > Artificial Intelligence (darpa.mil) > ) > and illustrated in the figure below. The idea is to identify objects based > on its parts. So, the figure below says that it?s a cat because it has fur, > whiskers, and claws plus an unlabeled visual feature. I am not sure if > DARPA got anything close to this from its funding of various entities. What > this means is that you need a parts model. And we do that. In the case of > MNIST handwritten digits that Geoff mentions, we ?teach? this parts model > what the top part of a digit ?3? looks like, what the bottom part looks > like and so on. And we also teach connectivity between parts and the > composition of objects from parts. And we do that for all digits. And we > get a symbolic model sitting on top of a CNN model that provides the > explanation that Geoff is referring to as the causal explanation. This > ?teaching? is similar to the way you would teach a kid to recognize > different digits. > > > > An advantage of this parts model, in addition to being in an explainable > symbolic form, is robustness to adversarial attack. We recently tested on > the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast > gradient method to 27%, our XAI model maintained an accuracy of 90%, > probably higher. In general, it would be hard to make a school bus look > like an ostrich, with a few pixel changes, if you can identify the parts of > a school bus and an ostrich. > > > > A parts model that DARPA wanted provides both a symbolic explanation and > adversarial protection. The problem that Geoffrey is referring to is solved. > > > > I am doing a tutorial on this at IEEE World Congress on Computational > Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy > 18-23 July > ). > I am copying the organizers and want to thank them for accepting the > tutorial proposal. The only other presentation I have done on this is at a > Military Operations Research Society (MORS) meeting last December. > > > > So, back to the future. Hybrid models might indeed save deep learning > models and let us deploy these models without concern. We might not even > need adversarial training of any kind. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > www.teuvonet.com > > > > > [image: Timeline Description automatically generated] > > > > *From:* Connectionists *On > Behalf Of *Geoffrey Hinton > *Sent:* Friday, February 4, 2022 1:24 PM > *To:* Dietterich, Thomas > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. We can introspect on how we do it > and this may or may not give some insight into how we check our answer, but > the immediate sense that a handwritten 2 is a 2 is computed by a neural net > that is not functionally equivalent to any simple and easily explainable > procedure. > > > > This does not mean that we should give up on trying to make artificial > neural nets work more like real ones. People can see a tilted square as > either an upright diamond or a tilted square and, so far as I know, a > convnet does not exhibit this type of alternate percept. People seem to > impose hierarchical structural descriptions on images and sound waves and > they clearly impose intrinsic coordinate frames on wholes and parts. If > this is what Gary means by symbolic then I don?t disagree that neural nets > should do symbol processing. However, there is a very different meaning of > "symbolic". A pure atomic symbol has no internal structure. The form of the > symbol itself tells you nothing about what it denotes. The only relevant > properties it has are that it's identical to other instances of the > same symbol and different from all other symbols. That's totally different > from a neural net that uses embedding vectors. Embedding vectors have a > rich internal structure that dictates how they interact with other > embedding vectors. What I really object to is the following approach: Start > with pure symbols and rules for how to manipulate structures made out of > pure symbols. These structures themselves can be denoted by symbols that > correspond to memory addresses where the bits in the address tell you > nothing about the content of the structure at that address. Then when the > rule-based approach doesn't work for dealing with the real world (e.g. > machine translation) try to use neural nets to convert the real world into > pure symbols and then carry on with the rule-based approach. That is like > using an electric motor to inject the gasoline into the same old gasoline > engine instead of just replacing the gasoline engine with an electric motor. > > > > > > On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: > > ?Understanding? is not a Boolean. It is a theorem that no system can > enumerate all of the consequences of a state of affairs in the world. > > > > For low-stakes application work, we can be satisfied by a system that > ?does the right thing?. If the system draws a good picture, that?s > sufficient. It ?understood? the request. > > > > But for higher-stakes applications---and for advancing the science---we > seek a causal account of how the components of a system cause it to do the > right thing. We are hoping that a small set of mechanisms can produce broad > coverage of intelligent behavior. This gives us confidence that the system > will respond correctly outside of the narrow tasks on which we have tested > it. > > > > --Tom > > > > Thomas G. Dietterich, Distinguished Professor Emeritus > > School of Electrical Engineering and Computer > Science > > US Mail: 1148 Kelley Engineering Center > > > > Office: 2067 Kelley Engineering Center > > Oregon State Univ., Corvallis, OR 97331-5501 > > Voice: 541-737-5559; FAX: 541-737-1300 > > URL: http://web.engr.oregonstate.edu/~tgd/ > > > > > *From:* Connectionists *On > Behalf Of *Gary Marcus > *Sent:* Thursday, February 3, 2022 8:26 AM > *To:* Danko Nikolic > *Cc:* connectionists at mailman.srv.cs.cmu.edu; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > [This email originated from outside of OSU. Use caution with links and > attachments.] > > Dear Danko, > > > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, > in which he said (paraphrasing from memory, because I don?t remember the > precise words) that the famous 200 Quoc Le unsupervised model [ > https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf > ] > had learned the concept of a ca. In reality the model had clustered > together some catlike images based on the image statistics that it had > extracted, but it was a long way from a full, counterfactual-supporting > concept of a cat, much as you describe below. > > > > I fully agree with you that the reason for even having a semantics is as > you put it, "to 1) learn with a few examples and 2) apply the knowledge to > a broad set of situations.? GPT-3 sometimes gives the appearance of having > done so, but it falls apart under close inspection, so the problem remains > unsolved. > > > > Gary > > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > > > G. Hinton wrote: "I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster wearing a red hat and > it draws such a picture, it understood the request." > > > > I would like to suggest why drawing a hamster with a red hat does not > necessarily imply understanding of the statement "hamster wearing a red > hat". > > To understand that "hamster wearing a red hat" would mean inferring, in > newly emerging situations of this hamster, all the real-life > implications that the red hat brings to the little animal. > > > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > ...and so on. > > > > Countless many questions can be asked. One has understood "hamster wearing > a red hat" only if one can answer reasonably well many of such real-life > relevant questions. Similarly, a student has understood materias in a class > only if they can apply the materials in real-life situations (e.g., > applying Pythagora's theorem). If a student gives a correct answer to a > multiple choice question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > > > I also suggest that understanding also comes together with effective > learning: We store new information in such a way that we can recall it > later and use it effectively i.e., make good inferences in newly emerging > situations based on this knowledge. > > > > In short: Understanding makes us humans able to 1) learn with a few > examples and 2) apply the knowledge to a broad set of situations. > > > > No neural network today has such capabilities and we don't know how to > give them such capabilities. Neural networks need large amounts of > training examples that cover a large variety of situations and then > the networks can only deal with what the training examples have already > covered. Neural networks cannot extrapolate in that 'understanding' sense. > > > > I suggest that understanding truly extrapolates from a piece of knowledge. > It is not about satisfying a task such as translation between languages or > drawing hamsters with hats. It is how you got the capability to complete > the task: Did you only have a few examples that covered something different > but related and then you extrapolated from that knowledge? If yes, this is > going in the direction of understanding. Have you seen countless examples > and then interpolated among them? Then perhaps it is not understanding. > > > > So, for the case of drawing a hamster wearing a red hat, understanding > perhaps would have taken place if the following happened before that: > > > > 1) first, the network learned about hamsters (not many examples) > > 2) after that the network learned about red hats (outside the context of > hamsters and without many examples) > > 3) finally the network learned about drawing (outside of the context of > hats and hamsters, not many examples) > > > > After that, the network is asked to draw a hamster with a red hat. If it > does it successfully, maybe we have started cracking the problem of > understanding. > > > > Note also that this requires the network to learn sequentially without > exhibiting catastrophic forgetting of the previous knowledge, which is > possibly also a consequence of human learning by understanding. > > > > > > Danko > > > > > > > > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > Virus-free. www.avast.com > > > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > > > Without getting into the specific dispute between Gary and Geoff, I think > with approaches similar to GLOM, we are finally headed in the right > direction. There?s plenty of neurophysiological evidence for single-cell > abstractions and multisensory neurons in the brain, which one might claim > correspond to symbols. And I think we can finally reconcile the decades old > dispute between Symbolic AI and Connectionism. > > > > > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > GARY: I have *never* called for dismissal of neural networks, but rather > for some hybrid between the two (as you yourself contemplated in 1991); the > point of the 2001 book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols could complement > them. > > > > > > Asim Roy > > > Professor, Information Systems > > > Arizona State University > > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > > > > > > > *From: Connectionists On > Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: > Geoffrey Hinton Cc: AIhub ; > connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen > Hanson in conversation with Geoff Hinton > * > > > > > Dear Geoff, and interested others, > > > > > > What, for example, would you make of a system that often drew the > red-hatted hamster you requested, and perhaps a fifth of the time gave you > utter nonsense? Or say one that you trained to create birds but sometimes > output stuff like this: > > > > > > > > > > > > One could > > > > > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > > or > > > b. wonder if it might be possible that sometimes the system gets the right > answer for the wrong reasons (eg partial historical contingency), and > wonder whether another approach might be indicated. > > > > > > Benchmarks are harder than they look; most of the field has come to > recognize that. The Turing Test has turned out to be a lousy measure of > intelligence, easily gamed. It has turned out empirically that the Winograd > Schema Challenge did not measure common sense as well as Hector might have > thought. (As it happens, I am a minor coauthor of a very recent review on > this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no > way means machines now have common sense; many people from many different > perspectives recognize that (including, e.g., Yann LeCun, who generally > tends to be more aligned with you than with me). > > > > > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote > me; but what you said about me and machine translation remains your > invention, and it is inexcusable that you simply ignored my 2019 > clarification. On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains unsolved. > > > > > > All of the problems LLMs have with coherence, reliability, truthfulness, > misinformation, etc stand witness to that fact. (Their persistent inability > to filter out toxic and insulting remarks stems from the same.) I am hardly > the only person in the field to see that progress on any given benchmark > does not inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > > > > > With respect to embeddings: Embeddings are very good for natural language > *processing*; but NLP is not the same as NL*U* ? when it comes to > *understanding*, their worth is still an open question. Perhaps they will > turn out to be necessary; they clearly aren?t sufficient. In their extreme, > they might even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a specific set of > numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > > > > > (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > > > > Notably absent from your email is any kind of apology for misrepresenting > my position. It?s fine to say that ?many people thirty years ago once > thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. > I have consistently felt throughout our interactions that you have mistaken > me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me > for having made that error. I am still not he. > > > > > > Which maybe connects to the last point; if you read my work, you would see > thirty years of arguments *for* neural networks, just not in the way that > you want them to exist. I have ALWAYS argued that there is a role for them; > characterizing me as a person ?strongly opposed to neural networks? misses > the whole point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > > > > > In the last two decades or so you have insisted (for reasons you have > never fully clarified, so far as I know) on abandoning symbol-manipulation, > but the reverse is not the case: I have *never* called for dismissal of > neural networks, but rather for some hybrid between the two (as you > yourself contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded and broke down, > and where symbols could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > > > > > Gary > > > > > > > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: > > > ? > > > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > > > > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > > > > > > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > > > > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > > > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > > > > > Geoff > > > > > > > > > > > > > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural > network community, > > > > > > There has been a lot of recent discussion on this list about framing and > scientific integrity. Often the first step in restructuring narratives is > to bully and dehumanize critics. The second is to misrepresent their > position. People in positions of power are sometimes tempted to do this. > > > > > > The Hinton-Hanson interview that you just published is a real-time example > of just that. It opens with a needless and largely content-free personal > attack on a single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive thing it says is > false. > > > > > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t > be able to do machine translation.? > > > > > > I never said any such thing. > > > > > > What I predicted, rather, was that multilayer perceptrons, as they existed > then, would not (on their own, absent other mechanisms) *understand* language. > Seven years later, they still haven?t, except in the most superficial way. > > > > > > > I made no comment whatsoever about machine translation, which I view as a > separate problem, solvable to a certain degree by correspondance without > semantics. > > > > > > I specifically tried to clarify Hinton?s confusion in 2019, but, > disappointingly, he has continued to purvey misinformation despite that > clarification. Here is what I wrote privately to him then, which should > have put the matter to rest: > > > > > > You have taken a single out of context quote [from 2015] and > misrepresented it. The quote, which you have prominently displayed at the > bottom on your own web page, says: > > > > > > Hierarchies of features are less suited to challenges such as language, > inference, and high-level planning. For example, as Noam Chomsky famously > pointed out, language is filled with sentences you haven't seen > before. Pure classifier systems don't know what to do with such sentences. > The talent of feature detectors -- in identifying which member of some > category something belongs to -- doesn't translate into understanding > novel sentences, in which each sentence has its own unique meaning. > > > > > > It does *not* say "neural nets would not be able to deal with novel > sentences"; it says that hierachies of features detectors (on their own, if > you read the context of the essay) would have trouble *understanding *novel sentences. > > > > > > > Google Translate does yet not *understand* the content of the sentences > is translates. It cannot reliably answer questions about who did what to > whom, or why, it cannot infer the order of the events in paragraphs, it > can't determine the internal consistency of those events, and so forth. > > > > > > Since then, a number of scholars, such as the the computational linguist > Emily Bender, have made similar points, and indeed current LLM difficulties > with misinformation, incoherence and fabrication all follow from these > concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter > with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, > also emphasizing issues of understanding and meaning: > > > > > > *The success of the large neural language models on many NLP tasks is > exciting. However, we find that these successes sometimes lead to hype in > which these models are being described as ?understanding? language or > capturing ?meaning?. In this position paper, we argue that a system trained > only on form has a priori no way to learn meaning. .. a clear understanding > of the distinction between form and meaning will help guide the field > towards better science around natural language understanding. > * > > > > > Her later article with Gebru on language models ?stochastic parrots? is in > some ways an extension of this point; machine translation requires mimicry, > true understanding (which is what I was discussing in 2015) requires > something deeper than that. > > > > > > Hinton?s intellectual error here is in equating machine translation with > the deeper comprehension that robust natural language understanding will > require; as Bender and Koller observed, the two appear not to be the same. > (There is a longer discussion of the relation between language > understanding and machine translation, and why the latter has turned out to > be more approachable than the former, in my 2019 book with Ernest Davis). > > > > > > More broadly, Hinton?s ongoing dismissiveness of research from > perspectives other than his own (e.g. linguistics) have done the field a > disservice. > > > > > > As Herb Simon once observed, science does not have to be zero-sum. > > > > > > Sincerely, > > > Gary Marcus > > > Professor Emeritus > > > New York University > > > > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > > ? > > > Stephen Hanson in conversation with Geoff Hinton > > > > > > In the latest episode of this video series for AIhub.org, Stephen Hanson > talks to Geoff Hinton about neural networks, backpropagation, > overparameterization, digit recognition, voxel cells, syntax and semantics, > Winograd sentences, and more. > > > > > > You can watch the discussion, and read the transcript, here: > > > > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > > > > About AIhub: > > > AIhub is a non-profit dedicated to connecting the AI community to the > public by providing free, high-quality information through AIhub.org ( > https://aihub.org/). We help researchers publish the latest AI news, > summaries of their work, opinion pieces, tutorials and more. We are > supported by many leading scientific organizations in AI, namely AAAI, > NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. > > > Twitter: @aihuborg > > > > > > Virus-free. www.avast.com > > > > > > > > > -- > > Juyang (John) Weng > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: not available URL: From gary.marcus at nyu.edu Mon Feb 7 00:05:34 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Sun, 6 Feb 2022 21:05:34 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <0596DA5A-3C76-40A4-A69A-6176F4E0A35F@nyu.edu> Dear Asim, I am not so sure. Walking has some cognitive components, but there are some fairly decent solutions (eg Boston Dynamics) and they haven?t told us much all about understanding in the sense I have been describing; some aspects of mouse cognition may be informative but most probably aren?t. Not every problem that has something to do with AI actually bears on the particular problems re understanding I have emphasized. More broadly, the field of AI has had a fair amount of success on narrow (& often simple) problems but in my view that hasn?t gotten us very far towards anything that really looks much like general intelligence. We may need some new approaches, less focused on short-term progress. Gary > On Feb 6, 2022, at 12:15, Asim Roy wrote: > > ? > Dear Gary, > > I don?t disagree with you. I think it would help research if we define several small, bounded ?understanding? problems rather than this humongous one that we think the human brain handles. May be define ?understanding? in the context of a robot learning to walk. Or define ?understanding? for an artificial mouse that?s simulated. That way, we have some well-defined problems to resolve and they could serve as building blocks for larger problems. In essence, we need to solve simple problems first before we can handle the more complicated ones. > > Asim > > From: Gary Marcus > Sent: Sunday, February 6, 2022 7:42 AM > To: Asim Roy > Cc: Ali Minai ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Asim, > > Sorry for a long answer to your short but rich questions. > Yes, memory in my view has to be part of the answer to the type-token problem. Symbol systems encoded in memory allow a natural way to set up records, and something akin to that seems necessary. Pure multilayer perceptrons struggle with type-token distinctions precisely because they lack such records. On the positive side, I see more and more movement towards recordlike stores (eg w key-value stores in memory networks), and I think that is an important and necessary step, very familiar from the symbol-manipulating playbook, sometimes implemented in new ways. > But ultimately, handling the type-token distinction requires considerable inferential overhead beyond the memory representation of a record per se. How do you determine when to denote something (e.g. Felix) as an instance, and of which kinds (cat, animal etc), and how do you leverage that knowledge once you determine it? > In the limit we reason about types vs tokens in fairly subtle ways, eg in guessing whether a glass that we put down at party is likely to be ours. The reverse is also important: we need to be learn particular traits for individuals and not erroneously generalize them to the class; if my aunt Esther wins the lottery, one shouldn?t infer that all of my aunts or all of my relatives or adult females have won the lottery. so you need both representational machinery that can distinguish eg my cat from cats in general and reasoning machinery to decide at what level certain learned knowledge should inhere. (I had a whole chapter about this sort of thing in The Algebraic Mind if you are interested, and Mike Mozer had a book about types and tokens in neural networks in the mid 1990s). > Yes, part (though not all!) of what we do when we set up cognitive models in our heads is to track particular individuals and their properties. If you only had to correlate kinds (cats) and their properties (have fur) you could maybe get away with a multilayer perceptron, but once you need to track individuals, yes, you really need some kind of memory-based records. > As far as I can tell, Transformers can sometimes approximate some of this for a few sentences, but not over long stretches. > > As a small terminological aside; for me cognitive models ? cognitive modeling. Cognitive modeling is about building psychological or computational models of how people think, whereas what I mean by a cognitive model is a representation of eg the entities in some situation and the relations between those entities. > > To your closing question, none of us yet really knows how to build understanding into machines. A solid type-token distinction, both in terms of representation and reasoning, is critical for general intelligence, but hardly sufficient. Personally, I think some minimal prerequisites would be: > representations of space, time, causality, individuals, kinds, persons, places, objects, etc. > representations of abstractions that can hold over all entities in a class > compositionality (if we are talking about human-like understanding) > capacity to construct and update cognitive models on the fly > capacity to reason over entities in those models > ability to learn about new entities and their properties > Much of my last book (Rebooting AI, w Ernie Davis) is about the above list. The section in the language chapter on a children?s story in which man has lost is wallet is an especially vivid worked example. Later chapters elaborate some of the challenges in representing space, time, and causality. > > Gary > > > > On Feb 5, 2022, at 18:58, Asim Roy wrote: > > ? > Gary, > > I don?t get much into the type of cognitive modeling you are talking about, but I would guess that the type problem can generally be handled by neural network models and tokens can be resolved with some memory-based system. But to the heart of the question, this is what so-called ?understanding? reduces to computation wise? > > Asim > > From: Gary Marcus > Sent: Saturday, February 5, 2022 8:39 AM > To: Asim Roy > Cc: Ali Minai ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > There is no magic in understanding, just computation that has been realized in the wetware of humans and that eventually can be realized in machines. But understanding is not (just) learning. > > Understanding incorporates (or works in tandem with) learning - but also, critically, in tandem with inference, and the development and maintenance of cognitive models. Part of developing an understanding of cats in general is to learn long term-knowledge about their properties, both directly (e.g., through observation) and indirectly (eg through learning facts about animals in general that can be extended to cats), often through inference (if all animals have DNA, and a cat is an animal, it must also have DNA). The understanding of a particular cat also involves direct observation, but also inference (eg one might surmise that the reason that Fluffy is running about the room is that Fluffy suspects there is a mouse stirring somewhere nearby). But all of that, I would say, is subservient to the construction of cognitive models that can be routinely updated (e.g., Fluffy is currently in the living room, skittering about, perhaps looking for a mouse). > > In humans, those dynamic, relational models, which form part of an understanding, can support inference (if Fluffy is in the living room, we can infer that Fluffy is not outside, not lost, etc). Without such models - which I think represent a core part of understanding - AGI is an unlikely prospect. > > Current neural networks, as it happens, are better at acquiring long-term knowledge (cats have whiskers) than they are at dynamically updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic model that I am describing. To a modest degree they can approximate it on the basis of large samples of texts, but their ultimate incoherence stems from the fact that they do not have robust internal cognitive models that they can update on the fly. > > Without such cognitive models you can still capture some aspects of understanding (eg predicting that cats are likely to be furry), but things fall apart quickly; inference is never reliable, and coherence is fleeting. > > As a final note, one of the most foundational challenges in constructing adequate cognitive models of the world is to have a clear distinction between individuals and kinds; as I emphasized 20 years ago (in The Algebraic Mind), this has always been a weakness in neural networks, and I don?t think that the type-token problem has yet been solved. > > Gary > > > > > On Feb 5, 2022, at 01:31, Asim Roy wrote: > > ? > All, > > I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? > > https://www.youtube.com/watch?v=gn4nRCC9TwQ > https://www.youtube.com/watch?v=8sO7VS3q8d0 > > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > > > From: Ali Minai > Sent: Friday, February 4, 2022 11:38 PM > To: Asim Roy > Cc: Gary Marcus ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Asim > > Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. > > FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! > > Ali > > Ali A. Minai, Ph.D. > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > > On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: > First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. > > DANKO: > What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) > What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) > What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) > What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > From: Gary Marcus > Sent: Thursday, February 3, 2022 9:26 AM > To: Danko Nikolic > Cc: Asim Roy ; Geoffrey Hinton ; AIhub ; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Danko, > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. > > I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. > > Gary > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." > > I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". > To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. > > What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) > What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) > What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) > What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) > > ...and so on. > > Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). > > I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. > > In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. > > No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. > > I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. > > So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: > > 1) first, the network learned about hamsters (not many examples) > 2) after that the network learned about red hats (outside the context of hamsters and without many examples) > 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) > > After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. > > Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. > > > Danko > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > --- A progress usually starts with an insight --- > > > > Virus-free. www.avast.com > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) > GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > From: Connectionists On Behalf Of Gary Marcus > Sent: Wednesday, February 2, 2022 1:26 PM > To: Geoffrey Hinton > Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Geoff, and interested others, > > What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: > > > > One could > > a. avert one?s eyes and deem the anomalous outputs irrelevant > or > b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. > > Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. > > All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. > > With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) > > Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. > > Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? > > In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. > > Gary > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: > > ? > Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. > > A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. > > > But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? > > Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. > I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. > > Geoff > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, > > There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. > > The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? > > I never said any such thing. > > What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. > > I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. > > I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: > > You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: > > Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. > > It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. > > Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. > > Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: > > The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. > > Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. > > Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). > > More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. > > As Herb Simon once observed, science does not have to be zero-sum. > > Sincerely, > Gary Marcus > Professor Emeritus > New York University > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > Stephen Hanson in conversation with Geoff Hinton > > In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. > > You can watch the discussion, and read the transcript, here: > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > About AIhub: > AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. > Twitter: @aihuborg > > > Virus-free. www.avast.com > -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Sun Feb 6 22:43:08 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Mon, 7 Feb 2022 03:43:08 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear John, We recognize whole objects, but at the same time we verify its parts. Best, Asim From: Juyang Weng Sent: Sunday, February 6, 2022 8:38 PM To: Asim Roy Cc: Geoffrey Hinton ; Dietterich, Thomas ; AIhub ; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; Danko Nikolic ; Stephen Jos? Hanson ; Marek Reformat ; MARCO GORI ; Alessandro Sperduti ; Xiaodong Li ; Hava Siegelmann ; Peter Tino ; Bing Xue ; Ali Minai ; Claudius Gros ; Jean-Philippe Thivierge ; Tsvi Achler ; Prof A Hussain Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, The brain does not assume a single object in a cluttered science. Thus, a simple explanation like "recognition by parts" (but without object abstraction) should be invalid. It is like a chicken and egg problem. Both chicken and egg are absent. We must not assume egg is there or chicken is there. Best regards, -John On Sun, Feb 6, 2022 at 2:42 PM Asim Roy > wrote: Dear John, You are right and I admit I am not solving all of the problems. It?s just in reference to this one problem that Geoffrey Hinton mentions that I think can be resolved: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? Best, Asim From: Juyang Weng > Sent: Sunday, February 6, 2022 10:06 AM To: Asim Roy > Cc: Geoffrey Hinton >; Dietterich, Thomas >; AIhub >; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus >; Danko Nikolic >; Stephen Jos? Hanson >; Marek Reformat >; MARCO GORI >; Alessandro Sperduti >; Xiaodong Li >; Hava Siegelmann >; Peter Tino >; Bing Xue >; Ali Minai >; Claudius Gros >; Jean-Philippe Thivierge >; Tsvi Achler >; Prof A Hussain > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I try to be brief so that I can explain why many of us have missed, and will continue to miss, the boat. In some of my talks, I have a ppt slide "The brain is like blindmen and an elephant". Unfortunately, your "identify objects based on its parts" is a good traditional idea from pattern recognition that is still a blindman. Your idea does not explain many other problems without which we will never understand a biological brain. For example, your idea does not explain how the brain learns planning and discovery in a cluttered world. We must solve many million-dollar problems holistically. Please watch my YouTube video: Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar Problems Solved https://youtu.be/Dgx1dLCdSKY Best regards, -John On Sat, Feb 5, 2022 at 12:01 AM Asim Roy > wrote: I am responding to this part of Geoffrey Hinton?s note: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? The causal explanation is actually done quite simply, and we are doing it currently. I can talk about this now because Arizona State University (ASU) has filed a provisional patent application on the technology. The basic idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable Artificial Intelligence (darpa.mil)) and illustrated in the figure below. The idea is to identify objects based on its parts. So, the figure below says that it?s a cat because it has fur, whiskers, and claws plus an unlabeled visual feature. I am not sure if DARPA got anything close to this from its funding of various entities. What this means is that you need a parts model. And we do that. In the case of MNIST handwritten digits that Geoff mentions, we ?teach? this parts model what the top part of a digit ?3? looks like, what the bottom part looks like and so on. And we also teach connectivity between parts and the composition of objects from parts. And we do that for all digits. And we get a symbolic model sitting on top of a CNN model that provides the explanation that Geoff is referring to as the causal explanation. This ?teaching? is similar to the way you would teach a kid to recognize different digits. An advantage of this parts model, in addition to being in an explainable symbolic form, is robustness to adversarial attack. We recently tested on the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast gradient method to 27%, our XAI model maintained an accuracy of 90%, probably higher. In general, it would be hard to make a school bus look like an ostrich, with a few pixel changes, if you can identify the parts of a school bus and an ostrich. A parts model that DARPA wanted provides both a symbolic explanation and adversarial protection. The problem that Geoffrey is referring to is solved. I am doing a tutorial on this at IEEE World Congress on Computational Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy 18-23 July). I am copying the organizers and want to thank them for accepting the tutorial proposal. The only other presentation I have done on this is at a Military Operations Research Society (MORS) meeting last December. So, back to the future. Hybrid models might indeed save deep learning models and let us deploy these models without concern. We might not even need adversarial training of any kind. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) www.teuvonet.com [Timeline Description automatically generated] From: Connectionists > On Behalf Of Geoffrey Hinton Sent: Friday, February 4, 2022 1:24 PM To: Dietterich, Thomas > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists > On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic > Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg Virus-free. www.avast.com -- Juyang (John) Weng -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: From ASIM.ROY at asu.edu Sat Feb 5 00:01:07 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sat, 5 Feb 2022 05:01:07 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: I am responding to this part of Geoffrey Hinton?s note: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? The causal explanation is actually done quite simply, and we are doing it currently. I can talk about this now because Arizona State University (ASU) has filed a provisional patent application on the technology. The basic idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable Artificial Intelligence (darpa.mil)) and illustrated in the figure below. The idea is to identify objects based on its parts. So, the figure below says that it?s a cat because it has fur, whiskers, and claws plus an unlabeled visual feature. I am not sure if DARPA got anything close to this from its funding of various entities. What this means is that you need a parts model. And we do that. In the case of MNIST handwritten digits that Geoff mentions, we ?teach? this parts model what the top part of a digit ?3? looks like, what the bottom part looks like and so on. And we also teach connectivity between parts and the composition of objects from parts. And we do that for all digits. And we get a symbolic model sitting on top of a CNN model that provides the explanation that Geoff is referring to as the causal explanation. This ?teaching? is similar to the way you would teach a kid to recognize different digits. An advantage of this parts model, in addition to being in an explainable symbolic form, is robustness to adversarial attack. We recently tested on the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast gradient method to 27%, our XAI model maintained an accuracy of 90%, probably higher. In general, it would be hard to make a school bus look like an ostrich, with a few pixel changes, if you can identify the parts of a school bus and an ostrich. A parts model that DARPA wanted provides both a symbolic explanation and adversarial protection. The problem that Geoffrey is referring to is solved. I am doing a tutorial on this at IEEE World Congress on Computational Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy 18-23 July). I am copying the organizers and want to thank them for accepting the tutorial proposal. The only other presentation I have done on this is at a Military Operations Research Society (MORS) meeting last December. So, back to the future. Hybrid models might indeed save deep learning models and let us deploy these models without concern. We might not even need adversarial training of any kind. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) www.teuvonet.com [Timeline Description automatically generated] From: Connectionists On Behalf Of Geoffrey Hinton Sent: Friday, February 4, 2022 1:24 PM To: Dietterich, Thomas Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists > On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic > Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg Virus-free. www.avast.com -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: From gary.marcus at nyu.edu Sun Feb 6 09:41:59 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Sun, 6 Feb 2022 06:41:59 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <981624DC-3FA6-440E-910F-055B1FB77C41@nyu.edu> Dear Asim, Sorry for a long answer to your short but rich questions. Yes, memory in my view has to be part of the answer to the type-token problem. Symbol systems encoded in memory allow a natural way to set up records, and something akin to that seems necessary. Pure multilayer perceptrons struggle with type-token distinctions precisely because they lack such records. On the positive side, I see more and more movement towards recordlike stores (eg w key-value stores in memory networks), and I think that is an important and necessary step, very familiar from the symbol-manipulating playbook, sometimes implemented in new ways. But ultimately, handling the type-token distinction requires considerable inferential overhead beyond the memory representation of a record per se. How do you determine when to denote something (e.g. Felix) as an instance, and of which kinds (cat, animal etc), and how do you leverage that knowledge once you determine it? In the limit we reason about types vs tokens in fairly subtle ways, eg in guessing whether a glass that we put down at party is likely to be ours. The reverse is also important: we need to be learn particular traits for individuals and not erroneously generalize them to the class; if my aunt Esther wins the lottery, one shouldn?t infer that all of my aunts or all of my relatives or adult females have won the lottery. so you need both representational machinery that can distinguish eg my cat from cats in general and reasoning machinery to decide at what level certain learned knowledge should inhere. (I had a whole chapter about this sort of thing in The Algebraic Mind if you are interested, and Mike Mozer had a book about types and tokens in neural networks in the mid 1990s). Yes, part (though not all!) of what we do when we set up cognitive models in our heads is to track particular individuals and their properties. If you only had to correlate kinds (cats) and their properties (have fur) you could maybe get away with a multilayer perceptron, but once you need to track individuals, yes, you really need some kind of memory-based records. As far as I can tell, Transformers can sometimes approximate some of this for a few sentences, but not over long stretches. As a small terminological aside; for me cognitive models ? cognitive modeling. Cognitive modeling is about building psychological or computational models of how people think, whereas what I mean by a cognitive model is a representation of eg the entities in some situation and the relations between those entities. To your closing question, none of us yet really knows how to build understanding into machines. A solid type-token distinction, both in terms of representation and reasoning, is critical for general intelligence, but hardly sufficient. Personally, I think some minimal prerequisites would be: representations of space, time, causality, individuals, kinds, persons, places, objects, etc. representations of abstractions that can hold over all entities in a class compositionality (if we are talking about human-like understanding) capacity to construct and update cognitive models on the fly capacity to reason over entities in those models ability to learn about new entities and their properties Much of my last book (Rebooting AI, w Ernie Davis) is about the above list. The section in the language chapter on a children?s story in which man has lost is wallet is an especially vivid worked example. Later chapters elaborate some of the challenges in representing space, time, and causality. Gary > On Feb 5, 2022, at 18:58, Asim Roy wrote: > > ? > Gary, > > I don?t get much into the type of cognitive modeling you are talking about, but I would guess that the type problem can generally be handled by neural network models and tokens can be resolved with some memory-based system. But to the heart of the question, this is what so-called ?understanding? reduces to computation wise? > > Asim > > From: Gary Marcus > Sent: Saturday, February 5, 2022 8:39 AM > To: Asim Roy > Cc: Ali Minai ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > There is no magic in understanding, just computation that has been realized in the wetware of humans and that eventually can be realized in machines. But understanding is not (just) learning. > > Understanding incorporates (or works in tandem with) learning - but also, critically, in tandem with inference, and the development and maintenance of cognitive models. Part of developing an understanding of cats in general is to learn long term-knowledge about their properties, both directly (e.g., through observation) and indirectly (eg through learning facts about animals in general that can be extended to cats), often through inference (if all animals have DNA, and a cat is an animal, it must also have DNA). The understanding of a particular cat also involves direct observation, but also inference (eg one might surmise that the reason that Fluffy is running about the room is that Fluffy suspects there is a mouse stirring somewhere nearby). But all of that, I would say, is subservient to the construction of cognitive models that can be routinely updated (e.g., Fluffy is currently in the living room, skittering about, perhaps looking for a mouse). > > In humans, those dynamic, relational models, which form part of an understanding, can support inference (if Fluffy is in the living room, we can infer that Fluffy is not outside, not lost, etc). Without such models - which I think represent a core part of understanding - AGI is an unlikely prospect. > > Current neural networks, as it happens, are better at acquiring long-term knowledge (cats have whiskers) than they are at dynamically updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic model that I am describing. To a modest degree they can approximate it on the basis of large samples of texts, but their ultimate incoherence stems from the fact that they do not have robust internal cognitive models that they can update on the fly. > > Without such cognitive models you can still capture some aspects of understanding (eg predicting that cats are likely to be furry), but things fall apart quickly; inference is never reliable, and coherence is fleeting. > > As a final note, one of the most foundational challenges in constructing adequate cognitive models of the world is to have a clear distinction between individuals and kinds; as I emphasized 20 years ago (in The Algebraic Mind), this has always been a weakness in neural networks, and I don?t think that the type-token problem has yet been solved. > > Gary > > > > On Feb 5, 2022, at 01:31, Asim Roy wrote: > > ? > All, > > I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? > > https://www.youtube.com/watch?v=gn4nRCC9TwQ > https://www.youtube.com/watch?v=8sO7VS3q8d0 > > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > > > From: Ali Minai > Sent: Friday, February 4, 2022 11:38 PM > To: Asim Roy > Cc: Gary Marcus ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Asim > > Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. > > FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! > > Ali > > Ali A. Minai, Ph.D. > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > > On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: > First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. > > DANKO: > What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) > What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) > What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) > What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > From: Gary Marcus > Sent: Thursday, February 3, 2022 9:26 AM > To: Danko Nikolic > Cc: Asim Roy ; Geoffrey Hinton ; AIhub ; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Danko, > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. > > I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. > > Gary > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." > > I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". > To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. > > What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) > What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) > What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) > What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) > > ...and so on. > > Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). > > I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. > > In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. > > No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. > > I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. > > So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: > > 1) first, the network learned about hamsters (not many examples) > 2) after that the network learned about red hats (outside the context of hamsters and without many examples) > 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) > > After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. > > Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. > > > Danko > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > --- A progress usually starts with an insight --- > > > > Virus-free. www.avast.com > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) > GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. > > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > From: Connectionists On Behalf Of Gary Marcus > Sent: Wednesday, February 2, 2022 1:26 PM > To: Geoffrey Hinton > Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Geoff, and interested others, > > What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: > > > > One could > > a. avert one?s eyes and deem the anomalous outputs irrelevant > or > b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. > > Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. > > All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. > > With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) > > Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. > > Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? > > In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. > > Gary > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: > > ? > Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. > > A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. > > > But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? > > Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. > I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. > > Geoff > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, > > There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. > > The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? > > I never said any such thing. > > What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. > > I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. > > I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: > > You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: > > Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. > > It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. > > Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. > > Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: > > The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. > > Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. > > Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). > > More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. > > As Herb Simon once observed, science does not have to be zero-sum. > > Sincerely, > Gary Marcus > Professor Emeritus > New York University > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > ? > Stephen Hanson in conversation with Geoff Hinton > > In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. > > You can watch the discussion, and read the transcript, here: > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > About AIhub: > AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. > Twitter: @aihuborg > > > Virus-free. www.avast.com > -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Sun Feb 6 11:47:13 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Sun, 6 Feb 2022 11:47:13 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Dear Geoff Hinton, I respect that you have been working on pattern recognition on isolated characters using neural networks. However, I am deeply disappointed that after receiving the Turing Award 2018, you are still falling behind your own award work by talking about "how you recognize that a handwritten 2 is a 2." You have fallen behind our group's Creceptron work in 1992, let alone our group's work on 3D-to-2D-to-3D Conscious Learning using DNs. Both deal with cluttered scenes. Specifically, you will never be able to get a correct causal explanation by looking at a single hand-written 2. Your problem is too small to explain a brain network. You must look at cluttered sciences, with many objects. Yours humbly, -John ---- Message: 7 Date: Fri, 4 Feb 2022 15:24:02 -0500 From: Geoffrey Hinton To: "Dietterich, Thomas" Cc: AIhub , "connectionists at mailman.srv.cs.cmu.edu" Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Content-Type: text/plain; charset="utf-8" I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Sun Feb 6 14:41:59 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sun, 6 Feb 2022 19:41:59 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear John, You are right and I admit I am not solving all of the problems. It?s just in reference to this one problem that Geoffrey Hinton mentions that I think can be resolved: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? Best, Asim From: Juyang Weng Sent: Sunday, February 6, 2022 10:06 AM To: Asim Roy Cc: Geoffrey Hinton ; Dietterich, Thomas ; AIhub ; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; Danko Nikolic ; Stephen Jos? Hanson ; Marek Reformat ; MARCO GORI ; Alessandro Sperduti ; Xiaodong Li ; Hava Siegelmann ; Peter Tino ; Bing Xue ; Ali Minai ; Claudius Gros ; Jean-Philippe Thivierge ; Tsvi Achler ; Prof A Hussain Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I try to be brief so that I can explain why many of us have missed, and will continue to miss, the boat. In some of my talks, I have a ppt slide "The brain is like blindmen and an elephant". Unfortunately, your "identify objects based on its parts" is a good traditional idea from pattern recognition that is still a blindman. Your idea does not explain many other problems without which we will never understand a biological brain. For example, your idea does not explain how the brain learns planning and discovery in a cluttered world. We must solve many million-dollar problems holistically. Please watch my YouTube video: Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar Problems Solved https://youtu.be/Dgx1dLCdSKY Best regards, -John On Sat, Feb 5, 2022 at 12:01 AM Asim Roy > wrote: I am responding to this part of Geoffrey Hinton?s note: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? The causal explanation is actually done quite simply, and we are doing it currently. I can talk about this now because Arizona State University (ASU) has filed a provisional patent application on the technology. The basic idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable Artificial Intelligence (darpa.mil)) and illustrated in the figure below. The idea is to identify objects based on its parts. So, the figure below says that it?s a cat because it has fur, whiskers, and claws plus an unlabeled visual feature. I am not sure if DARPA got anything close to this from its funding of various entities. What this means is that you need a parts model. And we do that. In the case of MNIST handwritten digits that Geoff mentions, we ?teach? this parts model what the top part of a digit ?3? looks like, what the bottom part looks like and so on. And we also teach connectivity between parts and the composition of objects from parts. And we do that for all digits. And we get a symbolic model sitting on top of a CNN model that provides the explanation that Geoff is referring to as the causal explanation. This ?teaching? is similar to the way you would teach a kid to recognize different digits. An advantage of this parts model, in addition to being in an explainable symbolic form, is robustness to adversarial attack. We recently tested on the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast gradient method to 27%, our XAI model maintained an accuracy of 90%, probably higher. In general, it would be hard to make a school bus look like an ostrich, with a few pixel changes, if you can identify the parts of a school bus and an ostrich. A parts model that DARPA wanted provides both a symbolic explanation and adversarial protection. The problem that Geoffrey is referring to is solved. I am doing a tutorial on this at IEEE World Congress on Computational Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy 18-23 July). I am copying the organizers and want to thank them for accepting the tutorial proposal. The only other presentation I have done on this is at a Military Operations Research Society (MORS) meeting last December. So, back to the future. Hybrid models might indeed save deep learning models and let us deploy these models without concern. We might not even need adversarial training of any kind. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) www.teuvonet.com [Timeline Description automatically generated] From: Connectionists > On Behalf Of Geoffrey Hinton Sent: Friday, February 4, 2022 1:24 PM To: Dietterich, Thomas > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists > On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic > Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg Virus-free. www.avast.com -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: From julia.verhoef at donders.ru.nl Mon Feb 7 00:56:08 2022 From: julia.verhoef at donders.ru.nl (Verhoef, J.P. (Julia)) Date: Mon, 7 Feb 2022 05:56:08 +0000 Subject: Connectionists: Postdoctoral Researcher for Dutch Research Consortium 'Language in Interaction' (Senior) Message-ID: <66cb039fb8bf4695a392656ade23548c@EXPRD08.hosting.ru.nl> Dear all, We would like to draw your attention to the vacancy available within the Language in Interaction consortium: Postdoctoral researcher for Dutch Research Consortium 'Language in Interaction' (Senior) Dutch Research Consortium 'Language in Interaction' Maximum salary: ? 5230 gross/month Vacancy number: 25.01.2022 Application deadline: March 1, 2022, 23:59.00 (CET) [Logo NWO] [Logo] Responsibilities The Language in Interaction research consortium invites applications for two senior postdoctoral positions. We are looking for highly motivated candidates to enrich a unique consortium of researchers that aims to unravel the neurocognitive mechanisms of language at multiple levels. The goal is to understand both the universality and the variability of the human language faculty from genes to behavior. You will contribute to the integration of empirical research in our consortium and will aid in the coordination of ?Theory of Language? meetings in which cutting edge research is discussed by consortium researchers. You will act in close collaboration with Peter Hagoort, programme director of the consortium. These positions provide the opportunity for conducting world-class research as a member of an interdisciplinary team. Moreover, it will provide the opportunity to contribute to developing a theoretical framework for our understanding of the human language faculty. Work environment The Netherlands has an outstanding track record in the language sciences. The research consortium ?Language in Interaction?, sponsored by a large grant from the Netherlands Organization for Scientific research (NWO), brings together many of the excellent research groups in the Netherlands with a research program on the foundations of language. In addition to excellence in the domain of language and related relevant fields of cognition, our consortium provides state-of-the-art research facilities and a research team with ample experience in the complex research methods that will be invoked to address the scientific questions at the highest level of methodological sophistication. These include methods from genetics, neuroimaging, computational modelling, and patient-related research. This consortium realizes both quality and critical mass for studying human language at a scale not easily found anywhere else. We have identified five Big Questions (BQ) that are central to our understanding of the human language faculty. These questions are interrelated at multiple levels. Teams of researchers will collaborate to collectively address these key questions of our field. Our five Big Questions are: BQ1: The nature of the mental lexicon: How to bridge neurobiology and psycholinguistic theory by computational modelling? BQ2: What are the characteristics and consequences of internal brain organization for language? BQ3: Creating a shared cognitive space: How is language grounded in and shaped by communicative settings of interacting people? BQ4: Variability in language processing and in language learning: Why does the ability to learn language change with age? How can we characterize and map individual language skills in relation to the population distribution? BQ5: How are other cognitive systems shaped by the presence of a language system in humans? You will be appointed at one of two participating institutes: The Max Planck Institute for Psycholinguistics in Nijmegen or the Donders Institute for Brain, Cognition and Behaviour (Donders Centre for Cognitive Neuroimaging), also located in Nijmegen. The research is conducted in an international setting at all participating institutions. English is the lingua franca. What we expect from you ? a PhD in an area related to the neurobiology of language and/or language sciences; ? expertise/interest in theoretical neuroscience and language; ? expertise / interest in theoretical neuroscience and language; ? an integrative mindset; ? a theory-driven approach; ? good communication skills; ? excellent proficiency in written and spoken English. What we have to offer The positions will be held at one of the participating institutes (Max Planck Institute Nijmegen or Donders Institute for Brain, Cognition and Behavior at the Donders Centre for Cognitive Neuroimaging). ? Employment: 1,0 fte; ? Employment at the Max Planck Institute or Donders Institute at the Radboud University; ? Starting date June 30th 2022 for a period of maximally 24 months; ? Starting salary will be based on previous work experience; ? Gross salary at the Max Planck Institute will be between ? 4187 and ? 4911 per month; ? Gross salary at the Radboud University will be between ? 3821 and ? 5230 per month (salary scale 11); ? In addition to the salary: an 8% holiday allowance; ? Each participating institute has a number of regulations that make it possible for employees to create a good work-life balance. Other Information The participating institutes are an equal opportunity employer, committed to building a culturally diverse intellectual community, and as such encourage applications from women and minorities. Would you like to know more? Further information on the Language in Interaction Consortium. For more information about this vacancy, please contact: Prof. dr. Peter Hagoort, programme director Language in Interaction E-mail: peter.hagoort at donders.ru.nl Tel: ++ 31-24-3610648/651 Fax: ++ 31-24-3610652 Are you interested? Applications can be submitted to Julia Verhoef (secretary of the Language in Interaction consortium; Julia.Verhoef at donders.ru.nl) in electronic form. Your application should include (and be limited to) the following attachments: ? a cover letter; ? your curriculum vitae, including, if applicable, a list of publications and the names of at least two persons who can provide references. Please apply before March 1, 2022, 23:59.00 (CET). Kind regards, Julia Verhoef Secretary - Language in Interaction Consortium Radboud University | Donders Centre for Cognitive Neuroimaging (DCCN) | room 0.026 Kapittelweg 29, 6525 EN Nijmegen, The Netherlands | P.O. Box 9101, 6500 HB, Nijmegen, The Netherlands | T: +31 (0)24 3666272 | E: J.Verhoef at donders.ru.nl|Office hours: 9-14 hr on Mon - Fri -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.jpg Type: image/jpeg Size: 2461 bytes Desc: image001.jpg URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image002.jpg Type: image/jpeg Size: 40202 bytes Desc: image002.jpg URL: From gary.marcus at nyu.edu Mon Feb 7 00:28:32 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Sun, 6 Feb 2022 21:28:32 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <148D10E4-0C1F-4FA4-B8AA-C80E1472A63A@nyu.edu> Ali, It?s useful to think about animals, but I really wouldn?t start with fish; it?s not clear that their ecological niche demands anything significant in the way of extrapolation, causal reasoning, or compositionality. There is good evidence elsewhere in the animal world for extrapolation of functions that may be innate (eg solar azimuth in bees), and causal reasoning (eg tool use in ravens, various primates, and octopi). It?s still not clear to me how much hierarchical representation (critical to AGI) exists outside of humans, though; the ability to construct rich new cognitive models may also be unique to us. In any case it matters not in the least whether the average cat or human cares about symbols, anymore that it matters whether the average animal understands digestion; only a tiny fraction of the creatures on this planet have any real understanding of their internal workings. My overall feeling is that we are a really, really long way from understanding the neural basis of higher-level cognition, and that AI is going to need muddle through on its own, for another decade or two, I do fully agree with your conclusion, though, that "AI today is driven more by habit and the incentives of the academic and corporate marketplaces than by a deep, long-term view of AI as a great exploratory project in fundamental science." Let's hope that changes. Gary > On Feb 6, 2022, at 13:19, Ali Minai wrote: > > ? > Gary, > > That?s a very interesting and accurate list of capabilities that a general intelligent system must have and that our AI does not. Of course, the list is familiar to me from having read your book. However, I have a somewhat different take on this whole thing. > > All the things we discuss here ? symbols/no symbols, parts/wholes, supervised/unsupervised, token/type, etc., are useful categories and distinctions for our analysis of the problem, and are partly a result of the historical evolution of the field of AI in particular and of philosophy in general. The categories are not wrong in any way, of course, but they are posterior to the actual system ? good for describing and analyzing it, and for validating our versions of it (which is how you use them). I think they are less useful as prescriptions for how to build our AI systems. If intelligent systems did not already exist and we were building them from scratch (please ignore the impossibility of that), having a list of ?must haves? would be great. But intelligent systems already exist ? from humans to fish ? and they already have these capacities to a greater or lesser degree because of the physics of their biology. A cat?s intelligence does not care whether it has symbols or not, and nor does mine or yours. Whatever we describe as symbolic processing post-facto has already been done by brains for at least tens of millions of years. Instead of getting caught up in ?how to add symbols into our neural models?, we should be investigating how what we see as symbolic processing emerges from animal brains, and then replicate those brains to the degree necessary. If we can do that, symbolic processing will already be present. But it cannot be done piece by piece. It must take the integrity of the whole brain and the body it is part of, and its environment, into account. That?s why I think that a much better ? though a very long ? route to AI is to start by understanding how a fish brain makes the intelligence of a fish possible, and then boot up our knowledge across phylogenetic stages: Bottom up reverse engineering rather than top-down engineering. That?s the way Nature built up to human intelligence, and we will succeed only by reverse engineering it. Of course, we can do it much faster and with shortcuts because we are intelligent, purposive agents, but working top-down by building piecewise systems that satisfy a list of attributes will not get us there. Among other things, those pieces will be impossible to integrate into the kind of intelligence that can have those general models of the world that you rightly point to as being necessary. > > I think that one thing that has been a great boon to the AI enterprise has also been one of the greatest impediments to its complete success, and that is the ?computationalization? of intelligence. On the one hand, thinking of intelligence computationally allows us to describe it abstractly and in a principled, formal way. It also resonates with the fact that we are trying to implement intelligence through computational machines. But, on the flip side, this view of intelligence divorces it from its physics ? from the fact that real intelligence in animals emerges from the physics of the physical system. That system is not a collection of its capabilities; rather, those capabilities are immanent in it by virtue of its physics. When we try to build those capabilities computationally, i.e., through code, we are making the same error that the practitioners of old-style ?symbolic AI? made ? what I call the ?professors are smarter than Nature? error, i.e., the idea that we are going to enumerate (or describe) all the things that underlie intelligence and implement them one by one until we get complete intelligence. We will never be able to enumerate all those capabilities, and will never be able to get to that complete intelligence. The only difference between us and the ?symbolists? of yore is that we are replacing giant LISP and Prolog programs with giant neural networks. Otherwise, we are using our models exactly as they were trying to use their models, and we will fail just as they did unless we get back to biology and the real thing. > > I will say again that the way we do AI today is driven more by habit and the incentives of the academic and corporate marketplaces than by a deep, long-term view of AI as a great exploratory project in fundamental science. We are just building AI to drive our cars, translate our documents, write our reports, and do our shopping. What that will teach us about actual intelligence is just incidental. > > My apologies too for a long response. > > Ali > > Ali A. Minai, Ph.D. > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > >> On Sun, Feb 6, 2022 at 9:42 AM Gary Marcus wrote: >> Dear Asim, >> >> Sorry for a long answer to your short but rich questions. >> Yes, memory in my view has to be part of the answer to the type-token problem. Symbol systems encoded in memory allow a natural way to set up records, and something akin to that seems necessary. Pure multilayer perceptrons struggle with type-token distinctions precisely because they lack such records. On the positive side, I see more and more movement towards recordlike stores (eg w key-value stores in memory networks), and I think that is an important and necessary step, very familiar from the symbol-manipulating playbook, sometimes implemented in new ways. >> But ultimately, handling the type-token distinction requires considerable inferential overhead beyond the memory representation of a record per se. How do you determine when to denote something (e.g. Felix) as an instance, and of which kinds (cat, animal etc), and how do you leverage that knowledge once you determine it? >> In the limit we reason about types vs tokens in fairly subtle ways, eg in guessing whether a glass that we put down at party is likely to be ours. The reverse is also important: we need to be learn particular traits for individuals and not erroneously generalize them to the class; if my aunt Esther wins the lottery, one shouldn?t infer that all of my aunts or all of my relatives or adult females have won the lottery. so you need both representational machinery that can distinguish eg my cat from cats in general and reasoning machinery to decide at what level certain learned knowledge should inhere. (I had a whole chapter about this sort of thing in The Algebraic Mind if you are interested, and Mike Mozer had a book about types and tokens in neural networks in the mid 1990s). >> Yes, part (though not all!) of what we do when we set up cognitive models in our heads is to track particular individuals and their properties. If you only had to correlate kinds (cats) and their properties (have fur) you could maybe get away with a multilayer perceptron, but once you need to track individuals, yes, you really need some kind of memory-based records. >> As far as I can tell, Transformers can sometimes approximate some of this for a few sentences, but not over long stretches. >> >> As a small terminological aside; for me cognitive models ? cognitive modeling. Cognitive modeling is about building psychological or computational models of how people think, whereas what I mean by a cognitive model is a representation of eg the entities in some situation and the relations between those entities. >> >> To your closing question, none of us yet really knows how to build understanding into machines. A solid type-token distinction, both in terms of representation and reasoning, is critical for general intelligence, but hardly sufficient. Personally, I think some minimal prerequisites would be: >> representations of space, time, causality, individuals, kinds, persons, places, objects, etc. >> representations of abstractions that can hold over all entities in a class >> compositionality (if we are talking about human-like understanding) >> capacity to construct and update cognitive models on the fly >> capacity to reason over entities in those models >> ability to learn about new entities and their properties >> Much of my last book (Rebooting AI, w Ernie Davis) is about the above list. The section in the language chapter on a children?s story in which man has lost is wallet is an especially vivid worked example. Later chapters elaborate some of the challenges in representing space, time, and causality. >> >> Gary >> >> >>>> On Feb 5, 2022, at 18:58, Asim Roy wrote: >>>> >>> ? >>> Gary, >>> >>> >>> >>> I don?t get much into the type of cognitive modeling you are talking about, but I would guess that the type problem can generally be handled by neural network models and tokens can be resolved with some memory-based system. But to the heart of the question, this is what so-called ?understanding? reduces to computation wise? >>> >>> >>> >>> Asim >>> >>> >>> >>> From: Gary Marcus >>> Sent: Saturday, February 5, 2022 8:39 AM >>> To: Asim Roy >>> Cc: Ali Minai ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> There is no magic in understanding, just computation that has been realized in the wetware of humans and that eventually can be realized in machines. But understanding is not (just) learning. >>> >>> >>> >>> Understanding incorporates (or works in tandem with) learning - but also, critically, in tandem with inference, and the development and maintenance of cognitive models. Part of developing an understanding of cats in general is to learn long term-knowledge about their properties, both directly (e.g., through observation) and indirectly (eg through learning facts about animals in general that can be extended to cats), often through inference (if all animals have DNA, and a cat is an animal, it must also have DNA). The understanding of a particular cat also involves direct observation, but also inference (eg one might surmise that the reason that Fluffy is running about the room is that Fluffy suspects there is a mouse stirring somewhere nearby). But all of that, I would say, is subservient to the construction of cognitive models that can be routinely updated (e.g., Fluffy is currently in the living room, skittering about, perhaps looking for a mouse). >>> >>> In humans, those dynamic, relational models, which form part of an understanding, can support inference (if Fluffy is in the living room, we can infer that Fluffy is not outside, not lost, etc). Without such models - which I think represent a core part of understanding - AGI is an unlikely prospect. >>> >>> Current neural networks, as it happens, are better at acquiring long-term knowledge (cats have whiskers) than they are at dynamically updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic model that I am describing. To a modest degree they can approximate it on the basis of large samples of texts, but their ultimate incoherence stems from the fact that they do not have robust internal cognitive models that they can update on the fly. >>> >>> Without such cognitive models you can still capture some aspects of understanding (eg predicting that cats are likely to be furry), but things fall apart quickly; inference is never reliable, and coherence is fleeting. >>> >>> As a final note, one of the most foundational challenges in constructing adequate cognitive models of the world is to have a clear distinction between individuals and kinds; as I emphasized 20 years ago (in The Algebraic Mind), this has always been a weakness in neural networks, and I don?t think that the type-token problem has yet been solved. >>> >>> >>> Gary >>> >>> >>> >>> >>> >>> >>> On Feb 5, 2022, at 01:31, Asim Roy wrote: >>> >>> ? >>> >>> All, >>> >>> >>> >>> I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? >>> >>> >>> >>> https://www.youtube.com/watch?v=gn4nRCC9TwQ >>> >>> https://www.youtube.com/watch?v=8sO7VS3q8d0 >>> >>> >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> From: Ali Minai >>> Sent: Friday, February 4, 2022 11:38 PM >>> To: Asim Roy >>> Cc: Gary Marcus ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> Asim >>> >>> >>> >>> Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. >>> >>> >>> >>> FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! >>> >>> >>> >>> Ali >>> >>> >>> >>> Ali A. Minai, Ph.D. >>> Professor and Graduate Program Director >>> Complex Adaptive Systems Lab >>> Department of Electrical Engineering & Computer Science >>> >>> 828 Rhodes Hall >>> >>> University of Cincinnati >>> Cincinnati, OH 45221-0030 >>> >>> >>> Phone: (513) 556-4783 >>> Fax: (513) 556-7326 >>> Email: Ali.Minai at uc.edu >>> minaiaa at gmail.com >>> >>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>> >>> >>> >>> >>> >>> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: >>> >>> First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. >>> >>> >>> >>> DANKO: >>> >>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> From: Gary Marcus >>> Sent: Thursday, February 3, 2022 9:26 AM >>> To: Danko Nikolic >>> Cc: Asim Roy ; Geoffrey Hinton ; AIhub ; connectionists at mailman.srv.cs.cmu.edu >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> Dear Danko, >>> >>> >>> >>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. >>> >>> >>> >>> I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. >>> >>> >>> >>> Gary >>> >>> >>> >>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: >>> >>> >>> >>> G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." >>> >>> >>> >>> I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". >>> >>> To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. >>> >>> >>> >>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>> >>> >>> >>> ...and so on. >>> >>> >>> >>> Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). >>> >>> >>> >>> I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. >>> >>> >>> >>> In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. >>> >>> >>> >>> No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. >>> >>> >>> >>> I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. >>> >>> >>> >>> So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: >>> >>> >>> >>> 1) first, the network learned about hamsters (not many examples) >>> >>> 2) after that the network learned about red hats (outside the context of hamsters and without many examples) >>> >>> 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) >>> >>> >>> >>> After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. >>> >>> >>> >>> Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. >>> >>> >>> >>> >>> >>> Danko >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> --- A progress usually starts with an insight --- >>> >>> >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>> >>> Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. >>> >>> >>> >>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>> >>> GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> From: Connectionists On Behalf Of Gary Marcus >>> Sent: Wednesday, February 2, 2022 1:26 PM >>> To: Geoffrey Hinton >>> Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> Dear Geoff, and interested others, >>> >>> >>> >>> What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: >>> >>> >>> >>> >>> >>> >>> >>> One could >>> >>> >>> >>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>> >>> or >>> >>> b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. >>> >>> >>> >>> Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). >>> >>> >>> >>> So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. >>> >>> >>> >>> All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. >>> >>> >>> >>> With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) >>> >>> >>> >>> (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>> >>> >>> >>> Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. >>> >>> >>> >>> Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? >>> >>> >>> >>> In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. >>> >>> >>> >>> Gary >>> >>> >>> >>> >>> >>> On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: >>> >>> ? >>> >>> Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. >>> >>> >>> >>> A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. >>> >>> >>> >>> But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? >>> >>> >>> >>> Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. >>> >>> I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. >>> >>> >>> >>> Geoff >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>> >>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, >>> >>> >>> >>> There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. >>> >>> >>> >>> The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. >>> >>> >>> >>> Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? >>> >>> >>> >>> I never said any such thing. >>> >>> >>> >>> What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. >>> >>> >>> >>> I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. >>> >>> >>> >>> I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: >>> >>> >>> >>> You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: >>> >>> >>> >>> Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. >>> >>> >>> >>> It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. >>> >>> >>> >>> Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. >>> >>> >>> >>> Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: >>> >>> >>> >>> The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. >>> >>> >>> >>> Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. >>> >>> >>> >>> Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). >>> >>> >>> >>> More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. >>> >>> >>> >>> As Herb Simon once observed, science does not have to be zero-sum. >>> >>> >>> >>> Sincerely, >>> >>> Gary Marcus >>> >>> Professor Emeritus >>> >>> New York University >>> >>> >>> >>> On Feb 2, 2022, at 06:12, AIhub wrote: >>> >>> ? >>> >>> Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. >>> >>> >>> >>> You can watch the discussion, and read the transcript, here: >>> >>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>> >>> >>> >>> About AIhub: >>> >>> AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. >>> >>> Twitter: @aihuborg >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> -------------- next part -------------- An HTML attachment was scrubbed... URL: From m.biehl at rug.nl Mon Feb 7 02:54:09 2022 From: m.biehl at rug.nl (Michael Biehl) Date: Mon, 7 Feb 2022 08:54:09 +0100 Subject: Connectionists: Fully funded PhD position (4 years) Message-ID: Apologies for multiple postings *Three weeks left to apply! * A *fully funded PhD position* (4 years) in the *Statistical **Physics of Neural Networks* is available at the University of Groningen, The Netherlands, see https://www.rug.nl/about-ug/work-with-us/job-opportunities/?details=00347-02S0008WFP for details and application details. Applications (before March 1) are only possible through this webpage. The title of the project is "The role of the activation function for feedforward learning systems (RAFFLES)". For further information please contact Michael Biehl. ---------------------------------------------------------- Prof. Dr. Michael Biehl Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence P.O. Box 407, 9700 AK Groningen The Netherlands Tel. +31 50 363 3997 https://www.cs.rug.nl/~biehl m.biehl at rug.nl -------------- next part -------------- An HTML attachment was scrubbed... URL: From eduardo.lopez at bitbrain.es Mon Feb 7 03:57:00 2022 From: eduardo.lopez at bitbrain.es (=?UTF-8?Q?Eduardo_L=C3=B3pez=2DLarraz?=) Date: Mon, 7 Feb 2022 09:57:00 +0100 Subject: Connectionists: 14 Open positions in neurotechnology R&D Message-ID: Bitbrain is a neurotechnology company that combines neuroscience, artificial intelligence, and hardware to develop innovative products. We are currently looking for 14 highly talented and enthusiastic individuals to join our R&D team. The profiles vary across the entire R&D pipeline of neurotechnology applications in healthcare: neuro/biomedical engineering, electronic engineering, software engineering, computer science, data science, project management, clinical neuropsychology, neuroscience lab assistants. More information and application procedure: https://www.bitbrain.com/careers -- Eduardo L?pez-Larraz, PhD Research Scientist R&D Department eduardo.lopez at bitbrain.com Twitter Scholar LinkedIn www.bitbrain.com +34 931 444 823 info at bitbrain.com Facebook Twitter LinkedIn -------------- next part -------------- An HTML attachment was scrubbed... URL: From M.Loog at tudelft.nl Mon Feb 7 03:10:33 2022 From: M.Loog at tudelft.nl (Marco Loog - EWI) Date: Mon, 7 Feb 2022 08:10:33 +0000 Subject: Connectionists: prediction modeling for hip OA, Ph.D. student Message-ID: <03a15e2904a245d580ec8273896ff9a4@tudelft.nl> Ph.D. student vacancy in prediction modeling using artificial intelligence for hip osteoarthritis with Erasmus MC [EMC] and Delft University of Technology [TU Delft]. See the vacancy text for details: https://www.werkenbijerasmusmc.nl/en/vacancy/64984/phd-student-project-hips-dont-lie-05.13.22.td From bucchiarone at fbk.eu Mon Feb 7 04:31:29 2022 From: bucchiarone at fbk.eu (Antonio Bucchiarone) Date: Mon, 7 Feb 2022 10:31:29 +0100 Subject: Connectionists: MODELS 2022 : ACM/IEEE 25th International Conference on Model Driven Engineering Languages and Systems Message-ID: --------------------------------------------------------------------------- MODELS 2022 ACM/IEEE 25th International Conference on Model Driven Engineering Languages and Systems October 16-21, 2022 Montreal, Canada https://conf.researchr.org/home/models-2022 --------------------------------------------------------------------------- MODELS is the premier conference series for model-based software and systems engineering. Since 1998, MODELS has covered all aspects of modeling, from languages and methods, to tools and applications. Attendees of MODELS come from diverse backgrounds, including researchers, academics, engineers, and industrial professionals. MODELS 2022 is a forum for participants to exchange cutting-edge research results and innovative practical experiences around modeling, modeling languages, and model-based software and systems engineering. This year?s edition will provide an opportunity for the modeling community to further advance the foundations of modeling, and come up with innovative applications of modeling in emerging areas of cyber-physical systems, embedded systems, socio-technical systems, cloud computing, big data, machine learning, security, open source, and sustainability. For this year?s edition, the conference has the special theme ?Modeling for social good? #MDE4SG. Thus, we especially encourage contributions where model-based engineering intersects with research and applications on, not exclusively, socio-technical systems, tools with social impact, integrating human values, data science, artificial intelligence, digital twins, Industry/Society 5.0, and intelligent systems in general. --------------------------------------------------------------------------- *Important Dates* Abstract Submission: May 11, 2022 Paper Submission: May 18, 2022 Author notification: July 12, 2022 Camera Ready Due: August 1, 2022 --------------------------------------------------------------------------- *Topics of Interest (but not restricted to):* MODELS 2022 seeks submissions on diverse topics related to modeling for software and systems engineering, including, but not limited to: ? Foundations of model-based engineering, including definition of syntax and semantics of modeling languages and model transformation languages. ? New paradigms, formalisms, applications, approaches, frameworks, or processes for model-based engineering such as low-code/no-code development, digital twins, etc. ? Definition, usage, and analysis of model-based generative and re-engineering approaches. ? Models at Runtime: model-based monitoring, analysis, and adaptation towards intelligent systems, e.g., with digital shadows or digital twins. ? Development of model-based systems engineering approaches and modeling-in-the-large including interdisciplinary engineering and coordination. ? Applications of AI to model-based engineering problems including e.g., search-based and machine learning approaches. ? Model-based engineering foundations for AI-based systems. ? Human and organizational factors in model-based engineering. ? Tools, meta-tools, and language workbenches for model-based engineering, including model management and scalable model repositories. ? Integration of modeling languages and tools (hybrid multi-modeling approaches). ? Evaluation and comparison of modeling languages, techniques, and tools. ? Quality assurance (analysis, testing, verification) for functional and non-functional properties of models and model transformations. ? Collaborative modeling research to address global and team management issues (e.g., browser-based and cloud-enabled collaboration). ? Evolution of modeling languages and related standards. ? Evidence-based education research for curricular concerns on modeling topics. ? Modeling in software engineering; applications of models to address general software engineering challenges. ? Modeling for specific challenges such as collaboration, scalability, security, interoperability, adaptability, flexibility, maintainability, dependability, reuse, energy efficiency, sustainability, and uncertainty. ? Modeling with, and for, new and emerging systems and paradigms such as security, cyber-physical systems (CPSs), the Internet of Things (IoTs), cloud computing, DevOps, data analytics, data science, machine learning, big data, systems engineering, socio-technical systems, critical infrastructures and services, robotics, mobile applications, conversational agents, open source software, sustainability and modeling for social good. ? Empirical studies of applying model-based engineering for domains such as smart production, smart cities, smart enterprises, smart mobility, smart society, etc. --------------------------------------------------------------------------- As in previous years, MODELS 2022 is offering two tracks for technical papers: the Foundations Track and the Practice & Innovation Track. A detailed description on the submission process for both tracks is provided at: https://conf.researchr.org/track/models-2022/models-2022-technical-track ***** FOUNDATIONS TRACK ***** We invite authors to submit high quality contributions describing significant, original, and unpublished results in the following categories: 1. Technical Papers Technical papers should describe innovative research in modeling or model-driven engineering activities. Papers in this submission category should describe a novel contribution to the field and should carefully support claims of novelty with citations to the relevant literature. Evaluation Criteria: Technical papers are evaluated on the basis of originality, soundness, relevance, importance of contribution, strength of validation, quality of presentation and appropriate comparison to related work. Where a submission builds upon previous work of the author(s), the novelty of the new contribution must be described clearly with respect to the previous work. Technical papers need to discuss clearly how the results were validated (e.g., formal proofs, controlled experiments, rigorous case studies, or simulations). Authors are strongly encouraged to make the artifacts used for the evaluation publicly accessible, e.g., through a Github repository or an alternative that is likely to remain available. There will be an artifact evaluation process, as discussed below. 2. New Ideas and Vision Papers We solicit short papers that present new ideas and visions. Such papers may describe new, non-conventional model-driven engineering research positions or approaches that depart from standard practice. They can describe well-defined research ideas that are at an early stage of investigation. They could also provide new evidence that common wisdom should be challenged, present new unifying theories about existing modeling research that provides novel insight or that can lead to the development of new technologies or approaches, or apply modeling technology to radically new application areas. Evaluation Criteria: New ideas and vision papers will be assessed primarily on their level of originality and potential for impact on the field in terms of promoting innovative thinking. Hence, inadequacies in the state-of-the-art and the pertinence, correctness, and impact of the idea/vision must be described clearly, even though the new idea need not be fully worked out, and a fully detailed roadmap need not be presented. Authors are strongly encouraged to make the artifacts used for the evaluation publicly accessible, e.g., through a Github repository or an alternative that is likely to remain available. There will be an artifact evaluation process, as discussed below. ***** PRACTICE AND INNOVATION TRACK* The goal of the Practice and Innovation (P&I) Track is to fill the gap between foundational research in model-based engineering (MBE) and industrial needs. We invite authors from academia and/or industry to submit original contributions reporting on the development of innovative MBE solutions in industries, public sector, or open-source settings, as well as innovative application of MBE in such contexts. Examples include: ? Scalable and cost-effective methodologies and tools ? Industrial case studies with valuable lessons learned ? Experience reports providing novel insights Each paper should provide clear take-away value by describing the context of a problem of practical importance, and the application of MBE that leads to a solution. A non-exclusive list of topics of interest is provided here. Evaluation Criteria: A paper in the P&I Track will be evaluated mainly from its practical take-away and the potential impact of the findings. More specifically, ? The paper should discuss why the solution to the problem is innovative (e.g., in terms of advancing the state-of-practice), effective, and/or efficient, and what likely practical impact it has or will have; ? The paper should provide a concise explanation of approaches, techniques, methodologies and tools employed; ? The paper should explain best practices that emerged, tools developed, and/or software processes involved. ? Studies reporting on negative findings must provide a thorough discussion of the potential causes of failure, and ideally a perspective on how to solve them. ? Authors are encouraged to make the artifacts publicly accessible, e.g., through a Github repository or an alternative that is likely to remain available. There will be an optional artifact evaluation process, as discussed below. --------------------------------------------------------------------------- ***** Artefact Evaluation* After the notification, the authors of accepted papers will be invited to submit their accompanying artifacts (e.g., software, datasets, proofs) to the Artifact Evaluation track to be evaluated by the Artifact Evaluation Committee. Participation in the Artifact Evaluation process is optional and does not affect the final decision regarding the acceptance of papers. Papers that successfully go through the Artifact Evaluation process will be rewarded with a seal of approval included in the papers. ***** Special Issue in SoSyM* Authors of best papers from the conference will be invited to revise and submit extended versions of their papers for publication in the Journal of Software and Systems Modeling. Dr. Antonio Bucchiarone Motivational Digital Systems (MoDiS) Fondazione Bruno Kessler (FBK), Trento, Italy -- -- Le informazioni contenute nella presente comunicazione sono di natura? privata e come tali sono da considerarsi riservate ed indirizzate? esclusivamente ai destinatari indicati e per le finalit? strettamente? legate al relativo contenuto. Se avete ricevuto questo messaggio per? errore, vi preghiamo di eliminarlo e di inviare una comunicazione? all?indirizzo e-mail del mittente. -- The information transmitted is intended only for the person or entity to which it is addressed and may contain confidential and/or privileged material. If you received this in error, please contact the sender and delete the material. -------------- next part -------------- An HTML attachment was scrubbed... URL: From bucchiarone at fbk.eu Mon Feb 7 04:34:14 2022 From: bucchiarone at fbk.eu (Antonio Bucchiarone) Date: Mon, 7 Feb 2022 10:34:14 +0100 Subject: Connectionists: MODELS 2022 - Call for Workshops Proposal Message-ID: ABOUT The MODELS series of conferences is the premier venue for the exchange of innovative technical ideas and experiences relating to model-driven approaches in the development of software-based systems. This year?s edition will provide an opportunity for the modeling community to further advance the foundations of modeling, and come up with innovative applications of modeling in emerging areas of cyber-physical systems, embedded systems, socio-technical systems, cloud computing, big data, machine learning, security, open source, and sustainability. Following the tradition of previous conferences, MODELS 2022 will host a number of workshops, during the three days before the main conference. The workshops will provide a collaborative forum for a group of typically 15 to 30 participants to exchange recent and/or preliminary results, to conduct intensive discussions on a particular topic, or to coordinate efforts between representatives of a technical community. They are intended as a forum for lively discussions of innovative ideas, recent progress, or practical experience on model-driven engineering for specific aspects, specific problems, or domain-specific needs. Each workshop should provide a balanced distribution of its time for both presentations of papers (favoring the attendance of young researchers) and discussions. The duration of these workshops is in general one day, but we encourage the submission of half-day workshop proposals on focused topics as well. NEW THIS YEAR: BEST THEME PAPER AWARD This year?s conference will feature a Best Theme Paper Award spanning across all tracks. The special theme of this year?s conference is "Modeling for social good" #MDE4SG. Workshops provide an excellent opportunity to contribute to this theme, as workshop papers tend to be more exploratory and easier to steer towards new avenues. We encourage researchers to submit papers related to the special theme, regardless of the workshop they are submitting their paper to. We also encourage prospective workshop organizers to reflect on this theme in their workshop proposals. Workshop proposals on related topics?such as socio-technical systems, tools with social impact, integrating human values, data science, and intelligent systems?are especially welcome. Workshop proposals should accommodate the following selection criteria and process for the Best Theme Paper Award. -A paper must be at least 5 pages long to be eligible. -Each workshop can nominate one candidate paper. This is optional, as not every workshop will be able to nominate a paper. We suggest you include the way of selection in your workshop proposals, and decide upon the actual nomination, or lack thereof, once the workshop PC has reviewed the submitted papers. -The workshop chairs will select one paper out of the candidate papers, and nominate it to the Selection Committee. SUBMISSION PROCESS Submit your workshop proposal electronically in PDF using the Springer LNCS style through the MODELS EasyChair submission site: https://easychair.org/conferences/?conf=models22workshops. Please adhere to the workshop proposal guidelines below, providing every requested information about the proposed workshop, using at most five pages. Please include the one-page draft of your planned Call for Papers to the proposal (not included in the five pages). In order to ensure proper coordination with the deadlines of the main conference, the deadlines specified in Important Dates below have to be respected by your plan for your workshop. An Overleaf template with the suggested structure is available here: https://www.overleaf.com/read/wbnjqkvthvhy. (This is a read-only link. In order to work with it, you need to make a copy of it.) PROCEEDINGS As in the previous years, a joint Companion Proceedings will be published by IEEE that includes papers from all workshops. This will include: -an opening message from the organizers (that should include the workshop program committee, if applicable); and -every peer-reviewed paper presented at the workshop. Submissions must adhere to the IEEE formatting instructions, which can be found at: https://www.ieee.org/conferences/publishing/templates.html. Papers should have at least 5 pages. We propose page limits of 5 pages for short papers and 10 pages for full papers following the same style and format of the main tracks of the Conference. IMPORTANT DATES -Workshop Proposal Submissions Deadline: March 25, 2022 -Workshop Proposal Acceptance Notification: April 18, 2022 -Submissions deadline: July 20, 2022 -Notification of authors: August 19, 2022 -Camera-ready deadline: August 26, 2022 -Workshop dates: October 16-18, 2022 ORGANIZATION AND CONTACT For any information, please visit the official website at https://conf.researchr.org/track/models-2022/models-2022-workshops, or contact the workshops co-chairs at: models2022workshops at easychair.org. Chairs: -Istvan David, Universit? de Montr?al, Canada -Jes?s S?nchez Cuadrado, Universidad de Murcia, Spain PROPOSAL GUIDELINES The guidelines?regarding the information you must include in your proposal and how the proposal document needs to be structured?are available on the website: https://conf.researchr.org/track/models-2022/models-2022-workshops#proposal-guidelines . *Dr. Antonio Bucchiarone* Motivational Digital Systems (MoDiS) Fondazione Bruno Kessler (FBK), Trento, Italy -- -- Le informazioni contenute nella presente comunicazione sono di natura? privata e come tali sono da considerarsi riservate ed indirizzate? esclusivamente ai destinatari indicati e per le finalit? strettamente? legate al relativo contenuto. Se avete ricevuto questo messaggio per? errore, vi preghiamo di eliminarlo e di inviare una comunicazione? all?indirizzo e-mail del mittente. -- The information transmitted is intended only for the person or entity to which it is addressed and may contain confidential and/or privileged material. If you received this in error, please contact the sender and delete the material. -------------- next part -------------- An HTML attachment was scrubbed... URL: From louisphil.lopes at gmail.com Mon Feb 7 05:01:34 2022 From: louisphil.lopes at gmail.com (Phil Lopes) Date: Mon, 7 Feb 2022 10:01:34 +0000 Subject: Connectionists: Post-Doc Junior Researcher in Signal Processing, Time Series Data Analysis and Machine Learning for Virtual Reality Message-ID: HEI-Lab is looking for a Junior Researcher proficient in Signal Processing, Time Series Data Analysis and Machine Learning. The person hired will be responsible for various phases of research and integrated in a multidisciplinary team of researchers at the crossroads of game design, virtual reality, biomedicine, psychology, and artificial intelligence. The core research field will be the construction of systems and experiences that leverage human physiological sensors for the adaptation of virtual content and capable of reporting detailed patient feedback for therapists during and post exposure. The majority of applications have a focus in using physiological monitoring technology such as ECG, EDA, EMG and EEG; as such applicants who already have some degree of experience with these technologies will be favoured. Responsibilities: - Take part of the experimental protocol design to optimize the data collection process and facilitate the processing methodology. - Clean, process and analyse time-series data, which can include the construction of tools and algorithms that aid this process for future projects. - Build statistical models through established techniques for the recognition and prediction of common human-based behaviours. - Become the signal processing ?expert? of the research unit, allowing fellow researchers to consult your expertise within the field. Remuneration: Gross monthly wage: 2134.73 euros, with correspondence to level 33 to the single remuneration table, which is updated based on the remuneration and the value table of the basic monthly remuneration as approved by Ordinance nr. 1553-C/2008, of December 31st, and combined with Decree-Law nr. 10/20201, of February 1st. Duration: The contract to be carried out is scheduled to start on April 1, 2022, ending on April 30, 2025. Location: The HEI-Lab itself is a research centre located at the Lus?fona University, in Lisbon, Portugal. Application: Applications are accepted until the 28th February 2022, and must be sent by e-mail to micaela.fonseca at ulusofona.pt with ana.mourato at ulusofona.pt CC?d, with ?COFAC/ULHT/HEI-LAB/JR/2022? (without quotations) as the email subject and with the following documents attached: - Presentation letter referring to the reasons that justified the application; - Curriculum "vitae" mentioning professional experience, accompanied by a list of scientific publications produced and participation in funded projects; - Doctoral certificate; - Identification and contacts with the respective ?email? addresses - of at least two academic personalities who attest to the displayed curriculum; - Work plan for the 3-year period to be developed in the 'Human-Computer Interaction' area, compiled by the HEI-Lab I&D Labs - (https://hei-lab.ulusofona.pt); - Link to their Portfolio Page (e.g. GitHub/GitLab/Bitbucket, Personal Website, etc.) - Other documents considered relevant by the candidate and which, in his perspective, seem relevant to prove and evaluate the respective Further Details: Additional information about the proposal can be found here: https://euraxess.ec.europa.eu/jobs/738448. Questions on the proposal can be directed to micaela.fonseca at ulusofona.pt. -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Mon Feb 7 08:12:13 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Mon, 7 Feb 2022 08:12:13 -0500 Subject: Connectionists: Please stop having conversations on this mailing list In-Reply-To: References: Message-ID: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> Pau, actually this was exactly what Connectionists was created for:? civil, thoughtful discussion about Neural Computation. It is very lightly moderated and provides a forum for discussion, promotion and potential collaborations. This discussion has been a bit redundant here and there, but for the most part reconstructing much of the arguments from the 1980s in the background of the DL revolution. Explainable AI, Neuro-Symbolic models and causal structure have been constant concerns of thoughtful folks using, building Neural network systems for 40 years. The threads have gone in predictable ways, but have interesting voices and real issues behind them.?? Some of this is dialetical (my friend Gary Marcus who is a master at this) and some of it hyperbolic.?? But all have made excellent and interesting points. Best Regards, Steve On 2/5/22 11:29 AM, pau wrote: > Dear connectionists, > > To the best of my understanding, the aim of this mailing list is to meet > the needs of working professionals. I understand by this sharing > information about events, publications, developments, etc. that can be > useful for someone who works in this field. > If someone wants to argue, discuss, praise or assign blame, agree or > disagree, please do so in private correspondence or in non-work-specific > social media channels. > > Best, > P. -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From jose at rubic.rutgers.edu Mon Feb 7 09:01:28 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Mon, 7 Feb 2022 09:01:28 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <537DF004-25CE-45A2-8155-D7E6018F4EE5@nyu.edu> References: <537DF004-25CE-45A2-8155-D7E6018F4EE5@nyu.edu> Message-ID: <3dff1e2e-52b6-2ff3-3152-c7cf919fb14e@rubic.rutgers.edu> Gary, This is one of the first posts of yours, that I can categorically agree with! I think building cognitive models through *some* training regime or focused sampling or architectures or something but not explicit, for example. The other fundamental cognitive/perceptual capability in this context is the ability of Neural Networks to do what Shepard (1970; Garner 1970s), had modeled as perceptual separable processing (finding parts) and perceptual integral process (finding covariance and structure). Shepard argued these fundamental perceptual processes were dependent on development and learning. A task was created with double dissociation of a categorization problem.? In one case: separable ( in effect, uncorrelated? features in the stimulus) were presented in categorization task that required you pay attention to at least 2 features at the same time to categorize correctly ("condensation").? in the other case: integral stimuli (in effect correlated features in stimuli) were presented? in a categorization task that required you to ignore the correlation and do categorize on 1 feature at a time ("filtration").?? This produced a? result that separable stimuli were more quickly learned in filtration tasks then integral stimuli in condensation tasks.? Non-intuitively,? Separable stimuli are learned more slowly in condensation tasks then integral stimuli then in filtration tasks.?? In other words attention to feature structure could cause improvement in learning or interference. Not that surprising.. however-- In the 1980s NN with single layers (Backprop) *could not* replicate this simple problem indicating that the cognitive model was somehow inadequate.???? Backprop simply learned ALL task/stimuli parings at the same rate, ignoring the subtle but critical difference.? It failed. Recently we (https://www.frontiersin.org/articles/10.3389/fpsyg.2018.00374/full?&utm_source=Email_to_authors_&utm_medium=Email&utm_content=T1_11.5e1_author&utm_campaign=Email_publication&field=&journalName=Frontiers_in_Psychology&id=284733) were able to show that JUST BY ADDING LAYERS the DL? does match to human performance. What are the layers doing??? We offer an possible explanation that needs testing.??? Layers, appear to create a type of buffer that allows the network to "curate", feature detectors that are spatially distant from the input (conv layer, for example), this curation comes in? various attention forms (something in that will appear in a new paper--not enough room here), which appears to qualitatively change the network processing states, and cognitive capabilities.???? Well, that's the claim. The larger point, is that apparently architectures interact with learning rules, in ways that can cross this symbolic/neural river of styx, without falling into it. Steve On 2/5/22 10:38 AM, Gary Marcus wrote: > There is no magic in understanding, just computation that has been > realized in the wetware of humans and that eventually can be realized > in machines. But understanding is not (just) learning. > > Understanding incorporates (or works in tandem with) learning - but > also, critically, in tandem with inference, /and the development and > maintenance of cognitive models/.Part of developing an understanding > of cats in general is to learn long term-knowledge about their > properties, both directly (e.g., through observation) and indirectly > (eg through learning factsabout animals in general that can be > extended to cats), often through inference (if all animals have DNA, > and a cat is an animal, it must also have DNA).The understanding of a > particular cat also involves direct observation, but also inference > (egone might surmise that the reason that Fluffy is running about the > room is that Fluffy suspects there is a mouse stirring somewhere > nearby). But all of that, I would say, is subservient to the > construction of cognitive models that can be routinely updated (e.g., > Fluffy is currently in the living room, skittering about, perhaps > looking for a mouse). > > > ?In humans, those dynamic, relational models, which form part of an > understanding, can support inference (if Fluffy is in the living room, > we can infer that Fluffy is not outside, not lost, etc). Without such > models - which I think represent a core part of understanding - AGI is > an unlikely prospect. > > > Current neural networks, as it happens, are better at acquiring > long-term knowledge (cats have whiskers) than they are at dynamically > updating cognitive models in real-time. LLMs like GPT-3 etc lack the > kind of dynamic model that I am describing. To a modest degree they > can approximate it on the basis of large samples of texts, but their > ultimate incoherence stems from the fact that they do not have robust > internal cognitive models that they can update on the fly. > > > Without such cognitive models you can still capture some aspects of > understanding (eg predicting that cats are likely to be furry), but > things fall apart quickly; inference is never reliable, and coherence > is fleeting. > > > As a final note, one of the most foundational challenges in > constructing adequate cognitive models of the world is to have a clear > distinction between individuals and kinds; as I emphasized 20 years > ago (in The Algebraic Mind), this has always been a weakness in neural > networks, and I don?t think that the type-token problem has yet been > solved. > > > Gary > > >> On Feb 5, 2022, at 01:31, Asim Roy wrote: >> >> ? >> >> All, >> >> I think the broader question was ?understanding.? Here are two >> Youtube videos showing simple robots ?learning? to walk. They are >> purely physical systems. Do they ?understand? anything ? such as the >> need to go around an obstacle, jumping over an obstacle, walking up >> and down stairs and so on? By the way, they ?learn? to do these >> things on their own, literally unsupervised, very much like babies. >> The basic question is: what is ?understanding? if not ?learning?? Is >> there some other mechanism (magic) at play in our brain that helps us >> ?understand?? >> >> https://www.youtube.com/watch?v=gn4nRCC9TwQ >> >> >> https://www.youtube.com/watch?v=8sO7VS3q8d0 >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> *From:* Ali Minai >> *Sent:* Friday, February 4, 2022 11:38 PM >> *To:* Asim Roy >> *Cc:* Gary Marcus ; Danko Nikolic >> ; Brad Wyble ; >> connectionists at mailman.srv.cs.cmu.edu; AIhub >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >> Geoff Hinton >> >> Asim >> >> Of course there's nothing magical about understanding, and the mind >> has to emerge from the physical system, but our AI models at this >> point are not even close to realizing how that happens. We are, at >> best, simulating a superficial approximation of a few parts of the >> real thing. A single, integrated system where all the aspects of >> intelligence emerge from the same deep, well-differentiated physical >> substrate is far beyond our capacity. Paying more attention to >> neurobiology will be essential to get there, but so will paying >> attention to development - both physical and cognitive - and >> evolution. The configuration of priors by evolution is key to >> understanding how real intelligence learns so quickly and from so >> little. This is not an argument for using genetic algorithms to >> design our systems, just for understanding the tricks evolution has >> used and replicating them by design. Development is more feasible to >> do computationally, but hardly any models have looked at it except in >> a superficial sense. Nature creates basic intelligence not so much by >> configuring functions by explicit training as by tweaking, >> modulating, ramifying, and combining existing ones in a multi-scale >> self-organization process. We then learn much more complicated things >> (like playing chess) by exploiting that substrate, and using explicit >> instruction or learning by practice. The fundamental lesson of >> complex systems is that complexity is built in stages - each level >> exploiting the organization of the level below it. We see it in >> evolution, development, societal evolution, the evolution of >> technology, etc. Our approach in AI, in contrast, is to initialize a >> giant, naive system and train it to do something really complicated - >> but really specific - by training the hell out of it. Sure, now we do >> build many systems on top of pre-trained models like GPT-3 and BERT, >> which is better, but those models were again trained by the same >> none-to-all process I decried above. Contrast that with how humans >> acquire language, and how they integrate it into their *entire* >> perceptual, cognitive, and behavioral repertoire, not focusing just >> on this or that task. The age of symbolic AI may have passed, but the >> reductionistic mindset has not. We cannot build minds by chopping it >> into separate verticals. >> >> FTR, I'd say that the emergence of models such as GLOM and Hawkins >> and Ahmed's "thousand brains" is a hopeful sign. They may not be >> "right", but they are, I think, looking in the right direction. With >> a million miles to go! >> >> Ali >> >> *Ali A. Minai, Ph.D.* >> Professor and Graduate Program Director >> Complex Adaptive Systems Lab >> Department of Electrical Engineering & Computer Science >> >> 828 Rhodes Hall >> >> University of Cincinnati >> Cincinnati, OH 45221-0030 >> >> >> Phone: (513) 556-4783 >> Fax: (513) 556-7326 >> Email: Ali.Minai at uc.edu >> minaiaa at gmail.com >> >> WWW: https://eecs.ceas.uc.edu/~aminai/ >> >> >> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy > > wrote: >> >> First of all, the brain is a physical system. There is no ?magic? >> inside the brain that does the ?understanding? part. Take for >> example learning to play tennis. You hit a few balls - some the >> right way and some wrong ? but you fairly quickly learn to hit >> them right most of the time. So there is obviously some >> simulation going on in the brain about hitting the ball in >> different ways and ?learning? its consequences. What you are >> calling ?understanding? is really these simulations about >> different scenarios. It?s also very similar to augmentation used >> to train image recognition systems where you rotate images, >> obscure parts and so on, so that you still can say it?s a cat >> even though you see only the cat?s face or whiskers or a cat >> flipped on its back. So, if the following questions relate to >> ?understanding,? you can easily resolve this by simulating such >> scenarios when ?teaching? the system. There?s nothing ?magical? >> about ?understanding.? As I said, bear in mind that the brain, >> after all, is a physical system and ?teaching? and >> ?understanding? is embodied in that physical system, not outside >> it. So ?understanding? is just part of ?learning,? nothing more. >> >> DANKO: >> >> What would happen to the hat if the hamster rolls on its back? >> (Would the hat fall off?) >> >> What would happen to the red hat when the hamster enters its >> lair? (Would the hat fall?off?) >> >> What would happen to that hamster when it goes foraging? (Would >> the red hat have an influence on finding food?) >> >> What would happen in a situation of being chased by a predator? >> (Would it be easier for predators to spot the hamster?) >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> *From:* Gary Marcus > > >> *Sent:* Thursday, February 3, 2022 9:26 AM >> *To:* Danko Nikolic > > >> *Cc:* Asim Roy >; >> Geoffrey Hinton > >; AIhub > >; >> connectionists at mailman.srv.cs.cmu.edu >> >> *Subject:* Re: Connectionists: Stephen Hanson in conversation >> with Geoff Hinton >> >> Dear Danko, >> >> Well said. I had a somewhat similar response to Jeff Dean?s 2021 >> TED talk, in which he said (paraphrasing from memory, because I >> don?t remember the precise words) that the famous 200 Quoc Le >> unsupervised model >> [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >> ] >> had learned the concept of a ca. In reality the model had >> clustered together some catlike images based on the image >> statistics that it had extracted, but it was a long way from a >> full, counterfactual-supporting concept of a cat, much as you >> describe below. >> >> I fully agree with you that the reason for even having a >> semantics is as you put it, "to 1) learn with a few examples and >> 2) apply the knowledge to a broad set of situations.? GPT-3 >> sometimes gives the appearance of having done so, but it falls >> apart under close inspection, so the problem remains unsolved. >> >> Gary >> >> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >> > wrote: >> >> G. Hinton wrote: "I believe that any reasonable person would >> admit that if you ask a neural net to draw a picture of a >> hamster wearing a red hat and it draws such a picture, it >> understood?the request." >> >> I would like to suggest why drawing a?hamster with a red?hat >> does not necessarily imply understanding of the statement >> "hamster wearing a red hat". >> >> To understand that "hamster wearing a red hat" would mean >> inferring, in newly?emerging situations of this hamster, all >> the real-life implications?that the red hat brings to the >> little animal. >> >> What would happen to the hat if the hamster rolls on its >> back? (Would the hat fall off?) >> >> What would happen to the red hat when the hamster enters its >> lair? (Would the hat fall?off?) >> >> What would happen to that hamster when it goes foraging? >> (Would the red hat have an influence on finding food?) >> >> What would happen in a situation of being chased by a >> predator? (Would it be easier for predators to spot the hamster?) >> >> ...and so on. >> >> Countless many questions can be asked. One has understood >> "hamster wearing a red hat" only if one can answer reasonably >> well many of such real-life relevant questions. Similarly, a >> student has?understood materias in a class only if they can >> apply the materials in real-life situations (e.g., applying >> Pythagora's theorem). If a student gives a correct answer to >> a multiple?choice question, we don't know whether the student >> understood the material or whether this was just rote >> learning (often, it is rote learning). >> >> I also suggest that understanding also comes together with >> effective learning: We store new information in such a way >> that we can recall it later and use it effectively? i.e., >> make good inferences in newly emerging situations based on >> this knowledge. >> >> In short: Understanding makes us humans able to 1) learn with >> a few examples and 2) apply the knowledge to a broad set of >> situations. >> >> No neural network?today has such capabilities and we don't >> know how to give them such capabilities. Neural networks need >> large amounts of training?examples that cover a large variety >> of situations and then the?networks can only deal with what >> the training examples have already covered. Neural networks >> cannot extrapolate in that 'understanding' sense. >> >> I suggest that understanding truly extrapolates from a piece >> of knowledge. It is not about satisfying a task such as >> translation between languages or drawing hamsters with hats. >> It is how you got the capability to complete the task: Did >> you only have a few examples that covered something different >> but related and then you extrapolated from that knowledge? If >> yes, this is going in the direction of understanding. Have >> you seen countless examples and then interpolated among them? >> Then perhaps it is not understanding. >> >> So, for the case of drawing a hamster wearing a red hat, >> understanding perhaps would have taken place if the following >> happened before that: >> >> 1) first, the network learned about hamsters (not many examples) >> >> 2) after that the network learned about red hats (outside the >> context of hamsters and without many examples) >> >> 3) finally the network learned about drawing (outside of the >> context of hats and hamsters, not many examples) >> >> After that, the network is asked to draw a hamster with a red >> hat. If it does it successfully, maybe we have started >> cracking the problem of understanding. >> >> Note also that this requires the network to learn >> sequentially without exhibiting catastrophic forgetting of >> the previous knowledge, which is possibly also a consequence >> of human learning by understanding. >> >> Danko >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> >> https://www.linkedin.com/in/danko-nikolic/ >> >> >> --- A progress usually starts with an insight --- >> >> >> >> >> >> Virus-free. www.avast.com >> >> >> >> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > > wrote: >> >> Without getting into the specific dispute between Gary >> and Geoff, I think with approaches similar to GLOM, we >> are finally headed in the right direction. There?s plenty >> of neurophysiological evidence for single-cell >> abstractions and multisensory neurons in the brain, which >> one might claim correspond to symbols. And I think we can >> finally reconcile the decades old dispute between >> Symbolic AI and Connectionism. >> >> GARY: (Your GLOM, which as you know I praised publicly, >> is in many ways an effort to wind up with encodings that >> effectively serve as symbols in exactly that way, >> guaranteed to serve as consistent representations of >> specific concepts.) >> >> GARY: I have /never/ called for dismissal of neural >> networks, but rather for some hybrid between the two (as >> you yourself contemplated in 1991); the point of the 2001 >> book was to characterize exactly where multilayer >> perceptrons succeeded and broke down, and where symbols >> could complement them. >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> *From:* Connectionists >> > > >> *On Behalf Of *Gary Marcus >> *Sent:* Wednesday, February 2, 2022 1:26 PM >> *To:* Geoffrey Hinton > > >> *Cc:* AIhub > >; >> connectionists at mailman.srv.cs.cmu.edu >> >> *Subject:* Re: Connectionists: Stephen Hanson in >> conversation with Geoff Hinton >> >> Dear Geoff, and interested others, >> >> What, for example, would you make of a system that >> often?drew the red-hatted hamster you requested, and >> perhaps a fifth of the time gave you utter nonsense?? Or >> say one that you trained to create birds but sometimes >> output stuff like this: >> >> >> >> One could >> >> a. avert one?s eyes and deem the anomalous outputs irrelevant >> >> or >> >> b. wonder if it might be possible that sometimes the >> system gets the right answer for the wrong reasons (eg >> partial historical contingency), and wonder whether >> another approach might be indicated. >> >> Benchmarks are harder than they look; most of the field >> has come to recognize that. The Turing Test has turned >> out to be a lousy measure of intelligence, easily gamed. >> It has turned out empirically that the Winograd Schema >> Challenge did not measure common sense as well as Hector >> might have thought. (As it happens, I am a minor coauthor >> of a very recent review on this very topic: >> https://arxiv.org/abs/2201.02387 >> ) >> But its conquest in no way means machines now have common >> sense; many people from many different perspectives >> recognize that (including, e.g., Yann LeCun, who >> generally tends to be more aligned with you than with me). >> >> So: on the goalpost of the Winograd schema, I was wrong, >> and you can quote me; but what you said about me and >> machine translation remains your invention, and it is >> inexcusable that you simply ignored my 2019 >> clarification. On the essential goal of trying to reach >> meaning and understanding, I remain unmoved; the problem >> remains unsolved. >> >> All of the problems LLMs have with coherence, >> reliability, truthfulness, misinformation, etc stand >> witness to that fact. (Their persistent inability to >> filter out toxic and insulting remarks stems from the >> same.) I am hardly the only person in the field to see >> that progress on any given benchmark does not inherently >> mean that the deep underlying problems have solved. You, >> yourself, in fact, have occasionally made that point. >> >> With respect to embeddings: Embeddings are very good for >> natural language /processing/; but NLP is not the same as >> NL/U/ ? when it comes to /understanding/, their worth is >> still an open question. Perhaps they will turn out to be >> necessary; they clearly aren?t sufficient. In their >> extreme, they might even collapse into being symbols, in >> the sense of uniquely identifiable encodings, akin to the >> ASCII code, in which a specific set of numbers stands for >> a specific word or concept. (Wouldn?t that be ironic?) >> >> (Your GLOM, which as you know I praised publicly, is in >> many ways an effort to wind up with encodings that >> effectively serve as symbols in exactly that way, >> guaranteed to serve as consistent representations of >> specific concepts.) >> >> Notably absent from your email is any kind of apology for >> misrepresenting my position. It?s fine to say that ?many >> people thirty years ago once thought X? and another to >> say ?Gary Marcus said X in 2015?, when I didn?t. I have >> consistently felt throughout our interactions that you >> have mistaken me for Zenon Pylyshyn; indeed, you once (at >> NeurIPS 2014) apologized to me for having made that >> error. I am still not he. >> >> Which maybe connects to the last point; if you read my >> work, you would see thirty years of arguments >> /for/?neural networks, just not in the way that you want >> them to exist. I have ALWAYS argued that there is a role >> for them; ?characterizing me as a person >> ?strongly?opposed to neural networks? misses the whole >> point of my 2001 book, which was subtitled ?Integrating >> Connectionism and Cognitive Science.? >> >> In the last two decades or so you have insisted (for >> reasons you have never fully clarified, so far as I know) >> on abandoning symbol-manipulation, but the reverse is not >> the case: I have /never/ called for dismissal of neural >> networks, but rather for some hybrid between the two (as >> you yourself contemplated in 1991); the point of the 2001 >> book was to characterize exactly where multilayer >> perceptrons succeeded and broke down, and where symbols >> could complement them. It?s a rhetorical trick (which is >> what the previous thread was about) to pretend otherwise. >> >> Gary >> >> On Feb 2, 2022, at 11:22, Geoffrey Hinton >> > > wrote: >> >> ? >> >> Embeddings are just vectors of soft feature detectors >> and they are very good for NLP.? The quote on my >> webpage from Gary's 2015 chapter implies the opposite. >> >> A few decades ago, everyone I knew then would?have >> agreed that the ability to translate a sentence into >> many different languages was strong evidence that you >> understood it. >> >> But once neural networks could do that, their critics >> moved the goalposts. An exception is Hector Levesque >> who defined the goalposts more sharply by saying that >> the ability to get pronoun references correct in >> Winograd sentences is a crucial test. Neural nets are >> improving at that but still have some way to go. Will >> Gary agree that when they can get pronoun >> references?correct in Winograd sentences they >> really?do understand? Or does he want to reserve the >> right to weasel out of that too? >> >> Some people, like Gary, appear to be strongly?opposed >> to neural networks because?they do not fit their >> preconceived notions of how the mind should work. >> >> I believe that any reasonable person would admit that >> if you ask a neural net to draw a picture of a >> hamster wearing a red hat and it draws such a >> picture, it understood?the request. >> >> Geoff >> >> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus >> > wrote: >> >> Dear AI Hub, cc: Steven Hanson and Geoffrey >> Hinton, and the larger neural network community, >> >> There has been a lot of recent discussion on this >> list about framing and scientific integrity. >> Often the first step in restructuring narratives >> is to bully and dehumanize critics. The second is >> to misrepresent their position. People in >> positions of power are sometimes tempted to do this. >> >> The Hinton-Hanson interview that you just >> published is a real-time example of just that. It >> opens with a needless and largely content-free >> personal attack on a single scholar (me), with >> the explicit intention of discrediting that >> person. Worse, the only substantive thing it says >> is false. >> >> Hinton says ?In 2015 he [Marcus] made a >> prediction that computers wouldn?t be able to do >> machine translation.? >> >> I never said any such thing. >> >> What I predicted, rather, was that multilayer >> perceptrons, as they existed then, would not (on >> their own, absent other mechanisms) >> /understand/?language. Seven years later, they >> still haven?t, except in the most superficial way. >> >> I made no comment whatsoever about machine >> translation, which I view as a separate problem, >> solvable to a certain degree by correspondance >> without semantics. >> >> I specifically tried to clarify Hinton?s >> confusion in 2019, but, disappointingly, he has >> continued to purvey misinformation despite that >> clarification. Here is what I wrote privately to >> him then, which should have put the matter to rest: >> >> You have taken a single out of context quote >> [from 2015] and misrepresented it. The quote, >> which you have prominently displayed at the >> bottom on your own web page, says: >> >> Hierarchies of features are less suited to >> challenges such as language, inference, and >> high-level planning. For example, as Noam Chomsky >> famously pointed out, language is filled with >> sentences you haven't seen before.?Pure >> classifier systems don't know what to do with >> such sentences. The talent of feature detectors >> -- in??identifying which member of some category >> something belongs to -- doesn't translate into >> understanding novel??sentences, in which each >> sentence has its own unique meaning. >> >> It does /not/?say "neural nets would not be able >> to deal with novel sentences"; it says that >> hierachies of features detectors (on their own, >> if you read the context of the essay) would have >> trouble /understanding /novel?sentences. >> >> Google Translate does yet not /understand/?the >> content of?the sentences is translates. It cannot >> reliably answer questions about who did what to >> whom, or why, it cannot infer the order of the >> events in paragraphs, it can't determine the >> internal consistency of those events, and so forth. >> >> Since then, a number of scholars, such as the the >> computational linguist Emily Bender, have made >> similar points, and indeed current LLM >> difficulties with misinformation, incoherence and >> fabrication all follow from these concerns. >> Quoting from Bender?s prizewinning 2020 ACL >> article on the matter with Alexander Koller, >> https://aclanthology.org/2020.acl-main.463.pdf >> , >> also emphasizing issues of understanding and meaning: >> >> /The success of the large neural language models >> on many NLP tasks is exciting. However, we find >> that these successes sometimes lead to hype in >> which these models are being described as >> ?understanding? language or capturing ?meaning?. >> In this position paper, we argue that a system >> trained only on form has a priori no way to learn >> meaning. .. a clear understanding of the >> distinction between form and meaning will help >> guide the field towards better science around >> natural language understanding. / >> >> Her later article with Gebru on language models >> ?stochastic parrots? is in some ways an extension >> of this point; machine translation requires >> mimicry, true understanding (which is what I was >> discussing in 2015) requires something deeper >> than that. >> >> Hinton?s intellectual error here is in equating >> machine translation with the deeper comprehension >> that robust natural language understanding will >> require; as Bender and Koller observed, the two >> appear not to be the same. (There is a longer >> discussion of the relation between language >> understanding and machine translation, and why >> the latter has turned out to be more approachable >> than the former, in my 2019 book with Ernest Davis). >> >> More broadly, Hinton?s ongoing dismissiveness of >> research from perspectives other than his own >> (e.g. linguistics) have done the field a disservice. >> >> As Herb Simon once observed, science does not >> have to be zero-sum. >> >> Sincerely, >> >> Gary Marcus >> >> Professor Emeritus >> >> New York University >> >> On Feb 2, 2022, at 06:12, AIhub >> > > wrote: >> >> ? >> >> Stephen Hanson in conversation with Geoff Hinton >> >> In the latest episode of this video series >> for AIhub.org >> , >> Stephen Hanson talks to Geoff Hinton?about >> neural networks, backpropagation, >> overparameterization, digit recognition, >> voxel cells, syntax and semantics, Winograd >> sentences, and more. >> >> You can watch the discussion, and read the >> transcript, here: >> >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> >> About AIhub: >> >> AIhub is a non-profit dedicated to connecting >> the AI community to the public by providing >> free, high-quality information through >> AIhub.org >> >> (https://aihub.org/ >> ). >> We help researchers publish the latest AI >> news, summaries of their work, opinion >> pieces, tutorials and more.? We are supported >> by many leading scientific organizations in >> AI, namely AAAI >> , >> NeurIPS >> , >> ICML >> , >> AIJ >> /IJCAI >> , >> ACM SIGAI >> , >> EurAI/AICOMM, CLAIRE >> >> and RoboCup >> . >> >> Twitter: @aihuborg >> >> >> >> >> >> Virus-free. www.avast.com >> >> >> -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From jose at rubic.rutgers.edu Mon Feb 7 08:26:02 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Mon, 7 Feb 2022 08:26:02 -0500 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: <6E8002C5-64E7-41DB-930F-B77BF78F600A@supsi.ch> References: <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> <7f97db1f-13e1-ba48-8b02-f3a2c4769df9@rubic.rutgers.edu> <6E8002C5-64E7-41DB-930F-B77BF78F600A@supsi.ch> Message-ID: <3c93a408-f085-4340-694e-b5aa2c9335de@rubic.rutgers.edu> Juergen, ignoring history by defining it as "fancy talk", is going to make your exegesis of Neural networks always lagging.? You need to embrace all the facts, not just the ones you like or are familiar with.? Your whole endeavor, is an attempt to destroy what you feel is the common false narrative on the origin of neural networks.? I am happy to chat more about this sometime, but I still think your mathematical lens, is preventing you from seeing the bigger conceptual picture. Best, Steve On 2/6/22 3:44 AM, Schmidhuber Juergen wrote: > Steve, it?s simple: the original ?shallow learning? (~1800) is much > older than your relatively recent ?shallow learning? references > (mostly from the 1900s). No need to mention all of them in this > report, which is really about "deep learning? (see title) with > adaptive hidden units, which started to work in the 1960s, first > through layer by layer training (USSR, 1965), then through stochastic > gradient descent (SGD) in relatively deep nets (Japan, 196I 7). The > reverse mode of automatic differentiation (now called backpropagation) > appeared 3 years later (Finland, 1970). No fancy talk about syntax vs > semantics can justify a revisionist history of deep learning that does > not mention these achievements. Cheers, J?rgen > > https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html > > > > > >> On 2 Feb 2022, at 00:48, Stephen Jos? Hanson > > wrote: >> >> Jeurgen:? Even some of us lowly psychologists know some math.?? And >> its not about the math.. its about the context (is this sounding like >> an echo?) >> >> Let me try again.. and I think your good intentions but misguided >> reconstruction of history is appearing to me, to be perverse. >> >> You tip your hand when you talk about "rebranding".??? Also that the >> PDP books were a "conspiracy".? But lets go point by point. >> >> (1) we already agreed that the Perceptron? was not linear >> regression--lets not go backwards. Closer to logistic regression.?? >> If you are talking about Widrow and Hoff, well it is the Delta Rule-- >> SSE kind of regression.?? But where did the Delta rule come from?? >> Lets look at Math.? So there is some nice papers by Gluck and >> Thompson (80s) showing how Pavlovian conditioning is exactly the >> Delta rule and even more relevant was shown to account for majority >> of classical (pavlovian) conditioning was the Rescorla-Wagner (1972) >> model-- \Delta V_A = [\alpha_A\beta_1](\lambda_1 - V_{AX}), which of >> course was Ivan Petrovich Pavlov discovery of classical conditioning >> (1880s).?? Why aren't you citing him??? What about John Brodeus >> Watson and Burris Fredrick Skinner???????? At least they were focused >> on learning albeit? *just* in biological systems.? But these were >> actual? natural world discoveries. >> >> (2) Function approximation.?? Ok Juergen, claims that everything?? is >> really?? just X, reminds me of the man with a Hammer to whom >> everything looks like a nail!????? To the point: its incidental.? >> Yes, Neural networks are function approximators, but that is >> incidental to the original more general context (PDP)? as a way to >> create "internal representations".?? The function approximation was a >> Bonus! >> >> (3) Branding. OMG.? So you seem to believe that everyone is cynical >> and will put their intellectual finger in the air to find out what to >> call what they are doing!?? Jeez, I hope this isn't true.? But the >> narrative you eschew is in fact something that Minsky would talk >> about (I remember this at lunch with him in the 90s at Thinking >> Machines), and he was quite clear that Perceptron was failing well >> before the 1969 book (trying to do speech recognition with a >> perceptron--yikes), but in a piling on kind of way Perceptrons killed >> the perceptron, but it was the linearity focus (as BAP points out) >> and the lack of depth. >> >> (4) Group Method of Handling Data.?? Frankly, the only one I can find >> that branded GMHD as a NeuroNet (as they call it) was you. >> There is a 2017 reference, but they reference you again. >> >> (5) Its just names,? fashion and preference..?? or no actual concepts >> matter.? Really? >> >> There was an french mathematician named Fourier in the 19th century >> who came up with an idea of periodic function decomposition into >> weighted trigonometric functions.. but he had no math.?? And Laplace >> Legendre and others said he had no math!? So they prevented him from >> publishing for FIFTEEN YEARS.. 150 years later after Tukey invented >> the FFT, its the most common transform used and misused? in general. >> >> Concepts lead to math.. and that may lead to further formalism.. but >> don't mistake the math for the concept behind it. The context matters >> and you are confusing syntax for semantics! >> >> Cheers, >> Steve >> >> >> >> On 1/31/22 11:38 AM, Schmidhuber Juergen wrote: >> >>> Steve, do you really want to erase the very origins of shallow learning (Gauss & Legendre ~1800) and deep learning (DL, Ivakhnenko & Lapa 1965) from the field's history? Why? Because they did not use modern terminology such as "artificial neural nets (NNs)" and "learning internal representations"? Names change all the time like fashions; the only thing that counts is the math. Not only mathematicians but also psychologists like yourself will agree. >>> >>> Again: the linear regressor of Legendre & Gauss is formally identical to what was much later called a linear NN for function approximation (FA), minimizing mean squared error, still widely used today. No history of "shallow learning" (without adaptive hidden layers) is complete without this original shallow learner of 2 centuries ago. Many NN courses actually introduce simple NNs in this mathematically and historically correct way, then proceed to DL NNs with several adaptive hidden layers. >>> >>> And of course, no DL history is complete without the origins of functional DL in 1965 [DEEP1-2]. Back then, Ivakhnenko and Lapa published the first general, working DL algorithm for supervised deep feedforward multilayer perceptrons (MLPs) with arbitrarily many layers of neuron-like elements, using nonlinear activation functions (actually Kolmogorov-Gabor polynomials) that combine both additions (like in linear NNs) and multiplications (basically they had deep NNs with gates, including higher order gates). They incrementally trained and pruned their DL networks layer by layer to learn internal representations, using regression and a separate validation set (network depth > 7 by 1971). They had standard justifications of DL such as: "a multilayered structure is a computationally feasible way to implement multinomials of very high degree" [DEEP2] (that cannot be approximated by simple linear NNs). Of course, their DL was automated, and many people have used it up to the 2000s ! >>> - just follow the numerous citations. >>> >>> I don't get your comments about Ivakhnenko's DL and function approximation (FA). FA is for all kinds of functions, including your "cognitive or perceptual or motor functions." NNs are used as FAs all the time. Like other NNs, Ivakhnenko's nets can be used as FAs for your motor control problems. You boldly claim: "This was not in the intellectual space" of Ivakhnenko's method. But obviously it was. >>> >>> Interestingly, 2 years later, Amari (1967-68) [GD1-2] trained his deep MLPs through a different DL method, namely, stochastic gradient descent (1951-52)[STO51-52]. His paper also did not contain the "modern" expression "learning internal representations in NNs." But that's what it was about. Math and algorithms are immune to rebranding. >>> >>> You may not like the fact that neither the original shallow learning (Gauss & Legendre ~1800) nor the original working DL (Ivakhnenko & Lapa 1965; Amari 1967) were biologically inspired. They were motivated through math and problem solving. The NN rebranding came later. Proper scientific credit assignment does not care for changes in terminology. >>> >>> BTW, unfortunately, Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions. But actually he had much more interesting MLPs with a non-learning randomized first layer and an adaptive output layer. So Rosenblatt basically had what much later was rebranded as "Extreme Learning Machines (ELMs)." The revisionist narrative of ELMs (see this web sitehttps://elmorigin.wixsite.com/originofelm) is a bit like the revisionist narrative of DL criticized by my report. Some ELM guys apparently thought they can get away with blatant improper credit assignment. After all, the criticized DL guys seemed to get away with it on an even grander scale. They called themselves the "DL conspiracy" [DLC]; the "ELM conspiracy" is similar. What an embarrassing lack of maturity of our field. >>> >>> Fortunately, more and more ML researchers are helping to set things straight. "In science, by definition, the facts will always win in the end. As long as the facts have not yet won it's not yet the end." [T21v1] >>> >>> References as always underhttps://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html >>> >>> J?rgen >>> >>> >>>> On 27 Jan 2022, at 17:37, Stephen Jos? Hanson wrote: >>>> >>>> >>>> >>>> Juergen, I have read through GMHD paper and a 1971 Review paper by Ivakhnenko. These are papers about function approximation. The method proposes to use series of polynomial functions that are stacked in filtered sets. The filtered sets are chosen based on best fit, and from what I can tell are manually grown.. so this must of been a tedious and slow process (I assume could be automated). So are the GMHDs "deep", in that they are stacked 4 deep in figure 1 (8 deep in another). Interestingly, they are using (with obvious FA justification) polynomials of various degree. Has this much to do with neural networks? Yes, there were examples initiated by Rumelhart (and me:https://www.routledge.com/Backpropagation-Theory-Architectures-and-Applications/Chauvin-Rumelhart/p/book/9780805812596), based on poly-synaptic dendrite complexity, but not in the GMHD paper.. which was specifically about function approximation. Ivakhnenko, lists four reasons for the approach t! >>> hey took: mainly reducing data size and being more efficient with data that one had. No mention of "internal representations" >>>> So when Terry, talks about "internal representations" --does he mean function approximation? Not so much. That of course is part of this, but the actual focus is on cognitive or perceptual or motor functions. Representation in the brain. Hidden units (which could be polynomials) cluster and project and model the input features wrt to the function constraints conditioned by training data. This is more similar to model specification through function space search. And the original Rumelhart meaning of internal representation in PDP vol 1, was in the case of representation certain binary functions (XOR), but more generally about the need for "neurons" (inter-neurons) explicitly between input (sensory) and output (motor). Consider NETTALK, in which I did the first hierarchical clustering of the hidden units over the input features (letters). What appeared wasn't probably surprising.. but without model specification, the network (w.hidden units), learned VOWELS and ! >>> CONSONANT distinctions just from training (Hanson & Burr, 1990). This would be a clear example of "internal representations" in the sense of Rumelhart. This was not in the intellectual space of Ivakhnenko's Group Method of Handling Data. (some of this is discussed in more detail in some recent conversations with Terry Sejnowski and another one to appear shortly with Geoff Hinton (AIHUB.org look in Opinions). >>>> Now I suppose one could be cynical and opportunistic, and even conclude if you wanted to get more clicks, rather than title your article GROUP METHOD OF HANDLING DATA, you should at least consider: NEURAL NETWORKS FOR HANDLING DATA, even if you didn't think neural networks had anything to do with your algorithm, after all everyone else is! Might get it published in this time frame, or even read. This is not scholarship. These publications threads are related but not dependent. And although they diverge they could be informative if one were to try and develop polynomial inductive growth networks (see Falhman, 1989; Cascade correlation and Hanson 1990: Meiosis nets) to motor control in the brain. But that's not what happened. I think, like Gauss, you need to drop this specific claim as well. >>>> >>>> With best regards, >>>> >>>> Steve >>> On 25 Jan 2022, at 20:03, Schmidhuber Juergen wrote: >>> >>> PS: Terry, you also wrote: "Our precious time is better spent moving the field forward.? However, it seems like in recent years much of your own precious time has gone to promulgating a revisionist history of deep learning (and writing the corresponding "amicus curiae" letters to award committees). For a recent example, your 2020 deep learning survey in PNAS [S20] claims that your 1985 Boltzmann machine [BM] was the first NN to learn internal representations. This paper [BM] neither cited the internal representations learnt by Ivakhnenko & Lapa's deep nets in 1965 [DEEP1-2] nor those learnt by Amari?s stochastic gradient descent for MLPs in 1967-1968 [GD1-2]. Nor did your recent survey [S20] attempt to correct this as good science should strive to do. On the other hand, it seems you celebrated your co-author's birthday in a special session while you were head of NeurIPS, instead of correcting these inaccuracies and celebrating the true pioneers of deep learning, such as ! >>> Ivakhnenko and Amari. Even your recent interviewhttps://blog.paperspace.com/terry-sejnowski-boltzmann-machines/ claims: "Our goal was to try to take a network with multiple layers - an input layer, an output layer and layers in between ? and make it learn. It was generally thought, because of early work that was done in AI in the 60s, that no one would ever find such a learning algorithm because it was just too mathematically difficult.? You wrote this although you knew exactly that such learning algorithms were first created in the 1960s, and that they worked. You are a well-known scientist, head of NeurIPS, and chief editor of a major journal. You must correct this. We must all be better than this as scientists. We owe it to both the past, present, and future scientists as well as those we ultimately serve. >>> >>> The last paragraph of my reporthttps://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html quotes Elvis Presley: "Truth is like the sun. You can shut it out for a time, but it ain't goin' away.? I wonder how the future will reflect on the choices we make now. >>> >>> J?rgen >>> >>> >>>> On 3 Jan 2022, at 11:38, Schmidhuber Juergen wrote: >>>> >>>> Terry, please don't throw smoke candles like that! >>>> >>>> This is not about basic math such as Calculus (actually first published by Leibniz; later Newton was also credited for his unpublished work; Archimedes already had special cases thereof over 2000 years ago; the Indian Kerala school made essential contributions around 1400). In fact, my report addresses such smoke candles in Sec. XII: "Some claim that 'backpropagation' is just the chain rule of Leibniz (1676) & L'Hopital (1696).' No, it is the efficient way of applying the chain rule to big networks with differentiable nodes (there are also many inefficient ways of doing this). It was not published until 1970 [BP1]." >>>> >>>> You write: "All these threads will be sorted out by historians one hundred years from now." To answer that, let me just cut and paste the last sentence of my conclusions: "However, today's scientists won't have to wait for AI historians to establish proper credit assignment. It is easy enough to do the right thing right now." >>>> >>>> You write: "let us be good role models and mentors" to the new generation. Then please do what's right! Your recent survey [S20] does not help. It's mentioned in my report as follows: "ACM seems to be influenced by a misleading 'history of deep learning' propagated by LBH & co-authors, e.g., Sejnowski [S20] (see Sec. XIII). It goes more or less like this: 'In 1969, Minsky & Papert [M69] showed that shallow NNs without hidden layers are very limited and the field was abandoned until a new generation of neural network researchers took a fresh look at the problem in the 1980s [S20].' However, as mentioned above, the 1969 book [M69] addressed a 'problem' of Gauss & Legendre's shallow learning (~1800)[DL1-2] that had already been solved 4 years prior by Ivakhnenko & Lapa's popular deep learning method [DEEP1-2][DL2] (and then also by Amari's SGD for MLPs [GD1-2]). Minsky was apparently unaware of this and failed to correct it later [HIN](Sec. I).... deep learning research was a! >>> live and kicking also in the 1970s, especially outside of the Anglosphere." >>>> Just follow ACM's Code of Ethics and Professional Conduct [ACM18] which states: "Computing professionals should therefore credit the creators of ideas, inventions, work, and artifacts, and respect copyrights, patents, trade secrets, license agreements, and other methods of protecting authors' works." No need to wait for 100 years. >>>> >>>> J?rgen >>>> >>>> >>>> >>>> >>>> >>>>> On 2 Jan 2022, at 23:29, Terry Sejnowski wrote: >>>>> >>>>> We would be remiss not to acknowledge that backprop would not be possible without the calculus, >>>>> so Isaac newton should also have been given credit, at least as much credit as Gauss. >>>>> >>>>> All these threads will be sorted out by historians one hundred years from now. >>>>> Our precious time is better spent moving the field forward. There is much more to discover. >>>>> >>>>> A new generation with better computational and mathematical tools than we had back >>>>> in the last century have joined us, so let us be good role models and mentors to them. >>>>> >>>>> Terry >> -- >> > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From iswc.conf at gmail.com Mon Feb 7 08:41:40 2022 From: iswc.conf at gmail.com (International Semantic Web Conference) Date: Mon, 7 Feb 2022 08:41:40 -0500 Subject: Connectionists: CfP ISWC 2022 - Call for Workshops and Tutorials Message-ID: *CfP: 21st International Semantic Web Conference (ISWC 2022)* Hangzhou, China, October 23-27, 2022 https://iswc2022.semanticweb.org/ The International Semantic Web Conference (ISWC) is the premier venue for presenting fundamental research, innovative technology, and applications concerning semantics, data, and the Web. It is the most important international venue to discuss and present the latest advances and applications of the semantic Web, knowledge graphs, linked data, ontologies and artificial intelligence (AI) on the Web. Follow us on social media: - Twitter: @iswc_conf #iswc_conf (https://twitter.com/iswc_conf) - LinkedIn: https://www.linkedin.com/groups/13612370 - Facebook: https://www.facebook.com/ISWConf - Instagram: https://www.instagram.com/iswc_conf/ Join ISWC 2022 by submitting to the following tracks and activities, or by attending the hybrid conference! Call for Workshops and Tutorials: https://iswc2022.semanticweb.org/index.php/workshops-and-tutorials/ Deadline: Friday, 18th March, 2022, 23:59 AoE (Anywhere on Earth) Workshops & Tutorials Chairs: - Marta Sabou, Vienna University of Economics and Business (WU), Austria marta.sabou at wu.ac.at - Raghava Mutharaju, IIIT-Delhi, India raghava.mutharaju at iiitd.ac.in The ISWC 2022 Organizing Team Organizing Committee ? ISWC 2022 (semanticweb.org) -------------- next part -------------- An HTML attachment was scrubbed... URL: From bwyble at gmail.com Mon Feb 7 09:40:37 2022 From: bwyble at gmail.com (Brad Wyble) Date: Mon, 7 Feb 2022 09:40:37 -0500 Subject: Connectionists: Please stop having conversations on this mailing list In-Reply-To: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> References: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> Message-ID: Agreed, I don't know what is the point of mailing lists except to encourage scientific exchange. The present discussions, while not always moving in a strictly forward direction, are a welcome break from the list of workshops, symposiums and opportunities that are the usual fare for connectionists. It's easy to skip over a thread if you don't find it useful. On Mon, Feb 7, 2022 at 9:23 AM Stephen Jos? Hanson wrote: > Pau, > > actually this was exactly what Connectionists was created for: civil, > thoughtful discussion about Neural Computation. > > It is very lightly moderated and provides a forum for discussion, > promotion and potential collaborations. > > This discussion has been a bit redundant here and there, but for the most > part reconstructing much of the arguments from the 1980s in the background > of the DL revolution. > > Explainable AI, Neuro-Symbolic models and causal structure have been > constant concerns of thoughtful folks using, building Neural network > systems for 40 years. > > The threads have gone in predictable ways, but have interesting voices and > real issues behind them. Some of this is dialetical (my friend Gary > Marcus who is a master at this) and some of it hyperbolic. But all have > made excellent and interesting points. > > Best Regards, > > Steve > On 2/5/22 11:29 AM, pau wrote: > > Dear connectionists, > > To the best of my understanding, the aim of this mailing list is to meet > the needs of working professionals. I understand by this sharing > information about events, publications, developments, etc. that can be > useful for someone who works in this field. > If someone wants to argue, discuss, praise or assign blame, agree or > disagree, please do so in private correspondence or in non-work-specific > social media channels. > > Best, > P. > > -- > -- Brad Wyble Associate Professor Psychology Department Penn State University http://wyblelab.com -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From tt at cs.dal.ca Mon Feb 7 09:29:05 2022 From: tt at cs.dal.ca (Thomas Trappenberg) Date: Mon, 7 Feb 2022 10:29:05 -0400 Subject: Connectionists: Please stop having conversations on this mailing list In-Reply-To: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> References: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> Message-ID: Just brief note (as an example) that I follow the discussion with interest and, admittedly, sometimes with amusement. I like that good points are raised in a forum that gives us a platform beyond announcements. It's easy to skip emails that are of less interest to us. Cheers, Thomas On Mon, Feb 7, 2022, 10:15 AM Stephen Jos? Hanson wrote: > Pau, > > actually this was exactly what Connectionists was created for: civil, > thoughtful discussion about Neural Computation. > > It is very lightly moderated and provides a forum for discussion, > promotion and potential collaborations. > > This discussion has been a bit redundant here and there, but for the most > part reconstructing much of the arguments from the 1980s in the background > of the DL revolution. > > Explainable AI, Neuro-Symbolic models and causal structure have been > constant concerns of thoughtful folks using, building Neural network > systems for 40 years. > > The threads have gone in predictable ways, but have interesting voices and > real issues behind them. Some of this is dialetical (my friend Gary > Marcus who is a master at this) and some of it hyperbolic. But all have > made excellent and interesting points. > > Best Regards, > > Steve > On 2/5/22 11:29 AM, pau wrote: > > Dear connectionists, > > To the best of my understanding, the aim of this mailing list is to meet > the needs of working professionals. I understand by this sharing > information about events, publications, developments, etc. that can be > useful for someone who works in this field. > If someone wants to argue, discuss, praise or assign blame, agree or > disagree, please do so in private correspondence or in non-work-specific > social media channels. > > Best, > P. > > -- > -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From i.jonathan.grainger at gmail.com Mon Feb 7 10:14:58 2022 From: i.jonathan.grainger at gmail.com (Jonathan Grainger) Date: Mon, 7 Feb 2022 16:14:58 +0100 Subject: Connectionists: Please stop having conversations on this mailing list In-Reply-To: References: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> Message-ID: I am a humble reading researcher (psychologist) interested in how much processing readers can perform in parallel - and I've found it interesting to see how having this connectionist thread open in a different window from my main work has helped me develop that kind of parallel processing :-) it's great stuff - keep going ... Jonathan Grainger On Mon, 7 Feb 2022 at 16:05, Brad Wyble wrote: > Agreed, I don't know what is the point of mailing lists except to > encourage scientific exchange. The present discussions, while not always > moving in a strictly forward direction, are a welcome break from the list > of workshops, symposiums and opportunities that are the usual fare for > connectionists. It's easy to skip over a thread if you don't find it > useful. > > On Mon, Feb 7, 2022 at 9:23 AM Stephen Jos? Hanson > wrote: > >> Pau, >> >> actually this was exactly what Connectionists was created for: civil, >> thoughtful discussion about Neural Computation. >> >> It is very lightly moderated and provides a forum for discussion, >> promotion and potential collaborations. >> >> This discussion has been a bit redundant here and there, but for the most >> part reconstructing much of the arguments from the 1980s in the background >> of the DL revolution. >> >> Explainable AI, Neuro-Symbolic models and causal structure have been >> constant concerns of thoughtful folks using, building Neural network >> systems for 40 years. >> >> The threads have gone in predictable ways, but have interesting voices >> and real issues behind them. Some of this is dialetical (my friend Gary >> Marcus who is a master at this) and some of it hyperbolic. But all have >> made excellent and interesting points. >> >> Best Regards, >> >> Steve >> On 2/5/22 11:29 AM, pau wrote: >> >> Dear connectionists, >> >> To the best of my understanding, the aim of this mailing list is to meet >> the needs of working professionals. I understand by this sharing >> information about events, publications, developments, etc. that can be >> useful for someone who works in this field. >> If someone wants to argue, discuss, praise or assign blame, agree or >> disagree, please do so in private correspondence or in non-work-specific >> social media channels. >> >> Best, >> P. >> >> -- >> > > > -- > Brad Wyble > Associate Professor > Psychology Department > Penn State University > > http://wyblelab.com > -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From cristian.grozea at fokus.fraunhofer.de Mon Feb 7 10:16:15 2022 From: cristian.grozea at fokus.fraunhofer.de (Cristian Grozea) Date: Mon, 7 Feb 2022 16:16:15 +0100 Subject: Connectionists: Please stop having conversations on this mailing list In-Reply-To: References: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> Message-ID: <896b92d4-363d-9aa7-5d1d-988816314add@fokus.fraunhofer.de> I monitor with much interest the discussion, and am glad I have the chance to get such nice summaries of interesting research work that I can then examine in details later. It is certainly the most interesting thing that happened in this list in a very long time, and I am aware of no equivalent or alternative forum where this mostly scientific discussion could take place. Therefore, please continue. best regards, Cristian Grozea -- Dr. Cristian Grozea Visual Computing (VISCOM) Tel: +49 30 3463 ? 7492 Fax: +49 30 3463 ? 99 - 7492 cristian.grozea at fokus.fraunhofer.de Fraunhofer Institute for Open Communication Systems FOKUS Kaiserin-Augusta-Allee 31 10589 Berlin Germany http://www.fokus.fraunhofer.de 2/7/22 3:29 PM, Thomas Trappenberg wrote: > Just brief note (as an example) that I follow the discussion with > interest and, admittedly, sometimes with amusement. I like that good > points are raised in a forum that gives us a platform beyond > announcements. It's easy to skip emails that are of less interest to us. > > Cheers, Thomas > > On Mon, Feb 7, 2022, 10:15 AM Stephen Jos? Hanson > > wrote: > > Pau, > > actually this was exactly what Connectionists was created for:? > civil, thoughtful discussion about Neural Computation. > > It is very lightly moderated and provides a forum for discussion, > promotion and potential collaborations. > > This discussion has been a bit redundant here and there, but for > the most part reconstructing much of the arguments from the 1980s > in the background of the DL revolution. > > Explainable AI, Neuro-Symbolic models and causal structure have > been constant concerns of thoughtful folks using, building Neural > network systems for 40 years. > > The threads have gone in predictable ways, but have interesting > voices and real issues behind them.?? Some of this is dialetical > (my friend Gary Marcus who is a master at this) and some of it > hyperbolic.?? But all have made excellent and interesting points. > > Best Regards, > > Steve > > On 2/5/22 11:29 AM, pau wrote: >> Dear connectionists, >> >> To the best of my understanding, the aim of this mailing list is to meet >> the needs of working professionals. I understand by this sharing >> information about events, publications, developments, etc. that can be >> useful for someone who works in this field. >> If someone wants to argue, discuss, praise or assign blame, agree or >> disagree, please do so in private correspondence or in non-work-specific >> social media channels. >> >> Best, >> P. > -- > -------------- next part -------------- An HTML attachment was scrubbed... URL: From mr.guang.yang at gmail.com Mon Feb 7 10:59:45 2022 From: mr.guang.yang at gmail.com (Guang Yang) Date: Mon, 7 Feb 2022 15:59:45 +0000 Subject: Connectionists: MIUA 2022 Call for Papers Message-ID: You are invited to submit your abstract paper to MIUA 2022 ( https://www.miua2022.com/) which will be held in Cambridge, UK. The abstract paper submission system is now open. https://cmt3.research.microsoft.com/User/Login?ReturnUrl=%2FMIUA2022 The full paper submission deadline will be 23:59, Greenwich Mean Time (GMT), on 1st April 2022. The abstract paper submission deadline will be 23:59, Greenwich Mean Time (GMT), on 31st May 2022. Selected extended papers will be invited for publication in two special issues in: COMPUTERIZED MEDICAL IMAGING AND GRAPHICS (IMPACT FACTOR: 4.790) INTERNATIONAL JOURNAL OF IMAGING SYSTEMS AND TECHNOLOGY (IMPACT FACTOR: 2.000) High-quality papers are requested, containing original contributions to the topics within the scope of MIUA. Paper Submissions: For the 26th MIUA conference, we welcome submissions, as regular conference papers and conference abstracts. Regular papers: Authors are invited to submit full papers of length between 8 and 15 pages (1 column ? the LNCS template) showing original research contributions under the topics of the conference. All submissions will be double-blind peer-reviewed and accepted articles will be published as MIUA Proceedings by the Springer Publishing Group. Conference abstracts: Authors are invited to submit short papers of length up to 3 pages excluding references (1 column ? the LNCS template) showing proof-of-concept research contributions under the topics of the conference. All submissions will be peer-reviewed and accepted articles will be published as MIUA Abstract Proceedings on the MIUA website. Accepted papers will be published in the MIUA proceedings in the Springer LNCS Series. Authors should consult Springer?s authors? guidelines and use their proceedings templates, either for LaTeX or for Word, for the preparation of their papers. Springer encourages authors to include their ORCIDs in their papers. In addition, the corresponding author of each paper, acting on behalf of all of the authors of that paper, must complete and sign a Consent-to-Publish form. The corresponding author signing the copyright form should match the corresponding author marked on the paper. Once the files have been sent to Springer, changes relating to the authorship of the papers cannot be made. MIUA continues to foster fairness, diversity, and inclusion within its community. Submissions from typically underrepresented groups are particularly encouraged. Review Process MIUA seeks your assistance as an expert reviewer for this annual conference. If you would like to review for MIUA, please contact us via: gateway at newton.ac.uk or twitter @miua2022. Please share this CfP with colleagues who might want to contribute to MIUA2022. Dr Guang Yang (Ph.D., M.Sc,, B.Eng., SM.IEEE, M.BMVA) Pronouns: he/him/his UKRI Future Leaders Fellow National Heart & Lung Institute, Imperial College London Group:// www.yanglab.fyi/ Email:// g.yang at imperial.ac.uk -------------- next part -------------- An HTML attachment was scrubbed... URL: From mr.guang.yang at gmail.com Mon Feb 7 10:59:45 2022 From: mr.guang.yang at gmail.com (Guang Yang) Date: Mon, 7 Feb 2022 15:59:45 +0000 Subject: Connectionists: MIUA 2022 Call for Papers Message-ID: You are invited to submit your abstract paper to MIUA 2022 ( https://www.miua2022.com/) which will be held in Cambridge, UK. The abstract paper submission system is now open. https://cmt3.research.microsoft.com/User/Login?ReturnUrl=%2FMIUA2022 The full paper submission deadline will be 23:59, Greenwich Mean Time (GMT), on 1st April 2022. The abstract paper submission deadline will be 23:59, Greenwich Mean Time (GMT), on 31st May 2022. Selected extended papers will be invited for publication in two special issues in: COMPUTERIZED MEDICAL IMAGING AND GRAPHICS (IMPACT FACTOR: 4.790) INTERNATIONAL JOURNAL OF IMAGING SYSTEMS AND TECHNOLOGY (IMPACT FACTOR: 2.000) High-quality papers are requested, containing original contributions to the topics within the scope of MIUA. Paper Submissions: For the 26th MIUA conference, we welcome submissions, as regular conference papers and conference abstracts. Regular papers: Authors are invited to submit full papers of length between 8 and 15 pages (1 column ? the LNCS template) showing original research contributions under the topics of the conference. All submissions will be double-blind peer-reviewed and accepted articles will be published as MIUA Proceedings by the Springer Publishing Group. Conference abstracts: Authors are invited to submit short papers of length up to 3 pages excluding references (1 column ? the LNCS template) showing proof-of-concept research contributions under the topics of the conference. All submissions will be peer-reviewed and accepted articles will be published as MIUA Abstract Proceedings on the MIUA website. Accepted papers will be published in the MIUA proceedings in the Springer LNCS Series. Authors should consult Springer?s authors? guidelines and use their proceedings templates, either for LaTeX or for Word, for the preparation of their papers. Springer encourages authors to include their ORCIDs in their papers. In addition, the corresponding author of each paper, acting on behalf of all of the authors of that paper, must complete and sign a Consent-to-Publish form. The corresponding author signing the copyright form should match the corresponding author marked on the paper. Once the files have been sent to Springer, changes relating to the authorship of the papers cannot be made. MIUA continues to foster fairness, diversity, and inclusion within its community. Submissions from typically underrepresented groups are particularly encouraged. Review Process MIUA seeks your assistance as an expert reviewer for this annual conference. If you would like to review for MIUA, please contact us via: gateway at newton.ac.uk or twitter @miua2022. Please share this CfP with colleagues who might want to contribute to MIUA2022. Dr Guang Yang (Ph.D., M.Sc,, B.Eng., SM.IEEE, M.BMVA) Pronouns: he/him/his UKRI Future Leaders Fellow National Heart & Lung Institute, Imperial College London Group:// www.yanglab.fyi/ Email:// g.yang at imperial.ac.uk -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Mon Feb 7 10:58:26 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Mon, 7 Feb 2022 10:58:26 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Asim, The biggest trap is to avoid brain problems and only work on so called "an engineering problem" or "a focused problem". In a cluttered scene, there maybe n>10 objects and m>> 1000 parts. A brain model must explain how the skull-closed brain segments objects and their parts. The traditional and also your "recognition by parts" approach does not explain (1) how different objects are segmented and (2) how different parts are segmented and group into objects. Note "skull-closed" condition. No supervision into the brain network is allowed. My conscious learning model using DN addresses such compounding problems while skull is closed (using unsupervised Hebbian learning) and without a given task. J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, USA, Jan.7-9, 2022. http://www.cse.msu.edu/~weng/research/ConsciousLearning-ICCE-2022-rvsd-cite.pdf YouTube talk: [image: image.png] My model must not address the above problems only. Best regards, -John On Sun, Feb 6, 2022 at 10:43 PM Asim Roy wrote: > Dear John, > > > > We recognize whole objects, but at the same time we verify its parts. > > > > Best, > > Asim > > > > *From:* Juyang Weng > *Sent:* Sunday, February 6, 2022 8:38 PM > *To:* Asim Roy > *Cc:* Geoffrey Hinton ; Dietterich, Thomas < > tgd at oregonstate.edu>; AIhub ; > connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; > Danko Nikolic ; Stephen Jos? Hanson < > jose at rubic.rutgers.edu>; Marek Reformat ; MARCO > GORI ; Alessandro Sperduti < > alessandro.sperduti at unipd.it>; Xiaodong Li ; > Hava Siegelmann ; Peter Tino < > P.Tino at cs.bham.ac.uk>; Bing Xue ; Ali Minai < > minaiaa at gmail.com>; Claudius Gros ; > Jean-Philippe Thivierge ; Tsvi Achler > ; Prof A Hussain > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Asim, > > The brain does not assume a single object in a cluttered science. Thus, a > simple explanation like "recognition by parts" (but without object > abstraction) should be invalid. It is like a chicken and egg problem. > Both chicken and egg are absent. We must not assume egg is there or > chicken is there. > > Best regards, > > -John > > > > On Sun, Feb 6, 2022 at 2:42 PM Asim Roy wrote: > > Dear John, > > > > You are right and I admit I am not solving all of the problems. It?s just > in reference to this one problem that Geoffrey Hinton mentions that I think > can be resolved: > > *?I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. We can introspect on how we do it > and this may or may not give some insight into how we check our answer, but > the immediate sense that a handwritten 2 is a 2 is computed by a neural net > that is not functionally equivalent to any simple and easily explainable > procedure.?* > > Best, > > Asim > > > > *From:* Juyang Weng > *Sent:* Sunday, February 6, 2022 10:06 AM > *To:* Asim Roy > *Cc:* Geoffrey Hinton ; Dietterich, Thomas < > tgd at oregonstate.edu>; AIhub ; > connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; > Danko Nikolic ; Stephen Jos? Hanson < > jose at rubic.rutgers.edu>; Marek Reformat ; MARCO > GORI ; Alessandro Sperduti < > alessandro.sperduti at unipd.it>; Xiaodong Li ; > Hava Siegelmann ; Peter Tino < > P.Tino at cs.bham.ac.uk>; Bing Xue ; Ali Minai < > minaiaa at gmail.com>; Claudius Gros ; > Jean-Philippe Thivierge ; Tsvi Achler > ; Prof A Hussain > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Asim, > > I try to be brief so that I can explain why many of us have missed, and > will continue to miss, the boat. > In some of my talks, I have a ppt slide "The brain is like blindmen and an > elephant". > > Unfortunately, your "identify objects based on its parts" is a good > traditional idea from pattern recognition that is still a blindman. > > Your idea does not explain many other problems without which we will never > understand a biological brain. > > For example, your idea does not explain how the brain learns planning and > discovery in a cluttered world. > > We must solve many million-dollar problems holistically. Please watch my > YouTube video: > Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar > Problems Solved > https://youtu.be/Dgx1dLCdSKY > > > Best regards, > > -John > > > > On Sat, Feb 5, 2022 at 12:01 AM Asim Roy wrote: > > I am responding to this part of Geoffrey Hinton?s note: > > > > *?I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. We can introspect on how we do it > and this may or may not give some insight into how we check our answer, but > the immediate sense that a handwritten 2 is a 2 is computed by a neural net > that is not functionally equivalent to any simple and easily explainable > procedure.?* > > > > The causal explanation is actually done quite simply, and we are doing it > currently. I can talk about this now because Arizona State University (ASU) > has filed a provisional patent application on the technology. The basic > idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable > Artificial Intelligence (darpa.mil) > ) > and illustrated in the figure below. The idea is to identify objects based > on its parts. So, the figure below says that it?s a cat because it has fur, > whiskers, and claws plus an unlabeled visual feature. I am not sure if > DARPA got anything close to this from its funding of various entities. What > this means is that you need a parts model. And we do that. In the case of > MNIST handwritten digits that Geoff mentions, we ?teach? this parts model > what the top part of a digit ?3? looks like, what the bottom part looks > like and so on. And we also teach connectivity between parts and the > composition of objects from parts. And we do that for all digits. And we > get a symbolic model sitting on top of a CNN model that provides the > explanation that Geoff is referring to as the causal explanation. This > ?teaching? is similar to the way you would teach a kid to recognize > different digits. > > > > An advantage of this parts model, in addition to being in an explainable > symbolic form, is robustness to adversarial attack. We recently tested on > the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast > gradient method to 27%, our XAI model maintained an accuracy of 90%, > probably higher. In general, it would be hard to make a school bus look > like an ostrich, with a few pixel changes, if you can identify the parts of > a school bus and an ostrich. > > > > A parts model that DARPA wanted provides both a symbolic explanation and > adversarial protection. The problem that Geoffrey is referring to is solved. > > > > I am doing a tutorial on this at IEEE World Congress on Computational > Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy > 18-23 July > ). > I am copying the organizers and want to thank them for accepting the > tutorial proposal. The only other presentation I have done on this is at a > Military Operations Research Society (MORS) meeting last December. > > > > So, back to the future. Hybrid models might indeed save deep learning > models and let us deploy these models without concern. We might not even > need adversarial training of any kind. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > www.teuvonet.com > > > > > [image: Timeline Description automatically generated] > > > > *From:* Connectionists *On > Behalf Of *Geoffrey Hinton > *Sent:* Friday, February 4, 2022 1:24 PM > *To:* Dietterich, Thomas > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. We can introspect on how we do it > and this may or may not give some insight into how we check our answer, but > the immediate sense that a handwritten 2 is a 2 is computed by a neural net > that is not functionally equivalent to any simple and easily explainable > procedure. > > > > This does not mean that we should give up on trying to make artificial > neural nets work more like real ones. People can see a tilted square as > either an upright diamond or a tilted square and, so far as I know, a > convnet does not exhibit this type of alternate percept. People seem to > impose hierarchical structural descriptions on images and sound waves and > they clearly impose intrinsic coordinate frames on wholes and parts. If > this is what Gary means by symbolic then I don?t disagree that neural nets > should do symbol processing. However, there is a very different meaning of > "symbolic". A pure atomic symbol has no internal structure. The form of the > symbol itself tells you nothing about what it denotes. The only relevant > properties it has are that it's identical to other instances of the > same symbol and different from all other symbols. That's totally different > from a neural net that uses embedding vectors. Embedding vectors have a > rich internal structure that dictates how they interact with other > embedding vectors. What I really object to is the following approach: Start > with pure symbols and rules for how to manipulate structures made out of > pure symbols. These structures themselves can be denoted by symbols that > correspond to memory addresses where the bits in the address tell you > nothing about the content of the structure at that address. Then when the > rule-based approach doesn't work for dealing with the real world (e.g. > machine translation) try to use neural nets to convert the real world into > pure symbols and then carry on with the rule-based approach. That is like > using an electric motor to inject the gasoline into the same old gasoline > engine instead of just replacing the gasoline engine with an electric motor. > > > > > > On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: > > ?Understanding? is not a Boolean. It is a theorem that no system can > enumerate all of the consequences of a state of affairs in the world. > > > > For low-stakes application work, we can be satisfied by a system that > ?does the right thing?. If the system draws a good picture, that?s > sufficient. It ?understood? the request. > > > > But for higher-stakes applications---and for advancing the science---we > seek a causal account of how the components of a system cause it to do the > right thing. We are hoping that a small set of mechanisms can produce broad > coverage of intelligent behavior. This gives us confidence that the system > will respond correctly outside of the narrow tasks on which we have tested > it. > > > > --Tom > > > > Thomas G. Dietterich, Distinguished Professor Emeritus > > School of Electrical Engineering and Computer > Science > > US Mail: 1148 Kelley Engineering Center > > > > Office: 2067 Kelley Engineering Center > > Oregon State Univ., Corvallis, OR 97331-5501 > > Voice: 541-737-5559; FAX: 541-737-1300 > > URL: http://web.engr.oregonstate.edu/~tgd/ > > > > > *From:* Connectionists *On > Behalf Of *Gary Marcus > *Sent:* Thursday, February 3, 2022 8:26 AM > *To:* Danko Nikolic > *Cc:* connectionists at mailman.srv.cs.cmu.edu; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > [This email originated from outside of OSU. Use caution with links and > attachments.] > > Dear Danko, > > > > Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, > in which he said (paraphrasing from memory, because I don?t remember the > precise words) that the famous 200 Quoc Le unsupervised model [ > https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf > ] > had learned the concept of a ca. In reality the model had clustered > together some catlike images based on the image statistics that it had > extracted, but it was a long way from a full, counterfactual-supporting > concept of a cat, much as you describe below. > > > > I fully agree with you that the reason for even having a semantics is as > you put it, "to 1) learn with a few examples and 2) apply the knowledge to > a broad set of situations.? GPT-3 sometimes gives the appearance of having > done so, but it falls apart under close inspection, so the problem remains > unsolved. > > > > Gary > > > > On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: > > > > G. Hinton wrote: "I believe that any reasonable person would admit that if > you ask a neural net to draw a picture of a hamster wearing a red hat and > it draws such a picture, it understood the request." > > > > I would like to suggest why drawing a hamster with a red hat does not > necessarily imply understanding of the statement "hamster wearing a red > hat". > > To understand that "hamster wearing a red hat" would mean inferring, in > newly emerging situations of this hamster, all the real-life > implications that the red hat brings to the little animal. > > > > What would happen to the hat if the hamster rolls on its back? (Would the > hat fall off?) > > What would happen to the red hat when the hamster enters its lair? (Would > the hat fall off?) > > What would happen to that hamster when it goes foraging? (Would the red > hat have an influence on finding food?) > > What would happen in a situation of being chased by a predator? (Would it > be easier for predators to spot the hamster?) > > > > ...and so on. > > > > Countless many questions can be asked. One has understood "hamster wearing > a red hat" only if one can answer reasonably well many of such real-life > relevant questions. Similarly, a student has understood materias in a class > only if they can apply the materials in real-life situations (e.g., > applying Pythagora's theorem). If a student gives a correct answer to a > multiple choice question, we don't know whether the student understood the > material or whether this was just rote learning (often, it is rote > learning). > > > > I also suggest that understanding also comes together with effective > learning: We store new information in such a way that we can recall it > later and use it effectively i.e., make good inferences in newly emerging > situations based on this knowledge. > > > > In short: Understanding makes us humans able to 1) learn with a few > examples and 2) apply the knowledge to a broad set of situations. > > > > No neural network today has such capabilities and we don't know how to > give them such capabilities. Neural networks need large amounts of > training examples that cover a large variety of situations and then > the networks can only deal with what the training examples have already > covered. Neural networks cannot extrapolate in that 'understanding' sense. > > > > I suggest that understanding truly extrapolates from a piece of knowledge. > It is not about satisfying a task such as translation between languages or > drawing hamsters with hats. It is how you got the capability to complete > the task: Did you only have a few examples that covered something different > but related and then you extrapolated from that knowledge? If yes, this is > going in the direction of understanding. Have you seen countless examples > and then interpolated among them? Then perhaps it is not understanding. > > > > So, for the case of drawing a hamster wearing a red hat, understanding > perhaps would have taken place if the following happened before that: > > > > 1) first, the network learned about hamsters (not many examples) > > 2) after that the network learned about red hats (outside the context of > hamsters and without many examples) > > 3) finally the network learned about drawing (outside of the context of > hats and hamsters, not many examples) > > > > After that, the network is asked to draw a hamster with a red hat. If it > does it successfully, maybe we have started cracking the problem of > understanding. > > > > Note also that this requires the network to learn sequentially without > exhibiting catastrophic forgetting of the previous knowledge, which is > possibly also a consequence of human learning by understanding. > > > > > > Danko > > > > > > > > > > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > Virus-free. www.avast.com > > > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: > > > Without getting into the specific dispute between Gary and Geoff, I think > with approaches similar to GLOM, we are finally headed in the right > direction. There?s plenty of neurophysiological evidence for single-cell > abstractions and multisensory neurons in the brain, which one might claim > correspond to symbols. And I think we can finally reconcile the decades old > dispute between Symbolic AI and Connectionism. > > > > > > GARY: (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > GARY: I have *never* called for dismissal of neural networks, but rather > for some hybrid between the two (as you yourself contemplated in 1991); the > point of the 2001 book was to characterize exactly where multilayer > perceptrons succeeded and broke down, and where symbols could complement > them. > > > > > > Asim Roy > > > Professor, Information Systems > > > Arizona State University > > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > > > > > > > *From: Connectionists On > Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: > Geoffrey Hinton Cc: AIhub ; > connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen > Hanson in conversation with Geoff Hinton > * > > > > > Dear Geoff, and interested others, > > > > > > What, for example, would you make of a system that often drew the > red-hatted hamster you requested, and perhaps a fifth of the time gave you > utter nonsense? Or say one that you trained to create birds but sometimes > output stuff like this: > > > > > > > > > > > > One could > > > > > > a. avert one?s eyes and deem the anomalous outputs irrelevant > > > or > > > b. wonder if it might be possible that sometimes the system gets the right > answer for the wrong reasons (eg partial historical contingency), and > wonder whether another approach might be indicated. > > > > > > Benchmarks are harder than they look; most of the field has come to > recognize that. The Turing Test has turned out to be a lousy measure of > intelligence, easily gamed. It has turned out empirically that the Winograd > Schema Challenge did not measure common sense as well as Hector might have > thought. (As it happens, I am a minor coauthor of a very recent review on > this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no > way means machines now have common sense; many people from many different > perspectives recognize that (including, e.g., Yann LeCun, who generally > tends to be more aligned with you than with me). > > > > > > So: on the goalpost of the Winograd schema, I was wrong, and you can quote > me; but what you said about me and machine translation remains your > invention, and it is inexcusable that you simply ignored my 2019 > clarification. On the essential goal of trying to reach meaning and > understanding, I remain unmoved; the problem remains unsolved. > > > > > > All of the problems LLMs have with coherence, reliability, truthfulness, > misinformation, etc stand witness to that fact. (Their persistent inability > to filter out toxic and insulting remarks stems from the same.) I am hardly > the only person in the field to see that progress on any given benchmark > does not inherently mean that the deep underlying problems have solved. > You, yourself, in fact, have occasionally made that point. > > > > > > With respect to embeddings: Embeddings are very good for natural language > *processing*; but NLP is not the same as NL*U* ? when it comes to > *understanding*, their worth is still an open question. Perhaps they will > turn out to be necessary; they clearly aren?t sufficient. In their extreme, > they might even collapse into being symbols, in the sense of uniquely > identifiable encodings, akin to the ASCII code, in which a specific set of > numbers stands for a specific word or concept. (Wouldn?t that be ironic?) > > > > > > (Your GLOM, which as you know I praised publicly, is in many ways an > effort to wind up with encodings that effectively serve as symbols in > exactly that way, guaranteed to serve as consistent representations of > specific concepts.) > > > > > > Notably absent from your email is any kind of apology for misrepresenting > my position. It?s fine to say that ?many people thirty years ago once > thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. > I have consistently felt throughout our interactions that you have mistaken > me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me > for having made that error. I am still not he. > > > > > > Which maybe connects to the last point; if you read my work, you would see > thirty years of arguments *for* neural networks, just not in the way that > you want them to exist. I have ALWAYS argued that there is a role for them; > characterizing me as a person ?strongly opposed to neural networks? misses > the whole point of my 2001 book, which was subtitled ?Integrating > Connectionism and Cognitive Science.? > > > > > > In the last two decades or so you have insisted (for reasons you have > never fully clarified, so far as I know) on abandoning symbol-manipulation, > but the reverse is not the case: I have *never* called for dismissal of > neural networks, but rather for some hybrid between the two (as you > yourself contemplated in 1991); the point of the 2001 book was to > characterize exactly where multilayer perceptrons succeeded and broke down, > and where symbols could complement them. It?s a rhetorical trick (which is > what the previous thread was about) to pretend otherwise. > > > > > > Gary > > > > > > > > > On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: > > > ? > > > Embeddings are just vectors of soft feature detectors and they are very > good for NLP. The quote on my webpage from Gary's 2015 chapter implies the > opposite. > > > > > > A few decades ago, everyone I knew then would have agreed that the ability > to translate a sentence into many different languages was strong evidence > that you understood it. > > > > > > But once neural networks could do that, their critics moved the goalposts. > An exception is Hector Levesque who defined the goalposts more sharply by > saying that the ability to get pronoun references correct in Winograd > sentences is a crucial test. Neural nets are improving at that but still > have some way to go. Will Gary agree that when they can get pronoun > references correct in Winograd sentences they really do understand? Or does > he want to reserve the right to weasel out of that too? > > > > > > Some people, like Gary, appear to be strongly opposed to neural networks > because they do not fit their preconceived notions of how the mind should > work. > > > I believe that any reasonable person would admit that if you ask a neural > net to draw a picture of a hamster wearing a red hat and it draws such a > picture, it understood the request. > > > > > > Geoff > > > > > > > > > > > > > > > > > > On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: > > > Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural > network community, > > > > > > There has been a lot of recent discussion on this list about framing and > scientific integrity. Often the first step in restructuring narratives is > to bully and dehumanize critics. The second is to misrepresent their > position. People in positions of power are sometimes tempted to do this. > > > > > > The Hinton-Hanson interview that you just published is a real-time example > of just that. It opens with a needless and largely content-free personal > attack on a single scholar (me), with the explicit intention of > discrediting that person. Worse, the only substantive thing it says is > false. > > > > > > Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t > be able to do machine translation.? > > > > > > I never said any such thing. > > > > > > What I predicted, rather, was that multilayer perceptrons, as they existed > then, would not (on their own, absent other mechanisms) *understand* language. > Seven years later, they still haven?t, except in the most superficial way. > > > > > > > I made no comment whatsoever about machine translation, which I view as a > separate problem, solvable to a certain degree by correspondance without > semantics. > > > > > > I specifically tried to clarify Hinton?s confusion in 2019, but, > disappointingly, he has continued to purvey misinformation despite that > clarification. Here is what I wrote privately to him then, which should > have put the matter to rest: > > > > > > You have taken a single out of context quote [from 2015] and > misrepresented it. The quote, which you have prominently displayed at the > bottom on your own web page, says: > > > > > > Hierarchies of features are less suited to challenges such as language, > inference, and high-level planning. For example, as Noam Chomsky famously > pointed out, language is filled with sentences you haven't seen > before. Pure classifier systems don't know what to do with such sentences. > The talent of feature detectors -- in identifying which member of some > category something belongs to -- doesn't translate into understanding > novel sentences, in which each sentence has its own unique meaning. > > > > > > It does *not* say "neural nets would not be able to deal with novel > sentences"; it says that hierachies of features detectors (on their own, if > you read the context of the essay) would have trouble *understanding *novel sentences. > > > > > > > Google Translate does yet not *understand* the content of the sentences > is translates. It cannot reliably answer questions about who did what to > whom, or why, it cannot infer the order of the events in paragraphs, it > can't determine the internal consistency of those events, and so forth. > > > > > > Since then, a number of scholars, such as the the computational linguist > Emily Bender, have made similar points, and indeed current LLM difficulties > with misinformation, incoherence and fabrication all follow from these > concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter > with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, > also emphasizing issues of understanding and meaning: > > > > > > *The success of the large neural language models on many NLP tasks is > exciting. However, we find that these successes sometimes lead to hype in > which these models are being described as ?understanding? language or > capturing ?meaning?. In this position paper, we argue that a system trained > only on form has a priori no way to learn meaning. .. a clear understanding > of the distinction between form and meaning will help guide the field > towards better science around natural language understanding. > * > > > > > Her later article with Gebru on language models ?stochastic parrots? is in > some ways an extension of this point; machine translation requires mimicry, > true understanding (which is what I was discussing in 2015) requires > something deeper than that. > > > > > > Hinton?s intellectual error here is in equating machine translation with > the deeper comprehension that robust natural language understanding will > require; as Bender and Koller observed, the two appear not to be the same. > (There is a longer discussion of the relation between language > understanding and machine translation, and why the latter has turned out to > be more approachable than the former, in my 2019 book with Ernest Davis). > > > > > > More broadly, Hinton?s ongoing dismissiveness of research from > perspectives other than his own (e.g. linguistics) have done the field a > disservice. > > > > > > As Herb Simon once observed, science does not have to be zero-sum. > > > > > > Sincerely, > > > Gary Marcus > > > Professor Emeritus > > > New York University > > > > > > On Feb 2, 2022, at 06:12, AIhub wrote: > > > ? > > > Stephen Hanson in conversation with Geoff Hinton > > > > > > In the latest episode of this video series for AIhub.org, Stephen Hanson > talks to Geoff Hinton about neural networks, backpropagation, > overparameterization, digit recognition, voxel cells, syntax and semantics, > Winograd sentences, and more. > > > > > > You can watch the discussion, and read the transcript, here: > > > > > https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ > > > > > > About AIhub: > > > AIhub is a non-profit dedicated to connecting the AI community to the > public by providing free, high-quality information through AIhub.org ( > https://aihub.org/). We help researchers publish the latest AI news, > summaries of their work, opinion pieces, tutorials and more. We are > supported by many leading scientific organizations in AI, namely AAAI, > NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. > > > Twitter: @aihuborg > > > > > > Virus-free. www.avast.com > > > > > > > > > -- > > Juyang (John) Weng > > > > > -- > > Juyang (John) Weng > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: not available URL: From n.langer at psychologie.uzh.ch Mon Feb 7 10:52:05 2022 From: n.langer at psychologie.uzh.ch (Nicolas Langer) Date: Mon, 7 Feb 2022 16:52:05 +0100 Subject: Connectionists: PhD position in Cognitive Neuroscience Message-ID: <0023663B-6BFB-4AC6-81C5-6EF4968AF690@psychologie.uzh.ch> PhD position in Cognitive Neuroscience Department: Department of Psychology at the University of Zurich (Switzerland), Methods of Plasticity Research (http://www.psychology.uzh.ch/en/chairs/plafor.html ) Position: PhD student (salary 54?000 USD/year) Description of UZH unit: Our lab develops novel methodological approaches to study variations in cognitive performance across the lifespan and along the continuum from healthy to pathological functioning. Specifically, we investigate the potential for plasticity, mechanisms for stabilization and compensation across the lifespan. For this, we acquire and analyze multimodal data sets, such as structural MRI, diffusion weighted data (DWI), simultaneous EEG and eye-tracking as well as behavioral data. From these rich data sets, we extract multivariate parameters and apply state-of-the-art methods, such as machine learning, functional network modelling, and longitudinal analyses. Responsibilities: The successful candidate will work on the Synapsis-foundation funded research project ?Real-life activity tracking as pre-screening tool for early stages of Alzheimer disease?. The aim of the project is to investigate whether real-life activity measures, derived from wearable technology (e.g. GPS and accelerometer data), are sensitive to identify early stages of Alzheimer?s disease. Further, we aim to provide evidence that these real-life activity measures are associated with current AD biomarkers (i.e. high Amyloid level and brain atrophy). The student will be expected to disseminate study results in peer reviewed journals, and to supervise Master?s students. The candidate will work in the team of Prof. Nicolas Langer, who is also part of the Neuroscience Center Zurich (ZNZ) (https://www.neuroscience.uzh.ch/en.html ), which offers a renowned international PhD programme in Neuroscience. The candidate will work closely with the Institute for Regenerative Medicine (https://www.irem.uzh.ch/en.html ), Geographic Information Systems (https://www.geo.uzh.ch/en/units/gis.html ), University Research Priority Programme from the University of Zurich ?Dynamics of Healthy Aging? (https://www.dynage.uzh.ch/en.html ), and the Department of Computer Science at the ETH Zurich (https://www.systems.ethz.ch/ ). Workload %: 80 - 100% Qualifications: ? MSc degree in a field related to cognitive neuroscience (e.g., cognitive neuroscience, (neuro-)psychology, computer science, biomedical or electrical engineering) ? Deep knowledge in data science (time-series data processing, feature engineering and analysis) ? Proficiency in programming (in Python, Matlab or R) is a must ? Knowledge in mobile and wearable digital technologies is desirable ? Experience with amyloid PET imaging and/or structural MRI analyses are a plus ? Training in machine learning is a plus ? Excellent verbal and written English skills Language requirements: English We offer: ? To work in a team of highly motivated young researchers who are passionate about neuroscience, psychology and computer science ? A very competitive salary (54?000 USD/year) and generous social benefits ? Employment 3 years with the possibility of extension ? Generous support for professional travel and research needs (~3?000 USD/year) ? An inspiring work environment within the Department of Psychology and the University of Zurich and part of the Neuroscience Center Zurich (ZNZ) with many high-caliber collaborations (Department of Computer Science (ETH Zurich; Ce Zhang), Institute for Regenerative Medicine (UZH; Prof. Christoph Hock), Geographic Information Systems (UZH; Prof. Robert Weibel) ? The opportunity to live in Zurich, one of the world?s most attractive cities Please visit?https://www.pa.uzh.ch/en/Willkommen-an-der-UZH.html for further information. This position opens on: 1.6.2022 (starting date) More information: Prof. Nicolas Langer, n.langer at psychologie.uzh.ch Application To be considered please stick to the following application format: ? CV including publication list and contact details of two referees (max. 3 pages) ? Statement describing motivations, personal qualifications and research interests (max. 2 pages) ? Save application in one single pdf file with the file name ?Methlab_[SURNAME]_[name].pdf? ? Send application by email to: n.langer at psychologie.uzh.ch Applications will be considered until the position is filled (ideally submit your application before 31st of March 2022). Description of the Project: Real-life activity tracking as pre-screening tool for early stages of Alzheimer disease Alzheimer's disease (AD) accounts for the majority of all dementia cases and represents a major and rapidly growing burden to the healthcare and economical systems. The current state of research indicates that therapies need to be administered as early as possible. Therefore, there is an urgent need for accelerating biomarker discovery for early stages of AD. Importantly, evidences suggest that neurodegenerative changes precede clinical manifestations of AD by 20-30 years. However, prevailing potential biomarkers for early stages of AD diagnosis, including genetic testing, molecular examination of CSF, structural MRI and PET imaging, are highly limited as they can only be applied to relatively small sample sizes due to their excessive costs and invasive nature. This prevents their usage in large epidemiological studies; yet such study designs are imperative for identifying the intra-individual progression from healthy ageing to AD. Thus, novel non-invasive and inexpensive biomarkers are urgently required to be administered at large scale with the aim to identify individuals with indications for early stages of AD. These identified subjects could then be referred to undergo more cost-intensive examinations with the currently available biomarker techniques to achieve the desired diagnostic accuracy. Mobile and wearable digital technologies have an unprecedented potential and could close this current gap as they permit abundant, continuous longitudinal data acquisition at low costs to investigate intra-individual changes as early markers of AD. In this project, we will investigate whether activity measures, derived from GPS and accelerometer data, are sensitive to identify early stages of AD (i.e. mild neurocognitive disorder due to AD). Further, we aim to provide evidence that real-life activity measures are associated with current AD biomarkers (i.e. high Amyloid level and brain atrophy). Because our long-term vision is to develop a smartphone application that is able to identify intra-individual trajectories in healthy adults to establish the transition point to mild neurocognitive disorder due to AD as early as possible, we will examine the sensitivity of the real-life activity measures for intra-individual changes. Therefore, the present proposal serves as a proof-of-concept to address the prerequisites for real-life activity tracking as a pre-screening assessment tool to identify potential mild neurocognitive disorder patients. Nicolas Langer Professor of Methods of Plasticity Research University of Zurich Andreasstrasse 15 (Office AND.4.56) 8050, Zurich, Switzerland n.langer at psychologie.uzh.ch phone: (+41) 44 635 34 14 http://www.psychology.uzh.ch/en/areas/nec/plafor.html -------------- next part -------------- An HTML attachment was scrubbed... URL: From achler at gmail.com Mon Feb 7 11:38:44 2022 From: achler at gmail.com (Tsvi Achler) Date: Mon, 7 Feb 2022 08:38:44 -0800 Subject: Connectionists: Please stop having conversations on this mailing list In-Reply-To: References: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> Message-ID: Haha it seems that is one thing we agree on, there needs to be more avenues for discussion whether it is arguing, venting or professing. It is more real and reveals our strengths and flaws. I also like the fact that it is not a part of a pay-to-play system (conferences and journals) that we seem to get ourselves into. In that vein I have also taken to making videos of my thoughts and ideas because it is a way to document that is more accessible.. I think those social media channels can actually be work-specific. I believe it is one of the best tools to fight against undesirable dictator-like behavior (which goes against the openness of science) is social media. . Speaking of venting, here is my latest video: https://youtu.be/9BIn_Vmiwz4 -Tsvi On Mon, Feb 7, 2022 at 7:26 AM Jonathan Grainger < i.jonathan.grainger at gmail.com> wrote: > I am a humble reading researcher (psychologist) interested in how much > processing readers can perform in parallel - and I've found it interesting > to see how having this connectionist thread open in a different window from > my main work has helped me develop that kind of parallel processing :-) > it's great stuff - keep going ... > Jonathan Grainger > > On Mon, 7 Feb 2022 at 16:05, Brad Wyble wrote: > >> Agreed, I don't know what is the point of mailing lists except to >> encourage scientific exchange. The present discussions, while not always >> moving in a strictly forward direction, are a welcome break from the list >> of workshops, symposiums and opportunities that are the usual fare for >> connectionists. It's easy to skip over a thread if you don't find it >> useful. >> >> On Mon, Feb 7, 2022 at 9:23 AM Stephen Jos? Hanson < >> jose at rubic.rutgers.edu> wrote: >> >>> Pau, >>> >>> actually this was exactly what Connectionists was created for: civil, >>> thoughtful discussion about Neural Computation. >>> >>> It is very lightly moderated and provides a forum for discussion, >>> promotion and potential collaborations. >>> >>> This discussion has been a bit redundant here and there, but for the >>> most part reconstructing much of the arguments from the 1980s in the >>> background of the DL revolution. >>> >>> Explainable AI, Neuro-Symbolic models and causal structure have been >>> constant concerns of thoughtful folks using, building Neural network >>> systems for 40 years. >>> >>> The threads have gone in predictable ways, but have interesting voices >>> and real issues behind them. Some of this is dialetical (my friend Gary >>> Marcus who is a master at this) and some of it hyperbolic. But all have >>> made excellent and interesting points. >>> >>> Best Regards, >>> >>> Steve >>> On 2/5/22 11:29 AM, pau wrote: >>> >>> Dear connectionists, >>> >>> To the best of my understanding, the aim of this mailing list is to meet >>> the needs of working professionals. I understand by this sharing >>> information about events, publications, developments, etc. that can be >>> useful for someone who works in this field. >>> If someone wants to argue, discuss, praise or assign blame, agree or >>> disagree, please do so in private correspondence or in non-work-specific >>> social media channels. >>> >>> Best, >>> P. >>> >>> -- >>> >> >> >> -- >> Brad Wyble >> Associate Professor >> Psychology Department >> Penn State University >> >> http://wyblelab.com >> > -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From mr.guang.yang at gmail.com Mon Feb 7 11:00:36 2022 From: mr.guang.yang at gmail.com (Guang Yang) Date: Mon, 7 Feb 2022 16:00:36 +0000 Subject: Connectionists: MIUA 2022 Call for Papers Message-ID: You are invited to submit your abstract paper to MIUA 2022 ( https://www.miua2022.com/) which will be held in Cambridge, UK. The abstract paper submission system is now open. https://cmt3.research.microsoft.com/User/Login?ReturnUrl=%2FMIUA2022 The full paper submission deadline will be 23:59, Greenwich Mean Time (GMT), on 1st April 2022. The abstract paper submission deadline will be 23:59, Greenwich Mean Time (GMT), on 31st May 2022. Selected extended papers will be invited for publication in two special issues in: COMPUTERIZED MEDICAL IMAGING AND GRAPHICS (IMPACT FACTOR: 4.790) INTERNATIONAL JOURNAL OF IMAGING SYSTEMS AND TECHNOLOGY (IMPACT FACTOR: 2.000) High-quality papers are requested, containing original contributions to the topics within the scope of MIUA. Paper Submissions: For the 26th MIUA conference, we welcome submissions, as regular conference papers and conference abstracts. Regular papers: Authors are invited to submit full papers of length between 8 and 15 pages (1 column ? the LNCS template) showing original research contributions under the topics of the conference. All submissions will be double-blind peer-reviewed and accepted articles will be published as MIUA Proceedings by the Springer Publishing Group. Conference abstracts: Authors are invited to submit short papers of length up to 3 pages excluding references (1 column ? the LNCS template) showing proof-of-concept research contributions under the topics of the conference. All submissions will be peer-reviewed and accepted articles will be published as MIUA Abstract Proceedings on the MIUA website. Accepted papers will be published in the MIUA proceedings in the Springer LNCS Series. Authors should consult Springer?s authors? guidelines and use their proceedings templates, either for LaTeX or for Word, for the preparation of their papers. Springer encourages authors to include their ORCIDs in their papers. In addition, the corresponding author of each paper, acting on behalf of all of the authors of that paper, must complete and sign a Consent-to-Publish form. The corresponding author signing the copyright form should match the corresponding author marked on the paper. Once the files have been sent to Springer, changes relating to the authorship of the papers cannot be made. MIUA continues to foster fairness, diversity, and inclusion within its community. Submissions from typically underrepresented groups are particularly encouraged. Review Process MIUA seeks your assistance as an expert reviewer for this annual conference. If you would like to review for MIUA, please contact us via: gateway at newton.ac.uk or twitter @miua2022. Please share this CfP with colleagues who might want to contribute to MIUA2022. Dr Guang Yang (Ph.D., M.Sc,, B.Eng., SM.IEEE, M.BMVA) Pronouns: he/him/his UKRI Future Leaders Fellow National Heart & Lung Institute, Imperial College London Group:// www.yanglab.fyi/ Email:// g.yang at imperial.ac.uk -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Mon Feb 7 10:57:34 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Mon, 7 Feb 2022 07:57:34 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear John, I agree with you that cluttered scenes are critical, but Geoff?s GLOM paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf] might actually have some relevance. It may well be that we need to do a better job with parts and whole before we can fully address clutter, and Geoff is certainly taking that question seriously. Geoff?s ?Stable islands of identical vectors? do sound suspiciously like symbols to me (in a good way!), but regardless, they seem to me to be a plausible candidate as a foundation for coping with clutter. And not just cluttered scenes, but also relations between multiple objects in a scene, which is another example of the broader issue you raise, challenging for pure MLPs but critical for deeper AI. Gary > On Feb 7, 2022, at 00:23, Juyang Weng wrote: > > ? > Dear Geoff Hinton, > I respect that you have been working on pattern recognition on isolated characters using neural networks. > > However, I am deeply disappointed that after receiving the Turing Award 2018, you are still falling behind your own award work by talking about "how you > recognize that a handwritten 2 is a 2." You have fallen behind our group's > Creceptron work in 1992, let alone our group's work on 3D-to-2D-to-3D Conscious Learning using DNs. Both deal with cluttered scenes. > > Specifically, you will never be able to get a correct causal explanation by looking at a single hand-written 2. Your problem is too small to explain a brain network. You must look at cluttered sciences, with many objects. > > Yours humbly, > -John > ---- > Message: 7 > Date: Fri, 4 Feb 2022 15:24:02 -0500 > From: Geoffrey Hinton > To: "Dietterich, Thomas" > Cc: AIhub , > "connectionists at mailman.srv.cs.cmu.edu" > > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > Message-ID: > > Content-Type: text/plain; charset="utf-8" > > I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. > > -- > Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From danny.silver at acadiau.ca Mon Feb 7 13:36:54 2022 From: danny.silver at acadiau.ca (Danny Silver) Date: Mon, 7 Feb 2022 18:36:54 +0000 Subject: Connectionists: Please stop having conversations on this mailing list In-Reply-To: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> References: <849cae35-ff2d-4e26-e08f-38f2c00831f0@rubic.rutgers.edu> Message-ID: Dear Stephen, Pau and Others .. The Connectionists list was created for the very exchange of ideas that we now see happening. It is wonderful that working professionals and those in the middle of academic research and applied work are using the list, but the vetting of new scientific ideas and subsequent discourse is what this forum has been about over 30 years. The current debate, and in some cases re-debate, of the symbolist versus connectionist positions harkens back to discussions in the 1980s and 90s. Of course, we now understand a lot more about the biological mechanisms of learning and have new theory concerning computational learning. Many of the emails reflect on this. The exchange is wonderful for the silverbacks to review and most certainly for the new researchers to share, discuss and add their own fresh ideas. This debate has rekindled some of my own thoughts from the past plus a few more recent perspectives. The representation of semantics and their composition into thought can be modeled by: (1) the symbolist perspective: symbolic representation and syntactic compositionality; and (2) the connectionist perspective: vectorial representation and composition through superposition. It can be argued that both can explain the systematicity of human thought and thinking, however ? symbols and syntactic compositionality is much simpler for human communication. And this has likely been to our detriment in terms of understanding the function of our nervous systems. It makes sense to me that early AI started with this choice of representation and compositionality, because it is how we exchange information between each other. But it is unlikely this is how our nervous systems function. Many simple reptiles and mammals can perform complex tasks without having to manipulate symbols. Although more complex to understand, it is likely that vectorial representation and composition is the basis of how our brains work, and that symbolic representation and syntactic compositionality developed as a method of sharing concepts and meaning between each other. The evolutionary step that allowed us to explain to a young person (using symbols) how to identify a dangerous insect or animal (using our senses and neural vectors) provided a huge advantage to humans and so it was inherited and expanded upon over many generations. Eventually, it also provided something else ? the ability to organize and ?think? about the things we sensed in the world, at least at a coarse level, using symbols and syntactic relations. My suspicions are that the duality of learning to sense and respond to the world affectively (using vectors) while also learning to explain those senses and actions (using symbols) creates a multiple task learning problem that takes longer to develop but results in significant benefits - the basis for an agent that can reason and plan over the things it senses and can affect. So currently I am in the camp that believes that our bodies manipulate sense and action vectors and not symbols. However, we have developed the ability to map ?manifolds? of sensory and action vectorial space onto symbols (also represented as vectors within our nervous system) that can be used for communications to others, or conscious reasoning about aspects of the world. Please see the following recent article by Jeff Mitchell and Mirella Lapata that dives much deep into aspects of this at https://onlinelibrary.wiley.com/doi/10.1111/j.1551-6709.2010.01106.x Interestingly, they sight Dedre Gentner?s article of 1989 which originally stimulated my thinking on this in the 1990s Gentner, D. (1989). The mechanisms of analogical learning. In S. Vosniadou & A. Ortony (Eds.), Similarity and analogical reasoning (pp. 199?241). Cambridge, England: Cambridge University Press. ========================== Daniel L. Silver Professor, Jodrey School of Computer Science Director, Acadia Institute for Data Analytics Acadia University, Office 314, Carnegie Hall, Wolfville, Nova Scotia Canada B4P 2R6 t. (902) 585-1413 f. (902) 585-1067 acadiau.ca Facebook Twitter YouTube LinkedIn Flickr [id:image001.png at 01D366AF.7F868A70] From: Connectionists on behalf of Stephen Jos? Hanson Date: Monday, February 7, 2022 at 10:14 AM To: pau , connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Please stop having conversations on this mailing list CAUTION: This email comes from outside Acadia. Verify the sender and use caution with any requests, links or attachments. Pau, actually this was exactly what Connectionists was created for: civil, thoughtful discussion about Neural Computation. It is very lightly moderated and provides a forum for discussion, promotion and potential collaborations. This discussion has been a bit redundant here and there, but for the most part reconstructing much of the arguments from the 1980s in the background of the DL revolution. Explainable AI, Neuro-Symbolic models and causal structure have been constant concerns of thoughtful folks using, building Neural network systems for 40 years. The threads have gone in predictable ways, but have interesting voices and real issues behind them. Some of this is dialetical (my friend Gary Marcus who is a master at this) and some of it hyperbolic. But all have made excellent and interesting points. Best Regards, Steve On 2/5/22 11:29 AM, pau wrote: Dear connectionists, To the best of my understanding, the aim of this mailing list is to meet the needs of working professionals. I understand by this sharing information about events, publications, developments, etc. that can be useful for someone who works in this field. If someone wants to argue, discuss, praise or assign blame, agree or disagree, please do so in private correspondence or in non-work-specific social media channels. Best, P. -- [cid:part1.AC909AA8.0241CDFF at rubic.rutgers.edu] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 20163 bytes Desc: signature.png URL: From juyang.weng at gmail.com Mon Feb 7 17:45:38 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Mon, 7 Feb 2022 17:45:38 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <78E1F023-F059-4762-AAC1-17F2387F3819@asu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <78E1F023-F059-4762-AAC1-17F2387F3819@asu.edu> Message-ID: Dear Asim, You wrote: "If it was skull-closed, Humans would barely learn anything." The skull is closed throughout life, but sensors and effectors are connected to the brain. Humans teach the brain as part of the environment. You wrote: "We teach a system about parts just like you would teach a kid or an adult about parts." When you teach, how does the kid's brain segment your 10 body parts and 1000-10=990 other parts in your classroom? You cannot assume that the kid recognizes and segments you. This is what I said about the chicken and egg problem. By skull-closed, I mean the following examples are invalid: (1) a human teacher tunes parameters in the brain-network like many CNNs and LSTMs have done because the human knows the data or (2) a human teacher assigns a neuron for a particular role (e.g., a symbol) as many symbolic networks have done (e.g. Joshua Tenenbaum's?), or (3) a human teacher assigns a group of neurons for a particular role (e.g., edge detectors), as many so called mental architectures have done. Sorry, Cresceptron and DN-1 did that, but not DN-2. Best regards, -John On Mon, Feb 7, 2022 at 4:39 PM Asim Roy wrote: > Dear John, > > We teach a system about parts just like you would teach a kid or an adult > about parts. There?s nothing ?skull-closed? about it. If it was > ?skull-closed,? Humans would barely learn anything. And dealing with a > thousand parts should not be a problem. Humans remember more than a > thousand parts. And cluttered images are not a problem. We have dealt with > satellite and other cluttered images already. So we are not looking at > Mickey Mouse problems. > > Asim > > Sent from my iPhone > > On Feb 7, 2022, at 8:59 AM, Juyang Weng wrote: > > ? > Dear Asim, > The biggest trap is to avoid brain problems and only work on so called "an > engineering problem" or "a focused problem". > In a cluttered scene, there maybe n>10 objects and m>> 1000 parts. A > brain model must explain how the skull-closed brain segments objects and > their parts. > The traditional and also your "recognition by parts" approach does not > explain > (1) how different objects are segmented and > (2) how different parts are segmented and group into objects. > Note "skull-closed" condition. No supervision into the brain network is > allowed. > My conscious learning model using DN addresses such compounding problems > while skull is closed (using unsupervised Hebbian learning) and without a > given task. > J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th > International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, > USA, Jan.7-9, 2022. > > http://www.cse.msu.edu/~weng/research/ConsciousLearning-ICCE-2022-rvsd-cite.pdf > > YouTube talk: > [image: image.png] > My model must not address the above problems only. > Best regards, > -John > > On Sun, Feb 6, 2022 at 10:43 PM Asim Roy wrote: > >> Dear John, >> >> >> >> We recognize whole objects, but at the same time we verify its parts. >> >> >> >> Best, >> >> Asim >> >> >> >> *From:* Juyang Weng >> *Sent:* Sunday, February 6, 2022 8:38 PM >> *To:* Asim Roy >> *Cc:* Geoffrey Hinton ; Dietterich, Thomas < >> tgd at oregonstate.edu>; AIhub ; >> connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; >> Danko Nikolic ; Stephen Jos? Hanson < >> jose at rubic.rutgers.edu>; Marek Reformat ; MARCO >> GORI ; Alessandro Sperduti < >> alessandro.sperduti at unipd.it>; Xiaodong Li ; >> Hava Siegelmann ; Peter Tino < >> P.Tino at cs.bham.ac.uk>; Bing Xue ; Ali Minai < >> minaiaa at gmail.com>; Claudius Gros ; >> Jean-Philippe Thivierge ; Tsvi >> Achler ; Prof A Hussain < >> hussain.doctor at gmail.com> >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Asim, >> >> The brain does not assume a single object in a cluttered science. Thus, >> a simple explanation like "recognition by parts" (but without object >> abstraction) should be invalid. It is like a chicken and egg problem. >> Both chicken and egg are absent. We must not assume egg is there or >> chicken is there. >> >> Best regards, >> >> -John >> >> >> >> On Sun, Feb 6, 2022 at 2:42 PM Asim Roy wrote: >> >> Dear John, >> >> >> >> You are right and I admit I am not solving all of the problems. It?s just >> in reference to this one problem that Geoffrey Hinton mentions that I think >> can be resolved: >> >> *?I agree that it's nice to have a causal explanations. But I am not >> convinced there will ever be a simple causal explanation for how you >> recognize that a handwritten 2 is a 2. We can introspect on how we do it >> and this may or may not give some insight into how we check our answer, but >> the immediate sense that a handwritten 2 is a 2 is computed by a neural net >> that is not functionally equivalent to any simple and easily explainable >> procedure.?* >> >> Best, >> >> Asim >> >> >> >> *From:* Juyang Weng >> *Sent:* Sunday, February 6, 2022 10:06 AM >> *To:* Asim Roy >> *Cc:* Geoffrey Hinton ; Dietterich, Thomas < >> tgd at oregonstate.edu>; AIhub ; >> connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; >> Danko Nikolic ; Stephen Jos? Hanson < >> jose at rubic.rutgers.edu>; Marek Reformat ; MARCO >> GORI ; Alessandro Sperduti < >> alessandro.sperduti at unipd.it>; Xiaodong Li ; >> Hava Siegelmann ; Peter Tino < >> P.Tino at cs.bham.ac.uk>; Bing Xue ; Ali Minai < >> minaiaa at gmail.com>; Claudius Gros ; >> Jean-Philippe Thivierge ; Tsvi >> Achler ; Prof A Hussain < >> hussain.doctor at gmail.com> >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Asim, >> >> I try to be brief so that I can explain why many of us have missed, and >> will continue to miss, the boat. >> In some of my talks, I have a ppt slide "The brain is like blindmen and >> an elephant". >> >> Unfortunately, your "identify objects based on its parts" is a good >> traditional idea from pattern recognition that is still a blindman. >> >> Your idea does not explain many other problems without which we will >> never understand a biological brain. >> >> For example, your idea does not explain how the brain learns planning and >> discovery in a cluttered world. >> >> We must solve many million-dollar problems holistically. Please watch my >> YouTube video: >> Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar >> Problems Solved >> https://youtu.be/Dgx1dLCdSKY >> >> >> Best regards, >> >> -John >> >> >> >> On Sat, Feb 5, 2022 at 12:01 AM Asim Roy wrote: >> >> I am responding to this part of Geoffrey Hinton?s note: >> >> >> >> *?I agree that it's nice to have a causal explanations. But I am not >> convinced there will ever be a simple causal explanation for how you >> recognize that a handwritten 2 is a 2. We can introspect on how we do it >> and this may or may not give some insight into how we check our answer, but >> the immediate sense that a handwritten 2 is a 2 is computed by a neural net >> that is not functionally equivalent to any simple and easily explainable >> procedure.?* >> >> >> >> The causal explanation is actually done quite simply, and we are doing it >> currently. I can talk about this now because Arizona State University (ASU) >> has filed a provisional patent application on the technology. The basic >> idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable >> Artificial Intelligence (darpa.mil) >> ) >> and illustrated in the figure below. The idea is to identify objects based >> on its parts. So, the figure below says that it?s a cat because it has fur, >> whiskers, and claws plus an unlabeled visual feature. I am not sure if >> DARPA got anything close to this from its funding of various entities. What >> this means is that you need a parts model. And we do that. In the case of >> MNIST handwritten digits that Geoff mentions, we ?teach? this parts model >> what the top part of a digit ?3? looks like, what the bottom part looks >> like and so on. And we also teach connectivity between parts and the >> composition of objects from parts. And we do that for all digits. And we >> get a symbolic model sitting on top of a CNN model that provides the >> explanation that Geoff is referring to as the causal explanation. This >> ?teaching? is similar to the way you would teach a kid to recognize >> different digits. >> >> >> >> An advantage of this parts model, in addition to being in an explainable >> symbolic form, is robustness to adversarial attack. We recently tested on >> the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast >> gradient method to 27%, our XAI model maintained an accuracy of 90%, >> probably higher. In general, it would be hard to make a school bus look >> like an ostrich, with a few pixel changes, if you can identify the parts of >> a school bus and an ostrich. >> >> >> >> A parts model that DARPA wanted provides both a symbolic explanation and >> adversarial protection. The problem that Geoffrey is referring to is solved. >> >> >> >> I am doing a tutorial on this at IEEE World Congress on Computational >> Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, >> Italy 18-23 July >> ). >> I am copying the organizers and want to thank them for accepting the >> tutorial proposal. The only other presentation I have done on this is at a >> Military Operations Research Society (MORS) meeting last December. >> >> >> >> So, back to the future. Hybrid models might indeed save deep learning >> models and let us deploy these models without concern. We might not even >> need adversarial training of any kind. >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> www.teuvonet.com >> >> >> >> >> [image: Timeline Description automatically generated] >> >> >> >> *From:* Connectionists *On >> Behalf Of *Geoffrey Hinton >> *Sent:* Friday, February 4, 2022 1:24 PM >> *To:* Dietterich, Thomas >> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> I agree that it's nice to have a causal explanations. But I am not >> convinced there will ever be a simple causal explanation for how you >> recognize that a handwritten 2 is a 2. We can introspect on how we do it >> and this may or may not give some insight into how we check our answer, but >> the immediate sense that a handwritten 2 is a 2 is computed by a neural net >> that is not functionally equivalent to any simple and easily explainable >> procedure. >> >> >> >> This does not mean that we should give up on trying to make artificial >> neural nets work more like real ones. People can see a tilted square as >> either an upright diamond or a tilted square and, so far as I know, a >> convnet does not exhibit this type of alternate percept. People seem to >> impose hierarchical structural descriptions on images and sound waves and >> they clearly impose intrinsic coordinate frames on wholes and parts. If >> this is what Gary means by symbolic then I don?t disagree that neural nets >> should do symbol processing. However, there is a very different meaning of >> "symbolic". A pure atomic symbol has no internal structure. The form of the >> symbol itself tells you nothing about what it denotes. The only relevant >> properties it has are that it's identical to other instances of the >> same symbol and different from all other symbols. That's totally different >> from a neural net that uses embedding vectors. Embedding vectors have a >> rich internal structure that dictates how they interact with other >> embedding vectors. What I really object to is the following approach: Start >> with pure symbols and rules for how to manipulate structures made out of >> pure symbols. These structures themselves can be denoted by symbols that >> correspond to memory addresses where the bits in the address tell you >> nothing about the content of the structure at that address. Then when the >> rule-based approach doesn't work for dealing with the real world (e.g. >> machine translation) try to use neural nets to convert the real world into >> pure symbols and then carry on with the rule-based approach. That is like >> using an electric motor to inject the gasoline into the same old gasoline >> engine instead of just replacing the gasoline engine with an electric motor. >> >> >> >> >> >> On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas >> wrote: >> >> ?Understanding? is not a Boolean. It is a theorem that no system can >> enumerate all of the consequences of a state of affairs in the world. >> >> >> >> For low-stakes application work, we can be satisfied by a system that >> ?does the right thing?. If the system draws a good picture, that?s >> sufficient. It ?understood? the request. >> >> >> >> But for higher-stakes applications---and for advancing the science---we >> seek a causal account of how the components of a system cause it to do the >> right thing. We are hoping that a small set of mechanisms can produce broad >> coverage of intelligent behavior. This gives us confidence that the system >> will respond correctly outside of the narrow tasks on which we have tested >> it. >> >> >> >> --Tom >> >> >> >> Thomas G. Dietterich, Distinguished Professor Emeritus >> >> School of Electrical Engineering and Computer >> Science >> >> US Mail: 1148 Kelley Engineering Center >> >> >> >> Office: 2067 Kelley Engineering Center >> >> Oregon State Univ., Corvallis, OR 97331-5501 >> >> Voice: 541-737-5559; FAX: 541-737-1300 >> >> URL: http://web.engr.oregonstate.edu/~tgd/ >> >> >> >> >> *From:* Connectionists *On >> Behalf Of *Gary Marcus >> *Sent:* Thursday, February 3, 2022 8:26 AM >> *To:* Danko Nikolic >> *Cc:* connectionists at mailman.srv.cs.cmu.edu; AIhub >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> [This email originated from outside of OSU. Use caution with links and >> attachments.] >> >> Dear Danko, >> >> >> >> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >> talk, in which he said (paraphrasing from memory, because I don?t remember >> the precise words) that the famous 200 Quoc Le unsupervised model [ >> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >> ] >> had learned the concept of a ca. In reality the model had clustered >> together some catlike images based on the image statistics that it had >> extracted, but it was a long way from a full, counterfactual-supporting >> concept of a cat, much as you describe below. >> >> >> >> I fully agree with you that the reason for even having a semantics is as >> you put it, "to 1) learn with a few examples and 2) apply the knowledge to >> a broad set of situations.? GPT-3 sometimes gives the appearance of having >> done so, but it falls apart under close inspection, so the problem remains >> unsolved. >> >> >> >> Gary >> >> >> >> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >> wrote: >> >> >> >> G. Hinton wrote: "I believe that any reasonable person would admit that >> if you ask a neural net to draw a picture of a hamster wearing a red hat >> and it draws such a picture, it understood the request." >> >> >> >> I would like to suggest why drawing a hamster with a red hat does not >> necessarily imply understanding of the statement "hamster wearing a red >> hat". >> >> To understand that "hamster wearing a red hat" would mean inferring, in >> newly emerging situations of this hamster, all the real-life >> implications that the red hat brings to the little animal. >> >> >> >> What would happen to the hat if the hamster rolls on its back? (Would the >> hat fall off?) >> >> What would happen to the red hat when the hamster enters its lair? (Would >> the hat fall off?) >> >> What would happen to that hamster when it goes foraging? (Would the red >> hat have an influence on finding food?) >> >> What would happen in a situation of being chased by a predator? (Would it >> be easier for predators to spot the hamster?) >> >> >> >> ...and so on. >> >> >> >> Countless many questions can be asked. One has understood "hamster >> wearing a red hat" only if one can answer reasonably well many of such >> real-life relevant questions. Similarly, a student has understood materias >> in a class only if they can apply the materials in real-life situations >> (e.g., applying Pythagora's theorem). If a student gives a correct answer >> to a multiple choice question, we don't know whether the student understood >> the material or whether this was just rote learning (often, it is rote >> learning). >> >> >> >> I also suggest that understanding also comes together with effective >> learning: We store new information in such a way that we can recall it >> later and use it effectively i.e., make good inferences in newly emerging >> situations based on this knowledge. >> >> >> >> In short: Understanding makes us humans able to 1) learn with a few >> examples and 2) apply the knowledge to a broad set of situations. >> >> >> >> No neural network today has such capabilities and we don't know how to >> give them such capabilities. Neural networks need large amounts of >> training examples that cover a large variety of situations and then >> the networks can only deal with what the training examples have already >> covered. Neural networks cannot extrapolate in that 'understanding' sense. >> >> >> >> I suggest that understanding truly extrapolates from a piece of >> knowledge. It is not about satisfying a task such as translation between >> languages or drawing hamsters with hats. It is how you got the capability >> to complete the task: Did you only have a few examples that covered >> something different but related and then you extrapolated from that >> knowledge? If yes, this is going in the direction of understanding. Have >> you seen countless examples and then interpolated among them? Then perhaps >> it is not understanding. >> >> >> >> So, for the case of drawing a hamster wearing a red hat, understanding >> perhaps would have taken place if the following happened before that: >> >> >> >> 1) first, the network learned about hamsters (not many examples) >> >> 2) after that the network learned about red hats (outside the context of >> hamsters and without many examples) >> >> 3) finally the network learned about drawing (outside of the context of >> hats and hamsters, not many examples) >> >> >> >> After that, the network is asked to draw a hamster with a red hat. If it >> does it successfully, maybe we have started cracking the problem of >> understanding. >> >> >> >> Note also that this requires the network to learn sequentially without >> exhibiting catastrophic forgetting of the previous knowledge, which is >> possibly also a consequence of human learning by understanding. >> >> >> >> >> >> Danko >> >> >> >> >> >> >> >> >> >> >> >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> >> https://www.linkedin.com/in/danko-nikolic/ >> >> >> --- A progress usually starts with an insight --- >> >> >> >> >> >> Virus-free. www.avast.com >> >> >> >> >> >> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >> >> >> Without getting into the specific dispute between Gary and Geoff, I think >> with approaches similar to GLOM, we are finally headed in the right >> direction. There?s plenty of neurophysiological evidence for single-cell >> abstractions and multisensory neurons in the brain, which one might claim >> correspond to symbols. And I think we can finally reconcile the decades old >> dispute between Symbolic AI and Connectionism. >> >> >> >> >> >> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >> an effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> >> GARY: I have *never* called for dismissal of neural networks, but rather >> for some hybrid between the two (as you yourself contemplated in 1991); the >> point of the 2001 book was to characterize exactly where multilayer >> perceptrons succeeded and broke down, and where symbols could complement >> them. >> >> >> >> >> >> Asim Roy >> >> >> Professor, Information Systems >> >> >> Arizona State University >> >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> >> >> >> >> >> >> >> *From: Connectionists On >> Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: >> Geoffrey Hinton Cc: AIhub ; >> connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen >> Hanson in conversation with Geoff Hinton >> * >> >> >> >> >> Dear Geoff, and interested others, >> >> >> >> >> >> What, for example, would you make of a system that often drew the >> red-hatted hamster you requested, and perhaps a fifth of the time gave you >> utter nonsense? Or say one that you trained to create birds but sometimes >> output stuff like this: >> >> >> >> >> >> >> >> >> >> >> >> One could >> >> >> >> >> >> a. avert one?s eyes and deem the anomalous outputs irrelevant >> >> >> or >> >> >> b. wonder if it might be possible that sometimes the system gets the >> right answer for the wrong reasons (eg partial historical contingency), and >> wonder whether another approach might be indicated. >> >> >> >> >> >> Benchmarks are harder than they look; most of the field has come to >> recognize that. The Turing Test has turned out to be a lousy measure of >> intelligence, easily gamed. It has turned out empirically that the Winograd >> Schema Challenge did not measure common sense as well as Hector might have >> thought. (As it happens, I am a minor coauthor of a very recent review on >> this very topic: https://arxiv.org/abs/2201.02387) But its conquest in >> no way means machines now have common sense; many people from many >> different perspectives recognize that (including, e.g., Yann LeCun, who >> generally tends to be more aligned with you than with me). >> >> >> >> >> >> So: on the goalpost of the Winograd schema, I was wrong, and you can >> quote me; but what you said about me and machine translation remains your >> invention, and it is inexcusable that you simply ignored my 2019 >> clarification. On the essential goal of trying to reach meaning and >> understanding, I remain unmoved; the problem remains unsolved. >> >> >> >> >> >> All of the problems LLMs have with coherence, reliability, truthfulness, >> misinformation, etc stand witness to that fact. (Their persistent inability >> to filter out toxic and insulting remarks stems from the same.) I am hardly >> the only person in the field to see that progress on any given benchmark >> does not inherently mean that the deep underlying problems have solved. >> You, yourself, in fact, have occasionally made that point. >> >> >> >> >> >> With respect to embeddings: Embeddings are very good for natural language >> *processing*; but NLP is not the same as NL*U* ? when it comes to >> *understanding*, their worth is still an open question. Perhaps they >> will turn out to be necessary; they clearly aren?t sufficient. In their >> extreme, they might even collapse into being symbols, in the sense of >> uniquely identifiable encodings, akin to the ASCII code, in which a >> specific set of numbers stands for a specific word or concept. (Wouldn?t >> that be ironic?) >> >> >> >> >> >> (Your GLOM, which as you know I praised publicly, is in many ways an >> effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> >> >> >> >> Notably absent from your email is any kind of apology for misrepresenting >> my position. It?s fine to say that ?many people thirty years ago once >> thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. >> I have consistently felt throughout our interactions that you have mistaken >> me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me >> for having made that error. I am still not he. >> >> >> >> >> >> Which maybe connects to the last point; if you read my work, you would >> see thirty years of arguments *for* neural networks, just not in the way >> that you want them to exist. I have ALWAYS argued that there is a role for >> them; characterizing me as a person ?strongly opposed to neural networks? >> misses the whole point of my 2001 book, which was subtitled ?Integrating >> Connectionism and Cognitive Science.? >> >> >> >> >> >> In the last two decades or so you have insisted (for reasons you have >> never fully clarified, so far as I know) on abandoning symbol-manipulation, >> but the reverse is not the case: I have *never* called for dismissal of >> neural networks, but rather for some hybrid between the two (as you >> yourself contemplated in 1991); the point of the 2001 book was to >> characterize exactly where multilayer perceptrons succeeded and broke down, >> and where symbols could complement them. It?s a rhetorical trick (which is >> what the previous thread was about) to pretend otherwise. >> >> >> >> >> >> Gary >> >> >> >> >> >> >> >> >> On Feb 2, 2022, at 11:22, Geoffrey Hinton >> wrote: >> >> >> ? >> >> >> Embeddings are just vectors of soft feature detectors and they are very >> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >> opposite. >> >> >> >> >> >> A few decades ago, everyone I knew then would have agreed that the >> ability to translate a sentence into many different languages was strong >> evidence that you understood it. >> >> >> >> >> >> But once neural networks could do that, their critics moved the >> goalposts. An exception is Hector Levesque who defined the goalposts more >> sharply by saying that the ability to get pronoun references correct in >> Winograd sentences is a crucial test. Neural nets are improving at that but >> still have some way to go. Will Gary agree that when they can get pronoun >> references correct in Winograd sentences they really do understand? Or does >> he want to reserve the right to weasel out of that too? >> >> >> >> >> >> Some people, like Gary, appear to be strongly opposed to neural networks >> because they do not fit their preconceived notions of how the mind should >> work. >> >> >> I believe that any reasonable person would admit that if you ask a neural >> net to draw a picture of a hamster wearing a red hat and it draws such a >> picture, it understood the request. >> >> >> >> >> >> Geoff >> >> >> >> >> >> >> >> >> >> >> >> >> >> >> >> >> >> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >> >> >> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural >> network community, >> >> >> >> >> >> There has been a lot of recent discussion on this list about framing and >> scientific integrity. Often the first step in restructuring narratives is >> to bully and dehumanize critics. The second is to misrepresent their >> position. People in positions of power are sometimes tempted to do this. >> >> >> >> >> >> The Hinton-Hanson interview that you just published is a real-time >> example of just that. It opens with a needless and largely content-free >> personal attack on a single scholar (me), with the explicit intention of >> discrediting that person. Worse, the only substantive thing it says is >> false. >> >> >> >> >> >> Hinton says ?In 2015 he [Marcus] made a prediction that computers >> wouldn?t be able to do machine translation.? >> >> >> >> >> >> I never said any such thing. >> >> >> >> >> >> What I predicted, rather, was that multilayer perceptrons, as they >> existed then, would not (on their own, absent other mechanisms) >> *understand* language. Seven years later, they still haven?t, except in >> the most superficial way. >> >> >> >> >> >> I made no comment whatsoever about machine translation, which I view as a >> separate problem, solvable to a certain degree by correspondance without >> semantics. >> >> >> >> >> >> I specifically tried to clarify Hinton?s confusion in 2019, but, >> disappointingly, he has continued to purvey misinformation despite that >> clarification. Here is what I wrote privately to him then, which should >> have put the matter to rest: >> >> >> >> >> >> You have taken a single out of context quote [from 2015] and >> misrepresented it. The quote, which you have prominently displayed at the >> bottom on your own web page, says: >> >> >> >> >> >> Hierarchies of features are less suited to challenges such as language, >> inference, and high-level planning. For example, as Noam Chomsky famously >> pointed out, language is filled with sentences you haven't seen >> before. Pure classifier systems don't know what to do with such sentences. >> The talent of feature detectors -- in identifying which member of some >> category something belongs to -- doesn't translate into understanding >> novel sentences, in which each sentence has its own unique meaning. >> >> >> >> >> >> It does *not* say "neural nets would not be able to deal with novel >> sentences"; it says that hierachies of features detectors (on their own, if >> you read the context of the essay) would have trouble *understanding *novel sentences. >> >> >> >> >> >> >> Google Translate does yet not *understand* the content of the sentences >> is translates. It cannot reliably answer questions about who did what to >> whom, or why, it cannot infer the order of the events in paragraphs, it >> can't determine the internal consistency of those events, and so forth. >> >> >> >> >> >> Since then, a number of scholars, such as the the computational linguist >> Emily Bender, have made similar points, and indeed current LLM difficulties >> with misinformation, incoherence and fabrication all follow from these >> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, >> also emphasizing issues of understanding and meaning: >> >> >> >> >> >> *The success of the large neural language models on many NLP tasks is >> exciting. However, we find that these successes sometimes lead to hype in >> which these models are being described as ?understanding? language or >> capturing ?meaning?. In this position paper, we argue that a system trained >> only on form has a priori no way to learn meaning. .. a clear understanding >> of the distinction between form and meaning will help guide the field >> towards better science around natural language understanding. >> * >> >> >> >> >> Her later article with Gebru on language models ?stochastic parrots? is >> in some ways an extension of this point; machine translation requires >> mimicry, true understanding (which is what I was discussing in 2015) >> requires something deeper than that. >> >> >> >> >> >> Hinton?s intellectual error here is in equating machine translation with >> the deeper comprehension that robust natural language understanding will >> require; as Bender and Koller observed, the two appear not to be the same. >> (There is a longer discussion of the relation between language >> understanding and machine translation, and why the latter has turned out to >> be more approachable than the former, in my 2019 book with Ernest Davis). >> >> >> >> >> >> More broadly, Hinton?s ongoing dismissiveness of research from >> perspectives other than his own (e.g. linguistics) have done the field a >> disservice. >> >> >> >> >> >> As Herb Simon once observed, science does not have to be zero-sum. >> >> >> >> >> >> Sincerely, >> >> >> Gary Marcus >> >> >> Professor Emeritus >> >> >> New York University >> >> >> >> >> >> On Feb 2, 2022, at 06:12, AIhub wrote: >> >> >> ? >> >> >> Stephen Hanson in conversation with Geoff Hinton >> >> >> >> >> >> In the latest episode of this video series for AIhub.org, Stephen Hanson >> talks to Geoff Hinton about neural networks, backpropagation, >> overparameterization, digit recognition, voxel cells, syntax and semantics, >> Winograd sentences, and more. >> >> >> >> >> >> You can watch the discussion, and read the transcript, here: >> >> >> >> >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> >> >> >> >> About AIhub: >> >> >> AIhub is a non-profit dedicated to connecting the AI community to the >> public by providing free, high-quality information through AIhub.org ( >> https://aihub.org/). We help researchers publish the latest AI news, >> summaries of their work, opinion pieces, tutorials and more. We are >> supported by many leading scientific organizations in AI, namely AAAI, >> NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. >> >> >> Twitter: @aihuborg >> >> >> >> >> >> Virus-free. www.avast.com >> >> >> >> >> >> >> >> >> -- >> >> Juyang (John) Weng >> >> >> >> >> -- >> >> Juyang (John) Weng >> > > > -- > Juyang (John) Weng > > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: not available URL: From ASIM.ROY at asu.edu Mon Feb 7 17:59:32 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Mon, 7 Feb 2022 22:59:32 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <78E1F023-F059-4762-AAC1-17F2387F3819@asu.edu> Message-ID: <0C8B57DC-3E91-4668-B3AE-268E92746232@asu.edu> Dear John, We are not doing any of that you are thinking. Come to our tutorial at WCCI 2022 in Padua, Italy. I love being in Italy. I have perhaps traveled more in Italy than any other EU country. And I love the people. Wish I knew the language. By the way, join the new Society of Explainable AI. We will do a conference in San Francisco in late July/ early August. All the best, Asim Sent from my iPhone On Feb 7, 2022, at 3:46 PM, Juyang Weng wrote: ? Dear Asim, You wrote: "If it was skull-closed, Humans would barely learn anything." The skull is closed throughout life, but sensors and effectors are connected to the brain. Humans teach the brain as part of the environment. You wrote: "We teach a system about parts just like you would teach a kid or an adult about parts." When you teach, how does the kid's brain segment your 10 body parts and 1000-10=990 other parts in your classroom? You cannot assume that the kid recognizes and segments you. This is what I said about the chicken and egg problem. By skull-closed, I mean the following examples are invalid: (1) a human teacher tunes parameters in the brain-network like many CNNs and LSTMs have done because the human knows the data or (2) a human teacher assigns a neuron for a particular role (e.g., a symbol) as many symbolic networks have done (e.g. Joshua Tenenbaum's?), or (3) a human teacher assigns a group of neurons for a particular role (e.g., edge detectors), as many so called mental architectures have done. Sorry, Cresceptron and DN-1 did that, but not DN-2. Best regards, -John On Mon, Feb 7, 2022 at 4:39 PM Asim Roy > wrote: Dear John, We teach a system about parts just like you would teach a kid or an adult about parts. There?s nothing ?skull-closed? about it. If it was ?skull-closed,? Humans would barely learn anything. And dealing with a thousand parts should not be a problem. Humans remember more than a thousand parts. And cluttered images are not a problem. We have dealt with satellite and other cluttered images already. So we are not looking at Mickey Mouse problems. Asim Sent from my iPhone On Feb 7, 2022, at 8:59 AM, Juyang Weng > wrote: ? Dear Asim, The biggest trap is to avoid brain problems and only work on so called "an engineering problem" or "a focused problem". In a cluttered scene, there maybe n>10 objects and m>> 1000 parts. A brain model must explain how the skull-closed brain segments objects and their parts. The traditional and also your "recognition by parts" approach does not explain (1) how different objects are segmented and (2) how different parts are segmented and group into objects. Note "skull-closed" condition. No supervision into the brain network is allowed. My conscious learning model using DN addresses such compounding problems while skull is closed (using unsupervised Hebbian learning) and without a given task. J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, USA, Jan.7-9, 2022. http://www.cse.msu.edu/~weng/research/ConsciousLearning-ICCE-2022-rvsd-cite.pdf YouTube talk: [image.png] My model must not address the above problems only. Best regards, -John On Sun, Feb 6, 2022 at 10:43 PM Asim Roy > wrote: Dear John, We recognize whole objects, but at the same time we verify its parts. Best, Asim From: Juyang Weng > Sent: Sunday, February 6, 2022 8:38 PM To: Asim Roy > Cc: Geoffrey Hinton >; Dietterich, Thomas >; AIhub >; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus >; Danko Nikolic >; Stephen Jos? Hanson >; Marek Reformat >; MARCO GORI >; Alessandro Sperduti >; Xiaodong Li >; Hava Siegelmann >; Peter Tino >; Bing Xue >; Ali Minai >; Claudius Gros >; Jean-Philippe Thivierge >; Tsvi Achler >; Prof A Hussain > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, The brain does not assume a single object in a cluttered science. Thus, a simple explanation like "recognition by parts" (but without object abstraction) should be invalid. It is like a chicken and egg problem. Both chicken and egg are absent. We must not assume egg is there or chicken is there. Best regards, -John On Sun, Feb 6, 2022 at 2:42 PM Asim Roy > wrote: Dear John, You are right and I admit I am not solving all of the problems. It?s just in reference to this one problem that Geoffrey Hinton mentions that I think can be resolved: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? Best, Asim From: Juyang Weng > Sent: Sunday, February 6, 2022 10:06 AM To: Asim Roy > Cc: Geoffrey Hinton >; Dietterich, Thomas >; AIhub >; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus >; Danko Nikolic >; Stephen Jos? Hanson >; Marek Reformat >; MARCO GORI >; Alessandro Sperduti >; Xiaodong Li >; Hava Siegelmann >; Peter Tino >; Bing Xue >; Ali Minai >; Claudius Gros >; Jean-Philippe Thivierge >; Tsvi Achler >; Prof A Hussain > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I try to be brief so that I can explain why many of us have missed, and will continue to miss, the boat. In some of my talks, I have a ppt slide "The brain is like blindmen and an elephant". Unfortunately, your "identify objects based on its parts" is a good traditional idea from pattern recognition that is still a blindman. Your idea does not explain many other problems without which we will never understand a biological brain. For example, your idea does not explain how the brain learns planning and discovery in a cluttered world. We must solve many million-dollar problems holistically. Please watch my YouTube video: Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar Problems Solved https://youtu.be/Dgx1dLCdSKY Best regards, -John On Sat, Feb 5, 2022 at 12:01 AM Asim Roy > wrote: I am responding to this part of Geoffrey Hinton?s note: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? The causal explanation is actually done quite simply, and we are doing it currently. I can talk about this now because Arizona State University (ASU) has filed a provisional patent application on the technology. The basic idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable Artificial Intelligence (darpa.mil)) and illustrated in the figure below. The idea is to identify objects based on its parts. So, the figure below says that it?s a cat because it has fur, whiskers, and claws plus an unlabeled visual feature. I am not sure if DARPA got anything close to this from its funding of various entities. What this means is that you need a parts model. And we do that. In the case of MNIST handwritten digits that Geoff mentions, we ?teach? this parts model what the top part of a digit ?3? looks like, what the bottom part looks like and so on. And we also teach connectivity between parts and the composition of objects from parts. And we do that for all digits. And we get a symbolic model sitting on top of a CNN model that provides the explanation that Geoff is referring to as the causal explanation. This ?teaching? is similar to the way you would teach a kid to recognize different digits. An advantage of this parts model, in addition to being in an explainable symbolic form, is robustness to adversarial attack. We recently tested on the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast gradient method to 27%, our XAI model maintained an accuracy of 90%, probably higher. In general, it would be hard to make a school bus look like an ostrich, with a few pixel changes, if you can identify the parts of a school bus and an ostrich. A parts model that DARPA wanted provides both a symbolic explanation and adversarial protection. The problem that Geoffrey is referring to is solved. I am doing a tutorial on this at IEEE World Congress on Computational Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy 18-23 July). I am copying the organizers and want to thank them for accepting the tutorial proposal. The only other presentation I have done on this is at a Military Operations Research Society (MORS) meeting last December. So, back to the future. Hybrid models might indeed save deep learning models and let us deploy these models without concern. We might not even need adversarial training of any kind. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) www.teuvonet.com [Timeline Description automatically generated] From: Connectionists > On Behalf Of Geoffrey Hinton Sent: Friday, February 4, 2022 1:24 PM To: Dietterich, Thomas > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists > On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic > Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg Virus-free. www.avast.com -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: image.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: image.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: From ASIM.ROY at asu.edu Mon Feb 7 18:42:56 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Mon, 7 Feb 2022 23:42:56 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <78E1F023-F059-4762-AAC1-17F2387F3819@asu.edu> <0C8B57DC-3E91-4668-B3AE-268E92746232@asu.edu> Message-ID: <11D154CE-8AED-4E5E-9C63-61CBD3CDCC52@asu.edu> Dear John, Our method incorporates many aspects of human learning. Come to WCCI 2022 and we can talk at length. Asim Sent from my iPhone On Feb 7, 2022, at 4:34 PM, Juyang Weng wrote: ? Dear Asim, I repeat my previous statement: The biggest trap is to avoid brain problems and only work on so-called "an engineering problem" or "a focused problem". AI is in a crisis caused by Deep Learning using CNN etc., because even people like you who are interested in brain problems are still saying "We are not doing any of that you are thinking". Raising the problems that the brain is facing is the first step toward a solution to brain problems. Best regard, -John On Mon, Feb 7, 2022 at 6:00 PM Asim Roy > wrote: Dear John, We are not doing any of that you are thinking. Come to our tutorial at WCCI 2022 in Padua, Italy. I love being in Italy. I have perhaps traveled more in Italy than any other EU country. And I love the people. Wish I knew the language. By the way, join the new Society of Explainable AI. We will do a conference in San Francisco in late July/ early August. All the best, Asim Sent from my iPhone On Feb 7, 2022, at 3:46 PM, Juyang Weng > wrote: ? Dear Asim, You wrote: "If it was skull-closed, Humans would barely learn anything." The skull is closed throughout life, but sensors and effectors are connected to the brain. Humans teach the brain as part of the environment. You wrote: "We teach a system about parts just like you would teach a kid or an adult about parts." When you teach, how does the kid's brain segment your 10 body parts and 1000-10=990 other parts in your classroom? You cannot assume that the kid recognizes and segments you. This is what I said about the chicken and egg problem. By skull-closed, I mean the following examples are invalid: (1) a human teacher tunes parameters in the brain-network like many CNNs and LSTMs have done because the human knows the data or (2) a human teacher assigns a neuron for a particular role (e.g., a symbol) as many symbolic networks have done (e.g. Joshua Tenenbaum's?), or (3) a human teacher assigns a group of neurons for a particular role (e.g., edge detectors), as many so called mental architectures have done. Sorry, Cresceptron and DN-1 did that, but not DN-2. Best regards, -John On Mon, Feb 7, 2022 at 4:39 PM Asim Roy > wrote: Dear John, We teach a system about parts just like you would teach a kid or an adult about parts. There?s nothing ?skull-closed? about it. If it was ?skull-closed,? Humans would barely learn anything. And dealing with a thousand parts should not be a problem. Humans remember more than a thousand parts. And cluttered images are not a problem. We have dealt with satellite and other cluttered images already. So we are not looking at Mickey Mouse problems. Asim Sent from my iPhone On Feb 7, 2022, at 8:59 AM, Juyang Weng > wrote: ? Dear Asim, The biggest trap is to avoid brain problems and only work on so called "an engineering problem" or "a focused problem". In a cluttered scene, there maybe n>10 objects and m>> 1000 parts. A brain model must explain how the skull-closed brain segments objects and their parts. The traditional and also your "recognition by parts" approach does not explain (1) how different objects are segmented and (2) how different parts are segmented and group into objects. Note "skull-closed" condition. No supervision into the brain network is allowed. My conscious learning model using DN addresses such compounding problems while skull is closed (using unsupervised Hebbian learning) and without a given task. J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, USA, Jan.7-9, 2022. http://www.cse.msu.edu/~weng/research/ConsciousLearning-ICCE-2022-rvsd-cite.pdf YouTube talk: [image.png] My model must not address the above problems only. Best regards, -John On Sun, Feb 6, 2022 at 10:43 PM Asim Roy > wrote: Dear John, We recognize whole objects, but at the same time we verify its parts. Best, Asim From: Juyang Weng > Sent: Sunday, February 6, 2022 8:38 PM To: Asim Roy > Cc: Geoffrey Hinton >; Dietterich, Thomas >; AIhub >; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus >; Danko Nikolic >; Stephen Jos? Hanson >; Marek Reformat >; MARCO GORI >; Alessandro Sperduti >; Xiaodong Li >; Hava Siegelmann >; Peter Tino >; Bing Xue >; Ali Minai >; Claudius Gros >; Jean-Philippe Thivierge >; Tsvi Achler >; Prof A Hussain > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, The brain does not assume a single object in a cluttered science. Thus, a simple explanation like "recognition by parts" (but without object abstraction) should be invalid. It is like a chicken and egg problem. Both chicken and egg are absent. We must not assume egg is there or chicken is there. Best regards, -John On Sun, Feb 6, 2022 at 2:42 PM Asim Roy > wrote: Dear John, You are right and I admit I am not solving all of the problems. It?s just in reference to this one problem that Geoffrey Hinton mentions that I think can be resolved: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? Best, Asim From: Juyang Weng > Sent: Sunday, February 6, 2022 10:06 AM To: Asim Roy > Cc: Geoffrey Hinton >; Dietterich, Thomas >; AIhub >; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus >; Danko Nikolic >; Stephen Jos? Hanson >; Marek Reformat >; MARCO GORI >; Alessandro Sperduti >; Xiaodong Li >; Hava Siegelmann >; Peter Tino >; Bing Xue >; Ali Minai >; Claudius Gros >; Jean-Philippe Thivierge >; Tsvi Achler >; Prof A Hussain > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I try to be brief so that I can explain why many of us have missed, and will continue to miss, the boat. In some of my talks, I have a ppt slide "The brain is like blindmen and an elephant". Unfortunately, your "identify objects based on its parts" is a good traditional idea from pattern recognition that is still a blindman. Your idea does not explain many other problems without which we will never understand a biological brain. For example, your idea does not explain how the brain learns planning and discovery in a cluttered world. We must solve many million-dollar problems holistically. Please watch my YouTube video: Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar Problems Solved https://youtu.be/Dgx1dLCdSKY Best regards, -John On Sat, Feb 5, 2022 at 12:01 AM Asim Roy > wrote: I am responding to this part of Geoffrey Hinton?s note: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? The causal explanation is actually done quite simply, and we are doing it currently. I can talk about this now because Arizona State University (ASU) has filed a provisional patent application on the technology. The basic idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable Artificial Intelligence (darpa.mil)) and illustrated in the figure below. The idea is to identify objects based on its parts. So, the figure below says that it?s a cat because it has fur, whiskers, and claws plus an unlabeled visual feature. I am not sure if DARPA got anything close to this from its funding of various entities. What this means is that you need a parts model. And we do that. In the case of MNIST handwritten digits that Geoff mentions, we ?teach? this parts model what the top part of a digit ?3? looks like, what the bottom part looks like and so on. And we also teach connectivity between parts and the composition of objects from parts. And we do that for all digits. And we get a symbolic model sitting on top of a CNN model that provides the explanation that Geoff is referring to as the causal explanation. This ?teaching? is similar to the way you would teach a kid to recognize different digits. An advantage of this parts model, in addition to being in an explainable symbolic form, is robustness to adversarial attack. We recently tested on the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast gradient method to 27%, our XAI model maintained an accuracy of 90%, probably higher. In general, it would be hard to make a school bus look like an ostrich, with a few pixel changes, if you can identify the parts of a school bus and an ostrich. A parts model that DARPA wanted provides both a symbolic explanation and adversarial protection. The problem that Geoffrey is referring to is solved. I am doing a tutorial on this at IEEE World Congress on Computational Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy 18-23 July). I am copying the organizers and want to thank them for accepting the tutorial proposal. The only other presentation I have done on this is at a Military Operations Research Society (MORS) meeting last December. So, back to the future. Hybrid models might indeed save deep learning models and let us deploy these models without concern. We might not even need adversarial training of any kind. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) www.teuvonet.com [Timeline Description automatically generated] From: Connectionists > On Behalf Of Geoffrey Hinton Sent: Friday, February 4, 2022 1:24 PM To: Dietterich, Thomas > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists > On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic > Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg Virus-free. www.avast.com -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: image.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: image.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: From bhammer at techfak.uni-bielefeld.de Mon Feb 7 12:44:07 2022 From: bhammer at techfak.uni-bielefeld.de (Barbara Hammer) Date: Mon, 7 Feb 2022 18:44:07 +0100 Subject: Connectionists: Talk by Max Welling in JAII Lecture Series Message-ID: <24934486-d301-d93d-2b5e-ee20cbbe3002@techfak.uni-bielefeld.de> Dear collegaues, We would like to announce a talk by Prof. Max Welling (Amsterdam University)[1] within the JAII Lecture Series on the 10th of February, 16:00-17:30 CET. The Joint Artificial Intelligence Institute (JAII) was set up by the universities of Bielefeld and Paderborn to promote joint research activities in the field of AI. The JAII has started a lecture series to invite interesting guests from this research field virtually and/or in person to Paderborn or Bielefeld. Max Welling will talk about "Solving PDEs with GNNs, and how to use symmetries". To register for the virtual lecture, click here: https://uni-bielefeld.zoom.us/meeting/register/tJEpdeiqqTIsHdwgc0lFoscot-NAamrEnmkv or visit https://jaii. Best regards, Barbara Hammer on behalf of the JAII board [1] https://staff.fnwi.uva.nl/m.welling/ -- Prof. Dr. Barbara Hammer Machine Learning Group, CITEC Bielefeld University D-33594 Bielefeld Phone: +49 521 / 106 12115 From minaiaa at gmail.com Tue Feb 8 00:43:22 2022 From: minaiaa at gmail.com (Ali Minai) Date: Tue, 8 Feb 2022 00:43:22 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <148D10E4-0C1F-4FA4-B8AA-C80E1472A63A@nyu.edu> References: <148D10E4-0C1F-4FA4-B8AA-C80E1472A63A@nyu.edu> Message-ID: Hi Gary Thanks for your reply. I'll think more about your points. I do think that, to understand the human mind, we should start with vertebrates, which is why I suggested fish. At least for the motor system - which is part of the mind - we have learned a lot from lampreys (e.g. Sten Grillner's work and that beautiful lamprey-salamander model by Ijspeert et al.), and it has taught us a lot about locomotion in other animals, including mammals. The principles clearly generalize, though the complexity increases a lot. Insects too are very interesting. After all, they are our ancestors too. I don't agree that we can posit a clear transition from deep cognitive models in humans to none below that in the phylogenetic tree. Chimpanzees and macaques certainly show some evidence, and there's no reason to think that it's a step change rather than a highly nonlinear continuum. And even though what we might (simplistically) call System 2 aspects of cognition are minimally present in other mammals, their precursors must be. My point about cats and symbols was not regarding whether cats are aware of symbols, but that symbols emerge naturally from the physics of their brains. Behaviors that require some small degree of symbolic processing exist in mammals other than humans (e.g., transitive inference and landmark-based navigation in rats), and it is seen better as an emergent property of brains than an attribute to be explicitly built-into neural models by us. Once we have a sufficiently brain-like neural model, symbolic processing will already be there. I agree with you completely that we are far from understanding some of the most fundamental principles of the brain, but even more importantly, we are not even looking in the right direction. I'm hoping to lay out my arguments about all this in more detail in some other form. Best Ali PS: I had inadvertently posted my reply of Gary's message only to him. Should have posted to everyone, so here it is. *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Mon, Feb 7, 2022 at 12:28 AM Gary Marcus wrote: > Ali, > > > It?s useful to think about animals, but I really wouldn?t start with fish; > it?s not clear that their ecological niche demands anything significant in > the way of extrapolation, causal reasoning, or compositionality. There is > good evidence elsewhere in the animal world for extrapolation of functions > that may be innate (eg solar azimuth in bees), and causal reasoning (eg > tool use in ravens, various primates, and octopi). It?s still not clear to > me how much hierarchical representation (critical to AGI) exists outside of > humans, though; the ability to construct rich new cognitive models may also > be unique to us. > > > In any case it matters not in the least whether the average cat or human > *cares* about symbols, anymore that it matters whether the average animal > understands digestion; only a tiny fraction of the creatures on this planet > have any real understanding of their internal workings. > > > My overall feeling is that we are a really, really long way from > understanding the neural basis of higher-level cognition, and that AI is > going to need muddle through on its own, for another decade or two, > > > I do fully agree with your conclusion, though, that "AI today is driven > more by habit and the incentives of the academic and corporate marketplaces > than by a deep, long-term view of AI as a great exploratory project in > fundamental science." Let's hope that changes. > > > Gary > > On Feb 6, 2022, at 13:19, Ali Minai wrote: > > ? > > Gary, > > That?s a very interesting and accurate list of capabilities that a general > intelligent system must have and that our AI does not. Of course, the list > is familiar to me from having read your book. However, I have a somewhat > different take on this whole thing. > > > > All the things we discuss here ? symbols/no symbols, parts/wholes, > supervised/unsupervised, token/type, etc., are useful categories and > distinctions for our analysis of the problem, and are partly a result of > the historical evolution of the field of AI in particular and of philosophy > in general. The categories are not wrong in any way, of course, but they > are posterior to the actual system ? good for describing and analyzing it, > and for validating our versions of it (which is how you use them). I think > they are less useful as prescriptions for how to build our AI systems. If > intelligent systems did not already exist and we were building them from > scratch (please ignore the impossibility of that), having a list of ?must > haves? would be great. But intelligent systems already exist ? from humans > to fish ? and they already have these capacities to a greater or lesser > degree because of the physics of their biology. A cat?s intelligence does > not care whether it has symbols or not, and nor does mine or yours. > Whatever we describe as symbolic processing post-facto has already been > done by brains for at least tens of millions of years. Instead of getting > caught up in ?how to add symbols into our neural models?, we should be > investigating how what we see as symbolic processing emerges from animal > brains, and then replicate those brains to the degree necessary. If we can > do that, symbolic processing will already be present. But it cannot be done > piece by piece. It must take the integrity of the whole brain and the body > it is part of, and its environment, into account. That?s why I think that a > much better ? though a very long ? route to AI is to start by understanding > how a fish brain makes the intelligence of a fish possible, and then boot > up our knowledge across phylogenetic stages: Bottom up reverse engineering > rather than top-down engineering. That?s the way Nature built up to human > intelligence, and we will succeed only by reverse engineering it. Of > course, we can do it much faster and with shortcuts because we are > intelligent, purposive agents, but working top-down by building piecewise > systems that satisfy a list of attributes will not get us there. Among > other things, those pieces will be impossible to integrate into the kind of > intelligence that can have those general models of the world that you > rightly point to as being necessary. > > > > I think that one thing that has been a great boon to the AI enterprise has > also been one of the greatest impediments to its complete success, and that > is the ?computationalization? of intelligence. On the one hand, thinking of > intelligence computationally allows us to describe it abstractly and in a > principled, formal way. It also resonates with the fact that we are trying > to implement intelligence through computational machines. But, on the flip > side, this view of intelligence divorces it from its physics ? from the > fact that real intelligence in animals emerges from the physics of the > physical system. That system is not a collection of its capabilities; > rather, those capabilities are immanent in it by virtue of its physics. > When we try to build those capabilities computationally, i.e., through > code, we are making the same error that the practitioners of old-style > ?symbolic AI? made ? what I call the ?professors are smarter than Nature? > error, i.e., the idea that we are going to enumerate (or describe) all the > things that underlie intelligence and implement them one by one until we > get complete intelligence. We will never be able to enumerate all those > capabilities, and will never be able to get to that complete intelligence. > The only difference between us and the ?symbolists? of yore is that we are > replacing giant LISP and Prolog programs with giant neural networks. > Otherwise, we are using our models exactly as they were trying to use their > models, and we will fail just as they did unless we get back to biology and > the real thing. > > > > I will say again that the way we do AI today is driven more by habit and > the incentives of the academic and corporate marketplaces than by a deep, > long-term view of AI as a great exploratory project in fundamental science. > We are just building AI to drive our cars, translate our documents, write > our reports, and do our shopping. What that will teach us about actual > intelligence is just incidental. > > > > My apologies too for a long response. > > Ali > > > *Ali A. Minai, Ph.D.* > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > > > On Sun, Feb 6, 2022 at 9:42 AM Gary Marcus wrote: > >> Dear Asim, >> >> >> Sorry for a long answer to your short but rich questions. >> >> - Yes, memory in my view has to be part of the answer to the >> type-token problem. Symbol systems encoded in memory allow a natural way to >> set up records, and something akin to that seems necessary. Pure multilayer >> perceptrons struggle with type-token distinctions precisely because they >> lack such records. On the positive side, I see more and more movement >> towards recordlike stores (eg w key-value stores in memory networks), and I >> think that is an important and necessary step, very familiar from the >> symbol-manipulating playbook, sometimes implemented in new ways. >> - But ultimately, handling the type-token distinction requires >> considerable inferential overhead beyond the memory representation of a >> record per se. How do you determine when to denote something (e.g. >> Felix) as an instance, and of which kinds (cat, animal etc), and how do you >> leverage that knowledge once you determine it? >> - In the limit we reason about types vs tokens in fairly subtle ways, >> eg in guessing whether a glass that we put down at party is likely to be >> ours. The reverse is also important: we need to be learn particular >> traits for individuals and not erroneously generalize them to the class; if >> my aunt Esther wins the lottery, one shouldn?t infer that all of my >> aunts or all of my relatives or adult females have won the lottery. so you >> need both representational machinery that can distinguish eg my cat from >> cats in general and reasoning machinery to decide at what level certain >> learned knowledge should inhere. (I had a whole chapter about this sort of >> thing in The Algebraic Mind if you are interested, and Mike Mozer had a >> book about types and tokens in neural networks in the mid 1990s). >> - Yes, part (though not all!) of what we do when we set up cognitive >> models in our heads is to track particular individuals and their >> properties. If you only had to correlate kinds (cats) and their properties >> (have fur) you could maybe get away with a multilayer perceptron, but once >> you need to track individuals, yes, you really need some kind of >> memory-based records. >> - As far as I can tell, Transformers can sometimes approximate some >> of this for a few sentences, but not over long stretches. >> >> >> As a small terminological aside; for me cognitive models ? cognitive >> modeling. Cognitive modeling is about building psychological or >> computational models of how people think, whereas what I mean by a cognitive >> model is a representation of eg the entities in some situation and the >> relations between those entities. >> >> >> To your closing question, none of us yet really knows how to build >> understanding into machines. A solid type-token distinction, both in >> terms of representation and reasoning, is critical for general >> intelligence, but hardly sufficient. Personally, I think some minimal >> prerequisites would be: >> >> - representations of space, time, causality, individuals, kinds, >> persons, places, objects, etc. >> - representations of abstractions that can hold over all entities in >> a class >> - compositionality (if we are talking about human-like understanding) >> - capacity to construct and update cognitive models on the fly >> - capacity to reason over entities in those models >> - ability to learn about new entities and their properties >> >> Much of my last book (*Rebooting AI*, w Ernie Davis) is about the above >> list. The section in the language chapter on a children?s story in which >> man has lost is wallet is an especially vivid worked example. Later >> chapters elaborate some of the challenges in representing space, time, and >> causality. >> >> >> Gary >> >> >> On Feb 5, 2022, at 18:58, Asim Roy wrote: >> >> ? >> >> Gary, >> >> >> >> I don?t get much into the type of cognitive modeling you are talking >> about, but I would guess that the type problem can generally be handled by >> neural network models and tokens can be resolved with some memory-based >> system. But to the heart of the question, this is what so-called >> ?understanding? reduces to computation wise? >> >> >> >> Asim >> >> >> >> *From:* Gary Marcus >> *Sent:* Saturday, February 5, 2022 8:39 AM >> *To:* Asim Roy >> *Cc:* Ali Minai ; Danko Nikolic < >> danko.nikolic at gmail.com>; Brad Wyble ; >> connectionists at mailman.srv.cs.cmu.edu; AIhub >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> There is no magic in understanding, just computation that has been >> realized in the wetware of humans and that eventually can be realized in >> machines. But understanding is not (just) learning. >> >> >> >> Understanding incorporates (or works in tandem with) learning - but >> also, critically, in tandem with inference, *and the development and >> maintenance of cognitive models*. Part of developing an understanding >> of cats in general is to learn long term-knowledge about their properties, >> both directly (e.g., through observation) and indirectly (eg through >> learning facts about animals in general that can be extended to cats), >> often through inference (if all animals have DNA, and a cat is an animal, >> it must also have DNA). The understanding of a particular cat also >> involves direct observation, but also inference (eg one might surmise >> that the reason that Fluffy is running about the room is that Fluffy >> suspects there is a mouse stirring somewhere nearby). *But all of that, >> I would say, is subservient to the construction of cognitive models that >> can be routinely updated *(e.g., Fluffy is currently in the living room, >> skittering about, perhaps looking for a mouse). >> >> >> >> In humans, those dynamic, relational models, which form part of an >> understanding, can support inference (if Fluffy is in the living room, we >> can infer that Fluffy is not outside, not lost, etc). Without such models - >> which I think represent a core part of understanding - AGI is an unlikely >> prospect. >> >> >> >> Current neural networks, as it happens, are better at acquiring long-term >> knowledge (cats have whiskers) than they are at dynamically updating >> cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic >> model that I am describing. To a modest degree they can approximate it on >> the basis of large samples of texts, but their ultimate incoherence stems >> from the fact that they do not have robust internal cognitive models that >> they can update on the fly. >> >> >> >> Without such cognitive models you can still capture some aspects of >> understanding (eg predicting that cats are likely to be furry), but things >> fall apart quickly; inference is never reliable, and coherence is fleeting. >> >> >> >> As a final note, one of the most foundational challenges in constructing >> adequate cognitive models of the world is to have a clear distinction >> between individuals and kinds; as I emphasized 20 years ago (in The >> Algebraic Mind), this has always been a weakness in neural networks, and I >> don?t think that the type-token problem has yet been solved. >> >> >> >> Gary >> >> >> >> >> >> On Feb 5, 2022, at 01:31, Asim Roy wrote: >> >> ? >> >> All, >> >> >> >> I think the broader question was ?understanding.? Here are two Youtube >> videos showing simple robots ?learning? to walk. They are purely physical >> systems. Do they ?understand? anything ? such as the need to go around an >> obstacle, jumping over an obstacle, walking up and down stairs and so on? >> By the way, they ?learn? to do these things on their own, literally >> unsupervised, very much like babies. The basic question is: what is >> ?understanding? if not ?learning?? Is there some other mechanism (magic) at >> play in our brain that helps us ?understand?? >> >> >> >> https://www.youtube.com/watch?v=gn4nRCC9TwQ >> >> >> https://www.youtube.com/watch?v=8sO7VS3q8d0 >> >> >> >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> >> >> >> >> >> *From:* Ali Minai >> *Sent:* Friday, February 4, 2022 11:38 PM >> *To:* Asim Roy >> *Cc:* Gary Marcus ; Danko Nikolic < >> danko.nikolic at gmail.com>; Brad Wyble ; >> connectionists at mailman.srv.cs.cmu.edu; AIhub >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Asim >> >> >> >> Of course there's nothing magical about understanding, and the mind has >> to emerge from the physical system, but our AI models at this point are not >> even close to realizing how that happens. We are, at best, simulating a >> superficial approximation of a few parts of the real thing. A single, >> integrated system where all the aspects of intelligence emerge from the >> same deep, well-differentiated physical substrate is far beyond our >> capacity. Paying more attention to neurobiology will be essential to get >> there, but so will paying attention to development - both physical and >> cognitive - and evolution. The configuration of priors by evolution is key >> to understanding how real intelligence learns so quickly and from so >> little. This is not an argument for using genetic algorithms to design our >> systems, just for understanding the tricks evolution has used and >> replicating them by design. Development is more feasible to do >> computationally, but hardly any models have looked at it except in a >> superficial sense. Nature creates basic intelligence not so much by >> configuring functions by explicit training as by tweaking, modulating, >> ramifying, and combining existing ones in a multi-scale self-organization >> process. We then learn much more complicated things (like playing chess) by >> exploiting that substrate, and using explicit instruction or learning by >> practice. The fundamental lesson of complex systems is that complexity is >> built in stages - each level exploiting the organization of the level below >> it. We see it in evolution, development, societal evolution, the evolution >> of technology, etc. Our approach in AI, in contrast, is to initialize a >> giant, naive system and train it to do something really complicated - but >> really specific - by training the hell out of it. Sure, now we do build >> many systems on top of pre-trained models like GPT-3 and BERT, which is >> better, but those models were again trained by the same none-to-all process >> I decried above. Contrast that with how humans acquire language, and how >> they integrate it into their *entire* perceptual, cognitive, and behavioral >> repertoire, not focusing just on this or that task. The age of symbolic AI >> may have passed, but the reductionistic mindset has not. We cannot build >> minds by chopping it into separate verticals. >> >> >> >> FTR, I'd say that the emergence of models such as GLOM and Hawkins and >> Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but >> they are, I think, looking in the right direction. With a million miles to >> go! >> >> >> >> Ali >> >> >> >> *Ali A. Minai, Ph.D.* >> Professor and Graduate Program Director >> Complex Adaptive Systems Lab >> Department of Electrical Engineering & Computer Science >> >> 828 Rhodes Hall >> >> University of Cincinnati >> Cincinnati, OH 45221-0030 >> >> >> Phone: (513) 556-4783 >> Fax: (513) 556-7326 >> Email: Ali.Minai at uc.edu >> minaiaa at gmail.com >> >> WWW: https://eecs.ceas.uc.edu/~aminai/ >> >> >> >> >> >> >> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: >> >> First of all, the brain is a physical system. There is no ?magic? inside >> the brain that does the ?understanding? part. Take for example learning to >> play tennis. You hit a few balls - some the right way and some wrong ? but >> you fairly quickly learn to hit them right most of the time. So there is >> obviously some simulation going on in the brain about hitting the ball in >> different ways and ?learning? its consequences. What you are calling >> ?understanding? is really these simulations about different scenarios. It?s >> also very similar to augmentation used to train image recognition systems >> where you rotate images, obscure parts and so on, so that you still can say >> it?s a cat even though you see only the cat?s face or whiskers or a cat >> flipped on its back. So, if the following questions relate to >> ?understanding,? you can easily resolve this by simulating such scenarios >> when ?teaching? the system. There?s nothing ?magical? about >> ?understanding.? As I said, bear in mind that the brain, after all, is a >> physical system and ?teaching? and ?understanding? is embodied in that >> physical system, not outside it. So ?understanding? is just part of >> ?learning,? nothing more. >> >> >> >> DANKO: >> >> What would happen to the hat if the hamster rolls on its back? (Would the >> hat fall off?) >> >> What would happen to the red hat when the hamster enters its lair? (Would >> the hat fall off?) >> >> What would happen to that hamster when it goes foraging? (Would the red >> hat have an influence on finding food?) >> >> What would happen in a situation of being chased by a predator? (Would it >> be easier for predators to spot the hamster?) >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> >> *From:* Gary Marcus >> *Sent:* Thursday, February 3, 2022 9:26 AM >> *To:* Danko Nikolic >> *Cc:* Asim Roy ; Geoffrey Hinton < >> geoffrey.hinton at gmail.com>; AIhub ; >> connectionists at mailman.srv.cs.cmu.edu >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Danko, >> >> >> >> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >> talk, in which he said (paraphrasing from memory, because I don?t remember >> the precise words) that the famous 200 Quoc Le unsupervised model [ >> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >> ] >> had learned the concept of a ca. In reality the model had clustered >> together some catlike images based on the image statistics that it had >> extracted, but it was a long way from a full, counterfactual-supporting >> concept of a cat, much as you describe below. >> >> >> >> I fully agree with you that the reason for even having a semantics is as >> you put it, "to 1) learn with a few examples and 2) apply the knowledge to >> a broad set of situations.? GPT-3 sometimes gives the appearance of having >> done so, but it falls apart under close inspection, so the problem remains >> unsolved. >> >> >> >> Gary >> >> >> >> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >> wrote: >> >> >> >> G. Hinton wrote: "I believe that any reasonable person would admit that >> if you ask a neural net to draw a picture of a hamster wearing a red hat >> and it draws such a picture, it understood the request." >> >> >> >> I would like to suggest why drawing a hamster with a red hat does not >> necessarily imply understanding of the statement "hamster wearing a red >> hat". >> >> To understand that "hamster wearing a red hat" would mean inferring, in >> newly emerging situations of this hamster, all the real-life >> implications that the red hat brings to the little animal. >> >> >> >> What would happen to the hat if the hamster rolls on its back? (Would the >> hat fall off?) >> >> What would happen to the red hat when the hamster enters its lair? (Would >> the hat fall off?) >> >> What would happen to that hamster when it goes foraging? (Would the red >> hat have an influence on finding food?) >> >> What would happen in a situation of being chased by a predator? (Would it >> be easier for predators to spot the hamster?) >> >> >> >> ...and so on. >> >> >> >> Countless many questions can be asked. One has understood "hamster >> wearing a red hat" only if one can answer reasonably well many of such >> real-life relevant questions. Similarly, a student has understood materias >> in a class only if they can apply the materials in real-life situations >> (e.g., applying Pythagora's theorem). If a student gives a correct answer >> to a multiple choice question, we don't know whether the student understood >> the material or whether this was just rote learning (often, it is rote >> learning). >> >> >> >> I also suggest that understanding also comes together with effective >> learning: We store new information in such a way that we can recall it >> later and use it effectively i.e., make good inferences in newly emerging >> situations based on this knowledge. >> >> >> >> In short: Understanding makes us humans able to 1) learn with a few >> examples and 2) apply the knowledge to a broad set of situations. >> >> >> >> No neural network today has such capabilities and we don't know how to >> give them such capabilities. Neural networks need large amounts of >> training examples that cover a large variety of situations and then >> the networks can only deal with what the training examples have already >> covered. Neural networks cannot extrapolate in that 'understanding' sense. >> >> >> >> I suggest that understanding truly extrapolates from a piece of >> knowledge. It is not about satisfying a task such as translation between >> languages or drawing hamsters with hats. It is how you got the capability >> to complete the task: Did you only have a few examples that covered >> something different but related and then you extrapolated from that >> knowledge? If yes, this is going in the direction of understanding. Have >> you seen countless examples and then interpolated among them? Then perhaps >> it is not understanding. >> >> >> >> So, for the case of drawing a hamster wearing a red hat, understanding >> perhaps would have taken place if the following happened before that: >> >> >> >> 1) first, the network learned about hamsters (not many examples) >> >> 2) after that the network learned about red hats (outside the context of >> hamsters and without many examples) >> >> 3) finally the network learned about drawing (outside of the context of >> hats and hamsters, not many examples) >> >> >> >> After that, the network is asked to draw a hamster with a red hat. If it >> does it successfully, maybe we have started cracking the problem of >> understanding. >> >> >> >> Note also that this requires the network to learn sequentially without >> exhibiting catastrophic forgetting of the previous knowledge, which is >> possibly also a consequence of human learning by understanding. >> >> >> >> >> >> Danko >> >> >> >> >> >> >> >> >> >> >> >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> >> https://www.linkedin.com/in/danko-nikolic/ >> >> >> --- A progress usually starts with an insight --- >> >> >> >> >> >> >> >> >> Virus-free. www.avast.com >> >> >> >> >> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >> >> Without getting into the specific dispute between Gary and Geoff, I think >> with approaches similar to GLOM, we are finally headed in the right >> direction. There?s plenty of neurophysiological evidence for single-cell >> abstractions and multisensory neurons in the brain, which one might claim >> correspond to symbols. And I think we can finally reconcile the decades old >> dispute between Symbolic AI and Connectionism. >> >> >> >> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >> an effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> GARY: I have *never* called for dismissal of neural networks, but rather >> for some hybrid between the two (as you yourself contemplated in 1991); the >> point of the 2001 book was to characterize exactly where multilayer >> perceptrons succeeded and broke down, and where symbols could complement >> them. >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> >> *From:* Connectionists *On >> Behalf Of *Gary Marcus >> *Sent:* Wednesday, February 2, 2022 1:26 PM >> *To:* Geoffrey Hinton >> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Geoff, and interested others, >> >> >> >> What, for example, would you make of a system that often drew the >> red-hatted hamster you requested, and perhaps a fifth of the time gave you >> utter nonsense? Or say one that you trained to create birds but sometimes >> output stuff like this: >> >> >> >> >> >> >> >> One could >> >> >> >> a. avert one?s eyes and deem the anomalous outputs irrelevant >> >> or >> >> b. wonder if it might be possible that sometimes the system gets the >> right answer for the wrong reasons (eg partial historical contingency), and >> wonder whether another approach might be indicated. >> >> >> >> Benchmarks are harder than they look; most of the field has come to >> recognize that. The Turing Test has turned out to be a lousy measure of >> intelligence, easily gamed. It has turned out empirically that the Winograd >> Schema Challenge did not measure common sense as well as Hector might have >> thought. (As it happens, I am a minor coauthor of a very recent review on >> this very topic: https://arxiv.org/abs/2201.02387 >> ) >> But its conquest in no way means machines now have common sense; many >> people from many different perspectives recognize that (including, e.g., >> Yann LeCun, who generally tends to be more aligned with you than with me). >> >> >> >> So: on the goalpost of the Winograd schema, I was wrong, and you can >> quote me; but what you said about me and machine translation remains your >> invention, and it is inexcusable that you simply ignored my 2019 >> clarification. On the essential goal of trying to reach meaning and >> understanding, I remain unmoved; the problem remains unsolved. >> >> >> >> All of the problems LLMs have with coherence, reliability, truthfulness, >> misinformation, etc stand witness to that fact. (Their persistent inability >> to filter out toxic and insulting remarks stems from the same.) I am hardly >> the only person in the field to see that progress on any given benchmark >> does not inherently mean that the deep underlying problems have solved. >> You, yourself, in fact, have occasionally made that point. >> >> >> >> With respect to embeddings: Embeddings are very good for natural language >> *processing*; but NLP is not the same as NL*U* ? when it comes to >> *understanding*, their worth is still an open question. Perhaps they >> will turn out to be necessary; they clearly aren?t sufficient. In their >> extreme, they might even collapse into being symbols, in the sense of >> uniquely identifiable encodings, akin to the ASCII code, in which a >> specific set of numbers stands for a specific word or concept. (Wouldn?t >> that be ironic?) >> >> >> >> (Your GLOM, which as you know I praised publicly, is in many ways an >> effort to wind up with encodings that effectively serve as symbols in >> exactly that way, guaranteed to serve as consistent representations of >> specific concepts.) >> >> >> >> Notably absent from your email is any kind of apology for misrepresenting >> my position. It?s fine to say that ?many people thirty years ago once >> thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. >> I have consistently felt throughout our interactions that you have mistaken >> me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me >> for having made that error. I am still not he. >> >> >> >> Which maybe connects to the last point; if you read my work, you would >> see thirty years of arguments *for* neural networks, just not in the way >> that you want them to exist. I have ALWAYS argued that there is a role for >> them; characterizing me as a person ?strongly opposed to neural networks? >> misses the whole point of my 2001 book, which was subtitled ?Integrating >> Connectionism and Cognitive Science.? >> >> >> >> In the last two decades or so you have insisted (for reasons you have >> never fully clarified, so far as I know) on abandoning symbol-manipulation, >> but the reverse is not the case: I have *never* called for dismissal of >> neural networks, but rather for some hybrid between the two (as you >> yourself contemplated in 1991); the point of the 2001 book was to >> characterize exactly where multilayer perceptrons succeeded and broke down, >> and where symbols could complement them. It?s a rhetorical trick (which is >> what the previous thread was about) to pretend otherwise. >> >> >> >> Gary >> >> >> >> >> >> On Feb 2, 2022, at 11:22, Geoffrey Hinton >> wrote: >> >> ? >> >> Embeddings are just vectors of soft feature detectors and they are very >> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >> opposite. >> >> >> >> A few decades ago, everyone I knew then would have agreed that the >> ability to translate a sentence into many different languages was strong >> evidence that you understood it. >> >> >> >> But once neural networks could do that, their critics moved the >> goalposts. An exception is Hector Levesque who defined the goalposts more >> sharply by saying that the ability to get pronoun references correct in >> Winograd sentences is a crucial test. Neural nets are improving at that but >> still have some way to go. Will Gary agree that when they can get pronoun >> references correct in Winograd sentences they really do understand? Or does >> he want to reserve the right to weasel out of that too? >> >> >> >> Some people, like Gary, appear to be strongly opposed to neural networks >> because they do not fit their preconceived notions of how the mind should >> work. >> >> I believe that any reasonable person would admit that if you ask a neural >> net to draw a picture of a hamster wearing a red hat and it draws such a >> picture, it understood the request. >> >> >> >> Geoff >> >> >> >> >> >> >> >> >> >> >> >> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >> >> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural >> network community, >> >> >> >> There has been a lot of recent discussion on this list about framing and >> scientific integrity. Often the first step in restructuring narratives is >> to bully and dehumanize critics. The second is to misrepresent their >> position. People in positions of power are sometimes tempted to do this. >> >> >> >> The Hinton-Hanson interview that you just published is a real-time >> example of just that. It opens with a needless and largely content-free >> personal attack on a single scholar (me), with the explicit intention of >> discrediting that person. Worse, the only substantive thing it says is >> false. >> >> >> >> Hinton says ?In 2015 he [Marcus] made a prediction that computers >> wouldn?t be able to do machine translation.? >> >> >> >> I never said any such thing. >> >> >> >> What I predicted, rather, was that multilayer perceptrons, as they >> existed then, would not (on their own, absent other mechanisms) >> *understand* language. Seven years later, they still haven?t, except in >> the most superficial way. >> >> >> >> I made no comment whatsoever about machine translation, which I view as a >> separate problem, solvable to a certain degree by correspondance without >> semantics. >> >> >> >> I specifically tried to clarify Hinton?s confusion in 2019, but, >> disappointingly, he has continued to purvey misinformation despite that >> clarification. Here is what I wrote privately to him then, which should >> have put the matter to rest: >> >> >> >> You have taken a single out of context quote [from 2015] and >> misrepresented it. The quote, which you have prominently displayed at the >> bottom on your own web page, says: >> >> >> >> Hierarchies of features are less suited to challenges such as language, >> inference, and high-level planning. For example, as Noam Chomsky famously >> pointed out, language is filled with sentences you haven't seen >> before. Pure classifier systems don't know what to do with such sentences. >> The talent of feature detectors -- in identifying which member of some >> category something belongs to -- doesn't translate into understanding >> novel sentences, in which each sentence has its own unique meaning. >> >> >> >> It does *not* say "neural nets would not be able to deal with novel >> sentences"; it says that hierachies of features detectors (on their own, if >> you read the context of the essay) would have trouble *understanding *novel sentences. >> >> >> >> >> Google Translate does yet not *understand* the content of the sentences >> is translates. It cannot reliably answer questions about who did what to >> whom, or why, it cannot infer the order of the events in paragraphs, it >> can't determine the internal consistency of those events, and so forth. >> >> >> >> Since then, a number of scholars, such as the the computational linguist >> Emily Bender, have made similar points, and indeed current LLM difficulties >> with misinformation, incoherence and fabrication all follow from these >> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf >> , >> also emphasizing issues of understanding and meaning: >> >> >> >> *The success of the large neural language models on many NLP tasks is >> exciting. However, we find that these successes sometimes lead to hype in >> which these models are being described as ?understanding? language or >> capturing ?meaning?. In this position paper, we argue that a system trained >> only on form has a priori no way to learn meaning. .. a clear understanding >> of the distinction between form and meaning will help guide the field >> towards better science around natural language understanding. * >> >> >> >> Her later article with Gebru on language models ?stochastic parrots? is >> in some ways an extension of this point; machine translation requires >> mimicry, true understanding (which is what I was discussing in 2015) >> requires something deeper than that. >> >> >> >> Hinton?s intellectual error here is in equating machine translation with >> the deeper comprehension that robust natural language understanding will >> require; as Bender and Koller observed, the two appear not to be the same. >> (There is a longer discussion of the relation between language >> understanding and machine translation, and why the latter has turned out to >> be more approachable than the former, in my 2019 book with Ernest Davis). >> >> >> >> More broadly, Hinton?s ongoing dismissiveness of research from >> perspectives other than his own (e.g. linguistics) have done the field a >> disservice. >> >> >> >> As Herb Simon once observed, science does not have to be zero-sum. >> >> >> >> Sincerely, >> >> Gary Marcus >> >> Professor Emeritus >> >> New York University >> >> >> >> On Feb 2, 2022, at 06:12, AIhub wrote: >> >> ? >> >> Stephen Hanson in conversation with Geoff Hinton >> >> >> >> In the latest episode of this video series for AIhub.org >> , >> Stephen Hanson talks to Geoff Hinton about neural networks, >> backpropagation, overparameterization, digit recognition, voxel cells, >> syntax and semantics, Winograd sentences, and more. >> >> >> >> You can watch the discussion, and read the transcript, here: >> >> >> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >> >> >> >> >> About AIhub: >> >> AIhub is a non-profit dedicated to connecting the AI community to the >> public by providing free, high-quality information through AIhub.org >> >> (https://aihub.org/ >> ). >> We help researchers publish the latest AI news, summaries of their work, >> opinion pieces, tutorials and more. We are supported by many leading >> scientific organizations in AI, namely AAAI >> , >> NeurIPS >> , >> ICML >> , >> AIJ >> >> /IJCAI >> , >> ACM SIGAI >> , >> EurAI/AICOMM, CLAIRE >> >> and RoboCup >> >> . >> >> Twitter: @aihuborg >> >> >> >> >> >> >> Virus-free. www.avast.com >> >> >> >> >> -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Mon Feb 7 21:33:45 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Tue, 8 Feb 2022 02:33:45 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <11D154CE-8AED-4E5E-9C63-61CBD3CDCC52@asu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <78E1F023-F059-4762-AAC1-17F2387F3819@asu.edu> <0C8B57DC-3E91-4668-B3AE-268E92746232@asu.edu> <11D154CE-8AED-4E5E-9C63-61CBD3CDCC52@asu.edu> Message-ID: <585F2E65-8710-4F0F-909B-FBC4236516B5@asu.edu> Dear John, You may be asking the wrong questions. We can talk at WCCI 2022. Asim Sent from my iPhone On Feb 7, 2022, at 4:42 PM, Asim Roy wrote: ? Dear John, Our method incorporates many aspects of human learning. Come to WCCI 2022 and we can talk at length. Asim Sent from my iPhone On Feb 7, 2022, at 4:34 PM, Juyang Weng wrote: ? Dear Asim, I repeat my previous statement: The biggest trap is to avoid brain problems and only work on so-called "an engineering problem" or "a focused problem". AI is in a crisis caused by Deep Learning using CNN etc., because even people like you who are interested in brain problems are still saying "We are not doing any of that you are thinking". Raising the problems that the brain is facing is the first step toward a solution to brain problems. Best regard, -John On Mon, Feb 7, 2022 at 6:00 PM Asim Roy > wrote: Dear John, We are not doing any of that you are thinking. Come to our tutorial at WCCI 2022 in Padua, Italy. I love being in Italy. I have perhaps traveled more in Italy than any other EU country. And I love the people. Wish I knew the language. By the way, join the new Society of Explainable AI. We will do a conference in San Francisco in late July/ early August. All the best, Asim Sent from my iPhone On Feb 7, 2022, at 3:46 PM, Juyang Weng > wrote: ? Dear Asim, You wrote: "If it was skull-closed, Humans would barely learn anything." The skull is closed throughout life, but sensors and effectors are connected to the brain. Humans teach the brain as part of the environment. You wrote: "We teach a system about parts just like you would teach a kid or an adult about parts." When you teach, how does the kid's brain segment your 10 body parts and 1000-10=990 other parts in your classroom? You cannot assume that the kid recognizes and segments you. This is what I said about the chicken and egg problem. By skull-closed, I mean the following examples are invalid: (1) a human teacher tunes parameters in the brain-network like many CNNs and LSTMs have done because the human knows the data or (2) a human teacher assigns a neuron for a particular role (e.g., a symbol) as many symbolic networks have done (e.g. Joshua Tenenbaum's?), or (3) a human teacher assigns a group of neurons for a particular role (e.g., edge detectors), as many so called mental architectures have done. Sorry, Cresceptron and DN-1 did that, but not DN-2. Best regards, -John On Mon, Feb 7, 2022 at 4:39 PM Asim Roy > wrote: Dear John, We teach a system about parts just like you would teach a kid or an adult about parts. There?s nothing ?skull-closed? about it. If it was ?skull-closed,? Humans would barely learn anything. And dealing with a thousand parts should not be a problem. Humans remember more than a thousand parts. And cluttered images are not a problem. We have dealt with satellite and other cluttered images already. So we are not looking at Mickey Mouse problems. Asim Sent from my iPhone On Feb 7, 2022, at 8:59 AM, Juyang Weng > wrote: ? Dear Asim, The biggest trap is to avoid brain problems and only work on so called "an engineering problem" or "a focused problem". In a cluttered scene, there maybe n>10 objects and m>> 1000 parts. A brain model must explain how the skull-closed brain segments objects and their parts. The traditional and also your "recognition by parts" approach does not explain (1) how different objects are segmented and (2) how different parts are segmented and group into objects. Note "skull-closed" condition. No supervision into the brain network is allowed. My conscious learning model using DN addresses such compounding problems while skull is closed (using unsupervised Hebbian learning) and without a given task. J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, USA, Jan.7-9, 2022. http://www.cse.msu.edu/~weng/research/ConsciousLearning-ICCE-2022-rvsd-cite.pdf YouTube talk: [image.png] My model must not address the above problems only. Best regards, -John On Sun, Feb 6, 2022 at 10:43 PM Asim Roy > wrote: Dear John, We recognize whole objects, but at the same time we verify its parts. Best, Asim From: Juyang Weng > Sent: Sunday, February 6, 2022 8:38 PM To: Asim Roy > Cc: Geoffrey Hinton >; Dietterich, Thomas >; AIhub >; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus >; Danko Nikolic >; Stephen Jos? Hanson >; Marek Reformat >; MARCO GORI >; Alessandro Sperduti >; Xiaodong Li >; Hava Siegelmann >; Peter Tino >; Bing Xue >; Ali Minai >; Claudius Gros >; Jean-Philippe Thivierge >; Tsvi Achler >; Prof A Hussain > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, The brain does not assume a single object in a cluttered science. Thus, a simple explanation like "recognition by parts" (but without object abstraction) should be invalid. It is like a chicken and egg problem. Both chicken and egg are absent. We must not assume egg is there or chicken is there. Best regards, -John On Sun, Feb 6, 2022 at 2:42 PM Asim Roy > wrote: Dear John, You are right and I admit I am not solving all of the problems. It?s just in reference to this one problem that Geoffrey Hinton mentions that I think can be resolved: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? Best, Asim From: Juyang Weng > Sent: Sunday, February 6, 2022 10:06 AM To: Asim Roy > Cc: Geoffrey Hinton >; Dietterich, Thomas >; AIhub >; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus >; Danko Nikolic >; Stephen Jos? Hanson >; Marek Reformat >; MARCO GORI >; Alessandro Sperduti >; Xiaodong Li >; Hava Siegelmann >; Peter Tino >; Bing Xue >; Ali Minai >; Claudius Gros >; Jean-Philippe Thivierge >; Tsvi Achler >; Prof A Hussain > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I try to be brief so that I can explain why many of us have missed, and will continue to miss, the boat. In some of my talks, I have a ppt slide "The brain is like blindmen and an elephant". Unfortunately, your "identify objects based on its parts" is a good traditional idea from pattern recognition that is still a blindman. Your idea does not explain many other problems without which we will never understand a biological brain. For example, your idea does not explain how the brain learns planning and discovery in a cluttered world. We must solve many million-dollar problems holistically. Please watch my YouTube video: Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar Problems Solved https://youtu.be/Dgx1dLCdSKY Best regards, -John On Sat, Feb 5, 2022 at 12:01 AM Asim Roy > wrote: I am responding to this part of Geoffrey Hinton?s note: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? The causal explanation is actually done quite simply, and we are doing it currently. I can talk about this now because Arizona State University (ASU) has filed a provisional patent application on the technology. The basic idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable Artificial Intelligence (darpa.mil)) and illustrated in the figure below. The idea is to identify objects based on its parts. So, the figure below says that it?s a cat because it has fur, whiskers, and claws plus an unlabeled visual feature. I am not sure if DARPA got anything close to this from its funding of various entities. What this means is that you need a parts model. And we do that. In the case of MNIST handwritten digits that Geoff mentions, we ?teach? this parts model what the top part of a digit ?3? looks like, what the bottom part looks like and so on. And we also teach connectivity between parts and the composition of objects from parts. And we do that for all digits. And we get a symbolic model sitting on top of a CNN model that provides the explanation that Geoff is referring to as the causal explanation. This ?teaching? is similar to the way you would teach a kid to recognize different digits. An advantage of this parts model, in addition to being in an explainable symbolic form, is robustness to adversarial attack. We recently tested on the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast gradient method to 27%, our XAI model maintained an accuracy of 90%, probably higher. In general, it would be hard to make a school bus look like an ostrich, with a few pixel changes, if you can identify the parts of a school bus and an ostrich. A parts model that DARPA wanted provides both a symbolic explanation and adversarial protection. The problem that Geoffrey is referring to is solved. I am doing a tutorial on this at IEEE World Congress on Computational Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy 18-23 July). I am copying the organizers and want to thank them for accepting the tutorial proposal. The only other presentation I have done on this is at a Military Operations Research Society (MORS) meeting last December. So, back to the future. Hybrid models might indeed save deep learning models and let us deploy these models without concern. We might not even need adversarial training of any kind. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) www.teuvonet.com [Timeline Description automatically generated] From: Connectionists > On Behalf Of Geoffrey Hinton Sent: Friday, February 4, 2022 1:24 PM To: Dietterich, Thomas > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists > On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic > Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg Virus-free. www.avast.com -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: image.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: image.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: From gary.marcus at nyu.edu Mon Feb 7 17:07:27 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Mon, 7 Feb 2022 14:07:27 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <3dff1e2e-52b6-2ff3-3152-c7cf919fb14e@rubic.rutgers.edu> References: <537DF004-25CE-45A2-8155-D7E6018F4EE5@nyu.edu> <3dff1e2e-52b6-2ff3-3152-c7cf919fb14e@rubic.rutgers.edu> Message-ID: Stephen, I don?t doubt for a minute that deep learning can characterize some aspects of psychology reasonably well; but either it needs to expands its borders or else be used in conjunction with other techniques. Take for example the name of the new Netflix show The Woman in the House Across the Street from the Girl in the Window Most of us can infer, compositionally, from that unusually long noun phrase, that the title is a description of particular person, that the title is not a complete sentence, and that the woman in question lives in a house; we also infer that there is a second, distinct person (likely a child) across the street, and so forth. We can also use some knowledge of pragmatics to infer that the woman in question is likely to be the protagonist in the show. Current systems still struggle with that sort of thing. We can then watch the show (I watched a few minutes of Episode 1) and quickly relate the title to the protagonist?s mental state, start to develop a mental model of the protagonist?s relation to her new neighbors, make inferences about whether certain choices appear to be ?within character?, empathize with character or question her judgements, etc, all with respect to a mental model that is rapidly encoded and quickly modified. I think that an understanding of how people build and modify such models would be extremely valuable (not just for fiction for everyday reality), but I don?t see how deep learning in its current form gives us much purchase on that. There is plenty of precedent for the kind of mental processes I am sketching (e.g Walter Kintsch?s work on text comprehension; Kamp/Kratzer/Heim work on discourse representation, etc) from psychological and linguistic perspectives, but almost no current contact in the neural network community with these well-attested psychological processes. Gary > On Feb 7, 2022, at 6:01 AM, Stephen Jos? Hanson wrote: > > Gary, > > This is one of the first posts of yours, that I can categorically agree with! > > I think building cognitive models through *some* training regime or focused sampling or architectures or something but not explicit, for example. > > The other fundamental cognitive/perceptual capability in this context is the ability of Neural Networks to do what Shepard (1970; Garner 1970s), had modeled as perceptual separable processing (finding parts) and perceptual integral process (finding covariance and structure). > > Shepard argued these fundamental perceptual processes were dependent on development and learning. > > A task was created with double dissociation of a categorization problem. In one case: separable ( in effect, uncorrelated features in the stimulus) were presented in categorization task that required you pay attention to at least 2 features at the same time to categorize correctly ("condensation"). in the other case: integral stimuli (in effect correlated features in stimuli) were presented in a categorization task that required you to ignore the correlation and do categorize on 1 feature at a time ("filtration"). This produced a result that separable stimuli were more quickly learned in filtration tasks then integral stimuli in condensation tasks. Non-intuitively, Separable stimuli are learned more slowly in condensation tasks then integral stimuli then in filtration tasks. In other words attention to feature structure could cause improvement in learning or interference. Not that surprising.. however-- > > In the 1980s NN with single layers (Backprop) *could not* replicate this simple problem indicating that the cognitive model was somehow inadequate. Backprop simply learned ALL task/stimuli parings at the same rate, ignoring the subtle but critical difference. It failed. > > Recently we (https://www.frontiersin.org/articles/10.3389/fpsyg.2018.00374/full?&utm_source=Email_to_authors_&utm_medium=Email&utm_content=T1_11.5e1_author&utm_campaign=Email_publication&field=&journalName=Frontiers_in_Psychology&id=284733 ) were able to show that JUST BY ADDING LAYERS the DL does match to human performance. > > What are the layers doing? We offer an possible explanation that needs testing. Layers, appear to create a type of buffer that allows the network to "curate", feature detectors that are spatially distant from the input (conv layer, for example), this curation comes in various attention forms (something in that will appear in a new paper--not enough room here), which appears to qualitatively change the network processing states, and cognitive capabilities. Well, that's the claim. > > The larger point, is that apparently architectures interact with learning rules, in ways that can cross this symbolic/neural river of styx, without falling into it. > > Steve > > > > On 2/5/22 10:38 AM, Gary Marcus wrote: >> There is no magic in understanding, just computation that has been realized in the wetware of humans and that eventually can be realized in machines. But understanding is not (just) learning. >> >> Understanding incorporates (or works in tandem with) learning - but also, critically, in tandem with inference, and the development and maintenance of cognitive models. Part of developing an understanding of cats in general is to learn long term-knowledge about their properties, both directly (e.g., through observation) and indirectly (eg through learning facts about animals in general that can be extended to cats), often through inference (if all animals have DNA, and a cat is an animal, it must also have DNA). The understanding of a particular cat also involves direct observation, but also inference (eg one might surmise that the reason that Fluffy is running about the room is that Fluffy suspects there is a mouse stirring somewhere nearby). But all of that, I would say, is subservient to the construction of cognitive models that can be routinely updated (e.g., Fluffy is currently in the living room, skittering about, perhaps looking for a mouse). >> >> In humans, those dynamic, relational models, which form part of an understanding, can support inference (if Fluffy is in the living room, we can infer that Fluffy is not outside, not lost, etc). Without such models - which I think represent a core part of understanding - AGI is an unlikely prospect. >> >> Current neural networks, as it happens, are better at acquiring long-term knowledge (cats have whiskers) than they are at dynamically updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic model that I am describing. To a modest degree they can approximate it on the basis of large samples of texts, but their ultimate incoherence stems from the fact that they do not have robust internal cognitive models that they can update on the fly. >> >> Without such cognitive models you can still capture some aspects of understanding (eg predicting that cats are likely to be furry), but things fall apart quickly; inference is never reliable, and coherence is fleeting. >> >> As a final note, one of the most foundational challenges in constructing adequate cognitive models of the world is to have a clear distinction between individuals and kinds; as I emphasized 20 years ago (in The Algebraic Mind), this has always been a weakness in neural networks, and I don?t think that the type-token problem has yet been solved. >> >> Gary >> >> >>> On Feb 5, 2022, at 01:31, Asim Roy wrote: >>> >>> ? >>> All, >>> >>> I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? >>> >>> https://www.youtube.com/watch?v=gn4nRCC9TwQ >>> https://www.youtube.com/watch?v=8sO7VS3q8d0 >>> >>> >>> Asim Roy >>> Professor, Information Systems >>> Arizona State University >>> Lifeboat Foundation Bios: Professor Asim Roy >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> From: Ali Minai >>> Sent: Friday, February 4, 2022 11:38 PM >>> To: Asim Roy >>> Cc: Gary Marcus ; Danko Nikolic ; Brad Wyble ;connectionists at mailman.srv.cs.cmu.edu ; AIhub >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> Asim >>> >>> Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. >>> >>> FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! >>> >>> Ali >>> >>> Ali A. Minai, Ph.D. >>> Professor and Graduate Program Director >>> Complex Adaptive Systems Lab >>> Department of Electrical Engineering & Computer Science >>> 828 Rhodes Hall >>> University of Cincinnati >>> Cincinnati, OH 45221-0030 >>> >>> Phone: (513) 556-4783 >>> Fax: (513) 556-7326 >>> Email: Ali.Minai at uc.edu >>> minaiaa at gmail.com >>> >>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>> >>> >>> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy > wrote: >>> First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. >>> >>> DANKO: >>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>> >>> Asim Roy >>> Professor, Information Systems >>> Arizona State University >>> Lifeboat Foundation Bios: Professor Asim Roy >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> From: Gary Marcus > >>> Sent: Thursday, February 3, 2022 9:26 AM >>> To: Danko Nikolic > >>> Cc: Asim Roy >; Geoffrey Hinton >; AIhub >;connectionists at mailman.srv.cs.cmu.edu >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> Dear Danko, >>> >>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf ] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. >>> >>> I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. >>> >>> Gary >>> >>> >>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: >>> >>> G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." >>> >>> I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". >>> To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. >>> >>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>> >>> ...and so on. >>> >>> Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). >>> >>> I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. >>> >>> In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. >>> >>> No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. >>> >>> I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. >>> >>> So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: >>> >>> 1) first, the network learned about hamsters (not many examples) >>> 2) after that the network learned about red hats (outside the context of hamsters and without many examples) >>> 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) >>> >>> After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. >>> >>> Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. >>> >>> >>> Danko >>> >>> >>> >>> >>> >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> https://www.linkedin.com/in/danko-nikolic/ >>> --- A progress usually starts with an insight --- >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy > wrote: >>> Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. >>> >>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>> GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. >>> >>> Asim Roy >>> Professor, Information Systems >>> Arizona State University >>> Lifeboat Foundation Bios: Professor Asim Roy >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> From: Connectionists > On Behalf Of Gary Marcus >>> Sent: Wednesday, February 2, 2022 1:26 PM >>> To: Geoffrey Hinton > >>> Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>> >>> Dear Geoff, and interested others, >>> >>> What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: >>> >>> >>> >>> One could >>> >>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>> or >>> b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. >>> >>> Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387 ) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). >>> >>> So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. >>> >>> All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. >>> >>> With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) >>> >>> (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>> >>> Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. >>> >>> Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? >>> >>> In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. >>> >>> Gary >>> >>> >>> On Feb 2, 2022, at 11:22, Geoffrey Hinton > wrote: >>> >>> ? >>> Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. >>> >>> A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. >>> >>> >>> But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? >>> >>> Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. >>> I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. >>> >>> Geoff >>> >>> >>> >>> >>> >>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus > wrote: >>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, >>> >>> There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. >>> >>> The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. >>> >>> Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? >>> >>> I never said any such thing. >>> >>> What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. >>> >>> I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. >>> >>> I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: >>> >>> You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: >>> >>> Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. >>> >>> It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. >>> >>> Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. >>> >>> Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf , also emphasizing issues of understanding and meaning: >>> >>> The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. >>> >>> Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. >>> >>> Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). >>> >>> More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. >>> >>> As Herb Simon once observed, science does not have to be zero-sum. >>> >>> Sincerely, >>> Gary Marcus >>> Professor Emeritus >>> New York University >>> >>> >>> On Feb 2, 2022, at 06:12, AIhub > wrote: >>> >>> ? >>> Stephen Hanson in conversation with Geoff Hinton >>> >>> In the latest episode of this video series for AIhub.org , Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. >>> >>> You can watch the discussion, and read the transcript, here: >>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>> >>> About AIhub: >>> AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/ ). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI , NeurIPS , ICML , AIJ /IJCAI , ACM SIGAI , EurAI/AICOMM, CLAIRE and RoboCup . >>> Twitter: @aihuborg >>> >>> >>> Virus-free. www.avast.com >>> > -- > -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Mon Feb 7 16:39:04 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Mon, 7 Feb 2022 21:39:04 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: <78E1F023-F059-4762-AAC1-17F2387F3819@asu.edu> Dear John, We teach a system about parts just like you would teach a kid or an adult about parts. There?s nothing ?skull-closed? about it. If it was ?skull-closed,? Humans would barely learn anything. And dealing with a thousand parts should not be a problem. Humans remember more than a thousand parts. And cluttered images are not a problem. We have dealt with satellite and other cluttered images already. So we are not looking at Mickey Mouse problems. Asim Sent from my iPhone On Feb 7, 2022, at 8:59 AM, Juyang Weng wrote: ? Dear Asim, The biggest trap is to avoid brain problems and only work on so called "an engineering problem" or "a focused problem". In a cluttered scene, there maybe n>10 objects and m>> 1000 parts. A brain model must explain how the skull-closed brain segments objects and their parts. The traditional and also your "recognition by parts" approach does not explain (1) how different objects are segmented and (2) how different parts are segmented and group into objects. Note "skull-closed" condition. No supervision into the brain network is allowed. My conscious learning model using DN addresses such compounding problems while skull is closed (using unsupervised Hebbian learning) and without a given task. J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, USA, Jan.7-9, 2022. http://www.cse.msu.edu/~weng/research/ConsciousLearning-ICCE-2022-rvsd-cite.pdf YouTube talk: [image.png] My model must not address the above problems only. Best regards, -John On Sun, Feb 6, 2022 at 10:43 PM Asim Roy > wrote: Dear John, We recognize whole objects, but at the same time we verify its parts. Best, Asim From: Juyang Weng > Sent: Sunday, February 6, 2022 8:38 PM To: Asim Roy > Cc: Geoffrey Hinton >; Dietterich, Thomas >; AIhub >; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus >; Danko Nikolic >; Stephen Jos? Hanson >; Marek Reformat >; MARCO GORI >; Alessandro Sperduti >; Xiaodong Li >; Hava Siegelmann >; Peter Tino >; Bing Xue >; Ali Minai >; Claudius Gros >; Jean-Philippe Thivierge >; Tsvi Achler >; Prof A Hussain > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, The brain does not assume a single object in a cluttered science. Thus, a simple explanation like "recognition by parts" (but without object abstraction) should be invalid. It is like a chicken and egg problem. Both chicken and egg are absent. We must not assume egg is there or chicken is there. Best regards, -John On Sun, Feb 6, 2022 at 2:42 PM Asim Roy > wrote: Dear John, You are right and I admit I am not solving all of the problems. It?s just in reference to this one problem that Geoffrey Hinton mentions that I think can be resolved: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? Best, Asim From: Juyang Weng > Sent: Sunday, February 6, 2022 10:06 AM To: Asim Roy > Cc: Geoffrey Hinton >; Dietterich, Thomas >; AIhub >; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus >; Danko Nikolic >; Stephen Jos? Hanson >; Marek Reformat >; MARCO GORI >; Alessandro Sperduti >; Xiaodong Li >; Hava Siegelmann >; Peter Tino >; Bing Xue >; Ali Minai >; Claudius Gros >; Jean-Philippe Thivierge >; Tsvi Achler >; Prof A Hussain > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I try to be brief so that I can explain why many of us have missed, and will continue to miss, the boat. In some of my talks, I have a ppt slide "The brain is like blindmen and an elephant". Unfortunately, your "identify objects based on its parts" is a good traditional idea from pattern recognition that is still a blindman. Your idea does not explain many other problems without which we will never understand a biological brain. For example, your idea does not explain how the brain learns planning and discovery in a cluttered world. We must solve many million-dollar problems holistically. Please watch my YouTube video: Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar Problems Solved https://youtu.be/Dgx1dLCdSKY Best regards, -John On Sat, Feb 5, 2022 at 12:01 AM Asim Roy > wrote: I am responding to this part of Geoffrey Hinton?s note: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? The causal explanation is actually done quite simply, and we are doing it currently. I can talk about this now because Arizona State University (ASU) has filed a provisional patent application on the technology. The basic idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable Artificial Intelligence (darpa.mil)) and illustrated in the figure below. The idea is to identify objects based on its parts. So, the figure below says that it?s a cat because it has fur, whiskers, and claws plus an unlabeled visual feature. I am not sure if DARPA got anything close to this from its funding of various entities. What this means is that you need a parts model. And we do that. In the case of MNIST handwritten digits that Geoff mentions, we ?teach? this parts model what the top part of a digit ?3? looks like, what the bottom part looks like and so on. And we also teach connectivity between parts and the composition of objects from parts. And we do that for all digits. And we get a symbolic model sitting on top of a CNN model that provides the explanation that Geoff is referring to as the causal explanation. This ?teaching? is similar to the way you would teach a kid to recognize different digits. An advantage of this parts model, in addition to being in an explainable symbolic form, is robustness to adversarial attack. We recently tested on the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast gradient method to 27%, our XAI model maintained an accuracy of 90%, probably higher. In general, it would be hard to make a school bus look like an ostrich, with a few pixel changes, if you can identify the parts of a school bus and an ostrich. A parts model that DARPA wanted provides both a symbolic explanation and adversarial protection. The problem that Geoffrey is referring to is solved. I am doing a tutorial on this at IEEE World Congress on Computational Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy 18-23 July). I am copying the organizers and want to thank them for accepting the tutorial proposal. The only other presentation I have done on this is at a Military Operations Research Society (MORS) meeting last December. So, back to the future. Hybrid models might indeed save deep learning models and let us deploy these models without concern. We might not even need adversarial training of any kind. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) www.teuvonet.com [Timeline Description automatically generated] From: Connectionists > On Behalf Of Geoffrey Hinton Sent: Friday, February 4, 2022 1:24 PM To: Dietterich, Thomas > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure. This does not mean that we should give up on trying to make artificial neural nets work more like real ones. People can see a tilted square as either an upright diamond or a tilted square and, so far as I know, a convnet does not exhibit this type of alternate percept. People seem to impose hierarchical structural descriptions on images and sound waves and they clearly impose intrinsic coordinate frames on wholes and parts. If this is what Gary means by symbolic then I don?t disagree that neural nets should do symbol processing. However, there is a very different meaning of "symbolic". A pure atomic symbol has no internal structure. The form of the symbol itself tells you nothing about what it denotes. The only relevant properties it has are that it's identical to other instances of the same symbol and different from all other symbols. That's totally different from a neural net that uses embedding vectors. Embedding vectors have a rich internal structure that dictates how they interact with other embedding vectors. What I really object to is the following approach: Start with pure symbols and rules for how to manipulate structures made out of pure symbols. These structures themselves can be denoted by symbols that correspond to memory addresses where the bits in the address tell you nothing about the content of the structure at that address. Then when the rule-based approach doesn't work for dealing with the real world (e.g. machine translation) try to use neural nets to convert the real world into pure symbols and then carry on with the rule-based approach. That is like using an electric motor to inject the gasoline into the same old gasoline engine instead of just replacing the gasoline engine with an electric motor. On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas > wrote: ?Understanding? is not a Boolean. It is a theorem that no system can enumerate all of the consequences of a state of affairs in the world. For low-stakes application work, we can be satisfied by a system that ?does the right thing?. If the system draws a good picture, that?s sufficient. It ?understood? the request. But for higher-stakes applications---and for advancing the science---we seek a causal account of how the components of a system cause it to do the right thing. We are hoping that a small set of mechanisms can produce broad coverage of intelligent behavior. This gives us confidence that the system will respond correctly outside of the narrow tasks on which we have tested it. --Tom Thomas G. Dietterich, Distinguished Professor Emeritus School of Electrical Engineering and Computer Science US Mail: 1148 Kelley Engineering Center Office: 2067 Kelley Engineering Center Oregon State Univ., Corvallis, OR 97331-5501 Voice: 541-737-5559; FAX: 541-737-1300 URL: http://web.engr.oregonstate.edu/~tgd/ From: Connectionists > On Behalf Of Gary Marcus Sent: Thursday, February 3, 2022 8:26 AM To: Danko Nikolic > Cc: connectionists at mailman.srv.cs.cmu.edu; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton [This email originated from outside of OSU. Use caution with links and attachments.] Dear Danko, Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. Gary On Feb 3, 2022, at 3:19 AM, Danko Nikolic > wrote: G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) ...and so on. Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: 1) first, the network learned about hamsters (not many examples) 2) after that the network learned about red hats (outside the context of hamsters and without many examples) 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- Virus-free. www.avast.com On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: Geoffrey Hinton Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Geoff, and interested others, What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: One could a. avert one?s eyes and deem the anomalous outputs irrelevant or b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. Gary On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: ? Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. Geoff On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? I never said any such thing. What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. As Herb Simon once observed, science does not have to be zero-sum. Sincerely, Gary Marcus Professor Emeritus New York University On Feb 2, 2022, at 06:12, AIhub wrote: ? Stephen Hanson in conversation with Geoff Hinton In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. You can watch the discussion, and read the transcript, here: https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ About AIhub: AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. Twitter: @aihuborg Virus-free. www.avast.com -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: image.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: image.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: image001.png URL: From juyang.weng at gmail.com Mon Feb 7 18:32:51 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Mon, 7 Feb 2022 18:32:51 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <0C8B57DC-3E91-4668-B3AE-268E92746232@asu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <78E1F023-F059-4762-AAC1-17F2387F3819@asu.edu> <0C8B57DC-3E91-4668-B3AE-268E92746232@asu.edu> Message-ID: Dear Asim, I repeat my previous statement: The biggest trap is to avoid brain problems and only work on so-called "an engineering problem" or "a focused problem". AI is in a crisis caused by Deep Learning using CNN etc., because even people like you who are interested in brain problems are still saying "We are not doing any of that you are thinking". Raising the problems that the brain is facing is the first step toward a solution to brain problems. Best regard, -John On Mon, Feb 7, 2022 at 6:00 PM Asim Roy wrote: > Dear John, > > We are not doing any of that you are thinking. Come to our tutorial at > WCCI 2022 in Padua, Italy. I love being in Italy. I have perhaps traveled > more in Italy than any other EU country. And I love the people. Wish I knew > the language. > > By the way, join the new Society of Explainable AI. We will do a > conference in San Francisco in late July/ early August. > > All the best, > Asim > > Sent from my iPhone > > On Feb 7, 2022, at 3:46 PM, Juyang Weng wrote: > > ? > Dear Asim, > > You wrote: "If it was skull-closed, Humans would barely learn anything." > The skull is closed throughout life, but sensors and effectors are > connected to the brain. Humans teach the brain as part of the > environment. > You wrote: "We teach a system about parts just like you would teach a kid > or an adult about parts." > When you teach, how does the kid's brain segment your 10 body parts and > 1000-10=990 other parts in your classroom? > You cannot assume that the kid recognizes and segments you. This is > what I said about the chicken and egg problem. > By skull-closed, I mean the following examples are invalid: > (1) a human teacher tunes parameters in the brain-network like many CNNs > and LSTMs have done because the human knows the data or > (2) a human teacher assigns a neuron for a particular role (e.g., a > symbol) as many symbolic networks have done (e.g. Joshua Tenenbaum's?), or > (3) a human teacher assigns a group of neurons for a particular role > (e.g., edge detectors), as many so called mental architectures have done. > Sorry, Cresceptron and DN-1 did that, but not DN-2. > Best regards, > -John > > On Mon, Feb 7, 2022 at 4:39 PM Asim Roy wrote: > >> Dear John, >> >> We teach a system about parts just like you would teach a kid or an adult >> about parts. There?s nothing ?skull-closed? about it. If it was >> ?skull-closed,? Humans would barely learn anything. And dealing with a >> thousand parts should not be a problem. Humans remember more than a >> thousand parts. And cluttered images are not a problem. We have dealt with >> satellite and other cluttered images already. So we are not looking at >> Mickey Mouse problems. >> >> Asim >> >> Sent from my iPhone >> >> On Feb 7, 2022, at 8:59 AM, Juyang Weng wrote: >> >> ? >> Dear Asim, >> The biggest trap is to avoid brain problems and only work on so called >> "an engineering problem" or "a focused problem". >> In a cluttered scene, there maybe n>10 objects and m>> 1000 parts. A >> brain model must explain how the skull-closed brain segments objects and >> their parts. >> The traditional and also your "recognition by parts" approach does not >> explain >> (1) how different objects are segmented and >> (2) how different parts are segmented and group into objects. >> Note "skull-closed" condition. No supervision into the brain network is >> allowed. >> My conscious learning model using DN addresses such compounding problems >> while skull is closed (using unsupervised Hebbian learning) and without a >> given task. >> J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th >> International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, >> USA, Jan.7-9, 2022. >> >> http://www.cse.msu.edu/~weng/research/ConsciousLearning-ICCE-2022-rvsd-cite.pdf >> >> YouTube talk: >> [image: image.png] >> My model must not address the above problems only. >> Best regards, >> -John >> >> On Sun, Feb 6, 2022 at 10:43 PM Asim Roy wrote: >> >>> Dear John, >>> >>> >>> >>> We recognize whole objects, but at the same time we verify its parts. >>> >>> >>> >>> Best, >>> >>> Asim >>> >>> >>> >>> *From:* Juyang Weng >>> *Sent:* Sunday, February 6, 2022 8:38 PM >>> *To:* Asim Roy >>> *Cc:* Geoffrey Hinton ; Dietterich, Thomas < >>> tgd at oregonstate.edu>; AIhub ; >>> connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; >>> Danko Nikolic ; Stephen Jos? Hanson < >>> jose at rubic.rutgers.edu>; Marek Reformat ; MARCO >>> GORI ; Alessandro Sperduti < >>> alessandro.sperduti at unipd.it>; Xiaodong Li ; >>> Hava Siegelmann ; Peter Tino < >>> P.Tino at cs.bham.ac.uk>; Bing Xue ; Ali Minai < >>> minaiaa at gmail.com>; Claudius Gros ; >>> Jean-Philippe Thivierge ; Tsvi >>> Achler ; Prof A Hussain < >>> hussain.doctor at gmail.com> >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Dear Asim, >>> >>> The brain does not assume a single object in a cluttered science. Thus, >>> a simple explanation like "recognition by parts" (but without object >>> abstraction) should be invalid. It is like a chicken and egg problem. >>> Both chicken and egg are absent. We must not assume egg is there or >>> chicken is there. >>> >>> Best regards, >>> >>> -John >>> >>> >>> >>> On Sun, Feb 6, 2022 at 2:42 PM Asim Roy wrote: >>> >>> Dear John, >>> >>> >>> >>> You are right and I admit I am not solving all of the problems. It?s >>> just in reference to this one problem that Geoffrey Hinton mentions that I >>> think can be resolved: >>> >>> *?I agree that it's nice to have a causal explanations. But I am not >>> convinced there will ever be a simple causal explanation for how you >>> recognize that a handwritten 2 is a 2. We can introspect on how we do it >>> and this may or may not give some insight into how we check our answer, but >>> the immediate sense that a handwritten 2 is a 2 is computed by a neural net >>> that is not functionally equivalent to any simple and easily explainable >>> procedure.?* >>> >>> Best, >>> >>> Asim >>> >>> >>> >>> *From:* Juyang Weng >>> *Sent:* Sunday, February 6, 2022 10:06 AM >>> *To:* Asim Roy >>> *Cc:* Geoffrey Hinton ; Dietterich, Thomas < >>> tgd at oregonstate.edu>; AIhub ; >>> connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; >>> Danko Nikolic ; Stephen Jos? Hanson < >>> jose at rubic.rutgers.edu>; Marek Reformat ; MARCO >>> GORI ; Alessandro Sperduti < >>> alessandro.sperduti at unipd.it>; Xiaodong Li ; >>> Hava Siegelmann ; Peter Tino < >>> P.Tino at cs.bham.ac.uk>; Bing Xue ; Ali Minai < >>> minaiaa at gmail.com>; Claudius Gros ; >>> Jean-Philippe Thivierge ; Tsvi >>> Achler ; Prof A Hussain < >>> hussain.doctor at gmail.com> >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Dear Asim, >>> >>> I try to be brief so that I can explain why many of us have missed, and >>> will continue to miss, the boat. >>> In some of my talks, I have a ppt slide "The brain is like blindmen and >>> an elephant". >>> >>> Unfortunately, your "identify objects based on its parts" is a good >>> traditional idea from pattern recognition that is still a blindman. >>> >>> Your idea does not explain many other problems without which we will >>> never understand a biological brain. >>> >>> For example, your idea does not explain how the brain learns planning >>> and discovery in a cluttered world. >>> >>> We must solve many million-dollar problems holistically. Please watch >>> my YouTube video: >>> Title: An Algorithmic Theory for Conscious Learning: 10 Million-Dollar >>> Problems Solved >>> https://youtu.be/Dgx1dLCdSKY >>> >>> >>> Best regards, >>> >>> -John >>> >>> >>> >>> On Sat, Feb 5, 2022 at 12:01 AM Asim Roy wrote: >>> >>> I am responding to this part of Geoffrey Hinton?s note: >>> >>> >>> >>> *?I agree that it's nice to have a causal explanations. But I am not >>> convinced there will ever be a simple causal explanation for how you >>> recognize that a handwritten 2 is a 2. We can introspect on how we do it >>> and this may or may not give some insight into how we check our answer, but >>> the immediate sense that a handwritten 2 is a 2 is computed by a neural net >>> that is not functionally equivalent to any simple and easily explainable >>> procedure.?* >>> >>> >>> >>> The causal explanation is actually done quite simply, and we are doing >>> it currently. I can talk about this now because Arizona State University >>> (ASU) has filed a provisional patent application on the technology. The >>> basic idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable >>> Artificial Intelligence (darpa.mil) >>> ) >>> and illustrated in the figure below. The idea is to identify objects based >>> on its parts. So, the figure below says that it?s a cat because it has fur, >>> whiskers, and claws plus an unlabeled visual feature. I am not sure if >>> DARPA got anything close to this from its funding of various entities. What >>> this means is that you need a parts model. And we do that. In the case of >>> MNIST handwritten digits that Geoff mentions, we ?teach? this parts model >>> what the top part of a digit ?3? looks like, what the bottom part looks >>> like and so on. And we also teach connectivity between parts and the >>> composition of objects from parts. And we do that for all digits. And we >>> get a symbolic model sitting on top of a CNN model that provides the >>> explanation that Geoff is referring to as the causal explanation. This >>> ?teaching? is similar to the way you would teach a kid to recognize >>> different digits. >>> >>> >>> >>> An advantage of this parts model, in addition to being in an explainable >>> symbolic form, is robustness to adversarial attack. We recently tested on >>> the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast >>> gradient method to 27%, our XAI model maintained an accuracy of 90%, >>> probably higher. In general, it would be hard to make a school bus look >>> like an ostrich, with a few pixel changes, if you can identify the parts of >>> a school bus and an ostrich. >>> >>> >>> >>> A parts model that DARPA wanted provides both a symbolic explanation and >>> adversarial protection. The problem that Geoffrey is referring to is solved. >>> >>> >>> >>> I am doing a tutorial on this at IEEE World Congress on Computational >>> Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, >>> Italy 18-23 July >>> ). >>> I am copying the organizers and want to thank them for accepting the >>> tutorial proposal. The only other presentation I have done on this is at a >>> Military Operations Research Society (MORS) meeting last December. >>> >>> >>> >>> So, back to the future. Hybrid models might indeed save deep learning >>> models and let us deploy these models without concern. We might not even >>> need adversarial training of any kind. >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> www.teuvonet.com >>> >>> >>> >>> >>> [image: Timeline Description automatically generated] >>> >>> >>> >>> *From:* Connectionists *On >>> Behalf Of *Geoffrey Hinton >>> *Sent:* Friday, February 4, 2022 1:24 PM >>> *To:* Dietterich, Thomas >>> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> I agree that it's nice to have a causal explanations. But I am not >>> convinced there will ever be a simple causal explanation for how you >>> recognize that a handwritten 2 is a 2. We can introspect on how we do it >>> and this may or may not give some insight into how we check our answer, but >>> the immediate sense that a handwritten 2 is a 2 is computed by a neural net >>> that is not functionally equivalent to any simple and easily explainable >>> procedure. >>> >>> >>> >>> This does not mean that we should give up on trying to make artificial >>> neural nets work more like real ones. People can see a tilted square as >>> either an upright diamond or a tilted square and, so far as I know, a >>> convnet does not exhibit this type of alternate percept. People seem to >>> impose hierarchical structural descriptions on images and sound waves and >>> they clearly impose intrinsic coordinate frames on wholes and parts. If >>> this is what Gary means by symbolic then I don?t disagree that neural nets >>> should do symbol processing. However, there is a very different meaning of >>> "symbolic". A pure atomic symbol has no internal structure. The form of the >>> symbol itself tells you nothing about what it denotes. The only relevant >>> properties it has are that it's identical to other instances of the >>> same symbol and different from all other symbols. That's totally different >>> from a neural net that uses embedding vectors. Embedding vectors have a >>> rich internal structure that dictates how they interact with other >>> embedding vectors. What I really object to is the following approach: Start >>> with pure symbols and rules for how to manipulate structures made out of >>> pure symbols. These structures themselves can be denoted by symbols that >>> correspond to memory addresses where the bits in the address tell you >>> nothing about the content of the structure at that address. Then when the >>> rule-based approach doesn't work for dealing with the real world (e.g. >>> machine translation) try to use neural nets to convert the real world into >>> pure symbols and then carry on with the rule-based approach. That is like >>> using an electric motor to inject the gasoline into the same old gasoline >>> engine instead of just replacing the gasoline engine with an electric motor. >>> >>> >>> >>> >>> >>> On Fri, Feb 4, 2022 at 2:32 AM Dietterich, Thomas >>> wrote: >>> >>> ?Understanding? is not a Boolean. It is a theorem that no system can >>> enumerate all of the consequences of a state of affairs in the world. >>> >>> >>> >>> For low-stakes application work, we can be satisfied by a system that >>> ?does the right thing?. If the system draws a good picture, that?s >>> sufficient. It ?understood? the request. >>> >>> >>> >>> But for higher-stakes applications---and for advancing the science---we >>> seek a causal account of how the components of a system cause it to do the >>> right thing. We are hoping that a small set of mechanisms can produce broad >>> coverage of intelligent behavior. This gives us confidence that the system >>> will respond correctly outside of the narrow tasks on which we have tested >>> it. >>> >>> >>> >>> --Tom >>> >>> >>> >>> Thomas G. Dietterich, Distinguished Professor Emeritus >>> >>> School of Electrical Engineering and Computer >>> Science >>> >>> US Mail: 1148 Kelley Engineering Center >>> >>> >>> >>> Office: 2067 Kelley Engineering Center >>> >>> Oregon State Univ., Corvallis, OR 97331-5501 >>> >>> Voice: 541-737-5559; FAX: 541-737-1300 >>> >>> URL: http://web.engr.oregonstate.edu/~tgd/ >>> >>> >>> >>> >>> *From:* Connectionists *On >>> Behalf Of *Gary Marcus >>> *Sent:* Thursday, February 3, 2022 8:26 AM >>> *To:* Danko Nikolic >>> *Cc:* connectionists at mailman.srv.cs.cmu.edu; AIhub >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> [This email originated from outside of OSU. Use caution with links and >>> attachments.] >>> >>> Dear Danko, >>> >>> >>> >>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >>> talk, in which he said (paraphrasing from memory, because I don?t remember >>> the precise words) that the famous 200 Quoc Le unsupervised model [ >>> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >>> ] >>> had learned the concept of a ca. In reality the model had clustered >>> together some catlike images based on the image statistics that it had >>> extracted, but it was a long way from a full, counterfactual-supporting >>> concept of a cat, much as you describe below. >>> >>> >>> >>> I fully agree with you that the reason for even having a semantics is as >>> you put it, "to 1) learn with a few examples and 2) apply the knowledge to >>> a broad set of situations.? GPT-3 sometimes gives the appearance of having >>> done so, but it falls apart under close inspection, so the problem remains >>> unsolved. >>> >>> >>> >>> Gary >>> >>> >>> >>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >>> wrote: >>> >>> >>> >>> G. Hinton wrote: "I believe that any reasonable person would admit that >>> if you ask a neural net to draw a picture of a hamster wearing a red hat >>> and it draws such a picture, it understood the request." >>> >>> >>> >>> I would like to suggest why drawing a hamster with a red hat does not >>> necessarily imply understanding of the statement "hamster wearing a red >>> hat". >>> >>> To understand that "hamster wearing a red hat" would mean inferring, in >>> newly emerging situations of this hamster, all the real-life >>> implications that the red hat brings to the little animal. >>> >>> >>> >>> What would happen to the hat if the hamster rolls on its back? (Would >>> the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? >>> (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red >>> hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would >>> it be easier for predators to spot the hamster?) >>> >>> >>> >>> ...and so on. >>> >>> >>> >>> Countless many questions can be asked. One has understood "hamster >>> wearing a red hat" only if one can answer reasonably well many of such >>> real-life relevant questions. Similarly, a student has understood materias >>> in a class only if they can apply the materials in real-life situations >>> (e.g., applying Pythagora's theorem). If a student gives a correct answer >>> to a multiple choice question, we don't know whether the student understood >>> the material or whether this was just rote learning (often, it is rote >>> learning). >>> >>> >>> >>> I also suggest that understanding also comes together with effective >>> learning: We store new information in such a way that we can recall it >>> later and use it effectively i.e., make good inferences in newly emerging >>> situations based on this knowledge. >>> >>> >>> >>> In short: Understanding makes us humans able to 1) learn with a few >>> examples and 2) apply the knowledge to a broad set of situations. >>> >>> >>> >>> No neural network today has such capabilities and we don't know how to >>> give them such capabilities. Neural networks need large amounts of >>> training examples that cover a large variety of situations and then >>> the networks can only deal with what the training examples have already >>> covered. Neural networks cannot extrapolate in that 'understanding' sense. >>> >>> >>> >>> I suggest that understanding truly extrapolates from a piece of >>> knowledge. It is not about satisfying a task such as translation between >>> languages or drawing hamsters with hats. It is how you got the capability >>> to complete the task: Did you only have a few examples that covered >>> something different but related and then you extrapolated from that >>> knowledge? If yes, this is going in the direction of understanding. Have >>> you seen countless examples and then interpolated among them? Then perhaps >>> it is not understanding. >>> >>> >>> >>> So, for the case of drawing a hamster wearing a red hat, understanding >>> perhaps would have taken place if the following happened before that: >>> >>> >>> >>> 1) first, the network learned about hamsters (not many examples) >>> >>> 2) after that the network learned about red hats (outside the context of >>> hamsters and without many examples) >>> >>> 3) finally the network learned about drawing (outside of the context of >>> hats and hamsters, not many examples) >>> >>> >>> >>> After that, the network is asked to draw a hamster with a red hat. If it >>> does it successfully, maybe we have started cracking the problem of >>> understanding. >>> >>> >>> >>> Note also that this requires the network to learn sequentially without >>> exhibiting catastrophic forgetting of the previous knowledge, which is >>> possibly also a consequence of human learning by understanding. >>> >>> >>> >>> >>> >>> Danko >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> >>> --- A progress usually starts with an insight --- >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> >>> >>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>> >>> >>> Without getting into the specific dispute between Gary and Geoff, I >>> think with approaches similar to GLOM, we are finally headed in the right >>> direction. There?s plenty of neurophysiological evidence for single-cell >>> abstractions and multisensory neurons in the brain, which one might claim >>> correspond to symbols. And I think we can finally reconcile the decades old >>> dispute between Symbolic AI and Connectionism. >>> >>> >>> >>> >>> >>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >>> an effort to wind up with encodings that effectively serve as symbols in >>> exactly that way, guaranteed to serve as consistent representations of >>> specific concepts.) >>> >>> >>> GARY: I have *never* called for dismissal of neural networks, but >>> rather for some hybrid between the two (as you yourself contemplated in >>> 1991); the point of the 2001 book was to characterize exactly where >>> multilayer perceptrons succeeded and broke down, and where symbols could >>> complement them. >>> >>> >>> >>> >>> >>> Asim Roy >>> >>> >>> Professor, Information Systems >>> >>> >>> Arizona State University >>> >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> *From: Connectionists On >>> Behalf Of Gary Marcus Sent: Wednesday, February 2, 2022 1:26 PM To: >>> Geoffrey Hinton Cc: AIhub ; >>> connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen >>> Hanson in conversation with Geoff Hinton >>> * >>> >>> >>> >>> >>> Dear Geoff, and interested others, >>> >>> >>> >>> >>> >>> What, for example, would you make of a system that often drew the >>> red-hatted hamster you requested, and perhaps a fifth of the time gave you >>> utter nonsense? Or say one that you trained to create birds but sometimes >>> output stuff like this: >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> One could >>> >>> >>> >>> >>> >>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>> >>> >>> or >>> >>> >>> b. wonder if it might be possible that sometimes the system gets the >>> right answer for the wrong reasons (eg partial historical contingency), and >>> wonder whether another approach might be indicated. >>> >>> >>> >>> >>> >>> Benchmarks are harder than they look; most of the field has come to >>> recognize that. The Turing Test has turned out to be a lousy measure of >>> intelligence, easily gamed. It has turned out empirically that the Winograd >>> Schema Challenge did not measure common sense as well as Hector might have >>> thought. (As it happens, I am a minor coauthor of a very recent review on >>> this very topic: https://arxiv.org/abs/2201.02387) But its conquest in >>> no way means machines now have common sense; many people from many >>> different perspectives recognize that (including, e.g., Yann LeCun, who >>> generally tends to be more aligned with you than with me). >>> >>> >>> >>> >>> >>> So: on the goalpost of the Winograd schema, I was wrong, and you can >>> quote me; but what you said about me and machine translation remains your >>> invention, and it is inexcusable that you simply ignored my 2019 >>> clarification. On the essential goal of trying to reach meaning and >>> understanding, I remain unmoved; the problem remains unsolved. >>> >>> >>> >>> >>> >>> All of the problems LLMs have with coherence, reliability, truthfulness, >>> misinformation, etc stand witness to that fact. (Their persistent inability >>> to filter out toxic and insulting remarks stems from the same.) I am hardly >>> the only person in the field to see that progress on any given benchmark >>> does not inherently mean that the deep underlying problems have solved. >>> You, yourself, in fact, have occasionally made that point. >>> >>> >>> >>> >>> >>> With respect to embeddings: Embeddings are very good for natural >>> language *processing*; but NLP is not the same as NL*U* ? when it comes >>> to *understanding*, their worth is still an open question. Perhaps they >>> will turn out to be necessary; they clearly aren?t sufficient. In their >>> extreme, they might even collapse into being symbols, in the sense of >>> uniquely identifiable encodings, akin to the ASCII code, in which a >>> specific set of numbers stands for a specific word or concept. (Wouldn?t >>> that be ironic?) >>> >>> >>> >>> >>> >>> (Your GLOM, which as you know I praised publicly, is in many ways an >>> effort to wind up with encodings that effectively serve as symbols in >>> exactly that way, guaranteed to serve as consistent representations of >>> specific concepts.) >>> >>> >>> >>> >>> >>> Notably absent from your email is any kind of apology for >>> misrepresenting my position. It?s fine to say that ?many people thirty >>> years ago once thought X? and another to say ?Gary Marcus said X in 2015?, >>> when I didn?t. I have consistently felt throughout our interactions that >>> you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) >>> apologized to me for having made that error. I am still not he. >>> >>> >>> >>> >>> >>> Which maybe connects to the last point; if you read my work, you would >>> see thirty years of arguments *for* neural networks, just not in the >>> way that you want them to exist. I have ALWAYS argued that there is a role >>> for them; characterizing me as a person ?strongly opposed to neural >>> networks? misses the whole point of my 2001 book, which was subtitled >>> ?Integrating Connectionism and Cognitive Science.? >>> >>> >>> >>> >>> >>> In the last two decades or so you have insisted (for reasons you have >>> never fully clarified, so far as I know) on abandoning symbol-manipulation, >>> but the reverse is not the case: I have *never* called for dismissal of >>> neural networks, but rather for some hybrid between the two (as you >>> yourself contemplated in 1991); the point of the 2001 book was to >>> characterize exactly where multilayer perceptrons succeeded and broke down, >>> and where symbols could complement them. It?s a rhetorical trick (which is >>> what the previous thread was about) to pretend otherwise. >>> >>> >>> >>> >>> >>> Gary >>> >>> >>> >>> >>> >>> >>> >>> >>> On Feb 2, 2022, at 11:22, Geoffrey Hinton >>> wrote: >>> >>> >>> ? >>> >>> >>> Embeddings are just vectors of soft feature detectors and they are very >>> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >>> opposite. >>> >>> >>> >>> >>> >>> A few decades ago, everyone I knew then would have agreed that the >>> ability to translate a sentence into many different languages was strong >>> evidence that you understood it. >>> >>> >>> >>> >>> >>> But once neural networks could do that, their critics moved the >>> goalposts. An exception is Hector Levesque who defined the goalposts more >>> sharply by saying that the ability to get pronoun references correct in >>> Winograd sentences is a crucial test. Neural nets are improving at that but >>> still have some way to go. Will Gary agree that when they can get pronoun >>> references correct in Winograd sentences they really do understand? Or does >>> he want to reserve the right to weasel out of that too? >>> >>> >>> >>> >>> >>> Some people, like Gary, appear to be strongly opposed to neural networks >>> because they do not fit their preconceived notions of how the mind should >>> work. >>> >>> >>> I believe that any reasonable person would admit that if you ask a >>> neural net to draw a picture of a hamster wearing a red hat and it draws >>> such a picture, it understood the request. >>> >>> >>> >>> >>> >>> Geoff >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>> >>> >>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger >>> neural network community, >>> >>> >>> >>> >>> >>> There has been a lot of recent discussion on this list about framing and >>> scientific integrity. Often the first step in restructuring narratives is >>> to bully and dehumanize critics. The second is to misrepresent their >>> position. People in positions of power are sometimes tempted to do this. >>> >>> >>> >>> >>> >>> The Hinton-Hanson interview that you just published is a real-time >>> example of just that. It opens with a needless and largely content-free >>> personal attack on a single scholar (me), with the explicit intention of >>> discrediting that person. Worse, the only substantive thing it says is >>> false. >>> >>> >>> >>> >>> >>> Hinton says ?In 2015 he [Marcus] made a prediction that computers >>> wouldn?t be able to do machine translation.? >>> >>> >>> >>> >>> >>> I never said any such thing. >>> >>> >>> >>> >>> >>> What I predicted, rather, was that multilayer perceptrons, as they >>> existed then, would not (on their own, absent other mechanisms) >>> *understand* language. Seven years later, they still haven?t, except in >>> the most superficial way. >>> >>> >>> >>> >>> >>> I made no comment whatsoever about machine translation, which I view as >>> a separate problem, solvable to a certain degree by correspondance without >>> semantics. >>> >>> >>> >>> >>> >>> I specifically tried to clarify Hinton?s confusion in 2019, but, >>> disappointingly, he has continued to purvey misinformation despite that >>> clarification. Here is what I wrote privately to him then, which should >>> have put the matter to rest: >>> >>> >>> >>> >>> >>> You have taken a single out of context quote [from 2015] and >>> misrepresented it. The quote, which you have prominently displayed at the >>> bottom on your own web page, says: >>> >>> >>> >>> >>> >>> Hierarchies of features are less suited to challenges such as language, >>> inference, and high-level planning. For example, as Noam Chomsky famously >>> pointed out, language is filled with sentences you haven't seen >>> before. Pure classifier systems don't know what to do with such sentences. >>> The talent of feature detectors -- in identifying which member of some >>> category something belongs to -- doesn't translate into understanding >>> novel sentences, in which each sentence has its own unique meaning. >>> >>> >>> >>> >>> >>> It does *not* say "neural nets would not be able to deal with novel >>> sentences"; it says that hierachies of features detectors (on their own, if >>> you read the context of the essay) would have trouble *understanding *novel sentences. >>> >>> >>> >>> >>> >>> >>> Google Translate does yet not *understand* the content of the sentences >>> is translates. It cannot reliably answer questions about who did what to >>> whom, or why, it cannot infer the order of the events in paragraphs, it >>> can't determine the internal consistency of those events, and so forth. >>> >>> >>> >>> >>> >>> Since then, a number of scholars, such as the the computational linguist >>> Emily Bender, have made similar points, and indeed current LLM difficulties >>> with misinformation, incoherence and fabrication all follow from these >>> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >>> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, >>> also emphasizing issues of understanding and meaning: >>> >>> >>> >>> >>> >>> *The success of the large neural language models on many NLP tasks is >>> exciting. However, we find that these successes sometimes lead to hype in >>> which these models are being described as ?understanding? language or >>> capturing ?meaning?. In this position paper, we argue that a system trained >>> only on form has a priori no way to learn meaning. .. a clear understanding >>> of the distinction between form and meaning will help guide the field >>> towards better science around natural language understanding. >>> * >>> >>> >>> >>> >>> Her later article with Gebru on language models ?stochastic parrots? is >>> in some ways an extension of this point; machine translation requires >>> mimicry, true understanding (which is what I was discussing in 2015) >>> requires something deeper than that. >>> >>> >>> >>> >>> >>> Hinton?s intellectual error here is in equating machine translation with >>> the deeper comprehension that robust natural language understanding will >>> require; as Bender and Koller observed, the two appear not to be the same. >>> (There is a longer discussion of the relation between language >>> understanding and machine translation, and why the latter has turned out to >>> be more approachable than the former, in my 2019 book with Ernest Davis). >>> >>> >>> >>> >>> >>> More broadly, Hinton?s ongoing dismissiveness of research from >>> perspectives other than his own (e.g. linguistics) have done the field a >>> disservice. >>> >>> >>> >>> >>> >>> As Herb Simon once observed, science does not have to be zero-sum. >>> >>> >>> >>> >>> >>> Sincerely, >>> >>> >>> Gary Marcus >>> >>> >>> Professor Emeritus >>> >>> >>> New York University >>> >>> >>> >>> >>> >>> On Feb 2, 2022, at 06:12, AIhub wrote: >>> >>> >>> ? >>> >>> >>> Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> >>> >>> In the latest episode of this video series for AIhub.org, Stephen >>> Hanson talks to Geoff Hinton about neural networks, backpropagation, >>> overparameterization, digit recognition, voxel cells, syntax and semantics, >>> Winograd sentences, and more. >>> >>> >>> >>> >>> >>> You can watch the discussion, and read the transcript, here: >>> >>> >>> >>> >>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>> >>> >>> >>> >>> >>> About AIhub: >>> >>> >>> AIhub is a non-profit dedicated to connecting the AI community to the >>> public by providing free, high-quality information through AIhub.org ( >>> https://aihub.org/). We help researchers publish the latest AI news, >>> summaries of their work, opinion pieces, tutorials and more. We are >>> supported by many leading scientific organizations in AI, namely AAAI, >>> NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. >>> >>> >>> Twitter: @aihuborg >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> >>> >>> >>> >>> >>> -- >>> >>> Juyang (John) Weng >>> >>> >>> >>> >>> -- >>> >>> Juyang (John) Weng >>> >> >> >> -- >> Juyang (John) Weng >> >> > > -- > Juyang (John) Weng > > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 259567 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 9661 bytes Desc: not available URL: From juyang.weng at gmail.com Mon Feb 7 19:39:17 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Mon, 7 Feb 2022 19:39:17 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear Gary, Thank you for the link to Geoff's GLOM paper. I quickly browsed it just now. Some fundamental comments, not all, to be concise. (1) You are right, Geoff's GLOM seems to be a symbolic network, which assigns a certain role to each group of neurons. (2) Geoff's part-whole problem is too narrow, a dead end to solving his part-whole problem. I quote: "Perhaps we can learn how to encode the information at each location in such a way that simply averaging the encodings from different locations is the only form of interaction we need." He got stuck with "locations". Like Convolution that I used in Cresceptron 1992 and Geoff also used much longer, as soon a feature representation is centered as a location, the system does not abstract as Michaal Jordan complained at an IJCNN conference. Michael Jordan did not sa=y what he meant by "does not abstract well", but his point is valid. (3) Two feed-forward networks in Geoff's GLOM, one bottom-up and the other top-down, are efficient pattern recognizers that do not abstract. The brain is not just a pattern recognizer. (4) It is very unfortunate that many neural network researchers including Alpha's DeepMinds have not dug deep into what a cell can do and what a cell cannot. Geoff's GLOM is an example. I have a paper about a brain model and I sent it to some people to pre-review. But like my Conscious Learning paper that was rejected by ICDL 2021 and AAAI 2022, this brain model would be rejected too. Your humbly, -John On Mon, Feb 7, 2022 at 10:57 AM Gary Marcus wrote: > Dear John, > > I agree with you that cluttered scenes are critical, but Geoff?s GLOM > paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf] might > actually have some relevance. It may well be that we need to do a better > job with parts and whole before we can fully address clutter, and Geoff is > certainly taking that question seriously. > > Geoff?s ?Stable islands of identical vectors? do sound suspiciously like > symbols to me (in a good way!), but regardless, they seem to me to be a > plausible candidate as a foundation for coping with clutter. > > And not just cluttered scenes, but also *relations between multiple > objects in a scene*, which is another example of the broader issue you > raise, challenging for pure MLPs but critical for deeper AI. > > Gary > > On Feb 7, 2022, at 00:23, Juyang Weng wrote: > > ? > Dear Geoff Hinton, > I respect that you have been working on pattern recognition on isolated > characters using neural networks. > > However, I am deeply disappointed that after receiving the Turing Award > 2018, you are still falling behind your own award work by talking about "how > you > recognize that a handwritten 2 is a 2." You have fallen behind our > group's > Creceptron work in 1992, let alone our group's work on 3D-to-2D-to-3D > Conscious Learning using DNs. Both deal with cluttered scenes. > > Specifically, you will never be able to get a correct causal explanation > by looking at a single hand-written 2. Your problem is too small to > explain a brain network. You must look at cluttered sciences, with many > objects. > > Yours humbly, > -John > ---- > Message: 7 > Date: Fri, 4 Feb 2022 15:24:02 -0500 > From: Geoffrey Hinton > To: "Dietterich, Thomas" > Cc: AIhub , > "connectionists at mailman.srv.cs.cmu.edu" > > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > Message-ID: > xCqaYmzu+pq5rnoh+p2mQ at mail.gmail.com> > Content-Type: text/plain; charset="utf-8" > > I agree that it's nice to have a causal explanations. But I am not > convinced there will ever be a simple causal explanation for how you > recognize that a handwritten 2 is a 2. > > -- > Juyang (John) Weng > > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From bhammer at techfak.uni-bielefeld.de Tue Feb 8 03:26:44 2022 From: bhammer at techfak.uni-bielefeld.de (Barbara Hammer) Date: Tue, 8 Feb 2022 09:26:44 +0100 Subject: Connectionists: link correction - talk by Max Welling link is available at JAII.EU Message-ID: <4a00599e-53cb-b5a0-e452-aadb3ab0df1f@techfak.uni-bielefeld.de> Dear colleagues, a small correction of the links (sorry for this inconvenience): Max Welling will talk about "Solving PDEs with GNNs, and how to use symmetries" at Thursday 10.2. at 4pm. Login at the link: https://jaii.eu/ or https://uni-bielefeld.zoom.us/meeting/register/tJEpdeiqqTIsHdwgc0lFoscot-NAamrEnmkv Best wishes Barbara Hammer -- Prof. Dr. Barbara Hammer Machine Learning Group, CITEC Bielefeld University D-33594 Bielefeld Phone: +49 521 / 106 12115 From marcella.cornia at unimore.it Tue Feb 8 04:29:56 2022 From: marcella.cornia at unimore.it (Marcella Cornia) Date: Tue, 8 Feb 2022 10:29:56 +0100 Subject: Connectionists: [CFP] Extedend Deadline "Attentive Models in Vision" (February 28, 2022) - Frontiers in Computer Vision Message-ID: ******************************** Research Topic ?Attentive Models in Vision? Computer Vision Section | Frontiers in Computer Science https://www.frontiersin.org/research-topics/23980/attentive-models-in-vision ******************************** === DEADLINE EXTENDED! ==== Paper Submission Deadline: February 28, 2022 Apologies for multiple posting Please distribute this call to interested parties TOPIC EDITORS ============== - Marcella Cornia, University of Modena and Reggio Emilia (Italy) - Luowei Zhou, Microsoft (United States) - Ramprasaath R. Selvaraju, Saleforce Research (United States) - Prof. Xavier Gir?-i-Nieto, Universitat Politecnica de Catalunya (Spain) - Prof. Jason Corso, Stevens Institute of Technology (United States) AIMS AND SCOPE =============== The modeling and replication of visual attention mechanisms have been extensively studied for more than 80 years by neuroscientists and more recently by computer vision researchers, contributing to the formation of various subproblems in the field. Among them, saliency estimation and human-eye fixation prediction have demonstrated their importance in improving many vision-based inference mechanisms: image segmentation and annotation, image and video captioning, and autonomous driving are some examples. Nowadays, with the surge of attentive and Transformer-based models, the modeling of attention has grown significantly and is a pillar of cutting-edge research in computer vision, multimedia, and natural language processing. In this context, current research efforts are also focused on new architectures which are candidates to replace the convolutional operator, as testified by recent works that perform image classification using attention-based architectures or that combine vision with other modalities, such as language, audio, and speech, by leveraging on fully-attentive solutions. Given the fundamental role of attention in the field of computer vision, the goal of this Research Topic is to contribute to the growth and development of attention-based solutions focusing on both traditional approaches and fully-attentive models. Moreover, the study of human attention has inspired models that leverage human gaze data to supervise machine attention. This Research Topic aims to present innovative research that relates to the study of human attention and to the usage of attention mechanisms in the development of deep learning architectures and enhancing model explainability. Research papers employing traditional attentive operations or employing novel Transformer-based architectures are encouraged, as well as works that apply attentive models to integrate vision and other modalities (e.g., language, audio, speech, etc.). We also welcome submissions on novel algorithms, datasets, literature reviews, and other innovations related to the scope of this Research Topic. TOPICS ======= The topics of interest include but are not limited to: - Saliency prediction and salient object detection - Applications of human attention in Vision - Visualization of attentive maps for Explainability of Deep Networks - Use of Explainable-AI techniques to improve any aspect of the network (generalization, robustness, and fairness) - Applications of attentive operators in the design of Deep Networks - Transformer-based or attention-based models for Computer Vision tasks (e.g. classification, detection, segmentation) - Transformer-based or attention-based models to combine Vision with other modalities (e.g. language, audio, speech) - Transformer-based or attention-based models for Vision-and-Language tasks (e.g., image and video captioning, visual question answering, cross-modal retrieval, textual grounding / referring expression localization, vision-and-language navigation) - Computational issues in attentive models - Applications of attentive models (e.g., robotics and embodied AI, medical imaging, document analysis, cultural heritage) IMPORTANT DATES ================= - Paper Submission Extended Deadline: February 28, 2022 Research topic page: https://www.frontiersin.org/research-topics/23980/attentive-models-in-vision Click here to participate: https://www.frontiersin.org/research-topics/23980/attentive-models-in-vision/participate-in-open-access-research-topic By expressing your interest in contributing to this collection, you will be registered as a contributing author and will receive regular updates regarding this Research Topic. SUBMISSION GUIDELINES ====================== All submitted articles are peer reviewed. All published articles are subject to article processing charges (APCs). Frontiers works with leading institutions to ensure researchers are supported when publishing open access. See if your institution has a payment plan with Frontiers or apply to the Frontiers Fee Support program. If you wish to know more about Frontiers publishing and contribution process, please head to the following sections: - Collaborative peer review - Author guidelines - Open Access, publishing fees, and waivers -- *Marcella Cornia*, PhD Tenure-Track Assistant Professor (RTD-B) Dipartimento di Educazione e Scienze Umane Universit? degli Studi di Modena e Reggio Emilia e-mail: marcella.cornia at unimore.it phone: +39 059 2058790 -------------- next part -------------- An HTML attachment was scrubbed... URL: From iswc.conf at gmail.com Tue Feb 8 08:50:20 2022 From: iswc.conf at gmail.com (International Semantic Web Conference) Date: Tue, 8 Feb 2022 08:50:20 -0500 Subject: Connectionists: CfP ISWC 2022 - Call for Semantic Web Challenge Proposals Message-ID: *CfP: 21st International Semantic Web Conference (ISWC 2022)* Hangzhou, China, October 23-27, 2022 https://iswc2022.semanticweb.org/ The International Semantic Web Conference (ISWC) is the premier venue for presenting fundamental research, innovative technology, and applications concerning semantics, data, and the Web. It is the most important international venue to discuss and present the latest advances and applications of the semantic Web, knowledge graphs, linked data, ontologies and artificial intelligence (AI) on the Web. A great way to advance the state of the art in a given domain is to create competition. We invite you to propose an ISWC 2022 Challenge, in which you define an open competition on a problem of your choice within the Semantic Web domain. Call for Semantic Web Challenge Proposals: https://iswc2022.semanticweb.org/index.php/semantic-web-challenge-proposals/ Deadline: Friday, 24th March, 2022, 23:59 AoE (Anywhere on Earth) Semantic Web Challenge Chairs: - Catia Pesquita, LASIGE, Faculdade de Ci?ncias, University of Lisbon clpesquita at fc.ul.pt - Daniele Dell?Aglio, Aalborg University, Denmark dade at cs.aau.dk Follow us on social media: - Twitter: @iswc_conf #iswc_conf (https://twitter.com/iswc_conf) - LinkedIn: https://www.linkedin.com/groups/13612370 - Facebook: https://www.facebook.com/ISWConf - Instagram: https://www.instagram.com/iswc_conf/ The ISWC 2022 Organizing Team Organizing Committee ? ISWC 2022 (semanticweb.org) -------------- next part -------------- An HTML attachment was scrubbed... URL: From daniel.polani at gmail.com Tue Feb 8 08:00:10 2022 From: daniel.polani at gmail.com (Daniel Polani) Date: Tue, 8 Feb 2022 13:00:10 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <148D10E4-0C1F-4FA4-B8AA-C80E1472A63A@nyu.edu> Message-ID: Hi Ali, and all, all excellent points. However, I think there is something more insidious and implicit about symbols that shows that they are not quite as "pure" as we would like them to be. We think of symbols as almost irreducible entities. But even when we embark on this most stringent symbol management game called "mathematics", it turns out that one can dig deeper and deeper to try to anchor them into some a prioristic axioms; but still these axioms are actually extracted through our expertise about the real world and are our crystallized expertise about (say, in the case of Euclidean geometry) how the real world workings can be simplified. And we know well that this game does not always a foregone conclusions (as non-Euclidean geometry ended up demonstrating) Even when we push further, at some point, when it seems we have reached the ultimate grounds, we get something like Goedel's theorems; where,while we think we intuitively understand what it means for something to be "true", it turns out it is not something that is ultimately grounded in the ground rules of the logical system that we impose on symbols. It is clear that we still have some "embodied" intuition to which the symbols, even in their purest form, do not give us proper access. In short, there is a clear tension/dichotomy between "pure" and "embodied" symbols. To vary Einstein's dictum: as far as symbols are "pure", they do not refer to reality; and as far as they refer to reality, the are not "pure". In practice, thus, I believe symbols in the brain of a living being always carry the "primeval sludge" with them. I think this is what makes symbols so tricky to handle in an artificial AI (and I am emphatically not restricting this argument to NNs). I personally do not think that we can do that satisfactorily without somehow getting the embodiment into the matter; but that's just my personal guess. - Daniel On Tue, Feb 8, 2022 at 7:25 AM Ali Minai wrote: > Hi Gary > > Thanks for your reply. I'll think more about your points. I do think that, > to understand the human mind, we should start with vertebrates, which is > why I suggested fish. At least for the motor system - which is part of the > mind - we have learned a lot from lampreys (e.g. Sten Grillner's work and > that beautiful lamprey-salamander model by Ijspeert et al.), and it has > taught us a lot about locomotion in other animals, including mammals. The > principles clearly generalize, though the complexity increases a lot. > Insects too are very interesting. After all, they are our ancestors too. > > I don't agree that we can posit a clear transition from deep cognitive > models in humans to none below that in the phylogenetic tree. Chimpanzees > and macaques certainly show some evidence, and there's no reason to think > that it's a step change rather than a highly nonlinear continuum. And even > though what we might (simplistically) call System 2 aspects of cognition > are minimally present in other mammals, their precursors must be. > > My point about cats and symbols was not regarding whether cats are aware > of symbols, but that symbols emerge naturally from the physics of their > brains. Behaviors that require some small degree of symbolic processing > exist in mammals other than humans (e.g., transitive inference and > landmark-based navigation in rats), and it is seen better as an emergent > property of brains than an attribute to be explicitly built-into neural > models by us. Once we have a sufficiently brain-like neural model, symbolic > processing will already be there. > > I agree with you completely that we are far from understanding some of the > most fundamental principles of the brain, but even more importantly, we are > not even looking in the right direction. I'm hoping to lay out my arguments > about all this in more detail in some other form. > > Best > Ali > > > PS: I had inadvertently posted my reply of Gary's message only to him. > Should have posted to everyone, so here it is. > > > *Ali A. Minai, Ph.D.* > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > > On Mon, Feb 7, 2022 at 12:28 AM Gary Marcus wrote: > >> Ali, >> >> >> It?s useful to think about animals, but I really wouldn?t start with >> fish; it?s not clear that their ecological niche demands anything >> significant in the way of extrapolation, causal reasoning, or >> compositionality. There is good evidence elsewhere in the animal world for >> extrapolation of functions that may be innate (eg solar azimuth in bees), >> and causal reasoning (eg tool use in ravens, various primates, and >> octopi). It?s still not clear to me how much hierarchical representation >> (critical to AGI) exists outside of humans, though; the ability to >> construct rich new cognitive models may also be unique to us. >> >> >> In any case it matters not in the least whether the average cat or human >> *cares* about symbols, anymore that it matters whether the average >> animal understands digestion; only a tiny fraction of the creatures on this >> planet have any real understanding of their internal workings. >> >> >> My overall feeling is that we are a really, really long way from >> understanding the neural basis of higher-level cognition, and that AI is >> going to need muddle through on its own, for another decade or two, >> >> >> I do fully agree with your conclusion, though, that "AI today is driven >> more by habit and the incentives of the academic and corporate marketplaces >> than by a deep, long-term view of AI as a great exploratory project in >> fundamental science." Let's hope that changes. >> >> >> Gary >> >> On Feb 6, 2022, at 13:19, Ali Minai wrote: >> >> ? >> >> Gary, >> >> That?s a very interesting and accurate list of capabilities that a >> general intelligent system must have and that our AI does not. Of course, >> the list is familiar to me from having read your book. However, I have a >> somewhat different take on this whole thing. >> >> >> >> All the things we discuss here ? symbols/no symbols, parts/wholes, >> supervised/unsupervised, token/type, etc., are useful categories and >> distinctions for our analysis of the problem, and are partly a result of >> the historical evolution of the field of AI in particular and of philosophy >> in general. The categories are not wrong in any way, of course, but they >> are posterior to the actual system ? good for describing and analyzing it, >> and for validating our versions of it (which is how you use them). I think >> they are less useful as prescriptions for how to build our AI systems. If >> intelligent systems did not already exist and we were building them from >> scratch (please ignore the impossibility of that), having a list of ?must >> haves? would be great. But intelligent systems already exist ? from humans >> to fish ? and they already have these capacities to a greater or lesser >> degree because of the physics of their biology. A cat?s intelligence does >> not care whether it has symbols or not, and nor does mine or yours. >> Whatever we describe as symbolic processing post-facto has already been >> done by brains for at least tens of millions of years. Instead of getting >> caught up in ?how to add symbols into our neural models?, we should be >> investigating how what we see as symbolic processing emerges from animal >> brains, and then replicate those brains to the degree necessary. If we can >> do that, symbolic processing will already be present. But it cannot be done >> piece by piece. It must take the integrity of the whole brain and the body >> it is part of, and its environment, into account. That?s why I think that a >> much better ? though a very long ? route to AI is to start by understanding >> how a fish brain makes the intelligence of a fish possible, and then boot >> up our knowledge across phylogenetic stages: Bottom up reverse engineering >> rather than top-down engineering. That?s the way Nature built up to human >> intelligence, and we will succeed only by reverse engineering it. Of >> course, we can do it much faster and with shortcuts because we are >> intelligent, purposive agents, but working top-down by building piecewise >> systems that satisfy a list of attributes will not get us there. Among >> other things, those pieces will be impossible to integrate into the kind of >> intelligence that can have those general models of the world that you >> rightly point to as being necessary. >> >> >> >> I think that one thing that has been a great boon to the AI enterprise >> has also been one of the greatest impediments to its complete success, and >> that is the ?computationalization? of intelligence. On the one hand, >> thinking of intelligence computationally allows us to describe it >> abstractly and in a principled, formal way. It also resonates with the fact >> that we are trying to implement intelligence through computational >> machines. But, on the flip side, this view of intelligence divorces it from >> its physics ? from the fact that real intelligence in animals emerges from >> the physics of the physical system. That system is not a collection of its >> capabilities; rather, those capabilities are immanent in it by virtue of >> its physics. When we try to build those capabilities computationally, i.e., >> through code, we are making the same error that the practitioners of >> old-style ?symbolic AI? made ? what I call the ?professors are smarter than >> Nature? error, i.e., the idea that we are going to enumerate (or describe) >> all the things that underlie intelligence and implement them one by one >> until we get complete intelligence. We will never be able to enumerate all >> those capabilities, and will never be able to get to that complete >> intelligence. The only difference between us and the ?symbolists? of yore >> is that we are replacing giant LISP and Prolog programs with giant neural >> networks. Otherwise, we are using our models exactly as they were trying to >> use their models, and we will fail just as they did unless we get back to >> biology and the real thing. >> >> >> >> I will say again that the way we do AI today is driven more by habit and >> the incentives of the academic and corporate marketplaces than by a deep, >> long-term view of AI as a great exploratory project in fundamental science. >> We are just building AI to drive our cars, translate our documents, write >> our reports, and do our shopping. What that will teach us about actual >> intelligence is just incidental. >> >> >> >> My apologies too for a long response. >> >> Ali >> >> >> *Ali A. Minai, Ph.D.* >> Professor and Graduate Program Director >> Complex Adaptive Systems Lab >> Department of Electrical Engineering & Computer Science >> 828 Rhodes Hall >> University of Cincinnati >> Cincinnati, OH 45221-0030 >> >> Phone: (513) 556-4783 >> Fax: (513) 556-7326 >> Email: Ali.Minai at uc.edu >> minaiaa at gmail.com >> >> WWW: https://eecs.ceas.uc.edu/~aminai/ >> >> >> >> On Sun, Feb 6, 2022 at 9:42 AM Gary Marcus wrote: >> >>> Dear Asim, >>> >>> >>> Sorry for a long answer to your short but rich questions. >>> >>> - Yes, memory in my view has to be part of the answer to the >>> type-token problem. Symbol systems encoded in memory allow a natural way to >>> set up records, and something akin to that seems necessary. Pure multilayer >>> perceptrons struggle with type-token distinctions precisely because they >>> lack such records. On the positive side, I see more and more movement >>> towards recordlike stores (eg w key-value stores in memory networks), and I >>> think that is an important and necessary step, very familiar from the >>> symbol-manipulating playbook, sometimes implemented in new ways. >>> - But ultimately, handling the type-token distinction requires >>> considerable inferential overhead beyond the memory representation of a >>> record per se. How do you determine when to denote something (e.g. >>> Felix) as an instance, and of which kinds (cat, animal etc), and how do you >>> leverage that knowledge once you determine it? >>> - In the limit we reason about types vs tokens in fairly subtle >>> ways, eg in guessing whether a glass that we put down at party is likely to >>> be ours. The reverse is also important: we need to be learn >>> particular traits for individuals and not erroneously generalize them to >>> the class; if my aunt Esther wins the lottery, one shouldn?t infer that >>> all of my aunts or all of my relatives or adult females have won the >>> lottery. so you need both representational machinery that can distinguish >>> eg my cat from cats in general and reasoning machinery to decide at what >>> level certain learned knowledge should inhere. (I had a whole chapter about >>> this sort of thing in The Algebraic Mind if you are interested, and Mike >>> Mozer had a book about types and tokens in neural networks in the mid >>> 1990s). >>> - Yes, part (though not all!) of what we do when we set up cognitive >>> models in our heads is to track particular individuals and their >>> properties. If you only had to correlate kinds (cats) and their properties >>> (have fur) you could maybe get away with a multilayer perceptron, but once >>> you need to track individuals, yes, you really need some kind of >>> memory-based records. >>> - As far as I can tell, Transformers can sometimes approximate some >>> of this for a few sentences, but not over long stretches. >>> >>> >>> As a small terminological aside; for me cognitive models ? cognitive >>> modeling. Cognitive modeling is about building psychological or >>> computational models of how people think, whereas what I mean by a cognitive >>> model is a representation of eg the entities in some situation and the >>> relations between those entities. >>> >>> >>> To your closing question, none of us yet really knows how to build >>> understanding into machines. A solid type-token distinction, both in >>> terms of representation and reasoning, is critical for general >>> intelligence, but hardly sufficient. Personally, I think some minimal >>> prerequisites would be: >>> >>> - representations of space, time, causality, individuals, kinds, >>> persons, places, objects, etc. >>> - representations of abstractions that can hold over all entities in >>> a class >>> - compositionality (if we are talking about human-like understanding) >>> - capacity to construct and update cognitive models on the fly >>> - capacity to reason over entities in those models >>> - ability to learn about new entities and their properties >>> >>> Much of my last book (*Rebooting AI*, w Ernie Davis) is about the above >>> list. The section in the language chapter on a children?s story in which >>> man has lost is wallet is an especially vivid worked example. Later >>> chapters elaborate some of the challenges in representing space, time, and >>> causality. >>> >>> >>> Gary >>> >>> >>> On Feb 5, 2022, at 18:58, Asim Roy wrote: >>> >>> ? >>> >>> Gary, >>> >>> >>> >>> I don?t get much into the type of cognitive modeling you are talking >>> about, but I would guess that the type problem can generally be handled by >>> neural network models and tokens can be resolved with some memory-based >>> system. But to the heart of the question, this is what so-called >>> ?understanding? reduces to computation wise? >>> >>> >>> >>> Asim >>> >>> >>> >>> *From:* Gary Marcus >>> *Sent:* Saturday, February 5, 2022 8:39 AM >>> *To:* Asim Roy >>> *Cc:* Ali Minai ; Danko Nikolic < >>> danko.nikolic at gmail.com>; Brad Wyble ; >>> connectionists at mailman.srv.cs.cmu.edu; AIhub >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> There is no magic in understanding, just computation that has been >>> realized in the wetware of humans and that eventually can be realized in >>> machines. But understanding is not (just) learning. >>> >>> >>> >>> Understanding incorporates (or works in tandem with) learning - but >>> also, critically, in tandem with inference, *and the development and >>> maintenance of cognitive models*. Part of developing an understanding >>> of cats in general is to learn long term-knowledge about their properties, >>> both directly (e.g., through observation) and indirectly (eg through >>> learning facts about animals in general that can be extended to cats), >>> often through inference (if all animals have DNA, and a cat is an animal, >>> it must also have DNA). The understanding of a particular cat also >>> involves direct observation, but also inference (eg one might surmise >>> that the reason that Fluffy is running about the room is that Fluffy >>> suspects there is a mouse stirring somewhere nearby). *But all of that, >>> I would say, is subservient to the construction of cognitive models that >>> can be routinely updated *(e.g., Fluffy is currently in the living >>> room, skittering about, perhaps looking for a mouse). >>> >>> >>> >>> In humans, those dynamic, relational models, which form part of an >>> understanding, can support inference (if Fluffy is in the living room, we >>> can infer that Fluffy is not outside, not lost, etc). Without such models - >>> which I think represent a core part of understanding - AGI is an unlikely >>> prospect. >>> >>> >>> >>> Current neural networks, as it happens, are better at acquiring >>> long-term knowledge (cats have whiskers) than they are at dynamically >>> updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind >>> of dynamic model that I am describing. To a modest degree they can >>> approximate it on the basis of large samples of texts, but their ultimate >>> incoherence stems from the fact that they do not have robust internal >>> cognitive models that they can update on the fly. >>> >>> >>> >>> Without such cognitive models you can still capture some aspects of >>> understanding (eg predicting that cats are likely to be furry), but things >>> fall apart quickly; inference is never reliable, and coherence is fleeting. >>> >>> >>> >>> As a final note, one of the most foundational challenges in constructing >>> adequate cognitive models of the world is to have a clear distinction >>> between individuals and kinds; as I emphasized 20 years ago (in The >>> Algebraic Mind), this has always been a weakness in neural networks, and I >>> don?t think that the type-token problem has yet been solved. >>> >>> >>> >>> Gary >>> >>> >>> >>> >>> >>> On Feb 5, 2022, at 01:31, Asim Roy wrote: >>> >>> ? >>> >>> All, >>> >>> >>> >>> I think the broader question was ?understanding.? Here are two Youtube >>> videos showing simple robots ?learning? to walk. They are purely physical >>> systems. Do they ?understand? anything ? such as the need to go around an >>> obstacle, jumping over an obstacle, walking up and down stairs and so on? >>> By the way, they ?learn? to do these things on their own, literally >>> unsupervised, very much like babies. The basic question is: what is >>> ?understanding? if not ?learning?? Is there some other mechanism (magic) at >>> play in our brain that helps us ?understand?? >>> >>> >>> >>> https://www.youtube.com/watch?v=gn4nRCC9TwQ >>> >>> >>> https://www.youtube.com/watch?v=8sO7VS3q8d0 >>> >>> >>> >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> *From:* Ali Minai >>> *Sent:* Friday, February 4, 2022 11:38 PM >>> *To:* Asim Roy >>> *Cc:* Gary Marcus ; Danko Nikolic < >>> danko.nikolic at gmail.com>; Brad Wyble ; >>> connectionists at mailman.srv.cs.cmu.edu; AIhub >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Asim >>> >>> >>> >>> Of course there's nothing magical about understanding, and the mind has >>> to emerge from the physical system, but our AI models at this point are not >>> even close to realizing how that happens. We are, at best, simulating a >>> superficial approximation of a few parts of the real thing. A single, >>> integrated system where all the aspects of intelligence emerge from the >>> same deep, well-differentiated physical substrate is far beyond our >>> capacity. Paying more attention to neurobiology will be essential to get >>> there, but so will paying attention to development - both physical and >>> cognitive - and evolution. The configuration of priors by evolution is key >>> to understanding how real intelligence learns so quickly and from so >>> little. This is not an argument for using genetic algorithms to design our >>> systems, just for understanding the tricks evolution has used and >>> replicating them by design. Development is more feasible to do >>> computationally, but hardly any models have looked at it except in a >>> superficial sense. Nature creates basic intelligence not so much by >>> configuring functions by explicit training as by tweaking, modulating, >>> ramifying, and combining existing ones in a multi-scale self-organization >>> process. We then learn much more complicated things (like playing chess) by >>> exploiting that substrate, and using explicit instruction or learning by >>> practice. The fundamental lesson of complex systems is that complexity is >>> built in stages - each level exploiting the organization of the level below >>> it. We see it in evolution, development, societal evolution, the evolution >>> of technology, etc. Our approach in AI, in contrast, is to initialize a >>> giant, naive system and train it to do something really complicated - but >>> really specific - by training the hell out of it. Sure, now we do build >>> many systems on top of pre-trained models like GPT-3 and BERT, which is >>> better, but those models were again trained by the same none-to-all process >>> I decried above. Contrast that with how humans acquire language, and how >>> they integrate it into their *entire* perceptual, cognitive, and behavioral >>> repertoire, not focusing just on this or that task. The age of symbolic AI >>> may have passed, but the reductionistic mindset has not. We cannot build >>> minds by chopping it into separate verticals. >>> >>> >>> >>> FTR, I'd say that the emergence of models such as GLOM and Hawkins and >>> Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but >>> they are, I think, looking in the right direction. With a million miles to >>> go! >>> >>> >>> >>> Ali >>> >>> >>> >>> *Ali A. Minai, Ph.D.* >>> Professor and Graduate Program Director >>> Complex Adaptive Systems Lab >>> Department of Electrical Engineering & Computer Science >>> >>> 828 Rhodes Hall >>> >>> University of Cincinnati >>> Cincinnati, OH 45221-0030 >>> >>> >>> Phone: (513) 556-4783 >>> Fax: (513) 556-7326 >>> Email: Ali.Minai at uc.edu >>> minaiaa at gmail.com >>> >>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>> >>> >>> >>> >>> >>> >>> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: >>> >>> First of all, the brain is a physical system. There is no ?magic? inside >>> the brain that does the ?understanding? part. Take for example learning to >>> play tennis. You hit a few balls - some the right way and some wrong ? but >>> you fairly quickly learn to hit them right most of the time. So there is >>> obviously some simulation going on in the brain about hitting the ball in >>> different ways and ?learning? its consequences. What you are calling >>> ?understanding? is really these simulations about different scenarios. It?s >>> also very similar to augmentation used to train image recognition systems >>> where you rotate images, obscure parts and so on, so that you still can say >>> it?s a cat even though you see only the cat?s face or whiskers or a cat >>> flipped on its back. So, if the following questions relate to >>> ?understanding,? you can easily resolve this by simulating such scenarios >>> when ?teaching? the system. There?s nothing ?magical? about >>> ?understanding.? As I said, bear in mind that the brain, after all, is a >>> physical system and ?teaching? and ?understanding? is embodied in that >>> physical system, not outside it. So ?understanding? is just part of >>> ?learning,? nothing more. >>> >>> >>> >>> DANKO: >>> >>> What would happen to the hat if the hamster rolls on its back? (Would >>> the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? >>> (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red >>> hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would >>> it be easier for predators to spot the hamster?) >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> *From:* Gary Marcus >>> *Sent:* Thursday, February 3, 2022 9:26 AM >>> *To:* Danko Nikolic >>> *Cc:* Asim Roy ; Geoffrey Hinton < >>> geoffrey.hinton at gmail.com>; AIhub ; >>> connectionists at mailman.srv.cs.cmu.edu >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Dear Danko, >>> >>> >>> >>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >>> talk, in which he said (paraphrasing from memory, because I don?t remember >>> the precise words) that the famous 200 Quoc Le unsupervised model [ >>> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >>> ] >>> had learned the concept of a ca. In reality the model had clustered >>> together some catlike images based on the image statistics that it had >>> extracted, but it was a long way from a full, counterfactual-supporting >>> concept of a cat, much as you describe below. >>> >>> >>> >>> I fully agree with you that the reason for even having a semantics is as >>> you put it, "to 1) learn with a few examples and 2) apply the knowledge to >>> a broad set of situations.? GPT-3 sometimes gives the appearance of having >>> done so, but it falls apart under close inspection, so the problem remains >>> unsolved. >>> >>> >>> >>> Gary >>> >>> >>> >>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >>> wrote: >>> >>> >>> >>> G. Hinton wrote: "I believe that any reasonable person would admit that >>> if you ask a neural net to draw a picture of a hamster wearing a red hat >>> and it draws such a picture, it understood the request." >>> >>> >>> >>> I would like to suggest why drawing a hamster with a red hat does not >>> necessarily imply understanding of the statement "hamster wearing a red >>> hat". >>> >>> To understand that "hamster wearing a red hat" would mean inferring, in >>> newly emerging situations of this hamster, all the real-life >>> implications that the red hat brings to the little animal. >>> >>> >>> >>> What would happen to the hat if the hamster rolls on its back? (Would >>> the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? >>> (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red >>> hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would >>> it be easier for predators to spot the hamster?) >>> >>> >>> >>> ...and so on. >>> >>> >>> >>> Countless many questions can be asked. One has understood "hamster >>> wearing a red hat" only if one can answer reasonably well many of such >>> real-life relevant questions. Similarly, a student has understood materias >>> in a class only if they can apply the materials in real-life situations >>> (e.g., applying Pythagora's theorem). If a student gives a correct answer >>> to a multiple choice question, we don't know whether the student understood >>> the material or whether this was just rote learning (often, it is rote >>> learning). >>> >>> >>> >>> I also suggest that understanding also comes together with effective >>> learning: We store new information in such a way that we can recall it >>> later and use it effectively i.e., make good inferences in newly emerging >>> situations based on this knowledge. >>> >>> >>> >>> In short: Understanding makes us humans able to 1) learn with a few >>> examples and 2) apply the knowledge to a broad set of situations. >>> >>> >>> >>> No neural network today has such capabilities and we don't know how to >>> give them such capabilities. Neural networks need large amounts of >>> training examples that cover a large variety of situations and then >>> the networks can only deal with what the training examples have already >>> covered. Neural networks cannot extrapolate in that 'understanding' sense. >>> >>> >>> >>> I suggest that understanding truly extrapolates from a piece of >>> knowledge. It is not about satisfying a task such as translation between >>> languages or drawing hamsters with hats. It is how you got the capability >>> to complete the task: Did you only have a few examples that covered >>> something different but related and then you extrapolated from that >>> knowledge? If yes, this is going in the direction of understanding. Have >>> you seen countless examples and then interpolated among them? Then perhaps >>> it is not understanding. >>> >>> >>> >>> So, for the case of drawing a hamster wearing a red hat, understanding >>> perhaps would have taken place if the following happened before that: >>> >>> >>> >>> 1) first, the network learned about hamsters (not many examples) >>> >>> 2) after that the network learned about red hats (outside the context of >>> hamsters and without many examples) >>> >>> 3) finally the network learned about drawing (outside of the context of >>> hats and hamsters, not many examples) >>> >>> >>> >>> After that, the network is asked to draw a hamster with a red hat. If it >>> does it successfully, maybe we have started cracking the problem of >>> understanding. >>> >>> >>> >>> Note also that this requires the network to learn sequentially without >>> exhibiting catastrophic forgetting of the previous knowledge, which is >>> possibly also a consequence of human learning by understanding. >>> >>> >>> >>> >>> >>> Danko >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> >>> --- A progress usually starts with an insight --- >>> >>> >>> >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> >>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>> >>> Without getting into the specific dispute between Gary and Geoff, I >>> think with approaches similar to GLOM, we are finally headed in the right >>> direction. There?s plenty of neurophysiological evidence for single-cell >>> abstractions and multisensory neurons in the brain, which one might claim >>> correspond to symbols. And I think we can finally reconcile the decades old >>> dispute between Symbolic AI and Connectionism. >>> >>> >>> >>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >>> an effort to wind up with encodings that effectively serve as symbols in >>> exactly that way, guaranteed to serve as consistent representations of >>> specific concepts.) >>> >>> GARY: I have *never* called for dismissal of neural networks, but >>> rather for some hybrid between the two (as you yourself contemplated in >>> 1991); the point of the 2001 book was to characterize exactly where >>> multilayer perceptrons succeeded and broke down, and where symbols could >>> complement them. >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> *From:* Connectionists *On >>> Behalf Of *Gary Marcus >>> *Sent:* Wednesday, February 2, 2022 1:26 PM >>> *To:* Geoffrey Hinton >>> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Dear Geoff, and interested others, >>> >>> >>> >>> What, for example, would you make of a system that often drew the >>> red-hatted hamster you requested, and perhaps a fifth of the time gave you >>> utter nonsense? Or say one that you trained to create birds but sometimes >>> output stuff like this: >>> >>> >>> >>> >>> >>> >>> >>> One could >>> >>> >>> >>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>> >>> or >>> >>> b. wonder if it might be possible that sometimes the system gets the >>> right answer for the wrong reasons (eg partial historical contingency), and >>> wonder whether another approach might be indicated. >>> >>> >>> >>> Benchmarks are harder than they look; most of the field has come to >>> recognize that. The Turing Test has turned out to be a lousy measure of >>> intelligence, easily gamed. It has turned out empirically that the Winograd >>> Schema Challenge did not measure common sense as well as Hector might have >>> thought. (As it happens, I am a minor coauthor of a very recent review on >>> this very topic: https://arxiv.org/abs/2201.02387 >>> ) >>> But its conquest in no way means machines now have common sense; many >>> people from many different perspectives recognize that (including, e.g., >>> Yann LeCun, who generally tends to be more aligned with you than with me). >>> >>> >>> >>> So: on the goalpost of the Winograd schema, I was wrong, and you can >>> quote me; but what you said about me and machine translation remains your >>> invention, and it is inexcusable that you simply ignored my 2019 >>> clarification. On the essential goal of trying to reach meaning and >>> understanding, I remain unmoved; the problem remains unsolved. >>> >>> >>> >>> All of the problems LLMs have with coherence, reliability, truthfulness, >>> misinformation, etc stand witness to that fact. (Their persistent inability >>> to filter out toxic and insulting remarks stems from the same.) I am hardly >>> the only person in the field to see that progress on any given benchmark >>> does not inherently mean that the deep underlying problems have solved. >>> You, yourself, in fact, have occasionally made that point. >>> >>> >>> >>> With respect to embeddings: Embeddings are very good for natural >>> language *processing*; but NLP is not the same as NL*U* ? when it comes >>> to *understanding*, their worth is still an open question. Perhaps they >>> will turn out to be necessary; they clearly aren?t sufficient. In their >>> extreme, they might even collapse into being symbols, in the sense of >>> uniquely identifiable encodings, akin to the ASCII code, in which a >>> specific set of numbers stands for a specific word or concept. (Wouldn?t >>> that be ironic?) >>> >>> >>> >>> (Your GLOM, which as you know I praised publicly, is in many ways an >>> effort to wind up with encodings that effectively serve as symbols in >>> exactly that way, guaranteed to serve as consistent representations of >>> specific concepts.) >>> >>> >>> >>> Notably absent from your email is any kind of apology for >>> misrepresenting my position. It?s fine to say that ?many people thirty >>> years ago once thought X? and another to say ?Gary Marcus said X in 2015?, >>> when I didn?t. I have consistently felt throughout our interactions that >>> you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) >>> apologized to me for having made that error. I am still not he. >>> >>> >>> >>> Which maybe connects to the last point; if you read my work, you would >>> see thirty years of arguments *for* neural networks, just not in the >>> way that you want them to exist. I have ALWAYS argued that there is a role >>> for them; characterizing me as a person ?strongly opposed to neural >>> networks? misses the whole point of my 2001 book, which was subtitled >>> ?Integrating Connectionism and Cognitive Science.? >>> >>> >>> >>> In the last two decades or so you have insisted (for reasons you have >>> never fully clarified, so far as I know) on abandoning symbol-manipulation, >>> but the reverse is not the case: I have *never* called for dismissal of >>> neural networks, but rather for some hybrid between the two (as you >>> yourself contemplated in 1991); the point of the 2001 book was to >>> characterize exactly where multilayer perceptrons succeeded and broke down, >>> and where symbols could complement them. It?s a rhetorical trick (which is >>> what the previous thread was about) to pretend otherwise. >>> >>> >>> >>> Gary >>> >>> >>> >>> >>> >>> On Feb 2, 2022, at 11:22, Geoffrey Hinton >>> wrote: >>> >>> ? >>> >>> Embeddings are just vectors of soft feature detectors and they are very >>> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >>> opposite. >>> >>> >>> >>> A few decades ago, everyone I knew then would have agreed that the >>> ability to translate a sentence into many different languages was strong >>> evidence that you understood it. >>> >>> >>> >>> But once neural networks could do that, their critics moved the >>> goalposts. An exception is Hector Levesque who defined the goalposts more >>> sharply by saying that the ability to get pronoun references correct in >>> Winograd sentences is a crucial test. Neural nets are improving at that but >>> still have some way to go. Will Gary agree that when they can get pronoun >>> references correct in Winograd sentences they really do understand? Or does >>> he want to reserve the right to weasel out of that too? >>> >>> >>> >>> Some people, like Gary, appear to be strongly opposed to neural networks >>> because they do not fit their preconceived notions of how the mind should >>> work. >>> >>> I believe that any reasonable person would admit that if you ask a >>> neural net to draw a picture of a hamster wearing a red hat and it draws >>> such a picture, it understood the request. >>> >>> >>> >>> Geoff >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>> >>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger >>> neural network community, >>> >>> >>> >>> There has been a lot of recent discussion on this list about framing and >>> scientific integrity. Often the first step in restructuring narratives is >>> to bully and dehumanize critics. The second is to misrepresent their >>> position. People in positions of power are sometimes tempted to do this. >>> >>> >>> >>> The Hinton-Hanson interview that you just published is a real-time >>> example of just that. It opens with a needless and largely content-free >>> personal attack on a single scholar (me), with the explicit intention of >>> discrediting that person. Worse, the only substantive thing it says is >>> false. >>> >>> >>> >>> Hinton says ?In 2015 he [Marcus] made a prediction that computers >>> wouldn?t be able to do machine translation.? >>> >>> >>> >>> I never said any such thing. >>> >>> >>> >>> What I predicted, rather, was that multilayer perceptrons, as they >>> existed then, would not (on their own, absent other mechanisms) >>> *understand* language. Seven years later, they still haven?t, except in >>> the most superficial way. >>> >>> >>> >>> I made no comment whatsoever about machine translation, which I view as >>> a separate problem, solvable to a certain degree by correspondance without >>> semantics. >>> >>> >>> >>> I specifically tried to clarify Hinton?s confusion in 2019, but, >>> disappointingly, he has continued to purvey misinformation despite that >>> clarification. Here is what I wrote privately to him then, which should >>> have put the matter to rest: >>> >>> >>> >>> You have taken a single out of context quote [from 2015] and >>> misrepresented it. The quote, which you have prominently displayed at the >>> bottom on your own web page, says: >>> >>> >>> >>> Hierarchies of features are less suited to challenges such as language, >>> inference, and high-level planning. For example, as Noam Chomsky famously >>> pointed out, language is filled with sentences you haven't seen >>> before. Pure classifier systems don't know what to do with such sentences. >>> The talent of feature detectors -- in identifying which member of some >>> category something belongs to -- doesn't translate into understanding >>> novel sentences, in which each sentence has its own unique meaning. >>> >>> >>> >>> It does *not* say "neural nets would not be able to deal with novel >>> sentences"; it says that hierachies of features detectors (on their own, if >>> you read the context of the essay) would have trouble *understanding *novel sentences. >>> >>> >>> >>> >>> Google Translate does yet not *understand* the content of the sentences >>> is translates. It cannot reliably answer questions about who did what to >>> whom, or why, it cannot infer the order of the events in paragraphs, it >>> can't determine the internal consistency of those events, and so forth. >>> >>> >>> >>> Since then, a number of scholars, such as the the computational linguist >>> Emily Bender, have made similar points, and indeed current LLM difficulties >>> with misinformation, incoherence and fabrication all follow from these >>> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >>> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf >>> , >>> also emphasizing issues of understanding and meaning: >>> >>> >>> >>> *The success of the large neural language models on many NLP tasks is >>> exciting. However, we find that these successes sometimes lead to hype in >>> which these models are being described as ?understanding? language or >>> capturing ?meaning?. In this position paper, we argue that a system trained >>> only on form has a priori no way to learn meaning. .. a clear understanding >>> of the distinction between form and meaning will help guide the field >>> towards better science around natural language understanding. * >>> >>> >>> >>> Her later article with Gebru on language models ?stochastic parrots? is >>> in some ways an extension of this point; machine translation requires >>> mimicry, true understanding (which is what I was discussing in 2015) >>> requires something deeper than that. >>> >>> >>> >>> Hinton?s intellectual error here is in equating machine translation with >>> the deeper comprehension that robust natural language understanding will >>> require; as Bender and Koller observed, the two appear not to be the same. >>> (There is a longer discussion of the relation between language >>> understanding and machine translation, and why the latter has turned out to >>> be more approachable than the former, in my 2019 book with Ernest Davis). >>> >>> >>> >>> More broadly, Hinton?s ongoing dismissiveness of research from >>> perspectives other than his own (e.g. linguistics) have done the field a >>> disservice. >>> >>> >>> >>> As Herb Simon once observed, science does not have to be zero-sum. >>> >>> >>> >>> Sincerely, >>> >>> Gary Marcus >>> >>> Professor Emeritus >>> >>> New York University >>> >>> >>> >>> On Feb 2, 2022, at 06:12, AIhub wrote: >>> >>> ? >>> >>> Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> In the latest episode of this video series for AIhub.org >>> , >>> Stephen Hanson talks to Geoff Hinton about neural networks, >>> backpropagation, overparameterization, digit recognition, voxel cells, >>> syntax and semantics, Winograd sentences, and more. >>> >>> >>> >>> You can watch the discussion, and read the transcript, here: >>> >>> >>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>> >>> >>> >>> >>> About AIhub: >>> >>> AIhub is a non-profit dedicated to connecting the AI community to the >>> public by providing free, high-quality information through AIhub.org >>> >>> (https://aihub.org/ >>> ). >>> We help researchers publish the latest AI news, summaries of their work, >>> opinion pieces, tutorials and more. We are supported by many leading >>> scientific organizations in AI, namely AAAI >>> , >>> NeurIPS >>> , >>> ICML >>> , >>> AIJ >>> >>> /IJCAI >>> , >>> ACM SIGAI >>> , >>> EurAI/AICOMM, CLAIRE >>> >>> and RoboCup >>> >>> . >>> >>> Twitter: @aihuborg >>> >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> >>> -------------- next part -------------- An HTML attachment was scrubbed... URL: From michele.ferrante at nih.gov Tue Feb 8 14:02:15 2022 From: michele.ferrante at nih.gov (Ferrante, Michele (NIH/NIMH) [E]) Date: Tue, 8 Feb 2022 19:02:15 +0000 Subject: Connectionists: New Funding Opportunity: Neuro-Glia Computational Mechanisms Governing Complex Behaviors Message-ID: New Funding Opportunity: Neuro-Glia Computational Mechanisms Governing Complex Behaviors https://grants.nih.gov/grants/guide/notice-files/NOT-MH-22-090.html Purpose Background and Rationale This Notice of Special Interest (NOSI) encourages projects to experimentally test mechanistic hypotheses on the role of neuro-glia activity coupling in modulating complex behaviors. The human brain regulates complex behavior by processing information across ~170 billion cells, including ~86 billion neurons and ~84 billion glial cells. The influence of glial cell types (i.e., astrocytes, oligodendrocytes, and microglia) on neural activity may explain behavioral processes across broad spatio-temporal scales and hierarchies. For example, astrocytes may regulate cognitive functions by releasing gliotransmitters that activate hundreds of neuronal synapses at once, regulating system-level short-/long-term plasticity. Activity changes in neuro-oligodendrocytes networks may dynamically regulate myelin axon-sheathing, which in turn may affect action potential conduction, neuronal spike timing, and oscillations linked to cognitive/social/affective processes. Finally, microglia activity-dependent synaptic pruning may alter behaviorally activated neural networks over long time scales. Discovering how mechanistic dysfunctions in neuro-glia interactions may alter behavioral phenotypes relevant to mental health is a challenge with potentially high translational impact. Studying how neuro-glia activity coupling affects complex behavior has been challenging in part due to technical barriers. For example, until recently, the field lacked reliable and selective tools to manipulate glial cells. Biotechnology is now expanding the range of possible investigations into glial function by providing new methods necessary to record and manipulate glial cells with mouse lines and viral methods expressing designer reporters, sensors, and actuators of glial activity. Advances in metabolic imaging and genetically encoded activity measurements allow for simultaneous observation of interactions in neurons and glial activity. Tools are also available for selectively stimulating and silencing glia cells' calcium signaling in-vivo through designer receptors exclusively activated by designer drugs (DREADDs) and optogenetics. In parallel, basic behavioral neuroscience has been rapidly advanced by integrating system-level neurotechnology with computational modeling. Computational models (e.g., reinforcement learning, drift-diffusion, Bayesian, biophysically realistic, dynamical systems, and deep neural networks) have directly linked behavioral parameters and the neural substrates that compute them, but the role of non-neuronal cells in these computations have been largely ignored. Thus, the integration of computational modeling approaches to investigate neuro-glia interactions could provide new perspectives on how they enable complex behavior and how they become altered in mental illnesses. Combining newly developed experimental methods for recording and controlling neuro-glia activity with rigorous computational approaches may inform mechanistic models of how neuro-glia interactions may compute or fail to compute cognitive and socio-affective functions relevant to neuropsychiatric disorders. Responsive Areas of Research Examples of research areas that may be included in applications submitted under this NOSI include, but are not limited to, the application of existing or novel: * Experimental methods to measure or control glial activity in behaving animals to elucidate the role of neuro-glia activity in cognitive or socio-affective behaviors. * Biologically inspired computational approaches to provide a mechanistic explanation of the consequences of neuro-glia interactions during cognitive or socio-affective behaviors. * Computational models able to map fine-grained cognitive or socio-affective behavioral parameters onto neuro-glia computations. * Neuro-glia computational principles or mechanistic knowledge derived from animal experiments in Basic Experimental Studies Involving Humans (BESH). Applications must include both: * A well-conceived scientific rationale, question, and/or hypothesis grounded in cognitive or socio-affective science, where neuro-glia activity may mechanistically explain a complex behavior. * Measurement or control neuro-glia activity during complex behavior. Applications proposing to exclusively investigate the effects of glial cells on behavior will be considered responsive to this NOSI, but applications including measurement and/or manipulation of both glia and neurons will be given higher priority. Applications Not Responsive to this NOSI * Applications that do not include all the items contained in the "applications must include" section above are not responsive to this NOSI and will not be reviewed. * Applications exclusively focused on in-vitro preparations - without an in-vivo behavioral component - are not responsive to this NOSI and will not be reviewed. * Applications proposing to apply animal models of mental disorders or use broad batteries of behavioral tests in animals to address constructs that are accessible only in humans by self-report, such as "depression" or "anxiety," are non-responsive to this NOSI. For additional information on NIMH's guidelines and priorities for animal neurobehavioral approaches applicants are strongly encouraged to review NOT-MH-19-053. Note: NIMH only accepts mechanistic studies that meet NIH's definition of a clinical trial through PA-20-183 and PA-20-184. Applications directed to NIMH for intervention development must be submitted through PAR-21-130. For further information on NIMH clinical trial policies, see NOT-MH-20-105 and NOT-MH-19-006. Application and Submission Information This notice applies to due dates on or after June 5, 2022 and subsequent receipt dates through May 8, 2025. Submit applications for this initiative using one of the following funding opportunity announcements (FOAs) or any reissues of these announcement through the expiration date of this notice. * PA-20-184 - Research Project Grant (Parent R01 Basic Experimental Studies with Humans Required) * PA-20-185 - NIH Research Project Grant (Parent R01 Clinical Trial Not Allowed) * PA-20-196 - NIH Exploratory/Developmental Research Grant Program (Parent R21 Basic Experimental Studies with Humans Required) * PA-21-219 - Joint NINDS/NIMH Exploratory Neuroscience Research Grant (R21 Clinical Trial Optional) * PA-21-235 - NIMH Exploratory/Developmental Research Grant (R21 Clinical Trial Not Allowed) * PAR-21-175 - Understanding and Modifying Temporal Dynamics of Coordinated Neural Activity (R01 Clinical Trial Optional * PAR-21-176 - Understanding and Modifying Temporal Dynamics of Coordinated Neural Activity (R21 Clinical Trial Optional) All instructions in the SF424 (R&R) Application Guide and the funding opportunity announcement used for submission must be followed, with the following additions: * For funding consideration, applicants must include "NOT-MH-22-090" (without quotation marks) in the Agency Routing Identifier field (box 4B) of the SF424 R&R form. Applications without this information in box 4B will not be considered for this initiative. Applications nonresponsive to terms of this NOSI will not be considered for the NOSI initiative. Inquiries Please direct all inquiries to the contacts in Section VII of the listed funding opportunity announcements with the following additions/substitutions: Scientific/Research Contact(s) Michele Ferrante, Ph.D. National Institute of Mental Health (NIMH) Telephone: 301-435-6782 Email: ferrantem at nih.gov Andrew Breeden, Ph.D. National Institute of Mental Health (NIMH) Telephone: 301-443-1576 Email: andrew.breeden at nih.gov Regards, Michele ~~~~ Michele Ferrante, PhD Program Director: Computational Neuroscience & Computational Psychiatry Schedule a call: 301-828-7365|WebEx FOAs: Behavior|RDoC|XAI|CRCNS|Supplements|Theories|ConvergentNS -------------- next part -------------- An HTML attachment was scrubbed... URL: From t.hauser at ucl.ac.uk Tue Feb 8 10:04:52 2022 From: t.hauser at ucl.ac.uk (Tobias U. Hauser) Date: Tue, 8 Feb 2022 15:04:52 +0000 Subject: Connectionists: PostDoc position at UCL: endogenous brain fluctuations and decision making Message-ID: PostDoc position in real-time fMRI at Max Planck UCL Centre for Computational Psychiatry We are looking for a PostDoc to work on an exciting project investigating the role of endogenous brain fluctuations on decision making, funded by my ERC Starting Grant. In this project, we will investigate how endogenous, spontaneous fluctuations (aka resting-state fluctuations) influence behaviour. In our previous work (https://pnas.org/content/116/37/18732.short) we showed that moment-to-moment fluctuations affect risk taking behaviour. Here, we want to build on this and see how wide-reaching these effects are, and what the pharmacological bases of these effects are. We are looking for someone with expertise in advanced functional neuroimaging and preferably experience in real-time fMRI. We think that a postdoc with advanced understanding of time-series analyses and in-depth fMRI experience would be best suited for this position. Even though the general topic is set, own ideas and project ideas will very much appreciated & encouraged! The position is for 2 years (at the first instance) and will be based at the Max Planck UCL Centre for Computational Psychiatry and Ageing Research and the Wellcome Centre for Human Neuroimaging, located in Central London. If you have further questions, please contact Tobias Hauser (t.hauser at ucl.ac.uk). To read more about the research group, please see https://devcompsy.org/ Job advert here: https://atsv7.wcn.co.uk/search_engine/jobs.cgi?SID=amNvZGU9MTg4MjA4NSZ2dF90ZW1wbGF0ZT05NjUmb3duZXI9NTA0MTE3OCZvd25lcnR5cGU9ZmFpciZicmFuZF9pZD0wJmpvYl9yZWZfY29kZT0xODgyMDg1JnBvc3RpbmdfY29kZT0yMjQ= Best, Tobias -- Play smartphone games and help us understand the brain: www.brainexplorer.net Dr Tobias U. Hauser Sir Henry Dale Fellow, Principal Research Fellow Developmental Computational Psychiatry lab Max Planck UCL Centre for Computational Psychiatry & Ageing Research Wellcome Centre for Human Neuroimaging University College London 10-12 Russell Square London WC1B 5EH +44 207 679 5264 (internal: 45264) t.hauser at ucl.ac.uk www.tobiasuhauser.com www.devcompsy.org From albagarciaseco at gmail.com Tue Feb 8 10:40:09 2022 From: albagarciaseco at gmail.com (=?UTF-8?B?QWxiYSBHYXJjw61h?=) Date: Tue, 8 Feb 2022 16:40:09 +0100 Subject: Connectionists: CFP- ImageCLEF Coral Annotation Challenge 2022: Training set set released Message-ID: Website: https://www.imageclef.org/2022/coral *** CALL FOR PARTICIPATION *** *The 4th Edition of the ImageCLEF Coral Annotation Challenge 2022* Data The increasing use of structure-from-motion photogrammetry for modelling large-scale environments from action cameras attached to drones has driven the next-generation of visualisation techniques that can be used in augmented and virtual reality headsets. It has also created a need to have such models labelled, with objects such as people, buildings, vehicles, terrain, etc. all essential for machine learning techniques to automatically identify as areas of interest and to label them appropriately. However, the complexity of the images makes impossible for human annotators to assess the contents of images on a large scale. Advances in automatically annotating images for complexity and benthic composition have been promising, and we are interested in automatically identify areas of interest and to label them appropriately for monitoring coral reefs. Coral reefs are in danger of being lost within the next 30 years, and with them the ecosystems they support. This catastrophe will not only see the extinction of many marine species, but also create a humanitarian crisis on a global scale for the billions of humans who rely on reef services. By monitoring the changes and composition of coral reefs we can help prioritise conservation efforts. *New for 2022:* Previous editions of ImageCLEFcoral in 2019 and 2020 have shown improvements in task performance and promising results on cross-learning between images from geographical regions. The 3rd edition in 2021 increased the complexity of the task and size of data available to participants through supplemental data, resulting in lower performance than previous years. The 4th edition plans to address these issues by targeting algorithms for geographical regions and raising the benchmark performance. As with the 3rd edition, the training and test data will form the complete set of images required to form 3D reconstructions of the marine environment. This will allow the participants to explore novel probabilistic computer vision techniques based around image overlap and transposition of data points. *Challenge description* Participants will be require to annotate and localise coral reef images by labelling the images with types of benthic substrate together. Each image is provided with possible class types. ImageCLEFcoral 2022 consists of two substaks: - Coral reef image annotation and localisation: https://www.aicrowd.com/challenges/imageclef-2022-coral-annotation-and-l... - Coral reef image pixel-wise parsing: https://www.aicrowd.com/challenges/imageclef-2022-coral-pixel-wise-parsing *Data* The data for this task originates from a growing, large-scale collection of images taken from coral reefs around the world as part of a coral reef monitoring project with the Marine Technology Research Unit at the University of Essex. The images partially overlap with each other and can be used to create 3D photogrammetric models of the marine environment. Substrates of the same type can have very different morphologies, coloUr variation and patterns. Some of the images contain a white line (scientific measurement tape) that may occlude part of the entity.The quality of the images is variable, some are blurry, and some have poor colour balance due to the cameras being used. This is representative of the Marine Technology Research Unit dataset and all images are useful for data analysis. The training set used for 2022 has undergone a significant review in order to rectify errors in classification and polygon shape. Additionally, the 13 substrate types have been refined to help participants understand the results of their analyses. *Important dates* - *07.02.2022*: development data released - *14.03.2022*: test data release starts - *06.05.2021*: deadline for submitting the participants runs - *27.05.2021*: deadline for submission of working notes papers by the participants - *5-8.09.2021*: CLEF 2022 , Bologna, Italy *Participant Registration* https://www.imageclef.org/2021#registration *Organizing Committee* - Jon Chamberlain ,University of Essex, UK - Antonio Campello ,Wellcome Trust, UK - Adrian Clark ,University of Essex, UK - Alba Garc?a Seco de Herrera ,University of Essex, UK For more details and updates, please visit the task website at: https://www.imageclef.org/2022/coral And join our mailing list: https://groups.google.com/d/forum/imageclefcoral -------------- next part -------------- An HTML attachment was scrubbed... URL: From manuel.baltieri at gmail.com Tue Feb 8 21:20:41 2022 From: manuel.baltieri at gmail.com (Manuel Baltieri) Date: Wed, 9 Feb 2022 11:20:41 +0900 Subject: Connectionists: [CFP] Call for Papers - Hybrid Life at ALIFE 2022, Deadline 1st March Message-ID: ********** Call for Papers - Submission Deadline 1st March ********** Special session - Hybrid Life V: Approaches to integrate biological, artificial and cognitive systems https://sites.google.com/view/hybridlife 2022 Artificial Life conference (ALIFE 2022) Online, 18th-22nd July - http://2022.alife.org/ --------------------------------------------------------------------------------------------- DESCRIPTION: The main focus of ALife research is the study of natural systems with the goal of understanding what life is. More concretely, ALife defines ways to investigate processes that contribute to the formation and proliferation of living organisms. In this session we focus on three common approaches used to tackle this investigation, proposing new and hybrid ways to integrate, extend and improve them. Traditionally, ALife has focused on areas including: 1) the formalisation of properties necessary for the definition of life, 2) the implementation and analysis of artificial agents, and 3) the study of the relation between life and cognition. For this special session we propose to start from these well-established Alife methodologies, and to create hybrid approaches that extend them through: the search for a unifying framework that spans across (models of) living, artificial and cognitive systems, overcoming the limitations of approaches focusing only on one type of systems; this area may include life-mind-continuity thesis, systems biology, theories of agency based on Bayesian inference, dynamical systems, information theory, etc. the exploration of biological creatures enhanced by artificial systems (or artificial systems augmented with organic parts) in order to investigate the boundaries between living and nonliving organisms; this includes work from bio-inspired robotics, human augmentation, synthetic biology, etc. the evaluation of coupled biological-artificial systems that could shed light on the importance of interactions among systems for the study of living and cognitive organisms; this approach welcomes contributions from the fields of human-agent interaction, animal-computer interaction, virtual / augmented reality systems, etc. The session focuses on hybrid methods, theoretical contributions that can shed new light on concepts common across artificial/ living/ cognitive systems (e.g., agency, goal-directed behaviour, self-organisation, adaptation and self-maintenance), and hybrid systems, where robotics and biology are combined to study areas in the cognitive domain. This special session aims to invite contributions from the fields of psychology, computational neuroscience, human-computer interaction (HCI), theoretical biology, artificial intelligence, robotics and cognitive science. Potential topics include, but are not limited to: Mathematical frameworks for life and cognition (e.g. dynamical systems theory, stochastic optimal control, Bayesian inference, etc.) Cognitive robotics Autopoiesis Life-mind continuity thesis Systems biology Origins-of-life theories with relationships to artificial and cognitive systems Animal-robot interaction Bio-inspired robotics Bio-integrated robotics Human-machine interaction Augmented cognition Sensory substitution Interactive evolutionary computation Artificial phenomenology Important Dates 1st March 2022 ? Paper submission deadline 15th April 2019 ? Paper acceptance notification 15th May 2019 ? Camera-ready version 18th-22nd July ? Artificial Life conference (ALIFE), Online Paper Submission Papers and abstracts submitted to this special sessions will be reviewed by a selected group of experts from the ALife community as well as from other areas key to our proposal, specifically chosen for this review process. If you are submitting to a special session you will be given the opportunity to select it during the submission process. Submissions to special sessions follow the same format, instructions and deadlines of regular ALife papers. Organizers Manuel Baltieri, Araya Inc., Tokyo, Japan - University of Sussex, Brighton, UK Keisuke Suzuki, Center for Human Nature, Artificial Intelligence, and Neuroscience, Hokkaido, Japan Olaf Witkowski, Cross Labs, Kyoto, Japan Contacts For questions, enquiries and more information please check our website https://sites.google.com/view/hybridlife or get in touch with using hybrid.alife [at] gmail.com (please notice the ?a? in alife). All the best, -- Manuel Baltieri, PhD Researcher, Araya Inc., Tokyo, Japan Visiting researcher, University of Sussex, Brighton, UK www.manuelbaltieri.com Twitter: @manuelbaltieri -------------- next part -------------- An HTML attachment was scrubbed... URL: From minaiaa at gmail.com Tue Feb 8 23:25:22 2022 From: minaiaa at gmail.com (Ali Minai) Date: Tue, 8 Feb 2022 23:25:22 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <4C768F69-06BE-48DA-AF37-4D2DBC035962@nyu.edu> References: <4C768F69-06BE-48DA-AF37-4D2DBC035962@nyu.edu> Message-ID: Hi Gary Thanks for letting me know about the book. I have a lot of reading in my future :-) The anecdote aboutTversky was reported in Michael Lewis' book, "The Undoing Project", about Kahneman and Tversky's work on irrationality. Apparently someone asked him if he worked on artificial intelligence. He replied, "No, I work on natural stupidity." I think that, the wordplay aside, it's a brilliant way to juxtapose the irrationality of the human mind against the desired rationality that we insist on in the artificial minds we are creating. Best Ali *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Tue, Feb 8, 2022 at 3:07 PM Gary Marcus wrote: > Hi, Ali, > > I don?t actually know where or when the term ?natural stupidity? was first > used, but McDermott?s points in this paper are still quite relevant, both > for symbolic and neural approaches: > > > https://cs.fit.edu/~kgallagher/Schtick/Serious/McDermott.AI.MeetsNaturalStupidity.pdf > > With respect to human cognitive limitations, if you make it through Marcus > 2004, there is also Marcus 2008, Kluge: The Haphazard Evolution of the > Human Mind to add to your list. > > Humans set a low bar; perhaps some day AI can exceed it. > > Gary > > On Feb 8, 2022, at 10:19, Ali Minai wrote: > > ? > Hi Gary > > Thanks for your reply. Much to think about in it. For one, it makes me > want to read your 2004 book right away (which I will). > > All human capacities - however unique they may seem to us - must > necessarily evolve from those of the precursor species. If we had > Australopithecus and Homo Erectus brains to study, I am sure we would find > less developed examples of symbolic processing, compositionality, etc., > that are recognizably of the same type as ours, and thus link our > capacities with earlier ancestors that may have been closer to chimpanzees > or bonobos. There's increasing reason to believe that Homo Neandertalis had > a pretty complex mind, though it is on a different evolutionary branch than > us (though it has contributed about 2% of our genome through > cross-breeding). That suggests that the common ancestor of both H. Sapiens > and H. Neandertalis also had these mental capacities to some degree. Moving > backwards like that, it's hard to decide where the chain should be cut. Of > course, we know that phenotypic attributes can emerge very rapidly in the > evolutionary process and then be in stasis for a long time, but they do > emerge from the available substrate through processes like > duplication-and-divergence, recombination, diversification in neutral > spaces, external pressures, etc.Understanding that process and leveraging > it for engineering is, in my opinion, a more reliable route to AI. Again, I > don't mean that we should replicate evolution - just that we should reverse > engineer its results. > > I think compositionality is a useful concept in the analysis of language > and inference, but not very useful as a prescriptive goal for building AI > systems. It's an attribute we see in human language now and would like to > see in our AI systems. But instead of asking "How can I get > compositionality (or any attribute X) into my neural network model?" I want > to ask, "How do the neural networks of the brain come to have the attribute > of compositionality (or X)?" or "What is it in the anatomy and physiology > of brain networks that allows for compositionality to be possible?" > Answering that question will also tell us how to achieve it in our models. > Otherwise, we are trying to get water from rocks. We don't know whether any > of our neural models are capable of the kind of symbolic processing, > compositionality, etc., that we want. We keep inventing new models to > incorporate such desired attributes purely by thinking them up based on our > previous models and their failures. This process is, in my opinion, a > futile quest that is, at best, likely to lead to bad, inadequate models > that will fall apart when we try to integrate them into larger intelligent > systems. Much better, I suggest, to look at the real-world systems that > exist, i.e., animal brains and bodies, and see how all the aspects of > intelligence emerge from them. When we do, we will have neural models > capable of symbolic processing, language, compositionality, etc. That's why > I think it's better to have models that are more grounded in learning from > biology with humility and less invested in top-down analysis and design > based on our formalisms. > > One place where we can learn a lot about "real intelligence" is in the > ways it "fails", i.e., by focusing more on what Amos Tversky flippantly > called "natural stupidity". In those imperfections, we may be able to see > how a real-time physical system like the brain achieves capacities such as > abstract reasoning, symbolic processing, language, etc. As my old friend > Daniel Polani suggested in another post in this thread (hi Daniel!), we > humans do all these things quite imperfectly, and far from being a bug, > that is probably a feature - an indicator of how, by sacrificing the goal > of being perfect, we (i.e., evolution) have been able to become merely > excellent. > > If we hope to understand intelligence, there are several inherited > preferences that we need to kill. High on that kill list are the obsessions > with optimality, stability, repeatability, explainability, etc., that are > the antithesis of how natural living systems grow, thrive, and evolve. > > Ali > > > *Ali A. Minai, Ph.D.* > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > > > On Tue, Feb 8, 2022 at 10:42 AM Gary Marcus wrote: > >> Hi, Ali, >> >> Just to be clear, I certainly think that animal models can be useful in >> principle, and eg. on motor-control and low-level vision they have been >> incredibly useful. On compositionality, they may turn out to be less so, at >> least in the near-term. My 2004 book The Birth of the Mind has a chapter on >> evolution where I try to reconcile what we know about the astonishing >> conservation across the biological world with the apparent uniqueness of >> human language, with a focus on how something that is apparently unique can >> rest partly but not entirely on inherited substrates. Emphasizing the >> biological process of duplication and divergence, I suggested that language >> might have hinged on multiple simultaneous mutations to existing >> cognitively-relevant genes that gave rise to advances in the capacity for >> hierarchical representation. Tecumseh Fitch has recently made a somewhat >> parallel argument, also suggesting that compositionality in strong form may >> be humanly unique, but arguing instead that a large expansion in Broca?s >> area is responsible. If either his speculation or mine were correct, animal >> models might be of relatively little direct help. (Though of course >> anything that helps us understand the brain better may be of substantial >> indirect help.) >> >> On your other point, saying that ?symbols emerge naturally from the >> physics of the brains [of the animals that have them]? just doesn?t tell >> us much; *everything* any animal does emerges from physics and biology >> and experience. But nature has a vast array of different solutions to a >> wide variety of problems, and it?s an empirical matter to understand each >> creature and how it handles different problems (digestion, circulation, >> navigation, planning, etc). All that is of course constrained by physics, >> but it doesn't tell us eg whether any specific aspect of a particular >> creature?s physiology ?emerges? or is hard-coded, learned, etc. >> (Empirically, my own experimental work with babies, since replicated by >> others with newborns, suggests that the capacity to acquire novel, >> rule-like abstractions emerges from a developmental program that does not >> depend on post-natal experience.) >> >> More broadly, I think that most brains on this planet have some capacity >> for symbol-manipulation (eg honey bees calculating the solar azimuth >> function and extrapolating to novel lighting conditions) but that >> compositionality is less prevalent. I don?t think that you have to have >> symbols or compositionality to have a hint of intelligence, but I do think >> you probably need them for AGI or for understanding how humans tick. >> >> Gary >> >> >> >> On Feb 7, 2022, at 23:22, Ali Minai wrote: >> >> ? >> Hi Gary >> >> Thanks for your reply. I'll think more about your points. I do think >> that, to understand the human mind, we should start with vertebrates, which >> is why I suggested fish. At least for the motor system - which is part of >> the mind - we have learned a lot from lampreys (e.g. Sten Grillner's work >> and that beautiful lamprey-salamander model by Ijspeert et al.), and it has >> taught us a lot about locomotion in other animals, including mammals. The >> principles clearly generalize, though the complexity increases a lot. >> Insects too are very interesting. After all, they are our ancestors too. >> >> I don't agree that we can posit a clear transition from deep cognitive >> models in humans to none below that in the phylogenetic tree. Chimpanzees >> and macaques certainly show some evidence, and there's no reason to think >> that it's a step change rather than a highly nonlinear continuum. And even >> though what we might (simplistically) call System 2 aspects of cognition >> are minimally present in other mammals, their precursors must be. >> >> My point about cats and symbols was not regarding whether cats are aware >> of symbols, but that symbols emerge naturally from the physics of their >> brains. Behaviors that require some small degree of symbolic processing >> exist in mammals other than humans (e.g., transitive inference and >> landmark-based navigation in rats), and it is seen better as an emergent >> property of brains than an attribute to be explicitly built-into neural >> models by us. Once we have a sufficiently brain-like neural model, symbolic >> processing will already be there. >> >> I agree with you completely that we are far from understanding some of >> the most fundamental principles of the brain, but even more importantly, we >> are not even looking in the right direction. I'm hoping to lay out my >> arguments about all this in more detail in some other form. >> >> Best >> Ali >> >> >> PS: I had inadvertently posted my reply of Gary's message only to him. >> Should have posted to everyone, so here it is. >> >> >> *Ali A. Minai, Ph.D.* >> Professor and Graduate Program Director >> Complex Adaptive Systems Lab >> Department of Electrical Engineering & Computer Science >> 828 Rhodes Hall >> University of Cincinnati >> Cincinnati, OH 45221-0030 >> >> >> Phone: (513) 556-4783 >> Fax: (513) 556-7326 >> Email: Ali.Minai at uc.edu >> minaiaa at gmail.com >> >> WWW: https://eecs.ceas.uc.edu/~aminai/ >> >> >> >> On Mon, Feb 7, 2022 at 12:28 AM Gary Marcus wrote: >> >>> Ali, >>> >>> >>> It?s useful to think about animals, but I really wouldn?t start with >>> fish; it?s not clear that their ecological niche demands anything >>> significant in the way of extrapolation, causal reasoning, or >>> compositionality. There is good evidence elsewhere in the animal world for >>> extrapolation of functions that may be innate (eg solar azimuth in bees), >>> and causal reasoning (eg tool use in ravens, various primates, and >>> octopi). It?s still not clear to me how much hierarchical representation >>> (critical to AGI) exists outside of humans, though; the ability to >>> construct rich new cognitive models may also be unique to us. >>> >>> >>> In any case it matters not in the least whether the average cat or human >>> *cares* about symbols, anymore that it matters whether the average >>> animal understands digestion; only a tiny fraction of the creatures on this >>> planet have any real understanding of their internal workings. >>> >>> >>> My overall feeling is that we are a really, really long way from >>> understanding the neural basis of higher-level cognition, and that AI is >>> going to need muddle through on its own, for another decade or two, >>> >>> >>> I do fully agree with your conclusion, though, that "AI today is driven >>> more by habit and the incentives of the academic and corporate marketplaces >>> than by a deep, long-term view of AI as a great exploratory project in >>> fundamental science." Let's hope that changes. >>> >>> >>> Gary >>> >>> On Feb 6, 2022, at 13:19, Ali Minai wrote: >>> >>> ? >>> >>> Gary, >>> >>> That?s a very interesting and accurate list of capabilities that a >>> general intelligent system must have and that our AI does not. Of course, >>> the list is familiar to me from having read your book. However, I have a >>> somewhat different take on this whole thing. >>> >>> >>> >>> All the things we discuss here ? symbols/no symbols, parts/wholes, >>> supervised/unsupervised, token/type, etc., are useful categories and >>> distinctions for our analysis of the problem, and are partly a result of >>> the historical evolution of the field of AI in particular and of philosophy >>> in general. The categories are not wrong in any way, of course, but they >>> are posterior to the actual system ? good for describing and analyzing it, >>> and for validating our versions of it (which is how you use them). I think >>> they are less useful as prescriptions for how to build our AI systems. If >>> intelligent systems did not already exist and we were building them from >>> scratch (please ignore the impossibility of that), having a list of ?must >>> haves? would be great. But intelligent systems already exist ? from humans >>> to fish ? and they already have these capacities to a greater or lesser >>> degree because of the physics of their biology. A cat?s intelligence does >>> not care whether it has symbols or not, and nor does mine or yours. >>> Whatever we describe as symbolic processing post-facto has already been >>> done by brains for at least tens of millions of years. Instead of getting >>> caught up in ?how to add symbols into our neural models?, we should be >>> investigating how what we see as symbolic processing emerges from animal >>> brains, and then replicate those brains to the degree necessary. If we can >>> do that, symbolic processing will already be present. But it cannot be done >>> piece by piece. It must take the integrity of the whole brain and the body >>> it is part of, and its environment, into account. That?s why I think that a >>> much better ? though a very long ? route to AI is to start by understanding >>> how a fish brain makes the intelligence of a fish possible, and then boot >>> up our knowledge across phylogenetic stages: Bottom up reverse engineering >>> rather than top-down engineering. That?s the way Nature built up to human >>> intelligence, and we will succeed only by reverse engineering it. Of >>> course, we can do it much faster and with shortcuts because we are >>> intelligent, purposive agents, but working top-down by building piecewise >>> systems that satisfy a list of attributes will not get us there. Among >>> other things, those pieces will be impossible to integrate into the kind of >>> intelligence that can have those general models of the world that you >>> rightly point to as being necessary. >>> >>> >>> >>> I think that one thing that has been a great boon to the AI enterprise >>> has also been one of the greatest impediments to its complete success, and >>> that is the ?computationalization? of intelligence. On the one hand, >>> thinking of intelligence computationally allows us to describe it >>> abstractly and in a principled, formal way. It also resonates with the fact >>> that we are trying to implement intelligence through computational >>> machines. But, on the flip side, this view of intelligence divorces it from >>> its physics ? from the fact that real intelligence in animals emerges from >>> the physics of the physical system. That system is not a collection of its >>> capabilities; rather, those capabilities are immanent in it by virtue of >>> its physics. When we try to build those capabilities computationally, i.e., >>> through code, we are making the same error that the practitioners of >>> old-style ?symbolic AI? made ? what I call the ?professors are smarter than >>> Nature? error, i.e., the idea that we are going to enumerate (or describe) >>> all the things that underlie intelligence and implement them one by one >>> until we get complete intelligence. We will never be able to enumerate all >>> those capabilities, and will never be able to get to that complete >>> intelligence. The only difference between us and the ?symbolists? of yore >>> is that we are replacing giant LISP and Prolog programs with giant neural >>> networks. Otherwise, we are using our models exactly as they were trying to >>> use their models, and we will fail just as they did unless we get back to >>> biology and the real thing. >>> >>> >>> >>> I will say again that the way we do AI today is driven more by habit and >>> the incentives of the academic and corporate marketplaces than by a deep, >>> long-term view of AI as a great exploratory project in fundamental science. >>> We are just building AI to drive our cars, translate our documents, write >>> our reports, and do our shopping. What that will teach us about actual >>> intelligence is just incidental. >>> >>> >>> >>> My apologies too for a long response. >>> >>> Ali >>> >>> >>> *Ali A. Minai, Ph.D.* >>> Professor and Graduate Program Director >>> Complex Adaptive Systems Lab >>> Department of Electrical Engineering & Computer Science >>> 828 Rhodes Hall >>> University of Cincinnati >>> Cincinnati, OH 45221-0030 >>> >>> Phone: (513) 556-4783 >>> Fax: (513) 556-7326 >>> Email: Ali.Minai at uc.edu >>> minaiaa at gmail.com >>> >>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>> >>> >>> >>> On Sun, Feb 6, 2022 at 9:42 AM Gary Marcus wrote: >>> >>>> Dear Asim, >>>> >>>> >>>> Sorry for a long answer to your short but rich questions. >>>> >>>> - Yes, memory in my view has to be part of the answer to the >>>> type-token problem. Symbol systems encoded in memory allow a natural way to >>>> set up records, and something akin to that seems necessary. Pure multilayer >>>> perceptrons struggle with type-token distinctions precisely because they >>>> lack such records. On the positive side, I see more and more movement >>>> towards recordlike stores (eg w key-value stores in memory networks), and I >>>> think that is an important and necessary step, very familiar from the >>>> symbol-manipulating playbook, sometimes implemented in new ways. >>>> - But ultimately, handling the type-token distinction requires >>>> considerable inferential overhead beyond the memory representation of a >>>> record per se. How do you determine when to denote something (e.g. >>>> Felix) as an instance, and of which kinds (cat, animal etc), and how do you >>>> leverage that knowledge once you determine it? >>>> - In the limit we reason about types vs tokens in fairly subtle >>>> ways, eg in guessing whether a glass that we put down at party is likely to >>>> be ours. The reverse is also important: we need to be learn >>>> particular traits for individuals and not erroneously generalize them to >>>> the class; if my aunt Esther wins the lottery, one shouldn?t infer that >>>> all of my aunts or all of my relatives or adult females have won the >>>> lottery. so you need both representational machinery that can distinguish >>>> eg my cat from cats in general and reasoning machinery to decide at what >>>> level certain learned knowledge should inhere. (I had a whole chapter about >>>> this sort of thing in The Algebraic Mind if you are interested, and Mike >>>> Mozer had a book about types and tokens in neural networks in the mid >>>> 1990s). >>>> - Yes, part (though not all!) of what we do when we set up >>>> cognitive models in our heads is to track particular individuals and their >>>> properties. If you only had to correlate kinds (cats) and their properties >>>> (have fur) you could maybe get away with a multilayer perceptron, but once >>>> you need to track individuals, yes, you really need some kind of >>>> memory-based records. >>>> - As far as I can tell, Transformers can sometimes approximate some >>>> of this for a few sentences, but not over long stretches. >>>> >>>> >>>> As a small terminological aside; for me cognitive models ? cognitive >>>> modeling. Cognitive modeling is about building psychological or >>>> computational models of how people think, whereas what I mean by a cognitive >>>> model is a representation of eg the entities in some situation and the >>>> relations between those entities. >>>> >>>> >>>> To your closing question, none of us yet really knows how to build >>>> understanding into machines. A solid type-token distinction, both in >>>> terms of representation and reasoning, is critical for general >>>> intelligence, but hardly sufficient. Personally, I think some minimal >>>> prerequisites would be: >>>> >>>> - representations of space, time, causality, individuals, kinds, >>>> persons, places, objects, etc. >>>> - representations of abstractions that can hold over all entities >>>> in a class >>>> - compositionality (if we are talking about human-like >>>> understanding) >>>> - capacity to construct and update cognitive models on the fly >>>> - capacity to reason over entities in those models >>>> - ability to learn about new entities and their properties >>>> >>>> Much of my last book (*Rebooting AI*, w Ernie Davis) is about the >>>> above list. The section in the language chapter on a children?s story in >>>> which man has lost is wallet is an especially vivid worked example. Later >>>> chapters elaborate some of the challenges in representing space, time, and >>>> causality. >>>> >>>> >>>> Gary >>>> >>>> >>>> On Feb 5, 2022, at 18:58, Asim Roy wrote: >>>> >>>> ? >>>> >>>> Gary, >>>> >>>> >>>> >>>> I don?t get much into the type of cognitive modeling you are talking >>>> about, but I would guess that the type problem can generally be handled by >>>> neural network models and tokens can be resolved with some memory-based >>>> system. But to the heart of the question, this is what so-called >>>> ?understanding? reduces to computation wise? >>>> >>>> >>>> >>>> Asim >>>> >>>> >>>> >>>> *From:* Gary Marcus >>>> *Sent:* Saturday, February 5, 2022 8:39 AM >>>> *To:* Asim Roy >>>> *Cc:* Ali Minai ; Danko Nikolic < >>>> danko.nikolic at gmail.com>; Brad Wyble ; >>>> connectionists at mailman.srv.cs.cmu.edu; AIhub >>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>> Geoff Hinton >>>> >>>> >>>> >>>> There is no magic in understanding, just computation that has been >>>> realized in the wetware of humans and that eventually can be realized in >>>> machines. But understanding is not (just) learning. >>>> >>>> >>>> >>>> Understanding incorporates (or works in tandem with) learning - but >>>> also, critically, in tandem with inference, *and the development and >>>> maintenance of cognitive models*. Part of developing an understanding >>>> of cats in general is to learn long term-knowledge about their properties, >>>> both directly (e.g., through observation) and indirectly (eg through >>>> learning facts about animals in general that can be extended to >>>> cats), often through inference (if all animals have DNA, and a cat is an >>>> animal, it must also have DNA). The understanding of a particular cat >>>> also involves direct observation, but also inference (eg one might >>>> surmise that the reason that Fluffy is running about the room is that >>>> Fluffy suspects there is a mouse stirring somewhere nearby). *But all >>>> of that, I would say, is subservient to the construction of cognitive >>>> models that can be routinely updated *(e.g., Fluffy is currently in >>>> the living room, skittering about, perhaps looking for a mouse). >>>> >>>> >>>> >>>> In humans, those dynamic, relational models, which form part of an >>>> understanding, can support inference (if Fluffy is in the living room, we >>>> can infer that Fluffy is not outside, not lost, etc). Without such models - >>>> which I think represent a core part of understanding - AGI is an unlikely >>>> prospect. >>>> >>>> >>>> >>>> Current neural networks, as it happens, are better at acquiring >>>> long-term knowledge (cats have whiskers) than they are at dynamically >>>> updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind >>>> of dynamic model that I am describing. To a modest degree they can >>>> approximate it on the basis of large samples of texts, but their ultimate >>>> incoherence stems from the fact that they do not have robust internal >>>> cognitive models that they can update on the fly. >>>> >>>> >>>> >>>> Without such cognitive models you can still capture some aspects of >>>> understanding (eg predicting that cats are likely to be furry), but things >>>> fall apart quickly; inference is never reliable, and coherence is fleeting. >>>> >>>> >>>> >>>> As a final note, one of the most foundational challenges in >>>> constructing adequate cognitive models of the world is to have a clear >>>> distinction between individuals and kinds; as I emphasized 20 years ago (in >>>> The Algebraic Mind), this has always been a weakness in neural networks, >>>> and I don?t think that the type-token problem has yet been solved. >>>> >>>> >>>> >>>> Gary >>>> >>>> >>>> >>>> >>>> >>>> On Feb 5, 2022, at 01:31, Asim Roy wrote: >>>> >>>> ? >>>> >>>> All, >>>> >>>> >>>> >>>> I think the broader question was ?understanding.? Here are two Youtube >>>> videos showing simple robots ?learning? to walk. They are purely physical >>>> systems. Do they ?understand? anything ? such as the need to go around an >>>> obstacle, jumping over an obstacle, walking up and down stairs and so on? >>>> By the way, they ?learn? to do these things on their own, literally >>>> unsupervised, very much like babies. The basic question is: what is >>>> ?understanding? if not ?learning?? Is there some other mechanism (magic) at >>>> play in our brain that helps us ?understand?? >>>> >>>> >>>> >>>> https://www.youtube.com/watch?v=gn4nRCC9TwQ >>>> >>>> >>>> https://www.youtube.com/watch?v=8sO7VS3q8d0 >>>> >>>> >>>> >>>> >>>> >>>> >>>> Asim Roy >>>> >>>> Professor, Information Systems >>>> >>>> Arizona State University >>>> >>>> Lifeboat Foundation Bios: Professor Asim Roy >>>> >>>> >>>> Asim Roy | iSearch (asu.edu) >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> *From:* Ali Minai >>>> *Sent:* Friday, February 4, 2022 11:38 PM >>>> *To:* Asim Roy >>>> *Cc:* Gary Marcus ; Danko Nikolic < >>>> danko.nikolic at gmail.com>; Brad Wyble ; >>>> connectionists at mailman.srv.cs.cmu.edu; AIhub >>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>> Geoff Hinton >>>> >>>> >>>> >>>> Asim >>>> >>>> >>>> >>>> Of course there's nothing magical about understanding, and the mind has >>>> to emerge from the physical system, but our AI models at this point are not >>>> even close to realizing how that happens. We are, at best, simulating a >>>> superficial approximation of a few parts of the real thing. A single, >>>> integrated system where all the aspects of intelligence emerge from the >>>> same deep, well-differentiated physical substrate is far beyond our >>>> capacity. Paying more attention to neurobiology will be essential to get >>>> there, but so will paying attention to development - both physical and >>>> cognitive - and evolution. The configuration of priors by evolution is key >>>> to understanding how real intelligence learns so quickly and from so >>>> little. This is not an argument for using genetic algorithms to design our >>>> systems, just for understanding the tricks evolution has used and >>>> replicating them by design. Development is more feasible to do >>>> computationally, but hardly any models have looked at it except in a >>>> superficial sense. Nature creates basic intelligence not so much by >>>> configuring functions by explicit training as by tweaking, modulating, >>>> ramifying, and combining existing ones in a multi-scale self-organization >>>> process. We then learn much more complicated things (like playing chess) by >>>> exploiting that substrate, and using explicit instruction or learning by >>>> practice. The fundamental lesson of complex systems is that complexity is >>>> built in stages - each level exploiting the organization of the level below >>>> it. We see it in evolution, development, societal evolution, the evolution >>>> of technology, etc. Our approach in AI, in contrast, is to initialize a >>>> giant, naive system and train it to do something really complicated - but >>>> really specific - by training the hell out of it. Sure, now we do build >>>> many systems on top of pre-trained models like GPT-3 and BERT, which is >>>> better, but those models were again trained by the same none-to-all process >>>> I decried above. Contrast that with how humans acquire language, and how >>>> they integrate it into their *entire* perceptual, cognitive, and behavioral >>>> repertoire, not focusing just on this or that task. The age of symbolic AI >>>> may have passed, but the reductionistic mindset has not. We cannot build >>>> minds by chopping it into separate verticals. >>>> >>>> >>>> >>>> FTR, I'd say that the emergence of models such as GLOM and Hawkins and >>>> Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but >>>> they are, I think, looking in the right direction. With a million miles to >>>> go! >>>> >>>> >>>> >>>> Ali >>>> >>>> >>>> >>>> *Ali A. Minai, Ph.D.* >>>> Professor and Graduate Program Director >>>> Complex Adaptive Systems Lab >>>> Department of Electrical Engineering & Computer Science >>>> >>>> 828 Rhodes Hall >>>> >>>> University of Cincinnati >>>> Cincinnati, OH 45221-0030 >>>> >>>> >>>> Phone: (513) 556-4783 >>>> Fax: (513) 556-7326 >>>> Email: Ali.Minai at uc.edu >>>> minaiaa at gmail.com >>>> >>>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>>> >>>> >>>> >>>> >>>> >>>> >>>> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: >>>> >>>> First of all, the brain is a physical system. There is no ?magic? >>>> inside the brain that does the ?understanding? part. Take for example >>>> learning to play tennis. You hit a few balls - some the right way and some >>>> wrong ? but you fairly quickly learn to hit them right most of the time. So >>>> there is obviously some simulation going on in the brain about hitting the >>>> ball in different ways and ?learning? its consequences. What you are >>>> calling ?understanding? is really these simulations about different >>>> scenarios. It?s also very similar to augmentation used to train image >>>> recognition systems where you rotate images, obscure parts and so on, so >>>> that you still can say it?s a cat even though you see only the cat?s face >>>> or whiskers or a cat flipped on its back. So, if the following questions >>>> relate to ?understanding,? you can easily resolve this by simulating such >>>> scenarios when ?teaching? the system. There?s nothing ?magical? about >>>> ?understanding.? As I said, bear in mind that the brain, after all, is a >>>> physical system and ?teaching? and ?understanding? is embodied in that >>>> physical system, not outside it. So ?understanding? is just part of >>>> ?learning,? nothing more. >>>> >>>> >>>> >>>> DANKO: >>>> >>>> What would happen to the hat if the hamster rolls on its back? (Would >>>> the hat fall off?) >>>> >>>> What would happen to the red hat when the hamster enters its lair? >>>> (Would the hat fall off?) >>>> >>>> What would happen to that hamster when it goes foraging? (Would the red >>>> hat have an influence on finding food?) >>>> >>>> What would happen in a situation of being chased by a predator? (Would >>>> it be easier for predators to spot the hamster?) >>>> >>>> >>>> >>>> Asim Roy >>>> >>>> Professor, Information Systems >>>> >>>> Arizona State University >>>> >>>> Lifeboat Foundation Bios: Professor Asim Roy >>>> >>>> >>>> Asim Roy | iSearch (asu.edu) >>>> >>>> >>>> >>>> >>>> >>>> >>>> *From:* Gary Marcus >>>> *Sent:* Thursday, February 3, 2022 9:26 AM >>>> *To:* Danko Nikolic >>>> *Cc:* Asim Roy ; Geoffrey Hinton < >>>> geoffrey.hinton at gmail.com>; AIhub ; >>>> connectionists at mailman.srv.cs.cmu.edu >>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>> Geoff Hinton >>>> >>>> >>>> >>>> Dear Danko, >>>> >>>> >>>> >>>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >>>> talk, in which he said (paraphrasing from memory, because I don?t remember >>>> the precise words) that the famous 200 Quoc Le unsupervised model [ >>>> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >>>> ] >>>> had learned the concept of a ca. In reality the model had clustered >>>> together some catlike images based on the image statistics that it had >>>> extracted, but it was a long way from a full, counterfactual-supporting >>>> concept of a cat, much as you describe below. >>>> >>>> >>>> >>>> I fully agree with you that the reason for even having a semantics is >>>> as you put it, "to 1) learn with a few examples and 2) apply the knowledge >>>> to a broad set of situations.? GPT-3 sometimes gives the appearance of >>>> having done so, but it falls apart under close inspection, so the problem >>>> remains unsolved. >>>> >>>> >>>> >>>> Gary >>>> >>>> >>>> >>>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >>>> wrote: >>>> >>>> >>>> >>>> G. Hinton wrote: "I believe that any reasonable person would admit that >>>> if you ask a neural net to draw a picture of a hamster wearing a red hat >>>> and it draws such a picture, it understood the request." >>>> >>>> >>>> >>>> I would like to suggest why drawing a hamster with a red hat does not >>>> necessarily imply understanding of the statement "hamster wearing a red >>>> hat". >>>> >>>> To understand that "hamster wearing a red hat" would mean inferring, in >>>> newly emerging situations of this hamster, all the real-life >>>> implications that the red hat brings to the little animal. >>>> >>>> >>>> >>>> What would happen to the hat if the hamster rolls on its back? (Would >>>> the hat fall off?) >>>> >>>> What would happen to the red hat when the hamster enters its lair? >>>> (Would the hat fall off?) >>>> >>>> What would happen to that hamster when it goes foraging? (Would the red >>>> hat have an influence on finding food?) >>>> >>>> What would happen in a situation of being chased by a predator? (Would >>>> it be easier for predators to spot the hamster?) >>>> >>>> >>>> >>>> ...and so on. >>>> >>>> >>>> >>>> Countless many questions can be asked. One has understood "hamster >>>> wearing a red hat" only if one can answer reasonably well many of such >>>> real-life relevant questions. Similarly, a student has understood materias >>>> in a class only if they can apply the materials in real-life situations >>>> (e.g., applying Pythagora's theorem). If a student gives a correct answer >>>> to a multiple choice question, we don't know whether the student understood >>>> the material or whether this was just rote learning (often, it is rote >>>> learning). >>>> >>>> >>>> >>>> I also suggest that understanding also comes together with effective >>>> learning: We store new information in such a way that we can recall it >>>> later and use it effectively i.e., make good inferences in newly emerging >>>> situations based on this knowledge. >>>> >>>> >>>> >>>> In short: Understanding makes us humans able to 1) learn with a few >>>> examples and 2) apply the knowledge to a broad set of situations. >>>> >>>> >>>> >>>> No neural network today has such capabilities and we don't know how to >>>> give them such capabilities. Neural networks need large amounts of >>>> training examples that cover a large variety of situations and then >>>> the networks can only deal with what the training examples have already >>>> covered. Neural networks cannot extrapolate in that 'understanding' sense. >>>> >>>> >>>> >>>> I suggest that understanding truly extrapolates from a piece of >>>> knowledge. It is not about satisfying a task such as translation between >>>> languages or drawing hamsters with hats. It is how you got the capability >>>> to complete the task: Did you only have a few examples that covered >>>> something different but related and then you extrapolated from that >>>> knowledge? If yes, this is going in the direction of understanding. Have >>>> you seen countless examples and then interpolated among them? Then perhaps >>>> it is not understanding. >>>> >>>> >>>> >>>> So, for the case of drawing a hamster wearing a red hat, understanding >>>> perhaps would have taken place if the following happened before that: >>>> >>>> >>>> >>>> 1) first, the network learned about hamsters (not many examples) >>>> >>>> 2) after that the network learned about red hats (outside the context >>>> of hamsters and without many examples) >>>> >>>> 3) finally the network learned about drawing (outside of the context of >>>> hats and hamsters, not many examples) >>>> >>>> >>>> >>>> After that, the network is asked to draw a hamster with a red hat. If >>>> it does it successfully, maybe we have started cracking the problem of >>>> understanding. >>>> >>>> >>>> >>>> Note also that this requires the network to learn sequentially without >>>> exhibiting catastrophic forgetting of the previous knowledge, which is >>>> possibly also a consequence of human learning by understanding. >>>> >>>> >>>> >>>> >>>> >>>> Danko >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> Dr. Danko Nikoli? >>>> www.danko-nikolic.com >>>> >>>> https://www.linkedin.com/in/danko-nikolic/ >>>> >>>> >>>> --- A progress usually starts with an insight --- >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> Virus-free. www.avast.com >>>> >>>> >>>> >>>> >>>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>>> >>>> Without getting into the specific dispute between Gary and Geoff, I >>>> think with approaches similar to GLOM, we are finally headed in the right >>>> direction. There?s plenty of neurophysiological evidence for single-cell >>>> abstractions and multisensory neurons in the brain, which one might claim >>>> correspond to symbols. And I think we can finally reconcile the decades old >>>> dispute between Symbolic AI and Connectionism. >>>> >>>> >>>> >>>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >>>> an effort to wind up with encodings that effectively serve as symbols in >>>> exactly that way, guaranteed to serve as consistent representations of >>>> specific concepts.) >>>> >>>> GARY: I have *never* called for dismissal of neural networks, but >>>> rather for some hybrid between the two (as you yourself contemplated in >>>> 1991); the point of the 2001 book was to characterize exactly where >>>> multilayer perceptrons succeeded and broke down, and where symbols could >>>> complement them. >>>> >>>> >>>> >>>> Asim Roy >>>> >>>> Professor, Information Systems >>>> >>>> Arizona State University >>>> >>>> Lifeboat Foundation Bios: Professor Asim Roy >>>> >>>> >>>> Asim Roy | iSearch (asu.edu) >>>> >>>> >>>> >>>> >>>> >>>> >>>> *From:* Connectionists *On >>>> Behalf Of *Gary Marcus >>>> *Sent:* Wednesday, February 2, 2022 1:26 PM >>>> *To:* Geoffrey Hinton >>>> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>> Geoff Hinton >>>> >>>> >>>> >>>> Dear Geoff, and interested others, >>>> >>>> >>>> >>>> What, for example, would you make of a system that often drew the >>>> red-hatted hamster you requested, and perhaps a fifth of the time gave you >>>> utter nonsense? Or say one that you trained to create birds but sometimes >>>> output stuff like this: >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> One could >>>> >>>> >>>> >>>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>>> >>>> or >>>> >>>> b. wonder if it might be possible that sometimes the system gets the >>>> right answer for the wrong reasons (eg partial historical contingency), and >>>> wonder whether another approach might be indicated. >>>> >>>> >>>> >>>> Benchmarks are harder than they look; most of the field has come to >>>> recognize that. The Turing Test has turned out to be a lousy measure of >>>> intelligence, easily gamed. It has turned out empirically that the Winograd >>>> Schema Challenge did not measure common sense as well as Hector might have >>>> thought. (As it happens, I am a minor coauthor of a very recent review on >>>> this very topic: https://arxiv.org/abs/2201.02387 >>>> ) >>>> But its conquest in no way means machines now have common sense; many >>>> people from many different perspectives recognize that (including, e.g., >>>> Yann LeCun, who generally tends to be more aligned with you than with me). >>>> >>>> >>>> >>>> So: on the goalpost of the Winograd schema, I was wrong, and you can >>>> quote me; but what you said about me and machine translation remains your >>>> invention, and it is inexcusable that you simply ignored my 2019 >>>> clarification. On the essential goal of trying to reach meaning and >>>> understanding, I remain unmoved; the problem remains unsolved. >>>> >>>> >>>> >>>> All of the problems LLMs have with coherence, reliability, >>>> truthfulness, misinformation, etc stand witness to that fact. (Their >>>> persistent inability to filter out toxic and insulting remarks stems from >>>> the same.) I am hardly the only person in the field to see that progress on >>>> any given benchmark does not inherently mean that the deep underlying >>>> problems have solved. You, yourself, in fact, have occasionally made that >>>> point. >>>> >>>> >>>> >>>> With respect to embeddings: Embeddings are very good for natural >>>> language *processing*; but NLP is not the same as NL*U* ? when it >>>> comes to *understanding*, their worth is still an open question. >>>> Perhaps they will turn out to be necessary; they clearly aren?t sufficient. >>>> In their extreme, they might even collapse into being symbols, in the sense >>>> of uniquely identifiable encodings, akin to the ASCII code, in which a >>>> specific set of numbers stands for a specific word or concept. (Wouldn?t >>>> that be ironic?) >>>> >>>> >>>> >>>> (Your GLOM, which as you know I praised publicly, is in many ways an >>>> effort to wind up with encodings that effectively serve as symbols in >>>> exactly that way, guaranteed to serve as consistent representations of >>>> specific concepts.) >>>> >>>> >>>> >>>> Notably absent from your email is any kind of apology for >>>> misrepresenting my position. It?s fine to say that ?many people thirty >>>> years ago once thought X? and another to say ?Gary Marcus said X in 2015?, >>>> when I didn?t. I have consistently felt throughout our interactions that >>>> you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) >>>> apologized to me for having made that error. I am still not he. >>>> >>>> >>>> >>>> Which maybe connects to the last point; if you read my work, you would >>>> see thirty years of arguments *for* neural networks, just not in the >>>> way that you want them to exist. I have ALWAYS argued that there is a role >>>> for them; characterizing me as a person ?strongly opposed to neural >>>> networks? misses the whole point of my 2001 book, which was subtitled >>>> ?Integrating Connectionism and Cognitive Science.? >>>> >>>> >>>> >>>> In the last two decades or so you have insisted (for reasons you have >>>> never fully clarified, so far as I know) on abandoning symbol-manipulation, >>>> but the reverse is not the case: I have *never* called for dismissal >>>> of neural networks, but rather for some hybrid between the two (as you >>>> yourself contemplated in 1991); the point of the 2001 book was to >>>> characterize exactly where multilayer perceptrons succeeded and broke down, >>>> and where symbols could complement them. It?s a rhetorical trick (which is >>>> what the previous thread was about) to pretend otherwise. >>>> >>>> >>>> >>>> Gary >>>> >>>> >>>> >>>> >>>> >>>> On Feb 2, 2022, at 11:22, Geoffrey Hinton >>>> wrote: >>>> >>>> ? >>>> >>>> Embeddings are just vectors of soft feature detectors and they are very >>>> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >>>> opposite. >>>> >>>> >>>> >>>> A few decades ago, everyone I knew then would have agreed that the >>>> ability to translate a sentence into many different languages was strong >>>> evidence that you understood it. >>>> >>>> >>>> >>>> But once neural networks could do that, their critics moved the >>>> goalposts. An exception is Hector Levesque who defined the goalposts more >>>> sharply by saying that the ability to get pronoun references correct in >>>> Winograd sentences is a crucial test. Neural nets are improving at that but >>>> still have some way to go. Will Gary agree that when they can get pronoun >>>> references correct in Winograd sentences they really do understand? Or does >>>> he want to reserve the right to weasel out of that too? >>>> >>>> >>>> >>>> Some people, like Gary, appear to be strongly opposed to neural >>>> networks because they do not fit their preconceived notions of how the mind >>>> should work. >>>> >>>> I believe that any reasonable person would admit that if you ask a >>>> neural net to draw a picture of a hamster wearing a red hat and it draws >>>> such a picture, it understood the request. >>>> >>>> >>>> >>>> Geoff >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>>> >>>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger >>>> neural network community, >>>> >>>> >>>> >>>> There has been a lot of recent discussion on this list about framing >>>> and scientific integrity. Often the first step in restructuring narratives >>>> is to bully and dehumanize critics. The second is to misrepresent their >>>> position. People in positions of power are sometimes tempted to do this. >>>> >>>> >>>> >>>> The Hinton-Hanson interview that you just published is a real-time >>>> example of just that. It opens with a needless and largely content-free >>>> personal attack on a single scholar (me), with the explicit intention of >>>> discrediting that person. Worse, the only substantive thing it says is >>>> false. >>>> >>>> >>>> >>>> Hinton says ?In 2015 he [Marcus] made a prediction that computers >>>> wouldn?t be able to do machine translation.? >>>> >>>> >>>> >>>> I never said any such thing. >>>> >>>> >>>> >>>> What I predicted, rather, was that multilayer perceptrons, as they >>>> existed then, would not (on their own, absent other mechanisms) >>>> *understand* language. Seven years later, they still haven?t, except >>>> in the most superficial way. >>>> >>>> >>>> >>>> I made no comment whatsoever about machine translation, which I view as >>>> a separate problem, solvable to a certain degree by correspondance without >>>> semantics. >>>> >>>> >>>> >>>> I specifically tried to clarify Hinton?s confusion in 2019, but, >>>> disappointingly, he has continued to purvey misinformation despite that >>>> clarification. Here is what I wrote privately to him then, which should >>>> have put the matter to rest: >>>> >>>> >>>> >>>> You have taken a single out of context quote [from 2015] and >>>> misrepresented it. The quote, which you have prominently displayed at the >>>> bottom on your own web page, says: >>>> >>>> >>>> >>>> Hierarchies of features are less suited to challenges such as language, >>>> inference, and high-level planning. For example, as Noam Chomsky famously >>>> pointed out, language is filled with sentences you haven't seen >>>> before. Pure classifier systems don't know what to do with such sentences. >>>> The talent of feature detectors -- in identifying which member of some >>>> category something belongs to -- doesn't translate into understanding >>>> novel sentences, in which each sentence has its own unique meaning. >>>> >>>> >>>> >>>> It does *not* say "neural nets would not be able to deal with novel >>>> sentences"; it says that hierachies of features detectors (on their own, if >>>> you read the context of the essay) would have trouble *understanding *novel sentences. >>>> >>>> >>>> >>>> >>>> Google Translate does yet not *understand* the content of the >>>> sentences is translates. It cannot reliably answer questions about who did >>>> what to whom, or why, it cannot infer the order of the events in >>>> paragraphs, it can't determine the internal consistency of those events, >>>> and so forth. >>>> >>>> >>>> >>>> Since then, a number of scholars, such as the the computational >>>> linguist Emily Bender, have made similar points, and indeed current LLM >>>> difficulties with misinformation, incoherence and fabrication all follow >>>> from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on >>>> the matter with Alexander Koller, >>>> https://aclanthology.org/2020.acl-main.463.pdf >>>> , >>>> also emphasizing issues of understanding and meaning: >>>> >>>> >>>> >>>> *The success of the large neural language models on many NLP tasks is >>>> exciting. However, we find that these successes sometimes lead to hype in >>>> which these models are being described as ?understanding? language or >>>> capturing ?meaning?. In this position paper, we argue that a system trained >>>> only on form has a priori no way to learn meaning. .. a clear understanding >>>> of the distinction between form and meaning will help guide the field >>>> towards better science around natural language understanding. * >>>> >>>> >>>> >>>> Her later article with Gebru on language models ?stochastic parrots? is >>>> in some ways an extension of this point; machine translation requires >>>> mimicry, true understanding (which is what I was discussing in 2015) >>>> requires something deeper than that. >>>> >>>> >>>> >>>> Hinton?s intellectual error here is in equating machine translation >>>> with the deeper comprehension that robust natural language understanding >>>> will require; as Bender and Koller observed, the two appear not to be the >>>> same. (There is a longer discussion of the relation between language >>>> understanding and machine translation, and why the latter has turned out to >>>> be more approachable than the former, in my 2019 book with Ernest Davis). >>>> >>>> >>>> >>>> More broadly, Hinton?s ongoing dismissiveness of research from >>>> perspectives other than his own (e.g. linguistics) have done the field a >>>> disservice. >>>> >>>> >>>> >>>> As Herb Simon once observed, science does not have to be zero-sum. >>>> >>>> >>>> >>>> Sincerely, >>>> >>>> Gary Marcus >>>> >>>> Professor Emeritus >>>> >>>> New York University >>>> >>>> >>>> >>>> On Feb 2, 2022, at 06:12, AIhub wrote: >>>> >>>> ? >>>> >>>> Stephen Hanson in conversation with Geoff Hinton >>>> >>>> >>>> >>>> In the latest episode of this video series for AIhub.org >>>> , >>>> Stephen Hanson talks to Geoff Hinton about neural networks, >>>> backpropagation, overparameterization, digit recognition, voxel cells, >>>> syntax and semantics, Winograd sentences, and more. >>>> >>>> >>>> >>>> You can watch the discussion, and read the transcript, here: >>>> >>>> >>>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>>> >>>> >>>> >>>> >>>> About AIhub: >>>> >>>> AIhub is a non-profit dedicated to connecting the AI community to the >>>> public by providing free, high-quality information through AIhub.org >>>> >>>> (https://aihub.org/ >>>> ). >>>> We help researchers publish the latest AI news, summaries of their work, >>>> opinion pieces, tutorials and more. We are supported by many leading >>>> scientific organizations in AI, namely AAAI >>>> , >>>> NeurIPS >>>> , >>>> ICML >>>> , >>>> AIJ >>>> >>>> /IJCAI >>>> , >>>> ACM SIGAI >>>> , >>>> EurAI/AICOMM, CLAIRE >>>> >>>> and RoboCup >>>> >>>> . >>>> >>>> Twitter: @aihuborg >>>> >>>> >>>> >>>> >>>> >>>> >>>> Virus-free. www.avast.com >>>> >>>> >>>> >>>> >>>> -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Tue Feb 8 10:42:15 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Tue, 8 Feb 2022 07:42:15 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <28F54308-3E98-42CA-A709-1F33F4FBDE00@nyu.edu> Hi, Ali, Just to be clear, I certainly think that animal models can be useful in principle, and eg. on motor-control and low-level vision they have been incredibly useful. On compositionality, they may turn out to be less so, at least in the near-term. My 2004 book The Birth of the Mind has a chapter on evolution where I try to reconcile what we know about the astonishing conservation across the biological world with the apparent uniqueness of human language, with a focus on how something that is apparently unique can rest partly but not entirely on inherited substrates. Emphasizing the biological process of duplication and divergence, I suggested that language might have hinged on multiple simultaneous mutations to existing cognitively-relevant genes that gave rise to advances in the capacity for hierarchical representation. Tecumseh Fitch has recently made a somewhat parallel argument, also suggesting that compositionality in strong form may be humanly unique, but arguing instead that a large expansion in Broca?s area is responsible. If either his speculation or mine were correct, animal models might be of relatively little direct help. (Though of course anything that helps us understand the brain better may be of substantial indirect help.) On your other point, saying that ?symbols emerge naturally from the physics of the brains [of the animals that have them]? just doesn?t tell us much; everything any animal does emerges from physics and biology and experience. But nature has a vast array of different solutions to a wide variety of problems, and it?s an empirical matter to understand each creature and how it handles different problems (digestion, circulation, navigation, planning, etc). All that is of course constrained by physics, but it doesn't tell us eg whether any specific aspect of a particular creature?s physiology ?emerges? or is hard-coded, learned, etc. (Empirically, my own experimental work with babies, since replicated by others with newborns, suggests that the capacity to acquire novel, rule-like abstractions emerges from a developmental program that does not depend on post-natal experience.) More broadly, I think that most brains on this planet have some capacity for symbol-manipulation (eg honey bees calculating the solar azimuth function and extrapolating to novel lighting conditions) but that compositionality is less prevalent. I don?t think that you have to have symbols or compositionality to have a hint of intelligence, but I do think you probably need them for AGI or for understanding how humans tick. Gary > On Feb 7, 2022, at 23:22, Ali Minai wrote: > > ? > Hi Gary > > Thanks for your reply. I'll think more about your points. I do think that, to understand the human mind, we should start with vertebrates, which is why I suggested fish. At least for the motor system - which is part of the mind - we have learned a lot from lampreys (e.g. Sten Grillner's work and that beautiful lamprey-salamander model by Ijspeert et al.), and it has taught us a lot about locomotion in other animals, including mammals. The principles clearly generalize, though the complexity increases a lot. Insects too are very interesting. After all, they are our ancestors too. > > I don't agree that we can posit a clear transition from deep cognitive models in humans to none below that in the phylogenetic tree. Chimpanzees and macaques certainly show some evidence, and there's no reason to think that it's a step change rather than a highly nonlinear continuum. And even though what we might (simplistically) call System 2 aspects of cognition are minimally present in other mammals, their precursors must be. > > My point about cats and symbols was not regarding whether cats are aware of symbols, but that symbols emerge naturally from the physics of their brains. Behaviors that require some small degree of symbolic processing exist in mammals other than humans (e.g., transitive inference and landmark-based navigation in rats), and it is seen better as an emergent property of brains than an attribute to be explicitly built-into neural models by us. Once we have a sufficiently brain-like neural model, symbolic processing will already be there. > > I agree with you completely that we are far from understanding some of the most fundamental principles of the brain, but even more importantly, we are not even looking in the right direction. I'm hoping to lay out my arguments about all this in more detail in some other form. > > Best > Ali > > > PS: I had inadvertently posted my reply of Gary's message only to him. Should have posted to everyone, so here it is. > > > Ali A. Minai, Ph.D. > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > >> On Mon, Feb 7, 2022 at 12:28 AM Gary Marcus wrote: >> Ali, >> >> It?s useful to think about animals, but I really wouldn?t start with fish; it?s not clear that their ecological niche demands anything significant in the way of extrapolation, causal reasoning, or compositionality. There is good evidence elsewhere in the animal world for extrapolation of functions that may be innate (eg solar azimuth in bees), and causal reasoning (eg tool use in ravens, various primates, and octopi). It?s still not clear to me how much hierarchical representation (critical to AGI) exists outside of humans, though; the ability to construct rich new cognitive models may also be unique to us. >> >> In any case it matters not in the least whether the average cat or human cares about symbols, anymore that it matters whether the average animal understands digestion; only a tiny fraction of the creatures on this planet have any real understanding of their internal workings. >> >> My overall feeling is that we are a really, really long way from understanding the neural basis of higher-level cognition, and that AI is going to need muddle through on its own, for another decade or two, >> >> I do fully agree with your conclusion, though, that "AI today is driven more by habit and the incentives of the academic and corporate marketplaces than by a deep, long-term view of AI as a great exploratory project in fundamental science." Let's hope that changes. >> >> Gary >> >>>> On Feb 6, 2022, at 13:19, Ali Minai wrote: >>>> >>> ? >>> Gary, >>> >>> That?s a very interesting and accurate list of capabilities that a general intelligent system must have and that our AI does not. Of course, the list is familiar to me from having read your book. However, I have a somewhat different take on this whole thing. >>> >>> All the things we discuss here ? symbols/no symbols, parts/wholes, supervised/unsupervised, token/type, etc., are useful categories and distinctions for our analysis of the problem, and are partly a result of the historical evolution of the field of AI in particular and of philosophy in general. The categories are not wrong in any way, of course, but they are posterior to the actual system ? good for describing and analyzing it, and for validating our versions of it (which is how you use them). I think they are less useful as prescriptions for how to build our AI systems. If intelligent systems did not already exist and we were building them from scratch (please ignore the impossibility of that), having a list of ?must haves? would be great. But intelligent systems already exist ? from humans to fish ? and they already have these capacities to a greater or lesser degree because of the physics of their biology. A cat?s intelligence does not care whether it has symbols or not, and nor does mine or yours. Whatever we describe as symbolic processing post-facto has already been done by brains for at least tens of millions of years. Instead of getting caught up in ?how to add symbols into our neural models?, we should be investigating how what we see as symbolic processing emerges from animal brains, and then replicate those brains to the degree necessary. If we can do that, symbolic processing will already be present. But it cannot be done piece by piece. It must take the integrity of the whole brain and the body it is part of, and its environment, into account. That?s why I think that a much better ? though a very long ? route to AI is to start by understanding how a fish brain makes the intelligence of a fish possible, and then boot up our knowledge across phylogenetic stages: Bottom up reverse engineering rather than top-down engineering. That?s the way Nature built up to human intelligence, and we will succeed only by reverse engineering it. Of course, we can do it much faster and with shortcuts because we are intelligent, purposive agents, but working top-down by building piecewise systems that satisfy a list of attributes will not get us there. Among other things, those pieces will be impossible to integrate into the kind of intelligence that can have those general models of the world that you rightly point to as being necessary. >>> >>> I think that one thing that has been a great boon to the AI enterprise has also been one of the greatest impediments to its complete success, and that is the ?computationalization? of intelligence. On the one hand, thinking of intelligence computationally allows us to describe it abstractly and in a principled, formal way. It also resonates with the fact that we are trying to implement intelligence through computational machines. But, on the flip side, this view of intelligence divorces it from its physics ? from the fact that real intelligence in animals emerges from the physics of the physical system. That system is not a collection of its capabilities; rather, those capabilities are immanent in it by virtue of its physics. When we try to build those capabilities computationally, i.e., through code, we are making the same error that the practitioners of old-style ?symbolic AI? made ? what I call the ?professors are smarter than Nature? error, i.e., the idea that we are going to enumerate (or describe) all the things that underlie intelligence and implement them one by one until we get complete intelligence. We will never be able to enumerate all those capabilities, and will never be able to get to that complete intelligence. The only difference between us and the ?symbolists? of yore is that we are replacing giant LISP and Prolog programs with giant neural networks. Otherwise, we are using our models exactly as they were trying to use their models, and we will fail just as they did unless we get back to biology and the real thing. >>> >>> I will say again that the way we do AI today is driven more by habit and the incentives of the academic and corporate marketplaces than by a deep, long-term view of AI as a great exploratory project in fundamental science. We are just building AI to drive our cars, translate our documents, write our reports, and do our shopping. What that will teach us about actual intelligence is just incidental. >>> >>> My apologies too for a long response. >>> >>> Ali >>> >>> Ali A. Minai, Ph.D. >>> Professor and Graduate Program Director >>> Complex Adaptive Systems Lab >>> Department of Electrical Engineering & Computer Science >>> 828 Rhodes Hall >>> University of Cincinnati >>> Cincinnati, OH 45221-0030 >>> >>> Phone: (513) 556-4783 >>> Fax: (513) 556-7326 >>> Email: Ali.Minai at uc.edu >>> minaiaa at gmail.com >>> >>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>> >>> >>>> On Sun, Feb 6, 2022 at 9:42 AM Gary Marcus wrote: >>>> Dear Asim, >>>> >>>> Sorry for a long answer to your short but rich questions. >>>> Yes, memory in my view has to be part of the answer to the type-token problem. Symbol systems encoded in memory allow a natural way to set up records, and something akin to that seems necessary. Pure multilayer perceptrons struggle with type-token distinctions precisely because they lack such records. On the positive side, I see more and more movement towards recordlike stores (eg w key-value stores in memory networks), and I think that is an important and necessary step, very familiar from the symbol-manipulating playbook, sometimes implemented in new ways. >>>> But ultimately, handling the type-token distinction requires considerable inferential overhead beyond the memory representation of a record per se. How do you determine when to denote something (e.g. Felix) as an instance, and of which kinds (cat, animal etc), and how do you leverage that knowledge once you determine it? >>>> In the limit we reason about types vs tokens in fairly subtle ways, eg in guessing whether a glass that we put down at party is likely to be ours. The reverse is also important: we need to be learn particular traits for individuals and not erroneously generalize them to the class; if my aunt Esther wins the lottery, one shouldn?t infer that all of my aunts or all of my relatives or adult females have won the lottery. so you need both representational machinery that can distinguish eg my cat from cats in general and reasoning machinery to decide at what level certain learned knowledge should inhere. (I had a whole chapter about this sort of thing in The Algebraic Mind if you are interested, and Mike Mozer had a book about types and tokens in neural networks in the mid 1990s). >>>> Yes, part (though not all!) of what we do when we set up cognitive models in our heads is to track particular individuals and their properties. If you only had to correlate kinds (cats) and their properties (have fur) you could maybe get away with a multilayer perceptron, but once you need to track individuals, yes, you really need some kind of memory-based records. >>>> As far as I can tell, Transformers can sometimes approximate some of this for a few sentences, but not over long stretches. >>>> >>>> As a small terminological aside; for me cognitive models ? cognitive modeling. Cognitive modeling is about building psychological or computational models of how people think, whereas what I mean by a cognitive model is a representation of eg the entities in some situation and the relations between those entities. >>>> >>>> To your closing question, none of us yet really knows how to build understanding into machines. A solid type-token distinction, both in terms of representation and reasoning, is critical for general intelligence, but hardly sufficient. Personally, I think some minimal prerequisites would be: >>>> representations of space, time, causality, individuals, kinds, persons, places, objects, etc. >>>> representations of abstractions that can hold over all entities in a class >>>> compositionality (if we are talking about human-like understanding) >>>> capacity to construct and update cognitive models on the fly >>>> capacity to reason over entities in those models >>>> ability to learn about new entities and their properties >>>> Much of my last book (Rebooting AI, w Ernie Davis) is about the above list. The section in the language chapter on a children?s story in which man has lost is wallet is an especially vivid worked example. Later chapters elaborate some of the challenges in representing space, time, and causality. >>>> >>>> Gary >>>> >>>> >>>>>> On Feb 5, 2022, at 18:58, Asim Roy wrote: >>>>>> >>>>> ? >>>>> Gary, >>>>> >>>>> >>>>> >>>>> I don?t get much into the type of cognitive modeling you are talking about, but I would guess that the type problem can generally be handled by neural network models and tokens can be resolved with some memory-based system. But to the heart of the question, this is what so-called ?understanding? reduces to computation wise? >>>>> >>>>> >>>>> >>>>> Asim >>>>> >>>>> >>>>> >>>>> From: Gary Marcus >>>>> Sent: Saturday, February 5, 2022 8:39 AM >>>>> To: Asim Roy >>>>> Cc: Ali Minai ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub >>>>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>>> >>>>> >>>>> >>>>> There is no magic in understanding, just computation that has been realized in the wetware of humans and that eventually can be realized in machines. But understanding is not (just) learning. >>>>> >>>>> >>>>> >>>>> Understanding incorporates (or works in tandem with) learning - but also, critically, in tandem with inference, and the development and maintenance of cognitive models. Part of developing an understanding of cats in general is to learn long term-knowledge about their properties, both directly (e.g., through observation) and indirectly (eg through learning facts about animals in general that can be extended to cats), often through inference (if all animals have DNA, and a cat is an animal, it must also have DNA). The understanding of a particular cat also involves direct observation, but also inference (eg one might surmise that the reason that Fluffy is running about the room is that Fluffy suspects there is a mouse stirring somewhere nearby). But all of that, I would say, is subservient to the construction of cognitive models that can be routinely updated (e.g., Fluffy is currently in the living room, skittering about, perhaps looking for a mouse). >>>>> >>>>> In humans, those dynamic, relational models, which form part of an understanding, can support inference (if Fluffy is in the living room, we can infer that Fluffy is not outside, not lost, etc). Without such models - which I think represent a core part of understanding - AGI is an unlikely prospect. >>>>> >>>>> Current neural networks, as it happens, are better at acquiring long-term knowledge (cats have whiskers) than they are at dynamically updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic model that I am describing. To a modest degree they can approximate it on the basis of large samples of texts, but their ultimate incoherence stems from the fact that they do not have robust internal cognitive models that they can update on the fly. >>>>> >>>>> Without such cognitive models you can still capture some aspects of understanding (eg predicting that cats are likely to be furry), but things fall apart quickly; inference is never reliable, and coherence is fleeting. >>>>> >>>>> As a final note, one of the most foundational challenges in constructing adequate cognitive models of the world is to have a clear distinction between individuals and kinds; as I emphasized 20 years ago (in The Algebraic Mind), this has always been a weakness in neural networks, and I don?t think that the type-token problem has yet been solved. >>>>> >>>>> >>>>> Gary >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> On Feb 5, 2022, at 01:31, Asim Roy wrote: >>>>> >>>>> ? >>>>> >>>>> All, >>>>> >>>>> >>>>> >>>>> I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? >>>>> >>>>> >>>>> >>>>> https://www.youtube.com/watch?v=gn4nRCC9TwQ >>>>> >>>>> https://www.youtube.com/watch?v=8sO7VS3q8d0 >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> Asim Roy >>>>> >>>>> Professor, Information Systems >>>>> >>>>> Arizona State University >>>>> >>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>> >>>>> Asim Roy | iSearch (asu.edu) >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> From: Ali Minai >>>>> Sent: Friday, February 4, 2022 11:38 PM >>>>> To: Asim Roy >>>>> Cc: Gary Marcus ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub >>>>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>>> >>>>> >>>>> >>>>> Asim >>>>> >>>>> >>>>> >>>>> Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. >>>>> >>>>> >>>>> >>>>> FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! >>>>> >>>>> >>>>> >>>>> Ali >>>>> >>>>> >>>>> >>>>> Ali A. Minai, Ph.D. >>>>> Professor and Graduate Program Director >>>>> Complex Adaptive Systems Lab >>>>> Department of Electrical Engineering & Computer Science >>>>> >>>>> 828 Rhodes Hall >>>>> >>>>> University of Cincinnati >>>>> Cincinnati, OH 45221-0030 >>>>> >>>>> >>>>> Phone: (513) 556-4783 >>>>> Fax: (513) 556-7326 >>>>> Email: Ali.Minai at uc.edu >>>>> minaiaa at gmail.com >>>>> >>>>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: >>>>> >>>>> First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. >>>>> >>>>> >>>>> >>>>> DANKO: >>>>> >>>>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>>>> >>>>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>>>> >>>>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>>>> >>>>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>>>> >>>>> >>>>> >>>>> Asim Roy >>>>> >>>>> Professor, Information Systems >>>>> >>>>> Arizona State University >>>>> >>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>> >>>>> Asim Roy | iSearch (asu.edu) >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> From: Gary Marcus >>>>> Sent: Thursday, February 3, 2022 9:26 AM >>>>> To: Danko Nikolic >>>>> Cc: Asim Roy ; Geoffrey Hinton ; AIhub ; connectionists at mailman.srv.cs.cmu.edu >>>>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>>> >>>>> >>>>> >>>>> Dear Danko, >>>>> >>>>> >>>>> >>>>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. >>>>> >>>>> >>>>> >>>>> I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. >>>>> >>>>> >>>>> >>>>> Gary >>>>> >>>>> >>>>> >>>>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: >>>>> >>>>> >>>>> >>>>> G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." >>>>> >>>>> >>>>> >>>>> I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". >>>>> >>>>> To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. >>>>> >>>>> >>>>> >>>>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>>>> >>>>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>>>> >>>>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>>>> >>>>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>>>> >>>>> >>>>> >>>>> ...and so on. >>>>> >>>>> >>>>> >>>>> Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). >>>>> >>>>> >>>>> >>>>> I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. >>>>> >>>>> >>>>> >>>>> In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. >>>>> >>>>> >>>>> >>>>> No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. >>>>> >>>>> >>>>> >>>>> I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. >>>>> >>>>> >>>>> >>>>> So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: >>>>> >>>>> >>>>> >>>>> 1) first, the network learned about hamsters (not many examples) >>>>> >>>>> 2) after that the network learned about red hats (outside the context of hamsters and without many examples) >>>>> >>>>> 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) >>>>> >>>>> >>>>> >>>>> After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. >>>>> >>>>> >>>>> >>>>> Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> Danko >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> Dr. Danko Nikoli? >>>>> www.danko-nikolic.com >>>>> https://www.linkedin.com/in/danko-nikolic/ >>>>> >>>>> --- A progress usually starts with an insight --- >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> Virus-free. www.avast.com >>>>> >>>>> >>>>> >>>>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>>>> >>>>> Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. >>>>> >>>>> >>>>> >>>>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>>>> >>>>> GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. >>>>> >>>>> >>>>> >>>>> Asim Roy >>>>> >>>>> Professor, Information Systems >>>>> >>>>> Arizona State University >>>>> >>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>> >>>>> Asim Roy | iSearch (asu.edu) >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> From: Connectionists On Behalf Of Gary Marcus >>>>> Sent: Wednesday, February 2, 2022 1:26 PM >>>>> To: Geoffrey Hinton >>>>> Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu >>>>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>>> >>>>> >>>>> >>>>> Dear Geoff, and interested others, >>>>> >>>>> >>>>> >>>>> What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> One could >>>>> >>>>> >>>>> >>>>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>>>> >>>>> or >>>>> >>>>> b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. >>>>> >>>>> >>>>> >>>>> Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). >>>>> >>>>> >>>>> >>>>> So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. >>>>> >>>>> >>>>> >>>>> All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. >>>>> >>>>> >>>>> >>>>> With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) >>>>> >>>>> >>>>> >>>>> (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>>>> >>>>> >>>>> >>>>> Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. >>>>> >>>>> >>>>> >>>>> Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? >>>>> >>>>> >>>>> >>>>> In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. >>>>> >>>>> >>>>> >>>>> Gary >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: >>>>> >>>>> ? >>>>> >>>>> Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. >>>>> >>>>> >>>>> >>>>> A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. >>>>> >>>>> >>>>> >>>>> But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? >>>>> >>>>> >>>>> >>>>> Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. >>>>> >>>>> I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. >>>>> >>>>> >>>>> >>>>> Geoff >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>>>> >>>>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, >>>>> >>>>> >>>>> >>>>> There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. >>>>> >>>>> >>>>> >>>>> The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. >>>>> >>>>> >>>>> >>>>> Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? >>>>> >>>>> >>>>> >>>>> I never said any such thing. >>>>> >>>>> >>>>> >>>>> What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. >>>>> >>>>> >>>>> >>>>> I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. >>>>> >>>>> >>>>> >>>>> I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: >>>>> >>>>> >>>>> >>>>> You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: >>>>> >>>>> >>>>> >>>>> Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. >>>>> >>>>> >>>>> >>>>> It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. >>>>> >>>>> >>>>> >>>>> Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. >>>>> >>>>> >>>>> >>>>> Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: >>>>> >>>>> >>>>> >>>>> The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. >>>>> >>>>> >>>>> >>>>> Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. >>>>> >>>>> >>>>> >>>>> Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). >>>>> >>>>> >>>>> >>>>> More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. >>>>> >>>>> >>>>> >>>>> As Herb Simon once observed, science does not have to be zero-sum. >>>>> >>>>> >>>>> >>>>> Sincerely, >>>>> >>>>> Gary Marcus >>>>> >>>>> Professor Emeritus >>>>> >>>>> New York University >>>>> >>>>> >>>>> >>>>> On Feb 2, 2022, at 06:12, AIhub wrote: >>>>> >>>>> ? >>>>> >>>>> Stephen Hanson in conversation with Geoff Hinton >>>>> >>>>> >>>>> >>>>> In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. >>>>> >>>>> >>>>> >>>>> You can watch the discussion, and read the transcript, here: >>>>> >>>>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>>>> >>>>> >>>>> >>>>> About AIhub: >>>>> >>>>> AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. >>>>> >>>>> Twitter: @aihuborg >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> Virus-free. www.avast.com >>>>> >>>>> -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Tue Feb 8 10:56:42 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Tue, 8 Feb 2022 07:56:42 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <2AFB6BFA-3340-47F3-A30C-84172C3877F5@nyu.edu> when we really full understand everything that individual neurons can do computationally, we will look back at the current era and shake our heads. > On Feb 7, 2022, at 16:39, Juyang Weng wrote: > > ? > Dear Gary, > Thank you for the link to Geoff's GLOM paper. I quickly browsed it just now. > Some fundamental comments, not all, to be concise. > (1) You are right, Geoff's GLOM seems to be a symbolic network, which assigns a certain role to each group of neurons. > (2) Geoff's part-whole problem is too narrow, a dead end to solving his part-whole problem. I quote: > "Perhaps we can learn how to encode the information at each location in such a way that simply averaging > the encodings from different locations is the only form of interaction we need." He got stuck with "locations". > Like Convolution that I used in Cresceptron 1992 and Geoff also used much longer, as soon a feature representation is centered as a location, the system does not abstract as Michaal Jordan complained at an IJCNN conference. Michael Jordan did not sa=y what he meant by "does not abstract well", but his point is valid. > (3) Two feed-forward networks in Geoff's GLOM, one bottom-up and the other top-down, are efficient pattern recognizers that do not abstract. The brain is not just a pattern recognizer. > (4) It is very unfortunate that many neural network researchers including Alpha's DeepMinds have not dug deep into what a cell can do and what a cell cannot. Geoff's GLOM is an example. > I have a paper about a brain model and I sent it to some people to pre-review. But like my Conscious Learning paper that was rejected by ICDL 2021 and AAAI 2022, this brain model would be rejected too. > Your humbly, > -John > >> On Mon, Feb 7, 2022 at 10:57 AM Gary Marcus wrote: >> Dear John, >> >> I agree with you that cluttered scenes are critical, but Geoff?s GLOM paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf] might actually have some relevance. It may well be that we need to do a better job with parts and whole before we can fully address clutter, and Geoff is certainly taking that question seriously. >> >> Geoff?s ?Stable islands of identical vectors? do sound suspiciously like symbols to me (in a good way!), but regardless, they seem to me to be a plausible candidate as a foundation for coping with clutter. >> >> And not just cluttered scenes, but also relations between multiple objects in a scene, which is another example of the broader issue you raise, challenging for pure MLPs but critical for deeper AI. >> >> Gary >> >>>> On Feb 7, 2022, at 00:23, Juyang Weng wrote: >>>> >>> ? >>> Dear Geoff Hinton, >>> I respect that you have been working on pattern recognition on isolated characters using neural networks. >>> >>> However, I am deeply disappointed that after receiving the Turing Award 2018, you are still falling behind your own award work by talking about "how you >>> recognize that a handwritten 2 is a 2." You have fallen behind our group's >>> Creceptron work in 1992, let alone our group's work on 3D-to-2D-to-3D Conscious Learning using DNs. Both deal with cluttered scenes. >>> >>> Specifically, you will never be able to get a correct causal explanation by looking at a single hand-written 2. Your problem is too small to explain a brain network. You must look at cluttered sciences, with many objects. >>> >>> Yours humbly, >>> -John >>> ---- >>> Message: 7 >>> Date: Fri, 4 Feb 2022 15:24:02 -0500 >>> From: Geoffrey Hinton >>> To: "Dietterich, Thomas" >>> Cc: AIhub , >>> "connectionists at mailman.srv.cs.cmu.edu" >>> >>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff >>> Hinton >>> Message-ID: >>> >>> Content-Type: text/plain; charset="utf-8" >>> >>> I agree that it's nice to have a causal explanations. But I am not >>> convinced there will ever be a simple causal explanation for how you >>> recognize that a handwritten 2 is a 2. >>> >>> -- >>> Juyang (John) Weng > > > -- > Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Tue Feb 8 15:07:06 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Tue, 8 Feb 2022 12:07:06 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <4C768F69-06BE-48DA-AF37-4D2DBC035962@nyu.edu> Hi, Ali, I don?t actually know where or when the term ?natural stupidity? was first used, but McDermott?s points in this paper are still quite relevant, both for symbolic and neural approaches: https://cs.fit.edu/~kgallagher/Schtick/Serious/McDermott.AI.MeetsNaturalStupidity.pdf With respect to human cognitive limitations, if you make it through Marcus 2004, there is also Marcus 2008, Kluge: The Haphazard Evolution of the Human Mind to add to your list. Humans set a low bar; perhaps some day AI can exceed it. Gary > On Feb 8, 2022, at 10:19, Ali Minai wrote: > > ? > Hi Gary > > Thanks for your reply. Much to think about in it. For one, it makes me want to read your 2004 book right away (which I will). > > All human capacities - however unique they may seem to us - must necessarily evolve from those of the precursor species. If we had Australopithecus and Homo Erectus brains to study, I am sure we would find less developed examples of symbolic processing, compositionality, etc., that are recognizably of the same type as ours, and thus link our capacities with earlier ancestors that may have been closer to chimpanzees or bonobos. There's increasing reason to believe that Homo Neandertalis had a pretty complex mind, though it is on a different evolutionary branch than us (though it has contributed about 2% of our genome through cross-breeding). That suggests that the common ancestor of both H. Sapiens and H. Neandertalis also had these mental capacities to some degree. Moving backwards like that, it's hard to decide where the chain should be cut. Of course, we know that phenotypic attributes can emerge very rapidly in the evolutionary process and then be in stasis for a long time, but they do emerge from the available substrate through processes like duplication-and-divergence, recombination, diversification in neutral spaces, external pressures, etc.Understanding that process and leveraging it for engineering is, in my opinion, a more reliable route to AI. Again, I don't mean that we should replicate evolution - just that we should reverse engineer its results. > > I think compositionality is a useful concept in the analysis of language and inference, but not very useful as a prescriptive goal for building AI systems. It's an attribute we see in human language now and would like to see in our AI systems. But instead of asking "How can I get compositionality (or any attribute X) into my neural network model?" I want to ask, "How do the neural networks of the brain come to have the attribute of compositionality (or X)?" or "What is it in the anatomy and physiology of brain networks that allows for compositionality to be possible?" Answering that question will also tell us how to achieve it in our models. Otherwise, we are trying to get water from rocks. We don't know whether any of our neural models are capable of the kind of symbolic processing, compositionality, etc., that we want. We keep inventing new models to incorporate such desired attributes purely by thinking them up based on our previous models and their failures. This process is, in my opinion, a futile quest that is, at best, likely to lead to bad, inadequate models that will fall apart when we try to integrate them into larger intelligent systems. Much better, I suggest, to look at the real-world systems that exist, i.e., animal brains and bodies, and see how all the aspects of intelligence emerge from them. When we do, we will have neural models capable of symbolic processing, language, compositionality, etc. That's why I think it's better to have models that are more grounded in learning from biology with humility and less invested in top-down analysis and design based on our formalisms. > > One place where we can learn a lot about "real intelligence" is in the ways it "fails", i.e., by focusing more on what Amos Tversky flippantly called "natural stupidity". In those imperfections, we may be able to see how a real-time physical system like the brain achieves capacities such as abstract reasoning, symbolic processing, language, etc. As my old friend Daniel Polani suggested in another post in this thread (hi Daniel!), we humans do all these things quite imperfectly, and far from being a bug, that is probably a feature - an indicator of how, by sacrificing the goal of being perfect, we (i.e., evolution) have been able to become merely excellent. > > If we hope to understand intelligence, there are several inherited preferences that we need to kill. High on that kill list are the obsessions with optimality, stability, repeatability, explainability, etc., that are the antithesis of how natural living systems grow, thrive, and evolve. > > Ali > > > Ali A. Minai, Ph.D. > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > >> On Tue, Feb 8, 2022 at 10:42 AM Gary Marcus wrote: >> Hi, Ali, >> >> Just to be clear, I certainly think that animal models can be useful in principle, and eg. on motor-control and low-level vision they have been incredibly useful. On compositionality, they may turn out to be less so, at least in the near-term. My 2004 book The Birth of the Mind has a chapter on evolution where I try to reconcile what we know about the astonishing conservation across the biological world with the apparent uniqueness of human language, with a focus on how something that is apparently unique can rest partly but not entirely on inherited substrates. Emphasizing the biological process of duplication and divergence, I suggested that language might have hinged on multiple simultaneous mutations to existing cognitively-relevant genes that gave rise to advances in the capacity for hierarchical representation. Tecumseh Fitch has recently made a somewhat parallel argument, also suggesting that compositionality in strong form may be humanly unique, but arguing instead that a large expansion in Broca?s area is responsible. If either his speculation or mine were correct, animal models might be of relatively little direct help. (Though of course anything that helps us understand the brain better may be of substantial indirect help.) >> >> On your other point, saying that ?symbols emerge naturally from the physics of the brains [of the animals that have them]? just doesn?t tell us much; everything any animal does emerges from physics and biology and experience. But nature has a vast array of different solutions to a wide variety of problems, and it?s an empirical matter to understand each creature and how it handles different problems (digestion, circulation, navigation, planning, etc). All that is of course constrained by physics, but it doesn't tell us eg whether any specific aspect of a particular creature?s physiology ?emerges? or is hard-coded, learned, etc. (Empirically, my own experimental work with babies, since replicated by others with newborns, suggests that the capacity to acquire novel, rule-like abstractions emerges from a developmental program that does not depend on post-natal experience.) >> >> More broadly, I think that most brains on this planet have some capacity for symbol-manipulation (eg honey bees calculating the solar azimuth function and extrapolating to novel lighting conditions) but that compositionality is less prevalent. I don?t think that you have to have symbols or compositionality to have a hint of intelligence, but I do think you probably need them for AGI or for understanding how humans tick. >> >> Gary >> >> >> >>>> On Feb 7, 2022, at 23:22, Ali Minai wrote: >>>> >>> ? >>> Hi Gary >>> >>> Thanks for your reply. I'll think more about your points. I do think that, to understand the human mind, we should start with vertebrates, which is why I suggested fish. At least for the motor system - which is part of the mind - we have learned a lot from lampreys (e.g. Sten Grillner's work and that beautiful lamprey-salamander model by Ijspeert et al.), and it has taught us a lot about locomotion in other animals, including mammals. The principles clearly generalize, though the complexity increases a lot. Insects too are very interesting. After all, they are our ancestors too. >>> >>> I don't agree that we can posit a clear transition from deep cognitive models in humans to none below that in the phylogenetic tree. Chimpanzees and macaques certainly show some evidence, and there's no reason to think that it's a step change rather than a highly nonlinear continuum. And even though what we might (simplistically) call System 2 aspects of cognition are minimally present in other mammals, their precursors must be. >>> >>> My point about cats and symbols was not regarding whether cats are aware of symbols, but that symbols emerge naturally from the physics of their brains. Behaviors that require some small degree of symbolic processing exist in mammals other than humans (e.g., transitive inference and landmark-based navigation in rats), and it is seen better as an emergent property of brains than an attribute to be explicitly built-into neural models by us. Once we have a sufficiently brain-like neural model, symbolic processing will already be there. >>> >>> I agree with you completely that we are far from understanding some of the most fundamental principles of the brain, but even more importantly, we are not even looking in the right direction. I'm hoping to lay out my arguments about all this in more detail in some other form. >>> >>> Best >>> Ali >>> >>> >>> PS: I had inadvertently posted my reply of Gary's message only to him. Should have posted to everyone, so here it is. >>> >>> >>> Ali A. Minai, Ph.D. >>> Professor and Graduate Program Director >>> Complex Adaptive Systems Lab >>> Department of Electrical Engineering & Computer Science >>> 828 Rhodes Hall >>> University of Cincinnati >>> Cincinnati, OH 45221-0030 >>> >>> >>> Phone: (513) 556-4783 >>> Fax: (513) 556-7326 >>> Email: Ali.Minai at uc.edu >>> minaiaa at gmail.com >>> >>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>> >>> >>>> On Mon, Feb 7, 2022 at 12:28 AM Gary Marcus wrote: >>>> Ali, >>>> >>>> It?s useful to think about animals, but I really wouldn?t start with fish; it?s not clear that their ecological niche demands anything significant in the way of extrapolation, causal reasoning, or compositionality. There is good evidence elsewhere in the animal world for extrapolation of functions that may be innate (eg solar azimuth in bees), and causal reasoning (eg tool use in ravens, various primates, and octopi). It?s still not clear to me how much hierarchical representation (critical to AGI) exists outside of humans, though; the ability to construct rich new cognitive models may also be unique to us. >>>> >>>> In any case it matters not in the least whether the average cat or human cares about symbols, anymore that it matters whether the average animal understands digestion; only a tiny fraction of the creatures on this planet have any real understanding of their internal workings. >>>> >>>> My overall feeling is that we are a really, really long way from understanding the neural basis of higher-level cognition, and that AI is going to need muddle through on its own, for another decade or two, >>>> >>>> I do fully agree with your conclusion, though, that "AI today is driven more by habit and the incentives of the academic and corporate marketplaces than by a deep, long-term view of AI as a great exploratory project in fundamental science." Let's hope that changes. >>>> >>>> Gary >>>> >>>>>> On Feb 6, 2022, at 13:19, Ali Minai wrote: >>>>>> >>>>> ? >>>>> Gary, >>>>> >>>>> That?s a very interesting and accurate list of capabilities that a general intelligent system must have and that our AI does not. Of course, the list is familiar to me from having read your book. However, I have a somewhat different take on this whole thing. >>>>> >>>>> All the things we discuss here ? symbols/no symbols, parts/wholes, supervised/unsupervised, token/type, etc., are useful categories and distinctions for our analysis of the problem, and are partly a result of the historical evolution of the field of AI in particular and of philosophy in general. The categories are not wrong in any way, of course, but they are posterior to the actual system ? good for describing and analyzing it, and for validating our versions of it (which is how you use them). I think they are less useful as prescriptions for how to build our AI systems. If intelligent systems did not already exist and we were building them from scratch (please ignore the impossibility of that), having a list of ?must haves? would be great. But intelligent systems already exist ? from humans to fish ? and they already have these capacities to a greater or lesser degree because of the physics of their biology. A cat?s intelligence does not care whether it has symbols or not, and nor does mine or yours. Whatever we describe as symbolic processing post-facto has already been done by brains for at least tens of millions of years. Instead of getting caught up in ?how to add symbols into our neural models?, we should be investigating how what we see as symbolic processing emerges from animal brains, and then replicate those brains to the degree necessary. If we can do that, symbolic processing will already be present. But it cannot be done piece by piece. It must take the integrity of the whole brain and the body it is part of, and its environment, into account. That?s why I think that a much better ? though a very long ? route to AI is to start by understanding how a fish brain makes the intelligence of a fish possible, and then boot up our knowledge across phylogenetic stages: Bottom up reverse engineering rather than top-down engineering. That?s the way Nature built up to human intelligence, and we will succeed only by reverse engineering it. Of course, we can do it much faster and with shortcuts because we are intelligent, purposive agents, but working top-down by building piecewise systems that satisfy a list of attributes will not get us there. Among other things, those pieces will be impossible to integrate into the kind of intelligence that can have those general models of the world that you rightly point to as being necessary. >>>>> >>>>> I think that one thing that has been a great boon to the AI enterprise has also been one of the greatest impediments to its complete success, and that is the ?computationalization? of intelligence. On the one hand, thinking of intelligence computationally allows us to describe it abstractly and in a principled, formal way. It also resonates with the fact that we are trying to implement intelligence through computational machines. But, on the flip side, this view of intelligence divorces it from its physics ? from the fact that real intelligence in animals emerges from the physics of the physical system. That system is not a collection of its capabilities; rather, those capabilities are immanent in it by virtue of its physics. When we try to build those capabilities computationally, i.e., through code, we are making the same error that the practitioners of old-style ?symbolic AI? made ? what I call the ?professors are smarter than Nature? error, i.e., the idea that we are going to enumerate (or describe) all the things that underlie intelligence and implement them one by one until we get complete intelligence. We will never be able to enumerate all those capabilities, and will never be able to get to that complete intelligence. The only difference between us and the ?symbolists? of yore is that we are replacing giant LISP and Prolog programs with giant neural networks. Otherwise, we are using our models exactly as they were trying to use their models, and we will fail just as they did unless we get back to biology and the real thing. >>>>> >>>>> I will say again that the way we do AI today is driven more by habit and the incentives of the academic and corporate marketplaces than by a deep, long-term view of AI as a great exploratory project in fundamental science. We are just building AI to drive our cars, translate our documents, write our reports, and do our shopping. What that will teach us about actual intelligence is just incidental. >>>>> >>>>> My apologies too for a long response. >>>>> >>>>> Ali >>>>> >>>>> Ali A. Minai, Ph.D. >>>>> Professor and Graduate Program Director >>>>> Complex Adaptive Systems Lab >>>>> Department of Electrical Engineering & Computer Science >>>>> 828 Rhodes Hall >>>>> University of Cincinnati >>>>> Cincinnati, OH 45221-0030 >>>>> >>>>> Phone: (513) 556-4783 >>>>> Fax: (513) 556-7326 >>>>> Email: Ali.Minai at uc.edu >>>>> minaiaa at gmail.com >>>>> >>>>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>>>> >>>>> >>>>>> On Sun, Feb 6, 2022 at 9:42 AM Gary Marcus wrote: >>>>>> Dear Asim, >>>>>> >>>>>> Sorry for a long answer to your short but rich questions. >>>>>> Yes, memory in my view has to be part of the answer to the type-token problem. Symbol systems encoded in memory allow a natural way to set up records, and something akin to that seems necessary. Pure multilayer perceptrons struggle with type-token distinctions precisely because they lack such records. On the positive side, I see more and more movement towards recordlike stores (eg w key-value stores in memory networks), and I think that is an important and necessary step, very familiar from the symbol-manipulating playbook, sometimes implemented in new ways. >>>>>> But ultimately, handling the type-token distinction requires considerable inferential overhead beyond the memory representation of a record per se. How do you determine when to denote something (e.g. Felix) as an instance, and of which kinds (cat, animal etc), and how do you leverage that knowledge once you determine it? >>>>>> In the limit we reason about types vs tokens in fairly subtle ways, eg in guessing whether a glass that we put down at party is likely to be ours. The reverse is also important: we need to be learn particular traits for individuals and not erroneously generalize them to the class; if my aunt Esther wins the lottery, one shouldn?t infer that all of my aunts or all of my relatives or adult females have won the lottery. so you need both representational machinery that can distinguish eg my cat from cats in general and reasoning machinery to decide at what level certain learned knowledge should inhere. (I had a whole chapter about this sort of thing in The Algebraic Mind if you are interested, and Mike Mozer had a book about types and tokens in neural networks in the mid 1990s). >>>>>> Yes, part (though not all!) of what we do when we set up cognitive models in our heads is to track particular individuals and their properties. If you only had to correlate kinds (cats) and their properties (have fur) you could maybe get away with a multilayer perceptron, but once you need to track individuals, yes, you really need some kind of memory-based records. >>>>>> As far as I can tell, Transformers can sometimes approximate some of this for a few sentences, but not over long stretches. >>>>>> >>>>>> As a small terminological aside; for me cognitive models ? cognitive modeling. Cognitive modeling is about building psychological or computational models of how people think, whereas what I mean by a cognitive model is a representation of eg the entities in some situation and the relations between those entities. >>>>>> >>>>>> To your closing question, none of us yet really knows how to build understanding into machines. A solid type-token distinction, both in terms of representation and reasoning, is critical for general intelligence, but hardly sufficient. Personally, I think some minimal prerequisites would be: >>>>>> representations of space, time, causality, individuals, kinds, persons, places, objects, etc. >>>>>> representations of abstractions that can hold over all entities in a class >>>>>> compositionality (if we are talking about human-like understanding) >>>>>> capacity to construct and update cognitive models on the fly >>>>>> capacity to reason over entities in those models >>>>>> ability to learn about new entities and their properties >>>>>> Much of my last book (Rebooting AI, w Ernie Davis) is about the above list. The section in the language chapter on a children?s story in which man has lost is wallet is an especially vivid worked example. Later chapters elaborate some of the challenges in representing space, time, and causality. >>>>>> >>>>>> Gary >>>>>> >>>>>> >>>>>>>> On Feb 5, 2022, at 18:58, Asim Roy wrote: >>>>>>>> >>>>>>> ? >>>>>>> Gary, >>>>>>> >>>>>>> >>>>>>> >>>>>>> I don?t get much into the type of cognitive modeling you are talking about, but I would guess that the type problem can generally be handled by neural network models and tokens can be resolved with some memory-based system. But to the heart of the question, this is what so-called ?understanding? reduces to computation wise? >>>>>>> >>>>>>> >>>>>>> >>>>>>> Asim >>>>>>> >>>>>>> >>>>>>> >>>>>>> From: Gary Marcus >>>>>>> Sent: Saturday, February 5, 2022 8:39 AM >>>>>>> To: Asim Roy >>>>>>> Cc: Ali Minai ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub >>>>>>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>>>>> >>>>>>> >>>>>>> >>>>>>> There is no magic in understanding, just computation that has been realized in the wetware of humans and that eventually can be realized in machines. But understanding is not (just) learning. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Understanding incorporates (or works in tandem with) learning - but also, critically, in tandem with inference, and the development and maintenance of cognitive models. Part of developing an understanding of cats in general is to learn long term-knowledge about their properties, both directly (e.g., through observation) and indirectly (eg through learning facts about animals in general that can be extended to cats), often through inference (if all animals have DNA, and a cat is an animal, it must also have DNA). The understanding of a particular cat also involves direct observation, but also inference (eg one might surmise that the reason that Fluffy is running about the room is that Fluffy suspects there is a mouse stirring somewhere nearby). But all of that, I would say, is subservient to the construction of cognitive models that can be routinely updated (e.g., Fluffy is currently in the living room, skittering about, perhaps looking for a mouse). >>>>>>> >>>>>>> In humans, those dynamic, relational models, which form part of an understanding, can support inference (if Fluffy is in the living room, we can infer that Fluffy is not outside, not lost, etc). Without such models - which I think represent a core part of understanding - AGI is an unlikely prospect. >>>>>>> >>>>>>> Current neural networks, as it happens, are better at acquiring long-term knowledge (cats have whiskers) than they are at dynamically updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind of dynamic model that I am describing. To a modest degree they can approximate it on the basis of large samples of texts, but their ultimate incoherence stems from the fact that they do not have robust internal cognitive models that they can update on the fly. >>>>>>> >>>>>>> Without such cognitive models you can still capture some aspects of understanding (eg predicting that cats are likely to be furry), but things fall apart quickly; inference is never reliable, and coherence is fleeting. >>>>>>> >>>>>>> As a final note, one of the most foundational challenges in constructing adequate cognitive models of the world is to have a clear distinction between individuals and kinds; as I emphasized 20 years ago (in The Algebraic Mind), this has always been a weakness in neural networks, and I don?t think that the type-token problem has yet been solved. >>>>>>> >>>>>>> >>>>>>> Gary >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Feb 5, 2022, at 01:31, Asim Roy wrote: >>>>>>> >>>>>>> ? >>>>>>> >>>>>>> All, >>>>>>> >>>>>>> >>>>>>> >>>>>>> I think the broader question was ?understanding.? Here are two Youtube videos showing simple robots ?learning? to walk. They are purely physical systems. Do they ?understand? anything ? such as the need to go around an obstacle, jumping over an obstacle, walking up and down stairs and so on? By the way, they ?learn? to do these things on their own, literally unsupervised, very much like babies. The basic question is: what is ?understanding? if not ?learning?? Is there some other mechanism (magic) at play in our brain that helps us ?understand?? >>>>>>> >>>>>>> >>>>>>> >>>>>>> https://www.youtube.com/watch?v=gn4nRCC9TwQ >>>>>>> >>>>>>> https://www.youtube.com/watch?v=8sO7VS3q8d0 >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> Asim Roy >>>>>>> >>>>>>> Professor, Information Systems >>>>>>> >>>>>>> Arizona State University >>>>>>> >>>>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>>>> >>>>>>> Asim Roy | iSearch (asu.edu) >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> From: Ali Minai >>>>>>> Sent: Friday, February 4, 2022 11:38 PM >>>>>>> To: Asim Roy >>>>>>> Cc: Gary Marcus ; Danko Nikolic ; Brad Wyble ; connectionists at mailman.srv.cs.cmu.edu; AIhub >>>>>>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>>>>> >>>>>>> >>>>>>> >>>>>>> Asim >>>>>>> >>>>>>> >>>>>>> >>>>>>> Of course there's nothing magical about understanding, and the mind has to emerge from the physical system, but our AI models at this point are not even close to realizing how that happens. We are, at best, simulating a superficial approximation of a few parts of the real thing. A single, integrated system where all the aspects of intelligence emerge from the same deep, well-differentiated physical substrate is far beyond our capacity. Paying more attention to neurobiology will be essential to get there, but so will paying attention to development - both physical and cognitive - and evolution. The configuration of priors by evolution is key to understanding how real intelligence learns so quickly and from so little. This is not an argument for using genetic algorithms to design our systems, just for understanding the tricks evolution has used and replicating them by design. Development is more feasible to do computationally, but hardly any models have looked at it except in a superficial sense. Nature creates basic intelligence not so much by configuring functions by explicit training as by tweaking, modulating, ramifying, and combining existing ones in a multi-scale self-organization process. We then learn much more complicated things (like playing chess) by exploiting that substrate, and using explicit instruction or learning by practice. The fundamental lesson of complex systems is that complexity is built in stages - each level exploiting the organization of the level below it. We see it in evolution, development, societal evolution, the evolution of technology, etc. Our approach in AI, in contrast, is to initialize a giant, naive system and train it to do something really complicated - but really specific - by training the hell out of it. Sure, now we do build many systems on top of pre-trained models like GPT-3 and BERT, which is better, but those models were again trained by the same none-to-all process I decried above. Contrast that with how humans acquire language, and how they integrate it into their *entire* perceptual, cognitive, and behavioral repertoire, not focusing just on this or that task. The age of symbolic AI may have passed, but the reductionistic mindset has not. We cannot build minds by chopping it into separate verticals. >>>>>>> >>>>>>> >>>>>>> >>>>>>> FTR, I'd say that the emergence of models such as GLOM and Hawkins and Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but they are, I think, looking in the right direction. With a million miles to go! >>>>>>> >>>>>>> >>>>>>> >>>>>>> Ali >>>>>>> >>>>>>> >>>>>>> >>>>>>> Ali A. Minai, Ph.D. >>>>>>> Professor and Graduate Program Director >>>>>>> Complex Adaptive Systems Lab >>>>>>> Department of Electrical Engineering & Computer Science >>>>>>> >>>>>>> 828 Rhodes Hall >>>>>>> >>>>>>> University of Cincinnati >>>>>>> Cincinnati, OH 45221-0030 >>>>>>> >>>>>>> >>>>>>> Phone: (513) 556-4783 >>>>>>> Fax: (513) 556-7326 >>>>>>> Email: Ali.Minai at uc.edu >>>>>>> minaiaa at gmail.com >>>>>>> >>>>>>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: >>>>>>> >>>>>>> First of all, the brain is a physical system. There is no ?magic? inside the brain that does the ?understanding? part. Take for example learning to play tennis. You hit a few balls - some the right way and some wrong ? but you fairly quickly learn to hit them right most of the time. So there is obviously some simulation going on in the brain about hitting the ball in different ways and ?learning? its consequences. What you are calling ?understanding? is really these simulations about different scenarios. It?s also very similar to augmentation used to train image recognition systems where you rotate images, obscure parts and so on, so that you still can say it?s a cat even though you see only the cat?s face or whiskers or a cat flipped on its back. So, if the following questions relate to ?understanding,? you can easily resolve this by simulating such scenarios when ?teaching? the system. There?s nothing ?magical? about ?understanding.? As I said, bear in mind that the brain, after all, is a physical system and ?teaching? and ?understanding? is embodied in that physical system, not outside it. So ?understanding? is just part of ?learning,? nothing more. >>>>>>> >>>>>>> >>>>>>> >>>>>>> DANKO: >>>>>>> >>>>>>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>>>>>> >>>>>>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>>>>>> >>>>>>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>>>>>> >>>>>>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>>>>>> >>>>>>> >>>>>>> >>>>>>> Asim Roy >>>>>>> >>>>>>> Professor, Information Systems >>>>>>> >>>>>>> Arizona State University >>>>>>> >>>>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>>>> >>>>>>> Asim Roy | iSearch (asu.edu) >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> From: Gary Marcus >>>>>>> Sent: Thursday, February 3, 2022 9:26 AM >>>>>>> To: Danko Nikolic >>>>>>> Cc: Asim Roy ; Geoffrey Hinton ; AIhub ; connectionists at mailman.srv.cs.cmu.edu >>>>>>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>>>>> >>>>>>> >>>>>>> >>>>>>> Dear Danko, >>>>>>> >>>>>>> >>>>>>> >>>>>>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED talk, in which he said (paraphrasing from memory, because I don?t remember the precise words) that the famous 200 Quoc Le unsupervised model [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf] had learned the concept of a ca. In reality the model had clustered together some catlike images based on the image statistics that it had extracted, but it was a long way from a full, counterfactual-supporting concept of a cat, much as you describe below. >>>>>>> >>>>>>> >>>>>>> >>>>>>> I fully agree with you that the reason for even having a semantics is as you put it, "to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations.? GPT-3 sometimes gives the appearance of having done so, but it falls apart under close inspection, so the problem remains unsolved. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Gary >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic wrote: >>>>>>> >>>>>>> >>>>>>> >>>>>>> G. Hinton wrote: "I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request." >>>>>>> >>>>>>> >>>>>>> >>>>>>> I would like to suggest why drawing a hamster with a red hat does not necessarily imply understanding of the statement "hamster wearing a red hat". >>>>>>> >>>>>>> To understand that "hamster wearing a red hat" would mean inferring, in newly emerging situations of this hamster, all the real-life implications that the red hat brings to the little animal. >>>>>>> >>>>>>> >>>>>>> >>>>>>> What would happen to the hat if the hamster rolls on its back? (Would the hat fall off?) >>>>>>> >>>>>>> What would happen to the red hat when the hamster enters its lair? (Would the hat fall off?) >>>>>>> >>>>>>> What would happen to that hamster when it goes foraging? (Would the red hat have an influence on finding food?) >>>>>>> >>>>>>> What would happen in a situation of being chased by a predator? (Would it be easier for predators to spot the hamster?) >>>>>>> >>>>>>> >>>>>>> >>>>>>> ...and so on. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Countless many questions can be asked. One has understood "hamster wearing a red hat" only if one can answer reasonably well many of such real-life relevant questions. Similarly, a student has understood materias in a class only if they can apply the materials in real-life situations (e.g., applying Pythagora's theorem). If a student gives a correct answer to a multiple choice question, we don't know whether the student understood the material or whether this was just rote learning (often, it is rote learning). >>>>>>> >>>>>>> >>>>>>> >>>>>>> I also suggest that understanding also comes together with effective learning: We store new information in such a way that we can recall it later and use it effectively i.e., make good inferences in newly emerging situations based on this knowledge. >>>>>>> >>>>>>> >>>>>>> >>>>>>> In short: Understanding makes us humans able to 1) learn with a few examples and 2) apply the knowledge to a broad set of situations. >>>>>>> >>>>>>> >>>>>>> >>>>>>> No neural network today has such capabilities and we don't know how to give them such capabilities. Neural networks need large amounts of training examples that cover a large variety of situations and then the networks can only deal with what the training examples have already covered. Neural networks cannot extrapolate in that 'understanding' sense. >>>>>>> >>>>>>> >>>>>>> >>>>>>> I suggest that understanding truly extrapolates from a piece of knowledge. It is not about satisfying a task such as translation between languages or drawing hamsters with hats. It is how you got the capability to complete the task: Did you only have a few examples that covered something different but related and then you extrapolated from that knowledge? If yes, this is going in the direction of understanding. Have you seen countless examples and then interpolated among them? Then perhaps it is not understanding. >>>>>>> >>>>>>> >>>>>>> >>>>>>> So, for the case of drawing a hamster wearing a red hat, understanding perhaps would have taken place if the following happened before that: >>>>>>> >>>>>>> >>>>>>> >>>>>>> 1) first, the network learned about hamsters (not many examples) >>>>>>> >>>>>>> 2) after that the network learned about red hats (outside the context of hamsters and without many examples) >>>>>>> >>>>>>> 3) finally the network learned about drawing (outside of the context of hats and hamsters, not many examples) >>>>>>> >>>>>>> >>>>>>> >>>>>>> After that, the network is asked to draw a hamster with a red hat. If it does it successfully, maybe we have started cracking the problem of understanding. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Note also that this requires the network to learn sequentially without exhibiting catastrophic forgetting of the previous knowledge, which is possibly also a consequence of human learning by understanding. >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> Danko >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> Dr. Danko Nikoli? >>>>>>> www.danko-nikolic.com >>>>>>> https://www.linkedin.com/in/danko-nikolic/ >>>>>>> >>>>>>> --- A progress usually starts with an insight --- >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> Virus-free. www.avast.com >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>>>>>> >>>>>>> Without getting into the specific dispute between Gary and Geoff, I think with approaches similar to GLOM, we are finally headed in the right direction. There?s plenty of neurophysiological evidence for single-cell abstractions and multisensory neurons in the brain, which one might claim correspond to symbols. And I think we can finally reconcile the decades old dispute between Symbolic AI and Connectionism. >>>>>>> >>>>>>> >>>>>>> >>>>>>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>>>>>> >>>>>>> GARY: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Asim Roy >>>>>>> >>>>>>> Professor, Information Systems >>>>>>> >>>>>>> Arizona State University >>>>>>> >>>>>>> Lifeboat Foundation Bios: Professor Asim Roy >>>>>>> >>>>>>> Asim Roy | iSearch (asu.edu) >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> From: Connectionists On Behalf Of Gary Marcus >>>>>>> Sent: Wednesday, February 2, 2022 1:26 PM >>>>>>> To: Geoffrey Hinton >>>>>>> Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu >>>>>>> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >>>>>>> >>>>>>> >>>>>>> >>>>>>> Dear Geoff, and interested others, >>>>>>> >>>>>>> >>>>>>> >>>>>>> What, for example, would you make of a system that often drew the red-hatted hamster you requested, and perhaps a fifth of the time gave you utter nonsense? Or say one that you trained to create birds but sometimes output stuff like this: >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> One could >>>>>>> >>>>>>> >>>>>>> >>>>>>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>>>>>> >>>>>>> or >>>>>>> >>>>>>> b. wonder if it might be possible that sometimes the system gets the right answer for the wrong reasons (eg partial historical contingency), and wonder whether another approach might be indicated. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Benchmarks are harder than they look; most of the field has come to recognize that. The Turing Test has turned out to be a lousy measure of intelligence, easily gamed. It has turned out empirically that the Winograd Schema Challenge did not measure common sense as well as Hector might have thought. (As it happens, I am a minor coauthor of a very recent review on this very topic: https://arxiv.org/abs/2201.02387) But its conquest in no way means machines now have common sense; many people from many different perspectives recognize that (including, e.g., Yann LeCun, who generally tends to be more aligned with you than with me). >>>>>>> >>>>>>> >>>>>>> >>>>>>> So: on the goalpost of the Winograd schema, I was wrong, and you can quote me; but what you said about me and machine translation remains your invention, and it is inexcusable that you simply ignored my 2019 clarification. On the essential goal of trying to reach meaning and understanding, I remain unmoved; the problem remains unsolved. >>>>>>> >>>>>>> >>>>>>> >>>>>>> All of the problems LLMs have with coherence, reliability, truthfulness, misinformation, etc stand witness to that fact. (Their persistent inability to filter out toxic and insulting remarks stems from the same.) I am hardly the only person in the field to see that progress on any given benchmark does not inherently mean that the deep underlying problems have solved. You, yourself, in fact, have occasionally made that point. >>>>>>> >>>>>>> >>>>>>> >>>>>>> With respect to embeddings: Embeddings are very good for natural language processing; but NLP is not the same as NLU ? when it comes to understanding, their worth is still an open question. Perhaps they will turn out to be necessary; they clearly aren?t sufficient. In their extreme, they might even collapse into being symbols, in the sense of uniquely identifiable encodings, akin to the ASCII code, in which a specific set of numbers stands for a specific word or concept. (Wouldn?t that be ironic?) >>>>>>> >>>>>>> >>>>>>> >>>>>>> (Your GLOM, which as you know I praised publicly, is in many ways an effort to wind up with encodings that effectively serve as symbols in exactly that way, guaranteed to serve as consistent representations of specific concepts.) >>>>>>> >>>>>>> >>>>>>> >>>>>>> Notably absent from your email is any kind of apology for misrepresenting my position. It?s fine to say that ?many people thirty years ago once thought X? and another to say ?Gary Marcus said X in 2015?, when I didn?t. I have consistently felt throughout our interactions that you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) apologized to me for having made that error. I am still not he. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Which maybe connects to the last point; if you read my work, you would see thirty years of arguments for neural networks, just not in the way that you want them to exist. I have ALWAYS argued that there is a role for them; characterizing me as a person ?strongly opposed to neural networks? misses the whole point of my 2001 book, which was subtitled ?Integrating Connectionism and Cognitive Science.? >>>>>>> >>>>>>> >>>>>>> >>>>>>> In the last two decades or so you have insisted (for reasons you have never fully clarified, so far as I know) on abandoning symbol-manipulation, but the reverse is not the case: I have never called for dismissal of neural networks, but rather for some hybrid between the two (as you yourself contemplated in 1991); the point of the 2001 book was to characterize exactly where multilayer perceptrons succeeded and broke down, and where symbols could complement them. It?s a rhetorical trick (which is what the previous thread was about) to pretend otherwise. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Gary >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Feb 2, 2022, at 11:22, Geoffrey Hinton wrote: >>>>>>> >>>>>>> ? >>>>>>> >>>>>>> Embeddings are just vectors of soft feature detectors and they are very good for NLP. The quote on my webpage from Gary's 2015 chapter implies the opposite. >>>>>>> >>>>>>> >>>>>>> >>>>>>> A few decades ago, everyone I knew then would have agreed that the ability to translate a sentence into many different languages was strong evidence that you understood it. >>>>>>> >>>>>>> >>>>>>> >>>>>>> But once neural networks could do that, their critics moved the goalposts. An exception is Hector Levesque who defined the goalposts more sharply by saying that the ability to get pronoun references correct in Winograd sentences is a crucial test. Neural nets are improving at that but still have some way to go. Will Gary agree that when they can get pronoun references correct in Winograd sentences they really do understand? Or does he want to reserve the right to weasel out of that too? >>>>>>> >>>>>>> >>>>>>> >>>>>>> Some people, like Gary, appear to be strongly opposed to neural networks because they do not fit their preconceived notions of how the mind should work. >>>>>>> >>>>>>> I believe that any reasonable person would admit that if you ask a neural net to draw a picture of a hamster wearing a red hat and it draws such a picture, it understood the request. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Geoff >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>>>>>> >>>>>>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger neural network community, >>>>>>> >>>>>>> >>>>>>> >>>>>>> There has been a lot of recent discussion on this list about framing and scientific integrity. Often the first step in restructuring narratives is to bully and dehumanize critics. The second is to misrepresent their position. People in positions of power are sometimes tempted to do this. >>>>>>> >>>>>>> >>>>>>> >>>>>>> The Hinton-Hanson interview that you just published is a real-time example of just that. It opens with a needless and largely content-free personal attack on a single scholar (me), with the explicit intention of discrediting that person. Worse, the only substantive thing it says is false. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Hinton says ?In 2015 he [Marcus] made a prediction that computers wouldn?t be able to do machine translation.? >>>>>>> >>>>>>> >>>>>>> >>>>>>> I never said any such thing. >>>>>>> >>>>>>> >>>>>>> >>>>>>> What I predicted, rather, was that multilayer perceptrons, as they existed then, would not (on their own, absent other mechanisms) understand language. Seven years later, they still haven?t, except in the most superficial way. >>>>>>> >>>>>>> >>>>>>> >>>>>>> I made no comment whatsoever about machine translation, which I view as a separate problem, solvable to a certain degree by correspondance without semantics. >>>>>>> >>>>>>> >>>>>>> >>>>>>> I specifically tried to clarify Hinton?s confusion in 2019, but, disappointingly, he has continued to purvey misinformation despite that clarification. Here is what I wrote privately to him then, which should have put the matter to rest: >>>>>>> >>>>>>> >>>>>>> >>>>>>> You have taken a single out of context quote [from 2015] and misrepresented it. The quote, which you have prominently displayed at the bottom on your own web page, says: >>>>>>> >>>>>>> >>>>>>> >>>>>>> Hierarchies of features are less suited to challenges such as language, inference, and high-level planning. For example, as Noam Chomsky famously pointed out, language is filled with sentences you haven't seen before. Pure classifier systems don't know what to do with such sentences. The talent of feature detectors -- in identifying which member of some category something belongs to -- doesn't translate into understanding novel sentences, in which each sentence has its own unique meaning. >>>>>>> >>>>>>> >>>>>>> >>>>>>> It does not say "neural nets would not be able to deal with novel sentences"; it says that hierachies of features detectors (on their own, if you read the context of the essay) would have trouble understanding novel sentences. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Google Translate does yet not understand the content of the sentences is translates. It cannot reliably answer questions about who did what to whom, or why, it cannot infer the order of the events in paragraphs, it can't determine the internal consistency of those events, and so forth. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Since then, a number of scholars, such as the the computational linguist Emily Bender, have made similar points, and indeed current LLM difficulties with misinformation, incoherence and fabrication all follow from these concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf, also emphasizing issues of understanding and meaning: >>>>>>> >>>>>>> >>>>>>> >>>>>>> The success of the large neural language models on many NLP tasks is exciting. However, we find that these successes sometimes lead to hype in which these models are being described as ?understanding? language or capturing ?meaning?. In this position paper, we argue that a system trained only on form has a priori no way to learn meaning. .. a clear understanding of the distinction between form and meaning will help guide the field towards better science around natural language understanding. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Her later article with Gebru on language models ?stochastic parrots? is in some ways an extension of this point; machine translation requires mimicry, true understanding (which is what I was discussing in 2015) requires something deeper than that. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Hinton?s intellectual error here is in equating machine translation with the deeper comprehension that robust natural language understanding will require; as Bender and Koller observed, the two appear not to be the same. (There is a longer discussion of the relation between language understanding and machine translation, and why the latter has turned out to be more approachable than the former, in my 2019 book with Ernest Davis). >>>>>>> >>>>>>> >>>>>>> >>>>>>> More broadly, Hinton?s ongoing dismissiveness of research from perspectives other than his own (e.g. linguistics) have done the field a disservice. >>>>>>> >>>>>>> >>>>>>> >>>>>>> As Herb Simon once observed, science does not have to be zero-sum. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Sincerely, >>>>>>> >>>>>>> Gary Marcus >>>>>>> >>>>>>> Professor Emeritus >>>>>>> >>>>>>> New York University >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Feb 2, 2022, at 06:12, AIhub wrote: >>>>>>> >>>>>>> ? >>>>>>> >>>>>>> Stephen Hanson in conversation with Geoff Hinton >>>>>>> >>>>>>> >>>>>>> >>>>>>> In the latest episode of this video series for AIhub.org, Stephen Hanson talks to Geoff Hinton about neural networks, backpropagation, overparameterization, digit recognition, voxel cells, syntax and semantics, Winograd sentences, and more. >>>>>>> >>>>>>> >>>>>>> >>>>>>> You can watch the discussion, and read the transcript, here: >>>>>>> >>>>>>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>>>>>> >>>>>>> >>>>>>> >>>>>>> About AIhub: >>>>>>> >>>>>>> AIhub is a non-profit dedicated to connecting the AI community to the public by providing free, high-quality information through AIhub.org (https://aihub.org/). We help researchers publish the latest AI news, summaries of their work, opinion pieces, tutorials and more. We are supported by many leading scientific organizations in AI, namely AAAI, NeurIPS, ICML, AIJ/IJCAI, ACM SIGAI, EurAI/AICOMM, CLAIRE and RoboCup. >>>>>>> >>>>>>> Twitter: @aihuborg >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> Virus-free. www.avast.com >>>>>>> >>>>>>> -------------- next part -------------- An HTML attachment was scrubbed... URL: From christos.dimitrakakis at gmail.com Tue Feb 8 12:03:18 2022 From: christos.dimitrakakis at gmail.com (Christos Dimitrakakis) Date: Tue, 8 Feb 2022 18:03:18 +0100 Subject: Connectionists: Postdoc position in reinforcement learning and decision making Message-ID: <1ea66c77-dc71-2f33-511d-950a74420b40@gmail.com> We are looking for a post-doctoral reseasrcher to join our group on reinforcement learning and decision making under uncertainty more generally, at the Chalmers university of Technology, Sweden. Within the area, we are looking for candidates with a strong research interest in the following fields - Reinforcement learning and decision making under uncertainty: 1. Exploration in reinforcement learning. 2. Decision making nuder partial information. 3. Representations of uncertainty in decision making. 4. Theory of reinforcement learning (e.g. PAC/regret bounds) 5. Bayesian inference and approximate Bayesian methods. - Social aspect of machine learning 1. Theory of differntial privacy. 2. Algorithms for differentially private machine learning. 3. Algorithms for fairness in machine learning. 4. Interactions between machine learning and game theory. 5. Inference of human models of fairness or privacy. The researcher will be mainly wirking with Christos Dimitrakakis? https://sites.google.com/site/christosdimitrakakis - they will have the opportunity to visit and work with other group members at the University of Oslo, Norway ( https://www.mn.uio.no/ifi/english/people/aca/chridim/index.html ), Chalmers University of Technology, Sweden ( http://www.cse.chalmers.se/~chrdimi/ )? and the University of Neuchatel, Switzerland. Examples of our group's past and current research can be found on arxiv: https://arxiv.org/search/?searchtype=author&query=Dimitrakakis%2C+C. The postdoctoral researcher must have a strong technical background, as evidenced by publications in conferences such as AAAI/AISTATS/ALT/COLT/IJCAI/ICML/ICLR/NeurIPS/UAI and journals such as JMLR/JAIR in the following areas: 1. Fairness. 2. Privacy. 3. Reinforcement learning 4. Statistical inference. >>>> Application Information <<<<< *Application deadline* 30 February 2021. Please apply here https://www.chalmers.se/en/about-chalmers/Working-at-Chalmers/Vacancies/Pages/default.aspx From minaiaa at gmail.com Tue Feb 8 13:18:48 2022 From: minaiaa at gmail.com (Ali Minai) Date: Tue, 8 Feb 2022 13:18:48 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <28F54308-3E98-42CA-A709-1F33F4FBDE00@nyu.edu> References: <28F54308-3E98-42CA-A709-1F33F4FBDE00@nyu.edu> Message-ID: Hi Gary Thanks for your reply. Much to think about in it. For one, it makes me want to read your 2004 book right away (which I will). All human capacities - however unique they may seem to us - must necessarily evolve from those of the precursor species. If we had Australopithecus and Homo Erectus brains to study, I am sure we would find less developed examples of symbolic processing, compositionality, etc., that are recognizably of the same type as ours, and thus link our capacities with earlier ancestors that may have been closer to chimpanzees or bonobos. There's increasing reason to believe that Homo Neandertalis had a pretty complex mind, though it is on a different evolutionary branch than us (though it has contributed about 2% of our genome through cross-breeding). That suggests that the common ancestor of both H. Sapiens and H. Neandertalis also had these mental capacities to some degree. Moving backwards like that, it's hard to decide where the chain should be cut. Of course, we know that phenotypic attributes can emerge very rapidly in the evolutionary process and then be in stasis for a long time, but they do emerge from the available substrate through processes like duplication-and-divergence, recombination, diversification in neutral spaces, external pressures, etc.Understanding that process and leveraging it for engineering is, in my opinion, a more reliable route to AI. Again, I don't mean that we should replicate evolution - just that we should reverse engineer its results. I think compositionality is a useful concept in the analysis of language and inference, but not very useful as a prescriptive goal for building AI systems. It's an attribute we see in human language now and would like to see in our AI systems. But instead of asking "How can I get compositionality (or any attribute X) into my neural network model?" I want to ask, "How do the neural networks of the brain come to have the attribute of compositionality (or X)?" or "What is it in the anatomy and physiology of brain networks that allows for compositionality to be possible?" Answering that question will also tell us how to achieve it in our models. Otherwise, we are trying to get water from rocks. We don't know whether any of our neural models are capable of the kind of symbolic processing, compositionality, etc., that we want. We keep inventing new models to incorporate such desired attributes purely by thinking them up based on our previous models and their failures. This process is, in my opinion, a futile quest that is, at best, likely to lead to bad, inadequate models that will fall apart when we try to integrate them into larger intelligent systems. Much better, I suggest, to look at the real-world systems that exist, i.e., animal brains and bodies, and see how all the aspects of intelligence emerge from them. When we do, we will have neural models capable of symbolic processing, language, compositionality, etc. That's why I think it's better to have models that are more grounded in learning from biology with humility and less invested in top-down analysis and design based on our formalisms. One place where we can learn a lot about "real intelligence" is in the ways it "fails", i.e., by focusing more on what Amos Tversky flippantly called "natural stupidity". In those imperfections, we may be able to see how a real-time physical system like the brain achieves capacities such as abstract reasoning, symbolic processing, language, etc. As my old friend Daniel Polani suggested in another post in this thread (hi Daniel!), we humans do all these things quite imperfectly, and far from being a bug, that is probably a feature - an indicator of how, by sacrificing the goal of being perfect, we (i.e., evolution) have been able to become merely excellent. If we hope to understand intelligence, there are several inherited preferences that we need to kill. High on that kill list are the obsessions with optimality, stability, repeatability, explainability, etc., that are the antithesis of how natural living systems grow, thrive, and evolve. Ali *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Tue, Feb 8, 2022 at 10:42 AM Gary Marcus wrote: > Hi, Ali, > > Just to be clear, I certainly think that animal models can be useful in > principle, and eg. on motor-control and low-level vision they have been > incredibly useful. On compositionality, they may turn out to be less so, at > least in the near-term. My 2004 book The Birth of the Mind has a chapter on > evolution where I try to reconcile what we know about the astonishing > conservation across the biological world with the apparent uniqueness of > human language, with a focus on how something that is apparently unique can > rest partly but not entirely on inherited substrates. Emphasizing the > biological process of duplication and divergence, I suggested that language > might have hinged on multiple simultaneous mutations to existing > cognitively-relevant genes that gave rise to advances in the capacity for > hierarchical representation. Tecumseh Fitch has recently made a somewhat > parallel argument, also suggesting that compositionality in strong form may > be humanly unique, but arguing instead that a large expansion in Broca?s > area is responsible. If either his speculation or mine were correct, animal > models might be of relatively little direct help. (Though of course > anything that helps us understand the brain better may be of substantial > indirect help.) > > On your other point, saying that ?symbols emerge naturally from the > physics of the brains [of the animals that have them]? just doesn?t tell > us much; *everything* any animal does emerges from physics and biology > and experience. But nature has a vast array of different solutions to a > wide variety of problems, and it?s an empirical matter to understand each > creature and how it handles different problems (digestion, circulation, > navigation, planning, etc). All that is of course constrained by physics, > but it doesn't tell us eg whether any specific aspect of a particular > creature?s physiology ?emerges? or is hard-coded, learned, etc. > (Empirically, my own experimental work with babies, since replicated by > others with newborns, suggests that the capacity to acquire novel, > rule-like abstractions emerges from a developmental program that does not > depend on post-natal experience.) > > More broadly, I think that most brains on this planet have some capacity > for symbol-manipulation (eg honey bees calculating the solar azimuth > function and extrapolating to novel lighting conditions) but that > compositionality is less prevalent. I don?t think that you have to have > symbols or compositionality to have a hint of intelligence, but I do think > you probably need them for AGI or for understanding how humans tick. > > Gary > > > > On Feb 7, 2022, at 23:22, Ali Minai wrote: > > ? > Hi Gary > > Thanks for your reply. I'll think more about your points. I do think that, > to understand the human mind, we should start with vertebrates, which is > why I suggested fish. At least for the motor system - which is part of the > mind - we have learned a lot from lampreys (e.g. Sten Grillner's work and > that beautiful lamprey-salamander model by Ijspeert et al.), and it has > taught us a lot about locomotion in other animals, including mammals. The > principles clearly generalize, though the complexity increases a lot. > Insects too are very interesting. After all, they are our ancestors too. > > I don't agree that we can posit a clear transition from deep cognitive > models in humans to none below that in the phylogenetic tree. Chimpanzees > and macaques certainly show some evidence, and there's no reason to think > that it's a step change rather than a highly nonlinear continuum. And even > though what we might (simplistically) call System 2 aspects of cognition > are minimally present in other mammals, their precursors must be. > > My point about cats and symbols was not regarding whether cats are aware > of symbols, but that symbols emerge naturally from the physics of their > brains. Behaviors that require some small degree of symbolic processing > exist in mammals other than humans (e.g., transitive inference and > landmark-based navigation in rats), and it is seen better as an emergent > property of brains than an attribute to be explicitly built-into neural > models by us. Once we have a sufficiently brain-like neural model, symbolic > processing will already be there. > > I agree with you completely that we are far from understanding some of the > most fundamental principles of the brain, but even more importantly, we are > not even looking in the right direction. I'm hoping to lay out my arguments > about all this in more detail in some other form. > > Best > Ali > > > PS: I had inadvertently posted my reply of Gary's message only to him. > Should have posted to everyone, so here it is. > > > *Ali A. Minai, Ph.D.* > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > > > On Mon, Feb 7, 2022 at 12:28 AM Gary Marcus wrote: > >> Ali, >> >> >> It?s useful to think about animals, but I really wouldn?t start with >> fish; it?s not clear that their ecological niche demands anything >> significant in the way of extrapolation, causal reasoning, or >> compositionality. There is good evidence elsewhere in the animal world for >> extrapolation of functions that may be innate (eg solar azimuth in bees), >> and causal reasoning (eg tool use in ravens, various primates, and >> octopi). It?s still not clear to me how much hierarchical representation >> (critical to AGI) exists outside of humans, though; the ability to >> construct rich new cognitive models may also be unique to us. >> >> >> In any case it matters not in the least whether the average cat or human >> *cares* about symbols, anymore that it matters whether the average >> animal understands digestion; only a tiny fraction of the creatures on this >> planet have any real understanding of their internal workings. >> >> >> My overall feeling is that we are a really, really long way from >> understanding the neural basis of higher-level cognition, and that AI is >> going to need muddle through on its own, for another decade or two, >> >> >> I do fully agree with your conclusion, though, that "AI today is driven >> more by habit and the incentives of the academic and corporate marketplaces >> than by a deep, long-term view of AI as a great exploratory project in >> fundamental science." Let's hope that changes. >> >> >> Gary >> >> On Feb 6, 2022, at 13:19, Ali Minai wrote: >> >> ? >> >> Gary, >> >> That?s a very interesting and accurate list of capabilities that a >> general intelligent system must have and that our AI does not. Of course, >> the list is familiar to me from having read your book. However, I have a >> somewhat different take on this whole thing. >> >> >> >> All the things we discuss here ? symbols/no symbols, parts/wholes, >> supervised/unsupervised, token/type, etc., are useful categories and >> distinctions for our analysis of the problem, and are partly a result of >> the historical evolution of the field of AI in particular and of philosophy >> in general. The categories are not wrong in any way, of course, but they >> are posterior to the actual system ? good for describing and analyzing it, >> and for validating our versions of it (which is how you use them). I think >> they are less useful as prescriptions for how to build our AI systems. If >> intelligent systems did not already exist and we were building them from >> scratch (please ignore the impossibility of that), having a list of ?must >> haves? would be great. But intelligent systems already exist ? from humans >> to fish ? and they already have these capacities to a greater or lesser >> degree because of the physics of their biology. A cat?s intelligence does >> not care whether it has symbols or not, and nor does mine or yours. >> Whatever we describe as symbolic processing post-facto has already been >> done by brains for at least tens of millions of years. Instead of getting >> caught up in ?how to add symbols into our neural models?, we should be >> investigating how what we see as symbolic processing emerges from animal >> brains, and then replicate those brains to the degree necessary. If we can >> do that, symbolic processing will already be present. But it cannot be done >> piece by piece. It must take the integrity of the whole brain and the body >> it is part of, and its environment, into account. That?s why I think that a >> much better ? though a very long ? route to AI is to start by understanding >> how a fish brain makes the intelligence of a fish possible, and then boot >> up our knowledge across phylogenetic stages: Bottom up reverse engineering >> rather than top-down engineering. That?s the way Nature built up to human >> intelligence, and we will succeed only by reverse engineering it. Of >> course, we can do it much faster and with shortcuts because we are >> intelligent, purposive agents, but working top-down by building piecewise >> systems that satisfy a list of attributes will not get us there. Among >> other things, those pieces will be impossible to integrate into the kind of >> intelligence that can have those general models of the world that you >> rightly point to as being necessary. >> >> >> >> I think that one thing that has been a great boon to the AI enterprise >> has also been one of the greatest impediments to its complete success, and >> that is the ?computationalization? of intelligence. On the one hand, >> thinking of intelligence computationally allows us to describe it >> abstractly and in a principled, formal way. It also resonates with the fact >> that we are trying to implement intelligence through computational >> machines. But, on the flip side, this view of intelligence divorces it from >> its physics ? from the fact that real intelligence in animals emerges from >> the physics of the physical system. That system is not a collection of its >> capabilities; rather, those capabilities are immanent in it by virtue of >> its physics. When we try to build those capabilities computationally, i.e., >> through code, we are making the same error that the practitioners of >> old-style ?symbolic AI? made ? what I call the ?professors are smarter than >> Nature? error, i.e., the idea that we are going to enumerate (or describe) >> all the things that underlie intelligence and implement them one by one >> until we get complete intelligence. We will never be able to enumerate all >> those capabilities, and will never be able to get to that complete >> intelligence. The only difference between us and the ?symbolists? of yore >> is that we are replacing giant LISP and Prolog programs with giant neural >> networks. Otherwise, we are using our models exactly as they were trying to >> use their models, and we will fail just as they did unless we get back to >> biology and the real thing. >> >> >> >> I will say again that the way we do AI today is driven more by habit and >> the incentives of the academic and corporate marketplaces than by a deep, >> long-term view of AI as a great exploratory project in fundamental science. >> We are just building AI to drive our cars, translate our documents, write >> our reports, and do our shopping. What that will teach us about actual >> intelligence is just incidental. >> >> >> >> My apologies too for a long response. >> >> Ali >> >> >> *Ali A. Minai, Ph.D.* >> Professor and Graduate Program Director >> Complex Adaptive Systems Lab >> Department of Electrical Engineering & Computer Science >> 828 Rhodes Hall >> University of Cincinnati >> Cincinnati, OH 45221-0030 >> >> Phone: (513) 556-4783 >> Fax: (513) 556-7326 >> Email: Ali.Minai at uc.edu >> minaiaa at gmail.com >> >> WWW: https://eecs.ceas.uc.edu/~aminai/ >> >> >> >> On Sun, Feb 6, 2022 at 9:42 AM Gary Marcus wrote: >> >>> Dear Asim, >>> >>> >>> Sorry for a long answer to your short but rich questions. >>> >>> - Yes, memory in my view has to be part of the answer to the >>> type-token problem. Symbol systems encoded in memory allow a natural way to >>> set up records, and something akin to that seems necessary. Pure multilayer >>> perceptrons struggle with type-token distinctions precisely because they >>> lack such records. On the positive side, I see more and more movement >>> towards recordlike stores (eg w key-value stores in memory networks), and I >>> think that is an important and necessary step, very familiar from the >>> symbol-manipulating playbook, sometimes implemented in new ways. >>> - But ultimately, handling the type-token distinction requires >>> considerable inferential overhead beyond the memory representation of a >>> record per se. How do you determine when to denote something (e.g. >>> Felix) as an instance, and of which kinds (cat, animal etc), and how do you >>> leverage that knowledge once you determine it? >>> - In the limit we reason about types vs tokens in fairly subtle >>> ways, eg in guessing whether a glass that we put down at party is likely to >>> be ours. The reverse is also important: we need to be learn >>> particular traits for individuals and not erroneously generalize them to >>> the class; if my aunt Esther wins the lottery, one shouldn?t infer that >>> all of my aunts or all of my relatives or adult females have won the >>> lottery. so you need both representational machinery that can distinguish >>> eg my cat from cats in general and reasoning machinery to decide at what >>> level certain learned knowledge should inhere. (I had a whole chapter about >>> this sort of thing in The Algebraic Mind if you are interested, and Mike >>> Mozer had a book about types and tokens in neural networks in the mid >>> 1990s). >>> - Yes, part (though not all!) of what we do when we set up cognitive >>> models in our heads is to track particular individuals and their >>> properties. If you only had to correlate kinds (cats) and their properties >>> (have fur) you could maybe get away with a multilayer perceptron, but once >>> you need to track individuals, yes, you really need some kind of >>> memory-based records. >>> - As far as I can tell, Transformers can sometimes approximate some >>> of this for a few sentences, but not over long stretches. >>> >>> >>> As a small terminological aside; for me cognitive models ? cognitive >>> modeling. Cognitive modeling is about building psychological or >>> computational models of how people think, whereas what I mean by a cognitive >>> model is a representation of eg the entities in some situation and the >>> relations between those entities. >>> >>> >>> To your closing question, none of us yet really knows how to build >>> understanding into machines. A solid type-token distinction, both in >>> terms of representation and reasoning, is critical for general >>> intelligence, but hardly sufficient. Personally, I think some minimal >>> prerequisites would be: >>> >>> - representations of space, time, causality, individuals, kinds, >>> persons, places, objects, etc. >>> - representations of abstractions that can hold over all entities in >>> a class >>> - compositionality (if we are talking about human-like understanding) >>> - capacity to construct and update cognitive models on the fly >>> - capacity to reason over entities in those models >>> - ability to learn about new entities and their properties >>> >>> Much of my last book (*Rebooting AI*, w Ernie Davis) is about the above >>> list. The section in the language chapter on a children?s story in which >>> man has lost is wallet is an especially vivid worked example. Later >>> chapters elaborate some of the challenges in representing space, time, and >>> causality. >>> >>> >>> Gary >>> >>> >>> On Feb 5, 2022, at 18:58, Asim Roy wrote: >>> >>> ? >>> >>> Gary, >>> >>> >>> >>> I don?t get much into the type of cognitive modeling you are talking >>> about, but I would guess that the type problem can generally be handled by >>> neural network models and tokens can be resolved with some memory-based >>> system. But to the heart of the question, this is what so-called >>> ?understanding? reduces to computation wise? >>> >>> >>> >>> Asim >>> >>> >>> >>> *From:* Gary Marcus >>> *Sent:* Saturday, February 5, 2022 8:39 AM >>> *To:* Asim Roy >>> *Cc:* Ali Minai ; Danko Nikolic < >>> danko.nikolic at gmail.com>; Brad Wyble ; >>> connectionists at mailman.srv.cs.cmu.edu; AIhub >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> There is no magic in understanding, just computation that has been >>> realized in the wetware of humans and that eventually can be realized in >>> machines. But understanding is not (just) learning. >>> >>> >>> >>> Understanding incorporates (or works in tandem with) learning - but >>> also, critically, in tandem with inference, *and the development and >>> maintenance of cognitive models*. Part of developing an understanding >>> of cats in general is to learn long term-knowledge about their properties, >>> both directly (e.g., through observation) and indirectly (eg through >>> learning facts about animals in general that can be extended to cats), >>> often through inference (if all animals have DNA, and a cat is an animal, >>> it must also have DNA). The understanding of a particular cat also >>> involves direct observation, but also inference (eg one might surmise >>> that the reason that Fluffy is running about the room is that Fluffy >>> suspects there is a mouse stirring somewhere nearby). *But all of that, >>> I would say, is subservient to the construction of cognitive models that >>> can be routinely updated *(e.g., Fluffy is currently in the living >>> room, skittering about, perhaps looking for a mouse). >>> >>> >>> >>> In humans, those dynamic, relational models, which form part of an >>> understanding, can support inference (if Fluffy is in the living room, we >>> can infer that Fluffy is not outside, not lost, etc). Without such models - >>> which I think represent a core part of understanding - AGI is an unlikely >>> prospect. >>> >>> >>> >>> Current neural networks, as it happens, are better at acquiring >>> long-term knowledge (cats have whiskers) than they are at dynamically >>> updating cognitive models in real-time. LLMs like GPT-3 etc lack the kind >>> of dynamic model that I am describing. To a modest degree they can >>> approximate it on the basis of large samples of texts, but their ultimate >>> incoherence stems from the fact that they do not have robust internal >>> cognitive models that they can update on the fly. >>> >>> >>> >>> Without such cognitive models you can still capture some aspects of >>> understanding (eg predicting that cats are likely to be furry), but things >>> fall apart quickly; inference is never reliable, and coherence is fleeting. >>> >>> >>> >>> As a final note, one of the most foundational challenges in constructing >>> adequate cognitive models of the world is to have a clear distinction >>> between individuals and kinds; as I emphasized 20 years ago (in The >>> Algebraic Mind), this has always been a weakness in neural networks, and I >>> don?t think that the type-token problem has yet been solved. >>> >>> >>> >>> Gary >>> >>> >>> >>> >>> >>> On Feb 5, 2022, at 01:31, Asim Roy wrote: >>> >>> ? >>> >>> All, >>> >>> >>> >>> I think the broader question was ?understanding.? Here are two Youtube >>> videos showing simple robots ?learning? to walk. They are purely physical >>> systems. Do they ?understand? anything ? such as the need to go around an >>> obstacle, jumping over an obstacle, walking up and down stairs and so on? >>> By the way, they ?learn? to do these things on their own, literally >>> unsupervised, very much like babies. The basic question is: what is >>> ?understanding? if not ?learning?? Is there some other mechanism (magic) at >>> play in our brain that helps us ?understand?? >>> >>> >>> >>> https://www.youtube.com/watch?v=gn4nRCC9TwQ >>> >>> >>> https://www.youtube.com/watch?v=8sO7VS3q8d0 >>> >>> >>> >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> *From:* Ali Minai >>> *Sent:* Friday, February 4, 2022 11:38 PM >>> *To:* Asim Roy >>> *Cc:* Gary Marcus ; Danko Nikolic < >>> danko.nikolic at gmail.com>; Brad Wyble ; >>> connectionists at mailman.srv.cs.cmu.edu; AIhub >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Asim >>> >>> >>> >>> Of course there's nothing magical about understanding, and the mind has >>> to emerge from the physical system, but our AI models at this point are not >>> even close to realizing how that happens. We are, at best, simulating a >>> superficial approximation of a few parts of the real thing. A single, >>> integrated system where all the aspects of intelligence emerge from the >>> same deep, well-differentiated physical substrate is far beyond our >>> capacity. Paying more attention to neurobiology will be essential to get >>> there, but so will paying attention to development - both physical and >>> cognitive - and evolution. The configuration of priors by evolution is key >>> to understanding how real intelligence learns so quickly and from so >>> little. This is not an argument for using genetic algorithms to design our >>> systems, just for understanding the tricks evolution has used and >>> replicating them by design. Development is more feasible to do >>> computationally, but hardly any models have looked at it except in a >>> superficial sense. Nature creates basic intelligence not so much by >>> configuring functions by explicit training as by tweaking, modulating, >>> ramifying, and combining existing ones in a multi-scale self-organization >>> process. We then learn much more complicated things (like playing chess) by >>> exploiting that substrate, and using explicit instruction or learning by >>> practice. The fundamental lesson of complex systems is that complexity is >>> built in stages - each level exploiting the organization of the level below >>> it. We see it in evolution, development, societal evolution, the evolution >>> of technology, etc. Our approach in AI, in contrast, is to initialize a >>> giant, naive system and train it to do something really complicated - but >>> really specific - by training the hell out of it. Sure, now we do build >>> many systems on top of pre-trained models like GPT-3 and BERT, which is >>> better, but those models were again trained by the same none-to-all process >>> I decried above. Contrast that with how humans acquire language, and how >>> they integrate it into their *entire* perceptual, cognitive, and behavioral >>> repertoire, not focusing just on this or that task. The age of symbolic AI >>> may have passed, but the reductionistic mindset has not. We cannot build >>> minds by chopping it into separate verticals. >>> >>> >>> >>> FTR, I'd say that the emergence of models such as GLOM and Hawkins and >>> Ahmed's "thousand brains" is a hopeful sign. They may not be "right", but >>> they are, I think, looking in the right direction. With a million miles to >>> go! >>> >>> >>> >>> Ali >>> >>> >>> >>> *Ali A. Minai, Ph.D.* >>> Professor and Graduate Program Director >>> Complex Adaptive Systems Lab >>> Department of Electrical Engineering & Computer Science >>> >>> 828 Rhodes Hall >>> >>> University of Cincinnati >>> Cincinnati, OH 45221-0030 >>> >>> >>> Phone: (513) 556-4783 >>> Fax: (513) 556-7326 >>> Email: Ali.Minai at uc.edu >>> minaiaa at gmail.com >>> >>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>> >>> >>> >>> >>> >>> >>> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy wrote: >>> >>> First of all, the brain is a physical system. There is no ?magic? inside >>> the brain that does the ?understanding? part. Take for example learning to >>> play tennis. You hit a few balls - some the right way and some wrong ? but >>> you fairly quickly learn to hit them right most of the time. So there is >>> obviously some simulation going on in the brain about hitting the ball in >>> different ways and ?learning? its consequences. What you are calling >>> ?understanding? is really these simulations about different scenarios. It?s >>> also very similar to augmentation used to train image recognition systems >>> where you rotate images, obscure parts and so on, so that you still can say >>> it?s a cat even though you see only the cat?s face or whiskers or a cat >>> flipped on its back. So, if the following questions relate to >>> ?understanding,? you can easily resolve this by simulating such scenarios >>> when ?teaching? the system. There?s nothing ?magical? about >>> ?understanding.? As I said, bear in mind that the brain, after all, is a >>> physical system and ?teaching? and ?understanding? is embodied in that >>> physical system, not outside it. So ?understanding? is just part of >>> ?learning,? nothing more. >>> >>> >>> >>> DANKO: >>> >>> What would happen to the hat if the hamster rolls on its back? (Would >>> the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? >>> (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red >>> hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would >>> it be easier for predators to spot the hamster?) >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> *From:* Gary Marcus >>> *Sent:* Thursday, February 3, 2022 9:26 AM >>> *To:* Danko Nikolic >>> *Cc:* Asim Roy ; Geoffrey Hinton < >>> geoffrey.hinton at gmail.com>; AIhub ; >>> connectionists at mailman.srv.cs.cmu.edu >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Dear Danko, >>> >>> >>> >>> Well said. I had a somewhat similar response to Jeff Dean?s 2021 TED >>> talk, in which he said (paraphrasing from memory, because I don?t remember >>> the precise words) that the famous 200 Quoc Le unsupervised model [ >>> https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >>> ] >>> had learned the concept of a ca. In reality the model had clustered >>> together some catlike images based on the image statistics that it had >>> extracted, but it was a long way from a full, counterfactual-supporting >>> concept of a cat, much as you describe below. >>> >>> >>> >>> I fully agree with you that the reason for even having a semantics is as >>> you put it, "to 1) learn with a few examples and 2) apply the knowledge to >>> a broad set of situations.? GPT-3 sometimes gives the appearance of having >>> done so, but it falls apart under close inspection, so the problem remains >>> unsolved. >>> >>> >>> >>> Gary >>> >>> >>> >>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >>> wrote: >>> >>> >>> >>> G. Hinton wrote: "I believe that any reasonable person would admit that >>> if you ask a neural net to draw a picture of a hamster wearing a red hat >>> and it draws such a picture, it understood the request." >>> >>> >>> >>> I would like to suggest why drawing a hamster with a red hat does not >>> necessarily imply understanding of the statement "hamster wearing a red >>> hat". >>> >>> To understand that "hamster wearing a red hat" would mean inferring, in >>> newly emerging situations of this hamster, all the real-life >>> implications that the red hat brings to the little animal. >>> >>> >>> >>> What would happen to the hat if the hamster rolls on its back? (Would >>> the hat fall off?) >>> >>> What would happen to the red hat when the hamster enters its lair? >>> (Would the hat fall off?) >>> >>> What would happen to that hamster when it goes foraging? (Would the red >>> hat have an influence on finding food?) >>> >>> What would happen in a situation of being chased by a predator? (Would >>> it be easier for predators to spot the hamster?) >>> >>> >>> >>> ...and so on. >>> >>> >>> >>> Countless many questions can be asked. One has understood "hamster >>> wearing a red hat" only if one can answer reasonably well many of such >>> real-life relevant questions. Similarly, a student has understood materias >>> in a class only if they can apply the materials in real-life situations >>> (e.g., applying Pythagora's theorem). If a student gives a correct answer >>> to a multiple choice question, we don't know whether the student understood >>> the material or whether this was just rote learning (often, it is rote >>> learning). >>> >>> >>> >>> I also suggest that understanding also comes together with effective >>> learning: We store new information in such a way that we can recall it >>> later and use it effectively i.e., make good inferences in newly emerging >>> situations based on this knowledge. >>> >>> >>> >>> In short: Understanding makes us humans able to 1) learn with a few >>> examples and 2) apply the knowledge to a broad set of situations. >>> >>> >>> >>> No neural network today has such capabilities and we don't know how to >>> give them such capabilities. Neural networks need large amounts of >>> training examples that cover a large variety of situations and then >>> the networks can only deal with what the training examples have already >>> covered. Neural networks cannot extrapolate in that 'understanding' sense. >>> >>> >>> >>> I suggest that understanding truly extrapolates from a piece of >>> knowledge. It is not about satisfying a task such as translation between >>> languages or drawing hamsters with hats. It is how you got the capability >>> to complete the task: Did you only have a few examples that covered >>> something different but related and then you extrapolated from that >>> knowledge? If yes, this is going in the direction of understanding. Have >>> you seen countless examples and then interpolated among them? Then perhaps >>> it is not understanding. >>> >>> >>> >>> So, for the case of drawing a hamster wearing a red hat, understanding >>> perhaps would have taken place if the following happened before that: >>> >>> >>> >>> 1) first, the network learned about hamsters (not many examples) >>> >>> 2) after that the network learned about red hats (outside the context of >>> hamsters and without many examples) >>> >>> 3) finally the network learned about drawing (outside of the context of >>> hats and hamsters, not many examples) >>> >>> >>> >>> After that, the network is asked to draw a hamster with a red hat. If it >>> does it successfully, maybe we have started cracking the problem of >>> understanding. >>> >>> >>> >>> Note also that this requires the network to learn sequentially without >>> exhibiting catastrophic forgetting of the previous knowledge, which is >>> possibly also a consequence of human learning by understanding. >>> >>> >>> >>> >>> >>> Danko >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> >>> --- A progress usually starts with an insight --- >>> >>> >>> >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> >>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy wrote: >>> >>> Without getting into the specific dispute between Gary and Geoff, I >>> think with approaches similar to GLOM, we are finally headed in the right >>> direction. There?s plenty of neurophysiological evidence for single-cell >>> abstractions and multisensory neurons in the brain, which one might claim >>> correspond to symbols. And I think we can finally reconcile the decades old >>> dispute between Symbolic AI and Connectionism. >>> >>> >>> >>> GARY: (Your GLOM, which as you know I praised publicly, is in many ways >>> an effort to wind up with encodings that effectively serve as symbols in >>> exactly that way, guaranteed to serve as consistent representations of >>> specific concepts.) >>> >>> GARY: I have *never* called for dismissal of neural networks, but >>> rather for some hybrid between the two (as you yourself contemplated in >>> 1991); the point of the 2001 book was to characterize exactly where >>> multilayer perceptrons succeeded and broke down, and where symbols could >>> complement them. >>> >>> >>> >>> Asim Roy >>> >>> Professor, Information Systems >>> >>> Arizona State University >>> >>> Lifeboat Foundation Bios: Professor Asim Roy >>> >>> >>> Asim Roy | iSearch (asu.edu) >>> >>> >>> >>> >>> >>> >>> *From:* Connectionists *On >>> Behalf Of *Gary Marcus >>> *Sent:* Wednesday, February 2, 2022 1:26 PM >>> *To:* Geoffrey Hinton >>> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Dear Geoff, and interested others, >>> >>> >>> >>> What, for example, would you make of a system that often drew the >>> red-hatted hamster you requested, and perhaps a fifth of the time gave you >>> utter nonsense? Or say one that you trained to create birds but sometimes >>> output stuff like this: >>> >>> >>> >>> >>> >>> >>> >>> One could >>> >>> >>> >>> a. avert one?s eyes and deem the anomalous outputs irrelevant >>> >>> or >>> >>> b. wonder if it might be possible that sometimes the system gets the >>> right answer for the wrong reasons (eg partial historical contingency), and >>> wonder whether another approach might be indicated. >>> >>> >>> >>> Benchmarks are harder than they look; most of the field has come to >>> recognize that. The Turing Test has turned out to be a lousy measure of >>> intelligence, easily gamed. It has turned out empirically that the Winograd >>> Schema Challenge did not measure common sense as well as Hector might have >>> thought. (As it happens, I am a minor coauthor of a very recent review on >>> this very topic: https://arxiv.org/abs/2201.02387 >>> ) >>> But its conquest in no way means machines now have common sense; many >>> people from many different perspectives recognize that (including, e.g., >>> Yann LeCun, who generally tends to be more aligned with you than with me). >>> >>> >>> >>> So: on the goalpost of the Winograd schema, I was wrong, and you can >>> quote me; but what you said about me and machine translation remains your >>> invention, and it is inexcusable that you simply ignored my 2019 >>> clarification. On the essential goal of trying to reach meaning and >>> understanding, I remain unmoved; the problem remains unsolved. >>> >>> >>> >>> All of the problems LLMs have with coherence, reliability, truthfulness, >>> misinformation, etc stand witness to that fact. (Their persistent inability >>> to filter out toxic and insulting remarks stems from the same.) I am hardly >>> the only person in the field to see that progress on any given benchmark >>> does not inherently mean that the deep underlying problems have solved. >>> You, yourself, in fact, have occasionally made that point. >>> >>> >>> >>> With respect to embeddings: Embeddings are very good for natural >>> language *processing*; but NLP is not the same as NL*U* ? when it comes >>> to *understanding*, their worth is still an open question. Perhaps they >>> will turn out to be necessary; they clearly aren?t sufficient. In their >>> extreme, they might even collapse into being symbols, in the sense of >>> uniquely identifiable encodings, akin to the ASCII code, in which a >>> specific set of numbers stands for a specific word or concept. (Wouldn?t >>> that be ironic?) >>> >>> >>> >>> (Your GLOM, which as you know I praised publicly, is in many ways an >>> effort to wind up with encodings that effectively serve as symbols in >>> exactly that way, guaranteed to serve as consistent representations of >>> specific concepts.) >>> >>> >>> >>> Notably absent from your email is any kind of apology for >>> misrepresenting my position. It?s fine to say that ?many people thirty >>> years ago once thought X? and another to say ?Gary Marcus said X in 2015?, >>> when I didn?t. I have consistently felt throughout our interactions that >>> you have mistaken me for Zenon Pylyshyn; indeed, you once (at NeurIPS 2014) >>> apologized to me for having made that error. I am still not he. >>> >>> >>> >>> Which maybe connects to the last point; if you read my work, you would >>> see thirty years of arguments *for* neural networks, just not in the >>> way that you want them to exist. I have ALWAYS argued that there is a role >>> for them; characterizing me as a person ?strongly opposed to neural >>> networks? misses the whole point of my 2001 book, which was subtitled >>> ?Integrating Connectionism and Cognitive Science.? >>> >>> >>> >>> In the last two decades or so you have insisted (for reasons you have >>> never fully clarified, so far as I know) on abandoning symbol-manipulation, >>> but the reverse is not the case: I have *never* called for dismissal of >>> neural networks, but rather for some hybrid between the two (as you >>> yourself contemplated in 1991); the point of the 2001 book was to >>> characterize exactly where multilayer perceptrons succeeded and broke down, >>> and where symbols could complement them. It?s a rhetorical trick (which is >>> what the previous thread was about) to pretend otherwise. >>> >>> >>> >>> Gary >>> >>> >>> >>> >>> >>> On Feb 2, 2022, at 11:22, Geoffrey Hinton >>> wrote: >>> >>> ? >>> >>> Embeddings are just vectors of soft feature detectors and they are very >>> good for NLP. The quote on my webpage from Gary's 2015 chapter implies the >>> opposite. >>> >>> >>> >>> A few decades ago, everyone I knew then would have agreed that the >>> ability to translate a sentence into many different languages was strong >>> evidence that you understood it. >>> >>> >>> >>> But once neural networks could do that, their critics moved the >>> goalposts. An exception is Hector Levesque who defined the goalposts more >>> sharply by saying that the ability to get pronoun references correct in >>> Winograd sentences is a crucial test. Neural nets are improving at that but >>> still have some way to go. Will Gary agree that when they can get pronoun >>> references correct in Winograd sentences they really do understand? Or does >>> he want to reserve the right to weasel out of that too? >>> >>> >>> >>> Some people, like Gary, appear to be strongly opposed to neural networks >>> because they do not fit their preconceived notions of how the mind should >>> work. >>> >>> I believe that any reasonable person would admit that if you ask a >>> neural net to draw a picture of a hamster wearing a red hat and it draws >>> such a picture, it understood the request. >>> >>> >>> >>> Geoff >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus wrote: >>> >>> Dear AI Hub, cc: Steven Hanson and Geoffrey Hinton, and the larger >>> neural network community, >>> >>> >>> >>> There has been a lot of recent discussion on this list about framing and >>> scientific integrity. Often the first step in restructuring narratives is >>> to bully and dehumanize critics. The second is to misrepresent their >>> position. People in positions of power are sometimes tempted to do this. >>> >>> >>> >>> The Hinton-Hanson interview that you just published is a real-time >>> example of just that. It opens with a needless and largely content-free >>> personal attack on a single scholar (me), with the explicit intention of >>> discrediting that person. Worse, the only substantive thing it says is >>> false. >>> >>> >>> >>> Hinton says ?In 2015 he [Marcus] made a prediction that computers >>> wouldn?t be able to do machine translation.? >>> >>> >>> >>> I never said any such thing. >>> >>> >>> >>> What I predicted, rather, was that multilayer perceptrons, as they >>> existed then, would not (on their own, absent other mechanisms) >>> *understand* language. Seven years later, they still haven?t, except in >>> the most superficial way. >>> >>> >>> >>> I made no comment whatsoever about machine translation, which I view as >>> a separate problem, solvable to a certain degree by correspondance without >>> semantics. >>> >>> >>> >>> I specifically tried to clarify Hinton?s confusion in 2019, but, >>> disappointingly, he has continued to purvey misinformation despite that >>> clarification. Here is what I wrote privately to him then, which should >>> have put the matter to rest: >>> >>> >>> >>> You have taken a single out of context quote [from 2015] and >>> misrepresented it. The quote, which you have prominently displayed at the >>> bottom on your own web page, says: >>> >>> >>> >>> Hierarchies of features are less suited to challenges such as language, >>> inference, and high-level planning. For example, as Noam Chomsky famously >>> pointed out, language is filled with sentences you haven't seen >>> before. Pure classifier systems don't know what to do with such sentences. >>> The talent of feature detectors -- in identifying which member of some >>> category something belongs to -- doesn't translate into understanding >>> novel sentences, in which each sentence has its own unique meaning. >>> >>> >>> >>> It does *not* say "neural nets would not be able to deal with novel >>> sentences"; it says that hierachies of features detectors (on their own, if >>> you read the context of the essay) would have trouble *understanding *novel sentences. >>> >>> >>> >>> >>> Google Translate does yet not *understand* the content of the sentences >>> is translates. It cannot reliably answer questions about who did what to >>> whom, or why, it cannot infer the order of the events in paragraphs, it >>> can't determine the internal consistency of those events, and so forth. >>> >>> >>> >>> Since then, a number of scholars, such as the the computational linguist >>> Emily Bender, have made similar points, and indeed current LLM difficulties >>> with misinformation, incoherence and fabrication all follow from these >>> concerns. Quoting from Bender?s prizewinning 2020 ACL article on the matter >>> with Alexander Koller, https://aclanthology.org/2020.acl-main.463.pdf >>> , >>> also emphasizing issues of understanding and meaning: >>> >>> >>> >>> *The success of the large neural language models on many NLP tasks is >>> exciting. However, we find that these successes sometimes lead to hype in >>> which these models are being described as ?understanding? language or >>> capturing ?meaning?. In this position paper, we argue that a system trained >>> only on form has a priori no way to learn meaning. .. a clear understanding >>> of the distinction between form and meaning will help guide the field >>> towards better science around natural language understanding. * >>> >>> >>> >>> Her later article with Gebru on language models ?stochastic parrots? is >>> in some ways an extension of this point; machine translation requires >>> mimicry, true understanding (which is what I was discussing in 2015) >>> requires something deeper than that. >>> >>> >>> >>> Hinton?s intellectual error here is in equating machine translation with >>> the deeper comprehension that robust natural language understanding will >>> require; as Bender and Koller observed, the two appear not to be the same. >>> (There is a longer discussion of the relation between language >>> understanding and machine translation, and why the latter has turned out to >>> be more approachable than the former, in my 2019 book with Ernest Davis). >>> >>> >>> >>> More broadly, Hinton?s ongoing dismissiveness of research from >>> perspectives other than his own (e.g. linguistics) have done the field a >>> disservice. >>> >>> >>> >>> As Herb Simon once observed, science does not have to be zero-sum. >>> >>> >>> >>> Sincerely, >>> >>> Gary Marcus >>> >>> Professor Emeritus >>> >>> New York University >>> >>> >>> >>> On Feb 2, 2022, at 06:12, AIhub wrote: >>> >>> ? >>> >>> Stephen Hanson in conversation with Geoff Hinton >>> >>> >>> >>> In the latest episode of this video series for AIhub.org >>> , >>> Stephen Hanson talks to Geoff Hinton about neural networks, >>> backpropagation, overparameterization, digit recognition, voxel cells, >>> syntax and semantics, Winograd sentences, and more. >>> >>> >>> >>> You can watch the discussion, and read the transcript, here: >>> >>> >>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>> >>> >>> >>> >>> About AIhub: >>> >>> AIhub is a non-profit dedicated to connecting the AI community to the >>> public by providing free, high-quality information through AIhub.org >>> >>> (https://aihub.org/ >>> ). >>> We help researchers publish the latest AI news, summaries of their work, >>> opinion pieces, tutorials and more. We are supported by many leading >>> scientific organizations in AI, namely AAAI >>> , >>> NeurIPS >>> , >>> ICML >>> , >>> AIJ >>> >>> /IJCAI >>> , >>> ACM SIGAI >>> , >>> EurAI/AICOMM, CLAIRE >>> >>> and RoboCup >>> >>> . >>> >>> Twitter: @aihuborg >>> >>> >>> >>> >>> >>> >>> Virus-free. www.avast.com >>> >>> >>> >>> >>> -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Wed Feb 9 09:03:35 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Wed, 9 Feb 2022 09:03:35 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <537DF004-25CE-45A2-8155-D7E6018F4EE5@nyu.edu> <3dff1e2e-52b6-2ff3-3152-c7cf919fb14e@rubic.rutgers.edu> Message-ID: <7790f7a1-e792-5ad5-ff31-772ddeada820@rubic.rutgers.edu> Gary, You're minimizing my example, and frankly,? I don't believe DL models will "characterize some aspects of Psychology reasonably well"...? and before getting to your netflix bingeing.. one of the serious problems with DL models, say comparing to neural recordings, using correlation matrices there seem to be some correspondence to make.. but sometimes little more than a roscharch test.??? In general, I think it will be hard to show good correspondence to decision processing, episodic memory, compound stimulus conditioning,? and various perceptual illusions and transformations.??? In general the DL focus on classification and translation has created models very unlikely to easily model cognitive and perceptual phenomena. Models like Grossberg are curated to account for specific effects and over a lifetime have done a better job at making sense of psychological/neural phenomena then any other neural models I know about, whether one subscribes to the details of the modeling is another issue. So in the perceptual/cognition abstraction task I discussed it is gob-smacking that JUST ADDING LAYERS solves this really critical failure of backpropagation, barely noted by most of the neural network community focused on better benchmarks. As to Netflix titles.?? I agree, cognitive models should be adaptable and responsive to updates that create more predictive outcomes for the agent using them.?? This in no ways means the cognitive model must be symbolic or rule based.?? This was something that was true in the 1980s and perhaps truer today. This is clearly a critical aspect of GPT models.. what are the cognitive models that they are building or are they just high dimensional, phrase-structure blobs that do similarity analysis and return a nearby phrase-structure response, which happens to sound good. Steve On 2/7/22 5:07 PM, Gary Marcus wrote: > Stephen, > > I don?t doubt for a minute that deep learning can characterize some > aspects of psychology reasonably well; but either it needs to expands > its borders or else be used ?in conjunction with other techniques. > Take for example the name of the new Netflix show > > The Woman in the House Across the Street from the Girl in the Window > > Most of us can infer, compositionally, from that unusually long noun > phrase, that the title is a description of particular person, that the > title is not a complete sentence, and that the woman in question lives > in a house; we also infer that there is a second, distinct person > (likely a child) across the street, and so forth. We can also use some > knowledge of pragmatics to infer that the woman in question is likely > to be the protagonist in the show. Current systems still struggle with > that sort of thing. > > We can then watch the show (I watched a few minutes of Episode 1) and > quickly relate the title to the protagonist?s mental state, start to > develop a mental model of the protagonist?s relation to her new > neighbors, make inferences about whether certain choices appear to be > ?within character?, empathize with character or question her > judgements, etc, all with respect to a mental model that is rapidly > encoded and quickly modified. > > I think that an understanding of how people build and modify such > models would be extremely valuable (not just for fiction for everyday > reality), but I don?t see how deep learning in its current form gives > us much purchase on that. There is plenty of precedent for the kind of > mental processes I am sketching (e.g ?Walter Kintsch?s work on text > comprehension; Kamp/Kratzer/Heim work on discourse representation, > etc) from psychological and linguistic perspectives, but almost no > current contact in the neural network community with these > well-attested psychological processes. > > Gary > >> On Feb 7, 2022, at 6:01 AM, Stephen Jos? Hanson >> > wrote: >> >> Gary, >> >> This is one of the first posts of yours, that I can categorically >> agree with! >> >> I think building cognitive models through *some* training regime or >> focused sampling or architectures or something but not explicit, for >> example. >> >> The other fundamental cognitive/perceptual capability in this context >> is the ability of Neural Networks to do what Shepard (1970; Garner >> 1970s), had modeled as perceptual separable processing (finding >> parts) and perceptual integral process (finding covariance and >> structure). >> >> Shepard argued these fundamental perceptual processes were dependent >> on development and learning. >> >> A task was created with double dissociation of a categorization >> problem.? In one case: separable ( in effect, uncorrelated? features >> in the stimulus) were presented in categorization task that required >> you pay attention to at least 2 features at the same time to >> categorize correctly ("condensation").? in the other case: integral >> stimuli (in effect correlated features in stimuli) were presented? in >> a categorization task that required you to ignore the correlation and >> do categorize on 1 feature at a time ("filtration").?? This produced >> a? result that separable stimuli were more quickly learned in >> filtration tasks then integral stimuli in condensation tasks.? >> Non-intuitively,? Separable stimuli are learned more slowly in >> condensation tasks then integral stimuli then in filtration tasks.?? >> In other words attention to feature structure could cause improvement >> in learning or interference.?? Not that surprising.. however-- >> >> In the 1980s NN with single layers (Backprop) *could not* replicate >> this simple problem indicating that the cognitive model was somehow >> inadequate.???? Backprop simply learned ALL task/stimuli parings at >> the same rate, ignoring the subtle but critical difference.? It failed. >> >> Recently we >> (https://www.frontiersin.org/articles/10.3389/fpsyg.2018.00374/full?&utm_source=Email_to_authors_&utm_medium=Email&utm_content=T1_11.5e1_author&utm_campaign=Email_publication&field=&journalName=Frontiers_in_Psychology&id=284733) >> were able to show that JUST BY ADDING LAYERS the DL does match to >> human performance. >> >> What are the layers doing??? We offer an possible explanation that >> needs testing.??? Layers, appear to create a type of buffer that >> allows the network to "curate", feature detectors that are spatially >> distant from the input (conv layer, for example), this curation comes >> in various attention forms (something in that will appear in a new >> paper--not enough room here), which appears to qualitatively change >> the network processing states, and cognitive capabilities. Well, >> that's the claim. >> >> The larger point, is that apparently architectures interact with >> learning rules, in ways that can cross this symbolic/neural river of >> styx, without falling into it. >> >> Steve >> >> >> On 2/5/22 10:38 AM, Gary Marcus wrote: >>> There is no magic in understanding, just computation that has been >>> realized in the wetware of humans and that eventually can be >>> realized in machines. But understanding is not (just) learning. >>> >>> Understanding incorporates (or works in tandem with) learning - but >>> also, critically, in tandem with inference, /and the development and >>> maintenance of cognitive models/.Part of developing an understanding >>> of cats in general is to learn long term-knowledge about their >>> properties, both directly (e.g., through observation) and indirectly >>> (eg through learning factsabout animals in general that can be >>> extended to cats), often through inference (if all animals have DNA, >>> and a cat is an animal, it must also have DNA).The understanding of >>> a particular cat also involves direct observation, but also >>> inference (egone might surmise that the reason that Fluffy is >>> running about the room is that Fluffy suspects there is a mouse >>> stirring somewhere nearby). But all of that, I would say, is >>> subservient to the construction of cognitive models that can be >>> routinely updated (e.g., Fluffy is currently in the living room, >>> skittering about, perhaps looking for a mouse). >>> >>> ?In humans, those dynamic, relational models, which form part of an >>> understanding, can support inference (if Fluffy is in the living >>> room, we can infer that Fluffy is not outside, not lost, etc). >>> Without such models - which I think represent a core part of >>> understanding - AGI is an unlikely prospect. >>> >>> Current neural networks, as it happens, are better at acquiring >>> long-term knowledge (cats have whiskers) than they are at >>> dynamically updating cognitive models in real-time. LLMs like GPT-3 >>> etc lack the kind of dynamic model that I am describing. To a modest >>> degree they can approximate it on the basis of large samples of >>> texts, but their ultimate incoherence stems from the fact that they >>> do not have robust internal cognitive models that they can update on >>> the fly. >>> >>> Without such cognitive models you can still capture some aspects of >>> understanding (eg predicting that cats are likely to be furry), but >>> things fall apart quickly; inference is never reliable, and >>> coherence is fleeting. >>> >>> As a final note, one of the most foundational challenges in >>> constructing adequate cognitive models of the world is to have a >>> clear distinction between individuals and kinds; as I emphasized 20 >>> years ago (in The Algebraic Mind), this has always been a weakness >>> in neural networks, and I don?t think that the type-token problem >>> has yet been solved. >>> >>> Gary >>> >>> >>>> On Feb 5, 2022, at 01:31, Asim Roy wrote: >>>> >>>> ? >>>> >>>> All, >>>> >>>> I think the broader question was ?understanding.? Here are two >>>> Youtube videos showing simple robots ?learning? to walk. They are >>>> purely physical systems. Do they ?understand? anything ? such as >>>> the need to go around an obstacle, jumping over an obstacle, >>>> walking up and down stairs and so on? By the way, they ?learn? to >>>> do these things on their own, literally unsupervised, very much >>>> like babies. The basic question is: what is ?understanding? if not >>>> ?learning?? Is there some other mechanism (magic) at play in our >>>> brain that helps us ?understand?? >>>> >>>> https://www.youtube.com/watch?v=gn4nRCC9TwQ >>>> >>>> >>>> https://www.youtube.com/watch?v=8sO7VS3q8d0 >>>> >>>> >>>> Asim Roy >>>> >>>> Professor, Information Systems >>>> >>>> Arizona State University >>>> >>>> Lifeboat Foundation Bios: Professor Asim Roy >>>> >>>> >>>> Asim Roy | iSearch (asu.edu) >>>> >>>> >>>> *From:* Ali Minai >>>> *Sent:* Friday, February 4, 2022 11:38 PM >>>> *To:* Asim Roy >>>> *Cc:* Gary Marcus ; Danko Nikolic >>>> ; Brad Wyble ; >>>> connectionists at mailman.srv.cs.cmu.edu; AIhub >>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>>> Geoff Hinton >>>> >>>> Asim >>>> >>>> Of course there's nothing magical about understanding, and the mind >>>> has to emerge from the physical system, but our AI models at this >>>> point are not even close to realizing how that happens. We are, at >>>> best, simulating a superficial approximation of a few parts of the >>>> real thing. A single, integrated system where all the aspects of >>>> intelligence emerge from the same deep, well-differentiated >>>> physical substrate is far beyond our capacity. Paying more >>>> attention to neurobiology will be essential to get there, but so >>>> will paying attention to development - both physical and cognitive >>>> - and evolution. The configuration of priors by evolution is key to >>>> understanding how real intelligence learns so quickly and from so >>>> little. This is not an argument for using genetic algorithms to >>>> design our systems, just for understanding the tricks evolution has >>>> used and replicating them by design. Development is more feasible >>>> to do computationally, but hardly any models have looked at it >>>> except in a superficial sense. Nature creates basic intelligence >>>> not so much by configuring functions by explicit training as by >>>> tweaking, modulating, ramifying, and combining existing ones in a >>>> multi-scale self-organization process. We then learn much more >>>> complicated things (like playing chess) by exploiting that >>>> substrate, and using explicit instruction or learning by practice. >>>> The fundamental lesson of complex systems is that complexity is >>>> built in stages - each level exploiting the organization of the >>>> level below it. We see it in evolution, development, societal >>>> evolution, the evolution of technology, etc. Our approach in AI, in >>>> contrast, is to initialize a giant, naive system and train it to do >>>> something really complicated - but really specific - by training >>>> the hell out of it. Sure, now we do build many systems on top of >>>> pre-trained models like GPT-3 and BERT, which is better, but those >>>> models were again trained by the same none-to-all process I decried >>>> above. Contrast that with how humans acquire language, and how they >>>> integrate it into their *entire* perceptual, cognitive, and >>>> behavioral repertoire, not focusing just on this or that task. The >>>> age of symbolic AI may have passed, but the reductionistic mindset >>>> has not. We cannot build minds by chopping it into separate verticals. >>>> >>>> FTR, I'd say that the emergence of models such as GLOM and Hawkins >>>> and Ahmed's "thousand brains" is a hopeful sign. They may not be >>>> "right", but they are, I think, looking in the right direction. >>>> With a million miles to go! >>>> >>>> Ali >>>> >>>> *Ali A. Minai, Ph.D.* >>>> Professor and Graduate Program Director >>>> Complex Adaptive Systems Lab >>>> Department of Electrical Engineering & Computer Science >>>> >>>> 828 Rhodes Hall >>>> >>>> University of Cincinnati >>>> Cincinnati, OH 45221-0030 >>>> >>>> >>>> Phone: (513) 556-4783 >>>> Fax: (513) 556-7326 >>>> Email: Ali.Minai at uc.edu >>>> minaiaa at gmail.com >>>> >>>> WWW: https://eecs.ceas.uc.edu/~aminai/ >>>> >>>> >>>> On Fri, Feb 4, 2022 at 2:42 AM Asim Roy >>> > wrote: >>>> >>>> First of all, the brain is a physical system. There is no >>>> ?magic? inside the brain that does the ?understanding? part. >>>> Take for example learning to play tennis. You hit a few balls - >>>> some the right way and some wrong ? but you fairly quickly >>>> learn to hit them right most of the time. So there is obviously >>>> some simulation going on in the brain about hitting the ball in >>>> different ways and ?learning? its consequences. What you are >>>> calling ?understanding? is really these simulations about >>>> different scenarios. It?s also very similar to augmentation >>>> used to train image recognition systems where you rotate >>>> images, obscure parts and so on, so that you still can say it?s >>>> a cat even though you see only the cat?s face or whiskers or a >>>> cat flipped on its back. So, if the following questions relate >>>> to ?understanding,? you can easily resolve this by simulating >>>> such scenarios when ?teaching? the system. There?s nothing >>>> ?magical? about ?understanding.? As I said, bear in mind that >>>> the brain, after all, is a physical system and ?teaching? and >>>> ?understanding? is embodied in that physical system, not >>>> outside it. So ?understanding? is just part of ?learning,? >>>> nothing more. >>>> >>>> DANKO: >>>> >>>> What would happen to the hat if the hamster rolls on its back? >>>> (Would the hat fall off?) >>>> >>>> What would happen to the red hat when the hamster enters its >>>> lair? (Would the hat fall?off?) >>>> >>>> What would happen to that hamster when it goes foraging? (Would >>>> the red hat have an influence on finding food?) >>>> >>>> What would happen in a situation of being chased by a predator? >>>> (Would it be easier for predators to spot the hamster?) >>>> >>>> Asim Roy >>>> >>>> Professor, Information Systems >>>> >>>> Arizona State University >>>> >>>> Lifeboat Foundation Bios: Professor Asim Roy >>>> >>>> >>>> Asim Roy | iSearch (asu.edu) >>>> >>>> >>>> *From:* Gary Marcus >>> > >>>> *Sent:* Thursday, February 3, 2022 9:26 AM >>>> *To:* Danko Nikolic >>> > >>>> *Cc:* Asim Roy >; >>>> Geoffrey Hinton >>> >; AIhub >>> >; >>>> connectionists at mailman.srv.cs.cmu.edu >>>> >>>> *Subject:* Re: Connectionists: Stephen Hanson in conversation >>>> with Geoff Hinton >>>> >>>> Dear Danko, >>>> >>>> Well said. I had a somewhat similar response to Jeff Dean?s >>>> 2021 TED talk, in which he said (paraphrasing from memory, >>>> because I don?t remember the precise words) that the famous 200 >>>> Quoc Le unsupervised model >>>> [https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf >>>> ] >>>> had learned the concept of a ca. In reality the model had >>>> clustered together some catlike images based on the image >>>> statistics that it had extracted, but it was a long way from a >>>> full, counterfactual-supporting concept of a cat, much as you >>>> describe below. >>>> >>>> I fully agree with you that the reason for even having a >>>> semantics is as you put it, "to 1) learn with a few examples >>>> and 2) apply the knowledge to a broad set of situations.? GPT-3 >>>> sometimes gives the appearance of having done so, but it falls >>>> apart under close inspection, so the problem remains unsolved. >>>> >>>> Gary >>>> >>>> On Feb 3, 2022, at 3:19 AM, Danko Nikolic >>>> > >>>> wrote: >>>> >>>> G. Hinton wrote: "I believe that any reasonable person >>>> would admit that if you ask a neural net to draw a picture >>>> of a hamster wearing a red hat and it draws such a picture, >>>> it understood?the request." >>>> >>>> I would like to suggest why drawing a?hamster with a >>>> red?hat does not necessarily imply understanding of the >>>> statement "hamster wearing a red hat". >>>> >>>> To understand that "hamster wearing a red hat" would mean >>>> inferring, in newly?emerging situations of this hamster, >>>> all the real-life implications?that the red hat brings to >>>> the little animal. >>>> >>>> What would happen to the hat if the hamster rolls on its >>>> back? (Would the hat fall off?) >>>> >>>> What would happen to the red hat when the hamster enters >>>> its lair? (Would the hat fall?off?) >>>> >>>> What would happen to that hamster when it goes foraging? >>>> (Would the red hat have an influence on finding food?) >>>> >>>> What would happen in a situation of being chased by a >>>> predator? (Would it be easier for predators to spot the >>>> hamster?) >>>> >>>> ...and so on. >>>> >>>> Countless many questions can be asked. One has understood >>>> "hamster wearing a red hat" only if one can answer >>>> reasonably well many of such real-life relevant questions. >>>> Similarly, a student has?understood materias in a class >>>> only if they can apply the materials in real-life >>>> situations (e.g., applying Pythagora's theorem). If a >>>> student gives a correct answer to a multiple?choice >>>> question, we don't know whether the student understood the >>>> material or whether this was just rote learning (often, it >>>> is rote learning). >>>> >>>> I also suggest that understanding also comes together with >>>> effective learning: We store new information in such a way >>>> that we can recall it later and use it effectively? i.e., >>>> make good inferences in newly emerging situations based on >>>> this knowledge. >>>> >>>> In short: Understanding makes us humans able to 1) learn >>>> with a few examples and 2) apply the knowledge to a broad >>>> set of situations. >>>> >>>> No neural network?today has such capabilities and we don't >>>> know how to give them such capabilities. Neural networks >>>> need large amounts of training?examples that cover a large >>>> variety of situations and then the?networks can only deal >>>> with what the training examples have already covered. >>>> Neural networks cannot extrapolate in that 'understanding' >>>> sense. >>>> >>>> I suggest that understanding truly extrapolates from a >>>> piece of knowledge. It is not about satisfying a task such >>>> as translation between languages or drawing hamsters with >>>> hats. It is how you got the capability to complete the >>>> task: Did you only have a few examples that covered >>>> something different but related and then you extrapolated >>>> from that knowledge? If yes, this is going in the direction >>>> of understanding. Have you seen countless examples and then >>>> interpolated among them? Then perhaps it is not understanding. >>>> >>>> So, for the case of drawing a hamster wearing a red hat, >>>> understanding perhaps would have taken place if the >>>> following happened before that: >>>> >>>> 1) first, the network learned about hamsters (not many >>>> examples) >>>> >>>> 2) after that the network learned about red hats (outside >>>> the context of hamsters and without many examples) >>>> >>>> 3) finally the network learned about drawing (outside of >>>> the context of hats and hamsters, not many examples) >>>> >>>> After that, the network is asked to draw a hamster with a >>>> red hat. If it does it successfully, maybe we have started >>>> cracking the problem of understanding. >>>> >>>> Note also that this requires the network to learn >>>> sequentially without exhibiting catastrophic forgetting of >>>> the previous knowledge, which is possibly also a >>>> consequence of human learning by understanding. >>>> >>>> Danko >>>> >>>> Dr. Danko Nikoli? >>>> www.danko-nikolic.com >>>> >>>> https://www.linkedin.com/in/danko-nikolic/ >>>> >>>> >>>> --- A progress usually starts with an insight --- >>>> >>>> >>>> >>>> >>>> >>>> Virus-free. www.avast.com >>>> >>>> >>>> >>>> On Thu, Feb 3, 2022 at 9:55 AM Asim Roy >>> > wrote: >>>> >>>> Without getting into the specific dispute between Gary >>>> and Geoff, I think with approaches similar to GLOM, we >>>> are finally headed in the right direction. There?s >>>> plenty of neurophysiological evidence for single-cell >>>> abstractions and multisensory neurons in the brain, >>>> which one might claim correspond to symbols. And I >>>> think we can finally reconcile the decades old dispute >>>> between Symbolic AI and Connectionism. >>>> >>>> GARY: (Your GLOM, which as you know I praised publicly, >>>> is in many ways an effort to wind up with encodings >>>> that effectively serve as symbols in exactly that way, >>>> guaranteed to serve as consistent representations of >>>> specific concepts.) >>>> >>>> GARY: I have /never/ called for dismissal of neural >>>> networks, but rather for some hybrid between the two >>>> (as you yourself contemplated in 1991); the point of >>>> the 2001 book was to characterize exactly where >>>> multilayer perceptrons succeeded and broke down, and >>>> where symbols could complement them. >>>> >>>> Asim Roy >>>> >>>> Professor, Information Systems >>>> >>>> Arizona State University >>>> >>>> Lifeboat Foundation Bios: Professor Asim Roy >>>> >>>> >>>> Asim Roy | iSearch (asu.edu) >>>> >>>> >>>> *From:* Connectionists >>>> >>> > >>>> *On Behalf Of *Gary Marcus >>>> *Sent:* Wednesday, February 2, 2022 1:26 PM >>>> *To:* Geoffrey Hinton >>> > >>>> *Cc:* AIhub >>> >; >>>> connectionists at mailman.srv.cs.cmu.edu >>>> >>>> *Subject:* Re: Connectionists: Stephen Hanson in >>>> conversation with Geoff Hinton >>>> >>>> Dear Geoff, and interested others, >>>> >>>> What, for example, would you make of a system that >>>> often?drew the red-hatted hamster you requested, and >>>> perhaps a fifth of the time gave you utter nonsense?? >>>> Or say one that you trained to create birds but >>>> sometimes output stuff like this: >>>> >>>> >>>> >>>> One could >>>> >>>> a. avert one?s eyes and deem the anomalous outputs >>>> irrelevant >>>> >>>> or >>>> >>>> b. wonder if it might be possible that sometimes the >>>> system gets the right answer for the wrong reasons (eg >>>> partial historical contingency), and wonder whether >>>> another approach might be indicated. >>>> >>>> Benchmarks are harder than they look; most of the field >>>> has come to recognize that. The Turing Test has turned >>>> out to be a lousy measure of intelligence, easily >>>> gamed. It has turned out empirically that the Winograd >>>> Schema Challenge did not measure common sense as well >>>> as Hector might have thought. (As it happens, I am a >>>> minor coauthor of a very recent review on this very >>>> topic: https://arxiv.org/abs/2201.02387 >>>> ) >>>> But its conquest in no way means machines now have >>>> common sense; many people from many different >>>> perspectives recognize that (including, e.g., Yann >>>> LeCun, who generally tends to be more aligned with you >>>> than with me). >>>> >>>> So: on the goalpost of the Winograd schema, I was >>>> wrong, and you can quote me; but what you said about me >>>> and machine translation remains your invention, and it >>>> is inexcusable that you simply ignored my 2019 >>>> clarification. On the essential goal of trying to reach >>>> meaning and understanding, I remain unmoved; the >>>> problem remains unsolved. >>>> >>>> All of the problems LLMs have with coherence, >>>> reliability, truthfulness, misinformation, etc stand >>>> witness to that fact. (Their persistent inability to >>>> filter out toxic and insulting remarks stems from the >>>> same.) I am hardly the only person in the field to see >>>> that progress on any given benchmark does not >>>> inherently mean that the deep underlying problems have >>>> solved. You, yourself, in fact, have occasionally made >>>> that point. >>>> >>>> With respect to embeddings: Embeddings are very good >>>> for natural language /processing/; but NLP is not the >>>> same as NL/U/ ? when it comes to /understanding/, their >>>> worth is still an open question. Perhaps they will turn >>>> out to be necessary; they clearly aren?t sufficient. In >>>> their extreme, they might even collapse into being >>>> symbols, in the sense of uniquely identifiable >>>> encodings, akin to the ASCII code, in which a specific >>>> set of numbers stands for a specific word or concept. >>>> (Wouldn?t that be ironic?) >>>> >>>> (Your GLOM, which as you know I praised publicly, is in >>>> many ways an effort to wind up with encodings that >>>> effectively serve as symbols in exactly that way, >>>> guaranteed to serve as consistent representations of >>>> specific concepts.) >>>> >>>> Notably absent from your email is any kind of apology >>>> for misrepresenting my position. It?s fine to say that >>>> ?many people thirty years ago once thought X? and >>>> another to say ?Gary Marcus said X in 2015?, when I >>>> didn?t. I have consistently felt throughout our >>>> interactions that you have mistaken me for Zenon >>>> Pylyshyn; indeed, you once (at NeurIPS 2014) apologized >>>> to me for having made that error. I am still not he. >>>> >>>> Which maybe connects to the last point; if you read my >>>> work, you would see thirty years of arguments >>>> /for/?neural networks, just not in the way that you >>>> want them to exist. I have ALWAYS argued that there is >>>> a role for them; ?characterizing me as a person >>>> ?strongly?opposed to neural networks? misses the whole >>>> point of my 2001 book, which was subtitled ?Integrating >>>> Connectionism and Cognitive Science.? >>>> >>>> In the last two decades or so you have insisted (for >>>> reasons you have never fully clarified, so far as I >>>> know) on abandoning symbol-manipulation, but the >>>> reverse is not the case: I have /never/ called for >>>> dismissal of neural networks, but rather for some >>>> hybrid between the two (as you yourself contemplated in >>>> 1991); the point of the 2001 book was to characterize >>>> exactly where multilayer perceptrons succeeded and >>>> broke down, and where symbols could complement them. >>>> It?s a rhetorical trick (which is what the previous >>>> thread was about) to pretend otherwise. >>>> >>>> Gary >>>> >>>> On Feb 2, 2022, at 11:22, Geoffrey Hinton >>>> >>> > wrote: >>>> >>>> ? >>>> >>>> Embeddings are just vectors of soft feature >>>> detectors and they are very good for NLP. The quote >>>> on my webpage from Gary's 2015 chapter implies the >>>> opposite. >>>> >>>> A few decades ago, everyone I knew then would?have >>>> agreed that the ability to translate a sentence >>>> into many different languages was strong evidence >>>> that you understood it. >>>> >>>> But once neural networks could do that, their >>>> critics moved the goalposts. An exception is Hector >>>> Levesque who defined the goalposts more sharply by >>>> saying that the ability to get pronoun references >>>> correct in Winograd sentences is a crucial test. >>>> Neural nets are improving at that but still have >>>> some way to go. Will Gary agree that when they can >>>> get pronoun references?correct in Winograd >>>> sentences they really?do understand? Or does he >>>> want to reserve the right to weasel out of that too? >>>> >>>> Some people, like Gary, appear to be >>>> strongly?opposed to neural networks because?they do >>>> not fit their preconceived notions of how the mind >>>> should work. >>>> >>>> I believe that any reasonable person would admit >>>> that if you ask a neural net to draw a picture of a >>>> hamster wearing a red hat and it draws such a >>>> picture, it understood?the request. >>>> >>>> Geoff >>>> >>>> On Wed, Feb 2, 2022 at 1:38 PM Gary Marcus >>>> > >>>> wrote: >>>> >>>> Dear AI Hub, cc: Steven Hanson and Geoffrey >>>> Hinton, and the larger neural network community, >>>> >>>> There has been a lot of recent discussion on >>>> this list about framing and scientific >>>> integrity. Often the first step in >>>> restructuring narratives is to bully and >>>> dehumanize critics. The second is to >>>> misrepresent their position. People in >>>> positions of power are sometimes tempted to do >>>> this. >>>> >>>> The Hinton-Hanson interview that you just >>>> published is a real-time example of just that. >>>> It opens with a needless and largely >>>> content-free personal attack on a single >>>> scholar (me), with the explicit intention of >>>> discrediting that person. Worse, the only >>>> substantive thing it says is false. >>>> >>>> Hinton says ?In 2015 he [Marcus] made a >>>> prediction that computers wouldn?t be able to >>>> do machine translation.? >>>> >>>> I never said any such thing. >>>> >>>> What I predicted, rather, was that multilayer >>>> perceptrons, as they existed then, would not >>>> (on their own, absent other mechanisms) >>>> /understand/?language. Seven years later, they >>>> still haven?t, except in the most superficial way. >>>> >>>> I made no comment whatsoever about machine >>>> translation, which I view as a separate >>>> problem, solvable to a certain degree by >>>> correspondance without semantics. >>>> >>>> I specifically tried to clarify Hinton?s >>>> confusion in 2019, but, disappointingly, he has >>>> continued to purvey misinformation despite that >>>> clarification. Here is what I wrote privately >>>> to him then, which should have put the matter >>>> to rest: >>>> >>>> You have taken a single out of context quote >>>> [from 2015] and misrepresented it. The quote, >>>> which you have prominently displayed at the >>>> bottom on your own web page, says: >>>> >>>> Hierarchies of features are less suited to >>>> challenges such as language, inference, and >>>> high-level planning. For example, as Noam >>>> Chomsky famously pointed out, language is >>>> filled with sentences you haven't seen >>>> before.?Pure classifier systems don't know what >>>> to do with such sentences. The talent of >>>> feature detectors -- in??identifying which >>>> member of some category something belongs to -- >>>> doesn't translate into understanding >>>> novel??sentences, in which each sentence has >>>> its own unique meaning. >>>> >>>> It does /not/?say "neural nets would not be >>>> able to deal with novel sentences"; it says >>>> that hierachies of features detectors (on their >>>> own, if you read the context of the essay) >>>> would have trouble /understanding >>>> /novel?sentences. >>>> >>>> Google Translate does yet not /understand/?the >>>> content of?the sentences is translates. It >>>> cannot reliably answer questions about who did >>>> what to whom, or why, it cannot infer the order >>>> of the events in paragraphs, it can't determine >>>> the internal consistency of those events, and >>>> so forth. >>>> >>>> Since then, a number of scholars, such as the >>>> the computational linguist Emily Bender, have >>>> made similar points, and indeed current LLM >>>> difficulties with misinformation, incoherence >>>> and fabrication all follow from these concerns. >>>> Quoting from Bender?s prizewinning 2020 ACL >>>> article on the matter with Alexander Koller, >>>> https://aclanthology.org/2020.acl-main.463.pdf >>>> , >>>> also emphasizing issues of understanding and >>>> meaning: >>>> >>>> /The success of the large neural language >>>> models on many NLP tasks is exciting. However, >>>> we find that these successes sometimes lead to >>>> hype in which these models are being described >>>> as ?understanding? language or capturing >>>> ?meaning?. In this position paper, we argue >>>> that a system trained only on form has a priori >>>> no way to learn meaning. .. a clear >>>> understanding of the distinction between form >>>> and meaning will help guide the field towards >>>> better science around natural language >>>> understanding. / >>>> >>>> Her later article with Gebru on language models >>>> ?stochastic parrots? is in some ways an >>>> extension of this point; machine translation >>>> requires mimicry, true understanding (which is >>>> what I was discussing in 2015) requires >>>> something deeper than that. >>>> >>>> Hinton?s intellectual error here is in equating >>>> machine translation with the deeper >>>> comprehension that robust natural language >>>> understanding will require; as Bender and >>>> Koller observed, the two appear not to be the >>>> same. (There is a longer discussion of the >>>> relation between language understanding and >>>> machine translation, and why the latter has >>>> turned out to be more approachable than the >>>> former, in my 2019 book with Ernest Davis). >>>> >>>> More broadly, Hinton?s ongoing dismissiveness >>>> of research from perspectives other than his >>>> own (e.g. linguistics) have done the field a >>>> disservice. >>>> >>>> As Herb Simon once observed, science does not >>>> have to be zero-sum. >>>> >>>> Sincerely, >>>> >>>> Gary Marcus >>>> >>>> Professor Emeritus >>>> >>>> New York University >>>> >>>> On Feb 2, 2022, at 06:12, AIhub >>>> >>> > wrote: >>>> >>>> ? >>>> >>>> Stephen Hanson in conversation with Geoff >>>> Hinton >>>> >>>> In the latest episode of this video series >>>> for AIhub.org >>>> , >>>> Stephen Hanson talks to? Geoff Hinton?about >>>> neural networks, backpropagation, >>>> overparameterization, digit recognition, >>>> voxel cells, syntax and semantics, Winograd >>>> sentences, and more. >>>> >>>> You can watch the discussion, and read the >>>> transcript, here: >>>> >>>> https://aihub.org/2022/02/02/what-is-ai-stephen-hanson-in-conversation-with-geoff-hinton/ >>>> >>>> >>>> About AIhub: >>>> >>>> AIhub is a non-profit dedicated to >>>> connecting the AI community to the public >>>> by providing free, high-quality information >>>> through AIhub.org >>>> >>>> (https://aihub.org/ >>>> ). >>>> We help researchers publish the latest AI >>>> news, summaries of their work, opinion >>>> pieces, tutorials and more.? We are >>>> supported by many leading scientific >>>> organizations in AI, namely AAAI >>>> , >>>> NeurIPS >>>> , >>>> ICML >>>> , >>>> AIJ >>>> /IJCAI >>>> , >>>> ACM SIGAI >>>> , >>>> EurAI/AICOMM, CLAIRE >>>> >>>> and RoboCup >>>> . >>>> >>>> Twitter: @aihuborg >>>> >>>> >>>> >>>> >>>> >>>> Virus-free. www.avast.com >>>> >>>> >>>> >> -- >> > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From el-ghazali.talbi at univ-lille.fr Wed Feb 9 15:02:02 2022 From: el-ghazali.talbi at univ-lille.fr (El-ghazali Talbi) Date: Wed, 9 Feb 2022 21:02:02 +0100 Subject: Connectionists: OLA'2022@Sicilia Extended deadline Message-ID: <3cfd9da2-e062-c31e-6af8-a2c75402f9ed@univ-lille.fr> Apologies for cross-posting. Appreciate if you can distribute this CFP to your network. **************************************************************************************** ????????????????????????? OLA'2022 ????????? International Conference on Optimization and Learning ????????????????????????? 18-20 July 2022 ????????????????????? Syracuse (Sicilia), Italy ??????????????? http://ola2022.sciencesconf.org/ ??????????????????? SCOPUS Springer Proceedings **************************************************************************************** OLA is a conference focusing on the future challenges of optimization and learning methods and their applications. The conference OLA'2022 will provide an opportunity to the international research community in optimization and learning to discuss recent research results and to develop new ideas and collaborations in a friendly and relaxed atmosphere. OLA'2022 welcomes presentations that cover any aspects of optimization and learning research such as big optimization and learning, optimization for learning, learning for optimization, optimization and learning under uncertainty, deep learning, new high-impact applications, parameter tuning, 4th industrial revolution, computer vision, hybridization issues, optimization-simulation, meta-modeling, high-performance computing, parallel and distributed optimization and learning, surrogate modeling, multi-objective optimization ... Submission papers: We will accept two different types of submissions: -?????? S1: Extended abstracts of work-in-progress and position papers of a maximum of 3 pages -?????? S2: Original research contributions of a maximum of 10 pages Important dates: =============== Paper submission extended deadline???? Feb 25, 2022 Notification of acceptance??? March 25, 2022 Proceedings: Accepted papers in categories S1 and S2 will be published in the proceedings. A SCOPUS and DBLP indexed Springer book will be published for accepted long papers. Proceedings will be available at the conference. -- ********************************************************************** OLA'2022 International Conference on Optimization and Learning (SCOPUS, Springer) 18-20 July 2022, Syracuse, Sicilia, Italy http://ola2022.sciencesconf.org *********************************************************************** Prof. El-ghazali TALBI Polytech'Lille, University Lille - INRIA CRISTAL - CNRS From publicity at acsos.org Wed Feb 9 16:51:39 2022 From: publicity at acsos.org (ACSOS Conference) Date: Wed, 9 Feb 2022 16:51:39 -0500 Subject: Connectionists: Autonomic Computing & Self-Organizing Systems: ACSOS 2022 CFP Message-ID: *** ACSOS 2022 - Call For Papers *** 3rd IEEE International Conference on Autonomic Computing and Self-Organizing Systems 19-23 September 2022 ? Online https://2022.acsos.org/https://twitter.com/ACSOSconf ************************************* The goal of the IEEE International Conference on Autonomic Computing and Self-Organizing Systems (ACSOS) is to provide a forum for sharing the latest research results, ideas and experiences in autonomic computing, self-adaptation and self-organization. ACSOS was founded in 2020 as a merger of the IEEE International Conference on Autonomic Computing (ICAC) and the IEEE International Conference on Self-Adaptive and Self-Organizing Systems (SASO). For more up-to-date news, follow us at https://twitter.com/ACSOSconf! *** Important Dates *** April 29, 2022: Abstract submission deadline May 6, 2022: Paper submission deadline July 2, 2022: Notification to authors August 5, 2022: Camera Ready Deadline September 19-23, 2022: ACSOS Conference! *** Challenge and Scope *** Emerging large-scale systems (including data centers, cloud computing, smart cities, cyber-physical systems, sensor networks, and embedded or pervasive environments) are becoming increasingly complex, heterogeneous, and difficult to manage. The challenges of designing, controlling, managing, monitoring, and evolving such complex systems in a principled way led the scientific community to look for inspiration in diverse fields, such as biology, biochemistry, physics, complex systems, control theory, artificial intelligence, and sociology. To address these challenges novel modeling and engineering techniques are needed that help to understand how local behavior and global behavior relate to each other. Such models and practices are a key condition for understanding, controlling, and designing the emergent behavior in autonomic and self-adaptive systems. The mission of ACSOS is to provide an interdisciplinary forum for researchers and industry practitioners to address these challenges to make resources, applications, and systems more autonomic, self-adaptive,? and self-organizing. ACSOS provides a venue to share and present their experiences, discuss challenges, and report state-of-the-art and in-progress research. The conference program will include technical research papers, in-practice experience reports, posters, demos, and a doctoral symposium.We invite novel contributions related to the fundamental understanding of autonomic computing, self-adaption and self-organization along with principles and practices of their engineering and application. The topics of interest include, but are not limited to: - Autonomic and Self-* system properties: robustness; resilience; resource efficiency; stability; anti-fragility; diversity; self-reference and reflection; emergent behavior; computational awareness and self-awareness; - Autonomic and Self-* systems theory: bio-inspired and socially-inspired paradigms and heuristics; theoretical frameworks and models; formal languages; queuing and control theory; requirement and goal expression techniques; uncertainty as a 1st class entity - Autonomic and Self-* systems engineering: reusable mechanisms and algorithms; design patterns; programming languages; architectures; operating systems and middleware; testing and validation methodologies; runtime models; techniques for assurance; platforms and toolkits; multi-agent systems; - Data-driven management and artificial intelligence: data mining; machine learning; in-network learning; distributed reinforcement learning; data science and other statistical techniques to analyze, understand, and manage the behavior of complex systems or establishing self-awareness; - Mechanisms and principles for self-organization and self-adaptation: inter-operation of self-* mechanisms; evolution, logic, and learning; addressing large-scale and decentralized systems; - Socio-technical self-* systems: human and social factors; visualization; crowdsourcing and collective awareness; - Autonomic and self-* concepts applied to hardware systems: self-* materials; self-construction; reconfigurable hardware, self-* properties for quantum computing; - Self-adaptive cybersecurity: intrusion detection, malware attribution, zero-trust networks and blockchain-based approaches, privacy in self-* systems; - Cross disciplinary research: approaches that draw inspiration from complex systems, organic computing, artificial intelligence, chemistry, psychology, sociology, and biology, and ethology. We invite research papers applying autonomic and self-* approaches to a wide range of application areas, including (but not limited to): - Smart environments: -grids, -cities, -homes, and -manufacturing; - Internet of things and cyber-physical systems;- Robotics, autonomous vehicles, and traffic management; - Cloud, fog/edge computing, High Performance Computing (HPC), quantum computing and data centers;- Internet of Things;- Hypervisors, containerization services, orchestration, operating systems, and middleware;- Biological and bio-inspired systems. *** Best Papers *** We intend to continue the tradition of giving the best papers of the conference an opportunity to publish an extended version in a special issue of ACM Transactions on Autonomous and Adaptive Systems (TAAS). The Karsten Schwan Best Paper Award will be awarded to a selected paper. *** Submission Instructions *** *Research papers* (up to 10 pages including images, tables, and references) should present novel ideas in the cross-disciplinary research context described in this call, motivated by problems from current practice or applied research. *Experience Reports* (up to 10 pages including images, tables, and references) cover innovative implementations, novel applications, interesting performance results and experience in applying recent research advance to practical situations on any topics of interest. *Vision Papers* (up to 6 pages including images, tables, and references) introduce ground-shaking, provocative, and even controversial ideas; discuss long term perspectives and challenges; focus on overlooked or underrepresented areas, and foster debate. Research papers and experience reports will be included in the conference proceedings that will be published by IEEE Computer Society Press and made available as a part of the IEEE Digital Library. Vision papers will be part of a separate proceedings volume (the ACSOS Companion).Please see https://2022.acsos.org/track/acsos-2022-papers#Call-for-Papers for full requirements and review criteria. -------------- next part -------------- An HTML attachment was scrubbed... URL: From phitzler at googlemail.com Wed Feb 9 10:57:05 2022 From: phitzler at googlemail.com (Pascal Hitzler) Date: Wed, 9 Feb 2022 09:57:05 -0600 Subject: Connectionists: Call for Papers: Special Issue on Neuro-Symbolic Artificial Intelligence and the Semantic Web In-Reply-To: References: Message-ID: Call for Papers: Special Issue on Neuro-Symbolic Artificial Intelligence and the Semantic Web Semantic Web journal. Deadline 1 August, 2022 Guest editors: Monireh Ebrahimi, IBM, USA Pascal Hitzler, Kansas State University, USA Md Kamruzzaman Sarker, University of Hartford, USA Daria Stepanova, Bosch Center for AI, Germany Further details: http://www.semantic-web-journal.net/blog/call-papers-special-issue-neuro-symbolic-artificial-intelligence-and-semantic-web Best Regards, Pascal. -- Pascal Hitzler Lloyd T. Smith Creativity in Engineering Chair Director, Center for AI and Data Science Kansas State University http://www.pascal-hitzler.de http://www.daselab.org http://www.semantic-web-journal.net From ioannakoroni at csd.auth.gr Thu Feb 10 01:30:22 2022 From: ioannakoroni at csd.auth.gr (Ioanna Koroni) Date: Thu, 10 Feb 2022 08:30:22 +0200 Subject: Connectionists: =?utf-8?q?Live_AIDA_e-Lecture_by_Prof=2E_Fredrik_?= =?utf-8?q?Heintz=3A_=E2=80=9CTowards_Trustworthy_AI_=E2=80=93_Inte?= =?utf-8?q?grating_Reasoning_and_Learning=E2=80=9D=2C_22nd_February?= =?utf-8?q?_2022_17=3A00-18=3A00_CET?= References: <003101d8182b$bea340e0$3be9c2a0$@csd.auth.gr> Message-ID: <00c401d81e47$af6eb8d0$0e4c2a70$@csd.auth.gr> Dear AI scientist/engineer/student/enthusiast, Prof. Fredrik Heintz, a prominent AI researcher internationally, will deliver the e-lecture: ?Towards Trustworthy AI - Integrating Reasoning and Learning?, on Tuesday 22nd February 2022 17:00-18:00 CET (8:00-9:00 am PST), (12:00 am-1:00am CST), see details in: http://www.i-aida.org/event_cat/ai-lectures/ You can join for free using the zoom link: https://authgr.zoom.us/j/91262043831 & Passcode: 148148 The International AI Doctoral Academy (AIDA), a joint initiative of the European R&D projects AI4Media, ELISE, Humane AI Net, TAILOR and VISION, is very pleased to offer you top quality scientific lectures on several current hot AI topics. Lectures are typically held once per week, Tuesdays 17:00-18:00 CET (8:00-9:00 am PST), (12:00 am-1:00am CST). Attendance is free. The lectures are disseminated through multiple channels and email lists (we apologize if you received it through various channels). If you want to stay informed on future lectures, you can register in the email lists AIDA email list and CVML email list. Best regards Profs. M. Chetouani, P. Flach, B. O?Sullivan, I. Pitas, N. Sebe -- This email has been checked for viruses by Avast antivirus software. https://www.avast.com/antivirus -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Wed Feb 9 17:18:43 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Wed, 9 Feb 2022 17:18:43 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Dear Gary, As my reply to Asim Roy indicated, the parts and whole problem that Geoff Hinton considered is ill-posed since it bypasses how a brain network segments the "whole" from 1000 parts in the cluttered scene. Only 10 parts belong to the whole. The relation problem has also been solved and mathematically proven if one understands emergent universal Turing machines using a Developmental Network (DN). The solution to relation is a special case of the solution to the compositionality problem which is a special case of the emergent universal Turing machine. I am not telling you "a son looks like his father because the father makes money to feed the son". The solution is supported by biology and a mathematical proof. Best regards, -John Date: Mon, 7 Feb 2022 07:57:34 -0800 From: Gary Marcus To: Juyang Weng Cc: Post Connectionists Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Content-Type: text/plain; charset="utf-8" Dear John, I agree with you that cluttered scenes are critical, but Geoff?s GLOM paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf] might actually have some relevance. It may well be that we need to do a better job with parts and whole before we can fully address clutter, and Geoff is certainly taking that question seriously. Geoff?s ?Stable islands of identical vectors? do sound suspiciously like symbols to me (in a good way!), but regardless, they seem to me to be a plausible candidate as a foundation for coping with clutter. And not just cluttered scenes, but also relations between multiple objects in a scene, which is another example of the broader issue you raise, challenging for pure MLPs but critical for deeper AI. Gary -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Wed Feb 9 16:22:29 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Wed, 9 Feb 2022 16:22:29 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Gary, The examples you mentioned in the email have already been COMPLETELY solved and mathematically proven by the emergent Turing machine theory using a Developmental Network (DN): J. Weng, "Brain as an Emergent Finite Automaton: A Theory and Three Theorems'' *International Journal of Intelligence Science*, vol. 5, no. 2, pp. 112-131, Jan. 2015. PDF file . (This is the journal version of the above IJCNN paper. It explains that the control of a Turing Machine, regular or universal, is an Finite Automaton and therefore a DN can learn any universal Turing Machine one transition at a time, immediately, error-free if there is sufficient neuronal resource, and optimal in the sense of maximum likelihood if there is no sufficient neuronal resource.) This is a good news. A bad news: Psychologists, neuroscientists and neural network researchers who have not learned the theory of encoding a universal Turing machine are not able to understand the fundamental solution. That is why you feel comfortable with only intuitive examples. Please suggest how to address this fundamental communication problem facing this community. You are not alone on this list. Best regards, -John ---- Date: Mon, 7 Feb 2022 14:07:27 -0800 From: Gary Marcus To: Stephen Jos? Hanson Cc: AIhub , connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Content-Type: text/plain; charset="utf-8" Stephen, I don?t doubt for a minute that deep learning can characterize some aspects of psychology reasonably well; but either it needs to expands its borders or else be used in conjunction with other techniques. Take for example the name of the new Netflix show The Woman in the House Across the Street from the Girl in the Window Most of us can infer, compositionally, from that unusually long noun phrase, that the title is a description of particular person, that the title is not a complete sentence, and that the woman in question lives in a house; we also infer that there is a second, distinct person (likely a child) across the street, and so forth. We can also use some knowledge of pragmatics to infer that the woman in question is likely to be the protagonist in the show. Current systems still struggle with that sort of thing. We can then watch the show (I watched a few minutes of Episode 1) and quickly relate the title to the protagonist?s mental state, start to develop a mental model of the protagonist?s relation to her new neighbors, make inferences about whether certain choices appear to be ?within character?, empathize with character or question her judgements, etc, all with respect to a mental model that is rapidly encoded and quickly modified. I think that an understanding of how people build and modify such models would be extremely valuable (not just for fiction for everyday reality), but I don?t see how deep learning in its current form gives us much purchase on that. There is plenty of precedent for the kind of mental processes I am sketching (e.g Walter Kintsch?s work on text comprehension; Kamp/Kratzer/Heim work on discourse representation, etc) from psychological and linguistic perspectives, but almost no current contact in the neural network community with these well-attested psychological processes. Gary -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From malini.vinita.samarasinghe at ini.ruhr-uni-bochum.de Thu Feb 10 02:59:04 2022 From: malini.vinita.samarasinghe at ini.ruhr-uni-bochum.de (Vinita Samarasinghe) Date: Thu, 10 Feb 2022 08:59:04 +0100 Subject: Connectionists: Postdoc position - Computational Neuroscience - Ruhr University Bochum Message-ID: <93031b71-9336-a02d-35fa-60e8dab25dd0@ini.ruhr-uni-bochum.de> Deadline Extended! Prof. Sen Cheng, Institute for Neural Computation at the Ruhr University Bochum, invites applications for a full time *Postdoctoral position* (TV-L E13) in Computational Neuroscience. The position starts on July 1, 2022 and is funded for three years. The successful applicant will work on a collaborative project (within the Collaborative Research Center ?Extinction Learning? (SFB 1280 )), together with experimentalists to: * analyze learning dynamics in behavioral, neural, and psychophysiological data, which will be collected by other projects within the SFB 1280, * compare the learning dynamics between individuals, species, learning phases and learning paradigms, * develop algorithms to analyze the learning dynamics, * develop and study computational models of learning dynamics, * coordinate research with other participating projects. Candidates must have: * a doctorate degree in neuroscience, physics, mathematics, electrical/biomedical engineering or a closely related field, * relevant experience in mathematical modeling, * excellent programming skills (e.g., Python, C/C++, Matlab), * excellent communication skills in English, * the ability to work well in a team. Research experience in neuroscience would be a further asset. The position is third party funded and does not have any formal teaching duties attached. The research group is highly dynamic and uses diverse computational modeling approaches including biological neural networks, cognitive modeling, and machine learning to investigate learning and memory in humans and animals. For further information see www.rub.de/cns. The Ruhr University Bochum is home to a vibrant research community in neuroscience and cognitive science. The Institute for Neural Computation is an independent research unit and combines different areas of expertise ranging from experimental and theoretical neuroscience to machine learning and robotics. The Institute for Neural Computation focuses on the dynamics and learning of perception and behavior on a functional level but is otherwise very diverse, ranging from neurophysiology and psychophysics over computational neuroscience to machine learning and technical applications. Please send your application, including CV, transcripts and research statement electronically, as a *single PDF file*, to *samarasinghe at ini.rub.de*. In addition, at least two academic references must be sent independently to the above email address. The deadline for applications is *February 20, 2022*. Travel costs for interviews will not be reimbursed. The Ruhr University Bochum is committed to equal opportunity. We strongly encourage applications from qualified women and persons with disabilities. We are committed to providing a supportive work environment for female researchers, in particular those with young children. Our university provides mentoring and coaching opportunities specifically aimed at women in research. We have a strong research network with female role models and will provide opportunities to network with them. Wherever possible, events will be scheduled during regular childcare hours. Special childcare will be arranged if events have to be scheduled outside of regular hours, in case of sickness and during school or daycare closures. Where childcare is not an option parents will be offered a home office solution. If you have any questions please feel free to get in touch with Vinita Samarasinghe (contact below) -- Vinita Samarasinghe M.Sc., M.A. Science Manager Arbeitsgruppe Computational Neuroscience Institut f?r Neuroinformatik Ruhr-Universit?t Bochum, NB 3/73 Postfachnummer 110 Universit?tstr. 150 D-44801 Bochum Tel: +49 (0)234 32 27996 Email:samarasinghe at ini.rub.de -------------- next part -------------- An HTML attachment was scrubbed... URL: From shu-chen.li at tu-dresden.de Thu Feb 10 03:07:45 2022 From: shu-chen.li at tu-dresden.de (Shu-Chen Li) Date: Thu, 10 Feb 2022 08:07:45 +0000 Subject: Connectionists: Postdoc position in Chair of Lifespan Dev Neuroscience at TU Dresden, Germany Message-ID: <2C9DBE82-43B9-4313-970F-22255F570A5B@tu-dresden.de> Dear all, We have an open postdoc position in the Chair of Lifespan Developmental Neuroscience in the Faculty of Psychology (starting from April 1, 2022 or later, for 3 years with the possibility of extension). The position is in the fields of neurocognitive development or aging. The specific research focus could be on the topics of perception, memory, cognitive control, decision making or motivation. Details about the position can be found in the announcement that is attached. The position may also be suitable for advanced PhD fellows who will soon receive their degrees. More details about the position can be found at this link: https://tu-dresden.de/mn/psychologie/ipep/epsy/ressourcen/dateien/pdfs-stellenangebote/Li_WIMI_eng_w036_202204.pdf?lang=en I would deeply appreciate if you could circulate the announcement to young researchers whom you think the position could be of interest. With best regards, Shu-Chen Li _______________________________________________________________________________________________________________________ Prof. Shu-Chen Li, Ph.D. Lehrstuhl Entwicklungspsychologie und Neurowissenschaft der Lebensspanne (Chair of Lifespan Developmental Neuroscience) Fakult?t Psychologie (Faculty of Psychology) Zellescher Weg 17, Rm. A233 D-01062 Dresden, Germany E-mail: Shu-Chen.Li at tu-dresden.de URL: http://tu-dresden.de/mn/psy/epsy Tel.: +49-351-46334162/Fax:+49-351-46342194 Excellence Cluster ? Centre for Tactile Internet with Human-in-the-Loop URL: https://www.ceti.one Technische Universit?t Dresden (TU Dresden) _________________________________________________________________________________________________________________________ -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Thu Feb 10 03:01:49 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Thu, 10 Feb 2022 08:01:49 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear John, We can deal with cluttered scenes. And we can also identify parts of wholes in these scenes. Here are some example scenes. In the first two scenes, we can identify the huskies along with the ears, eyes, legs, faces and so on. In the satellite image below, we can identify parts of the planes like the fuselage, tail, wing and so on. That?s the fundamental part of DARPA?s XAI model ? to be able to identify the parts to confirm the whole object. And if you can identify the parts, a school bus will never become an ostrich with change of a few pixels. So you get a lot of things with Explainable models of this form ? a symbolic XAI model, robustness against adversarial attacks, and a model that you can trust. Explainable AI of this form can become the best defense against adversarial attacks. You may not need any adversarial training of any kind. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) [A dog and a cat lying on a bed Description automatically generated with low confidence] [A wolf walking in the snow Description automatically generated with medium confidence] [An aerial view of a city Description automatically generated with medium confidence] From: Connectionists On Behalf Of Juyang Weng Sent: Wednesday, February 9, 2022 3:19 PM To: Post Connectionists Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary, As my reply to Asim Roy indicated, the parts and whole problem that Geoff Hinton considered is ill-posed since it bypasses how a brain network segments the "whole" from 1000 parts in the cluttered scene. Only 10 parts belong to the whole. The relation problem has also been solved and mathematically proven if one understands emergent universal Turing machines using a Developmental Network (DN). The solution to relation is a special case of the solution to the compositionality problem which is a special case of the emergent universal Turing machine. I am not telling you "a son looks like his father because the father makes money to feed the son". The solution is supported by biology and a mathematical proof. Best regards, -John Date: Mon, 7 Feb 2022 07:57:34 -0800 From: Gary Marcus > To: Juyang Weng > Cc: Post Connectionists > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: > Content-Type: text/plain; charset="utf-8" Dear John, I agree with you that cluttered scenes are critical, but Geoff?s GLOM paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf] might actually have some relevance. It may well be that we need to do a better job with parts and whole before we can fully address clutter, and Geoff is certainly taking that question seriously. Geoff?s ?Stable islands of identical vectors? do sound suspiciously like symbols to me (in a good way!), but regardless, they seem to me to be a plausible candidate as a foundation for coping with clutter. And not just cluttered scenes, but also relations between multiple objects in a scene, which is another example of the broader issue you raise, challenging for pure MLPs but critical for deeper AI. Gary -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.jpg Type: image/jpeg Size: 27597 bytes Desc: image001.jpg URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image002.jpg Type: image/jpeg Size: 91893 bytes Desc: image002.jpg URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image004.jpg Type: image/jpeg Size: 399723 bytes Desc: image004.jpg URL: From mtkostecki at gmail.com Thu Feb 10 04:38:49 2022 From: mtkostecki at gmail.com (Mateusz Kostecki) Date: Thu, 10 Feb 2022 10:38:49 +0100 Subject: Connectionists: Transatlantic Behavioural Neuroscience Summer School Message-ID: Hello! We are super happy to announce our flagship event for 2022 - *Transatlantic Behavioural Neuroscience Summer School *(Sept 1-10 2022) that we are organizing in cooperation with our friends from the *University of Buenos Aires and NeuroElectronics Research Flanders.* The field of behavioural neuroscience is changing. It?s changing on the conceptual level ? ?neuroscience needs behaviour?, a bold claim in 2017, has become a truism. Behavioural paradigms became much more complex. This change is driven by the recent explosion of new tools. With stunning precision, we can record both the behaviour and neural activity of an animal performing various complex tasks. And the most important aspect of the tool revolution is its open and worldwide character ? the field of behavioural neuroscience is being democratised on an unprecedented scale. Since money is becoming less of a constraint, the exploration of new theories and ideas is not limited anymore to privileged institutions. Our Transatlantic Behavioural Neuroscience Summer School is designed to embrace and foster these changes. *Our school will be composed of intense, hands-on tutorials on behavioural experiment design and data analysis*. The main project of the school will be the assembly and configuration of a setup for fruit flies. This project will show how to design a setup to be versatile, and how to build it from scratch. These experiments will also provide data for more complex offline analysis using machine learning tools like DeepLabCut. We want to give the students the knowledge and tools to modify and improve their own setups and remove obstacles when designing future experiments. We will also host fantastic lecturers specializing in different aspects of behavioural studies: - Anne von Philipsborn (Aarhus University) - Ahmed el Hady (University of Konstanz) - Daniel Tomsic (University of Buenos Aires) - Karolina Socha (Uiversity College London) - Basil el Jundi (University of Wuerzburg) - Irene Jacobsen (Kavli Institute) - Alicja Pu?cian (Nencki Institute) - Ofer Yizhar (Weizmann Instiute) The school will be organized on Sept 1-10th 2022 in Wojciechy, in a beautiful lake region of Masuria. The fee is EUR 750 (we are now also securing funds for fee waivers). The deadline for applications will be* March 29th 2022*. If you have any questions, contact us at openlab at nencki.edu.pl. Please find the application form and kmore info here *- https://nenckiopenlab.org/tbnss/ .* Please help us to spread the news! See you there! Nencki Open Lab Team -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: TBNSS poster.png Type: image/png Size: 2537544 bytes Desc: not available URL: From ASIM.ROY at asu.edu Fri Feb 11 00:54:30 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Fri, 11 Feb 2022 05:54:30 +0000 Subject: Connectionists: A New Society for Explainability, Safety and Trust in AI + a possible inaugural conference in San Francisco in late July, early August Message-ID: Dear Colleagues, Some of us are thinking of forming a new Society for Explainability, Safety and Trust in AI. We already have some members of this community show interest in such a society. We plan to have our first conference in late July or early August in San Francisco. And a separate journal will follow that. You can email me if you have interest. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) -------------- next part -------------- An HTML attachment was scrubbed... URL: From has168 at eng.ucsd.edu Thu Feb 10 16:29:01 2022 From: has168 at eng.ucsd.edu (Hao Su) Date: Thu, 10 Feb 2022 13:29:01 -0800 Subject: Connectionists: =?utf-8?q?=5BCall_for_Paper=5D_ICLR_2022_workshop?= =?utf-8?q?_on_=E2=80=9CGeneralizable_Policy_Learning_in_the_Physic?= =?utf-8?b?YWwgV29ybGTigJ0=?= Message-ID: We are excited to announce the first workshop on ?Generalizable Policy Learning in the Physical World? at ICLR 2022! We are calling for papers. If you work on robotics, learning, vision, and other relevant fields, please consider submitting. We accept both 4-page short papers and 8-page long papers. The submission deadline is Feb 25, 2022. Check our workshop website for more information: https://ai-workshops.github.io/generalizable-policy-learning-in-the-physical-world/ Introduction to our workshop: Generalization is particularly important when learning policies to interact with the physical world. The spectrum of such policies is broad: the policies can be high-level, such as action plans that concern temporal dependencies and casualties of environment states; or low-level, such as object manipulation skills to transform objects that are rigid, articulated, soft, or even fluid. In the physical world, an embodied agent can face a number of changing factors such as physical parameters, action spaces, tasks, visual appearances of the scenes, geometry and topology of the objects, etc. And many important real-world tasks involving generalizable policy learning, e.g., visual navigation, object manipulation, and autonomous driving. Therefore, learning generalizable policies is crucial to developing intelligent embodied agents in the real world. Our main targeted participants are researchers interested in applying learning methods to develop intelligent embodied agents in the physical world. More specifically, target communities include, but are not limited to: robotics, reinforcement learning, learning from demonstrations, offline reinforcement learning, meta-learning, multi-task learning, 3D vision, computer vision, computer graphics, and physical simulation. In affiliation to this workshop, we are also organizing the ManiSkill Challenge, which focuses on learning to manipulate unseen objects in simulation with 3D visual inputs. We will announce winners and host winner presentations in this workshop. Have a wonderful day and stay safe throughout the pandemic! Best, Hao Su on behalf of the organizing committee -- Hao Su Assistant Professor Department of Computer Science and Engineering Jacobs School of Engineering University of California, San Diego -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Thu Feb 10 10:23:03 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Thu, 10 Feb 2022 07:23:03 -0800 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <873A06DA-D7BC-4090-B2DE-49179685DBF8@nyu.edu> Try your algorithm on datasets such as this https://ai.facebook.com/blog/introducing-unidentified-video-objects-a-new-benchmark-for-open-world-object-segmentation/ If you produce strong empirical results, the community will take notice. > On Feb 9, 2022, at 10:54 PM, Juyang Weng wrote: > > ? > Dear Gary, > > As my reply to Asim Roy indicated, the parts and whole problem that Geoff Hinton considered is ill-posed since it bypasses how a brain network segments the "whole" from 1000 parts in the cluttered scene. Only 10 parts belong to the whole. > > The relation problem has also been solved and mathematically proven if one understands emergent universal Turing machines using a Developmental Network (DN). The solution to relation is a special case of the solution to the compositionality problem which is a special case of the emergent universal Turing machine. > > I am not telling you "a son looks like his father because the father makes money to feed the son". The solution is supported by biology and a mathematical proof. > > Best regards, > -John > > Date: Mon, 7 Feb 2022 07:57:34 -0800 > From: Gary Marcus > To: Juyang Weng > Cc: Post Connectionists > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > Message-ID: > Content-Type: text/plain; charset="utf-8" > > Dear John, > > I agree with you that cluttered scenes are critical, but Geoff?s GLOM paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf] might actually have some relevance. It may well be that we need to do a better job with parts and whole before we can fully address clutter, and Geoff is certainly taking that question seriously. > > Geoff?s ?Stable islands of identical vectors? do sound suspiciously like symbols to me (in a good way!), but regardless, they seem to me to be a plausible candidate as a foundation for coping with clutter. > > And not just cluttered scenes, but also relations between multiple objects in a scene, which is another example of the broader issue you raise, challenging for pure MLPs but critical for deeper AI. > > Gary > > -- > Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From malini.vinita.samarasinghe at ini.ruhr-uni-bochum.de Fri Feb 11 07:35:02 2022 From: malini.vinita.samarasinghe at ini.ruhr-uni-bochum.de (Vinita Samarasinghe) Date: Fri, 11 Feb 2022 13:35:02 +0100 Subject: Connectionists: 12 Postdoctoral positions - Ruhr University Bochum Message-ID: <8274b070-b65e-a094-4459-c3d6622f31cb@ini.rub.de> *12 postdoctoral positions (min. 3 years) in Neuroscience, Psychology, and Philosophy of Mind*** The Ruhr University Bochum and the University of Duisburg-Essen are among Germany?s leading research universities. They draw their strength from both the diversity and the proximity of natural sciences, humanities, and engineering disciplines on coherent campuses. These highly dynamic settings enable researchers to work across traditional boundaries of academic subjects and faculties. *We are searching for advanced postdocs who ideally already have ca. 2 years of research experience*and want to continue their careers in an inspiring environment that fosters interdisciplinary approaches. We aim to install an outstanding group of postdocs, starting on 1^st of April 2022 at the earliest, who are tightly integrated with the research focus ?THINK at Ruhr?. Details of this research program as well as all further procedural details are outlined on the website: _*www.thinkatruhr.de* _. Deadline of application is February 27, 2022. *Applicants should have *an excellent PhD in a research topic related to neuroscience, psychology, or philosophy of mind and should already have published in high-ranking peer-reviewed journals. The postdoc positions offer a focus on research only (no teaching or administration). It is expected that the successful candidate will submit a grant proposal to establish her/his own research group (e.g. ERC Starting Grant, Emmy Noether Research Grant, Sofja Kovalevskaja Award, etc.) during the postdoctoral period. Candidate selection will be based on whether submission of such a proposal is a realistic goal in addition to usual measures of scientific excellence. The candidates will be associated with faculty members who are named on the website thinkatruhr.de. THINK at Ruhr will foster and support any research activities of the postdocs. Our universities offer attractive academic career track options and family friendly conditions. For more information including application procedure and contact information please visit https://thinkatruhr.de/what-we-offer/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From mpavone at dmi.unict.it Fri Feb 11 06:51:21 2022 From: mpavone at dmi.unict.it (Mario Pavone DMI) Date: Fri, 11 Feb 2022 12:51:21 +0100 Subject: Connectionists: =?utf-8?q?Metaheuristics_International_Conference?= =?utf-8?q?_2022_=40=C2=A0Ortigia-Syracuse=2C_Italy?= Message-ID: Apologies for cross-posting. Appreciate if you can distribute this CFP to your network. ********************************************************* MIC 2022 - 14th Metaheuristics International Conference 11-14 July 2022, Ortigia-Syracuse, Italy https://www.ANTs-lab.it/mic2022/ mic2022 at ANTs-lab.it ********************************************************* ** Submission deadline: 30th March 2022 ** ** Proceedings in LNCS Volume, Springer ** ** Special Issue in ITOR journal ** ** 5 Plenary Speakers ** ** SUBMISSION SYSTEM ALREADY OPEN ** https://www.easychair.org/conferences/?conf=mic2022 *Scope of the Conference ======================== The Metaheuristics International Conference (MIC) conference series was established in 1995 and this is its 14th edition! MIC is nowadays the main event focusing on the progress of the area of Metaheuristics and their applications. As in all previous editions, provides an opportunity to the international research community in Metaheuristics to discuss recent research results, to develop new ideas and collaborations, and to meet old and make new friends in a friendly and relaxed atmosphere. Considering the particular moment, the conference will be held in presence and online mode. Of course, in case the conference will be held in presence, the organizing committee will ensure compliance of all safety conditions. MIC 2022 is focus on presentations that cover different aspects of metaheuristic research such as new algorithmic developments, high-impact and original applications, new research challenges, theoretical developments, implementation issues, and in-depth experimental studies. MIC 2022 strives a high-quality program that will be completed by a number of invited talks, tutorials, workshops and special sessions. *Plenary Speakers ======================== + Christian Blum, Artificial Intelligence Research Institute (IIIA), Spanish National Research Council (CSIC) + Salvatore Greco, University of Catania, Italy + Kalyanmoy Deb, Michigan State University, USA + Holger H. Hoos, Leiden University, The Netherlands + El-Ghazali Talbi, University of Lille, France Important Dates ================ Submission deadline March 30th, 2022 Notification of acceptance May 10th, 2022 Camera ready copy May 25th, 2022 Early registration May 25th , 2022 Submission Details =================== MIC 2022 accepts submissions in three different formats: S1) Regular paper: novel and original research contributions of a maximum of 15 pages (LNCS format) S2) Short paper: extended abstract of novel research works of 6 pages (LNCS format) S3) Oral/Poster presentation: high-quality manuscripts that have recently, within the last year, been submitted or accepted for journal publication. All papers must be prepared using Lecture Notes in Computer Science (LNCS) template, and must be submitted in PDF at the link: https://www.easychair.org/conferences/?conf=mic2022 Proceedings and special issue ============================ Accepted papers in categories S1 and S2 will be published as post-proceedings in Lecture Notes in Computer Science series by Springer. Accepted contributions of category S3 will be considered for oral or poster presentations at the conference based on the number received and the slots available, and will not be included into the LNCS proceedings. An electronic book instead will be prepared by the MIC 2022 organizing committee, and made available on the website. In addition, a post-conference special issue in International Transactions in Operational Research (ITOR) will be considered for the significantly extended and revised versions of selected accepted papers from categories S1 and S2. Conference Location ==================== MIC 2022 will be held in the beautiful Ortigia island, the historical centre of the city of Syracuse, Sicily-Italy. Syracuse is very famous for its ancient ruins, with particular reference to the Roman Amphitheater, Greek Theatre, and the Orecchio di Dionisio (Ear of Dionisio) that is a limestone cave shaped like a human ear. Syracuse is also the city where the greatest mathematician Archimede was born. https://www.siracusaturismo.net/multimedia_lista.asp MIC'2022 Conference Chairs ============================== Conference Chairs - Luca Di Gaspero, University of Undine, Italy - Paola Festa, University of Naples, Italy - Amir Nakib, Universit? Paris Est Cr?teil, France - Mario Pavone, University of Catania, Italy -- Mario F. Pavone, PhD Associate Professor Dept of Mathematics and Computer Science University of Catania V.le A. Doria 6 - 95125 Catania, Italy --------------------------------------------- tel: +39 095 7383034 mobile: +39 3384342147 Email: mpavone at dmi.unict.it http://www.dmi.unict.it/mpavone/ FB: https://www.facebook.com/mfpavone Skype: mpavone ========================================================= MIC 2022 - 14th International Metaheuristics Conference 11-14 July 2022, Ortigia-Syracuse, Italy https://www.ants-lab.it/mic2022/ ========================================================= -------------- next part -------------- An HTML attachment was scrubbed... URL: From laurent.perrinet at univ-amu.fr Fri Feb 11 09:11:10 2022 From: laurent.perrinet at univ-amu.fr (PERRINET Laurent) Date: Fri, 11 Feb 2022 14:11:10 +0000 Subject: Connectionists: crowd-sourcing COSYNE post-review score sheet In-Reply-To: <36CC6CD1-C12C-4E5E-BA9C-37E6FE5999F8@univ-amu.fr> References: <36CC6CD1-C12C-4E5E-BA9C-37E6FE5999F8@univ-amu.fr> Message-ID: <2E15398E-1E86-4695-A4F8-B23595E1187E@univ-amu.fr> Dear community, As of today, I have received N = 82 answers from the google form (out of them, 79 are valid) out of the 881 submitted abstracts. In short, the total score is simply the linear sum of the scores relatively weighted by the confidence levels (as stated in the email we received from the chairs) and the threshold is close to 6.05 this year: [2022-02-11_COSYNE-razor] More details in the notebook (or directly in this post) which can also be forked here and interactively modified on binder. cheers, Laurent -- Laurent Perrinet - INT (UMR 7289) AMU/CNRS https://laurentperrinet.github.io/ On 4 Feb 2022, at 09:19, PERRINET Laurent > wrote: Dear community COSYNE is a great conference which plays a pivotal role in our field. Raw numbers we were given are * 881 submitted abstracts * 215 independent reviewers * 2639 reviews If you have submitted an abstract (or several) you have recently received your scores. I am not affiliated to COSYNE - yet I would like to contribute in some way and would like to ask one minute of your time to report the raw scores from your reviewers: https://forms.gle/p7eG1p6dJAkr4Cyg7 (Do one form per abstract.) For this crowd-sourcing effort to have a most positive impact, I will share the results and summarize in a few lines them in one week time (11/02). The more numerous your feedbacks, the higher the precision of results! Thanks in advance for your action, Laurent PS: if any similar initiative already exists, I'll be more than willing to receive feedback -- Laurent Perrinet - INT (UMR 7289) AMU/CNRS https://laurentperrinet.github.io/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From iswc.conf at gmail.com Fri Feb 11 11:20:34 2022 From: iswc.conf at gmail.com (International Semantic Web Conference) Date: Fri, 11 Feb 2022 11:20:34 -0500 Subject: Connectionists: CfP ISWC 2022- Call for Doctoral Consortium Papers Message-ID: *CfP: 21st International Semantic Web Conference (ISWC 2022)* Hangzhou, China, October 23-27, 2022 https://iswc2022.semanticweb.org/ Call for Doctoral Consortium papers ***************************************** The ISWC 2022 Doctoral Consortium (DC) will take place virtually as part of the 21st International Semantic Web Conference. This forum will provide PhD students an opportunity to: - present and discuss their research ideas in a supportive, formative and yet critical environment; - receive feedback from mentors, typically senior members of the Semantic Web research community, and peers; - explore career pathways available after completing their PhD degree, and finally - network and build collaborations with other members of the community. The event is intended for students who have articulated a reasonably detailed research proposal, preferably supported by some preliminary results. The aim is to support the students in refining their proposal and suggest possible ways to improve their research plan and achieve results with prospective greater impact. While doctoral degrees can vary in format and conduct, we aim this Call for Papers to PhD candidates who will have already partially investigated some specific problems. Students will be required to submit a paper to the doctoral consortium, structured like a research proposal (see ?Submission Details? below). All proposals submitted to the Doctoral Consortium will undergo a rigorous review process by the International Programme Committee, who will provide detailed and constructive feedback and select those submissions to be presented at the Doctoral Consortium. If accepted, students will have to register and attend the event, which will include a range of interactive activities. Call for Doctoral Consortium Papers: https://iswc2022.semanticweb.org/index.php/doctoral-consortium/ Deadline: Friday, 13th May, 2022, 23:59 AoE (Anywhere on Earth) DC Chairs: - Olaf Hartig, Link?ping University, Sweden - Oshani Seneviratne, Rensselaer Polytechnic Institute, USA Contact: DoctoralConsortium-iswc2022 at easychair.org Follow us on social media: - Twitter: @iswc_conf #iswc_conf (https://twitter.com/iswc_conf) - LinkedIn: https://www.linkedin.com/groups/13612370 - Facebook: https://www.facebook.com/ISWConf - Instagram: https://www.instagram.com/iswc_conf/ The ISWC 2022 Organizing Team Organizing Committee ? ISWC 2022 (semanticweb.org) -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Fri Feb 11 16:58:59 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Fri, 11 Feb 2022 21:58:59 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear John, 1. Let?s start with a simple case, say a dog, and enumerate how many possible parts and objects a dog would need to remember or recognize. 2. How many possible combinations did you use in your calculation for the DN? Best, Asim From: Juyang Weng Sent: Friday, February 11, 2022 12:33 PM To: Asim Roy ; John K Tsotsos Cc: connectionists at mailman.srv.cs.cmu.edu; Gary Marcus Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, Thank you for saying "we can". Please provide: (1) a neural network that does all you said "we can" and (2) the complexity analysis for all possible combinations among all possible parts and all possible objects This chain of conversations is very useful for those who are not yet familiar with the "complexity of vision" (NP hard) that John Tsotso wrote papers argued about. John Tsotso: Our DN solves this problem like a brain in a constant time (frame time)! The solution simply pops up. Best regards, -John On Thu, Feb 10, 2022 at 3:01 AM Asim Roy > wrote: Dear John, We can deal with cluttered scenes. And we can also identify parts of wholes in these scenes. Here are some example scenes. In the first two scenes, we can identify the huskies along with the ears, eyes, legs, faces and so on. In the satellite image below, we can identify parts of the planes like the fuselage, tail, wing and so on. That?s the fundamental part of DARPA?s XAI model ? to be able to identify the parts to confirm the whole object. And if you can identify the parts, a school bus will never become an ostrich with change of a few pixels. So you get a lot of things with Explainable models of this form ? a symbolic XAI model, robustness against adversarial attacks, and a model that you can trust. Explainable AI of this form can become the best defense against adversarial attacks. You may not need any adversarial training of any kind. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) [A dog and a cat lying on a bed Description automatically generated with low confidence] [A wolf walking in the snow Description automatically generated with medium confidence] [An aerial view of a city Description automatically generated with medium confidence] From: Connectionists > On Behalf Of Juyang Weng Sent: Wednesday, February 9, 2022 3:19 PM To: Post Connectionists > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary, As my reply to Asim Roy indicated, the parts and whole problem that Geoff Hinton considered is ill-posed since it bypasses how a brain network segments the "whole" from 1000 parts in the cluttered scene. Only 10 parts belong to the whole. The relation problem has also been solved and mathematically proven if one understands emergent universal Turing machines using a Developmental Network (DN). The solution to relation is a special case of the solution to the compositionality problem which is a special case of the emergent universal Turing machine. I am not telling you "a son looks like his father because the father makes money to feed the son". The solution is supported by biology and a mathematical proof. Best regards, -John Date: Mon, 7 Feb 2022 07:57:34 -0800 From: Gary Marcus > To: Juyang Weng > Cc: Post Connectionists > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: > Content-Type: text/plain; charset="utf-8" Dear John, I agree with you that cluttered scenes are critical, but Geoff?s GLOM paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf] might actually have some relevance. It may well be that we need to do a better job with parts and whole before we can fully address clutter, and Geoff is certainly taking that question seriously. Geoff?s ?Stable islands of identical vectors? do sound suspiciously like symbols to me (in a good way!), but regardless, they seem to me to be a plausible candidate as a foundation for coping with clutter. And not just cluttered scenes, but also relations between multiple objects in a scene, which is another example of the broader issue you raise, challenging for pure MLPs but critical for deeper AI. Gary -- Juyang (John) Weng -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.jpg Type: image/jpeg Size: 27597 bytes Desc: image001.jpg URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image002.jpg Type: image/jpeg Size: 91893 bytes Desc: image002.jpg URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image003.jpg Type: image/jpeg Size: 399723 bytes Desc: image003.jpg URL: From ASIM.ROY at asu.edu Fri Feb 11 21:49:04 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sat, 12 Feb 2022 02:49:04 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear John, If I understand correctly, all learning systems do something along the lines of maximum likelihood learning or error minimization, like your DN. What?s your point? JOHN: ?Of course, the brain network does not remember all shapes and all configurations of parts. That is why our DN must do maximum likelihood optimality, using a limited number of resources to best estimate such a huge space of cluttered scenes.? So, can your DN model identify the parts of objects in the cluttered images below? Here was my note: ASIM: ?And we can also identify parts of wholes in these scenes. Here are some example scenes. In the first two scenes, we can identify the huskies along with the ears, eyes, legs, faces and so on. In the satellite image below, we can identify parts of the planes like the fuselage, tail, wing and so on. That?s the fundamental part of DARPA?s XAI model ? to be able to identify the parts to confirm the whole object. And if you can identify the parts, a school bus will never become an ostrich with change of a few pixels. So you get a lot of things with Explainable models of this form ? a symbolic XAI model, robustness against adversarial attacks, and a model that you can trust. Explainable AI of this form can become the best defense against adversarial attacks. You may not need any adversarial training of any kind.? Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) [A dog and a cat lying on a bed Description automatically generated with low confidence] [A wolf walking in the snow Description automatically generated with medium confidence] [An aerial view of a city Description automatically generated with medium confidence] From: Juyang Weng Sent: Friday, February 11, 2022 3:40 PM To: Asim Roy Cc: John K Tsotsos ; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, We should not just assume your 1, a dog, since there are many articulated objects and each articulated object looks very different under many situations. There are further many other articulated objects other than the dog. You will see below. Since you want explainable AI, you must not start with a single object symbol like "dog". A symbol is "abstract" already. That is way Michael Jordan complained that neural networks do not abstract well. Let us do your 2. How many possible combinations? The number of shapes of a part: Suppose a part of a uniform color. It has m pixels along its boundary; each pixel has n possible positions, the number of shapes of a part is O(m^n), already exponential in n. We assumed that each pixel has the same color, which is not true. Suppose that each part is centered at location l, the number of combinations of p parts of your object (dog) is O(p^l), another exponential complexity. Thus the number of combinations of your object (dog) in a clean background, like Jeffe Hinton proposed to do, is already a product two two exponential complexities: O(m^n)O(p^l)=O(m^{n}p^{l}). Suppose that there are b objects in a cluttered scene, the number of combination of objects in a cluttered scene is [O(m^{n}p^{l})]^b=O(m^{nb}p^{lb}). Of course, the brain network does not remember all shapes and all configurations of parts. That is why our DN must do maximum likelihood optimality, using a limited number of resources to best estimate such a huge space of cluttered scenes. I am not talking about abstraction yet, which is another subject about "pop up" in the brain. I guess that many people on this list are not familiar with such a complexity analysis. Gary Marcus, sorry to overload you with this. That is what I said in several talks that the brain is an elephant and all disciplines are blind men. Even people with a PhD in computer science may not be skillful in such an exponential complexity in vision, since many computer science programs have dropped automata theory from their required course lists. I asked Gary Marcus to suggest how to solve this huge problem. But instead, he asked me to try a data set which is a dead end. Many have tried and are dead, like ImageNet. Best regards, -Joh On Fri, Feb 11, 2022 at 4:59 PM Asim Roy > wrote: Dear John, 1. Let?s start with a simple case, say a dog, and enumerate how many possible parts and objects a dog would need to remember or recognize. 2. How many possible combinations did you use in your calculation for the DN? Best, Asim From: Juyang Weng > Sent: Friday, February 11, 2022 12:33 PM To: Asim Roy >; John K Tsotsos > Cc: connectionists at mailman.srv.cs.cmu.edu; Gary Marcus > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, Thank you for saying "we can". Please provide: (1) a neural network that does all you said "we can" and (2) the complexity analysis for all possible combinations among all possible parts and all possible objects This chain of conversations is very useful for those who are not yet familiar with the "complexity of vision" (NP hard) that John Tsotso wrote papers argued about. John Tsotso: Our DN solves this problem like a brain in a constant time (frame time)! The solution simply pops up. Best regards, -John On Thu, Feb 10, 2022 at 3:01 AM Asim Roy > wrote: Dear John, We can deal with cluttered scenes. And we can also identify parts of wholes in these scenes. Here are some example scenes. In the first two scenes, we can identify the huskies along with the ears, eyes, legs, faces and so on. In the satellite image below, we can identify parts of the planes like the fuselage, tail, wing and so on. That?s the fundamental part of DARPA?s XAI model ? to be able to identify the parts to confirm the whole object. And if you can identify the parts, a school bus will never become an ostrich with change of a few pixels. So you get a lot of things with Explainable models of this form ? a symbolic XAI model, robustness against adversarial attacks, and a model that you can trust. Explainable AI of this form can become the best defense against adversarial attacks. You may not need any adversarial training of any kind. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) [A dog and a cat lying on a bed Description automatically generated with low confidence] [A wolf walking in the snow Description automatically generated with medium confidence] [An aerial view of a city Description automatically generated with medium confidence] From: Connectionists > On Behalf Of Juyang Weng Sent: Wednesday, February 9, 2022 3:19 PM To: Post Connectionists > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary, As my reply to Asim Roy indicated, the parts and whole problem that Geoff Hinton considered is ill-posed since it bypasses how a brain network segments the "whole" from 1000 parts in the cluttered scene. Only 10 parts belong to the whole. The relation problem has also been solved and mathematically proven if one understands emergent universal Turing machines using a Developmental Network (DN). The solution to relation is a special case of the solution to the compositionality problem which is a special case of the emergent universal Turing machine. I am not telling you "a son looks like his father because the father makes money to feed the son". The solution is supported by biology and a mathematical proof. Best regards, -John Date: Mon, 7 Feb 2022 07:57:34 -0800 From: Gary Marcus > To: Juyang Weng > Cc: Post Connectionists > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: > Content-Type: text/plain; charset="utf-8" Dear John, I agree with you that cluttered scenes are critical, but Geoff?s GLOM paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf] might actually have some relevance. It may well be that we need to do a better job with parts and whole before we can fully address clutter, and Geoff is certainly taking that question seriously. Geoff?s ?Stable islands of identical vectors? do sound suspiciously like symbols to me (in a good way!), but regardless, they seem to me to be a plausible candidate as a foundation for coping with clutter. And not just cluttered scenes, but also relations between multiple objects in a scene, which is another example of the broader issue you raise, challenging for pure MLPs but critical for deeper AI. Gary -- Juyang (John) Weng -- Juyang (John) Weng -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.jpg Type: image/jpeg Size: 27597 bytes Desc: image001.jpg URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image002.jpg Type: image/jpeg Size: 91893 bytes Desc: image002.jpg URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image003.jpg Type: image/jpeg Size: 399723 bytes Desc: image003.jpg URL: From timvogels at gmail.com Fri Feb 11 10:16:08 2022 From: timvogels at gmail.com (Tim Vogels) Date: Fri, 11 Feb 2022 16:16:08 +0100 Subject: Connectionists: crowd-sourcing COSYNE post-review score sheet In-Reply-To: <2E15398E-1E86-4695-A4F8-B23595E1187E@univ-amu.fr> References: <2E15398E-1E86-4695-A4F8-B23595E1187E@univ-amu.fr> Message-ID: Dear all, I shall neither confirm nor deny the validity of these results, but squinting my eyes, and not taking the various dimensions that we had to account for to make for a hopefully well-balanced conference, then yeah that looks like a very rough average of the several different thresholds we used. Once the dust settles and I?m not totally overwhelmed anymore (right!) Laura and I will sit down and write some sort of account of how things went down this year (follow us on Twitter - @visioncircuits and @tpvogels for updates on that). Thanks Laurent, for taking a first stab at this though, and keeping us on our toes. Did we get it right? Probably not in every way, and what we?ll write will be more more the next year than anything else. Do you have feedback? Please let us know, though I?d like to receive your criticism offline, if possible. It makes it much easier to digest. I hope this helps. All the best, Tim (sent from my phone) > On 11 Feb 2022, at 15:28, PERRINET Laurent wrote: > ? > Dear community, > > As of today, I have received N = 82 answers from the google form (out of them, 79 are valid) out of the 881 submitted abstracts. In short, the total score is simply the linear sum of the scores relatively weighted by the confidence levels (as stated in the email we received from the chairs) and the threshold is close to 6.05 this year: > > > > More details in the notebook (or directly in this post) which can also be forked here and interactively modified on binder. > > cheers, > > Laurent > > -- > Laurent Perrinet - INT (UMR 7289) AMU/CNRS > https://laurentperrinet.github.io/ > > > > >> On 4 Feb 2022, at 09:19, PERRINET Laurent wrote: >> >> Dear community >> >> COSYNE is a great conference which plays a pivotal role in our field. Raw numbers we were given are >> >> * 881 submitted abstracts >> * 215 independent reviewers >> * 2639 reviews >> >> If you have submitted an abstract (or several) you have recently received your scores. >> >> I am not affiliated to COSYNE - yet I would like to contribute in some way and would like to ask one minute of your time to report the raw scores from your reviewers: >> >> https://forms.gle/p7eG1p6dJAkr4Cyg7 >> >> (Do one form per abstract.) >> >> For this crowd-sourcing effort to have a most positive impact, I will share the results and summarize in a few lines them in one week time (11/02). The more numerous your feedbacks, the higher the precision of results! >> >> Thanks in advance for your action, >> Laurent >> >> >> PS: if any similar initiative already exists, I'll be more than willing to receive feedback >> >> >> -- >> Laurent Perrinet - INT (UMR 7289) AMU/CNRS >> https://laurentperrinet.github.io/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From sina.tootoonian at gmail.com Fri Feb 11 12:13:00 2022 From: sina.tootoonian at gmail.com (Sina Tootoonian) Date: Fri, 11 Feb 2022 17:13:00 +0000 Subject: Connectionists: Poster for Group Leader Position at the Francis Crick Institute Message-ID: The poster advertising the Group Leader position at the Crick was missing from the original message, please see attached. Thanks, Sina -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: FLIER Crick GL search 2022.pdf Type: application/pdf Size: 146951 bytes Desc: not available URL: From juyang.weng at gmail.com Fri Feb 11 14:16:27 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Fri, 11 Feb 2022 14:16:27 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Dear Gary, Sorry, you are still not aware of, or did not read, my work: No data sets are valid for conscious learning. Conscious learning must be learned on the fly. Please do not ask anybody to try any data sets any more since they are a dead end in AI. If you are always spoon fed, you will never be able to understand how to get food yourself. Please read my reasoning and proofs in my paper about the post-selection protocol flaw and the paper about conscious learning. J. Weng, "On Post Selections Using Test Sets (PSUTS) in AI", in Proc. International Joint Conference on Neural Networks, pp. 1-8, Shengzhen, China, July 18-22, 2021. PDF file . J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, USA, Jan.7-9, 2022. PDF file . Best regards, -John Date: Thu, 10 Feb 2022 07:23:03 -0800 From: Gary Marcus To: Juyang Weng Cc: Post Connectionists Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: <873A06DA-D7BC-4090-B2DE-49179685DBF8 at nyu.edu> Content-Type: text/plain; charset="utf-8" Try your algorithm on datasets such as this https://ai.facebook.com/blog/introducing-unidentified-video-objects-a-new-benchmark-for-open-world-object-segmentation/ If you produce strong empirical results, the community will take notice. > On Feb 9, 2022, at 10:54 PM, Juyang Weng wrote: > > ? > Dear Gary, > > As my reply to Asim Roy indicated, the parts and whole problem that Geoff Hinton considered is ill-posed since it bypasses how a brain network segments the "whole" from 1000 parts in the cluttered scene. Only 10 parts belong to the whole. > > The relation problem has also been solved and mathematically proven if one understands emergent universal Turing machines using a Developmental Network (DN). The solution to relation is a special case of the solution to the compositionality problem which is a special case of the emergent universal Turing machine. > > I am not telling you "a son looks like his father because the father makes money to feed the son". The solution is supported by biology and a mathematical proof. > > Best regards, > -John > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Fri Feb 11 14:33:06 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Fri, 11 Feb 2022 14:33:06 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear Asim, Thank you for saying "we can". Please provide: (1) a neural network that does all you said "we can" and (2) the complexity analysis for all possible combinations among all possible parts and all possible objects This chain of conversations is very useful for those who are not yet familiar with the "complexity of vision" (NP hard) that John Tsotso wrote papers argued about. John Tsotso: Our DN solves this problem like a brain in a constant time (frame time)! The solution simply pops up. Best regards, -John On Thu, Feb 10, 2022 at 3:01 AM Asim Roy wrote: > Dear John, > > > > We can deal with cluttered scenes. And we can also identify parts of > wholes in these scenes. Here are some example scenes. In the first two > scenes, we can identify the huskies along with the ears, eyes, legs, faces > and so on. In the satellite image below, we can identify parts of the > planes like the fuselage, tail, wing and so on. That?s the fundamental part > of DARPA?s XAI model ? to be able to identify the parts to confirm the > whole object. And if you can identify the parts, a school bus will never > become an ostrich with change of a few pixels. So you get a lot of things > with Explainable models of this form ? a symbolic XAI model, robustness > against adversarial attacks, and a model that you can trust. Explainable AI > of this form can become the best defense against adversarial attacks. You > may not need any adversarial training of any kind. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > [image: A dog and a cat lying on a bed Description automatically > generated with low confidence] [image: A wolf walking in the snow > Description automatically generated with medium confidence] [image: An > aerial view of a city Description automatically generated with medium > confidence] > > > > > > *From:* Connectionists *On > Behalf Of *Juyang Weng > *Sent:* Wednesday, February 9, 2022 3:19 PM > *To:* Post Connectionists > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Gary, > > > > As my reply to Asim Roy indicated, the parts and whole problem that Geoff > Hinton considered is ill-posed since it bypasses how a brain network > segments the "whole" from 1000 parts in the cluttered scene. Only 10 parts > belong to the whole. > > > > The relation problem has also been solved and mathematically proven if one > understands emergent universal Turing machines using a > Developmental Network (DN). The solution to relation is a special case of > the solution to the compositionality problem which is a special case of the > emergent universal Turing machine. > > > > I am not telling you "a son looks like his father because the father makes > money to feed the son". The solution is supported by biology and a > mathematical proof. > > > Best regards, > > -John > > > > Date: Mon, 7 Feb 2022 07:57:34 -0800 > From: Gary Marcus > To: Juyang Weng > Cc: Post Connectionists > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > Message-ID: > Content-Type: text/plain; charset="utf-8" > > Dear John, > > I agree with you that cluttered scenes are critical, but Geoff?s GLOM > paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf > ] > might actually have some relevance. It may well be that we need to do a > better job with parts and whole before we can fully address clutter, and > Geoff is certainly taking that question seriously. > > Geoff?s ?Stable islands of identical vectors? do sound suspiciously like > symbols to me (in a good way!), but regardless, they seem to me to be a > plausible candidate as a foundation for coping with clutter. > > And not just cluttered scenes, but also relations between multiple objects > in a scene, which is another example of the broader issue you raise, > challenging for pure MLPs but critical for deeper AI. > > Gary > > > > -- > > Juyang (John) Weng > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.jpg Type: image/jpeg Size: 27597 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image002.jpg Type: image/jpeg Size: 91893 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image004.jpg Type: image/jpeg Size: 399723 bytes Desc: not available URL: From juyang.weng at gmail.com Fri Feb 11 17:39:50 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Fri, 11 Feb 2022 17:39:50 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear Asim, We should not just assume your 1, a dog, since there are many articulated objects and each articulated object looks very different under many situations. There are further many other articulated objects other than the dog. You will see below. Since you want explainable AI, you must not start with a single object symbol like "dog". A symbol is "abstract" already. That is way Michael Jordan complained that neural networks do not abstract well. Let us do your 2. How many possible combinations? The number of shapes of a part: Suppose a part of a uniform color. It has m pixels along its boundary; each pixel has n possible positions, the number of shapes of a part is O(m^n), already exponential in n. We assumed that each pixel has the same color, which is not true. Suppose that each part is centered at location l, the number of combinations of p parts of your object (dog) is O(p^l), another exponential complexity. Thus the number of combinations of your object (dog) in a clean background, like Jeffe Hinton proposed to do, is already a product two two exponential complexities: O(m^n)O(p^l)=O(m^{n}p^{l}). Suppose that there are b objects in a cluttered scene, the number of combination of objects in a cluttered scene is [O(m^{n}p^{l})]^b=O(m^{nb}p^{lb}). Of course, the brain network does not remember all shapes and all configurations of parts. That is why our DN must do maximum likelihood optimality, using a limited number of resources to best estimate such a huge space of cluttered scenes. I am not talking about abstraction yet, which is another subject about "pop up" in the brain. I guess that many people on this list are not familiar with such a complexity analysis. Gary Marcus, sorry to overload you with this. That is what I said in several talks that the brain is an elephant and all disciplines are blind men. Even people with a PhD in computer science may not be skillful in such an exponential complexity in vision, since many computer science programs have dropped automata theory from their required course lists. I asked Gary Marcus to suggest how to solve this huge problem. But instead, he asked me to try a data set which is a dead end. Many have tried and are dead, like ImageNet. Best regards, -Joh On Fri, Feb 11, 2022 at 4:59 PM Asim Roy wrote: > Dear John, > > > > 1. Let?s start with a simple case, say a dog, and enumerate how many > possible parts and objects a dog would need to remember or recognize. > 2. How many possible combinations did you use in your calculation for > the DN? > > > > Best, > > Asim > > > > *From:* Juyang Weng > *Sent:* Friday, February 11, 2022 12:33 PM > *To:* Asim Roy ; John K Tsotsos > *Cc:* connectionists at mailman.srv.cs.cmu.edu; Gary Marcus < > gary.marcus at nyu.edu> > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Asim, > > > > Thank you for saying "we can". > Please provide: > > (1) a neural network that does all you said "we can" and > > (2) the complexity analysis for all possible combinations among all > possible parts and all possible objects > > This chain of conversations is very useful for those who are not yet > familiar with the "complexity of vision" (NP hard) that John Tsotso wrote > papers argued about. > > John Tsotso: > > Our DN solves this problem like a brain in a constant time (frame time)! > The solution simply pops up. > > > > Best regards, > > -John > > > > On Thu, Feb 10, 2022 at 3:01 AM Asim Roy wrote: > > Dear John, > > > > We can deal with cluttered scenes. And we can also identify parts of > wholes in these scenes. Here are some example scenes. In the first two > scenes, we can identify the huskies along with the ears, eyes, legs, faces > and so on. In the satellite image below, we can identify parts of the > planes like the fuselage, tail, wing and so on. That?s the fundamental part > of DARPA?s XAI model ? to be able to identify the parts to confirm the > whole object. And if you can identify the parts, a school bus will never > become an ostrich with change of a few pixels. So you get a lot of things > with Explainable models of this form ? a symbolic XAI model, robustness > against adversarial attacks, and a model that you can trust. Explainable AI > of this form can become the best defense against adversarial attacks. You > may not need any adversarial training of any kind. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > [image: A dog and a cat lying on a bed Description automatically > generated with low confidence] [image: A wolf walking in the snow > Description automatically generated with medium confidence] [image: An > aerial view of a city Description automatically generated with medium > confidence] > > > > > > *From:* Connectionists *On > Behalf Of *Juyang Weng > *Sent:* Wednesday, February 9, 2022 3:19 PM > *To:* Post Connectionists > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Gary, > > > > As my reply to Asim Roy indicated, the parts and whole problem that Geoff > Hinton considered is ill-posed since it bypasses how a brain network > segments the "whole" from 1000 parts in the cluttered scene. Only 10 parts > belong to the whole. > > > > The relation problem has also been solved and mathematically proven if one > understands emergent universal Turing machines using a > Developmental Network (DN). The solution to relation is a special case of > the solution to the compositionality problem which is a special case of the > emergent universal Turing machine. > > > > I am not telling you "a son looks like his father because the father makes > money to feed the son". The solution is supported by biology and a > mathematical proof. > > > Best regards, > > -John > > > > Date: Mon, 7 Feb 2022 07:57:34 -0800 > From: Gary Marcus > To: Juyang Weng > Cc: Post Connectionists > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > Message-ID: > Content-Type: text/plain; charset="utf-8" > > Dear John, > > I agree with you that cluttered scenes are critical, but Geoff?s GLOM > paper [https://www.cs.toronto.edu/~hinton/absps/glomfinal.pdf > ] > might actually have some relevance. It may well be that we need to do a > better job with parts and whole before we can fully address clutter, and > Geoff is certainly taking that question seriously. > > Geoff?s ?Stable islands of identical vectors? do sound suspiciously like > symbols to me (in a good way!), but regardless, they seem to me to be a > plausible candidate as a foundation for coping with clutter. > > And not just cluttered scenes, but also relations between multiple objects > in a scene, which is another example of the broader issue you raise, > challenging for pure MLPs but critical for deeper AI. > > Gary > > > > -- > > Juyang (John) Weng > > > > > -- > > Juyang (John) Weng > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.jpg Type: image/jpeg Size: 27597 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image002.jpg Type: image/jpeg Size: 91893 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image003.jpg Type: image/jpeg Size: 399723 bytes Desc: not available URL: From sina.tootoonian at gmail.com Fri Feb 11 11:53:40 2022 From: sina.tootoonian at gmail.com (Sina Tootoonian) Date: Fri, 11 Feb 2022 16:53:40 +0000 Subject: Connectionists: Group Leader Positions Francis Crick Institute (theoretical & computational) - deadline March 10 Message-ID: Dear colleagues, The Francis Crick Institute is recruiting Early Career Group Leaders again. Please see the attached poster / advert. We are recruiting very broadly across all areas of theoretical and computational biology (*including all facets of computational and theoretical neuroscience*). The conditions at the Crick are phenomenal with excellent, collaborative, interactive and interdisciplinary colleagues, core funded positions, core facilities / platforms including state-of-the art scientific computing, and essentially no teaching obligations and close ties to the London research institutes, universities and organisations with a location at the heart of London. For a computational and theoretical neuroscientist who envisions close collaboration with experimental and theoretical colleagues alike, this is the perfect place to start their lab. More information can be found under https://www.crick.ac.uk/careers-and-study/faculty https://www.crick.ac.uk/research/research-topics/neurosciences Most importantly - spread the word! Please retweet as well. The *deadline is March 10.* Thanks and best, Sina Tootoonian -------------- next part -------------- An HTML attachment was scrubbed... URL: From david at irdta.eu Sat Feb 12 06:10:09 2022 From: david at irdta.eu (David Silva - IRDTA) Date: Sat, 12 Feb 2022 12:10:09 +0100 (CET) Subject: Connectionists: DeepLearn 2022 Spring / Summer / Autumn Message-ID: <1860776725.102172.1644664209115@webmail.strato.com> Dear all, DeepLearn, the International School on Deep Learning, is running since 2017 successfully. Please note the next editions of the program in 2022: https://irdta.eu/deeplearn/2022sp/ https://irdta.eu/deeplearn/2022su/ https://irdta.eu/deeplearn/2022au/ Best regards, DeepLearn organizing team -------------- next part -------------- An HTML attachment was scrubbed... URL: From martaruizcostajussa at gmail.com Fri Feb 11 12:50:28 2022 From: martaruizcostajussa at gmail.com (Marta Ruiz) Date: Fri, 11 Feb 2022 18:50:28 +0100 Subject: Connectionists: 1st CFP NAACL 2022 4th Workshop on Gender Bias for Natural Language Processing In-Reply-To: References: Message-ID: 1st CFP NAACL 2022 4th Workshop on Gender Bias for Natural Language Processing http://genderbiasnlp.talp.cat Gender bias, among other demographic biases (e.g. race, nationality, religion), in machine-learned models is of increasing interest to the scientific community and industry. Models of natural language are highly affected by such biases, which are present in widely used products and can lead to poor user experiences. There is a growing body of research into improved representations of gender in NLP models. Key example approaches are to build and use balanced training and evaluation datasets (e.g. Webster et al., 2018), and to change the learning algorithms themselves (e.g. Bolukbasi et al., 2016). While these approaches show promising results, there is more to do to solve identified and future bias issues. In order to make progress as a field, we need to create widespread awareness of bias and a consensus on how to work against it, for instance by developing standard tasks and metrics. Our workshop provides a forum to achieve this goal. Our proposal follows up three successful previous editions of the Workshop collocated with ACL 2019, COLING 2020, and ACL-IJCNLP 2021, respectively. As in the two previous years (2020 and 2021), special efforts will be made this year to encourage a careful and reflective approach to gender bias by the means of separately reviewed bias statements (Blodgett et al., 2020; Hardmeier et al., 2021). This helps to make clear (a) what system behaviors are considered as bias in the work, and (b) why those behaviors are harmful, in what ways, and to whom. We encourage authors to engage with definitions of bias and other relevant concepts such as prejudice, harm, discrimination from outside NLP, especially from social sciences and normative ethics, in this statement and in their work in general. We will keep pushing the integration of several communities such as social sciences as well as a wider representation of approaches dealing with bias. Topics of interest We invite submissions of technical work exploring the detection, measurement, and mediation of gender bias in NLP models and applications. Other important topics are the creation of datasets, identifying and assessing relevant biases or focusing on fairness in NLP systems. Finally, the workshop is also open to non-technical work addressing sociological perspectives, and we strongly encourage critical reflections on the sources and implications of bias throughout all types of work. Paper Submission Information Submissions will be accepted as short papers (4-6 pages) and as long papers (8-10 pages), plus additional pages for references, following the NAACL 2022 guidelines. Supplementary material can be added, but should not be central to the argument of the paper. Blind submission is required. Each paper should include a statement which explicitly defines (a) what system behaviours are considered as bias in the work and (b) why those behaviours are harmful, in what ways, and to whom (cf. Blodgett et al. (2020)). More information on this requirement, which was successfully introduced at GeBNLP 2020, can be found on the workshop website. We also encourage authors to engage with definitions of bias and other relevant concepts such as prejudice, harm, discrimination from outside NLP, especially from social sciences and normative ethics, in this statement and in their work in general.Non-archival option The authors have the option of submitting research as non-archival, meaning that the paper will not be published in the conference proceedings. We expect these submissions to describe the same quality of work and format as archival submissions. Important dates Apr 8, 2022: Workshop Paper Due Date May 6, 2022: Notification of Acceptance May 20, 2022: Camera-ready papers due July 14 or 15, 2022: Workshop Dates Keynote Kellie Webster, Research Scientist at Google Research Organizers Marta R. Costa-juss?, Meta AI, Paris Christian Hardmeier, Uppsala University Hila Gonen, FAIR and University of Washington Christine Basta, Universitat Polit?cnica de Catalunya, Barcelona Gabriel Stanovsky, Hebrew University of Jerusalem -------------- next part -------------- An HTML attachment was scrubbed... URL: From amir.kalfat at gmail.com Sat Feb 12 13:02:16 2022 From: amir.kalfat at gmail.com (Amir Aly) Date: Sat, 12 Feb 2022 18:02:16 +0000 Subject: Connectionists: [Jobs] Available Position: Professor in Artificial Intelligence (AI) at the University of Plymouth, UK Message-ID: Dear All, *Apologies for cross-posting* *The School of Engineering, Computing, and Mathematics (SECaM)*, at the University of Plymouth has a unique opportunity for an outstanding academic and their team to join the School and strengthen our research, leadership, and collaboration in *Artificial Intelligence*. This is an opportunity to establish senior academic direction and growth in this core area of Artificial Intelligence and to build on multi-disciplinary collaboration, supporting the School?s strategic objectives to create a critical mass of expertise in the area of Artificial Intelligence aligning directly with Industry 4.0 and future national and international research opportunities and skills development needs. You will work together with colleagues to create a vision for our existing Centre for Robotics and Neural Systems (CRNS), which builds on the world-leading and international excellence performance in the field of computer science, social and cognitive robotics, neural computation, and others. You will join a community of over 100 staff and 2000 students who work alongside one another in world-leading facilities. This synergistic mix of academic disciplines has research strength aligned with challenges of the fourth Industrial Revolution, clean growth, digitalization, autonomy, and health technology, helping to develop pioneering, innovative solutions to real-world problems that have a genuinely positive impact upon society. The deadline is 25th February. For more information and to apply, please visit: https://hrservices.plymouth.ac.uk/tlive_webrecruitment/wrd/run/ETREC107GF.open?VACANCY_ID=121682Kdw9&WVID=1602750fTZ&LANG=USA Regards ---------------- *Dr. Amir Aly* Lecturer in Artificial Intelligence and Robotics Center for Robotics and Neural Systems (CRNS) School of Engineering, Computing, and Mathematics Room B332, Portland Square, Drake Circus, PL4 8AA University of Plymouth, UK -------------- next part -------------- An HTML attachment was scrubbed... URL: From r.pascanu at gmail.com Fri Feb 11 12:16:10 2022 From: r.pascanu at gmail.com (Razvan Pascanu) Date: Fri, 11 Feb 2022 17:16:10 +0000 Subject: Connectionists: Eastern European Machine Learning summer school (EEML), Hybrid in Vilnius, Lithuania, 6-14 July 2022, DEADLINE FOR APPLICATIONS April 7, 2022 Message-ID: Call for Participation (apologies for crossposting) Eastern European Machine Learning summer school (**with online and in-person sections**) July 6-14, 2021, Vilnius, Lithuania Web: https://www.eeml.eu Email: contact at eeml.eu Applications are open! Details about the application process https://www.eeml.eu/application. Application closes: April 7, 2022 Notification of acceptance: Early May 2022. **Registration will be free for all accepted participants, for both online and in-person attendance.** Motivation and description EEML is a machine learning summer school that aims to democratise access to education and research in AI, and improve diversity in the field. The summer school is held yearly in Eastern Europe ? this year it will be held in Vilnius, Lithuania. Because of the pandemic, the school will use a hybrid format: first 3 days fully online, last 4 days online and in-person for those who wish to travel to Vilnius; check details on our webpage https://www.eeml.eu/program. By bringing together (virtually or in-person) high quality lecturers and participants from all over the world, we strive to enable communication and networking among the Eastern European AI communities as well as with researchers from around the world. The school is open to participants from all over the world. The selection process has equal opportunities and diversity at heart, and will assess interest and knowledge in machine learning. We encourage applications from candidates at all levels of expertise in Machine Learning (beginner, intermediate, advanced). Details about the application process are available online at https://www.eeml.eu/application. The programme consists of lectures, reading groups, hands-on practical sessions, panel discussions, and more. Some of the core topics to be covered include Reinforcement Learning, Natural Language Processing, Computer Vision, Theory of Deep Learning, Causal Inference. List of confirmed speakers (so far) Doina Precup, McGill University & DeepMind Ferenc Huszar, University of Cambridge Finale Doshi-Velez, Harvard University Gintare Karolina Dziugaite, Google Research Michal Valko, DeepMind Razvan Pascanu, DeepMind Suriya Gunasekar, Microsoft Research Redmond Victor Lempitsky, Skoltech & Samsung Yee Whye Teh, University of Oxford & DeepMind Poster session Participants will have the opportunity to present their research work and interests during virtual poster sessions. The work described does not have to be novel. For example, participants can present their experience of reproducing published work. Organizers Doina Precup, McGill University & DeepMind Razvan Pascanu, DeepMind Viorica Patraucean, DeepMind Ferenc Huszar, University of Cambridge Gintare Karolina Dziugaite, Google Research Jevgenij Gamper, Vinted Linas Petkevi?ius, Vilnius University Dovydas ?eilutka, Vinted Linas Baltr?nas, Wayfair Technical support Gabriel Marchidan, IasiAI & Feel IT Services ZoomTV Partners Artificial Intelligence Association of Lithuania Faculty of Mathematics and Informatics, Vilnius University Local Sponsors Go Vilnius More info https://www.eeml.eu contact at eeml.eu Follow us on Twitter https://twitter.com/EEMLcommunity -------------- next part -------------- An HTML attachment was scrubbed... URL: From iccc22.conference at gmail.com Sun Feb 13 07:58:04 2022 From: iccc22.conference at gmail.com (ICCC 2022) Date: Sun, 13 Feb 2022 23:58:04 +1100 Subject: Connectionists: CFP: Final Call for Full Papers for the 13th International Conference on Computational Creativity (ICCC'22) Message-ID: *** Apologies for cross-posting *** ------------------------------------------------ *Final Call for Full Papers* *13th International Conference on Computational Creativity, ICCC?22!* https://computationalcreativity.net/iccc22/full-papers/ ======================================================================= *Abstracts due February 18, 2022* *You must submit an abstract to be able to submit your full paper.* The abstracts are used to allocate reviewer workload. The abstract itself can be updated with the full paper submission deadline on February 25. ======================================================================= Computational Creativity (or CC) is a discipline with its roots in Artificial Intelligence, Cognitive Science, Engineering, Design, Psychology and Philosophy that explores the potential for computers to be autonomous creators in their own right. ICCC is an annual conference that welcomes papers on different aspects of CC, on systems that exhibit varying degrees of creative autonomy, on frameworks that offer greater clarity or computational felicity for thinking about machine (and human) creativity, on methodologies for building or evaluating CC systems, on approaches to teaching CC in schools and universities or to promoting societal uptake of CC as a field and as a technology, and so on. ***** *Themes and Topics ** **** Original research contributions are solicited in all areas related to Computational Creativity research and practice, including, but not limited to: - Applications that address creativity in specific domains such as music, language, narrative, poetry, games, visual arts, graphic design, product design, architecture, entertainment, education, mathematical invention, scientific discovery, and programming. - Applications and frameworks that allow for co-creativity between humans and machines, in which the machine is more than a mere tool and takes on significant creative responsibility for itself. - Metrics, frameworks, formalisms and methodologies for the evaluation of creativity in computational systems, and for the evaluation of how systems are perceived in society. - Syntheses of AI/CC treatments of very different genres or domains of creativity (e.g. art and science, humour and mathematics, language and image, etc.) - Computational paradigms for understanding creativity, including heuristic search, analogical and meta-level reasoning, and representation. - Resource development and data gathering/knowledge curation for creative systems, especially resources and data collections that are scalable, extensible and freely available as open-source materials. - Ethical considerations in the design, deployment or testing of CC systems, as well as studies that explore the societal impact of CC systems. - Cognitive and psychological computational models of creativity, and their relation with existing cognitive architectures and psychological accounts - Innovation, improvisation, virtuosity and related pursuits investigating the production of novel experiences and artefacts within a CC context. - Computational accounts of factors that enhance creativity, including emotion, surprise(unexpectedness), reflection, conflict, diversity, motivation, knowledge, intuition, reward structures. - Computational models of social aspects of creativity, including the relationship between individual and social creativity, diffusion of ideas, collaboration and creativity, formation of creative teams, and creativity in social settings. - Perspectives on computational creativity which draw from philosophical and/or sociological studies in a context of creative intelligent systems. - Computational creativity in the cloud, including how web services can be used to foster unexpected creative behaviour in computational systems. - Big data approaches to computational creativity. - Debate papers that raise new issues or reopen seemingly settled ones. Provocations that question the foundations of the discipline or throw new light on old work are also welcome. **** Important Dates **** Abstracts due: February 18, 2022 Submissions due: February 25, 2022 Acceptance notification: April 8, 2022 Camera-ready copies due: May 13, 2022 Conference: June 27-July 1, 2022 We are working to a tighter schedule this year, as we shift the conference from September back to June, and so authors should not expect any extension to the above deadlines. Rather, these will be strictly enforced to give the program committee sufficient time for their review work. All deadlines given are 23:59 anywhere on Earth time. We expect the submission deadline for short papers to be set a week after the full-paper notification, allowing a short period for authors to retool their long-paper submissions for this call. Please watch for future announcements of the short-paper call. **** Paper Types **** We welcome the submission of five different types of long papers: Technical papers, System or Resource description papers, Study papers, Cultural application papers and Position papers. Please indicate in your submission which category (or categories) your paper broadly fits into: - *Technical papers:* These are papers posing and addressing hypotheses about aspects of creative behaviour in computational systems. The emphasis here is on using solid experimentation, computational models, formal proof, and/or argumentation that clearly demonstrates advancement in the state of the art or current thinking in CC research. Strong evaluation of approaches through comparative, statistical, social, or other means is essential. - *System or Resource description papers:* These are papers describing the building and deployment of a creative system or resource to produce artefacts of potential cultural value in one or more domains. The emphasis here is on presenting engineering achievement, technical difficulties encountered and overcome, techniques employed, reusable resources built, and general findings about how to get computational systems to produce valuable results. Presentation of results from the system or resource is expected. While full evaluation of the approaches employed is not essential if the technical achievement is very high, some evaluation is expected to show the contribution to CC of this work. - *Study papers:* These are papers which draw on allied fields such as psychology, philosophy, cognitive science, mathematics, humanities, the arts, and so on; or which appeal to broader areas of AI and Computer Science in general; or which appeal to studies of the field of CC as a whole. The emphasis here is on presenting enlightening novel perspectives related to the building, assessment, or deployment of systems ranging from autonomously creative systems to creativity support tools. Such perspectives can be presented through a variety of approaches including ethnographic studies, thought experiments, comparisons with studies of human creativity, and surveys. The contribution of the paper to CC should be made clear in every case. - *Cultural application papers:* These are papers presenting the use of creative software in a cultural setting, for example via art exhibitions/books, concerts/recordings/scores, poetry or story readings/anthologies, cookery nights/books, results for scientific journals or scientific practice, released games/game jam entries, and so on. The emphasis here is on a clear description of the role of the system in the given context, the results of the system in the setting, technical details of inclusion of the system, and evaluative feedback from the experience garnered from public audiences, critics, experts, stakeholders, and other interested parties. - *Position papers:* These are papers presenting an opinion on some aspect of the culture of CC research, including discussions of future directions, past triumphs or mistakes, and issues of the day. The emphasis here is on carefully arguing a position; highlighting or exposing previously hidden or misunderstood issues or ideas; and providing thought leadership for the field, either in a general fashion or in a specific setting. While opinions need not be substantiated through formalisation or experimentation, any justification of a point of view will need to draw on a thorough knowledge of the field of CC and of overlapping areas, and provide relevant motivations and arguments. All submissions will be reviewed in terms of quality, impact, and relevance to the area of Computational Creativity. A call for short papers will be announced shortly. **** Presentation **** In order to ensure the highest level of quality, all submissions will be evaluated in terms of their scientific, technical, artistic, and/or cultural contribution, and therefore there will be only one format for submission. The program committee will decide the best format for presenting accepted manuscripts in the conference. To be included in the proceedings, each paper must be presented at the conference by one of the authors. This implies that at least one author will have to register and will have to participate live in the session in which their paper is presented, including the designated question-and-answer period. *** All authors of accepted papers can opt to also show a demo of their system or prototype during the conference. You will be asked if you are interested in this option during the submission process *** **** Submission Instructions **** This year the submission process has two stages: initial submission of a title and abstract, and subsequent submission of the full paper a week later. - Recommended length for the abstract is 100-200 words. - The long paper page limit is 8 pages + up to 2 pages of references. The reference pages may include the above mentioned Contributions-section as well as the Acknowledgement section. - The manuscript submission date given below is a hard deadline. Even though it has become customary in recent years, do not expect the submission deadline for ICCC 2022 to be extended. - Papers will be reviewed in a double-blind fashion, which necessitates that authors take appropriate steps to remain anonymous. You are responsible for making your papers anonymous to allow for double-blind review. Remove all references to your home institution(s), refer to your past work in the third person, etc. - The final, de-anonymized version of multi-author papers should include a Contributions section in which the contribution of each author is explicitly stated. - To be considered, papers must be submitted as a PDF document formatted according to ICCC style (which is similar to AAAI and IJCAI formats). You can download the updated ICCC?22 template [ https://computationalcreativity.net/iccc22/wp-content/uploads/2022/01/ICCC-22-author-kit.zip ]. - Abstracts are to be returned one week before the full paper deadline. Submit your abstract via the Easy Chair system. You are required to fill out authors, a title, abstract and keywords. You can include the same information in a pdf. - Submit your full paper by updating the EasyChair Abstract with your manuscript file. Abstract submissions that do not contain a manuscript will be automatically rejected at the beginning of the review time. - Papers must be submitted through the EasyChair platform at the ICCC 2022 site [https://easychair.org/conferences/?conf=iccc20220]. - *Double submissions policy:* The work submitted to ICCC should not be under review in another scientific conference or journal at the time of submission. -------------- next part -------------- An HTML attachment was scrubbed... URL: From trentin at dii.unisi.it Sun Feb 13 06:50:23 2022 From: trentin at dii.unisi.it (Edmondo Trentin) Date: Sun, 13 Feb 2022 12:50:23 +0100 Subject: Connectionists: Formal Analysis of Deep Artificial Neural Networks (Entropy SI) Message-ID: Dear colleagues, dear friends, The Entropy Journal (https://www.mdpi.com/journal/entropy) is currently running a Special Isuue on Formal Analysis of Deep Artificial Neural Networks (https://www.mdpi.com/journal/entropy/special_issues/DANN) Submission deadline: 31 May 2022 This Special Issue welcomes original research papers on the analysis of ANNs based on mathematically founded methods in general. Review articles describing the current state of the art of ANNs in the aforementioned contexts are highly encouraged. All submissions to this Special Issue must include substantial theoretical aspects of ANN research. Keywords * ANN architectures and learning in approximation and complexity theories * Cost functions and constraints in information-theoretic learning algorithms for ANNs * Complexity of deep, recurrent, or quantum ANN learning * Information-theoretic principles for sampling and feature extraction * Analysis of learning based on information-theoretic methods (e.g., information bottleneck approach) in deep, recurrent, or quantum ANNs * Applications of ANNs based on information-theoretic principles or quantum computing * Theoretical advances in quantum ANNs Feel free to contact us: Edmondo Trentin or? Friedhelm Schwenker if you are interested in submitting your work. Best, Friedhelm Schwenker Edmondo Trentin ----------------------------------------------- Edmondo Trentin, PhD Dip. Ingegneria dell'Informazione e Scienze MM. V. Roma, 56 - I-53100 Siena (Italy) E-mail: trentin at dii.unisi.it Voice: +39-0577-234636 Fax: +39-0577-233602 WWW: http://www.dii.unisi.it/~trentin/HomePage.html -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Sun Feb 13 14:58:39 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Sun, 13 Feb 2022 14:58:39 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear Asim, The following information might be very useful to Gary Marcus and many others on this list. I try to make my complexity analysis a little more complete, as John Tsotsos seems to have considered (2) below (?), if my memory serves me correctly. (1) The exponential complexity of recognizing and segmenting a part: each pixel has c colors, a part with e pixel-elements has O(c^e) complexity. (2) Group "parts" into an attended object: "Suppose that each part is centered at location l, the number of combinations of p parts of your object (dog) is O(p^l), another exponential complexity." This exponential O(p^l) has never been addressed by any neural networks other than our DN. (3) Segment an object from a cluttered background. Suppose a cluttered scene has m parts, m>> p. Segmenting an object from a cluttered scene (many parts!) has a complexity O(2^m) where 2 is belonging or not-belonging to the object. The real complexity is at least a product of above three exponential complexities. O(c^e p^l 2^m). In other words, what you wrote "*we can also identify parts of wholes in these scenes" is an illusion, since you have not discussed how your network deals with* *NP hard problems. Just three examples are an illusion. It is a toy illusion. * *Of course, our DN can do all above and more, with a constant (ML) frame complexity, but the network size is of a brain-size. * *I am not saying that we solved the NP completeness problem. The NO completeness problem is pure symbolic. The problem of the brain is not symbolic (e.g., pixels).We should not expect to do better than humans, unlike Li Fei-Fei incorrectly claimed. * *Best regards,* *-John* On Fri, Feb 11, 2022, 9:49 PM Asim Roy wrote: > Dear John, > > > > If I understand correctly, all learning systems do something along the > lines of maximum likelihood learning or error minimization, like your DN. > What?s your point? > > > > JOHN: *?Of course, the brain network does not remember all shapes and all > configurations of parts. That is why our DN must do maximum likelihood > optimality, using a limited number of resources to best estimate such a > huge space of cluttered scenes.?* > > > > So, can your DN model identify the parts of objects in the cluttered > images below? Here was my note: > > > > ASIM: *?And we can also identify parts of wholes in these scenes. Here > are some example scenes. In the first two scenes, we can identify the > huskies along with the ears, eyes, legs, faces and so on. In the satellite > image below, we can identify parts of the planes like the fuselage, tail, > wing and so on. That?s the fundamental part of DARPA?s XAI model ? to be > able to **identify the parts to confirm the whole object. And if you can > identify the parts, a school bus will never become an ostrich with change > of a few pixels. So you get a lot of things with Explainable models of this > form ? a symbolic XAI model, robustness against adversarial attacks, and a > model that you can trust. Explainable AI of this form can become the best > defense against adversarial attacks. You may not need any adversarial > training of any kind.?* > > > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > [image: A dog and a cat lying on a bed Description automatically generated > with low confidence] [image: A wolf walking in the snow Description > automatically generated with medium confidence] [image: An aerial view > of a city Description automatically generated with medium confidence] > > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.jpg Type: image/jpeg Size: 27597 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image002.jpg Type: image/jpeg Size: 91893 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image003.jpg Type: image/jpeg Size: 399723 bytes Desc: not available URL: From ASIM.ROY at asu.edu Sun Feb 13 18:13:44 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sun, 13 Feb 2022 23:13:44 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear John, We don?t use a segmentation approach to finding parts. For high-speed image processing, an engineering problem of course, you cannot afford to have a human looking at segmented parts for verification. With symbolic outputs for the parts, as in the DARPA figure, the verification for parts can be done at an extremely high speed by a separate computer program. We can also output ?Don?t Know? if we can?t verify the parts and, in that case, a human can take a look to verify objects. I kind of explained what we do in my response to Geoffrey Hinton?s comment ?But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2.? I am adding that response one more time. We teach the system composition of objects from parts and also the connectivity between the parts. It?s similar to how we teach humans about parts of objects. Here are responses to some additional comments of yours: John: ?I am not saying that we solved the NP completeness problem.? I have not made any claim about training these systems in polynomial time. When we have a large number of objects and parts to deal with, we can break up the problem to consist of several smaller sets of objects and parts. Simple divide and conquer principle. John: ?The problem of the brain is not symbolic (e.g., pixels).? I am not sure what you mean here. There?s plenty of neurophysiological evidence that the brain uses multimodal abstractions (multisensory neurons). Again, to reiterate, you get a lot with Explainable models of the type envisioned by DARPA ? a symbolic XAI model, robustness against adversarial attacks, a model you can trust, and high-speed processing. XAI of this form can become the best defense against adversarial attacks. You may not need any adversarial training of any kind. Hope to see you at the IEEE World Congress on Computational Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy 18-23 July) where I have a tutorial on this subject. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) www.teuvonet.com ================================================================================================================================================== I am responding to this part of Geoffrey Hinton?s note: ?I agree that it's nice to have a causal explanations. But I am not convinced there will ever be a simple causal explanation for how you recognize that a handwritten 2 is a 2. We can introspect on how we do it and this may or may not give some insight into how we check our answer, but the immediate sense that a handwritten 2 is a 2 is computed by a neural net that is not functionally equivalent to any simple and easily explainable procedure.? The causal explanation is actually done quite simply, and we are doing it currently. I can talk about this now because Arizona State University (ASU) has filed a provisional patent application on the technology. The basic idea was laid out by DARPA in their Explainable AI (XAI) program (Explainable Artificial Intelligence (darpa.mil)) and illustrated in the figure below. The idea is to identify objects based on its parts. So, the figure below says that it?s a cat because it has fur, whiskers, and claws plus an unlabeled visual feature. I am not sure if DARPA got anything close to this from its funding of various entities. What this means is that you need a parts model. And we do that. In the case of MNIST handwritten digits that Geoff mentions, we ?teach? this parts model what the top part of a digit ?3? looks like, what the bottom part looks like and so on. And we also teach connectivity between parts and the composition of objects from parts. And we do that for all digits. And we get a symbolic model sitting on top of a CNN model that provides the explanation that Geoff is referring to as the causal explanation. This ?teaching? is similar to the way you would teach a kid to recognize different digits. An advantage of this parts model, in addition to being in an explainable symbolic form, is robustness to adversarial attack. We recently tested on the MNIST data. Where a regular CNN model?s accuracy was reduced by a fast gradient method to 27%, our XAI model maintained an accuracy of 90%, probably higher. In general, it would be hard to make a school bus look like an ostrich, with a few pixel changes, if you can identify the parts of a school bus and an ostrich. A parts model that DARPA wanted provides both a symbolic explanation and adversarial protection. The problem that Geoffrey is referring to is solved. I am doing a tutorial on this at IEEE World Congress on Computational Intelligence in Padua, Italy, July 2022 (WCCI2022 ? WCCI2022 Padua, Italy 18-23 July). I am copying the organizers and want to thank them for accepting the tutorial proposal. The only other presentation I have done on this is at a Military Operations Research Society (MORS) meeting last December. So, back to the future. Hybrid models might indeed save deep learning models and let us deploy these models without concern. We might not even need adversarial training of any kind. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) www.teuvonet.com From: Juyang Weng Sent: Sunday, February 13, 2022 12:59 PM To: Asim Roy Cc: John K Tsotsos ; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, The following information might be very useful to Gary Marcus and many others on this list. I try to make my complexity analysis a little more complete, as John Tsotsos seems to have considered (2) below (?), if my memory serves me correctly. (1) The exponential complexity of recognizing and segmenting a part: each pixel has c colors, a part with e pixel-elements has O(c^e) complexity. (2) Group "parts" into an attended object: "Suppose that each part is centered at location l, the number of combinations of p parts of your object (dog) is O(p^l), another exponential complexity." This exponential O(p^l) has never been addressed by any neural networks other than our DN. (3) Segment an object from a cluttered background. Suppose a cluttered scene has m parts, m>> p. Segmenting an object from a cluttered scene (many parts!) has a complexity O(2^m) where 2 is belonging or not-belonging to the object. The real complexity is at least a product of above three exponential complexities. O(c^e p^l 2^m). In other words, what you wrote "we can also identify parts of wholes in these scenes" is an illusion, since you have not discussed how your network deals with NP hard problems. Just three examples are an illusion. It is a toy illusion. Of course, our DN can do all above and more, with a constant (ML) frame complexity, but the network size is of a brain-size. I am not saying that we solved the NP completeness problem. The NO completeness problem is pure symbolic. The problem of the brain is not symbolic (e.g., pixels). We should not expect to do better than humans, unlike Li Fei-Fei incorrectly claimed. Best regards, -John On Fri, Feb 11, 2022, 9:49 PM Asim Roy > wrote: Dear John, If I understand correctly, all learning systems do something along the lines of maximum likelihood learning or error minimization, like your DN. What?s your point? JOHN: ?Of course, the brain network does not remember all shapes and all configurations of parts. That is why our DN must do maximum likelihood optimality, using a limited number of resources to best estimate such a huge space of cluttered scenes.? So, can your DN model identify the parts of objects in the cluttered images below? Here was my note: ASIM: ?And we can also identify parts of wholes in these scenes. Here are some example scenes. In the first two scenes, we can identify the huskies along with the ears, eyes, legs, faces and so on. In the satellite image below, we can identify parts of the planes like the fuselage, tail, wing and so on. That?s the fundamental part of DARPA?s XAI model ? to be able to identify the parts to confirm the whole object. And if you can identify the parts, a school bus will never become an ostrich with change of a few pixels. So you get a lot of things with Explainable models of this form ? a symbolic XAI model, robustness against adversarial attacks, and a model that you can trust. Explainable AI of this form can become the best defense against adversarial attacks. You may not need any adversarial training of any kind.? Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) -------------- next part -------------- An HTML attachment was scrubbed... URL: From m.biehl at rug.nl Mon Feb 14 02:41:27 2022 From: m.biehl at rug.nl (Michael Biehl) Date: Mon, 14 Feb 2022 08:41:27 +0100 Subject: Connectionists: 4 year PhD position in Groningen/NL Message-ID: *Two weeks left to apply! * A *fully funded PhD position* (4 years) in the *Statistical **Physics of Neural Networks* is available at the University of Groningen, The Netherlands, see https://www.rug.nl/about-ug/work-with-us/job-opportunities/?details=00347-02S0008WFP for details and application details. Applications *(before March 1)* are only possible through this webpage. The title of the project is "The role of the activation function for feedforward learning systems (RAFFLES)". For further information please contact Michael Biehl. ---------------------------------------------------------- Prof. Dr. Michael Biehl Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence P.O. Box 407, 9700 AK Groningen The Netherlands Tel. +31 50 363 3997 https://www.cs.rug.nl/~biehl m.biehl at rug.nl -------------- next part -------------- An HTML attachment was scrubbed... URL: From timvogels at gmail.com Mon Feb 14 02:55:41 2022 From: timvogels at gmail.com (Tim Vogels) Date: Mon, 14 Feb 2022 08:55:41 +0100 Subject: Connectionists: Computational Neuroscience Summer school in South Africa, DEADLINE TOMORROW Message-ID: Dear all, This is a reminder that the application DEADLINE for the next IMBIZO is TOMORROW. If you, or anyone you know who would benefit from this opportunity please consider applying: IBRO-SIMONS COMPUTATIONAL NEUROSCIENCE IMBIZO #isiCNI2022 12 August - 4 September 2022, Noordhoek Beach, Cape Town, South Africa http://imbizo.africa/ Application deadline: 15th February 2022 The #isiCNI2022 is a southern hemisphere summer school aiming to promote computational neuroscience in Africa. It will bring together international and local students under the tutelage of the world's leading experts in the field. Like its international sister courses, this four-week summer school aims to teach central ideas, methods, and practices of modern computational neuroscience through a combination of lectures and hands-on project work. Mornings will be devoted to lectures on topics across the breadth of computational neuroscience, including experimental underpinnings and machine learning analogues. The rest of the day will be spent working on research projects under the close supervision of expert tutors and faculty. Individual research projects will focus on the modelling of neurons, neural systems, behaviour, the analysis of state-of-the-art neural data, and the development of theories to explain experimental observations. New this year is a week focused on neuroscience-inspired machine learning! Who should apply? This course is aimed at Masters and early-PhD level students though Honours or advanced undergraduates may also apply. Postdoctoral students who can motivate why the course would benefit them are also encouraged to apply, Students should have sufficient quantitative skills, (e.g. a background in mathematics, physics, computer science, statistics, engineering or related field). Some knowledge of neural biology will be useful but not essential. Experimental neuroscience students are encouraged to apply, but should ensure that they have a reasonable level of quantitative proficiency (i.e. at least second-year level mathematics or statistics and have done at least one course in computer programming). Please distribute this information as widely as you can. Essential details * Fee (which covers tuition, lodging, and meals): 1100 EUR Thanks to our generous sponsors, significant financial assistance is available to reduce and waiver fees for students, particularly for African applicants. We also hope to provide some travel bursaries for international students. If you are in need of financial assistance to attend the Imbizo, please state so clearly in the relevant section of your application. The Imbizo is planned to be hosted in person, with everyone following our COVID policy . If it is unsafe to hold the summer school, we will follow up with further information at the appropriate time. * Application deadline: 15th February 2022 * Notification of results: late April 2022 Information and application https://imbizo.africa/ Questions? isicn.imbizo at gmail.com What is an Imbizo? \?m?bi?z?\ | Xhosa - Zulu A gathering of the people to share knowledge. FACULTY Demba Ba - Harvard University Adrienne Fairhall - University of Washington Peter Latham - University College London Daphne Bavelier - University of Geneva Timothy Lillicrap - DeepMind Jonathan Pillow - Princeton University Joseph Raimondo - University of Cape Town Mackenzie Mathis - EPFL Lausanne Athanassia Papoutsi - IMBB-FORTH Evan Schaffer - Columbia University Henning Sprekeler - Technical University of Berlin Thomas Tagoe - University of Ghana Tim Vogels - Institute of Science and Technology Austria Blake Richards - McGill University Alex Pouget - University of Geneva TUTORS Mohamed Abdelhack - Krembil Centre for Neuroinformatics Annik Carson - McGill University Spiros Chavlis - IMBB-FORTH Christopher Currin - Institute of Science and Technology Austria Sanjukta Krishnaopal - University College London Marjorie Xie - Columbia University ORGANISERS Demba Ba (Harvard University) Christopher Currin (Institute of Science and Technology Austria) Peter Latham (Gatsby Unit for Computational Neuroscience) Joseph Raimondo (University of Cape Town) Emma Vaughan (Imbizo Logistics) Tim Vogels (Institute of Science and Technology Austria) Sponsors The isiCNI is made possible by the generous support from the Simons Foundation and the International Brain Research Organisation (IBRO) , as well as the?Wellcome Trust , Deep Mind and Wits University Organizational Affiliates University of Cape Town, University College London, Institute of Science and Technology Austria, TReND in Africa , Neuroscience Institute , Gatsby Foundation , IBRO African Center for Advanced Training in Neurosciences at UCT -------------- next part -------------- An HTML attachment was scrubbed... URL: From geoffrey.hinton at gmail.com Mon Feb 14 03:02:23 2022 From: geoffrey.hinton at gmail.com (Geoffrey Hinton) Date: Mon, 14 Feb 2022 03:02:23 -0500 Subject: Connectionists: Weird beliefs about consciousness Message-ID: Many AI researchers are very unclear about what consciousness is and also very sure that GPT-3 doesn?t have it. It?s a strange combination. -------------- next part -------------- An HTML attachment was scrubbed... URL: From edwige.pissaloux at univ-rouen.fr Mon Feb 14 03:42:45 2022 From: edwige.pissaloux at univ-rouen.fr (Edwige Pissaloux) Date: Mon, 14 Feb 2022 09:42:45 +0100 Subject: Connectionists: =?utf-8?q?3-year_PhD_position_open_at_the_Univers?= =?utf-8?q?ity_of_Rouen=2C_France?= Message-ID: <3501-620a1600-15d-3f9462c@27178999> Dear Colleague, There is a 3-year PhD position open in my research group as soon as possible (April 2022 latest) - please see the attached file. The profile of the candidate : vision for robotics, mechatronics, navigation, human navigation assistance. Thank you for your help in advertising this position Kind regards Edwige Pissaloux -------------- next part -------------- A non-text attachment was scrubbed... Name: Annonce These ENG janv 2022 -VF-.doc Type: application/msword Size: 43008 bytes Desc: not available URL: From benabbessarra at gmail.com Mon Feb 14 04:24:59 2022 From: benabbessarra at gmail.com (=?UTF-8?Q?Sarra_Ben_Abb=C3=A8s?=) Date: Mon, 14 Feb 2022 10:24:59 +0100 Subject: Connectionists: [CFP] 2nd International workshop on Ontology Uses and Contribution to Artificial Intelligence @PAKDD-2022 Message-ID: Dear colleagues and researchers, Please consider contributing to the 2nd edition of the international workshop "* Ontology Uses and Contribution to Artificial Intelligence* ", in conjunction with *PAKDD 2022* which will be held online or in Chengdu, China - May 16 - 19, 2022. ================================================================== The deadline for paper submissions is *March 11, 2022* ================================================================== *OnUCAI-2022* 2nd International workshop on Ontology Uses and Contribution to Artificial Intelligence at *PAKDD 2022* , Chengdu, China - May 16 - 19, 2022 Workshop website: https://sites.google.com/view/onucai-pakdd-2022 ================================================================== *Context* An ontology is well known to be the best way to represent knowledge in a domain of interest. It is defined by Gruber and Guarino as ?an explicit specification of a conceptualization?. It allows us to represent explicitly and formally existing entities, their relationships, and their constraints in an application domain. This representation is the most suitable and beneficial way to solve many challenging problems related to the information domain (e.g., knowledge representation, knowledge sharing, knowledge reusing, automated reasoning, knowledge capitalizing, and ensuring semantic interoperability among heterogeneous systems). Using ontology has many advantages, among them we can cite ontology reusing, reasoning, explanation, commitment, and agreement on a domain of discourse, ontology evolution, mapping, etc. As a field of artificial intelligence (AI), ontology aims at representing knowledge based on declarative and symbolic formalization. Combining this symbolic field with computational fields of IA such as Machine Learning (ML), Deep Learning (DL), Uncertainty and Probabilistic Graphical Models (PGMs), Computer Vision (CV), Multi-Agent Systems (SMA) and Natural Languages Processing (NLP) is a promising association. Indeed, ontological modeling plays a vital role to help AI reduce the complexity of the studied domain and organizing information inside it. It broadens AI?s scope allowing it to include any data type as it supports unstructured, semi-structured, or structured data format which enables smoother data integration. The ontology also assists AI for the interpretation process, learning, enrichment, prediction, semantic disambiguation, and discovery of complex inferences. Finally, the ultimate goal of ontologies is the ability to be integrated into the software to make sense of all information. In the last decade, ontologies are increasingly being used to provide background knowledge for several AI domains in different sectors (e.g. energy, transport, health, banking, insurance, etc.). Some of these AI domains are: - Machine learning and deep learning: semantic data selection, semantic data pre-processing, semantic data transformation, semantic data prediction, semantic clustering correction of the outputs, semantic enrichment with ontological concepts, use the semantic structure for promoting distance measure, etc. - Uncertainty and Probabilistic Graphical Models: learning PGM (structure or parameters) using ontologies, probabilistic semantic reasoning, semantic causality, probability, etc. - Computer Vision: semantic image processing, semantic image classification, semantic object recognition/classification, etc. - Blockchain: semantic transactions, interoperable blockchain systems, etc. - Natural Language Processing: semantic text mining, semantic text classification, semantic role labeling, semantic machine translation, semantic question answering, ontology-based text summarizing, semantic recommendation systems, etc. - Multi-Agent Systems and Robotics: semantic task composition, task assignment, communication, cooperation, coordination, plans, and plannification, etc. - Voice-video-speech: semantic voice recognition, semantic speech annotation, etc. - Game Theory: semantic definition of specific games, semantic rules and goals definition, etc. - etc. *Objective* This workshop aims at highlighting recent and future advances on the role of ontologies and knowledge graphs in different domains of AI and how they can be used in order to reduce the semantic gap between the data, applications, machine learning process, etc., in order to obtain semantic-aware approaches. In addition, the goal of this workshop is to bring together an area for experts from industry, science, and academia to exchange ideas and discuss the results of ongoing research in ontologies and AI approaches. ======================================================================== We invite the submission of original works that are related -- but are not limited to -- the topics below. *Topics of interest:* * Ontology for Machine Learning/Deep Learning * Ontology for Uncertainty and Probabilistic Graphical Models * Ontology for Edge Computing * Ontology for Federated Machine Learning * Ontology for Smart Contracts * Ontology for Computer Vision * Ontology for Natural Language Processing * Ontology for Robotics and Multi-agent Systems * Ontology for Voice-video-speech * Ontology for Game Theory * and so on. *Submission:* The workshop is open to submitting unpublished work resulting from research that presents original scientific results, methodological aspects, concepts, and approaches. All submissions are not anonymous and must be PDF documents written in English and formatted using the following style files: PAKDD2022_authors_kit Papers are to be submitted through the workshop's CMT3 submission page. We welcome the following types of contributions: * *Full papers* of up to 9 pages, including abstract, figures, and appendices (if any), but excluding references and acknowledgments: Finished or consolidated R&D works, to be included in one of the Workshop topics. ** Short papers* of up to 4 pages, excluding references and acknowledgments: Ongoing works with relevant preliminary results, opened to discussion. Submitting a paper to the workshop means that the authors agree that at least one author should attend the workshop to present the paper if the paper is accepted. For no-show authors, their affiliations will receive a notification. For further instructions, please refer to the PAKDD 2022 page. *Important dates:* * Workshop paper submission due: *March 11, 2022* * Workshop paper notifications: March 31, 2022 * Workshop paper camera-ready versions due: April 15, 2022 * Workshop: May 16-19, 2022 (Half-Day) All deadlines are 23:59 anywhere on earth (UTC-12). *Publication:* The accepted papers of this workshop may be included in the Proceedings of PAKDD 2022 Workshops published by Springer. ================================================================= *Workshop Chairs* * Sarra Ben Abb?s, Engie, France * Lynda Temal, Engie, France * Nada Mimouni, CNAM, France * Ahmed Mabrouk, Engie, France * Philippe Calvez, Engie, France *Program Committee* * Shridhar Devamane, Physical Design Engineer, Tecsec Technologies, Bangalore, India * Oudom Kem, Researcher at Engie, France * Philippe Leray, Professor at University of Nantes * Stefan Fenz, key researcher at SBA Research and Senior Scientist at Vienna University of Technology * Olivier Dameron, Professor at Universit? de Rennes I, Dyliss team, Irisa / Inria Rennes-Bretagne Atlantique * Ammar Mechouche, Data Science expert at AIRBUS Helicopters * Aar?n Ayll?n Benitez, PhD in bioinformatics and Ontology Lead at BASF Digital Solutions S.L. * Fran?ois Scharffe, Researcher on Knowledge-based AI, New York, United States * Maxime Lefran?ois, Associate Professor at Saint Etienne University, France * Pierre Maret, The QA Company & Saint Etienne University, France * Sanju Tiwari, Universidad Autonoma de Tamaulipas, Mexico -------------- next part -------------- An HTML attachment was scrubbed... URL: From benabbessarra at gmail.com Mon Feb 14 04:24:53 2022 From: benabbessarra at gmail.com (=?UTF-8?Q?Sarra_Ben_Abb=C3=A8s?=) Date: Mon, 14 Feb 2022 10:24:53 +0100 Subject: Connectionists: [CFP] 3rd International workshop on Deep Learning meets Ontologies and Natural Language Processing @ESWC-2022 Message-ID: Dear colleagues and researchers, Please consider contributing to the 3rd edition of the international workshop "*Deep Learning meets Ontologies and Natural Language Processing*" which will be held online or in Hersonissos, Greece - May 29 - June 2, 2022. ========================================================================= The deadline for paper submissions is *March 18th, 2022* ========================================================================= *DeepOntoNLP-2022* 3rd International workshop on Deep Learning meets Ontologies and Natural Language Processing at ESWC 2022 , Hersonissos, Greece - May 29 - June 2, 2022 Workshop website: https://sites.google.com/view/deepontonlp2022/ ========================================================================= *Context* In recent years, deep learning has been applied successfully and achieved state-of-the-art performance in a variety of domains, such as image analysis. Despite this success, deep learning models remain hard to analyze data and understand what knowledge is represented in them, and how they generate decisions. Deep Learning (DL) meets Natural Language Processing (NLP) to solve human language problems for further applications, such as information extraction, machine translation, search, and summarization. Previous works have attested the positive impact of domain knowledge on data analysis and vice versa, for example pre-processing data, searching data, redundancy and inconsistency data, knowledge engineering, domain concepts, and relationships extraction, etc. Ontology is a structured knowledge representation that facilitates data access (data sharing and reuse) and assists the DL process as well. DL meets recent ontologies and tries to model data representations with many layers of non-linear transformations. The combination of DL, ontologies, and NLP might be beneficial for different tasks: - Deep Learning for Ontologies: ontology population, ontology extension, ontology learning, ontology alignment, and integration, - Ontologies for Deep Learning: semantic graph embeddings, latent semantic representation, hybrid embeddings (symbolic and semantic representations), - Deep Learning for NLP: summarization, translation, named entity recognition, question answering, document classification, etc. - NLP for Deep Learning: parsing (part-of-speech tagging), tokenization, sentence detection, dependency parsing, semantic role labeling, semantic dependency parsing, etc. *Objective* This workshop aims at demonstrating recent and future advances in semantic rich deep learning by using Semantic Web and NLP techniques which can reduce the semantic gap between the data, applications, machine learning process, in order to obtain semantic-aware approaches. In addition, the goal of this workshop is to bring together an area for experts from industry, science, and academia to exchange ideas and discuss the results of ongoing research in natural language processing, structured knowledge, and deep learning approaches. ======================================================================== We invite the submission of original works that are related -- but are not limited to -- the topics below. Topics of interest: - Construction ontology embeddings - Ontology-based text classification - Learning ontology embeddings - Semantic role labeling - Ontology reasoning with Deep Neural Networks - Deep learning for ontological semantic annotations - Spatial and temporal ontology embeddings - Ontology alignment and matching based on deep learning models - Ontology learning from text using deep learning models - Unsupervised Learning - Text classification using deep models - Neural machine translation - Deep question answering - Deep text summarization - Deep speech recognition - and so on. Submission: The workshop is open to submitting unpublished work resulting from research that presents original scientific results, methodological aspects, concepts, and approaches. All submissions must be PDF documents written in English and formatted according to LNCS instructions for authors . Papers are to be submitted through the workshop's EasyChair submission page. We welcome the following types of contributions: - Full research papers (8-10 pages): Finished or consolidated R&D works, to be included in one of the Workshop topics - Short papers (4-6 pages): Ongoing works with relevant preliminary results, opened to discussion. At least one author of each accepted paper must register for the workshop, in order to present the paper, there, and at the conference. For further instructions please refer to the ESWC 2022 page. Important dates: - Workshop paper submission due: March 18th, 2022 - Workshop paper notifications: April 15th, 2022 - Workshop paper camera-ready versions due: April 22th, 2022 - Workshop: 30th of May, 2022 (afternoon-Half-Day) All deadlines are 23:59 anywhere on earth (UTC-12). Publication: The best papers from this workshop may be included in the supplementary proceedings of ESWC 2022. ======================================================================== *Workshop Chairs* Sarra Ben Abb?s, Engie, France Rim Hantach, Engie, France Philippe Calvez, Engie, France *Program Committee* *Nada Mimouni**,* CNAM, France Lynda Temal, Engie, France Davide Buscaldi, LIPN, Universit? Sorbonne Paris Nord, France Valentina Janev, Mihajlo Pupin Institute, Serbia Mohamed Hedi Karray, LGP-INP-ENIT, Universit? de Toulouse, France -------------- next part -------------- An HTML attachment was scrubbed... URL: From vito.trianni at istc.cnr.it Mon Feb 14 04:37:36 2022 From: vito.trianni at istc.cnr.it (Vito Trianni) Date: Mon, 14 Feb 2022 10:37:36 +0100 Subject: Connectionists: [jobs] postdoc position: active monitoring for UAV swarms in precision agriculture - ISTC-CNR, Rome, Italy Message-ID: A postdoc position is now open in the context of the AGR-o-RAMA research project (http://www.agrorama.it) The research will be conducted at the ISTC-CNR (https://istc.cnr.it) in Rome, Italy, under the supervision of Vito Trianni, and in collaboration with Sapienza and RomaTre university. The starting date should be around May the 1st, 2022, with some flexibility. A valid working permit in Europe/Italy is an asset. Candidates nearing the end of the PhD (up to 6 months from the defence) are also welcome. _________________________________________________ Applications Only for inquiries, you can contact Vito Trianni by email (vito.trianni _at_ istc.cnr.it). Your official candidature must instead be submitted following the procedure below, which also states eligibility conditions and other relevant information about the position: Information for applicants ? English: http://bandi.urp.cnr.it/doc-assegni/documentazione/12193_DOC_EN.pdf Information for applicants ? Italian: http://bandi.urp.cnr.it/doc-assegni/documentazione/12193_DOC_IT.pdf Note that, despite mentioned in the eligibility conditions, knowledge of Italian IS NOT required. _________________________________________________ Research Description AGR-o-RAMA aims at the study of autonomous intelligent drones capable of actively monitoring a field in order to identify and map features of interests (e.g., weed or pests) that could be distributed heterogeneously within the field. Moving beyond the classical approach of a uniform coverage of the field with predetermined mission plans for remote sensing, AGR-o-RAMA proposes to adaptively define the sampling frequency and resolution in order to focus on the areas of interest, while mildly monitoring areas that are devoid of relevant features. Active monitoring strategies will be developed and tested both with a single drone and with groups of drones. In the latter case, the drones will coordinate to explore the field in parallel and collaborate to identify the areas of interest, maximising both efficiency and accuracy. _________________________________________________ Skills: - Programming abilities in C++ and Python (required) - Knowledge of ROS/ROS2 (required) - Knowledge of deep learning frameworks (asset) - Knowledge of Unity (asset) - UAV Piloting abilities (asset) _________________________________________________ Background knowledge: - multi-robot systems - Information-based motion planning - Bayesian inference - Information theory - Active vision ======================================================================== Vito Trianni, Ph.D. vito.trianni@(no_spam)istc.cnr.it ISTC-CNR http://www.istc.cnr.it/people/vito-trianni Via San Martino della Battaglia 44 Tel: +39 06 44595277 00185 Roma Fax: +39 06 44595243 Italy ======================================================================== From davidjrohde at gmail.com Mon Feb 14 04:43:22 2022 From: davidjrohde at gmail.com (david rohde) Date: Mon, 14 Feb 2022 10:43:22 +0100 Subject: Connectionists: Laplace's Causal Demon 28 Feb - 9 Mar Message-ID: We are very pleased to announce a new series of webinars on the interface of Bayes and causality starting the week of 28th of February. Laplace's Causal Demon - Criteo AI Lab 28 February 2022 Christopher Sims 2011 Nobel Prize Winner in Economics, Princeton Large Parameter Spaces and Weighted Data: A Bayesian Perspective 2 March 2022 Yixin Wang University of Michigan Representation Learning: A Causal Perspective 3 March 2022 Fan Li Duke University Propensity score in Bayesian causal inference: why, why not and how? 7 March 2022 Andrew Gelman Columbia University Bayesian Methods in Causal Inference and Decision Making 9 March 2022 David Rohde Criteo Causal Inference is (Bayesian) Inference - A beautifully simple idea that not everyone accepts -------------- next part -------------- An HTML attachment was scrubbed... URL: From cognitivium at sciencebeam.com Mon Feb 14 08:58:53 2022 From: cognitivium at sciencebeam.com (Mary) Date: Mon, 14 Feb 2022 17:28:53 +0330 Subject: Connectionists: NFB, QEEG, and tDCS workshop Message-ID: <202202141358.21EDwtfd085511@scs-mx-02.ANDREW.cmu.edu> Dear researchers and clinicians, This would be our pleasure to invite you to join us for the upcoming hands-on Neurofeedback, QEEG and tDCS workshop, on February 26-27, 2022. This meeting will take place in Istanbul, Turkey.? This workshop will be a deep look into the QEEG assessment and Neurofeedback and tDCS treatment methods. It will provide the interested researchers and clinicians with an opportunity to learn all the details of Neurofeedback, tDCS, and QEEG through simultaneous theoretical and practical sessions. After this comprehensive workshop, all the attendees will be able to do QEEG assessment, and Neurofeedback and tDCS treatment methods on their own. Due to the current situation of COVID-19, only limited seats are available and the registration deadline is February 22nd. Save your seat NOW: https://sciencebeam.com/neurofeedback-qeeg-workshop4/ Besides, we are holding a free introductory webinar to the QEEG (Brain mapping) technique, on February 20th. Register now: https://sciencebeam.com/qeeg-webinar-2/ Should you require any further information about the webinar and workshop, or the registration process, do not hesitate to contact us: Email: workshop at sciencebeam.com WhatsApp: +90 (535) 6498587 Mary Reae Human Neuroscience Dept. Manager www.sciencebeam.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Mon Feb 14 08:59:19 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Mon, 14 Feb 2022 14:59:19 +0100 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Geoffrey Hinton via mailman.srv.cs.cmu.edu 9:25 AM (5 hours ago) to Connectionists By that logic, one could say that the following is a strange combination too: Many AI researchers are very unclear about what consciousness is and are also very sure that a refrigerator doesn?t have it. And it is not only AI researchers. Many mind and brain researchers and philosophers of mind think that GPT-3 doesn't have consciousness. These groups should be slightly more clear about what consciousness is than AI researchers, albeit far from full clarity. Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Mon, Feb 14, 2022 at 9:25 AM Geoffrey Hinton wrote: > Many AI researchers are very unclear about what consciousness is and also > very sure that GPT-3 doesn?t have it. It?s a strange combination. > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Mon Feb 14 07:57:52 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Mon, 14 Feb 2022 07:57:52 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Ah proof by ignorance.??? Yes, another way to deny the need for experimental tests. Interesting PNAS article.. on consciousness as near critical slow cortical dynamics.. https://www.pnas.org/content/119/7/e2024455119 (Barak Pearlmutter had some similar theory concerning sleep a while back). Steve On 2/14/22 3:02 AM, Geoffrey Hinton wrote: > Many AI researchers are very unclear about what consciousness is and > also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From gary.marcus at nyu.edu Mon Feb 14 09:14:50 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Mon, 14 Feb 2022 06:14:50 -0800 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Also true: Many AI researchers are very unclear about what consciousness is and also very sure that ELIZA doesn?t have it. Neither ELIZA nor GPT-3 have - anything remotely related to embodiment - any capacity to reflect upon themselves Hypothesis: neither keyword matching nor tensor manipulation, even at scale, suffice in themselves to qualify for consciousness. - Gary > On Feb 14, 2022, at 00:24, Geoffrey Hinton wrote: > > ?Many AI researchers are very unclear about what consciousness is and also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > From minaiaa at gmail.com Mon Feb 14 10:42:18 2022 From: minaiaa at gmail.com (Ali Minai) Date: Mon, 14 Feb 2022 10:42:18 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> References: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Message-ID: Gary I agree with all of that completely. However, I do have a further question. How will we ever know whether even a fully embodied, high performing intelligent machine that can ?explain? its decisions has the capacity for actual self-reflection? How can we be sure that other humans have the capacity for self-reflection, beyond the fact that they are like us, give us understandable reports of their inner thinking, and tickle our mirror system in the right way? This is not to deny that humans can do self-reflection; it is about whether we have an objective method to know that another species not like us is doing it. At some point in what I think will be a fairly distant future, an nth- generation, descendent of GPT-3 may be so sophisticated that the question will become moot. That is the inevitable consequence of a non-dualistic view of intelligence that we all share, and will come to pass unless the current narrow application-driven view of AI derails the entire project. I do think that that that system will have a very different neural architecture much closer to the brain?s, will be embodied, will not need to learn using billions of data points and millions of gradient-descent iterations, and will be capable of actual mental growth (not just learning more stuff but learning new modes of understanding) over its lifetime. It will also not be able to explain its inner thoughts perfectly, nor feel the need to do so because it will not be our obedient servant. In other words, it will not be anything like GPT-3. Ali On Mon, Feb 14, 2022 at 10:06 AM Gary Marcus wrote: > Also true: Many AI researchers are very unclear about what consciousness > is and also very sure that ELIZA doesn?t have it. > > Neither ELIZA nor GPT-3 have > - anything remotely related to embodiment > - any capacity to reflect upon themselves > > Hypothesis: neither keyword matching nor tensor manipulation, even at > scale, suffice in themselves to qualify for consciousness. > > - Gary > > > On Feb 14, 2022, at 00:24, Geoffrey Hinton > wrote: > > > > ?Many AI researchers are very unclear about what consciousness is and > also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > > > > > -- *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Past-President (2015-2016) International Neural Network Society Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From iam.palat at gmail.com Mon Feb 14 10:46:28 2022 From: iam.palat at gmail.com (Iam Palatnik) Date: Mon, 14 Feb 2022 12:46:28 -0300 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> References: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Message-ID: A somewhat related question, just out of curiosity. Imagine the following: - An automatic solar panel that tracks the position of the sun. - A group of single celled microbes with phototaxis that follow the sunlight. - A jellyfish (animal without a brain) that follows/avoids the sunlight. - A cockroach (animal with a brain) that avoids the sunlight. - A drone with onboard AI that flies to regions of more intense sunlight to recharge its batteries. - A human that dislikes sunlight and actively avoids it. Can any of these, beside the human, be said to be aware or conscious of the sunlight, and why? What is most relevant? Being a biological life form, having a brain, being able to make decisions based on the environment? Being taxonomically close to humans? On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: > Also true: Many AI researchers are very unclear about what consciousness > is and also very sure that ELIZA doesn?t have it. > > Neither ELIZA nor GPT-3 have > - anything remotely related to embodiment > - any capacity to reflect upon themselves > > Hypothesis: neither keyword matching nor tensor manipulation, even at > scale, suffice in themselves to qualify for consciousness. > > - Gary > > > On Feb 14, 2022, at 00:24, Geoffrey Hinton > wrote: > > > > ?Many AI researchers are very unclear about what consciousness is and > also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Mon Feb 14 10:06:42 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Mon, 14 Feb 2022 10:06:42 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> References: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Message-ID: <50c5876c-b52a-13f3-f69c-ff487ee10f78@rubic.rutgers.edu> Seems glib.. I don't know what consciousness is either.. but I am sure the logs burning in my fireplace don't have it. S On 2/14/22 9:14 AM, Gary Marcus wrote: > Also true: Many AI researchers are very unclear about what consciousness is and also very sure that ELIZA doesn?t have it. > > Neither ELIZA nor GPT-3 have > - anything remotely related to embodiment > - any capacity to reflect upon themselves > > Hypothesis: neither keyword matching nor tensor manipulation, even at scale, suffice in themselves to qualify for consciousness. > > - Gary > >> On Feb 14, 2022, at 00:24, Geoffrey Hinton wrote: >> >> ?Many AI researchers are very unclear about what consciousness is and also very sure that GPT-3 doesn?t have it. It?s a strange combination. >> >> -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From achler at gmail.com Mon Feb 14 18:26:56 2022 From: achler at gmail.com (Tsvi Achler) Date: Mon, 14 Feb 2022 15:26:56 -0800 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Message-ID: A huge part of the problem in any discussion about consciousness is there isn't even a clear definition of consciousness for humans (and animals), as consciousness is perceptual and not possible to measure externally. Moreover it is possible to do complex tasks without being conscious, e.g. have conversations, go (e.g. drive) from one place to another without conscious awareness if you are: tired, sick, on medications that will keep you awake and responsive but have you forget the surgery, bump your head, etc... So consciousness is not necessary or sufficient for complex thoughts or behavior. Subsequently consciousness may be a behavioral mechanism for self preservation labeling *My* memories and experience gathering is important thus I must preserve myself and not do things I know may put an end to them.. In other words a figment to behaviorally motivate us to focus on our survival. The situation becomes worse when applying to algorithms something that is not well defined in the first place. That is why I stay away. But that is not to say there are not huge differences between organisms and current AI. Organisms can learn and update without catastrophic interference and forgetting. They dont require the methods currently needed to store all bits of information and present-rehearse it in fixed frequencies and random order (iid) whenever something new is to be learned. Most natural environments are not iid so online learning does not apply either. There is also a huge amount of feedback back to the inputs that has been hypothesized to be associated with consciousness. I think the secret to organism-like flexibility lies in those connections, but the association of these connections with consciousness (as opposed to recognition) is hugely misleading, not the least because of the reasons above. -Tsvi On Mon, Feb 14, 2022 at 8:20 AM Iam Palatnik wrote: > A somewhat related question, just out of curiosity. > > Imagine the following: > > - An automatic solar panel that tracks the position of the sun. > - A group of single celled microbes with phototaxis that follow the > sunlight. > - A jellyfish (animal without a brain) that follows/avoids the sunlight. > - A cockroach (animal with a brain) that avoids the sunlight. > - A drone with onboard AI that flies to regions of more intense sunlight > to recharge its batteries. > - A human that dislikes sunlight and actively avoids it. > > Can any of these, beside the human, be said to be aware or conscious of > the sunlight, and why? > What is most relevant? Being a biological life form, having a brain, being > able to make decisions based on the environment? Being taxonomically close > to humans? > > > > > > > > On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: > >> Also true: Many AI researchers are very unclear about what consciousness >> is and also very sure that ELIZA doesn?t have it. >> >> Neither ELIZA nor GPT-3 have >> - anything remotely related to embodiment >> - any capacity to reflect upon themselves >> >> Hypothesis: neither keyword matching nor tensor manipulation, even at >> scale, suffice in themselves to qualify for consciousness. >> >> - Gary >> >> > On Feb 14, 2022, at 00:24, Geoffrey Hinton >> wrote: >> > >> > ?Many AI researchers are very unclear about what consciousness is and >> also very sure that GPT-3 doesn?t have it. It?s a strange combination. >> > >> > >> >> -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Mon Feb 14 12:45:00 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Mon, 14 Feb 2022 09:45:00 -0800 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> Stephen, On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more plausible as a consciousness than GPT-3. 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. 2. Roomba makes choices based on an explicit representation of its location relative to a mapped space. GPT lacks any consistent reflection of self; eg if you ask it, as I have, if you are you person, and then ask if it is a computer, it?s liable to say yes to both, showing no stable knowledge of self. 3. Roomba has explicit, declarative knowledge eg of walls and other boundaries, as well its own location. GPT has no systematically interrogable explicit representations. All this is said with tongue lodged partway in cheek, but I honestly don?t see what criterion would lead anyone to believe that GPT is a more plausible candidate for consciousness than any other AI program out there. ELIZA long ago showed that you could produce fluent speech that was mildly contextually relevant, and even convincing to the untutored; just because GPT is a better version of that trick doesn?t mean it?s any more conscious. Gary > On Feb 14, 2022, at 08:56, Stephen Jos? Hanson wrote: > > ? > this is a great list of behavior.. > > Some biologically might be termed reflexive, taxes, classically conditioned, implicit (memory/learning)... all however would not be > conscious in the several senses: (1) wakefulness-- sleep (2) self aware (3) explicit/declarative. > > I think the term is used very loosely, and I believe what GPT3 and other AI are hoping to show signs of is "self-awareness".. > > In response to : "why are you doing that?", "What are you doing now", "what will you be doing in 2030?" > > Steve > > > > On 2/14/22 10:46 AM, Iam Palatnik wrote: >> A somewhat related question, just out of curiosity. >> >> Imagine the following: >> >> - An automatic solar panel that tracks the position of the sun. >> - A group of single celled microbes with phototaxis that follow the sunlight. >> - A jellyfish (animal without a brain) that follows/avoids the sunlight. >> - A cockroach (animal with a brain) that avoids the sunlight. >> - A drone with onboard AI that flies to regions of more intense sunlight to recharge its batteries. >> - A human that dislikes sunlight and actively avoids it. >> >> Can any of these, beside the human, be said to be aware or conscious of the sunlight, and why? >> What is most relevant? Being a biological life form, having a brain, being able to make decisions based on the environment? Being taxonomically close to humans? >> >> >> >> >> >> >> >> On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: >>> Also true: Many AI researchers are very unclear about what consciousness is and also very sure that ELIZA doesn?t have it. >>> >>> Neither ELIZA nor GPT-3 have >>> - anything remotely related to embodiment >>> - any capacity to reflect upon themselves >>> >>> Hypothesis: neither keyword matching nor tensor manipulation, even at scale, suffice in themselves to qualify for consciousness. >>> >>> - Gary >>> >>> > On Feb 14, 2022, at 00:24, Geoffrey Hinton wrote: >>> > >>> > ?Many AI researchers are very unclear about what consciousness is and also very sure that GPT-3 doesn?t have it. It?s a strange combination. >>> > >>> > >>> > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From rloosemore at susaro.com Mon Feb 14 11:47:06 2022 From: rloosemore at susaro.com (Richard Loosemore) Date: Mon, 14 Feb 2022 11:47:06 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: <8B9EABD4-D27B-404E-91B8-D78ED4C02F36@susaro.com> Geoff, Necessary and sufficient conditions. For some people, they have a list of necessary conditions for consciousness. GPT-3 fails that test. Those people don?t know the sufficient conditions for consciousness. Your summary of that is that those people ?are very unclear about what consciousness is.? That should clarify what seems to be a strange combination. (I would add that am not one of those people, but I felt like speaking up on their behalf). Richard Loosemore > On Feb 14, 2022, at 3:31 AM, Geoffrey Hinton wrote: > > ?Many AI researchers are very unclear about what consciousness is and also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > From tt at cs.dal.ca Mon Feb 14 11:39:06 2022 From: tt at cs.dal.ca (Thomas Trappenberg) Date: Mon, 14 Feb 2022 12:39:06 -0400 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Message-ID: ... the avoiding sun part is not the key, but disliking it consciously as human might be. On Mon, Feb 14, 2022 at 12:20 PM Iam Palatnik wrote: > A somewhat related question, just out of curiosity. > > Imagine the following: > > - An automatic solar panel that tracks the position of the sun. > - A group of single celled microbes with phototaxis that follow the > sunlight. > - A jellyfish (animal without a brain) that follows/avoids the sunlight. > - A cockroach (animal with a brain) that avoids the sunlight. > - A drone with onboard AI that flies to regions of more intense sunlight > to recharge its batteries. > - A human that dislikes sunlight and actively avoids it. > > Can any of these, beside the human, be said to be aware or conscious of > the sunlight, and why? > What is most relevant? Being a biological life form, having a brain, being > able to make decisions based on the environment? Being taxonomically close > to humans? > > > > > > > > On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: > >> Also true: Many AI researchers are very unclear about what consciousness >> is and also very sure that ELIZA doesn?t have it. >> >> Neither ELIZA nor GPT-3 have >> - anything remotely related to embodiment >> - any capacity to reflect upon themselves >> >> Hypothesis: neither keyword matching nor tensor manipulation, even at >> scale, suffice in themselves to qualify for consciousness. >> >> - Gary >> >> > On Feb 14, 2022, at 00:24, Geoffrey Hinton >> wrote: >> > >> > ?Many AI researchers are very unclear about what consciousness is and >> also very sure that GPT-3 doesn?t have it. It?s a strange combination. >> > >> > >> >> -------------- next part -------------- An HTML attachment was scrubbed... URL: From daniel.polani at gmail.com Mon Feb 14 12:04:02 2022 From: daniel.polani at gmail.com (Daniel Polani) Date: Mon, 14 Feb 2022 17:04:02 +0000 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Message-ID: There are quite a few researchers spending a lot of effort trying to understand the origins of consciousness and to understand whether and how the subjective experience of consciousness can be captured in a descriptive and ideally mathematical manner. Tononi, Albantakis, Seth, O'Regan, just to name a few; one does not have to agree with them, but this question has been given a lot of attention and it's worth having a look before discussing it in a vacuum. Also worth reading, amongst other, Dennett and Chalmers (just as side remark: some of you may remember the the latter as he had actually a nice Evolutionary Algorithm experiment in the 90s showing how the Widrow-Hoff rule emerged as "optimal" learning rule in a neural-type learning scenario). The issue about consciousness being an exclusively human ability (as is often insinuated) is probably not anymore seriously discussed; it is pretty clear that even self-awareness extends significantly beyond humans, not even mentioning the subjective experience which does away with the requirement of self-reflection. It is certainly far safer to estimate that it will be a matter of degree of consciousness in the animal kingdom than to claim that it is either present or not. It seems that even our understanding of elementary experiences must be redefined as e.g. lobsters actually may feel pain. Thus we should be careful making sweeping statements about the presence of consciousness in the biological realm. It is indeed a very interesting question to understand to which extent (if at all) an artificial system can experience that, too; what if the artificial system is a specifically designed, but growing biological neuron culture on an agar plate? If the response is yes for the latter, but no for the former, what is the core difference? Is it the recurrence that matters? The embodiment? Some aspect of its biological makeup? Something else? I do not think we have good answers for this at this stage, but only some vague hints. On Mon, Feb 14, 2022 at 4:22 PM Iam Palatnik wrote: > A somewhat related question, just out of curiosity. > > Imagine the following: > > - An automatic solar panel that tracks the position of the sun. > - A group of single celled microbes with phototaxis that follow the > sunlight. > - A jellyfish (animal without a brain) that follows/avoids the sunlight. > - A cockroach (animal with a brain) that avoids the sunlight. > - A drone with onboard AI that flies to regions of more intense sunlight > to recharge its batteries. > - A human that dislikes sunlight and actively avoids it. > > Can any of these, beside the human, be said to be aware or conscious of > the sunlight, and why? > What is most relevant? Being a biological life form, having a brain, being > able to make decisions based on the environment? Being taxonomically close > to humans? > > > > > > > > On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: > >> Also true: Many AI researchers are very unclear about what consciousness >> is and also very sure that ELIZA doesn?t have it. >> >> Neither ELIZA nor GPT-3 have >> - anything remotely related to embodiment >> - any capacity to reflect upon themselves >> >> Hypothesis: neither keyword matching nor tensor manipulation, even at >> scale, suffice in themselves to qualify for consciousness. >> >> - Gary >> >> > On Feb 14, 2022, at 00:24, Geoffrey Hinton >> wrote: >> > >> > ?Many AI researchers are very unclear about what consciousness is and >> also very sure that GPT-3 doesn?t have it. It?s a strange combination. >> > >> > >> >> -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Mon Feb 14 14:30:31 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Mon, 14 Feb 2022 14:30:31 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> Message-ID: <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Gary,? these weren't criterion.???? Let me try again. I wasn't talking about wake-sleep cycles... I was talking about being awake or asleep and the transition that ensues.. Rooba's don't sleep.. they turn off, I have two of them.? They turn on once (1) their batteries are recharged (2) a timer has been set for being turned on. GPT3 is essentially a CYC that actually works.. by reading Wikipedia (which of course is a terribly biased sample). I was indicating the difference between implicit and explicit learning/problem solving.??? Implicit learning/memory is unconscious and similar to a habit.. (good or bad). I believe that when someone says "is gpt3 conscious?"? they are asking: is gpt3 self-aware? ???? Roombas know about vacuuming and they are unconscious. S On 2/14/22 12:45 PM, Gary Marcus wrote: > Stephen, > > On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more > plausible as a consciousness than GPT-3. > > 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. > 2. Roomba makes choices based on an explicit representation of its > location relative to a mapped space. GPT lacks any consistent > reflection of self; eg if you ask it, as I have, if you are you > person, and then ask if it is a computer, it?s liable to say yes to > both, showing no stable knowledge of self. > 3. Roomba has explicit, declarative knowledge eg of walls and other > boundaries, as well its own location. GPT has no systematically > interrogable explicit representations. > > All this is said with tongue lodged partway in cheek, but I honestly > don?t see what criterion would lead anyone to believe that GPT is a > more plausible candidate for consciousness than any other AI program > out there. > > ELIZA long ago showed that you could produce fluent speech that was > mildly contextually relevant, and even convincing to the untutored; > just because GPT is a better version of that trick doesn?t mean it?s > any more conscious. > > Gary > >> On Feb 14, 2022, at 08:56, Stephen Jos? Hanson >> wrote: >> >> ? >> >> this is a great list of behavior.. >> >> Some biologically might be termed reflexive, taxes, classically >> conditioned, implicit (memory/learning)... all however would not be >> conscious in the several senses:? (1)? wakefulness-- sleep? (2) self >> aware (3) explicit/declarative. >> >> I think the term is used very loosely, and I believe what GPT3 and >> other AI are hoping to show signs of is "self-awareness".. >> >> In response to :? "why are you doing that?",? "What are you doing >> now", "what will you be doing in 2030?" >> >> Steve >> >> >> On 2/14/22 10:46 AM, Iam Palatnik wrote: >>> A somewhat related question, just out of curiosity. >>> >>> Imagine the following: >>> >>> - An automatic solar panel that tracks the position of the sun. >>> - A group of single celled microbes with phototaxis that follow the >>> sunlight. >>> - A jellyfish (animal without a brain) that follows/avoids the sunlight. >>> - A cockroach (animal with a brain) that avoids the sunlight. >>> - A drone with onboard AI that flies to regions of more intense >>> sunlight to recharge its batteries. >>> - A human that dislikes sunlight and actively avoids it. >>> >>> Can any of these, beside the human, be said to be aware or conscious >>> of the sunlight, and why? >>> What is most relevant? Being a biological life form, having a brain, >>> being able to make decisions based on the environment? Being >>> taxonomically close to humans? >>> >>> >>> >>> >>> >>> >>> >>> On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus >> > wrote: >>> >>> Also true: Many AI researchers are very unclear about what >>> consciousness is and also very sure that ELIZA doesn?t have it. >>> >>> Neither ELIZA nor GPT-3 have >>> - anything remotely related to embodiment >>> - any capacity to reflect upon themselves >>> >>> Hypothesis: neither keyword matching nor tensor manipulation, >>> even at scale, suffice in themselves to qualify for consciousness. >>> >>> - Gary >>> >>> > On Feb 14, 2022, at 00:24, Geoffrey Hinton >>> > >>> wrote: >>> > >>> > ?Many AI researchers are very unclear about what consciousness >>> is and also very sure that GPT-3 doesn?t have it. It?s a strange >>> combination. >>> > >>> > >>> >> -- -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From jose at rubic.rutgers.edu Mon Feb 14 11:56:29 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Mon, 14 Feb 2022 11:56:29 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Message-ID: this is a great list of behavior.. Some biologically might be termed reflexive, taxes, classically conditioned, implicit (memory/learning)... all however would not be conscious in the several senses:? (1)? wakefulness-- sleep? (2) self aware (3) explicit/declarative. I think the term is used very loosely, and I believe what GPT3 and other AI are hoping to show signs of is "self-awareness".. In response to :? "why are you doing that?", "What are you doing now", "what will you be doing in 2030?" Steve On 2/14/22 10:46 AM, Iam Palatnik wrote: > A somewhat related question, just out of curiosity. > > Imagine the following: > > - An automatic solar panel that tracks the position of the sun. > - A group of single celled microbes with phototaxis that follow the > sunlight. > - A jellyfish (animal without a brain) that follows/avoids the sunlight. > - A cockroach (animal with a brain) that avoids the sunlight. > - A drone with onboard AI that flies to regions of more intense > sunlight to recharge its batteries. > - A human that dislikes sunlight and actively avoids it. > > Can any of these, beside the human, be said to be aware or conscious > of the sunlight, and why? > What is most relevant? Being a biological life form, having a brain, > being able to make decisions based on the environment? Being > taxonomically close to humans? > > > > > > > > On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus > wrote: > > Also true: Many AI researchers are very unclear about what > consciousness is and also very sure that ELIZA doesn?t have it. > > Neither ELIZA nor GPT-3 have > - anything remotely related to embodiment > - any capacity to reflect upon themselves > > Hypothesis: neither keyword matching nor tensor manipulation, even > at scale, suffice in themselves to qualify for consciousness. > > - Gary > > > On Feb 14, 2022, at 00:24, Geoffrey Hinton > > wrote: > > > > ?Many AI researchers are very unclear about what consciousness > is and also very sure that GPT-3 doesn?t have it. It?s a strange > combination. > > > > > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From giacomo.cabri at unimore.it Mon Feb 14 04:32:07 2022 From: giacomo.cabri at unimore.it (Giacomo Cabri) Date: Mon, 14 Feb 2022 10:32:07 +0100 Subject: Connectionists: CfP - Adaptive Computing (and Agents) for Enhanced Collaboration (ACEC) at WETICE 2022 Message-ID: * * *20th Adaptive Computing (and Agents) for Enhanced Collaboration (ACEC)* Track?at IEEE WETICE 2022 Stuttgart, Germany, June 29-July 1, 2022 http://didattica.agentgroup.unimore.it/ACEC2022 Call for Papers *Aims and Scope* Over its 19 years of existence, ACEC has focused on works?that explore?the adaptability, autonomy and intelligence of software agents for the collaboration across the enterprise. In 2022,?organizers would like to continue to explore the research on agent-based computing, but they would also?welcome works that leverage advanced adaptive techniques, non necessarily based on software agents. In addition to the?traditional domain areas, i.e., Computer Supported Collaborative Work, Workflow and Supply Chain Management, Automation in Virtual Enterprises, and Automated Distributed Service Composition, ACEC is?also interested in new adaptive techniques, e.g.,?Cloud Computing, Crowd-Sourcing, and?Social Networking. In addition to traditional papers, the forthcoming 18th episode of ACEC welcomes papers from two focus areas: * Adaptive and Agent-based Services * Adaptive Techniques for Organizational/Enterprise Use of Emerging Web Paradigms (Cloud, Crowd-sourcing, Mobile Apps) Such two themes represent important areas where software agents can leverage their distributed nature,?along with their proactive and autonomous characteristics, to provide solutions to?complex problems, which are difficult to solve?using traditional/existing technologies. *Topics of Interest* Topics of interest include, but are not limited to: * Adaptive and/or agent-mediated workflow, supply chain, and virtual?enterprises * Methodologies, languages and tools to support agent collaboration * Agent architectures and infrastructures for dynamic collaboration * Adaptive and/or Agent-based service architectures and infrastructures * Service Oriented Architectures (SOAs) based on agents * Services for dynamic agent collaboration * Agent-to-Human service interactions * Autonomous, Adaptive and/or Agent-mediated service integration * Organizational and enterprise systems that leverage the Web 2.0 * Adaptive and Agent-mediated cloud environments *Important Dates* * Paper submission: *February 25th, 2022* * Notification: *March 25th, 2022* * Camera ready: *May 6th, 2022 * * Conference: *June?29-July 1, 2022* *Paper Submission* Papers should contain original contributions (not published or submitted elsewhere) and references to related state-of-the art work. Please submit your papers in PDF or PS format. Papers up to 6 pages (including figures, tables and references) can be submitted. Papers should follow the IEEE format, which is single-spaced, two columns, 10pt Times/Roman font. Papers should include a title, the name and affiliation of each author, an abstract of up to 150 words and no more than eight keywords. All submitted papers will be peer-reviewed by a minimum of three program committee members. The accepted papers will be published in the post-conference proceedings (to be published by the IEEE Computer Society Press). Authors of accepted papers must present their papers at the conference.?At least one author for each accepted paper should register and attend WETICE 2022 to have the paper published in the proceedings. The paper submission procedure is carried out using the EasyChair conference management system: https://wetice2022.svccomp.de/#submissions *Track Chairs* * Federico Bergenti, Universit? degli Studi di Parma, Italy * M. Brian Blake, George Washington University, USA * Giacomo Cabri, Universit? degli Studi di Modena e Reggio Emilia, Italy * Stefania Monica, Universit? degli Studi di Modena e Reggio Emilia, Italy * Usman Wajid, The University of Manchester, UK -- |----------------------------------------------------| | Prof. Giacomo Cabri - Ph.D., Full Professor | Rector's Delegate for Teaching | Dip. di Scienze Fisiche, Informatiche e Matematiche | Universita' di Modena e Reggio Emilia - Italia | e-mail giacomo.cabri at unimore.it | tel. +39-059-2058320 fax +39-059-2055216 |----------------------------------------------------| From juyang.weng at gmail.com Mon Feb 14 20:02:07 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Mon, 14 Feb 2022 20:02:07 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Dear Asim, You wrote: "We teach the system composition of objects from parts and also the connectivity between the parts. It?s similar to how we teach humans about parts of objects." You are doing manual image annotations like many have done. Unfortunately, we have been on the wrong track of manually annotating images for too long. Sorry, I started it all, i.e., the annotation practice. From the end of 1990, we at UIUC started Cresceptron on learning from natural images of cluttered scenes (published at IJCNN 1992, ICCV 1993 and IJCV 1997). Nobody did that before us, as far as I know. We at UIUC were trying to solve the general vision problem using a learning neural network called Cresceptron. Namely, detection, recognition and segmentation of 3D objects from all cluttered natural scenes! A flood of similar works followed Crescetron, slowly and after several years, but many did not cite our Cresceptron. (I do not want to mention those big shot names.) I do not understand why. I chatted with Narendra Ahuja about this unethical plagiarism. He explained well. The Cresceptron appeared in arguably the "best" neural network conference, the "best" computer vision conference and the "best" computer vision journal. However, enough is enough. We must go beyond manual annotation of images, although this line has created a lot of business. Many AI companies contracted with large companies to do just manual image annotations. We must cut it out! No more image annotations! In a paper I just submitted to IJCNN 2022 today, the first million-dollar problem solved is: (1) the image annotation problem (e.g., retina is without bounding box to learn, unlike ImageNet) Let me list them all: (2) the sensorimotor recurrence problem (e.g., all big data sets are invalid), (3) the motor-supervision problem (e.g., impractical to supervise motors throughout lifetime), (4) the sensor calibration problem (e.g., a life calibrates the eyes automatically), (5) the inverse kinematics problem (e.g., a life calibrates all redundant limbs automatically), (6) the government-free problem (i.e., no intelligent homunculus inside a brain), (7) the closed-skull problem (e.g., supervising hidden neurons is biologically implausible), (8) the nonlinear controller problem (e.g., a brain is a nonlinear controller but task-nonspecific), (9) the curse of dimensionality problem (e.g., a set of global features is insufficient for a life), (10) the under-sample problem (i.e., few available examples in a life), (11) the distributed vs. local representations problem (i.e., how both representations emerge), (12) the frame problem (also called symbol grounding problem, thus must be free from any symbols), (13) the local minima problem (so, avoid error-backprop learning and Post-Selections), (14) the abstraction problem (i.e., require various invariances and transfers), (15) the compositionality problem (e.g., metonymy beyond those composable from sentences), (16) the smooth representations problem (e.g., brain representations are globally smooth), (17) the motivation problem (e.g., including reinforcements and various emotions), (18) the global optimality problem (e.g., avoid catastrophic memory loss and Post-Selections), (19) the auto-programming for general purposes (APFGP) problem, (20) the brain-thinking problem. The paper discusses also why the proposed holistic solution of conscious learning solves each. Best regards, -John -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From janet.hsiao at gmail.com Mon Feb 14 21:39:13 2022 From: janet.hsiao at gmail.com (Janet Hsiao) Date: Tue, 15 Feb 2022 10:39:13 +0800 Subject: Connectionists: [Job] Postdoc Position: Attention Brain & Cognition Lab, University of Hong Kong Message-ID: Applications are invited for appointment as *Post-doctoral Fellow at the Attention Brain & Cognition Laboratory in the Department of Psychology* at University of Hong Kong (Ref.:511720), to commence on June 14, 2022 or as soon as possible for one year. Applicants should possess a Ph.D. degree in Psychology, Cognitive Science, Computer Science, or a related field. They should have good communication skills and computer skills, and be curious, self-motivated and willing to participate in innovative and interdisciplinary research topics. Those with experience in conducting eye tracking experiments or performing advanced data analysis using machine-learning-related methods are preferred. The appointee will work at the Attention Brain & Cognition Laboratory led by Dr. Janet Hsiao in the Department of Psychology. He/She will lead a team in developing a project related to understanding human attention in object detection and identification through both behavioral, eye tracking, and computational methods. In particular, he/she will have the opportunity to learn and perform advanced data analysis using Eye Movement analysis with Hidden Markov Models (EMHMM; http://abc.psy.hku.hk/emhmm). There are also opportunities to participate in projects related to current research in the Laboratory including electroencephalogram (EEG) and computational modeling. To know more about the research in the Laboratory, please see the Laboratory website at http://abc.psy.hku.hk/, or the publications via http://web.hku.hk/~jhsiao/pubs/.The project is in collaboration with Huawei Hong Kong Research Center (HKRC). The partnering team from Huawei HKRC consists of researchers from various research units, whose research areas include deep learning framework, trustworthy AI software, fundamental AI theory and so on. For more information about the post, please contact Dr. Janet Hsiao at jhsiao at hku.hk with the email subject line ?Postdoctoral Fellow?. A highly competitive salary commensurate with qualifications and experience will be offered, in addition to annual leave and medical benefits. The University only accepts online application for the above post. Applicants should apply online and upload an up-to-date CV including academic qualifications, research experience, and publications, copies of transcripts, a cover letter describing interests in this post, and provide the names and contact information of at least three referees on the online application form via https://jobs.hku.hk/cw/en/job/511720/postdoctoral-fellow. Review of applications will start as soon as possible and continue until *February 28, 2022*, or until the post is filled, whichever is earlier. -------------- next part -------------- An HTML attachment was scrubbed... URL: From maanakg at gmail.com Mon Feb 14 22:14:15 2022 From: maanakg at gmail.com (Maanak Gupta) Date: Mon, 14 Feb 2022 21:14:15 -0600 Subject: Connectionists: FINAL: Call for Papers: 27th ACM Symposium on Access Control Models and Technologies Message-ID: ?ACM SACMAT 2022 ----------------------------------------------- | ONLINE | ----------------------------------------------- Call for Research Papers ============================================================== Papers offering novel research contributions are solicited for submission. Accepted papers will be presented at the symposium and published by the ACM in the symposium proceedings. In addition to the regular research track, this year SACMAT will again host the special track -- "Blue Sky/Vision Track". Researchers are invited to submit papers describing promising new ideas and challenges of interest to the community as well as access control needs emerging from other fields. We are particularly looking for potentially disruptive and new ideas which can shape the research agenda for the next 10 years. We also encourage submissions to the "Work-in-progress Track" to present ideas that may have not been completely developed and experimentally evaluated. Topics of Interest ============================================================== Submissions to the regular track covering any relevant area of access control are welcomed. Areas include, but are not limited to, the following: * Systems: * Operating systems * Cloud systems and their security * Distributed systems * Fog and Edge-computing systems * Cyber-physical and Embedded systems * Mobile systems * Autonomous systems (e.g., UAV security, autonomous vehicles, etc) * IoT systems (e.g., home-automation systems) * WWW * Design for resiliency * Designing systems with zero-trust architecture * Network: * Network systems (e.g., Software-defined network, Network function virtualization) * Corporate and Military-grade Networks * Wireless and Cellular Networks * Opportunistic Network (e.g., delay-tolerant network, P2P) * Overlay Network * Satellite Network * Privacy and Privacy-enhancing Technologies: * Mixers and Mixnets * Anonymous protocols (e.g., Tor) * Online social networks (OSN) * Anonymous communication and censorship resistance * Access control and identity management with privacy * Cryptographic tools for privacy * Data protection technologies * Attacks on Privacy and their defenses * Authentication: * Password-based Authentication * Biometric-based Authentication * Location-based Authentication * Identity management * Usable authentication * Mechanisms: * Blockchain Technologies * AI/ML Technologies * Cryptographic Technologies * Programming-language based Technologies * Hardware-security Technologies (e.g., Intel SGX, ARM TrustZone) * Economic models and game theory * Trust Management * Usable mechanisms * Data Security: * Big data * Databases and data management * Data leakage prevention * Data protection on untrusted infrastructure * Policies and Models: * Novel policy language design * New Access Control Models * Extension of policy languages * Extension of Models * Analysis of policy languages * Analysis of Models * Policy engineering and policy mining * Verification of policy languages * Efficient enforcement of policies * Usable access control policy New in ACM SACMAT 2022 ============================================================== We are moving ACM SACMAT 2022 to have two submission cycles. Authors submitting papers in the first submission cycle will have the opportunity to receive a major revision verdict in addition to the usual accept and reject verdicts. Authors can decide to prepare a revised version of the paper and submit it to the second submission cycle for consideration. Major revision papers will be reviewed by the program committee members based on the criteria set forward by them in the first submission cycle. Regular Track Paper Submission and Format ============================================================== Papers must be written in?English. Authors are required to use the ACM format for papers, using the two-column SIG Proceedings Template (the sigconf template for LaTex) available in the following link: https://www.acm.org/publications/authors/submissions The length of the paper in the proceedings format must not exceed?twelve?US letter pages formatted for 8.5" x 11" paper and be no more than 5MB in size. It is the responsibility of the authors to ensure that their submissions will print easily on simple default configurations. The submission must be anonymous, so information that might identify the authors - including author names, affiliations, acknowledgments, or obvious self-citations - must be excluded. It is the authors' responsibility to ensure that their anonymity is preserved when citing their work. Submissions should be made to the EasyChair conference management system by the paper submission deadline of: February 18th, 2022 (Submission Cycle 2) All submissions must contain a?significant original contribution. That is, submitted papers must not substantially overlap papers that have been published or that are simultaneously submitted to a journal, conference, or workshop. In particular, simultaneous submission of the same work is not allowed. Wherever appropriate, relevant related work, including that of the authors, must be cited. Submissions that are not accepted as full papers may be invited to appear as short papers. At least one author from each accepted paper must register for the conference before the camera-ready deadline. Blue Sky Track Paper Submission and Format ============================================================== All submissions to this track should be in the same format as for the regular track, but the length must not exceed ten US letter pages, and the submissions are not required to be anonymized (optional). Submissions to this track should be submitted to the EasyChair conference management system by the same deadline as for the regular track. Work-in-progress Track Paper Submission and Format ============================================================== Authors are invited to submit papers in the newly introduced work-in-progress track. This track is introduced for (junior) authors, ideally, Ph.D. and Master's students, to obtain early, constructive feedback on their work. Submissions in this track should follow the same format as for the regular track papers while limiting the total number of pages to six US letter pages. Paper submitted in this track should be anonymized and can be submitted to the EasyChair conference management system by the same deadline as for the regular track. Call for Lightning Talk ============================================================== Participants are invited to submit proposals for 5-minute lightning talks describing recently published results, work in progress, wild ideas, etc. Lightning talks are a new feature of SACMAT, introduced this year to partially replace the informal sharing of ideas at in-person meetings. Submissions are expected??by May 27, 2022. Notification of acceptance will be on June 3, 2022. Call for Posters ============================================================== SACMAT 2022 will include a poster session to promote discussion of ongoing projects among researchers in the field of access control and computer security. Posters can cover preliminary or exploratory work with interesting ideas, or research projects in the early stages with promising results in all aspects of access control and computer security. Authors interested in displaying a poster must submit a poster abstract in the same format as for the regular track, but the length must not exceed three US letter pages, and the submission should not be anonymized. The title should start with "Poster:". Accepted poster abstracts will be included in the conference proceedings. Submissions should be emailed to the poster chair by Apr 15th, 2022. The subject line should include "SACMAT 2022 Poster:" followed by the poster title. Call for Demos ============================================================== A demonstration proposal should clearly describe (1) the overall architecture of the system or technology to be demonstrated, and (2) one or more demonstration scenarios that describe how the audience, interacting with the demonstration system or the demonstrator, will gain an understanding of the underlying technology. Submissions will be evaluated based on the motivation of the work behind the use of the system or technology to be demonstrated and its novelty. The subject line should include "SACMAT 2022 Demo:" followed by the demo title. Demonstration proposals should be in the same format as for the regular track, but the length must not exceed four US letter pages, and the submission should not be anonymized. A two-page description of the demonstration will be included in the conference proceedings. Submissions should be emailed to the Demonstrations Chair by Apr 15th, 2022. Financial Conflict of Interest (COI) Disclosure: ============================================================== In the interests of transparency and to help readers form their own judgments of potential bias, ACM SACMAT requires authors and PC members to declare any competing financial and/or non-financial interests in relation to the work described. Definition ------------------------- For the purposes of this policy, competing interests are defined as financial and non-financial interests that could directly undermine, or be perceived to undermine the objectivity, integrity, and value of a publication, through a potential influence on the judgments and actions of authors with regard to objective data presentation, analysis, and interpretation. Financial competing interests include any of the following: Funding: Research support (including salaries, equipment, supplies, and other expenses) by organizations that may gain or lose financially through this publication. A specific role for the funding provider in the conceptualization, design, data collection, analysis, decision to publish, or preparation of the manuscript, should be disclosed. Employment: Recent (while engaged in the research project), present or anticipated employment by any organization that may gain or lose financially through this publication. Personal financial interests: Ownership or contractual interest in stocks or shares of companies that may gain or lose financially through publication; consultation fees or other forms of remuneration (including reimbursements for attending symposia) from organizations that may gain or lose financially; patents or patent applications (awarded or pending) filed by the authors or their institutions whose value may be affected by publication. For patents and patent applications, disclosure of the following information is requested: patent applicant (whether author or institution), name of the inventor(s), application number, the status of the application, specific aspect of manuscript covered in the patent application. It is difficult to specify a threshold at which a financial interest becomes significant, but note that many US universities require faculty members to disclose interests exceeding $10,000 or 5% equity in a company. Any such figure is necessarily arbitrary, so we offer as one possible practical alternative guideline: "Any undeclared competing financial interests that could embarrass you were they to become publicly known after your work was published." We do not consider diversified mutual funds or investment trusts to constitute a competing financial interest. Also, for employees in non-executive or leadership positions, we do not consider financial interest related to stocks or shares in their company to constitute a competing financial interest, as long as they are publishing under their company affiliation. Non-financial competing interests: Non-financial competing interests can take different forms, including personal or professional relations with organizations and individuals. We would encourage authors and PC members to declare any unpaid roles or relationships that might have a bearing on the publication process. Examples of non-financial competing interests include (but are not limited to): * Unpaid membership in a government or non-governmental organization * Unpaid membership in an advocacy or lobbying organization * Unpaid advisory position in a commercial organization * Writing or consulting for an educational company * Acting as an expert witness Conference Code of Conduct and Etiquette ============================================================== ACM SACMAT will follow the ACM Policy Against Harassment at ACM Activities. Please familiarize yourself with the ACM Policy Against Harassment (available at https://www.acm.org/special-interest-groups/volunteer-resources/officers-manual/ policy-against-discrimination-and-harassment) and guide to Reporting Unacceptable Behavior (available at https://www.acm.org/about-acm/reporting-unacceptable-behavior). AUTHORS TAKE NOTE ============================================================== The official publication date is the date the proceedings are made available in the ACM Digital Library. This date may be up to two weeks before the first day of your conference. The official publication date affects the deadline for any patent filings related to published work. (For those rare conferences whose proceedings are published in the ACM Digital Library after the conference is over, the official publication date remains the first day of the conference.) Important dates ============================================================== **Note that, these dates are currently only tentative and subject to change.** * Paper submission: February 18th, 2022 (Submission Cycle 2) * Rebuttal: March 24th - March 28th, 2022 (Submission Cycle 2) * Notifications: April 8th, 2022 (Submission Cycle 2) * Systems demo and Poster submissions: April 15th, 2022 * Systems demo and Poster notifications: April 22nd, 2022 * Panel Proposal: March 18th, 2022 * Camera-ready paper submission: April 29th, 2022 * Conference date: June 8 - June 10, 2022 -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Mon Feb 14 22:24:06 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Tue, 15 Feb 2022 03:24:06 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear John, If I had the kind of technology that you list here, I wouldn?t waste time on the Connectionist list. That kind of technology could be worth billions of dollars, if not trillions. So, look for investors. Best of luck, Asim From: Juyang Weng Sent: Monday, February 14, 2022 6:02 PM To: Asim Roy Cc: Post Connectionists Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, You wrote: "We teach the system composition of objects from parts and also the connectivity between the parts. It?s similar to how we teach humans about parts of objects." You are doing manual image annotations like many have done. Unfortunately, we have been on the wrong track of manually annotating images for too long. Sorry, I started it all, i.e., the annotation practice. From the end of 1990, we at UIUC started Cresceptron on learning from natural images of cluttered scenes (published at IJCNN 1992, ICCV 1993 and IJCV 1997). Nobody did that before us, as far as I know. We at UIUC were trying to solve the general vision problem using a learning neural network called Cresceptron. Namely, detection, recognition and segmentation of 3D objects from all cluttered natural scenes! A flood of similar works followed Crescetron, slowly and after several years, but many did not cite our Cresceptron. (I do not want to mention those big shot names.) I do not understand why. I chatted with Narendra Ahuja about this unethical plagiarism. He explained well. The Cresceptron appeared in arguably the "best" neural network conference, the "best" computer vision conference and the "best" computer vision journal. However, enough is enough. We must go beyond manual annotation of images, although this line has created a lot of business. Many AI companies contracted with large companies to do just manual image annotations. We must cut it out! No more image annotations! In a paper I just submitted to IJCNN 2022 today, the first million-dollar problem solved is: (1) the image annotation problem (e.g., retina is without bounding box to learn, unlike ImageNet) Let me list them all: (2) the sensorimotor recurrence problem (e.g., all big data sets are invalid), (3) the motor-supervision problem (e.g., impractical to supervise motors throughout lifetime), (4) the sensor calibration problem (e.g., a life calibrates the eyes automatically), (5) the inverse kinematics problem (e.g., a life calibrates all redundant limbs automatically), (6) the government-free problem (i.e., no intelligent homunculus inside a brain), (7) the closed-skull problem (e.g., supervising hidden neurons is biologically implausible), (8) the nonlinear controller problem (e.g., a brain is a nonlinear controller but task-nonspecific), (9) the curse of dimensionality problem (e.g., a set of global features is insufficient for a life), (10) the under-sample problem (i.e., few available examples in a life), (11) the distributed vs. local representations problem (i.e., how both representations emerge), (12) the frame problem (also called symbol grounding problem, thus must be free from any symbols), (13) the local minima problem (so, avoid error-backprop learning and Post-Selections), (14) the abstraction problem (i.e., require various invariances and transfers), (15) the compositionality problem (e.g., metonymy beyond those composable from sentences), (16) the smooth representations problem (e.g., brain representations are globally smooth), (17) the motivation problem (e.g., including reinforcements and various emotions), (18) the global optimality problem (e.g., avoid catastrophic memory loss and Post-Selections), (19) the auto-programming for general purposes (APFGP) problem, (20) the brain-thinking problem. The paper discusses also why the proposed holistic solution of conscious learning solves each. Best regards, -John -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From rskhorshidi at gmail.com Tue Feb 15 04:02:56 2022 From: rskhorshidi at gmail.com (Reza Khorshidi) Date: Tue, 15 Feb 2022 09:02:56 +0000 Subject: Connectionists: Call for application -- Oxford ML Summer School (OxML 2022) Message-ID: Dates: August 7-14, 2022 (St Catherine's College, Oxford + Virtual) Application deadline: 15 April, 2022 For more info, please visit the school?s website: www.oxfordml.school Link to the application form: https://forms.gle/EqvC3qxKmoGKJgGs5 About OxML 2022 OxML is organised by AI for Global Goals, in partnership with CIFAR and The University of Oxford?s Deep Medicine Program. OxML schools have a special focus on ML and SDG s. That is, in addition to theoretical ML lectures, there will be lectures on the application of ML in various SDGs areas. OxML 2022 will have two separate 4-day schools: (1) ML x Health, and (2) ML x Finance. Furthermore, based on the success of last year's program, and in order to provide all participants with the necessary background -- particularly for those who are new to the theory and fundamentals of modern ML -- the program will also have an online ML Fundamentals module (June 27-29), which will be open to both schools' accepted participants. The schools will take place in St Catherine's College, Oxford (UK). There will also be a virtual option for those who cannot (or prefer not to) travel to the UK. During each school, in addition to applied and theoretical lectures (taking place in the main hall, with ~250 seats), there will be multiple workshops and sessions on Advanced ML topics, ML Ops, ML Products, and ML Career (taking place in the 4x smaller halls, that have ~50-100 seats). We aim to host ~200 participants in person (plus 100-200 virtually) in each school. Note that, while our current plan is to have a hybrid format, in the worst case COVID scenarios, we have (and are ready to execute) a plan B to go fully virtual. You can find out more about the previous schools ? including the previous speakers ? here . The schools' theoretical tutorials on modern ML (including DL) will cover topics such as: Neural networks, deep learning / representation learning (with, with little, or without supervision), ... Statistical/probabilistic ML (e.g., Bayesian ML, Gaussian processes, variational inference, Bayesian neural networks) Neural sequence models (e.g., Transformers) NLP, computer vision, and multi-modal representation learning Reinforcement learning Causal ML Graph neural nets and Geometric DL Learning theory Optimisation ... On the applied side, the school will cover topics such as: ML x Health: The applications of ML in imaging, genomics, electric health records (EHR), drug discovery, ... ML x Finance: The applications of ML in sustainable investing (e.g., ESG), insurance and emerging risks, banking , hedging and options trading, ... Speakers The school?s world-renowned speakers are from top ML research groups (e.g., Oxford, Cambridge, CIFAR, Amazon, DeepMind, Microsoft Research, and more). The speakers? bios and more details on their talks will be announced in the coming weeks (you can see our previous speakers here , and follow the updates via the school?s website , or Twitter and LinkedIn accounts). Target audience Everyone is welcome to apply to OxML 2022 regardless of their origin, nationality, and country of residence. Our target audience are (1) PhD students with a good technical background whose research topics are related to ML, plus (2) researchers and engineers in both academia and industry with similar/advanced levels of technical knowledge. All applicants are subject to a selection process; we aim to select strongly motivated participants, who are interested in broadening their knowledge of the advanced topics in the field of ML/DL and their applications. Application You can find the application for for the school here . Given the overwhelming number of applications we received in previous years, the application portal may close earlier than the deadline if the number of applications exceeds our capacity to review. For any queries, you can contact us using this email address: contact at oxfordml.school Best, ? Reza Khorshidi, D.Phil. (Oxon) Deep Medicine Program, The University of Oxford -------------- next part -------------- An HTML attachment was scrubbed... URL: From Marie.PELLEAU at univ-cotedazur.fr Tue Feb 15 06:03:54 2022 From: Marie.PELLEAU at univ-cotedazur.fr (Marie Pelleau) Date: Tue, 15 Feb 2022 11:03:54 +0000 Subject: Connectionists: CPAIOR 2022: Extended Deadline Message-ID: <3791F3B6-216C-4D7A-B191-2B7663F5ED86@univ-cotedazur.fr> ****************** Our apologies for multiple reception of this announcement ***************** ****** This call can be seen online: https://sites.google.com/usc.edu/cpaior-2022/cfp ****** Call for extended abstracts The 19th International Conference on the Integration of Constraint Programming, Artificial Intelligence, and Operations Research will be held in Los Angeles, US, June 20th-June 23th, 2022. The aim of the conference is to bring together interested researchers from Constraint Programming (CP), Artificial Intelligence (AI), and Operations Research (OR) to present new techniques or applications and to provide an opportunity for researchers in one area to learn about techniques in the others. A main objective of this conference series is also to give these researchers the opportunity to show how the integration of techniques from different fields can lead to interesting results on large and complex problems. Therefore, papers that actively combine, integrate, or contrast approaches from more than one of the areas are especially solicited. High quality papers from a single area are also welcome, if they are of interest to other communities involved. Application papers showcasing CP/AI/OR techniques on novel and challenging applications or experience reports on such applications are strongly encouraged. The program committee invites submissions that include but are not limited to the following topics: - Inference and relaxation methods: constraint propagation, cutting planes, global constraints, graph algorithms, dynamic programming, Lagrangian and convex relaxations, heuristic functions based on constraint relaxation. - Search methods: branch and bound, intelligent backtracking, incomplete search, randomized search, portfolios, column generation, Benders decompositions or any other decomposition methods, local search, meta-heuristics. - AI and Machine Learning techniques applied to solve optimization and Operations Research problems or CP/OR techniques to solve AI and machine learning problems. - Integration methods: solver communication, model transformations and solver selection, parallel and distributed solving, combining machine learning with combinatorial optimization. - Modeling methods: comparison of models, symmetry breaking, uncertainty, dominance relationships. - Innovative applications of CP/AI/OR techniques. - Implementation of CP/AI/OR techniques and optimization systems. Extended abstracts should be 1 or 2 pages in length and may present preliminary work or work already published in other outlets. The extended abstracts are submitted for presentation only (if accepted), and will not be formally published in the LNCS conference volume. A collection of the accepted extended abstracts will be published on the conference website. A submission representing work submitted or published in another outlet should state that outlet. Extended abstracts will be reviewed to ensure appropriateness for the conference. Submission Process All extended abstracts are to be submitted electronically in PDF format by following the instructions at the URL https://easychair.org/conferences/?conf=cpaior2022 Submission schedule for extended abstracts: Abstract: March 3, 2022 (AoE) Notification: April 4th, 2022 Questions For any queries on the submission process, please contact the Program Chair Pierre Schaus at: cpaior2022 at gmail.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Tue Feb 15 07:18:52 2022 From: jose at rubic.rutgers.edu (=?UTF-8?Q?Stephen_Jos=c3=a9_Hanson?=) Date: Tue, 15 Feb 2022 07:18:52 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: <22ca6e4a-8c12-021a-7b85-70552389dbdb@rubic.rutgers.edu> Mostly from an experimental point of view. Neurologists study consciousness in patients in comas for example, as it seems clear that while we are asleep or knocked out through anesthesia or coma, we are not conscious. Self-awareness is also a form of consciousness in that we are paying attention to ourselves and can self-report-- emotional awareness, some individuals are not aware of their own states and others are exquisitely aware..? and so on.. Mostly these are explicit (in the sense of learning/memory) as the individuals can provide some verbal evidence of their self-awareness (but brain imaging is plausible here as a measure). Obviously more difficult call in an AI robot, roomba or an oyster. Steve On 2/15/22 5:34 AM, Adam Kosiorek wrote: > Stephen, > > It's curious to me that wake-sleep cycles should be included in the > notion of consciousness, in the sense that I see no problems with a > conscious creature that does not sleep. Could you tell me a little > more about your thinking here? > > Thanks, > > Adam R. Kosiorek > > > On Tue, 15 Feb 2022 at 07:12, Stephen Jos? Hanson > > wrote: > > Gary,? these weren't criterion.???? Let me try again. > > I wasn't talking about wake-sleep cycles... I was talking about > being awake or asleep and the transition that ensues.. > > Rooba's don't sleep.. they turn off, I have two of them.? They > turn on once (1) their batteries are recharged (2) a timer has > been set for being turned on. > > GPT3 is essentially a CYC that actually works.. by reading > Wikipedia (which of course is a terribly biased sample). > > I was indicating the difference between implicit and explicit > learning/problem solving. Implicit learning/memory is unconscious > and similar to a habit.. (good or bad). > > I believe that when someone says "is gpt3 conscious?"? they are > asking: is gpt3 self-aware? Roombas know about vacuuming and they > are unconscious. > > S > > On 2/14/22 12:45 PM, Gary Marcus wrote: >> Stephen, >> >> On criteria (1)-(3), a high-end, mapping-equippped Roomba is far >> more plausible as a consciousness than GPT-3. >> >> 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. >> 2. Roomba makes choices based on an explicit representation of >> its location relative to a mapped space. GPT lacks any consistent >> reflection of self; eg if you ask it, as I have, if you are you >> person, and then ask if it is a computer, it?s liable to say yes >> to both, showing no stable knowledge of self. >> 3. Roomba has explicit, declarative knowledge eg of walls and >> other boundaries, as well its own location. GPT has no >> systematically interrogable explicit representations. >> >> All this is said with tongue lodged partway in cheek, but I >> honestly don?t see what criterion would lead anyone to believe >> that GPT is a more plausible candidate for consciousness than any >> other AI program out there. >> >> ELIZA long ago showed that you could produce fluent speech that >> was mildly contextually relevant, and even convincing to the >> untutored; just because GPT is a better version of that trick >> doesn?t mean it?s any more conscious. >> >> Gary >> >>> On Feb 14, 2022, at 08:56, Stephen Jos? Hanson >>> wrote: >>> >>> ? >>> >>> this is a great list of behavior.. >>> >>> Some biologically might be termed reflexive, taxes, classically >>> conditioned, implicit (memory/learning)... all however would not be >>> conscious in the several senses:? (1) wakefulness-- sleep? (2) >>> self aware (3) explicit/declarative. >>> >>> I think the term is used very loosely, and I believe what GPT3 >>> and other AI are hoping to show signs of is "self-awareness".. >>> >>> In response to :? "why are you doing that?",? "What are you >>> doing now", "what will you be doing in 2030?" >>> >>> Steve >>> >>> >>> On 2/14/22 10:46 AM, Iam Palatnik wrote: >>>> A somewhat related question, just out of curiosity. >>>> >>>> Imagine the following: >>>> >>>> - An automatic solar panel that tracks the position of the sun. >>>> - A group of single celled microbes with phototaxis that follow >>>> the sunlight. >>>> - A jellyfish (animal without a brain) that follows/avoids the >>>> sunlight. >>>> - A cockroach (animal with a brain) that avoids the sunlight. >>>> - A drone with onboard AI that flies to regions of more intense >>>> sunlight to recharge its batteries. >>>> - A human that dislikes sunlight and actively avoids it. >>>> >>>> Can any of these, beside the human, be said to be aware or >>>> conscious of the sunlight, and why? >>>> What is most relevant? Being a biological life form, having a >>>> brain, being able to make decisions based on the environment? >>>> Being taxonomically close to humans? >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus >>>> > wrote: >>>> >>>> Also true: Many AI researchers are very unclear about what >>>> consciousness is and also very sure that ELIZA doesn?t have it. >>>> >>>> Neither ELIZA nor GPT-3 have >>>> - anything remotely related to embodiment >>>> - any capacity to reflect upon themselves >>>> >>>> Hypothesis: neither keyword matching nor tensor >>>> manipulation, even at scale, suffice in themselves to >>>> qualify for consciousness. >>>> >>>> - Gary >>>> >>>> > On Feb 14, 2022, at 00:24, Geoffrey Hinton >>>> >>> > wrote: >>>> > >>>> > ?Many AI researchers are very unclear about what >>>> consciousness is and also very sure that GPT-3 doesn?t have >>>> it. It?s a strange combination. >>>> > >>>> > >>>> >>> -- > -- > -- -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From kosiorek.adam at gmail.com Tue Feb 15 05:34:53 2022 From: kosiorek.adam at gmail.com (Adam Kosiorek) Date: Tue, 15 Feb 2022 10:34:53 +0000 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: Stephen, It's curious to me that wake-sleep cycles should be included in the notion of consciousness, in the sense that I see no problems with a conscious creature that does not sleep. Could you tell me a little more about your thinking here? Thanks, Adam R. Kosiorek On Tue, 15 Feb 2022 at 07:12, Stephen Jos? Hanson wrote: > Gary, these weren't criterion. Let me try again. > > I wasn't talking about wake-sleep cycles... I was talking about being > awake or asleep and the transition that ensues.. > > Rooba's don't sleep.. they turn off, I have two of them. They turn on > once (1) their batteries are recharged (2) a timer has been set for being > turned on. > > GPT3 is essentially a CYC that actually works.. by reading Wikipedia > (which of course is a terribly biased sample). > > I was indicating the difference between implicit and explicit > learning/problem solving. Implicit learning/memory is unconscious and > similar to a habit.. (good or bad). > > I believe that when someone says "is gpt3 conscious?" they are asking: is > gpt3 self-aware? Roombas know about vacuuming and they are unconscious. > > S > On 2/14/22 12:45 PM, Gary Marcus wrote: > > Stephen, > > On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more > plausible as a consciousness than GPT-3. > > 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. > 2. Roomba makes choices based on an explicit representation of its > location relative to a mapped space. GPT lacks any consistent reflection of > self; eg if you ask it, as I have, if you are you person, and then ask if > it is a computer, it?s liable to say yes to both, showing no stable > knowledge of self. > 3. Roomba has explicit, declarative knowledge eg of walls and other > boundaries, as well its own location. GPT has no systematically > interrogable explicit representations. > > All this is said with tongue lodged partway in cheek, but I honestly don?t > see what criterion would lead anyone to believe that GPT is a more > plausible candidate for consciousness than any other AI program out there. > > ELIZA long ago showed that you could produce fluent speech that was mildly > contextually relevant, and even convincing to the untutored; just because > GPT is a better version of that trick doesn?t mean it?s any more conscious. > > Gary > > On Feb 14, 2022, at 08:56, Stephen Jos? Hanson > wrote: > > ? > > this is a great list of behavior.. > > Some biologically might be termed reflexive, taxes, classically > conditioned, implicit (memory/learning)... all however would not be > conscious in the several senses: (1) wakefulness-- sleep (2) self aware > (3) explicit/declarative. > > I think the term is used very loosely, and I believe what GPT3 and other > AI are hoping to show signs of is "self-awareness".. > > In response to : "why are you doing that?", "What are you doing now", > "what will you be doing in 2030?" > > Steve > > > On 2/14/22 10:46 AM, Iam Palatnik wrote: > > A somewhat related question, just out of curiosity. > > Imagine the following: > > - An automatic solar panel that tracks the position of the sun. > - A group of single celled microbes with phototaxis that follow the > sunlight. > - A jellyfish (animal without a brain) that follows/avoids the sunlight. > - A cockroach (animal with a brain) that avoids the sunlight. > - A drone with onboard AI that flies to regions of more intense sunlight > to recharge its batteries. > - A human that dislikes sunlight and actively avoids it. > > Can any of these, beside the human, be said to be aware or conscious of > the sunlight, and why? > What is most relevant? Being a biological life form, having a brain, being > able to make decisions based on the environment? Being taxonomically close > to humans? > > > > > > > > On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: > >> Also true: Many AI researchers are very unclear about what consciousness >> is and also very sure that ELIZA doesn?t have it. >> >> Neither ELIZA nor GPT-3 have >> - anything remotely related to embodiment >> - any capacity to reflect upon themselves >> >> Hypothesis: neither keyword matching nor tensor manipulation, even at >> scale, suffice in themselves to qualify for consciousness. >> >> - Gary >> >> > On Feb 14, 2022, at 00:24, Geoffrey Hinton >> wrote: >> > >> > ?Many AI researchers are very unclear about what consciousness is and >> also very sure that GPT-3 doesn?t have it. It?s a strange combination. >> > >> > >> >> -- > > -- > -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From lorincz at inf.elte.hu Tue Feb 15 08:49:53 2022 From: lorincz at inf.elte.hu (Andras Lorincz) Date: Tue, 15 Feb 2022 13:49:53 +0000 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: Dear Steve and Gary: This is how I see (try to understand) consciousness and the related terms: (Our) consciousness seems to be related to the close-to-deterministic nature of the episodes on from few hundred millisecond to a few second domain. Control instructions may leave our brain 200 ms earlier than the action starts and they become conscious only by that time. In addition, observations of those may also be delayed by a similar amount. (It then follows that the launching of the control actions is not conscious and -- therefore -- free will can be debated in this very limited context.) On the other hand, model-based synchronization is necessary for timely observation, planning, decision making, and execution in a distributed and slow computational system. If this model-based synchronization is not working properly, then the observation of the world breaks and schizophrenic symptoms appear. As an example, individuals with pronounced schizotypal traits are particularly successful in self-tickling (source: https://philpapers.org/rec/LEMIWP, and a discussion on Asperger and schizophrenia: https://www.frontiersin.org/articles/10.3389/fpsyt.2020.503462/full) a manifestation of improper binding. The internal model enables and the synchronization requires the internal model and thus a certain level of consciousness can appear in a time interval around the actual time instant and its length depends on the short-term memory. Other issues, like separating the self from the rest of the world are more closely related to the soft/hard style interventions (as called in the recent deep learning literature), i.e., those components (features) that can be modified/controlled, e.g., color and speed, and the ones that are Lego-like and can be separated/amputed/occluded/added. Best, Andras ------------------------------------ Andras Lorincz http://nipg.inf.elte.hu/ Fellow of the European Association for Artificial Intelligence https://scholar.google.com/citations?user=EjETXQkAAAAJ&hl=en Department of Artificial Intelligence Faculty of Informatics Eotvos Lorand University Budapest, Hungary ________________________________ From: Connectionists on behalf of Stephen Jos? Hanson Sent: Monday, February 14, 2022 8:30 PM To: Gary Marcus Cc: Connectionists Subject: Re: Connectionists: Weird beliefs about consciousness Gary, these weren't criterion. Let me try again. I wasn't talking about wake-sleep cycles... I was talking about being awake or asleep and the transition that ensues.. Rooba's don't sleep.. they turn off, I have two of them. They turn on once (1) their batteries are recharged (2) a timer has been set for being turned on. GPT3 is essentially a CYC that actually works.. by reading Wikipedia (which of course is a terribly biased sample). I was indicating the difference between implicit and explicit learning/problem solving. Implicit learning/memory is unconscious and similar to a habit.. (good or bad). I believe that when someone says "is gpt3 conscious?" they are asking: is gpt3 self-aware? Roombas know about vacuuming and they are unconscious. S On 2/14/22 12:45 PM, Gary Marcus wrote: Stephen, On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more plausible as a consciousness than GPT-3. 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. 2. Roomba makes choices based on an explicit representation of its location relative to a mapped space. GPT lacks any consistent reflection of self; eg if you ask it, as I have, if you are you person, and then ask if it is a computer, it?s liable to say yes to both, showing no stable knowledge of self. 3. Roomba has explicit, declarative knowledge eg of walls and other boundaries, as well its own location. GPT has no systematically interrogable explicit representations. All this is said with tongue lodged partway in cheek, but I honestly don?t see what criterion would lead anyone to believe that GPT is a more plausible candidate for consciousness than any other AI program out there. ELIZA long ago showed that you could produce fluent speech that was mildly contextually relevant, and even convincing to the untutored; just because GPT is a better version of that trick doesn?t mean it?s any more conscious. Gary On Feb 14, 2022, at 08:56, Stephen Jos? Hanson wrote: ? this is a great list of behavior.. Some biologically might be termed reflexive, taxes, classically conditioned, implicit (memory/learning)... all however would not be conscious in the several senses: (1) wakefulness-- sleep (2) self aware (3) explicit/declarative. I think the term is used very loosely, and I believe what GPT3 and other AI are hoping to show signs of is "self-awareness".. In response to : "why are you doing that?", "What are you doing now", "what will you be doing in 2030?" Steve On 2/14/22 10:46 AM, Iam Palatnik wrote: A somewhat related question, just out of curiosity. Imagine the following: - An automatic solar panel that tracks the position of the sun. - A group of single celled microbes with phototaxis that follow the sunlight. - A jellyfish (animal without a brain) that follows/avoids the sunlight. - A cockroach (animal with a brain) that avoids the sunlight. - A drone with onboard AI that flies to regions of more intense sunlight to recharge its batteries. - A human that dislikes sunlight and actively avoids it. Can any of these, beside the human, be said to be aware or conscious of the sunlight, and why? What is most relevant? Being a biological life form, having a brain, being able to make decisions based on the environment? Being taxonomically close to humans? On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus > wrote: Also true: Many AI researchers are very unclear about what consciousness is and also very sure that ELIZA doesn?t have it. Neither ELIZA nor GPT-3 have - anything remotely related to embodiment - any capacity to reflect upon themselves Hypothesis: neither keyword matching nor tensor manipulation, even at scale, suffice in themselves to qualify for consciousness. - Gary > On Feb 14, 2022, at 00:24, Geoffrey Hinton > wrote: > > ?Many AI researchers are very unclear about what consciousness is and also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > -- [cid:part3.4CFD88BF.56E17955 at rubic.rutgers.edu] -- [cid:part3.4CFD88BF.56E17955 at rubic.rutgers.edu] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: signature.png URL: From hava.siegelmann at gmail.com Tue Feb 15 09:38:51 2022 From: hava.siegelmann at gmail.com (Hava Siegelmann) Date: Tue, 15 Feb 2022 09:38:51 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: <8B9EABD4-D27B-404E-91B8-D78ED4C02F36@susaro.com> References: <8B9EABD4-D27B-404E-91B8-D78ED4C02F36@susaro.com> Message-ID: I guess some researchers believe that there is a huge difference between man made technology and natural. (Evolutionary made / created). creatures. On Tue, Feb 15, 2022 at 1:53 AM Richard Loosemore wrote: > > Geoff, > > Necessary and sufficient conditions. > > For some people, they have a list of necessary conditions for > consciousness. GPT-3 fails that test. > > Those people don?t know the sufficient conditions for consciousness. Your > summary of that is that those people ?are very unclear about what > consciousness is.? > > That should clarify what seems to be a strange combination. > > (I would add that am not one of those people, but I felt like speaking up > on their behalf). > > Richard Loosemore > > > > On Feb 14, 2022, at 3:31 AM, Geoffrey Hinton > wrote: > > > > ?Many AI researchers are very unclear about what consciousness is and > also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From minaiaa at gmail.com Tue Feb 15 09:58:07 2022 From: minaiaa at gmail.com (Ali Minai) Date: Tue, 15 Feb 2022 09:58:07 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> Message-ID: I agree with what Gary is saying. However, I do have a further question. How will we ever know whether even a fully embodied, high performing intelligent machine that can ?explain? its decisions has the capacity for actual self-reflection? How can we be sure that other humans have the capacity for self-reflection, beyond the fact that they are like us, give us understandable reports of their inner thinking, and tickle our mirror system in the right way? This is not to deny that humans can do self-reflection; it is about whether we have an objective method to know that another species not like us is doing it. At some point in what I think will be a fairly distant future, an nth- generation, descendent of GPT-3 may be so sophisticated that the question will become moot. That is the inevitable consequence of a non-dualistic view of intelligence that we all share, and will come to pass unless the current narrow application-driven view of AI derails the entire project. I do think that that that system will have a very different neural architecture much closer to the brain?s, will be embodied, will not need to learn using billions of data points and millions of gradient-descent iterations, and will be capable of actual mental growth (not just learning more stuff but learning new modes of understanding) over its lifetime. It will also not be able to explain its inner thoughts perfectly, nor feel the need to do so because it will not be our obedient servant. In other words, it will not be anything like GPT-3. Ali On Tue, Feb 15, 2022 at 2:11 AM Gary Marcus wrote: > Stephen, > > On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more > plausible as a consciousness than GPT-3. > > 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. > 2. Roomba makes choices based on an explicit representation of its > location relative to a mapped space. GPT lacks any consistent reflection of > self; eg if you ask it, as I have, if you are you person, and then ask if > it is a computer, it?s liable to say yes to both, showing no stable > knowledge of self. > 3. Roomba has explicit, declarative knowledge eg of walls and other > boundaries, as well its own location. GPT has no systematically > interrogable explicit representations. > > All this is said with tongue lodged partway in cheek, but I honestly don?t > see what criterion would lead anyone to believe that GPT is a more > plausible candidate for consciousness than any other AI program out there. > > ELIZA long ago showed that you could produce fluent speech that was mildly > contextually relevant, and even convincing to the untutored; just because > GPT is a better version of that trick doesn?t mean it?s any more conscious. > > Gary > > On Feb 14, 2022, at 08:56, Stephen Jos? Hanson > wrote: > > ? > > this is a great list of behavior.. > > Some biologically might be termed reflexive, taxes, classically > conditioned, implicit (memory/learning)... all however would not be > conscious in the several senses: (1) wakefulness-- sleep (2) self aware > (3) explicit/declarative. > > I think the term is used very loosely, and I believe what GPT3 and other > AI are hoping to show signs of is "self-awareness".. > > In response to : "why are you doing that?", "What are you doing now", > "what will you be doing in 2030?" > > Steve > > > On 2/14/22 10:46 AM, Iam Palatnik wrote: > > A somewhat related question, just out of curiosity. > > Imagine the following: > > - An automatic solar panel that tracks the position of the sun. > - A group of single celled microbes with phototaxis that follow the > sunlight. > - A jellyfish (animal without a brain) that follows/avoids the sunlight. > - A cockroach (animal with a brain) that avoids the sunlight. > - A drone with onboard AI that flies to regions of more intense sunlight > to recharge its batteries. > - A human that dislikes sunlight and actively avoids it. > > Can any of these, beside the human, be said to be aware or conscious of > the sunlight, and why? > What is most relevant? Being a biological life form, having a brain, being > able to make decisions based on the environment? Being taxonomically close > to humans? > > > > > > > > On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: > >> Also true: Many AI researchers are very unclear about what consciousness >> is and also very sure that ELIZA doesn?t have it. >> >> Neither ELIZA nor GPT-3 have >> - anything remotely related to embodiment >> - any capacity to reflect upon themselves >> >> Hypothesis: neither keyword matching nor tensor manipulation, even at >> scale, suffice in themselves to qualify for consciousness. >> >> - Gary >> >> > On Feb 14, 2022, at 00:24, Geoffrey Hinton >> wrote: >> > >> > ?Many AI researchers are very unclear about what consciousness is and >> also very sure that GPT-3 doesn?t have it. It?s a strange combination. >> > >> > >> > >> -- > > -- *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Past-President (2015-2016) International Neural Network Society Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.png Type: image/png Size: 19957 bytes Desc: not available URL: From juyang.weng at gmail.com Tue Feb 15 12:32:49 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Tue, 15 Feb 2022 12:32:49 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear Asim, (A) You wrote: "you mean a certain kind of mathematical formulation can give rise to consciousness?" The maximum likelihood mathematical formulation is not a sufficient condition for conscious learning, but a necessary condition. This local minima issue is a CENTRAL issue for people on this list. The local minima problems, including the Turing Aware 2018, have been giving neural networks a lot of doubts . For example, minimizing supervised motor errors in backprop deep learning have two consequences: (1) It has violated the sensorimotor recurrence that is necessary for conscious learning (all big data violated it) and (2) It requires Post-Selections which amounts to a rarely disclosed protocol flaw: Any such products have a big uncertainty: Each customer of CNN, LSTM and ELM systems has to cast a dice. One of (1) and (2) above is sufficient for such neural networks to become impossible to learn consciousness. Of course, as I posted yesterday, there are about 20 million-dollar problems that prevent such neural networks to earn consciousness. All these 20 million-dollar problems must be solved in order to claim to learn consciousness. (B) You need to spend more time, as you are a mathematician (you can understand). The ML optimality in DN and minimizing fitting errors in CNN, LSTM and ELM are greatly different. The former optimality has been mathematically proven (Weng IJIS 2015). The latter optimality is never proven. The formulation is superficial (only about the fitting error, not studying distributions of system weights). I have proved below that error-backprop depends on casting a dice. J. Weng, "On Post Selections Using Test Sets (PSUTS) in AI", in Proc. International Joint Conference on Neural Networks, pp. 1-8, Shengzhen, China, July 18-22, 2021. PDF file . Best regards, -John On Mon, Feb 14, 2022 at 11:16 PM Asim Roy wrote: > Dear John, > > > > 1. On your statement that ?Maximum likelihood: DN formulation that gives rise > to brain-like consciousness? ? you mean a certain kind of mathematical > formulation can give rise to consciousness? I wish you were right. That > solves our consciousness problem and I don?t know why others are arguing > about it in a different Connectionists email chain. You should claim > this on that chain and get some feedback. > > > > 1. On your statement that ?Do you mean the difference between maximum > likelihood and a specially defined minimization of a cost function is not a > whole lot?? ? I have not studied this deeply, but did a quick search. > For some distributions, they can be equivalent. Here are a few blogs. > Again, I didn?t go through their mathematics. > > > > Why Squared Error Minimization = Maximum Likelihood Estimation | Abhimanyu > (abhimanyu08.github.io) > > > Linear Regression. A unification of Maximum Likelihood? | by William > Fleshman | Towards Data Science > > > > > > > But you consciousness claim is really an eye-opener. I didn?t know about > it. You should claim it on the other Connectionists email chain. > > > > Best, > > Asim > > > > *From:* Juyang Weng > *Sent:* Monday, February 14, 2022 7:26 AM > *To:* Asim Roy > *Cc:* Gary Marcus ; John K Tsotsos < > tsotsos at cse.yorku.ca> > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Asim, > > Do you mean the difference between maximum likelihood and a specially > defined minimization of a cost function is not a whole lot? > > Maximum likelihood: DN formulation that gives rise to brain-like > consciousness. > > Deep Learning: minimize an error rate with supervised class labels. > > John > > > > On Sun, Feb 13, 2022 at 6:17 PM Asim Roy wrote: > > John, > > > > I don?t think this needs a response. There are some difference, but I > don?t think a whole lot. > > > > Asim > > > > *From:* Juyang Weng > *Sent:* Sunday, February 13, 2022 1:22 PM > *To:* Asim Roy ; Gary Marcus > *Cc:* John K Tsotsos > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Aim: > You wrote "If I understand correctly, all learning systems do something > along the lines of maximum likelihood learning or error minimization, like > your DN. What?s your point?" > > "Maximum likelihood learning" (ML) is EXTREMELY different from "error > minimization" (EM) like what Geoff Hinton's group did. > > ML incrementally estimates the best solution from the distribution of a > huge number of parameters such as weights, agees, connections patterns etc. > conditioned on (I) Incremental Learning, (II) a learning experience, and > (III) a limited computations resource, such as the number of neurons. > > > > EM (like what Geoff Hinton's group did) only finds a luckiest network from > multiple trained networks without condition (III) above. All such trained > networks do not estimate the distribution of a huge number of parameters. > Thus, they are all local minima, actually very bad local minima because > their error-backprop does not have competition as I explained in my YouTube > talk: > BMTalk 3D Episode 6: Did Turing Awards Go to Fraud? > https://youtu.be/Rz6CFlKrx2k > > > > > I am writing a paper in which I have proved that without condition (III) > above, a special nearest neighbor classifier I designed can give any > non-zero verification error rate and any non-zero test error rate! > > > > Best regards, > > -John > > > > -- > > Juyang (John) Weng > > > > > -- > > Juyang (John) Weng > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Tue Feb 15 12:53:59 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Tue, 15 Feb 2022 12:53:59 -0500 Subject: Connectionists: Weird beliefs about consciousness Message-ID: Dear Tsvi, You wrote "A huge part of the problem in any discussion about consciousness is there isn't even a clear definition of consciousness". Look at the 5 level definition of consciousness: https://www.merriam-webster.com/dictionary/consciousness You wrote: "So consciousness is not necessary or sufficient for complex thoughts or behavior." I was thinking that way too, until recently. I found consciousness IS REQUIRED for even learning basic intelligence. To put it in a short way so that people on this list can benefit: The motors (as context/actions) in the brain require consciousness in order to learn correctly in the physical world. Please read the first model about conscious learning: J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, USA, Jan.7-9, 2022. PDF file . Best regards, -John ---- From: Tsvi Achler To: Iam Palatnik Cc: Connectionists Subject: Re: Connectionists: Weird beliefs about consciousness -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From arthur.vancamp at bristol.ac.uk Tue Feb 15 15:52:43 2022 From: arthur.vancamp at bristol.ac.uk (Arthur Van Camp) Date: Tue, 15 Feb 2022 20:52:43 +0000 Subject: Connectionists: =?utf-8?q?Tenth_SIPTA_School_on_Imprecise_Probabi?= =?utf-8?q?lities_in_Bristol=2C_UK=2C_August_15_=E2=80=93_19=2C_2022?= References: <44064474-2808-4c76-83d2-bb3a3591880b@Spark> Message-ID: Tenth SIPTA School on Imprecise Probabilities https://ep-imp.gitlab.io/sipta-summer-school-2022/ Imprecise Probabilities The members of SIPTA, the Society for Imprecise Probabilities: Theories and Applications, are convinced that there is more to uncertainty than probabilities. There is much more, in fact. Do you like priors but don't love precise distributions? What about incomparability without indifference? Would you like to model uncertainty without any probabilities at all? You're in luck! There are numerous mathematical models that can measure chance or uncertainty without sharp numerical probabilities. We refer to these as imprecise probabilities. Tenth SIPTA School The tenth SIPTA School will take place in Bristol, UK, from 15 to 19 of August 2022. It will introduce the main theoretical aspects of imprecise probability models, decision making using such models, statistics in imprecise probability theory, the use and applications of imprecise probabilities in artificial intelligence, and evaluating imprecise forecasts. Leading experts in these topics will lecture on the main concepts and techniques associated with their area of expertise, in a friendly environment favouring interaction between participants. Tentative schedule: * Monday, 15 August, first session (1 hour): "Why imprecise probabilities? by Ben Levinstein * Monday, 15 August, second session (all day except first hour): ?Introduction to imprecise probabilities? by Gert de Cooman * Tuesday, 16 August (all day): ?Decision making? by Teddy Seidenfeld * Wednesday, 17 August (morning): ?Statistics and imprecise probabilities? by Thomas Augustin * Wednesday, 17 August (afternoon): excursion * Thursday, 18 August (all day): ?Imprecise probabilities in artificial intelligence? by Alessandro Antonucci, Cassio de Campos and Fabio Cozman * Friday, 19 August (all day): ?Accuracy for imprecise probabilities? by Jason Konek and Ben Levinstein Participation and registration We will have a limited number of seats available for in-person attendance (~40 people), but virtually unlimited virtual capacity. The fee for virtual registration will be ?25. We have two options for in-person attendance: Room and Board (?250) and Lunch Only (?25). * attending virtually: ?25 * attending in-person Room and Board: ?250 * attending in-person Lunch Only: ?25 To register, please fill out this this registration form by 15 May 2022. You will need a small statement explaining your research interests (in text format) and a short CV (as a pdf or document file). Accommodation and meals For in-person participants, we have two registration options: Room and Board: the fee is ?250. This includes student accommodation at Goldney Hall for the nights of 14 ? 19 August (i.e., ending the morning of the 20th). Breakfast will be included (at Goldney Hall), and we will also cater lunch every day of the summer school. Lunch Only: the fee is ?25. This option includes the same daily catered lunches as the Room and Board option, but no accommodation or breakfast. Note on Accommodation at Goldney Hall: Not all of the available rooms have an en suite bathroom; some rooms will share a communal bathroom. If you need a private bathroom, please indicate this on the form and give a brief explanation of why. We will do our best to assign rooms to accommodate everyone who has a compelling need. Scholarship/Bursaries If you are a student who would like to attend but would find the cost to be a hardship, we would like to help. In addition to waiving the fee for registration, we may be able to defray travel expenses. When filling out the registration form, please take care to: (1) indicate that you are a student, and (2) check the box expressing interest in being considered. Important dates The registration via this form closes 15 May 2022. The registration fee should be paid before 1 June 2022. We are looking forward to a great summer school! Kind regards, Jason Konek, chair Fabio Cozman, external advisor Kevin Blackwell Catrin Campbell-Moore Giacomo Molinari Richard Pettigrew -------------- next part -------------- An HTML attachment was scrubbed... URL: From achler at gmail.com Tue Feb 15 22:05:45 2022 From: achler at gmail.com (Tsvi Achler) Date: Tue, 15 Feb 2022 19:05:45 -0800 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: After studying the brain from a multidisciplinary perspective I am well aware of the difficulties speaking and understanding each other across disciplines. There are many terms that are defined differently in different fields... and unfortunately things are not as simple as looking them up in a dictionary. For example the term recurrent connections have different meanings in the computational neuroscience, neural networks, and cognitive psychology communities. In neural networks recurrent means an output used back as an input within a paradigm of delayed inputs. It is a method of representing time or sequences. Often recurrent connections in neural networks are confused with feedback back to the same inputs which are actually never used in neural networks because it forms an infinite loop and is not possible to rewind in order to generate an error signal. In computational neuroscience recurrent connections are used to describe lateral connections. In cognitive psychology the term re-entrant connections are used to describe feedback back to the same inputs. I believe in order to truly appreciate "brain-like" ideas, members of this group need to familiarize themselves with these brain-focused fields. For example in cognitive psychology there is a rich literature on salience (which again is a bit different from salience in the neural network community). Salience is a dynamic process which determines how well a certain input or input feature is processed. Salience changes in the brain depending on what other inputs or features are concurrently present or what the person is instructed to focus on. There is very little appreciation, integration or implementation of these findings in neural networks, yet salience plays a factor in every recognition decision and modality including smell and touch. The term Consciousness is a particularly problematic minefield which also adds in philosophy, metaphysics and subjectivity into the mix. Juyang, I think we both agree about the basics: the need for more realistic real world recognition and to move beyond the rehearsal limitations of neural networks. I believe scientists not seeing eye-to-eye with each other and other members of the community is in no small part due to some of these terms. Sincerely, -Tsvi On Tue, Feb 15, 2022 at 9:54 AM Juyang Weng wrote: > Dear Tsvi, > You wrote "A huge part of the problem in any discussion about > consciousness is there isn't even a clear definition of consciousness". > Look at the 5 level definition of consciousness: > https://www.merriam-webster.com/dictionary/consciousness > > You wrote: "So consciousness is not necessary or sufficient for complex > thoughts or behavior." > I was thinking that way too, until recently. > I found consciousness IS REQUIRED for even learning basic intelligence. > To put it in a short way so that people on this list can benefit: > The motors (as context/actions) in the brain require consciousness in > order to learn correctly in the physical world. Please read the first > model about conscious learning: > J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th > International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, > USA, Jan.7-9, 2022. PDF file > > . > > Best regards, > -John > ---- > From: Tsvi Achler > To: Iam Palatnik > Cc: Connectionists > Subject: Re: Connectionists: Weird beliefs about consciousness > > -- > Juyang (John) Weng > -------------- next part -------------- An HTML attachment was scrubbed... URL: From cgf at isep.ipp.pt Tue Feb 15 18:26:27 2022 From: cgf at isep.ipp.pt (Carlos) Date: Tue, 15 Feb 2022 23:26:27 +0000 Subject: Connectionists: CFP: MLJ special issue on Foundations of Data Science Message-ID: <7241144e-40ba-ea50-d040-7e30346a6ed7@isep.ipp.pt> Data science is a hot topic with an extensive scope, both in terms of theory and applications. Machine Learning forms one of its core foundational pillars. Simultaneously, Data Science applications provide important challenges that can often be addressed only with innovative Machine Learning algorithms and methodologies. This special issue will highlight the latest development of the Machine Learning foundations of data science and on the synergy of data science and machine learning. We welcome new developments in statistics, mathematics, informatics and computing-driven machine learning for data science, including foundations, algorithms and models, systems, innovative applications and other research contributions. Following the great success of the 2021 MLJ special issue with DSAA'2021, this 2022 special issue will further capture the state-of-the-art machine learning advances for data science. Accepted papers will be published in MLJ and presented at a journal track of the 2022 IEEE International Conference on Data Science and Advanced Analytics (DSAA'2022) in Shenzhen, October 2022. ==================== Topics of Interest ==================== We welcome original and well-grounded research papers on all aspects of foundations of data science including but not limited to the following topics: Machine Learning Foundations for Data Science * Auto-ML * Information fusion from disparate sources * Feature engineering, embedding, mining and representation * Learning from network and graph data * Learning from data with domain knowledge * Reinforcement learning * Non-IID learning, nonstationary, coupled and entangled learning * Heterogeneous, mixed, multimodal, multi-view and multi-distributional learning * Online, streaming, dynamic and real-time learning * Causality and learning causal models * Multi-instance, multi-label, multi-class and multi-target learning * Semi-supervised and weakly supervised learning * Representation learning of complex interactions, couplings, relations * Deep learning theories and models * Evaluation of data science systems * Open domain/set learning Emerging Impactful Machine Learning Applications * Data preprocessing, manipulation and augmentation * Autonomous learning and optimization systems * Digital, social, economic and financial (finance, FinTech, blockchains and cryptocurrencies) analytics * Graph and network embedding and mining * Machine learning for recommender systems, marketing, online and e-commerce * Augmented reality, computer vision and image processing * Risk, compliance, regulation, anomaly, debt, failure and crisis * Cybersecurity and information disorder, misinformation/fake detection * Human-centered and domain-driven data science and learning * Privacy, ethics, transparency, accountability, responsibility, trust, reproducibility and retractability * Fairness, explainability and algorithm bias * Green and energy-efficient, scalable, cloud/distributed and parallel analytics and infrastructures * IoT, smart city, smart home, telecommunications, 5G and mobile data science and learning * Government and enterprise data science * Transportation, manufacturing, procurement, and Industry 4.0 * Energy, smart grids and renewable energies * Agricultural, environmental and spatio-temporal analytics and climate change Contributions must contain new, unpublished, original and fundamental work relating to the Machine Learning Journal's mission. All submissions will be reviewed using rigorous scientific criteria whereby the novelty of the contribution will be crucial. ==================== Submission Instructions ==================== Submit manuscripts to: http://MACH.edmgr.com. Select this special issue as the article type. Papers must be prepared in accordance with the Journal guidelines: https://www.springer.com/journal/10994 All papers will be reviewed following standard reviewing procedures for the Journal. ==================== Key Dates ==================== We will have a continuous submission/review process starting in Oct. 2021. Last paper submission deadline: 1 March 2022 Paper acceptance: 1 June 2022 Camera-ready: 15 June 2022 ==================== Guest Editors ==================== Longbing Cao, University of Technology Sydney, Australia Jo?o Gama, University of Porto, Portugal Nitesh Chawla, University of Notre Dame, United States Joshua Huang, Shenzhen University, China Carlos Ferreira ISEP | Instituto Superior de Engenharia do Porto Rua Dr. Ant?nio Bernardino de Almeida, 431 4249-015 Porto - PORTUGAL tel. +351 228 340 500 | fax +351 228 321 159 mail at isep.ipp.pt | www.isep.ipp.pt From iswc.conf at gmail.com Tue Feb 15 20:13:38 2022 From: iswc.conf at gmail.com (International Semantic Web Conference) Date: Tue, 15 Feb 2022 20:13:38 -0500 Subject: Connectionists: Joint CFP : 21st International Semantic Web Conference Message-ID: *Joint CfP: 21st International Semantic Web Conference (ISWC 2022)* Hangzhou, China, October 23-27, 2022 https://iswc2022.semanticweb.org/ In this announcement: 1. Call for Research papers 2. Call for In-Use papers 3. Call for Resource papers 1. Call for Research papers ******************************************* The* research track* of ISWC 2022 solicits novel and significant research contributions addressing theoretical, analytical and empirical aspects of the Semantic Web. We welcome work that relates to the W3C Semantic Web recommendations (e.g., RDF, OWL, SPARQL, SHACL, etc.) that may also lie at the intersection of the Semantic Web and other scientific disciplines. Submissions to the research track should describe original, significant, and replicable research on the Semantic Web. Topics of interest and further details: https://iswc2022.semanticweb.org/index.php/research-track/ Deadline: Friday, 28th April, 2022, 23:59 AoE (Anywhere on Earth) Track Chairs: - Aidan Hogan, Department of Computer Science, Universidad de Chile ahogan at dcc.uchile.cl - Uli Sattler, Department of Computer Science, University of Manchester Uli.Sattler at manchester.ac.uk 2. Call for In-Use papers ******************************************* The track aims to give a stage to solutions for real-world problems in which Semantic Web and Knowledge Graph technologies play a crucial role. Real-world applications of these technologies in combination with machine learning, deep learning and other AI techniques are of particular interest. The In-Use track thus seeks submissions describing applied and validated solutions such as software tools, systems or architectures that benefit from the use of Semantic Web and Knowledge Graph technologies (including, but not limited to, technologies based on the Semantic Web standards). Importantly, submitted papers should provide convincing evidence of use of the proposed application or tool by the target user group, preferably outside the group that conducted the development and, more broadly, outside the Semantic Web and Knowledge Graph research communities. Topics of interest and further details: https://iswc2022.semanticweb.org/index.php/in-use-track/ Deadline: Friday, 28th April, 2022, 23:59 AoE (Anywhere on Earth) Track Chairs: - Jo?o Paulo A. Almeida, Federal University of Esp?rito Santo, Brazil - Hideaki Takeda, National Institute of Informatics, Japan Contact: In-UseTrack-iswc2022 at easychair.org 3. Call for Resource papers ******************************************* The ISWC 2022 Resources Track aims to promote the sharing of resources which support, enable or utilise semantic web research. Resources include, but not restricted to: datasets, ontologies/vocabularies, ontology design patterns, evaluation benchmarks or methods, software tools/services, APIs and software frameworks, workflows, crowdsourcing task designs, protocols, methodologies and metrics, that have contributed or may contribute to the generation of novel scientific work. In particular, we encourage the sharing of such resources following best and well-established practices within the Semantic Web community. As such, this track calls for contributions that provide a concise and clear description of a resource and its usage. Topics of interest and further details: https://iswc2022.semanticweb.org/index.php/resources-track/ Deadline: Friday, 28th April, 2022, 23:59 AoE (Anywhere on Earth) Track Chairs: - Valentina Presutti, LILEC Department, University of Bologna, Italy valentina.presutti at unibo.it - Maria Keet, Department of Computer Science, University of Cape Town, South Africa mkeet at cs.uct.ac.za Follow us on social media: - Twitter: @iswc_conf #iswc_conf (https://twitter.com/iswc_conf) - LinkedIn: https://www.linkedin.com/groups/13612370 - Facebook: https://www.facebook.com/ISWConf - Instagram: https://www.instagram.com/iswc_conf/ The ISWC 2022 Organizing Team Organizing Committee ? ISWC 2022 (semanticweb.org) -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Tue Feb 15 13:05:52 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Tue, 15 Feb 2022 13:05:52 -0500 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton Message-ID: Dear Asim, You wrote "If I had the kind of technology that you list here, I wouldn?t waste time on the Connectionist list. That kind of technology could be worth billions of dollars, if not trillions. So, look for investors." Many people who have issued "call for paper" on brain or consciousness in fact are not capable of understanding a true solution to brain or consciousness. That is why I am not wasting time here. I am helping many people on this list by spending my precious time. Through this list, I hope some investors looking for great investment opportunities can read my posts. This is not too good to be true. Best regards, -John -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From ioannakoroni at csd.auth.gr Wed Feb 16 04:32:17 2022 From: ioannakoroni at csd.auth.gr (Ioanna Koroni) Date: Wed, 16 Feb 2022 11:32:17 +0200 Subject: Connectionists: Participation in the free International AI Doctoral Academy (AIDA) Short course "Human Behavior Analysis with a Social Signal Processing Perspective" by Prof. Cigdem Beyan, February 23-24, 2022 References: <002601d8224f$f9af0fb0$ed0d2f10$@csd.auth.gr> Message-ID: <008101d82318$17584e00$4608ea00$@csd.auth.gr> COURSE TITLE: Human Behavior Analysis with a Social Signal Processing Perspective LECTURER: Cigdem Beyan, cigdem.beyan at unitn.it ORGANIZER: Department of Information Engineering and Computer Science, University of Trento, Italy CONTENT AND ORGANIZATION: Social Signal Processing (SSP) is a cross-disciplinary field that covers studying social signals that 1) are produced during human-human interaction, 2) play a part in the formation and adjustment of relationships between agents (being humans and embodied interfaces) or provide information about the interlocutors, and 3) can be synthetized in embodied interfaces. They can be related to social (inter)actions, social emotions, social evaluations, social attitudes and social relations. The significant part of SSP deals with automatic detection of social attitudes and relations, such as leadership or hostile attitude from the data captured with a wide array of sensors. During this lecture, I will present several datasets and applications (in small groups, mingling scenarios, egocentric vision based) using multimodal data (audio, video, text) and the state-of-the-art machine/deep learning methods. REGISTRATION: Free of charge WHEN: 23-24 February 2022 13.30- 15.30 CET WHERE: Online (telco link to be provided by the Lecturer after registration/enrollment) HOW TO REGISTER and ENROLL: Both AIDA and non-AIDA students are encouraged to participate in this short course. If you are an AIDA Student* already, please: Step (a): Register in the course by following the instructions given in "Teaching" tab of cbeyan.github.io or alternatively by sending an email to cigdem.beyan[at]unitn.it for your registration. AND Step (b): Enroll in the same course in the AIDA system using the button below, so that this course enters your AIDA Certificate of Course Attendance. If you are not an AIDA Student do only step (a). *The International AI Doctoral Academy (AIDA) has 73 members, which are top AI Universities, Research centers and Industries: https://www.i-aida.org/ AIDA Students should have been registered in the AIDA system already (they are PhD students or PostDocs that belong only to the AIDA Members listed in this page: Members ) -- This email has been checked for viruses by Avast antivirus software. https://www.avast.com/antivirus -------------- next part -------------- An HTML attachment was scrubbed... URL: From marcin at amu.edu.pl Wed Feb 16 05:24:12 2022 From: marcin at amu.edu.pl (Marcin Paprzycki) Date: Wed, 16 Feb 2022 11:24:12 +0100 Subject: Connectionists: CFP -- FedCSIS 2022 -- Hybrid Meeting -- CORE B ranked -- IEEE DL -- IEEE: #52320 In-Reply-To: References: Message-ID: CALL FOR PAPERS 17th Conference on Computer Science and Intelligence Systems (FedCSIS?2022) HYBRID CONFERENCE -- Sofia, Bulgaria, 4-7 September, 2022 (CORE B Ranked, IEEE: #52320, 70 punkt?w parametrycznych MEiN) www.fedcsis.org Strict submission deadline: May 10, 2022, 23:59:59 AOE (no extensions) ************************** COVID-19 Information ************************ Conference will take place in Sofia Bulgaria, for those who will be able to make it there. For those who will not be able to reach Sofia, online participation will be made available. ************************************************************************ Please feel free to forward this announcement to your colleagues and associates who could be interested in it. KEY FACTS: Proceedings: submitted to IEEE Digital Library; indexing: DBLP, Scopus and Web of Science; 21 Technical Sessions in 5 Tracks; Doctoral Symposium; Data Mining Competition. FedCSIS is an annual international conference, this year organized jointly by the Polish Information Processing Society (PTI), IEEE Poland Section Computer Society Chapter and Bulgarian Academy of Sciences. It is technically sponsored by a number of IEEE units and other professional organizations (for the full list, see the conference site). The mission of the FedCSIS Conference Series is to provide a highly acclaimed forum in computer science and intelligence systems. We invite researchers from around the world to contribute their research results and participate in Technical Sessions, focused on their scientific and professional interests in computer science and intelligence systems. Since June 2021, FedCSIS is ranked CORE B. This resulted in FedCSIS delivering 70 parametric points to Polish researchers (since November 2021). Since 2012, Proceedings of the FedCSIS conference are indexed in SCOPUS, DBLP and other indexing services. Information about FedCSIS indexing / bibliometry / rankings can be found at: https://www.fedcsis.org/2022/indexation. FedCSIS TRACKS AND TECHNICAL SESSIONS The FedCSIS 2022 consists of five conference Tracks, hosting Technical Sessions: Track 1: Advanced Artificial Intelligence in Applications (17th Symposium AAIA'22) * Artificial Intelligence for Next-Generation Diagnostis Imaging (1st Workshop AI4NextGenDI'22) * Intelligence for Patient Empowerment with Sensor Systems (1st Workshop AI4Empowerment'22) * Intelligence in Machine Vision and Graphics (4th Workshop AIMaViG'22) * Ambient Assisted Living Systems (1st Workshop IntelligentAAL'22) * Personalization and Recommender Systems (1st Workshop PeRS'22) * Rough Sets: Theory and Applications (4th International Symposium RSTA'22) * Computational Optimization (15th Workshop WCO'22) Track 2: Computer Science & Systems (CSS'22) * Actors, Agents, Assistants, Avatars (1st Workshop 4A'22) * Computer Aspects of Numerical Algorithms (15th Workshop CANA'22) * Concurrency, Specification and Programming (30th Symposium CS&P'22) * Multimedia Applications and Processing (15th International Symposium MMAP'22) * Scalable Computing (12th Workshop WSC'22) Track 3: Network Systems and Applications (NSA'22) * Complex Networks - Theory and Application (1st Workshop CN-TA'22) * Internet of Things - Enablers, Challenges and Applications (6th Workshop IoT-ECAW'22) * Cyber Security, Privacy and Trust (3rd International Forum NEMESIS'22) Track 4: Advances in Information Systems and Technologies (AIST'22) * Data Science in Health, Ecology and Commerce (4th Workshop DSH'22) * Information Systems Management (17th Conference ISM'22) * Knowledge Acquisition and Management (28th Conference KAM'22) Track 5: Software, System and Service Engineering (S3E'22) * Cyber-Physical Systems (9th Workshop IWCPS-8) * Model Driven Approaches in System Development (7th Workshop MDASD'22) * Software Engineering (42th IEEE Workshop SEW-42) Recent Advances in Information Technology (7th Symposium DS-RAIT'22) KEYNOTE SPEAKERS * Krassimir Atanassov Bulgarian Academy of Sciences, Sofia, Bulgaria https://scholar.google.com/citations?hl=pl&user=K-vuWKsAAAAJ * Thomas Blaschke University of Salzburg, Salzburg, Austria https://scholar.google.com/citations?user=kMroJzUAAAAJ * Chris Cornelis Ghent University, Ghent, Belgium https://scholar.google.com/citations?hl=pl&user=ln46HlkAAAAJ * Franco Zambonelli University of Modena e Reggio Emilia, Bologna, Italy https://scholar.google.com/citations?hl=pl&user=zxulxcoAAAAJ DATA MINING COMPETITION -- Predicting the Costs of Forwarding Contracts The topic of this year's data mining competition is prediction of costs related to the execution of forwarding contracts. The data sets will contain multiple years of history of orders appearing on the transport exchange, along with details such as the type of order, basic characteristics of the shipped goods (e.g., dimensions, special requirements), as well as the expected route that a driver will have to cover. The task will be to develop a predictive model that will assess the actual cost of individual orders as accurately as possible (for detailed information see: https://fedcsis.org/2022/aaia/dm). Prizes for best solutions are: ? 1,000 USD for the winning solution (+ the cost of one FedCSIS?2022 registration) ? 500 USD for the 2nd place solution (+ the cost of one FedCSIS?2022 registration) ? 250 USD for the 3rd place solution (+ the cost of one FedCSIS?2022 registration) ZDZISLAW PAWLAK BEST PAPER AWARD The Professor Zdzislaw Pawlak Awards are given in the following categories: ? Best Paper Award (?600) ? Young Researcher Paper Award (?400) ? Industry Cooperation Award (?400) ? International Cooperation Award (?400) All papers accepted to FedCSIS 2022 are eligible to be considered as the award winners. This award will be granted independently from awards given by individual FedCSIS events (Tracks and/or Technical Sessions). Past Award winners can be found here: https://fedcsis.org/2022/zp_award PAPER SUBMISSION AND PUBLICATION + Papers should be submitted by May 10, 2022 (strict deadline, no extensions, submission system is open, via EasyChair https://easychair.org/my/conference?conf=fedcsis2022). + Preprints will be published online. + Only papers presented during the conference will be submitted to the IEEE for inclusion in the Xplore Digital Library. Furthermore, proceedings, published in a volume with ISBN, ISSN and DOI numbers, will be posted within the conference Web portal. Moreover, most Technical Session organizers arrange quality journals, edited volumes, etc., and may invite selected extended and revised papers for post-conference publications (information can be found at the websites of individual events, or by contacting Chairs of said events). IMPORTANT DATES + Paper submission (strict deadline): May 10, 2022, 23:59:59 (AOE; there will be no extension) + Position paper submission: June 7, 2022 + Author notification: July 6, 2022 + Final paper submission and registration: July 12, 2022 + Payment (early fee deadline): August 2, 2022 + Conference date: September 4-7, 2022 CHAIRS OF FedCSIS CONFERENCE SERIES Maria Ganzha, Marcin Paprzycki, Dominik Slezak CONTACT FedCSIS at: secretariat at fedcsis.org FedCSIS in Social Media: FedCSIS on Facebook: http://tinyurl.com/FedCSISFacebook FedCSIS on LinkedIN: https://tinyurl.com/FedCSISonLinkedIN FedCSIS on Twitter: https://twitter.com/FedCSIS FedCSIS on XING: http://preview.tinyurl.com/FedCSISonXING -- Ta wiadomo?? zosta?a sprawdzona na obecno?? wirus?w przez oprogramowanie antywirusowe Avast. https://www.avast.com/antivirus From bwyble at gmail.com Wed Feb 16 08:30:54 2022 From: bwyble at gmail.com (Brad Wyble) Date: Wed, 16 Feb 2022 08:30:54 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Tsvi you wrote: > > For example in cognitive psychology there is a rich literature on > salience (which again is a bit different from salience in the neural > network community). Salience is a dynamic process which determines how > well a certain input or input feature is processed. Salience changes in the > brain depending on what other inputs or features are concurrently present > or what the person is instructed to focus on. There is very little > appreciation, integration or implementation of these findings in neural > networks, yet salience plays a factor in every recognition decision and > modality including smell and touch. > > I'm having trouble understanding what you mean by this, since computational modelling of salience is a major thrust of computer vision. Itti Koch & Niebur (1998) has been cited 13,000 times and there are hundreds of papers that have elaborated on this ANN approach to salience computation in vision. Is this not what you're asking for? If not, what am I misunderstanding? kind regards -Brad -------------- next part -------------- An HTML attachment was scrubbed... URL: From amartino at luiss.it Wed Feb 16 09:13:40 2022 From: amartino at luiss.it (Alessio Martino) Date: Wed, 16 Feb 2022 14:13:40 +0000 Subject: Connectionists: =?utf-8?q?=5BExtended_Deadline=5D_CfP_SI_on_Compl?= =?utf-8?q?ex_Systems_Modelling_via_Hypergraphs=C2=A0=5BEntropy=2C_MDPI=5D?= Message-ID: Dear Colleagues, I am contacting you in my capacity as Guest Editor for a Special Issue titled: "Advances in Complex Systems Modelling via Hypergraphs" to appear in ?Entropy? MDPI journal: https://www.mdpi.com/journal/entropy/special_issues/Hypergraphs With this call for papers, I invite you and/or your co-authors to submit an original research paper, or a focused review, for our special issue. Deadline for manuscript submissions: 31 August 2022. Submitted papers will be peer reviewed and, upon acceptance, the paper will be published in open access form soon after professional editing. Thank you in advance for your consideration and I sincerely hope that you will accept this invitation to contribute to this Special Issue. If you believe that you will be able to submit a manuscript, I would also greatly appreciate if you could respond to this invitation at your earliest convenience. ?Entropy? (ISSN 1099-4300, IF: 2.494) is an EI, Scopus and ESCI indexed, Open Access journal published online monthly by MDPI. Best Regards ________________________________________ Alessio Martino, PhD Assistant Professor of Computer Science LUISS Guido Carli University Department of Business and Management Viale Romania 32, 00197 Rome, Italy (Room 539) Phone: (+39) 06-85225957 E-mail: amartino at luiss.it Web: La presente e-mail proviene da Luiss Guido Carli e s'intende inviata per scopi lavorativi. Tutte le informazioni ivi contenute, compresi eventuali allegati, sono da ritenersi esclusivamente confidenziali e riservati secondo i termini del vigente D.Lgs. 196/2003 in materia di privacy e del Regolamento europeo 679/2016 - GDPR. ? vietato qualsiasi ulteriore utilizzo non autorizzato. Qualora la stessa Le fosse pervenuta per errore, La preghiamo di eliminarla immediatamente e di darcene tempestiva comunicazione. Grazie. This e-mail message is sent by Luiss Guido Carli for business purposes. All informations contained therein, including any attachments, are for the sole use of the intended recipient and may contain confidential and privileged information pursuant to Legislative Decree 196/2003 and the European General Data Protection Regulation 679/2016 - GDPR -. Any unauthorized review, use, disclosure or distribution is prohibited. If you are not the intended recipient, please contact the sender by soon reply this e-mail and destroy all copies of the original message. Thanks -------------- next part -------------- An HTML attachment was scrubbed... URL: From fabio.bellavia at unifi.it Wed Feb 16 08:45:59 2022 From: fabio.bellavia at unifi.it (Fabio Bellavia) Date: Wed, 16 Feb 2022 14:45:59 +0100 Subject: Connectionists: A month left ! - Fine Art Pattern Extraction and Recognition (FAPER2022) @ICIAP2021 Message-ID: ??????????? ---===== Apologies for cross-postings =====--- ?????????? Please distribute this call to interested parties ________________________________________________________________________ ?International Workshop on Fine Art Pattern Extraction and Recognition ????????????????????????? F A P E R?? 2 0 2 2 ??????? in conjunction with the 21st International Conference on ?????????????? Image Analysis and Processing (ICIAP 2021) ???????????????????? Lecce, Italy, MAY 23-27, 2022 ????????????? +++ SUBMISSION DEADLINE: March 15, 2022 +++ ??????????? >>> https://sites.google.com/view/faper2022 <<< -> Submission link: https://easychair.org/conferences/?conf=faper2022 <- ????????????? [[[ both virtual and in presence event ]]] ________________________________________________________________________ === Aim & Scope === Cultural heritage, especially fine arts, plays an invaluable role in the cultural, historical and economic growth of our societies. Fine arts are primarily developed for aesthetic purposes and are mainly expressed through painting, sculpture and architecture. In recent years, thanks to technological improvements and drastic cost reductions, a large-scale digitization effort has been made, which has led to an increasing availability of large digitized fine art collections. This availability, coupled with recent advances in pattern recognition and computer vision, has disclosed new opportunities, especially for researchers in these fields, to assist the art community with automatic tools to further analyze and understand fine arts. Among other benefits, a deeper understanding of fine arts has the potential to make them more accessible to a wider population, both in terms of fruition and creation, thus supporting the spread of culture. Following the success of the first edition, organized in conjunction with ICPR 2020, the aim of the workshop is to provide an international forum for those wishing to present advancements in the state-of-the-art, innovative research, ongoing projects, and academic and industrial reports on the application of visual pattern extraction and recognition for a better understanding and fruition of fine arts. The workshop solicits contributions from diverse areas such as pattern recognition, computer vision, artificial intelligence and image processing. === Topics === Topics of interest include, but are not limited to: - Application of machine learning and deep learning to cultural heritage and digital humanities - Computer vision and multimedia data processing for fine arts - Generative adversarial networks for artistic data - Augmented and virtual reality for cultural heritage - 3D reconstruction of historical artifacts - Point cloud segmentation and classification for cultural heritage - Historical document analysis - Content-based retrieval in the art domain - Speech, audio and music analysis from historical archives - Digitally enriched museum visits - Smart interactive experiences in cultural sites - Projects, products or prototypes for cultural heritage restoration, preservation and fruition - Visual question answering and artwork captioning - Art history and computer vision === Invited speaker === Eva Cetinic (Digital Visual Studies, University of Zurich, Switzerland) - "Beyond Similarity: From Stylistic Concepts to Computational Metrics" Dr. Eva Cetinic is currently working as a postdoctoral fellow at the Center for Digital Visual Studies at the University of Zurich. She previously worked as a postdoc in Digital Humanities and Machine Learning at the Department of Computer Science, Durham University, and as a postdoctoral researcher and professional associate at the Ru?er Bo?kovic Institute in Zagreb. She obtained her Ph.D. in Computer science from the Faculty of Electrical Engineering and Computing, University of Zagreb in 2019 with the thesis titled "Computational detection of stylistic properties of paintings based on high-level image feature analysis". Besides being generally interested in the interdisciplinary field of digital humanities, her specific interests focus on studying new research methodologies rooted in the intersection of artificial intelligence and art history. Particularly, she is interested in exploring deep learning techniques for computational image understanding and multi-modal reasoning in the context of visual art. === Workshop modality === The workshop will be held in a hybrid form, both virtual and in presence participation will be allowed. === Submission guidelines === Accepted manuscripts will be included in the ICIAP 2021 proceedings, which will be published by Springer as Lecture Notes in Computer Science series (LNCS). Authors of selected papers will be invited to extend and improve their contributions for a Special Issue on IET Image Processing. Please follow the guidelines provided by Springer when preparing your contribution. The maximum number of pages is 10 + 2 pages for references. Each contribution will be reviewed on the basis of originality, significance, clarity, soundness, relevance and technical content. Once accepted, the presence of at least one author at the event and the oral presentation of the paper are expected. Please submit your manuscript through EasyChair: https://easychair.org/conferences/?conf=faper2022 === Important Dates === - Workshop submission deadline: March 15, 2022 - Author notification: April 1, 2022 - Camera-ready submission and registration: April 15, 2022 - Workshop day: May 23-24, 2022 === Organizing committee === Gennaro Vessio (University of Bari, Italy) Giovanna Castellano (University of Bari, Italy) Fabio Bellavia (University of Palermo, Italy) Sinem Aslan (University of Venice, Italy | Ege University, Turkey) === Venue === The workshop will be hosted at Convitto Palmieri, which is located in Piazzetta di Giosue' Carducci, Lecce, Italy ____________________________________________________ ?Contacts: gennaro.vessio at uniba.it ?????????? giovanna.castellano at uniba.it ?????????? fabio.bellavia at unipa.it ?????????? sinem.aslan at unive.it ?Workshop: https://sites.google.com/view/faper2022 ICIAP2021: https://www.iciap2021.org/ From franciscocruzhh at gmail.com Wed Feb 16 09:55:58 2022 From: franciscocruzhh at gmail.com (Francisco Cruz) Date: Wed, 16 Feb 2022 11:55:58 -0300 Subject: Connectionists: Last CfP -- Special Issue on Human-aligned Reinforcement Learning for Autonomous Agents and Robots @NCAA journal Message-ID: ** Last Call for Papers ** The deadline for submission has been extended until 28th February 2022. However, the review process for manuscripts will start as soon as they are being received. If you have further questions, please do not hesitate to contact the guest editors. See relevant details at: https://www.springer.com/journal/521/updates/19055662 ** Call for papers ** Topical Collection on Human-aligned Reinforcement Learning for Autonomous Agents and Robots at the Springer journal Neural Computing and Applications. ** Topics ** The main topics of interest in the call for submissions are explainability, interactivity, safety, and ethics in social robotics and autonomous agents, especially from a reinforcement learning perspective. In this regard, approaches with special interest for this topical collection are (but not limited to): - Explainability, interpretability, and transparency methods for feature-oriented and goal-driven RL. - Explainable robotic systems with RL approaches. - Assisted and interactive RL in human-robot and human-agent scenarios. - Human-in-the-loop RL and applications. - RL from demonstrations and imperfect demonstrations. - Robot and agent learning from multiple human sources. - Multi-robot systems with human collaboration. - Safe exploration during learning. - Ethical reasoning and moral uncertainty. - Fairness in RL and multi-agent systems. - Theory of mind based RL frameworks. - Use of human priors in RL. ** Provisional deadlines ** - Deadline for submissions: December 15, 2021 (extended to 28th February, 2022) - Decisions: April, 2022 - Revised manuscript submission: May, 2022 - Deadline for second review: June, 2022 - Final decisions: July, 2022 ** Guest editors ** Dr. Francisco Cruz (Lead guest editor) School of Information Technology Deakin University, Australia Dr. Thommen George Karimpanal Applied Arti?cial Intelligence Institute (A2I2) Deakin University, Australia Dr. Miguel Solis, Facultad de Ingenieria Universidad Andres Bello, Chile Dr. Pablo Barros Cognitive Architecture for Collaborative Technologies Unit Italian Institute of Technology (IIT), Italy A/Prof. Richard Dazeley School of Information Technology Deakin University, Australia ** More details at: ** https://www.springer.com/journal/521/updates/19055662 -------------- next part -------------- An HTML attachment was scrubbed... URL: From m.plumbley at surrey.ac.uk Wed Feb 16 10:23:58 2022 From: m.plumbley at surrey.ac.uk (Mark Plumbley) Date: Wed, 16 Feb 2022 15:23:58 +0000 Subject: Connectionists: JOB: Research Engineer (Research Fellow) in Sound Sensing, University of Surrey, UK Message-ID: Dear List, (with apologies for cross-posting) We are recruiting for a Research Engineer in Sound Sensing, as part of the project AI for Sound (https://ai4s.surrey.ac.uk/). Please forward the information below to any colleagues who may be interested. I would particularly like to encourage applications from women, disabled and Black, Asian and minority ethnic candidates, since these groups are currently underrepresented in our area. Many thanks, Mark --- Research Engineer (Research Fellow) in Sound Sensing Location: University of Surrey, Guildford, UK Closing Date: Monday 14 March 2022 (23:59 GMT) Applications are invited for a 2.5-year (30-month) Research Engineer (Research Fellow) in Sound Sensing, to work full-time on an EPSRC-funded Fellowship project ?AI for Sound? (https://ai4s.surrey.ac.uk/), to start on 1 April 2022 or as soon as possible thereafter. The aim of the project is to undertake research in computational analysis of everyday sounds, in the context of a set of real-world use cases in assisted living in the home, smart buildings, smart cities, and the creative sector. The postholder will be responsible for designing and building the hardware and software to be developed in the fellowship, including sound sensor systems, open-source software libraries and datasets to be released from the project. The postholder will be based in the Centre for Vision, Speech and Signal Processing (CVSSP) and work under the direction of PI (EPSRC Fellow) Prof Mark Plumbley. The successful applicant is expected to have a postgraduate qualification in electronic engineering, computer science or a related subject, or equivalent professional experience; experience in software and hardware development relevant to signal processing or sensor devices, and experience in software development in topics such as audio signal processing, machine learning, deep learning, and/or sensor systems. Experience in development and deployment of hardware sensors, Internet-of-Things (IoT) devices, or audio systems; and programming experience using Python, C++, MATLAB, or other tools for signal processing, machine learning or deep learning is desirable. Direct research experience, or experience of hardware or software development while working closely with researchers, is also desirable. CVSSP is an International Centre of Excellence for research in Audio-Visual Machine Perception, with 180 researchers, a grant portfolio of ?26M (?17.5M EPSRC), and a turnover of ?7M/annum. The Centre has state-of-the-art acoustic capture and analysis facilities and a Visual Media Lab with video and audio capture facilities supporting research in real-time video and audio processing and visualisation. CVSSP has a compute facility with 120 GPUs for deep learning and >1PB of high-speed secure storage. The University is located in Guildford, a picturesque market town with excellent schools and amenities, and set in the beautiful Surrey Hills, an area of Outstanding Natural Beauty. London is just 35 minutes away by train, while both Heathrow and Gatwick airports are readily accessible. For more information about the post and how to apply, please visit: https://jobs.surrey.ac.uk/009722 Deadline: Monday 14 March 2022 (23:59 GMT) For informal inquiries about the position, please contact Prof Mark Plumbley (m.plumbley at surrey.ac.uk). -- Prof Mark D Plumbley Head of School of Computer Science and Electronic Engineering Email: Head-of-School-CSEE at surrey.ac.uk Professor of Signal Processing University of Surrey, Guildford, Surrey, GU2 7XH, UK Email: m.plumbley at surrey.ac.uk PA: Kelly Green Email: k.d.green at surrey.ac.uk From achler at gmail.com Wed Feb 16 15:37:03 2022 From: achler at gmail.com (Tsvi Achler) Date: Wed, 16 Feb 2022 12:37:03 -0800 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Salience is a much more fundamental phenomena within recognition than the spotlight attention type map suggested by Itti et al and Treisman et al 1980 (the cognitive psychology-equivalent reference). It is also integrated into non-spatial modalities and occurs even when the display is too fast to form an attention map in fast-masking experiments eg (Francis & Cho 2008). It occurs from a bottom up (through input interactions) way before there is a chance to select a spatial region focus and is a source of "pop-out". Salience is associated with a signal-to-noise ratio during processing which can be measured by the speed of processing given different inputs. These effects of salience can be measured both in spatial processing and by reaction times and errors in humans given fast stimuli. Salience kicks in immediately while processing information so it is an integral part of processing, not an attention spatial filter after-effect as hypothesized in the old cognitive and not very much updated current neural network literatures. Pop-out and difficulty with similarity (Duncan & Humphreys 1989; Wolfe 2001) which are analogous signal-to-noise effects (Rosenholtz 2001) are observed in non-visual modalities with poor spatial resolution such as olfaction (e.g. Rinberg et al 2006). Salience seems generated ?on-the-fly? as an inseparable part of recognition mechanisms and thus my opinion (and computational findings) are that top-down connections back to inputs are very important for all recognition (not an aftereffect of spatial attention). Here are some references I have accumulated over my studies: Francis, G. & Cho, Y. (2008). Effects of temporal integration on the shape of visual backward masking functions. Journal of Experimental Psychology: Human Perception & Performance, 34, 1116-1128. Treisman, A.M. and G. Gelade, A feature-integration theory of attention. Cogn Psychol, 1980. 12(1): p. 97-136. Macknik S. L., Martinez-Conde S. (2007). The role of feedback in visual masking and visual processing. Advances in Cognitive Psychology, 3, 125?152. Enns, J.T., & Di Lollo, V. (1997). Object substitution: A new form of visual masking in unattended visual locations. Psychological Science, 8, 135-139. Duncan, J. and G. W. Humphreys (1989). "Visual-Search and Stimulus Similarity." Psychological Review 96(3): 433-458. Breitmeyer, B. G., & ??men, H. (2006). Visual Masking: Time Slices Through Conscious and Unconscious Vision. Oxford: Oxford University Press. Bichot, N. P., A. F. Rossi, et al. (2005). "Parallel and serial neural mechanisms for visual search in macaque area V4." Science 308(5721): 529-34. Wolfe, J.M., Asymmetries in visual search: An introduction. Perception & Psychophysics, 2001. 63(3): p. 381-389. Rinberg D, Koulakov A, Gelperin A (2006) Speed accuracy tradeoff in olfaction. Neuron, 51(3), pp.351-358 Rosenholtz R (2001) Search asymmetries? What search asymmetries? Perception & Psychophysics, 63(3), 476-489 P.S. In order not to offend as much (but dont worry I believe every field deserves criticisms) I have put my opinion about the state of the field here after the references. I find the neural network community is stuck with 1950's feedforward neurons and 1980's attention mechanisms and its associated computer science community is stuck using data sets and paradigms that promote feedforward methods but are unrealistic paradigms for real life environments. The computational neuroscience community is also generally bogged down with a large number of parameters but additionally with statistical models (not really connectionist) with predominantly feedforward and lateral inhibition structures. The cognitive community sits on the most interesting data but is also stuck with (either) overparameterized rate models or abstract non-computational models. The cognitive community is more open to feedback back to inputs but trying to publish or get funds by doing something that covers all three communities gets bogged down by sometimes conflicting requirements, nomenclature and politics in each one. In my opinion and experience this is why there is little progress even if there are new ideas. Thus brain science progress suffers a lot because of these separations. -Tsvi On Wed, Feb 16, 2022 at 9:39 AM Brad Wyble wrote: > Hi Balazs, >> > > You wrote: > >> That is a very interesting question and I would love to know more about >> the reconciliation of the two views. From what I understand, saliency in >> cognitive science is dependent on both 1) the scene represented by pixels >> (or other sensors) and 2) the state of mind of the perceiver (focus, goal, >> memory, etc.). Whereas the current paradigm in computer vision seems to me >> that perception is bottom up, the "true" salience of various image parts >> are a function of the image, and the goal is to learn it from examples. >> Furthermore, it seems to me that there is a consensus that salience >> detection is pre-inferential, so it cannot be learned in the classical >> supervised way: to select and label the data to learn salience, one would >> need to have the very faculty that determines salience, leading to a loop. >> >> I'm very cautious on all this since it's far from my main expertise, so >> my aim is to ask for information rather than to state anything with >> certainty. I'm reading all these discussions with a lot of interest, I find >> that this channel has a space between twitter and formal scientific papers. >> >> > Very good point and it's absolutely true that computational approaches to > salience are a shallow version of how humans compute salience. A great > example I like to use is that if you show someone a picture with a Sun in > it, noone looks at the sun, regardless of how salient it is according > Itti-et al. 1998. We incorporate meaning into our assessment of what is > important, and this controls even the very first eye movements in response > to viewing a new visual scene. > > However, my point was that using NN's to compute salience is a very active > area of research with a wide variety of approaches being used, including > more recently the involvement of meaning. Recent work is starting to tease > apart what recent approaches to salience are missing, e.g. > > > https://www.nature.com/articles/s41598-021-97879-z#:~:text=Deep%20saliency%20models%20represent%20the,look%20in%20real%2Dworld%20scenes.&text=We%20found%20that%20all%20three,feature%20weightings%20and%20interaction%20patterns > . > > So while these approaches are still far from getting it right (just like > the rest of AI), I just wanted to highlight that there is a lot of work in > active progress. > > Thanks! > -Brad > > > > > > > > > -- > Brad Wyble > Associate Professor > Psychology Department > Penn State University > > http://wyblelab.com > -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Wed Feb 16 22:45:30 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Wed, 16 Feb 2022 22:45:30 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Dear Tsvi: You wrote: "I believe scientists not seeing eye-to-eye with each other and other members of the community is in no small part due to these terms." I agree. This is a HUGE problem, as the attached figure "Blind Men and an Elephant" indicates. What should this multidisciplinary community do? Senior people do not want to get a PhD in all 6 disciplines in the attached figure: biology, neuroscience, psychology, computer science, electrical engineering, mathematics. Best regards, -John On Tue, Feb 15, 2022 at 10:00 PM Tsvi Achler wrote: > After studying the brain from a multidisciplinary perspective I am well > aware of the difficulties speaking and understanding each other across > disciplines. There are many terms that are defined differently in > different fields... and unfortunately things are not as simple as looking > them up in a dictionary. > > For example the term recurrent connections have different meanings in the > computational neuroscience, neural networks, and cognitive psychology > communities. > In neural networks recurrent means an output used back as an input within > a paradigm of delayed inputs. It is a method of representing time or > sequences. Often recurrent connections in neural networks are confused > with feedback back to the same inputs which are actually never used in > neural networks because it forms an infinite loop and is not possible to > rewind in order to generate an error signal. > In computational neuroscience recurrent connections are used to describe > lateral connections. > In cognitive psychology the term re-entrant connections are used to > describe feedback back to the same inputs. > > I believe in order to truly appreciate "brain-like" ideas, members of this > group need to familiarize themselves with these brain-focused fields. For > example in cognitive psychology there is a rich literature on salience > (which again is a bit different from salience in the neural network > community). Salience is a dynamic process which determines how well a > certain input or input feature is processed. Salience changes in the brain > depending on what other inputs or features are concurrently present or what > the person is instructed to focus on. There is very little appreciation, > integration or implementation of these findings in feedforward networks, > yet salience plays a factor in every recognition decision and modality > including smell and touch. > > Consciousness is a particularly problematic minefield which also adds in > philosophy, metaphysics and subjectivity into the mix. > > Juyang, I think we both agree about the basics: the need for more > realistic real world recognition and to move beyond the rehearsal > limitations of neural networks. I believe scientists not seeing eye-to-eye > with each other and other members of the community is in no small part due > to these terms. > > Sincerely, > -Tsvi > > > > On Tue, Feb 15, 2022 at 9:54 AM Juyang Weng wrote: > >> Dear Tsvi, >> You wrote "A huge part of the problem in any discussion about >> consciousness is there isn't even a clear definition of consciousness". >> Look at the 5 level definition of consciousness: >> https://www.merriam-webster.com/dictionary/consciousness >> >> You wrote: "So consciousness is not necessary or sufficient for complex >> thoughts or behavior." >> I was thinking that way too, until recently. >> I found consciousness IS REQUIRED for even learning basic intelligence. >> To put it in a short way so that people on this list can benefit: >> The motors (as context/actions) in the brain require consciousness in >> order to learn correctly in the physical world. Please read the first >> model about conscious learning: >> J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th >> International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, >> USA, Jan.7-9, 2022. PDF file >> >> . >> >> Best regards, >> -John >> ---- >> From: Tsvi Achler >> To: Iam Palatnik >> Cc: Connectionists >> Subject: Re: Connectionists: Weird beliefs about consciousness >> >> -- >> Juyang (John) Weng >> > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: BlindMenElephant.jpg Type: image/jpeg Size: 424448 bytes Desc: not available URL: From balazskegl at gmail.com Wed Feb 16 12:03:38 2022 From: balazskegl at gmail.com (Balazs Kegl) Date: Wed, 16 Feb 2022 18:03:38 +0100 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: On Wed, Feb 16, 2022 at 2:46 PM Brad Wyble wrote: > Tsvi you wrote: > >> >> For example in cognitive psychology there is a rich literature on >> salience (which again is a bit different from salience in the neural >> network community). Salience is a dynamic process which determines how >> well a certain input or input feature is processed. Salience changes in the >> brain depending on what other inputs or features are concurrently present >> or what the person is instructed to focus on. There is very little >> appreciation, integration or implementation of these findings in neural >> networks, yet salience plays a factor in every recognition decision and >> modality including smell and touch. >> >> > I'm having trouble understanding what you mean by this, since > computational modelling of salience is a major thrust of computer vision. > Itti Koch & Niebur (1998) has been cited 13,000 times and there are > hundreds of papers that have elaborated on this ANN approach to salience > computation in vision. Is this not what you're asking for? If not, what > am I misunderstanding? > That is a very interesting question and I would love to know more about the reconciliation of the two views. From what I understand, saliency in cognitive science is dependent on both 1) the scene represented by pixels (or other sensors) and 2) the state of mind of the perceiver (focus, goal, memory, etc.). Whereas the current paradigm in computer vision seems to me that perception is bottom up, the "true" salience of various image parts are a function of the image, and the goal is to learn it from examples. Furthermore, it seems to me that there is a consensus that salience detection is pre-inferential, so it cannot be learned in the classical supervised way: to select and label the data to learn salience, one would need to have the very faculty that determines salience, leading to a loop. I'm very cautious on all this since it's far from my main expertise, so my aim is to ask for information rather than to state anything with certainty. I'm reading all these discussions with a lot of interest, I find that this channel has a space between twitter and formal scientific papers. Best regards, Balazs -------------- next part -------------- An HTML attachment was scrubbed... URL: From terry at salk.edu Wed Feb 16 19:24:32 2022 From: terry at salk.edu (Terry Sejnowski) Date: Wed, 16 Feb 2022 16:24:32 -0800 Subject: Connectionists: NEURAL COMPUTATION - February 1, 2022 In-Reply-To: Message-ID: Neural Computation - Volume 34, Number 2 - February 1, 2022 Available online for download now: http://www.mitpressjournals.org/toc/neco/34/2 http://cognet.mit.edu/content/neural-computation ----- View Bridging the Gap Between Neurons and Cognition Through Assemblies of Neurons Christos Papadimitriou, Angela Friederici Article A Normative Account of Confirmation Bias During Reinforcement Learning Christopher Summerfield, Germain Lefebvre, and Rafal Bogacz Letters Model-based or Model-free ? Comparing Adaptive Methods for Estimating Sensitivity Thresholds in Neuroscience Julien Audiffren, Jean-Pierre Bresciani Implicit Contact Dynamics Modeling With Explicit Inertia Matrix Representation for Real-Time Model-Based Control in Physical Environment Takeshi Itoh, Koji Ishihara, and Jun Morimoto A Neurodynamic Model of Saliency Prediction in V1 David Berga, Xavier Otazu Distributed Phase Oscillatory Excitation Efficiently Produces Attractors Using Spike Timing Dependent Plasticity Eric Wong Categorical Perception: A Groundwork for Deep Learning Laurent Bonnasse-Gahot, Jean-Pierre Nadal Nonconvex Sparse Regularization for Deep Neural Networks and Its Optimality Ilsang Ohn, Yongdai Kim Convolution-based Model Solving Method for Three-dimensional Unsteady Partial Differential Equations Wenshu Zha, Wen Zhang, Daolun Li, Yan Xing, Lei He, and Jieqing Tan ----- ON-LINE -- http://www.mitpressjournals.org/neco MIT Press Journals, One Rogers Street, Cambridge, MA 02142-1209 Tel: (617) 253-2889 FAX: (617) 577-1545 journals-cs at mit.edu ----- From bwyble at gmail.com Wed Feb 16 12:38:29 2022 From: bwyble at gmail.com (Brad Wyble) Date: Wed, 16 Feb 2022 12:38:29 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: > > Hi Balazs, > You wrote: > That is a very interesting question and I would love to know more about > the reconciliation of the two views. From what I understand, saliency in > cognitive science is dependent on both 1) the scene represented by pixels > (or other sensors) and 2) the state of mind of the perceiver (focus, goal, > memory, etc.). Whereas the current paradigm in computer vision seems to me > that perception is bottom up, the "true" salience of various image parts > are a function of the image, and the goal is to learn it from examples. > Furthermore, it seems to me that there is a consensus that salience > detection is pre-inferential, so it cannot be learned in the classical > supervised way: to select and label the data to learn salience, one would > need to have the very faculty that determines salience, leading to a loop. > > I'm very cautious on all this since it's far from my main expertise, so my > aim is to ask for information rather than to state anything with certainty. > I'm reading all these discussions with a lot of interest, I find that this > channel has a space between twitter and formal scientific papers. > > Very good point and it's absolutely true that computational approaches to salience are a shallow version of how humans compute salience. A great example I like to use is that if you show someone a picture with a Sun in it, noone looks at the sun, regardless of how salient it is according Itti-et al. 1998. We incorporate meaning into our assessment of what is important, and this controls even the very first eye movements in response to viewing a new visual scene. However, my point was that using NN's to compute salience is a very active area of research with a wide variety of approaches being used, including more recently the involvement of meaning. Recent work is starting to tease apart what recent approaches to salience are missing, e.g. https://www.nature.com/articles/s41598-021-97879-z#:~:text=Deep%20saliency%20models%20represent%20the,look%20in%20real%2Dworld%20scenes.&text=We%20found%20that%20all%20three,feature%20weightings%20and%20interaction%20patterns . So while these approaches are still far from getting it right (just like the rest of AI), I just wanted to highlight that there is a lot of work in active progress. Thanks! -Brad -- Brad Wyble Associate Professor Psychology Department Penn State University http://wyblelab.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From giacomo.cabri at unimore.it Wed Feb 16 13:49:47 2022 From: giacomo.cabri at unimore.it (Giacomo Cabri) Date: Wed, 16 Feb 2022 19:49:47 +0100 Subject: Connectionists: CfP - 4th International Workshop on Key Enabling Technologies for Digital Factories (KET4DF) @ CAiSE 2022 Message-ID: <284a9540-d431-969b-e92a-0eacccf5613f@unimore.it> [ apologies for potential cross-postings ] ============================================= ???????? Call For Papers ?? 4th International Workshop on ?Key Enabling Technologies for Digital Factories ???? in conjunction with CAiSE 2022 ???? ???? Leuven, Belgium https://sites.google.com/view/ket4df2022/ ============================================= Paper submission deadline March 8, 2022, 11:59pm Hawaii Time ** IMPORTANT NEWS ** Selected papers of the workshop will be invited to provide extended versions for submissions to a Special Issue of the"Sensors" journal https://www.mdpi.com/journal/sensors/about. More details are available on the Workshop Web site. Scope --------------------- The manufacturing industry is entering a new digital era in which advanced information systems and especially (big) data-driven methods and Artificial Intelligence techniques allow companies to move beyond distributed and supervisory control systems to support significant operational improvements and allow dynamic adaptability. Moreover, human-centered trustworthy AI systems provide the ability to augment human work and extend human capabilities in order to solve problems and achieve goals that were unreachable by either humans or machines alone. Such efforts may allow the manufacturing industry in the post-COVID to "go back to normal? and? optimize, scale and ensure their processes and services towards a circular and resilient economy. This workshop seeks at providing the opportunity for inspiration and cross-fertilization for the research groups working on technological solutions for digital factories and smart manufacturing. We welcome innovative papers from academic and industrial researchers covering a wide range of topics of interests in the computer science and computer engineering fields. Topics of interest --------------------- The topics include but are not limited to: * Advanced Information Systems for Smart Manufacturing * Information Systems Engineering for Production Management * Big Data Technologies and Analytics for Smart Manufacturing * Artificial Intelligence and Machine Learning methods for production systems * AI and digital twin applications in production and operations management * Design issues for multimodal human-AI interactions in the industrial environment * Explainable and transparent AI in manufacturing * AI approaches to support and leverage circular economy production models * AI approaches to predict and minimize the effects of unpredictable events * Digital intelligent assistants, software robots (softbots) and chatbots * Applications of Human-Centered AI systems in manufacturing Important Dates --------------------- Paper submission: 8 March 2022 Acceptance notification: 8 April 2022 Workshop: 6 or 7 June 2022 --------------------- Workshop Co-chairs: --------------------- Giacomo Cabri Universit? di Modena e Reggio Emilia, Italy Federica Mandreoli Universit? di Modena e Reggio Emilia, Italy Gregoris Mentzas National Technical University of Athens, Greece Karl Hribernik BIBA - Bremer Institut f?r Produktion und Logistik GmbH, Germany --------------------- Workshop web site: https://sites.google.com/view/ket4df2022/ --------------------- --------------------- -- |----------------------------------------------------| | Prof. Giacomo Cabri - Ph.D., Full Professor | Rector's Delegate for Teaching | Dip. di Scienze Fisiche, Informatiche e Matematiche | Universita' di Modena e Reggio Emilia - Italia | e-mail giacomo.cabri at unimore.it | tel. +39-059-2058320 fax +39-059-2055216 |----------------------------------------------------| From juyang.weng at gmail.com Wed Feb 16 22:53:22 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Wed, 16 Feb 2022 22:53:22 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Brad, You wrote "Itti Koch & Niebur (1998) ... salience computation in vision" With due respect, I think salience is a very small piece of attention. Maybe 10%? Based on studies in six disciplines, attention is mainly top-down, guided by self-generated intent. That is how our DN models a general framework of attention, represented in the state/action that the network self-generated. In this model, salience is about 10%, when state/action is "none". Sorry, the meth is a lot more complex. I am trying to communicate in an intuitive way. Best regards, -John On Wed, Feb 16, 2022 at 8:31 AM Brad Wyble wrote: > Tsvi you wrote: > >> >> For example in cognitive psychology there is a rich literature on >> salience (which again is a bit different from salience in the neural >> network community). Salience is a dynamic process which determines how >> well a certain input or input feature is processed. Salience changes in the >> brain depending on what other inputs or features are concurrently present >> or what the person is instructed to focus on. There is very little >> appreciation, integration or implementation of these findings in neural >> networks, yet salience plays a factor in every recognition decision and >> modality including smell and touch. >> >> > I'm having trouble understanding what you mean by this, since > computational modelling of salience is a major thrust of computer vision. > Itti Koch & Niebur (1998) has been cited 13,000 times and there are > hundreds of papers that have elaborated on this ANN approach to salience > computation in vision. Is this not what you're asking for? If not, what > am I misunderstanding? > > kind regards > -Brad > > > > > > > > > > > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Thu Feb 17 03:04:41 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Thu, 17 Feb 2022 09:04:41 +0100 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Hi all, There are even more cool and relevant facts about attention than what Tsvi listed. The capability to effectively apply attention is tightly correlated to our knowledge about what we are attending to and to the capacity of working memory; there is a linear relationship between the capacity of working memory for a certain type of inputs and the ability to effectively attend to those inputs. For example, non Chinese readers have poor attention capacity and poor working memory for Chinese characters. Chinese readers, in contrast, have great capacities for both. The part of the brain that drives attention to X is also the part of the brain that stores X into working memory. A tight relationship indeed. The capability to attend also determines the time it takes to learn a visual pattern -- which also seems to be a linear relationship. Attention seems to require concepts. If you have a concept of A, you can effectively attend to A. Attention may not be a separate mechanism but a shadow of something else in the brain that also does working memory and other stuff like decision making and intelligence in general. We don't know what that is. (I have given my two cents on the properties of this mechanism behind, which I named anapoiesis.) Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Feb 17, 2022 at 8:24 AM Tsvi Achler wrote: > > Salience is a much more fundamental phenomena within recognition than the > spotlight attention type map suggested by Itti et al and Treisman et al > 1980 (the cognitive psychology-equivalent reference). > It is also integrated into non-spatial modalities and occurs even when the > display is too fast to form an attention map in fast-masking experiments eg > (Francis & Cho 2008). > It occurs from a bottom up (through input interactions) way before there > is a chance to select a spatial region focus and is a source of "pop-out". > Salience is associated with a signal-to-noise ratio during processing which > can be measured by the speed of processing given different inputs. > These effects of salience can be measured both in spatial processing and > by reaction times and errors in humans given fast stimuli. Salience kicks > in immediately while processing information so it is an integral part of > processing, not an attention spatial filter after-effect as hypothesized in > the old cognitive and not very much updated current neural network > literatures. > > Pop-out and difficulty with similarity (Duncan & Humphreys 1989; Wolfe > 2001) which are analogous signal-to-noise effects (Rosenholtz 2001) are > observed in non-visual modalities with poor spatial resolution such as > olfaction (e.g. Rinberg et al 2006). > Salience seems generated ?on-the-fly? as an inseparable part of > recognition mechanisms and thus my opinion (and computational findings) are > that top-down connections back to inputs are very important for all > recognition (not an aftereffect of spatial attention). > > Here are some references I have accumulated over my studies: > Francis, G. & Cho, Y. (2008). Effects of temporal integration on the shape > of visual backward masking functions. Journal of Experimental Psychology: > Human Perception & Performance, 34, 1116-1128. > Treisman, A.M. and G. Gelade, A feature-integration theory of attention. > Cogn Psychol, 1980. 12(1): p. 97-136. > Macknik S. L., Martinez-Conde S. (2007). The role of feedback in visual > masking and visual processing. Advances in Cognitive Psychology, 3, 125?152. > Enns, J.T., & Di Lollo, V. (1997). Object substitution: A new form of > visual masking in unattended visual locations. Psychological Science, 8, > 135-139. > Duncan, J. and G. W. Humphreys (1989). "Visual-Search and Stimulus > Similarity." Psychological Review 96(3): 433-458. > Breitmeyer, B. G., & ??men, H. (2006). Visual Masking: Time Slices Through > Conscious and Unconscious Vision. Oxford: Oxford University Press. > Bichot, N. P., A. F. Rossi, et al. (2005). "Parallel and serial neural > mechanisms for visual search in macaque area V4." Science 308(5721): 529-34. > Wolfe, J.M., Asymmetries in visual search: An introduction. Perception & > Psychophysics, 2001. 63(3): p. 381-389. > Rinberg D, Koulakov A, Gelperin A (2006) Speed accuracy tradeoff in > olfaction. Neuron, 51(3), pp.351-358 > Rosenholtz R (2001) Search asymmetries? What search asymmetries? > Perception & Psychophysics, 63(3), 476-489 > > P.S. In order not to offend as much (but dont worry I believe every field > deserves criticisms) I have put my opinion about the state of the field > here after the references. > I find the neural network community is stuck with 1950's feedforward > neurons and 1980's attention mechanisms and its associated computer science > community is stuck using data sets and paradigms that promote feedforward > methods but are unrealistic paradigms for real life environments. > The computational neuroscience community is also generally bogged down > with a large number of parameters but additionally with statistical models > (not really connectionist) with predominantly feedforward and lateral > inhibition structures. > The cognitive community sits on the most interesting data but is also > stuck with (either) overparameterized rate models or abstract > non-computational models. > The cognitive community is more open to feedback back to inputs but trying > to publish or get funds by doing something that covers all three > communities gets bogged down by sometimes conflicting requirements, > nomenclature and politics in each one. In my opinion and experience this > is why there is little progress even if there are new ideas. > Thus brain science progress suffers a lot because of these separations. > -Tsvi > > On Wed, Feb 16, 2022 at 9:39 AM Brad Wyble wrote: > >> Hi Balazs, >>> >> >> You wrote: >> >>> That is a very interesting question and I would love to know more about >>> the reconciliation of the two views. From what I understand, saliency in >>> cognitive science is dependent on both 1) the scene represented by pixels >>> (or other sensors) and 2) the state of mind of the perceiver (focus, goal, >>> memory, etc.). Whereas the current paradigm in computer vision seems to me >>> that perception is bottom up, the "true" salience of various image parts >>> are a function of the image, and the goal is to learn it from examples. >>> Furthermore, it seems to me that there is a consensus that salience >>> detection is pre-inferential, so it cannot be learned in the classical >>> supervised way: to select and label the data to learn salience, one would >>> need to have the very faculty that determines salience, leading to a loop. >>> >>> I'm very cautious on all this since it's far from my main expertise, so >>> my aim is to ask for information rather than to state anything with >>> certainty. I'm reading all these discussions with a lot of interest, I find >>> that this channel has a space between twitter and formal scientific papers. >>> >>> >> Very good point and it's absolutely true that computational approaches to >> salience are a shallow version of how humans compute salience. A great >> example I like to use is that if you show someone a picture with a Sun in >> it, noone looks at the sun, regardless of how salient it is according >> Itti-et al. 1998. We incorporate meaning into our assessment of what is >> important, and this controls even the very first eye movements in response >> to viewing a new visual scene. >> >> However, my point was that using NN's to compute salience is a very >> active area of research with a wide variety of approaches being used, >> including more recently the involvement of meaning. Recent work is >> starting to tease apart what recent approaches to salience are missing, e.g. >> >> >> https://www.nature.com/articles/s41598-021-97879-z#:~:text=Deep%20saliency%20models%20represent%20the,look%20in%20real%2Dworld%20scenes.&text=We%20found%20that%20all%20three,feature%20weightings%20and%20interaction%20patterns >> . >> >> So while these approaches are still far from getting it right (just like >> the rest of AI), I just wanted to highlight that there is a lot of work in >> active progress. >> >> Thanks! >> -Brad >> >> >> >> >> >> >> >> >> -- >> Brad Wyble >> Associate Professor >> Psychology Department >> Penn State University >> >> http://wyblelab.com >> > -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Thu Feb 17 03:15:43 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Thu, 17 Feb 2022 09:15:43 +0100 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Dear Juyang, You wrote "Senior people do not want to get a PhD in all 6 disciplines in the attached figure: biology, neuroscience, psychology, computer science, electrical engineering, mathematics." I would cross electrical engineering from that list. It seems to me that the contribution of electrical engineering is minor. But then I would add philosophy of mind and cybernetics. These two seem a lot more important to acquire a PhD-level knowledge in. Best, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Feb 17, 2022 at 8:22 AM Juyang Weng wrote: > Dear Tsvi: > > You wrote: "I believe scientists not seeing eye-to-eye with each other and > other members of the community is in no small part due to these terms." > > I agree. This is a HUGE problem, as the attached figure "Blind Men and an > Elephant" indicates. What should this multidisciplinary community do? > Senior people do not want to get a PhD in all 6 disciplines in the > attached figure: biology, neuroscience, psychology, computer science, > electrical engineering, mathematics. > > Best regards, > -John > > On Tue, Feb 15, 2022 at 10:00 PM Tsvi Achler wrote: > >> After studying the brain from a multidisciplinary perspective I am well >> aware of the difficulties speaking and understanding each other across >> disciplines. There are many terms that are defined differently in >> different fields... and unfortunately things are not as simple as looking >> them up in a dictionary. >> >> For example the term recurrent connections have different meanings in the >> computational neuroscience, neural networks, and cognitive psychology >> communities. >> In neural networks recurrent means an output used back as an input within >> a paradigm of delayed inputs. It is a method of representing time or >> sequences. Often recurrent connections in neural networks are confused >> with feedback back to the same inputs which are actually never used in >> neural networks because it forms an infinite loop and is not possible to >> rewind in order to generate an error signal. >> In computational neuroscience recurrent connections are used to describe >> lateral connections. >> In cognitive psychology the term re-entrant connections are used to >> describe feedback back to the same inputs. >> >> I believe in order to truly appreciate "brain-like" ideas, members of >> this group need to familiarize themselves with these brain-focused fields. >> For example in cognitive psychology there is a rich literature on salience >> (which again is a bit different from salience in the neural network >> community). Salience is a dynamic process which determines how well a >> certain input or input feature is processed. Salience changes in the brain >> depending on what other inputs or features are concurrently present or what >> the person is instructed to focus on. There is very little appreciation, >> integration or implementation of these findings in feedforward networks, >> yet salience plays a factor in every recognition decision and modality >> including smell and touch. >> >> Consciousness is a particularly problematic minefield which also adds in >> philosophy, metaphysics and subjectivity into the mix. >> >> Juyang, I think we both agree about the basics: the need for more >> realistic real world recognition and to move beyond the rehearsal >> limitations of neural networks. I believe scientists not seeing eye-to-eye >> with each other and other members of the community is in no small part due >> to these terms. >> >> Sincerely, >> -Tsvi >> >> >> >> On Tue, Feb 15, 2022 at 9:54 AM Juyang Weng >> wrote: >> >>> Dear Tsvi, >>> You wrote "A huge part of the problem in any discussion about >>> consciousness is there isn't even a clear definition of consciousness". >>> Look at the 5 level definition of consciousness: >>> https://www.merriam-webster.com/dictionary/consciousness >>> >>> You wrote: "So consciousness is not necessary or sufficient for complex >>> thoughts or behavior." >>> I was thinking that way too, until recently. >>> I found consciousness IS REQUIRED for even learning basic intelligence. >>> To put it in a short way so that people on this list can benefit: >>> The motors (as context/actions) in the brain require consciousness in >>> order to learn correctly in the physical world. Please read the first >>> model about conscious learning: >>> J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th >>> International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, >>> USA, Jan.7-9, 2022. PDF file >>> >>> . >>> >>> Best regards, >>> -John >>> ---- >>> From: Tsvi Achler >>> To: Iam Palatnik >>> Cc: Connectionists >>> Subject: Re: Connectionists: Weird beliefs about consciousness >>> >>> -- >>> Juyang (John) Weng >>> >> > > -- > Juyang (John) Weng > -------------- next part -------------- An HTML attachment was scrubbed... URL: From georgios.yannakakis at um.edu.mt Thu Feb 17 03:13:20 2022 From: georgios.yannakakis at um.edu.mt (Georgios N Yannakakis) Date: Thu, 17 Feb 2022 09:13:20 +0100 Subject: Connectionists: AI and Games Summer School - Registration Open! Chania, Greece, 29/8-2/9 [Hybrid] In-Reply-To: References: Message-ID: We are happy to announce that the 4th edition of the AI and Games Summer School returns to *Chania, Greece *and will will take place *from 29 August to 2 September, 2022*! https://school.gameaibook.org/ Our event last year attracted over 230 participants who had fun learning about the various uses of AI in games and implementing their own cool projects! To be able to support participants who have challenges travelling but also deliver on a full summer school experience, this year we are going for a hybrid format!* Both online and in-person tickets are available!* The summer school is dedicated to the uses of AI in and for games including AI for playing and testing games, for generating game content, and for modeling players. Our program already features talks and workshops from* Microsoft, Meta AI, EA, Askblu.ai, Modl.ai *with more joining soon. Graduate students, research staff and faculty, and industry professionals are welcome to register by *March 31* (early bird) by visiting the summer school page! *https://school.gameaibook.org/#registration * If you have any questions please don't hesitate to contact us at school at gameaibook.org! We hope to see you all in August in sunny Chania, Crete! Best regards, Summer School on AI and Games Organisers -------------- next part -------------- An HTML attachment was scrubbed... URL: From achler at gmail.com Thu Feb 17 05:25:30 2022 From: achler at gmail.com (Tsvi Achler) Date: Thu, 17 Feb 2022 02:25:30 -0800 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Hi Danko, Indeed there are so many top down effects based on what has been learned, which is further evidence that salience is inseparable from recognition. However as a note about education, I dont buy into defining especially STEM fields that need to be studied. If what is needed to be known is known, it wouldn't be science. My undergraduate degree is in (ahem) Electrical Engineering Computer Science. The electrical engineering provided useful tools through control theory to understand and manage the dynamics of top-down feedback. This includes the mathematical tools to determine stability and steady state endpoints of dynamical systems. In the recognition model I developed, regulatory feedback networks, each input is regulated by the outputs it activates which subsequently creates a dynamical salience which makes sure inputs are not over or under represented within the network and lets the system evaluate all the inputs together. This model inherently displays the signal-to-noise salience phenomena that includes difficulty with similarity and asymmetry seen in humans even if the inputs are not spatial. It shows Excitation-Inhibition balance findings in neuroscience (Xue et al 2014) or what I call network-wide-bursting. It also does not require the iid rehearsal to learn like feedforward models because it uses the salience-feedback dynamics to determine the best relevance of each input given what has been learned and what is present at recognition time. This dynamic salience is inseparable from the mechanism of recognition. Unfortunately I dont seem to be able to convince the academic cognitive psychology and neuroscience communities that while they can model any of their phenomena with enough parameters, models with huge parameter spaces are less desirable. Scalability and minimizing degrees of freedom are important in models: models that show the most amount of phenomena (across multiple disciplines) with the least amount of free parameters are better. Neither do I seem able to convince the academic connectionist community that the iid rehearsal requirements are killing the ability to use models in a natural way and can be avoided using feedback back to the inputs during recognition. Nor does there seem to be cross interest within the academic computational communities where this model can show top-down attention effects and controls similar to Bayesian models (with dynamic priors, priming, and weights similar to likelihoods helping with explainability) but is completely connectionist. Sincerely, -Tsvi Xue, M, Atallah, BV, Scanziani, M. (2014). Equalizing excitation-inhibition ratios across visual cortical neurons. Nature 511, 596?600 https://www.youtube.com/watch?v=F-GBIZoZ1mI&list=PL4nMP8F3B7bg3cNWWwLG8BX-wER2PeB-3&index=1 On Thu, Feb 17, 2022 at 12:15 AM Danko Nikolic wrote: > > Dear Juyang, > > You wrote "Senior people do not want to get a PhD in all 6 disciplines in > the attached figure: biology, neuroscience, psychology, computer science, > electrical engineering, mathematics." > > I would cross electrical engineering from that list. It seems to me that > the contribution of electrical engineering is minor. But then I would add > philosophy of mind and cybernetics. These two seem a lot more important > to acquire a PhD-level knowledge in. > > Best, > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > --- A progress usually starts with an insight --- > > > On Thu, Feb 17, 2022 at 8:22 AM Juyang Weng wrote: > >> Dear Tsvi: >> >> You wrote: "I believe scientists not seeing eye-to-eye with each other >> and other members of the community is in no small part due to these terms." >> >> I agree. This is a HUGE problem, as the attached figure "Blind Men and >> an Elephant" indicates. What should this multidisciplinary community >> do? Senior people do not want to get a PhD in all 6 disciplines in the >> attached figure: biology, neuroscience, psychology, computer science, >> electrical engineering, mathematics. >> >> Best regards, >> -John >> >> On Tue, Feb 15, 2022 at 10:00 PM Tsvi Achler wrote: >> >>> After studying the brain from a multidisciplinary perspective I am well >>> aware of the difficulties speaking and understanding each other across >>> disciplines. There are many terms that are defined differently in >>> different fields... and unfortunately things are not as simple as looking >>> them up in a dictionary. >>> >>> For example the term recurrent connections have different meanings in >>> the computational neuroscience, neural networks, and cognitive psychology >>> communities. >>> In neural networks recurrent means an output used back as an input >>> within a paradigm of delayed inputs. It is a method of representing time >>> or sequences. Often recurrent connections in neural networks are confused >>> with feedback back to the same inputs which are actually never used in >>> neural networks because it forms an infinite loop and is not possible to >>> rewind in order to generate an error signal. >>> In computational neuroscience recurrent connections are used to describe >>> lateral connections. >>> In cognitive psychology the term re-entrant connections are used to >>> describe feedback back to the same inputs. >>> >>> I believe in order to truly appreciate "brain-like" ideas, members of >>> this group need to familiarize themselves with these brain-focused fields. >>> For example in cognitive psychology there is a rich literature on salience >>> (which again is a bit different from salience in the neural network >>> community). Salience is a dynamic process which determines how well a >>> certain input or input feature is processed. Salience changes in the brain >>> depending on what other inputs or features are concurrently present or what >>> the person is instructed to focus on. There is very little appreciation, >>> integration or implementation of these findings in feedforward networks, >>> yet salience plays a factor in every recognition decision and modality >>> including smell and touch. >>> >>> Consciousness is a particularly problematic minefield which also adds in >>> philosophy, metaphysics and subjectivity into the mix. >>> >>> Juyang, I think we both agree about the basics: the need for more >>> realistic real world recognition and to move beyond the rehearsal >>> limitations of neural networks. I believe scientists not seeing eye-to-eye >>> with each other and other members of the community is in no small part due >>> to these terms. >>> >>> Sincerely, >>> -Tsvi >>> >>> >>> >>> On Tue, Feb 15, 2022 at 9:54 AM Juyang Weng >>> wrote: >>> >>>> Dear Tsvi, >>>> You wrote "A huge part of the problem in any discussion about >>>> consciousness is there isn't even a clear definition of consciousness". >>>> Look at the 5 level definition of consciousness: >>>> https://www.merriam-webster.com/dictionary/consciousness >>>> >>>> You wrote: "So consciousness is not necessary or sufficient for complex >>>> thoughts or behavior." >>>> I was thinking that way too, until recently. >>>> I found consciousness IS REQUIRED for even learning basic intelligence. >>>> To put it in a short way so that people on this list can benefit: >>>> The motors (as context/actions) in the brain require consciousness in >>>> order to learn correctly in the physical world. Please read the first >>>> model about conscious learning: >>>> J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th >>>> International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, >>>> USA, Jan.7-9, 2022. PDF file >>>> >>>> . >>>> >>>> Best regards, >>>> -John >>>> ---- >>>> From: Tsvi Achler >>>> To: Iam Palatnik >>>> Cc: Connectionists >>>> Subject: Re: Connectionists: Weird beliefs about consciousness >>>> >>>> -- >>>> Juyang (John) Weng >>>> >>> >> >> -- >> Juyang (John) Weng >> > -------------- next part -------------- An HTML attachment was scrubbed... URL: From hans.ekkehard.plesser at nmbu.no Thu Feb 17 05:49:23 2022 From: hans.ekkehard.plesser at nmbu.no (Hans Ekkehard Plesser) Date: Thu, 17 Feb 2022 10:49:23 +0000 Subject: Connectionists: NEST Conference 2022 - registration is open Message-ID: <15B0ABCF-4A9C-4B3E-A787-B77C3C13AEF7@nmbu.no> Dear Colleagues, The NEST Initiative is excited to invite everyone interested in Neural Simulation Technology and the NEST Simulator to the NEST Conference 2022. The NEST Conference provides an opportunity for the NEST Community to meet, exchange success stories, swap advice, learn about current developments in and around NEST spiking network simulation and its application. We particularly encourage young scientists to participate in the conference! This year's conference will again take place as a virtual event on Thursday/Friday 23/24 June 2022. Register now! For more information please visit the conference website https://nest-simulator.org/conference We are looking forward to seeing you all in June! Hans Ekkehard Plesser and colleagues -- Prof. Dr. Hans Ekkehard Plesser Head, Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway Phone +47 6723 1560 Email hans.ekkehard.plesser at nmbu.no Home http://arken.nmbu.no/~plesser -------------- next part -------------- An HTML attachment was scrubbed... URL: From francisco.pereira at gmail.com Thu Feb 17 08:36:22 2022 From: francisco.pereira at gmail.com (Francisco Pereira) Date: Thu, 17 Feb 2022 08:36:22 -0500 Subject: Connectionists: job: staff scientist/team leader at NIMH Message-ID: The National Institute of Mental Health (NIMH), National Institutes of Health (NIH), Department of Health and Human Services (DHHS), is recruiting for a Staff Scientist/Team Leader position to direct the Scientific and Statistical Computing Core/AFNI Team (https://afni.nimh.nih.gov).The primary function of this group is to support multimodal neuroimaging research in the intramural research program (IRP) at NIH. This includes (i) applying, extending, and developing new and existing software tools and techniques to conduct pioneering analyses on a diverse array of neuroimaging modalities (e.g. MRI, DTI, fMRI, MEG, PET, and EEG), (ii) advising and training IRP researchers on neuroimaging analysis methods, and (iii) leading the future growth and development of the AFNI code base ( https://github.com/afni/afni). Primary responsibilities of the Team Leader will include serving as a partner in and facilitator of neuroimaging research within the NIMH IRP; providing guidance on study design; and ensuring that appropriate standards of quality, rigor, and transparency are maintained. They will also act in coordination with the leaders of the Data Science and Sharing ( https://cmn.nimh.nih.gov/dsst) and Machine Learning Teams ( https://cmn.nimh.nih.gov/mlt). Finally, they will be responsible for management and supervision of approximately 9 full time staff members including software engineers, statisticians, and other biomedical science professionals. The successful candidate must be committed to scientific excellence and highly collaborative research, ensuring an immense impact on the fields of brain imaging and mental health research. Candidates must have earned a Ph.D. or M.D. and demonstrate experience as an outstanding scientist. Applicants must possess the ability to analyze, prioritize, and delegate resources while managing multiple projects. Expertise and experience should include demonstration of positive and collegial interactions with staff at all levels and the ability to communicate and promote their research, software, and scholarly contributions. Applicants should also have knowledge and experience in software engineering, including writing, distributing, maintaining, and supporting scientific software for a large user base while adhering to community recognized standards and practices. Please see the official posting https://cmn.nimh.nih.gov/compcore-job for more details, including instructions on how to apply. -------------- next part -------------- An HTML attachment was scrubbed... URL: From constantin.rothkopf at cogsci.tu-darmstadt.de Thu Feb 17 09:03:42 2022 From: constantin.rothkopf at cogsci.tu-darmstadt.de (Constantin Rothkopf) Date: Thu, 17 Feb 2022 15:03:42 +0100 Subject: Connectionists: KogWis2022 | 15th Biannual Conference of the German Cognitive Science Society +++ The call for symposia is open Message-ID: Event: The 15th Biannual Conference of the German Cognitive Science Society, KogWis2022 Theme: Understanding Minds Date: 5.-7. September 2022 Location: Freiburg im Breisgau, Germany Host: Albert-Ludwigs University Freiburg Web: https://www.kogwis2022.uni-freiburg.de/ Abstract: The 15. biannual Conference of the German Cognitive Science Society (KogWis2022) will take place from 5.-7. September 2022 in Freiburg i.Br., Germany, organized by the Center for Cognitive Science. The theme ?Understanding Minds? reflects two perspectives: The conference provides a forum for all topics in the study of how minds ? both human and artificial ? operate.? The theme also puts a specific spotlight on how cognitive systems make sense of the world, in particular in language comprehension and communication. The Call for Symposia for the 15th Biannual Conference of the German Cognitive Science Society, KogWis2022, is now open! https://www.kogwis2022.uni-freiburg.de/ Confirmed keynote speakers are: Dedre Gentner, Northwestern University, Illinois Seana Coulson, University of California, San Diego Marcel Brass, Humboldt University, Berlin Matthew Crocker, Saarland University, Saarbr?cken For any inquiries please use the conference e-mail-address kogwis2022 at cognition.uni-freiburg.de. Further information is available on the conference website: www.kogwis2022.uni-freiburg.de. Organising Committee: Evelyn Ferstl, Lars Konieczny, Rul von St?lpnagel, Lisa Zacharski, Judith Beck From hava.siegelmann at gmail.com Thu Feb 17 09:54:00 2022 From: hava.siegelmann at gmail.com (Hava Siegelmann) Date: Thu, 17 Feb 2022 09:54:00 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Message-ID: I agree with Daniel. I would like to add that before consciousness is well defined it is meaningless to discuss whether it exists in artificial systems. We r scientists and require clarity. On Tue, Feb 15, 2022 at 2:38 AM Daniel Polani wrote: > There are quite a few researchers spending a lot of effort trying to > understand the origins of consciousness and to understand whether and how > the subjective experience of consciousness can be captured in a descriptive > and ideally mathematical manner. Tononi, Albantakis, Seth, O'Regan, just to > name a few; one does not have to agree with them, but this question has > been given a lot of attention and it's worth having a look before > discussing it in a vacuum. Also worth reading, amongst other, Dennett and > Chalmers (just as side remark: some of you may remember the the latter as > he had actually a nice Evolutionary Algorithm experiment in the 90s showing > how the Widrow-Hoff rule emerged as "optimal" learning rule in a > neural-type learning scenario). > > The issue about consciousness being an exclusively human ability (as is > often insinuated) is probably not anymore seriously discussed; it is pretty > clear that even self-awareness extends significantly beyond humans, not > even mentioning the subjective experience which does away with the > requirement of self-reflection. It is certainly far safer to estimate that > it will be a matter of degree of consciousness in the animal kingdom than > to claim that it is either present or not. It seems that even our > understanding of elementary experiences must be redefined as e.g. lobsters > actually may feel pain. > > Thus we should be careful making sweeping statements about the presence of > consciousness in the biological realm. It is indeed a very interesting > question to understand to which extent (if at all) an artificial system can > experience that, too; what if the artificial system is a specifically > designed, but growing biological neuron culture on an agar plate? If the > response is yes for the latter, but no for the former, what is the core > difference? Is it the recurrence that matters? The embodiment? Some aspect > of its biological makeup? Something else? > > I do not think we have good answers for this at this stage, but only some > vague hints. > > > > > > On Mon, Feb 14, 2022 at 4:22 PM Iam Palatnik wrote: > >> A somewhat related question, just out of curiosity. >> >> Imagine the following: >> >> - An automatic solar panel that tracks the position of the sun. >> - A group of single celled microbes with phototaxis that follow the >> sunlight. >> - A jellyfish (animal without a brain) that follows/avoids the sunlight. >> - A cockroach (animal with a brain) that avoids the sunlight. >> - A drone with onboard AI that flies to regions of more intense sunlight >> to recharge its batteries. >> - A human that dislikes sunlight and actively avoids it. >> >> Can any of these, beside the human, be said to be aware or conscious of >> the sunlight, and why? >> What is most relevant? Being a biological life form, having a brain, >> being able to make decisions based on the environment? Being taxonomically >> close to humans? >> >> >> >> >> >> >> >> On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: >> >>> Also true: Many AI researchers are very unclear about what consciousness >>> is and also very sure that ELIZA doesn?t have it. >>> >>> Neither ELIZA nor GPT-3 have >>> - anything remotely related to embodiment >>> - any capacity to reflect upon themselves >>> >>> Hypothesis: neither keyword matching nor tensor manipulation, even at >>> scale, suffice in themselves to qualify for consciousness. >>> >>> - Gary >>> >>> > On Feb 14, 2022, at 00:24, Geoffrey Hinton >>> wrote: >>> > >>> > ?Many AI researchers are very unclear about what consciousness is and >>> also very sure that GPT-3 doesn?t have it. It?s a strange combination. >>> > >>> > >>> >>> -------------- next part -------------- An HTML attachment was scrubbed... URL: From maanakg at gmail.com Thu Feb 17 21:51:53 2022 From: maanakg at gmail.com (Maanak Gupta) Date: Thu, 17 Feb 2022 20:51:53 -0600 Subject: Connectionists: DEADLINE EXTENSION: Call for Papers: 27th ACM Symposium on Access Control Models and Technologies Message-ID: ACM SACMAT 2022 ----------------------------------------------- | ONLINE | ----------------------------------------------- Call for Research Papers ============================================================== Papers offering novel research contributions are solicited for submission. Accepted papers will be presented at the symposium and published by the ACM in the symposium proceedings. In addition to the regular research track, this year SACMAT will again host the special track -- "Blue Sky/Vision Track". Researchers are invited to submit papers describing promising new ideas and challenges of interest to the community as well as access control needs emerging from other fields. We are particularly looking for potentially disruptive and new ideas which can shape the research agenda for the next 10 years. We also encourage submissions to the "Work-in-progress Track" to present ideas that may have not been completely developed and experimentally evaluated. Topics of Interest ============================================================== Submissions to the regular track covering any relevant area of access control are welcomed. Areas include, but are not limited to, the following: * Systems: * Operating systems * Cloud systems and their security * Distributed systems * Fog and Edge-computing systems * Cyber-physical and Embedded systems * Mobile systems * Autonomous systems (e.g., UAV security, autonomous vehicles, etc) * IoT systems (e.g., home-automation systems) * WWW * Design for resiliency * Designing systems with zero-trust architecture * Network: * Network systems (e.g., Software-defined network, Network function virtualization) * Corporate and Military-grade Networks * Wireless and Cellular Networks * Opportunistic Network (e.g., delay-tolerant network, P2P) * Overlay Network * Satellite Network * Privacy and Privacy-enhancing Technologies: * Mixers and Mixnets * Anonymous protocols (e.g., Tor) * Online social networks (OSN) * Anonymous communication and censorship resistance * Access control and identity management with privacy * Cryptographic tools for privacy * Data protection technologies * Attacks on Privacy and their defenses * Authentication: * Password-based Authentication * Biometric-based Authentication * Location-based Authentication * Identity management * Usable authentication * Mechanisms: * Blockchain Technologies * AI/ML Technologies * Cryptographic Technologies * Programming-language based Technologies * Hardware-security Technologies (e.g., Intel SGX, ARM TrustZone) * Economic models and game theory * Trust Management * Usable mechanisms * Data Security: * Big data * Databases and data management * Data leakage prevention * Data protection on untrusted infrastructure * Policies and Models: * Novel policy language design * New Access Control Models * Extension of policy languages * Extension of Models * Analysis of policy languages * Analysis of Models * Policy engineering and policy mining * Verification of policy languages * Efficient enforcement of policies * Usable access control policy New in ACM SACMAT 2022 ============================================================== We are moving ACM SACMAT 2022 to have two submission cycles. Authors submitting papers in the first submission cycle will have the opportunity to receive a major revision verdict in addition to the usual accept and reject verdicts. Authors can decide to prepare a revised version of the paper and submit it to the second submission cycle for consideration. Major revision papers will be reviewed by the program committee members based on the criteria set forward by them in the first submission cycle. Regular Track Paper Submission and Format ============================================================== Papers must be written in?English. Authors are required to use the ACM format for papers, using the two-column SIG Proceedings Template (the sigconf template for LaTex) available in the following link: https://www.acm.org/publications/authors/submissions The length of the paper in the proceedings format must not exceed?twelve?US letter pages formatted for 8.5" x 11" paper and be no more than 5MB in size. It is the responsibility of the authors to ensure that their submissions will print easily on simple default configurations. The submission must be anonymous, so information that might identify the authors - including author names, affiliations, acknowledgments, or obvious self-citations - must be excluded. It is the authors' responsibility to ensure that their anonymity is preserved when citing their work. Submissions should be made to the EasyChair conference management system by the paper submission deadline of: February 25th, 2022 (DEADLINE EXTENSION Submission Cycle 2) All submissions must contain a?significant original contribution. That is, submitted papers must not substantially overlap papers that have been published or that are simultaneously submitted to a journal, conference, or workshop. In particular, simultaneous submission of the same work is not allowed. Wherever appropriate, relevant related work, including that of the authors, must be cited. Submissions that are not accepted as full papers may be invited to appear as short papers. At least one author from each accepted paper must register for the conference before the camera-ready deadline. Blue Sky Track Paper Submission and Format ============================================================== All submissions to this track should be in the same format as for the regular track, but the length must not exceed ten US letter pages, and the submissions are not required to be anonymized (optional). Submissions to this track should be submitted to the EasyChair conference management system by the same deadline as for the regular track. Work-in-progress Track Paper Submission and Format ============================================================== Authors are invited to submit papers in the newly introduced work-in-progress track. This track is introduced for (junior) authors, ideally, Ph.D. and Master's students, to obtain early, constructive feedback on their work. Submissions in this track should follow the same format as for the regular track papers while limiting the total number of pages to six US letter pages. Paper submitted in this track should be anonymized and can be submitted to the EasyChair conference management system by the same deadline as for the regular track. Call for Lightning Talk ============================================================== Participants are invited to submit proposals for 5-minute lightning talks describing recently published results, work in progress, wild ideas, etc. Lightning talks are a new feature of SACMAT, introduced this year to partially replace the informal sharing of ideas at in-person meetings. Submissions are expected??by May 27, 2022. Notification of acceptance will be on June 3, 2022. Call for Posters ============================================================== SACMAT 2022 will include a poster session to promote discussion of ongoing projects among researchers in the field of access control and computer security. Posters can cover preliminary or exploratory work with interesting ideas, or research projects in the early stages with promising results in all aspects of access control and computer security. Authors interested in displaying a poster must submit a poster abstract in the same format as for the regular track, but the length must not exceed three US letter pages, and the submission should not be anonymized. The title should start with "Poster:". Accepted poster abstracts will be included in the conference proceedings. Submissions should be emailed to the poster chair by Apr 15th, 2022. The subject line should include "SACMAT 2022 Poster:" followed by the poster title. Call for Demos ============================================================== A demonstration proposal should clearly describe (1) the overall architecture of the system or technology to be demonstrated, and (2) one or more demonstration scenarios that describe how the audience, interacting with the demonstration system or the demonstrator, will gain an understanding of the underlying technology. Submissions will be evaluated based on the motivation of the work behind the use of the system or technology to be demonstrated and its novelty. The subject line should include "SACMAT 2022 Demo:" followed by the demo title. Demonstration proposals should be in the same format as for the regular track, but the length must not exceed four US letter pages, and the submission should not be anonymized. A two-page description of the demonstration will be included in the conference proceedings. Submissions should be emailed to the Demonstrations Chair by Apr 15th, 2022. Financial Conflict of Interest (COI) Disclosure: ============================================================== In the interests of transparency and to help readers form their own judgments of potential bias, ACM SACMAT requires authors and PC members to declare any competing financial and/or non-financial interests in relation to the work described. Definition ------------------------- For the purposes of this policy, competing interests are defined as financial and non-financial interests that could directly undermine, or be perceived to undermine the objectivity, integrity, and value of a publication, through a potential influence on the judgments and actions of authors with regard to objective data presentation, analysis, and interpretation. Financial competing interests include any of the following: Funding: Research support (including salaries, equipment, supplies, and other expenses) by organizations that may gain or lose financially through this publication. A specific role for the funding provider in the conceptualization, design, data collection, analysis, decision to publish, or preparation of the manuscript, should be disclosed. Employment: Recent (while engaged in the research project), present or anticipated employment by any organization that may gain or lose financially through this publication. Personal financial interests: Ownership or contractual interest in stocks or shares of companies that may gain or lose financially through publication; consultation fees or other forms of remuneration (including reimbursements for attending symposia) from organizations that may gain or lose financially; patents or patent applications (awarded or pending) filed by the authors or their institutions whose value may be affected by publication. For patents and patent applications, disclosure of the following information is requested: patent applicant (whether author or institution), name of the inventor(s), application number, the status of the application, specific aspect of manuscript covered in the patent application. It is difficult to specify a threshold at which a financial interest becomes significant, but note that many US universities require faculty members to disclose interests exceeding $10,000 or 5% equity in a company. Any such figure is necessarily arbitrary, so we offer as one possible practical alternative guideline: "Any undeclared competing financial interests that could embarrass you were they to become publicly known after your work was published." We do not consider diversified mutual funds or investment trusts to constitute a competing financial interest. Also, for employees in non-executive or leadership positions, we do not consider financial interest related to stocks or shares in their company to constitute a competing financial interest, as long as they are publishing under their company affiliation. Non-financial competing interests: Non-financial competing interests can take different forms, including personal or professional relations with organizations and individuals. We would encourage authors and PC members to declare any unpaid roles or relationships that might have a bearing on the publication process. Examples of non-financial competing interests include (but are not limited to): * Unpaid membership in a government or non-governmental organization * Unpaid membership in an advocacy or lobbying organization * Unpaid advisory position in a commercial organization * Writing or consulting for an educational company * Acting as an expert witness Conference Code of Conduct and Etiquette ============================================================== ACM SACMAT will follow the ACM Policy Against Harassment at ACM Activities. Please familiarize yourself with the ACM Policy Against Harassment (available at https://www.acm.org/special-interest-groups/volunteer-resources/officers-manual/ policy-against-discrimination-and-harassment) and guide to Reporting Unacceptable Behavior (available at https://www.acm.org/about-acm/reporting-unacceptable-behavior). AUTHORS TAKE NOTE ============================================================== The official publication date is the date the proceedings are made available in the ACM Digital Library. This date may be up to two weeks before the first day of your conference. The official publication date affects the deadline for any patent filings related to published work. (For those rare conferences whose proceedings are published in the ACM Digital Library after the conference is over, the official publication date remains the first day of the conference.) Important dates ============================================================== **Note that, these dates are currently only tentative and subject to change.** * Paper submission: February 25th, 2022 (Submission Cycle 2) * Rebuttal: March 24th - March 28th, 2022 (Submission Cycle 2) * Notifications: April 8th, 2022 (Submission Cycle 2) * Systems demo and Poster submissions: April 15th, 2022 * Systems demo and Poster notifications: April 22nd, 2022 * Panel Proposal: March 18th, 2022 * Camera-ready paper submission: April 29th, 2022 * Conference date: June 8 - June 10, 2022 -------------- next part -------------- An HTML attachment was scrubbed... URL: From lorincz at inf.elte.hu Fri Feb 18 00:49:22 2022 From: lorincz at inf.elte.hu (Andras Lorincz) Date: Fri, 18 Feb 2022 05:49:22 +0000 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: Dear Asim: This is how I see it ? without warranties? Consciousness at a low level is necessary for a distributed system with computational delays and learned predictive capabilities. It can arise as a consequence of having synchronous internal predictive and generative representations and external observations about the actions of the self. All components of this loop; (a) observation of the actions, (b) representation of the actions, (c) actions should be synchronous. That has consequences on free will in the 200 ms domain. Since model-based prediction can achieve synchronicity, the same system can be developed further for more extended time scales. Then -- in agreement with Hava Siegelmann's proposal -- the problem disappears. The curse of dimensionality remains, and, from the point of view of intelligence, there is a need for algorithms that can overcome the curse and find the relevant (and few) components for problem-solving. One such component is the self, its own features, including its own capabilities and desires. Self-consciousness emerges if the self is separated from the rest of the world (being a more challenging task and developing slowly for many autistic people, for example). Constraints of the self can be established, and free will can play a role within these constraints. It is very different from low-level consciousness. Best, Andr?s ------------------------------------ Andras Lorincz http://nipg.inf.elte.hu/ Fellow of the European Association for Artificial Intelligence https://scholar.google.com/citations?user=EjETXQkAAAAJ&hl=en Department of Artificial Intelligence Faculty of Informatics Eotvos Lorand University Budapest, Hungary ________________________________ From: Asim Roy Sent: Friday, February 18, 2022 5:22 AM To: Andras Lorincz ; Stephen Jos? Hanson ; Gary Marcus Cc: Connectionists Subject: RE: Connectionists: Weird beliefs about consciousness In 1998, after our debate about the brain at the WCCI in Anchorage, Alaska, I asked Walter Freeman if he thought the brain controls the body. His answer was, you can also say that the body controls the brain. I then asked him if the driver controls a car, or the pilot controls an airplane. His answer was the same, that you can also say that the car controls the driver, or the plane controls the pilot. I then realized that Walter was also a philosopher and believed in the No-free Will theory and what he was arguing for is that the world is simply made of interacting systems. However, both Walter, and his close friend John Taylor, were into consciousness. I have argued with Walter on many different topics over nearly two decades and have utmost respect for him as a scholar, but this first argument I will always remember. Obviously, there?s a conflict between consciousness and the No-free Will theory. Wonder where we stand with regard to this conflict. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Andras Lorincz Sent: Tuesday, February 15, 2022 6:50 AM To: Stephen Jos? Hanson ; Gary Marcus Cc: Connectionists Subject: Re: Connectionists: Weird beliefs about consciousness Dear Steve and Gary: This is how I see (try to understand) consciousness and the related terms: (Our) consciousness seems to be related to the close-to-deterministic nature of the episodes on from few hundred millisecond to a few second domain. Control instructions may leave our brain 200 ms earlier than the action starts and they become conscious only by that time. In addition, observations of those may also be delayed by a similar amount. (It then follows that the launching of the control actions is not conscious and -- therefore -- free will can be debated in this very limited context.) On the other hand, model-based synchronization is necessary for timely observation, planning, decision making, and execution in a distributed and slow computational system. If this model-based synchronization is not working properly, then the observation of the world breaks and schizophrenic symptoms appear. As an example, individuals with pronounced schizotypal traits are particularly successful in self-tickling (source: https://philpapers.org/rec/LEMIWP, and a discussion on Asperger and schizophrenia: https://www.frontiersin.org/articles/10.3389/fpsyt.2020.503462/full) a manifestation of improper binding. The internal model enables and the synchronization requires the internal model and thus a certain level of consciousness can appear in a time interval around the actual time instant and its length depends on the short-term memory. Other issues, like separating the self from the rest of the world are more closely related to the soft/hard style interventions (as called in the recent deep learning literature), i.e., those components (features) that can be modified/controlled, e.g., color and speed, and the ones that are Lego-like and can be separated/amputed/occluded/added. Best, Andras ------------------------------------ Andras Lorincz http://nipg.inf.elte.hu/ Fellow of the European Association for Artificial Intelligence https://scholar.google.com/citations?user=EjETXQkAAAAJ&hl=en Department of Artificial Intelligence Faculty of Informatics Eotvos Lorand University Budapest, Hungary ________________________________ From: Connectionists > on behalf of Stephen Jos? Hanson > Sent: Monday, February 14, 2022 8:30 PM To: Gary Marcus > Cc: Connectionists > Subject: Re: Connectionists: Weird beliefs about consciousness Gary, these weren't criterion. Let me try again. I wasn't talking about wake-sleep cycles... I was talking about being awake or asleep and the transition that ensues.. Rooba's don't sleep.. they turn off, I have two of them. They turn on once (1) their batteries are recharged (2) a timer has been set for being turned on. GPT3 is essentially a CYC that actually works.. by reading Wikipedia (which of course is a terribly biased sample). I was indicating the difference between implicit and explicit learning/problem solving. Implicit learning/memory is unconscious and similar to a habit.. (good or bad). I believe that when someone says "is gpt3 conscious?" they are asking: is gpt3 self-aware? Roombas know about vacuuming and they are unconscious. S On 2/14/22 12:45 PM, Gary Marcus wrote: Stephen, On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more plausible as a consciousness than GPT-3. 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. 2. Roomba makes choices based on an explicit representation of its location relative to a mapped space. GPT lacks any consistent reflection of self; eg if you ask it, as I have, if you are you person, and then ask if it is a computer, it?s liable to say yes to both, showing no stable knowledge of self. 3. Roomba has explicit, declarative knowledge eg of walls and other boundaries, as well its own location. GPT has no systematically interrogable explicit representations. All this is said with tongue lodged partway in cheek, but I honestly don?t see what criterion would lead anyone to believe that GPT is a more plausible candidate for consciousness than any other AI program out there. ELIZA long ago showed that you could produce fluent speech that was mildly contextually relevant, and even convincing to the untutored; just because GPT is a better version of that trick doesn?t mean it?s any more conscious. Gary On Feb 14, 2022, at 08:56, Stephen Jos? Hanson wrote: ? this is a great list of behavior.. Some biologically might be termed reflexive, taxes, classically conditioned, implicit (memory/learning)... all however would not be conscious in the several senses: (1) wakefulness-- sleep (2) self aware (3) explicit/declarative. I think the term is used very loosely, and I believe what GPT3 and other AI are hoping to show signs of is "self-awareness".. In response to : "why are you doing that?", "What are you doing now", "what will you be doing in 2030?" Steve On 2/14/22 10:46 AM, Iam Palatnik wrote: A somewhat related question, just out of curiosity. Imagine the following: - An automatic solar panel that tracks the position of the sun. - A group of single celled microbes with phototaxis that follow the sunlight. - A jellyfish (animal without a brain) that follows/avoids the sunlight. - A cockroach (animal with a brain) that avoids the sunlight. - A drone with onboard AI that flies to regions of more intense sunlight to recharge its batteries. - A human that dislikes sunlight and actively avoids it. Can any of these, beside the human, be said to be aware or conscious of the sunlight, and why? What is most relevant? Being a biological life form, having a brain, being able to make decisions based on the environment? Being taxonomically close to humans? On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus > wrote: Also true: Many AI researchers are very unclear about what consciousness is and also very sure that ELIZA doesn?t have it. Neither ELIZA nor GPT-3 have - anything remotely related to embodiment - any capacity to reflect upon themselves Hypothesis: neither keyword matching nor tensor manipulation, even at scale, suffice in themselves to qualify for consciousness. - Gary > On Feb 14, 2022, at 00:24, Geoffrey Hinton > wrote: > > ?Many AI researchers are very unclear about what consciousness is and also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > -- [cid:image001.png at 01D82444.67C0E320] -- [cid:image001.png at 01D82444.67C0E320] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 19957 bytes Desc: image001.png URL: From juyang.weng at gmail.com Thu Feb 17 19:42:04 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Thu, 17 Feb 2022 19:42:04 -0500 Subject: Connectionists: Weird beliefs about consciousness Message-ID: Danko: You wrote "Attention may not be a separate mechanism but a shadow of something else in the brain that also does working memory and other stuff like decision making and intelligence in general. We don't know what that is. " I urge you to learn universal Turing machines (UTMs). Although UTMs were not for brains, I found that they are extremely useful for modeling and understanding the brain, because so much formalism has been built rigorously on them. In my model, state and actions must be considered the same but UTMs treat them separately. I hope many people on this list do not overlook UTMs. Best regards, -John -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Thu Feb 17 19:22:46 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Thu, 17 Feb 2022 19:22:46 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Dear Danko, Electrical engineering probably contributes the most papers in neural networks among these six disciplines. Of course, many of these papers do not yet aim to model biological brains. Some of the 20 million-dollar problems I solved were raised from electrical engineering, such as the inverse kinematics problem and the nonlinear controller problem. I wish electrical engineering would have a revolution soon from my brain model. Most of my students are from electrical engineering. Best regards, -John On Thu, Feb 17, 2022 at 3:15 AM Danko Nikolic wrote: > > Dear Juyang, > > You wrote "Senior people do not want to get a PhD in all 6 disciplines in > the attached figure: biology, neuroscience, psychology, computer science, > electrical engineering, mathematics." > > I would cross electrical engineering from that list. It seems to me that > the contribution of electrical engineering is minor. But then I would add > philosophy of mind and cybernetics. These two seem a lot more important > to acquire a PhD-level knowledge in. > > Best, > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > --- A progress usually starts with an insight --- > > > On Thu, Feb 17, 2022 at 8:22 AM Juyang Weng wrote: > >> Dear Tsvi: >> >> You wrote: "I believe scientists not seeing eye-to-eye with each other >> and other members of the community is in no small part due to these terms." >> >> I agree. This is a HUGE problem, as the attached figure "Blind Men and >> an Elephant" indicates. What should this multidisciplinary community >> do? Senior people do not want to get a PhD in all 6 disciplines in the >> attached figure: biology, neuroscience, psychology, computer science, >> electrical engineering, mathematics. >> >> Best regards, >> -John >> >> On Tue, Feb 15, 2022 at 10:00 PM Tsvi Achler wrote: >> >>> After studying the brain from a multidisciplinary perspective I am well >>> aware of the difficulties speaking and understanding each other across >>> disciplines. There are many terms that are defined differently in >>> different fields... and unfortunately things are not as simple as looking >>> them up in a dictionary. >>> >>> For example the term recurrent connections have different meanings in >>> the computational neuroscience, neural networks, and cognitive psychology >>> communities. >>> In neural networks recurrent means an output used back as an input >>> within a paradigm of delayed inputs. It is a method of representing time >>> or sequences. Often recurrent connections in neural networks are confused >>> with feedback back to the same inputs which are actually never used in >>> neural networks because it forms an infinite loop and is not possible to >>> rewind in order to generate an error signal. >>> In computational neuroscience recurrent connections are used to describe >>> lateral connections. >>> In cognitive psychology the term re-entrant connections are used to >>> describe feedback back to the same inputs. >>> >>> I believe in order to truly appreciate "brain-like" ideas, members of >>> this group need to familiarize themselves with these brain-focused fields. >>> For example in cognitive psychology there is a rich literature on salience >>> (which again is a bit different from salience in the neural network >>> community). Salience is a dynamic process which determines how well a >>> certain input or input feature is processed. Salience changes in the brain >>> depending on what other inputs or features are concurrently present or what >>> the person is instructed to focus on. There is very little appreciation, >>> integration or implementation of these findings in feedforward networks, >>> yet salience plays a factor in every recognition decision and modality >>> including smell and touch. >>> >>> Consciousness is a particularly problematic minefield which also adds in >>> philosophy, metaphysics and subjectivity into the mix. >>> >>> Juyang, I think we both agree about the basics: the need for more >>> realistic real world recognition and to move beyond the rehearsal >>> limitations of neural networks. I believe scientists not seeing eye-to-eye >>> with each other and other members of the community is in no small part due >>> to these terms. >>> >>> Sincerely, >>> -Tsvi >>> >>> >>> >>> On Tue, Feb 15, 2022 at 9:54 AM Juyang Weng >>> wrote: >>> >>>> Dear Tsvi, >>>> You wrote "A huge part of the problem in any discussion about >>>> consciousness is there isn't even a clear definition of consciousness". >>>> Look at the 5 level definition of consciousness: >>>> https://www.merriam-webster.com/dictionary/consciousness >>>> >>>> You wrote: "So consciousness is not necessary or sufficient for complex >>>> thoughts or behavior." >>>> I was thinking that way too, until recently. >>>> I found consciousness IS REQUIRED for even learning basic intelligence. >>>> To put it in a short way so that people on this list can benefit: >>>> The motors (as context/actions) in the brain require consciousness in >>>> order to learn correctly in the physical world. Please read the first >>>> model about conscious learning: >>>> J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th >>>> International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, >>>> USA, Jan.7-9, 2022. PDF file >>>> >>>> . >>>> >>>> Best regards, >>>> -John >>>> ---- >>>> From: Tsvi Achler >>>> To: Iam Palatnik >>>> Cc: Connectionists >>>> Subject: Re: Connectionists: Weird beliefs about consciousness >>>> >>>> -- >>>> Juyang (John) Weng >>>> >>> >> >> -- >> Juyang (John) Weng >> > -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From marius.bilasco at univ-lille.fr Thu Feb 17 16:03:24 2022 From: marius.bilasco at univ-lille.fr (Ioan Marius BILASCO) Date: Thu, 17 Feb 2022 22:03:24 +0100 Subject: Connectionists: [jobs] PhD opportunity - Spatio-temporal augmentation models for motion pattern learning - AI_PhD@Lille, France Message-ID: <3caa0ba6-eed1-f845-2f18-e0c0ca4b294d@univ-lille.fr> The FOX team from the CRIStAL laboratory (UMR CNRS), Lille France is looking to recruit a PhD student starting on October 1st 2022 on the following subject : Spatio-temporal data augmentation models for motion pattern learning using deep learning: applications to facial analysis in the wild The FOX research group is part of the CRIStAL laboratory (University of Lille, CNRS), located in Lille, France. We focus on video analysis for human behavior understanding. Specifically, we develop spatio-temporal models of motions for tasks such as abnormal event detection, emotion recognition, and face alignment. Our work is published in major journals (Pattern Recognition, IEEE Trans. on Affective Computing) and conferences (WACV, IJCNN). Abstract: Facial expression analysis is a well-studied field when dealing with segmented and constrained data captured in lab conditions. However, many challenges must still be addressed for building in-the-wild solutions that account for various motion intensities, strong head movements during expressions, the spotting of the subsequence containing the expression, partially occluded faces, etc. In recent years, learned features based on deep learning architectures were proposed in order to deal with these challenges. Deep learning is characterized by neural architectures that depend on a huge number of parameters. The convergence of these neural networks and the estimation of optimal parameters require large amounts of training data, especially when dealing with spatio-temporal data, particulary adequate for facial expression recognition. The quantity, but also the quality, of the data and its capacity to reflect the addressed challenges are key elements for training properly the networks. Augmenting the data artificially in an intelligent and controlled way is an interesting solution. The augmentation techniques identified in the literature are mainly focused on image augmentation and consist of scaling, rotation, and flipping operations, or they make use of more complex adversarial training. These techniques can be applied at the frame level, but there is a need for sequence level augmentation in order to better control the augmentation process and ensure the absence of temporal artifacts that might bias the learning process. The generation of dynamic frontal facial expressions has already been addressed in the literature. The goal of this Ph.D. is to conceive new space-time augmentation methods for unconstrained facial analysis (involving head movements, occultations, etc.). Attention should be paid in assessing the quality standards related to facial expression requirements: stability over time, absence of facial artifacts, etc. More specifically, the Ph.D. can didate is expected to conceive augmentation architectures that address various challenges (motion intensities, head movements) while maintaining temporal stability and eliminating facial artifacts. More details are available here : https://bit.ly/staugm_motion Candidates must hold a Master degree in Computer Science, Statistics, Applied Mathematics or a related field. Experience in one or more of the following is a plus: ? image processing, computer vision; ? machine learning; ? research methodology (literature review, experimentation?). Candidates should have the following skills: ? good proficiency in English, both spoken and written; ? scientific writing; ? programming (experience in C++ is a plus, but not mandatory). This PHD thesis will be funded in the framework of the AI_PhD at Lilleprogram. http://www.isite-ulne.fr/index.php/en/phd-in-artificial-intelligence/ The candidate will be funded for 3 years; he/she is expected to defend his/her thesis and graduate by the end of the contract. The monthly gross salary is around 2000?, including benefits (health insurance, retirement fund, and paid vacations). Additional financial support is expected in the framework of the AI_PhD at Lille program. The position is located in Lille, France. With over 110 000 students, the metropolitan area of Lille is one France's top education student cities. The European Doctoral College Lille Nord-Pas de Calais is headquartered in Lille Metropole and includes 3,000 PhD Doctorate students supported by university research laboratories. Lille has a convenient location in the European high-speed rail network. It lies on the Eurostar line to London (1:20 hour journey). The French TGV network also puts it only 1 hour from Paris, 35 mn from Brussels, and a short trips to other major centres in France such as Paris, Marseille and Lyon. We look forward to receiving your application as soon as possible, but no later than 26.03.2021. -------------- next part -------------- An HTML attachment was scrubbed... URL: From marius.bilasco at univ-lille.fr Thu Feb 17 16:12:53 2022 From: marius.bilasco at univ-lille.fr (Ioan Marius BILASCO) Date: Thu, 17 Feb 2022 22:12:53 +0100 Subject: Connectionists: [jobs] Phd opportunity in Lille, France - Spiking Neural Networks for Video Analysis Message-ID: <88130C53-44F0-4CC5-B36E-84A96BD2A95D@univ-lille.fr> The FOX team from the CRIStAL laboratory (UMR CNRS), Lille France is looking to recruit a PhD student starting as soon as possible on the following subject : Spiking Neural Networks for Video Analysis The FOX research group is part of the CRIStAL laboratory (University of Lille, CNRS), located in Lille, France. We focus on video analysis for human behavior understanding. Specifically, we develop spatio-temporal models of motions for tasks such as abnormal event detection, emotion recognition, and face alignment. We are also involved in IRCICA (CNRS), a research institute promoting multidisciplanary research. At IRCICA, we collaborate with computer scientists and experts in electronics engineering to create new models of neural networks that can be implemented on low-power hardware architectures. Recently, we designed state-of-the-art models for image recognition with single and multi-layer unsupervised spiking neural networks. We were among the first to succesfully apply unsupervised SNNs on modern datasets of computer vision. We also developed our own SNN simulator to support experiments with SNN on computer vision problems. Our work is published in major journals (Pattern Recognition, IEEE Trans. on Affective Computing) and conferences (WACV, IJCNN) in the field. Abstract: Spiking Neural Network have recently been evaluated on classical image recognition tasks [1]. This work has highlighted their promising performances in this domain and have identified ways to improve them to be competitive with comparable deep learning approaches. In particular, it demonstrated the ability of SNN architectures to learn relevant patterns for static pattern recognition in an unsupervised manner. However, dealing with static images is not enough, and the computer vision community is increasingly interested in video analysis, for two reasons. First, video data is more and more common and corresponds to a wide range of applications (video surveillance, audio-visual productions, autonomous vehicles...). Second, this data is richer than isolated static images, and thus offers the possibility to develop more effective systems, e.g. using motion information. Thus, it is recognized in the community that modeling motion in videos is more relevant than studying visual appearance alone for tasks such as action or emotion recognition. The next step for SNNs is therefore to study their ability to model motion rather than, or in addition to, image appearance. The goal of the Ph.D. candidate will be to explore the use of SNNs for space-time modeling in videos. This work will be targeted towards applications in human behavior understanding and especially action recognition. More specifically, the Ph.D. candidate is expected to: * identify what issues may prevent space-time modeling with SNNs and how they can be circumvented; * propose new supervised and unsupervised SNN models for motion modeling, which are compatible with hardware implementations on ultra-low power devices; * evaluate the proposed models on standard datasets for video analysis. Detailed subject: https://bit.ly/stssnnfox Candidates must hold a Master degree (or an equivalent degree) in Computer Science, Statistics, Applied Mathematics or a related field. Experience in one or more of the following is a plus: ? image processing, computer vision; ? machine learning; ? bio-inspired computing; ? research methodology (literature review, experimentation?). Candidates should have the following skills: ? good proficiency in English, both spoken and written; ? scientific writing; ? programming (experience in C++ is a plus, but not mandatory). This PHD thesis will be funded in the framework of the ANVI-Luxant industrial chair. The general objective of the Chair is to make a scientific and technological progress in the mastery of emerging information processing architectures such as neuromorphic architectures as an embedded artificial intelligence technique. The use-case studies will come from video protection in the context of retail and transportation. The candidate will be funded for 3 years; he/she is expected to defend his/her thesis and graduate by the end of the contract. The monthly gross salary is around 2000?, including benefits (health insurance, retirement fund, and paid vacations). The position is located in Lille, France. With over 110 000 students, the metropolitan area of Lille is one France's top education student cities. The European Doctoral College Lille Nord-Pas de Calais is headquartered in Lille Metropole and includes 3,000 PhD Doctorate students supported by university research laboratories. Lille has a convenient location in the European high-speed rail network. It lies on the Eurostar line to London (1:20 hour journey). The French TGV network also puts it only 1 hour from Paris, 35 mn from Brussels, and a short trips to other major centres in France such as Paris, Marseille and Lyon. For application, please send the following information in a single PDF file to Dr. Marius Bilasco (marius.bilasco at univ-lille.fr ) with subject [PhD_Luxant-ANVI]: * A cover letter. * A curriculum vitae, including a list of publications, if any. * Transcripts of grades of Master's degree. * The contact information of two references (and any letters if available). We look forward to receiving your application as soon as possible, but no later than 26.3.2022 -------------- next part -------------- An HTML attachment was scrubbed... URL: From juyang.weng at gmail.com Thu Feb 17 19:31:51 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Thu, 17 Feb 2022 19:31:51 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: Message-ID: Dear Tsvi, You wrote " Nor does there seem to be cross interest within the academic computational communities where this model can show top-down attention effects and controls similar to Bayesian models (with dynamic priors, priming, and weights similar to likelihoods helping with explainability) but is completely connectionist." As far as I can see, in terms of top-down modeling, it is more an issue of "lack of computational knowledge" than a lack of "cross interest". The landscape of AI should greatly change soon from my brain model of conscious learning, sooner than the wave that Cresceptron created. I hope those who will use the DN brain model do not intentionally plagiarize DN as many papers did to Cresceptron. Best regards, -John On Thu, Feb 17, 2022 at 5:25 AM Tsvi Achler wrote: > Hi Danko, > Indeed there are so many top down effects based on what has been learned, > which is further evidence that salience is inseparable from recognition. > However as a note about education, I dont buy into defining > especially STEM fields that need to be studied. If what is needed to be > known is known, it wouldn't be science. > My undergraduate degree is in (ahem) Electrical Engineering Computer > Science. > The electrical engineering provided useful tools through control theory to > understand and manage the dynamics of top-down feedback. This includes the > mathematical tools to determine stability and steady state endpoints of > dynamical systems. > > In the recognition model I developed, regulatory feedback networks, each > input is regulated by the outputs it activates which subsequently creates a > dynamical salience which makes sure inputs are not over or under > represented within the network and lets the system evaluate all the inputs > together. > > This model inherently displays the signal-to-noise salience phenomena that > includes difficulty with similarity and asymmetry seen in humans even if > the inputs are not spatial. > It shows Excitation-Inhibition balance findings in neuroscience (Xue et al > 2014) or what I call network-wide-bursting. > It also does not require the iid rehearsal to learn like feedforward > models because it uses the salience-feedback dynamics to determine the best > relevance of each input given what has been learned and what is present at > recognition time. This dynamic salience is inseparable from the mechanism > of recognition. > > Unfortunately I dont seem to be able to convince the academic cognitive > psychology and neuroscience communities that while they can model any of > their phenomena with enough parameters, models with huge parameter spaces > are less desirable. Scalability and minimizing degrees of freedom are > important in models: models that show the most amount of phenomena > (across multiple disciplines) with the least amount of free parameters are > better. > Neither do I seem able to convince the academic connectionist community > that the iid rehearsal requirements are killing the ability to use models > in a natural way and can be avoided using feedback back to the inputs > during recognition. Nor does there seem to be cross interest within the > academic computational communities where this model can show top-down > attention effects and controls similar to Bayesian models (with dynamic > priors, priming, and weights similar to likelihoods helping with > explainability) but is completely connectionist. > Sincerely, > -Tsvi > Xue, M, Atallah, BV, Scanziani, M. (2014). Equalizing > excitation-inhibition ratios across visual cortical neurons. Nature 511, > 596?600 > > https://www.youtube.com/watch?v=F-GBIZoZ1mI&list=PL4nMP8F3B7bg3cNWWwLG8BX-wER2PeB-3&index=1 > > > > On Thu, Feb 17, 2022 at 12:15 AM Danko Nikolic > wrote: > >> >> Dear Juyang, >> >> You wrote "Senior people do not want to get a PhD in all 6 disciplines in >> the attached figure: biology, neuroscience, psychology, computer science, >> electrical engineering, mathematics." >> >> I would cross electrical engineering from that list. It seems to me that >> the contribution of electrical engineering is minor. But then I would add >> philosophy of mind and cybernetics. These two seem a lot more important >> to acquire a PhD-level knowledge in. >> >> Best, >> >> Danko >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> https://www.linkedin.com/in/danko-nikolic/ >> --- A progress usually starts with an insight --- >> >> >> On Thu, Feb 17, 2022 at 8:22 AM Juyang Weng >> wrote: >> >>> Dear Tsvi: >>> >>> You wrote: "I believe scientists not seeing eye-to-eye with each other >>> and other members of the community is in no small part due to these terms." >>> >>> I agree. This is a HUGE problem, as the attached figure "Blind Men and >>> an Elephant" indicates. What should this multidisciplinary community >>> do? Senior people do not want to get a PhD in all 6 disciplines in the >>> attached figure: biology, neuroscience, psychology, computer science, >>> electrical engineering, mathematics. >>> >>> Best regards, >>> -John >>> >>> On Tue, Feb 15, 2022 at 10:00 PM Tsvi Achler wrote: >>> >>>> After studying the brain from a multidisciplinary perspective I am well >>>> aware of the difficulties speaking and understanding each other across >>>> disciplines. There are many terms that are defined differently in >>>> different fields... and unfortunately things are not as simple as looking >>>> them up in a dictionary. >>>> >>>> For example the term recurrent connections have different meanings in >>>> the computational neuroscience, neural networks, and cognitive psychology >>>> communities. >>>> In neural networks recurrent means an output used back as an input >>>> within a paradigm of delayed inputs. It is a method of representing time >>>> or sequences. Often recurrent connections in neural networks are confused >>>> with feedback back to the same inputs which are actually never used in >>>> neural networks because it forms an infinite loop and is not possible to >>>> rewind in order to generate an error signal. >>>> In computational neuroscience recurrent connections are used to >>>> describe lateral connections. >>>> In cognitive psychology the term re-entrant connections are used to >>>> describe feedback back to the same inputs. >>>> >>>> I believe in order to truly appreciate "brain-like" ideas, members of >>>> this group need to familiarize themselves with these brain-focused fields. >>>> For example in cognitive psychology there is a rich literature on salience >>>> (which again is a bit different from salience in the neural network >>>> community). Salience is a dynamic process which determines how well a >>>> certain input or input feature is processed. Salience changes in the brain >>>> depending on what other inputs or features are concurrently present or what >>>> the person is instructed to focus on. There is very little appreciation, >>>> integration or implementation of these findings in feedforward networks, >>>> yet salience plays a factor in every recognition decision and modality >>>> including smell and touch. >>>> >>>> Consciousness is a particularly problematic minefield which also adds >>>> in philosophy, metaphysics and subjectivity into the mix. >>>> >>>> Juyang, I think we both agree about the basics: the need for more >>>> realistic real world recognition and to move beyond the rehearsal >>>> limitations of neural networks. I believe scientists not seeing eye-to-eye >>>> with each other and other members of the community is in no small part due >>>> to these terms. >>>> >>>> Sincerely, >>>> -Tsvi >>>> >>>> >>>> >>>> On Tue, Feb 15, 2022 at 9:54 AM Juyang Weng >>>> wrote: >>>> >>>>> Dear Tsvi, >>>>> You wrote "A huge part of the problem in any discussion about >>>>> consciousness is there isn't even a clear definition of consciousness". >>>>> Look at the 5 level definition of consciousness: >>>>> https://www.merriam-webster.com/dictionary/consciousness >>>>> >>>>> You wrote: "So consciousness is not necessary or sufficient for >>>>> complex thoughts or behavior." >>>>> I was thinking that way too, until recently. >>>>> I found consciousness IS REQUIRED for even learning basic >>>>> intelligence. >>>>> To put it in a short way so that people on this list can benefit: >>>>> The motors (as context/actions) in the brain require consciousness in >>>>> order to learn correctly in the physical world. Please read the first >>>>> model about conscious learning: >>>>> J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th >>>>> International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, >>>>> USA, Jan.7-9, 2022. PDF file >>>>> >>>>> . >>>>> >>>>> Best regards, >>>>> -John >>>>> ---- >>>>> From: Tsvi Achler >>>>> To: Iam Palatnik >>>>> Cc: Connectionists >>>>> Subject: Re: Connectionists: Weird beliefs about consciousness >>>>> >>>>> -- >>>>> Juyang (John) Weng >>>>> >>>> >>> >>> -- >>> Juyang (John) Weng >>> >> -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From pfbaldi at ics.uci.edu Fri Feb 18 00:35:23 2022 From: pfbaldi at ics.uci.edu (Baldi,Pierre) Date: Thu, 17 Feb 2022 21:35:23 -0800 Subject: Connectionists: The Quarks of Attention Message-ID: <9ae35848-ea6e-3c1f-ce4f-626daf2fb3b5@ics.uci.edu> For those of you interested in the most fundamental building blocks of attention, we have a new technical report: The Quarks of Attention (P. Baldi and R. Vershynin) http://arxiv.org/abs/2202.08371 From ASIM.ROY at asu.edu Thu Feb 17 23:22:03 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Fri, 18 Feb 2022 04:22:03 +0000 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: In 1998, after our debate about the brain at the WCCI in Anchorage, Alaska, I asked Walter Freeman if he thought the brain controls the body. His answer was, you can also say that the body controls the brain. I then asked him if the driver controls a car, or the pilot controls an airplane. His answer was the same, that you can also say that the car controls the driver, or the plane controls the pilot. I then realized that Walter was also a philosopher and believed in the No-free Will theory and what he was arguing for is that the world is simply made of interacting systems. However, both Walter, and his close friend John Taylor, were into consciousness. I have argued with Walter on many different topics over nearly two decades and have utmost respect for him as a scholar, but this first argument I will always remember. Obviously, there?s a conflict between consciousness and the No-free Will theory. Wonder where we stand with regard to this conflict. Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Andras Lorincz Sent: Tuesday, February 15, 2022 6:50 AM To: Stephen Jos? Hanson ; Gary Marcus Cc: Connectionists Subject: Re: Connectionists: Weird beliefs about consciousness Dear Steve and Gary: This is how I see (try to understand) consciousness and the related terms: (Our) consciousness seems to be related to the close-to-deterministic nature of the episodes on from few hundred millisecond to a few second domain. Control instructions may leave our brain 200 ms earlier than the action starts and they become conscious only by that time. In addition, observations of those may also be delayed by a similar amount. (It then follows that the launching of the control actions is not conscious and -- therefore -- free will can be debated in this very limited context.) On the other hand, model-based synchronization is necessary for timely observation, planning, decision making, and execution in a distributed and slow computational system. If this model-based synchronization is not working properly, then the observation of the world breaks and schizophrenic symptoms appear. As an example, individuals with pronounced schizotypal traits are particularly successful in self-tickling (source: https://philpapers.org/rec/LEMIWP, and a discussion on Asperger and schizophrenia: https://www.frontiersin.org/articles/10.3389/fpsyt.2020.503462/full) a manifestation of improper binding. The internal model enables and the synchronization requires the internal model and thus a certain level of consciousness can appear in a time interval around the actual time instant and its length depends on the short-term memory. Other issues, like separating the self from the rest of the world are more closely related to the soft/hard style interventions (as called in the recent deep learning literature), i.e., those components (features) that can be modified/controlled, e.g., color and speed, and the ones that are Lego-like and can be separated/amputed/occluded/added. Best, Andras ------------------------------------ Andras Lorincz http://nipg.inf.elte.hu/ Fellow of the European Association for Artificial Intelligence https://scholar.google.com/citations?user=EjETXQkAAAAJ&hl=en Department of Artificial Intelligence Faculty of Informatics Eotvos Lorand University Budapest, Hungary ________________________________ From: Connectionists > on behalf of Stephen Jos? Hanson > Sent: Monday, February 14, 2022 8:30 PM To: Gary Marcus > Cc: Connectionists > Subject: Re: Connectionists: Weird beliefs about consciousness Gary, these weren't criterion. Let me try again. I wasn't talking about wake-sleep cycles... I was talking about being awake or asleep and the transition that ensues.. Rooba's don't sleep.. they turn off, I have two of them. They turn on once (1) their batteries are recharged (2) a timer has been set for being turned on. GPT3 is essentially a CYC that actually works.. by reading Wikipedia (which of course is a terribly biased sample). I was indicating the difference between implicit and explicit learning/problem solving. Implicit learning/memory is unconscious and similar to a habit.. (good or bad). I believe that when someone says "is gpt3 conscious?" they are asking: is gpt3 self-aware? Roombas know about vacuuming and they are unconscious. S On 2/14/22 12:45 PM, Gary Marcus wrote: Stephen, On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more plausible as a consciousness than GPT-3. 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. 2. Roomba makes choices based on an explicit representation of its location relative to a mapped space. GPT lacks any consistent reflection of self; eg if you ask it, as I have, if you are you person, and then ask if it is a computer, it?s liable to say yes to both, showing no stable knowledge of self. 3. Roomba has explicit, declarative knowledge eg of walls and other boundaries, as well its own location. GPT has no systematically interrogable explicit representations. All this is said with tongue lodged partway in cheek, but I honestly don?t see what criterion would lead anyone to believe that GPT is a more plausible candidate for consciousness than any other AI program out there. ELIZA long ago showed that you could produce fluent speech that was mildly contextually relevant, and even convincing to the untutored; just because GPT is a better version of that trick doesn?t mean it?s any more conscious. Gary On Feb 14, 2022, at 08:56, Stephen Jos? Hanson wrote: ? this is a great list of behavior.. Some biologically might be termed reflexive, taxes, classically conditioned, implicit (memory/learning)... all however would not be conscious in the several senses: (1) wakefulness-- sleep (2) self aware (3) explicit/declarative. I think the term is used very loosely, and I believe what GPT3 and other AI are hoping to show signs of is "self-awareness".. In response to : "why are you doing that?", "What are you doing now", "what will you be doing in 2030?" Steve On 2/14/22 10:46 AM, Iam Palatnik wrote: A somewhat related question, just out of curiosity. Imagine the following: - An automatic solar panel that tracks the position of the sun. - A group of single celled microbes with phototaxis that follow the sunlight. - A jellyfish (animal without a brain) that follows/avoids the sunlight. - A cockroach (animal with a brain) that avoids the sunlight. - A drone with onboard AI that flies to regions of more intense sunlight to recharge its batteries. - A human that dislikes sunlight and actively avoids it. Can any of these, beside the human, be said to be aware or conscious of the sunlight, and why? What is most relevant? Being a biological life form, having a brain, being able to make decisions based on the environment? Being taxonomically close to humans? On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus > wrote: Also true: Many AI researchers are very unclear about what consciousness is and also very sure that ELIZA doesn?t have it. Neither ELIZA nor GPT-3 have - anything remotely related to embodiment - any capacity to reflect upon themselves Hypothesis: neither keyword matching nor tensor manipulation, even at scale, suffice in themselves to qualify for consciousness. - Gary > On Feb 14, 2022, at 00:24, Geoffrey Hinton > wrote: > > ?Many AI researchers are very unclear about what consciousness is and also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > -- [cid:image001.png at 01D82444.67C0E320] -- [cid:image001.png at 01D82444.67C0E320] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 19957 bytes Desc: image001.png URL: From wduch at umk.pl Thu Feb 17 15:15:03 2022 From: wduch at umk.pl (Wlodzislaw Duch) Date: Thu, 17 Feb 2022 21:15:03 +0100 Subject: Connectionists: AI and Neuroscience track, at the Third Polish Conferences on Artificial Intelligence. In-Reply-To: References: <4E206109-EE28-442B-BC3B-DFCEC2D344BD@nyu.edu> Message-ID: <0a3a0862-72fa-8d20-aa20-d37de8dcb923@umk.pl> An HTML attachment was scrubbed... URL: From minaiaa at gmail.com Fri Feb 18 02:47:16 2022 From: minaiaa at gmail.com (Ali Minai) Date: Fri, 18 Feb 2022 02:47:16 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: Jerry Coyne recently did a very nice post on an essay by Massimo Pigliucci on the topic of free-will. Also links to the remarkable recent paper by Maoz et al on the difference between random and deliberate choice and its relationship to readiness potential experiments. https://whyevolutionistrue.com/2022/02/09/massimo-pigliucci-free-will-is-incoherent/?fbclid=IwAR1-U_Z34z3-3Uu0L7siMTzyj6Q0v7ey77Gho40V9OW618cPRzEZhR13qSk I have to agree with Pigliucci and Coyne: Free-will in the conventional sense is logically incoherent. However, a lot probably remains to be discovered about volition. Ali *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Fri, Feb 18, 2022 at 2:20 AM Asim Roy wrote: > In 1998, after our debate about the brain at the WCCI in Anchorage, > Alaska, I asked Walter Freeman if he thought the brain controls the body. > His answer was, you can also say that the body controls the brain. I then > asked him if the driver controls a car, or the pilot controls an airplane. > His answer was the same, that you can also say that the car controls the > driver, or the plane controls the pilot. I then realized that Walter was > also a philosopher and believed in the No-free Will theory and what he was > arguing for is that the world is simply made of interacting systems. > However, both Walter, and his close friend John Taylor, were into > consciousness. > > > > I have argued with Walter on many different topics over nearly two decades > and have utmost respect for him as a scholar, but this first argument I > will always remember. > > > > Obviously, there?s a conflict between consciousness and the No-free Will > theory. Wonder where we stand with regard to this conflict. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Andras Lorincz > *Sent:* Tuesday, February 15, 2022 6:50 AM > *To:* Stephen Jos? Hanson ; Gary Marcus < > gary.marcus at nyu.edu> > *Cc:* Connectionists > *Subject:* Re: Connectionists: Weird beliefs about consciousness > > > > Dear Steve and Gary: > > This is how I see (try to understand) consciousness and the related terms: > > (Our) consciousness seems to be related to the close-to-deterministic > nature of the episodes on from few hundred millisecond to a few second > domain. Control instructions may leave our brain 200 ms earlier than the > action starts and they become conscious only by that time. In addition, > observations of those may also be delayed by a similar amount. (It then > follows that the launching of the control actions is not conscious and -- > therefore -- free will can be debated in this very limited context.) On the > other hand, model-based synchronization is necessary for timely > observation, planning, decision making, and execution in a distributed and > slow computational system. If this model-based synchronization is not > working properly, then the observation of the world breaks and > schizophrenic symptoms appear. As an example, individuals with pronounced > schizotypal traits are particularly successful in self-tickling (source: > https://philpapers.org/rec/LEMIWP > , > and a discussion on Asperger and schizophrenia: > https://www.frontiersin.org/articles/10.3389/fpsyt.2020.503462/full > ) > a manifestation of improper binding. The internal model enables and the > synchronization requires the internal model and thus a certain level of > consciousness can appear in a time interval around the actual time instant > and its length depends on the short-term memory. > > Other issues, like separating the self from the rest of the world are more > closely related to the soft/hard style interventions (as called in the > recent deep learning literature), i.e., those components (features) that > can be modified/controlled, e.g., color and speed, and the ones that are > Lego-like and can be separated/amputed/occluded/added. > > Best, > > Andras > > > > ------------------------------------ > > Andras Lorincz > > http://nipg.inf.elte.hu/ > > > Fellow of the European Association for Artificial Intelligence > > https://scholar.google.com/citations?user=EjETXQkAAAAJ&hl=en > > > Department of Artificial Intelligence > > Faculty of Informatics > > Eotvos Lorand University > > Budapest, Hungary > > > > > > > ------------------------------ > > *From:* Connectionists on > behalf of Stephen Jos? Hanson > *Sent:* Monday, February 14, 2022 8:30 PM > *To:* Gary Marcus > *Cc:* Connectionists > *Subject:* Re: Connectionists: Weird beliefs about consciousness > > > > Gary, these weren't criterion. Let me try again. > > I wasn't talking about wake-sleep cycles... I was talking about being > awake or asleep and the transition that ensues.. > > Rooba's don't sleep.. they turn off, I have two of them. They turn on > once (1) their batteries are recharged (2) a timer has been set for being > turned on. > > GPT3 is essentially a CYC that actually works.. by reading Wikipedia > (which of course is a terribly biased sample). > > I was indicating the difference between implicit and explicit > learning/problem solving. Implicit learning/memory is unconscious and > similar to a habit.. (good or bad). > > I believe that when someone says "is gpt3 conscious?" they are asking: is > gpt3 self-aware? Roombas know about vacuuming and they are unconscious. > > S > > On 2/14/22 12:45 PM, Gary Marcus wrote: > > Stephen, > > > > On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more > plausible as a consciousness than GPT-3. > > > > 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. > > 2. Roomba makes choices based on an explicit representation of its > location relative to a mapped space. GPT lacks any consistent reflection of > self; eg if you ask it, as I have, if you are you person, and then ask if > it is a computer, it?s liable to say yes to both, showing no stable > knowledge of self. > > 3. Roomba has explicit, declarative knowledge eg of walls and other > boundaries, as well its own location. GPT has no systematically > interrogable explicit representations. > > > > All this is said with tongue lodged partway in cheek, but I honestly don?t > see what criterion would lead anyone to believe that GPT is a more > plausible candidate for consciousness than any other AI program out there. > > > > ELIZA long ago showed that you could produce fluent speech that was mildly > contextually relevant, and even convincing to the untutored; just because > GPT is a better version of that trick doesn?t mean it?s any more conscious. > > > > Gary > > > > On Feb 14, 2022, at 08:56, Stephen Jos? Hanson > wrote: > > ? > > this is a great list of behavior.. > > Some biologically might be termed reflexive, taxes, classically > conditioned, implicit (memory/learning)... all however would not be > conscious in the several senses: (1) wakefulness-- sleep (2) self aware > (3) explicit/declarative. > > I think the term is used very loosely, and I believe what GPT3 and other > AI are hoping to show signs of is "self-awareness".. > > In response to : "why are you doing that?", "What are you doing now", > "what will you be doing in 2030?" > > Steve > > > > On 2/14/22 10:46 AM, Iam Palatnik wrote: > > A somewhat related question, just out of curiosity. > > > > Imagine the following: > > > > - An automatic solar panel that tracks the position of the sun. > > - A group of single celled microbes with phototaxis that follow the > sunlight. > > - A jellyfish (animal without a brain) that follows/avoids the sunlight. > > - A cockroach (animal with a brain) that avoids the sunlight. > > - A drone with onboard AI that flies to regions of more intense sunlight > to recharge its batteries. > > - A human that dislikes sunlight and actively avoids it. > > > > Can any of these, beside the human, be said to be aware or conscious of > the sunlight, and why? > > What is most relevant? Being a biological life form, having a brain, being > able to make decisions based on the environment? Being taxonomically close > to humans? > > > > > > > > > > > > > > > > On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: > > Also true: Many AI researchers are very unclear about what consciousness > is and also very sure that ELIZA doesn?t have it. > > Neither ELIZA nor GPT-3 have > - anything remotely related to embodiment > - any capacity to reflect upon themselves > > Hypothesis: neither keyword matching nor tensor manipulation, even at > scale, suffice in themselves to qualify for consciousness. > > - Gary > > > On Feb 14, 2022, at 00:24, Geoffrey Hinton > wrote: > > > > ?Many AI researchers are very unclear about what consciousness is and > also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > > > > > -- > > -- > -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 19957 bytes Desc: not available URL: From EPNSugan at ntu.edu.sg Fri Feb 18 04:13:59 2022 From: EPNSugan at ntu.edu.sg (Ponnuthurai Nagaratnam Suganthan) Date: Fri, 18 Feb 2022 09:13:59 +0000 Subject: Connectionists: (Impact Factor=6.2) EAAI Journal Special Issue on "randomization-based learning methods" Message-ID: Call for Papers Journal: Engineering Applications of Artificial Intelligence, Elsevier (IF: 6.2) Special Issue Title: Randomized Deep and Shallow Learning Algorithms We welcome submissions on non-randomized algorithms with significant comparisons against randomized algorithms too. Aim and Scope: Randomization-based learning algorithms have received considerable attention from academics, researchers, and domain workers because randomization-based neural networks can be trained by non-iterative approaches possessing closed-form solutions. Those methods are in general computationally faster than iterative solutions and less sensitive to parameter settings. Even though randomization-based non-iterative methods have attracted much attention in recent years, their deep structures have not been sufficiently developed nor benchmarked. This special issue aims to bridge this gap. The first target of this special issue is to present the recent advances of randomization- based learning methods. Randomization based neural networks usually offer non-iterative closed form solutions. Secondly, the focus is on promoting the concepts of non-iterative optimization with respect to counterparts, such as gradient-based methods and derivative-free iterative optimization techniques. Besides the dissemination of the latest research results on randomization-based and/or non-iterative algorithms, it is also expected that this special issue will cover some practical applications, present some new ideas and identify directions for future studies. Original contributions as well as comparative studies among randomization-based and non-randomized-based methods are welcome with unbiased literature review and comparative studies. Typical deep/shallow paradigms include (but not limited to) random vector functional link (RVFL), echo state networks (ESN), liquid state networks (LSN), kernel ridge regression (KRR) with randomization, extreme learning machines (ELM), random forests (RF), CNN with randomization, broad learning system (BLS), stochastic configuration network (SCN) and so on. All contributions must include sufficient application contents. Topics: The topics of the special issue include (with randomization-based methods), but are not limited to: l Randomized convolutional neural networks l Randomized internal representation learning l Regression, classification and time series analysis by randomization-based methods l Kernel methods such as kernel ridge regression, kernel adaptive filters, etc. with randomization l Feedforward, recurrent, multilayer, deep and other structures with randomization l Ensemble learning with randomization l Moore-Penrose pseudo inverse, SVD and other solution procedures. l Gaussian process regression l Randomization-based methods using novel fuzzy approaches l Randomization-based methods for large-scale problems with and without kernels l Theoretical analysis of randomization-based methods l Comparative studies with competing methods without randomization l Applications of randomized methods and information fusion in areas such as power systems, biomedical, finance, economics, signal processing, big data and all other relevant areas Submission Guideline: Papers should be submitted using EAAI?s online submission system: https://www.journals.elsevier.com/engineering-applications-of-artificial-intelligence. When submitting your manuscript please select the article type "SI: Randomization-based learning algorithms". Important Dates ? Manuscript submission due: April 30, 2022 ? First review completed: July 15, 2022 ? Revised manuscript due: Aug 15, 2022 ? Final decisions to authors: Oct 30, 2022 Guest Editors: Dr. P.N. Suganthan, IEEE Fellow School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore Email: epnsugan at ntu.edu.sg Website: https://www3.ntu.edu.sg/home/epnsugan/ Dr. M. Tanveer, IEEE Senior Member Department of Mathematics, Indian Institute of Technology Indore Email: mtanveer at iiti.ac.in Website: ?http://iiti.ac.in/people/~mtanveer/ Prof. Chin-Teng Lin, IEEE Fellow, IFSA Fellow Director, Computational Intelligence and Brain Computer Interface Centre Co-Director, Centre for AI (CAI) University of Technology Sydney, Australia Email: Chin-Teng.Lin at uts.edu.au ________________________________ CONFIDENTIALITY: This email is intended solely for the person(s) named and may be confidential and/or privileged. If you are not the intended recipient, please delete it, notify us and do not copy, use, or disclose its contents. Towards a sustainable earth: Print only when necessary. Thank you. -------------- next part -------------- An HTML attachment was scrubbed... URL: From gros at itp.uni-frankfurt.de Fri Feb 18 03:33:04 2022 From: gros at itp.uni-frankfurt.de (Claudius Gros) Date: Fri, 18 Feb 2022 09:33:04 +0100 Subject: Connectionists: consciousness debate on connectionist In-Reply-To: Message-ID: <1bb8e-620f5a00-1a7-5497f180@142079208> Hi Everybody I do not want to interfere with the interesting debate regarding consciousness. Just to point out, that there has been a recent Twitter discussion on the matter between Google and Facebook AI: https://twitter.com/ilyasut/status/1491554478243258368 Cheers, Claudius ### ### Prof. Dr. Claudius Gros ### http://itp.uni-frankfurt.de/~gros ### ### Complex and Adaptive Dynamical Systems, A Primer ### A graduate-level textbook, Springer (2008/10/13/15) ### ### Life for barren exoplanets: The Genesis project ### https://link.springer.com/article/10.1007/s10509-016-2911-0 ### From xtai at informatik.uni-wuerzburg.de Fri Feb 18 05:10:43 2022 From: xtai at informatik.uni-wuerzburg.de (Extended Artificial Intelligence Contact) Date: Fri, 18 Feb 2022 10:10:43 +0000 Subject: Connectionists: New MSc course: EXtended Artificial Intelligence (XtAI) Message-ID: <20220218101043.Horde.gTWxxhPTZlK_yK87xrue1gw@webmail.uni-wuerzburg.de> The University of W?rzburg, Germany, and the Center for Artificial Intelligence and Data Science (CAIDAS) now offer a new, specialized Master's course "eXtended Artificial Intelligence". The tuition-free course focuses on the design and development of complex AI systems, with an emphasis on combining AI with extended (virtual and augmented) reality. A brief overview of XtAI can be found in the flyer: https://go.uniwue.de/xtaiflyer For more detailed information and requirements, please visit: https://go.uniwue.de/xtai The application period for winter semester 2022/23 has already started and runs until the 15th of March. Forwarding this message to interested students is much appreciated. Kind regards, XtAI From a.crimi at sanoscience.org Fri Feb 18 06:41:57 2022 From: a.crimi at sanoscience.org (Alex Crimi) Date: Fri, 18 Feb 2022 11:41:57 +0000 Subject: Connectionists: Post-doc positions at BAMlab In-Reply-To: References: Message-ID: Dear all, We are looking for post-docs working in a brand new vibrant lab focused on neuroimaging https://bam.sano.science/ More specifically, we are looking for talented graduated PHD (or close to graduation) working on functional and structural connectivity from MRI, EEG and fNIRS in daily use, as investigating changes in daily use allows to see neurological, psychiatric effects in a broader sense. Apply even if dates on call is in the past, the positions will be filled when suitable candidates are found: https://sano.science/job-offers/eeg-fnirs-brain-analysis-in-real-world/ Required background and skills of the candidate: * recent or pending PhD in relevant field of science (computer science/biomedical engineering or related fields); * at least one first-author research publication in a peer-reviewed scientific journal, top conference or currently in press (e.g. Neuroimage, Nature Methods, Nature Communications, Nature Scientific Reports, PloS, IEEE TMI, ?); * Python programming, medical imaging experience (preferably in 3D or 4D), statistics; * knowledge of time series signal as fMRI, EEG or fNIRS; * excellent written and oral English communication skills; * additional assets: knowledge of brain anatomy and related tools (Dipy, FSL, etc), experince in circuit design or electronics DIY (Arduino, RaspberryPi, Jetson, etc). We offer a fixed term contract for 40 hours per week for the duration of 2 years. This will be supported by an educational plan that includes attendance of courses and (international) meetings. The contract will include opportunities to participate in teaching and supervision of undergraduate and master students. The salary, depending on relevant experience before the beginning of the employment contract, will be up to 12.000-14.000 PLN (2.6-3.1k EUR) gross per month, based on a full-time contract (40 hours a week) for the duration of 2 years with private medical care and a sports card. Sano offers excellent opportunities for study and development, an access to many international conferences on computational medicine and a possibility to grow in a scientific society. For scientific inquiries contact a.crimi at sanoscience.org For organizational inquiries contact e.sokalska at sanoscience.org ----------------------------------------------------------------------------------------- I send emails at unusual times, I don't expect immediate response ----------------------------------------------------------------------------------------- Dr. Alessandro Crimi Research Group Leader https://bam.sano.science [cid:4e3c8bf0-3dae-4fce-be13-cd9e605c5094] Centre for Computational Medicine Czarnowiejska 36, building C5. 30-072 Krak?w, Poland Phone: +48 575 455 016 www.sano.science ----------------------------------------------- Visiting Lecturer African Institute for Mathematical Sciences www.aims.edu.gh --------------------------------------------------------- https://twitter.com/Dr_Alex_Crimi -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: Outlook-i4ajhvir.png Type: image/png Size: 12229 bytes Desc: Outlook-i4ajhvir.png URL: From wduch at umk.pl Fri Feb 18 06:43:21 2022 From: wduch at umk.pl (Wlodzislaw Duch) Date: Fri, 18 Feb 2022 12:43:21 +0100 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: An HTML attachment was scrubbed... URL: From ioannakoroni at csd.auth.gr Fri Feb 18 08:03:11 2022 From: ioannakoroni at csd.auth.gr (Ioanna Koroni) Date: Fri, 18 Feb 2022 15:03:11 +0200 Subject: Connectionists: =?utf-8?q?Live_AIDA_e-Lecture_by_Prof=2E_Fredrik_?= =?utf-8?q?Heintz=3A_=E2=80=9CTowards_Trustworthy_AI_=E2=80=93_Inte?= =?utf-8?q?grating_Reasoning_and_Learning=E2=80=9D=2C_22nd_February?= =?utf-8?q?_2022_17=3A00-18=3A00_CET?= References: <003101d8182b$bea340e0$3be9c2a0$@csd.auth.gr> Message-ID: <038501d824c7$e27c7a10$a7756e30$@csd.auth.gr> Dear AI scientist/engineer/student/enthusiast, Prof. Fredrik Heintz, a prominent AI researcher internationally, will deliver the e-lecture: ?Towards Trustworthy AI - Integrating Reasoning and Learning?, on Tuesday 22nd February 2022 17:00-18:00 CET (8:00-9:00 am PST), (12:00 am-1:00am CST), see details in: http://www.i-aida.org/event_cat/ai-lectures/ You can join for free using the zoom link: https://authgr.zoom.us/j/91262043831 & Passcode: 148148 The International AI Doctoral Academy (AIDA), a joint initiative of the European R&D projects AI4Media, ELISE, Humane AI Net, TAILOR and VISION, is very pleased to offer you top quality scientific lectures on several current hot AI topics. Lectures are typically held once per week, Tuesdays 17:00-18:00 CET (8:00-9:00 am PST), (12:00 am-1:00am CST). Attendance is free. The lectures are disseminated through multiple channels and email lists (we apologize if you received it through various channels). If you want to stay informed on future lectures, you can register in the email lists AIDA email list and CVML email list. Best regards Profs. M. Chetouani, P. Flach, B. O?Sullivan, I. Pitas, N. Sebe -- This email has been checked for viruses by Avast antivirus software. https://www.avast.com/antivirus -------------- next part -------------- An HTML attachment was scrubbed... URL: From george at cs.ucy.ac.cy Fri Feb 18 09:02:34 2022 From: george at cs.ucy.ac.cy (George A. Papadopoulos) Date: Fri, 18 Feb 2022 16:02:34 +0200 Subject: Connectionists: ACM International Conference on Information Technology for Social Good (GoodIT 2022): Fourth Call for Contributions Message-ID: *** Fourth Call for Contributions *** ACM International Conference on Information Technology for Social Good (GoodIT 2022) 7?9 September, 2022, 5* St. Raphael Resort & Marina, Limassol, Cyprus https://cyprusconferences.org/goodit2022/ Scope The ACM GoodIT conference seeks papers describing significant research contributions related to the application of information technologies (IT) to social good. Social good is typically defined as something that provides a benefit to the general public. In this context, clean air and water, Internet connection, education, and healthcare are all good examples of social goods. However, new media innovations and the explosion of online communities have added new meaning to the term. Social good is now about global citizens uniting to unlock the potential of individuals, technology, and collaboration to create a positive societal impact. GoodIT solicits papers that address important research challenges related to, but not limited to: ? Citizen science ? Civic intelligence ? Decentralized approaches to IT ? Digital solutions for Cultural Heritage ? Environmental monitoring ? Ethical computing ? Frugal solutions for IT ? Game, entertainment, and multimedia applications ? Health and social care ? IT for automotive ? IT for development ? IT for education ? IT for smart living ? Privacy, trust and ethical issues in ICT solutions ? Smart governance and e-administration ? Social informatics ? Socially responsible IT solutions ? Sustainable cities and transportation ? Sustainable IT ? Technology addressing the digital divide Paper Submission The papers should not exceed six (6) pages (US letter size) double-column, including figures, tables, and references in standard ACM format (https://cyprusconferences.org/goodit2022/index.php/authors/ ). They must be original works and must not have been previously published. At least one of the authors of all accepted papers must register and present the work at the conference; otherwise, the paper will not be published in the proceedings. All accepted and presented papers will be included in the conference proceedings published in the ACM Digital Library. In addition to a Main Track, the conference features 9 additional Special Tracks. More information on these tracks is available on the conference web site. Journal Special Issues and Best Paper Award Selected papers will be invited to submit an extended version to two special journal issues: ? FGCS, Elsevier ? MDPI Sensors, where the theme of the special issue will be "Application of Information Technology (IT) to Social Good" (https://www.mdpi.com/journal/sensors/special_issues/topical_collection_goodit ). Specifically 5 papers will be invited free of charge and another 5 papers will get a 20% discount on the publication fees. Furthermore, MDPI Sensors will sponsor a Best Paper Award with the amount of 400 CHF. Work-in-Progress and PhD Track Inside ACM GoodIT, the Work-in-Progress and PhD Track provides an opportunity to showcase interesting new work that is still at an early stage. We encourage practitioners and researchers to submit to the Work-in-Progress venue as it provides a unique opportunity for sharing valuable ideas, eliciting feedback on early-stage work, and fostering discussions and collaborations among colleagues. Moreover, this track provides a platform for PhD students to present and receive feedback on their ongoing research. Students at different stages of their research will have the opportunity to present and discuss their research questions, goals, methods and results. This is an opportunity to obtain guidance on various aspects of their research from established researchers and other PhD students working in research areas related to technologies for social good. Important: For this specific track, papers must not exceed four (4) pages (US letter size) double column, including figures, tables, and references in standard ACM format (https://cyprusconferences.org/goodit2022/index.php/authors/ ). Submission Instructions All papers must be submitted electronically via the hotcrp web site (https://goodit2022.hotcrp.com ). Once on the submission page, you will be able to select the track where to submit your paper. Important Dates ? Submission deadline for all types of contributions: 23 May 2022 ? Notification of acceptance: 20 June 2022 ? Camera-ready submission and author registration: 11 July 2022 Conference Committees https://cyprusconferences.org/goodit2022/committees/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From wduch at umk.pl Fri Feb 18 12:45:07 2022 From: wduch at umk.pl (Wlodzislaw Duch) Date: Fri, 18 Feb 2022 18:45:07 +0100 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: An HTML attachment was scrubbed... URL: From minaiaa at gmail.com Fri Feb 18 10:36:59 2022 From: minaiaa at gmail.com (Ali Minai) Date: Fri, 18 Feb 2022 10:36:59 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: Wlodek I think that the debate about consciousness in the strong sense of having a conscious experience like we do is sterile. We will never have a measuring device for whether another entity is "conscious", and at some point, we will get to an AI that is sufficiently complex in its observable behavior that we will either accept its inner state of consciousness on trust - just as we do with humans and other animals - or admit that we will never believe that a machine that is "not like us" can ever be conscious. The "like us" part is more important than many of us in the AI field think: A big part of why we believe other humans and our dogs are conscious is because we know that they are "like us", and assume that they must share our capacity for inner conscious experience. We already see this at a superficial level where, as ordinary humans, we have a much easier time identifying with an embodied, humanoid AI like Wall-E or the Terminator than with a disembodied one like Hal or Skynet. This is also why so many people find the Boston Dynamics "dog" so disconcerting. The question of embodiment is a complex one, as you know, of course, but I am with those who think that it is necessary for grounding mental representations - that it is the only way that the internal representations of the system are linked directly to its experience. For example, if an AI system trained only on text (like GPT-3) comes to learn that touching something hot results in the fact of getting burned, we cannot accept that as sufficient because it is based only on the juxtaposition of abstractions, not the actual painful experience of getting burned. For that, you need a body with sensors and a brain with a state corresponding to pain - something that can be done in an embodied robot. This is why I think that all language systems trained purely on the assumption of the distributional hypothesis of meaning will remain superficial; they lack the grounding that can only be supplied by experience. This does not mean that systems based on the distributional hypothesis cannot learn a lot, or even develop brain-like representations, as the following extremely interesting paper shows: Y. Zhang, K. Han, R. Worth, and Z. Liu. Connecting concepts in the brain by mapping cortical representations of semantic relations. Nature Communications, 11(1):1877, Apr 2020. In a formal sense, however, embodiment could be in any space, including very abstract ones. We can think of text data as GPT-3's world and, in that world, it is "embodied" and its fundamentally distributional learning, though superficial and lacking in experience to us, is grounded for it within its world. Of course, this is not a very useful view of embodiment and grounding since we want to create AI that is grounded in our sense, but one of the most under-appreciated risks of AI is that, as we develop systems that live in worlds very different than ours, they will - implicitly and emergently - embody values completely alien to us. The proverbial loan-processing AI that learns to be racially biased in just a caricature of this hazard, but one that should alert us to deeper issues. Our quaintly positivistic and reductionistic notion that we can deal with such things by removing biases from data, algorithms, etc., is misplaced. The world is too complicated for that. Ali *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Fri, Feb 18, 2022 at 7:27 AM Wlodzislaw Duch wrote: > Asim, > > I was on the Anchorage panel, and asked others what could be a great > achievement in computational intelligence. Steve Grossberg replied, that > symbolic AI is meaningless, but creation of artificial rat that could > survive in hostile environment would be something. Of course this is still > difficult, but perhaps DARPA autonomous machines are not that far? > > I also had similar discussions with Walter and support his position: you > cannot separate tightly coupled systems. Any external influence will create > activation in both, linear causality looses its meaning. This is clear if > both systems adjust to each other. But even if only one system learns > (brain) and the other is mechanical but responds to human actions it may > behave as one system. Every musician knows that: piano becomes a part of > our body, responding in so many ways to actions, not only by producing > sounds but also providing haptic feedback. > > This simply means that brains of locked-in people worked in somehow > different way than brains of healthy people. Why do we consider them > conscious? Because they can reflect on their mind states, imagine things > and describe their inner states. If GPT-3 was coupled with something like > DALL-E that creates images from text, and could describe what they see in > their inner world, create some kind of episodic memory, we would have hard > time to deny that this thing is not conscious of what it has in its mind. > Embodiment helps to create inner world and changes it, but it is not > necessary for consciousness. Can we find a good argument that such system > is not conscious of its own states? It may not have all qualities of human > consciousness, but that is a matter of more detailed approximation of > missing functions. > > I have made this argument a long time ago (ex. in "*Brain-inspired > conscious computing architecture" *written over 20 years ago, see more > papers on this on my web page). > > Wlodek > > Prof. W?odzis?aw Duch > Fellow, International Neural Network Society > Past President, European Neural Network Society > Head, Neurocognitive Laboratory, CMIT NCU, Poland > > Google: Wlodzislaw Duch > > On 18/02/2022 05:22, Asim Roy wrote: > > In 1998, after our debate about the brain at the WCCI in Anchorage, > Alaska, I asked Walter Freeman if he thought the brain controls the body. > His answer was, you can also say that the body controls the brain. I then > asked him if the driver controls a car, or the pilot controls an airplane. > His answer was the same, that you can also say that the car controls the > driver, or the plane controls the pilot. I then realized that Walter was > also a philosopher and believed in the No-free Will theory and what he was > arguing for is that the world is simply made of interacting systems. > However, both Walter, and his close friend John Taylor, were into > consciousness. > > > > I have argued with Walter on many different topics over nearly two decades > and have utmost respect for him as a scholar, but this first argument I > will always remember. > > > > Obviously, there?s a conflict between consciousness and the No-free Will > theory. Wonder where we stand with regard to this conflict. > > > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists > *On Behalf Of *Andras > Lorincz > *Sent:* Tuesday, February 15, 2022 6:50 AM > *To:* Stephen Jos? Hanson > ; Gary Marcus > > *Cc:* Connectionists > > *Subject:* Re: Connectionists: Weird beliefs about consciousness > > > > Dear Steve and Gary: > > This is how I see (try to understand) consciousness and the related terms: > > (Our) consciousness seems to be related to the close-to-deterministic > nature of the episodes on from few hundred millisecond to a few second > domain. Control instructions may leave our brain 200 ms earlier than the > action starts and they become conscious only by that time. In addition, > observations of those may also be delayed by a similar amount. (It then > follows that the launching of the control actions is not conscious and -- > therefore -- free will can be debated in this very limited context.) On the > other hand, model-based synchronization is necessary for timely > observation, planning, decision making, and execution in a distributed and > slow computational system. If this model-based synchronization is not > working properly, then the observation of the world breaks and > schizophrenic symptoms appear. As an example, individuals with pronounced > schizotypal traits are particularly successful in self-tickling (source: > https://philpapers.org/rec/LEMIWP > , > and a discussion on Asperger and schizophrenia: > https://www.frontiersin.org/articles/10.3389/fpsyt.2020.503462/full > ) > a manifestation of improper binding. The internal model enables and the > synchronization requires the internal model and thus a certain level of > consciousness can appear in a time interval around the actual time instant > and its length depends on the short-term memory. > > Other issues, like separating the self from the rest of the world are more > closely related to the soft/hard style interventions (as called in the > recent deep learning literature), i.e., those components (features) that > can be modified/controlled, e.g., color and speed, and the ones that are > Lego-like and can be separated/amputed/occluded/added. > > Best, > > Andras > > > > ------------------------------------ > > Andras Lorincz > > http://nipg.inf.elte.hu/ > > > Fellow of the European Association for Artificial Intelligence > > https://scholar.google.com/citations?user=EjETXQkAAAAJ&hl=en > > > Department of Artificial Intelligence > > Faculty of Informatics > > Eotvos Lorand University > > Budapest, Hungary > > > > > > > ------------------------------ > > *From:* Connectionists on > behalf of Stephen Jos? Hanson > *Sent:* Monday, February 14, 2022 8:30 PM > *To:* Gary Marcus > *Cc:* Connectionists > *Subject:* Re: Connectionists: Weird beliefs about consciousness > > > > Gary, these weren't criterion. Let me try again. > > I wasn't talking about wake-sleep cycles... I was talking about being > awake or asleep and the transition that ensues.. > > Rooba's don't sleep.. they turn off, I have two of them. They turn on > once (1) their batteries are recharged (2) a timer has been set for being > turned on. > > GPT3 is essentially a CYC that actually works.. by reading Wikipedia > (which of course is a terribly biased sample). > > I was indicating the difference between implicit and explicit > learning/problem solving. Implicit learning/memory is unconscious and > similar to a habit.. (good or bad). > > I believe that when someone says "is gpt3 conscious?" they are asking: is > gpt3 self-aware? Roombas know about vacuuming and they are unconscious. > > S > > On 2/14/22 12:45 PM, Gary Marcus wrote: > > Stephen, > > > > On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more > plausible as a consciousness than GPT-3. > > > > 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. > > 2. Roomba makes choices based on an explicit representation of its > location relative to a mapped space. GPT lacks any consistent reflection of > self; eg if you ask it, as I have, if you are you person, and then ask if > it is a computer, it?s liable to say yes to both, showing no stable > knowledge of self. > > 3. Roomba has explicit, declarative knowledge eg of walls and other > boundaries, as well its own location. GPT has no systematically > interrogable explicit representations. > > > > All this is said with tongue lodged partway in cheek, but I honestly don?t > see what criterion would lead anyone to believe that GPT is a more > plausible candidate for consciousness than any other AI program out there. > > > > ELIZA long ago showed that you could produce fluent speech that was mildly > contextually relevant, and even convincing to the untutored; just because > GPT is a better version of that trick doesn?t mean it?s any more conscious. > > > > Gary > > > > On Feb 14, 2022, at 08:56, Stephen Jos? Hanson > wrote: > > ? > > this is a great list of behavior.. > > Some biologically might be termed reflexive, taxes, classically > conditioned, implicit (memory/learning)... all however would not be > conscious in the several senses: (1) wakefulness-- sleep (2) self aware > (3) explicit/declarative. > > I think the term is used very loosely, and I believe what GPT3 and other > AI are hoping to show signs of is "self-awareness".. > > In response to : "why are you doing that?", "What are you doing now", > "what will you be doing in 2030?" > > Steve > > > > On 2/14/22 10:46 AM, Iam Palatnik wrote: > > A somewhat related question, just out of curiosity. > > > > Imagine the following: > > > > - An automatic solar panel that tracks the position of the sun. > > - A group of single celled microbes with phototaxis that follow the > sunlight. > > - A jellyfish (animal without a brain) that follows/avoids the sunlight. > > - A cockroach (animal with a brain) that avoids the sunlight. > > - A drone with onboard AI that flies to regions of more intense sunlight > to recharge its batteries. > > - A human that dislikes sunlight and actively avoids it. > > > > Can any of these, beside the human, be said to be aware or conscious of > the sunlight, and why? > > What is most relevant? Being a biological life form, having a brain, being > able to make decisions based on the environment? Being taxonomically close > to humans? > > > > > > > > > > > > > > > > On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: > > Also true: Many AI researchers are very unclear about what consciousness > is and also very sure that ELIZA doesn?t have it. > > Neither ELIZA nor GPT-3 have > - anything remotely related to embodiment > - any capacity to reflect upon themselves > > Hypothesis: neither keyword matching nor tensor manipulation, even at > scale, suffice in themselves to qualify for consciousness. > > - Gary > > > On Feb 14, 2022, at 00:24, Geoffrey Hinton > wrote: > > > > ?Many AI researchers are very unclear about what consciousness is and > also very sure that GPT-3 doesn?t have it. It?s a strange combination. > > > > > > -- > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From m.lengyel at eng.cam.ac.uk Fri Feb 18 09:58:43 2022 From: m.lengyel at eng.cam.ac.uk (=?utf-8?B?TcOhdMOpIExlbmd5ZWw=?=) Date: Fri, 18 Feb 2022 14:58:43 +0000 Subject: Connectionists: postdoc in Computational Neuroscience, Cambridge-Columbia collaboration Message-ID: Postdoc in Computational Neuroscience, Cambridge-Columbia collaboration We are seeking a highly creative and motivated postdoctoral fellow to work on a collaborative project between the labs of Guillaume Hennequin and M?t? Lengyel at the Computational and Biological Learning Lab, Department of Engineering, University of Cambridge, and Daniel Wolpert at the Zuckerman Mind Brain Behavior Institute, Columbia University. The project studies neural network mechanisms underlying the context-dependent (continual) learning of motor repertoires and ties together several threads of research recently developed in our labs, based on the following key publications (note that several are collaborative between our groups): - JB Heald, M Lengyel, DM Wolpert (2021) Contextual inference underlies the learning of sensorimotor repertoires. Nature 600:489-493. - TC Kao, MS Sadabadi, G Hennequin (2021). Optimal anticipatory control as a theory of motor preparation: a thalamo-cortical circuit model. Neuron 109:1567-1581. - TC Kao, KT Jensen, GM van de Ven, A Bernacchia, G Hennequin (2021) Natural continual learning: success is a journey, not (just) a destination. NeurIPS, https://tinyurl.com/2jfyss8c - Echeveste R, Aitchison L, Hennequin G, Lengyel M (2020) Cortical-like dynamics in recurrent circuits optimized for sampling-based probabilistic inference. Nature Neuroscience 23:1138-1149. The postdoc will be based in Cambridge, with an opportunity for regular visits to Columbia. The successful candidate will have - a strong quantitative background - demonstrable interest in theoretical neuroscience - obtained (or be close to the completion of) a PhD or equivalent in computational neuroscience, physics, mathematics, computer science, machine learning or a related field Preference will be given to candidates with - previous experience in computational neuroscience, especially with the dynamics of recurrent neural network, and function-optimized neural networks - sufficient programming skills to run numerical simulations and to use large scale optimization packages - expertise with advanced data analysis and Bayesian techniques For further details and to send your application, please go to https://www.jobs.cam.ac.uk/job/33691/. For informal queries, please contact Guillaume Hennequin , M?t? Lengyel or Daniel Wolpert . M?t? Lengyel -- Professor of Computational Neuroscience Computational and Biological Learning Lab Cambridge University Engineering Department Trumpington Street, Cambridge CB2 1PZ, UK tel: +44 (0)1223 748 532, fax: +44 (0)1223 765 587 email: m.lengyel at eng.cam.ac.uk web: lengyellab.org From gangluo at cs.wisc.edu Fri Feb 18 15:52:21 2022 From: gangluo at cs.wisc.edu (GANG LUO) Date: Fri, 18 Feb 2022 20:52:21 +0000 Subject: Connectionists: Call for Papers - International Workshop on Data Management and Analytics for Medicine and Healthcare (DMAH 2022) Message-ID: -- Call for Papers -- The Eighth International Workshop on Data Management and Analytics for Medicine and Healthcare (DMAH 2022) In Conjunction with VLDB 2022 Sydney, Australia - virtual event September 9, 2022 https://sites.google.com/view/vldbdmah2022/ The format of the workshop is to be decided, but a virtual format will be provided for remote participation. Healthcare enterprises are producing large amounts of data through electronic medical records, medical imaging, health insurance claims, surveillance, and others. Such data have high potential to transform current healthcare to improve healthcare quality and prevent diseases, and advance biomedical research. Medical Informatics is an interdisciplinary field that studies and pursues the effective use of medical data, information, and knowledge for scientific inquiry, problem solving and decision making, driven by efforts to improve human health and well being. The goal of the workshop is to bring people in the field cross-cutting information management and medical informatics to discuss innovative data management and analytics technologies highlighting end-to-end applications, systems, and methods to address problems in healthcare, public health, and everyday wellness, with clinical, physiological, imaging, behavioral, environmental, and omic- data, and data from social media and the Web. It will provide a unique opportunity for interaction between information management researchers and biomedical researchers for the interdisciplinary field. This workshop welcomes papers that address fundamental research issues for complex medical data environments, data management and analytical methods, systems and applications. Topics of interest include, but not limited to: Big data management for medical data; Blockchain for healthcare; Biomedical data integration; Biomedical knowledge management and decision support; Semantics and interoperability for healthcare data; Clinical natural language processing and text mining; Predictive modeling for diagnosis and treatment; Visual analytics for medical data; Medical image analytics; Data privacy and security for healthcare data; Hospital readmission analytics; Medical fraud detection; Social media and Web data analytics for public health (public health 2.0); Data analytics for pervasive computing for medical care. DMAH 2022 accept two types of papers: 1) Regular research papers reporting original research results or significant case studies (18 pages). 2) Short papers will be 6 pages. 3) Abstracts will be 2 pages Important Dates: Abstract (optional): May 10, 2022 Individual Workshop Papers: May 16, 2022 Notification of Acceptance: June 10, 2022 Camera Ready: June 25, 2022 Workshop date: September 9, 2022 All submitted papers will be rigorously reviewed. All accepted papers will be made available as a workshop proceedings to be published by Springer LNCS. Workshop Chairs: Chairs: Fusheng Wang, Stony Brook University, USA Gang Luo, University of Washington, USA Dejun Teng, Alibaba Inc., China Jun Kong, Georgia State University, USA From juyang.weng at gmail.com Fri Feb 18 21:21:09 2022 From: juyang.weng at gmail.com (Juyang Weng) Date: Fri, 18 Feb 2022 21:21:09 -0500 Subject: Connectionists: Weird beliefs about consciousness Message-ID: Dear Hava, Your wrote: " I would like to add that before consciousness is well defined it is meaningless to discuss whether it exists in artificial systems. We r scientists and require clarity." Do you have a chance to read my Conscious Learning paper: J. Weng, "3D-to-2D-to-3D Conscious Learning", in Proc. IEEE 40th International Conference on Consumer Electronics, pp. 1-6, Las Vegas NV, USA, Jan.7-9, 2022. PDF file . I drafted a paper about biological brain and natural conscious learning, timidly for Nature (but I am afraid that it will be rejected). Do you mind pre-reading it like a reviewer? If you are so kind, I will send it to you (for you to reject I am afraid). Best regards, -John -- Juyang (John) Weng -------------- next part -------------- An HTML attachment was scrubbed... URL: From george at cs.ucy.ac.cy Sat Feb 19 05:55:35 2022 From: george at cs.ucy.ac.cy (George A. Papadopoulos) Date: Sat, 19 Feb 2022 12:55:35 +0200 Subject: Connectionists: Second International Conference on ICT for Health, Accessibility and Wellbeing (IHAW 2022): First Call for Papers Message-ID: *** First Call for Papers *** Second International Conference on ICT for Health, Accessibility and Wellbeing (IHAW 2022) December 5-7, 2022, Golden Bay Beach Hotel 5*, Larnaca, Cyprus https://cyprusconferences.org/ihaw2022 ICT for Health, Accessibility and Wellbeing (IHAW 2022) is the second of the series of International Conferences on "ICT for Societal Challenges". It is a showcase for high quality oral and poster presentations and demonstrations sessions. This conference aims to be a platform for multi and interdisciplinary research at the interplay between Information and Communication Technologies, Biomedical, Neuro-cognitive, and Experimental research. This research includes the design, experimental evaluation and standardization of new ICT scalable systems and in-silico systems for new and future inclusive and sustainable technologies that benefit all: healthy people, people with disabilities or other impairments, people having chronic diseases, etc. User-centered design and innovation, new intuitive ways of human -computer interaction, and user acceptance are the topics of particular interest. Conference Topics Relevant topics include (but are not limited to) the following: Artificial Intelligence, Computation and Data Analytics ? Artificial Intelligence methods for medical device testing. ? Algorithms, methods and services for condition-specific intervention (e.g., diabetes, obesity, dementia, post cancer treatment, allergies, mental health). ? Algorithms, methods and services for predicting and monitoring infectious disease. ? Crowd-sourcing and social media analysis for predicting and monitoring infectious disease. ? Medical Data and/or Medical Image Analysis. ? Electronic Medical Records Analysis. ? Computational methods for medical devices. Human Computer Interaction and Cognition ? Human-Machine Interaction for healthcare and well-being. ? Cognitive Mechatronics for healthcare and well-being. ? Models for human-device interaction for medicine. ? Cobotics for healthcare and well-being. ? Model-based design and configuration tools for healthcare and well-being. Assistive Devices ? Precision medicine. ? ICT for in-silicon trials. ? Implantable medical devices. ? Multimodal assistive ICT devices to empower people with sensory, cognitive, motor, balance and spatial impairments. ICT & Wellbeing ? Age-friendly systems for active and healthy ageing (telepresence, robotics solutions, innovative solutions for independent living, innovative elderly care, integrated care, age-related risks prevention/detection). ? ICT systems to improve the quality of life and for daily life activities assistance (education, recreation, and nutrition). ? Smart living homes and wearables (Intelligent and personalized digital solutions for sustaining and extending healthy and independent living; personalized early risk detection and intervention). ? Smart Systems and services promoting access to the socio-economical and cultural environment. ? IoT and smart real-time surveillance systems for monitoring, auditing and control to prevent the spread of the pandemic. ? eHealth smart solutions in the fight against a COVID-19 like pandemic. ? IoT and Smart Healthcare systems with an environmentally friendly and sustainable footprint. Health Infrastructure and Healthcare Operation Services ? Distributed and connected digital healthcare services. ? IoT services for real-time monitoring of health data and status of patients and/or older adults. ? Wearable devices and IoT systems for remote monitoring of health data and status of patients and/or older adults. ? mHealth services and applications using mobile and wearable devices to collect community and clinical health data, and deliver healthcare information to practitioners, researchers and patients. ? Sustainable city environments for emergency health management. ? 5G and beyond for healthcare in sustainable smart cities. ? Wireless Sensor Networks for advanced smart healthcare in sustainable cities. Quality in Healthcare Systems ? New experimental validation methods with end-users. ? Systems and services for ensuring patient?s commitment to the medication schedule. ? Digital health systems and tools for health care professional training and workforce development. ? Communication systems and services improving the quality of patient and healthcare provider contact before, during and after admission. ? Methods and Technology for Improving the quality of services-oriented care delivery systems. ? Methods, Digital Tools and/or Services for inclusive-for-all healthcare systems. ? Co-Creation of healthcare systems for social well being of people with special needs, older adults and/or deprivileged or disadvantaged people. ? Systems for management of health and care (mental health, pain, neurological disorders, sight, hearing, balance, space awareness; sensory based physiological and psychological non-invasive measurements, preventive healthcare, m-healthcare, e-healthcare, integrated care, serious games, electronic health record, self- management, patient-centered systems for survivorship, palliation and/or end-of-life care). Privacy, Security & Standardization ? Standardization, certification, labelling, and communication issues (related to ageing well, to sensory impairment). ? Privacy and Security/Regulation compliant services in health care systems (e.g., HIPAA). ? Security and privacy of digital health systems and service. ? Socio-economic issues of smart healthcare in sustainable societies. ? Privacy, security and ethics in eHealth smart solutions and surveillance at scale in the fight against a COVID-19 pandemic. High-quality original submissions that address such future issues, show the design and evaluation in (near-) real scenarios, explain how to benchmark systems, and outline the education and training procedures for acquiring new perceptual skills while using such systems are welcome. Research and technical papers are expected to present significant and original contributions validated with the targeted end-users. Submissions should clearly state the progress beyond the existing state-of- the-art and the expected societal benefits of the developed technology. When possible, validate scenarios with the target user groups and well-identified technology readiness levels (https://en.wikipedia.org/wiki/Technology_readiness_level) should be at least outlined. Submissions We invite Research and Technical papers, up to 15 pages, describing original unpublished research, making a substantial contribution to the research field. All submissions will be reviewed by the Program Committee. As was the case for IHAW 2021, the proceedings of IHAW 2022 will be published by Springer in the Communications in Computer and Information Science (CCIS) series (https://www.springer.com/series/7899) and will be presented in the technical sessions of the conference. As was also the case for IHAW 2021, the authors of the best papers accepted for IHAW2022 will be invited to submit extended versions for a special issue in a high quality journal (currently under negotiation). Submissions of all types should be carefully formatted according to the Springer format for conference proceedings: https://www.springer.com/gp/computer-science/lncs/conference-proceedings-guidelines . The submission process will be handled through Easy Chair and the submission link is: https://easychair.org/my/conference?conf=icihaw2022 . Important Dates ? Submission Deadline: July 18, 2022 (AoE) ? Notification: September 26, 2022 ? Camera-Ready Submission Deadline: October 10, 2022 ? Author Registration Deadline: October 10, 2022 Organizers Honorary General Chair ? Edwige Pissaloux, University of Rouen Normandy, France General Chair ? George A. Papadopoulos, University of Cyprus, Cyprus Scientific Chair ? Achilleas Achilleos, Frederick University, Cyprus Scientific Vice-Chair ? Ramiro Velazquez, Universidad Panamericana, Mexico Publicity Chair ? Jessica Allingham, Lakehead University, Canada Finance Chair ? Petros Stratis, Easy Conferences LTD, Cyprus -------------- next part -------------- An HTML attachment was scrubbed... URL: From phitzler at googlemail.com Sun Feb 20 00:52:02 2022 From: phitzler at googlemail.com (Pascal Hitzler) Date: Sat, 19 Feb 2022 23:52:02 -0600 Subject: Connectionists: Neuro-Symbolic Artificial Intelligence: The State of the Art (book announcement) Message-ID: <200f4d4e-9463-1c77-1cec-3193a1af616f@googlemail.com> Pascal Hitzler, Md Kamruzzaman Sarker Neuro-Symbolic Artificial Intelligence: The State of the Art Frontiers in Artificial Intelligence and Applications Vol. 342 IOS Press, Amsterdam, 2022 ISBN 978-1-64368-244-0 (print) 978-1-64368-245-7 (online) Available here: https://www.iospress.com/catalog/books/neuro-symbolic-artificial-intelligence-the-state-of-the-art Table of Contents: https://ebooks.iospress.nl/ISBN/978-1-64368-244-0 Preface Frank van Harmelen Chapter 1. Neural-Symbolic Learning and Reasoning: A Survey and Interpretation Tarek R. Besold, Artur d?Avila Garcez, Sebastian Bader, Howard Bowman, Pedro Domingos, Pascal Hitzler, Kai-Uwe K?hnberger, Luis C. Lamb, Priscila Machado Vieira Lima, Leo de Penning, Gadi Pinkas, Hoifung Poon, Gerson Zaverucha Chapter 2. Symbolic Reasoning in Latent Space: Classical Planning as an Example Masataro Asai, Hiroshi Kajino, Alex Fukunaga, Christian Muise Chapter 3. Logic Meets Learning: From Aristotle to Neural Networks Vaishak Belle Chapter 4. Graph Reasoning Networks and Applications Qingxing Cao, Wentao Wan, Xiaodan Liang, Liang Lin Chapter 5. Answering Natural-Language Questions with Neuro-Symbolic Knowledge Bases Haitian Sun, Pat Verga, William W. Cohen Chapter 6. Tractable Boolean and Arithmetic Circuits Adnan Darwiche Chapter 7. Neuro-Symbolic AI = Neural + Logical + Probabilistic AI Robin Manhaeve, Giuseppe Marra, Thomas Demeester, Sebastijan Duman?i?, Angelika Kimmig, Luc De Raedt Chapter 8. A Constraint-Based Approach to Learning and Reasoning Michelangelo Diligenti, Francesco Giannini, Marco Gori, Marco Maggini, Giuseppe Marra Chapter 9. Spike-Based Symbolic Computations on Bit Strings and Numbers Ceca Krai?nikovi?, Wolfgang Maass, Robert Legenstein Chapter 10. Explainable Neuro-Symbolic Hierarchical Reinforcement LearningDaoming Lyu, Fangkai Yang, Hugh Kwon, Bo Liu, Wen Dong, Levent Yilmaz Chapter 11. Neuro-Symbolic Semantic Reasoning Bassem Makni, Monireh Ebrahimi, Dagmar Gromann, Aaron Eberhart Chapter 12. Learning Reasoning Strategies in End-to-End Differentiable Proving Pasquale Minervini, Sebastian Riedel, Pontus Stenetorp, Edward Grefenstette, Tim Rockt?schel Chapter 13. Generalizable Neuro-Symbolic Systems for Commonsense Question Answering Alessandro Oltramari, Jonathan Francis, Filip Ilievski, Kaixin Ma, Roshanak Mirzaee Chapter 14. Combining Probabilistic Logic and Deep Learning for Self-Supervised Learning Hoifung Poon, Hai Wang, Hunter Lang Chapter 15. Human-Centered Concept Explanations for Neural Networks Chih-Kuan Yeh, Been Kim, Pradeep Ravikumar Chapter 16. Abductive Learning Zhi-Hua Zhou, Yu-Xuan Huang Chapter 17. Logic Tensor Networks: Theory and Applications Luciano Serafini, Artur d?Avila Garcez, Samy Badreddine, Ivan Donadello, Michael Spranger, Federico Bianchi -- Pascal Hitzler Lloyd T. Smith Creativity in Engineering Chair Director, Center for AI and Data Science Kansas State University http://www.pascal-hitzler.de http://www.daselab.org http://www.semantic-web-journal.net From david at irdta.eu Sat Feb 19 09:47:46 2022 From: david at irdta.eu (David Silva - IRDTA) Date: Sat, 19 Feb 2022 15:47:46 +0100 (CET) Subject: Connectionists: DeepLearn 2022 Spring: early registration March 16 Message-ID: <906615605.747330.1645282066336@webmail.strato.com> ****************************************************************** 5th INTERNATIONAL SCHOOL ON DEEP LEARNING DeepLearn 2022 Spring Guimar?es, Portugal April 18-22, 2022 https://irdta.eu/deeplearn/2022sp/ ***************** Co-organized by: Algoritmi Center University of Minho, Guimar?es Institute for Research Development, Training and Advice ? IRDTA Brussels/London ****************************************************************** Early registration: March 16, 2022 ****************************************************************** SCOPE: DeepLearn 2022 Spring will be a research training event with a global scope aiming at updating participants on the most recent advances in the critical and fast developing area of deep learning. Previous events were held in Bilbao, Genova, Warsaw, Las Palmas de Gran Canaria, and Bournemouth. Deep learning is a branch of artificial intelligence covering a spectrum of current frontier research and industrial innovation that provides more efficient algorithms to deal with large-scale data in a huge variety of environments: computer vision, neurosciences, speech recognition, language processing, human-computer interaction, drug discovery, biomedical informatics, image analysis, recommender systems, advertising, fraud detection, robotics, games, finance, biotechnology, physics experiments, etc. etc. Renowned academics and industry pioneers will lecture and share their views with the audience. Most deep learning subareas will be displayed, and main challenges identified through 22 four-hour and a half courses and 3 keynote lectures, which will tackle the most active and promising topics. The organizers are convinced that outstanding speakers will attract the brightest and most motivated students. Face to face interaction and networking will be main ingredients of the event. It will be also possible to fully participate in vivo remotely. An open session will give participants the opportunity to present their own work in progress in 5 minutes. Moreover, there will be two special sessions with industrial and recruitment profiles. ADDRESSED TO: Graduate students, postgraduate students and industry practitioners will be typical profiles of participants. However, there are no formal pre-requisites for attendance in terms of academic degrees, so people less or more advanced in their career will be welcome as well. Since there will be a variety of levels, specific knowledge background may be assumed for some of the courses. Overall, DeepLearn 2022 Spring is addressed to students, researchers and practitioners who want to keep themselves updated about recent developments and future trends. All will surely find it fruitful to listen to and discuss with major researchers, industry leaders and innovators. VENUE: DeepLearn 2022 Spring will take place in Guimar?es, in the north of Portugal, listed as UNESCO World Heritage Site and often referred to as the birthplace of the country. The venue will be: Hotel de Guimar?es Eduardo Manuel de Almeida 202 4810-440 Guimar?es http://www.hotel-guimaraes.com/ STRUCTURE: 3 courses will run in parallel during the whole event. Participants will be able to freely choose the courses they wish to attend as well as to move from one to another. Full in vivo online participation will be possible. However, the organizers highlight the importance of face to face interaction and networking in this kind of research training event. KEYNOTE SPEAKERS: Kate Smith-Miles (University of Melbourne), Stress-testing Algorithms via Instance Space Analysis Mihai Surdeanu (University of Arizona), Explainable Deep Learning for Natural Language Processing Zhongming Zhao (University of Texas, Houston), Deep Learning Approaches for Predicting Virus-Host Interactions and Drug Response PROFESSORS AND COURSES: Eneko Agirre (University of the Basque Country), [introductory/intermediate] Natural Language Processing in the Pretrained Language Model Era Altan ?ak?r (Istanbul Technical University), [introductory] Introduction to Deep Learning with Apache Spark Rylan Conway (Amazon), [introductory/intermediate] Deep Learning for Digital Assistants Jianfeng Gao (Microsoft Research), [introductory/intermediate] An Introduction to Conversational Information Retrieval Daniel George (JPMorgan Chase), [introductory] An Introductory Course on Machine Learning and Deep Learning with Mathematica/Wolfram Language Bohyung Han (Seoul National University), [introductory/intermediate] Robust Deep Learning Lina J. Karam (Lebanese American University), [introductory/intermediate] Deep Learning for Quality Robust Visual Recognition Xiaoming Liu (Michigan State University), [intermediate] Deep Learning for Trustworthy Biometrics Jennifer Ngadiuba (Fermi National Accelerator Laboratory), [intermediate] Ultra Low-latency and Low-area Machine Learning Inference at the Edge Lucila Ohno-Machado (University of California, San Diego), [introductory] Use of Predictive Models in Medicine and Biomedical Research Bhiksha Raj (Carnegie Mellon University), [introductory] Quantum Computing and Neural Networks Bart ter Haar Romenij (Eindhoven University of Technology), [intermediate] Deep Learning and Perceptual Grouping Kaushik Roy (Purdue University), [intermediate] Re-engineering Computing with Neuro-inspired Learning: Algorithms, Architecture, and Devices Walid Saad (Virginia Polytechnic Institute and State University), [intermediate/advanced] Machine Learning for Wireless Communications: Challenges and Opportunities Yvan Saeys (Ghent University), [introductory/intermediate] Interpreting Machine Learning Models Martin Schultz (J?lich Research Centre), [intermediate] Deep Learning for Air Quality, Weather and Climate Richa Singh (Indian Institute of Technology, Jodhpur), [introductory/intermediate] Trusted AI Sofia Vallecorsa (European Organization for Nuclear Research), [introductory/intermediate] Deep Generative Models for Science: Example Applications in Experimental Physics Michalis Vazirgiannis (?cole Polytechnique), [intermediate/advanced] Machine Learning with Graphs and Applications Guowei Wei (Michigan State University), [introductory/advanced] Integrating AI and Advanced Mathematics with Experimental Data for Forecasting Emerging SARS-CoV-2 Variants Xiaowei Xu (University of Arkansas, Little Rock), [intermediate/advanced] Deep Learning for NLP and Causal Inference Guoying Zhao (University of Oulu), [introductory/intermediate] Vision-based Emotion AI OPEN SESSION: An open session will collect 5-minute voluntary presentations of work in progress by participants. They should submit a half-page abstract containing the title, authors, and summary of the research to david at irdta.eu by April 10, 2022. INDUSTRIAL SESSION: A session will be devoted to 10-minute demonstrations of practical applications of deep learning in industry. Companies interested in contributing are welcome to submit a 1-page abstract containing the program of the demonstration and the logistics needed. People in charge of the demonstration must register for the event. Expressions of interest have to be submitted to david at irdta.eu by April 10, 2022. EMPLOYER SESSION: Firms searching for personnel well skilled in deep learning will have a space reserved for one-to-one contacts. It is recommended to produce a 1-page .pdf leaflet with a brief description of the company and the profiles looked for to be circulated among the participants prior to the event. People in charge of the search must register for the event. Expressions of interest have to be submitted to david at irdta.eu by April 10, 2022. ORGANIZING COMMITTEE: Dalila Dur?es (Braga, co-chair) Jos? Machado (Braga, co-chair) Carlos Mart?n-Vide (Tarragona, program chair) Sara Morales (Brussels) Paulo Novais (Braga, co-chair) David Silva (London, co-chair) REGISTRATION: It has to be done at https://irdta.eu/deeplearn/2022sp/registration/ The selection of 8 courses requested in the registration template is only tentative and non-binding. For the sake of organization, it will be helpful to have an estimation of the respective demand for each course. During the event, participants will be free to attend the courses they wish. Since the capacity of the venue is limited, registration requests will be processed on a first come first served basis. The registration period will be closed and the on-line registration tool disabled when the capacity of the venue will get exhausted. It is highly recommended to register prior to the event. FEES: Fees comprise access to all courses and lunches. There are several early registration deadlines. Fees depend on the registration deadline. ACCOMMODATION: Accommodation suggestions are available at https://irdta.eu/deeplearn/2022sp/accommodation/ CERTIFICATE: A certificate of successful participation in the event will be delivered indicating the number of hours of lectures. QUESTIONS AND FURTHER INFORMATION: david at irdta.eu ACKNOWLEDGMENTS: Centro Algoritmi, University of Minho, Guimar?es School of Engineering, University of Minho Intelligent Systems Associate Laboratory, University of Minho Rovira i Virgili University Municipality of Guimar?es Institute for Research Development, Training and Advice ? IRDTA, Brussels/London -------------- next part -------------- An HTML attachment was scrubbed... URL: From trentin at dii.unisi.it Sun Feb 20 17:27:14 2022 From: trentin at dii.unisi.it (Edmondo Trentin) Date: Sun, 20 Feb 2022 23:27:14 +0100 Subject: Connectionists: CFP: 10th IAPR TC3 International Workshop on Artificial Neural Networks in Pattern Recognition (ANNPR 2022, Dubai) Message-ID: <34206952646ed2d2194644742a841bf8.squirrel@mailsrv.diism.unisi.it> 10th IAPR TC3 International Workshop on Artificial Neural Networks in Pattern Recognition (ANNPR 2022) November 24th - 26th, 2022 Dubai Campus of Heriot-Watt University, Dubai, UAE URL: https://annpr2022.com/ Sponsored by the International Association for Pattern Recognition (IAPR) The Workshop proceedings will be published in the Springer LNAI series ANNPR 2022 invites papers that present original work in the areas of neural networks and machine learning oriented to pattern recognition, focusing on their algorithmic, theoretical, and applied aspects. Topics of interest include, but are not limited to: Methodological Issues - Supervised, semi-supervised, unsupervised, and reinforcement learning - Deep learning and deep reinforcement learning - Feed-forward, recurrent, and convolutional neural networks - Hierarchical modular architectures and hybrid systems - Interpretability and explainability of neural networks - Generative models - Robustness & generalization of neural networks - Meta-learning, Auto-ML - Multiple classifier systems and ensemble methods - Kernel machines - Probabilistic graphical models Applications to Pattern Recognition - Image processing and segmentation - Object detection - NLP and conversational agents - Sensor-fusion and multi-modal processing - Biometrics, including sspeech and speaker recognition and segmentation - Data, text, and social media analytics - Bioinformatics/Cheminformatics and medical applications - Industrial applications, e.g. quality control and predictive maintenance - Data clustering ANNPR 2022 is organized by the Technical Committee 3 (IAPR TC3) on Neural Networks & Computational Intelligence of the International Association for Pattern Recognition (IAPR). As such, we particularly encourage submissions that fit the Manifesto and the research directions of the TC3 (see http://iapr-tc3.diism.unisi.it/Research.html). The Workshop is officially sponsored by the IAPR. ANNPR 2022 is being planned as an in-presence event, and we encourage perspective attendees to join us in Dubai. Nevertheless, depending on the status of the ongoing Covid-19 pandemic, we are ready to go hybrid (or, even entirely virtual if necessary). Important Dates March 1st, 2022 ? Special session proposal June 5th, 2022 ? Paper submission July 29th, 2022 ? Notification of acceptance Sept 4th, 2022 ? Camera-ready Sept 4th, 2022 ? Early registration Workshop dates: Nov 24-26, 2022 Paper Submission: Perspective Authors shall submit their paper in Springer LNCS/LNAI format. Instructions for Authors, LaTeX templates, etc. are available at the Springer LNCS/LNAI web-site (see http://www.springer.com/it/computer-science/lncs/conference-proceedings-guidelines). The maximum paper length is 12 pages. Submission of a paper constitutes a commitment that, if accepted, at least one of the Authors will complete an early registration to the workshop. On-line submission via EasyChair will be made available in due time through the ANNPR 2022 website. For more information, please visit us at: https://annpr2022.com/ Do not hesitate to contact the ANNPR 2022 Chairs for any inquiries. ANNPR 2022 Chairs: Neamat El Gayar, Heriot-Watt University, Dubai (UAE) Mirco Ravanelli, Concordia University, Montr?al (Canada) Hazem Abbas, Ain Shams University, Cairo (Egypt) Edmondo Trentin, University of Siena, Siena (Italy) ----------------------------------------------- Edmondo Trentin, PhD Dip. Ingegneria dell'Informazione e Scienze MM. V. Roma, 56 - I-53100 Siena (Italy) E-mail: trentin at dii.unisi.it Voice: +39-0577-234636 Fax: +39-0577-233602 WWW: http://www.dii.unisi.it/~trentin/HomePage.html From mgalle at gmail.com Sun Feb 20 20:53:55 2022 From: mgalle at gmail.com (=?UTF-8?Q?Matthias_Gall=C3=A9?=) Date: Sun, 20 Feb 2022 22:53:55 -0300 Subject: Connectionists: Final CfP: Challenges & Perspectives in Creating Large Language Models Message-ID: *****Submission Deadline: Feb 28th***** *Call for Papers: Workshop on Challenges & Perspectives in Creating Large Language Models* May 27th 2022 (w/ ACL) https://bigscience.huggingface.co/acl-2022 Two years after the appearance of GPT-3, large language models seem to have taken over NLP. Their capabilities, limitations, societal impact and the potential new applications they unlocked have been discussed and debated at length. A handful of replication studies have been published since then, confirming some of the initial findings and discovering new limitations. This workshop aims to gather researchers and practitioners involved in the creation of these models in order to: 1. Share ideas on the next directions of research in this field, including ? but not limited to ? grounding, multi-modal models, continuous updates and reasoning capabilities. 2. Share best-practices, brainstorm solutions to identified limitations and discuss challenges, such as: - ?*Infrastructure*. What are the infrastructure and software challenges involved in scaling models to billions or trillions of parameters, and deploying training and inference on distributed servers when each model replicas is itself larger than a single node capacity? - ?*Data*. While the self-supervised setting dispenses with human annotation, the importance of cleaning, filtering and the bias and limitation in existing or reported corpora has become more and more apparent over the last years. - ?*Ethical & Legal frameworks*. What type of data can/should be used, what type of access should be provided, what filters are or should be necessary? - ?*Evaluation*. Investigating the diversity of intrinsic and extrinsic evaluation measures, how do they correlate and how the performances of a very large pretrained language model should be evaluated. - *?Training efficiency.* Discussing the practical scaling approaches, practical questions around large scale training hyper-parameters and early-stopping conditions. Discussing measures to reduce the associated energy consumption. This workshop is organized by the BigScience initiative and will also serve as the closing session of this one year-long initiative aimed at developing a multilingual large language model, which is currently gathering 900 researchers from more than 60 countries and 250 institutions. Its goal is to investigate the creation of a large scale dataset and model from a very wide diversity of angles. *Submissions* We call for relevant contributions, either in long (8 pages) or short (4 pages) format. Accepted papers will be presented during a poster session. Submissions can be archival or non-archival. Submissions should be made via OpenReview (https ://openreview.net/group?id=aclweb.org/ACL/2022/Workshop/BigScience). *Dates* Feb. 28, 2022: Submission Deadline ?March 26, 2022: Notification of Acceptance ?April 10, 2022: Camera-ready papers due -------------- next part -------------- An HTML attachment was scrubbed... URL: From a.crimi at sanoscience.org Sun Feb 20 18:22:59 2022 From: a.crimi at sanoscience.org (Alex Crimi) Date: Sun, 20 Feb 2022 23:22:59 +0000 Subject: Connectionists: First International Summer School in Italy 4th-9th July Message-ID: Dear Colleagues, apologies for cross-posting. I would like to invite you to 1st international Summer School on Neuroimaging, NeuroScience, Neuroncology 4?9 July, 2022, Aktea Hotel, Lipari, Italy https://www.neurosummerschool.org A first edition of the Sano Summer School dedicated to bridging global neuroscience world. Scope: The school is aimed at providing lectures and workshops about Neuroimaging, NeuroScience, Neuroncology for PhD students, post-doctoral researchers, and faculty/company members. Program comprises lectures and workshops with * Karl Friston * Alain Chedotal * Danielle Bassett * Luiz Pessoa * Bozena Kaminska * Randy Mcintosh * Walter Riviera * Marcin Wierzbinski * Mara Cercignani * Martijn van den Heuvel * Andras Jakab Organizing committee * Alessandro Crimi, Brian&More lab of Sano Science (Krakow, Poland) * Spyridon Bakas, Pereleman School of Medicine of University of Pennsylvania (Philadelphia, USA) Except further COVID-19 restrictions, with the idyllic landscape of the Aeolian islands, we invite you to lectures by well-known worldwide speakers, workshops, and tutorials. Complete program will be posted. Climbing the sulphuric crater of Vulcano is planned, as well as a boat trip to Panarea lagoons and around the active volcano of Stromboli. Due to Covid restrictions we will limit to 70 participants, with no particular pre-selection (First-Registered-First-Served) See you in Lipari. -------------- next part -------------- An HTML attachment was scrubbed... URL: From minaiaa at gmail.com Mon Feb 21 02:57:16 2022 From: minaiaa at gmail.com (Ali Minai) Date: Mon, 21 Feb 2022 02:57:16 -0500 Subject: Connectionists: Weird beliefs about consciousness In-Reply-To: References: <385C35DD-C428-4FAB-AA12-346DFDBD3B71@nyu.edu> <1abd4fb7-ff99-3494-48e2-6b9a28a3abcc@rubic.rutgers.edu> Message-ID: Hi Wlodek Thanks for your very thought-provoking reply and the great reading suggestions. We have known for a long time that the brain has both modal and amodal representations of concepts. There is also evidence that abstract concepts are built on the scaffolding of concrete ones (such as directions and shapes in physical space), even in non-human animals. This is just a conjecture but I think that the ability to build abstractions is just meta-representation made possible by hierarchical depth with the evolution of the cortex. So the representations such as hippocampal place codes, built as a direct result of embodied experience, become "the world" for higher layers of processing doing essentially the same thing but with a different level of grounding - with abstract concepts as "place codes" in a more abstract space. I call it "multi-level grounding". When you became grounded at the level of group theory, your grounding in embodiment was temporarily obscured because it was several level down. Of course, this is hardly a new idea, but worth keeping in mind. To go on a bit of a tangent but not unrelated, did you see this new paper: https://arxiv.org/abs/2202.07206 Apparently, GPT-3 does rely more than people admit on regurgitation. I don't think any language model build on the distributional hypothesis can ever be sufficiently grounded to have "understanding", but it should be possible in highly formalized domains such as computer programming, where the truth is so constrained and present wholly in the patterns of symbols. Natura; language less so. Actual experience, hardly at all. Ali *Ali A. Minai, Ph.D.* Professor and Graduate Program Director Complex Adaptive Systems Lab Department of Electrical Engineering & Computer Science 828 Rhodes Hall University of Cincinnati Cincinnati, OH 45221-0030 Past-President (2015-2016) International Neural Network Society Phone: (513) 556-4783 Fax: (513) 556-7326 Email: Ali.Minai at uc.edu minaiaa at gmail.com WWW: https://eecs.ceas.uc.edu/~aminai/ On Fri, Feb 18, 2022 at 12:45 PM Wlodzislaw Duch wrote: > Ali, > > certainly for many people identification with "being like us" is important > - this covers fertilized eggs and embryos, but not orangutans. John Locke > wrote 300 years ago: "Consciousness is the perception of what passes in a > Man's own mind". Physical states and processes that represent imagery, and > the ability to create symbolic narratives describing what goes on inside > cognitive system, should be the hallmark of consciousness. Of course more > people will accept it if we put it in a baby robot -:) > > This is why I prefer to focus on a simple requirement: inner world and the > ability to describe it. > The road to create robots that can feel has been described by Kevin > O'Regan in the book: > > O?Regan, J.K. (2011). Why Red Doesn?t Sound Like a Bell: Understanding the > Feel of Consciousness. Oxford University Press, USA. > > Inner worlds may be based on different representations, not always deeply > grounded in experience. Binder made a step toward a brain-based semantics: > Binder, J. R., Conant, L. L., Humphries, C. J., Fernandino, L., Simons, S. > B., Aguilar, M., & Desai, R. H. (2016). Toward a brain-based componential > semantic representation. Cognitive Neuropsychology, 33(3?4), 130?174. > Fernandino, L., Tong, J.-Q., Conant, L. L., Humphries, C. J., & Binder, J. > R. (2022). Decoding the information structure underlying the neural > representation of concepts. PNAS 119(6). > > This does not solve the symbol grounding problem (Harnad, 1990), but goes > half the way, mimicking embodiment by decomposing symbolic concepts into > attributes that are relevant to the brain. It should be sufficient to add > human-like semantics to bots. As you mention yourself, embodiment could > be more abstract, and I can imagine that a copy of a robot brain that has > grounded its representations in interactions with environment will endow a > new robot with similar experience. Can we simply implant it in the > network? > > I wonder if absorption in abstract thinking can leave space for the use of > experientially grounded concepts. I used to focus on group theory for hours > and was not able to understand what was said to me for brief moments. Was I > not conscious? Or should we consider continuous transition from abstract > semantics to fully embodied, human-like semantics in artificial systems? > > Wlodek > > > On 18/02/2022 16:36, Ali Minai wrote: > > Wlodek > > I think that the debate about consciousness in the strong sense of having > a conscious experience like we do is sterile. We will never have a > measuring device for whether another entity is "conscious", and at some > point, we will get to an AI that is sufficiently complex in its observable > behavior that we will either accept its inner state of consciousness on > trust - just as we do with humans and other animals - or admit that we will > never believe that a machine that is "not like us" can ever be conscious. > The "like us" part is more important than many of us in the AI field think: > A big part of why we believe other humans and our dogs are conscious is > because we know that they are "like us", and assume that they must share > our capacity for inner conscious experience. We already see this at a > superficial level where, as ordinary humans, we have a much easier time > identifying with an embodied, humanoid AI like Wall-E or the Terminator > than with a disembodied one like Hal or Skynet. This is also why so many > people find the Boston Dynamics "dog" so disconcerting. > > The question of embodiment is a complex one, as you know, of course, but I > am with those who think that it is necessary for grounding mental > representations - that it is the only way that the internal representations > of the system are linked directly to its experience. For example, if an AI > system trained only on text (like GPT-3) comes to learn that touching > something hot results in the fact of getting burned, we cannot accept that > as sufficient because it is based only on the juxtaposition of > abstractions, not the actual painful experience of getting burned. For > that, you need a body with sensors and a brain with a state corresponding > to pain - something that can be done in an embodied robot. This is why I > think that all language systems trained purely on the assumption of the > distributional hypothesis of meaning will remain superficial; they lack the > grounding that can only be supplied by experience. This does not mean that > systems based on the distributional hypothesis cannot learn a lot, or even > develop brain-like representations, as the following extremely interesting > paper shows: > > Y. Zhang, K. Han, R. Worth, and Z. Liu. Connecting concepts in the brain > by mapping cortical representations of semantic relations. Nature > Communications, 11(1):1877, Apr 2020. > > In a formal sense, however, embodiment could be in any space, including > very abstract ones. We can think of text data as GPT-3's world and, in that > world, it is "embodied" and its fundamentally distributional learning, > though superficial and lacking in experience to us, is grounded for it > within its world. Of course, this is not a very useful view of embodiment > and grounding since we want to create AI that is grounded in our sense, but > one of the most under-appreciated risks of AI is that, as we develop > systems that live in worlds very different than ours, they will - > implicitly and emergently - embody values completely alien to us. The > proverbial loan-processing AI that learns to be racially biased in just a > caricature of this hazard, but one that should alert us to deeper issues. > Our quaintly positivistic and reductionistic notion that we can deal with > such things by removing biases from data, algorithms, etc., is misplaced. > The world is too complicated for that. > > Ali > > *Ali A. Minai, Ph.D.* > Professor and Graduate Program Director > Complex Adaptive Systems Lab > Department of Electrical Engineering & Computer Science > 828 Rhodes Hall > University of Cincinnati > Cincinnati, OH 45221-0030 > > Phone: (513) 556-4783 > Fax: (513) 556-7326 > Email: Ali.Minai at uc.edu > minaiaa at gmail.com > > WWW: https://eecs.ceas.uc.edu/~aminai/ > > > On Fri, Feb 18, 2022 at 7:27 AM Wlodzislaw Duch wrote: > >> Asim, >> >> I was on the Anchorage panel, and asked others what could be a great >> achievement in computational intelligence. Steve Grossberg replied, that >> symbolic AI is meaningless, but creation of artificial rat that could >> survive in hostile environment would be something. Of course this is still >> difficult, but perhaps DARPA autonomous machines are not that far? >> >> I also had similar discussions with Walter and support his position: you >> cannot separate tightly coupled systems. Any external influence will create >> activation in both, linear causality looses its meaning. This is clear if >> both systems adjust to each other. But even if only one system learns >> (brain) and the other is mechanical but responds to human actions it may >> behave as one system. Every musician knows that: piano becomes a part of >> our body, responding in so many ways to actions, not only by producing >> sounds but also providing haptic feedback. >> >> This simply means that brains of locked-in people worked in somehow >> different way than brains of healthy people. Why do we consider them >> conscious? Because they can reflect on their mind states, imagine things >> and describe their inner states. If GPT-3 was coupled with something like >> DALL-E that creates images from text, and could describe what they see in >> their inner world, create some kind of episodic memory, we would have hard >> time to deny that this thing is not conscious of what it has in its mind. >> Embodiment helps to create inner world and changes it, but it is not >> necessary for consciousness. Can we find a good argument that such system >> is not conscious of its own states? It may not have all qualities of human >> consciousness, but that is a matter of more detailed approximation of >> missing functions. >> >> I have made this argument a long time ago (ex. in "*Brain-inspired >> conscious computing architecture" *written over 20 years ago, see more >> papers on this on my web page). >> >> Wlodek >> >> Prof. W?odzis?aw Duch >> Fellow, International Neural Network Society >> Past President, European Neural Network Society >> Head, Neurocognitive Laboratory, CMIT NCU, Poland >> >> Google: Wlodzislaw Duch >> >> On 18/02/2022 05:22, Asim Roy wrote: >> >> In 1998, after our debate about the brain at the WCCI in Anchorage, >> Alaska, I asked Walter Freeman if he thought the brain controls the body. >> His answer was, you can also say that the body controls the brain. I then >> asked him if the driver controls a car, or the pilot controls an airplane. >> His answer was the same, that you can also say that the car controls the >> driver, or the plane controls the pilot. I then realized that Walter was >> also a philosopher and believed in the No-free Will theory and what he was >> arguing for is that the world is simply made of interacting systems. >> However, both Walter, and his close friend John Taylor, were into >> consciousness. >> >> >> >> I have argued with Walter on many different topics over nearly two >> decades and have utmost respect for him as a scholar, but this first >> argument I will always remember. >> >> >> >> Obviously, there?s a conflict between consciousness and the No-free Will >> theory. Wonder where we stand with regard to this conflict. >> >> >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> >> *From:* Connectionists >> *On Behalf Of *Andras >> Lorincz >> *Sent:* Tuesday, February 15, 2022 6:50 AM >> *To:* Stephen Jos? Hanson >> ; Gary Marcus >> >> *Cc:* Connectionists >> >> *Subject:* Re: Connectionists: Weird beliefs about consciousness >> >> >> >> Dear Steve and Gary: >> >> This is how I see (try to understand) consciousness and the related >> terms: >> >> (Our) consciousness seems to be related to the close-to-deterministic >> nature of the episodes on from few hundred millisecond to a few second >> domain. Control instructions may leave our brain 200 ms earlier than the >> action starts and they become conscious only by that time. In addition, >> observations of those may also be delayed by a similar amount. (It then >> follows that the launching of the control actions is not conscious and -- >> therefore -- free will can be debated in this very limited context.) On the >> other hand, model-based synchronization is necessary for timely >> observation, planning, decision making, and execution in a distributed and >> slow computational system. If this model-based synchronization is not >> working properly, then the observation of the world breaks and >> schizophrenic symptoms appear. As an example, individuals with pronounced >> schizotypal traits are particularly successful in self-tickling (source: >> https://philpapers.org/rec/LEMIWP >> , >> and a discussion on Asperger and schizophrenia: >> https://www.frontiersin.org/articles/10.3389/fpsyt.2020.503462/full >> ) >> a manifestation of improper binding. The internal model enables and the >> synchronization requires the internal model and thus a certain level of >> consciousness can appear in a time interval around the actual time instant >> and its length depends on the short-term memory. >> >> Other issues, like separating the self from the rest of the world are >> more closely related to the soft/hard style interventions (as called in the >> recent deep learning literature), i.e., those components (features) that >> can be modified/controlled, e.g., color and speed, and the ones that are >> Lego-like and can be separated/amputed/occluded/added. >> >> Best, >> >> Andras >> >> >> >> ------------------------------------ >> >> Andras Lorincz >> >> http://nipg.inf.elte.hu/ >> >> >> Fellow of the European Association for Artificial Intelligence >> >> https://scholar.google.com/citations?user=EjETXQkAAAAJ&hl=en >> >> >> Department of Artificial Intelligence >> >> Faculty of Informatics >> >> Eotvos Lorand University >> >> Budapest, Hungary >> >> >> >> >> >> >> ------------------------------ >> >> *From:* Connectionists >> on behalf of Stephen Jos? Hanson >> *Sent:* Monday, February 14, 2022 8:30 PM >> *To:* Gary Marcus >> *Cc:* Connectionists >> *Subject:* Re: Connectionists: Weird beliefs about consciousness >> >> >> >> Gary, these weren't criterion. Let me try again. >> >> I wasn't talking about wake-sleep cycles... I was talking about being >> awake or asleep and the transition that ensues.. >> >> Rooba's don't sleep.. they turn off, I have two of them. They turn on >> once (1) their batteries are recharged (2) a timer has been set for being >> turned on. >> >> GPT3 is essentially a CYC that actually works.. by reading Wikipedia >> (which of course is a terribly biased sample). >> >> I was indicating the difference between implicit and explicit >> learning/problem solving. Implicit learning/memory is unconscious and >> similar to a habit.. (good or bad). >> >> I believe that when someone says "is gpt3 conscious?" they are asking: >> is gpt3 self-aware? Roombas know about vacuuming and they are >> unconscious. >> >> S >> >> On 2/14/22 12:45 PM, Gary Marcus wrote: >> >> Stephen, >> >> >> >> On criteria (1)-(3), a high-end, mapping-equippped Roomba is far more >> plausible as a consciousness than GPT-3. >> >> >> >> 1. The Roomba has a clearly defined wake-sleep cycle; GPT does not. >> >> 2. Roomba makes choices based on an explicit representation of its >> location relative to a mapped space. GPT lacks any consistent reflection of >> self; eg if you ask it, as I have, if you are you person, and then ask if >> it is a computer, it?s liable to say yes to both, showing no stable >> knowledge of self. >> >> 3. Roomba has explicit, declarative knowledge eg of walls and other >> boundaries, as well its own location. GPT has no systematically >> interrogable explicit representations. >> >> >> >> All this is said with tongue lodged partway in cheek, but I honestly >> don?t see what criterion would lead anyone to believe that GPT is a more >> plausible candidate for consciousness than any other AI program out there. >> >> >> >> ELIZA long ago showed that you could produce fluent speech that was >> mildly contextually relevant, and even convincing to the untutored; just >> because GPT is a better version of that trick doesn?t mean it?s any more >> conscious. >> >> >> >> Gary >> >> >> >> On Feb 14, 2022, at 08:56, Stephen Jos? Hanson >> wrote: >> >> ? >> >> this is a great list of behavior.. >> >> Some biologically might be termed reflexive, taxes, classically >> conditioned, implicit (memory/learning)... all however would not be >> conscious in the several senses: (1) wakefulness-- sleep (2) self >> aware (3) explicit/declarative. >> >> I think the term is used very loosely, and I believe what GPT3 and other >> AI are hoping to show signs of is "self-awareness".. >> >> In response to : "why are you doing that?", "What are you doing now", >> "what will you be doing in 2030?" >> >> Steve >> >> >> >> On 2/14/22 10:46 AM, Iam Palatnik wrote: >> >> A somewhat related question, just out of curiosity. >> >> >> >> Imagine the following: >> >> >> >> - An automatic solar panel that tracks the position of the sun. >> >> - A group of single celled microbes with phototaxis that follow the >> sunlight. >> >> - A jellyfish (animal without a brain) that follows/avoids the sunlight. >> >> - A cockroach (animal with a brain) that avoids the sunlight. >> >> - A drone with onboard AI that flies to regions of more intense sunlight >> to recharge its batteries. >> >> - A human that dislikes sunlight and actively avoids it. >> >> >> >> Can any of these, beside the human, be said to be aware or conscious of >> the sunlight, and why? >> >> What is most relevant? Being a biological life form, having a brain, >> being able to make decisions based on the environment? Being taxonomically >> close to humans? >> >> >> >> >> >> >> >> >> >> >> >> >> >> >> >> On Mon, Feb 14, 2022 at 12:06 PM Gary Marcus wrote: >> >> Also true: Many AI researchers are very unclear about what consciousness >> is and also very sure that ELIZA doesn?t have it. >> >> Neither ELIZA nor GPT-3 have >> - anything remotely related to embodiment >> - any capacity to reflect upon themselves >> >> Hypothesis: neither keyword matching nor tensor manipulation, even at >> scale, suffice in themselves to qualify for consciousness. >> >> - Gary >> >> > On Feb 14, 2022, at 00:24, Geoffrey Hinton >> wrote: >> > >> > ?Many AI researchers are very unclear about what consciousness is and >> also very sure that GPT-3 doesn?t have it. It?s a strange combination. >> > >> > >> >> -- >> >> >> -- > Prof. W?odzis?aw Duch > Fellow, International Neural Network Society > Past President, European Neural Network Society > Head, Neurocognitive Laboratory, CMIT NCU, Poland > Google: Wlodzislaw Duch > -------------- next part -------------- An HTML attachment was scrubbed... URL: From sahidullahmd at gmail.com Mon Feb 21 02:58:12 2022 From: sahidullahmd at gmail.com (Md Sahidullah) Date: Mon, 21 Feb 2022 13:28:12 +0530 Subject: Connectionists: First Spoofing-Aware Speaker Verification (SASV) 2022 Challenge Message-ID: *Apologies if you have received multiple copies of this announcement.* Dear all, We are thrilled to announce the Spoofing-Aware Speaker Verification Challenge. While spoofing countermeasures, promoted within the sphere of the ASVspoof challenge series, can help to protect reliability in the face of spoofing, they have been developed as independent subsystems for a fixed ASV subsystem. Better performance can be expected when countermeasures and ASV subsystems are both optimised to operate in tandem. The first Spoofing-Aware Speaker Verification (SASV) 2022 challenge aims to encourage the development of original solutions involving, but not limited to: - back-end fusion of pre-trained automatic speaker verification and pre-trained audio spoofing countermeasure subsystems; - integrated spoofing-aware automatic speaker verification systems that have the capacity to reject both non-target and spoofed trials. We warmly invite the submission of general contributions in this direction. The Interspeech 2022 Spoofing-Aware Automatic Speaker Verification special session also incorporates a challenge ? SASV 2022. Participants are encouraged to evaluate their solutions using the SASV benchmarking framework which comprises a common database, protocol, and evaluation metric. Further details and resources can be found on the SASV challenge website. URL: https://sasv-challenge.github.io Schedule: -January 19, 2022: Release of the evaluation plan - March 10, 2022: Results submission - March 14, 2022: Release of participant ranks - March 21, 2022: INTERSPEECH Paper submission deadline - March 28, 2022: INTERSPEECH Paper update deadline - June 13, 2022: INTERSPEECH Author notification - September 18-22, 2022: SASV challenge special session at INTERSPEECH To participate, please register your interest at https://forms.gle/htoVnog34kvs3as56 For further information, please contact us at sasv.challenge at gmail.com. We are looking forward to hearing from you. Kind regards, The SASV Challenge 2022 Organisers ------ Organisers: Jee-weon Jung, Naver Corporation, South Korea Hemlata Tak, EURECOM, France Hye-jin Shim, University of Seoul, South Korea Hee-Soo Heo, Naver Corporation, South Korea Bong-Jin Lee, Naver Corporation, South Korea Soo-Whan Chung, Naver Corporation, South Korea Hong-Goo Kang, Yonsei University, South Korea Ha-Jin Yu, University of Seoul, South Korea Nicholas Evans, EURECOM, France Tomi H. Kinnunen, University of Eastern Finland, Finland -- Md Sahidullah website: *https://sites.google.com/site/iitkgpsahi/ * -------------- next part -------------- An HTML attachment was scrubbed... URL: From angelo.ciaramella at uniparthenope.it Mon Feb 21 03:18:32 2022 From: angelo.ciaramella at uniparthenope.it (ANGELO CIARAMELLA) Date: Mon, 21 Feb 2022 08:18:32 +0000 Subject: Connectionists: =?windows-1252?q?=5BSpecial_Issue_on_=93Human-Cen?= =?windows-1252?q?tered_Intelligent_System=5D_-_Soft_Computing_deadline_ap?= =?windows-1252?q?proaching=3A_1_March_2022?= Message-ID: ... Apologize for cross-posting ... ------------------------------------------------------------------------------------------------------ Special issue on ?Human-Centered Intelligent System? Soft Computing - A Fusion of Foundations, Methodologies and Applications ------------------------------------------------------------------------------------------------------ Website of the call for papers https://www.springer.com/journal/500/updates/19724572 Aims and scope Nowadays, Artificial Intelligence has become an enabling technology that pervades many aspects of our daily life. At the forefront of this advancement are data-driven technologies like machine learning. However, as the role of Artificial Intelligence becomes more and more important, so does the need for reliable solutions to several issues that go well beyond technological aspects: How can we make automated agents justify their actions? and how to make them accountable for these actions? What will be the social acceptance of intelligent systems, possibly embodied (e.g. in robots), in their interaction with people? How will automated agents be made aware of the whole spectrum of human nonverbal communication, so as to take it into account and avoid missing crucial messages? Is it possible to avoid amplifying human biases and ensure fairness in decisions that have been taken automatically? How can we enable collaborative intelligence amongst humans and machines? Purely data-driven technologies are showing their limits precisely in these areas. There is a growing need for methods that, in a tight interaction with them, allow different degrees of control over the several facets of automated knowledge processing. The diversity and complementarity of Soft Computing techniques in addressing these issues is playing a crucial role. This Special Issue aims to collect the most recent advancements in the research on Human-Centered Intelligent Systems with special focus on Soft Computing methods, techniques and applications on the following and related topics: * Trustworthiness, explainability, accountability and social acceptance of intelligent systems; * Human-computer interaction to foster collaboration with intelligent systems * Affective computing and sentiment analysis to promote nonverbal communication in intelligent systems * Fighting algorithmic bias and ensuring fairness in intelligent systems; * Real world applications in health-care, justice, education, digital marketing, biology, hard and natural sciences, autonomous vehicles, etc. Proposals related to further topics are welcome as long as they fall within the general scope of this special issue, which is computational intelligence methods in human-centered computing. Submission guidelines and review process Papers must be submitted according to the standard procedure of Soft Computing, selecting the S.I. ?Human-Centered Intelligent Systems?. All submitted papers should report original work and make a meaningful contribution to the state of the art. Each submitted paper will undergo a first screening by the Guest Editors. If the submission falls within the scope of the SI, it will undergo a regular revision process. Acceptance criteria are the same of regular issues of the journal. Important dates * Submissions open: December 1, 2021 * Paper submission deadline: March 1, 2022 * Final decision: May 1, 2022 * Tentative period for final publication: Fall 2022 Authors guidelines and journal information can be found at https://www.springer.com/journal/500 Guest Editors - Angelo Ciaramella - Universit? degli Studi di Napoli Parthenope, Italy - Corrado Mencar - Universit? degli Studi di Bari Aldo Moro, Italy - Susana Montes - Universidad de Oviedo, Spain - Stefano Rovetta - Universit? degli Studi di Genova, Italy For any information, please contact Angelo Ciaramella > ??????????????????????? Angelo Ciaramella - School of Science, Engineering and Health, Department of Science and Technology, University of Naples Parthenope - Room 431, C4 Island, Centro Direzionale di Napoli - I-80143, Naples, Italy - tel.: 0815476674 - e-mail: angelo.ciaramella at uniparthenope.it [signature_1686498466] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 8513 bytes Desc: image001.png URL: From joanna.zapisek at eng.ox.ac.uk Mon Feb 21 05:51:57 2022 From: joanna.zapisek at eng.ox.ac.uk (Joanna Zapisek) Date: Mon, 21 Feb 2022 10:51:57 +0000 Subject: Connectionists: University of Oxford, Postdoctoral Research Assistant in Computer Vision Message-ID: Hello, We are seeking an experienced full-time Postdoctoral Research Assistant's in Computer Vision to join Prof Victor Adrian Prisacariu's and Prof Philip Torr's research groups at the Department of Engineering Science (University of Oxford). This is a fixed term role for 18 months with funding provided by industry. Both groups are internationally leading research groups that have numerous scientific awards and have close links with some of the top industrial research labs. This is a research project that aims to link 3D object reconstruction using SfM and/or depth fusion with 3D deep learning, with a strong focus on accuracy. You will be required to develop and implement novel computer vision and learning algorithms to reconstruct 3D object geometry whilst also contributing documented software to the group library. You will also write reports on your research, and publish these as papers in leading conferences and journals. You should possess a doctorate, or be near completion of a doctorate, in Computer Vision, Machine Learning or a relevant subject, together with a strong publication record at principal computer vision conferences (CVPR, ECCV, ICCV and BMVC), with a background 3D geometry (depth fusion/SLAM/object tracking and reconstruction) and/or deep learning. Informal enquiries may be addressed to victor at robots.ox.ac.uk or philip.torr at eng.ox.ac.uk Only applications received before midday on the 28th February 2022 can be considered. You will be required to upload a covering letter/supporting statement, including a brief statement of research interests (describing how past experience and future plans fit with the advertised position), CV and the details of two referees as part of your online application. The Department holds an Athena Swan Bronze award, highlighting its commitment to promoting women in Science, Engineering and Technology. The University of Oxford is committed to equality, diversity and inclusivity. We welcome applications from everyone, regardless of the categories they put themselves in. We are keen to offer opportunities, where possible, to people from groups who are currently underrepresented in science, technology, engineering and mathematics (STEM) in the UK, such as women and individuals from black or Asian backgrounds. To apply go to https://my.corehr.com/pls/uoxrecruit/erq_jobspec_version_4.display_form?p_company=10&p_internal_external=E&p_display_in_irish=N&p_process_type=&p_applicant_no=&p_form_profile_detail=&p_display_apply_ind=Y&p_refresh_search=Y&p_recruitment_id=155875 Thank you, Joanna [cid:694DA3CF-6F66-4FBC-A3E2-A0586029DE1B]Joanna Zapisek Senior Research Manager Professor Torr Vision Group University of Oxford Department of Engineering Science Parks Road, Oxford, OX1 3PJ t: +44 (0) 1865 273130 e: Joanna.zapisek at eng.ox.ac.uk w: https://torrvision.com/ I'm working part time (Mon, Wed, Fri full day; Tue and Thurs mornings) -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 19989 bytes Desc: image001.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image003.jpg Type: image/jpeg Size: 2411 bytes Desc: image003.jpg URL: From george at cs.ucy.ac.cy Mon Feb 21 05:08:15 2022 From: george at cs.ucy.ac.cy (George A. Papadopoulos) Date: Mon, 21 Feb 2022 12:08:15 +0200 Subject: Connectionists: Faculty Opening -- Department of Computer Science -- University of Cyprus Message-ID: VACANCY ANNOUNCEMENT FOR AN ACADEMIC POSITION The University of Cyprus was founded in 1989 and admitted its first students in 1992. Within a short time, the University of Cyprus achieved international distinctions. Today, it is ranked as the 108th young university (under 50 years) and among the #401-500 best universities worldwide by the Times Higher Education Rankings. These notable distinctions, achieved in the highly competitive field of research, are the result of commitment and dedication to continuous development and research excellence, the promotion and strengthening of which constitutes a key strategic objective of the University of Cyprus. Moreover, the University continually extends and develops its undergraduate and graduate programs of studies. To best serve its research and educational aims, the University recruits high-caliber academic staff who can make significant contributions to the development of internationally competitive research projects and to the design and delivery of new curricula. DEPARTMENT OF COMPUTER SCIENCE The University of Cyprus invites applications for one (1) tenure-track academic position at the rank of Lecturer or Assistant Professor in the Department of Computer Science, in the field of ?Artificial Intelligence with emphasis in machine learning?. For all academic ranks, a Doctoral degree from an accredited University is required. The minimum requirements for each academic rank are available at http://www.cs.ucy.ac.cy/~george/GPLists_2021/lm.php?tk=CQlGb3JtYWwJY29ubmVjdGlvbmlzdHNAbWFpbG1hbi5zcnYuY3MuY211LmVkdQlGYWN1bHR5IE9wZW5pbmcgLS0gRGVwYXJ0bWVudCBvZiBDb21wdXRlciBTY2llbmNlIC0tIFVuaXZlcnNpdHkgb2YgQ3lwcnVzCTQ4NwlHZW9yZ2UJMTUJY2xpY2sJeWVzCW5v&url=https%3A%2F%2Fwww.ucy.ac.cy%2Facad.staff.procedures and include: previous academic experience, outstanding research achievements and notable scientific contributions, experience in developing and teaching of high quality undergraduate and graduate curricula. Candidates do not need to be citizens of the Republic of Cyprus. The official languages of instruction are Greek and Turkish. For the above position, fluency in the Greek language is necessary. In case the selected candidate is not proficient in the Greek language, the candidate and the Department shall ensure that the former acquires sufficient knowledge of the Greek language within 3 years from the date of appointment. Each Department sets its own criteria concerning the required level of fluency in the Greek language. In accordance with the applicable legislation, the annual gross salary (including the 13th salary) for full-time employment is: Assistant Professor (Scale A13-A14). ?59.173,18- ?79.802,36 Lecturer (Scale A12-A13) ?44.975,98- ?73.185,96 Employee contributions to the various State funds will be deducted from the above amounts. Candidates are invited to submit their applications electronically by uploading the following documents in English and in PDF format at the following link: http://www.cs.ucy.ac.cy/~george/GPLists_2021/lm.php?tk=CQlGb3JtYWwJY29ubmVjdGlvbmlzdHNAbWFpbG1hbi5zcnYuY3MuY211LmVkdQlGYWN1bHR5IE9wZW5pbmcgLS0gRGVwYXJ0bWVudCBvZiBDb21wdXRlciBTY2llbmNlIC0tIFVuaXZlcnNpdHkgb2YgQ3lwcnVzCTQ4NwlHZW9yZ2UJMTUJY2xpY2sJeWVzCW5v&url=https%3A%2F%2Fapplications.ucy.ac.cy%2Frecruitment 1. Cover Letter 2. Curriculum Vitae 3. Copy of ID/Passport 4. Copies of degree certificates 5. Review of previous research work and a brief description of future research projects (up to 3 pages) 6. List of publications 7. Representative publications (up to 3 publications which should be submitted separately). For the rank of Lecturer, the submission of representative publications is optional. 8. The names and email addresses of three academic referees, who, upon submission of the application, will be automatically notified to provide recommendation letters (in English), up to seven days following the deadline for submission of applications. The above documents (1-7) must be uploaded as separate PDF documents. No change can be possible upon submission of your application. The deadline for applications is on Wednesday 18th of May 2022. The selected applicants will be required to submit copies of degree certificates certified by the Ministry of Education (if the degrees were obtained from universities n Cyprus) or from the Issuing Authority (if the degrees were obtained from foreign universities). Applications, supporting documents and recommendation letters submitted in response to previous vacancy announcements will not be considered and must be resubmitted. Applications not providing all the required documents specified in the online application form at the above link will not be considered. The applicant shall ensure that their application has been successfully submitted. Upon submission, the candidate will receive an automated confirmation email. For more information, please contact the Human Resources Services (tel.: 00357 22 89 4146) or the Department of Computer Science (+357 22 89 2669). -------------- next part -------------- An HTML attachment was scrubbed... URL: From calendarsites at insticc.org Mon Feb 21 08:32:09 2022 From: calendarsites at insticc.org (calendarsites at insticc.org) Date: Mon, 21 Feb 2022 13:32:09 -0000 Subject: Connectionists: [CFP]19th Int. Conf. on Smart Business Technologies :: Submission Deadline - 2nd of March Message-ID: <002801d82727$7148bdb0$53da3910$@insticc.org> CALL FOR PAPERS 19th International Conference on Smart Business Technologies **Submission Deadline: March 2, 2022** https://icsbt.scitevents.org/ July 14 - 16, 2022 Lisbon, Portugal Important Note: The conference will be held in Lisbon but we are open to accept online presentations in case the participants can't attend the conference. The International Conference on Smart Business Technologies (formerly known as ICE-B - International Conference on e-Business), aims at bringing together researchers and practitioners who are interested in e-Business technology and its current applications. The scope of the conference covers low-level technological issues, such as technology platforms, internet of things and web services, but also higher-level issues, such as business processes, business intelligence, value setting and business strategy. Furthermore, it covers different approaches to address these issues and different possible applications with their own specific needs and requirements on technology. These are all areas of theoretical and practical importance within the broad scope of e-Business, whose growing importance can be seen from the increasing interest of the IT research community. ICSBT is organized in 4 major tracks: 1 - Business Intelligence 2 - Business Models and Business Processes 3 - Collaboration and Interoperability 4 - Technologies and Applications Conference Chair(s) Marten van Sinderen, University of Twente, Netherlands Program Chair(s) Fons Wijnhoven, University of Twente, Netherlands In the last years, the proceedings have been fully indexed by SCOPUS. Beside this index all the proceedings have also been submitted to Google Scholar, The DBLP Computer Science Bibliography, Semantic Scholar, Engineering Index (EI) and Web of Science / Conference Proceedings Citation Index. A short list of presented papers will be selected so that revised and extended versions of these papers will be published by Springer in a CCIS Series book. All papers presented at the conference venue will also be available at the SCITEPRESS Digital Library. Kind regards, M?nica Saramago ICSBT Secretariat Web: https://icsbt.scitevents.org/ e-mail: ice-b.secretariat at insticc.org -------------- next part -------------- An HTML attachment was scrubbed... URL: From jerchen at bu.edu Mon Feb 21 10:28:50 2022 From: jerchen at bu.edu (Chen, Jerry) Date: Mon, 21 Feb 2022 15:28:50 +0000 Subject: Connectionists: NSF Neuronex Big Brain Imaging Hybrid Workshop @ Boston University - April 7-8, 2022 Message-ID: We are proud to announce the 2022 Big Brain Imaging Workshop hosted jointly by the NSF NeuroNex Nemonic Hub as well as the Boston University Neurophotonics Center! This two-day hybrid workshop will take place at Boston University and online on Thursday, April 7th and continue on Friday, April 8th, 2022. The 2022 Big Brain Imaging Workshop is an opportunity to join in the discussion on optical approaches for cellular resolution imaging of neuronal populations that have been highly instrumental for uncovering the principles of neural circuit function. To date, the microscope technology development for imaging the nervous system has largely been focused on applications on a subset of model species with smaller, optically accessible brains (ex. c. elegans, zebrafish, mice). The ability to translate these technologies to species with brains that are both physically larger and less optically accessible remains a challenge. The topic of the workshop is focused on advancing cellular resolution brain imaging techniques in "larger" brains of mammalian species (ex. rats, cats, ferrets, marmosets, macaques). The goal is to bring neuroscientists and optical engineers together to discuss challenges and come up with potential solutions that will pave the way for the next generation of microscopes that can be applied to a greater number of model organisms. Invited Speakers Daniel Aharoni - University of California, Los Angeles Adam Charles - Johns Hopkins University Anna Devor - Boston University David Fitzpatrick - Max Planck Institute Florida Emily Gibson - University of Colorado Anschutz Prakash Kara - University of Minnesota Jerome Mertz - Boston University Kristina Nielsen - Johns Hopkins University Anitha Pasupathy - University of Washington Bijan Perasan - New York University Nicholas Priebe - University of Texas, Austin Eyal Seidemann - University of Texas, Austin Lei Tian - Boston University Alipasha Vaziri - Rockefeller University Chris Xu - Cornell University In addition to the invited speakers, spots are available for short talks by visiting the registration page and submitting a brief abstract: For more information and free registration, visit: https://nemonic.ece.ucsb.edu/index.php/Big_Brain Jerry Chen, PhD Assistant Professor Department of Biology Boston University www.chen-lab.org jerry at chen-lab.org -------------- next part -------------- An HTML attachment was scrubbed... URL: From papaleon at sch.gr Mon Feb 21 12:31:52 2022 From: papaleon at sch.gr (Papaleonidas Antonios) Date: Mon, 21 Feb 2022 19:31:52 +0200 Subject: Connectionists: 18th AIAI 2022 Hybrid @ Crete, Greece - LAST Call for Papers Message-ID: <030601d82748$ea223ab0$be66b010$@sch.gr> 18th AIAI 2022, 17 - 20 June 2022 Hybrid@ Web & Aldemar Knossos Royal, Crete, Greece www.ifipaiai.org/2022 CALL FOR PAPERS for 18th AIAI 2022 Hybrid @ Web & Crete, Greece Dear Antonios Papaleonidas We would like to invite you to submit your work at the 18th International Conference on Artificial Intelligence Applications and Innovations (AIAI2022 ) 18th International Conference on Artificial Intelligence Applications and Innovations, AIAI 2022, is technically sponsored by IFIP Artificial Intelligence Applications WG12.5. It is going to be co-organized as a Joint event with 23rd Conference on Engineering Applications of Neural Networks, EANN 2022, which is technically sponsored by the INNS (International Neural Network Society). SPECIAL ISSUES - PROCEEDINGS: Selected papers will be published in 4 special issues of high quality international scientific Journals: * World Scientific journal, International Journal of Neural Systems , Impact factor 5.87 * Springer journal , Neural Computing and Applications , Impact Factor 5.61 * ???? journal, International Journal of Biomedical and Health Informatics , Impact factor 5.772 * Springer journal, AI & Ethics PROCEEDINGS will be published SPRINGER IFIP AICT Series and they are INDEXED BY SCOPUS, DBLP, Google Scholar, ACM Digital Library, IO-Port, MAthSciNet, CPCI, Zentralblatt MATH and EI Engineering Index Papers submissions will be up to 12 pages long and not less than 6 pages. BIBLIOMETRIC DETAILS: We proudly announce that according to Springer?s statistics, the last 15 AIAI conferences have been downloaded 1,719,00 times! IFIP AIAI series has reached h-index of 29 and published papers have been Cited more than 6000 times! For more Bibliometric Details please click at AIAI BIBLIOMETRIC DETAILS page IMPORTANT DATES: * Paper Submission Deadline: 25th of February 2022 * Notification of Acceptance: 26th of March 2022 * Camera ready Submission: 22th of April 2022 * Early / Authors Registration Deadline: 22th of April 2022 * Conference: 17 - 20 of June 2022 WORKSHOPS & SPECIAL SESSIONS: So far, the following 8 high quality Workshops & Speccail Sessions have been accepted and scheduled: * 11th Mining Humanistic Data Workshop (MHDW 2022) * 7th Workshop on ?5G ? Putting Intelligence to the Network Edge? (5G-PINE 2021) * 2nd Defense Applications of AI Workshop (DAAI) an EDA ? EU Workshop * 2nd Distributed AI for Resource-Constrained Platforms Workshop (DARE 2022) * 2nd Artificial Intelligence in Biomedical Engineering and Informatics (AI-BEI 2022) * 2nd Artificail Intelligence & Ethics Workshop (AIETH 2022) * AI in Energy, Buildings and Micro-Grids Workshop (??BMG) * Machine Learning and Big Data in Health Care (ML at HC) For more info please visit AIAI 2022 workshop info page KEYNOTE SPEAKERS: So far two Plenary Lectures have been announced, both by distinguished Professors with an important imprint in AI and Machine Learning. * Professor Hojjat Adeli Ohio State University, Columbus, USA, Fellow of the Institute of Electrical and Electronics Engineers (IEEE) (IEEE), Honorary Professor, Southeast University, Nanjing, China, Member, Polish and Lithuanian Academy of Sciences, Elected corresponding member of the Spanish Royal Academy of Engineering. Visit Google Scholar profile , h-index: 114 * Professor Riitta Salmelin Department of Neuroscience and Biomedical Engineering Aalto University, Finland Visit Google Scholar profile , h-index: 65 * Professor Dr. Elisabeth Andr? Human-Centered Artificial Intelligence, Institute for Informatics, University of Augsburg, Germany Visit Google Scholar profile , h-index: 61 * Professor Verena Reiser School of Mathematical and Computer Sciences (MACS) at Heriot Watt University, Edinburgh Visit Google Scholar profile , h-index: 31 For more info please visit AIAI 2022 Keynote info page VENUE: ALDEMAR KNOSSOS ROYAL Beach Resort in Hersonisso Peninsula, Crete, Greece. Special Half Board prices have been arranged for the conference delegates in the Aldemar Knossos Royal Beach Resort. For details please see: https://ifipaiai.org/2022/venue/ Conference topics, CFPs, Submissions & Registration details can be found at: * ifipaiai.org/2022/calls-for-papers/ * ifipaiai.org/2022/paper-submission/ * ifipaiai.org/2022/registration/ We are expecting Submissions on all topics related to Artificial and Computational Intelligence and their Applications. Detailed Guidelines on the Topics and the submission details can be found at the links above General co-Chairs: * Ilias Maglogiannis, University of Piraeus, Greece * John Macintyre, University of Sunderland, United Kingdom Program co-Chairs: * Lazaros Iliadis, School of Engineering, Democritus University of Thrace, Greece * Konstantinos Votis, Information Technologies Institute, ITI Thessaloniki, Greece * Vangelis Metsis, Texas State University, USA -------------- next part -------------- An HTML attachment was scrubbed... URL: From timothee.proix at unige.ch Mon Feb 21 12:46:45 2022 From: timothee.proix at unige.ch (=?UTF-8?Q?Timoth=C3=A9e_Proix?=) Date: Mon, 21 Feb 2022 18:46:45 +0100 Subject: Connectionists: Postdoctoral position in natural language processing and neuroscience, University of Geneva, Switzerland Message-ID: <43327f0e-8ae9-4055-a8f4-0adfcb631286@www.fastmail.com> Dear all, a postdoctoral position is available under the supervision of Prof. Pierre M?gevand (https://www.unige.ch/medecine/neucli/en/groupes-de-recherche/1034megevand/) and Timoth?e Proix (https://ndlab.ch/) at the University of Geneva, Switzerland. The project lies at the interface between natural language processing and intracranial EEG data analysis. The goal of the project is to understand the neural representations of speech production and perception in conversations. The candidate will contribute to acquisition and analysis of intracranial EEG data and speech recordings. Applicants should have prior experience with machine learning methods for natural language processing, as well as in advanced data analysis techniques. Experience in electrophysiological recordings and knowledge in neuroscience is a significant advantage. The research will be performed at the campus Biotech in the university of Geneva, in collaboration with the Geneva University Hospital. The positions are funded by the private foundation of the Geneva University Hospital, and salaries are in compliance with guidelines of the University of Geneva. The ideal start is May 2022, with some flexibility. To apply, send a motivation letter, CV, and contact information for 2-3 references, as well as any other relevant documents to the Prof. Pierre M?gevand (pierre.megevand (at) unige.ch) as well as Timoth?e Proix (timothee.proix (at) unige.ch). Screening of applications will start by March 2022 and will continue until the position is filled. Timoth?e Proix From ludovico.montalcini at gmail.com Tue Feb 22 03:24:20 2022 From: ludovico.montalcini at gmail.com (Ludovico Montalcini) Date: Tue, 22 Feb 2022 09:24:20 +0100 Subject: Connectionists: CfP ACDL 2022, 5th Online & Onsite Advanced Course on Data Science & Machine Learning | August 22-26, 2022 | Certosa di Pontignano, Italy - Early Registration: by March 23 In-Reply-To: <030601d82748$ea223ab0$be66b010$@sch.gr> References: <030601d82748$ea223ab0$be66b010$@sch.gr> Message-ID: * Apologies for multiple copies. Please forward to anybody who might be interested * #ACDL2022, An Interdisciplinary Course: #BigData, #DeepLearning & #ArtificialIntelligence without Borders (Online attendance available) Certosa di Pontignano, Castelnuovo Berardenga (Siena) - #Tuscany, Italy August 22-26 https://acdl2022.icas.cc acdl at icas.cc ACDL 2022 (as ACDL 2021 and ACDL 2020): an #OnlineAndOnsiteCourse https://acdl2022.icas.cc/acdl-2022-as-acdl-2021-and-acdl-2020-an-online-onsite-course/ EARLY REGISTRATION: by March 23 https://acdl2022.icas.cc/registration/ DEADLINES: Early Registration: by Wednesday March 23 (AoE) Oral/Poster Presentation Submission Deadline: Wednesday March 23 (AoE) Late Registration: from Thursday March 24 Accommodation Reservation at Reservation at the Certosa di Pontignano: by July 20 Notification of Decision for Oral/Poster Presentation: by Thursday June 23 LECTURERS: Each Lecturer will hold three/four lessons on a specific topic. https://acdl2022.icas.cc/lecturers/ ?iga Avsec, DeepMind, London, UK Roman Belavkin, Middlesex University London, UK Alfredo Canziani, New York University, USA Alex Davies, DeepMind, London, UK Edith Elkind, University of Oxford, UK Marco Gori, University of Siena, Italy Danica Kragic Jensfel, Royal Institute of Technology, Sweden Yukie Nagai, The University of Tokyo, Japan Panos Pardalos, University of Florida, USA Silvio Savarese Salesforce & Stanford University, USA Mihaela van der Schaar, University of Cambridge, UK More Keynote Speakers to be announced soon. PAST LECTURERS: https://acdl2022.icas.cc/past-lecturers/ VENUE: The venue of ACDL 2022 will be The Certosa di Pontignano ? Siena The Certosa di Pontignano Localit? Pontignano, 5 ? 53019, Castelnuovo Berardenga (Siena) ? Tuscany ? Italy phone: +39-0577-1521104 fax: +39-0577-1521098 info at lacertosadipontignano.com https://www.lacertosadipontignano.com/en/index.php Contact person: Dr. Lorenzo Pasquinuzzi https://acdl2022.icas.cc/venue/ PAST EDITIONS: https://acdl2022.icas.cc/past-editions/ https://acdl2018.icas.xyz https://acdl2019.icas.xyz https://acdl2020.icas.xyz https://acdl2021.icas.cc REGISTRATION: https://acdl2022.icas.cc/registration/ ACDL 2022 POSTER: https://acdl2022.icas.cc/wp-content/uploads/sites/19/2022/02/poster-ACDL-2022.png Anyone interested in participating in ACDL 2022 should register as soon as possible. Similarly for accommodation at the Certosa di Pontignano (the Course Venue), book your full board accommodation at the Certosa as soon as possible. All course participants must stay at the Certosa di Pontignano. See you in 3D or 2D :) in Tuscany in August! ACDL 2022 Directors. https://acdl2022.icas.cc/category/news/ https://acdl2022.icas.cc/faq/ acdl at icas.cc https://acdl2022.icas.cc https://www.facebook.com/groups/204310640474650/ https://twitter.com/TaoSciences * Apologies for multiple copies. Please forward to anybody who might be interested * -------------- next part -------------- An HTML attachment was scrubbed... URL: From m.fairbank at essex.ac.uk Tue Feb 22 08:52:39 2022 From: m.fairbank at essex.ac.uk (Fairbank, Michael H) Date: Tue, 22 Feb 2022 13:52:39 +0000 Subject: Connectionists: Deep Learning in Target Space Message-ID: Dear Connectionists, We'd like to highlight to you our new paper "Deep Learning in Target Space" at JMLR: https://jmlr.org/papers/v23/20-040.html It offers a potentially new paradigm on how to train neural networks - to search through the space of activations of the hidden nodes (i.e. "target space") as opposed to the conventional backpropagation of the gradient in "weight space". We argue that this stabilises the gradient descent process, a process we call "cascade untangling", meaning deeper neural networks can be trained with less training data and potentially less CPU time. The work is an updated form of the Moving Targets algorithm by Rohwer (1990) and some follow up work, where we have fixed a lot of technical issues that existing in those earlier works. For some quick highlights of the work, see our paper's Figs 2 and 4. Results are included on deep RNN problems, CNN standard benchmarks, and some MLP experiments. The RNN experimental results are especially interesting in that we achieve performance of LSTM networks without using any memory gates - this hopefully motivates the paradigm shift from weight space to target space. Admittedly though, the target space training steps come with an extra computational cost. We would welcome discussion. Dr Michael Fairbank, and co-authors University of Essex UK -------------- next part -------------- An HTML attachment was scrubbed... URL: From suashdeb at gmail.com Tue Feb 22 09:17:53 2022 From: suashdeb at gmail.com (Suash Deb) Date: Tue, 22 Feb 2022 19:47:53 +0530 Subject: Connectionists: Soliciting your submission for ISCMI22 (Toronto) Message-ID: Dear friends and esteemed colleagues, This is abt ISCMI22, our 2022 9th flagship event to be held (hopefully on-site) in Toronto with technical co sponsorships of IEEE Toronto Section and IEEE WIE Toronto chapter. I solicit your own submissions from your end as well as your help in disseminating the info among your peers and help us in receiving a few quality manuscripts from them too. Apart from IEEE Conference Proceedings, all the previous ones of which have been successfully uploaded at lEEE Xplore, scope also exists for publication of the extended versions of some Conference papers at "Neural Computing and Applications", a reputed Springer Publication. For more information, pls. visit http://iscmi.us/ Thanks and with best regards, Suash Deb General Chair, ISCMI22 -------------- next part -------------- An HTML attachment was scrubbed... URL: From T.Nowotny at sussex.ac.uk Tue Feb 22 12:58:29 2022 From: T.Nowotny at sussex.ac.uk (Thomas Nowotny) Date: Tue, 22 Feb 2022 17:58:29 +0000 Subject: Connectionists: Call for proposals for workshops and tutorials at CNS*2022 Melbourne Message-ID: [cid:image001.png at 01D82815.659E0A30] Dear Connectionists, We are requesting proposals for workshops and tutorials at CNS*2022 from the international community of computational neuroscientists. The 31st Annual Computational Neuroscience Meeting, CNS*2022, will be held at Melbourne, Australia, 16-20 July 2022. Proposals from all levels of faculty as well as advanced postdoctoral fellows are welcome. Tutorials are to be held on the first day and workshops on the last two days of the CNS*2022 Meeting. This is an excellent opportunity to reach international audience of computational neuroscience researchers participating in this meeting. Please see the links below for detailed information. Questions regarding workshops and tutorials should be addressed to workshops at cnsorg.org or tutorials at cnsorg.org as appropriate. We are looking forward to your submissions. CALL FOR WORKSHOPS: https://www.cnsorg.org/cns-2022-call-for-workshops Priority Deadline for proposals: April 15 Jorge Mejias, Workshops Chair for CNS*2022 CALL FOR TUTORIALS: https://ocns.memberclicks.net/cns-2022-call-for-tutorials Deadline for proposals: April 15 Anca Doloc-Mihu, Tutorials Chair for CNS*2022 With kind regards, Thomas Nowotny, Vice President, on behalf of The OCNS Board of Directors -- Prof Thomas Nowotny Head of AI Research Group School of Engineering and Informatics University of Sussex Falmer, Brighton BN1 9QJ, UK Phone: +44-1273-678593 FAX: +44-1273-877873 -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 300500 bytes Desc: image001.png URL: From andreas.rowald at fau.de Tue Feb 22 11:00:25 2022 From: andreas.rowald at fau.de (Rowald, Andreas) Date: Tue, 22 Feb 2022 17:00:25 +0100 Subject: Connectionists: PhD opportunity at Chair of Digital Health FAU, Germany Message-ID: <0b9e4cda66c2c9cd4b647aec84bb1569@fau.de> To whom it may concern, the Chair of Digital Health at the Friedrich-Alexander-University Erlangen-Nuremberg, Germany invites applications for a Doctoral Researcher/PhD Student in computational modeling for personalized neuromodulation therapies. For more information please refer to the attached PDF document. Best regards, Andreas Rowald -- Andreas Rowald, Ph.D. Research Group Leader Chair of Digital Health Friedrich-Alexander University Erlangen-N?rnberg (FAU) Henkestra?e 91, 91052 Erlangen Tel.: +49 1590 6120662 http://www.cdh.med.fau.de -------------- next part -------------- A non-text attachment was scrubbed... Name: PhD_Call_CdH_FAU.pdf Type: application/pdf Size: 196335 bytes Desc: not available URL: From amir.kalfat at gmail.com Tue Feb 22 15:57:52 2022 From: amir.kalfat at gmail.com (Amir Aly) Date: Tue, 22 Feb 2022 20:57:52 +0000 Subject: Connectionists: =?utf-8?q?=5BMeetings=5D_CRNS_Talk_Series_=283=29?= =?utf-8?q?_-_Live_Talk_by_Dr=2E_S=C3=A9verin_Lemaignan_-_PAL_Robot?= =?utf-8?q?ics=2C_Spain?= Message-ID: Dear All * Apologies for cross-posting* The *Center for Robotics and Neural Systems* (CRNS) is pleased to announce the talk of *Dr. S?verin Lemaignan* who is a senior scientist at *PAL Robotics*, Spain on Wednesday, March 2nd from *11:00 am* to *12:30* *pm* (*London time*) over *Zoom*. >> *Events*: The CRNS talk series will cover a wide range of topics including social and cognitive robotics, computational neuroscience, computational linguistics, cognitive vision, machine learning, AI, and applications to autism. More details are available here: https://www.plymouth.ac.uk/research/robotics-neural-systems/whats-on >> *Link for the next event (No Registration is Required)*: Join Zoom Meeting https://plymouth.zoom.us/j/97451532624?pwd=ZmVSUzkrT3lUNWdXNmFEekNYSjRQUT09&from=addon >> *Title of the talk*: Teaching robots autonomy in social situations *Abstract*: Participatory methodologies are now well established in social robotics to generate blueprints of what robots should do to assist humans. The actual implementation of these blueprints, however, remains a technical challenge for us, roboticists, and the end-users are not usually involved at that stage. In two recent studies, we have however shown that, under the right conditions, robots can directly learn their behaviours from domain experts, replacing the traditional heuristic-based or plan-based robot controllers by autonomously learnt social policies. We have derived from these studies a novel 'end-to-end' participatory methodology called LEADOR, that I will introduce during the seminar. I will also discuss recent progress on human perception and modeling in a ROS environment with the emerging ROS4HRI standard. >> If you have any questions, please don't hesitate to contact me, Regards ---------------- *Dr. Amir Aly* Lecturer in Artificial Intelligence and Robotics Center for Robotics and Neural Systems (CRNS) School of Engineering, Computing, and Mathematics Room B332, Portland Square, Drake Circus, PL4 8AA University of Plymouth, UK -------------- next part -------------- An HTML attachment was scrubbed... URL: From oliver at roesler.co.uk Tue Feb 22 18:13:44 2022 From: oliver at roesler.co.uk (Oliver Roesler) Date: Tue, 22 Feb 2022 23:13:44 +0000 Subject: Connectionists: CFP Special Issue on Socially Acceptable Robot Behavior: Approaches for Learning, Adaptation and Evaluation Message-ID: <658dbe10-4015-1bea-280f-111591002006@roesler.co.uk> *CALL FOR PAPERS* **Apologies for cross-posting** * * *Special Issue* on *Socially Acceptable Robot Behavior: Approaches for Learning, Adaptation and Evaluation* in Interaction Studies *I. Aim and Scope* A key factor for the acceptance of robots as regular partners in human-centered environments is the appropriateness and predictability of their behavior. The behavior of human-human interactions is governed by customary rules that define how people should behave in different situations, thereby governing their expectations. Socially compliant behavior is usually rewarded by group acceptance, while non-compliant behavior might have consequences including isolation from a social group. Making robots able to understand human social norms allows for improving the naturalness and effectiveness of human-robot interaction and collaboration. Since social norms can differ greatly between different cultures and social groups, it is essential that robots are able to learn and adapt their behavior based on feedback and observations from the environment. This special issue in Interaction Studies aims to attract the latest research aiming at learning, producing, and evaluating human-aware robot behavior, thereby, following the recent RO-MAN 2021 Workshop on Robot Behavior Adaptation to Human Social Norms (TSAR) in providing a venue to discuss the limitations of the current approaches and future directions towards intelligent human-aware robot behaviors. *II. Submission* 1. Before submitting, please check the official journal guidelines . 2. For paper submission, please use the online submission system . 3. After logging into the submission system, please click on "Submit a manuscript" and select "Original article". 4. Please ensure that you select "Special Issue: Socially Acceptable Robot Behavior" under "General information". ??? The primary list of topics covers the following points (but not limited to): * Human-human vs human-robot social norms * Influence of cultural and social background on robot behavior perception * Learning of socially accepted behavior * Behavior adaptation based on social feedback * Transfer learning of social norms experience * The role of robot appearance on applied social norms * Perception of socially normative robot behavior * Human-aware collaboration and navigation * Social norms and trust in human-robot interaction * Representation and modeling techniques for social norms * Metrics and evaluation criteria for socially compliant robot behavior *III. Timeline* 1. Deadline for paper submission: *March 31, 2022*** 2. First notification for authors: *June 15, 2022* 3. Deadline for revised papers submission: *July 31, 2022* 4. Final notification for authors: *September 15, 2022* 5. Deadline for submission of camera-ready manuscripts: *October 15, 2022* ??? Please note that these deadlines are only indicative and that all submitted papers will be reviewed as soon as they are received. *IV. Guest Editors* 1. *Oliver Roesler* ? Vrije Universiteit Brussel ? Belgium 2. *Elahe Bagheri* ? Vrije Universiteit Brussel ? Belgium 3. *Amir Aly* ? University of Plymouth ? UK 4. *Silvia Rossi* ? University of Naples Federico II ? Italy 5. *Rachid Alami* ? CNRS-LAAS ? France -------------- next part -------------- An HTML attachment was scrubbed... URL: From louisphil.lopes at gmail.com Tue Feb 22 15:27:01 2022 From: louisphil.lopes at gmail.com (Phil Lopes) Date: Tue, 22 Feb 2022 20:27:01 +0000 Subject: Connectionists: [REMINDER] Post-Doc Junior Researcher in Signal Processing, Time Series Data Analysis and Machine Learning for Virtual Reality Message-ID: HEI-Lab is looking for a Junior Researcher proficient in Signal Processing, Time Series Data Analysis and Machine Learning. The person hired will be responsible for various phases of research and integrated in a multidisciplinary team of researchers at the crossroads of game design, virtual reality, biomedicine, psychology, and artificial intelligence. The core research field will be the construction of systems and experiences that leverage human physiological sensors for the adaptation of virtual content and capable of reporting detailed patient feedback for therapists during and post exposure. The majority of applications have a focus in using physiological monitoring technology such as ECG, EDA, EMG and EEG; as such applicants who already have some degree of experience with these technologies will be favoured. Responsibilities: - Take part of the experimental protocol design to optimize the data collection process and facilitate the processing methodology. - Clean, process and analyse time-series data, which can include the construction of tools and algorithms that aid this process for future projects. - Build statistical models through established techniques for the recognition and prediction of common human-based behaviours. - Become the signal processing ?expert? of the research unit, allowing fellow researchers to consult your expertise within the field. Remuneration: Gross monthly wage: 2134.73 euros, with correspondence to level 33 to the single remuneration table, which is updated based on the remuneration and the value table of the basic monthly remuneration as approved by Ordinance nr. 1553-C/2008, of December 31st, and combined with Decree-Law nr. 10/20201, of February 1st. Duration: The contract to be carried out is scheduled to start on April 1, 2022, ending on April 30, 2025. Location: The HEI-Lab itself is a research centre located at the Lus?fona University, in Lisbon, Portugal. *Applications are accepted until the 28th February 2022*, and must be sent by e-mail to micaela.fonseca at ulusofona.pt with ana.mourato at ulusofona.pt CC?d, with ?COFAC/ULHT/HEI-LAB/JR/2022? (without quotations) as the email subject and with the following documents attached: - Presentation letter referring to the reasons that justified the application; - Curriculum "vitae" mentioning professional experience, accompanied by a list of scientific publications produced and participation in funded projects; - Doctoral certificate; - Identification and contacts with the respective ?email? addresses - of at least two academic personalities who attest to the displayed curriculum; - Work plan for the 3-year period to be developed in the 'Human-Computer Interaction' area, compiled by the HEI-Lab I&D Labs - (https://hei-lab.ulusofona.pt); - Link to their Portfolio Page (e.g. GitHub/GitLab/Bitbucket, Personal Website, etc.) - Other documents considered relevant by the candidate and which, in his perspective, seem relevant to prove and evaluate the respective Further Details: Additional information about the proposal can be found here: https://euraxess.ec.europa.eu/jobs/738448. Questions on the proposal can be directed to micaela.fonseca at ulusofona.pt. -------------- next part -------------- An HTML attachment was scrubbed... URL: From antona at alleninstitute.org Tue Feb 22 19:55:03 2022 From: antona at alleninstitute.org (Anton Arkhipov) Date: Wed, 23 Feb 2022 00:55:03 +0000 Subject: Connectionists: Survey of the Allen Institute modeling software Message-ID: <4F2616DE-66A6-494E-A2B5-5E2DB61BFD96@alleninstitute.org> Dear Colleagues, We are conducting a survey about the Allen Institute?s neuroscience modeling software: https://forms.gle/R1ThdxBE98FhG1uQ6 The survey covers the Brain Modeling ToolKit, the SONATA file format, and the Visual Neuronal Dynamics visualization tool. It only takes several minutes, and we will appreciate to hear from you! Whether you are using our software, may be interested in using it, or are happily using something else, we will be glad to hear your opinion about our tools and modeling software in neuroscience in general. Please let us know what works, what doesn?t, and what new features would be helpful. Results of the survey will be used to improve our modeling software at the service of the community. This work is conducted under the NIH grant U24NS124001. Thank you, and we are looking forward to hearing from you! Anton. Anton Arkhipov Associate Investigator T: 206.548.8414 E: antona at alleninstitute.org [Text Description automatically generated] alleninstitute.org brain-map.org -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image001.png Type: image/png Size: 26603 bytes Desc: image001.png URL: From donatello.conte at univ-tours.fr Wed Feb 23 04:08:10 2022 From: donatello.conte at univ-tours.fr (Donatello Conte) Date: Wed, 23 Feb 2022 10:08:10 +0100 Subject: Connectionists: [jobs] Ph.D. position at University of Tours, France: Learning Spatio-temporal data by Graph Representations Message-ID: <005001d82894$e128fbe0$a37af3a0$@univ-tours.fr> Dear colleagues, We are accepting applications for a Ph. D. Position at LIFAT Laboratory in Tours, France. You can find all information here: http://www.rfai.lifat.univ-tours.fr/ph-d-position-10-2022-learning-spatio-te mporal-data-by-graph-representations/ ** Deadline for applications: April 8, 2022 ** Regards, ------------------------------------------------------------------- Donatello Conte Associate Professor Head of Computer Science Departement Ecole Polytechnique de l'Universit? de Tours Laboratoire d'Informatique Fondamentale et Appliqu?e de Tours (LIFAT) ------------------------------------------------------------------- -------------- next part -------------- An HTML attachment was scrubbed... URL: From avinashsingh214 at gmail.com Tue Feb 22 20:15:59 2022 From: avinashsingh214 at gmail.com (Avinash K Singh) Date: Wed, 23 Feb 2022 12:15:59 +1100 Subject: Connectionists: Special issue: Advances and Challenges to Bridge Computational Intelligence and Neuroscience for Brain-computer Interface Message-ID: Dear All, Brain-computer interfaces (BCIs) are a highly interdisciplinary research field consisting of researchers from neuroscience, mathematics, computer science, engineering, etc. They contribute to BCIs by developing and proposing new methods, techniques, BCI paradigms, brain signals recording methods, devices, and generating lots of information and data. Although most of the information is regularly available from peer-reviewed platforms and shared over open access data and method repository, it faces an incredible challenge in interpreting, reusing, comparing, and benchmarking. This challenge is growing significantly between computational intelligence and neuroscience research due to the current flow of readily available tools and devices. Most methods, data, techniques, etc., are available openly for the researcher to reuse. However, although openly available, such a multidisciplinary inclusion and their generated information create significant gaps in sharing methods and datasets, comparing results, and reproducing experiments. Such a gap exists because researchers only share domain-specific information that is not easy to interpret by researchers from other disciplines. Consider what would be needed to reproduce a steady-state visually evoked potential (SSVEP)-BCIs, besides sharing data. There is a need for information like the number of unique flickering stimuli presented to the user, the flickering rate, and signal processing specific details such as the impedance of electrodes, type of reference used, applied signal filters, appropriate labels, etc. These details are usually available in related publications but are hard to interpret for non-domain-specific researchers. This special issue aims to attract researchers from the multidisciplinary domain of BCI, particularly focused on computational intelligence and neuroscience, to provide their advances and challenges in solving the problem of bridging such an interdisciplinary research field. In this context, we welcome studies that help find a unique approach to solve the problem of unifying computational intelligence and the neuroscience community for BCI development. Therefore, we are looking for research studies on different techniques in machine learning, novel framework in BCI, unified format for terminologies and representation, automated tool to convert large open-source datasets, case studies of a converted dataset. We welcome original research, review, methods, and perspective articles that cover, but are not limited to, the following topics: - Novel frameworks for BCI data resharing - Unified functional models of BCI - Automated machine learning tools and pipelines to populate metadata in BCI - Methods, techniques, and tools to convert large open BCI dataset into a unified format - Benchmarking approaches for BCI Topic Editor(s): - Avinash Kumar Singh, University of Technology Sydney, Sydney, Australia - Luigi Bianchi, University of Rome Tor Vergata, Roma, Italy - Davide Valeriani, Neurable Inc., Boston, United States - Masaki Nakanishi, Institute for Neural Computation, University of California, San Diego, San Diego, United States Journal/Specialty: Frontiers in Neuroergonomics - section Neurotechnology and Systems Neuroergonomics Research Topic Title: Advances and Challenges to Bridge Computational Intelligence and Neuroscience for Brain-computer Interface Manuscripts can be submitted directly here: https://www.frontiersin.org/research-topics/27045 Here are quick links to: - Author guidelines: https://www.frontiersin.org/about/author-guidelines - List of article types and publishing fees: https://www.frontiersin.org/about/publishing-fees Kind Regards, Avinash Kumar Singh Topic Editor, Neurotechnology and Systems Neuroergonomics Section, Frontiers in Neuroergonomics -------------- next part -------------- An HTML attachment was scrubbed... URL: From r.pascanu at gmail.com Wed Feb 23 04:17:34 2022 From: r.pascanu at gmail.com (Razvan Pascanu) Date: Wed, 23 Feb 2022 09:17:34 +0000 Subject: Connectionists: CFP: 1st Conference on Lifelong Learning Agents (CoLLAs) 2022 - Deadline March 04 Message-ID: Dear all, *Apologies for crossposting* We invite submissions to the 1st Conference on Lifelong Learning Agents ( CoLLAs) that describe new theory, methodology or *new insights into existing algorithms and/or benchmarks*. Accepted papers will be published in the Proceedings of Machine Learning Research (PMLR). Topics of submission may include, but are not limited to, reinforcement learning, supervised learning or unsupervised learning approaches for: - Lifelong Learning / Continual Learning - Meta-Learning - Multi-task learning - Transfer Learning - Domain adaptation - Few-shot learning - Out-of-distribution generalization - Online Learning The conference also welcomes submissions at the intersection of machine learning and neuroscience and applications of the topics of interest to real-world problems. Submitted papers will be evaluated based on their novelty, technical quality, and potential impact. Experimental methods and results are expected to be reproducible, and authors are strongly encouraged to make code and data available. We also encourage submissions of proof-of-concept research that puts forward novel ideas and demonstrates potential, as well as in-depth analysis of existing methods and concepts. Key datesThe planned dates are as follows: - Abstract deadline: March 01, 2022, 11:59 pm (Anywhere on Earth, AoE) - Paper submission deadline: March 04, 2022, 11:59 pm (AoE) - Review released: April 08, 2022 - Author rebuttals due: April 15, 2022, 11:59 pm (AoE) - Notifications: May 06, 2022 - Resubmissions*: July 06, 2022, 11:59 pm (AoE) - Decisions on resubmissions* : August 06, 2022 - Tentative conference dates: August 29-31, 2022 *For more information, see the review process section. Review Process Papers will be selected via rigorous double-blind peer-review process. All accepted papers will be presented at the Conference as contributed talks or as posters and will be published in the Proceedings. The reviews process will be hosted on OpenReview with submissions and reviews being private until a decision is made. Reviews and discussions of the accepted papers will be made available after acceptance. In addition to accept/reject, a paper can be marked for revision and resubmission. In this case, the authors have a fixed amount of time to update the work and resubmit, to get a final accept/reject decision. If the paper was marked for resubmission but after resubmission it is rejected, then the work will automatically be accepted to a non-archival Workshop track of CoLLAs. The authors will still be able to present a poster on their work as part of this track. This system is aimed to produce a fairer treatment of borderline papers and to save the time spent in going through the entire reviewing process from scratch when resubmitting to a future edition of the conference or a different relevant conference. During the rebuttal period, authors are allowed to update their paper once. However, reviewers are not required to read the new version. Any paper which had a substantial update during this period will automatically go to the resubmission stage for a detailed re-review. Formatting and Supplementary Material Submissions should have a recommended length of 9 single-column CoLLAs-formatted pages, plus unlimited pages for references and appendices. We enforce a maximum length of 10 pages, where the 10th page can be used if it helps with the formatting of the paper. The appendices should be within the same pdf file as the main publication, however an additional zip file can be submitted that can include multiple files of different formats (e.g. videos or code). Note that reviewers are under no obligation to examine the appendix and the supplementary material. Please format the paper using the official LaTeX style files that can be found here: https://www.overleaf.com/read/grrjqdpztnpb. We do not support submissions in formats other than LaTeX. Please do not modify the layout given by the style file. For any questions, you can reach us at . Submissions will be through OpenReview ( https://openreview.net/group?id=lifelong-ml.cc/CoLLAs/2022/Conference). Complete CFP can be found here: https://lifelong-ml.cc/call Regards, Doina Precup, Sarath Chandar, Razvan Pascanu CoLLAs 2022 General and Program Chairs https://lifelong-ml.cc/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From giacomo.cabri at unimore.it Wed Feb 23 04:08:42 2022 From: giacomo.cabri at unimore.it (Giacomo Cabri) Date: Wed, 23 Feb 2022 10:08:42 +0100 Subject: Connectionists: CfP (EXTENDED) - Adaptive Computing (and Agents) for Enhanced Collaboration (ACEC) at WETICE 2022 Message-ID: <9fa80152-c410-cd79-2a33-118689ca6543@unimore.it> *20th Adaptive Computing (and Agents) for Enhanced Collaboration (ACEC)* Track?at IEEE WETICE 2022 Stuttgart, Germany, June 29-July 1, 2022 http://didattica.agentgroup.unimore.it/ACEC2022 Call for Papers **EXTENDED DEALINE: ***March 11th, 2022*** ** *Aims and Scope* Over its 19 years of existence, ACEC has focused on works?that explore?the adaptability, autonomy, and intelligence of software agents for the collaboration across the enterprise. In 2022,?organizers would like to continue to explore the research on agent-based computing, but they would also?welcome works that leverage advanced adaptive techniques, not necessarily based on software agents. In addition to the?traditional domains, i.e., Computer Supported Collaborative Work, Workflow and Supply Chain Management, Automation in Virtual Enterprises, and Automated Distributed Service Composition, ACEC is?also interested in new adaptive techniques, e.g.,?Cloud Computing, Crowd-Sourcing, and?Social Networking. In addition to traditional papers, the forthcoming 20th episode of ACEC welcomes papers from two focus areas: * Adaptive and Agent-based Services * Adaptive Techniques for Organizational/Enterprise Use of Emerging Web Paradigms (Cloud, Crowd sourcing, Mobile Apps) Such two themes represent important areas where software agents can leverage their distributed nature,?along with their proactive and autonomous characteristics, to provide solutions to?complex problems, which are difficult to solve?using traditional/existing technologies. *Topics of Interest* Topics of interest include, but are not limited to: * Adaptive and/or agent-mediated workflow, supply chain, and virtual?enterprises * Methodologies, languages, and tools to support agent collaboration * Agent architectures and infrastructures for dynamic collaboration * Adaptive and/or agent-based service architectures and infrastructures * Service-Oriented Architectures (SOAs) based on agents * Services for dynamic agent collaboration * Agent-to-human service interactions * Autonomous, adaptive and/or agent-mediated service integration * Organizational and enterprise systems that leverage the Web 2.0 * Adaptive and agent-mediated cloud environments *Important Dates* * Paper submission: *(extended)?March 11th, 2022* * Notification: *(extended)?April 1st, 2022* * Camera ready: *May 6th, 2022 * * Conference: *June?29-July 1, 2022* *Paper Submission* Papers should contain original contributions (not published or submitted elsewhere) and references to related state of the art. Please submit your papers in PDF or PS format. Papers up to 6 pages (including figures, tables, and references) can be submitted. Papers should follow the IEEE format, which is single-spaced, two columns, 10pt Times/Roman font. Papers should include a title, the name and affiliation of each author, an abstract of up to 150 words and no more than eight keywords. All submitted papers will be peer-reviewed by a minimum of three members of the Program Committee. Accepted papers will be published in the post-conference proceedings (to be published by IEEE Computer Society Press). Authors of accepted papers must present their papers at the conference.?At least one author for each accepted paper should register and attend WETICE 2022 to have the paper published in the proceedings. The paper submission procedure is carried out using the EasyChair conference management system: https://wetice2022.svccomp.de/#submissions *Track Chairs* * Federico Bergenti, Universit? degli Studi di Parma, Italy * M. Brian Blake, George Washington University, USA * Giacomo Cabri, Universit? degli Studi di Modena e Reggio Emilia, Italy * Stefania Monica, Universit? degli Studi di Modena e Reggio Emilia, Italy * Usman Wajid, The University of Manchester, UK From mtista at gmail.com Wed Feb 23 04:50:22 2022 From: mtista at gmail.com (Massimo Tistarelli) Date: Wed, 23 Feb 2022 10:50:22 +0100 Subject: Connectionists: Extended deadline: Call for participation to the 19th Int.l Summer School on Biometrics In-Reply-To: References: Message-ID: ?? Please accept our sincere apologies for receiving multiple copies of this announcement ?? *19th Int.l Summer School for Advanced Studies on Biometrics for Secure Authentication:* * CONTINUALLY LEARNING BIOMETRICS* *Alghero, Italy ? June 6 - 10 2022* http://biometrics.uniss.it To face the expected travel limitations due to the Covid-19 outbreak, the school is planned as a mixed virtual event, allowing participation both in person and remotely with videoconference facilities *Contact:* tista at uniss.it *EXTENDED Application deadline: March 1* *st 2022 *(download the application form at: http://biometrics.uniss.it) >From the early days, when security was the driving force behind biometric research, today?s challenges go far beyond security. Machine learning, Image understanding, Signal analysis, Neuroscience, Robotics, Forensic science, Digital forensics and other disciplines, converged in a truly multidisciplinary effort to devise and build advanced systems to facilitate the interpretation of signals recorded from individuals acting in a given environment. This is what we simply call today ?Biometrics?. For the last nineteen years, the International Summer School on Biometrics has been closely following the developments in science and technology to offer a cutting edge, intensive training course, always up to date with the current state-of-the-art. What are the most up-to-date core biometric technologies developed in the field? What is the potential impact of biometrics in forensic investigation and crime prevention? What can we learn from human perception? How to deploy current Machine Learning approaches? How to deal with *adversarial attacks* in biometric recognition? How can a biometric system learn continually? This school follows the successful track of the International Summer Schools on Biometrics held since 2003. In this 19th edition, the courses will mainly focus on new and emerging issues: ? *The impact of AI and advanced learning techniques in Biometrics;* ? *How to make ?Deep Biometrics? systems explainable;* ? *The advantages of continual learning for biometrics;* ? *How to exploit new biometric technologies in forensic and emerging applications.* The courses will provide a clear and in-depth picture on the state-of-the-art in biometric verification/identification technology, both under the theoretical and scientific point of view as well as in diverse application domains. The lectures will be given by 18 outstanding experts in the field, from both academia and industry. An advanced feature of this summer school will be some practical sessions to better understand, ?hands on?, the real potential of today?s biometric technologies. *Participant application* The expected school fees will be in the order of 1,600 ? (400 ? in videoconference) for students and 2,200 ? (800 ? in videoconference) for others. The fees will include full board accommodation, all courses and handling material. A limited number of scholarships, partially covering the fees, will be awarded to Phd students, selected on the basis of their scientific background and on-going research work. The scholarship request form can be downloaded from the school web site http://biometrics.uniss.it Send a filled application form (download from http://biometrics.uniss.it) together with a short resume to: *Prof. Massimo Tistarelli ?* e-mail: *biometricsummerschool at gmail.com* - Submission of applications: February 15th, 2022 - Notification of acceptance: March 20th, 2022 - Registration: April 25th, 2022 *Advance pre-registration is strictly required by March 1st 2022* *School location* The school will be hosted by Hotel Dei Pini (https://www.hoteldeipini.com/ ) in the Capo Caccia bay, near Alghero, Sardinia. This is one of the most beautiful resorts in the Mediterranean Sea. The structure is beautifully immersed into the Capo Caccia bay. The hotel Dei Pini has a recently renovated conference centre, fully equipped for scientific events. The school venue, as well as the surroundings, proved to be a perfect environment for the school activities. The school lectures will be delivered as a *mixed mode event*, allowing both physical (if the medical advice at the time of the school will allow it in full security) and remote attendance with videoconferencing facilities. - For participants attending in videoconference, live lectures will be delivered online with full sharing of the lecturing material and allowing live interaction with the participants. Private and group meetings with each lecturer will be organised to deepen the discussion started in the class. - Ad-hoc teleconferencing and communication tools will be also set up for practical hands-on sessions and to allow a good engagement of the participants and the lecturers. - Open sessions will be organised with questions and answers moderated by the leading experts in the field. The organisers and all lecturers are fully committed to make this year's school as successful, instructing and inspiring as in the past years. *School Committee:* *Massimo Tistarelli* Computer Vision Laboratory ? University of Sassari, Italy *Josef Bigun * Department of Computer Science ? Halmstad University, Sweden *Enrico Grosso* Computer Vision Laboratory ? University of Sassari, Italy *Anil K. Jain* Biometrics laboratory ? Michigan State University, USA *Distinguished lecturers from past school editions* *Josef **Bigun* Halmstad University ? Sweden *David Meuwly* Netherlands Forensic Institute ? NL *Thirimachos Bourlai* West Virginia University ? USA *Emilio Mordini MD* Responsible Technologies ? Italy *Vincent Bouatou* Safran Morpho ? France *Mark Nixon* University of Southampton ? UK *Kevin Bowyer * University of Notre Dame ? USA *Alice O?Toole* University of Texas ? USA *Deepak Chandra * Google Inc. ? USA *Maja Pantic* Imperial College ? UK *Rama Chellappa* University of Maryland ? USA *Johnathon Phillips* NIST ? USA *John Daugman* University of Cambridge ? UK *Tomaso Poggio* MIT ? USA *Farzin Deravi* University of Kent ? UK *Nalini Ratha* IBM ? USA *James Haxby* Dartmouth University ? USA *Arun Ross* Michigan State University ? USA *Anil K. Jain* Michigan State University ? USA *Tieniu Tan * CASIA-NLPR ? China *Joseph Kittler* University of Surrey ? UK *Massimo Tistarelli* Universit? di Sassari ? Italy *Davide Maltoni* Universit? di Bologna ? Italy *Alessandro Verri* Universit? di Genova ? Italy *John Mason* Swansea University ? UK *James Wayman* University of San Jos? ? USA *Aldo Mattei* Arma dei Carabinieri ? Italy *Lior Wolf* Tel Aviv University ? Israel -- *Dona il 5x1000* all'Universit? degli Studi di Sassari codice fiscale: 00196350904 -- Sent from Gmail Mobile -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: Summer School 2022 Announcement.pdf Type: application/pdf Size: 377615 bytes Desc: not available URL: From max.garagnani at gmail.com Wed Feb 23 05:28:42 2022 From: max.garagnani at gmail.com (Max Garagnani) Date: Wed, 23 Feb 2022 10:28:42 +0000 Subject: Connectionists: Applications for 2022-23 entry OPEN:: MSc in Computational Cognitive Neuroscience, Goldsmiths (London, UK) References: Message-ID: Apologies for cross-posting ? Please could you forward the following announcement to any potentially interested party: ********************************************************************************** The MSc in COMPUTATIONAL COGNITIVE NEUROSCIENCE at Goldsmiths, University of London (UK) ********************************************************************************** is now ACCEPTING APPLICATIONS for 2022-23 ENTRY. Note that places on this programme are limited and will be allocated on a first-come first-served basis. If you are considering this MSc, we recommend applying now rather than later to avoid disappointment. The course builds on the multi-disciplinary and strong research profiles of our Computing and Psychology Departments staff. It equips students with a solid theoretical basis and experimental techniques in computational cognitive neuroscience, providing them also with an opportunity to apply their newly acquired knowledge in a practical research project, which may be carried out in collaboration with one of our industry partners (see below). Applications range from computational neuroscience and machine learning to brain-computer interfaces to experimental and clinical research. For more INFORMATION ABOUT THE COURSE please visit: https://www.gold.ac.uk/pg/msc-computational-cognitive-neuroscience/ HOW TO APPLY: ============= Submitting an online application is easy and free of cost. Simply visit https://bit.ly/2Fi86SB and follow the instructions. COURSE OUTLINE: =============== This is a one-year full-time or two-years part-time Masters programme, consisting of taught courses (120 credits) plus research project and dissertation (60 credits). (Note: students who need a Tier-4 VISA to study in the UK can only register for the full-time pathway). It is designed for students with a good degree in the biological / life sciences (psychology, neuroscience, biology, medicine, etc.) or physical sciences (computer science, mathematics, physics, engineering); however, individuals with different backgrounds but equivalent experience will also be considered. The core contents of this course include (i) fundamentals of cognitive neuroscience (cortical and subcortical mechanisms and structures underlying cognition and behaviour, plus experimental and neuroimaging techniques), and (ii) concepts and methods of computational modelling of biological neurons, simple neuronal circuits, and higher brain functions. Students are trained with a rich variety of computational and advanced methodological skills, taught in the four core modules of the course (Modelling Cognitive Functions, Cognitive Neuroscience, Cortical Modelling, and Advanced Quantitative Methods). Unlike other standard computational neuroscience programmes (which focus predominantly on modelling low-level aspects of brain function), one of the distinctive features of this course is that it includes the study of biologically constrained models of cognitive processes (including, e.g., language and decision making). The final research project can be carried out 'in house' or in collaboration with an external partner, either from academia or industry. For samples of previous students' MSc projects, visit: https://coconeuro.com/index.php/student-projects/ For information about funding opportunities and tuition fees, please visit: https://www.gold.ac.uk/pg/fees-funding/ LINKS WITH INDUSTRY: ==================== The programme benefits from ongoing collaborative partnerships with several companies having headquarters in UK, USA, Germany, Italy, and Japan. Carrying out your final research project with one of our industry partners will enable you to acquire cutting-edge skills which are in demand, providing you with a competitive profile on the job market and paving the way towards post-Masters internships and job opportunities. Here are examples of career pathways that our alumni have taken: https://coconeuro.com/index.php/alumni/ For any other specific questions, please do not hesitate to get in touch. Kind regards, Max Garagnani -- Joint Programme Leader, MSc in Computational Cognitive Neuroscience Senior Lecturer in Computer Science Department of Computing Goldsmiths, University of London Lewisham Way, New Cross London SE14 6NW, UK https://www.gold.ac.uk/computing/people/garagnani-max/ ******************************************************************************* -------------- next part -------------- An HTML attachment was scrubbed... URL: From m.biehl at rug.nl Wed Feb 23 06:20:02 2022 From: m.biehl at rug.nl (Michael Biehl) Date: Wed, 23 Feb 2022 12:20:02 +0100 Subject: Connectionists: Fully funded PhD position, University of Groningen / NL Message-ID: *Last reminder, j**ust a few days left to apply! * A *fully funded PhD position* (4 years) in the *Statistical **Physics of Neural Networks* is available at the University of Groningen, The Netherlands, see https://www.rug.nl/about-ug/work-with-us/job-opportunities/?details=00347-02S0008WFP for details and application details. Applications *(before March 1)* are only possible through this webpage. The title of the project is "The role of the activation function for feedforward learning systems (RAFFLES)". For further information please contact Michael Biehl. ---------------------------------------------------------- Prof. Dr. Michael Biehl Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence P.O. Box 407, 9700 AK Groningen The Netherlands Tel. +31 50 363 3997 https://www.cs.rug.nl/~biehl m.biehl at rug.nl -- ---------------------------------------------------------- Prof. Dr. Michael Biehl Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence P.O. Box 407, 9700 AK Groningen The Netherlands Tel. +31 50 363 3997 https://www.cs.rug.nl/~biehl m.biehl at rug.nl -------------- next part -------------- An HTML attachment was scrubbed... URL: From ioannakoroni at csd.auth.gr Wed Feb 23 06:33:28 2022 From: ioannakoroni at csd.auth.gr (Ioanna Koroni) Date: Wed, 23 Feb 2022 13:33:28 +0200 Subject: Connectionists: Asynchronous Web e-Courses on Machine Learning and Neural Networks/Deep Learning on offer. Free access to course material References: <000601d82362$a8755aa0$f9600fe0$@csd.auth.gr> Message-ID: <040a01d828a9$2de5ab40$89b101c0$@csd.auth.gr> Dear Machine Learning, Computer Vision and Autonomous Systems Engineers, Scientists and Enthusiasts, you are welcomed to register and attend the Web e-Course on Computer Vision consisting of the following two CVML Web e-Course Modules on offer (total 22 lectures): Machine Learning (12 Lectures), http://icarus.csd.auth.gr/machine-learning-web-module/ 1) Introduction to Machine Learning 2) Data Clustering 3) Distance-based Classification 4) Decision Surfaces. Support Vector Machines 5) Label Propagation 6) Dimensionality Reduction 7) Graph-based Dimensionality Reduction 8) Kernel Methods 9) Bayesian Learning 10) Parameter Estimation 11) Hypothesis Testing 12) Syntactic Pattern Recognition Neural Networks/Deep Learning (14 Lectures), http://icarus.csd.auth.gr/neural-networks-and-deep-learning-web-module/ 1. Artificial Neural Networks. Perceptron 2. Multilayer Perceptron. Backpropagation. 3. Convolutional Neural Networks 4. 1D Convolutional Neural Networks 5. Deep Autoencoders 6. Attention and Transformer Networks 7. Recurrent Neural Networks. LSTMs 8. Deep Object Detection 9. Special topics in Object Detection 10. Few-Shot Object Recognition 11. Deep Semantic Image Segmentation 12. Adversarial Machine Learning 13. Generative Adversarial Networks in Multimedia Creation 14. Mathematical Brain Modeling Around 50% of the lectures provide free access to the full lecture pdf! You can find sample Web e-Course Module material to make up your mind and/or can perform CVML Web e-Course registration in: http://icarus.csd.auth.gr/cvml-web-lecture-series/ For questions, please contact: Ioanna Koroni > More information on this Web e-course: This Web e-Course Computer Vision material that can cover a semester course, but you can master it in approximately 1 month. Course materials are at senior undergraduate/MSc level in a CS, CSE, EE or ECE or related Engineering or Science Department. Their structure, level and offer are completely different from what you can find in either Coursera or Udemy. CVML Web e-Course Module materials typically consist of: a) a lecture pdf/ppt, b) lecture self-assessment understanding questionnaire and lecture video, programming exercises, tutorial exercises (for several modules/lectures) and overall course module satisfaction questionnaire. Asynchronous tutor support will be provided in case of questions. Course materials have been very successfully used in many top conference keynote speeches/tutorials worldwide and in short courses, summer schools, semester courses delivered by AIIA Lab physically or on-line from 2018 onwards, attracting many hundreds of registrants. More information on other CVML Web e-course: Several other Web e-Course Modules are offered on Deep Learning, Computer Vision, Autonomous Systems, Signal/Image/Video Processing, Human-centered Computing, Social Media, Mathematical Foundations, CVML SW tools. See: http://icarus.csd.auth.gr/cvml-web-lecture-series/ You can combine CVML Web e-Course Modules to create CVML Web e-Courses (typically consisting of 16 lectures) of your own choice that cater your personal education needs. Each CVML Web e-Course you will create (typically 16 lectures) provides you material that can cover a semester course, but you can master it in approximately 1 month. Academic/Research/Industry offer and arrangements Special arrangements can be made to offer the material of these CVML Web e-Course Modules at University/Department/Company level: * by granting access to the material to University/research/industry lecturers to be used as an aid in their teaching, * by enabling class registration in CVML Web e-Courses * by delivering such live short courses physically or on-line by Prof. Ioannis Pitas * by combinations of the above. The CVML Web e-Course is organized by Prof. I. Pitas, IEEE and EURASIP fellow, Coordinator of International AI Doctoral Academy (AIDA), past Chair of the IEEE SPS Autonomous Systems Initiative, Director of the Artificial Intelligence and Information analysis Lab (AIIA Lab), Aristotle University of Thessaloniki, Greece, Coordinator of the European Horizon2020 R&D project Multidrone. He is ranked 249-top Computer Science and Electronics scientist internationally by Guide2research (2018). He has 34100+ citations to his work and h-index 87+. The informatics Department at AUTH ranked 106th internationally in the field of Computer Science for 2019 in the Leiden Ranking list Relevant links: 1. Prof. I. Pitas: https://scholar.google.gr/citations?user=lWmGADwAAAAJ &hl=el 2. International AI Doctoral Academy (AIDA): https://www.i-aida.org/ 3. Horizon2020 EU funded R&D project Aerial-Core: https://aerial-core.eu/ 4. Horizon2020 EU funded R&D project Multidrone: https://multidrone.eu/ 5. Horizon2020 EU funded R&D project AI4Media: https://ai4media.eu/ 6. AIIA Lab: https://aiia.csd.auth.gr/ Sincerely yours Prof. I. Pitas Director of the Artificial Intelligence and Information analysis Lab (AIIA Lab) Aristotle University of Thessaloniki, Greece Post scriptum: To stay current on CVML matters, you may want to register to the CVML email list, following instructions in https://lists.auth.gr/sympa/info/cvml -- This email has been checked for viruses by Avast antivirus software. https://www.avast.com/antivirus -------------- next part -------------- An HTML attachment was scrubbed... URL: From jakobjordan at posteo.de Wed Feb 23 07:46:27 2022 From: jakobjordan at posteo.de (Jakob Jordan) Date: Wed, 23 Feb 2022 12:46:27 +0000 Subject: Connectionists: =?utf-8?q?14=E1=B5=97=CA=B0_Advanced_Scientific_P?= =?utf-8?q?rogramming_in_Python_in_Bilbao_Spain=2C_5=E2=80=9311_September?= =?utf-8?q?=2C_2022?= Message-ID: ASPP2022: 14?? Advanced Scientific Programming in Python ======================================================== a Summer School by the ASPP faculty and the Faculty of Engineering of the Mondragon University, Bilbao https://aspp.school Scientists spend more and more time writing, maintaining, and debugging software. While techniques for doing this efficiently have evolved, only few scientists have been trained to use them. As a result, instead of doing their research, they spend far too much time writing deficient code and reinventing the wheel. In this course we will present a selection of advanced programming techniques and best practices which are standard in the industry, but especially tailored to the needs of a programming scientist. Lectures are devised to be interactive and to give the students enough time to acquire direct hands-on experience with the materials. Students will work in pairs throughout the school and will team up to practice the newly learned skills in a real programming project ? an entertaining computer game. We use the Python programming language for the entire course. Python works as a simple programming language for beginners, but more importantly, it also works great in scientific simulations and data analysis. We show how clean language design, ease of extensibility, and the great wealth of open source libraries for scientific computing and data visualization are driving Python to become a standard tool for the programming scientist. This school is targeted at Master or PhD students and Post-docs from all areas of science. Competence in Python or in another language such as Java, C/C++, MATLAB, or R is absolutely required. Basic knowledge of Python and of a version control system such as git, subversion, mercurial, or bazaar is assumed. Participants without any prior experience with Python and/or git should work through the proposed introductory material before the course. We are striving hard to get a pool of students which is international and gender-balanced. Date & Location =============== 5?11 September, 2022. Bilbao, Spain. Application =========== You can apply online: https://aspp.school Application deadline: 23:59 UTC, Sunday 1 May, 2022. There will be no deadline extension, so be sure to apply on time. Be sure to read the FAQ before applying: https://aspp.school/wiki/faq Participation is for free, i.e. no fee is charged! Participants however should take care of travel, living, and accommodation expenses by themselves. We are in the process of securing some funds for supporting students with accommodation and living costs. Program ======= ? Version control with git and how to contribute to open source projects with GitHub ? Best practices in data visualization ? Testing and debugging scientific code ? Advanced NumPy ? Organizing, documenting, and distributing scientific code ? Advanced scientific Python: context managers and generators ? Writing parallel applications in Python ? Profiling and speeding up scientific code with Cython and numba ? Programming in teams Faculty ======= ? Jakob Jordan, Department of Physiology, University of Bern Switzerland ? Jenni Rinker, Department of Wind Energy, Technical University of Denmark, Roskilde Denmark ? Lisa Schwetlick, Experimental and Biological Psychology, Universit?t Potsdam Germany ? Nicolas Rougier, Inria Bordeaux Sud-Ouest, Institute of Neurodegenerative Diseases, University of Bordeaux France ? Pamela Hathway, GfK, Nuremberg Germany ? Pietro Berkes, NAGRA Kudelski, Lausanne Switzerland ? Rike-Benjamin Schuppner, Institute for Theoretical Biology, Humboldt-Universit?t zu Berlin Germany ? Tiziano Zito, Department of Psychology, Humboldt-Universit?t zu Berlin Germany ? Zbigniew J?drzejewski-Szmek, Red Hat Inc., Warsaw Poland Organizers ========== Head of the organization for ASPP and responsible for the scientific program: ? Tiziano Zito, Department of Psychology, Humboldt-Universit?t zu Berlin Germany Organization team in Bilbao: ? Aitor Morales-Gregorio, Theoretical Neuroanatomy, Institute of Neuroscience and Medicine (INM-6), Forschungszentrum J?lich, Germany ? Carlos Cernuda, Data Analysis & Cybersecurity, Faculty of Engineering, Mondragon Unibertsitatea, Bilbao Spain Website: https://aspp.school Contact: info at aspp.school From angelo.cangelosi at manchester.ac.uk Wed Feb 23 08:12:20 2022 From: angelo.cangelosi at manchester.ac.uk (Angelo Cangelosi) Date: Wed, 23 Feb 2022 13:12:20 +0000 Subject: Connectionists: Call for Papers - IEEE International Conference on Development and Learning ICDL 2022, (12th edition of the ICDL-EPIROB Conference). Message-ID: IEEE International Conference on Development and Learning ICDL 2022, (12th edition of the ICDL-EPIROB Conference). Conference dates: September 12-15, 2022 Web page: http://icdl-2022.org An IEEE conference sponsored by CIS (Computational Intelligence Society) Location: London, UK Deadline for Submission of Regular Papers, Tutorial and Workshops Proposals, and Journal Track Papers: March 18th , 2022 (24:00 PST)" ==== Overview ==== The IEEE International Conference on Development and Learning (ICDL), previously referred to as ICDL-EpiRob, is the premier gathering of professionals dedicated to the advancement of cognitive and developmental learning. As such, ICDL is a unique conference gathering researchers from computer science, robotics, psychology and developmental science. We invite submissions for the conference in 2022 to explore, extend, and consolidate the interdisciplinary boundaries of this exciting research field, in the form of full papers, workshop and tutorials and with submissions to the journal track. We also invite your participation in the SmartBot Challenge and in the REAL 2022 competition. More details on the call can be found at: https://icdl2022.qmul.ac.uk/?page_id=107 ==== Topics ==== The primary list of topics of interest include (but are not limited to): ? General principles of development and learning; ? Development of skills in biological systems and robots; ? Nature VS nurture, critical periods and developmental stages; ? Models on active learning; ? Architectures for cognitive development and life-long learning; ? Emergence of body knowledge and affordance perception; ? Analysis and modelling of human motion and state; ? Models for prediction, planning and problem solving; ? Models of human-human and human-robot interaction; ? Emergence of verbal and non-verbal communication skills; ? Epistemological foundations and philosophical issues; ? Models of child development from experimental psychology. ==== Full Papers ==== Papers of at most 6 pages in IEEE double column format will undergo peer-review, and accepted and presented submissions will be included in the conference proceedings published by IEEE Xplore. The authors of the best conference papers will be invited to extend their contributions for a Special Issue of IEEE Transaction on Cognitive and Developmental Systems (IEEE TCDS). Maximum two-extra pages can be acceptable for a publication fee of $100 per page. Contributed papers can participate to the SmartBot Challenge, if they meet the eligibility requirements (see below). Authors should tick the appropriate box during submission if they want their paper to be considered for the SmartBot Challenge. The detailed instructions and templates for the submission can be found here: https://icdl2022.qmul.ac.uk/?page_id=209 ==== Workshops and Tutorials === We invite experts in different areas to organize either a tutorial or a workshop to be held on the first day of the conference (September 12, 2022). Tutorials are meant to provide insights into specific topics through hands-on training and interactive experiences. Workshops are exciting opportunities to present a focused research topic cumulatively. Tutorials and workshops can be half- or full-day in duration including oral presentations, posters and live demonstrations. Submission format: two double-column pages in standard IEEE format including title, duration (half day or full day), concept, target audience, list of speakers, open to paper/poster submission, website link. The detailed instructions and templates for the submission can be found here: https://icdl2022.qmul.ac.uk/?page_id=108 ==== Journal Track ==== Authors are encouraged to submit applications for Journal Track talks. The application must be about a journal paper that has been published already, in the March 2021 ? March 2022 period, on a topic relevant to ICDL. All submissions must be made by email before March 18th 2022, including (in a single PDF): ? Title of the original journal paper. ? Abstract of the original journal paper. ? A complete reference to the original paper in APA format. ? URL where the paper can be shown to be formally published by the publisher (even if early access). ? URL where the paper with its final camera-ready contents can be freely download for the evaluation process. ? A brief description (no more than half-page) to explain why the authors believe that the paper is relevant to ICDL. The ICDL committee will select a limited number of applications through an expedite evaluation process that will consider the quality of the journal paper and the relevance to ICDL. The authors of the selected applications will be invited to present their work (oral presentation) during a dedicated session at the conference. The detailed instructions for the submission can be found here: https://icdl2022.qmul.ac.uk/?page_id=109 ==== Competitions ==== We also strongly expect your participation in the SmartBot and the REAL competitions. The SmartBot Challenge is designed to help strengthen the bridge between two research communities: those who study learning and development in humans and those who study comparable processes in artificial systems. When submitting your 6-pages paper to ICDL 2022 you can specify that you want it to participate in the challenge. To be eligible, your paper should describe a computational or robotic model that explains one or several studies from the infant development literature. More details on the criteria to be eligible for the challenge here: https://icdl2022.qmul.ac.uk/?page_id=241 The REAL competition addresses open-ended learning with a focus on "Robot open-Ended Autonomous Learning" (REAL)", that is, on systems that: (a) acquire sensorimotor competence that allows them to interact with objects and physical environments; (b) learn in a fully autonomous way, i.e. with no human intervention. The competition has a two-phase structure: during the first ?intrinsic phase? your model has a certain time to explore and learn in the environment freely. Then during the second ?extrinsic phase? the quality of the knowledge acquired in the intrinsic phase is measured with tasks unknown to the robot during this autonomous phase. Find all the details to participate here: https://icdl2022.qmul.ac.uk/?page_id=243 ==== Important Dates ==== Conference dates: September 12-15, 2022 Workshops/Tutorials Date: September 12th, 2022 Submission deadline: March 18th, 2022 (24:00 PST) (applies to Papers, Workshop and Tutorial proposals and Journal Track submissions). Paper Author Notification: May 15th, 2022 Paper Camera Ready Due: June 12th, 2022 Workshop and Tutorial Notification: April 24th, 2022 Workshop and Tutorial Camera Ready Due: May 22th, 2022 Journal Track Papers Notification: May 15th, 2022 SmartBot Challenge Submission Deadline: March 18th, 2022 (24:00 PST) REAL Competition: See dedicated website - https://eval.ai/web/challenges/challenge-page/1134/overview ================= Best regards, The ICDL 2022 Organizing Committee. https://icdl2022.qmul.ac.uk/?page_id=37 General Chairs Lorenzo Jamone (Queen Mary Univ. of London, UK) Yukie Nagai (Univ. of Tokyo, Japan) -------------- next part -------------- An HTML attachment was scrubbed... URL: From ludovico.montalcini at gmail.com Wed Feb 23 09:23:27 2022 From: ludovico.montalcini at gmail.com (Ludovico Montalcini) Date: Wed, 23 Feb 2022 15:23:27 +0100 Subject: Connectionists: CfP ACAIN 2022, 2nd Int. Online & Onsite Advanced Course & Symposium on Artificial Intelligence & Neuroscience, Sept 18-22, Certosa di Pontignano, Tuscany - Italy In-Reply-To: References: Message-ID: ________________________________________________________________________ Call for Participation & Call for Papers (apologies for cross-postings) Please distribute this call to interested parties, thanks ________________________________________________________________________ The 2nd International Online & Onsite Advanced Course & Symposium on #ArtificialIntelligence & #Neuroscience - #ACAIN2022 September 18-22, 2022 Certosa di Pontignano, Castelnuovo Berardenga (Siena), #Tuscany - Italy LECTURERS: * Marvin M. Chun, Yale University, USA * Ila Fiete, MIT, USA * Karl Friston, University College London, UK & Wellcome Trust Centre for Neuroimaging * Wulfram Gerstner, EPFL, Switzerland * M?t? Lengyel, Cambridge University, UK * Max Erik Tegmark, MIT, USA & Future of Life Institute * Michail Tsodyks, Institute for Advanced Study, USA More Lecturers and Speakers to be announced soon! W: https://acain2022.artificial-intelligence-sas.org E: acain at icas.cc NEWS: https://acain2022.artificial-intelligence-sas.org/category/news/ Past Edition: https://acain2021.artificial-intelligence-sas.org Early Registration (Course): by March 23, 2022 (AoE) https://acain2022.artificial-intelligence-sas.org/registration/ Paper Submission (Symposium) : by Saturday April 23, 2022 (AoE) https://acain2022.artificial-intelligence-sas.org/symposium-call-for-papers/ https://easychair.org/conferences/?conf=acain2022 SCOPE & MOTIVATION: The ACAIN 2022 symposium is an interdisciplinary event featuring leading scientists from AI and Neuroscience, providing a special opportunity to learn about cutting-edge research in the fields. While the Advanced Course and Symposium on Artificial Intelligence & Neuroscience (ACAIN) is a full-immersion residential (or online) Course and Symposium at the Certosa di Pontignano (Tuscany - Italy) on cutting-edge advances in Artificial Intelligence and Neuroscience with lectures delivered by world-renowned experts. The Course provides a stimulating environment for academics, early career researchers, Post-Docs, PhD students and industry leaders. Participants will also have the chance to present their results with oral talks or posters, and to interact with their peers, in a friendly and constructive environment. Two days of keynote talks and oral presentations, the ACAIN Symposium, (September 21-22), will be preceded by lectures of leading scientists, the ACAIN Course, (September 18-20). Bringing together AI and neuroscience promises to yield benefits for both fields. The future impact and progress in both AI and Neuroscience will strongly depend on continuous synergy and efficient cooperation between the two research communities. These are the goals of the International Course and Symposium ? ACAIN 2022, which is aimed both at AI experts with interests in Neuroscience and at neuroscientists with an interest in AI. ACAIN 2022 accepts rigorous research that promotes and fosters multidisciplinary interactions between artificial intelligence and neuroscience. The Event (Course and Symposium) will involve a total of 36-40 hours of lectures. Academically, this will be equivalent to 8 ECTS points for the PhD Students and the Master Students attending the Event. COURSE DESCRIPTION: https://acain2022.artificial-intelligence-sas.org/course-description/ LECTURERS: https://acain2022.artificial-intelligence-sas.org/course-lecturers/ * Marvin M. Chun, Yale University, USA * Ila Fiete, MIT, USA * Karl Friston, University College London, UK & Wellcome Trust Centre for Neuroimaging * Wulfram Gerstner, EPFL, Switzerland * M?t? Lengyel, Cambridge University, UK * Max Erik Tegmark, MIT, USA & Future of Life Institute * Michail Tsodyks, Institute for Advanced Study, USA More Lecturers and Speakers to be announced soon! ORGANIZING COMMITTEE: https://acain2022.artificial-intelligence-sas.org/organizing-committee/ VENUE & ACCOMMODATION: https://acain2022.artificial-intelligence-sas.org/venue/ https://acain2022.artificial-intelligence-sas.org/accommodation/ The venue of ACAIN 2022 will be The Certosa di Pontignano ? Siena The Certosa di Pontignano Localit? Pontignano, 5 ? 53019, Castelnuovo Berardenga (Siena) ? Tuscany ? Italy phone: +39-0577-1521104 fax: +39-0577-1521098 info at lacertosadipontignano.com https://www.lacertosadipontignano.com/en/index.php Contact person: Dr. Lorenzo Pasquinuzzi You need to book your accommodation at the venue and pay the amount for accommodation directly to the Certosa di Pontignano. ACTIVITIES: https://acain2022.artificial-intelligence-sas.org/activities/ REGISTRATION: https://acain2022.artificial-intelligence-sas.org/registration/ See you in 3D or 2D :) in Tuscany in September! ACAIN 2022 Directors. POSTER: https://acain2022.artificial-intelligence-sas.org/wp-content/uploads/sites/21/2022/02/poster-ACAIN-2022.png NEWS: https://acain2022.artificial-intelligence-sas.org/category/news/ E: acain at icas.cc W: https://acain2022.artificial-intelligence-sas.org Past Edition, ACAIN 2021: https://acain2021.artificial-intelligence-sas.org * Apologies for multiple copies. Please forward to anybody who might be interested * -------------- next part -------------- An HTML attachment was scrubbed... URL: From kkuehnbe at uos.de Wed Feb 23 10:39:42 2022 From: kkuehnbe at uos.de (=?UTF-8?Q?Kai-Uwe_K=c3=bchnberger?=) Date: Wed, 23 Feb 2022 16:39:42 +0100 Subject: Connectionists: Research Positions available at the Institute of Cognitive Science, Onsabrueck University Message-ID: The Artificial IntelligenceResearch Group (Prof. Dr. Kai-Uwe K?hnberger) of the? Institute of Cognitive Science of Osnabr?ck University invites applications for *3 Research Associates (m/f/d)****** ??? ??? ??? ??? ??? ??? ?? ?? ??? ?? ??? ??? (Salary level E 13 TV-L, up to 100%)*** to be filled as soon as possible for a period of initially three years. The positions allow for further scientific qualification and can be held full-time orpart-time. *Description of Responsibilities:*** The position involves participation in the research activities of the ArtificialIntelligence Group with an emphasis on the areas of cognitively inspired approachesto knowledge representation, machine learning, computational creativity, e-learning,and cognitive architectures. Both, a full-time and part-time position includes teaching Cognitive Science courses at B.Sc. and M.Sc. level (4 hours/week for a full-timeposition). *Required Qualifications:*** - Applicants are expected to have an excellent academic degree (Master/Diploma) and may also have a PhD degree, - experience and interest in several of the domains listed above, - basic knowledge in at least two of the following areas: formal logic andreasoning, development of algorithms, (deep) neural networks and machinelearning, probabilistic approaches to AI, - good programming skills and experience with libraries (e.g. Prolog, Scheme,Python, TensorFlow, Keras) are mandatory, - a good command of the English language is imperative. Osnabr?ck University as a family-friendly university is committed to helpingworking/studying parents and carers balance their family and work life.Osnabr?ck University seeks to guarantee equality of opportunity for women andmen and strives to correct any gender imbalance in its schools and departments. If two candidates are equally qualified, preference will be given to the candidate withdisability status. Applications with the usual documentation should be submitted by e-mail in a*single***PDF-file to Prof. Dr. K.-U. K?hnberger (kkuehnbe at uni-osnabrueck.de) with a cc tooffice at ikw.uni-osnabrueck.deno later than*March 11, 2022*. Further informationcan be obtained from Prof. Dr. Kai-Uwe K?hnberger (kkuehnbe at uni-osnabrueck.de) -------------- next part -------------- An HTML attachment was scrubbed... URL: From calendarsites at insticc.org Wed Feb 23 10:15:30 2022 From: calendarsites at insticc.org (calendarsites at insticc.org) Date: Wed, 23 Feb 2022 15:15:30 -0000 Subject: Connectionists: [CFP] 19th Int. Conf. on Signal Processing and Multimedia Applications :: Submission Deadline - 2nd of March Message-ID: <034301d828c8$36ba9f40$a42fddc0$@insticc.org> CALL FOR PAPERS 19th International Conference on Signal Processing and Multimedia Applications **Submission Deadline: March 2, 2022** https://sigmap.scitevents.org July 14 - 16, 2022 Lisbon, Portugal Important Note: The conference will be held in Lisbon but we are open to accept online presentations in case the participants can't attend the conference. The purpose of SIGMAP 2022, the International Conference on Signal Processing and Multimedia Applications, is to bring together researchers, engineers and practitioners interested on information systems and applications, including theory and practice in various heterogeneous and interrelated fields including image, video and audio data processing, new sources of multimodal data (text, social, health, etc.) and Multimedia Applications related to representation, storage, authentication and communication of multimedia information. Multimedia is a research field that includes computing methods in which different modalities are integrated and combined, with the aim to take advantage from each data source. SIGMAP is organized in 6 major tracks: 1 - Multimedia Networking and Communication 2 - Multimedia Signal Processing 3 - Multimedia Systems and Applications 4 - Multimedia and Deep Learning 5 - Multimedia Indexing and Retrieval 6 - Social Multimedia Conference Chair(s) Andrew Sung, University of Southern Mississippi, United States Program Chair(s) Simone Santini, Universidad Aut?noma de Madrid, Spain In the last years, the proceedings have been fully indexed by SCOPUS. Beside this index all the proceedings have also been submitted to Google Scholar, The DBLP Computer Science Bibliography, Semantic Scholar, Engineering Index (EI) and Web of Science / Conference Proceedings Citation Index. A short list of presented papers will be selected so that revised and extended versions of these papers will be published by Springer in a CCIS Series book. Also, a short list of best papers will be invited for a post-conference special issue of the Springer Nature Computer Science journal. All papers presented at the conference venue will also be available at the SCITEPRESS Digital Library. Kind regards, M?nica Saramago SIGMAP Secretariat Web: https://sigmap.scitevents.org e-mail: sigmap.secretariat at insticc.org -------------- next part -------------- An HTML attachment was scrubbed... URL: From RoigNoguera at em.uni-frankfurt.de Wed Feb 23 10:47:51 2022 From: RoigNoguera at em.uni-frankfurt.de (Gemma Roig) Date: Wed, 23 Feb 2022 16:47:51 +0100 Subject: Connectionists: Postdoc position On Computational Vision and Artificial Intelligence at Frankfurt Institute for Advance Studies Message-ID: <64DCD384-1AD6-481C-87E1-F0A2ADF0DA52@em.uni-frankfurt.de> Hello community, There is a job opening as a postdoc, at Centre for multi-scale modelling, analysis and simulation of biological processes at the Frankfurt Institute for Advance Studies. The PI of the project is Prof. Gemma Roig, and the duration of the position is for up to 2 years. The position will involve: - Developing cutting-edge research work on computational models of human visual cognition and artificial intelligence. - Linking AI models to brain data, including functional MRI, MEG and EEG - The outcome of the research pursued is expected to be published in top tier conferences and journals. Key requirements for the position: - PhD in Computer Science, applied mathematics, electrical engineering or related fields. - Strong knowledge in machine learning (deep learning, RNN, LSTMs) - Excellent programming skills in python, and extensive experience in Tensorflow, pyTorch or related libraries. - Knowledge of high performance distributed computing with GPUs. - At least two publications in tier 1 conferences or journals as first author in related topics to computer science and cognitive computational neuroscience. - Basic knowledge and high motivation to learn about computational neuroscience and cognitive science. - Fluent in English (writing, speaking, oral) Self-motivated candidates with a hands-on, proactive approach and problem-solving skills are strongly encouraged to apply. Application Instructions: Applicants are asked to send their documents (motivational letter, CV, 2 publications, diploma and certificates) to Gemma Roig at roig at cs.uni-frankfurt.de until 15th of March 2022, in electronic form. Please include in the subject [CMMS postdoc]. Interviews will be held in Frankfurt or conducted electronically. -------------- next part -------------- An HTML attachment was scrubbed... URL: From jorgecbalmeida at gmail.com Wed Feb 23 13:05:40 2022 From: jorgecbalmeida at gmail.com (Jorge Almeida) Date: Wed, 23 Feb 2022 18:05:40 +0000 Subject: Connectionists: [open now] 1 Tenured Assistant Professor Positions in the area of Forensic Psychology (Cognitive Forensics; Forensic Cognitive Neuroscience) at the Faculty of Psychology and Educational Sciences of the University of Coimbra Portugal Message-ID: The Faculty of Psychology and Educational Sciences of the University of Coimbra, Portugal, is looking for Assistant Professor in the area of Forensic Psychology. Researchers in the area of Fundamental research in forensic cognition and forensic cognitive neuroscience are welcome to apply. Attached in the formal text of the call and all the requirements. The deadline for application is April 5, 2022, and you should apply here: https://apply.uc.pt/P053-22-11440 This is a tenured position with a competitive salary for Portugal. The University has a 3T MRI, and the faculty has facilities for EEG, and other labs in the field. The Faculty of Psychology and Educational Science has several funded projects on-going including one ERC starting grant in Cognitive Neuroscience and object recognition. The University of Coimbra is a 700 year old University and has been selected as a UNESCO world Heritage site. Coimbra is one of the most lively university cities in the world, and it is a beautiful city with easy access to the beach and mountain. The applicants should have their Diplomas registered in Portugal. Depending on where you obtained your PhD, and whether the diploma is not in English (or Portuguese) you may need to have a notarized translation and you may need an Apostille. You can register your Diplomas online at the ministry of education website: https://www.dges.gov.pt/pt/content/recautomatico This process takes some time and it is a necessary process for all applications in Portugal, so you should start it right away. In terms of the language requirements, if you are not a native speaker of Portuguese OR English, you need to prove that you have a C1 level in English OR Portuguese (or both obviously, but one is enough). One option to do so is to have a statement under oath that you have the required level of English (or Portuguese). Although I am not in the jury, and will not be involved in any decisions, if you have questions about this, I can try and help. Hope to see you in Coimbra! Jorge Almeida -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: Edital_prof_auxiliar_en_P053-22-11440_signed.pdf Type: application/pdf Size: 691371 bytes Desc: not available URL: From has168 at eng.ucsd.edu Thu Feb 24 00:08:10 2022 From: has168 at eng.ucsd.edu (Hao Su) Date: Wed, 23 Feb 2022 21:08:10 -0800 Subject: Connectionists: Job Opportunity: Postdoc and senior visiting Ph.D. researcher in Prof. Hao Su's Lab of UCSD Message-ID: SU Lab at UCSD (directed by Prof. Hao Su, https://cseweb.ucsd.edu/~haosu/index.html) has an opening to hire postdoc researcher and senior visiting Ph.D. student working on learning-based 3D capturing techniques. The researcher is expected to conduct independent research and lead junior graduate researchers. Applicants should have an excellent Ph.D. record in a research topic related to applying machine learning for 3D reconstruction and/or rendering/inverse rendering. Welcome to apply by directly contacting Prof. Su via email ( haosu at eng.ucsd.edu, mark [Application] in email subject). If you know anyone who may fit the position, please help to spread the message, help appreciated! -- Hao Su Assistant Professor Department of Computer Science and Engineering Jacobs School of Engineering University of California, San Diego -------------- next part -------------- An HTML attachment was scrubbed... URL: From haim.dub at gmail.com Thu Feb 24 10:21:27 2022 From: haim.dub at gmail.com (Haim Dubossarsky) Date: Thu, 24 Feb 2022 15:21:27 +0000 Subject: Connectionists: =?utf-8?q?=5BDeadline_extension_and_final_CfP=5D_?= =?utf-8?q?3rd_International_Workshop_on_Computational_Approaches_t?= =?utf-8?q?o_Historical_Language_Change_2022_=28LChange=E2=80=9922?= =?utf-8?q?=29?= Message-ID: Deadline extension and Final Call for Papers 3rd International Workshop on Computational Approaches to Historical Language Change 2022 (LChange?22) New deadline for submission, March 2nd. May 26-27, co-located with ACL https://languagechange.org/events/2022-acl-lchange/ Contact email: PC-ACLws2022 at languagechange.org Workshop description The third LChange workshop will be co-located with ACL (2022) to be held in Dublin, during May 26-27, 2022 as a hybrid event. - All aspects around computational approaches to historical language change with the focus on digital text corpora are welcome. LChange explores state-of-the-art computational methodologies, theories and digital text resources on exploring the time-varying nature of human language. - The aim of this workshop is to provide pioneering researchers who work on computational methods, evaluation, and large-scale modelling of language change an outlet for disseminating research on topics concerning language change. Besides these goals, this workshop will also support discussion on the evaluation of computational methodologies for uncovering language change. - LChange?22 will feature a shared task on semantic change detection for Spanish as one track of the workshop. This year we will offer mentoring for PhD students and young researchers in one-on-one meetings during the workshop. If you are interested, send us a short description of your work and we will set you up with one of the organizers of this workshop. If your paper is rejected from the workshop, we can also provide advice on improving it for future submission. This offer is limited, and will be chosen based on topical fit and availability of appropriate mentors. Deadline for applying for mentorship is May 30th via . Via our sponsor, Iguanodon.ai , we can offer one free registration for a PhD student! Apply by emailing us your short cv and why you need your registration paid. Important Dates * March 2, 2022: Paper submission * March 14, 2022: Task description papers * March 26, 2022: Notification of acceptance * March 30, 2022: Deadline for mentorship application * April 10, 2022: Camera-ready papers due * May 26-27, 2022: Workshop date Keynote Talks We can announce that Prof. Dirk Geeraerts and Dominik Schlechtweg will give keynotes. More information to come. Submissions We accept three types of submissions, long and short papers, following the ACL2022 style, and the ACL submission policy, and shared task papers. See ACL submission policy: https://www.aclweb.org/adminwiki/index.php?title=ACL_Policies_for_Submission,_Review_and_Citation Long and short papers may consist of up to eight (8) and four (4) pages of content, respectively, plus unlimited references; final versions will be given one additional page of content so that reviewers' comments can be taken into account. Shared task papers may consist of up to four (4) pages plus unlimited references, but without an additional page upon acceptance. Overleaf templates are available here: https://www.overleaf.com/latex/templates/acl-rolling-review-template/jxbhdzhmcpdm Submission is electronic, using the ACL Rolling Review (ARR), and is now open: https://openreview.net/group?id=aclweb.org/ACL/2022/Workshop/LChange We invite original research papers from a wide range of topics, including but not limited to: * Novel methods for detecting diachronic semantic change and lexical replacement * Automatic discovery and quantitative evaluation of laws of language change * Computational theories and generative models of language change * Sense-aware (semantic) change analysis * Diachronic word sense disambiguation * Novel methods for diachronic analysis of low-resource languages * Novel methods for diachronic linguistic data visualization * Novel applications and implications of language change detection * Quantification of sociocultural influences on language change * Cross-linguistic, phylogenetic, and developmental approaches to language change * Novel datasets for cross-linguistic and diachronic analyses of language Submissions are open to all, and are to be submitted anonymously. All papers will be refereed through a double-blind peer review process by at least three reviewers with final acceptance decisions made by the workshop organizers. The workshop is scheduled for May 26-27. Contact us at PC-ACLws2022 at languagechange.org if you have any questions. Workshop organizers: Nina Tahmasebi, University of Gothenburg Lars Borin, University of Gothenburg Simon Hengchen, University of Gothenburg Syrielle Montariol, University Paris-Saclay Haim Dubossarsky, Queen Mary University of London Andrey Kutuzov, University of Oslo -------------- next part -------------- An HTML attachment was scrubbed... URL: From franrruiz87 at gmail.com Thu Feb 24 13:21:41 2022 From: franrruiz87 at gmail.com (=?UTF-8?Q?Francisco_J=2E_Rodr=C3=ADguez_Ruiz?=) Date: Thu, 24 Feb 2022 18:21:41 +0000 Subject: Connectionists: ICBINB Monthly Seminar Series Talk: Sebastian Nowozin (Mar 3rd 10am EST) Message-ID: Dear all, We?re very excited to host *Sebastian Nowozin** (Microsoft Research)* for the second talk of the newly created* ?I Can?t Believe It?s Not Better!? (ICBINB) virtual seminar series*. More details about this series are below. The *"I Can't Believe It's Not Better!" (ICBINB) monthly online seminar series* seeks to shine a light on the "stuck" phase of research. Speakers will tell us about their most beautiful ideas that didn't "work", about when theory didn't match practice, or perhaps just when the going got tough. These talks will let us peek inside the file drawer of unexpected results and peer behind the curtain to see the real story of *how real researchers did real research*. *When: *March 3rd, 2022 at 10am EST / 4pm CET *Where: *RSVP for the Zoom link here: https://us02web.zoom.us/meeting/register/tZcucuCvqz4tGta1B5lCos-JGiA5YsZIyD6p *Title:* *I Can?t Believe Bayesian Deep Learning is not Better* *Abstract:* *Bayesian deep learning is seductive: it combines the simplicity, coherence, and beauty of the Bayesian approach to problem solving together with the expressivity and compositional flexibility of deep neural networks. Yes, inference can be challenging, but the promises of improved uncertainty quantification, better out-of-distribution behaviour, and improved sample efficiency are worth it. Or is it? In this talk I will tell a personal story of being seduced by, then frustrated with, and now recovering from Bayesian deep learning. I will present the context of our work on the cold posterior effect, (Wenzel et al., 2020) and it?s main findings, as well as some more recent work that tries to explain the effect. I will also offer some personal reflections on research practice and narratives that contributed to the lack of progress in Bayesian deep learning.* *Bio: **Sebastian Nowozin is a deep learning researcher at Microsoft Research Cambridge, UK, where he currently leads the Machine Intelligence research theme. His research interests are in probabilistic deep learning and applications of machine learning models to real-world problems. He completed his PhD in 2009 at the Max Planck Institute in T?bingen, and has since worked on domains as varied as computer vision, computational imaging, cloud-based machine learning, and approximate inference.* *--* For more information and for ways to get involved, please visit us at http://icbinb.cc/, Tweet to us @ICBINBWorkhop , or email us at cant.believe.it.is.not.better at gmail.com. -- Best wishes, The ICBINB Organizers -------------- next part -------------- An HTML attachment was scrubbed... URL: From marcella.cornia at unimore.it Fri Feb 25 04:44:07 2022 From: marcella.cornia at unimore.it (Marcella Cornia) Date: Fri, 25 Feb 2022 10:44:07 +0100 Subject: Connectionists: [CFP] AI4DH: Artificial Intelligence for Digital Humanities - International Workshop at ICIAP 2021 Message-ID: ******************************** Call for Papers ?AI4DH: Artificial Intelligence for Digital Humanities? International Workshop at ICIAP 2021 https://ailb-web.ing.unimore.it/ai4dh2021/ ******************************** === SUBMISSIONS ARE OPEN!!! ==== Apologies for multiple posting Please distribute this call to interested parties AIMS AND SCOPE =============== Researchers have explored the benefits and applications of modern Artificial Intelligence (AI) algorithms in different scenarios. Digital Humanities (DH) is a newly-emerging field that brings together humanities, social, and computer scientists to work both on fundamental and applied research in humanities. The large-scale cultural and societal implications of these changes and the ethical questions that raise offer an important challenge as techniques in Artificial Intelligence and Data Learning have matured. Thus, there has been a wide range of computational tools, methods, and models that have enabled humanities to conduct research at a scale once thought impossible. The goal of this workshop is to encourage and highlight novel strategies and original research in applying Artificial Intelligence techniques in digital humanities research such as data discovery, digital data creation, management, data analytics (including text mining, image mining and data visualization) in literature, linguistics, culture heritage, media, social science, history, music and acoustics, and Artificial Intelligence for Digital Humanities in pedagogy and academic curricula. This Workshop aims not only to serve as a venue for presenting work in this area but also to build a community and share information in this new field. TOPICS ======= The workshop calls for submissions addressing, but not limited to, the following topics: - AI for Textual DH - AI in Natural Language Processing and Cultural Heritage applications - Handwritten Text Recognition of historical documents - Layout Analysis for historical documents - Keyword Spotting on historical documents - Writer Identification in historical documents - AI for Archeological DH - AI in Digital Archeology, digitization and on-site documentation - AI in Archaeometry and Data Analysis - Computational Archeology - AI and simulations in Archeology and Cultural Heritage - 3D artifacts reconstruction - AI for Visual DH - Artworks cross-modal retrieval - AI for art generation - Automatic metadata extraction from artworks - AI for video analysis - AI for artistic images analysis - AI for DH Preservation and Enhancement - AI in virtual systems for education and tourism - AI in museums and cultural tourism - Museum and Service Robotics - Music and visual art-induced emotion recognition - Automatic storytelling of art - Fundation AI Applications to DH - AI in digital cultural content/object analysis - AI in content?based classification and retrieval - AI in Semantics and Knowledge Representation - Intelligent methods in Spatial and Temporal Analysis - Intelligent crowdsourcing approaches IMPORTANT DATES ================= - Paper Submission Deadline: March 24th, 2022 - Decision to Authors: April 15th, 2022 - Camera ready papers due: April 22nd, 2021 - Workshop date: TBA SUBMISSION GUIDELINES ====================== All the papers should be submitted at: https://cmt3.research.microsoft.com/AI4DH2022 The maximum number of pages is 10 + 2 pages for references. While preparing their contributions, authors must follow guidelines and technical instructions provided by Springer that can be found at: https://www.springer.com/gp/computer-science/lncs/conference-proceedings-guidelines . Each accepted paper must be covered by at least one registered author. Registration can be either for the full event (5 days) at a regular rate or just for workshops and tutorials (2 days). WORKSHOP MODALITY ==================== The workshop will be held in conjunction with the International Conference on Image Analysis and Processing (ICIAP 2021). Both virtual and in presence participation will be allowed. ORGANIZING COMMITTEE ====================== - Lorenzo Baraldi, University of Modena and Reggio Emilia, Italy - Silvia Cascianelli, University of Modena and Reggio Emilia, Italy - Marcella Cornia, University of Modena and Reggio Emilia, Modena, Italy - Francesca Matrone, Politecnico di Torino, Italy - Marina Paolanti, University of Macerata, Italy -- *Marcella Cornia*, PhD Tenure-Track Assistant Professor (RTD-B) Dipartimento di Educazione e Scienze Umane Universit? degli Studi di Modena e Reggio Emilia e-mail: marcella.cornia at unimore.it phone: +39 059 2058790 -------------- next part -------------- An HTML attachment was scrubbed... URL: From erik at oist.jp Fri Feb 25 08:26:11 2022 From: erik at oist.jp (Erik De Schutter) Date: Fri, 25 Feb 2022 13:26:11 +0000 Subject: Connectionists: Okinawa/OIST Computational Neuroscience Course 2022: application deadline extended In-Reply-To: References: <367D1A2B-0339-4E63-9148-A6F5A9D25207@oist.jp> Message-ID: Because of impending changes to Japanese border controls we want to give potential applicants more time to consider spending a few summer weeks in Japan: the deadline for applications has been extended to Sunday March 2022. On Jan 25, 2022, at 11:31 , Erik De Schutter wrote: OKINAWA/OIST COMPUTATIONAL NEUROSCIENCE COURSE 2022 Methods, Neurons, Networks and Behaviors June 13 to June 29, 2022 Okinawa Institute of Science and Technology Graduate University, Japan https://groups.oist.jp/ocnc After two consecutive cancelations due to COVID-19, OCNC 2022 will take place on June 13-29, preceding Neuro2022 (https://neuro2022.jnss.org) in Okinawa. Depending on the immigration situation in June the course will be either pure on-site or hybrid: a mixture of on-site and remote. The aim of the Okinawa/OIST Computational Neuroscience Course is to provide opportunities for young researchers with theoretical backgrounds to learn the latest advances in neuroscience, and for those with experimental backgrounds to have hands-on experience in computational modeling. We invite graduate students and postgraduate researchers to participate in the course, held from June June 13 through June 29, 2022 at an oceanfront seminar house of the Okinawa Institute of Science and Technology Graduate University. Applications are through the course web page (https://groups.oist.jp/ocnc) only: January 26 - March 20, 2022. Applicants will receive confirmation of acceptance end of April. The 17th OCNC will be a shorter course than in the past: a two-week course covering single neurons, networks, and behaviors with time for student projects. Teaching will focus on methods with hands-on tutorials during the afternoons, and lectures by international experts. The course has a strong hands-on component based on student proposed modeling or data analysis projects, which are further refined with the help of a dedicated tutor. Applicants are required to propose their project at the time of application. However in the case of a hybrid format course, only on-site students will receive support for student projects. There is no tuition fee. The sponsor will provide lodging and meals during the course and may provide partial travel support. We hope that this course will be a good opportunity for theoretical and experimental neuroscientists to meet each other and to explore the attractive nature and culture of Okinawa, the southernmost island prefecture of Japan. Invited faculty: ? Michael Berry II (Princeton University, USA) ? Anne Churchland, Cold Spring Harbor Labs, USA ? Erik De Schutter (OIST) ? Kenji Doya (OIST) ? Gaute Einevoll, Norwegian University of Life Sciences (online) ? Tomoki Fukai (OIST) ? Boris Gutkin (Ecole Normale Sup?rieure, Paris, France) ? Yukiyasu Kamitani, Kyoto University, Japan ? Bernd Kuhn (OIST) ? Sang Wan Lee, KAIST, South Korea ? Devika Narain, Erasmus Medical Center, Rotterdam, Netherlands ? Viola Priesemann, MPG Go?ttingen, Germany (online) ? Ivan Soltesz, Stanford University, USA ? Greg Stuart, Australian National University, Australia ? Greg Stephens (OIST) ? Saori Tanaka, (ATR, Japan) ? Marylka Yoe Uusisaari (OIST) -------------- next part -------------- A non-text attachment was scrubbed... Name: smime.p7s Type: application/pkcs7-signature Size: 3766 bytes Desc: not available URL: From avinashsingh214 at gmail.com Fri Feb 25 08:45:18 2022 From: avinashsingh214 at gmail.com (Avinash K Singh) Date: Sat, 26 Feb 2022 00:45:18 +1100 Subject: Connectionists: =?utf-8?q?=5BCFP=5D_QoMEX=E2=80=9922_14th_Interna?= =?utf-8?q?tional_Conference_on_Quality_of_Multimedia_Experience?= Message-ID: Call For Papers QoMEX?22 14th International Conference on Quality of Multimedia Experience September 5.-7. 2022 ? Lippstadt, Germany Full Paper Submission: March 31, 2022 https://qomex2022.itec.aau.at/ ----------------------------------------------------------------- The 14th International Conference on Quality of Multimedia Experience will be held from September 5th to 7th, 2022 in Lippstadt, Germany. It will bring together leading experts from academia and industry to present and discuss current and future research on multimedia quality, quality of experience (QoE) and user experience (UX). This way, it will contribute to excellence in developing multimedia technology, towards user well-being, and it will foster the exchange between multidisciplinary communities. The QoMEX 2022 team solicits contributions including but not limited to topics: - Immersive experiences and technologies - QoE, Big Data and Artificial Intelligence - Games User Research and Experience - New assessment and evaluation methods - Quality, experience, and user state - Quality of Life and Well-being - Multimodal perception & quality - Databases for QoE research - Audio/Visual user experience - QoE-aware networks and services management Prospective authors are invited to submit full (maximum of 6 pages) or short papers (3 +1 page of references) to the general track and to special sessions. Each paper will undergo a double-blind review process. Full and short papers will be included in the conference proceedings and published in IEEExplore (approval pending). Important Dates and Details --------------------------- Full Paper Submission: 31 March 2022 Full Paper Notification: 31 May 2022 Short Paper Submission: 07 June 2022 Short Paper Notification: 15 July 2022 Conference: September 5-7, 2022 Website: https://qomex2022.itec.aau.at Twitter: @QoMEXconf General Chair ------------- - Jan-Niklas Voigt-Antons, HSHL, Germany Technical Program Chairs ----------------------------- - Luigi Atzori, Univ. Cagliari, Italy - Sebastian M?ller, TU Berlin/DFKI Berlin, Germany - Alexander Raake, TU Ilmenau, Germany -- Regards, Avinash K Singh -------------- next part -------------- An HTML attachment was scrubbed... URL: From angelo.cangelosi at manchester.ac.uk Fri Feb 25 09:11:24 2022 From: angelo.cangelosi at manchester.ac.uk (Angelo Cangelosi) Date: Fri, 25 Feb 2022 14:11:24 +0000 Subject: Connectionists: Launch of new Editorial Board and Call for Submissions: Interaction Studies Message-ID: Interaction Studies is an international, peer-reviewed journal aiming to advance knowledge in the growing and strongly interdisciplinary area of Interaction Studies in biological and artificial systems. Understanding social behaviour and communication in bifBarological and artificial systems requires knowledge of evolutionary, developmental and neurobiological aspects of social behaviour and communication; the embodied nature of interactions; origins and characteristics of social and narrative intelligence; perception, action and communication in the context of dynamic and social environments; social learning, adaptation and imitation; social behaviour in human-machine interactions; the nature of empathic understanding, behaviour and intention reading; minimal requirements and systems exhibiting social behaviour; the role of cultural factors in shaping social behaviour and communication in biological or artificial societies. We welcome submissions in those, as well as related areas. The journal has recently refreshed its editorial board, with a new team of Associate Editors and Editorial Board Members. This team includes the senior leaders in the field, as well as a group of rising stars. Associate Editors * Tony Belpaeme | Ghent University * Emilia Barakova | Eindhoven University of Technology * Justine Cassell | HCII, Carnegie Mellon University & Inria Paris * Vicky Charisi | European Commission, Centre for Advanced Studies * Kerstin Fischer | University of Southern Denmark * Hatice Gunes | University of Cambridge * Takayuki Kanda | Kyoto University * Katja Liebal | University of Leipzig * Xiaofeng Liu | Hohai University * Dingsheng Luo | Peking University * Gary Lupyan | University of Wisconsin-Madison * Tetsuro Matsuzawa | California Institute of Technology * Robert W. Mitchell | Eastern Kentucky University * Chrystopher L. Nehaniv | University of Waterloo * Jacqueline Nadel | H?pital de la Salp?tri?re * Katerina Pastra | Institute for Language and Speech Processing (ILSP) * Simone Pika | University of Osnabr?ck * Gnanathusharan Rajendran | Heriot-Watt University, Edinburgh * Silvia Rossi | Universit? degli Studi di Napoli Federico II * Giulio Sandini | University of Genova * Alessandra Sciutti | Istituto Italiano di Tecnologia * Gentiane Venture | Tokyo University of Agriculture and Technology * S?awomir Wacewicz | Nicolaus Copernicus University * Astrid Weiss | TU Wien * Gert Westermann | Lancaster University * Chenguang Yang | University of the West of England Editorial Board Members * Amir Aly | University of Plymouth * Daniela Conti | University of Catania * Cinzia Di Dio | Universit? Cattolica del Sacro Cuore * Monica Gori | Italian Institute of Technology * Moojan Ghafurian | University of Waterloo * Kirsty E. Graham | University of St Andrews * Yuko Hattori | Kyoto University * Catherine Hobaiter | University of St Andrews * Patrick Holthaus | University of Hertfordshire * Kheng Lee Koay | University of Hertfordshire * Gabriella Lakatos | University of Hertfordshire * Sarah Ita Levitan | Hunter College - CUNY * Angelica Lim | Simon Fraser University * Wing-Yue Geoffrey Louie | Oakland University * AJung Moon | McGill University * Francesco Rea | Istituto Italiano di Tecnologia * Alessandra Rossi | University of Naples Federico II * Harold Soh | National University of Singapore * Katherine E. Twomey | University of Manchester * Alan R. Wagner | Penn State Interaction Studies publishes regular research articles, research reports, and book reviews. For any new submission, see guidelines in the journal website https://benjamins.com/catalog/is/submission The journal currently has the following Special Issues open for submission. Socially Acceptable Robot Behavior: Approaches for Learning, Adaptation and Evaluation. Deadline 31 March 2022 Interaction Studies Editors-In-Chief Kerstin Dautenhahn, University of Waterloo, Canada Angelo Cangelosi, University of Manchester, UK -------------- next part -------------- An HTML attachment was scrubbed... URL: From stephan.petrone at gmail.com Fri Feb 25 15:19:14 2022 From: stephan.petrone at gmail.com (Stephan Petrone) Date: Fri, 25 Feb 2022 21:19:14 +0100 Subject: Connectionists: Metaheuristics International Conference 2022 @ Ortigia-Syracuse, Italy Message-ID: Apologies for cross-posting. Appreciate if you can distribute this CFP to your network. ********************************************************* MIC 2022 - 14th Metaheuristics International Conference 11-14 July 2022, Ortigia-Syracuse, Italy https://www.ANTs-lab.it/mic2022/ mic2022 at ANTs-lab.it ********************************************************* ** Submission deadline: 30th March 2022 ** ** Proceedings in LNCS Volume, Springer ** ** Special Issue in ITOR journal ** ** 7 Plenary Speakers ** ** SUBMISSION SYSTEM ALREADY OPEN ** https://www.easychair.org/conferences/?conf=mic2022 *Scope of the Conference ======================== The *Metaheuristics International Conference* (MIC) conference series was established in 1995 and this is its 14th edition! MIC is nowadays the main event focusing on the progress of the area of Metaheuristics and their applications. As in all previous editions, provides an opportunity to the international research community in Metaheuristics to discuss recent research results, to develop new ideas and collaborations, and to meet old and make new friends in a friendly and relaxed atmosphere. Considering the particular moment, *the conference will be held in presence and online mode*. Of course, in case the conference will be held in presence, the organizing committee will ensure compliance of all safety conditions. MIC 2022 is focus on presentations that cover different aspects of metaheuristic research such as new algorithmic developments, high-impact and original applications, new research challenges, theoretical developments, implementation issues, and in-depth experimental studies. MIC 2022 strives a high-quality program that will be completed by a number of invited talks, tutorials, workshops and special sessions. *Plenary Speakers ======================== + Christian Blum, Artificial Intelligence Research Institute (IIIA), Spanish National Research Council (CSIC) + Kalyanmoy Deb, Michigan State University, USA + Fred Glover, Meta-Analytics, Inc., USA + Salvatore Greco, University of Catania, Italy + Holger H. Hoos, Leiden University, The Netherlands + Gary Kochenberger, University of Colorado, USA + El-Ghazali Talbi, University of Lille, France Important Dates ================ Submission deadline March 30th, 2022 Notification of acceptance May 10th, 2022 Camera ready copy May 25th, 2022 Early registration May 25th , 2022 Submission Details =================== MIC 2022 accepts submissions in three different formats: S1) *Regular paper*: novel and original research contributions of a *maximum of 15 pages* (LNCS format) S2) *Short paper*: extended abstract of novel research works of *6 pages* (LNCS format) S3) *Oral/Poster presentation*: high-quality manuscripts that have recently, within the last year, been submitted or accepted for journal publication. All papers must be prepared *using Lecture Notes in Computer Science (LNCS) template*, and must be submitted in PDF at the link: https://www.easychair.org/conferences/?conf=mic2022 Proceedings and special issue ============================ Accepted papers in *categories S1 and S2* will be published as *post-proceedings* in *Lecture Notes in Computer Science* series by Springer. Accepted contributions of *category S3* will be considered for oral or poster presentations at the conference based on the number received and the slots available, and *will not be included into the LNCS proceedings*. An electronic book instead will be prepared by the MIC 2022 organizing committee, and made available on the website. In addition, a post-conference special issue in *International Transactions in Operational Research (ITOR)* will be considered for the significantly extended and revised versions of selected accepted papers from categories S1 and S2. Conference Location ==================== MIC 2022 will be held in the beautiful Ortigia island, the historical centre of the city of Syracuse, Sicily-Italy. Syracuse is very famous for its ancient ruins, with particular reference to the Roman Amphitheater, Greek Theatre, and the Orecchio di Dionisio (Ear of Dionisio) that is a limestone cave shaped like a human ear. Syracuse is also the city where the greatest mathematician Archimede was born. https://www.siracusaturismo.net/multimedia_lista.asp MIC'2022 Conference Chairs ============================== Conference Chairs - Luca Di Gaspero, University of Undine, Italy - Paola Festa, University of Naples, Italy - Amir Nakib, Universit? Paris Est Cr?teil, France - Mario Pavone, University of Catania, Italy -------------- next part -------------- An HTML attachment was scrubbed... URL: From nasim.sonboli at gmail.com Sun Feb 27 23:45:53 2022 From: nasim.sonboli at gmail.com (Nasim Sonboli) Date: Sun, 27 Feb 2022 21:45:53 -0700 Subject: Connectionists: =?utf-8?q?First_Call_for_UMAP=E2=80=9922_Doctoral?= =?utf-8?q?_Consortium_Papers?= Message-ID: --- Please forward to anyone who might be interested --- --- Apologies for cross-posting --- ----------------------------------------------------------------------------- 30th ACM International Conference on User Modeling, Adaptation and Personalization (ACM UMAP'22) Barcelona, Spain, and Online 4 - 7 July 2022 https://www.um.org/umap2022/ ----------------------------------------------------------------------------- Important Dates: - Doctoral Consortium Submission: March 31, 2022; - Notification to authors: April 28, 2022; - Camera-ready submission: May 2, 2022; - Conference: July 4-7, 2022; - DC day: July TBD, 2022; Note: All the deadlines must be intended at 11:59 PM AoE time (Anywhere on Earth) ----------------------------------------------------------------------------- Background and Scope The UMAP 2022 Doctoral Consortium (DC) will take place as part of the 30th International Conference on User Modeling, Adaptation, and Personalization. Doctoral students are invited to apply to present their research to experienced scholars who will provide constructive feedback and advice. Students should consider participating in the DC if they are at least fifteen months away from completing their dissertation at the time of the event, but after having settled on a research area or dissertation topic. This forum will provide Ph.D. students an opportunity to: - Present and discuss their research ideas to experienced scholars in a supportive, formative, and yet critical environment; - Explore and develop their research interests under the guidance of distinguished researchers and industry practitioners from the field who will provide constructive feedback and advice; - Explore career pathways available after completing their Ph.D. degree, and finally - Network and build collaborations with other members of the community. Students are asked to submit a brief proposal outlining their doctoral research (see detailed requirements below), which will be evaluated by the consortium committee. Good quality applications will be selected for presentation at a DC Session as part of the conference. Each student with an accepted submission will be assigned a mentor who will provide feedback on the student?s work and will discuss the doctoral research with the student and the audience at the consortium. ----------------------------------------------------------------------------- A Brief Review of The General Guidelines for a Doctoral Consortium Paper Candidates must submit their DC proposals ?aligning with topics pertaining to the ACM UMAP 2022 key areas? by the submission deadline. Each proposal should be a single PDF, with four required components: a six-page description of your Ph.D., an expected benefits statement, a recommendation letter, and a two-page curriculum vitae. If you are unable to obtain a letter of recommendation from your dissertation advisor, please include a short explanation. For more details please refer to the website: https://www.um.org/umap2022/call-for-doctoral-consortium-papers/ ----------------------------------------------------------------------------- Financial Support ACM UMAP has a history of supporting students who want to attend the conference. Students of accepted DC proposals will have higher priority for UMAP grant applications. ----------------------------------------------------------------------------- DC Chairs: umap2022-doctoral at um.org - Rita Orji (Dalhousie University, Canada) - Giovanni Stilo (University of L?Aquila, Italy) ----------------------------------------------------------------------------- Follow UMAP 2022 on Social Media - Twitter: @UMAPconf #umap2022 (https://twitter.com/UMAPconf) - Facebook: https://www.facebook.com/acmumap -------------- next part -------------- An HTML attachment was scrubbed... URL: From david at irdta.eu Sat Feb 26 15:34:49 2022 From: david at irdta.eu (David Silva - IRDTA) Date: Sat, 26 Feb 2022 21:34:49 +0100 (CET) Subject: Connectionists: DeepLearn 2022 Autumn: early registration March 18 Message-ID: <1635278022.2007957.1645907689820@webmail.strato.com> ****************************************************************** 7th INTERNATIONAL SCHOOL ON DEEP LEARNING DeepLearn 2022 Autumn Lule?, Sweden October 17-21, 2022 https://irdta.eu/deeplearn/2022au/ ***************** Co-organized by: Lule? University of Technology EISLAB Machine Learning Institute for Research Development, Training and Advice ? IRDTA Brussels/London ****************************************************************** Early registration: March 18, 2022 ****************************************************************** SCOPE: DeepLearn 2022 Autumn will be a research training event with a global scope aiming at updating participants on the most recent advances in the critical and fast developing area of deep learning. Previous events were held in Bilbao, Genova, Warsaw, Las Palmas de Gran Canaria, Guimar?es and Las Palmas de Gran Canaria. Deep learning is a branch of artificial intelligence covering a spectrum of current frontier research and industrial innovation that provides more efficient algorithms to deal with large-scale data in a huge variety of environments: computer vision, neurosciences, speech recognition, language processing, human-computer interaction, drug discovery, health informatics, medical image analysis, recommender systems, advertising, fraud detection, robotics, games, finance, biotechnology, physics experiments, biometrics, communications, climate sciences, bioinformatics, etc. etc. Renowned academics and industry pioneers will lecture and share their views with the audience. Most deep learning subareas will be displayed, and main challenges identified through 24 four-hour and a half courses and 3 keynote lectures, which will tackle the most active and promising topics. The organizers are convinced that outstanding speakers will attract the brightest and most motivated students. Face to face interaction and networking will be main ingredients of the event. It will be also possible to fully participate in vivo remotely. An open session will give participants the opportunity to present their own work in progress in 5 minutes. Moreover, there will be two special sessions with industrial and recruitment profiles. ADDRESSED TO: Graduate students, postgraduate students and industry practitioners will be typical profiles of participants. However, there are no formal pre-requisites for attendance in terms of academic degrees, so people less or more advanced in their career will be welcome as well. Since there will be a variety of levels, specific knowledge background may be assumed for some of the courses. Overall, DeepLearn 2022 Autumn is addressed to students, researchers and practitioners who want to keep themselves updated about recent developments and future trends. All will surely find it fruitful to listen to and discuss with major researchers, industry leaders and innovators. VENUE: DeepLearn 2022 Autumn will take place in Lule?, on the coast of northern Sweden, hosting a large steel industry and the northernmost university in the country. The venue will be: Lule? University of Technology https://www.ltu.se/?l=en STRUCTURE: 3 courses will run in parallel during the whole event. Participants will be able to freely choose the courses they wish to attend as well as to move from one to another. Full live online participation will be possible. However, the organizers highlight the importance of face to face interaction and networking in this kind of research training event. KEYNOTE SPEAKERS: (to be completed) Wolfram Burgard (University of Freiburg), Probabilistic and Deep Learning Techniques for Robot Navigation and Automated Driving Tommaso Dorigo (Italian National Institute for Nuclear Physics), Deep-Learning-Optimized Design of Experiments: Challenges and Opportunities PROFESSORS AND COURSES: (to be completed) Sean Benson (Netherlands Cancer Institute), [intermediate] Deep Learning for a Better Understanding of Cancer Daniele Bonacorsi (University of Bologna), [intermediate/advanced] Applied ML for High-Energy Physics Thomas Breuel (Nvidia), [intermediate/advanced] Large Scale Deep Learning and Self-Supervision in Vision and NLP Hao Chen (Hong Kong University of Science and Technology), [introductory/intermediate] Label-Efficient Deep Learning for Medical Image Analysis Jianlin Cheng (University of Missouri), [introductory/intermediate] Deep Learning for Bioinformatics Peng Cui (Tsinghua University), [introductory/advanced] Towards Out-Of-Distribution Generalization: Causality, Stability and Invariance S?bastien Fabbro (University of Victoria), [introductory/intermediate] Learning with Astronomical Data Quanquan Gu (University of California Los Angeles), [intermediate/advanced] Benign Overfitting in Machine Learning: From Linear Models to Neural Networks Jiawei Han (University of Illinois Urbana-Champaign), [advanced] Text Mining and Deep Learning: Exploring the Power of Pretrained Language Models Awni Hannun (Zoom), [intermediate] An Introduction to Weighted Finite-State Automata in Machine Learning Shirley Ho (Flatiron Institute), [intermediate] Structured Machine Learning for Simulations Timothy Hospedales (University of Edinburgh), [introductory/intermediate] Deep Learning with Limited Data Shih-Chieh Hsu (University of Washington), [intermediate/advanced] Real-Time Artificial Intelligence for Science and Engineering Andrew Laine (Columbia University), [introductory/intermediate] Applications of AI in Medical Imaging Tatiana Likhomanenko (Apple), [intermediate/advanced] Self-, Weakly-, Semi-Supervised Learning in Speech Recognition Peter Richt?rik (King Abdullah University of Science and Technology), [intermediate/advanced] Introduction to Federated Learning Othmane Rifki (Spectrum Labs), [introductory/advanced] Speech and Language Processing in Modern Applications Mayank Vatsa (Indian Institute of Technology Jodhpur), [introductory/intermediate] Small Sample Size Deep Learning Zichen Wang (Amazon Web Services), [introductory/intermediate] Graph Machine Learning for Healthcare and Life Sciences Alper Yilmaz (Ohio State University), [introductory/intermediate] Deep Learning and Deep Reinforcement Learning for Geospatial Localization OPEN SESSION: An open session will collect 5-minute voluntary presentations of work in progress by participants. They should submit a half-page abstract containing the title, authors, and summary of the research to david at irdta.eu by October 9, 2022. INDUSTRIAL SESSION: A session will be devoted to 10-minute demonstrations of practical applications of deep learning in industry. Companies interested in contributing are welcome to submit a 1-page abstract containing the program of the demonstration and the logistics needed. People in charge of the demonstration must register for the event. Expressions of interest have to be submitted to david at irdta.eu by October 9, 2022. EMPLOYER SESSION: Organizations searching for personnel well skilled in deep learning will have a space reserved for one-to-one contacts. It is recommended to produce a 1-page .pdf leaflet with a brief description of the organization and the profiles looked for to be circulated among the participants prior to the event. People in charge of the search must register for the event. Expressions of interest have to be submitted to david at irdta.eu by October 9, 2022. ORGANIZING COMMITTEE: Sana Sabah Al-Azzawi (Lule?) Lama Alkhaled (Lule?) Prakash Chandra Chhipa (Lule?) Saleha Javed (Lule?) Marcus Liwicki (Lule?, organization co-chair) Carlos Mart?n-Vide (Tarragona, program chair) Hamam Mokayed (Lule?) Sara Morales (Brussels) Mia Oldenburg (Lule?) Maryam Pahlavan (Lule?) David Silva (London, organization co-chair) Richa Upadhyay (Lule?) REGISTRATION: It has to be done at https://irdta.eu/deeplearn/2022au/registration/ The selection of 8 courses requested in the registration template is only tentative and non-binding. For logistical reasons, it will be helpful to have an estimation of the respective demand for each course. During the event, participants will be free to attend the courses they wish. Since the capacity of the venue is limited, registration requests will be processed on a first come first served basis. The registration period will be closed and the on-line registration tool disabled when the capacity of the venue will have got exhausted. It is highly recommended to register prior to the event. FEES: Fees comprise access to all courses and lunches. There are several early registration deadlines. Fees depend on the registration deadline. The fees for on site and for online participants are the same. ACCOMMODATION: Accommodation suggestions will be available in due time at https://irdta.eu/deeplearn/2022au/accommodation/ CERTIFICATE: A certificate of successful participation in the event will be delivered indicating the number of hours of lectures. QUESTIONS AND FURTHER INFORMATION: david at irdta.eu ACKNOWLEDGMENTS: Lule? University of Technology, EISLAB Machine Learning Rovira i Virgili University Institute for Research Development, Training and Advice ? IRDTA, Brussels/London -------------- next part -------------- An HTML attachment was scrubbed... URL: From juergen at idsia.ch Mon Feb 28 03:03:35 2022 From: juergen at idsia.ch (Schmidhuber Juergen) Date: Mon, 28 Feb 2022 08:03:35 +0000 Subject: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc. In-Reply-To: <3c93a408-f085-4340-694e-b5aa2c9335de@rubic.rutgers.edu> References: <2293D07C-A5E3-4E66-9120-C14DE15239A7@supsi.ch> <29BC825D-F353-457A-A9FD-9F25F3D1A6DB@supsi.ch> <3155202C-080E-4BE7-84B6-A567E306AC1D@supsi.ch> <58AC5011-BF6A-453F-9A5E-FAE0F63E2B02@supsi.ch> <7f97db1f-13e1-ba48-8b02-f3a2c4769df9@rubic.rutgers.edu> <6E8002C5-64E7-41DB-930F-B77BF78F600A@supsi.ch> <3c93a408-f085-4340-694e-b5aa2c9335de@rubic.rutgers.edu> Message-ID: <4AE801EB-34BD-4C89-B81F-3F19AC26637A@supsi.ch> Steve, surely you can agree with me that plagiarism cannot win here? As you said, let's "embrace all the facts? and "the bigger conceptual picture,? but always credit those who did things first. You mention the ?common false narrative? (promulgated by certain self-aggrandizing psychologists and neuroscientists since the 1980s). Indeed, this narrative is simply incompatible with the historic facts, ignoring the very origins of deep learning in the mid 1960s (and shallow learning in the 1800s). We have a duty as academics and scientists to ensure the facts win. J?rgen https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html > On 7 Feb 2022, at 16:26, Stephen Jos? Hanson wrote: > > Juergen, > > ignoring history by defining it as "fancy talk", is going to make your exegesis of Neural networks always lagging. You need to embrace all the facts, not just the ones you like or are familiar with. Your whole endeavor, is an attempt to destroy what you feel is the common false narrative on the origin of neural networks. I am happy to chat more about this sometime, but I still think your mathematical lens, is preventing you from seeing the bigger conceptual picture. > > Best, > > Steve > > On 2/6/22 3:44 AM, Schmidhuber Juergen wrote: >> Steve, it?s simple: the original ?shallow learning? (~1800) is much older than your relatively recent ?shallow learning? references (mostly from the 1900s). No need to mention all of them in this report, which is really about "deep learning? (see title) with adaptive hidden units, which started to work in the 1960s, first through layer by layer training (USSR, 1965), then through stochastic gradient descent (SGD) in relatively deep nets (Japan, 196I 7). The reverse mode of automatic differentiation (now called backpropagation) appeared 3 years later (Finland, 1970). No fancy talk about syntax vs semantics can justify a revisionist history of deep learning that does not mention these achievements. Cheers, J?rgen >> >> https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html >> >> >> >> >>> On 2 Feb 2022, at 00:48, Stephen Jos? Hanson wrote: >>> >>> Jeurgen: Even some of us lowly psychologists know some math. And its not about the math.. its about the context (is this sounding like an echo?) >>> >>> Let me try again.. and I think your good intentions but misguided reconstruction of history is appearing to me, to be perverse. >>> >>> You tip your hand when you talk about "rebranding". Also that the PDP books were a "conspiracy". But lets go point by point. >>> >>> (1) we already agreed that the Perceptron was not linear regression--lets not go backwards. Closer to logistic regression. If you are talking about Widrow and Hoff, well it is the Delta Rule-- SSE kind of regression. But where did the Delta rule come from? Lets look at Math. So there is some nice papers by Gluck and Thompson (80s) showing how Pavlovian conditioning is exactly the Delta rule and even more relevant was shown to account for majority of classical (pavlovian) conditioning was the Rescorla-Wagner (1972) model-- \Delta V_A = [\alpha_A\beta_1](\lambda_1 - V_{AX}), which of course was Ivan Petrovich Pavlov discovery of classical conditioning (1880s). Why aren't you citing him? What about John Brodeus Watson and Burris Fredrick Skinner? At least they were focused on learning albeit *just* in biological systems. But these were actual natural world discoveries. >>> >>> (2) Function approximation. Ok Juergen, claims that everything is really just X, reminds me of the man with a Hammer to whom everything looks like a nail! To the point: its incidental. Yes, Neural networks are function approximators, but that is incidental to the original more general context (PDP) as a way to create "internal representations". The function approximation was a Bonus! >>> >>> (3) Branding. OMG. So you seem to believe that everyone is cynical and will put their intellectual finger in the air to find out what to call what they are doing! Jeez, I hope this isn't true. But the narrative you eschew is in fact something that Minsky would talk about (I remember this at lunch with him in the 90s at Thinking Machines), and he was quite clear that Perceptron was failing well before the 1969 book (trying to do speech recognition with a perceptron--yikes), but in a piling on kind of way Perceptrons killed the perceptron, but it was the linearity focus (as BAP points out) and the lack of depth. >>> >>> (4) Group Method of Handling Data. Frankly, the only one I can find that branded GMHD as a NeuroNet (as they call it) was you. >>> There is a 2017 reference, but they reference you again. >>> >>> (5) Its just names, fashion and preference.. or no actual concepts matter. Really? >>> >>> There was an french mathematician named Fourier in the 19th century who came up with an idea of periodic function decomposition into weighted trigonometric functions.. but he had no math. And Laplace Legendre and others said he had no math! So they prevented him from publishing for FIFTEEN YEARS.. 150 years later after Tukey invented the FFT, its the most common transform used and misused in general. >>> >>> Concepts lead to math.. and that may lead to further formalism.. but don't mistake the math for the concept behind it. The context matters and you are confusing syntax for semantics! >>> >>> Cheers, >>> Steve >>> >>> >>> >>> >>> >>> On 1/31/22 11:38 AM, Schmidhuber Juergen wrote: >>> >>>> Steve, do you really want to erase the very origins of shallow learning (Gauss & Legendre ~1800) and deep learning (DL, Ivakhnenko & Lapa 1965) from the field's history? Why? Because they did not use modern terminology such as "artificial neural nets (NNs)" and "learning internal representations"? Names change all the time like fashions; the only thing that counts is the math. Not only mathematicians but also psychologists like yourself will agree. >>>> >>>> Again: the linear regressor of Legendre & Gauss is formally identical to what was much later called a linear NN for function approximation (FA), minimizing mean squared error, still widely used today. No history of "shallow learning" (without adaptive hidden layers) is complete without this original shallow learner of 2 centuries ago. Many NN courses actually introduce simple NNs in this mathematically and historically correct way, then proceed to DL NNs with several adaptive hidden layers. >>>> >>>> And of course, no DL history is complete without the origins of functional DL in 1965 [DEEP1-2]. Back then, Ivakhnenko and Lapa published the first general, working DL algorithm for supervised deep feedforward multilayer perceptrons (MLPs) with arbitrarily many layers of neuron-like elements, using nonlinear activation functions (actually Kolmogorov-Gabor polynomials) that combine both additions (like in linear NNs) and multiplications (basically they had deep NNs with gates, including higher order gates). They incrementally trained and pruned their DL networks layer by layer to learn internal representations, using regression and a separate validation set (network depth > 7 by 1971). They had standard justifications of DL such as: "a multilayered structure is a computationally feasible way to implement multinomials of very high degree" [DEEP2] (that cannot be approximated by simple linear NNs). Of course, their DL was automated, and many people have used it up to the 2000s ! >>>> - just follow the numerous citations. >>>> >>>> I don't get your comments about Ivakhnenko's DL and function approximation (FA). FA is for all kinds of functions, including your "cognitive or perceptual or motor functions." NNs are used as FAs all the time. Like other NNs, Ivakhnenko's nets can be used as FAs for your motor control problems. You boldly claim: "This was not in the intellectual space" of Ivakhnenko's method. But obviously it was. >>>> >>>> Interestingly, 2 years later, Amari (1967-68) [GD1-2] trained his deep MLPs through a different DL method, namely, stochastic gradient descent (1951-52)[STO51-52]. His paper also did not contain the "modern" expression "learning internal representations in NNs." But that's what it was about. Math and algorithms are immune to rebranding. >>>> >>>> You may not like the fact that neither the original shallow learning (Gauss & Legendre ~1800) nor the original working DL (Ivakhnenko & Lapa 1965; Amari 1967) were biologically inspired. They were motivated through math and problem solving. The NN rebranding came later. Proper scientific credit assignment does not care for changes in terminology. >>>> >>>> BTW, unfortunately, Minsky & Papert [M69] made some people think that Rosenblatt [R58-62] had only linear NNs plus threshold functions. But actually he had much more interesting MLPs with a non-learning randomized first layer and an adaptive output layer. So Rosenblatt basically had what much later was rebranded as "Extreme Learning Machines (ELMs)." The revisionist narrative of ELMs (see this web site >>>> https://elmorigin.wixsite.com/originofelm >>>> ) is a bit like the revisionist narrative of DL criticized by my report. Some ELM guys apparently thought they can get away with blatant improper credit assignment. After all, the criticized DL guys seemed to get away with it on an even grander scale. They called themselves the "DL conspiracy" [DLC]; the "ELM conspiracy" is similar. What an embarrassing lack of maturity of our field. >>>> >>>> Fortunately, more and more ML researchers are helping to set things straight. "In science, by definition, the facts will always win in the end. As long as the facts have not yet won it's not yet the end." [T21v1] >>>> >>>> References as always under >>>> https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html >>>> >>>> >>>> J?rgen >>>> >>>> >>>> >>>>> On 27 Jan 2022, at 17:37, Stephen Jos? Hanson >>>>> wrote: >>>>> >>>>> >>>>> >>>>> Juergen, I have read through GMHD paper and a 1971 Review paper by Ivakhnenko. These are papers about function approximation. The method proposes to use series of polynomial functions that are stacked in filtered sets. The filtered sets are chosen based on best fit, and from what I can tell are manually grown.. so this must of been a tedious and slow process (I assume could be automated). So are the GMHDs "deep", in that they are stacked 4 deep in figure 1 (8 deep in another). Interestingly, they are using (with obvious FA justification) polynomials of various degree. Has this much to do with neural networks? Yes, there were examples initiated by Rumelhart (and me: >>>>> https://www.routledge.com/Backpropagation-Theory-Architectures-and-Applications/Chauvin-Rumelhart/p/book/9780805812596 >>>>> ), based on poly-synaptic dendrite complexity, but not in the GMHD paper.. which was specifically about function approximation. Ivakhnenko, lists four reasons for the approach t! >>>>> >>>> hey took: mainly reducing data size and being more efficient with data that one had. No mention of "internal representations" >>>> >>>>> So when Terry, talks about "internal representations" --does he mean function approximation? Not so much. That of course is part of this, but the actual focus is on cognitive or perceptual or motor functions. Representation in the brain. Hidden units (which could be polynomials) cluster and project and model the input features wrt to the function constraints conditioned by training data. This is more similar to model specification through function space search. And the original Rumelhart meaning of internal representation in PDP vol 1, was in the case of representation certain binary functions (XOR), but more generally about the need for "neurons" (inter-neurons) explicitly between input (sensory) and output (motor). Consider NETTALK, in which I did the first hierarchical clustering of the hidden units over the input features (letters). What appeared wasn't probably surprising.. but without model specification, the network (w.hidden units), learned VOWELS and ! >>>>> >>>> CONSONANT distinctions just from training (Hanson & Burr, 1990). This would be a clear example of "internal representations" in the sense of Rumelhart. This was not in the intellectual space of Ivakhnenko's Group Method of Handling Data. (some of this is discussed in more detail in some recent conversations with Terry Sejnowski and another one to appear shortly with Geoff Hinton (AIHUB.org >>>> look in Opinions). >>>> >>>>> Now I suppose one could be cynical and opportunistic, and even conclude if you wanted to get more clicks, rather than title your article GROUP METHOD OF HANDLING DATA, you should at least consider: NEURAL NETWORKS FOR HANDLING DATA, even if you didn't think neural networks had anything to do with your algorithm, after all everyone else is! Might get it published in this time frame, or even read. This is not scholarship. These publications threads are related but not dependent. And although they diverge they could be informative if one were to try and develop polynomial inductive growth networks (see Falhman, 1989; Cascade correlation and Hanson 1990: Meiosis nets) to motor control in the brain. But that's not what happened. I think, like Gauss, you need to drop this specific claim as well. >>>>> >>>>> With best regards, >>>>> >>>>> Steve >>>>> >>>> On 25 Jan 2022, at 20:03, Schmidhuber Juergen >>>> wrote: >>>> >>>> PS: Terry, you also wrote: "Our precious time is better spent moving the field forward.? However, it seems like in recent years much of your own precious time has gone to promulgating a revisionist history of deep learning (and writing the corresponding "amicus curiae" letters to award committees). For a recent example, your 2020 deep learning survey in PNAS [S20] claims that your 1985 Boltzmann machine [BM] was the first NN to learn internal representations. This paper [BM] neither cited the internal representations learnt by Ivakhnenko & Lapa's deep nets in 1965 [DEEP1-2] nor those learnt by Amari?s stochastic gradient descent for MLPs in 1967-1968 [GD1-2]. Nor did your recent survey [S20] attempt to correct this as good science should strive to do. On the other hand, it seems you celebrated your co-author's birthday in a special session while you were head of NeurIPS, instead of correcting these inaccuracies and celebrating the true pioneers of deep learning, such as ! >>>> Ivakhnenko and Amari. Even your recent interview >>>> https://blog.paperspace.com/terry-sejnowski-boltzmann-machines/ >>>> claims: "Our goal was to try to take a network with multiple layers - an input layer, an output layer and layers in between ? and make it learn. It was generally thought, because of early work that was done in AI in the 60s, that no one would ever find such a learning algorithm because it was just too mathematically difficult.? You wrote this although you knew exactly that such learning algorithms were first created in the 1960s, and that they worked. You are a well-known scientist, head of NeurIPS, and chief editor of a major journal. You must correct this. We must all be better than this as scientists. We owe it to both the past, present, and future scientists as well as those we ultimately serve. >>>> >>>> The last paragraph of my report >>>> https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html >>>> quotes Elvis Presley: "Truth is like the sun. You can shut it out for a time, but it ain't goin' away.? I wonder how the future will reflect on the choices we make now. >>>> >>>> J?rgen >>>> >>>> >>>> >>>>> On 3 Jan 2022, at 11:38, Schmidhuber Juergen >>>>> wrote: >>>>> >>>>> Terry, please don't throw smoke candles like that! >>>>> >>>>> This is not about basic math such as Calculus (actually first published by Leibniz; later Newton was also credited for his unpublished work; Archimedes already had special cases thereof over 2000 years ago; the Indian Kerala school made essential contributions around 1400). In fact, my report addresses such smoke candles in Sec. XII: "Some claim that 'backpropagation' is just the chain rule of Leibniz (1676) & L'Hopital (1696).' No, it is the efficient way of applying the chain rule to big networks with differentiable nodes (there are also many inefficient ways of doing this). It was not published until 1970 [BP1]." >>>>> >>>>> You write: "All these threads will be sorted out by historians one hundred years from now." To answer that, let me just cut and paste the last sentence of my conclusions: "However, today's scientists won't have to wait for AI historians to establish proper credit assignment. It is easy enough to do the right thing right now." >>>>> >>>>> You write: "let us be good role models and mentors" to the new generation. Then please do what's right! Your recent survey [S20] does not help. It's mentioned in my report as follows: "ACM seems to be influenced by a misleading 'history of deep learning' propagated by LBH & co-authors, e.g., Sejnowski [S20] (see Sec. XIII). It goes more or less like this: 'In 1969, Minsky & Papert [M69] showed that shallow NNs without hidden layers are very limited and the field was abandoned until a new generation of neural network researchers took a fresh look at the problem in the 1980s [S20].' However, as mentioned above, the 1969 book [M69] addressed a 'problem' of Gauss & Legendre's shallow learning (~1800)[DL1-2] that had already been solved 4 years prior by Ivakhnenko & Lapa's popular deep learning method [DEEP1-2][DL2] (and then also by Amari's SGD for MLPs [GD1-2]). Minsky was apparently unaware of this and failed to correct it later [HIN](Sec. I).... deep learning research was a! >>>>> >>>> live and kicking also in the 1970s, especially outside of the Anglosphere." >>>> >>>>> Just follow ACM's Code of Ethics and Professional Conduct [ACM18] which states: "Computing professionals should therefore credit the creators of ideas, inventions, work, and artifacts, and respect copyrights, patents, trade secrets, license agreements, and other methods of protecting authors' works." No need to wait for 100 years. >>>>> >>>>> J?rgen >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>>> On 2 Jan 2022, at 23:29, Terry Sejnowski >>>>>> wrote: >>>>>> >>>>>> We would be remiss not to acknowledge that backprop would not be possible without the calculus, >>>>>> so Isaac newton should also have been given credit, at least as much credit as Gauss. >>>>>> >>>>>> All these threads will be sorted out by historians one hundred years from now. >>>>>> Our precious time is better spent moving the field forward. There is much more to discover. >>>>>> >>>>>> A new generation with better computational and mathematical tools than we had back >>>>>> in the last century have joined us, so let us be good role models and mentors to them. >>>>>> >>>>>> Terry >>>>>> >>> -- >>> >> > -- > From benabbessarra at gmail.com Mon Feb 28 04:19:19 2022 From: benabbessarra at gmail.com (=?UTF-8?Q?Sarra_Ben_Abb=C3=A8s?=) Date: Mon, 28 Feb 2022 10:19:19 +0100 Subject: Connectionists: [CFP] 2nd International workshop on Ontology Uses and Contribution to Artificial Intelligence @PAKDD-2022 Message-ID: Dear colleagues and researchers, Please consider contributing to the 2nd edition of the international workshop "* Ontology Uses and Contribution to Artificial Intelligence* ", in conjunction with *PAKDD 2022* which will be held online or in Chengdu, China - May 16 - 19, 2022. ================================================================== The deadline for paper submissions is *March 11, 2022* ================================================================== *OnUCAI-2022* 2nd International workshop on Ontology Uses and Contribution to Artificial Intelligence at *PAKDD 2022* , Chengdu, China - May 16 - 19, 2022 Workshop website: https://sites.google.com/view/onucai-pakdd-2022 ================================================================== *Context* An ontology is well known to be the best way to represent knowledge in a domain of interest. It is defined by Gruber and Guarino as ?an explicit specification of a conceptualization?. It allows us to represent explicitly and formally existing entities, their relationships, and their constraints in an application domain. This representation is the most suitable and beneficial way to solve many challenging problems related to the information domain (e.g., knowledge representation, knowledge sharing, knowledge reusing, automated reasoning, knowledge capitalizing, and ensuring semantic interoperability among heterogeneous systems). Using ontology has many advantages, among them we can cite ontology reusing, reasoning, explanation, commitment, and agreement on a domain of discourse, ontology evolution, mapping, etc. As a field of artificial intelligence (AI), ontology aims at representing knowledge based on declarative and symbolic formalization. Combining this symbolic field with computational fields of IA such as Machine Learning (ML), Deep Learning (DL), Uncertainty and Probabilistic Graphical Models (PGMs), Computer Vision (CV), Multi-Agent Systems (SMA) and Natural Languages Processing (NLP) is a promising association. Indeed, ontological modeling plays a vital role to help AI reduce the complexity of the studied domain and organizing information inside it. It broadens AI?s scope allowing it to include any data type as it supports unstructured, semi-structured, or structured data format which enables smoother data integration. The ontology also assists AI for the interpretation process, learning, enrichment, prediction, semantic disambiguation, and discovery of complex inferences. Finally, the ultimate goal of ontologies is the ability to be integrated into the software to make sense of all information. In the last decade, ontologies are increasingly being used to provide background knowledge for several AI domains in different sectors (e.g. energy, transport, health, banking, insurance, etc.). Some of these AI domains are: - Machine learning and deep learning: semantic data selection, semantic data pre-processing, semantic data transformation, semantic data prediction, semantic clustering correction of the outputs, semantic enrichment with ontological concepts, use the semantic structure for promoting distance measure, etc. - Uncertainty and Probabilistic Graphical Models: learning PGM (structure or parameters) using ontologies, probabilistic semantic reasoning, semantic causality, probability, etc. - Computer Vision: semantic image processing, semantic image classification, semantic object recognition/classification, etc. - Blockchain: semantic transactions, interoperable blockchain systems, etc. - Natural Language Processing: semantic text mining, semantic text classification, semantic role labeling, semantic machine translation, semantic question answering, ontology-based text summarizing, semantic recommendation systems, etc. - Multi-Agent Systems and Robotics: semantic task composition, task assignment, communication, cooperation, coordination, plans, and plannification, etc. - Voice-video-speech: semantic voice recognition, semantic speech annotation, etc. - Game Theory: semantic definition of specific games, semantic rules and goals definition, etc. - etc. *Objective* This workshop aims at highlighting recent and future advances on the role of ontologies and knowledge graphs in different domains of AI and how they can be used in order to reduce the semantic gap between the data, applications, machine learning process, etc., in order to obtain semantic-aware approaches. In addition, the goal of this workshop is to bring together an area for experts from industry, science, and academia to exchange ideas and discuss the results of ongoing research in ontologies and AI approaches. ======================================================================== We invite the submission of original works that are related -- but are not limited to -- the topics below. *Topics of interest:* * Ontology for Machine Learning/Deep Learning * Ontology for Uncertainty and Probabilistic Graphical Models * Ontology for Edge Computing * Ontology for Federated Machine Learning * Ontology for Smart Contracts * Ontology for Computer Vision * Ontology for Natural Language Processing * Ontology for Robotics and Multi-agent Systems * Ontology for Voice-video-speech * Ontology for Game Theory * and so on. *Submission:* The workshop is open to submitting unpublished work resulting from research that presents original scientific results, methodological aspects, concepts, and approaches. All submissions are not anonymous and must be PDF documents written in English and formatted using the following style files: PAKDD2022_authors_kit Papers are to be submitted through the workshop's easychair submission page. We welcome the following types of contributions: * *Full papers* of up to 9 pages, including abstract, figures, and appendices (if any), but excluding references and acknowledgments: Finished or consolidated R&D works, to be included in one of the Workshop topics. ** Short papers* of up to 4 pages, excluding references and acknowledgments: Ongoing works with relevant preliminary results, opened to discussion. Submitting a paper to the workshop means that the authors agree that at least one author should attend the workshop to present the paper if the paper is accepted. For no-show authors, their affiliations will receive a notification. For further instructions, please refer to the PAKDD 2022 page. *Important dates:* * Workshop paper submission due: *March 11, 2022* * Workshop paper notifications: March 31, 2022 * Workshop paper camera-ready versions due: April 15, 2022 * Workshop: May 16-19, 2022 (Half-Day) All deadlines are 23:59 anywhere on earth (UTC-12). *Publication:* The accepted papers of this workshop may be included in the Proceedings of PAKDD 2022 Workshops published by Springer. ================================================================= *Workshop Chairs* * Sarra Ben Abb?s, Engie, France * Lynda Temal, Engie, France * Nada Mimouni, CNAM, France * Ahmed Mabrouk, Engie, France * Philippe Calvez, Engie, France *Program Committee* * Shridhar Devamane, Physical Design Engineer, Tecsec Technologies, Bangalore, India * Oudom Kem, Researcher at Engie, France * Philippe Leray, Professor at University of Nantes * Stefan Fenz, key researcher at SBA Research and Senior Scientist at Vienna University of Technology * Olivier Dameron, Professor at Universit? de Rennes I, Dyliss team, Irisa / Inria Rennes-Bretagne Atlantique * Ammar Mechouche, Data Science expert at AIRBUS Helicopters * Aar?n Ayll?n Benitez, PhD in bioinformatics and Ontology Lead at BASF Digital Solutions S.L. * Fran?ois Scharffe, Researcher on Knowledge-based AI, New York, United States * Maxime Lefran?ois, Associate Professor at Saint Etienne University, France * Pierre Maret, The QA Company & Saint Etienne University, France * Sanju Tiwari, Universidad Autonoma de Tamaulipas, Mexico -------------- next part -------------- An HTML attachment was scrubbed... URL: From benabbessarra at gmail.com Mon Feb 28 04:21:27 2022 From: benabbessarra at gmail.com (=?UTF-8?Q?Sarra_Ben_Abb=C3=A8s?=) Date: Mon, 28 Feb 2022 10:21:27 +0100 Subject: Connectionists: [CFP] 3rd International workshop on Deep Learning meets Ontologies and Natural Language Processing @ESWC-2022 Message-ID: Dear colleagues and researchers, Please consider contributing to the 3rd edition of the international workshop "*Deep Learning meets Ontologies and Natural Language Processing*" which will be held online or in Hersonissos, Greece - May 29 - June 2, 2022. ========================================================================= The deadline for paper submissions is *March 18th, 2022* ========================================================================= *DeepOntoNLP-2022* 3rd International workshop on Deep Learning meets Ontologies and Natural Language Processing at ESWC 2022 , Hersonissos, Greece - May 29 - June 2, 2022 Workshop website: https://sites.google.com/view/deepontonlp2022/ ========================================================================= *Context* In recent years, deep learning has been applied successfully and achieved state-of-the-art performance in a variety of domains, such as image analysis. Despite this success, deep learning models remain hard to analyze data and understand what knowledge is represented in them, and how they generate decisions. Deep Learning (DL) meets Natural Language Processing (NLP) to solve human language problems for further applications, such as information extraction, machine translation, search, and summarization. Previous works have attested the positive impact of domain knowledge on data analysis and vice versa, for example pre-processing data, searching data, redundancy and inconsistency data, knowledge engineering, domain concepts, and relationships extraction, etc. Ontology is a structured knowledge representation that facilitates data access (data sharing and reuse) and assists the DL process as well. DL meets recent ontologies and tries to model data representations with many layers of non-linear transformations. The combination of DL, ontologies, and NLP might be beneficial for different tasks: - Deep Learning for Ontologies: ontology population, ontology extension, ontology learning, ontology alignment, and integration, - Ontologies for Deep Learning: semantic graph embeddings, latent semantic representation, hybrid embeddings (symbolic and semantic representations), - Deep Learning for NLP: summarization, translation, named entity recognition, question answering, document classification, etc. - NLP for Deep Learning: parsing (part-of-speech tagging), tokenization, sentence detection, dependency parsing, semantic role labeling, semantic dependency parsing, etc. *Objective* This workshop aims at demonstrating recent and future advances in semantic rich deep learning by using Semantic Web and NLP techniques which can reduce the semantic gap between the data, applications, machine learning process, in order to obtain semantic-aware approaches. In addition, the goal of this workshop is to bring together an area for experts from industry, science, and academia to exchange ideas and discuss the results of ongoing research in natural language processing, structured knowledge, and deep learning approaches. ======================================================================== We invite the submission of original works that are related -- but are not limited to -- the topics below. Topics of interest: - Construction ontology embeddings - Ontology-based text classification - Learning ontology embeddings - Semantic role labeling - Ontology reasoning with Deep Neural Networks - Deep learning for ontological semantic annotations - Spatial and temporal ontology embeddings - Ontology alignment and matching based on deep learning models - Ontology learning from text using deep learning models - Unsupervised Learning - Text classification using deep models - Neural machine translation - Deep question answering - Deep text summarization - Deep speech recognition - and so on. Submission: The workshop is open to submitting unpublished work resulting from research that presents original scientific results, methodological aspects, concepts, and approaches. All submissions must be PDF documents written in English and formatted according to LNCS instructions for authors . Papers are to be submitted through the workshop's EasyChair submission page. We welcome the following types of contributions: - Full research papers (8-10 pages): Finished or consolidated R&D works, to be included in one of the Workshop topics - Short papers (4-6 pages): Ongoing works with relevant preliminary results, opened to discussion. At least one author of each accepted paper must register for the workshop, in order to present the paper, there, and at the conference. For further instructions please refer to the ESWC 2022 page. Important dates: - Workshop paper submission due: March 18th, 2022 - Workshop paper notifications: April 15th, 2022 - Workshop paper camera-ready versions due: April 22th, 2022 - Workshop: 30th of May, 2022 (afternoon-Half-Day) All deadlines are 23:59 anywhere on earth (UTC-12). Publication: The best papers from this workshop may be included in the supplementary proceedings of ESWC 2022. ======================================================================== *Workshop Chairs* Sarra Ben Abb?s, Engie, France Rim Hantach, Engie, France Philippe Calvez, Engie, France *Program Committee* *Nada Mimouni**,* CNAM, France Lynda Temal, Engie, France Davide Buscaldi, LIPN, Universit? Sorbonne Paris Nord, France Valentina Janev, Mihajlo Pupin Institute, Serbia Mohamed Hedi Karray, LGP-INP-ENIT, Universit? de Toulouse, France -------------- next part -------------- An HTML attachment was scrubbed... URL: From pkoenig at uos.de Mon Feb 28 05:56:43 2022 From: pkoenig at uos.de (=?UTF-8?Q?Peter_K=c3=b6nig?=) Date: Mon, 28 Feb 2022 11:56:43 +0100 Subject: Connectionists: =?utf-8?q?Computational_Neuroscience=3A_Assistant?= =?utf-8?q?_Professor_in_Computational_Neuroscience_=28W1=29_with_tenure_t?= =?utf-8?q?rack_option_=28to_W2=29_at_the_Institute_for_Cognitive_Science?= =?utf-8?q?=2C_University_of_Osnabr=C3=BCck=2E?= Message-ID: Great opportunity in Computational Neuroscience: Assistant Professor in Computational Neuroscience (W1) with tenure track option (to W2) at the Institute for Cognitive Science, University of Osnabr?ck. (apologies for multiple postings) Full details at: https://www.uni-osnabrueck.de/universitaet/stellenangebote/stellenangebote-detail/41-fb-8-ikw-assistant-professor-computational-neuroscience-w1-with-tt-w2/ --- This tenure-track Professorship is funded by the German Federal Government and the L?nder under its Program for the Promotion of Young Academics (Tenure-Track Program). Upon meeting the general administrative requirements, you will be employed as a civil servant on limited tenure for an initial period of three years. If you receive a positive evaluation after the initial three years, this period can be extended by up to three further years. If you meet Osnabr?ck University?s standards with respect to ability, competence and academic achievement, you will be offered a tenured W2 Professorship in accordance with the relevant legal provisions without further application. Tasks and Responsibilities: Research and teaching should be in the area of computational neuroscience. The professorship and research should be focused on understanding principles of neuronal information processing and cognitive processes. The professorship complements and cooperates with the existing research groups of the Institute of Cognitive Science at Osnabr?ck University. Further, the holder of the position should actively contribute to existing and future collaborative research projects (e.g. RTG Situated Cognition, RTG Computational Cognition) and should participate in the respective continuation research proposal of the Research Training Group Computational Cognition. In addition, the professorship should engage in teaching in all degree programs of Cognitive Science (BSc, MSc, PhD). Teaching is in English. Conditions of Employment: Condition of employment is: - Research experience in the field of neuronal information processing with the focus of understanding principles of neuronal information processing or cognitive processes. - Research focus on at least one of the two following topics: - Methods of Experimental cognitive neuroscience like EEG, MEG, fMRI, TMS, cell recordings, eyetracking, or virtual reality - Models of information processing in neurons, networks of neurons like for example deep neuronal networks and machine learning. - A coherent research profile, must be motivated by or explain experimental data and there must be a realistic perspective to implement all parts of the own research, including the experimental components or data collection, at the university of Osnabr?ck. - An excellent command of the English language. - Willingness to participate in university administrative processes and commission work. For non-German speaking applicants, willingness is expected to learn the German language with a sufficient proficiency within two years. In addition, it would be welcome: - Teaching experience in the field of the professorship - Experience with acquiring third party research funding. Legal Conditions of Employment: You will hold a first degree, have a strong commitment to teaching, and have demonstrated your ability to engage independently in advanced academic research, as a rule by obtaining an outstanding PhD (in accordance with Section 30 subsection 2 of the Lower Saxony Higher Education Act [NHG]) in a relevant subject. The time between the applicant?s final PhD examination or the completion of other equivalent qualifications which qualify the applicant for the position in accordance with Section 30 para. 2 sentence 1 No. 3 Lower Saxony Higher Education Act and the applicant?s application for the assistant professorship should be no longer than four years. This period of time does not include periods caring for a child or several children under 18 years of age or periods caring for a dependent relative and increases by up to two years per child or period of care to a maximum four years in multiple cases of care (Section 30 para. 5 Lower Saxony Higher Education Act). The position is available on a full-time or part-time basis. Osnabr?ck University is a family-friendly university and is committed to helping working/studying parents balance their family and working lives. Osnabr?ck University is actively seeking to increase the number of female Professors in its employ. Applications from women are therefore particularly welcome. If two candidates are equally qualified, preference will be given to the candidate with disability status. For further information, please contact Prof. Dr Gordon Pipa, Tel. +49 541-969-2277, E-Mail: gpipa at uni-osnabrueck.de. Please submit your application (including a resume with full details of your scholarly and scientific employment history, list of publications and courses taught as well as planned research) in electronic form (as one pdf file) together with the ?Bewerbungsprofil? [?Applicant Profile?] (DOCX, 13,58 kB) to the Dean of the School of Human Sciences, Prof. Dr. Susanne Boshammer, Universit?t Osnabr?ck, 49069 Osnabr?ck (bewerbungfb08 at uni-osnabrueck.de) to arrive by March 27, 2022. Please enter the code word "CNtt" in the subject of your e-mail. We look forward to receiving your application. -- Prof. Dr. Peter K?nig, Institute of Cognitive Science, University Osnabr?ck -- Prof. Dr. Peter K?nig, Institute of Cognitive Science, University Osnabr?ck From papaleon at sch.gr Mon Feb 28 06:13:56 2022 From: papaleon at sch.gr (Papaleonidas Antonios) Date: Mon, 28 Feb 2022 13:13:56 +0200 Subject: Connectionists: 18th AIAI 2022 Hybrid @ Crete, Greece - Submission extension Message-ID: <011001d82c94$46fac090$d4f041b0$@sch.gr> 18th AIAI 2022, 17 - 20 June 2022 Hybrid@ Web & Aldemar Knossos Royal, Crete, Greece www.ifipaiai.org/2022 CALL FOR PAPERS for 18th AIAI 2022 Hybrid @ Web & Crete, Greece Extended Paper Submission deadline: 10th of March 2022 Dear Antonios Papaleonidas We would like to invite you to submit your work at the 18th International Conference on Artificial Intelligence Applications and Innovations ( AIAI2022) 18th International Conference on Artificial Intelligence Applications and Innovations, AIAI 2022, is technically sponsored by IFIP Artificial Intelligence Applications WG12.5. It is going to be co-organized as a Joint event with 23rd Conference on Engineering Applications of Neural Networks, EANN 2022, which is technically sponsored by the INNS (International Neural Network Society). SPECIAL ISSUES - PROCEEDINGS: Selected papers will be published in 4 special issues of high quality international scientific Journals: * World Scientific journal, International Journal of Neural Systems, Impact factor 5.87 * Springer journal , Neural Computing and Applications, Impact Factor 5.61 * ???? journal, International Journal of Biomedical and Health Informatics, Impact factor 5.772 * Springer journal, AI & Ethics PROCEEDINGS will be published SPRINGER IFIP AICT Series and they are INDEXED BY SCOPUS, DBLP, Google Scholar, ACM Digital Library, IO-Port, MAthSciNet, CPCI, Zentralblatt MATH and EI Engineering Index Papers submissions will be up to 12 pages long and not less than 6 pages. BIBLIOMETRIC DETAILS: We proudly announce that according to Springer?s statistics, the last 15 AIAI conferences have been downloaded 1,719,00 times! IFIP AIAI series has reached h-index of 29 and published papers have been Cited more than 6000 times! For more Bibliometric Details please click at AIAI BIBLIOMETRIC DETAILS page IMPORTANT DATES: * Paper Submission Deadline: 25th of February 2022, 10th of March 2022 * Notification of Acceptance: 26th of March 2022 * Camera ready Submission: 22th of April 2022 * Early / Authors Registration Deadline: 22th of April 2022 * Conference: 17 - 20 of June 2022 WORKSHOPS & SPECIAL SESSIONS: So far, the following 8 high quality Workshops & Speccail Sessions have been accepted and scheduled: * 11th Mining Humanistic Data Workshop (MHDW 2022) * 7th Workshop on ?5G ? Putting Intelligence to the Network Edge? (5G-PINE 2021) * 2nd Defense Applications of AI Workshop (DAAI) an EDA ? EU Workshop * 2nd Distributed AI for Resource-Constrained Platforms Workshop (DARE 2022) * 2nd Artificial Intelligence in Biomedical Engineering and Informatics (AI-BEI 2022) * 2nd Artificail Intelligence & Ethics Workshop (AIETH 2022) * AI in Energy, Buildings and Micro-Grids Workshop (??BMG) * Machine Learning and Big Data in Health Care (ML at HC) For more info please visit AIAI 2022 workshop info page KEYNOTE SPEAKERS: So far two Plenary Lectures have been announced, both by distinguished Professors with an important imprint in AI and Machine Learning. * Professor Hojjat Adeli Ohio State University, Columbus, USA, Fellow of the Institute of Electrical and Electronics Engineers (IEEE) (IEEE), Honorary Professor, Southeast University, Nanjing, China, Member, Polish and Lithuanian Academy of Sciences, Elected corresponding member of the Spanish Royal Academy of Engineering. Visit Google Scholar profile , h-index: 114 * Professor Riitta Salmelin Department of Neuroscience and Biomedical Engineering Aalto University, Finland Visit Google Scholar profile, h-index: 65 * Professor Dr. Elisabeth Andr? Human-Centered Artificial Intelligence, Institute for Informatics, University of Augsburg, Germany Visit Google Scholar profile, h-index: 61 * Professor Verena Reiser School of Mathematical and Computer Sciences (MACS) at Heriot Watt University, Edinburgh Visit Google Scholar profile, h-index: 31 For more info please visit AIAI 2022 Keynote info page VENUE: ALDEMAR KNOSSOS ROYAL Beach Resort in Hersonisso Peninsula, Crete, Greece. Special Half Board prices have been arranged for the conference delegates in the Aldemar Knossos Royal Beach Resort. For details please see: https://ifipaiai.org/2022/venue/ Conference topics, CFPs, Submissions & Registration details can be found at: * ifipaiai.org/2022/calls-for-papers/ * ifipaiai.org/2022/paper-submission/ * ifipaiai.org/2022/registration/ We are expecting Submissions on all topics related to Artificial and Computational Intelligence and their Applications. Detailed Guidelines on the Topics and the submission details can be found at the links above General co-Chairs: * Ilias Maglogiannis, University of Piraeus, Greece * John Macintyre, University of Sunderland, United Kingdom Program co-Chairs: * Lazaros Iliadis, School of Engineering, Democritus University of Thrace, Greece * Konstantinos Votis, Information Technologies Institute, ITI Thessaloniki, Greece * Vangelis Metsis, Texas State University, USA *** Apologies for cross-posting *** Dr Papaleonidas Antonios Organizing - Publication & Publicity co-Chair of 23rd EANN 2022 & 18th AIAI 2022 Civil Engineering Department Democritus University of Thrace papaleon at civil.duth.gr papaleon at sch.gr -------------- next part -------------- An HTML attachment was scrubbed... URL: From francesco.piccialli at unina.it Mon Feb 28 07:20:21 2022 From: francesco.piccialli at unina.it (Francesco Piccialli) Date: Mon, 28 Feb 2022 13:20:21 +0100 Subject: Connectionists: IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE - Call For Paper In-Reply-To: <7a0189c0-08f7-0c57-2c23-ed50fc9d1aae@unina.it> References: <7a0189c0-08f7-0c57-2c23-ed50fc9d1aae@unina.it> Message-ID: Dear Colleague, In review of your expertise it is a pleasure for me to invite you and advise you about the following OPEN Call-For-Paper: IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE: Special Issue on Physics-Informed Machine Learning Link:http://wpage.unina.it/francesco.piccialli/CfP_IEEE_TAI_2022.html Link:https://cis.ieee.org/publications/ieee-transactions-on-artificial-intelligence/special-issues ??? DEADLINE for manuscript submission is: April 30, 2022 Looking forward to receive your valuable contribution. Kind Regards -- Francesco Piccialli, Ph.D. DMA - Department of Mathematics and Applications "R. Caccioppoli" University of Naples Federico II, Italy Tel. +39 081675852 Web: http://wpage.unina.it/francesco.piccialli/ M.O.D.A.L group: http://www.labdma.unina.it Institutional web: https://www.docenti.unina.it/francesco.piccialli