From suashdeb at gmail.com Fri Jul 1 02:47:26 2022 From: suashdeb at gmail.com (Suash Deb) Date: Fri, 1 Jul 2022 12:17:26 +0530 Subject: Connectionists: Extension of deadline, ISCMI22 (in memory of life and work of Prof. Lotfi Zadeh) Message-ID: Dear friends and esteemed colleagues, On numerous requests, the deadline for submission of manuscripts for ISCMI22, had been extended till 30th July'22 http://iscmi.us Hope in case you missed the earlier deadline, this will help you to proceed and submit your paper(s). Thank you and best regards, Suash Deb General Chair, ISCMI22 -------------- next part -------------- An HTML attachment was scrubbed... URL: From school at utia.cas.cz Fri Jul 1 05:04:44 2022 From: school at utia.cas.cz (Miroslav Karny) Date: Fri, 1 Jul 2022 11:04:44 +0200 Subject: Connectionists: [JOB] Postodoc positions in cybernetics, AI & applied math (Prague, UTIA , research institute of the Czech Academy of Sciences ) Message-ID: <1656666284351642493@utia.cas.cz> Dear all, I believe somebody around you could be interested in postodoc positions concerning various areas of cybernetics, AI, and applied mathematics. They offer work in domains: ? artificial intelligence and machine learning,? probabilistic graphical models,? statistics and stochastics,? image, video, and signal processing,? control theory,? adaptive decision intelligence and human-centric intelligence,? modelling economic and financial problems,? non-smooth analysis,? PDEs, calculus of variations, and continuum mechanics. They are advertised in the attached file and I am ready to provide additional information., especially, about the boldfaced topic. Best regards Miroslav K?rn? https://www.utia.cas.cz/people/karny -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: postdocposition-UTIA-2023b.pdf Type: application/pdf Size: 63870 bytes Desc: not available URL: From eloise.zehnder at univ-lorraine.fr Fri Jul 1 04:43:14 2022 From: eloise.zehnder at univ-lorraine.fr (Eloise Zehnder) Date: Fri, 1 Jul 2022 10:43:14 +0200 (CEST) Subject: Connectionists: =?utf-8?q?=5BExtended_deadline=5D=5BCFP=5D_RoMan2?= =?utf-8?q?022_online_Workshop_CFP_-The_=E2=80=9CTowards_Socially_Intellig?= =?utf-8?q?ent_Robots_In_Real-World_Applications=3A_Challenges_And_Intrica?= =?utf-8?b?Y2llc+KAnSAoU0lSUlcp?= Message-ID: <213068046.445934.1656664994345.JavaMail.zimbra@univ-lorraine.fr> [Apologies for cross-posting] Hello Everyone, We are happy to invite you to submit a paper to our upcoming full-day online workshop at IEEE Ro-Man 2022 . The ?Towards Socially Intelligent Robots In Real-World Applications: Challenges And Intricacies? (SIRRW) workshop is your opportunity to present your work and get an overview of state-of-the-art research in human-centered, real-life settings. Researchers from all HRI/Cognitive robotics/Machine Learning-relevant fields are invited to submit a 2-pages abstract or a 4 pages short-paper, before June 30th July 15th , 2022 . Selected papers will be presented as either an oral presentation or in a poster session. For more information, please see our website: [ https://sirrw-2022.github.io/ | https://sirrw-2022.github.io ] Overview ================ Being proactive, trustworthy, and dealing with uncertainty are the current challenges faced by social robots in real-world applications. To be seamlessly integrated into human-populated environments, robots will be expected to have intelligent social capabilities on top of their physical abilities. Human-Centered Artificial Intelligence has a critical role in providing robots with such capabilities by addressing some of the complex computational challenges that naturalistic human-robot interactions introduce. For a robot to behave proactively and in a manner that is appropriate to the context of interaction, it should cope with uncertainty when dealing with elements that are not fully observable and often hard to predict, such as the states representing the dynamic environment and humans. To build trustworthy interactions with humans, intelligent social robots need to successfully address challenging issues such as predicting human intentions, goals, expectations, understanding, and reasoning about the dynamic states of objects and other smart devices in the surroundings, and previous actions and their consequences while performing in complex situations. Trust is an important construct for evaluating adaptive social robot behaviors that could be inferred and evaluated through objective measures using computational models and subjective measures by human users. Such measures can be used to assess humans? disposition to be vulnerable around robots. Hence, addressing uncertainty is a key factor in developing trustworthy AI solutions and endowing robots with intelligent social capabilities. Important Dates ================ June 30th July 15th : Paper submission deadline August 5th : Acceptance notification August 17th : Camera-ready submission August 24th : Workshop Submissions ================ * Papers will be submitted through [ https://easychair.org/account/signin?l=BbbUuHtL2mSjVI7LN8iC06 | EasyChair ] * The manuscript should be of 2 or 4 pages in [ https://ras.papercept.net/conferences/support/support.php | IEEE double-column format ] excluding references. Relevant topics will include (but not restricted to): - Decision-making under uncertainty - Modeling human behavior - Communicative robot behavior generation - Automatic adaptation and personalization of robot behavior - Human-interactive robot learning - Planning methods for interactive robot behaviors - Perception for HRI - Cognitive architectures for interactive robots - Robot curiosity - Safety/Trust-critical applications for HRI - Human-robot collaboration - Reliability and explainability of robot decisions/actions Invited Speakers ================ * Prof. Agnieszka Wykowska (Instituto Italiano di Technologia (IIT)) * Prof. Tony Belpaeme (Ghent University and Plymouth University) * Dr. Dan Bohus (Microsoft) * Prof. Alan R. Wagner (Pennsylvania State University) Organizers and Contact ================ Dr. Melanie Jouaiti ( [ mailto:mjouaiti at uwaterloo.ca | mjouaiti at uwaterloo.ca ] , University of Waterloo) Dr. Sera Buyukgoz ( [ mailto:serabuyukgoz at gmail.com | serabuyukgoz at gmail.com ] , Sorbonne Universit? and Softbank Robotics Europe) Eloise Zehnder ( [ mailto:eloise.zehnder at univ-lorraine.fr | eloise.zehnder at univ-lorraine.fr ] , Universit? de Lorraine and Inria, Loria) Dr. Amir Aly ( [ mailto:amir.aly at plymouth.ac.uk | amir.aly at plymouth.ac.uk ] , Plymouth University) Pr. Kerstin Dautenhahn ( [ mailto:kerstin.dautenhahn at uwaterloo.ca | kerstin.dautenhahn at uwaterloo.ca ] , University of Waterloo) -------------- next part -------------- An HTML attachment was scrubbed... URL: From kerstin.ritter at bccn-berlin.de Fri Jul 1 04:58:29 2022 From: kerstin.ritter at bccn-berlin.de (Kerstin Ritter) Date: Fri, 1 Jul 2022 10:58:29 +0200 (CEST) Subject: Connectionists: =?utf-8?q?Fully-funded_PhD_and_PostDoc_position_?= =?utf-8?q?=284_years=29_in_Machine_/_Deep_Learning_in_Translational_Psych?= =?utf-8?q?iatry=2C_Charit=C3=A9_-_Universit=C3=A4tsmedizin_Berlin=2C_Germ?= =?utf-8?q?any?= Message-ID: <1304760499.688489.1656665909812.JavaMail.zimbra@bccn-berlin.de> Dear all, we offer one PostDoc position (4 years, 100%) and one PhD position (4 years, 75%) at Charit? - Universit?tsmedizin (Berlin, Germany) in Machine / Deep Learning in Translational Psychiatry in a newly established Research Unit 5187 "Precision Psychotherapy" (headed by Prof. Ulrike L?ken). Please see the job advertisements below. Please forward to potential candidates. Thanks, and best wishes, Kerstin (Ritter) ______________________________________________________________ Explainable Machine Learning / Deep Learning in Clinical Neuroimaging and Translational Psychiatry 1) PostDoc position (starting date: autumn 2022, 4 years; 100 %) at Charit? - Universit?tsmedizin (Berlin, Germany) At Charit? - Universit?tsmedizin Berlin and the Bernstein Center for Computational Neuroscience, we are looking for a motivated and highly qualified PostDoc for methods development at the intersection of explainable machine learning / deep learning and clinical neuroimaging / translational psychiatry. The position will be located in the research groups of Prof. Kerstin Ritter and Prof. John-Dylan Haynes at Charit? Berlin. The main task will be to predict response to cognitive-behavioral psychotherapy in retrospective data and a prospective cohort of patients with internalizing disorders including depression and anxiety from a complex, multimodal data set comprising tabular data as well as imaging data (e.g., clinical data, smartphone data, EEG, structural and functional MRI data). An additional task will be to contribute to the organization and maintenance of the prospective cohort. This study will be one of several projects in the newly established Research Unit 5187 "Precision Psychotherapy" (headed by Prof. Ulrike L?ken). Requirements for the Postdoc: - Excellent master and PhD degree in relevant field (e.g. computer science, mathematics, physics, psychology, computational neuroscience or related). - Excellent methodological skills (esp. machine learning, deep learning, explainable AI, medical image analysis) - Excellent programming skills in python - Excellent writing and communication skills (in English) - Strong interest in neuroscientific and psychiatric research 2) PhD position (starting date: autumn 2022, 4 years; 75 %) at Charit? - Universit?tsmedizin (Berlin, Germany) At Charit? - Universit?tsmedizin Berlin and the Bernstein Center for Computational Neuroscience, we are looking for a motivated and highly qualified PhD student for analyzing structural and functional MRI data in patients with internalizing disorders (e.g., depression and anxiety) using machine learning methods. An additional task will be the acquisition of fMRI data together with other PhD students in the newly established Research Unit 5187 "Precision Psychotherapy" (headed by Prof. Ulrike L?ken). The PhD student will be supervised by Prof. Kerstin Ritter, Prof. Ulrike L?ken and Prof. Norbert Kathmann. Requirements for the PhD student: - Excellent master in relevant field (e.g. computer science, mathematics, physics, psychology, computational neuroscience or related). - Very good methodological skills (esp. machine learning, deep learning, statistics, neuroimaging) - Very good programming skills in python - Excellent writing and communication skills (in English and German) - Strong interest in neuroscientific and psychiatric research Please send your application (motivation+CV+references) in one pdf-file (in English or German) asap to: Prof. Dr. rer. nat. Kerstin Ritter Charit? - Universit?tsmedizin Berlin Department of Psychiatry and Psychotherapy Bernstein Center for Computational Neuroscience Sauerbruchweg 4, Charit?platz 1, 10117 Berlin E-Mail: kerstin.ritter at charite.de Webpage: https://psychiatrie-psychotherapie.charite.de/en/research/translation_and_neurotechnology/machine_learning/ From vomlel at utia.cas.cz Fri Jul 1 08:14:03 2022 From: vomlel at utia.cas.cz (Jirka Vomlel) Date: Fri, 1 Jul 2022 14:14:03 +0200 Subject: Connectionists: Postdoc positions in Czech Academy of Sciences, Prague, EU Message-ID: <539fb059-7f9c-c2a5-c7f0-e9ee78e6edc2@utia.cas.cz> Institute of Information Theory and Automation, Czech Academy of Sciences, Prague, Czechia, EU invites applications? for two-year postdoctoral positions in the institute beginning in January 2023 with possibility to move to the tenure track in the institute. Candidates are expected to work in one of these areas: ? artificial intelligence and machine learning, ? probabilistic graphical models, ? statistics and stochastics, ? image, video, and signal processing, ? control theory, ? adaptive decision intelligence and human-centric intelligence, ? modelling economic and financial problems, ? non-smooth analysis, ? PDEs, calculus of variations, and continuum mechanics. The candidates are also expected to have a strong record of, or outstanding potential for, significant research and have no more than two years since being awarded a PhD, Dr. or equivalent title (as of September 30). Moreover, experience in obtaining third-party funds is advantageous. The Institute offers a monthly salary of CZK 50 000 (about 2 000 EURO) and yearly benefits supporting e.g. recreational and sport activities, as well as health care programs. Complete applications must be received by August 31, 2022. In case of interest, please send your application via email to utia at utia.cas.cz . The application should include a CV , a research statement, a motivation letter, and a copy of the PhD diploma. Letter(s) of recommendation is/are welcome. They should be sent by their authors directly to the email above. For further information see http://www.utia.cas.cz/news/3572 and the institute page at http://www.utia.cas.cz/ Jirka Vomlel From franrruiz87 at gmail.com Fri Jul 1 10:20:52 2022 From: franrruiz87 at gmail.com (=?UTF-8?Q?Francisco_J=2E_Rodr=C3=ADguez_Ruiz?=) Date: Fri, 1 Jul 2022 15:20:52 +0100 Subject: Connectionists: ICBINB Monthly Seminar Series Talk: Thomas Dietterich Message-ID: Dear all, We are pleased to announce that the next speaker of the *?I Can?t Believe It?s Not Better!? (**ICBINB)* virtual seminar series will be *Thomas Dietterich** (**Oregon State University**)*. More details about this series and the talk 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: *July 7th, 2022 at 11am EDT / 5pm CEST (*Note*: time different from usual one.) *Where: *RSVP for the Zoom link here: https://us02web.zoom.us/meeting/register/tZUuf?hpzgvEtxEIOcuo1-PJ8wDkvqmR8L6 *Title:* *Struggling to Achieve Novelty Detection in Deep Learning* *Abstract: **In 2005, motivated by an open world computer vision application, I became interested in novelty detection. However, there were few methods available in computer vision at that time, and my research turned to studying anomaly detection in standard feature vector data. In that arena, many good algorithms were being published. Fundamentally, these methods rely on a notion of distance or density in feature space and detect anomalies as outliers in that space.* *Returning to deep learning 10 years later, my students and I attempted, without much success. to apply these methods to the latent representations in deep learning. Other groups attempted to apply deep density models, again with limited success. Summary: I couldn?t believe it was not better. In the meantime, simple anomaly scores such as the maximum softmax probability of the max logit score were shown to be doing very well.We decided that we had reached the limits of what macro-level analysis (error rates, AUC scores) could tell us about these techniques. It was time to look closely at the actual feature values. In this talk, I?ll show our analysis of feature activations and introduce the Familiarity Hypothesis, which states that the max logit/max softmax score is measuring the amount of familiarity in an image rather than the amount of novelty. This is a direct consequence of the fact that the only features that are learned are ones that capture variability in the training data. Hence, deep nets can only represent images that fall within this variability. Novel images are mapped into this representation, and hence cannot be detected as outliers.I?ll close with some potential directions to overcome this limitation.* *Bio:* *Dr. Dietterich (AB Oberlin College 1977; MS University of Illinois 1979; PhD Stanford University 1984) is Distinguished Professor Emeritus in the School of Electrical Engineering and Computer Science at Oregon State University. Dietterich is one of the pioneers of the field of Machine Learning and has authored more than 225 refereed publications and two books. His current research topics include robust artificial intelligence, robust human-AI systems, and applications in sustainability.Dietterich has devoted many years of service to the research community and was recently given the ACML and AAAI distinguished service awards. He is a former President of the Association for the Advancement of Artificial Intelligence and the founding president of the International Machine Learning Society. Other major roles include Executive Editor of the journal Machine Learning, co-founder of the Journal for Machine Learning Research, and program chair of AAAI 1990 and NIPS 2000. He currently serves as one of the moderators for the cs.LG category on arXiv.* 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 xavier.hinaut at inria.fr Sun Jul 3 09:38:48 2022 From: xavier.hinaut at inria.fr (Xavier Hinaut) Date: Sun, 3 Jul 2022 15:38:48 +0200 Subject: Connectionists: SMILES workshop call for abstracts Message-ID: <52E5805A-21D6-4886-8B83-30D87EA0B203@inria.fr> The SMILES (Sensorimotor Interaction, Language and Embodiment of Symbols) Workshop will take place both on site and virtually at the ICDL 2022 (International Conference on Developmental Learning). * Call for abstracts : - Deadline: July 18th - Abstracts call: from 1/2 page to 2 pages (onsite and virtual participation are possible) - Abstract format: same as ICDL conference https://www.ieee.org/conferences/publishing/templates.html - Submissions: smiles.conf at gmail.com + indicate if you will be onsite or online - Workshop dates: September 12, 2022 - Venue onsite: Queen Mary University of London, UK. - Venue online: via Zoom and Discord group. Accepted abstract will be asked to make a short video or poster for the workshop. * Workshop Short Description On the one hand, models of sensorimotor interaction are embodied in the environment and in the interaction with other agents. On the other hand, recent Deep Learning development of Natural Language Processing (NLP) models allow to capture increasing language complexity (e.g. compositional representations, word embedding, long term dependencies). However, those NLP models are disembodied in the sense that they are learned from static datasets of text or speech. How can we bridge the gap from low-level sensorimotor interaction to high-level compositional symbolic communication? The SMILES workshop will address this issue through an interdisciplinary approach involving researchers from (but not limited to): - Sensori-motor learning, - Symbol grounding and symbol emergence, - Emergent communication in multi-agent systems, - Chunking of perceptuo-motor gestures (gestures in a general sense: motor, vocal, ...), - Compositional representations for communication and action sequence, - Hierarchical representations of temporal information, - Language processing and language acquisition in brains and machines, - Models of animal communication, - Understanding composition and temporal processing in neural network models, and - Enaction, active perception, perception-action loop. * More info - contact: smiles.conf at gmail.com - organizers: Xavier Hinaut, Cl?ment Moulin-Frier, Silvia Pagliarini, Joni Zhong, Michael Spranger, Tadahiro Taniguchi, Anne Warlaumont. - invited speakers (coming soon) - workshop website (updated regularly): https://sites.google.com/view/smiles-workshop/ - ICDL conference website: https://icdl2022.qmul.ac.uk/ Xavier Hinaut Inria Research Scientist www.xavierhinaut.com -- +33 5 33 51 48 01 Mnemosyne team, Inria, Bordeaux, France -- https://team.inria.fr/mnemosyne & LaBRI, Bordeaux University -- https://www4.labri.fr/en/formal-methods-and-models & IMN (Neurodegeneratives Diseases Institute) -- http://www.imn-bordeaux.org/en --- Our new release of Reservoir Computing library: https://github.com/reservoirpy/reservoirpy -------------- next part -------------- An HTML attachment was scrubbed... URL: From aruneshsinha at gmail.com Mon Jul 4 00:50:34 2022 From: aruneshsinha at gmail.com (Arunesh Sinha) Date: Mon, 4 Jul 2022 12:50:34 +0800 Subject: Connectionists: [Deadline Extended] Joint International Conference on Data Science & Management of Data (CODS-COMAD 2023) Message-ID: --------------------------------------------------------------------------------------------------- CODS-COMAD 2023 10th ACM IKDD CODS and 28th COMAD 6th Joint International Conference on Data Science & Management of Data January 4-7, 2023 IIT Bombay https://cods-comad.in/ ---------------------------------------------------------------------------------------------------- We are pleased to announce The 6th Joint International Conference on Data Sciences and Management of Data to be held in Mumbai, India, on January 4-7, 2023 as a hybrid event. The conference will be flexible with respect to in-person vs. online presentation of papers, though we encourage physical presence to facilitate deeper interactions. CODS-COMAD is a premier international conference focusing on scientific work in Databases, Data Sciences and their applications. Being held for the 6th time as a common conference bringing together the COMAD and the CODS communities, the conference invites researchers in the field of databases, data sciences and their applications to submit their original work. - CALL FOR PAPERS - RESEARCH TRACK The research track invites full as well as short papers describing innovative and original research contributions in the areas of data management, data science, machine learning and AI. Papers can range from theoretical contributions to systems and algorithms to experimental research and benchmarking. The goal of the short papers is to provide a venue for innovative ideas such as engineered solutions, exciting work-in-progress or even negative results that would be interesting to the broader community. Authors of accepted papers will get an opportunity to showcase their work as an oral presentation. For more details visit: https://cods-comad.in/call-for-research-track-papers.php APPLIED DATA SCIENCE TRACK The Applied Data Sciences (ADS) track invites both full as well as short papers describing the design, implementation and results of solutions and systems for application of data science techniques to real-world problems. Technical approaches can include data science, data mining, applied machine learning, testing and governance of data science models and solutions, and practical MLOps approaches. Accepted papers will be given the opportunity to present their work as an oral presentation. For more details, visit: https://cods-comad.in/call-for-applied-data-science-track-papers.php TUTORIALS Tutorials at CODS-COMAD offer a platform to showcase state-of-the-art tools and technologies to support research, development and applications of data systems, data science, data management and platforms, data-driven applications, ethics of data science etc. We solicit tutorial proposals on all topics of interest to the CODS-COMAD conference. For more details, visit: https://cods-comad.in/call-for-tutorial-proposals.php DEMO TRACK The demonstration (demo) track invites submissions from academia and industry where a live or pre-recorded demonstration of the software system, prototype, conceptual design or library adds significant value to the work. Submissions should not exceed 4 pages and (unlimited) references. Topics of interest include data systems, data science, data management and platforms, data-driven applications etc. The scope for demos is broad and includes all topics of interest to the CODS-COMAD conference. The demo should be the result of innovative work including solving novel technical or research problems and/or creating novel UI/UX. Authors of accepted demo track papers will have the option to do either a recorded demo or a live demo at the conference site. For more details, visit: https://cods-comad.in/call-for-demo-track-papers.php YOUNG RESEARCHERS' SYMPOSIUM The Young Researchers? Symposium at CODS-COMAD 2023 invites submissions from students and postdoctoral fellows. This is a unique opportunity for young researchers to have fruitful peer-to-peer discussions and to get feedback from leading senior researchers about their current research work. Submissions to the Young Researchers' Symposium (YRS) are in the form of a 2-page extended abstract, and are invited on the topics such as data systems, data science, data management and platforms, data-driven applications etc. We welcome many kinds of papers, such as, but not limited to: novel research papers, work-in-progress papers, appraisal papers of existing methods and tools (e.g., lessons learned). The first author of any submission should be an undergraduate/postgraduate/Ph.D. student or a Postdoctoral researcher affiliated with any academic institution. Authors of accepted papers will get a chance to present their work both as a poster and a quick lightning talk at the conference. For more details, visit: https://cods-comad.in/call-for-young-researchers-symposium-papers.php DIFFERENCE BETWEEN RESEARCH AND APPLIED DATA SCIENCE TRACK Check the link below to help you decide which track to submit your paper to. It is the authors? responsibility to submit their paper into the appropriate track. Papers that do not satisfy the requirements (e.g., a research track paper) might be rejected without a formal review. https://cods-comad.in/call-for-young-researchers-symposium-papers.php IMPORTANT DATES - July 10, 2022: Abstract submission deadline in Research and Applied Data Science tracks (Revised Date) - July 17, 2022: Paper submission deadline in Research and Applied Data Science tracks (Revised Date) - September 11, 2022: First Stage Accept/Major Revision/ Reject decisions in Research and Data Science tracks - September 15, 2022: Paper submission deadline in Demo and Young Researchers' Symposium tracks. Tutorial proposal deadline - October 9, 2022: Re-submission of revised manuscripts in Research and Applied Data Science tracks - October 30, 2022: Final Notifications of Accept / Reject decisions across all tracks - November 15, 2022: Camera ready submission deadline across all tracks PAPER SUBMISSION LINK https://easychair.org/conferences/?conf=codscomad2023 SUBMISSION INSTRUCTIONS AND POLICIES Detailed submission instructions, submission format, page limits, conflict of interest, dual submission, plagiarism and other policies are available at the link below. Do check them out before submitting your paper. https://cods-comad.in/common-instructions-policies.php AWARDS The best paper in each track will receive an award citation TRAVEL GRANTS Conference will provide travel assistance to a reasonable number of students whose papers are accepted. The travel grant includes free accommodation and monetary travel support partially covering the travel cost. Details of the grant will be made available at the conference website in due course. ORGANIZING COMMITTEE General Chairs - Pushpak Bhattacharyya (IIT Bombay) - Amr El Abbadi (University of California, Santa Barbara) PC Chairs (Research Track) - Praneeth Nethrapalli (Google Research) - Louiqa Raschid (University of Maryland) PC Chairs (Applied Data Science Track) - Tanuja Ganu (Microsoft) - Ponnurangam Kumaraguru (IIIT Hyderabad) Tutorial Chairs - Animesh Mukherjee (IIT Kharagpur) - Kalapriya Kannan (Hewlett Packard Enterprise) Demo Chairs - Mayank Vatsa (IIT Jodhpur) - Anoop Kunchukuttan (Microsoft) Young Researchers' Symposium (YRS) Chairs - Preethi Jyothi (IIT Bombay) - Bivas Mitra (IIT Kharagpur) Diversity & Inclusion Chairs - Chitra Babu (SSN Institutions) - Balaraman Ravindran (IIT Madras) Proceedings Chair - Abhinandan SP (NIE Mysore) - Charu Sharma (IIIT Hyderabad) Publicity Chairs - Rajiv Ratn Shah (IIIT Delhi) - Arunesh Sinha (Singapore Management University) Sponsorship Chairs - Shourya Roy (Flipkart) - Yogesh Simmhan (IISc Bangalore) Local Organizing Chairs - Biplab Banerjee (IIT Bombay) - Manoj Nambiar (TCS Research) Finance Chair - Chandrashekhar Sahasrabudhe (ACM India) - Raj Sharma (Goldman Sachs) Thanks, Arunesh Sinha -------------- next part -------------- An HTML attachment was scrubbed... URL: From marcin at amu.edu.pl Sat Jul 2 04:58:47 2022 From: marcin at amu.edu.pl (Marcin Paprzycki) Date: Sat, 2 Jul 2022 10:58:47 +0200 Subject: Connectionists: Table of Contents; Scalable Computing: Practice and Experience (SCPE); Vol 22 No 2 (2021) In-Reply-To: <55da44a6-c6e5-2800-386c-c3ce5ad11fec@pti.org.pl> References: <55da44a6-c6e5-2800-386c-c3ce5ad11fec@pti.org.pl> Message-ID: Continuing process of catching up (though issues till Vol 23, No 1 have been published ;-); Table of Contents; Scalable Computing: Practice and Experience (SCPE); Vol 22 No 2 (2021): ALL papers available (OPEN ACCESS) via: http://scpe.org/index.php/scpe/issue/view/153 Introduction to the Special Issue on Artificial Intelligence for Smart Cities and Industries Ashutosh Sharma, Pradeep Kumar Singh, Wei-Chiang Hong, Gaurav Dhiman, Adam Slowik 89-91 Special Issue Papers Research on Construction Cost Estimation based on Artificial Intelligence Technology Bin Wang, Jianjun Yuan, Kayhan Zrar Ghafoor 93-104 An IOT and Blockchain Approach for the Smart Water Management System in Agriculture Yunyan Chang, Jian Xu, Kayhan Zrar Ghafoor 105?116 Design and Research on the Intelligent System of Urban Rail Transit Project based on BIM+GIS Yan Liu, Mohd Asif Shah, Anton Pljonkin, Mohammad Asif Ikbal, Mohammad Shabaz 117?126 An IOT and Blockchain Approach for Food Traceability System in Agriculture Jianli Guo, Korhan Cengiz, Ravi Tomar 127?137 Design and Application of College Online Education Platform Based on WebRTC Guoliang Li, Rixing Wang, Qikun Zhou 139?148 Research on Data Security Detection Algorithm in IoT Based on K-means Jianxing Zhu, Lina Huo, Mohd Dilshad Ansari, Mohammad Asif Ikbal 149?159 Network Virus and Computer Network Security Detection Technology Optimization Zhifeng Hu, Feng Zhao, Lina Qin, Hongkai Lin 161?170 A Detailed Study on GPS and GIS Enabled Agricultural Equipment Field Position Monitoring system for Smart Farming Jianbo Nie, Bin Yang 171?181 Design of Intelligent Building Scheduling System for Internet of Things and Cloud Computing Tiangang Wang, Zhe Mi 183?192 Research on TCP Performance Model and Transport Agent Architecture in Broadband Wireless Network Lintao Li, Parv Sharma, Mehdi Gheisari, Amit Sharma 193?201 Research on Multi-Agent Systems in a Smart Small Grid for Resource Apportionment and Planning Zhixian Yang, Kshuangchen Fu, Jhon Paul 203?213 Study and Research on IoT and Big Data Analysis for Smart City Development Haixia Yu, Ion Cosmin Mihai, Anand Srivastava 215?225 Cloud based Resource Scheduling Methodology for Data-Intensive Smart Cities and Industrial Applications Shiming Ma, Jichang Chen, Yang Zhang, Anand Shrivastava, Hari Mohan 227?235 Research on Mobile User Interface for Robot Arm Remote Control in Industrial Application Jiangnan Ni, Vipin Balyan 237?245 A Cluster based Intelligent Method to Manage Load of Controllers in SDN-IoT Networks for Smart Cities Surendra Kumar Keshari, Vineet Kansal, Sumit Kumar 247?257 Emergency Rapid Response to Epileptic Seizures - A Novel IOT Framework for Smart Cities Shabana R Ziyad, Armaan Ziyad 259-272 Enhanced Secure ATM authentication using NFC Technology and Iris Verification Smita S Agrawal, Parita Oza, Mahima Biswas, Neer Choksi 273-282 * SCPE does NOT charge any fees for publishing Open Access papers and is Indexed in SCOPUS and in Clarivate Analytics (former Thompson Reuters) Emerging Sources Citation Index * From david at irdta.eu Sat Jul 2 05:57:21 2022 From: david at irdta.eu (David Silva - IRDTA) Date: Sat, 2 Jul 2022 11:57:21 +0200 (CEST) Subject: Connectionists: DeepLearn 2023 Winter: early registration July 4 Message-ID: <1550563570.205598.1656755841234@webmail.strato.com> ****************************************************************** 8th INTERNATIONAL SCHOOL ON DEEP LEARNING DeepLearn 2023 Winter Bournemouth, UK January 16-20, 2023 https://irdta.eu/deeplearn/2023wi/ *********** Co-organized by: Department of Computing and Informatics Bournemouth University Institute for Research Development, Training and Advice ? IRDTA Brussels/London ****************************************************************** Early registration: July 4, 2022 ****************************************************************** SCOPE: DeepLearn 2022 Winter 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, Las Palmas de Gran Canaria and Lule?. Deep learning is a branch of artificial intelligence covering a spectrum of current exciting 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 2023 Winter 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 2023 Winter will take place in Bournemouth, a coastal resort town on the south coast of England. The venue will be: Talbot Campus Bournemouth University https://www.bournemouth.ac.uk/about/contact-us/directions-maps/directions-our-talbot-campus 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: Yi Ma (University of California, Berkeley), CTRL: Closed-Loop Data Transcription via Rate Reduction Daphna Weinshall (Hebrew University of Jerusalem), Curriculum Learning in Deep Networks Eric P. Xing (Carnegie Mellon University), It Is Time for Deep Learning to Understand Its Expense Bills PROFESSORS AND COURSES: (to be completed) Mohammed Bennamoun (University of Western Australia), [intermediate/advanced] Deep Learning for 3D Vision Matias Carrasco Kind (University of Illinois, Urbana-Champaign), [intermediate] Anomaly Detection Nitesh Chawla (University of Notre Dame), [introductory/intermediate] Graph Representation Learning Seungjin Choi (Intellicode), [introductory/intermediate] Bayesian Optimization over Continuous, Discrete, or Hybrid Spaces Sumit Chopra (New York University), [intermediate] Deep Learning in Healthcare Luc De Raedt (KU Leuven), [introductory/intermediate] Statistical Relational and Neurosymbolic AI Marco Duarte (University of Massachusetts, Amherst), [introductory/intermediate] Explainable Machine Learning Jo?o Gama (University of Porto), [introductory] Learning from Data Streams: Challenges, Issues, and Opportunities Claus Horn (Zurich University of Applied Sciences), [intermediate] Deep Learning for Biotechnology Zhiting Hu (University of California, San Diego) & Eric P. Xing (Carnegie Mellon University), [intermediate/advanced] A "Standard Model" for Machine Learning with All Experiences Nathalie Japkowicz (American University), [intermediate/advanced] Learning from Class Imbalances Gregor Kasieczka (University of Hamburg), [introductory/intermediate] Deep Learning Fundamental Physics: Rare Signals, Unsupervised Anomaly Detection, and Generative Models Karen Livescu (Toyota Technological Institute at Chicago), [intermediate/advanced] Speech Processing: Automatic Speech Recognition and beyond (to be confirmed) David McAllester (Toyota Technological Institute at Chicago), [intermediate/advanced] Information Theory for Deep Learning Dhabaleswar K. Panda (Ohio State University), [intermediate] Exploiting High-performance Computing for Deep Learning: Why and How? Fabio Roli (University of Cagliari), [introductory/intermediate] Adversarial Machine Learning Richa Singh (Indian Institute of Technology Jodhpur), [introductory/intermediate] Trusted AI Kunal Talwar (Apple), [introductory/intermediate] Foundations of Differentially Private Learning Tinne Tuytelaars (KU Leuven), [introductory/intermediate] Continual Learning in Deep Neural Networks Lyle Ungar (University of Pennsylvania), [intermediate] Natural Language Processing using Deep Learning Bram van Ginneken (Radboud University Medical Center), [introductory/intermediate] Deep Learning for Medical Image Analysis Yu-Dong Zhang (University of Leicester), [introductory/intermediate] Convolutional Neural Networks and Their Applications to COVID-19 Diagnosis 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 January 8, 2023. 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 January 8, 2023. 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 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 January 8, 2023. ORGANIZING COMMITTEE: Rashid Bakirov (Bournemouth, local co-chair) Marcin Budka (Bournemouth) Vegard Engen (Bournemouth) Nan Jiang (Bournemouth, local co-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/2023wi/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 are available at https://irdta.eu/deeplearn/2023wi/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: Bournemouth University Rovira i Virgili University Institute for Research Development, Training and Advice ? IRDTA, Brussels/London -------------- next part -------------- An HTML attachment was scrubbed... URL: From oliver at roesler.co.uk Sat Jul 2 12:34:51 2022 From: oliver at roesler.co.uk (Oliver Roesler) Date: Sat, 2 Jul 2022 16:34:51 +0000 Subject: Connectionists: Deadline Extension - CFP RO-MAN 2022 Workshop on Machine Learning for HRI: Bridging the Gap between Action and Perception Message-ID: <3ad1c82d-fc30-b3ce-a002-411e706da317@roesler.co.uk> *DEADLINE EXTENSION* **Apologies for cross-posting** We are happy to announce that the deadline for submissions has been extended until _*July 15*_. *CALL FOR PAPERS* The *full-day virtual* workshop: *Machine Learning for HRI: Bridging the Gap between Action and Perception (ML-HRI)* In conjunction with the *31st IEEE International Conference on Robot and**Human Interactive Communication (RO-MAN) - August 22, 2022??? * Webpage:?https://ml-hri2022.ivai.onl/ *I. Aim and Scope* A key factor for the acceptance of robots as partners in complex and dynamic human-centered environments is their ability to continuously adapt their behavior. This includes learning the most appropriate behavior for each encountered situation based on its specific characteristics as perceived through the robots senors. To determine the correct actions the robot has to take into account prior experiences with the same agents, their current emotional and mental states, as well as their specific characteristics, e.g. personalities and preferences. Since every encountered situation is unique, the appropriate behavior cannot be hard-coded in advance but must be learned over time through interactions. Therefore, artificial agents need to be able to learn continuously what behaviors are most appropriate for certain situations and people based on feedback and observations received from the environment to enable more natural, enjoyful, and effective interactions between humans and robots. This workshop aims to attract the latest research studies and expertise in human-robot interaction and machine learning at the intersection of rapidly growing communities, including social and cognitive robotics, machine learning, and artificial intelligence, to present novel approaches aiming at integrating and evaluating machine learning in HRI. Furthermore, it will provide a venue to discuss the limitations of the current approaches and future directions towards creating robots that utilize machine learning to improve their interaction with humans. *II. Keynote Speakers and Panelists* 1. *Dorsa Sadigh* ? Stanford University ? USA 2. *Oya Celiktutan* ? King's College London ? UK 3. *Sean Andrist *??Microsoft ? USA 4. *Stefan Wermter* ? University of Hamburg ? Germany *III. Submission* 1. For paper submission, use the following EasyChair web link: Paper Submission . 2. Use the RO-MAN 2022 format: RO-MAN Papers Templates . 3. Submitted papers should be 4-6 pages for regular papers and 2 pages for position papers. ??? The primary list of topics covers the following points (but not limited to): * Autonomous robot behavior adaptation * Interactive learning approaches for HRI * Continual learning * Meta-learning * Transfer learning * Learning for multi-agent systems * User adaptation of interactive learning approaches * Architectures, frameworks, and tools for learning in HRI * Metrics and evaluation criteria for learning systems in HRI * Legal and ethical considerations for real-word deployment of learning approaches *IV. Important Dates* 1. Paper submission: *June 17, 2022**July 15, 2022 (AoE)* 2. Notification of acceptance: *August 1, 2022* *August 7, 2022 (AoE)* 3. Camera ready: *August 14, 2022 (AoE)* 4. Workshop: *August 22, 2022* *V. Organizers* 1. *Oliver Roesler* ? IVAI ? Germany 2. *Elahe Bagheri* ? IVAI ? Germany 3. *Amir Aly* ? University of Plymouth ? UK -------------- next part -------------- An HTML attachment was scrubbed... URL: From jncor at dei.uc.pt Sun Jul 3 03:00:00 2022 From: jncor at dei.uc.pt (=?UTF-8?Q?Jo=C3=A3o_Nuno_Correia?=) Date: Sun, 3 Jul 2022 08:00:00 +0100 Subject: Connectionists: Call for Nominations - Julian Francis Miller Award Message-ID: Dear Colleague(s), Below you will find the call for nominations for the "Julian Francis Miller Award". Feel free to distribute. Thank you for your time! ---- Jo?o Correia ------------------------------------------------ (Apologies for cross-posting) ************************************************************************************************** CALL FOR NOMINATIONS JULIAN FRANCIS MILLER AWARD ************************************************************************************************** SPECIES, the 'Society for the Promotion of Evolutionary Computation in Europe and its Surroundings' has announced a new award, the Julian Francis Miller Award, for important contributions to the algorithmic exploration and embodiment of evolution, development and/or learning. The award will be handed out every year at the EvoStar Conference, beginning in 2023. Please have a read and consider to submit a nomination for an appropriate candidate to be considered. Deadline for submission is July 31st of this year. Please visit the SPECIES website at http://species-society.org/julian-francis-miller-award/ for further details on the submission process. Contact: Wolfgang Banzhaf (banzhafw at msu.edu) Penousal Machado (machado at dei.uc.pt) Thank you. -------------- next part -------------- An HTML attachment was scrubbed... URL: From ludovico.montalcini at gmail.com Sun Jul 3 04:06:23 2022 From: ludovico.montalcini at gmail.com (Ludovico Montalcini) Date: Sun, 3 Jul 2022 10:06:23 +0200 Subject: Connectionists: CfP: The 8th Int. Online & Onsite Conf. On Machine Learning, Optimization & Data Science - LOD 2022, September 18-22, Certosa di Pontignano, Tuscany - Italy - Late Breaking Paper Submission Deadline: July 15 In-Reply-To: References: Message-ID: CfP: The 8th Int. Online & Onsite Conf. On Machine Learning, Optimization & Data Science - LOD 2022, September 18-22, Certosa di Pontignano, Tuscany - Italy - Late Breaking Paper Submission Deadline: July 15 Dear Colleague, Apologies if you receive multiple copies of this announcement. Please kindly help forward it to potentially interested authors/attendees, thanks! The 8th International Online & Onsite Conference on Machine Learning, Optimization, and Data Science? #LOD2022 - September 18-22, Certosa di Pontignano, #Tuscany - Italy LOD 2022 An Interdisciplinary Conference: #MachineLearning, #Optimization, #BigData & #ArtificialIntelligence, #DeepLearning without Borders https://lod2022.icas.cc lod at icas.cc Late Breaking PAPERS SUBMISSION: July 15 (Anywhere on Earth) All papers must be submitted using EasyChair: https://easychair.org/conferences/?conf=lod2022 LOD 2022 KEYNOTE SPEAKER(S): * Pierre Baldi, University of California Irvine, USA * Jurgen Bajorath, University of Bonn, Germany * Ross King, University of Cambridge, UK & The Alan Turing Institute, UK * Rema Padman, Carnegie Mellon University, USA *Panos Pardalos, University of Florida, USA LOD 2022 TUTORIAL SPEAKER: * Simone Scardapane, University of Rome "La Sapienza", Italy ACAIN 2022 KEYNOTE SPEAKERS: * 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 * Mate 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! https://acain2022.artificial-intelligence-sas.org/course-lecturers/ PAPER FORMAT: Please prepare your paper using the Springer Nature ? Lecture Notes in Computer Science (LNCS) template. Papers must be submitted in PDF. TYPES OF SUBMISSIONS: When submitting a paper to LOD 2022, authors are required to select one of the following four types of papers: * long paper: original novel and unpublished work (max. 15 pages in Springer LNCS format); * short paper: an extended abstract of novel work (max. 5 pages); * work for oral presentation only (no page restriction; any format). For example, work already published elsewhere, which is relevant, and which may solicit fruitful discussion at the conference; * abstract for poster presentation only (max 2 pages; any format). The poster format for the presentation is A0 (118.9 cm high and 84.1 cm wide, respectively 46.8 x 33.1 inch). For research work which is relevant, and which may solicit fruitful discussion at the conference. Each paper submitted will be rigorously evaluated. The evaluation will ensure the high interest and expertise of reviewers. Following the tradition of LOD, we expect high-quality papers in terms of their scientific contribution, rigor, correctness, novelty, clarity, quality of presentation and reproducibility of experiments. Accepted papers must contain significant novel results. Results can be either theoretical or empirical. Results will be judged on the degree to which they have been objectively established and/or their potential for scientific and technological impact. It is also possible to present the talk virtually (Zoom). LOD 2022 Special Sessions: https://lod2022.icas.cc/special-sessions/ https://easychair.org/my/conference?conf=lod2022 PAST LOD KEYNOTE SPEAKERS: https://lod2022.icas.cc/past-keynote-speakers/ Yoshua Bengio, Head of the Montreal Institute for Learning Algorithms (MILA) & University of Montreal, Canada Bettina Berendt, TU Berlin, Germany & KU Leuven, Belgium, and Weizenbaum Institute for the Networked Society, Germany Jorg Bornschein, DeepMind, London, UK Michael Bronstein, Imperial College London, UK Nello Cristianini, University of Bristol, UK Peter Flach, University of Bristol, UK, and EiC of the Machine Learning Journal Marco Gori, University of Siena, Italy Arthur Gretton, UCL, UK Arthur Guez, Google DeepMind, Montreal, UK Yi-Ke Guo, Imperial College London, UK George Karypis, University of Minnesota, USA Vipin Kumar, University of Minnesota, USA Marta Kwiatkowska, University of Oxford, UK George Michailidis, University of Florida, USA Kaisa Miettinen, University of Jyvskyl, Finland Stephen Muggleton, Imperial College London, UK Panos Pardalos, University of Florida, USA Jan Peters, Technische Universitaet Darmstadt & Max-Planck Institute for Intelligent Systems, Germany Tomaso Poggio, MIT, USA Andrey Raygorodsky, Moscow Institute of Physics and Technology, Russia Mauricio G. C. Resende, Amazon.com Research and University of Washington Seattle, Washington, USA Ruslan Salakhutdinov, Carnegie Mellon University, USA, and AI Research at Apple Maria Schuld, Xanadu & University of KwaZulu-Natal, South Africa Richard E. Turner, Department of Engineering, University of Cambridge, UK Ruth Urner, York University, Toronto, Canada Isabel Valera, Saarland University, Saarbrucken & Max Planck Institute for Intelligent Systems, Tubingen, Germany TRACKS & SPECIAL SESSIONS: https://lod2022.icas.cc/special-sessions/ BEST PAPER AWARD: Springer sponsors the LOD 2022 Best Paper Award https://lod2022.icas.cc/best-paper-award/ PROGRAM COMMITTEE: https://lod2022.icas.cc/program-committee/ SCHEDULE: https://lod2022.icas.cc/wp-content/uploads/sites/20/2022/02/LOD-2022-Schedule-Ver-1.pdf VENUE: https://lod2022.icas.cc/venue/ The venue of LOD 2022 will be The Certosa di Pontignano? Siena The Certosa di Pontignano Localita? 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://lod2022.icas.cc/activities/ POSTER: https://lod2022.icas.cc/wp-content/uploads/sites/20/2022/02/poster-LOD-2022-1.png Submit your research work today! https://easychair.org/conferences/?conf=lod2022 See you in the beautiful Tuscany in September! Best regards, LOD 2022 Organizing Committee LOD 2022 NEWS: https://lod2022.icas.cc/category/news/ Past Editions https://lod2022.icas.cc/past-editions/ LOD 2021, The Seventh International Conference on Machine Learning, Optimization and Big Data Grasmere ? Lake District ? England, UK. Nature Springer ? LNCS volumes 13163 and 13164. LOD 2020, The Sixth International Conference on Machine Learning, Optimization and Big Data Certosa di Pontignano ? Siena ? Tuscany ? Italy. Nature Springer ? LNCS volumes 12565 and 12566. LOD 2019, The Fifth International Conference on Machine Learning, Optimization and Big Data Certosa di Pontignano ? Siena ? Tuscany ? Italy. Nature Springer ? LNCS volume 11943. LOD 2018, The Fourth International Conference on Machine Learning, Optimization and Big Data Volterra ? Tuscany ? Italy. Nature Springer ? LNCS volume 11331. MOD 2017, The Third International Conference on Machine Learning, Optimization and Big Data Volterra ? Tuscany ? Italy. Springer ? LNCS volume 10710. MOD 2016, The Second International Workshop on Machine learning, Optimization and big Data Volterra ? Tuscany ? Italy. Springer ? LNCS volume 10122. MOD 2015, International Workshop on Machine learning, Optimization and big Data Taormina ? Sicily ? Italy. Springer ? LNCS volume 9432. https://www.facebook.com/groups/2236577489686309/ https://twitter.com/TaoSciences https://www.linkedin.com/groups/12092025/ lod at icas.cc https://lod2022.icas.cc * Apologies for multiple copies. Please forward to anybody who might be interested * -------------- next part -------------- An HTML attachment was scrubbed... URL: From elscdigital at savion.huji.ac.il Mon Jul 4 03:26:10 2022 From: elscdigital at savion.huji.ac.il (digital elsc) Date: Mon, 4 Jul 2022 07:26:10 +0000 Subject: Connectionists: Postdoc positions at ELSC Message-ID: Call for applications - ELSC's post-doctoral brain sciences fellowships! ELSC invites exceptional researchers with a Ph.D. to join our flourishing community and participate in outstanding research. This postdoctoral program reaches across several disciplines to provide the tools and skills for the theoretical and experimental study of the dynamics and computational capabilities of neural networks. Our goal is to develop new insights into the mechanisms underlying both brain function and dysfunction. [cid:bfffbe36-5527-4b14-9e62-58fc039e0208] For more info, click here: https://elsc.huji.ac.il/opportunities/postdoctoral-program/about-the-program/ Gaia Niv Digital & Social Media Coordinator Edmond and Lily Safra Center for Brain Sciences Edmond J. Safra Campus, Jerusalem Hebrew University of Jerusalem 9190401 Mobile: 054-722-1880 [https://ci3.googleusercontent.com/mail-sig/AIorK4xvgoghUXyR2IYy7NoezUCwCcX5T2iRJiR1k8UMwh-whLRLolXT4Q1jfwO2MvFmKDXOVo_lNT4] [https://ci4.googleusercontent.com/proxy/AXrVIefSMsmLJIR7uBqkEbJ5mWimhTZllJgb6dxdbvFY0Y1EFLf2fie5qhb0K022JzX3ym_9HtrstWJEBtYsyAQzU4Wvvq2T-BV9p-jvlVbfaJBHnZ5HxypJwgbY4HqM2_8UFeOTyk5SG2ijBJm3I_tV6HPKJ3CuTJe8SU0_voKDPqRNP13LDIDeQSjpnhVj2Kv5-Rc14KhbSuK-jg=s0-d-e1-ft#https://docs.google.com/uc?export=download&id=1W6Ti3LH1BhIo8Y-SGu2YfhU4m9jtZkna&revid=0ByX1PVCIx8ZWd3FsWnpYTmRQa1dQb094NTNFcWlCTjgvSG5RPQ] [https://ci4.googleusercontent.com/proxy/KbfUurOrc0DFajmKsMbi8FXkxm-kZWKB7-jnHYcar4bXaiILMb6xZXCkb9BzxZ5m_65xrDd5GPq5kLVG2qw10Jvtl-LaYwWCTXxDZFPDyV9-tM1vA8FS8vek4rphPAkfsQtE7aVRrAlMlD9HBs__R0cj4Q68sIuUqk2leo8t-JgGVWqcUvZTnIGZakspSUfJ7MDUNOww0LQR8V6kVg=s0-d-e1-ft#https://docs.google.com/uc?export=download&id=1J_rELi6nOVHMR2SO5M6Ftwv6YR0L6yMG&revid=0ByX1PVCIx8ZWemhSYUxCbWJ4N3lxRlEvNDBONWF4a1IwYWc0PQ] [https://ci5.googleusercontent.com/proxy/-8wy5g4XqGjzEozfclykLmS-luLWl1Jx8m2ABkqsH-A7R5wDdjCnFVCwaV-AeAR_vGv7W1YP2btgC2SXP7iwbBe4a8GYDh-SV7LfJiBImqg3QXs8dylMnLvxpahuXhwdfgRW4Olhxz94OMgqvvMVXX7ZgBZjJ1Vzm9-MtAmrSlwROtemu8xozzdktuDqX80QJMVUaFzhqobHAIYhEg=s0-d-e1-ft#https://docs.google.com/uc?export=download&id=1OTSjg23rcbgEmCiW2t9-SZ46s1fzrfdm&revid=0ByX1PVCIx8ZWblpFSUloV1IyNE9xQVhxd1FJMEgxbExRVlVvPQ] [https://ci3.googleusercontent.com/proxy/O6bmsFGDX0QZZKw1zhtzcYbQYmMA5rUagjWVMNB7zqtxZMPNPobqy0CYmRqnboKmvw94cTjQkja84ckWc4ka_A9RXW-V30LMervm_zM1i3uAPLbjlEbmzCuGjaLJdHSUMteQEQI7WgWIgMzSsjzVJdBh5b_pvmSAYk0o83Iig8WbJBz_1ZZvhT2DFUpl-mdXIsl4IATfDIb2SrpP2w=s0-d-e1-ft#https://docs.google.com/uc?export=download&id=1Zqakn0mYSuPPWKH5WSJ25iW1OfKlvEdT&revid=0ByX1PVCIx8ZWallERkRNYUVRRFdDTGVwbzhBRUF3WHJ0RmJFPQ] [https://ci4.googleusercontent.com/proxy/stDK5fjELD4yryy5oDWaSSeIHNrUZvuqSFetsYst7pLba9jh6slhnm9h-6tb-7kLioGZClY29tigS4NN9hmucCmrl6gV6O0dHCQkqXXJTK4r2SryV19zVsNdULJNG5cGh1FC7MyfJe75AFyiAEGzBkoUwd5pdfZiIIQZ1FBBzEwLlWCsr2BLnXjEZ_IN5izL3c8tLHrV4Xug1bac6Q=s0-d-e1-ft#https://docs.google.com/uc?export=download&id=17wl5KwV5wDPkux5M_aPnCzH0c9wizrKv&revid=0ByX1PVCIx8ZWY0xHbXpUUFU5N0Roc1dMWGYvMWpicUxRNHBJPQ] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 772905 bytes Desc: image.png URL: From ioannakoroni at csd.auth.gr Mon Jul 4 04:17:15 2022 From: ioannakoroni at csd.auth.gr (Ioanna Koroni) Date: Mon, 4 Jul 2022 11:17:15 +0300 Subject: Connectionists: =?iso-8859-1?q?Live_e-Lecture_by_Prof=2E_A=2E_del?= =?iso-8859-1?q?_Bimbo=3A_=22Social_interaction_in_trajectory_predi?= =?iso-8859-1?q?ction_with_Memory_Augmented_Networks=22=2C_5th_July?= =?iso-8859-1?q?_2022_17=3A00-18=3A00_CEST=2E_Upcoming_AIDA_AI_exce?= =?iso-8859-1?q?llence_lectures?= Message-ID: <169001d88f7e$78f9d5c0$6aed8140$@csd.auth.gr> Dear AI scientist/engineer/student/enthusiast, Prof. A. del Bimbo (Universit? di Firenze, Italy), a prominent AI & Digital Media researcher internationally, will deliver the e-lecture: ?Social interaction in trajectory prediction with Memory Augmented Networks?, on Tuesday 5th July 2022 17:00-18:00 CEST (8:00-9:00 am PST), (12:00 am-1:00am CST), see details in: http://www.i-aida.org/ai-lectures/ You can join for free using the zoom link: https://authgr.zoom.us/j/95605045574 & Passcode: 148148 The International AI Doctoral Academy (AIDA), a joint initiative of the European R&D projects AI4Media, ELISE , Humane AI Net , TAILOR , VISION , currently in the process of formation, is very pleased to offer you top quality scientific lectures on several current hot AI topics. Lectures will be offered alternatingly by: Top highly-cited senior AI scientists internationally or Young AI scientists with promise of excellence (AI sprint lectures) Other upcoming lecture: Assoc. Prof. Negar Kiyavash: ?Causal Inference in Complex Networks?, 11th July 2022 17:00 ? 18:00 CEST. More lecture infos in: https://www.i-aida.org/events/causal-inference-in-complex-networks These 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, J. Stefanowski -------------- next part -------------- An HTML attachment was scrubbed... URL: From tanya.brown at ae.mpg.de Mon Jul 4 06:33:13 2022 From: tanya.brown at ae.mpg.de (Brown, Tanya) Date: Mon, 4 Jul 2022 10:33:13 +0000 Subject: Connectionists: Mindvoyage | Dr RANDY MCINTOSH on July 11 (hybrid event) Message-ID: <8a8ee828cf5b46ce87c6a2c4a858a9d6@ae.mpg.de> [cid14046*image001.png at 01D8816C.151EF0F0] The Mindvoyage lecture series features prominent scholars from different disciplines including the humanities, biology, neuroscience and physics. Talks are dedicated to engaging in discussions related to novel, distinct and often controversial topics. The next talk, featuring Dr Randy McIntosh on Monday July 11 @ 10:00-11:30 CET This special edition talk will take place in-person at Max Planck Institute for Empirical Aesthetics (Artlab) If you are not able to join us in Frankfurt, take part in this hybrid event via zoom! All are welcome, but registration is required. CLICK HERE TO REGISTER [Inline image OWAPstImg964049] Abstract | Maybe we can access the brain?s hidden repertoire through music Music is culturally ubiquitous, supporting social and personal functions. Unlike language, music listening and performing seem to engage several brain networks. The broad engagement opens the possibility of identifying key personal brain signatures that reflect the capacity of brain systems to work together. This potential meshes well with the evolving theory of Structured Flow on Manifolds (SFM), where the manifolds define potentials and the flow represents actual expressions of network dynamics. My talk will lay the foundation for these ideas and the link to music listening. When we consider music as having similar SFM properties as the brain, a connection may be formed by linking the music and brain flows. I will present some preliminary data from EEG, where we start linking flows using Hidden Markov modelling. I will finish with ideas for an extension to ageing and dementia. REGISTRATION LINK https://www.aesthetics.mpg.de/en/the-institute/events/events/article/mindvoyage-03-randy-mcintosh.html Hosted by Lucia Melloni and Tanya Brown from the Max Planck Institute for Empirical Aesthetics, on behalf of the ARC-COGITATE Consortium [cid14046*image003.png at 01D8816C.151EF0F0] My working hours may not be yours, respond in your own time Tanya Brown Scientific Coordinator | ARC-Cogitate Max Planck Institute for Empirical Aesthetics Gr?neburgweg 14, 60322 Frankfurt am Main, Germany tanya.brown at ae.mpg.de [1622818538171] -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: OutlookEmoji-16228185381717b8a25a5-7efc-4c17-8da8-7dbfbb701a3e.png Type: image/png Size: 19053 bytes Desc: OutlookEmoji-16228185381717b8a25a5-7efc-4c17-8da8-7dbfbb701a3e.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: OutlookEmoji-cid14046*image001.png at 01D8816C.151EF0F0adf565c9-338f-450b-b10a-516c41c2102e.png Type: image/png Size: 222062 bytes Desc: OutlookEmoji-cid14046*image001.png at 01D8816C.151EF0F0adf565c9-338f-450b-b10a-516c41c2102e.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: OutlookEmoji-Inline image OWAPstImg9640499e14b9b4-fead-4367-828f-634e76ab78bb.png Type: image/png Size: 255428 bytes Desc: OutlookEmoji-Inline image OWAPstImg9640499e14b9b4-fead-4367-828f-634e76ab78bb.png URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: OutlookEmoji-cid14046*image003.png at 01D8816C.151EF0F0f79d9b97-195b-4654-a5d8-d54653693557.png Type: image/png Size: 8443 bytes Desc: OutlookEmoji-cid14046*image003.png at 01D8816C.151EF0F0f79d9b97-195b-4654-a5d8-d54653693557.png URL: From ksharma.raj at gmail.com Mon Jul 4 12:09:50 2022 From: ksharma.raj at gmail.com (Raj Sharma) Date: Mon, 4 Jul 2022 21:39:50 +0530 Subject: Connectionists: Deadline Extended: The 2nd International Conference on AI-ML-Systems (AIMLSystems'22) Message-ID: ============================================================== *AIMLSystems 2022* The 2nd International Conference on AI-ML-Systems (An Initiative of the COMSNETS Association) 12 - 15 October 2022 Bangalore, India https://www.aimlsystems.org/2022/ https://cmt3.research.microsoft.com/AIMLSystems2022/ In-Cooperation With: ACM, ACM SIGKDD, ACM SIGMOD, ACM SIGAI ============================================================== *ANNOUNCEMENT: Paper submission deadline is extended to July 12, 2022, 11:59 pm AET* AIMLSystems is a new conference targeting research at the intersection of AI/ML techniques and systems engineering. Through this conference we plan to bring out and highlight the natural connections between these two fields and their application to socio-economic systems. Specifically we explore how immense strides in AI/ML techniques are made possible through computational systems research (e.g., improvements in CPU/GPU architectures, data-intensive infrastructure, and communications ), how the use of AI/ML can help in the continuous and workload-driven design space exploration of computational systems (e.g., self-tuning databases, learning compiler optimisers, and learnable network systems ), and the use of AI/ML in the design of socio-economic systems such as public healthcare, and security. The goal is to bring together these diverse communities and elicit connections between them. Contributions are invited under Research, Industry & Applications, and Demonstration Tracks of the conference. Authors are encouraged to submit previously unpublished research at the intersection of computational / socio-economic systems and AI/ML. *------------------------Topics of Interest------------------------* The areas of interest are broadly categorized into the following three streams: ** Systems for AI/ML, including but not limited to:* - CPU/GPU architectures for AI/ML - Specialized/Embedded hardware for AI/ML workloads - Data intensive systems for efficient and distributed training - Challenges in production deployment of ML systems - ML programming models, languages, and abstractions, - ML compilers and runtime - Efficient systems for data preparation and processing - Systems for visualization of data, models, and predictions - Testing, debugging, and monitoring of ML applications - Cloud-computing for machine and deep learning - Machine and deep learning ?as-a-service? - Efficient model training, optimization and inference - Hardware efficient ML methods - Resource-constrained ML - Tiny Machine Learning - Embedded and Edge Artificial Intelligence - Distributed and parallel learning algorithms - MLOps (data collection, monitoring and re-training) ** AI/ML for Systems, including but not limited to:* - AI/ML for VLSI and architecture design - AI/ML in compiler optimization - AI/ML in data management - including database optimizations, virtualization, etc. - AI/ML for networks - design of networks, load modeling, etc. - AI/ML for power management - green computing, power models, etc. - AI/ML for Cloud Computing - AI/ML for IOT networks ** AI/ML for Socio-Economic Systems Design, which includes, but not limited to:* - Computational design and analysis of socio-economic systems - Fair and bias-free systems for social welfare, business platforms - Applications of AI/ML in the design, short-/long-term analysis of cyber-physical systems - Mechanism design for socio-economic systems - Fairness, interpretability and explainability for ML applications - Privacy and security in AI/ML systems - Sustainability in AI/ML systems - Ethics in AI/ML systems - Applications of AI/ML in financial systems -------------- *Key Dates* -------------- *Paper submissions due: July 12, 2022 (Firm Deadline)* Author notifications: August 30, 2022 Camera ready deadline: September 12, 2022 Conference dates: October 12-15, 2022 *---------Venue---------* The Chancery Pavilion, Residency Road, Bangalore, India | Hybrid Conference *----------------------------Submission Instruction----------------------------* Research papers must not exceed 8 pages, excluding appendix, acknowledgments and bibliography. Only electronic submissions in PDF format using the ACM sigconf template (see https://www.acm.org/publications/proceedings-template) will be considered. Papers can be submitted under any of the three main topics listed above. Authors are required to make a primary topic selection, with optional secondary topics for each paper. Number of papers accepted under each topic is not capped. We will accept all papers that meet the high quality and innovation levels required by the AIMLSystems conference. All papers that are accepted will appear in the proceedings. All accepted papers will be presented as posters at AIMLSystems 2022, but a select subset of them will be given a ?conventional? (oral) presentation slot during the conference. However, all accepted papers will be treated equally in the conference proceedings, which are the persistent, archival record of the conference. *----------------------------------Dual Submission Policy----------------------------------* A paper submitted to AIMLSystems can not be under review at any other conference or journal during the entire time it is considered for review at AIMLSystems, and it must be substantially different from any previously published work or any work under review. After submission and during the review period, submissions to AIMLSystems must not be submitted to other conferences/journals for consideration. However, authors may publish at non-archival venues, such as workshops without proceedings, or as technical reports (including arXiv). *---------Ethics---------* Plagiarism Policy: Submission of papers to AIMLSystems 2022 carries with it the implied agreement that the paper represents original work. We will follow the ACM Policy on Plagiarism, Misrepresentation, and Falsification ? see https://www.acm.org/publications/policies/plagiarism-overview. All submitted papers will be subjected to a ?similarity test?. Papers achieving a high similarity score will be examined and those that are deemed unacceptable will be rejected without a formal review. We also expect to report such unacceptable submissions to the superiors of each of the authors. Submission of papers to AIMLSystems 2022 also carries with it the implied agreement that one or more of the listed authors will register for and attend the conference and present the paper. Papers not presented at the conference will not be included in the final program or in the digital proceedings. Therefore, authors are strongly encouraged to plan accordingly before deciding to submit a paper. *------------------------------------------* *Keynote Speakers* *------------------------------------------* - Carlos Guestrin, Stanford University, USA - Thorsten Joachims, Cornell University, USA - Sunita Sarawagi, IIT Bombay, India - Partha Pratim Talukdar, IISc Bangalore & Google Research, India - Cynthia Rudin, Duke University, USA - Max Welling, University of Amsterdam and Microsoft Research, Netherlands *----------------------------------------------------------* *Conference Chairs & Contact Information* *---------------------------------------------------------* General Chairs: - Ralf Herbrich (Hasso Plattner Institute, Germany) - Rajeev Rastogi (Amazon, India) - Dan Roth (University of Pennsylvania, USA) TPC Chairs: - Sumohana Channappayya (IIT Hyderabad, India) - Srujana Merugu (Amazon, India) - Manuel Roveri (Politecnico di Milano, Italy) For general inquiries: aimlsys.conference at gmail.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From a.passarella at iit.cnr.it Tue Jul 5 07:00:01 2022 From: a.passarella at iit.cnr.it (Andrea Passarella) Date: Tue, 5 Jul 2022 13:00:01 +0200 (CEST) Subject: Connectionists: Scouting for PhDs/Post-docs at Ubiquitous Internet, IIT-CNR, Pisa, Italy - DEADLINE July 31st Message-ID: <20220705110001.511AD163008F@magneto.iit.cnr.it> The Ubiquitous Internet Research Unit of IIT-CNR (Pisa, Italy) is scouting for talented candidates for the following PhD and post-doc areas: * Human-centric explainable and efficient pervasive Artificial Intelligence [3 topics] * Architectures, Algorithms, and Protocols for the Quantum Internet [2 topics] * 6G networks and edge computing [4 topics] * Mobile health and digital phenotyping for personalized health services [1 topic] * ICT-enabled production optimization in Industry 4.0 [1 topic] Complete information about research areas, associated topics, reference contact points are provided at https://turig.iit.cnr.it/ui-positions/. Expressions of interest will be continuously considered upon reception. The very final date for sending EoIs is 31st July 2022, but topics may be closed earlier, depending on the received EoIs. Interested people are strongly encouraged to send expressions of interest as soon as possible. Details on the procedure are available here https://turig.iit.cnr.it/ui-positions/apply/. PhD and PostDocs will be carried out in the framework of one of the following European and National Projects. This will allow ample opportunities for international collaboration: * AI and BigData: H2020 HumanAI-Net, SoBigData++, CHIST-ERA SAI (Social Explainable AI) * Edge computing & decentralised AI: HE RE4DY, H2020 MARVEL, PON-MIUR OK-INSAID * Next-Generation Internet Infrastructures: ESFRI-SLICES, HE SLICES-PP, H2020 SLICES-DS, SLICES-SC * Quantum Computing & Networking: H2020 HPCQS, PON-MIUR QUANCOM Research Group: Ubiquitous Internet @ IIT-CNR http://ui.iit.cnr.it General contact: Andrea Passarella Specific contacts: see topics description at https://turig.iit.cnr.it/ui-positions/positions/ From benoit.frenay at unamur.be Tue Jul 5 07:50:54 2022 From: benoit.frenay at unamur.be (=?UTF-8?B?QmVub8OudCBGcsOpbmF5?=) Date: Tue, 5 Jul 2022 13:50:54 +0200 Subject: Connectionists: =?utf-8?q?Fwd=3A_=5Bannonces-AFIHM=5D_Postdoc_de_?= =?utf-8?q?2_ans_=C3=A0_Namur_=28Belgique=29_en_interaction_multimodale_ou?= =?utf-8?b?IHLDqWFsaXTDqSBhdWdtZW50w6ll?= In-Reply-To: <97FD8410-E7DB-4E3B-824A-DB6C4387F332@unamur.be> References: <97FD8410-E7DB-4E3B-824A-DB6C4387F332@unamur.be> Message-ID: <6e15e464-ecc4-f950-0dee-51654a09eeff@unamur.be> Bonjour ? tous, Dans le cadre d?un projet, je suis ? la recherche d?un chercheur de niveau post-doctoral qui ait une expertise en interaction multimodale, ou en interaction en r?alit? augment?e. Le but du projet de recherche (une collaboration avec plusieurs entreprises wallonnes et fran?aises) est la conception d?un syst?me de pilotage de v?hicule int?gr? ? un casque ? affichage augment?, utilisant parole, regard, niveau d?attention de l?utilisateur, et possiblement d?autres modalit?s. Au niveau des probl?matiques de recherche identifi?es sur lesquelles Namur travaillera, au-del? de la d?finition d?un dialogue multimodal naturel et efficace, ainsi que de certains aspects logiciels, il y aura des questions li?es ? l?adaptation de l?interface multimodale. Quelques petites remarques : le poste est un temps plein sur 2 ans, avec un salaire comp?titif. L?intitul? exact du poste est ??charg? de recherche avec th?se??. Le poste d?marrerait dans l?id?al au 1er septembre, mais la date de d?marrage effective peut ?tre un peu retard?e. Les candidatures sont ? remettre *d?ici au 15 juillet*. Plus d'infos + candidature sur la page suivante : https://jobs.unamur.be/emploi.2022-05-18.6143812019 N?h?sitez pas ? revenir vers moi pour toute question, et ? faire tourner l?annonce si vous pensez ? quelqu?un ?! Bonne journ?e ? tous, Bruno Dumas * * *Bruno DUMAS* Professeur en Facult? d'informatique Namur Digital Institute (NaDI)?- /EXtended User Interactions (EXUI)/?team leader *T.*?+32 (0)81 724 975 *F.*?+32 (0)81 724 967 bruno.dumas at unamur.be http://directory.unamur.be/staff/bdumas http://be.linkedin.com/in/brunodumas *Universit? de Namur ASBL* Rue de Bruxelles 61 - 5000 Namur -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: egbaedbg.png Type: image/png Size: 9356 bytes Desc: not available URL: From ioannakoroni at csd.auth.gr Wed Jul 6 02:35:45 2022 From: ioannakoroni at csd.auth.gr (Ioanna Koroni) Date: Wed, 6 Jul 2022 09:35:45 +0300 Subject: Connectionists: Live ELLIS/AIDA e-Lecture by Assoc. Prof. Negar Kiyavash: "Causal Inference in Complex Networks", 11 July 2022 17:00-18:00 CEST Message-ID: <1b7901d89102$9ff81f30$dfe85d90$@csd.auth.gr> Dear AI scientist/engineer/student/enthusiast, Assoc. Prof. Negar Kiyavash (University of Illinois at Urbana-Champaign, USA), a prominent AI researcher internationally, will deliver the e-lecture: 'Causal Inference in Complex Networks', on Monday 11th July 2022 17:00-18:00 CEST (8:00-9:00 am PST), (12:00 am-1:00am CST), see details in: http://www.i-aida.org/ai-lectures/ You can join for free using the zoom link: https://authgr.zoom.us/j/96116246838 & Passcode: 148148 Attendance is free. This is the closing keynote lecture for the ELLIS PhD and Postdoc Summit. The International AI Doctoral Academy (AIDA), a joint initiative of the European R&D projects AI4Media, ELISE , Humane AI Net , TAILOR , VISION , currently in the process of formation, is very pleased to offer you top quality scientific lectures on several current hot AI topics. Lectures will be offered alternatingly by: Top highly-cited senior AI scientists internationally or Young AI scientists with promise of excellence (AI sprint lectures) This is the last lecture from a series of successful lectures by AIDA and the next round of lectures will start in September 2022. These 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, J. Stefanowski -------------- next part -------------- An HTML attachment was scrubbed... URL: From timofte.radu at gmail.com Wed Jul 6 11:21:19 2022 From: timofte.radu at gmail.com (Radu Timofte) Date: Wed, 6 Jul 2022 17:21:19 +0200 Subject: Connectionists: Open Positions for Doctoral and PostDoctoral Researchers in AI, Computer Vision and Machine Learning Message-ID: *PostDoctoral and Doctoral Researcher Open Positions in * *Artificial Intelligence, Computer Vision, and Machine Learning* (Apologies for cross-postings.) Computer Vision Laboratory led by *Prof.Dr. Radu Timofte *, from the newly established *Center for Artificial Intelligence and Data Science , University of Wurzburg*, is looking for outstanding candidates to fill several postdoctoral and doctoral researcher fully-funded positions in the AI, computer vision, and machine learning fields. *Julius Maximilians University of W?rzburg (JMU), *founded in 1402, is one of the leading institutions of higher education in Germany and well-known on the international stage for delivering research excellence with a global impact. The University of W?rzburg is proud to be the home of outstanding researchers and fourteen Nobel Prize Laureates. W?rzburg is a vibrant city in Bavaria, Germany?s economically strongest state and home base to many international companies. We look forward to welcoming you to the University of W?rzburg! *Computer Vision Laboratory* and University of W?rzburg in general are an exciting environment for research, for independent thinking. Prof. Radu Timofte?s team is highly international, with people from about 12 countries, and the members have already won awards at top conferences (ICCV, CVPR, ICRA, NeurIPS, ...), founded successful spinoffs, and/or collaborated with industry. Prof. Radu Timofte is *a 2022 winner of the prestigious Humboldt Professorship for Artificial Intelligence Award.* Prof. Radu Timofte also leads the *Augmented Perception Group* at ETH Zurich. Depending on the position, the successful candidate will focus on a subset of the following *Research Topics:* ? deep learning ? computational photography ? domain translation ? learned image/video compression ? image/video super-resolution ? learning paradigms ? 3D ? image/video understanding ? augmented and mixed reality ? edge inference and mobile AI ? super-resolution microscopy *The tasks* will involve designing, developing, and testing novel ideas and solutions in cutting-edge research, as well as coordinating and conducting data collection for their evaluation when necessary. The successful candidate will conduct research on deep learning machines and a new cluster with hundreds of GPUs. *Profile* ? Master's degree in AI, computer science, electrical engineering, physics or applied mathematics/ statistics. ? Good programming skills, experience with Python / C++ and deep learning frameworks (PyTorch/TensorFlow). ? Interest, prior knowledge and experience in one or more of the following is a plus: computer vision, deep learning, machine learning, image processing, artificial intelligence. ? Enthusiasm for leading-edge research, team spirit, and capability of independent problem-solving. ? Fluent written and spoken English is a must. ? Postdoctoral applicants are expected to have a strong track of published research, including top, high impact, journal (such as PAMI, IJCV, TIP, NEUCOM, JMLR, CVIU) or conference (such as ICCV, CVPR, ECCV, ICRA, NeurIPS, ICLR, AAAI) papers. *Timeline* The positions are open immediately, fully funded, the salaries of the doctoral students and postdocs are competitive on the German scales TV-L E13 and E14, up to 70k euros per year, before tax. Typically a PhD takes ~4 years to complete and a postdoc position is for at least 1 year. The applications received by 15.07.2022 will be reviewed by 31.07.2022. Only the selected applicants will be contacted by email for interviews. After 15.07.2022 the applications will be reviewed on a rolling basis until all positions are filled. *Application* Interested applicants should email asap their PDF documents (including full CV, motivation letter, diplomas, transcripts of records, links to master or PhD thesis, referees / recommendation letters, etc.) to Prof. Dr. Radu Timofte at *radu.timofte at uni-wuerzburg.de* or *radu.timofte at vision.ee.ethz.ch* -------------- next part -------------- An HTML attachment was scrubbed... URL: From sayan.mukherjee at mis.mpg.de Wed Jul 6 08:11:47 2022 From: sayan.mukherjee at mis.mpg.de (sayan.mukherjee at mis.mpg.de) Date: Wed, 6 Jul 2022 14:11:47 +0200 (CEST) Subject: Connectionists: postdoc positions at University of Leipzig and the MPI MiS Message-ID: <596471391.47859.1657109507208.JavaMail.zimbra@mis.mpg.de> All: There are four PostDoc Positions available in the Statistical Learning Department headed by Sayan Mukherjee, Alexander von Humboldt Professor in AI. The Department is joint between the Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) at the University of Leipzig and the Max Planck Institute for Mathematics in the Sciences. The English version of the add is at https://www.uni-leipzig.de/en/university/working-at-leipzig-university/job-opportunities/detailed-view-job-description/artikel/4-wissenschaftliche-mitarbeiter-postdoc-im-bereich-probabilistisches-maschinelles-lernen-und-computergestuetzte-biologie-m-w-d-2022-05-30 For more information on the research of Dr. Mukherjee see https://sayanmuk.github.io/ Please contact me if you are interested. cheers, sayan -- Sayan Mukherjee Alexander von Humboldt Professor Universit?t Leipzig Center for Scalable Data Analytics and Artificial Intelligence Max Planck Institute for Mathematics in the Sciences he/him/his From axel.hutt at inria.fr Wed Jul 6 05:46:22 2022 From: axel.hutt at inria.fr (Axel Hutt) Date: Wed, 6 Jul 2022 11:46:22 +0200 (CEST) Subject: Connectionists: Announcement of Tutorial at CNS 2022 in Melbourne Message-ID: <430633324.147215484.1657100782604.JavaMail.zimbra@inria.fr> **************************************** CNS 2022 in Melbourne, Australia **************************************** Tutorial on Saturday, July 16, 1:30pm AEST on "Spectral Analysis of Neural Signals" The spectral analysis of observed neural activity is essential in a large part of experimental research. To apply successfully the meanwhile large number and different types of spectral analysis techniques, it is important to understand in detail fundamental aspects of spectral analysis methods. The tutorial is targeted at experimentalists at all levels and will just touch theoretical details. It will be a hands-on tutorial based on practical problems. As an additional support, Python source code scripts will be provided for several analysis problems discussed in the tutorial. These scripts permit the participant to implement herself/himself different techniques discussed and support further understanding. Content: ** Fundamentals in sampling theory, Fourier theory and related artifact (aliasing, spectral leakage) ** Linear filters and spectral power in stationary signals ** Spectral power in non-stationary signals: windowed Fourier transform and time-frequency spectral analysis. ** The concept of analytical signal: Hilbert Transform, phase synchronization, Empirical Mode Decomposition More information at https://www.cnsorg.org/cns-2022-tutorials#T7 --------------------------------- Axel Hutt Directeur de Recherche Equipe MIMESIS INRIA Nancy Grand Est B?timent IHU 1, Place de l'Hopital 67000 Strasbourg, France https://mimesis.inria.fr/speaker/axel-hutt/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From perusquia at ieee.org Wed Jul 6 07:17:11 2022 From: perusquia at ieee.org (Monica Perusquia Hernandez) Date: Wed, 6 Jul 2022 20:17:11 +0900 Subject: Connectionists: [CfP][deadline 22Jul] 3rd Momentary Emotion Elicitation & Capture (MEEC) at ACII 2022 Message-ID: ======================================================= 3rd Momentary Emotion Elicitation & Capture (MEEC) at ACII 2022 Call for Papers 17 Oct 2022, Hybrid event, Nara, Japan https://cwi-dis.github.io/meec-ws/index.html ======================================================= To train machines to detect and recognise human emotions sensibly, we need valid emotion ground truths that consider dynamic changes over time. A fundamental challenge here is the momentary emotion elicitation and capture (MEEC) from groups and individuals continuously and in real-time, without adversely affecting user experience. In this half-day virtual ACII 2022 workshop, we will (1) have a keynote presentation about ambulatory sampling methods by Prof. Ulrich Ebner-Priemer from the Karlsruhe Institute of Technology; (2) have participant talks showcasing their submissions; (3) brainstorm on techniques to understand dynamic changes given different temporal measurement resolutions; and (4) create a battery of methods relevant to diverse affective contexts. We seek contributions across disciplines that explore how emotions can be naturally elicited and captured in the moment. Topics include: 1. Elicitation: 1. multi-modal (e.g., film, music) and multi-sensory (e.g., smell, taste, thermal) elicitation 2. emotion elicitation across domains (e.g., automotive, healthcare) 3. elicitation and immersiveness (e.g., AR/VR/MR) 4. elicitation over time (e.g., mood) 5. elicitation through human-robot interaction 6. ethical considerations 2. Capture: 1. emotion models (dimensional, discrete) 2. annotation modalities (e.g., speech) and (remote) tools (e.g., ESMs) 3. devices (e.g., mobile, wearable) and sensors (e.g., RGB / thermal cameras, EEG, eye-tracking) 4. attention considerations (e.g., interruptions) 5. ethical issues in tracking and detection We invite position papers, research results papers, and demos (2-9 pages, including references) that describe/showcase emotion elicitation and/or capture methods. Submissions will be peer-reviewed by two peers and selected on their potential to spark discussion. Submissions should be prepared according to the IEEE conference template and submitted in PDF through Easychair ( https://easychair.org/conferences/?conf=meec2022). The templates can be found here at this link: LaTeX/Word Templates . Accepted submissions will be made available on the workshop proceedings of ACII2022. They will be published and indexed on IEEE Xplore. At least one author must register for the workshop and one day of the conference. Submission Deadline (EXTENDED): 22 July 2022 23:59 AoE Notification of Acceptance: 5 August 2022 Camera-ready Deadline: 12 August 2022 Workshop Day: 17 October 2022 -------------- next part -------------- An HTML attachment was scrubbed... URL: From sigurd.lokse at uit.no Thu Jul 7 08:27:37 2022 From: sigurd.lokse at uit.no (=?Windows-1252?Q?Sigurd_Eivindson_L=F8kse?=) Date: Thu, 7 Jul 2022 12:27:37 +0000 Subject: Connectionists: =?windows-1252?q?2nd_Call_for_Contributions=3A_6t?= =?windows-1252?q?h_Northern_Lights_Deep_Learning_Conference=2C_10-12_Janu?= =?windows-1252?q?ary_2023=2C_Troms=F8_=28=93North_Pole=94=29=2C_Norway?= Message-ID: Please join for the 6th Northern Lights Deep Learning Conference (NLDL) on 10-12 January 2023 in Troms?, Norway, organized by Visual Intelligence and the UiT Machine Learning Group. We look forward to gathering the deep learning community again in the cool arctic air for a physical conference, after two years online. In addition, the NLDL winter school, which is a part of the NORA research school http://nora.ai, starts at Jan 9 and ends at Jan 13 and incorporates events during the main conference days. The winter school includes scientific topics, industry event, women in AI event, and transferrable skills. More information coming soon at http://www.nldl.org/winter-school. We invite submissions presenting new and original research on all aspects of Deep Learning. The topics include but are not limited to the following: * Architecture, concepts and optimization * Deep learning for structured and unstructured data * Graph neural networks * Generative models * Bayesian Deep Learning * Lightweight / frugal Deep Learning * Explainability and interpretability of Deep Learning models * Computer vision * Natural language processing * Deep Learning for signals, images, 3D and hyperspectral images * Deep Learning applications to biology and medicine * Deep Learning application to environment and ecology * Deep Learning applications to Physics * Deep Learning for industrial applications As always, we are happy to have top international speakers. This year, for instance * Mark Girolami ? University of Cambridge/Alan Turing Institute * Mihaela van der Shaar ? University of Cambridge/Alan Turing Institute * Polina Golland ? MIT * Christian Igel ? University of Copenhagen We are accepting two alternatives for contributions: (1) Full paper submissions (6 pages) will be presented either as orals or as posters and will be published in the conference proceedings. The proceedings are approved as a level 1 publication in the Norwegian national list of authorized research publication channels; (2) Extended abstracts (2 pages) will be presented either as orals or as posters (but not published in the conference proceedings). The review process is double-blind. Deadline for both types of submissions: September 16th, 2022. Instructions on template etc. can be found on http://www.nldl.org. A tentative program will be available soon at http://www.nldl.org/ and will include keynotes, scientific talks, an industry event, a Women in AI event and social events. We hope to see many participants for a nice scientific gathering on the ?north pole?, including social events, and hopefully some northern lights Kind regards, The NLDL 2023 organizing committee http://visual-intelligence.no http://machine-learning.uit.no -------------- next part -------------- An HTML attachment was scrubbed... URL: From q.huys at ucl.ac.uk Thu Jul 7 17:42:52 2022 From: q.huys at ucl.ac.uk (Quentin Huys) Date: Thu, 7 Jul 2022 22:42:52 +0100 Subject: Connectionists: postdoc position in computational psychiatry at UCL Message-ID: <20220707214252.6d23w3khace5elvm@Qh> Postdoctoral position available at the UCL Applied Computational Psychiatry Lab (www.acplab.org, PI Quentin Huys). The Wellcome Trust-funded project will combine cognitive probes with computational modelling and MEG imaging to better understand the algorithmic structure of maladaptive thinking patterns in depression, and how they relate to antidepressants, relapse and serotonin. These projects are an exciting possibility to apply advanced computational and neuroimaging methods to clinically relevant problems. The Applied Computational Psychiatry lab is situated within the UCL Max Planck Centre for Computational Psychiatry and Ageing Research and the Division of Psychiatry. The official job advertisement is here: https://atsv7.wcn.co.uk/search_engine/jobs.cgi?SID=amNvZGU9MTg4NTU5NSZ2dF90ZW1wbGF0ZT05NjUmb3duZXI9NTA0MTE3OCZvd25lcnR5cGU9ZmFpciZicmFuZF9pZD0wJmpvYl9yZWZfY29kZT0xODg1NTk1JnBvc3RpbmdfY29kZT0yMjQ= From hussain.doctor at gmail.com Fri Jul 8 15:36:26 2022 From: hussain.doctor at gmail.com (Amir Hussain) Date: Fri, 8 Jul 2022 20:36:26 +0100 Subject: Connectionists: Postdoctoral Research fellow vacancy (10 July deadline) & invite to 1st AVSEC Challenge (part of IEEE SLT 2022) Message-ID: Dear all: Please see *two* key updates below on behalf of the UK EPSRC funded COG-MHEAR research programme (http://cogmhear.org), and kindly help circulate as appropriate - thank you very much in advance. *(1) *COG-MHEAR is recruiting for a *postdoctoral research fellow* based at Edinburgh Napier University - please see application details below. COG-MHEAR Research Fellow post (for up to 30 months - closing date: 10th July 2022): https://www.jobs.ac.uk/job/CRB059/research-fellow-cog-mhear-full-time-up-to-30-months-fixed-term *(2) * Invite to the *first COG-MHEAR Audio-visual Speech Enhancement Challenge (AVSEC)* - http://challenge.cogmhear.org Participants will work on a large dataset derived from TED talks to enhance speech in extremely challenging noisy environments and with competing speakers. The performance will be evaluated using human listening tests as well as with objective measures. We hope that the Challenge will create a benchmark for AVSEC research that will be useful for years to come. The challenge data and development tools are now available - for details see the challenge website: https://challenge.cogmhear.org/#/ and our github repository: https://github.com/cogmhear/avse_challenge AVSEC has been accepted as an official challenge at the *IEEE Spoken Language Technology (SLT) Workshop* (https://slt2022.org/) to be held in Doha, Qatar, 9-12 Jan 2023, where a special session will be run. *Important Dates* 1st May 2022: Challenge website launch 31st May 2022: Release of the full toolset, training/development data and baseline system *1st June 2022: Registration for challenge entrants opens* 25th July 2022: Evaluation data released 1st Sept 2022: Submission deadline for evaluation (by objective and subjective measures) 9th Jan 2023: Results announced at IEEE SLT 2022 *Background: *Human performance in everyday noisy situations is known to be dependent upon both aural and visual senses that are contextually combined by the brain?s multi-level integration strategies. The multimodal nature of speech is well established, with listeners known to unconsciously lip-read to improve the intelligibility of speech in a real noisy environment. It has been shown that the visual aspect of speech has a potentially strong impact on the ability of humans to focus their auditory attention on a particular stimulus. The aim of the first AVSEC is to bring together the wider computer vision, hearing and speech research communities to explore novel approaches to multimodal speech-in-noise processing. Both raw and pre-processed AV datasets ? derived from TED talk videos ? will be made available to participants for training and development of audio-visual models to perform speech enhancement and speaker separation at SNR levels that will be significantly more challenging than those typically used in audio-only scenarios. Baseline neural network models and a training recipe will be provided. In addition to participation at IEEE SLT, Challenge participants will be invited to contribute to a *Journal Special Issue* on the topic of Audio-Visual Speech Enhancement that will be announced later this year. *Registration/further information*: If you are interested in participating and wish to receive further information, please sign up here: https://challenge.cogmhear.org/#/getting-started/register If you have questions, contact us directly at: cogmhear at napier.ac.uk *Organising Team*: Amir Hussain, Edinburgh Napier University, UK (co-Chair) Peter Bell, University of Edinburgh, UK (co-Chair) Mandar Gogate, Edinburgh Napier University, UK Cassia Valentini Botinhao, University of Edinburgh, UK Kia Dashtipour, Edinburgh Napier University, UK Lorena Aldana, University of Edinburgh, UK Evaluation Panel Chair: John Hansen, University of Texas in Dallas, USA Scientific Committee Chair: Michael Akeroyd, University of Nottingham, UK Industry co-ordinator: Peter Derleth, Sonova AG Funded by the UK Engineering and Physical Sciences Research Council (EPSRC) programme grant: COG-MHEAR (http://cogmhear.org ) Supported by RNID (formerly Action on Hearing Loss), Deaf Scotland, Sonova AG We hope to see you soon. Kindest regards Amir --- Professor Amir Hussain Programme Director: EPSRC COG-MHEAR (http://cogmhear.org) School of Computing, Engineering & Built Environment, Edinburgh Napier University, Edinburgh EH10 5DT, Scotland, UK E-mail: a.hussain at napier.ac.uk -------------- next part -------------- An HTML attachment was scrubbed... URL: From triesch at fias.uni-frankfurt.de Fri Jul 8 15:45:12 2022 From: triesch at fias.uni-frankfurt.de (Jochen Triesch) Date: Fri, 8 Jul 2022 21:45:12 +0200 Subject: Connectionists: Open PhD student position in the area of plasticity and learning in recurrent spiking neural networks Message-ID: <0DDF19C9-A08A-4BB0-8B14-0AFBA55F22D1@fias.uni-frankfurt.de> We solicit applications for a PhD student position in my lab (http://fias.uni-frankfurt.de/~triesch/ ) at the Frankfurt Institute for Advanced Studies (FIAS) to study plasticity and learning in recurrent spiking neural networks. The PhD project will address how learning induces connectome changes that implement the formation of a synaptic memory trace in cortical networks and how such a distributed memory trace becomes robust against alterations of individual synaptic connections in the presence of constant synaptic turnover. In particular, we will develop computational models of the dynamics of the excitatory and inhibitory connectome in the mouse cortex during learning to explain learning dynamics in terms of fundamental underlying plasticity mechanisms and characterize the conditions for the formation of lasting memories. The research will be performed in close collaboration with the labs of Simon Rumpel (University of Mainz, experimental collaborator) and Matthias Kaschube (FIAS, Frankfurt, theoretical collaborator) and J?rgen Jost (MPI for Mathematics in the Sciences, Leipzig, theoretical collaborator). The project is embedded in Germany?s priority program ?Computational Connectomics? (https://spp2041.de). We are seeking an outstanding and highly motivated PhD student for this project. Applicants should have obtained a Master Degree in Computational Neuroscience or a related field (Physics, Computer Science, Mathematics, Engineering, etc.). The ideal candidate will have excellent programming and analytic skills, experience with spiking neural network simulations, and a broad knowledge of Computational Neuroscience and neural plasticity. A strong interest in collaborating with experimental labs is a plus. The Frankfurt Institute for Advanced Studies (https://fias.institute/en/) is a research institution dedicated to fundamental theoretical research in various areas of science. The city of Frankfurt is the hub of one of the most vibrant metropolitan areas in Europe. It boasts a rich culture and arts community and repeatedly earns high rankings in worldwide surveys of quality of living. Most recently, Frankfurt achieved 7th place worldwide in a ranking by the Economist. Funding is available initially for three years. Renumeration is according to the German E13 pay scale at 65% full time equivalent (FTE). Applications should consist of a single pdf file. Please include a brief statement of research interests, CV, and contact information for at least two references. The position can be filled immediately. Applications will reviewed on a continuing basis. Upload your document using the application platform at: https://pm.fias.science/projects/application For applicants participating in next week's FENS meeting in Paris: I?ll also be at FENS. Please get in touch via email to schedule an informal in person meeting. Regards, Jochen Triesch -- Prof. Dr. Jochen Triesch Johanna Quandt Chair for Theoretical Life Sciences Frankfurt Institute for Advanced Studies and Goethe University Frankfurt http://fias.uni-frankfurt.de/~triesch/ Tel: +49 (0)69 798-47531 Fax: +49 (0)69 798-47611 From marius.pedersen at ntnu.no Sat Jul 9 03:36:38 2022 From: marius.pedersen at ntnu.no (Marius Pedersen) Date: Sat, 9 Jul 2022 07:36:38 +0000 Subject: Connectionists: PhD Candidate in Deep Learning for Capsule Video Endoscopy Message-ID: A three-year fully-funded PhD position in Deep Learning for Capsule Video Endoscopy is available at the Department of Computer Science, NTNU, in Gj?vik, Norway. Digestive system diseases like Crohn's disease, inflammatory bowel disease, and cancer, are affecting a large population across the world. Wireless capsule endoscope is a good alternative for screening colorectal cancer and other digestive diseases which is both relatively pain free and eliminates the fear of traditional colonoscopy with foreign object insertion into body. The position will focus on machine learning techniques for improving diagnosis cancer and other digestive diseases in wireless capsule endoscopy. The PhD position is part of the project "CapsNetwork - International Network for Capsule Imaging in Endoscopy" funded by the Research Council of Norway. For a position as a PhD Candidate, the goal is a completed doctoral education up to an obtained doctoral degree. Application deadline: 01.08.2022 Application portal: https://www.jobbnorge.no/en/available-jobs/job/229495/phd-candidate-in-deep-learning-for-capsule-video-endoscopy -------------- next part -------------- An HTML attachment was scrubbed... URL: From david at irdta.eu Sat Jul 9 05:23:27 2022 From: david at irdta.eu (David Silva - IRDTA) Date: Sat, 9 Jul 2022 11:23:27 +0200 (CEST) Subject: Connectionists: DeepLearn 2022 Autumn: early registration July 16 Message-ID: <1443714799.591996.1657358607526@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: July 16, 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 23 four-hour and a half courses and 2 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: Tommaso Dorigo (Italian National Institute for Nuclear Physics), Deep-Learning-Optimized Design of Experiments: Challenges and Opportunities Elaine O. Nsoesie (Boston University), AI and Health Equity PROFESSORS AND COURSES: 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 Nadya Chernyavskaya (European Organization for Nuclear Research), [intermediate] Graph Networks for Scientific Applications with Examples from Particle Physics 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 Efstratios Gavves (University of Amsterdam), [advanced] Advanced Deep Learning 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 Tin Kam Ho (IBM Thomas J. Watson Research Center), [introductory/intermediate] Deep Learning Applications in Natural Language Understanding Timothy Hospedales (University of Edinburgh), [intermediate/advanced] Deep Meta-Learning 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 Yao Wang (New York University), [introductory/intermediate] Deep Learning for Computer Vision 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: Nosheen Abid (Lule?) Sana Sabah Al-Azzawi (Lule?) Lama Alkhaled (Lule?) Prakash Chandra Chhipa (Lule?) Saleha Javed (Lule?) Marcus Liwicki (Lule?, local chair) Carlos Mart?n-Vide (Tarragona, program chair) Hamam Mokayed (Lule?) Sara Morales (Brussels) Mia Oldenburg (Lule?) Maryam Pahlavan (Lule?) David Silva (London, organization 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 are available 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 boubchir at ai.univ-paris8.fr Sun Jul 10 06:32:43 2022 From: boubchir at ai.univ-paris8.fr (Larbi Boubchir) Date: Sun, 10 Jul 2022 12:32:43 +0200 Subject: Connectionists: [CfP] The 3rd international workshop on Machine Learning for EEG Signal Processing (MLESP) In-Reply-To: <8f51ad5e-b276-8a29-df0f-b2da8e672395@ai.univ-paris8.fr> References: <8f51ad5e-b276-8a29-df0f-b2da8e672395@ai.univ-paris8.fr> Message-ID: <71c6d4bc-d6ba-a308-63ed-b5c7e205e14e@ai.univ-paris8.fr> [Apologies for multiple postings] ** *CALL FOR PAPERS* The 3^rd international workshop on Machine Learning for EEG Signal Processing (MLESP 2022, https://mlesp2022.sciencesconf.org/) will be held in Las Vegas, USA, from 6 to 9 december 2022, in conjunction with the IEEE International Conference on Bioinformatics and Biomedicine (IEEE BIBM 2022, https://ieeebibm.org/BIBM2022/) *Overview* EEG signal processing involves the analysis and treatment of the electrical activity of the brain measured with Electroencephalography, or EEG, in order to provide useful information on which decisions can be made. The recent advances in signal processing and machine learning for EEG data processing have brought an impressive progress to solve several practical and challenging problems in many areas such as healthcare, biomedicine, biomedical engineering, BCI and biometrics. The aim of this workshop is to present and discuss the recent advances in machine learning for EEG signal analysis and processing. We are inviting original research work, as well as significant work-in-progress, covering novel theories, innovative methods, and meaningful applications that can potentially lead to significant advances in EEG data analytics. This workshop is an opportunity to bring together academic and industrial scientists to discuss the recent advances. The topics of interest include but not limited to: - EEG signal processing and analysis - Time-frequency EEG signal analysis - Signal processing for EEG Data - EEG feature extraction and selection - Machine learning for EEG signal processing - EEG classification and Hierarchical clustering - EEG abnormalities detection (e.g. Epileptic seizure, Alzheimer's disease, etc.) - Machine learning in EEG Big Data - Deep Learning for EEG Big Data - Neural Rehabilitation Engineering - Brain-Computer Interface - Neurofeedback - EEG-based Biometrics - Related applications Important Dates *Aug. 31, 2022* (11:59 pm CST): Due date for full workshop papers submission Oct. 31, 2022: Notification of paper acceptance to authors Nov. 12, 2022: Camera-ready of accepted papers Dec 6-9, 2022: Workshops Paper Submission - Please submit a full-length paper (up to 8 page IEEE 2-column format) through the online submission system. You can download the format instruction here: https://www.ieee.org/conferences/publishing/templates.html - Electronic submissions in PDF format are required. Online Submission https://wi-lab.com/cyberchair/2022/bibm22/scripts/submit.php?subarea=S02&undisplay_detail=1&wh=/cyberchair/2022/bibm22/scripts/ws_submit.php Publication All accepted papers will be published in the BIBM proceedings and IEEE Xplore Digital Library. *Contact* Prof. Larbi Boubchir /(//Workshop Chair/),University of Paris 8, France E-mail: larbi.boubchir at univ-paris8.fr -------------- next part -------------- An HTML attachment was scrubbed... URL: From massimiliano.pontil at gmail.com Sun Jul 10 16:18:52 2022 From: massimiliano.pontil at gmail.com (massimiliano.pontil at gmail.com) Date: Sun, 10 Jul 2022 22:18:52 +0200 Subject: Connectionists: Job Openings in Machine Learning @IIT in Genoa Message-ID: We have two Postdoc and Researcher positions at Istituto Italiano di Tecnologia. Outstanding candidates will be considered in all areas of Machine Learning with a preference to: - Statistical learning theory - High dimensional statistics - Online learning and bandits algorithms - Stochastic and numerical optimization. The researchers will work at IIT within the Computational Statistics and Machine Learning group ( https://www.iit.it/it/web/computational-statistics-and-machine-learning) led by Prof. Massimiliano Pontil, and also be part of the ELLIS Unit Genoa (ellisgenoa.eu). The group's research is in the areas of machine learning theory and algorithms, with a focus on multitask and meta-learning, online and interactive learning, kernel methods and statistical learning theory, dynamical systems and time series analysis. We are also interested in techniques from mathematical statistics, numerical linear algebra and optimization. These positions are part of a broader plan at IIT to recruit scientists in Artificial Intelligence, Machine Learning and related application areas of science and engineering. *Deadline: August 15, 2022. * *For more information and instruction about how to apply see:* https://iit.taleo.net/careersection/ex/jobdetail.ftl?lang=en&job=22000052 https://iit.taleo.net/careersection/ex/jobdetail.ftl?lang=en&job=22000053 ESSENTIAL REQUIREMENTS - A PhD in Mathematics, Computer Science or related disciplines. - Documented experience on Machine Learning Theory and Algorithms - A strong track record of research publications in top tier conferences and journals (COLT, ICML, NIPS, Annals of Statistics, JMLR, etc.) is a plus. - Strong problem-solving attitude - The ability to properly report, organize and publish your research results. - Good command in spoken and written English. ADDITIONAL SKILLS - Experience in coaching PhD students and master students. - Experience on writing research proposal - Good knowledge of software tools (Python, TensorFlow, etc.) - Spirit of innovation and creativity - Good in time and priority management - Ability to work independently and collaboratively in a challenging and international environment COMPENSATION PACKAGE - Competitive salary package for international standards - Private health care coverage (depending on your role and contract) - Wide range of staff discounts -------------- next part -------------- An HTML attachment was scrubbed... URL: From ioannakoroni at csd.auth.gr Mon Jul 11 02:25:17 2022 From: ioannakoroni at csd.auth.gr (Ioanna Koroni) Date: Mon, 11 Jul 2022 09:25:17 +0300 Subject: Connectionists: Live ELLIS/AIDA e-Lecture by Assoc. Prof. Negar Kiyavash: "Causal Inference in Complex Networks", 11 July 2022 17:00-18:00 CEST Message-ID: <235201d894ee$fdc8eb40$f95ac1c0$@csd.auth.gr> Dear AI scientist/engineer/student/enthusiast, Assoc. Prof. Negar Kiyavash (University of Illinois at Urbana-Champaign, USA), a prominent AI researcher internationally, will deliver the e-lecture: 'Causal Inference in Complex Networks', on Monday 11th July 2022 17:00-18:00 CEST (8:00-9:00 am PST), (12:00 am-1:00am CST), see details in: http://www.i-aida.org/ai-lectures/ You can join for free using the zoom link: https://authgr.zoom.us/j/96116246838 & Passcode: 148148 Attendance is free. This is the closing keynote lecture for the ELLIS PhD and Postdoc Summit. The International AI Doctoral Academy (AIDA), a joint initiative of the European R&D projects AI4Media, ELISE , Humane AI Net , TAILOR , VISION , currently in the process of formation, is very pleased to offer you top quality scientific lectures on several current hot AI topics. Lectures will be offered alternatingly by: Top highly-cited senior AI scientists internationally or Young AI scientists with promise of excellence (AI sprint lectures) This is the last lecture from a series of successful lectures by AIDA and the next round of lectures will start in September 2022. These 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, J. Stefanowski -------------- next part -------------- An HTML attachment was scrubbed... URL: From mostafa.sadeghi at inria.fr Mon Jul 11 05:54:51 2022 From: mostafa.sadeghi at inria.fr (Mostafa Sadeghi) Date: Mon, 11 Jul 2022 11:54:51 +0200 (CEST) Subject: Connectionists: [Research engineer/post-doc offer] Robust and generalizable audiovisual speech enhancement Message-ID: <1590908849.4947693.1657533291866.JavaMail.zimbra@inria.fr> Dear all, We are looking for a research engineer or post-doc to work on audiovisual speech enhancement in the Multispeech team at Inria, France. More details can be found below or in the attached flyer. We appreciate it if you could forward this advertisement to interested candidates. Best regards, -- Mostafa Sadeghi, Researcher, Multispeech Team, Inria, Nancy - Grand Est, France. **************************************************************************************** Context: The [ https://team.inria.fr/multispeech | Multispeech team ] , at Inria Nancy, France, seeks a qualified candidate to work on signal processing and machine learning techniques for robust audiovisual speech enhancement . The candidate will be working under the co-supervision of [ https://msaadeghii.github.io/ | Mostafa Sadeghi ] (researcher, [ https://team.inria.fr/multispeech | Multispeech team ] ), [ http://xavirema.eu/ | Xavier Alameda-Pineda ] (researcher and team leader of [ https://team.inria.fr/robotlearn/ | RobotLearn team ] ), and [ https://team.inria.fr/robotlearn/team-members/radu-patrice-horaud/ | Radu Horaud ] (senior researcher, [ https://team.inria.fr/robotlearn/ | RobotLearn team ] ). Starting date & duration: October 2022 (flexible), for a duration of one year (renewable depending on funding availability and performance). Background: Audio-visual speech enhancement (AVSE) refers to the task of improving the intelligibility and quality of a noisy speech signal utilizing the complementary information of visual modality (lip movements of the speaker) [1], which could be very helpful in highly noisy environments. Recently, and due to the great success and progress of deep neural network (DNN) architectures, AVSE has been extensively revisited [1]. Existing DNN-based AVSE methods are categorized into supervised and unsupervised approaches. In the former category, a DNN is trained on a large audiovisual corpus, e.g., AVSpeech [2], with diverse enough noise instances, to directly map the noisy speech signal and the associated video frames of the speaker into a clean estimate of the target speech signal. The trained models are usually very complex and contain millions of parameters. The unsupervised methods [3] follow a statistical modeling-based approach combined with the expressive power of DNNs, which involves learning the prior distribution of clean speech using deep generative models, e.g., variational autoencoders (VAEs) [4], on clean corpora such as TCD-TIMIT [5], and estimating clean speech signal in a probabilistic way. As there is no training on noise, the models are much lighter than those of supervised methods. Furthermore, the unsupervised methods have potentially better generalization performance and robustness to visual noise thanks to their probabilistic nature [6-8]. Nevertheless, these methods are very recent and significantly less explored compared to the supervised approaches. Project description: In this project, we plan to devise a robust and efficient AVSE framework by thoroughly investigating the coupling between the recently proposed deep learning architectures for speech enhancement, both supervised and unsupervised, benefiting from the best of both worlds, along with the state-of-the-art generative modeling approaches. This will include, e.g., the use of dynamical VAEs [9], temporal convolutional networks (TCNs) [10], and attention-based strategies [11,12]. The main objectives of this project are summarized as follows: 1. Developing a neural architecture that identifies reliable (either frontal or non-frontal) and unreliable (occluded, extreme poses, missing) lip images by providing a normalized score at the output; 2. Developing deep generative models that efficiently exploit the sequential nature of data; 3. Integrating the developed visual reliability analysis network within the deep generative model that accordingly decides whether to utilize the visual data or not. This will provide a flexible and robust audiovisual fusion and enhancement framework. Requirements & skills: The preferred profile is described below. * M.Sc. or Ph.D. degree in speech/audio processing, computer vision, machine learning, or in a related field, * Ability to work independently as well as in a team, * Solid programming skills (Python, PyTorch), and deep learning knowledge, * Good level of written and spoken English. How to apply: Interested candidates are encouraged to contact Mostafa Sadeghi, Xavier Alameda-Pineda, and Radu Horaud ([first name].[last name]@inria.fr), with the required documents (CV, transcripts, motivation letter). References: [1] D. Michelsanti, Z. H. Tan, S. X. Zhang, Y. Xu, M. Yu, D. Yu, and J. Jensen, ?An overview of deep learning-based audio-visual speech enhancement and separation,? IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, 2021. [2] A. Ephrat, I. Mosseri, O. Lang, T. Dekel, K. Wilson, A. Hassidim, W.T. Freeman, M. Rubinstein, ?Looking-to-Listen at the Cocktail Party: A Speaker-Independent Audio-Visual Model for Speech Separation,? SIGGRAPH 2018. [3] M. Sadeghi, S. Leglaive, X. Alameda-Pineda, L. Girin, and R. Horaud, ?Audio-visual speech enhancement using conditional variational auto-encoders,? IEEE/ACM Transactions on Audio, Speech and Language Processing, vol. 28, pp. 1788 ?1800, 2020. [4] D. P. Kingma and M. Welling, ?Auto-encoding variational Bayes,? in International Conference on Learning Representations (ICLR), 2014. [5] N. Harte and E. Gillen, ?TCD-TIMIT: An Audio-Visual Corpus of Continuous Speech,? IEEE Transactions on Multimedia, vol.17, no.5, pp.603-615, May 2015. [6] M. Sadeghi and X. Alameda-Pineda, ?Switching variational autoencoders for noise-agnostic audio-visual speech enhancement,? in ICASSP, 2021. [7] Z. Kang, M. Sadeghi, R. Horaud, ?Face Frontalization Based on Robustly Fitting a Deformable Shape Model to 3D Landmarks,? in International Conference on Computer Vision (ICCV) Workshops, Montreal ? Virtual, Canada, Oct. 2021, pp. 1?16. [8] Z. Kang, M. Sadeghi, R. Horaud, X. Alameda-Pineda, J. Donley, and A. Kumar, ?The impact of removing head movements on audio-visual speech enhancement,? in ICASSP, 2022, pp. 1?5. [9] L. Girin, S. Leglaive, X. Bie, J. Diard, T. Hueber, and X. Alameda-Pineda, ?Dynamical variational autoencoders: A comprehensive review,? Foundations and Trends in Machine Learning, vol. 15, no. 1-2, 2021. [10] C. Lea, R. Vidal, A. Reiter, and G. D. Hager. ?Temporal convolutional networks: A unified approach to action segmentation.? In European Conference on Computer Vision, pp. 47-54. Springer, Cham, 2016. [11] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N Gomez, L. Kaiser, and I. Polosukhin, ?Attention is all you need,? in Advances in neural information processing systems, 2017, pp. 5998?6008. [12] J. Jiang, G. G Xia, D. B Carlton, C. N Anderson, and R. H Miyakawa, ?Transformer VAE: A hierarchical model for structure-aware and interpretable music representation learning,? in ICASSP, 2020, pp. 516?520. -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: research_engineer_postdoc_AVSE.pdf Type: application/pdf Size: 76148 bytes Desc: not available URL: From boubchir at ai.univ-paris8.fr Mon Jul 11 07:10:36 2022 From: boubchir at ai.univ-paris8.fr (Larbi Boubchir) Date: Mon, 11 Jul 2022 13:10:36 +0200 Subject: Connectionists: Special Issue on "Advanced Machine Learning Algorithms for Biometrics and Its Applications" In-Reply-To: <422d09db-8edf-e87d-f145-7ff9b82db981@ai.univ-paris8.fr> References: <422d09db-8edf-e87d-f145-7ff9b82db981@ai.univ-paris8.fr> Message-ID: Dear colleagues, You are cordially invited to submit a manuscript to the Applied Sciences special issue on "Advanced Machine Learning Algorithms for Biometrics and Its Applications". Please find further information below. https://www.mdpi.com/journal/applsci/special_issues/Machine_Learning_Biometrics With best regards, Prof. Larbi Boubchir From antona at alleninstitute.org Mon Jul 11 13:51:06 2022 From: antona at alleninstitute.org (Anton Arkhipov) Date: Mon, 11 Jul 2022 17:51:06 +0000 Subject: Connectionists: Scientist - Neural Data Analysis at the Allen Institute Message-ID: <83C4FEA6-9AC2-4A5D-84B4-1D40F8924963@alleninstitute.org> Dear colleagues, An exciting opportunity is open at the Allen Institute (Seattle, WA): https://alleninstitute.hrmdirect.com/employment/job-opening.php?req=1937290 Scientist - Neural Data Analysis ? Systems Neuroscience We are seeking a Scientist to be an essential team member in our ongoing efforts to link transcriptomically defined cell types with the types of functional neural activity in vivo. We are generating systematic datasets of neural activity using 2-photon calcium imaging in the mouse cortex, followed up by spatial transcriptomics, in order to better understand cell-type-specific contributions to neural coding and circuit dynamics. The core responsibility for this position will be the computational analysis of these datasets, including co-registration between the calcium imaging and spatial transcriptomics data, linking these modalities for individual cells. The ideal candidate will have experience with analysis of imaging data, sensory physiology, machine learning, computer vision, and/or spatial transcriptomics. This position sits at the interface of neuroscience and data science, and good coding and analytical skills are a high priority. The Allen Institute believes that team science significantly benefits from the participation of diverse voices, experiences, and backgrounds. High-quality science can only be produced when it includes different perspectives. We are committed to increasing diversity across every team and encourage people from all backgrounds to apply for this role. Please see details and apply here: https://alleninstitute.hrmdirect.com/employment/job-opening.php?req=1937290 Best regards, 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 benoit.frenay at unamur.be Mon Jul 11 16:34:11 2022 From: benoit.frenay at unamur.be (=?UTF-8?B?QmVub8OudCBGcsOpbmF5?=) Date: Mon, 11 Jul 2022 22:34:11 +0200 Subject: Connectionists: AIMLAI@CIKM2022: International Workshop on Advances in Interpretable Machine Learning and Artificial Intelligence Message-ID: <9536ecff-f842-7bc7-4e3f-68eae3a7b7fd@unamur.be> We invite researchers working on interpretability and explainability in ML/AI, and related topics, to submit regular (8 pages, single column) or short (3 pages, single column) papers to the AIMLAI workshop that will be held at CIKM 2022. Website: https://project.inria.fr/aimlai/ Submission link: https://easychair.org/conferences/?conf=aimlaicikm22 Submission deadline: August 15, 2022 The purpose of AIMLAI (Advances in Interpretable Machine Learning and Artificial Intelligence) is to encourage principled research that will lead to the advancement of explainable, transparent, ethical and fair data mining, machine learning, and artificial intelligence. AIMLAI is a workshop that seeks top-quality submissions addressing uncovered important issues related to explainable and interpretable data mining and machine learning models. Papers should present novel research results in any of the topics of interest for the workshop as well as application experiences, tools and promising preliminary ideas. AIMLAI asks for contributions from researchers, academia and industry, working on topics addressing these challenges primarily from a technical point of view, but also from a legal, ethical or sociological perspective. Besides the central topic of interpretable algorithms and explanation methods, we also welcome submissions that answer research questions like "how to measure and evaluate interpretability and explainability?" and "how to integrate humans in the machine learning pipeline for interpretability purposes?". This year's edition of AIMLAI is open to two kinds of submissions: regular papers (6 pages, double column) and extended abstracts (2 pages, double column). A non-exhaustive list of topics that are of interest for AIMLAI are the following: * Interpretable ML ?? * Supervised and Unsupervised ML ?? * Explaining recommendation models ?? * Multimodal explanations * Transparency in AI and ML ?? * Ethical aspects ?? * Legal aspects ?? * Fairness issues * Methodology and formalization of interpretability ?? * Formal measures of interpretability ?? * Interpretability/complexity trade-offs ?? * How to evaluate interpretability * User-centric interpretability ?? * Explanation modules ?? * Interpretability and Semantics: how to add semantics to explanations? ?? * Human-in-the-loop to construct and/or evaluate interpretable models ?? * Integration of ML algorithms, infovis and man-machine interfaces - Submission Guidelines Papers must be written in English and formatted according to the ACM sigconf template. Regular papers must be 6 pages long maximum. Extended abstracts are restricted to a maximum of 2 pages. Overlength papers will be rejected without review (papers with smaller page margins and font sizes than specified in the author instructions and set in the style files will also be treated as overlength). Authors who submit their work to AIMLAI 2022 commit themselves to present their paper at the workshop in case of acceptance. AIMLAI 2022 considers the author list submitted with the paper as final. No additions or deletions to this list may be made after paper submission, either during the review period, or in case of acceptance, at the final camera ready stage. Condition for inclusion in the post-proceedings is that at least one of the co-authors has (in-person or virtually) presented the paper at the workshop. All papers for AIMLAI 2022 must be submitted by using the online submission system at https://easychair.org/conferences/?conf=aimlaicikm22. - Program Chairs * Adrien Bibal, University of Louvain, Belgium * Tassadit Bouadi, University of Rennes/IRISA, France * Beno?t Fr?nay, University of Namur, Belgium * Luis Gal?rraga, Inria/IRISA, France * Jos? Oramas, University of Antwerp/imec-IDLab, Belgium - Important dates All dates are given in Central European Standard Time (CEST). * Paper submission deadline: August 15, 2022 * Paper reviewing period: August 18, 2022 - September 10, 2022 * Paper Notifications: September 15, 2022 * Camera-ready deadline: September 30, 2022 - Publication All accepted papers will be published as post-proceedings at https://proceedings.mlr.press/. - Venue The workshop will be co-located with the CIKM conference, which will be held in Atlanta, Georgia, USA from the 17th to the 22nd of October, 2022. Contact All questions about submissions should be emailed to aimlaicikm22 at easychair.org. From h.bilen at ed.ac.uk Tue Jul 12 04:47:32 2022 From: h.bilen at ed.ac.uk (Hakan Bilen) Date: Tue, 12 Jul 2022 09:47:32 +0100 Subject: Connectionists: PhD Studentship in Computer Vision at University of Edinburgh In-Reply-To: <86b819bc-a61e-a783-dab0-5f647ec60abc@ed.ac.uk> References: <86b819bc-a61e-a783-dab0-5f647ec60abc@ed.ac.uk> Message-ID: Dear all, We are seeking an exceptional PhD candidate to study in the Visual Computing Group (http://groups.inf.ed.ac.uk/vico/) at the University of Edinburgh. The successful candidate will have an opportunity to work on cutting-edge computer vision and machine learning research project in the intersection of 3D understanding, anomaly detection and few-shot learning from images. The goal of this project is to develop the next generation of deep learning systems for computer vision that can learn to detect anomalies based on 3D interpretation of visual scenes from limited number of images. The position is available from October 2022. PhD candidate requirements We are looking for creative and motivated applicants with, or expected to obtain soon, ideally a MSc or BSc degree in a relevant discipline, including Informatics, Computer Science, Electrical Engineering but not limited to. Background in machine learning and computer vision, and good programming skills (in python) are required. Funding This is an industry funded award and will provide an annual stipend for three and half years plus University fees. There is no restriction on the nationality. Women candidates and candidates from underrepresented backgrounds are especially encouraged to apply. Application If you are interested in the position, please provide a CV, a personal statement detailing your research interests and reasons for applying (max 1 page), marks for your degree(s) and an email address for one academic reference. There is no application deadline, the position will be open till the candidate is appointed. All documents should be in electronic format and sent via e-mail me as soon as possible (Email: h.bilen at ed.ac.uk). Best wishes, Hakan Bilen The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336. Is e buidheann carthannais a th? ann an Oilthigh Dh?n ?ideann, cl?raichte an Alba, ?ireamh cl?raidh SC005336. -------------- next part -------------- An HTML attachment was scrubbed... URL: From samuel.kaski at manchester.ac.uk Tue Jul 12 14:25:52 2022 From: samuel.kaski at manchester.ac.uk (Samuel Kaski) Date: Tue, 12 Jul 2022 18:25:52 +0000 Subject: Connectionists: 16 funded positions in artificial intelligence & machine learning for postdocs, research fellows, PhD students, at FCAI Message-ID: Join us to work on new machine learning techniques at the Finnish Center for Artificial Intelligence FCAI! We have exciting topics available around the following areas of research: (1) reinforcement learning, (2) probabilistic methods, (3) simulator-based inference, (4) privacy and federated learning, and (5) multi-agent modeling. Your work can be theoretical or applied, or both. The deadline for the postdoc/research fellow applications is on August 21 and for the PhD student applications on August 28, 2022. Read more and apply here: https://fcai.fi/we-are-hiring -------------- next part -------------- An HTML attachment was scrubbed... URL: From Donald.Adjeroh at mail.wvu.edu Tue Jul 12 11:02:06 2022 From: Donald.Adjeroh at mail.wvu.edu (Donald Adjeroh) Date: Tue, 12 Jul 2022 15:02:06 +0000 Subject: Connectionists: Call for Participation: Summer School on AI & Smart Health & Mini-workshop series (July 25-27, 2022) In-Reply-To: References: , Message-ID: We apologize if you receive multiple copies. As part of the NSF Track-2 project on "Multi-Scale Integrative Approach to Digital Health: Collaborative Research and Education in Smart Health in West Virginia and Arkansas" (http://community.wvu.edu/~daadjeroh/projects/cresh/ ) and the NSF BridgesDH NRT on "Bridges in Digital Health" (https://community.wvu.edu/~daadjeroh/projects/bridges/ ) we would like to announce a Summer School on AI & Smart Health, and related mini-workshop series. The summer school and mini-workshop series will be from July 25 - 27, 2022. This will be held remotely as a virtual event. The activities are free, with no registration fee. Please visit the website: https://community.wvu.edu/~daadjeroh/projects/cresh/activity/summersch2022/ for more information on the summer school and mini-workshop series, and to register. Please help us to forward this to students, colleagues, and others that may be interested. Regards, Don Adjeroh, PhD Professor and Associate Chair Graduate Coordinator of Computer Science West Virginia University Morgantown, WV 26506 http://community.wvu.edu/~daadjeroh/ Tel: 304-293-9681 Fax: 304-293-8602 -------------- next part -------------- An HTML attachment was scrubbed... URL: From T.E.Weyde at city.ac.uk Tue Jul 12 15:44:13 2022 From: T.E.Weyde at city.ac.uk (Weyde, Tillman) Date: Tue, 12 Jul 2022 19:44:13 +0000 Subject: Connectionists: 2 PhD positions on NLP and Knowledge Graphs at City Uni London & Rolls Royce Message-ID: We are offering two fully funded PhD studentships in Computer Science on applications of NLP combined with Knowledge graphs. The projects will be hosted by City, University of London in collaboration with Rolls Royce (R^2 Factory). More information is available at https://www.city.ac.uk/prospective-students/finance/funding/rolls-royce-studentship Please redistribute and if you are interested, please get in touch with me. -- Dr Tillman Weyde Department of Computer Science City, University of London -------------- next part -------------- An HTML attachment was scrubbed... URL: From federico.becattini at unifi.it Tue Jul 12 11:13:09 2022 From: federico.becattini at unifi.it (Federico Becattini) Date: Tue, 12 Jul 2022 17:13:09 +0200 Subject: Connectionists: [CFP] WCPA 2022 - International Workshop at ECCV 2022 Message-ID: ******************************** Call for Papers ?1st International Workshop and Challenge on People Analysis: >From Face, Body and Fashion To 3d Virtual Avatars? International Workshop at ECCV 2022 (1st Edition) https://sites.google.com/view/wcpa2022/ ******************************** === DEADLINE EXTENDED TO JULY 24!!! ==== (11.59 p.m. CET) Apologies for multiple posting Please distribute this call to interested parties AIMS AND SCOPE =============== Human-centered data are extremely widespread and have been intensely investigated by researchers belonging to even very different fields, including Computer Vision, Machine Learning, and Artificial Intelligence. These research efforts are motivated by the several highly-informative aspects of humans that can be investigated, ranging from corporal elements (e.g. bodies, faces, hands, anthropometric measurements) to emotions and outward appearance (e.g. human garments and accessories). The huge amount and the extreme variety of this kind of data make the analysis and the use of learning approaches extremely challenging. In this context, several interesting problems can be addressed, such as the reliable detection and tracking of people, the estimation of the body pose, the development of new human-computer interaction paradigms based on expression and sentiment analysis. Furthermore, considering the crucial impact of human-centered technologies in many industrial application domains, the demand for accurate models able also to run on mobile and embedded solutions is constantly increasing. For instance, the analysis and manipulation of garments and accessories worn by people can play a crucial role in the fashion business. Also, the human pose estimation can be used to monitor and guarantee the safety between workers and industrial robotic arms. The goal of this workshop is to improve the communication between researchers and companies and to develop novel ideas that can shape the future of this area, in terms of motivations, methodologies, prospective trends, and potential industrial applications. Finally, a consideration about the privacy issues behind the acquisition and the use of human-centered data must be addressed for both the academia and companies. TOPICS ======= The topics of interest include but are not limited to: - Human Body - People Detection and Tracking - 2D/3D Human Pose Estimation - Action and Gesture Recognition - Anthropometric Measurements Estimation - Gait Analysis - Person Re-identification - 3D Body Reconstruction - Human Face - Facial Landmarks Detection - Head Pose Estimation - Facial Expression and Emotion Recognition - Outward Appearance and Fashion - Garment-based Virtual Try-On - Human-centered Image and Video Synthesis - Generative Clothing - Human Clothing and Attribute Recognition - Fashion Image Manipulation - Outfit Recommendation - Human-centered Data - Novel Datasets with Human Data - Fairness and Biases in Human Analysis - Privacy Preserving and Data Anonymization - First Person Vision for Human Behavior Understanding - Multimodal Data Fusion for Human Analysis - Computational Issues in Human Analysis Architectures - Biometrics - Face Recognition and Verification - Fingerprint and Iris Recognition - Morphing Attack Detection IMPORTANT DATES ================= - Paper Submission Deadline: July 10th, 2022 July 24th, 2022 (11.59 p.m. CET) - Decision to Authors: August 10th, 2022 - Camera ready papers due: August 22nd, 2022 SUBMISSION GUIDELINES ====================== All the papers should be submitted at: https://cmt3.research.microsoft.com/WCPA2022. All papers will be reviewed by at least two reviewers with double-blind peer-review policy. Accepted submissions will be published in the ECCV 2022 Workshops proceedings. Papers must be prepared according to the ECCV guidelines. Papers are limited to 14 pages, including figures and tables, in the ECCV style. Additional pages containing only cited references are allowed. Papers that are not properly anonymized, or do not use the template, or have more than 14 pages (excluding references) will be rejected without review. Papers shorter than 4 pages (excluding references) will not be published in the ECCV 2022 Workshops proceedings. Note also that the template has changed since ECCV 2020. We therefore strongly urge authors to use this new template instead of templates from older conferences. WORKSHOP MODALITY ==================== The workshop will be held in conjunction with the European Conference on Computer Vision (ECCV 2022). The workshop will take place in an entirely virtual mode. WORKSHOP ORGANIZERS ======================= - Alberto del Bimbo, University of Florence, Italy - Mohamed Daoudi, IMT Lille Douai, France - Roberto Vezzani, University of Modena and Reggio Emilia, Italy - Xavier Alameda-Pineda, INRIA Grenoble, France - Guido Borghi, University of Bologna, Italy - Marcella Cornia, University of Modena and Reggio Emilia, Italy - Claudio Ferrari, University of Parma, Italy - Federico Becattini, University of Florence, Italy - Andrea Pilzer, NVIDIA, Italy CHALLENGE ORGANIZERS ======================= - Zhiwen Chen, Alibaba Group, China - Xiangyu Zhu, Institute of Automation, Chinese Academy of Sciences, China - Ye Pan, Shanghai Jiao Tong University, China - Xiaoming Liu, Michigan State University, USA -- Federico Becattini, Ph.D. Universit? di Firenze - MICC Tel.: +39 055 275 1394 https://www.micc.unifi.it/people/federico-becattini/ https://fedebecat.github.io/ federico.becattini at unifi.it -------------- next part -------------- An HTML attachment was scrubbed... URL: From snooles at gmail.com Wed Jul 13 03:00:51 2022 From: snooles at gmail.com (Gilles Vanwalleghem) Date: Wed, 13 Jul 2022 09:00:51 +0200 Subject: Connectionists: 16 PhD fellowships in neuroscience - Denmark Message-ID: Hi everyone, please circulate in your network this first call to join the new Neuroscience Academy Denmark program. There are 16 fully funded PhD fellowships available, with one year of rotations before making a decision on which lab to join. Information on the program can be found here: https://neuroscienceacademydenmark.dk/fellowship-2022/ Denmark is a great place to do science, and to live, PhD students are full employees with all the benefits that come with it. Please reach out if you have any questions, Gilles Vanwalleghem Assistant Professor in neurobiology Department of Molecular Biology and Genetics Team leader at Danish Research Institute of Translational Neuroscience DANDRITE, Nordic-EMBL Partnership for Molecular Medicine Aarhus University Denmark https://mbg.au.dk/en/gilles-vanwalleghem -------------- next part -------------- An HTML attachment was scrubbed... URL: From terry at salk.edu Wed Jul 13 20:19:45 2022 From: terry at salk.edu (Terry Sejnowski) Date: Wed, 13 Jul 2022 17:19:45 -0700 Subject: Connectionists: NEURAL COMPUTATION - July 1, 2022 In-Reply-To: Message-ID: Neural Computation - Volume 34, Number 7 - July 1, 2022 Available online for download now: http://www.mitpressjournals.org/toc/neco/34/7 http://cognet.mit.edu/content/neural-computation ----- Articles A Predictive Processing Model of Episodic Memory and Time Perception Zafeirios Fountas, Anastasia Sylaidi, Kyriacos Nikiforou, Anil K. Seth, Murray Shanahan, and Warrick Roseboom Reduced Dimension, Biophysical Neuron Models Constructed From Observed Data Randall Clark, Lawson Fuller, Jason A Platt, and Henry Abarbanel Differential Dopamine Receptor-dependent Sensitivity Improves the Switch Between Hard and Soft Selection in a Model of the Basal Ganglia Olivier Codol, Paul L. Gribble, and Kevin N. Gurney Letter Sensitivity of Sparse Codes to Image Distortions Kyle Luther, H. Sebastian Seung ----- 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 nagai.yukie at mail.u-tokyo.ac.jp Wed Jul 13 21:07:36 2022 From: nagai.yukie at mail.u-tokyo.ac.jp (nagai.yukie at mail.u-tokyo.ac.jp) Date: Thu, 14 Jul 2022 01:07:36 +0000 Subject: Connectionists: Call for submissions to the ICDL 2022 workshop on neurodiversity of cognitive feelings Message-ID: <1FFDB6A8-21B3-4FCB-A468-CE5D4C04B9EA@mail.u-tokyo.ac.jp> Dear colleagues, We are going to organize the following workshop at IEEE International Conference on Development and Learning (ICDL 2022), which will be held in London, UK on September 12-15. We invite submissions of extended abstracts (1-2 pages) of ongoing research that focuses on one or several of our listed topics below. Accepted submission will be considered for symposium talks or as part of a poster session. Workshop on Neurodiversity of Cognitive Feelings https://sites.google.com/view/icdl-cognitive-feeling ICDL 2022 https://icdl2022.qmul.ac.uk [ Theme of the workshop ] >From the general conception of our own subjective experiences, and from behavioural and neuroimaging research we know that affective emotional states can influence our cognitive functions. While the existence of this relationship is generally accepted, understanding the neural mechanisms that lead to individual diversity remains an open question that requires continued development of understanding and assessment of individual cognitive abilities. One of the suggested pathways includes the perception of emotional sensations that are experienced towards multisensory perception systems with a cognitive feeling. Cognitive feeling can be described as a sense of confidence, knowing, familiarity, distinguishing reality and fluency of information from multisensory perceptual experiences (Clore and Parrott, 1994; Arango-Mu?oz 2014). Where emotional sensations, or affective feelings, are the result of perceived multisensory integration such as liking, disliking, or fear, arousal. Each kind of feeling is then providing information about different system. The affective feeling provides information on whether what is perceived is positive or negative with regards to our expectations. The cognitive feeling is the informational source of our expectations, knowing and understanding the multisensory information. The principles of cognitive feeling can be applied in developmental and cognitive robotics (Asada et al. 2009; Tani 2016), predictive coding theories (Friston & Kiebel 2009; Friston 2010), computational psychiatry (Montague et al. 2012), and behavioural studies through the verification of virtual reality-based approaches. [ Welcomed topics ] Affective Neuroscience Affective Computation Interoception Self-awareness and self-recognition Rubber hand illusion and bodily illusion Sense of agency Sense of presence Psychiatric disorder and developmental disorder Consciousness Brain-body interactions Metacognitive feeling [ Keynote speakers ] ? Professor Sarah Garfinkel, Institute of Cognitive Neuroscience, University College London, UK ? Dr Sophie Betka, The ?cole polytechnique f?d?rale de Lausanne (EPFL), Switzerland ? Professor Tony Prescott, Department of Computer Science, The University of Sheffield, UK [ Submission ] Extended abstract submission guidelines can be found on the workshop website (https://sites.google.com/view/icdl-cognitive-feeling), which should be sent to Sean D. Lynch (lynch.sean at mail.u-tokyo.ac.jp) by August 13, 2022. Please write in the subject line ?ICDL 2022 neurodiversity workshop submission?. [ Important dates ] Submission deadline: August 13, 2022 Notification of acceptance: August 27, 2022 Workshop: September 12, 2022 [ Organizers ] Sean Lynch (The University of Tokyo) Keisuke Suzuki (Hokkaido University) Yukie Nagai (The University of Tokyo) ? 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 G.C.H.E.deCroon at tudelft.nl Wed Jul 13 14:56:12 2022 From: G.C.H.E.deCroon at tudelft.nl (Guido de Croon - LR) Date: Wed, 13 Jul 2022 18:56:12 +0000 Subject: Connectionists: =?utf-8?q?=5BJobs=5D_Open_Assistant_Professor_Pos?= =?utf-8?q?ition_=22Flying_Artificial_Intelligence_at_the_Edge=E2=80=9D_at?= =?utf-8?q?_TU_Delft=2C_the_Netherlands=2E?= Message-ID: Open Assistant Professor Position "Flying Artificial Intelligence at the Edge? - Tenure Track at TU Delft, the Netherlands, Faculty of Aerospace Engineering Job description Artificial Intelligence (AI) is disrupting various domains and industries, and the aerospace domain is no exception. There has been an accelerated increase in the number of drones. Their number has surpassed the number of aircraft already years ago. Currently, most drones have a limited form of autonomy, taking care of stabilization and the following of GPS waypoints. However, there are now more and more professional products on the market that are also able to avoid obstacles and follow a person to make the best possible selfie or action video. AI plays an enormous role in these capabilities. The demands on autonomy and hence AI on drones will increase, while it is enormously important not to add too much weight and processing, as the energy of drones is a very limiting factor for their flight duration and usefulness. This calls for innovations in AI for which we need to prepare a new generation of engineers and researchers. The proposed position will involve aiding in current AI-related courses at the faculty of Aerospace Engineering, and contributing to additional ones, as well as performing research on relevant topics, such as deep neural networks, neuromorphic sensing and processing, bio-inspired algorithms, swarming, reinforcement learning, computer vision, or any other topics that are relevant to AI at the edge for autonomous drones. You will be expected to: * develop courses and conduct teaching at undergraduate and post-graduate levels * perform top-quality research in Artificial Intelligence for robotics * establish and execute an externally funded research programme * interact and collaborate with other researchers and specialists in academia and / or industry * be an inspiring member of our staff and have excellent communication skills Requirements We expect the candidate to have: * a PhD degree in Computer Science, Artificial Intelligence, Control Systems Theory, or a related discipline and several years of post-doctoral experience * excellent track record in scientific research, as evidenced by publications in academic journals * experience in the acquisition of external funding and an entrepreneurial mindset * proven ability to provide inspiring teaching in English It is desired for the candidate to have experience in designing / working with robotic systems, (embedded) programming and software development. Conditions of employment A tenure-track position is offered for six years. In the fifth year we?ll decide if you will be offered a permanent faculty position, based on performance indicators agreed upon at the start of the appointment. We expect that you have the potential to grow towards an Associate Professor and/or Full Professor role in the future. Inspiring, excellent education is our central aim. We expect you to obtain a University Teaching Qualification (UTQ) within three years if you have less than five years of teaching experience. This is provided by the TU Delft UTQ programme. TU Delft sets high standards for the English competency of the teaching staff. The TU Delft offers training to improve English competency. If you do not speak Dutch, we offer courses to learn the Dutch language within three years. Salary and benefits are in accordance with the Collective Labour Agreement for Dutch Universities. The TU Delft offers a customisable compensation package, a discount on health insurance and sport memberships, and a monthly work costs contribution. Flexible work schedules can be arranged. For international applicants we offer the Coming to Delft Service and Partner Career Advice to assist you with your relocation. An International Children's Centre offers childcare and there is an international primary school. TU Delft (Delft University of Technology) Delft University of Technology is built on strong foundations. As creators of the world-famous Dutch waterworks and pioneers in biotech, TU Delft is a top international university combining science, engineering and design. It delivers world class results in education, research and innovation to address challenges in the areas of energy, climate, mobility, health and digital society. For generations, our engineers have proven to be entrepreneurial problem-solvers, both in business and in a social context. At TU Delft we embrace diversity and aim to be as inclusive as possible (see our Code of Conduct). Together, we imagine, invent and create solutions using technology to have a positive impact on a global scale. Faculty Aerospace Engineering The Faculty of Aerospace Engineering at Delft University of Technology is one of the world?s most highly ranked (and most comprehensive) research, education and innovation communities devoted entirely to aerospace engineering. More than 200 science staff, around 250 PhD candidates and over 2,700 BSc and MSc students apply aerospace engineering disciplines to address the global societal challenges that threaten us today, climate change without doubt being the most important. Our focal subjects: sustainable aerospace, big data and artificial intelligence, bio-inspired engineering and smart instruments and systems. Working at the faculty means working together. With partners in other faculties, knowledge institutes, governments and industry, both aerospace and non-aerospace. Working in field labs and innovation hubs on our university campus and beyond. Additional information The selected candidate will be employed in the section Control and Simulation (C&S) in the Faculty of Aerospace of Delft University of Technology. The candidate will be associated to the Micro Air Vehicle lab (MAVLab: http://mavlab.tudelft.nl/). This lab brings together knowledge and experience in many fields including MAV flight dynamics and control, aerodynamics, electronics, and artificial intelligence. The lab includes all the machines necessary for going from the design of a novel MAV to its realization, including 3D printers, milling machines, wire cutters, etc. Furthermore, the lab runs the ?Cyberzoo? facility, a 10x10x7m motion tracking arena, and has the possibility for outdoor testing in various test locations. Application procedure Are you interested in this vacancy? Please apply before 31-07-2022 via the following link and upload your motivation letter and CV and three email addresses that can be contacted for additional information: https://www.tudelft.nl/over-tu-delft/werken-bij-tu-delft/vacatures/details?jobId=7200&jobTitle=Assistant%20Professor%20%26quot%3BFlying%20Artificial%20Intelligence%20at%20the%20Edge%20-%20Tenure%20Track A pre-employment screening can be part of the selection procedure. Applying for an exemption for specific research and educational areas is an obligatory part of the selection procedure for this vacancy. This exemption must be obtained from the Ministry of Education, Culture and Science (OCW) before an employment contract is agreed upon. More information can be found here: What technology courses require an exemption from the knowledge embargo, and how can I apply for an exemption? | Secondary vocational education (MBO) and higher education | Government.nl For more information on the position you can contact g.c.h.e.decroon at tudelft, mentioning the vacancy number (TUD02404). -------------- next part -------------- An HTML attachment was scrubbed... URL: From domenico.maisto at cnr.it Thu Jul 14 04:05:43 2022 From: domenico.maisto at cnr.it (Domenico Maisto) Date: Thu, 14 Jul 2022 10:05:43 +0200 Subject: Connectionists: CfP: Special Issue on Information Theory and Cognitive Agents - Entropy by MDPI. Message-ID: Dear Colleagues, You are cordially invited to submit a manuscript to the Special Issue "Information Theory and Cognitive Agents" in /Entropy/. https://www.mdpi.com/journal/entropy/special_issues/IT_Cognitive_Agents Deadline for manuscript submissions: 30 November 2022. Please find further information below. My very best regards, Domenico Maisto ================================================================ Open Access Journal ENTROPY by MDPI Special Issue on Information Theory and Cognitive Agents ----------------------------------------------------------------------------------------------------------------- This Special Issue aims to focus on recent advances in Information Theory (IT) and their applications on Cognitive Agents. We warmly welcome contributions that: - Propose models of Cognitive Agents in which perception, control, and learning are addressed using one among the various IT frameworks; - Present innovative IT methods, or introduce progresses in consolidated approaches (e.g., information bottleneck, empowerment, free energy principle, algorithmic probabilistic induction, etc); - Illustrate IT-based Cognitive Agents designed as hypotheses and conjectures about every aspects of cognition; - Show stimulating employment of IT-based Cognitive Agents to solve complex problems in specific domains. Guest Editors Dr. Domenico Maisto Prof. Daniel Polani Entropy (ISSN 1099-4300) is an Open Access Journal by MDPI with high visibility (indexed within Scopus, SCIE (Web of Science), MathSciNet, Inspec, PubMed, PMC, and many other databases), excellent citescore (IF 2021: 2.738; JCR: Q1 of ?Mathematical Physics?, Q2 of ?Physics and Astronomy?, Q2 of Information Systems and Q2 of ?Electrical and Electronic Engineering?, Q2 "Physics, Multidisciplinary"), and rapid publication (18.5 days from submission to first decision in 2021, acceptance to publication undertaken in 3.4 days). -------------- next part -------------- An HTML attachment was scrubbed... URL: From marcel.van.gerven at gmail.com Thu Jul 14 03:58:38 2022 From: marcel.van.gerven at gmail.com (Marcel van Gerven) Date: Thu, 14 Jul 2022 09:58:38 +0200 Subject: Connectionists: Postdoctoral researcher in the AI Department of the Donders Institute for Brain, Cognition and Behaviour Message-ID: <138E714E-F270-4F86-A1DC-DB9B943BBDF4@gmail.com> The AI Department of the Donders Centre for Cognition (DCC), embedded in the Donders Institute for Brain, Cognition and Behaviour, and the School of Artificial Intelligence at Radboud University is looking for a postdoctoral researcher with excellent communication skills and a background in machine learning, scientific computing and/or computational neuroscience. The objective of the position is to develop advanced computational models of learning, inference and control in the context of neuroscience and neurotechnology based on dynamical systems theory, neural differential equations, reinforcement learning and control theory. You will be part of the DBI2 consortium (https://www.ru.nl/donders/@1354007/gravitation-proposal-awarded-research-brain/ ) which investigates mechanisms of brain function in natural agents. DCC and the Donders Institute provide excellent facilities such as computing facilities, a robot lab, a virtual reality lab, behavioural labs, and a technical support group. The AI Department is also a founding member of Radboud AI and the ELLIS Unit Nijmegen (European Excellence Network in Machine Learning). You will join the Artificial Cognitive Systems (ACS) group and interact closely with other machine learning researchers and computational neuroscientists. If you are interested, please check out https://www.ru.nl/en/working-at/job-opportunities/postdoctoral-researcher-in-modelling-and-control-of-neural-systems-at-the-donders-center-for-cognition and feel encouraged to apply (deadline August 28). Kind regards, Marcel van Gerven ?? Marcel van Gerven | Professor of Artificial Intelligence | PI of the Artificial Cognitive?Systems lab (@artcogsys ) | Chair of the Department of Artificial?Intelligence | Donders Institute for Brain, Cognition?and Behaviour | ELLIS Fellow | Director of ELLIS Nijmegen | Radboud University | Maria Montessoribuilding, Room 02.368 | Thomas van Aquinostraat 4, 6525 GD Nijmegen, the Netherlands -------------- next part -------------- An HTML attachment was scrubbed... URL: From cgf at isep.ipp.pt Thu Jul 14 04:51:32 2022 From: cgf at isep.ipp.pt (Carlos) Date: Thu, 14 Jul 2022 09:51:32 +0100 Subject: Connectionists: CFP: BDL 2022 - IEEE SBAC-PAD 2022 - Extended Submission Deadline: 30 of July Message-ID: <502f0e88-ed6e-3bed-c226-c159435d9e7b@isep.ipp.pt> --------------- CALL FOR PAPERS --------------- BDL 2022 Workshop on Big Data & Deep Learning in High Performance Computing in conjunction with IEEE SBAC-PAD 2022 Bordeaux, France, November 2-5, 2022 https://www.dcc.fc.up.pt/bdl2022/ ---------------------- Aims and scope of BDL ---------------------- The number of very large data repositories (big data) is increasing in a rapid pace. Analysis of such repositories using the traditional sequential implementations of Machine Learning (ML) and emerging techniques, like deep learning, that model high-level abstractions in data by using multiple processing layers, requires expensive computational resources and long running times. Parallel or distributed computing are possible approaches that can make analysis of very large repositories and exploration of high-level representations feasible. Taking advantage of a parallel or a distributed execution of a ML/statistical system may: i) increase its speed; ii) learn hidden representations; iii) search a larger space and reach a better solution or; iv) increase the range of applications where it can be used (because it can process more data, for example). Parallel and distributed computing is therefore of high importance to extract knowledge from massive amounts of data and learn hidden representations. The workshop will be concerned with the exchange of experience among academics, researchers and the industry whose work in big data and deep learning require high performance computing to achieve goals. Participants will present recently developed algorithms/systems, on going work and applications taking advantage of such parallel or distributed environments. ------ Topics ------ BDL 2022 invites papers on all topics in novel data-intensive computing techniques, data storage and integration schemes, and algorithms for cutting-edge high performance computing architectures which targets Big Data and Deep Learning are of interest to the workshop. Examples of topics include but not limited to: * parallel algorithms for data-intensive applications; * scalable data and text mining and information retrieval; * using Hadoop, MapReduce, Spark, Storm, Streaming to analyze Big Data; * energy-efficient data-intensive computing; * deep-learning with massive-scale datasets; * querying and visualization of large network datasets; * processing large-scale datasets on clusters of multicore and manycore processors, and accelerators; * heterogeneous computing for Big Data architectures; * Big Data in the Cloud; * processing and analyzing high-resolution images using high-performance computing; * using hybrid infrastructures for Big Data analysis; * new algorithms for parallel/distributed execution of ML systems; * applications of big data and deep learning to real-life problems. ------------------ Program Chairs ------------------ Jo?o Gama, University of Porto, Portugal Carlos Ferreira, Polytechnic Institute of Porto, Portugal Miguel Areias, University of Porto, Portugal ----------------- Program Committee ----------------- TBA ---------------- Important dates ---------------- Submission deadline: July 30, 2022(AoE) Author notification: September 2, 2022 Camera-ready: September 12, 2022 Registration deadline: September 14, 2022 ---------------- Paper submission ---------------- Papers submitted to BDL 2022 must describe original research results and must not have been published or simultaneously submitted anywhere else. Manuscripts must follow the IEEE conference formatting guidelines and submitted via the EasyChair Conference Management System as one pdf file. The strict page limit for initial submission and camera-ready version is 8 pages in the aforementioned format. Each paper will receive a minimum of three reviews by members of the international technical program committee. Papers will be selected based on their originality, relevance, technical clarity and quality of presentation. At least one author of each accepted paper must register for the BDL 2022 workshop and present the paper. ----------- Proceedings ----------- All accepted papers will be published at IEEE Xplore. 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 danko.nikolic at gmail.com Thu Jul 14 06:05:28 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Thu, 14 Jul 2022 12:05:28 +0200 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 and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 school at utia.cas.cz Thu Jul 14 08:03:51 2022 From: school at utia.cas.cz (Miroslav Karny) Date: Thu, 14 Jul 2022 14:03:51 +0200 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: <1657800231454062568@utia.cas.cz> Dear all, I am an external observer of your interesting discussions. It has been a bit surprising to me that the work of prof. Nikolic does not comment the work @book{Fus:05, title={Cortex and mind: Unifying cognition}, author={J.M. Fuster}, year={2005}, publisher={Oxford university press}}, which I feel as the highly relevant predecessor of his work. Best regards Miroslav Karny https://www.utia.cas.cz/people/karny Danko Nikolic wrote: > Dear Gary and everyone, > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > I am happy to announce that I have now finished a draft of a paper in which > I propose how the brain is able to achieve that. The manuscript requires a > bit of patience for two reasons: one is that the reader may be exposed for > the first time to certain aspects of brain physiology. The second reason is > that it may take some effort to understand the counterintuitive > implications of the new ideas (this requires a different way of thinking > than what we are used to based on connectionism). > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > (I apologize for the neuroscience lingo, which I tried to minimize.) > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > Thanks. > > Danko > > > Dr. Danko Nikoli? > http://www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > > > On Thu, Feb 3, 2022 at 5:25 PM Gary Marcus " target="_blank"> 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 " target="_blank"> 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? > > http://www.danko-nikolic.com > > > > https://www.linkedin.com/in/danko-nikolic/ > > > > --- A progress usually starts with an insight --- > > > > > > > > Virus-free. > > http://www.avast.com > > > > > > On Thu, Feb 3, 2022 at 9:55 AM Asim Roy " target="_blank"> 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 " target="_blank"> *On > >> Behalf Of *Gary Marcus > >> *Sent:* Wednesday, February 2, 2022 1:26 PM > >> *To:* Geoffrey Hinton " target="_blank"> > >> *Cc:* AIhub " target="_blank">; 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 " target="_blank"> > >> 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 " target="_blank"> 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 " target="_blank"> 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. > > http://www.avast.com > > > > > > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Thu Jul 14 08:08:19 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Thu, 14 Jul 2022 14:08:19 +0200 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <1657800231454062568@utia.cas.cz> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <1657800231454062568@utia.cas.cz> Message-ID: This can still be improved on. Always happy to cite relevant predecessors. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Thu, Jul 14, 2022 at 2:03 PM Miroslav Karny wrote: > Dear all, > > I am an external observer of your interesting discussions. It has been a > bit surprising to me that the work of prof. Nikolic does not comment > the work @book{Fus:05, title={Cortex and mind: Unifying > cognition}, author={J.M. Fuster}, year={2005}, publisher={Oxford university > press}}, which* I feel *as the highly relevant predecessor of his work. > > Best regards Miroslav Karny > > https://www.utia.cas.cz/people/karny > > Danko Nikolic wrote: > > Dear Gary and everyone, > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > I am happy to announce that I have now finished a draft of a paper in which > I propose how the brain is able to achieve that. The manuscript requires a > bit of patience for two reasons: one is that the reader may be exposed for > the first time to certain aspects of brain physiology. The second reason is > that it may take some effort to understand the counterintuitive > implications of the new ideas (this requires a different way of thinking > than what we are used to based on connectionism). > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > (I apologize for the neuroscience lingo, which I tried to minimize.) > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > Thanks. > > Danko > > > Dr. Danko Nikoli? > http://www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > > > On Thu, Feb 3, 2022 at 5:25 PM 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? > > http://www.danko-nikolic.com > > < > https://urldefense.proofpoint.com/v2/url?u=http-3A__www.danko-2Dnikolic.com&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=HwOLDw6UCRzU5-FPSceKjtpNm7C6sZQU5kuGAMVbPaI&e= > > > > https://www.linkedin.com/in/danko-nikolic/ > > < > https://urldefense.proofpoint.com/v2/url?u=https-3A__www.linkedin.com_in_danko-2Dnikolic_&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=b70c8lokmxM3Kz66OfMIM4pROgAhTJOAlp205vOmCQ8&e= > > > > --- A progress usually starts with an insight --- > > > > > > > > < > https://urldefense.proofpoint.com/v2/url?u=https-3A__www.avast.com_sig-2Demail-3Futm-5Fmedium-3Demail-26utm-5Fsource-3Dlink-26utm-5Fcampaign-3Dsig-2Demail-26utm-5Fcontent-3Dwebmail&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=Ao9QQWtO62go0hx1tb3NU6xw2FNBadjj8q64-hl5Sx4&e=> > Virus-free. > > http://www.avast.com > > < > https://urldefense.proofpoint.com/v2/url?u=https-3A__www.avast.com_sig-2Demail-3Futm-5Fmedium-3Demail-26utm-5Fsource-3Dlink-26utm-5Fcampaign-3Dsig-2Demail-26utm-5Fcontent-3Dwebmail&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=Ao9QQWtO62go0hx1tb3NU6xw2FNBadjj8q64-hl5Sx4&e= > > > > > > 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 > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__lifeboat.com_ex_bios.asim.roy&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=oDRJmXX22O8NcfqyLjyu4Ajmt8pcHWquTxYjeWahfuw&e= > > > >> > >> Asim Roy | iSearch (asu.edu) > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__isearch.asu.edu_profile_9973&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=jCesWT7oGgX76_y7PFh4cCIQ-Ife-esGblJyrBiDlro&e= > > > >> > >> > >> > >> > >> > >> *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 > >> < > https://urldefense.com/v3/__https:/arxiv.org/abs/2201.02387__;!!IKRxdwAv5BmarQ!INA0AMmG3iD1B8MDtLfjWCwcBjxO-e-eM2Ci9KEO_XYOiIEgiywK-G_8j6L3bHA$ > >) > >> 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 > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__aclanthology.org_2020.acl-2Dmain.463.pdf&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=xnFSVUARkfmiXtiTP_uXfFKv4uNEGgEeTluRFR7dnUpay2BM5EiLz-XYCkBNJLlL&s=K-Vl6vSvzuYtRMi-s4j7mzPkNRTb-I6Zmf7rbuKEBpk&e= > >, > >> 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 > >> < > https://urldefense.proofpoint.com/v2/url?u=http-3A__AIhub.org&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=xnFSVUARkfmiXtiTP_uXfFKv4uNEGgEeTluRFR7dnUpay2BM5EiLz-XYCkBNJLlL&s=eOtzMh8ILIH5EF7K20Ks4Fr27XfNV_F24bkj-SPk-2A&e= > >, > >> 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/ > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__aihub.org_2022_02_02_what-2Dis-2Dai-2Dstephen-2Dhanson-2Din-2Dconversation-2Dwith-2Dgeoff-2Dhinton_&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=OY_RYGrfxOqV7XeNJDHuzE--aEtmNRaEyQ0VJkqFCWw&e= > > > >> > >> > >> > >> 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://urldefense.proofpoint.com/v2/url?u=http-3A__AIhub.org&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=xnFSVUARkfmiXtiTP_uXfFKv4uNEGgEeTluRFR7dnUpay2BM5EiLz-XYCkBNJLlL&s=eOtzMh8ILIH5EF7K20Ks4Fr27XfNV_F24bkj-SPk-2A&e= > > > >> (https://aihub.org/ > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__aihub.org_&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=IKFanqeMi73gOiS7yD-X_vRx_OqDAwv1Il5psrxnhIA&e= > >). > >> 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 > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__aaai.org_&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=wBvjOWTzEkbfFAGNj9wOaiJlXMODmHNcoWO5JYHugS0&e= > >, > >> NeurIPS > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__neurips.cc_&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=3-lOHXyu8171pT_UE9hYWwK6ft4I-cvYkuX7shC00w0&e= > >, > >> ICML > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__icml.cc_imls_&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=JJyjwIpPy9gtKrZzBMbW3sRMh3P3Kcw-SvtxG35EiP0&e= > >, > >> AIJ > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__www.journals.elsevier.com_artificial-2Dintelligence&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=eWrRCVWlcbySaH3XgacPpi0iR0-NDQYCLJ1x5yyMr8U&e= > > > >> /IJCAI > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__www.journals.elsevier.com_artificial-2Dintelligence&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=eWrRCVWlcbySaH3XgacPpi0iR0-NDQYCLJ1x5yyMr8U&e= > >, > >> ACM SIGAI > >> < > https://urldefense.proofpoint.com/v2/url?u=http-3A__sigai.acm.org_&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=7rC6MJFaMqOms10EYDQwfnmX-zuVNhu9fz8cwUwiLGQ&e= > >, > >> EurAI/AICOMM, CLAIRE > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__claire-2Dai.org_&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=66ZofDIhuDba6Fb0LhlMGD3XbBhU7ez7dc3HD5-pXec&e= > > > >> and RoboCup > >> < > https://urldefense.proofpoint.com/v2/url?u=https-3A__www.robocup.org__&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=yl7-VPSvMrHWYKZFtKdFpThQ9UTb2jW14grhVOlAwV21R4FwPri0ROJ-uFdMqHy1&s=bBI6GRq--MHLpIIahwoVN8iyXXc7JAeH3kegNKcFJc0&e= > > > >> . > >> > >> Twitter: @aihuborg > >> > >> > > > > < > https://urldefense.proofpoint.com/v2/url?u=https-3A__www.avast.com_sig-2Demail-3Futm-5Fmedium-3Demail-26utm-5Fsource-3Dlink-26utm-5Fcampaign-3Dsig-2Demail-26utm-5Fcontent-3Dwebmail&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=Ao9QQWtO62go0hx1tb3NU6xw2FNBadjj8q64-hl5Sx4&e=> > Virus-free. > > http://www.avast.com > > < > https://urldefense.proofpoint.com/v2/url?u=https-3A__www.avast.com_sig-2Demail-3Futm-5Fmedium-3Demail-26utm-5Fsource-3Dlink-26utm-5Fcampaign-3Dsig-2Demail-26utm-5Fcontent-3Dwebmail&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=Ao9QQWtO62go0hx1tb3NU6xw2FNBadjj8q64-hl5Sx4&e= > > > > > > > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From steve at bu.edu Thu Jul 14 09:30:16 2022 From: steve at bu.edu (Grossberg, Stephen) Date: Thu, 14 Jul 2022 13:30:16 +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 Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus Cc: connectionists at mailman.srv.cs.cmu.edu ; AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 ksharma.raj at gmail.com Thu Jul 14 10:31:37 2022 From: ksharma.raj at gmail.com (Raj Sharma) Date: Thu, 14 Jul 2022 20:01:37 +0530 Subject: Connectionists: ICMI 2022: Call for Demos and Exhibits (Deadline Extended) Message-ID: **apologies if you have received multiple copies of this email* ********************************************************************* ICMI 2022 24th ACM International Conference on Multimodal Interaction https://icmi.acm.org/2022/ 7-11 Nov 2022, Bengaluru, India ********************************************************************* Call for Demonstrations and Exhibits https://icmi.acm.org/2022/call-for-demonstrations-and-exhibits/ We invite you to submit your proposals for demonstrations and exhibits to be held during the 24th ACM International Conference on Multimodal Interaction (ICMI 2022), located in Bengaluru (Bangalore), India, November 7-11th, 2022. This year?s conference theme is ?Intelligent and responsible Embodied Conversational Agents (ECAs) in the multi-lingual real world?. Demonstrations and Exhibits The ICMI 2022 Demonstrations & Exhibits session is intended to provide a forum to showcase innovative implementations, systems and technologies demonstrating new ideas about interactive multimodal interfaces. It can also serve as a platform to introduce commercial products. Proposals may be of two types: demonstrations or exhibits. The main difference is that demonstrations include a 2-3 page paper in one column, which will be included in the ICMI main proceedings, while the exhibits only need to include a brief outline (no more than two pages in one column; not included in ICMI proceedings). We encourage both the submission of early research prototypes and interesting mature systems. In addition, authors of accepted regular research papers may be invited to participate in the demonstration sessions as well. Demonstration Submission Please submit a 2-3 page description of the demonstration in a single column format through the main ICMI conference management system ( new.precisionconference.com/sigchi). Demonstration description(s) must be in PDF format, according to the ACM conference format, of no more than 3 pages in a single column format including references. For instructions and links to the templates, please see the Guidelines for Authors . Demonstration proposals should include a description with photographs and/or screen captures of the demonstration. Demonstration submissions should be accompanied by a video of the proposed demo (no larger than 200MB), which can include a set of slides (no more than 10 slides) in PowerPoint format. The demo and exhibit paper submissions are not anonymous. However, all ACM rules and guidelines related to paper submission should be followed (e.g. plagiarism, including self-plagiarism). The demonstration submissions will be peer reviewed, according to the following criteria: suitability as a demo, scientific or engineering feasibility of the proposed demo system, application, or interactivity, alignment with the conference focus, potential to engage the audience, and overall quality and presentation of the written proposal. Authors are encouraged to address such criteria in their proposals, along with preparing the papers mindful of the quality and rigorous scientific expectations of an ACM publication. The demo program will include the accepted proposals and may additionally include invited demos from among regular papers accepted for presentation at the conference. Please note that the accepted demos will be included in the ICMI main proceedings. Exhibit Submission Exhibit proposals should be submitted following the same guidelines, formatting, and due dates as for demonstration proposals. Exhibit proposals must be shorter in length (up to two pages), and are more suitable for showcasing mature systems. Like demos, submissions for exhibits should be accompanied by a video (no larger than 200MB), which can include a set of slides (no more than 10 slides) in PowerPoint format. Exhibits will not have a paper published in the ICMI 2022 proceedings. Facilities Once accepted, demonstrators and video presenters will be provided with a table, poster board, power outlet and wireless (shared) Internet. Demo and video presenters are expected to bring with them everything else needed for their demo and video presentations, such as hardware, laptops, sensors, PCs, etc. However, if you have special requests such as a larger space, special lighting conditions and so on, we will do our best to arrange them. Important note for the authors: 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 prior to the first day of the conference. The official publication date affects the deadline for any patent filings related to published work. Attendance At least one author of all accepted Demonstrations and Exhibits submissions must register for and attend the conference, including the conference demonstrations and exhibits session(s). Important Dates Submission of demo and exhibit proposals: July 26, 2022 (Extended Deadline) Demo and exhibit notification of acceptance: August 8, 2022 Submission of demo final papers: August 19, 2022 Questions? For further questions, contact the Demonstrations and Exhibits co-chairs: Dan Bohus and Ramanathan Subramanian (icmi2022-demo-chairs at acm.org). -------------- next part -------------- An HTML attachment was scrubbed... URL: From steve at bu.edu Thu Jul 14 13:14:21 2022 From: steve at bu.edu (Grossberg, Stephen) Date: Thu, 14 Jul 2022 17:14:21 +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 Danko, I will respond to your comment below to the entire list. I recommend that future interactions be done between just the two of us. Everything that I write below is summarized in my new book: In brief, within Adaptive Resonance Theory, or ART, there IS no problem of scaling, in the sense that you describe it below, from learning (say) to correctly recognizing 100 objects to doing the same for 200. In fact, one of the properties for which I introduced ART in 1976 was to enable incremental learning in real time of arbitrary numbers of objects or events without experiencing catastrophic forgetting. I have called this a solution of the stability-plasticity dilemma: How our brains, and models like ART, can rapidly learn (plasticity) arbitrary numbers of objects without experiencing catastrophic forgetting (stability). I also derived ART in 1980, in an article within Psychological Review, from a thought experiment about how ANY system can AUTONOMOUSLY correct predictive errors in a changing world that is filled with unexpected events. This thought is experiment was derived from a few facts that are familiar to us all. They are familiar because they represent ubiquitous environmental pressures that we all experience. The thought experiment clarifies why, when they act together during the evolutionary process, models like ART are the unique result. Moveover, at no point during the thought experiment are the words mind or brain mentioned. ART is thus a universal solution of this learning, classification, and prediction problem. You write below about "connectionist systems". ART is a connectionist system. What do you mean by a "connectionist system"? What you write below about them is not true in general. Best, Steve ________________________________ From: Danko Nikolic Sent: Thursday, July 14, 2022 12:16 PM To: Grossberg, Stephen Cc: Gary Marcus ; connectionists at mailman.srv.cs.cmu.edu ; AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 Jul 14 09:36:53 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Thu, 14 Jul 2022 06:36:53 -0700 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <6F562DD8-5B77-4132-A304-EEBAEA69DA3D@nyu.edu> Dear Danko, Thanks for sharing this; the brain surely does something like what you say?the transient selection of subnetworks?and it is very interesting to consider both how the brain does it and what implications that might have for questions about AI and learning. An avenue well worth exploring. - Gary > On Jul 14, 2022, at 3:05 AM, Danko Nikolic wrote: > > ? > Dear Gary and everyone, > > I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. > > As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. > > I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). > > In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. > > I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. > > The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug > (I apologize for the neuroscience lingo, which I tried to minimize.) > > It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? > > Thanks. > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > > >> On Thu, Feb 3, 2022 at 5:25 PM 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 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 danko.nikolic at gmail.com Thu Jul 14 12:16:48 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Thu, 14 Jul 2022 18:16:48 +0200 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 Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: > Dear Danko, > > I have just read your new article and would like to comment briefly about > it. > > In your introductory remarks, you write: > > "However, connectionism did not yet produce a satisfactory explanation of > how the mental emerges from the physical. A number of open problems remain > s ( 5,6,7,8). As a result, the explanatory gap between the mind and the > brain remains wide open." > > I certainly believe that no theoretical explanation in science is ever > complete. However, I also believe that "the explanatory gap between the > mind and the brain" does not remain "wide open". > > My Magnum Opus, that was published in 2021, makes that belief clear in its > title: > > *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* > > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > The book provides a self-contained and non-technical exposition in a > conversational tone of many principled and unifying explanations of > psychological and neurobiological data. > > In particular, it explains roles for the metabotropic glutamate receptors > that you mention in your own work. See the text and figures around p. 521. > This explanation unifies psychological, anatomical, neurophysiological, > biophysical, and biochemical data about the processes under discussion. > > I have a very old-fashioned view about how to understand scientific > theories. I get excited by theories that explain and predict more data than > previous theories. > > Which of the data that I explain in my book, and support with quantitative > computer simulations, can you also explain? > > What data can you explain, in the same quantitative sense, that you do not > think the neural models in my book can explain? > > I would be delighted to discuss these issues further with you. > > If you are in touch with my old friend and esteemed colleague, Wolf > Singer, please send him my warm regards. I cite the superb work that he and > various of his collaborators have done in many places in my book. > > Best, > > Steve > > 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 Danko Nikolic > *Sent:* Thursday, July 14, 2022 6:05 AM > *To:* Gary Marcus > *Cc:* connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > Dear Gary and everyone, > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > I am happy to announce that I have now finished a draft of a paper in > which I propose how the brain is able to achieve that. The manuscript > requires a bit of patience for two reasons: one is that the reader may be > exposed for the first time to certain aspects of brain physiology. The > second reason is that it may take some effort to understand the > counterintuitive implications of the new ideas (this requires a different > way of thinking than what we are used to based on connectionism). > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > (I apologize for the neuroscience lingo, which I tried to minimize.) > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > Thanks. > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > > > On Thu, Feb 3, 2022 at 5:25 PM 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 krallinger.martin at gmail.com Thu Jul 14 09:52:44 2022 From: krallinger.martin at gmail.com (Martin Krallinger) Date: Thu, 14 Jul 2022 15:52:44 +0200 Subject: Connectionists: Research Engineer Position: NLP applied to biomedicine and biomaterials at BSC, Barcelona, Spain In-Reply-To: References: Message-ID: Dear all, We are looking for motivated candidates to work on biomedical text mining and NLP at my group, mainly on topics related to biomedical NLP (biomaterials) in the context of two large Horizon Europe international projects which enable direct collaboration with clinical sites, research groups as well as technological partners. *Reference:* 273_22_LS_TM_RE2 *Job title:* Research Engineer - Text Mining (RE2) *Applications procedure and process**:* https://www.bsc.es/join-us/job-opportunities/27322lstmre2 *Context And Mission* The Text Mining Unit (PI Martin Krallinger) has several research lines related to biomedical and clinical NLP and Text Mining activities with particular interest in implementing advanced deep learning/transformer based approaches as well as building and exploiting domain specific language models. Through collaboration with research groups, hospitals as well as project partners (Horizon Projects) our aim is to develop real world applications related to semantic annotation (NER) components, entity linking, semantic text similarity, building and exploiting text-derived knowledge graph as well as domain specific machine translation technologies. We search for a Research Engineer in text mining / Natural Language Processing (NLP), medical informatics, computational linguistics or language engineering with a strong background in Machine Learning (ML), who will be responsible for the implementation of text mining tools for the Horizon BIOMATDB (https://biomatdb.eu) project. *Key Duties* - Design, implementation, and evaluation of text mining, NLP, deep learning and ML tools and models applied to the biomaterials, chemistry, biology and biomedical application domain. - Coordination and organization of shared tasks and evaluation campaigns (like BioASQ, IberLEF,..). - Technical coordination and supervision of annotation projects to generate high-quality text corpora. - Integration of results and components into a biomaterials knowledgebase. *Requirements* Education - University degree in computer science, mathematics, statistics, Chemical engineering, materials engineering, 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. - Experience with named entity recognition and entity linking methodologies. Additional Knowledge and Professional Experience - Strong programming skills in at least one of the following languages: Python, C++, Scala, R, Java. - Experience and skills related to bash, Docker, Kubernetes, Unity testing, Collab Competences - Interest in biomaterial sciences, biomedicine and related application domains - 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 *Conditions* - The position will be located at BSC within the Life Sciences Department - We offer a full-time contract, a good working environment, a highly stimulating environment with state-of-the-art infrastructure, flexible working hours, extensive training plan, tickets restaurant, private health insurance, fully support to the relocation procedures - Duration: Open-ended contract due to technical and scientific activities linked to the project and budget duration - Salary: we offer a competitive salary commensurate with the qualifications and experience of the candidate and according to the cost of living in Barcelona - Starting date: ASAP *About BSC* The Barcelona Supercomputing Center - Centro Nacional de Supercomputaci?n (BSC-CNS) is the leading supercomputing center in Spain. It houses MareNostrum, one of the most powerful supercomputers in Europe, and is a hosting member of the PRACE European distributed supercomputing infrastructure. The mission of BSC is to research, develop and manage information technologies in order to facilitate scientific progress. BSC combines HPC service provision and R&D into both computer and computational science (life, earth and engineering sciences) under one roof, and currently has over 770 staff from 55 countries. -- ======================================= Martin Krallinger, Dr. Head of Biological Text Mining Unit Barcelona Supercomputing Center (BSC-CNS) https://www.linkedin.com/in/martin-krallinger-85495920/ ======================================= -------------- next part -------------- An HTML attachment was scrubbed... URL: From tgd at oregonstate.edu Fri Jul 15 04:01:05 2022 From: tgd at oregonstate.edu (Dietterich, Thomas) Date: Fri, 15 Jul 2022 08:01:05 +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 Danko, In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. --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: Connectionists On Behalf Of Danko Nikolic Sent: Thursday, July 14, 2022 09:17 To: Grossberg, Stephen 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.] Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 ASIM.ROY at asu.edu Thu Jul 14 17:48:05 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Thu, 14 Jul 2022 21:48:05 +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 Steve Grossberg is generally correct. ART, RBF nets and similar flexible architectures can scale up almost linearly to problem size. My conjecture is that architectures that use distributed representation run into the scaling issue. On the other hand, however, distributed representation produces a more compact architecture compared to the shallow architectures of ART and RBFs. But, in terms of adding concepts, it?s easy to add a new object or concept to the shallow architectures. Perhaps someone can provide more insights on the architectural differences and the corresponding pros and cons of each. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Grossberg, Stephen Sent: Thursday, July 14, 2022 10:14 AM To: Danko Nikolic Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, I will respond to your comment below to the entire list. I recommend that future interactions be done between just the two of us. Everything that I write below is summarized in my new book: In brief, within Adaptive Resonance Theory, or ART, there IS no problem of scaling, in the sense that you describe it below, from learning (say) to correctly recognizing 100 objects to doing the same for 200. In fact, one of the properties for which I introduced ART in 1976 was to enable incremental learning in real time of arbitrary numbers of objects or events without experiencing catastrophic forgetting. I have called this a solution of the stability-plasticity dilemma: How our brains, and models like ART, can rapidly learn (plasticity) arbitrary numbers of objects without experiencing catastrophic forgetting (stability). I also derived ART in 1980, in an article within Psychological Review, from a thought experiment about how ANY system can AUTONOMOUSLY correct predictive errors in a changing world that is filled with unexpected events. This thought is experiment was derived from a few facts that are familiar to us all. They are familiar because they represent ubiquitous environmental pressures that we all experience. The thought experiment clarifies why, when they act together during the evolutionary process, models like ART are the unique result. Moveover, at no point during the thought experiment are the words mind or brain mentioned. ART is thus a universal solution of this learning, classification, and prediction problem. You write below about "connectionist systems". ART is a connectionist system. What do you mean by a "connectionist system"? What you write below about them is not true in general. Best, Steve ________________________________ From: Danko Nikolic > Sent: Thursday, July 14, 2022 12:16 PM To: Grossberg, Stephen > Cc: Gary Marcus >; connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 steve at bu.edu Thu Jul 14 19:54:38 2022 From: steve at bu.edu (Grossberg, Stephen) Date: Thu, 14 Jul 2022 23:54:38 +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 Asim, Variants of ART, such as distributed ARTMAP, have more compact representations than winner-take-all ART examples. Please keep in mind the distinction between ART as a cognitive and neural theory, whose foundational hypotheses now have copious and compelling empirical and conceptual support, and ART as a series of algorithms with ever more powerful computational capabilities; e.g., ART 1, 2, 3A, ARTMAP, distributed ARTMAP, gaussian ARTMAP, etc. Each of the algorithms is supported by a combination of mathematical theorems and/or extensive parametric simulations and comparative benchmark studies. The two activities are mutually reinforcing and energizing, but not the same. I will reply to any further comments and/or questions about this topic in one-to-one emails. Best, Steve ________________________________ From: Asim Roy Sent: Thursday, July 14, 2022 5:48 PM To: Grossberg, Stephen ; Danko Nikolic ; Gary Marcus Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: RE: Connectionists: Stephen Hanson in conversation with Geoff Hinton I think Steve Grossberg is generally correct. ART, RBF nets and similar flexible architectures can scale up almost linearly to problem size. My conjecture is that architectures that use distributed representation run into the scaling issue. On the other hand, however, distributed representation produces a more compact architecture compared to the shallow architectures of ART and RBFs. But, in terms of adding concepts, it?s easy to add a new object or concept to the shallow architectures. Perhaps someone can provide more insights on the architectural differences and the corresponding pros and cons of each. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Grossberg, Stephen Sent: Thursday, July 14, 2022 10:14 AM To: Danko Nikolic Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, I will respond to your comment below to the entire list. I recommend that future interactions be done between just the two of us. Everything that I write below is summarized in my new book: In brief, within Adaptive Resonance Theory, or ART, there IS no problem of scaling, in the sense that you describe it below, from learning (say) to correctly recognizing 100 objects to doing the same for 200. In fact, one of the properties for which I introduced ART in 1976 was to enable incremental learning in real time of arbitrary numbers of objects or events without experiencing catastrophic forgetting. I have called this a solution of the stability-plasticity dilemma: How our brains, and models like ART, can rapidly learn (plasticity) arbitrary numbers of objects without experiencing catastrophic forgetting (stability). I also derived ART in 1980, in an article within Psychological Review, from a thought experiment about how ANY system can AUTONOMOUSLY correct predictive errors in a changing world that is filled with unexpected events. This thought is experiment was derived from a few facts that are familiar to us all. They are familiar because they represent ubiquitous environmental pressures that we all experience. The thought experiment clarifies why, when they act together during the evolutionary process, models like ART are the unique result. Moveover, at no point during the thought experiment are the words mind or brain mentioned. ART is thus a universal solution of this learning, classification, and prediction problem. You write below about "connectionist systems". ART is a connectionist system. What do you mean by a "connectionist system"? What you write below about them is not true in general. Best, Steve ________________________________ From: Danko Nikolic > Sent: Thursday, July 14, 2022 12:16 PM To: Grossberg, Stephen > Cc: Gary Marcus >; connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 danko.nikolic at gmail.com Fri Jul 15 03:18:31 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Fri, 15 Jul 2022 09:18:31 +0200 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 agree about the potential for linear scaling of ART and.other connectionist systems. However, there are two problems. The problem number one kills it already and this is that the real brain scales a lot better than linearly: For each new object learned, we are able to resolve countless many new situations in which this object takes part (e.g., finding various uses for a knife, many of which may be new, ad hoc -- this is a great ability of biological minds often referred to as 'understanding'). Hence, simple linear scaling by adding more neurons for additional objects is not good enough to match biological intelligence. The second problem becomes an overkill, and this is that linear scaling in connectionist systems works only in theory, under idealized conditions. In real life, say if working with ImageNet, the scaling turns into a power-law with an exponent much larger than one: We need something like 500x more resources just to double the number of objects. Hence, in practice, the demands for resources explode if you want to add more categories whilst not losing the accuracy. To summarize, there is no linear scaling in practice nor would linear scaling suffice, even if we found one. This should be a strong enough argument to search for another paradigm, something that scales better than connectionism. I discuss both problems in the new manuscript, and even track a bit deeper the problem of why connectionism lacks linear scaling in practice (I provide some revealing computations in the Supplementary Materials (with access to the code), although much more work needs to be done). Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Thu, Jul 14, 2022 at 11:48 PM Asim Roy wrote: > I think Steve Grossberg is generally correct. ART, RBF nets and similar > flexible architectures can scale up almost linearly to problem size. My > conjecture is that architectures that use distributed representation run > into the scaling issue. On the other hand, however, distributed > representation produces a more compact architecture compared to the shallow > architectures of ART and RBFs. But, in terms of adding concepts, it?s easy > to add a new object or concept to the shallow architectures. > > > > Perhaps someone can provide more insights on the architectural differences > and the corresponding pros and cons of each. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Grossberg, Stephen > *Sent:* Thursday, July 14, 2022 10:14 AM > *To:* Danko Nikolic > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Danko, > > > > I will respond to your comment below to the entire list. I recommend that > future interactions be done between just the two of us. > > > > Everything that I write below is summarized in my new book: > > > > In brief, within Adaptive Resonance Theory, or ART, there IS no problem of > scaling, in the sense that you describe it below, from learning (say) to > correctly recognizing 100 objects to doing the same for 200. In fact, one > of the properties for which I introduced ART in 1976 was to enable > incremental learning in real time of arbitrary numbers of objects or events > without experiencing catastrophic forgetting. > > > > I have called this a solution of the *stability-plasticity dilemma*: How > our brains, and models like ART, can rapidly learn (plasticity) arbitrary > numbers of objects without experiencing catastrophic forgetting (stability). > > > > I also derived ART in 1980, in an article within Psychological Review, > from a *thought experiment* about how ANY system can AUTONOMOUSLY correct > predictive errors in a changing world that is filled with unexpected events. > > > > This thought is experiment was derived from a few facts that are familiar > to us all. They are familiar because they represent ubiquitous > environmental pressures that we all experience. The thought experiment > clarifies why, when they act together during the evolutionary process, > models like ART are the unique result. > > > > Moveover, at no point during the thought experiment are the words mind or > brain mentioned. ART is thus a *universal solution* of this learning, > classification, and prediction problem. > > > > You write below about "connectionist systems". ART is a connectionist > system. > > > > What do you mean by a "connectionist system"? What you write below about > them is not true in general. > > > > Best, > > > > Steve > ------------------------------ > > *From:* Danko Nikolic > *Sent:* Thursday, July 14, 2022 12:16 PM > *To:* Grossberg, Stephen > *Cc:* Gary Marcus ; > connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Steve, > > > > Thank you very much for your message and for the greetings. I will pass > them on if an occasion arises. > > > > Regarding your question: The key problem I am trying to address and that, > to the best of my knowledge, no connectionist system was able to solve so > far is that of scaling the system's intelligence. For example, if the > system is able to correctly recognize 100 different objects, how many > additional resources are needed to double that to 200? All the empirical > data show that connectionist systems scale poorly: Some of the best systems > we have require 500x more resources in order to increase the intelligence > by only 2x. I document this problem in the manuscript and even run some > simulations to show that the worst performance is if connectionist systems > need to solve a generalized XOR problem. > > > > In contrast, the biological brain scales well. This I also quantify in the > paper. > > > > I will look at the publication that you mentioned. However, so far, I > haven't seen a solution that scales well in intelligence. > > > > My argument is that transient selection of subnetworks by the help of the > mentioned proteins is how intelligence scaling is achieved in biological > brains. > > > > In short, intelligence scaling is the key problem that concerns me. I > describe the intelligence scaling problem in more detail in this book that > just came out a few weeks ago and that is written for practitioners in Data > Scientist and AI: https://amzn.to/3IBxUpL > > > > > I hope that this at least partly answers where I see the problems and what > I am trying to solve. > > > > Greetings from Germany, > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: > > Dear Danko, > > > > I have just read your new article and would like to comment briefly about > it. > > > > In your introductory remarks, you write: > > > > "However, connectionism did not yet produce a satisfactory explanation of > how the mental emerges from the physical. A number of open problems > remains ( 5,6,7,8). As a result, the explanatory gap between the mind and > the brain remains wide open." > > > > I certainly believe that no theoretical explanation in science is ever > complete. However, I also believe that "the explanatory gap between the > mind and the brain" does not remain "wide open". > > > > My Magnum Opus, that was published in 2021, makes that belief clear in its > title: > > > > *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* > > > > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > > > > The book provides a self-contained and non-technical exposition in a > conversational tone of many principled and unifying explanations of > psychological and neurobiological data. > > > > In particular, it explains roles for the metabotropic glutamate receptors > that you mention in your own work. See the text and figures around p. 521. > This explanation unifies psychological, anatomical, neurophysiological, > biophysical, and biochemical data about the processes under discussion. > > > > I have a very old-fashioned view about how to understand scientific > theories. I get excited by theories that explain and predict more data than > previous theories. > > > > Which of the data that I explain in my book, and support with quantitative > computer simulations, can you also explain? > > > > What data can you explain, in the same quantitative sense, that you do not > think the neural models in my book can explain? > > > > I would be delighted to discuss these issues further with you. > > > > If you are in touch with my old friend and esteemed colleague, Wolf > Singer, please send him my warm regards. I cite the superb work that he and > various of his collaborators have done in many places in my book. > > > > Best, > > > > Steve > > > > 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 Danko Nikolic > *Sent:* Thursday, July 14, 2022 6:05 AM > *To:* Gary Marcus > *Cc:* connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Gary and everyone, > > > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > > > I am happy to announce that I have now finished a draft of a paper in > which I propose how the brain is able to achieve that. The manuscript > requires a bit of patience for two reasons: one is that the reader may be > exposed for the first time to certain aspects of brain physiology. The > second reason is that it may take some effort to understand the > counterintuitive implications of the new ideas (this requires a different > way of thinking than what we are used to based on connectionism). > > > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > > > (I apologize for the neuroscience lingo, which I tried to minimize.) > > > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > > > Thanks. > > > > Danko > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > > > > > On Thu, Feb 3, 2022 at 5:25 PM 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 danko.nikolic at gmail.com Fri Jul 15 04:51:09 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Fri, 15 Jul 2022 10:51:09 +0200 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 Thomas, Thank you for reading the paper and for the comments. I cite: "In my experience, supervised classification scales linearly in the number of classes." This would be good to quantify as a plot. Maybe a research paper would be a good idea. The reason is that it seems that everyone else who tried to quantify that relation found a power law. At this point, it would be surprising to find a linear relationship. And it would probably make a well read paper. But please do not forget that my argument states that even a linear relationship is not good enough to match bilogical brains. We need something more similar to a power law with exponent zero when it comes to the model size i.e., a constant number of parameters in the model. And we need linear relationship when it comes to learning time: Each newly learned object should needs about as much of learning effort as was needed for each previous object. I cite: "The real world is not dominated by generalized XOR problems." Agreed. And it is good so because generalize XOR scales worse than power law. It scales exponentially! This a more agressive form of explosion than power law. Importantly, a generalized AND operation also scales exponentially (with a smaller exponent, though). I guess we would agree that the real world probably encouners a lot of AND problems. The only logical operaiton that could be learned with a linear increase in the number of parameters was a generalized OR. Finally, I foiund that a mixure of AND and OR resulted in a power law-like scaling of the number of parameters. So, a mixture of AND and OR seemed to scale as good (or as bad) as the real world. I have put this information into Supplementary Materials. The conclusion that I derived from those analyses is: connectionism is not sustainable to reach human (or animal) levels of intelligence. Therefore, I hunted for an alternative pradigm. Greetings, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Fri, Jul 15, 2022 at 10:01 AM Dietterich, Thomas wrote: > Dear Danko, > > > > In my experience, supervised classification scales linearly in the number > of classes. Of course it depends to some extent on how subtle the > distinctions are between the different categories. The real world is not > dominated by generalized XOR problems. > > > > --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:* Connectionists *On > Behalf Of *Danko Nikolic > *Sent:* Thursday, July 14, 2022 09:17 > *To:* Grossberg, Stephen > *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.] > > Dear Steve, > > > > Thank you very much for your message and for the greetings. I will pass > them on if an occasion arises. > > > > Regarding your question: The key problem I am trying to address and that, > to the best of my knowledge, no connectionist system was able to solve so > far is that of scaling the system's intelligence. For example, if the > system is able to correctly recognize 100 different objects, how many > additional resources are needed to double that to 200? All the empirical > data show that connectionist systems scale poorly: Some of the best systems > we have require 500x more resources in order to increase the intelligence > by only 2x. I document this problem in the manuscript and even run some > simulations to show that the worst performance is if connectionist systems > need to solve a generalized XOR problem. > > > > In contrast, the biological brain scales well. This I also quantify in the > paper. > > > > I will look at the publication that you mentioned. However, so far, I > haven't seen a solution that scales well in intelligence. > > > > My argument is that transient selection of subnetworks by the help of the > mentioned proteins is how intelligence scaling is achieved in biological > brains. > > > > In short, intelligence scaling is the key problem that concerns me. I > describe the intelligence scaling problem in more detail in this book that > just came out a few weeks ago and that is written for practitioners in Data > Scientist and AI: https://amzn.to/3IBxUpL > > > > > I hope that this at least partly answers where I see the problems and what > I am trying to solve. > > > > Greetings from Germany, > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: > > Dear Danko, > > > > I have just read your new article and would like to comment briefly about > it. > > > > In your introductory remarks, you write: > > > > "However, connectionism did not yet produce a satisfactory explanation of > how the mental emerges from the physical. A number of open problems > remains ( 5,6,7,8). As a result, the explanatory gap between the mind and > the brain remains wide open." > > > > I certainly believe that no theoretical explanation in science is ever > complete. However, I also believe that "the explanatory gap between the > mind and the brain" does not remain "wide open". > > > > My Magnum Opus, that was published in 2021, makes that belief clear in its > title: > > > > *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* > > > > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > > > > The book provides a self-contained and non-technical exposition in a > conversational tone of many principled and unifying explanations of > psychological and neurobiological data. > > > > In particular, it explains roles for the metabotropic glutamate receptors > that you mention in your own work. See the text and figures around p. 521. > This explanation unifies psychological, anatomical, neurophysiological, > biophysical, and biochemical data about the processes under discussion. > > > > I have a very old-fashioned view about how to understand scientific > theories. I get excited by theories that explain and predict more data than > previous theories. > > > > Which of the data that I explain in my book, and support with quantitative > computer simulations, can you also explain? > > > > What data can you explain, in the same quantitative sense, that you do not > think the neural models in my book can explain? > > > > I would be delighted to discuss these issues further with you. > > > > If you are in touch with my old friend and esteemed colleague, Wolf > Singer, please send him my warm regards. I cite the superb work that he and > various of his collaborators have done in many places in my book. > > > > Best, > > > > Steve > > > > 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 Danko Nikolic > *Sent:* Thursday, July 14, 2022 6:05 AM > *To:* Gary Marcus > *Cc:* connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Gary and everyone, > > > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > > > I am happy to announce that I have now finished a draft of a paper in > which I propose how the brain is able to achieve that. The manuscript > requires a bit of patience for two reasons: one is that the reader may be > exposed for the first time to certain aspects of brain physiology. The > second reason is that it may take some effort to understand the > counterintuitive implications of the new ideas (this requires a different > way of thinking than what we are used to based on connectionism). > > > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > > > (I apologize for the neuroscience lingo, which I tried to minimize.) > > > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > > > Thanks. > > > > Danko > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > > > > > On Thu, Feb 3, 2022 at 5:25 PM 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.marcus at nyu.edu Fri Jul 15 09:51:37 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Fri, 15 Jul 2022 06:51:37 -0700 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <615FCABF-1610-49FF-9468-6BF8F6512585@nyu.edu> I am with Danko here: he said ?resolve? not ?anticipate in advance?. I doubt any human is perfect in anticipating all uses of a knife but eg audiences had little trouble interpreting and enjoying all the weird repurposings that the TV character Macgyver was known for. On Jul 15, 2022, at 6:36 AM, Asim Roy wrote: > > ? > Dear Danko, > > I am not sure if I myself know all the uses of a knife, leave aside countless ones. Given a particular situation, I might simulate in my mind about the potential usage, but I doubt our minds explore all the countless situations of usage of an object as soon as it learns about it. > I am not sure if a 2 or 3 year old child, after having ?learnt? about a knife, knows very many uses of it. I doubt the kid is awake all night and day simulating in the brain how and where to use such a knife. > ?Understanding? is a loaded term. I think it needs a definition. > I am copying Max Versace, a student of Steve Grossberg. His company markets a software that can learn quickly from a few examples. Not exactly one-shot learning, it needs a few shots. I believe it?s a variation of ART. But Max can clarify the details. And Tsvi is doing similar work. So, what you are asking for may already exist. So linear scaling may be the worst case scenario. > > Best, > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > > From: Danko Nikolic > Sent: Friday, July 15, 2022 12:19 AM > To: Asim Roy > Cc: Grossberg, Stephen ; Gary Marcus ; AIhub ; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Asim, > > I agree about the potential for linear scaling of ART and.other connectionist systems. However, there are two problems. > > The problem number one kills it already and this is that the real brain scales a lot better than linearly: For each new object learned, we are able to resolve countless many new situations in which this object takes part (e.g., finding various uses for a knife, many of which may be new, ad hoc -- this is a great ability of biological minds often referred to as 'understanding'). Hence, simple linear scaling by adding more neurons for additional objects is not good enough to match biological intelligence. > > The second problem becomes an overkill, and this is that linear scaling in connectionist systems works only in theory, under idealized conditions. In real life, say if working with ImageNet, the scaling turns into a power-law with an exponent much larger than one: We need something like 500x more resources just to double the number of objects. Hence, in practice, the demands for resources explode if you want to add more categories whilst not losing the accuracy. > > To summarize, there is no linear scaling in practice nor would linear scaling suffice, even if we found one. > > This should be a strong enough argument to search for another paradigm, something that scales better than connectionism. > > I discuss both problems in the new manuscript, and even track a bit deeper the problem of why connectionism lacks linear scaling in practice (I provide some revealing computations in the Supplementary Materials (with access to the code), although much more work needs to be done). > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > -- I wonder, how is the brain able to generate insight? -- > > > On Thu, Jul 14, 2022 at 11:48 PM Asim Roy wrote: > I think Steve Grossberg is generally correct. ART, RBF nets and similar flexible architectures can scale up almost linearly to problem size. My conjecture is that architectures that use distributed representation run into the scaling issue. On the other hand, however, distributed representation produces a more compact architecture compared to the shallow architectures of ART and RBFs. But, in terms of adding concepts, it?s easy to add a new object or concept to the shallow architectures. > > Perhaps someone can provide more insights on the architectural differences and the corresponding pros and cons of each. > > Best, > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim Roy > Asim Roy | iSearch (asu.edu) > > > From: Connectionists On Behalf Of Grossberg, Stephen > Sent: Thursday, July 14, 2022 10:14 AM > To: Danko Nikolic > Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Danko, > > I will respond to your comment below to the entire list. I recommend that future interactions be done between just the two of us. > > Everything that I write below is summarized in my new book: > > In brief, within Adaptive Resonance Theory, or ART, there IS no problem of scaling, in the sense that you describe it below, from learning (say) to correctly recognizing 100 objects to doing the same for 200. In fact, one of the properties for which I introduced ART in 1976 was to enable incremental learning in real time of arbitrary numbers of objects or events without experiencing catastrophic forgetting. > > I have called this a solution of the stability-plasticity dilemma: How our brains, and models like ART, can rapidly learn (plasticity) arbitrary numbers of objects without experiencing catastrophic forgetting (stability). > > I also derived ART in 1980, in an article within Psychological Review, from a thought experiment about how ANY system can AUTONOMOUSLY correct predictive errors in a changing world that is filled with unexpected events. > > This thought is experiment was derived from a few facts that are familiar to us all. They are familiar because they represent ubiquitous environmental pressures that we all experience. The thought experiment clarifies why, when they act together during the evolutionary process, models like ART are the unique result. > > Moveover, at no point during the thought experiment are the words mind or brain mentioned. ART is thus a universal solution of this learning, classification, and prediction problem. > > You write below about "connectionist systems". ART is a connectionist system. > > What do you mean by a "connectionist system"? What you write below about them is not true in general. > > Best, > > Steve > From: Danko Nikolic > Sent: Thursday, July 14, 2022 12:16 PM > To: Grossberg, Stephen > Cc: Gary Marcus ; connectionists at mailman.srv.cs.cmu.edu ; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Steve, > > Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. > > Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. > > In contrast, the biological brain scales well. This I also quantify in the paper. > > I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. > > My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. > > In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL > > I hope that this at least partly answers where I see the problems and what I am trying to solve. > > Greetings from Germany, > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > --- A progress usually starts with an insight --- > > > On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: > Dear Danko, > > I have just read your new article and would like to comment briefly about it. > > In your introductory remarks, you write: > > "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." > > I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". > > My Magnum Opus, that was published in 2021, makes that belief clear in its title: > > Conscious Mind, Resonant Brain: How Each Brain Makes a Mind > > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. > > In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. > > I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. > > Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? > > What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? > > I would be delighted to discuss these issues further with you. > > If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. > > Best, > > Steve > > 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM > To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu ; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Gary and everyone, > > I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. > > As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. > > I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). > > In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. > > I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. > > The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug > (I apologize for the neuroscience lingo, which I tried to minimize.) > > It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? > > Thanks. > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > > > On Thu, Feb 3, 2022 at 5:25 PM 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 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 junfeng989 at gmail.com Fri Jul 15 20:00:28 2022 From: junfeng989 at gmail.com (Jun Feng) Date: Fri, 15 Jul 2022 20:00:28 -0400 Subject: Connectionists: The 2022 IEEE International Conference on Privacy Computing (PriComp 2022) Message-ID: CFP: The 2022 IEEE International Conference on Privacy Computing (IEEE PriComp 2022) Dec. 15-18, Haikou, China [Submission Deadline: Sep. 1] http://www.ieee-smart-world.org/2022/pricomp/ PriComp 2022 is the 8th in this series of conferences started in 2015 that are devoted to algorithms and architectures for Privacy Computing. PriComp conference provides a forum for academics and practitioners from countries around the world to exchange ideas for improving the efficiency, performance, reliability, security and interoperability of Privacy Computing systems and applications. Following the traditions of the previous successful PriComp conferences held in Fuzhou, China (2015); Qingdao, China (2016); Melbourne, Australia (2017); Boppard, Germany (2018); Canterbury, UK (2019); Hainan, China (2020) and Xi'an, Shanghai, China (online, 2021); PriComp 2022 will be held in Haikou, China. PriComp 2022 will focus on an evolving pathway from privacy protection to privacy computing, by serving as an international premier forum for engineers and scientists in academia, industry, and government to address the resulting profound challenges and to present and discuss their new ideas, research results, applications and experience on all aspects of privacy computing. The conference of PriComp 2022 is co-organized by Chinese Information Processing Society of China, Institute of Information Engineering, CAS, and Hainan University. ================== Important Dates ================== Workshop Proposal: July 15, 2022 Paper Submission: September 01, 2022 Author Notification: October 01, 2022 Camera-Ready Submission: October 31, 2022 Conference Date: December 15-18, 2022 ================== Topics of interest include, but are not limited to ================== - Theories and foundations for privacy computing - Programming languages and compilers for privacy computing - Privacy computing models - Privacy metrics and formalization - Privacy taxonomies and ontologies - Privacy information management and engineering - Privacy operation and modeling - Data utility and privacy loss - Cryptography for privacy protection - Privacy protection based information hiding and sharing - Data analytics oriented privacy control and protection - Privacy-aware information collection - Privacy sensing and distribution - Combined and comprehensive privacy protection - Privacy-preserving data publishing - Private information storage - Private integration and synergy - Private information exchange and sharing - Privacy inference and reasoning - Internet and web privacy - Cloud privacy - Social media privacy - Mobile privacy - Location privacy - IoT privacy - Behavioral advertising - Privacy in large ecosystems such as smart cities - Privacy of AI models and systems - AI for privacy computing - Privacy and blockchain - User-centric privacy protection solutions - Human factors in privacy computing - Privacy nudging - Automated solutions for privacy policies and notices - Legal issues in privacy computing and other interdisciplinary topics ================== Paper Submission ================== All papers need to be submitted electronically through the conference submission website (https://edas.info/N29960) with PDF format. The materials presented in the papers should not be published or under submission elsewhere. Each paper is limited to 8 pages (or 10 pages with over length charge) including figures and references using IEEE Computer Society Proceedings Manuscripts style (two columns, single-spaced, 10 fonts). You can confirm the IEEE Computer Society Proceedings Author Guidelines at the following web page: http://www.computer.org/web/cs-cps/ Manuscript Templates for Conference Proceedings can be found at: https://www.ieee.org/conferences_events/conferences/publishing/templates.html Once accepted, the paper will be included into the IEEE conference proceedings published by IEEE Computer Society Press (indexed by EI). At least one of the authors of any accepted paper is requested to register the paper at the conference. ================== Special Issues ================== All accepted papers will be submitted to IEEE Xplore and Engineering Index (EI). Best Paper Awards will be presented to high quality papers. Distinguished papers, after further revisions, will be published in SCI & EI indexed prestigious journals. 1. Special issue on ?Dark side of the Socio-Cyber World: Media Manipulation, Fake News, and Misinformation?, IEEE Transactions on Computational Social Systems http://cyber-science.org/2022/assets/files/si/220412_IEEE%20TCSS_SI.pdf 2. Special issue on ?Responsible AI in Social Computing?, IEEE Transactions on Computational Social Systems https://www.ieeesmc.org/images/publications/pdfs/Call_for_Paper_-_SI_on_Responsible_AI_in_Social_Computing.pdf 3. Special issue on ?Decentralized Trust Management with Intelligence?, Information Science https://www.journals.elsevier.com/information-sciences/call-for-papers/decentralized-trust-management-with-intelligence 4. Special issue on ?Resource Sustainable Computational and Artificial Intelligence?, IEEE Transactions on Emerging Topics in Computational Intelligence 5. Special issue on ?Smart Blockchain for IoT Trust, Security and Privacy?, IEEE IoT Journal https://ieee-iotj.org/wp-content/uploads/2022/05/IEEEIoT-SmartBlockchain-TSP.pdf 6. Special issue on ?Edge Computing Optimization and Security?, Journal of Systems Architecture https://www.journals.elsevier.com/journal-of-systems-architecture/call-for-papers/edge-computing-optimization-and-security-vsi-edgeos2022 7. Special issue on ?Distributed Learning and Blockchain Enabled Infrastructures for Next Generation of Big Data Driven Cyber-Physical Systems?, Journal of Systems Architecture http://cyber-science.org/2022/assets/files/si/JSA_SI_0331.pdf 8. Special issue on ?Distributed and Collaborative Learning Empowered Edge Intelligence in Smart City?, ACM Transactions on Sensor Networks https://dl.acm.org/pb-assets/static_journal_pages/tosn/pdf/ACM_TOSN_CFP1210-1640635690003.pdf 9. Special issue on ?Robustness, Privacy, and Forensics in Intelligent Multimedia Systems? Information Science https://www.journals.elsevier.com/information-sciences/forthcoming-special-issues/robustness-privacy-and-forensics-in-intelligent-multimedia-systems * More special issues will be added later. http://www.ieee-smart-world.org/2022/pricomp/si.php ================== Organizing Committee ================== General Chairs - Fenghua Li, Institute of Information Engineering, CAS, China - Laurence T. Yang, Hainan University, China - Willy Susilo, University of Wollongong, Australia Program Chairs - Hui Li, Xidian University, China - Mamoun Alazab, Charles Darwin University, Australia - Jun Feng, Huazhong University of Science and Technology, China Local Chairs - Weidong Qiu, Shanghai Jiaotong University, China - Jieren Cheng, Hainan University, China Publicity Chairs - Bocheng Ren, Huazhong University of Science and Technology, China - Xin Nie, Huazhong University of Science and Technology, China - Peng Tang, Shanghai Jiao Tong University, China -- Dr. Jun Feng Huazhong University of Science and Technology Mobile: +86-18827365073 WeChat: junfeng10001000 E-Mail: junfeng989 at gmail.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From steve at bu.edu Fri Jul 15 11:19:23 2022 From: steve at bu.edu (Grossberg, Stephen) Date: Fri, 15 Jul 2022 15:19:23 +0000 Subject: Connectionists: Neural architectures that embody biological intelligence In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Asim and Danko, A lot of your concerns about scaling do not apply to the kinds of biological neural networks that my colleagues and I have developed over the years. You can find a self-contained summary of many of them in my Magnum Opus: https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 As Asim notes below, it is indeed the case that ART can often make good predictions based on small amounts of learned data. This applies as well to large-scale applications naturalistic data. Gail Carpenter and her colleagues have, for example, shown how this works in learning complicated maps of multiple vegetation classes during remote sensing; e.g. Carpenter, G.A., Gopal, S., Macomber, S., Martens, S., & Woodcock, C.E. (1999). A neural network method for mixture estimation for vegetation mapping. Remote Sensing of Environment, 70(2), 138-152. http://techlab.bu.edu/members/gail/articles/127_Mixtures_RSE_1999_.pdf or in learning medical database predictions in response to incomplete, probabilistic, and even incorrect data. In this regard, Gail et al. have also shown how an ART system can incrementally learn a cognitive hierarchy of rules whereby to understand such data; i.e., converts information into knowledge; e.g., Carpenter, G.A., & Ravindran, A. (2008). Unifying multiple knowledge domains using the ARTMAP information fusion system. Proceedings of the 11th International Conference on Information Fusion, Cologne, Germany, June 30 - July 3, 2008. http://techlab.bu.edu/members/gail/articles/155_Fusion2008_CarpenterRavindran.pdf Carpenter, G.A. & Markuzon, N. (1998). ARTMAP?IC and medical diagnosis: Instance counting and inconsistent cases. Neural Networks, 11(2), 323-336. http://techlab.bu.edu/members/gail/articles/117_ARTMAP-IC_1998_.pdf My own work is filled with models that incrementally learn to carry out goal-oriented tasks without regard to scaling concerns. This work develops neural architectures that involve the coordinated actions of many brain regions, not just learned classification. These architectures are supported by unified and principled explanations of lots of psychological and neurobiological data; e.g., Chang, H.-C., Grossberg, S., and Cao, Y. (2014) Where?s Waldo? How perceptual cognitive, and emotional brain processes cooperate during learning to categorize and find desired objects in a cluttered scene. Frontiers in Integrative Neuroscience, doi: 10.3389/fnint.2014.0043, https://www.frontiersin.org/articles/10.3389/fnint.2014.00043/full Grossberg, S., and Vladusich, T. (2010). How do children learn to follow gaze, share joint attention, imitate their teachers, and use tools during social interactions? Neural Networks, 23, 940-965. https://sites.bu.edu/steveg/files/2016/06/GrossbergVladusichNN2010.pdf See Figure 1 in the following article to get a sense of how many brain processes other than classification are needed to realize true biological intelligence: Grossberg, S. (2018). Desirability, availability, credit assignment, category learning, and attention: Cognitive-emotional and working memory dynamics of orbitofrontal, ventrolateral, and dorsolateral prefrontal cortices. Brain and Neuroscience Advances, May 8, 2018. https://journals.sagepub.com/doi/full/10.1177/2398212818772179 Best, Steve ________________________________ From: Asim Roy Sent: Friday, July 15, 2022 9:35 AM To: Danko Nikolic Cc: Grossberg, Stephen ; Gary Marcus ; AIhub ; connectionists at mailman.srv.cs.cmu.edu ; 'maxversace at gmail.com' Subject: RE: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, 1. I am not sure if I myself know all the uses of a knife, leave aside countless ones. Given a particular situation, I might simulate in my mind about the potential usage, but I doubt our minds explore all the countless situations of usage of an object as soon as it learns about it. 2. I am not sure if a 2 or 3 year old child, after having ?learnt? about a knife, knows very many uses of it. I doubt the kid is awake all night and day simulating in the brain how and where to use such a knife. 3. ?Understanding? is a loaded term. I think it needs a definition. 4. I am copying Max Versace, a student of Steve Grossberg. His company markets a software that can learn quickly from a few examples. Not exactly one-shot learning, it needs a few shots. I believe it?s a variation of ART. But Max can clarify the details. And Tsvi is doing similar work. So, what you are asking for may already exist. So linear scaling may be the worst case scenario. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Danko Nikolic Sent: Friday, July 15, 2022 12:19 AM To: Asim Roy Cc: Grossberg, Stephen ; Gary Marcus ; AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I agree about the potential for linear scaling of ART and.other connectionist systems. However, there are two problems. The problem number one kills it already and this is that the real brain scales a lot better than linearly: For each new object learned, we are able to resolve countless many new situations in which this object takes part (e.g., finding various uses for a knife, many of which may be new, ad hoc -- this is a great ability of biological minds often referred to as 'understanding'). Hence, simple linear scaling by adding more neurons for additional objects is not good enough to match biological intelligence. The second problem becomes an overkill, and this is that linear scaling in connectionist systems works only in theory, under idealized conditions. In real life, say if working with ImageNet, the scaling turns into a power-law with an exponent much larger than one: We need something like 500x more resources just to double the number of objects. Hence, in practice, the demands for resources explode if you want to add more categories whilst not losing the accuracy. To summarize, there is no linear scaling in practice nor would linear scaling suffice, even if we found one. This should be a strong enough argument to search for another paradigm, something that scales better than connectionism. I discuss both problems in the new manuscript, and even track a bit deeper the problem of why connectionism lacks linear scaling in practice (I provide some revealing computations in the Supplementary Materials (with access to the code), although much more work needs to be done). Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Thu, Jul 14, 2022 at 11:48 PM Asim Roy > wrote: I think Steve Grossberg is generally correct. ART, RBF nets and similar flexible architectures can scale up almost linearly to problem size. My conjecture is that architectures that use distributed representation run into the scaling issue. On the other hand, however, distributed representation produces a more compact architecture compared to the shallow architectures of ART and RBFs. But, in terms of adding concepts, it?s easy to add a new object or concept to the shallow architectures. Perhaps someone can provide more insights on the architectural differences and the corresponding pros and cons of each. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Grossberg, Stephen Sent: Thursday, July 14, 2022 10:14 AM To: Danko Nikolic > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, I will respond to your comment below to the entire list. I recommend that future interactions be done between just the two of us. Everything that I write below is summarized in my new book: In brief, within Adaptive Resonance Theory, or ART, there IS no problem of scaling, in the sense that you describe it below, from learning (say) to correctly recognizing 100 objects to doing the same for 200. In fact, one of the properties for which I introduced ART in 1976 was to enable incremental learning in real time of arbitrary numbers of objects or events without experiencing catastrophic forgetting. I have called this a solution of the stability-plasticity dilemma: How our brains, and models like ART, can rapidly learn (plasticity) arbitrary numbers of objects without experiencing catastrophic forgetting (stability). I also derived ART in 1980, in an article within Psychological Review, from a thought experiment about how ANY system can AUTONOMOUSLY correct predictive errors in a changing world that is filled with unexpected events. This thought is experiment was derived from a few facts that are familiar to us all. They are familiar because they represent ubiquitous environmental pressures that we all experience. The thought experiment clarifies why, when they act together during the evolutionary process, models like ART are the unique result. Moveover, at no point during the thought experiment are the words mind or brain mentioned. ART is thus a universal solution of this learning, classification, and prediction problem. You write below about "connectionist systems". ART is a connectionist system. What do you mean by a "connectionist system"? What you write below about them is not true in general. Best, Steve ________________________________ From: Danko Nikolic > Sent: Thursday, July 14, 2022 12:16 PM To: Grossberg, Stephen > Cc: Gary Marcus >; connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 vishal.pup at gmail.com Fri Jul 15 12:39:02 2022 From: vishal.pup at gmail.com (=?UTF-8?B?VmlzaGFsIEdveWFsKOCkteCkv+CktuCkvuCksiDgpJfgpYvgpK/gpLIp?=) Date: Fri, 15 Jul 2022 22:09:02 +0530 Subject: Connectionists: Emerging Educational technologies for multilingual societies" on 16-07-2022 (Saturday) at 11:00 AM In-Reply-To: References: Message-ID: Centre for E-Learning and Teaching Excellence is going to organize e-talk by Dr. Girish Nath Jha,Chairman, Commission for Scientific and Technical Terminology,Ministry of Education, Government of India on the topic "Emerging Educational technologies for multilingual societies" on 16-07-2022 (Saturday) at 11:00 AM. Lecture will be streamed live on https://www.youtube.com/channel/UCcL-bhN7IgytVIIgep0_4XA E-Certificates will be issued after filling feedback form to be shared at the end of the lecture in Youtube Chat box. -------------- next part -------------- An HTML attachment was scrubbed... URL: From mtanveer at iiti.ac.in Fri Jul 15 06:11:33 2022 From: mtanveer at iiti.ac.in (M Tanveer) Date: Fri, 15 Jul 2022 15:41:33 +0530 Subject: Connectionists: ICONIP 2022 | Call for Papers | Submission Deadline: July 28, 2022 Message-ID: CALL FOR PAPERS *Paper submission deadline extended to July 28, 2022 (Firm deadline)* 29th International Conference on Neural Information Processing (ICONIP 2022) November 22-26, 2022 New Delhi, India https://www.iconip2022.apnns.org/ Dear Colleague, We would like to invite you to submit your paper to the 29th International Conference on Neural Information Processing (ICONIP 2022), New Delhi, India, November 22-26, 2022. Updated flyer of the call for paper is attached with this email. The conference website: https://iconip2022.apnns.org/index.php Submission Page: https://easychair.org/conferences/?conf=iconip2022 Paper Submission: *July 28, 2022* Paper Notification Date: *Sept. 15, 2022* Camera Ready Submission: *Sept. 30, 2022* Thank you so much for your kind support in this conference. Kindly share among your contacts. If you have any queries, please contact us: iconip2022 at gmail.com. Sincerely, General Chairs - ICONIP 2022 ---------------------------------------------------------- Dr. M. Tanveer (General Chair - ICONIP 2022, IEEE CIS SS 2022) Associate Professor and Ramanujan Fellow Department of Mathematics Indian Institute of Technology Indore Email: mtanveer at iiti.ac.in Mobile: +91-9413259268 Homepage: http://iiti.ac.in/people/~mtanveer/ Associate Editor: IEEE TNNLS (IF: 14.25). Action Editor: Neural Networks, Elsevier (IF: 9.65). Associate Editor: Pattern Recognition, Elsevier (IF: 8.52). Associate Editor: Cognitive Computation, Springer (IF: 8.26). Board of Editors: Engineering Applications of AI, Elsevier (IF: 7.80). Associate Editor: Neurocomputing, Elsevier (IF: 5.78). Editorial Board: Applied Soft Computing, Elsevier (IF: 6.72). Associate Editor: International Journal of Machine Learning & Cybernetics (IF: 4.37). -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: CallForPapers-ICONIP 2022v3_final.pdf Type: application/pdf Size: 500290 bytes Desc: not available URL: From tiako at ieee.org Fri Jul 15 17:37:12 2022 From: tiako at ieee.org (Pierre F. Tiako) Date: Fri, 15 Jul 2022 16:37:12 -0500 Subject: Connectionists: [CFP Due Extended Sept 03] CAIS/SERD/CESP2022 Intelligent Systems/Sofware/Process, Oct 3-6, Hybrid, USA Message-ID: [Due to numerous requests (Heatwave, Summer Break, Pandemic, Start of Semester), the paper submission deadline is extended to Sept 03] --- Call for Abstracts and Papers ------------- CAIS/SERD/CESP2022 Intelligent Systems/Sofware/Process Downtown Oklahoma City, OK, USA & Online October 3-6, 2022 OkIP Published & WoS/SCOPUS Indexed Submission Deadline Extended: September 03, 2022 Extended versions of the best papers will be considered for journal publication. >>> Contribution Types (One-Column IEEE Format Style) - Full Paper: Accomplished research results (10 pages) - Short Paper: Work in progress/fresh developments (6 pages) - Extended Abstract/Poster/Journal First: Displayed/Oral presented (3 pages) >>> Automated and Intelligent Systems (CAIS) https://eventutor.com/e/CAIS002 * Areas - AI, Machine Learning (ML), and Applications - Agent-based, Automated, and Distributed Supports - Intelligent Systems and Applications - Knowledge-based and Control Supports - Robotics and Vehicles * Technical Program Committee https://eventutor.com/event/19/page/56-committee >>> Software Engineering Research & Development (SERD) https://eventutor.com/e/SERD002 * Areas: - General and Social Aspects of Software Engineering(SE) - Software Design, Testing, Evolution, and Maintenance - Formal Methods and Theoretical Foundations - SE Service Orientation and Human Interactions - AI in SE, Web-Based Environments, and Adaptive Systems - Emerging SE Technologies and Dependability - SE Distribution, Componentization, and Collaboration * Technical Program Committee https://eventutor.com/event/18/page/54-committee >>> Enterprise and Software Process (CESP) https://eventutor.com/e/CESP002 * Areas: - Agile, Hybrid, and Traditional Process - Model, Method, Standard, and Architecture - AI and Knowledge Management Process - Process, Application, and Tool - Human Factor and Communication Technology * Technical Program Committee https://eventutor.com/event/24/page/66-committee Please feel free to contact us for any inquiries at: info at okipublishing.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Fri Jul 15 18:37:27 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Fri, 15 Jul 2022 22:37:27 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <615FCABF-1610-49FF-9468-6BF8F6512585@nyu.edu> Message-ID: Dear Danko, 1. ?Figure it out once the situation emerges? and ?we do not need to learn? ?upfront? sounds a bit magical, even for biological systems. A professional tennis player practices for years and years so that he/she knows exactly how to respond to each and every situation as much as possible. Such learning does not end with just 10 days of training. Such a player would prefer to know as much as possible ?upfront? the various situations that can arise. That?s the meaning of training and learning. And that also means hitting tennis balls millions of times (countless??) perhaps. And that?s learning from a lot of data. 2. You might want to rethink your definition of ?understanding? given the above example. Understanding for a tennis player is knowing about the different situations that can arise. Ones ability ?to resolve? different situations comes from ones experience with similar situations. A tennis player?s understanding indeed comes from that big ?data set? of responses to different situations. 3. In general, biological learning may not be that magical as you think. Wish it was. Best, Asim From: Danko Nikolic Sent: Friday, July 15, 2022 11:39 AM To: Gary Marcus Cc: Asim Roy ; Grossberg, Stephen ; AIhub ; Post Connectionists ; maxversace at gmail.com Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Thanks Gary and Asim, Gary, yes, that is what I meant: recognizing a new situation in which a knife is being used or needs to be used, or could be used. We do not need to learn those at the time when learning about knives. We figure it out once the situation emerges. This is what is countless: the number of situations that may emerge. We do not need to know them upfront. Asim, it is interesting that you assumed that everything needs to be learned upfront. This is maybe exactly the difference between what connectionism assumes and what the human brain can actually do. The biological brain needs not to learn things upfront and yet 'understands' them once they happen. Also, as you asked for a definition of understanding, perhaps we can start exactly from that point: Understanding is when you do not have to learn different applications of knife (or object X, in general) and yet you are able to resolve the use of the knife once a relevant situation emerges. Understanding is great because the number of possible situations is countless and one cannot possibly prepare them as a learning data set. Transient selection of subnetworks based on MRs and GPGICs may do that 'understanding' job in the brain. That is my best guess after a long search for an appropriate mechanism. The scaling problem that I am talking about is about those countless situations. To be able to resolve them, linear scaling would not be enough. Even if there are connectionist systems that can scale linearly (albeit unlikely as the research stands now), the linearity would not be enough to fix the problem. Greetings, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Fri, Jul 15, 2022 at 3:51 PM Gary Marcus > wrote: I am with Danko here: he said ?resolve? not ?anticipate in advance?. I doubt any human is perfect in anticipating all uses of a knife but eg audiences had little trouble interpreting and enjoying all the weird repurposings that the TV character Macgyver was known for. On Jul 15, 2022, at 6:36 AM, Asim Roy > wrote: ? Dear Danko, 1. I am not sure if I myself know all the uses of a knife, leave aside countless ones. Given a particular situation, I might simulate in my mind about the potential usage, but I doubt our minds explore all the countless situations of usage of an object as soon as it learns about it. 2. I am not sure if a 2 or 3 year old child, after having ?learnt? about a knife, knows very many uses of it. I doubt the kid is awake all night and day simulating in the brain how and where to use such a knife. 3. ?Understanding? is a loaded term. I think it needs a definition. 4. I am copying Max Versace, a student of Steve Grossberg. His company markets a software that can learn quickly from a few examples. Not exactly one-shot learning, it needs a few shots. I believe it?s a variation of ART. But Max can clarify the details. And Tsvi is doing similar work. So, what you are asking for may already exist. So linear scaling may be the worst case scenario. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Danko Nikolic > Sent: Friday, July 15, 2022 12:19 AM To: Asim Roy > Cc: Grossberg, Stephen >; Gary Marcus >; AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I agree about the potential for linear scaling of ART and.other connectionist systems. However, there are two problems. The problem number one kills it already and this is that the real brain scales a lot better than linearly: For each new object learned, we are able to resolve countless many new situations in which this object takes part (e.g., finding various uses for a knife, many of which may be new, ad hoc -- this is a great ability of biological minds often referred to as 'understanding'). Hence, simple linear scaling by adding more neurons for additional objects is not good enough to match biological intelligence. The second problem becomes an overkill, and this is that linear scaling in connectionist systems works only in theory, under idealized conditions. In real life, say if working with ImageNet, the scaling turns into a power-law with an exponent much larger than one: We need something like 500x more resources just to double the number of objects. Hence, in practice, the demands for resources explode if you want to add more categories whilst not losing the accuracy. To summarize, there is no linear scaling in practice nor would linear scaling suffice, even if we found one. This should be a strong enough argument to search for another paradigm, something that scales better than connectionism. I discuss both problems in the new manuscript, and even track a bit deeper the problem of why connectionism lacks linear scaling in practice (I provide some revealing computations in the Supplementary Materials (with access to the code), although much more work needs to be done). Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Thu, Jul 14, 2022 at 11:48 PM Asim Roy > wrote: I think Steve Grossberg is generally correct. ART, RBF nets and similar flexible architectures can scale up almost linearly to problem size. My conjecture is that architectures that use distributed representation run into the scaling issue. On the other hand, however, distributed representation produces a more compact architecture compared to the shallow architectures of ART and RBFs. But, in terms of adding concepts, it?s easy to add a new object or concept to the shallow architectures. Perhaps someone can provide more insights on the architectural differences and the corresponding pros and cons of each. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Grossberg, Stephen Sent: Thursday, July 14, 2022 10:14 AM To: Danko Nikolic > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, I will respond to your comment below to the entire list. I recommend that future interactions be done between just the two of us. Everything that I write below is summarized in my new book: In brief, within Adaptive Resonance Theory, or ART, there IS no problem of scaling, in the sense that you describe it below, from learning (say) to correctly recognizing 100 objects to doing the same for 200. In fact, one of the properties for which I introduced ART in 1976 was to enable incremental learning in real time of arbitrary numbers of objects or events without experiencing catastrophic forgetting. I have called this a solution of the stability-plasticity dilemma: How our brains, and models like ART, can rapidly learn (plasticity) arbitrary numbers of objects without experiencing catastrophic forgetting (stability). I also derived ART in 1980, in an article within Psychological Review, from a thought experiment about how ANY system can AUTONOMOUSLY correct predictive errors in a changing world that is filled with unexpected events. This thought is experiment was derived from a few facts that are familiar to us all. They are familiar because they represent ubiquitous environmental pressures that we all experience. The thought experiment clarifies why, when they act together during the evolutionary process, models like ART are the unique result. Moveover, at no point during the thought experiment are the words mind or brain mentioned. ART is thus a universal solution of this learning, classification, and prediction problem. You write below about "connectionist systems". ART is a connectionist system. What do you mean by a "connectionist system"? What you write below about them is not true in general. Best, Steve ________________________________ From: Danko Nikolic > Sent: Thursday, July 14, 2022 12:16 PM To: Grossberg, Stephen > Cc: Gary Marcus >; connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 achler at gmail.com Fri Jul 15 04:32:06 2022 From: achler at gmail.com (Tsvi Achler) Date: Fri, 15 Jul 2022 01:32:06 -0700 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 the core problem is feedforward networks. In order for it to work correctly all weights need to be adjusted. This then requires iid rehearsal which is unnatural. Systems with massive feedback (much more than ART's vigilance) can avoid this problem. -Tsvi On Fri, Jul 15, 2022 at 1:21 AM Asim Roy wrote: > I think Steve Grossberg is generally correct. ART, RBF nets and similar > flexible architectures can scale up almost linearly to problem size. My > conjecture is that architectures that use distributed representation run > into the scaling issue. On the other hand, however, distributed > representation produces a more compact architecture compared to the shallow > architectures of ART and RBFs. But, in terms of adding concepts, it?s easy > to add a new object or concept to the shallow architectures. > > > > Perhaps someone can provide more insights on the architectural differences > and the corresponding pros and cons of each. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Grossberg, Stephen > *Sent:* Thursday, July 14, 2022 10:14 AM > *To:* Danko Nikolic > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Danko, > > > > I will respond to your comment below to the entire list. I recommend that > future interactions be done between just the two of us. > > > > Everything that I write below is summarized in my new book: > > > > In brief, within Adaptive Resonance Theory, or ART, there IS no problem of > scaling, in the sense that you describe it below, from learning (say) to > correctly recognizing 100 objects to doing the same for 200. In fact, one > of the properties for which I introduced ART in 1976 was to enable > incremental learning in real time of arbitrary numbers of objects or events > without experiencing catastrophic forgetting. > > > > I have called this a solution of the *stability-plasticity dilemma*: How > our brains, and models like ART, can rapidly learn (plasticity) arbitrary > numbers of objects without experiencing catastrophic forgetting (stability). > > > > I also derived ART in 1980, in an article within Psychological Review, > from a *thought experiment* about how ANY system can AUTONOMOUSLY correct > predictive errors in a changing world that is filled with unexpected events. > > > > This thought is experiment was derived from a few facts that are familiar > to us all. They are familiar because they represent ubiquitous > environmental pressures that we all experience. The thought experiment > clarifies why, when they act together during the evolutionary process, > models like ART are the unique result. > > > > Moveover, at no point during the thought experiment are the words mind or > brain mentioned. ART is thus a *universal solution* of this learning, > classification, and prediction problem. > > > > You write below about "connectionist systems". ART is a connectionist > system. > > > > What do you mean by a "connectionist system"? What you write below about > them is not true in general. > > > > Best, > > > > Steve > ------------------------------ > > *From:* Danko Nikolic > *Sent:* Thursday, July 14, 2022 12:16 PM > *To:* Grossberg, Stephen > *Cc:* Gary Marcus ; > connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Steve, > > > > Thank you very much for your message and for the greetings. I will pass > them on if an occasion arises. > > > > Regarding your question: The key problem I am trying to address and that, > to the best of my knowledge, no connectionist system was able to solve so > far is that of scaling the system's intelligence. For example, if the > system is able to correctly recognize 100 different objects, how many > additional resources are needed to double that to 200? All the empirical > data show that connectionist systems scale poorly: Some of the best systems > we have require 500x more resources in order to increase the intelligence > by only 2x. I document this problem in the manuscript and even run some > simulations to show that the worst performance is if connectionist systems > need to solve a generalized XOR problem. > > > > In contrast, the biological brain scales well. This I also quantify in the > paper. > > > > I will look at the publication that you mentioned. However, so far, I > haven't seen a solution that scales well in intelligence. > > > > My argument is that transient selection of subnetworks by the help of the > mentioned proteins is how intelligence scaling is achieved in biological > brains. > > > > In short, intelligence scaling is the key problem that concerns me. I > describe the intelligence scaling problem in more detail in this book that > just came out a few weeks ago and that is written for practitioners in Data > Scientist and AI: https://amzn.to/3IBxUpL > > > > > I hope that this at least partly answers where I see the problems and what > I am trying to solve. > > > > Greetings from Germany, > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: > > Dear Danko, > > > > I have just read your new article and would like to comment briefly about > it. > > > > In your introductory remarks, you write: > > > > "However, connectionism did not yet produce a satisfactory explanation of > how the mental emerges from the physical. A number of open problems > remains ( 5,6,7,8). As a result, the explanatory gap between the mind and > the brain remains wide open." > > > > I certainly believe that no theoretical explanation in science is ever > complete. However, I also believe that "the explanatory gap between the > mind and the brain" does not remain "wide open". > > > > My Magnum Opus, that was published in 2021, makes that belief clear in its > title: > > > > *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* > > > > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > > > > The book provides a self-contained and non-technical exposition in a > conversational tone of many principled and unifying explanations of > psychological and neurobiological data. > > > > In particular, it explains roles for the metabotropic glutamate receptors > that you mention in your own work. See the text and figures around p. 521. > This explanation unifies psychological, anatomical, neurophysiological, > biophysical, and biochemical data about the processes under discussion. > > > > I have a very old-fashioned view about how to understand scientific > theories. I get excited by theories that explain and predict more data than > previous theories. > > > > Which of the data that I explain in my book, and support with quantitative > computer simulations, can you also explain? > > > > What data can you explain, in the same quantitative sense, that you do not > think the neural models in my book can explain? > > > > I would be delighted to discuss these issues further with you. > > > > If you are in touch with my old friend and esteemed colleague, Wolf > Singer, please send him my warm regards. I cite the superb work that he and > various of his collaborators have done in many places in my book. > > > > Best, > > > > Steve > > > > 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 Danko Nikolic > *Sent:* Thursday, July 14, 2022 6:05 AM > *To:* Gary Marcus > *Cc:* connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Gary and everyone, > > > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > > > I am happy to announce that I have now finished a draft of a paper in > which I propose how the brain is able to achieve that. The manuscript > requires a bit of patience for two reasons: one is that the reader may be > exposed for the first time to certain aspects of brain physiology. The > second reason is that it may take some effort to understand the > counterintuitive implications of the new ideas (this requires a different > way of thinking than what we are used to based on connectionism). > > > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > > > (I apologize for the neuroscience lingo, which I tried to minimize.) > > > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > > > Thanks. > > > > Danko > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > > > > > On Thu, Feb 3, 2022 at 5:25 PM 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 ASIM.ROY at asu.edu Fri Jul 15 09:35:53 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Fri, 15 Jul 2022 13:35:53 +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 Danko, 1. I am not sure if I myself know all the uses of a knife, leave aside countless ones. Given a particular situation, I might simulate in my mind about the potential usage, but I doubt our minds explore all the countless situations of usage of an object as soon as it learns about it. 2. I am not sure if a 2 or 3 year old child, after having ?learnt? about a knife, knows very many uses of it. I doubt the kid is awake all night and day simulating in the brain how and where to use such a knife. 3. ?Understanding? is a loaded term. I think it needs a definition. 4. I am copying Max Versace, a student of Steve Grossberg. His company markets a software that can learn quickly from a few examples. Not exactly one-shot learning, it needs a few shots. I believe it?s a variation of ART. But Max can clarify the details. And Tsvi is doing similar work. So, what you are asking for may already exist. So linear scaling may be the worst case scenario. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Danko Nikolic Sent: Friday, July 15, 2022 12:19 AM To: Asim Roy Cc: Grossberg, Stephen ; Gary Marcus ; AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I agree about the potential for linear scaling of ART and.other connectionist systems. However, there are two problems. The problem number one kills it already and this is that the real brain scales a lot better than linearly: For each new object learned, we are able to resolve countless many new situations in which this object takes part (e.g., finding various uses for a knife, many of which may be new, ad hoc -- this is a great ability of biological minds often referred to as 'understanding'). Hence, simple linear scaling by adding more neurons for additional objects is not good enough to match biological intelligence. The second problem becomes an overkill, and this is that linear scaling in connectionist systems works only in theory, under idealized conditions. In real life, say if working with ImageNet, the scaling turns into a power-law with an exponent much larger than one: We need something like 500x more resources just to double the number of objects. Hence, in practice, the demands for resources explode if you want to add more categories whilst not losing the accuracy. To summarize, there is no linear scaling in practice nor would linear scaling suffice, even if we found one. This should be a strong enough argument to search for another paradigm, something that scales better than connectionism. I discuss both problems in the new manuscript, and even track a bit deeper the problem of why connectionism lacks linear scaling in practice (I provide some revealing computations in the Supplementary Materials (with access to the code), although much more work needs to be done). Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Thu, Jul 14, 2022 at 11:48 PM Asim Roy > wrote: I think Steve Grossberg is generally correct. ART, RBF nets and similar flexible architectures can scale up almost linearly to problem size. My conjecture is that architectures that use distributed representation run into the scaling issue. On the other hand, however, distributed representation produces a more compact architecture compared to the shallow architectures of ART and RBFs. But, in terms of adding concepts, it?s easy to add a new object or concept to the shallow architectures. Perhaps someone can provide more insights on the architectural differences and the corresponding pros and cons of each. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Grossberg, Stephen Sent: Thursday, July 14, 2022 10:14 AM To: Danko Nikolic > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, I will respond to your comment below to the entire list. I recommend that future interactions be done between just the two of us. Everything that I write below is summarized in my new book: In brief, within Adaptive Resonance Theory, or ART, there IS no problem of scaling, in the sense that you describe it below, from learning (say) to correctly recognizing 100 objects to doing the same for 200. In fact, one of the properties for which I introduced ART in 1976 was to enable incremental learning in real time of arbitrary numbers of objects or events without experiencing catastrophic forgetting. I have called this a solution of the stability-plasticity dilemma: How our brains, and models like ART, can rapidly learn (plasticity) arbitrary numbers of objects without experiencing catastrophic forgetting (stability). I also derived ART in 1980, in an article within Psychological Review, from a thought experiment about how ANY system can AUTONOMOUSLY correct predictive errors in a changing world that is filled with unexpected events. This thought is experiment was derived from a few facts that are familiar to us all. They are familiar because they represent ubiquitous environmental pressures that we all experience. The thought experiment clarifies why, when they act together during the evolutionary process, models like ART are the unique result. Moveover, at no point during the thought experiment are the words mind or brain mentioned. ART is thus a universal solution of this learning, classification, and prediction problem. You write below about "connectionist systems". ART is a connectionist system. What do you mean by a "connectionist system"? What you write below about them is not true in general. Best, Steve ________________________________ From: Danko Nikolic > Sent: Thursday, July 14, 2022 12:16 PM To: Grossberg, Stephen > Cc: Gary Marcus >; connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 jose at rubic.rutgers.edu Fri Jul 15 07:27:51 2022 From: jose at rubic.rutgers.edu (=?utf-8?B?U3RlcGhlbiBKb3PDqSBIYW5zb24=?=) Date: Fri, 15 Jul 2022 11:27:51 +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: So there are a number of issues here. In the first one, the brain obviously has a specific structure. DL supervised modeling of classification, NOT psychological categorization-- may scale linearly with resources.. however, its clear just adding more hidden layers has diminishing returns, adding various kinds of classification structure (GANs etc) or doing unsupervised learning or reinforcement learning is not going to end up with AGI or really any sort of final AI system and certainly not biological. Humans btw, scale as sub-linear systems.. but with hard limits. We can only recognize about 6000 faces (and the variance is about 4k across individuals), most undergraduates can recognize 100k words again with a large variance., so despite the fact that the brain has trillions of synapses, "loading" in explicit (MTL) memory is complex and somehow localized. Memory systems involve implicit (most DL) or explicit (almost noone), working memory is tiny--like 3 or 4+- 2 (sorry George). Many (Bengio, Le Cun etc) are experimenting with more complex structures, but keeping the system homogeneous, no extra symbolic baggage and forcing the exercise to have to learn the configuration rather than engineer it. In some ways, Grossberg has been out in front in all this, but with his own special and situated neural systems. He has been consistently and persistently creating focused neural modeling, which as an art form has always been a matter of taste. Connectionist models tend toward minimalist forms and Steve and others follow more Baroque or perhaps Victorian forms. What is clear is that the future is going to be blends of these forms and more complex, without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet. Steve On 7/15/22 4:01 AM, Dietterich, Thomas wrote: Dear Danko, In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. --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: Connectionists On Behalf Of Danko Nikolic Sent: Thursday, July 14, 2022 09:17 To: Grossberg, Stephen 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.] Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Fri Jul 15 14:38:58 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Fri, 15 Jul 2022 20:38:58 +0200 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <615FCABF-1610-49FF-9468-6BF8F6512585@nyu.edu> References: <615FCABF-1610-49FF-9468-6BF8F6512585@nyu.edu> Message-ID: Thanks Gary and Asim, Gary, yes, that is what I meant: recognizing a new situation in which a knife is being used or needs to be used, or could be used. We do not need to learn those at the time when learning about knives. We figure it out once the situation emerges. This is what is countless: the number of situations that may emerge. We do not need to know them upfront. Asim, it is interesting that you assumed that everything needs to be learned upfront. This is maybe exactly the difference between what connectionism assumes and what the human brain can actually do. The biological brain needs not to learn things upfront and yet 'understands' them once they happen. Also, as you asked for a definition of understanding, perhaps we can start exactly from that point: Understanding is when you do not have to learn different applications of knife (or object X, in general) and yet you are able to resolve the use of the knife once a relevant situation emerges. Understanding is great because the number of possible situations is countless and one cannot possibly prepare them as a learning data set. Transient selection of subnetworks based on MRs and GPGICs may do that 'understanding' job in the brain. That is my best guess after a long search for an appropriate mechanism. The scaling problem that I am talking about is about those countless situations. To be able to resolve them, linear scaling would not be enough. Even if there are connectionist systems that can scale linearly (albeit unlikely as the research stands now), the linearity would not be enough to fix the problem. Greetings, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Fri, Jul 15, 2022 at 3:51 PM Gary Marcus wrote: > I am with Danko here: he said ?resolve? not ?anticipate in advance?. > > I doubt any human is perfect in anticipating all uses of a knife but eg > audiences had little trouble interpreting and enjoying all the weird > repurposings that the TV character Macgyver was known for. > > On Jul 15, 2022, at 6:36 AM, Asim Roy wrote: > > > ? > > Dear Danko, > > > > 1. I am not sure if I myself know all the uses of a knife, leave aside > countless ones. Given a particular situation, I might simulate in my mind > about the potential usage, but I doubt our minds explore all the countless > situations of usage of an object as soon as it learns about it. > 2. I am not sure if a 2 or 3 year old child, after having ?learnt? > about a knife, knows very many uses of it. I doubt the kid is awake all > night and day simulating in the brain how and where to use such a knife. > 3. ?Understanding? is a loaded term. I think it needs a definition. > 4. I am copying Max Versace, a student of Steve Grossberg. His company > markets a software that can learn quickly from a few examples. Not exactly > one-shot learning, it needs a few shots. I believe it?s a variation of ART. > But Max can clarify the details. And Tsvi is doing similar work. So, what > you are asking for may already exist. So linear scaling may be the worst > case scenario. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > > > *From:* Danko Nikolic > *Sent:* Friday, July 15, 2022 12:19 AM > *To:* Asim Roy > *Cc:* Grossberg, Stephen ; Gary Marcus ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Asim, > > > > I agree about the potential for linear scaling of ART and.other > connectionist systems. However, there are two problems. > > > > The problem number one kills it already and this is that the real brain > scales a lot better than linearly: For each new object learned, we are able > to resolve countless many new situations in which this object takes part > (e.g., finding various uses for a knife, many of which may be new, ad hoc > -- this is a great ability of biological minds often referred to as > 'understanding'). Hence, simple linear scaling by adding more neurons for > additional objects is not good enough to match biological intelligence. > > > > The second problem becomes an overkill, and this is that linear scaling in > connectionist systems works only in theory, under idealized conditions. In > real life, say if working with ImageNet, the scaling turns into a power-law > with an exponent much larger than one: We need something like 500x more > resources just to double the number of objects. Hence, in practice, the > demands for resources explode if you want to add more categories whilst not > losing the accuracy. > > > > To summarize, there is no linear scaling in practice nor would linear > scaling suffice, even if we found one. > > > > This should be a strong enough argument to search for another paradigm, > something that scales better than connectionism. > > > > I discuss both problems in the new manuscript, and even track a bit deeper > the problem of why connectionism lacks linear scaling in practice (I > provide some revealing computations in the Supplementary Materials (with > access to the code), although much more work needs to be done). > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > -- I wonder, how is the brain able to generate insight? -- > > > > > > On Thu, Jul 14, 2022 at 11:48 PM Asim Roy wrote: > > I think Steve Grossberg is generally correct. ART, RBF nets and similar > flexible architectures can scale up almost linearly to problem size. My > conjecture is that architectures that use distributed representation run > into the scaling issue. On the other hand, however, distributed > representation produces a more compact architecture compared to the shallow > architectures of ART and RBFs. But, in terms of adding concepts, it?s easy > to add a new object or concept to the shallow architectures. > > > > Perhaps someone can provide more insights on the architectural differences > and the corresponding pros and cons of each. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Grossberg, Stephen > *Sent:* Thursday, July 14, 2022 10:14 AM > *To:* Danko Nikolic > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Danko, > > > > I will respond to your comment below to the entire list. I recommend that > future interactions be done between just the two of us. > > > > Everything that I write below is summarized in my new book: > > > > In brief, within Adaptive Resonance Theory, or ART, there IS no problem of > scaling, in the sense that you describe it below, from learning (say) to > correctly recognizing 100 objects to doing the same for 200. In fact, one > of the properties for which I introduced ART in 1976 was to enable > incremental learning in real time of arbitrary numbers of objects or events > without experiencing catastrophic forgetting. > > > > I have called this a solution of the *stability-plasticity dilemma*: How > our brains, and models like ART, can rapidly learn (plasticity) arbitrary > numbers of objects without experiencing catastrophic forgetting (stability). > > > > I also derived ART in 1980, in an article within Psychological Review, > from a *thought experiment* about how ANY system can AUTONOMOUSLY correct > predictive errors in a changing world that is filled with unexpected events. > > > > This thought is experiment was derived from a few facts that are familiar > to us all. They are familiar because they represent ubiquitous > environmental pressures that we all experience. The thought experiment > clarifies why, when they act together during the evolutionary process, > models like ART are the unique result. > > > > Moveover, at no point during the thought experiment are the words mind or > brain mentioned. ART is thus a *universal solution* of this learning, > classification, and prediction problem. > > > > You write below about "connectionist systems". ART is a connectionist > system. > > > > What do you mean by a "connectionist system"? What you write below about > them is not true in general. > > > > Best, > > > > Steve > ------------------------------ > > *From:* Danko Nikolic > *Sent:* Thursday, July 14, 2022 12:16 PM > *To:* Grossberg, Stephen > *Cc:* Gary Marcus ; > connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Steve, > > > > Thank you very much for your message and for the greetings. I will pass > them on if an occasion arises. > > > > Regarding your question: The key problem I am trying to address and that, > to the best of my knowledge, no connectionist system was able to solve so > far is that of scaling the system's intelligence. For example, if the > system is able to correctly recognize 100 different objects, how many > additional resources are needed to double that to 200? All the empirical > data show that connectionist systems scale poorly: Some of the best systems > we have require 500x more resources in order to increase the intelligence > by only 2x. I document this problem in the manuscript and even run some > simulations to show that the worst performance is if connectionist systems > need to solve a generalized XOR problem. > > > > In contrast, the biological brain scales well. This I also quantify in the > paper. > > > > I will look at the publication that you mentioned. However, so far, I > haven't seen a solution that scales well in intelligence. > > > > My argument is that transient selection of subnetworks by the help of the > mentioned proteins is how intelligence scaling is achieved in biological > brains. > > > > In short, intelligence scaling is the key problem that concerns me. I > describe the intelligence scaling problem in more detail in this book that > just came out a few weeks ago and that is written for practitioners in Data > Scientist and AI: https://amzn.to/3IBxUpL > > > > > I hope that this at least partly answers where I see the problems and what > I am trying to solve. > > > > Greetings from Germany, > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: > > Dear Danko, > > > > I have just read your new article and would like to comment briefly about > it. > > > > In your introductory remarks, you write: > > > > "However, connectionism did not yet produce a satisfactory explanation of > how the mental emerges from the physical. A number of open problems > remains ( 5,6,7,8). As a result, the explanatory gap between the mind and > the brain remains wide open." > > > > I certainly believe that no theoretical explanation in science is ever > complete. However, I also believe that "the explanatory gap between the > mind and the brain" does not remain "wide open". > > > > My Magnum Opus, that was published in 2021, makes that belief clear in its > title: > > > > *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* > > > > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > > > > The book provides a self-contained and non-technical exposition in a > conversational tone of many principled and unifying explanations of > psychological and neurobiological data. > > > > In particular, it explains roles for the metabotropic glutamate receptors > that you mention in your own work. See the text and figures around p. 521. > This explanation unifies psychological, anatomical, neurophysiological, > biophysical, and biochemical data about the processes under discussion. > > > > I have a very old-fashioned view about how to understand scientific > theories. I get excited by theories that explain and predict more data than > previous theories. > > > > Which of the data that I explain in my book, and support with quantitative > computer simulations, can you also explain? > > > > What data can you explain, in the same quantitative sense, that you do not > think the neural models in my book can explain? > > > > I would be delighted to discuss these issues further with you. > > > > If you are in touch with my old friend and esteemed colleague, Wolf > Singer, please send him my warm regards. I cite the superb work that he and > various of his collaborators have done in many places in my book. > > > > Best, > > > > Steve > > > > 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 Danko Nikolic > *Sent:* Thursday, July 14, 2022 6:05 AM > *To:* Gary Marcus > *Cc:* connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Gary and everyone, > > > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > > > I am happy to announce that I have now finished a draft of a paper in > which I propose how the brain is able to achieve that. The manuscript > requires a bit of patience for two reasons: one is that the reader may be > exposed for the first time to certain aspects of brain physiology. The > second reason is that it may take some effort to understand the > counterintuitive implications of the new ideas (this requires a different > way of thinking than what we are used to based on connectionism). > > > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > > > (I apologize for the neuroscience lingo, which I tried to minimize.) > > > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > > > Thanks. > > > > Danko > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > > > > > On Thu, Feb 3, 2022 at 5:25 PM 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 achler at gmail.com Sat Jul 16 10:51:56 2022 From: achler at gmail.com (Tsvi Achler) Date: Sat, 16 Jul 2022 07:51:56 -0700 Subject: Connectionists: Neural architectures that embody biological intelligence In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Stephen, I have a connectionist model where feedback takes on a much greater role than the Resonance and other theories. I also have a richer background than many researchers in this field. I have a degree in electrical engineering & computer science, I did my PhD work doing neurophysiology recording neurons and working in a cognitive lab recording differential human reaction times to visual stimulation. I also got an MD focusing on neurology and patients. Consistently throughout the years established academics and their associates have blocked this theory's publication and funding in favor of their own. Since academia is mostly political, this is a big deal. Moreover, it bothers me seeing this done to others. Unfortunately you are by far NOT the worst at doing so, you are just the most transparent about it. I came to the conclusion that academia is not a place to innovate, especially if you come from a multidisciplinary background because (analogous to some of the models) the politics multiply exponentially. Although your work was innovative in the grand scheme of things, what you and other well established academics are doing is not ok. Sincerely, -Tsvi On Sat, Jul 16, 2022 at 12:04 AM Grossberg, Stephen wrote: > Dear Asim and Danko, > > A lot of your concerns about scaling do not apply to the kinds of > biological neural networks that my colleagues and I have developed over the > years. You can find a self-contained summary of many of them in my Magnum > Opus: > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > As Asim notes below, it is indeed the case that ART can often make good > predictions based on small amounts of learned data. This applies as well to > large-scale applications naturalistic data. > > Gail Carpenter and her colleagues have, for example, shown how this works > in learning complicated maps of multiple vegetation classes during remote > sensing; e.g. > > Carpenter, G.A., Gopal, S., Macomber, S., Martens, S., & Woodcock, C.E. > (1999). A neural network method for mixture estimation for vegetation > mapping. Remote Sensing of Environment, 70(2), 138-152. > http://techlab.bu.edu/members/gail/articles/127_Mixtures_RSE_1999_.pdf > > or in learning medical database predictions in response to incomplete, > probabilistic, and even incorrect data. > > In this regard, Gail et al. have also shown how an ART system can > incrementally learn a cognitive hierarchy of rules whereby to understand > such data; i.e., converts information into knowledge; e.g., > > Carpenter, G.A., & Ravindran, A. (2008). Unifying multiple knowledge > domains using the ARTMAP information fusion system. Proceedings of the 11th > International Conference on Information Fusion, Cologne, Germany, June 30 - > July 3, 2008. > > http://techlab.bu.edu/members/gail/articles/155_Fusion2008_CarpenterRavindran.pdf > > Carpenter, G.A. & Markuzon, N. (1998). ARTMAP?IC and medical diagnosis: > Instance counting and inconsistent cases. Neural Networks, 11(2), 323-336. > http://techlab.bu.edu/members/gail/articles/117_ARTMAP-IC_1998_.pdf > > My own work is filled with models that incrementally learn to carry out > goal-oriented tasks without regard to scaling concerns. This work develops > neural architectures that involve the coordinated actions of many brain > regions, not just learned classification. These architectures are supported > by unified and principled explanations of lots of psychological and > neurobiological data; e.g., > > Chang, H.-C., Grossberg, S., and Cao, Y. (2014) Where?s Waldo? How > perceptual cognitive, and emotional brain processes cooperate during > learning to categorize and find desired objects in a cluttered scene. > Frontiers in Integrative Neuroscience, doi: 10.3389/fnint.2014.0043, > https://www.frontiersin.org/articles/10.3389/fnint.2014.00043/full > > Grossberg, S., and Vladusich, T. (2010). How do children learn to follow > gaze, share joint attention, imitate their teachers, and use tools during > social interactions? Neural Networks, 23, 940-965. > > https://sites.bu.edu/steveg/files/2016/06/GrossbergVladusichNN2010.pdf > > See Figure 1 in the following article to get a sense of how many brain > processes other than classification are needed to realize true biological > intelligence: > > Grossberg, S. (2018). Desirability, availability, credit assignment, > category learning, and attention: Cognitive-emotional and working memory > dynamics of orbitofrontal, ventrolateral, and dorsolateral prefrontal > cortices. Brain and Neuroscience Advances, May 8, 2018. > https://journals.sagepub.com/doi/full/10.1177/2398212818772179 > > Best, > > Steve > ------------------------------ > *From:* Asim Roy > *Sent:* Friday, July 15, 2022 9:35 AM > *To:* Danko Nikolic > *Cc:* Grossberg, Stephen ; Gary Marcus ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; 'maxversace at gmail.com' < > maxversace at gmail.com> > *Subject:* RE: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > Dear Danko, > > > > 1. I am not sure if I myself know all the uses of a knife, leave aside > countless ones. Given a particular situation, I might simulate in my mind > about the potential usage, but I doubt our minds explore all the countless > situations of usage of an object as soon as it learns about it. > 2. I am not sure if a 2 or 3 year old child, after having ?learnt? > about a knife, knows very many uses of it. I doubt the kid is awake all > night and day simulating in the brain how and where to use such a knife. > 3. ?Understanding? is a loaded term. I think it needs a definition. > 4. I am copying Max Versace, a student of Steve Grossberg. His company > markets a software that can learn quickly from a few examples. Not exactly > one-shot learning, it needs a few shots. I believe it?s a variation of ART. > But Max can clarify the details. And Tsvi is doing similar work. So, what > you are asking for may already exist. So linear scaling may be the worst > case scenario. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > > > *From:* Danko Nikolic > *Sent:* Friday, July 15, 2022 12:19 AM > *To:* Asim Roy > *Cc:* Grossberg, Stephen ; Gary Marcus ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Asim, > > > > I agree about the potential for linear scaling of ART and.other > connectionist systems. However, there are two problems. > > > > The problem number one kills it already and this is that the real brain > scales a lot better than linearly: For each new object learned, we are able > to resolve countless many new situations in which this object takes part > (e.g., finding various uses for a knife, many of which may be new, ad hoc > -- this is a great ability of biological minds often referred to as > 'understanding'). Hence, simple linear scaling by adding more neurons for > additional objects is not good enough to match biological intelligence. > > > > The second problem becomes an overkill, and this is that linear scaling in > connectionist systems works only in theory, under idealized conditions. In > real life, say if working with ImageNet, the scaling turns into a power-law > with an exponent much larger than one: We need something like 500x more > resources just to double the number of objects. Hence, in practice, the > demands for resources explode if you want to add more categories whilst not > losing the accuracy. > > > > To summarize, there is no linear scaling in practice nor would linear > scaling suffice, even if we found one. > > > > This should be a strong enough argument to search for another paradigm, > something that scales better than connectionism. > > > > I discuss both problems in the new manuscript, and even track a bit deeper > the problem of why connectionism lacks linear scaling in practice (I > provide some revealing computations in the Supplementary Materials (with > access to the code), although much more work needs to be done). > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > -- I wonder, how is the brain able to generate insight? -- > > > > > > On Thu, Jul 14, 2022 at 11:48 PM Asim Roy wrote: > > I think Steve Grossberg is generally correct. ART, RBF nets and similar > flexible architectures can scale up almost linearly to problem size. My > conjecture is that architectures that use distributed representation run > into the scaling issue. On the other hand, however, distributed > representation produces a more compact architecture compared to the shallow > architectures of ART and RBFs. But, in terms of adding concepts, it?s easy > to add a new object or concept to the shallow architectures. > > > > Perhaps someone can provide more insights on the architectural differences > and the corresponding pros and cons of each. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Grossberg, Stephen > *Sent:* Thursday, July 14, 2022 10:14 AM > *To:* Danko Nikolic > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Danko, > > > > I will respond to your comment below to the entire list. I recommend that > future interactions be done between just the two of us. > > > > Everything that I write below is summarized in my new book: > > > > In brief, within Adaptive Resonance Theory, or ART, there IS no problem of > scaling, in the sense that you describe it below, from learning (say) to > correctly recognizing 100 objects to doing the same for 200. In fact, one > of the properties for which I introduced ART in 1976 was to enable > incremental learning in real time of arbitrary numbers of objects or events > without experiencing catastrophic forgetting. > > > > I have called this a solution of the *stability-plasticity dilemma*: How > our brains, and models like ART, can rapidly learn (plasticity) arbitrary > numbers of objects without experiencing catastrophic forgetting (stability). > > > > I also derived ART in 1980, in an article within Psychological Review, > from a *thought experiment* about how ANY system can AUTONOMOUSLY correct > predictive errors in a changing world that is filled with unexpected events. > > > > This thought is experiment was derived from a few facts that are familiar > to us all. They are familiar because they represent ubiquitous > environmental pressures that we all experience. The thought experiment > clarifies why, when they act together during the evolutionary process, > models like ART are the unique result. > > > > Moveover, at no point during the thought experiment are the words mind or > brain mentioned. ART is thus a *universal solution* of this learning, > classification, and prediction problem. > > > > You write below about "connectionist systems". ART is a connectionist > system. > > > > What do you mean by a "connectionist system"? What you write below about > them is not true in general. > > > > Best, > > > > Steve > ------------------------------ > > *From:* Danko Nikolic > *Sent:* Thursday, July 14, 2022 12:16 PM > *To:* Grossberg, Stephen > *Cc:* Gary Marcus ; > connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Steve, > > > > Thank you very much for your message and for the greetings. I will pass > them on if an occasion arises. > > > > Regarding your question: The key problem I am trying to address and that, > to the best of my knowledge, no connectionist system was able to solve so > far is that of scaling the system's intelligence. For example, if the > system is able to correctly recognize 100 different objects, how many > additional resources are needed to double that to 200? All the empirical > data show that connectionist systems scale poorly: Some of the best systems > we have require 500x more resources in order to increase the intelligence > by only 2x. I document this problem in the manuscript and even run some > simulations to show that the worst performance is if connectionist systems > need to solve a generalized XOR problem. > > > > In contrast, the biological brain scales well. This I also quantify in the > paper. > > > > I will look at the publication that you mentioned. However, so far, I > haven't seen a solution that scales well in intelligence. > > > > My argument is that transient selection of subnetworks by the help of the > mentioned proteins is how intelligence scaling is achieved in biological > brains. > > > > In short, intelligence scaling is the key problem that concerns me. I > describe the intelligence scaling problem in more detail in this book that > just came out a few weeks ago and that is written for practitioners in Data > Scientist and AI: https://amzn.to/3IBxUpL > > > > > I hope that this at least partly answers where I see the problems and what > I am trying to solve. > > > > Greetings from Germany, > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: > > Dear Danko, > > > > I have just read your new article and would like to comment briefly about > it. > > > > In your introductory remarks, you write: > > > > "However, connectionism did not yet produce a satisfactory explanation of > how the mental emerges from the physical. A number of open problems > remains ( 5,6,7,8). As a result, the explanatory gap between the mind and > the brain remains wide open." > > > > I certainly believe that no theoretical explanation in science is ever > complete. However, I also believe that "the explanatory gap between the > mind and the brain" does not remain "wide open". > > > > My Magnum Opus, that was published in 2021, makes that belief clear in its > title: > > > > *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* > > > > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > > > > The book provides a self-contained and non-technical exposition in a > conversational tone of many principled and unifying explanations of > psychological and neurobiological data. > > > > In particular, it explains roles for the metabotropic glutamate receptors > that you mention in your own work. See the text and figures around p. 521. > This explanation unifies psychological, anatomical, neurophysiological, > biophysical, and biochemical data about the processes under discussion. > > > > I have a very old-fashioned view about how to understand scientific > theories. I get excited by theories that explain and predict more data than > previous theories. > > > > Which of the data that I explain in my book, and support with quantitative > computer simulations, can you also explain? > > > > What data can you explain, in the same quantitative sense, that you do not > think the neural models in my book can explain? > > > > I would be delighted to discuss these issues further with you. > > > > If you are in touch with my old friend and esteemed colleague, Wolf > Singer, please send him my warm regards. I cite the superb work that he and > various of his collaborators have done in many places in my book. > > > > Best, > > > > Steve > > > > 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 Danko Nikolic > *Sent:* Thursday, July 14, 2022 6:05 AM > *To:* Gary Marcus > *Cc:* connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Gary and everyone, > > > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > > > I am happy to announce that I have now finished a draft of a paper in > which I propose how the brain is able to achieve that. The manuscript > requires a bit of patience for two reasons: one is that the reader may be > exposed for the first time to certain aspects of brain physiology. The > second reason is that it may take some effort to understand the > counterintuitive implications of the new ideas (this requires a different > way of thinking than what we are used to based on connectionism). > > > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > > > (I apologize for the neuroscience lingo, which I tried to minimize.) > > > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > > > Thanks. > > > > Danko > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > > > > > On Thu, Feb 3, 2022 at 5:25 PM 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 Stefano.Rovetta at unige.it Sat Jul 16 06:49:45 2022 From: Stefano.Rovetta at unige.it (Stefano Rovetta) Date: Sat, 16 Jul 2022 12:49:45 +0200 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: <615FCABF-1610-49FF-9468-6BF8F6512585@nyu.edu> Message-ID: <20220716124945.Horde.b0sg9EKX-3keilzmIVmOKgu@posta.unige.it> Dear Asim what do you mean by "similar situations"? --Stefano Rovetta Asim Roy ha scritto: > Dear Danko, > > > 1. ?Figure it out once the situation emerges? and ?we do not need > to learn? ?upfront? sounds a bit magical, even for biological > systems. A professional tennis player practices for years and years > so that he/she knows exactly how to respond to each and every > situation as much as possible. Such learning does not end with just > 10 days of training. Such a player would prefer to know as much as > possible ?upfront? the various situations that can arise. That?s the > meaning of training and learning. And that also means hitting tennis > balls millions of times (countless??) perhaps. And that?s learning > from a lot of data. > 2. You might want to rethink your definition of ?understanding? > given the above example. Understanding for a tennis player is > knowing about the different situations that can arise. Ones ability > ?to resolve? different situations comes from ones experience with > similar situations. A tennis player?s understanding indeed comes > from that big ?data set? of responses to different situations. > 3. In general, biological learning may not be that magical as you > think. Wish it was. > > Best, > Asim > > From: Danko Nikolic > Sent: Friday, July 15, 2022 11:39 AM > To: Gary Marcus > Cc: Asim Roy ; Grossberg, Stephen ; > AIhub ; Post Connectionists > ; maxversace at gmail.com > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Thanks Gary and Asim, > > Gary, yes, that is what I meant: recognizing a new situation in > which a knife is being used or needs to be used, or could be used. > We do not need to learn those at the time when learning about > knives. We figure it out once the situation emerges. This is what is > countless: the number of situations that may emerge. We do not need > to know them upfront. > > Asim, it is interesting that you assumed that everything needs to be > learned upfront. This is maybe exactly the difference between what > connectionism assumes and what the human brain can actually do. The > biological brain needs not to learn things upfront and yet > 'understands' them once they happen. > > Also, as you asked for a definition of understanding, perhaps we can > start exactly from that point: Understanding is when you do not have > to learn different applications of knife (or object X, in general) > and yet you are able to resolve the use of the knife once a relevant > situation emerges. Understanding is great because the number of > possible situations is countless and one cannot possibly prepare > them as a learning data set. > > Transient selection of subnetworks based on MRs and GPGICs may do > that 'understanding' job in the brain. That is my best guess after a > long search for an appropriate mechanism. > > The scaling problem that I am talking about is about those countless > situations. To be able to resolve them, linear scaling would not be > enough. Even if there are connectionist systems that can scale > linearly (albeit unlikely as the research stands now), the linearity > would not be enough to fix the problem. > > Greetings, > > Danko > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > -- I wonder, how is the brain able to generate insight? -- > > > On Fri, Jul 15, 2022 at 3:51 PM Gary Marcus > > wrote: > I am with Danko here: he said ?resolve? not ?anticipate in advance?. > > I doubt any human is perfect in anticipating all uses of a knife but > eg audiences had little trouble interpreting and enjoying all the > weird repurposings that the TV character Macgyver was known for. > > On Jul 15, 2022, at 6:36 AM, Asim Roy > > wrote: > > ? > Dear Danko, > > > 1. I am not sure if I myself know all the uses of a knife, leave > aside countless ones. Given a particular situation, I might simulate > in my mind about the potential usage, but I doubt our minds explore > all the countless situations of usage of an object as soon as it > learns about it. > 2. I am not sure if a 2 or 3 year old child, after having > ?learnt? about a knife, knows very many uses of it. I doubt the kid > is awake all night and day simulating in the brain how and where to > use such a knife. > 3. ?Understanding? is a loaded term. I think it needs a definition. > 4. I am copying Max Versace, a student of Steve Grossberg. His > company markets a software that can learn quickly from a few > examples. Not exactly one-shot learning, it needs a few shots. I > believe it?s a variation of ART. But Max can clarify the details. > And Tsvi is doing similar work. So, what you are asking for may > already exist. So linear scaling may be the worst case scenario. > > Best, > Asim Roy > Professor, Information Systems > Arizona State University > Lifeboat Foundation Bios: Professor Asim > Roy > Asim Roy | iSearch > (asu.edu) > > > > From: Danko Nikolic > > Sent: Friday, July 15, 2022 12:19 AM > To: Asim Roy > > Cc: Grossberg, Stephen >; Gary > Marcus >; AIhub > >; > connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton > > Dear Asim, > > I agree about the potential for linear scaling of ART and.other > connectionist systems. However, there are two problems. > > The problem number one kills it already and this is that the real > brain scales a lot better than linearly: For each new object > learned, we are able to resolve countless many new situations in > which this object takes part (e.g., finding various uses for a > knife, many of which may be new, ad hoc -- this is a great ability > of biological minds often referred to as 'understanding'). Hence, > simple linear scaling by adding more neurons for additional objects > is not good enough to match biological intelligence. > > The second problem becomes an overkill, and this is that linear > scaling in connectionist systems works only in theory, under > idealized conditions. In real life, say if working with ImageNet, > the scaling turns into a power-law with an exponent much larger than > one: We need something like 500x more resources just to double the > number of objects. Hence, in practice, the demands for resources > explode if you want to add more categories whilst not losing the > accuracy. > > To summarize, there is no linear scaling in practice nor would > linear scaling suffice, even if we found one. > > This should be a strong enough argument to search for another > paradigm, something that scales better than connectionism. > > I discuss both problems in the new manuscript, and even track a bit > deeper the problem of why connectionism lacks linear scaling in > practice (I provide some revealing computations in the Supplementary > Materials (with access to the code), although much more work needs > to be done). > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ From achler at gmail.com Sun Jul 17 10:55:35 2022 From: achler at gmail.com (Tsvi Achler) Date: Sun, 17 Jul 2022 07:55:35 -0700 Subject: Connectionists: Neural architectures that embody biological intelligence In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Steve, What motivated me to write was your response a couple of messages ago to someone who is not established in their field describing their model. Studies on academics show that researchers who are not established but do original work do not get published and cited as much. Please see article: www.nber.org/papers/w22180 Moreover established researchers tend to push their theories and increments of their theories so strongly that it significantly affects progress in the field. Please see article: www.nber.org/papers/w21788 Since you mention it, the personal instance I am referring to is a conference where I got the following review (and I am paraphrasing): I* dont really understand this model but it must be ART, and if it is ART this is wrong and that is wrong so I recommend rejecting it.* And in a box for reviewer certainty the review was listed as *100% certain*. The consequence was that I had only 3 minutes to talk about a model that is counterintuitive given today's notions, as someone who exhausted all their meager resources just to get there. This summarizes my experiences in academia trying to put forward something new. I am happy to pull up the specific text but that distracts from the point. The point is that at least this review was transparent. Most reviewers are not likely to be as transparent when something is counterintuitive, not normative and thus harder to understand. What I am saying is that given this knowledge about academia, established researchers should be very careful as they can easily stifle new research without realizing it. If established academics push too strongly then academica can become a political club, not a place for progress. I believe this is a major contributor to why so little progress has been made in the field of understanding the brain through connectionist models. Sincerely, -Tsvi On Sat, Jul 16, 2022 at 8:45 AM Grossberg, Stephen wrote: > Dear Tsvi, > > I have no idea why you are writing to me. > > I would prefer that you did not engage the entire connectionists mailing > list. However, since you did, I need to include everyone in my reply. > > For starters, I have not been an editor of any journal since 2010. > > When I was Editor-in-Chief of *Neural Networks* before that, and a new > article was submitted, I assigned it to one of over 70 action editors who > was a specialist in the topic of the article. > > That action editor then took full responsibility for getting three reviews > of the article. If any reviewer disagreed with other reviewers for a > potentially serious reason, then yet another reviewer was typically sought > by the action editor in order to try to resolve the difference. > > Almost always, the reviewers agreed about publication recommendations, so > this was not needed. > > I always followed the recommendations of action editors to publish or not, > based upon the above process. > > I only entered any decision if the action editor solicited my help for a > problem for which he/she needed advice. This hardly ever happened. > > Best, > > Steve > > > ------------------------------ > *From:* Tsvi Achler > *Sent:* Saturday, July 16, 2022 10:51 AM > *To:* Grossberg, Stephen > *Cc:* Asim Roy ; Danko Nikolic ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu> > *Subject:* Re: Connectionists: Neural architectures that embody > biological intelligence > > Dear Stephen, > I have a connectionist model where feedback takes on a much greater role > than the Resonance and other theories. > I also have a richer background than many researchers in this field. I > have a degree in electrical engineering & computer science, I did my PhD > work doing neurophysiology recording neurons and working in a cognitive > lab recording differential human reaction times to visual stimulation. I > also got an MD focusing on neurology and patients. > > Consistently throughout the years established academics and their > associates have blocked this theory's publication and funding in favor of > their own. > Since academia is mostly political, this is a big deal. Moreover, it > bothers me seeing this done to others. > > Unfortunately you are by far NOT the worst at doing so, you are just the > most transparent about it. > > I came to the conclusion that academia is not a place to innovate, > especially if you come from a multidisciplinary background because > (analogous to some of the models) the politics multiply exponentially. > > Although your work was innovative in the grand scheme of things, what you > and other well established academics are doing is not ok. > Sincerely, > -Tsvi > > > On Sat, Jul 16, 2022 at 12:04 AM Grossberg, Stephen wrote: > > Dear Asim and Danko, > > A lot of your concerns about scaling do not apply to the kinds of > biological neural networks that my colleagues and I have developed over the > years. You can find a self-contained summary of many of them in my Magnum > Opus: > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > As Asim notes below, it is indeed the case that ART can often make good > predictions based on small amounts of learned data. This applies as well to > large-scale applications naturalistic data. > > Gail Carpenter and her colleagues have, for example, shown how this works > in learning complicated maps of multiple vegetation classes during remote > sensing; e.g. > > Carpenter, G.A., Gopal, S., Macomber, S., Martens, S., & Woodcock, C.E. > (1999). A neural network method for mixture estimation for vegetation > mapping. Remote Sensing of Environment, 70(2), 138-152. > http://techlab.bu.edu/members/gail/articles/127_Mixtures_RSE_1999_.pdf > > or in learning medical database predictions in response to incomplete, > probabilistic, and even incorrect data. > > In this regard, Gail et al. have also shown how an ART system can > incrementally learn a cognitive hierarchy of rules whereby to understand > such data; i.e., converts information into knowledge; e.g., > > Carpenter, G.A., & Ravindran, A. (2008). Unifying multiple knowledge > domains using the ARTMAP information fusion system. Proceedings of the 11th > International Conference on Information Fusion, Cologne, Germany, June 30 - > July 3, 2008. > > http://techlab.bu.edu/members/gail/articles/155_Fusion2008_CarpenterRavindran.pdf > > Carpenter, G.A. & Markuzon, N. (1998). ARTMAP?IC and medical diagnosis: > Instance counting and inconsistent cases. Neural Networks, 11(2), 323-336. > http://techlab.bu.edu/members/gail/articles/117_ARTMAP-IC_1998_.pdf > > My own work is filled with models that incrementally learn to carry out > goal-oriented tasks without regard to scaling concerns. This work develops > neural architectures that involve the coordinated actions of many brain > regions, not just learned classification. These architectures are supported > by unified and principled explanations of lots of psychological and > neurobiological data; e.g., > > Chang, H.-C., Grossberg, S., and Cao, Y. (2014) Where?s Waldo? How > perceptual cognitive, and emotional brain processes cooperate during > learning to categorize and find desired objects in a cluttered scene. > Frontiers in Integrative Neuroscience, doi: 10.3389/fnint.2014.0043, > https://www.frontiersin.org/articles/10.3389/fnint.2014.00043/full > > Grossberg, S., and Vladusich, T. (2010). How do children learn to follow > gaze, share joint attention, imitate their teachers, and use tools during > social interactions? Neural Networks, 23, 940-965. > > https://sites.bu.edu/steveg/files/2016/06/GrossbergVladusichNN2010.pdf > > See Figure 1 in the following article to get a sense of how many brain > processes other than classification are needed to realize true biological > intelligence: > > Grossberg, S. (2018). Desirability, availability, credit assignment, > category learning, and attention: Cognitive-emotional and working memory > dynamics of orbitofrontal, ventrolateral, and dorsolateral prefrontal > cortices. Brain and Neuroscience Advances, May 8, 2018. > https://journals.sagepub.com/doi/full/10.1177/2398212818772179 > > Best, > > Steve > ------------------------------ > *From:* Asim Roy > *Sent:* Friday, July 15, 2022 9:35 AM > *To:* Danko Nikolic > *Cc:* Grossberg, Stephen ; Gary Marcus ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; 'maxversace at gmail.com' < > maxversace at gmail.com> > *Subject:* RE: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > Dear Danko, > > > > 1. I am not sure if I myself know all the uses of a knife, leave aside > countless ones. Given a particular situation, I might simulate in my mind > about the potential usage, but I doubt our minds explore all the countless > situations of usage of an object as soon as it learns about it. > 2. I am not sure if a 2 or 3 year old child, after having ?learnt? > about a knife, knows very many uses of it. I doubt the kid is awake all > night and day simulating in the brain how and where to use such a knife. > 3. ?Understanding? is a loaded term. I think it needs a definition. > 4. I am copying Max Versace, a student of Steve Grossberg. His company > markets a software that can learn quickly from a few examples. Not exactly > one-shot learning, it needs a few shots. I believe it?s a variation of ART. > But Max can clarify the details. And Tsvi is doing similar work. So, what > you are asking for may already exist. So linear scaling may be the worst > case scenario. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > > > *From:* Danko Nikolic > *Sent:* Friday, July 15, 2022 12:19 AM > *To:* Asim Roy > *Cc:* Grossberg, Stephen ; Gary Marcus ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Asim, > > > > I agree about the potential for linear scaling of ART and.other > connectionist systems. However, there are two problems. > > > > The problem number one kills it already and this is that the real brain > scales a lot better than linearly: For each new object learned, we are able > to resolve countless many new situations in which this object takes part > (e.g., finding various uses for a knife, many of which may be new, ad hoc > -- this is a great ability of biological minds often referred to as > 'understanding'). Hence, simple linear scaling by adding more neurons for > additional objects is not good enough to match biological intelligence. > > > > The second problem becomes an overkill, and this is that linear scaling in > connectionist systems works only in theory, under idealized conditions. In > real life, say if working with ImageNet, the scaling turns into a power-law > with an exponent much larger than one: We need something like 500x more > resources just to double the number of objects. Hence, in practice, the > demands for resources explode if you want to add more categories whilst not > losing the accuracy. > > > > To summarize, there is no linear scaling in practice nor would linear > scaling suffice, even if we found one. > > > > This should be a strong enough argument to search for another paradigm, > something that scales better than connectionism. > > > > I discuss both problems in the new manuscript, and even track a bit deeper > the problem of why connectionism lacks linear scaling in practice (I > provide some revealing computations in the Supplementary Materials (with > access to the code), although much more work needs to be done). > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > -- I wonder, how is the brain able to generate insight? -- > > > > > > On Thu, Jul 14, 2022 at 11:48 PM Asim Roy wrote: > > I think Steve Grossberg is generally correct. ART, RBF nets and similar > flexible architectures can scale up almost linearly to problem size. My > conjecture is that architectures that use distributed representation run > into the scaling issue. On the other hand, however, distributed > representation produces a more compact architecture compared to the shallow > architectures of ART and RBFs. But, in terms of adding concepts, it?s easy > to add a new object or concept to the shallow architectures. > > > > Perhaps someone can provide more insights on the architectural differences > and the corresponding pros and cons of each. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Grossberg, Stephen > *Sent:* Thursday, July 14, 2022 10:14 AM > *To:* Danko Nikolic > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Danko, > > > > I will respond to your comment below to the entire list. I recommend that > future interactions be done between just the two of us. > > > > Everything that I write below is summarized in my new book: > > > > In brief, within Adaptive Resonance Theory, or ART, there IS no problem of > scaling, in the sense that you describe it below, from learning (say) to > correctly recognizing 100 objects to doing the same for 200. In fact, one > of the properties for which I introduced ART in 1976 was to enable > incremental learning in real time of arbitrary numbers of objects or events > without experiencing catastrophic forgetting. > > > > I have called this a solution of the *stability-plasticity dilemma*: How > our brains, and models like ART, can rapidly learn (plasticity) arbitrary > numbers of objects without experiencing catastrophic forgetting (stability). > > > > I also derived ART in 1980, in an article within Psychological Review, > from a *thought experiment* about how ANY system can AUTONOMOUSLY correct > predictive errors in a changing world that is filled with unexpected events. > > > > This thought is experiment was derived from a few facts that are familiar > to us all. They are familiar because they represent ubiquitous > environmental pressures that we all experience. The thought experiment > clarifies why, when they act together during the evolutionary process, > models like ART are the unique result. > > > > Moveover, at no point during the thought experiment are the words mind or > brain mentioned. ART is thus a *universal solution* of this learning, > classification, and prediction problem. > > > > You write below about "connectionist systems". ART is a connectionist > system. > > > > What do you mean by a "connectionist system"? What you write below about > them is not true in general. > > > > Best, > > > > Steve > ------------------------------ > > *From:* Danko Nikolic > *Sent:* Thursday, July 14, 2022 12:16 PM > *To:* Grossberg, Stephen > *Cc:* Gary Marcus ; > connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Steve, > > > > Thank you very much for your message and for the greetings. I will pass > them on if an occasion arises. > > > > Regarding your question: The key problem I am trying to address and that, > to the best of my knowledge, no connectionist system was able to solve so > far is that of scaling the system's intelligence. For example, if the > system is able to correctly recognize 100 different objects, how many > additional resources are needed to double that to 200? All the empirical > data show that connectionist systems scale poorly: Some of the best systems > we have require 500x more resources in order to increase the intelligence > by only 2x. I document this problem in the manuscript and even run some > simulations to show that the worst performance is if connectionist systems > need to solve a generalized XOR problem. > > > > In contrast, the biological brain scales well. This I also quantify in the > paper. > > > > I will look at the publication that you mentioned. However, so far, I > haven't seen a solution that scales well in intelligence. > > > > My argument is that transient selection of subnetworks by the help of the > mentioned proteins is how intelligence scaling is achieved in biological > brains. > > > > In short, intelligence scaling is the key problem that concerns me. I > describe the intelligence scaling problem in more detail in this book that > just came out a few weeks ago and that is written for practitioners in Data > Scientist and AI: https://amzn.to/3IBxUpL > > > > > I hope that this at least partly answers where I see the problems and what > I am trying to solve. > > > > Greetings from Germany, > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: > > Dear Danko, > > > > I have just read your new article and would like to comment briefly about > it. > > > > In your introductory remarks, you write: > > > > "However, connectionism did not yet produce a satisfactory explanation of > how the mental emerges from the physical. A number of open problems > remains ( 5,6,7,8). As a result, the explanatory gap between the mind and > the brain remains wide open." > > > > I certainly believe that no theoretical explanation in science is ever > complete. However, I also believe that "the explanatory gap between the > mind and the brain" does not remain "wide open". > > > > My Magnum Opus, that was published in 2021, makes that belief clear in its > title: > > > > *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* > > > > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > > > > The book provides a self-contained and non-technical exposition in a > conversational tone of many principled and unifying explanations of > psychological and neurobiological data. > > > > In particular, it explains roles for the metabotropic glutamate receptors > that you mention in your own work. See the text and figures around p. 521. > This explanation unifies psychological, anatomical, neurophysiological, > biophysical, and biochemical data about the processes under discussion. > > > > I have a very old-fashioned view about how to understand scientific > theories. I get excited by theories that explain and predict more data than > previous theories. > > > > Which of the data that I explain in my book, and support with quantitative > computer simulations, can you also explain? > > > > What data can you explain, in the same quantitative sense, that you do not > think the neural models in my book can explain? > > > > I would be delighted to discuss these issues further with you. > > > > If you are in touch with my old friend and esteemed colleague, Wolf > Singer, please send him my warm regards. I cite the superb work that he and > various of his collaborators have done in many places in my book. > > > > Best, > > > > Steve > > > > 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 Danko Nikolic > *Sent:* Thursday, July 14, 2022 6:05 AM > *To:* Gary Marcus > *Cc:* connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Gary and everyone, > > > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > > > I am happy to announce that I have now finished a draft of a paper in > which I propose how the brain is able to achieve that. The manuscript > requires a bit of patience for two reasons: one is that the reader may be > exposed for the first time to certain aspects of brain physiology. The > second reason is that it may take some effort to understand the > counterintuitive implications of the new ideas (this requires a different > way of thinking than what we are used to based on connectionism). > > > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > > > (I apologize for the neuroscience lingo, which I tried to minimize.) > > > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > > > Thanks. > > > > Danko > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > > > > > On Thu, Feb 3, 2022 at 5:25 PM 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 pstone at cs.utexas.edu Sat Jul 16 22:51:52 2022 From: pstone at cs.utexas.edu (Peter Stone) Date: Sat, 16 Jul 2022 21:51:52 -0500 Subject: Connectionists: Postdoc opportunity - UT Austin Message-ID: <23062.1658026312@cs.utexas.edu> DEPARTMENT OF COMPUTER SCIENCE, THE UNIVERSITY OF TEXAS AT AUSTIN, USA POSITION: Post-doctoral fellow on Robotics, Multiagent Systems, and Reinforcement Learning CONTACT: Prof. Peter Stone The University of Texas at Austin 2317 Speedway, Stop D9500 Austin, TX 78712 USA pstone at cs.utexas.edu www.cs.utexas.edu/~pstone Applications are invited for a postdoctoral fellow of one year, possibly renewable for additional years, in the Department of Computer Science in the Learning Agents Research Group headed by Prof. Peter Stone. Primary responsibilities include performing cutting-edge research in collaboration with faculty, Ph.D. students, and other researchers. The research will focus on developing and testing novel algorithms for in connection with a range of projects related to robotics and multiagent reinforcement learning. Motivating use cases include long-term autonomous service robots, robot soccer, and adaptive autonomous driving and traffic management. QUALIFICATIONS: Applicants should have a Ph.D. in Computer Science or related field. Experience with machine learning and intelligent robotics is essential. Experience in deep reinforcement learning, multiagent systems, and/or ROS is desired. TO APPLY: Applicants should send by email to pstone at cs.utexas.edu - a curriculum vitae - names of two references with contact information - a two-page summary of past research and relevant qualifications - a personal Web page, if available, where further details can be found This position is to start as early as September of 2022 or at any agreed upon later date. Applications will be reviewed as they are received. From steve at bu.edu Sat Jul 16 11:45:05 2022 From: steve at bu.edu (Grossberg, Stephen) Date: Sat, 16 Jul 2022 15:45:05 +0000 Subject: Connectionists: Neural architectures that embody biological intelligence In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Tsvi, I have no idea why you are writing to me. I would prefer that you did not engage the entire connectionists mailing list. However, since you did, I need to include everyone in my reply. For starters, I have not been an editor of any journal since 2010. When I was Editor-in-Chief of Neural Networks before that, and a new article was submitted, I assigned it to one of over 70 action editors who was a specialist in the topic of the article. That action editor then took full responsibility for getting three reviews of the article. If any reviewer disagreed with other reviewers for a potentially serious reason, then yet another reviewer was typically sought by the action editor in order to try to resolve the difference. Almost always, the reviewers agreed about publication recommendations, so this was not needed. I always followed the recommendations of action editors to publish or not, based upon the above process. I only entered any decision if the action editor solicited my help for a problem for which he/she needed advice. This hardly ever happened. Best, Steve ________________________________ From: Tsvi Achler Sent: Saturday, July 16, 2022 10:51 AM To: Grossberg, Stephen Cc: Asim Roy ; Danko Nikolic ; AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Neural architectures that embody biological intelligence Dear Stephen, I have a connectionist model where feedback takes on a much greater role than the Resonance and other theories. I also have a richer background than many researchers in this field. I have a degree in electrical engineering & computer science, I did my PhD work doing neurophysiology recording neurons and working in a cognitive lab recording differential human reaction times to visual stimulation. I also got an MD focusing on neurology and patients. Consistently throughout the years established academics and their associates have blocked this theory's publication and funding in favor of their own. Since academia is mostly political, this is a big deal. Moreover, it bothers me seeing this done to others. Unfortunately you are by far NOT the worst at doing so, you are just the most transparent about it. I came to the conclusion that academia is not a place to innovate, especially if you come from a multidisciplinary background because (analogous to some of the models) the politics multiply exponentially. Although your work was innovative in the grand scheme of things, what you and other well established academics are doing is not ok. Sincerely, -Tsvi On Sat, Jul 16, 2022 at 12:04 AM Grossberg, Stephen > wrote: Dear Asim and Danko, A lot of your concerns about scaling do not apply to the kinds of biological neural networks that my colleagues and I have developed over the years. You can find a self-contained summary of many of them in my Magnum Opus: https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 As Asim notes below, it is indeed the case that ART can often make good predictions based on small amounts of learned data. This applies as well to large-scale applications naturalistic data. Gail Carpenter and her colleagues have, for example, shown how this works in learning complicated maps of multiple vegetation classes during remote sensing; e.g. Carpenter, G.A., Gopal, S., Macomber, S., Martens, S., & Woodcock, C.E. (1999). A neural network method for mixture estimation for vegetation mapping. Remote Sensing of Environment, 70(2), 138-152. http://techlab.bu.edu/members/gail/articles/127_Mixtures_RSE_1999_.pdf or in learning medical database predictions in response to incomplete, probabilistic, and even incorrect data. In this regard, Gail et al. have also shown how an ART system can incrementally learn a cognitive hierarchy of rules whereby to understand such data; i.e., converts information into knowledge; e.g., Carpenter, G.A., & Ravindran, A. (2008). Unifying multiple knowledge domains using the ARTMAP information fusion system. Proceedings of the 11th International Conference on Information Fusion, Cologne, Germany, June 30 - July 3, 2008. http://techlab.bu.edu/members/gail/articles/155_Fusion2008_CarpenterRavindran.pdf Carpenter, G.A. & Markuzon, N. (1998). ARTMAP?IC and medical diagnosis: Instance counting and inconsistent cases. Neural Networks, 11(2), 323-336. http://techlab.bu.edu/members/gail/articles/117_ARTMAP-IC_1998_.pdf My own work is filled with models that incrementally learn to carry out goal-oriented tasks without regard to scaling concerns. This work develops neural architectures that involve the coordinated actions of many brain regions, not just learned classification. These architectures are supported by unified and principled explanations of lots of psychological and neurobiological data; e.g., Chang, H.-C., Grossberg, S., and Cao, Y. (2014) Where?s Waldo? How perceptual cognitive, and emotional brain processes cooperate during learning to categorize and find desired objects in a cluttered scene. Frontiers in Integrative Neuroscience, doi: 10.3389/fnint.2014.0043, https://www.frontiersin.org/articles/10.3389/fnint.2014.00043/full Grossberg, S., and Vladusich, T. (2010). How do children learn to follow gaze, share joint attention, imitate their teachers, and use tools during social interactions? Neural Networks, 23, 940-965. https://sites.bu.edu/steveg/files/2016/06/GrossbergVladusichNN2010.pdf See Figure 1 in the following article to get a sense of how many brain processes other than classification are needed to realize true biological intelligence: Grossberg, S. (2018). Desirability, availability, credit assignment, category learning, and attention: Cognitive-emotional and working memory dynamics of orbitofrontal, ventrolateral, and dorsolateral prefrontal cortices. Brain and Neuroscience Advances, May 8, 2018. https://journals.sagepub.com/doi/full/10.1177/2398212818772179 Best, Steve ________________________________ From: Asim Roy > Sent: Friday, July 15, 2022 9:35 AM To: Danko Nikolic > Cc: Grossberg, Stephen >; Gary Marcus >; AIhub >; connectionists at mailman.srv.cs.cmu.edu >; 'maxversace at gmail.com' > Subject: RE: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, 1. I am not sure if I myself know all the uses of a knife, leave aside countless ones. Given a particular situation, I might simulate in my mind about the potential usage, but I doubt our minds explore all the countless situations of usage of an object as soon as it learns about it. 2. I am not sure if a 2 or 3 year old child, after having ?learnt? about a knife, knows very many uses of it. I doubt the kid is awake all night and day simulating in the brain how and where to use such a knife. 3. ?Understanding? is a loaded term. I think it needs a definition. 4. I am copying Max Versace, a student of Steve Grossberg. His company markets a software that can learn quickly from a few examples. Not exactly one-shot learning, it needs a few shots. I believe it?s a variation of ART. But Max can clarify the details. And Tsvi is doing similar work. So, what you are asking for may already exist. So linear scaling may be the worst case scenario. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Danko Nikolic > Sent: Friday, July 15, 2022 12:19 AM To: Asim Roy > Cc: Grossberg, Stephen >; Gary Marcus >; AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I agree about the potential for linear scaling of ART and.other connectionist systems. However, there are two problems. The problem number one kills it already and this is that the real brain scales a lot better than linearly: For each new object learned, we are able to resolve countless many new situations in which this object takes part (e.g., finding various uses for a knife, many of which may be new, ad hoc -- this is a great ability of biological minds often referred to as 'understanding'). Hence, simple linear scaling by adding more neurons for additional objects is not good enough to match biological intelligence. The second problem becomes an overkill, and this is that linear scaling in connectionist systems works only in theory, under idealized conditions. In real life, say if working with ImageNet, the scaling turns into a power-law with an exponent much larger than one: We need something like 500x more resources just to double the number of objects. Hence, in practice, the demands for resources explode if you want to add more categories whilst not losing the accuracy. To summarize, there is no linear scaling in practice nor would linear scaling suffice, even if we found one. This should be a strong enough argument to search for another paradigm, something that scales better than connectionism. I discuss both problems in the new manuscript, and even track a bit deeper the problem of why connectionism lacks linear scaling in practice (I provide some revealing computations in the Supplementary Materials (with access to the code), although much more work needs to be done). Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Thu, Jul 14, 2022 at 11:48 PM Asim Roy > wrote: I think Steve Grossberg is generally correct. ART, RBF nets and similar flexible architectures can scale up almost linearly to problem size. My conjecture is that architectures that use distributed representation run into the scaling issue. On the other hand, however, distributed representation produces a more compact architecture compared to the shallow architectures of ART and RBFs. But, in terms of adding concepts, it?s easy to add a new object or concept to the shallow architectures. Perhaps someone can provide more insights on the architectural differences and the corresponding pros and cons of each. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Grossberg, Stephen Sent: Thursday, July 14, 2022 10:14 AM To: Danko Nikolic > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, I will respond to your comment below to the entire list. I recommend that future interactions be done between just the two of us. Everything that I write below is summarized in my new book: In brief, within Adaptive Resonance Theory, or ART, there IS no problem of scaling, in the sense that you describe it below, from learning (say) to correctly recognizing 100 objects to doing the same for 200. In fact, one of the properties for which I introduced ART in 1976 was to enable incremental learning in real time of arbitrary numbers of objects or events without experiencing catastrophic forgetting. I have called this a solution of the stability-plasticity dilemma: How our brains, and models like ART, can rapidly learn (plasticity) arbitrary numbers of objects without experiencing catastrophic forgetting (stability). I also derived ART in 1980, in an article within Psychological Review, from a thought experiment about how ANY system can AUTONOMOUSLY correct predictive errors in a changing world that is filled with unexpected events. This thought is experiment was derived from a few facts that are familiar to us all. They are familiar because they represent ubiquitous environmental pressures that we all experience. The thought experiment clarifies why, when they act together during the evolutionary process, models like ART are the unique result. Moveover, at no point during the thought experiment are the words mind or brain mentioned. ART is thus a universal solution of this learning, classification, and prediction problem. You write below about "connectionist systems". ART is a connectionist system. What do you mean by a "connectionist system"? What you write below about them is not true in general. Best, Steve ________________________________ From: Danko Nikolic > Sent: Thursday, July 14, 2022 12:16 PM To: Grossberg, Stephen > Cc: Gary Marcus >; connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 jose at rubic.rutgers.edu Sat Jul 16 12:55:25 2022 From: jose at rubic.rutgers.edu (=?utf-8?B?U3RlcGhlbiBKb3PDqSBIYW5zb24=?=) Date: Sat, 16 Jul 2022 16:55:25 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <19582f9f-e0b0-fb86-5e11-6fc8f6c9dca4@rubic.rutgers.edu> yeah, yeah yeah.. Gary.. read them both out-loud, carefully, to say, a non-linguistic sentient being.. maybe a cat, and it will dis-habituate to the second one! ? Steve On 7/16/22 12:47 PM, Gary Marcus wrote: I can?t help but note a profound tension between these two very recent quotes: Hanson, 2022, below: as we go forward, let?s avoid at all costs breaking ?the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet.? LeCun and Browning, 2022: ?everyone working in DL agrees that symbolic manipulation is a necessary feature for creating human-like AI.? https://www.noemamag.com/what-ai-can-tell-us-about-intelligence On Jul 16, 2022, at 12:11 AM, Stephen Jos? Hanson wrote: ? So there are a number of issues here. In the first one, the brain obviously has a specific structure. DL supervised modeling of classification, NOT psychological categorization-- may scale linearly with resources.. however, its clear just adding more hidden layers has diminishing returns, adding various kinds of classification structure (GANs etc) or doing unsupervised learning or reinforcement learning is not going to end up with AGI or really any sort of final AI system and certainly not biological. Humans btw, scale as sub-linear systems.. but with hard limits. We can only recognize about 6000 faces (and the variance is about 4k across individuals), most undergraduates can recognize 100k words again with a large variance., so despite the fact that the brain has trillions of synapses, "loading" in explicit (MTL) memory is complex and somehow localized. Memory systems involve implicit (most DL) or explicit (almost noone), working memory is tiny--like 3 or 4+- 2 (sorry George). Many (Bengio, Le Cun etc) are experimenting with more complex structures, but keeping the system homogeneous, no extra symbolic baggage and forcing the exercise to have to learn the configuration rather than engineer it. In some ways, Grossberg has been out in front in all this, but with his own special and situated neural systems. He has been consistently and persistently creating focused neural modeling, which as an art form has always been a matter of taste. Connectionist models tend toward minimalist forms and Steve and others follow more Baroque or perhaps Victorian forms. What is clear is that the future is going to be blends of these forms and more complex, without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet. Steve On 7/15/22 4:01 AM, Dietterich, Thomas wrote: Dear Danko, In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. --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: Connectionists On Behalf Of Danko Nikolic Sent: Thursday, July 14, 2022 09:17 To: Grossberg, Stephen 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.] Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -------------- next part -------------- An HTML attachment was scrubbed... URL: From nemanja at temple.edu Sat Jul 16 20:20:04 2022 From: nemanja at temple.edu (Nemanja Djuric) Date: Sun, 17 Jul 2022 00:20:04 +0000 Subject: Connectionists: 2nd CfP - AVVision @ ECCV2022 Message-ID: Call for Workshop Papers AVVision 2022 in conjunction with 17th European Conference on Computer Vision (ECCV 2022) Tel-Aviv, Israel, October 23rd-27th, 2022 https://avvision.xyz/eccv22/ The 3rd Autonomous Vehicle Vision (AVVision) Workshop aims to bring together industry professionals and academics to brainstorm and exchange ideas on the advancement of computer vision techniques for autonomous driving. In this one-day workshop, we will have several keynote talks and regular paper presentations (oral and poster) to discuss the state-of-the-art as well as existing challenges in the field of autonomous driving. The workshop webpage is at https://avvision.xyz/eccv22. Call for papers: With a number of breakthroughs in autonomous system technology over the past decade, the race to commercialize self-driving cars has become fiercer than ever. The integration of advanced sensing, computer vision, signal/image processing, and machine/deep learning into autonomous vehicles enables them to perceive the environment intelligently and navigate safely. Autonomous driving is required to ensure safe, reliable, and efficient automated mobility in complex uncontrolled real-world environments. Various applications range from automated transportation and farming to public safety and environment exploration. Visual perception is a critical component of autonomous driving. Enabling technologies include: a) affordable sensors that can acquire useful data under varying environmental conditions, b) reliable simultaneous localization and mapping, c) machine learning that can effectively handle varying real-world conditions and unforeseen events, as well as ?machine-learning friendly? signal processing to enable more effective classification and decision making, d) hardware and software co-design for efficient real-time performance, e) resilient and robust platforms that can withstand adversarial attacks and failures, and f) end-to-end system integration of sensing, computer vision, signal/image processing and machine/deep learning. The 3rd AVVision workshop will cover all these topics. Research papers are solicited in, but not limited to, the following topics: * 3D road/environment reconstruction and understanding; * Mapping and localization for autonomous cars; * Semantic/instance driving scene segmentation and semantic mapping; * Self-supervised/unsupervised visual environment perception; * Car/pedestrian/object/obstacle detection/tracking and 3D localization; * Car/license plate/road sign detection and recognition; * Driver status monitoring and human-car interfaces; * Deep/machine learning and image analysis for car perception; * Adversarial domain adaptation for autonomous driving; * On-board embedded visual perception systems; * Bio-inspired vision sensing for car perception; * Real-time deep learning inference. Important Dates: * Paper submission deadline: July 31th, 2022 * Review feedback release date: August 16th, 2022 * Camera-ready submission: August 22nd, 2022 * Workshop date: Oct. 23-24th, 2022 (exact date TBD) Submission Guidelines: Full papers: Authors are encouraged to submit high-quality, original (i.e., not previously published or accepted for publication in substantially similar form in any peer-reviewed venue including journal, conference, or workshop) research. The paper template is identical to the one used at ECCV 2022 main conference (see the ?Submission Guidelines? section). Papers are limited to 14 pages, including figures and tables, in the ECCV style. Additional pages containing only cited references are allowed. ??Extended abstracts: We encourage participants to submit preliminary ideas that have not been published before as extended abstracts. These submissions would benefit from additional exposure and discussion that can shape a better future publication. We also invite papers that have been published at other venues to spark discussions and foster new collaborations. Submissions may consist of up to 7 pages plus one additional page solely for references (using the template detailed above). The extended abstracts do NOT need to be anonymized and will NOT be published in the workshop proceedings. Papers that are not properly anonymized, or do not use the template, or have more than the allowed number of pages will be rejected without review. The submission site is now open. Organizers: * Rui Ranger Fan, Tongji University * Nemanja Djuric, Aurora Innovation * Wenshuo Wang, McGill University * Peter Ondruska, Lyft * Jie Li, Toyota Research Institute -------------- next part -------------- An HTML attachment was scrubbed... URL: From david at irdta.eu Sat Jul 16 05:30:01 2022 From: david at irdta.eu (David Silva - IRDTA) Date: Sat, 16 Jul 2022 11:30:01 +0200 (CEST) Subject: Connectionists: DeepLearn 2022 Summer: regular registration July 22 Message-ID: <128631297.2076280.1657963802052@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 ****************************************************************** Regular registration: July 22, 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 21 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 [virtual] PROFESSORS AND COURSES: 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 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 Tor Lattimore (DeepMind), [intermediate/advanced] Tools and Techniques of Reinforcement Learning to Overcome Bellman's Curse of Dimensionality Vincent Lepetit (Paris Institute of Technology), [intermediate] Deep Learning and 3D Reasoning for 3D Scene Understanding 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 Clarisa 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] Learning a Representation of the Visual World with Neural Networks [virtual] Johan Suykens (KU Leuven), [introductory/intermediate] Deep Learning, Neural Networks and Kernel Machines 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 are available 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 - Fundaci?n Parque Cient?fico Tecnol?gico Universitat Rovira i Virgili Institute for Research Development, Training and Advice ? IRDTA, Brussels/London -------------- next part -------------- An HTML attachment was scrubbed... URL: From eloise.zehnder at univ-lorraine.fr Mon Jul 18 02:17:07 2022 From: eloise.zehnder at univ-lorraine.fr (Eloise Zehnder) Date: Mon, 18 Jul 2022 08:17:07 +0200 (CEST) Subject: Connectionists: =?utf-8?q?=5BExtended_deadline=5D=5BCFP=5D=5BMeet?= =?utf-8?q?ings=5D_RoMan2022_ONLINE_WORKSHOP_CFP_-The_=E2=80=9CTowards_Soc?= =?utf-8?q?ially_Intelligent_Robots_In_Real-World_Applications=3A_Challeng?= =?utf-8?q?es_And_Intricacies=E2=80=9D_=28SIRRW=29?= In-Reply-To: <874769966.44208.1658124151510.JavaMail.zimbra@univ-lorraine.fr> References: <213068046.445934.1656664994345.JavaMail.zimbra@univ-lorraine.fr> <874769966.44208.1658124151510.JavaMail.zimbra@univ-lorraine.fr> Message-ID: <2021398983.57053.1658125027467.JavaMail.zimbra@univ-lorraine.fr> [Apologies for cross-posting] Hello Everyone, We are happy to invite you to submit a paper to our upcoming full-day ONLINE WORKSHOP at IEEE Ro-Man 2022 . The ?Towards Socially Intelligent Robots In Real-World Applications: Challenges And Intricacies? (SIRRW) workshop is your opportunity to present your work and get an overview of state-of-the-art research in human-centered, real-life settings. Researchers from all HRI/Cognitive robotics/Machine Learning-relevant fields are invited to submit a 2-pages abstract or a 4 pages short-paper, before July 15th July 31st , 2022 . Selected papers will be presented as either an oral presentation or in a poster session. For more information, please see our website: [ https://sirrw-2022.github.io/ | https://sirrw-2022.github.io ] Overview ================ Being proactive, trustworthy, and dealing with uncertainty are the current challenges faced by social robots in real-world applications. To be seamlessly integrated into human-populated environments, robots will be expected to have intelligent social capabilities on top of their physical abilities. Human-Centered Artificial Intelligence has a critical role in providing robots with such capabilities by addressing some of the complex computational challenges that naturalistic human-robot interactions introduce. For a robot to behave proactively and in a manner that is appropriate to the context of interaction, it should cope with uncertainty when dealing with elements that are not fully observable and often hard to predict, such as the states representing the dynamic environment and humans. To build trustworthy interactions with humans, intelligent social robots need to successfully address challenging issues such as predicting human intentions, goals, expectations, understanding, and reasoning about the dynamic states of objects and other smart devices in the surroundings, and previous actions and their consequences while performing in complex situations. Trust is an important construct for evaluating adaptive social robot behaviors that could be inferred and evaluated through objective measures using computational models and subjective measures by human users. Such measures can be used to assess humans? disposition to be vulnerable around robots. Hence, addressing uncertainty is a key factor in developing trustworthy AI solutions and endowing robots with intelligent social capabilities. Important Dates ================ July 15th July 31st : Paper submission deadline August 5th : Acceptance notification August 17th : Camera-ready submission August 24th : Workshop Submissions ================ * Papers will be submitted through [ https://easychair.org/account/signin?l=BbbUuHtL2mSjVI7LN8iC06 | EasyChair ] * The manuscript should be of 2 or 4 pages in [ https://ras.papercept.net/conferences/support/support.php | IEEE double-column format ] excluding references. Relevant topics will include (but not restricted to): - Decision-making under uncertainty - Modeling human behavior - Communicative robot behavior generation - Automatic adaptation and personalization of robot behavior - Human-interactive robot learning - Planning methods for interactive robot behaviors - Perception for HRI - Cognitive architectures for interactive robots - Robot curiosity - Safety/Trust-critical applications for HRI - Human-robot collaboration - Reliability and explainability of robot decisions/actions Invited Speakers ================ * Prof. Agnieszka Wykowska (Instituto Italiano di Technologia (IIT)) * Prof. Tony Belpaeme (Ghent University and Plymouth University) * Dr. Dan Bohus (Microsoft) * Prof. Alan R. Wagner (Pennsylvania State University) Organizers and Contact ================ Dr. Melanie Jouaiti ( [ mailto:mjouaiti at uwaterloo.ca | mjouaiti at uwaterloo.ca ] , University of Waterloo) Dr. Sera Buyukgoz ( [ mailto:serabuyukgoz at gmail.com | serabuyukgoz at gmail.com ] , Sorbonne Universit? and Softbank Robotics Europe) Eloise Zehnder ( [ mailto:eloise.zehnder at univ-lorraine.fr | eloise.zehnder at univ-lorraine.fr ] , Universit? de Lorraine and Inria, Loria) Dr. Amir Aly ( [ mailto:amir.aly at plymouth.ac.uk | amir.aly at plymouth.ac.uk ] , Plymouth University) Pr. Kerstin Dautenhahn ( [ mailto:kerstin.dautenhahn at uwaterloo.ca | kerstin.dautenhahn at uwaterloo.ca ] , University of Waterloo) -------------- next part -------------- An HTML attachment was scrubbed... URL: From battleday at princeton.edu Mon Jul 18 03:24:21 2022 From: battleday at princeton.edu (Ruairidh McLennan Battleday) Date: Mon, 18 Jul 2022 00:24:21 -0700 Subject: Connectionists: Reminder: Mathematics of Neuroscience Symposium, Crete, Greece, 24-25th September 2022 Message-ID: Dear All, Just a reminder that there is under a week left to apply to present at the Mathematics of Neuroscience symposium to be held this year in September on Crete, Greece. We have a great list of confirmed speakers and would love to hear from you too. Deadline for submission is the 23rd of July. ------------------ Symposium on the Mathematics of Neuroscience. Crete, Greece, 24-25th September 2022 (www.neuromonster.org). Two decades into the 21st century, can we claim to be any closer to a unified model of the brain? In this exploratory symposium, we invite submissions for short talks and posters presenting general mathematical models of brain function. We give priority to those models that account for brain or behavioural data, or provide simulations to that effect. This year?s theme is life-long learning and discovery. Keynote Speakers Professor Peter Dayan (Director, Department of Computational Neuroscience, Max Planck Institute for Biological Cybernetics, T?bingen) ?Learning from scratch: Non-parametric models of task acquisition over the long run? Professor Andrew Adamatzky (Director, Unconventional Computing Laboratory, University of the West of England) ?Fungal Brain? Symposium Chairs Professor Dan V. Nicolau (King?s College London) Dr Ruairidh McLennan Battleday (Princeton University) Confirmed Talks Professor Michael Levin (TUFTS) Professor Kobi Kremnitzer (University of Oxford) Professor Marc Howard (Boston University) Professor Kevin Burrage (Queensland University of Technology) Professor Carina Curto (Penn State) Professor Rahul Bhui (MIT) Dr Jonathan Mason (University of Oxford) Dr James Whittington (University of Oxford / Stanford) Dr Ilia Sucholutsky (Princeton University) Dr Sophia Sanborn (UC Berkeley, UC Santa Barbara, University of British Columbia) Dr Christina Merrick (UC San Francisco) Dr Timothy Muller (UCL) Dr Kamila Maria J??wik (University of Cambridge) Dr Chris Hillar (Awecom, Inc) Prize Talks Dr Aenne Brielmann (Max Planck Institute, T?bingen) Andrew Ligeralde (Redwood Center for Theoretical Neuroscience, University of California, Berkeley) The symposium will be held virtually or in-person on the island of Crete, Greece from the 24-25th of September 2022 (www.neuromonster.org). Submission is by 250-word abstract before the 23rd July 2022, emailed to the organizers Professor Dan V. Nicolau Jr (dan.nicolau at kcl.ac.uk) and Dr Ruairidh M. Battleday (battleday at princeton.edu). -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Sun Jul 17 12:23:14 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sun, 17 Jul 2022 16:23:14 +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> <9B59973F-F4F1-4935-AE5B-B59039ADA7F8@asu.edu> Message-ID: Steve, I meant the hybrid systems are the way forward. And it does use explicit representation at some point. And it has enormous benefits. Asim From: Stephen Jos? Hanson Sent: Sunday, July 17, 2022 9:17 AM To: Asim Roy Cc: Dietterich, Thomas ; Danko Nikolic ; Grossberg, Stephen ; AIhub ; connectionists at mailman.srv.cs.cmu.edu; Gary Marcus ; Yoshua Bengio ; Geoffrey Hinton Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Asim, with all due respect, I think you are missing the point. The enormous recent progress of DL is what's issue. The way forward, is already happening. Hybrid system integration is simply an admission that the explicit representations being developed for DL have failed. We already know how our invented symbolic systems can add 2 and 2.. not so clear how the brain does it. I think this is where the confusion lay. Steve (Btw you may not know, but I am an Alum of ASU). On 7/17/22 12:05 PM, Asim Roy wrote: It is not a slippery slope. It already has good traction and produces great results. It?s the way forward. Sent from my iPhone On Jul 17, 2022, at 6:00 PM, Stephen Jos? Hanson wrote: ? Not really. the distinction is in the details.. neural modeling involves biological constraints, learning.. OTOH, hybrid systems, as Gary and some others are promoting is more than a slippery slope.. its already off the cliff. Someone said once to me ..if you have a perfectly good electric car.. why would you have it push a car with a combustion engine around?" why not improve the EV without giving up on the energy source? S On 7/17/22 11:10 AM, Asim Roy wrote: ?without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet.? Those who are aware, Gary Marcus? symbolic parts are already tumbling out of the closet. There are numerous efforts, even by Hinton and Bengio, that try to encode a symbolic concept using a set of neurons. All the best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Stephen Jos? Hanson Sent: Friday, July 15, 2022 4:28 AM To: Dietterich, Thomas ; Danko Nikolic ; Grossberg, Stephen Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton So there are a number of issues here. In the first one, the brain obviously has a specific structure. DL supervised modeling of classification, NOT psychological categorization-- may scale linearly with resources.. however, its clear just adding more hidden layers has diminishing returns, adding various kinds of classification structure (GANs etc) or doing unsupervised learning or reinforcement learning is not going to end up with AGI or really any sort of final AI system and certainly not biological. Humans btw, scale as sub-linear systems.. but with hard limits. We can only recognize about 6000 faces (and the variance is about 4k across individuals), most undergraduates can recognize 100k words again with a large variance., so despite the fact that the brain has trillions of synapses, "loading" in explicit (MTL) memory is complex and somehow localized. Memory systems involve implicit (most DL) or explicit (almost noone), working memory is tiny--like 3 or 4+- 2 (sorry George). Many (Bengio, Le Cun etc) are experimenting with more complex structures, but keeping the system homogeneous, no extra symbolic baggage and forcing the exercise to have to learn the configuration rather than engineer it. In some ways, Grossberg has been out in front in all this, but with his own special and situated neural systems. He has been consistently and persistently creating focused neural modeling, which as an art form has always been a matter of taste. Connectionist models tend toward minimalist forms and Steve and others follow more Baroque or perhaps Victorian forms. What is clear is that the future is going to be blends of these forms and more complex, without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet. Steve On 7/15/22 4:01 AM, Dietterich, Thomas wrote: Dear Danko, In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. --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: Connectionists On Behalf Of Danko Nikolic Sent: Thursday, July 14, 2022 09:17 To: Grossberg, Stephen 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.] Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -------------- next part -------------- An HTML attachment was scrubbed... URL: From m.biehl at rug.nl Mon Jul 18 03:01:47 2022 From: m.biehl at rug.nl (Michael Biehl) Date: Mon, 18 Jul 2022 09:01:47 +0200 Subject: Connectionists: International workshop AMALEA, 12-16 Sept 2022, Cetraro, Italy Message-ID: On behalf of Prof. Dr. Nicolai Petkov (www.cs.rug.nl/~petkov) *International workshop on Applications of Machine Learning and Neural Networks* *AMALEA, 12-16 Sept 2022, Cetraro, Calabria, Italy* Dear colleagues and friends, There are a few rooms still available for AMALEA - APPLICATIONS OF MACHINE LEARNING International Workshop, Cetraro, Italy, September 12-16, 2022 http://amalea.web.rug.nl/index.html Venue: Grand Hotel San Michele on the West, Tyrrhenian sea coast of Southern Italy with a surrounding own land estate, golf course, private beach and a conference center. Photo gallery Aim: to provide a discussion forum for scientists from different disciplines where they can inspire and learn from each other, meet new people and form new research alliances. Format: a single track for talks from 10:00-12:30H and 17:00-19:30H in the conference center. Enough time for discussions outside the conference room: during breakfast on the terrace overlooking the Tyrrhenian sea, at lunch in the private beach restaurant, in the afternoon break on the private beach, at dinner on the hotel terrace and after dinner listening to the life piano music in the grand salon. Please visit the workshop website for further information and contact. -- --------------------------------------------------- Prof. Dr. Michael Biehl Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence P.O. Box 407, 9700 AK Groningen The Netherlands https://www.cs.rug.nl/~biehl m.biehl at rug.nl -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Mon Jul 18 03:28:26 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Mon, 18 Jul 2022 09:28:26 +0200 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, Thank you for inquiring about the generalized XOR? To answer such question, I wrote a paper. So, please read the paper. Everything should be explained over there better than what I can do in a short message (there are additional details in the supplementary materials; also the code on github). In short, learning mechanisms cannot discover generalized XOR functions with simple connectivity -- only with complex connectivity. This problem results in exponential growth of needed resources as the number of bits in the generalized XOR increases. Remember, we are talking about a generalized XOR (more than two bits as inputs). The paper: https://bit.ly/3IFs8Ug As you will see, although connectionism can scale linearly in theory, in practice no learning mechanism seems to exist that could discover the needed connections. This would require super-smart learning mechanisms, but such mechanisms do not exist. As a result, the whole thing fails. And, as I mentioned before, even the linear scaling would not be enough to match the biological brain. We need something a lot more powerful than linear scaling. This is my argument on why connectionism fails; it fails even with linear scaling, which it cannot achieve anyway in practice. Again, these are not just empty works; I provide evidence for that in the manuscript. The good news is that there seems to be a solution: transient selection of subnetworks, which I characterize in the same paper. So, the future of AI looks nevertheless bright, I think. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Mon, Jul 18, 2022 at 12:59 AM gary at ucsd.edu wrote: > Sorry, I can't let this go by: > > And it is good so because generalize XOR scales worse than power law. It > scales exponentially! This a more agressive form of explosion than power > law. > > I'm not sure exactly what you mean by this, but a single-hidden layer > network with N inputs and N hidden units can solve N-bit parity. Each unit > has an increasing threshold, so, one turns on if there is one unit on in > the input, and then turns on the output with a weight of +1. If two units > are on in the input, then a second unit comes on and cancels the activation > of the first unit via a weight of -1. Etc. > > g. > > > On Sat, Jul 16, 2022 at 12:03 AM Danko Nikolic > wrote: > >> Dear Thomas, >> >> Thank you for reading the paper and for the comments. >> >> I cite: "In my experience, supervised classification scales linearly in >> the number of classes." >> This would be good to quantify as a plot. Maybe a research paper would be >> a good idea. The reason is that it seems that everyone else who tried to >> quantify that relation found a power law. At this point, it would be >> surprising to find a linear relationship. And it would probably make a well >> read paper. >> >> But please do not forget that my argument states that even a linear >> relationship is not good enough to match bilogical brains. We need >> something more similar to a power law with exponent zero when it comes to >> the model size i.e., a constant number of parameters in the model. And we >> need linear relationship when it comes to learning time: Each newly learned >> object should needs about as much of learning effort as was needed for each >> previous object. >> >> I cite: "The real world is not dominated by generalized XOR problems." >> Agreed. And it is good so because generalize XOR scales worse than power >> law. It scales exponentially! This a more agressive form of explosion than >> power law. >> Importantly, a generalized AND operation also scales exponentially (with >> a smaller exponent, though). I guess we would agree that the real world >> probably encouners a lot of AND problems. The only logical operaiton that >> could be learned with a linear increase in the number of parameters was a >> generalized OR. Finally, I foiund that a mixure of AND and OR resulted in a >> power law-like scaling of the number of parameters. So, a mixture of AND >> and OR seemed to scale as good (or as bad) as the real world. I have put >> this information into Supplementary Materials. >> >> The conclusion that I derived from those analyses is: connectionism is >> not sustainable to reach human (or animal) levels of intelligence. >> Therefore, I hunted for an alternative pradigm. >> >> Greetings, >> >> Danko >> >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> https://www.linkedin.com/in/danko-nikolic/ >> -- I wonder, how is the brain able to generate insight? -- >> >> >> On Fri, Jul 15, 2022 at 10:01 AM Dietterich, Thomas >> wrote: >> >>> Dear Danko, >>> >>> >>> >>> In my experience, supervised classification scales linearly in the >>> number of classes. Of course it depends to some extent on how subtle the >>> distinctions are between the different categories. The real world is not >>> dominated by generalized XOR problems. >>> >>> >>> >>> --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:* Connectionists *On >>> Behalf Of *Danko Nikolic >>> *Sent:* Thursday, July 14, 2022 09:17 >>> *To:* Grossberg, Stephen >>> *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.] >>> >>> Dear Steve, >>> >>> >>> >>> Thank you very much for your message and for the greetings. I will pass >>> them on if an occasion arises. >>> >>> >>> >>> Regarding your question: The key problem I am trying to address and >>> that, to the best of my knowledge, no connectionist system was able to >>> solve so far is that of scaling the system's intelligence. For example, if >>> the system is able to correctly recognize 100 different objects, how many >>> additional resources are needed to double that to 200? All the empirical >>> data show that connectionist systems scale poorly: Some of the best systems >>> we have require 500x more resources in order to increase the intelligence >>> by only 2x. I document this problem in the manuscript and even run some >>> simulations to show that the worst performance is if connectionist systems >>> need to solve a generalized XOR problem. >>> >>> >>> >>> In contrast, the biological brain scales well. This I also quantify in >>> the paper. >>> >>> >>> >>> I will look at the publication that you mentioned. However, so far, I >>> haven't seen a solution that scales well in intelligence. >>> >>> >>> >>> My argument is that transient selection of subnetworks by the help of >>> the mentioned proteins is how intelligence scaling is achieved in >>> biological brains. >>> >>> >>> >>> In short, intelligence scaling is the key problem that concerns me. I >>> describe the intelligence scaling problem in more detail in this book that >>> just came out a few weeks ago and that is written for practitioners in Data >>> Scientist and AI: https://amzn.to/3IBxUpL >>> >>> >>> >>> >>> I hope that this at least partly answers where I see the problems and >>> what I am trying to solve. >>> >>> >>> >>> Greetings from Germany, >>> >>> >>> >>> Danko >>> >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> >>> --- A progress usually starts with an insight --- >>> >>> >>> >>> >>> >>> On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: >>> >>> Dear Danko, >>> >>> >>> >>> I have just read your new article and would like to comment briefly >>> about it. >>> >>> >>> >>> In your introductory remarks, you write: >>> >>> >>> >>> "However, connectionism did not yet produce a satisfactory explanation >>> of how the mental emerges from the physical. A number of open problems >>> remains ( 5,6,7,8). As a result, the explanatory gap between the mind and >>> the brain remains wide open." >>> >>> >>> >>> I certainly believe that no theoretical explanation in science is ever >>> complete. However, I also believe that "the explanatory gap between the >>> mind and the brain" does not remain "wide open". >>> >>> >>> >>> My Magnum Opus, that was published in 2021, makes that belief clear in >>> its title: >>> >>> >>> >>> *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* >>> >>> >>> >>> https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 >>> >>> >>> >>> >>> The book provides a self-contained and non-technical exposition in a >>> conversational tone of many principled and unifying explanations of >>> psychological and neurobiological data. >>> >>> >>> >>> In particular, it explains roles for the metabotropic glutamate >>> receptors that you mention in your own work. See the text and figures >>> around p. 521. This explanation unifies psychological, anatomical, >>> neurophysiological, biophysical, and biochemical data about the processes >>> under discussion. >>> >>> >>> >>> I have a very old-fashioned view about how to understand scientific >>> theories. I get excited by theories that explain and predict more data than >>> previous theories. >>> >>> >>> >>> Which of the data that I explain in my book, and support with >>> quantitative computer simulations, can you also explain? >>> >>> >>> >>> What data can you explain, in the same quantitative sense, that you do >>> not think the neural models in my book can explain? >>> >>> >>> >>> I would be delighted to discuss these issues further with you. >>> >>> >>> >>> If you are in touch with my old friend and esteemed colleague, Wolf >>> Singer, please send him my warm regards. I cite the superb work that he and >>> various of his collaborators have done in many places in my book. >>> >>> >>> >>> Best, >>> >>> >>> >>> Steve >>> >>> >>> >>> 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 Danko Nikolic >>> *Sent:* Thursday, July 14, 2022 6:05 AM >>> *To:* Gary Marcus >>> *Cc:* connectionists at mailman.srv.cs.cmu.edu < >>> connectionists at mailman.srv.cs.cmu.edu>; AIhub >>> *Subject:* Re: Connectionists: Stephen Hanson in conversation with >>> Geoff Hinton >>> >>> >>> >>> Dear Gary and everyone, >>> >>> >>> >>> I am continuing the discussion from where we left off a few months ago. >>> Back then, some of us agreed that the problem of understanding remains >>> unsolved. >>> >>> >>> >>> As a reminder, the challenge for connectionism was to 1) learn with few >>> examples and 2) apply the knowledge to a broad set of situations. >>> >>> >>> >>> I am happy to announce that I have now finished a draft of a paper in >>> which I propose how the brain is able to achieve that. The manuscript >>> requires a bit of patience for two reasons: one is that the reader may be >>> exposed for the first time to certain aspects of brain physiology. The >>> second reason is that it may take some effort to understand the >>> counterintuitive implications of the new ideas (this requires a different >>> way of thinking than what we are used to based on connectionism). >>> >>> >>> >>> In short, I am suggesting that instead of the connectionist paradigm, we >>> adopt transient selection of subnetworks. The mechanisms that transiently >>> select brain subnetworks are distributed all over the nervous system and, I >>> argue, are our main machinery for thinking/cognition. The surprising >>> outcome is that neural activation, which was central in connectionism, now >>> plays only a supportive role, while the real 'workers' within the brain are >>> the mechanisms for transient selection of subnetworks. >>> >>> >>> >>> I also explain how I think transient selection achieves learning with >>> only a few examples and how the learned knowledge is possible to apply to a >>> broad set of situations. >>> >>> >>> >>> The manuscript is made available to everyone and can be downloaded here: >>> https://bit.ly/3IFs8Ug >>> >>> >>> (I apologize for the neuroscience lingo, which I tried to minimize.) >>> >>> >>> >>> It will likely take a wide effort to implement these concepts as an AI >>> technology, provided my ideas do not have a major flaw in the first place. >>> Does anyone see a flaw? >>> >>> >>> >>> Thanks. >>> >>> >>> >>> Danko >>> >>> >>> >>> >>> Dr. Danko Nikoli? >>> www.danko-nikolic.com >>> >>> https://www.linkedin.com/in/danko-nikolic/ >>> >>> >>> >>> >>> >>> >>> On Thu, Feb 3, 2022 at 5:25 PM 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.marcus at nyu.edu Sat Jul 16 12:47:51 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Sat, 16 Jul 2022 09:47:51 -0700 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: I can?t help but note a profound tension between these two very recent quotes: Hanson, 2022, below: as we go forward, let?s avoid at all costs breaking ?the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet.? LeCun and Browning, 2022: ?everyone working in DL agrees that symbolic manipulation is a necessary feature for creating human-like AI.? https://www.noemamag.com/what-ai-can-tell-us-about-intelligence > On Jul 16, 2022, at 12:11 AM, Stephen Jos? Hanson wrote: > > ? > So there are a number of issues here. In the first one, the brain obviously has a specific structure. DL supervised modeling of classification, NOT psychological categorization-- may scale linearly with resources.. however, its clear just adding more hidden layers has diminishing returns, adding various kinds of classification structure (GANs etc) or doing unsupervised learning or reinforcement learning is not going to end up with AGI or really any sort of final AI system and certainly not biological. Humans btw, scale as sub-linear systems.. but with hard limits. We can only recognize about 6000 faces (and the variance is about 4k across individuals), most undergraduates can recognize 100k words again with a large variance., so despite the fact that the brain has trillions of synapses, "loading" in explicit (MTL) memory is complex and somehow localized. Memory systems involve implicit (most DL) or explicit (almost noone), working memory is tiny--like 3 or 4+- 2 (sorry George). Many (Bengio, Le Cun etc) are experimenting with more complex structures, but keeping the system homogeneous, no extra symbolic baggage and forcing the exercise to have to learn the configuration rather than engineer it. In some ways, Grossberg has been out in front in all this, but with his own special and situated neural systems. He has been consistently and persistently creating focused neural modeling, which as an art form has always been a matter of taste. Connectionist models tend toward minimalist forms and Steve and others follow more Baroque or perhaps Victorian forms. What is clear is that the future is going to be blends of these forms and more complex, without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet. > > Steve > >> On 7/15/22 4:01 AM, Dietterich, Thomas wrote: >> Dear Danko, >> >> In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. >> >> --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: Connectionists On Behalf Of Danko Nikolic >> Sent: Thursday, July 14, 2022 09:17 >> To: Grossberg, Stephen >> 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.] >> >> Dear Steve, >> >> Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. >> >> Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. >> >> In contrast, the biological brain scales well. This I also quantify in the paper. >> >> I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. >> >> My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. >> >> In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL >> >> I hope that this at least partly answers where I see the problems and what I am trying to solve. >> >> Greetings from Germany, >> >> Danko >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> https://www.linkedin.com/in/danko-nikolic/ >> --- A progress usually starts with an insight --- >> >> >> On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: >> Dear Danko, >> >> I have just read your new article and would like to comment briefly about it. >> >> In your introductory remarks, you write: >> >> "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." >> >> I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". >> >> My Magnum Opus, that was published in 2021, makes that belief clear in its title: >> >> Conscious Mind, Resonant Brain: How Each Brain Makes a Mind >> >> https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 >> >> The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. >> >> In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. >> >> I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. >> >> Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? >> >> What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? >> >> I would be delighted to discuss these issues further with you. >> >> If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. >> >> Best, >> >> Steve >> >> 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 Danko Nikolic >> Sent: Thursday, July 14, 2022 6:05 AM >> To: Gary Marcus >> Cc: connectionists at mailman.srv.cs.cmu.edu ; AIhub >> Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton >> >> Dear Gary and everyone, >> >> I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. >> >> As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. >> >> I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). >> >> In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. >> >> I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. >> >> The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug >> (I apologize for the neuroscience lingo, which I tried to minimize.) >> >> It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? >> >> Thanks. >> >> Danko >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> https://www.linkedin.com/in/danko-nikolic/ >> >> >> On Thu, Feb 3, 2022 at 5:25 PM 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 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 >> > -- > Stephen Jose Hanson > Professor, Director > Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -------------- next part -------------- An HTML attachment was scrubbed... URL: From steve at bu.edu Sun Jul 17 11:52:28 2022 From: steve at bu.edu (Grossberg, Stephen) Date: Sun, 17 Jul 2022 15:52:28 +0000 Subject: Connectionists: Neural architectures that embody biological intelligence In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Tsvi, If this chat continues, please do write to me one-on-one. It is true that some conferences are run by self-interested cliques who do not run an open meeting, despite protestations to the contrary. Other, more open, conferences do not always have enough reviewers to expertly review articles on all the topics that are covered by our interdisciplinary field. My advice if this happens is simply to find another conference where your work might be better appreciated. That such conferences exist is another advantage of working in such an interdisciplinary field. All of us have had disappointing experiences. My first experience was to send a series of articles to a single journal about discoveries that I had made over a period of many years. I was just starting out and naive enough to send them all to the same journal. As a result, they were all rejected without even being reviewed. The editor-in-chief, who years later became a friend, told me that they simply didn't know how to handle so many articles at once. Sometime later, I submitted two articles close in time to two different journals. I tried to anticipate which one might more likely reject the article that I sent to them. I was very proud of one article. It was, at least to my mind, quite a deep result. The other article was reasonably good craft, but more a continuation of earlier work than a breakthrough. Perhaps not surprisingly, the deep article got rejected while the less important article was accepted. I submitted the rejected article almost immediately to another journal and it was eventually published in a good place. The moral of the story is that, if you believe in your work, and the criticisms of it are not valid, do not give up. If, however, the criticisms are valid, one must not let your love of a result blind you to its weaknesses. I came to believe that all criticisms by reviewers are valuable and should be taken into account in your revision. Even if a reviewer's criticisms are, to your mind, wrong-headed, they represent the viewpoint of a more-than-usually-qualified reader who has given you the privilege of taking enough time to read your article. If you want to reach the maximum number of readers, then you should revise your article accordingly. Good luck! Steve ________________________________ From: Tsvi Achler Sent: Sunday, July 17, 2022 10:55 AM To: Grossberg, Stephen Cc: Asim Roy ; Danko Nikolic ; AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Neural architectures that embody biological intelligence Dear Steve, What motivated me to write was your response a couple of messages ago to someone who is not established in their field describing their model. Studies on academics show that researchers who are not established but do original work do not get published and cited as much. Please see article: www.nber.org/papers/w22180 Moreover established researchers tend to push their theories and increments of their theories so strongly that it significantly affects progress in the field. Please see article: www.nber.org/papers/w21788 Since you mention it, the personal instance I am referring to is a conference where I got the following review (and I am paraphrasing): I dont really understand this model but it must be ART, and if it is ART this is wrong and that is wrong so I recommend rejecting it. And in a box for reviewer certainty the review was listed as 100% certain. The consequence was that I had only 3 minutes to talk about a model that is counterintuitive given today's notions, as someone who exhausted all their meager resources just to get there. This summarizes my experiences in academia trying to put forward something new. I am happy to pull up the specific text but that distracts from the point. The point is that at least this review was transparent. Most reviewers are not likely to be as transparent when something is counterintuitive, not normative and thus harder to understand. What I am saying is that given this knowledge about academia, established researchers should be very careful as they can easily stifle new research without realizing it. If established academics push too strongly then academica can become a political club, not a place for progress. I believe this is a major contributor to why so little progress has been made in the field of understanding the brain through connectionist models. Sincerely, -Tsvi On Sat, Jul 16, 2022 at 8:45 AM Grossberg, Stephen > wrote: Dear Tsvi, I have no idea why you are writing to me. I would prefer that you did not engage the entire connectionists mailing list. However, since you did, I need to include everyone in my reply. For starters, I have not been an editor of any journal since 2010. When I was Editor-in-Chief of Neural Networks before that, and a new article was submitted, I assigned it to one of over 70 action editors who was a specialist in the topic of the article. That action editor then took full responsibility for getting three reviews of the article. If any reviewer disagreed with other reviewers for a potentially serious reason, then yet another reviewer was typically sought by the action editor in order to try to resolve the difference. Almost always, the reviewers agreed about publication recommendations, so this was not needed. I always followed the recommendations of action editors to publish or not, based upon the above process. I only entered any decision if the action editor solicited my help for a problem for which he/she needed advice. This hardly ever happened. Best, Steve ________________________________ From: Tsvi Achler > Sent: Saturday, July 16, 2022 10:51 AM To: Grossberg, Stephen > Cc: Asim Roy >; Danko Nikolic >; AIhub >; connectionists at mailman.srv.cs.cmu.edu > Subject: Re: Connectionists: Neural architectures that embody biological intelligence Dear Stephen, I have a connectionist model where feedback takes on a much greater role than the Resonance and other theories. I also have a richer background than many researchers in this field. I have a degree in electrical engineering & computer science, I did my PhD work doing neurophysiology recording neurons and working in a cognitive lab recording differential human reaction times to visual stimulation. I also got an MD focusing on neurology and patients. Consistently throughout the years established academics and their associates have blocked this theory's publication and funding in favor of their own. Since academia is mostly political, this is a big deal. Moreover, it bothers me seeing this done to others. Unfortunately you are by far NOT the worst at doing so, you are just the most transparent about it. I came to the conclusion that academia is not a place to innovate, especially if you come from a multidisciplinary background because (analogous to some of the models) the politics multiply exponentially. Although your work was innovative in the grand scheme of things, what you and other well established academics are doing is not ok. Sincerely, -Tsvi On Sat, Jul 16, 2022 at 12:04 AM Grossberg, Stephen > wrote: Dear Asim and Danko, A lot of your concerns about scaling do not apply to the kinds of biological neural networks that my colleagues and I have developed over the years. You can find a self-contained summary of many of them in my Magnum Opus: https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 As Asim notes below, it is indeed the case that ART can often make good predictions based on small amounts of learned data. This applies as well to large-scale applications naturalistic data. Gail Carpenter and her colleagues have, for example, shown how this works in learning complicated maps of multiple vegetation classes during remote sensing; e.g. Carpenter, G.A., Gopal, S., Macomber, S., Martens, S., & Woodcock, C.E. (1999). A neural network method for mixture estimation for vegetation mapping. Remote Sensing of Environment, 70(2), 138-152. http://techlab.bu.edu/members/gail/articles/127_Mixtures_RSE_1999_.pdf or in learning medical database predictions in response to incomplete, probabilistic, and even incorrect data. In this regard, Gail et al. have also shown how an ART system can incrementally learn a cognitive hierarchy of rules whereby to understand such data; i.e., converts information into knowledge; e.g., Carpenter, G.A., & Ravindran, A. (2008). Unifying multiple knowledge domains using the ARTMAP information fusion system. Proceedings of the 11th International Conference on Information Fusion, Cologne, Germany, June 30 - July 3, 2008. http://techlab.bu.edu/members/gail/articles/155_Fusion2008_CarpenterRavindran.pdf Carpenter, G.A. & Markuzon, N. (1998). ARTMAP?IC and medical diagnosis: Instance counting and inconsistent cases. Neural Networks, 11(2), 323-336. http://techlab.bu.edu/members/gail/articles/117_ARTMAP-IC_1998_.pdf My own work is filled with models that incrementally learn to carry out goal-oriented tasks without regard to scaling concerns. This work develops neural architectures that involve the coordinated actions of many brain regions, not just learned classification. These architectures are supported by unified and principled explanations of lots of psychological and neurobiological data; e.g., Chang, H.-C., Grossberg, S., and Cao, Y. (2014) Where?s Waldo? How perceptual cognitive, and emotional brain processes cooperate during learning to categorize and find desired objects in a cluttered scene. Frontiers in Integrative Neuroscience, doi: 10.3389/fnint.2014.0043, https://www.frontiersin.org/articles/10.3389/fnint.2014.00043/full Grossberg, S., and Vladusich, T. (2010). How do children learn to follow gaze, share joint attention, imitate their teachers, and use tools during social interactions? Neural Networks, 23, 940-965. https://sites.bu.edu/steveg/files/2016/06/GrossbergVladusichNN2010.pdf See Figure 1 in the following article to get a sense of how many brain processes other than classification are needed to realize true biological intelligence: Grossberg, S. (2018). Desirability, availability, credit assignment, category learning, and attention: Cognitive-emotional and working memory dynamics of orbitofrontal, ventrolateral, and dorsolateral prefrontal cortices. Brain and Neuroscience Advances, May 8, 2018. https://journals.sagepub.com/doi/full/10.1177/2398212818772179 Best, Steve ________________________________ From: Asim Roy > Sent: Friday, July 15, 2022 9:35 AM To: Danko Nikolic > Cc: Grossberg, Stephen >; Gary Marcus >; AIhub >; connectionists at mailman.srv.cs.cmu.edu >; 'maxversace at gmail.com' > Subject: RE: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, 1. I am not sure if I myself know all the uses of a knife, leave aside countless ones. Given a particular situation, I might simulate in my mind about the potential usage, but I doubt our minds explore all the countless situations of usage of an object as soon as it learns about it. 2. I am not sure if a 2 or 3 year old child, after having ?learnt? about a knife, knows very many uses of it. I doubt the kid is awake all night and day simulating in the brain how and where to use such a knife. 3. ?Understanding? is a loaded term. I think it needs a definition. 4. I am copying Max Versace, a student of Steve Grossberg. His company markets a software that can learn quickly from a few examples. Not exactly one-shot learning, it needs a few shots. I believe it?s a variation of ART. But Max can clarify the details. And Tsvi is doing similar work. So, what you are asking for may already exist. So linear scaling may be the worst case scenario. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Danko Nikolic > Sent: Friday, July 15, 2022 12:19 AM To: Asim Roy > Cc: Grossberg, Stephen >; Gary Marcus >; AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Asim, I agree about the potential for linear scaling of ART and.other connectionist systems. However, there are two problems. The problem number one kills it already and this is that the real brain scales a lot better than linearly: For each new object learned, we are able to resolve countless many new situations in which this object takes part (e.g., finding various uses for a knife, many of which may be new, ad hoc -- this is a great ability of biological minds often referred to as 'understanding'). Hence, simple linear scaling by adding more neurons for additional objects is not good enough to match biological intelligence. The second problem becomes an overkill, and this is that linear scaling in connectionist systems works only in theory, under idealized conditions. In real life, say if working with ImageNet, the scaling turns into a power-law with an exponent much larger than one: We need something like 500x more resources just to double the number of objects. Hence, in practice, the demands for resources explode if you want to add more categories whilst not losing the accuracy. To summarize, there is no linear scaling in practice nor would linear scaling suffice, even if we found one. This should be a strong enough argument to search for another paradigm, something that scales better than connectionism. I discuss both problems in the new manuscript, and even track a bit deeper the problem of why connectionism lacks linear scaling in practice (I provide some revealing computations in the Supplementary Materials (with access to the code), although much more work needs to be done). Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Thu, Jul 14, 2022 at 11:48 PM Asim Roy > wrote: I think Steve Grossberg is generally correct. ART, RBF nets and similar flexible architectures can scale up almost linearly to problem size. My conjecture is that architectures that use distributed representation run into the scaling issue. On the other hand, however, distributed representation produces a more compact architecture compared to the shallow architectures of ART and RBFs. But, in terms of adding concepts, it?s easy to add a new object or concept to the shallow architectures. Perhaps someone can provide more insights on the architectural differences and the corresponding pros and cons of each. Best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists > On Behalf Of Grossberg, Stephen Sent: Thursday, July 14, 2022 10:14 AM To: Danko Nikolic > Cc: AIhub >; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Danko, I will respond to your comment below to the entire list. I recommend that future interactions be done between just the two of us. Everything that I write below is summarized in my new book: In brief, within Adaptive Resonance Theory, or ART, there IS no problem of scaling, in the sense that you describe it below, from learning (say) to correctly recognizing 100 objects to doing the same for 200. In fact, one of the properties for which I introduced ART in 1976 was to enable incremental learning in real time of arbitrary numbers of objects or events without experiencing catastrophic forgetting. I have called this a solution of the stability-plasticity dilemma: How our brains, and models like ART, can rapidly learn (plasticity) arbitrary numbers of objects without experiencing catastrophic forgetting (stability). I also derived ART in 1980, in an article within Psychological Review, from a thought experiment about how ANY system can AUTONOMOUSLY correct predictive errors in a changing world that is filled with unexpected events. This thought is experiment was derived from a few facts that are familiar to us all. They are familiar because they represent ubiquitous environmental pressures that we all experience. The thought experiment clarifies why, when they act together during the evolutionary process, models like ART are the unique result. Moveover, at no point during the thought experiment are the words mind or brain mentioned. ART is thus a universal solution of this learning, classification, and prediction problem. You write below about "connectionist systems". ART is a connectionist system. What do you mean by a "connectionist system"? What you write below about them is not true in general. Best, Steve ________________________________ From: Danko Nikolic > Sent: Thursday, July 14, 2022 12:16 PM To: Grossberg, Stephen > Cc: Gary Marcus >; connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 jose at rubic.rutgers.edu Sun Jul 17 12:00:45 2022 From: jose at rubic.rutgers.edu (=?utf-8?B?U3RlcGhlbiBKb3PDqSBIYW5zb24=?=) Date: Sun, 17 Jul 2022 16:00:45 +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: Not really. the distinction is in the details.. neural modeling involves biological constraints, learning.. OTOH, hybrid systems, as Gary and some others are promoting is more than a slippery slope.. its already off the cliff. Someone said once to me ..if you have a perfectly good electric car.. why would you have it push a car with a combustion engine around?" why not improve the EV without giving up on the energy source? S On 7/17/22 11:10 AM, Asim Roy wrote: ?without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet.? Those who are aware, Gary Marcus? symbolic parts are already tumbling out of the closet. There are numerous efforts, even by Hinton and Bengio, that try to encode a symbolic concept using a set of neurons. All the best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Stephen Jos? Hanson Sent: Friday, July 15, 2022 4:28 AM To: Dietterich, Thomas ; Danko Nikolic ; Grossberg, Stephen Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton So there are a number of issues here. In the first one, the brain obviously has a specific structure. DL supervised modeling of classification, NOT psychological categorization-- may scale linearly with resources.. however, its clear just adding more hidden layers has diminishing returns, adding various kinds of classification structure (GANs etc) or doing unsupervised learning or reinforcement learning is not going to end up with AGI or really any sort of final AI system and certainly not biological. Humans btw, scale as sub-linear systems.. but with hard limits. We can only recognize about 6000 faces (and the variance is about 4k across individuals), most undergraduates can recognize 100k words again with a large variance., so despite the fact that the brain has trillions of synapses, "loading" in explicit (MTL) memory is complex and somehow localized. Memory systems involve implicit (most DL) or explicit (almost noone), working memory is tiny--like 3 or 4+- 2 (sorry George). Many (Bengio, Le Cun etc) are experimenting with more complex structures, but keeping the system homogeneous, no extra symbolic baggage and forcing the exercise to have to learn the configuration rather than engineer it. In some ways, Grossberg has been out in front in all this, but with his own special and situated neural systems. He has been consistently and persistently creating focused neural modeling, which as an art form has always been a matter of taste. Connectionist models tend toward minimalist forms and Steve and others follow more Baroque or perhaps Victorian forms. What is clear is that the future is going to be blends of these forms and more complex, without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet. Steve On 7/15/22 4:01 AM, Dietterich, Thomas wrote: Dear Danko, In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. --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: Connectionists On Behalf Of Danko Nikolic Sent: Thursday, July 14, 2022 09:17 To: Grossberg, Stephen 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.] Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -------------- next part -------------- An HTML attachment was scrubbed... URL: From ASIM.ROY at asu.edu Sun Jul 17 11:10:20 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sun, 17 Jul 2022 15:10:20 +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: ?without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet.? Those who are aware, Gary Marcus? symbolic parts are already tumbling out of the closet. There are numerous efforts, even by Hinton and Bengio, that try to encode a symbolic concept using a set of neurons. All the best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Stephen Jos? Hanson Sent: Friday, July 15, 2022 4:28 AM To: Dietterich, Thomas ; Danko Nikolic ; Grossberg, Stephen Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton So there are a number of issues here. In the first one, the brain obviously has a specific structure. DL supervised modeling of classification, NOT psychological categorization-- may scale linearly with resources.. however, its clear just adding more hidden layers has diminishing returns, adding various kinds of classification structure (GANs etc) or doing unsupervised learning or reinforcement learning is not going to end up with AGI or really any sort of final AI system and certainly not biological. Humans btw, scale as sub-linear systems.. but with hard limits. We can only recognize about 6000 faces (and the variance is about 4k across individuals), most undergraduates can recognize 100k words again with a large variance., so despite the fact that the brain has trillions of synapses, "loading" in explicit (MTL) memory is complex and somehow localized. Memory systems involve implicit (most DL) or explicit (almost noone), working memory is tiny--like 3 or 4+- 2 (sorry George). Many (Bengio, Le Cun etc) are experimenting with more complex structures, but keeping the system homogeneous, no extra symbolic baggage and forcing the exercise to have to learn the configuration rather than engineer it. In some ways, Grossberg has been out in front in all this, but with his own special and situated neural systems. He has been consistently and persistently creating focused neural modeling, which as an art form has always been a matter of taste. Connectionist models tend toward minimalist forms and Steve and others follow more Baroque or perhaps Victorian forms. What is clear is that the future is going to be blends of these forms and more complex, without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet. Steve On 7/15/22 4:01 AM, Dietterich, Thomas wrote: Dear Danko, In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. --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: Connectionists On Behalf Of Danko Nikolic Sent: Thursday, July 14, 2022 09:17 To: Grossberg, Stephen 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.] Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -------------- next part -------------- An HTML attachment was scrubbed... URL: From achler at gmail.com Sun Jul 17 12:58:12 2022 From: achler at gmail.com (Tsvi Achler) Date: Sun, 17 Jul 2022 09:58:12 -0700 Subject: Connectionists: Neural architectures that embody biological intelligence In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: Dear Steven, As long as it seems you are: 1) belittling my experience yet pointing me to the same things 2) show no interest in the articles about academia 3) address me as someone who has not gotten a deeper multidisciplinary background than most 4) address me as someone who has not worked on this for decades, I respectfully prefer to keep things public. Moreover I have not even pointed out how because of this system and its hegemony of control, most peer reviewed articles are not repeatable (see repeatability crisis) and this includes computer science. I came to the conclusion that it is not worth the effort to keep on plugging through academia. There are better ways. Besides, my point here is to make others aware about the systemic issues, thus it is not directly about you. Sincerely, -Tsvi On Sun, Jul 17, 2022 at 8:52 AM Grossberg, Stephen wrote: > Dear Tsvi, > > If this chat continues, please do write to me one-on-one. > > It is true that some conferences are run by self-interested cliques who do > not run an open meeting, despite protestations to the contrary. > > Other, more open, conferences do not always have enough reviewers to > expertly review articles on all the topics that are covered by our > interdisciplinary field. > > My advice if this happens is simply to find another conference where your > work might be better appreciated. > > That such conferences exist is another advantage of working in such an > interdisciplinary field. > > All of us have had disappointing experiences. > > My first experience was to send a series of articles to a single journal > about discoveries that I had made over a period of many years. I was just > starting out and naive enough to send them all to the same journal. > > As a result, they were all rejected without even being reviewed. The > editor-in-chief, who years later became a friend, told me that they simply > didn't know how to handle so many articles at once. > > Sometime later, I submitted two articles close in time to two different > journals. I tried to anticipate which one might more likely reject the > article that I sent to them. > > I was very proud of one article. It was, at least to my mind, quite a deep > result. The other article was reasonably good craft, but more a > continuation of earlier work than a breakthrough. > > Perhaps not surprisingly, the deep article got rejected while the less > important article was accepted. > > I submitted the rejected article almost immediately to another journal and > it was eventually published in a good place. > > The moral of the story is that, if you believe in your work, and the > criticisms of it are not valid, do not give up. > > If, however, the criticisms are valid, one must not let your love of a > result blind you to its weaknesses. > > I came to believe that all criticisms by reviewers are valuable and should > be taken into account in your revision. > > Even if a reviewer's criticisms are, to your mind, wrong-headed, they > represent the viewpoint of a more-than-usually-qualified reader who has > given you the privilege of taking enough time to read your article. > > If you want to reach the maximum number of readers, then you should revise > your article accordingly. > > Good luck! > > Steve > > ------------------------------ > *From:* Tsvi Achler > *Sent:* Sunday, July 17, 2022 10:55 AM > *To:* Grossberg, Stephen > *Cc:* Asim Roy ; Danko Nikolic ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu> > *Subject:* Re: Connectionists: Neural architectures that embody > biological intelligence > > > Dear Steve, > > What motivated me to write was your response a couple of messages ago to > someone who is not established in their field describing their model. > > Studies on academics show that researchers who are not established but do > original work do not get published and cited as much. Please see article: > www.nber.org/papers/w22180 > > Moreover established researchers tend to push their theories and > increments of their theories so strongly that it significantly affects > progress in the field. Please see article: www.nber.org/papers/w21788 > > Since you mention it, the personal instance I am referring to is a > conference where I got the following review (and I am paraphrasing): > I* dont really understand this model but it must be ART, and if it is ART > this is wrong and that is wrong so I recommend rejecting it.* And in a > box for reviewer certainty the review was listed as *100% certain*. > > The consequence was that I had only 3 minutes to talk about a model that > is counterintuitive given today's notions, as someone who exhausted all > their meager resources just to get there. This summarizes my experiences > in academia trying to put forward something new. > > I am happy to pull up the specific text but that distracts from the point. > The point is that at least this review was transparent. > Most reviewers are not likely to be as transparent when something is > counterintuitive, not normative and thus harder to understand. > > What I am saying is that given this knowledge about academia, established > researchers should be very careful as they can easily stifle new research > without realizing it. > > If established academics push too strongly then academica can become a > political club, not a place for progress. > I believe this is a major contributor to why so little progress has been > made in the field of understanding the brain through connectionist models. > > Sincerely, > -Tsvi > > > On Sat, Jul 16, 2022 at 8:45 AM Grossberg, Stephen wrote: > > Dear Tsvi, > > I have no idea why you are writing to me. > > I would prefer that you did not engage the entire connectionists mailing > list. However, since you did, I need to include everyone in my reply. > > For starters, I have not been an editor of any journal since 2010. > > When I was Editor-in-Chief of *Neural Networks* before that, and a new > article was submitted, I assigned it to one of over 70 action editors who > was a specialist in the topic of the article. > > That action editor then took full responsibility for getting three reviews > of the article. If any reviewer disagreed with other reviewers for a > potentially serious reason, then yet another reviewer was typically sought > by the action editor in order to try to resolve the difference. > > Almost always, the reviewers agreed about publication recommendations, so > this was not needed. > > I always followed the recommendations of action editors to publish or not, > based upon the above process. > > I only entered any decision if the action editor solicited my help for a > problem for which he/she needed advice. This hardly ever happened. > > Best, > > Steve > > > ------------------------------ > *From:* Tsvi Achler > *Sent:* Saturday, July 16, 2022 10:51 AM > *To:* Grossberg, Stephen > *Cc:* Asim Roy ; Danko Nikolic ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu> > *Subject:* Re: Connectionists: Neural architectures that embody > biological intelligence > > Dear Stephen, > I have a connectionist model where feedback takes on a much greater role > than the Resonance and other theories. > I also have a richer background than many researchers in this field. I > have a degree in electrical engineering & computer science, I did my PhD > work doing neurophysiology recording neurons and working in a cognitive > lab recording differential human reaction times to visual stimulation. I > also got an MD focusing on neurology and patients. > > Consistently throughout the years established academics and their > associates have blocked this theory's publication and funding in favor of > their own. > Since academia is mostly political, this is a big deal. Moreover, it > bothers me seeing this done to others. > > Unfortunately you are by far NOT the worst at doing so, you are just the > most transparent about it. > > I came to the conclusion that academia is not a place to innovate, > especially if you come from a multidisciplinary background because > (analogous to some of the models) the politics multiply exponentially. > > Although your work was innovative in the grand scheme of things, what you > and other well established academics are doing is not ok. > Sincerely, > -Tsvi > > > On Sat, Jul 16, 2022 at 12:04 AM Grossberg, Stephen wrote: > > Dear Asim and Danko, > > A lot of your concerns about scaling do not apply to the kinds of > biological neural networks that my colleagues and I have developed over the > years. You can find a self-contained summary of many of them in my Magnum > Opus: > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > As Asim notes below, it is indeed the case that ART can often make good > predictions based on small amounts of learned data. This applies as well to > large-scale applications naturalistic data. > > Gail Carpenter and her colleagues have, for example, shown how this works > in learning complicated maps of multiple vegetation classes during remote > sensing; e.g. > > Carpenter, G.A., Gopal, S., Macomber, S., Martens, S., & Woodcock, C.E. > (1999). A neural network method for mixture estimation for vegetation > mapping. Remote Sensing of Environment, 70(2), 138-152. > http://techlab.bu.edu/members/gail/articles/127_Mixtures_RSE_1999_.pdf > > or in learning medical database predictions in response to incomplete, > probabilistic, and even incorrect data. > > In this regard, Gail et al. have also shown how an ART system can > incrementally learn a cognitive hierarchy of rules whereby to understand > such data; i.e., converts information into knowledge; e.g., > > Carpenter, G.A., & Ravindran, A. (2008). Unifying multiple knowledge > domains using the ARTMAP information fusion system. Proceedings of the 11th > International Conference on Information Fusion, Cologne, Germany, June 30 - > July 3, 2008. > > http://techlab.bu.edu/members/gail/articles/155_Fusion2008_CarpenterRavindran.pdf > > Carpenter, G.A. & Markuzon, N. (1998). ARTMAP?IC and medical diagnosis: > Instance counting and inconsistent cases. Neural Networks, 11(2), 323-336. > http://techlab.bu.edu/members/gail/articles/117_ARTMAP-IC_1998_.pdf > > My own work is filled with models that incrementally learn to carry out > goal-oriented tasks without regard to scaling concerns. This work develops > neural architectures that involve the coordinated actions of many brain > regions, not just learned classification. These architectures are supported > by unified and principled explanations of lots of psychological and > neurobiological data; e.g., > > Chang, H.-C., Grossberg, S., and Cao, Y. (2014) Where?s Waldo? How > perceptual cognitive, and emotional brain processes cooperate during > learning to categorize and find desired objects in a cluttered scene. > Frontiers in Integrative Neuroscience, doi: 10.3389/fnint.2014.0043, > https://www.frontiersin.org/articles/10.3389/fnint.2014.00043/full > > Grossberg, S., and Vladusich, T. (2010). How do children learn to follow > gaze, share joint attention, imitate their teachers, and use tools during > social interactions? Neural Networks, 23, 940-965. > > https://sites.bu.edu/steveg/files/2016/06/GrossbergVladusichNN2010.pdf > > See Figure 1 in the following article to get a sense of how many brain > processes other than classification are needed to realize true biological > intelligence: > > Grossberg, S. (2018). Desirability, availability, credit assignment, > category learning, and attention: Cognitive-emotional and working memory > dynamics of orbitofrontal, ventrolateral, and dorsolateral prefrontal > cortices. Brain and Neuroscience Advances, May 8, 2018. > https://journals.sagepub.com/doi/full/10.1177/2398212818772179 > > Best, > > Steve > ------------------------------ > *From:* Asim Roy > *Sent:* Friday, July 15, 2022 9:35 AM > *To:* Danko Nikolic > *Cc:* Grossberg, Stephen ; Gary Marcus ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; 'maxversace at gmail.com' < > maxversace at gmail.com> > *Subject:* RE: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > Dear Danko, > > > > 1. I am not sure if I myself know all the uses of a knife, leave aside > countless ones. Given a particular situation, I might simulate in my mind > about the potential usage, but I doubt our minds explore all the countless > situations of usage of an object as soon as it learns about it. > 2. I am not sure if a 2 or 3 year old child, after having ?learnt? > about a knife, knows very many uses of it. I doubt the kid is awake all > night and day simulating in the brain how and where to use such a knife. > 3. ?Understanding? is a loaded term. I think it needs a definition. > 4. I am copying Max Versace, a student of Steve Grossberg. His company > markets a software that can learn quickly from a few examples. Not exactly > one-shot learning, it needs a few shots. I believe it?s a variation of ART. > But Max can clarify the details. And Tsvi is doing similar work. So, what > you are asking for may already exist. So linear scaling may be the worst > case scenario. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > > > *From:* Danko Nikolic > *Sent:* Friday, July 15, 2022 12:19 AM > *To:* Asim Roy > *Cc:* Grossberg, Stephen ; Gary Marcus ; > AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Asim, > > > > I agree about the potential for linear scaling of ART and.other > connectionist systems. However, there are two problems. > > > > The problem number one kills it already and this is that the real brain > scales a lot better than linearly: For each new object learned, we are able > to resolve countless many new situations in which this object takes part > (e.g., finding various uses for a knife, many of which may be new, ad hoc > -- this is a great ability of biological minds often referred to as > 'understanding'). Hence, simple linear scaling by adding more neurons for > additional objects is not good enough to match biological intelligence. > > > > The second problem becomes an overkill, and this is that linear scaling in > connectionist systems works only in theory, under idealized conditions. In > real life, say if working with ImageNet, the scaling turns into a power-law > with an exponent much larger than one: We need something like 500x more > resources just to double the number of objects. Hence, in practice, the > demands for resources explode if you want to add more categories whilst not > losing the accuracy. > > > > To summarize, there is no linear scaling in practice nor would linear > scaling suffice, even if we found one. > > > > This should be a strong enough argument to search for another paradigm, > something that scales better than connectionism. > > > > I discuss both problems in the new manuscript, and even track a bit deeper > the problem of why connectionism lacks linear scaling in practice (I > provide some revealing computations in the Supplementary Materials (with > access to the code), although much more work needs to be done). > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > -- I wonder, how is the brain able to generate insight? -- > > > > > > On Thu, Jul 14, 2022 at 11:48 PM Asim Roy wrote: > > I think Steve Grossberg is generally correct. ART, RBF nets and similar > flexible architectures can scale up almost linearly to problem size. My > conjecture is that architectures that use distributed representation run > into the scaling issue. On the other hand, however, distributed > representation produces a more compact architecture compared to the shallow > architectures of ART and RBFs. But, in terms of adding concepts, it?s easy > to add a new object or concept to the shallow architectures. > > > > Perhaps someone can provide more insights on the architectural differences > and the corresponding pros and cons of each. > > > > Best, > > Asim Roy > > Professor, Information Systems > > Arizona State University > > Lifeboat Foundation Bios: Professor Asim Roy > > > Asim Roy | iSearch (asu.edu) > > > > > > > *From:* Connectionists *On > Behalf Of *Grossberg, Stephen > *Sent:* Thursday, July 14, 2022 10:14 AM > *To:* Danko Nikolic > *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Danko, > > > > I will respond to your comment below to the entire list. I recommend that > future interactions be done between just the two of us. > > > > Everything that I write below is summarized in my new book: > > > > In brief, within Adaptive Resonance Theory, or ART, there IS no problem of > scaling, in the sense that you describe it below, from learning (say) to > correctly recognizing 100 objects to doing the same for 200. In fact, one > of the properties for which I introduced ART in 1976 was to enable > incremental learning in real time of arbitrary numbers of objects or events > without experiencing catastrophic forgetting. > > > > I have called this a solution of the *stability-plasticity dilemma*: How > our brains, and models like ART, can rapidly learn (plasticity) arbitrary > numbers of objects without experiencing catastrophic forgetting (stability). > > > > I also derived ART in 1980, in an article within Psychological Review, > from a *thought experiment* about how ANY system can AUTONOMOUSLY correct > predictive errors in a changing world that is filled with unexpected events. > > > > This thought is experiment was derived from a few facts that are familiar > to us all. They are familiar because they represent ubiquitous > environmental pressures that we all experience. The thought experiment > clarifies why, when they act together during the evolutionary process, > models like ART are the unique result. > > > > Moveover, at no point during the thought experiment are the words mind or > brain mentioned. ART is thus a *universal solution* of this learning, > classification, and prediction problem. > > > > You write below about "connectionist systems". ART is a connectionist > system. > > > > What do you mean by a "connectionist system"? What you write below about > them is not true in general. > > > > Best, > > > > Steve > ------------------------------ > > *From:* Danko Nikolic > *Sent:* Thursday, July 14, 2022 12:16 PM > *To:* Grossberg, Stephen > *Cc:* Gary Marcus ; > connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Steve, > > > > Thank you very much for your message and for the greetings. I will pass > them on if an occasion arises. > > > > Regarding your question: The key problem I am trying to address and that, > to the best of my knowledge, no connectionist system was able to solve so > far is that of scaling the system's intelligence. For example, if the > system is able to correctly recognize 100 different objects, how many > additional resources are needed to double that to 200? All the empirical > data show that connectionist systems scale poorly: Some of the best systems > we have require 500x more resources in order to increase the intelligence > by only 2x. I document this problem in the manuscript and even run some > simulations to show that the worst performance is if connectionist systems > need to solve a generalized XOR problem. > > > > In contrast, the biological brain scales well. This I also quantify in the > paper. > > > > I will look at the publication that you mentioned. However, so far, I > haven't seen a solution that scales well in intelligence. > > > > My argument is that transient selection of subnetworks by the help of the > mentioned proteins is how intelligence scaling is achieved in biological > brains. > > > > In short, intelligence scaling is the key problem that concerns me. I > describe the intelligence scaling problem in more detail in this book that > just came out a few weeks ago and that is written for practitioners in Data > Scientist and AI: https://amzn.to/3IBxUpL > > > > > I hope that this at least partly answers where I see the problems and what > I am trying to solve. > > > > Greetings from Germany, > > > > Danko > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > --- A progress usually starts with an insight --- > > > > > > On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: > > Dear Danko, > > > > I have just read your new article and would like to comment briefly about > it. > > > > In your introductory remarks, you write: > > > > "However, connectionism did not yet produce a satisfactory explanation of > how the mental emerges from the physical. A number of open problems > remains ( 5,6,7,8). As a result, the explanatory gap between the mind and > the brain remains wide open." > > > > I certainly believe that no theoretical explanation in science is ever > complete. However, I also believe that "the explanatory gap between the > mind and the brain" does not remain "wide open". > > > > My Magnum Opus, that was published in 2021, makes that belief clear in its > title: > > > > *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* > > > > https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 > > > > > The book provides a self-contained and non-technical exposition in a > conversational tone of many principled and unifying explanations of > psychological and neurobiological data. > > > > In particular, it explains roles for the metabotropic glutamate receptors > that you mention in your own work. See the text and figures around p. 521. > This explanation unifies psychological, anatomical, neurophysiological, > biophysical, and biochemical data about the processes under discussion. > > > > I have a very old-fashioned view about how to understand scientific > theories. I get excited by theories that explain and predict more data than > previous theories. > > > > Which of the data that I explain in my book, and support with quantitative > computer simulations, can you also explain? > > > > What data can you explain, in the same quantitative sense, that you do not > think the neural models in my book can explain? > > > > I would be delighted to discuss these issues further with you. > > > > If you are in touch with my old friend and esteemed colleague, Wolf > Singer, please send him my warm regards. I cite the superb work that he and > various of his collaborators have done in many places in my book. > > > > Best, > > > > Steve > > > > 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 Danko Nikolic > *Sent:* Thursday, July 14, 2022 6:05 AM > *To:* Gary Marcus > *Cc:* connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu>; AIhub > *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff > Hinton > > > > Dear Gary and everyone, > > > > I am continuing the discussion from where we left off a few months ago. > Back then, some of us agreed that the problem of understanding remains > unsolved. > > > > As a reminder, the challenge for connectionism was to 1) learn with few > examples and 2) apply the knowledge to a broad set of situations. > > > > I am happy to announce that I have now finished a draft of a paper in > which I propose how the brain is able to achieve that. The manuscript > requires a bit of patience for two reasons: one is that the reader may be > exposed for the first time to certain aspects of brain physiology. The > second reason is that it may take some effort to understand the > counterintuitive implications of the new ideas (this requires a different > way of thinking than what we are used to based on connectionism). > > > > In short, I am suggesting that instead of the connectionist paradigm, we > adopt transient selection of subnetworks. The mechanisms that transiently > select brain subnetworks are distributed all over the nervous system and, I > argue, are our main machinery for thinking/cognition. The surprising > outcome is that neural activation, which was central in connectionism, now > plays only a supportive role, while the real 'workers' within the brain are > the mechanisms for transient selection of subnetworks. > > > > I also explain how I think transient selection achieves learning with only > a few examples and how the learned knowledge is possible to apply to a > broad set of situations. > > > > The manuscript is made available to everyone and can be downloaded here: > https://bit.ly/3IFs8Ug > > > (I apologize for the neuroscience lingo, which I tried to minimize.) > > > > It will likely take a wide effort to implement these concepts as an AI > technology, provided my ideas do not have a major flaw in the first place. > Does anyone see a flaw? > > > > Thanks. > > > > Danko > > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > > https://www.linkedin.com/in/danko-nikolic/ > > > > > > > On Thu, Feb 3, 2022 at 5:25 PM 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 ASIM.ROY at asu.edu Sun Jul 17 12:05:08 2022 From: ASIM.ROY at asu.edu (Asim Roy) Date: Sun, 17 Jul 2022 16:05:08 +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: <9B59973F-F4F1-4935-AE5B-B59039ADA7F8@asu.edu> It is not a slippery slope. It already has good traction and produces great results. It?s the way forward. Sent from my iPhone On Jul 17, 2022, at 6:00 PM, Stephen Jos? Hanson wrote: ? Not really. the distinction is in the details.. neural modeling involves biological constraints, learning.. OTOH, hybrid systems, as Gary and some others are promoting is more than a slippery slope.. its already off the cliff. Someone said once to me ..if you have a perfectly good electric car.. why would you have it push a car with a combustion engine around?" why not improve the EV without giving up on the energy source? S On 7/17/22 11:10 AM, Asim Roy wrote: ?without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet.? Those who are aware, Gary Marcus? symbolic parts are already tumbling out of the closet. There are numerous efforts, even by Hinton and Bengio, that try to encode a symbolic concept using a set of neurons. All the best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Stephen Jos? Hanson Sent: Friday, July 15, 2022 4:28 AM To: Dietterich, Thomas ; Danko Nikolic ; Grossberg, Stephen Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton So there are a number of issues here. In the first one, the brain obviously has a specific structure. DL supervised modeling of classification, NOT psychological categorization-- may scale linearly with resources.. however, its clear just adding more hidden layers has diminishing returns, adding various kinds of classification structure (GANs etc) or doing unsupervised learning or reinforcement learning is not going to end up with AGI or really any sort of final AI system and certainly not biological. Humans btw, scale as sub-linear systems.. but with hard limits. We can only recognize about 6000 faces (and the variance is about 4k across individuals), most undergraduates can recognize 100k words again with a large variance., so despite the fact that the brain has trillions of synapses, "loading" in explicit (MTL) memory is complex and somehow localized. Memory systems involve implicit (most DL) or explicit (almost noone), working memory is tiny--like 3 or 4+- 2 (sorry George). Many (Bengio, Le Cun etc) are experimenting with more complex structures, but keeping the system homogeneous, no extra symbolic baggage and forcing the exercise to have to learn the configuration rather than engineer it. In some ways, Grossberg has been out in front in all this, but with his own special and situated neural systems. He has been consistently and persistently creating focused neural modeling, which as an art form has always been a matter of taste. Connectionist models tend toward minimalist forms and Steve and others follow more Baroque or perhaps Victorian forms. What is clear is that the future is going to be blends of these forms and more complex, without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet. Steve On 7/15/22 4:01 AM, Dietterich, Thomas wrote: Dear Danko, In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. --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: Connectionists On Behalf Of Danko Nikolic Sent: Thursday, July 14, 2022 09:17 To: Grossberg, Stephen 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.] Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -------------- next part -------------- An HTML attachment was scrubbed... URL: From xavier.hinaut at inria.fr Mon Jul 18 03:40:12 2022 From: xavier.hinaut at inria.fr (Xavier Hinaut) Date: Mon, 18 Jul 2022 09:40:12 +0200 Subject: Connectionists: SMILES workshop call extended 28th July Message-ID: <2935E7AF-BA08-4177-B5DB-2F4EDB32E67D@inria.fr> The SMILES (Sensorimotor Interaction, Language and Embodiment of Symbols) Workshop will take place both on site and virtually at the ICDL 2022 (International Conference on Developmental Learning). * Call for abstracts : - Deadline extended: July 28th - Abstracts call: from 1/2 page to 2 pages (onsite and virtual participation are possible) - Abstract format: same as ICDL conference https://www.ieee.org/conferences/publishing/templates.html - Submissions: smiles.conf at gmail.com + indicate if you will be onsite or online - Workshop dates: September 12, 2022 - Venue onsite: Queen Mary University of London, UK. - Venue online: via Zoom and Discord group. https://discord.gg/B8xbemQS Accepted abstract will be asked to make a short video or poster for the workshop. * Workshop Short Description On the one hand, models of sensorimotor interaction are embodied in the environment and in the interaction with other agents. On the other hand, recent Deep Learning development of Natural Language Processing (NLP) models allow to capture increasing language complexity (e.g. compositional representations, word embedding, long term dependencies). However, those NLP models are disembodied in the sense that they are learned from static datasets of text or speech. How can we bridge the gap from low-level sensorimotor interaction to high-level compositional symbolic communication? The SMILES workshop will address this issue through an interdisciplinary approach involving researchers from (but not limited to): - Sensori-motor learning, - Symbol grounding and symbol emergence, - Emergent communication in multi-agent systems, - Chunking of perceptuo-motor gestures (gestures in a general sense: motor, vocal, ...), - Compositional representations for communication and action sequence, - Hierarchical representations of temporal information, - Language processing and language acquisition in brains and machines, - Models of animal communication, - Understanding composition and temporal processing in neural network models, and - Enaction, active perception, perception-action loop. * More info - contact: smiles.conf at gmail.com - organizers: Xavier Hinaut, Cl?ment Moulin-Frier, Silvia Pagliarini, Joni Zhong, Michael Spranger, Tadahiro Taniguchi, Anne Warlaumont. - invited speakers (coming soon) - workshop website (updated regularly): https://sites.google.com/view/smiles-workshop/ - ICDL conference website: https://icdl2022.qmul.ac.uk/ Xavier Hinaut on behalf of the SMILES workshop organisers: - Cl?ment Moulin-Frier, Inria and Ensta ParisTech, Bordeaux, France - Anne Warlaumont, UCLA, Los Angeles, USA - Silvia Pagliarini, UCLA, Los Angeles, USA - Michael Spranger, Sony AI and Sony CSL, Tokyo, Japan - Tadahiro Taniguchi, Ritsumeikan University, Kyoto, Japan - Junpei Zhong, South China University of Technology, Guangzhou, China Xavier Hinaut Inria Research Scientist www.xavierhinaut.com -- +33 5 33 51 48 01 Mnemosyne team, Inria, Bordeaux, France -- https://team.inria.fr/mnemosyne & LaBRI, Bordeaux University -- https://www4.labri.fr/en/formal-methods-and-models & IMN (Neurodegeneratives Diseases Institute) -- http://www.imn-bordeaux.org/en --- Our new release of Reservoir Computing library: https://github.com/reservoirpy/reservoirpy -------------- next part -------------- An HTML attachment was scrubbed... URL: From jose at rubic.rutgers.edu Sun Jul 17 12:17:14 2022 From: jose at rubic.rutgers.edu (=?utf-8?B?U3RlcGhlbiBKb3PDqSBIYW5zb24=?=) Date: Sun, 17 Jul 2022 16:17:14 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <9B59973F-F4F1-4935-AE5B-B59039ADA7F8@asu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <9B59973F-F4F1-4935-AE5B-B59039ADA7F8@asu.edu> Message-ID: Asim, with all due respect, I think you are missing the point. The enormous recent progress of DL is what's issue. The way forward, is already happening. Hybrid system integration is simply an admission that the explicit representations being developed for DL have failed. We already know how our invented symbolic systems can add 2 and 2.. not so clear how the brain does it. I think this is where the confusion lay. Steve (Btw you may not know, but I am an Alum of ASU). On 7/17/22 12:05 PM, Asim Roy wrote: It is not a slippery slope. It already has good traction and produces great results. It?s the way forward. Sent from my iPhone On Jul 17, 2022, at 6:00 PM, Stephen Jos? Hanson wrote: ? Not really. the distinction is in the details.. neural modeling involves biological constraints, learning.. OTOH, hybrid systems, as Gary and some others are promoting is more than a slippery slope.. its already off the cliff. Someone said once to me ..if you have a perfectly good electric car.. why would you have it push a car with a combustion engine around?" why not improve the EV without giving up on the energy source? S On 7/17/22 11:10 AM, Asim Roy wrote: ?without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet.? Those who are aware, Gary Marcus? symbolic parts are already tumbling out of the closet. There are numerous efforts, even by Hinton and Bengio, that try to encode a symbolic concept using a set of neurons. All the best, Asim Roy Professor, Information Systems Arizona State University Lifeboat Foundation Bios: Professor Asim Roy Asim Roy | iSearch (asu.edu) From: Connectionists On Behalf Of Stephen Jos? Hanson Sent: Friday, July 15, 2022 4:28 AM To: Dietterich, Thomas ; Danko Nikolic ; Grossberg, Stephen Cc: AIhub ; connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton So there are a number of issues here. In the first one, the brain obviously has a specific structure. DL supervised modeling of classification, NOT psychological categorization-- may scale linearly with resources.. however, its clear just adding more hidden layers has diminishing returns, adding various kinds of classification structure (GANs etc) or doing unsupervised learning or reinforcement learning is not going to end up with AGI or really any sort of final AI system and certainly not biological. Humans btw, scale as sub-linear systems.. but with hard limits. We can only recognize about 6000 faces (and the variance is about 4k across individuals), most undergraduates can recognize 100k words again with a large variance., so despite the fact that the brain has trillions of synapses, "loading" in explicit (MTL) memory is complex and somehow localized. Memory systems involve implicit (most DL) or explicit (almost noone), working memory is tiny--like 3 or 4+- 2 (sorry George). Many (Bengio, Le Cun etc) are experimenting with more complex structures, but keeping the system homogeneous, no extra symbolic baggage and forcing the exercise to have to learn the configuration rather than engineer it. In some ways, Grossberg has been out in front in all this, but with his own special and situated neural systems. He has been consistently and persistently creating focused neural modeling, which as an art form has always been a matter of taste. Connectionist models tend toward minimalist forms and Steve and others follow more Baroque or perhaps Victorian forms. What is clear is that the future is going to be blends of these forms and more complex, without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet. Steve On 7/15/22 4:01 AM, Dietterich, Thomas wrote: Dear Danko, In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. --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: Connectionists On Behalf Of Danko Nikolic Sent: Thursday, July 14, 2022 09:17 To: Grossberg, Stephen 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.] Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -------------- next part -------------- An HTML attachment was scrubbed... URL: From junfeng989 at gmail.com Sun Jul 17 21:19:22 2022 From: junfeng989 at gmail.com (Jun Feng) Date: Sun, 17 Jul 2022 21:19:22 -0400 Subject: Connectionists: The 2022 IEEE International Conference on Privacy Enhancing Technologies [10 Special Issues] Message-ID: CFP: The 2022 IEEE International Conference on Privacy Enhancing Technologies The 2022 IEEE International Conference on Privacy Computing (IEEE PriComp 2022) Dec. 15-18, Haikou, China [Submission Deadline: Sep. 1] http://www.ieee-smart-world.org/2022/pricomp/ PriComp 2022 is the 8th in this series of conferences started in 2015 that are devoted to algorithms and architectures for Privacy Computing. PriComp conference provides a forum for academics and practitioners from countries around the world to exchange ideas for improving the efficiency, performance, reliability, security and interoperability of Privacy Computing systems and applications. Following the traditions of the previous successful PriComp conferences held in Fuzhou, China (2015); Qingdao, China (2016); Melbourne, Australia (2017); Boppard, Germany (2018); Canterbury, UK (2019); Hainan, China (2020) and Xi'an, Shanghai, China (online, 2021); PriComp 2022 will be held in Haikou, China. PriComp 2022 will focus on an evolving pathway from privacy protection to privacy computing, by serving as an international premier forum for engineers and scientists in academia, industry, and government to address the resulting profound challenges and to present and discuss their new ideas, research results, applications and experience on all aspects of privacy computing. The conference of PriComp 2022 is co-organized by Chinese Information Processing Society of China, Institute of Information Engineering, CAS, and Hainan University. ================== Important Dates ================== Workshop Proposal: July 15, 2022 Paper Submission: September 01, 2022 Author Notification: October 01, 2022 Camera-Ready Submission: October 31, 2022 Conference Date: December 15-18, 2022 ================== Topics of interest include, but are not limited to ================== - Theories and foundations for privacy computing - Programming languages and compilers for privacy computing - Privacy computing models - Privacy metrics and formalization - Privacy taxonomies and ontologies - Privacy information management and engineering - Privacy operation and modeling - Data utility and privacy loss - Cryptography for privacy protection - Privacy protection based information hiding and sharing - Data analytics oriented privacy control and protection - Privacy-aware information collection - Privacy sensing and distribution - Combined and comprehensive privacy protection - Privacy-preserving data publishing - Private information storage - Private integration and synergy - Private information exchange and sharing - Privacy inference and reasoning - Internet and web privacy - Cloud privacy - Social media privacy - Mobile privacy - Location privacy - IoT privacy - Behavioral advertising - Privacy in large ecosystems such as smart cities - Privacy of AI models and systems - AI for privacy computing - Privacy and blockchain - User-centric privacy protection solutions - Human factors in privacy computing - Privacy nudging - Automated solutions for privacy policies and notices - Legal issues in privacy computing and other interdisciplinary topics ================== Paper Submission ================== All papers need to be submitted electronically through the conference submission website (https://edas.info/N29960) with PDF format. The materials presented in the papers should not be published or under submission elsewhere. Each paper is limited to 8 pages (or 10 pages with over length charge) including figures and references using IEEE Computer Society Proceedings Manuscripts style (two columns, single-spaced, 10 fonts). You can confirm the IEEE Computer Society Proceedings Author Guidelines at the following web page: http://www.computer.org/web/cs-cps/ Manuscript Templates for Conference Proceedings can be found at: https://www.ieee.org/conferences_events/conferences/publishing/templates.html Once accepted, the paper will be included into the IEEE conference proceedings published by IEEE Computer Society Press (indexed by EI). At least one of the authors of any accepted paper is requested to register the paper at the conference. ================== Special Issues ================== All accepted papers will be submitted to IEEE Xplore and Engineering Index (EI). Best Paper Awards will be presented to high quality papers. Distinguished papers, after further revisions, will be published in SCI & EI indexed prestigious journals. 1. Special issue on ?Dark side of the Socio-Cyber World: Media Manipulation, Fake News, and Misinformation?, IEEE Transactions on Computational Social Systems http://cyber-science.org/2022/assets/files/si/220412_IEEE%20TCSS_SI.pdf 2. Special issue on ?Responsible AI in Social Computing?, IEEE Transactions on Computational Social Systems https://www.ieeesmc.org/images/publications/pdfs/Call_for_Paper_-_SI_on_Responsible_AI_in_Social_Computing.pdf 3. Special issue on ?Decentralized Trust Management with Intelligence?, Information Science https://www.journals.elsevier.com/information-sciences/call-for-papers/decentralized-trust-management-with-intelligence 4. Special issue on ?Resource Sustainable Computational and Artificial Intelligence?, IEEE Transactions on Emerging Topics in Computational Intelligence 5. Special issue on ?Smart Blockchain for IoT Trust, Security and Privacy?, IEEE IoT Journal https://ieee-iotj.org/wp-content/uploads/2022/05/IEEEIoT-SmartBlockchain-TSP.pdf 6. Special issue on ?Edge Computing Optimization and Security?, Journal of Systems Architecture https://www.journals.elsevier.com/journal-of-systems-architecture/call-for-papers/edge-computing-optimization-and-security-vsi-edgeos2022 7. Special issue on ?Distributed Learning and Blockchain Enabled Infrastructures for Next Generation of Big Data Driven Cyber-Physical Systems?, Journal of Systems Architecture http://cyber-science.org/2022/assets/files/si/JSA_SI_0331.pdf 8. Special issue on ?Distributed and Collaborative Learning Empowered Edge Intelligence in Smart City?, ACM Transactions on Sensor Networks https://dl.acm.org/pb-assets/static_journal_pages/tosn/pdf/ACM_TOSN_CFP1210-1640635690003.pdf 9. Special issue on ?Robustness, Privacy, and Forensics in Intelligent Multimedia Systems? Information Science https://www.journals.elsevier.com/information-sciences/forthcoming-special-issues/robustness-privacy-and-forensics-in-intelligent-multimedia-systems * More special issues will be added later. http://www.ieee-smart-world.org/2022/pricomp/si.php ================== Organizing Committee ================== General Chairs - Fenghua Li, Institute of Information Engineering, CAS, China - Laurence T. Yang, Hainan University, China - Willy Susilo, University of Wollongong, Australia Program Chairs - Hui Li, Xidian University, China - Mamoun Alazab, Charles Darwin University, Australia - Jun Feng, Huazhong University of Science and Technology, China Local Chairs - Weidong Qiu, Shanghai Jiaotong University, China - Jieren Cheng, Hainan University, China Publicity Chairs - Bocheng Ren, Huazhong University of Science and Technology, China - Xin Nie, Huazhong University of Science and Technology, China - Peng Tang, Shanghai Jiao Tong University, China -- Dr. Jun Feng Huazhong University of Science and Technology Mobile: +86-18827365073 WeChat: junfeng10001000 E-Mail: junfeng989 at gmail.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From avinashsingh214 at gmail.com Sun Jul 17 10:38:05 2022 From: avinashsingh214 at gmail.com (Avinash K Singh) Date: Mon, 18 Jul 2022 00:38:05 +1000 Subject: Connectionists: Last date is approaching for SI: Advances and Challenges to Bridge Computational Intelligence and Neuroscience for Brain-computer Interface In-Reply-To: References: 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 *Submission date: 31st July* 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 -- Regards, Avinash K Singh -------------- next part -------------- An HTML attachment was scrubbed... URL: From iam.palat at gmail.com Sat Jul 16 11:22:13 2022 From: iam.palat at gmail.com (Iam Palatnik) Date: Sat, 16 Jul 2022 12:22:13 -0300 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: It's not clear to me how we could argue that large NLP models don't have an understanding of language. It may not be the best understanding of language (by whatever arbitrary metric), but it seems to me that it is definitely beyond that of other animals besides humans (many of which can be trained to react consistently to certain English words and even respond in the case of some birds). I'm keeping hand sign apes out of this one because there seems to be controversy regarding their real language skills, although I think it's just another facet of the same issue: "Are the birds/apes/machines just 'parroting' us without understanding?" When we are faced with anything that is not human presenting some level of understanding of language we seem to go into defensive mode about how that's not true understanding, and it's not clear to me that this isn't a 'human bias' so to speak. If an AI model is capable of creating english sentences vastly different from the ones in the training data, and it is capable of influencing this generation with theme prompts, how is this not some degree of understanding of english? To me it seems congruent with the knife usage example. SOTA NLP models seem to be quite decent at improvising with the linguistic knife. Multimodal image+text models are also gaining traction in image generation. I'd agree that it's not by any means the same type or perhaps level of human understanding, but is it really no 'understanding' at all? On Sat, Jul 16, 2022 at 4:56 AM Tsvi Achler wrote: > I think the core problem is feedforward networks. > In order for it to work correctly all weights need to be adjusted. This > then requires iid rehearsal which is unnatural. > Systems with massive feedback (much more than ART's vigilance) can avoid > this problem. > -Tsvi > > On Fri, Jul 15, 2022 at 1:21 AM Asim Roy wrote: > >> I think Steve Grossberg is generally correct. ART, RBF nets and similar >> flexible architectures can scale up almost linearly to problem size. My >> conjecture is that architectures that use distributed representation run >> into the scaling issue. On the other hand, however, distributed >> representation produces a more compact architecture compared to the shallow >> architectures of ART and RBFs. But, in terms of adding concepts, it?s easy >> to add a new object or concept to the shallow architectures. >> >> >> >> Perhaps someone can provide more insights on the architectural >> differences and the corresponding pros and cons of each. >> >> >> >> Best, >> >> Asim Roy >> >> Professor, Information Systems >> >> Arizona State University >> >> Lifeboat Foundation Bios: Professor Asim Roy >> >> >> Asim Roy | iSearch (asu.edu) >> >> >> >> >> >> >> *From:* Connectionists *On >> Behalf Of *Grossberg, Stephen >> *Sent:* Thursday, July 14, 2022 10:14 AM >> *To:* Danko Nikolic >> *Cc:* AIhub ; connectionists at mailman.srv.cs.cmu.edu >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Danko, >> >> >> >> I will respond to your comment below to the entire list. I recommend that >> future interactions be done between just the two of us. >> >> >> >> Everything that I write below is summarized in my new book: >> >> >> >> In brief, within Adaptive Resonance Theory, or ART, there IS no problem >> of scaling, in the sense that you describe it below, from learning (say) to >> correctly recognizing 100 objects to doing the same for 200. In fact, one >> of the properties for which I introduced ART in 1976 was to enable >> incremental learning in real time of arbitrary numbers of objects or events >> without experiencing catastrophic forgetting. >> >> >> >> I have called this a solution of the *stability-plasticity dilemma*: How >> our brains, and models like ART, can rapidly learn (plasticity) arbitrary >> numbers of objects without experiencing catastrophic forgetting (stability). >> >> >> >> I also derived ART in 1980, in an article within Psychological Review, >> from a *thought experiment* about how ANY system can AUTONOMOUSLY >> correct predictive errors in a changing world that is filled with >> unexpected events. >> >> >> >> This thought is experiment was derived from a few facts that are familiar >> to us all. They are familiar because they represent ubiquitous >> environmental pressures that we all experience. The thought experiment >> clarifies why, when they act together during the evolutionary process, >> models like ART are the unique result. >> >> >> >> Moveover, at no point during the thought experiment are the words mind or >> brain mentioned. ART is thus a *universal solution* of this learning, >> classification, and prediction problem. >> >> >> >> You write below about "connectionist systems". ART is a connectionist >> system. >> >> >> >> What do you mean by a "connectionist system"? What you write below about >> them is not true in general. >> >> >> >> Best, >> >> >> >> Steve >> ------------------------------ >> >> *From:* Danko Nikolic >> *Sent:* Thursday, July 14, 2022 12:16 PM >> *To:* Grossberg, Stephen >> *Cc:* Gary Marcus ; >> connectionists at mailman.srv.cs.cmu.edu < >> connectionists at mailman.srv.cs.cmu.edu>; AIhub >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Steve, >> >> >> >> Thank you very much for your message and for the greetings. I will pass >> them on if an occasion arises. >> >> >> >> Regarding your question: The key problem I am trying to address and that, >> to the best of my knowledge, no connectionist system was able to solve so >> far is that of scaling the system's intelligence. For example, if the >> system is able to correctly recognize 100 different objects, how many >> additional resources are needed to double that to 200? All the empirical >> data show that connectionist systems scale poorly: Some of the best systems >> we have require 500x more resources in order to increase the intelligence >> by only 2x. I document this problem in the manuscript and even run some >> simulations to show that the worst performance is if connectionist systems >> need to solve a generalized XOR problem. >> >> >> >> In contrast, the biological brain scales well. This I also quantify in >> the paper. >> >> >> >> I will look at the publication that you mentioned. However, so far, I >> haven't seen a solution that scales well in intelligence. >> >> >> >> My argument is that transient selection of subnetworks by the help of the >> mentioned proteins is how intelligence scaling is achieved in biological >> brains. >> >> >> >> In short, intelligence scaling is the key problem that concerns me. I >> describe the intelligence scaling problem in more detail in this book that >> just came out a few weeks ago and that is written for practitioners in Data >> Scientist and AI: https://amzn.to/3IBxUpL >> >> >> >> >> I hope that this at least partly answers where I see the problems and >> what I am trying to solve. >> >> >> >> Greetings from Germany, >> >> >> >> Danko >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> >> https://www.linkedin.com/in/danko-nikolic/ >> >> >> --- A progress usually starts with an insight --- >> >> >> >> >> >> On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: >> >> Dear Danko, >> >> >> >> I have just read your new article and would like to comment briefly about >> it. >> >> >> >> In your introductory remarks, you write: >> >> >> >> "However, connectionism did not yet produce a satisfactory explanation of >> how the mental emerges from the physical. A number of open problems >> remains ( 5,6,7,8). As a result, the explanatory gap between the mind and >> the brain remains wide open." >> >> >> >> I certainly believe that no theoretical explanation in science is ever >> complete. However, I also believe that "the explanatory gap between the >> mind and the brain" does not remain "wide open". >> >> >> >> My Magnum Opus, that was published in 2021, makes that belief clear in >> its title: >> >> >> >> *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* >> >> >> >> https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 >> >> >> >> >> The book provides a self-contained and non-technical exposition in a >> conversational tone of many principled and unifying explanations of >> psychological and neurobiological data. >> >> >> >> In particular, it explains roles for the metabotropic glutamate receptors >> that you mention in your own work. See the text and figures around p. 521. >> This explanation unifies psychological, anatomical, neurophysiological, >> biophysical, and biochemical data about the processes under discussion. >> >> >> >> I have a very old-fashioned view about how to understand scientific >> theories. I get excited by theories that explain and predict more data than >> previous theories. >> >> >> >> Which of the data that I explain in my book, and support with >> quantitative computer simulations, can you also explain? >> >> >> >> What data can you explain, in the same quantitative sense, that you do >> not think the neural models in my book can explain? >> >> >> >> I would be delighted to discuss these issues further with you. >> >> >> >> If you are in touch with my old friend and esteemed colleague, Wolf >> Singer, please send him my warm regards. I cite the superb work that he and >> various of his collaborators have done in many places in my book. >> >> >> >> Best, >> >> >> >> Steve >> >> >> >> 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 Danko Nikolic >> *Sent:* Thursday, July 14, 2022 6:05 AM >> *To:* Gary Marcus >> *Cc:* connectionists at mailman.srv.cs.cmu.edu < >> connectionists at mailman.srv.cs.cmu.edu>; AIhub >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Gary and everyone, >> >> >> >> I am continuing the discussion from where we left off a few months ago. >> Back then, some of us agreed that the problem of understanding remains >> unsolved. >> >> >> >> As a reminder, the challenge for connectionism was to 1) learn with few >> examples and 2) apply the knowledge to a broad set of situations. >> >> >> >> I am happy to announce that I have now finished a draft of a paper in >> which I propose how the brain is able to achieve that. The manuscript >> requires a bit of patience for two reasons: one is that the reader may be >> exposed for the first time to certain aspects of brain physiology. The >> second reason is that it may take some effort to understand the >> counterintuitive implications of the new ideas (this requires a different >> way of thinking than what we are used to based on connectionism). >> >> >> >> In short, I am suggesting that instead of the connectionist paradigm, we >> adopt transient selection of subnetworks. The mechanisms that transiently >> select brain subnetworks are distributed all over the nervous system and, I >> argue, are our main machinery for thinking/cognition. The surprising >> outcome is that neural activation, which was central in connectionism, now >> plays only a supportive role, while the real 'workers' within the brain are >> the mechanisms for transient selection of subnetworks. >> >> >> >> I also explain how I think transient selection achieves learning with >> only a few examples and how the learned knowledge is possible to apply to a >> broad set of situations. >> >> >> >> The manuscript is made available to everyone and can be downloaded here: >> https://bit.ly/3IFs8Ug >> >> >> (I apologize for the neuroscience lingo, which I tried to minimize.) >> >> >> >> It will likely take a wide effort to implement these concepts as an AI >> technology, provided my ideas do not have a major flaw in the first place. >> Does anyone see a flaw? >> >> >> >> Thanks. >> >> >> >> Danko >> >> >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> >> https://www.linkedin.com/in/danko-nikolic/ >> >> >> >> >> >> >> On Thu, Feb 3, 2022 at 5:25 PM 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 Sun Jul 17 18:58:54 2022 From: gary at eng.ucsd.edu (gary@ucsd.edu) Date: Sun, 17 Jul 2022 15:58:54 -0700 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: Sorry, I can't let this go by: And it is good so because generalize XOR scales worse than power law. It scales exponentially! This a more agressive form of explosion than power law. I'm not sure exactly what you mean by this, but a single-hidden layer network with N inputs and N hidden units can solve N-bit parity. Each unit has an increasing threshold, so, one turns on if there is one unit on in the input, and then turns on the output with a weight of +1. If two units are on in the input, then a second unit comes on and cancels the activation of the first unit via a weight of -1. Etc. g. On Sat, Jul 16, 2022 at 12:03 AM Danko Nikolic wrote: > Dear Thomas, > > Thank you for reading the paper and for the comments. > > I cite: "In my experience, supervised classification scales linearly in > the number of classes." > This would be good to quantify as a plot. Maybe a research paper would be > a good idea. The reason is that it seems that everyone else who tried to > quantify that relation found a power law. At this point, it would be > surprising to find a linear relationship. And it would probably make a well > read paper. > > But please do not forget that my argument states that even a linear > relationship is not good enough to match bilogical brains. We need > something more similar to a power law with exponent zero when it comes to > the model size i.e., a constant number of parameters in the model. And we > need linear relationship when it comes to learning time: Each newly learned > object should needs about as much of learning effort as was needed for each > previous object. > > I cite: "The real world is not dominated by generalized XOR problems." > Agreed. And it is good so because generalize XOR scales worse than power > law. It scales exponentially! This a more agressive form of explosion than > power law. > Importantly, a generalized AND operation also scales exponentially (with a > smaller exponent, though). I guess we would agree that the real world > probably encouners a lot of AND problems. The only logical operaiton that > could be learned with a linear increase in the number of parameters was a > generalized OR. Finally, I foiund that a mixure of AND and OR resulted in a > power law-like scaling of the number of parameters. So, a mixture of AND > and OR seemed to scale as good (or as bad) as the real world. I have put > this information into Supplementary Materials. > > The conclusion that I derived from those analyses is: connectionism is not > sustainable to reach human (or animal) levels of intelligence. Therefore, I > hunted for an alternative pradigm. > > Greetings, > > Danko > > > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > -- I wonder, how is the brain able to generate insight? -- > > > On Fri, Jul 15, 2022 at 10:01 AM Dietterich, Thomas > wrote: > >> Dear Danko, >> >> >> >> In my experience, supervised classification scales linearly in the number >> of classes. Of course it depends to some extent on how subtle the >> distinctions are between the different categories. The real world is not >> dominated by generalized XOR problems. >> >> >> >> --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:* Connectionists *On >> Behalf Of *Danko Nikolic >> *Sent:* Thursday, July 14, 2022 09:17 >> *To:* Grossberg, Stephen >> *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.] >> >> Dear Steve, >> >> >> >> Thank you very much for your message and for the greetings. I will pass >> them on if an occasion arises. >> >> >> >> Regarding your question: The key problem I am trying to address and that, >> to the best of my knowledge, no connectionist system was able to solve so >> far is that of scaling the system's intelligence. For example, if the >> system is able to correctly recognize 100 different objects, how many >> additional resources are needed to double that to 200? All the empirical >> data show that connectionist systems scale poorly: Some of the best systems >> we have require 500x more resources in order to increase the intelligence >> by only 2x. I document this problem in the manuscript and even run some >> simulations to show that the worst performance is if connectionist systems >> need to solve a generalized XOR problem. >> >> >> >> In contrast, the biological brain scales well. This I also quantify in >> the paper. >> >> >> >> I will look at the publication that you mentioned. However, so far, I >> haven't seen a solution that scales well in intelligence. >> >> >> >> My argument is that transient selection of subnetworks by the help of the >> mentioned proteins is how intelligence scaling is achieved in biological >> brains. >> >> >> >> In short, intelligence scaling is the key problem that concerns me. I >> describe the intelligence scaling problem in more detail in this book that >> just came out a few weeks ago and that is written for practitioners in Data >> Scientist and AI: https://amzn.to/3IBxUpL >> >> >> >> >> I hope that this at least partly answers where I see the problems and >> what I am trying to solve. >> >> >> >> Greetings from Germany, >> >> >> >> Danko >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> >> https://www.linkedin.com/in/danko-nikolic/ >> >> >> --- A progress usually starts with an insight --- >> >> >> >> >> >> On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen wrote: >> >> Dear Danko, >> >> >> >> I have just read your new article and would like to comment briefly about >> it. >> >> >> >> In your introductory remarks, you write: >> >> >> >> "However, connectionism did not yet produce a satisfactory explanation of >> how the mental emerges from the physical. A number of open problems >> remains ( 5,6,7,8). As a result, the explanatory gap between the mind and >> the brain remains wide open." >> >> >> >> I certainly believe that no theoretical explanation in science is ever >> complete. However, I also believe that "the explanatory gap between the >> mind and the brain" does not remain "wide open". >> >> >> >> My Magnum Opus, that was published in 2021, makes that belief clear in >> its title: >> >> >> >> *Conscious Mind, Resonant Brain: How Each Brain Makes a Mind* >> >> >> >> https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 >> >> >> >> >> The book provides a self-contained and non-technical exposition in a >> conversational tone of many principled and unifying explanations of >> psychological and neurobiological data. >> >> >> >> In particular, it explains roles for the metabotropic glutamate receptors >> that you mention in your own work. See the text and figures around p. 521. >> This explanation unifies psychological, anatomical, neurophysiological, >> biophysical, and biochemical data about the processes under discussion. >> >> >> >> I have a very old-fashioned view about how to understand scientific >> theories. I get excited by theories that explain and predict more data than >> previous theories. >> >> >> >> Which of the data that I explain in my book, and support with >> quantitative computer simulations, can you also explain? >> >> >> >> What data can you explain, in the same quantitative sense, that you do >> not think the neural models in my book can explain? >> >> >> >> I would be delighted to discuss these issues further with you. >> >> >> >> If you are in touch with my old friend and esteemed colleague, Wolf >> Singer, please send him my warm regards. I cite the superb work that he and >> various of his collaborators have done in many places in my book. >> >> >> >> Best, >> >> >> >> Steve >> >> >> >> 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 Danko Nikolic >> *Sent:* Thursday, July 14, 2022 6:05 AM >> *To:* Gary Marcus >> *Cc:* connectionists at mailman.srv.cs.cmu.edu < >> connectionists at mailman.srv.cs.cmu.edu>; AIhub >> *Subject:* Re: Connectionists: Stephen Hanson in conversation with Geoff >> Hinton >> >> >> >> Dear Gary and everyone, >> >> >> >> I am continuing the discussion from where we left off a few months ago. >> Back then, some of us agreed that the problem of understanding remains >> unsolved. >> >> >> >> As a reminder, the challenge for connectionism was to 1) learn with few >> examples and 2) apply the knowledge to a broad set of situations. >> >> >> >> I am happy to announce that I have now finished a draft of a paper in >> which I propose how the brain is able to achieve that. The manuscript >> requires a bit of patience for two reasons: one is that the reader may be >> exposed for the first time to certain aspects of brain physiology. The >> second reason is that it may take some effort to understand the >> counterintuitive implications of the new ideas (this requires a different >> way of thinking than what we are used to based on connectionism). >> >> >> >> In short, I am suggesting that instead of the connectionist paradigm, we >> adopt transient selection of subnetworks. The mechanisms that transiently >> select brain subnetworks are distributed all over the nervous system and, I >> argue, are our main machinery for thinking/cognition. The surprising >> outcome is that neural activation, which was central in connectionism, now >> plays only a supportive role, while the real 'workers' within the brain are >> the mechanisms for transient selection of subnetworks. >> >> >> >> I also explain how I think transient selection achieves learning with >> only a few examples and how the learned knowledge is possible to apply to a >> broad set of situations. >> >> >> >> The manuscript is made available to everyone and can be downloaded here: >> https://bit.ly/3IFs8Ug >> >> >> (I apologize for the neuroscience lingo, which I tried to minimize.) >> >> >> >> It will likely take a wide effort to implement these concepts as an AI >> technology, provided my ideas do not have a major flaw in the first place. >> Does anyone see a flaw? >> >> >> >> Thanks. >> >> >> >> Danko >> >> >> >> >> Dr. Danko Nikoli? >> www.danko-nikolic.com >> >> https://www.linkedin.com/in/danko-nikolic/ >> >> >> >> >> >> >> On Thu, Feb 3, 2022 at 5:25 PM 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 krallinger.martin at gmail.com Mon Jul 18 05:25:23 2022 From: krallinger.martin at gmail.com (Martin Krallinger) Date: Mon, 18 Jul 2022 11:25:23 +0200 Subject: Connectionists: CFP: ClinSpEn subtask at Biomedical WMT Shared Task (WMT/EMNLP 2022): translation of clinical cases, entities, terminologies and ontologies In-Reply-To: References: Message-ID: Call for Participation ClinSpEn @ Biomedical WMT Shared Task (WMT/EMNLP 2022) Automatic Translation of Clinical cases, ontologies & medical entities: Spanish - English https://temu.bsc.es/clinspen/ ClinSpEn is part of the Biomedical WMT 2022 shared task, having the aim to promote the development and evaluation of machine translation systems adapted to the medical domain with three highly relevant sub-tracks: clinical cases, medical controlled vocabularies/ontologies, and clinical terms and entities extracted from medical content. Key information: - ClinSpEn sub-track: https://temu.bsc.es/clinspen/ - Biomedical WMT: https://statmt.org/wmt22/biomedical-translation-task.html - Main WMT: https://statmt.org/wmt22/ - EMNLP conference: https://2022.emnlp.org/ - Sample/Training Data: - Clinical Cases: https://doi.org/10.5281/zenodo.6497350 - Clinical Terms: https://doi.org/10.5281/zenodo.6497372 - Ontology Concepts: https://doi.org/10.5281/zenodo.6497388 - Registration/support: https://temu.bsc.es/clinspen/registration/ Motivation Machine translation applied to the clinical domain is a specially challenging task due to the complexity of medical language and the heavy use of health-related technical terms and medical expressions. Therefore there is a large community of specialized medical translators, able to deal with medical narratives, terminologies or the use of ambiguous abbreviations and acronyms. Taking into account the relevance, impact and diversity of health-related content, as well as the rapidly growing number of publications, EHRs, clinical trials, informed consent documents and medical terminologies there is a pressing need to be able to generate more robust medical machine translation resources together with independent quality evaluation scenarios. Recent advances in machine translation technologies together with the use of other NLP components are showing promising results, thus domain adaptation of MT approaches can have a significant impact in unlocking key information from medical content. The ClinSpEn sub-task of Biomedical WMT proposes three different highly relevant sub-tracks, each associated with highly relevant medical machine translation application scenarios:: - ClinSpEn-CC (Clinical Cases) subtask: translation of clinical case documents from English to Spanish, a type of document relevant both for processing medical literature as well as clinical records. - ClinSpEn-CT (Clinical Terms): translation of clinical terms and entity mentions from Spanish to English. The use terms were directly extracted from medical literature and clinical records, with particular focus on diseases, symptoms, findings, procedures and professions. - ClinSpEn-OC (Ontology Concepts): translation of clinical controlled vocabularies and ontology concepts from English to Spanish. Ontologies and structured vocabularies represent a key resource for semantic interoperability, entity linking, biomedical knlwedgebases and precision medicine, and thus there is a pressing need to generate multilingual biomedical ontologies for a range of clinicla applications. . A decently-sized sample set for each data type has been released. Participants may use it to test their existing systems or try out new ones. In addition to the manually translated test set by professional medical translators, participants will also have access to a larger background collection for each of the three substracks, which might serve as additional resources and to promote scalability and robustness assessment of machine translation technology. Schedule - Test and Background Set Release: July 21st, 2022 - Participant Predictions Due: July 28th, 2022 - Paper Submission Deadline: September 7th, 2022 - Notification of Acceptance (peer-reviews): October 9th, 2022 - Camera-ready Version Due: October 16th, 2022 - WMT @ EMNLP: December 7th and 8th, 2022 [All deadlines are in AoE (Anywhere on Earth)] Registration For the time being, participants may register using the ClinSpEn registration form at: https://temu.bsc.es/clinspen/registration/. This form will be used to support teams during their participation and keep them updated on the official WMT/EMNLP registration, as well as on all related deadlines and important news. Publications and WMT workshop Teams participating in the ClinSpEn subtrack of Biomedical WMT will be invited to contribute a systems description paper for the WMT 2022 Working Notes proceedings. More information on the paper?s specifications, formatting guidelines and review process at: https://statmt.org/wmt22/index.html. If you are interested in Machine Translation, the biomedical domain or other language combinations, remember to check out the Biomedical WMT site and the rest of this year?s sub-tracks and language pairs: https://statmt.org/wmt22/biomedical-translation-task.html ClinSpEn Organizers - Salvador Lima-L?pez (Barcelona Supercomputing Center, Spain) - Darryl Johan Estrada (Barcelona Supercomputing Center, Spain) - Eul?lia Farr?-Maduell (Barcelona Supercomputing Center, Spain) - Martin Krallinger (Barcelona Supercomputing Center, Spain) Biomedical WMT Organizers - Rachel Bawden (University of Edinburgh, UK) - Giorgio Maria Di Nunzio (University of Padua, Italy) - Darryl Johan Estrada (Barcelona Supercomputing Center, Spain) - Eul?lia Farr?-Maduell (Barcelona Supercomputing Center, Spain) - Cristian Grozea (Fraunhofer Institute, Germany) - Antonio Jimeno Yepes (University of Melbourne, Australia) - Salvador Lima-L?pez (Barcelona Supercomputing Center, Spain) - Martin Krallinger (Barcelona Supercomputing Center, Spain) - Aur?lie N?v?ol (Universit? Paris Saclay, CNRS, LISN, France) - Mariana Neves (German Federal Institute for Risk Assessment, Germany) - Roland Roller (DFKI, Germany) - Amy Siu (Beuth University of Applied Sciences, Germany) - Philippe Thomas (DFKI, Germany) - Federica Vezzani (University of Padua, Italy) - Maika Vicente Navarro, Maika Spanish Translator, Melbourne, Australia - Dina Wiemann (Novartis, Switzerland) - Lana Yeganova (NCBI/NLM/NIH, USA -------------- next part -------------- An HTML attachment was scrubbed... URL: From barak at pearlmutter.net Mon Jul 18 07:12:37 2022 From: barak at pearlmutter.net (Barak A. Pearlmutter) Date: Mon, 18 Jul 2022 12:12:37 +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: On Mon, 18 Jul 2022 at 08:28, Danko Nikolic wrote: > In short, learning mechanisms cannot discover generalized XOR functions with simple connectivity -- only with complex connectivity. This problem results in exponential growth of needed resources as the number of bits in the generalized XOR increases. Assuming that "generalized XOR" means parity, this must rely on some unusual definitions which you should probably state in order to avoid confusion. Parity is a poster boy for an *easy* function to learn, albeit a nonlinear one. This is because in the (boolean) Fourier domain its spectrum consists of a single nonzero coefficient, and functions that are sparse in that domain are very easy to learn. See N. Linial, Y. Mansour, and N. Nisan, "Constant depth circuits, Fourier Transform and learnability", FOCS 1989, or Mansour, Y. (1994). Learning Boolean Functions via the Fourier Transform. Theoretical Advances in Neural Computation and Learning, 391?424. doi:10.1007/978-1-4615-2696-4_11 --Barak Pearlmutter From gary.marcus at nyu.edu Mon Jul 18 08:59:51 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Mon, 18 Jul 2022 05:59:51 -0700 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: identity is just as ?easy?. but as I showed long ago (1998 and 2001), what is learned is specific to a space of training examples. there is interpolation in that space, but no reliable extrapolation outside that space. eg if you arrange the problem as training over even numbers represented in binary digits in a standard multi-layer perceptron, the system will not generalize properly to odd numbers. nowadays Bengio and others call this the problem of distribution shift, and you would get the same sort of thing with parity (with a slightly different example), because what is learned and described as ?parity? is fairly superficial, example-based rather than fully abstract. gary > On Jul 18, 2022, at 05:46, Barak A. Pearlmutter wrote: > > ?On Mon, 18 Jul 2022 at 08:28, Danko Nikolic wrote: >> In short, learning mechanisms cannot discover generalized XOR functions with simple connectivity -- only with complex connectivity. This problem results in exponential growth of needed resources as the number of bits in the generalized XOR increases. > > Assuming that "generalized XOR" means parity, this must rely on some > unusual definitions which you should probably state in order to avoid > confusion. > > Parity is a poster boy for an *easy* function to learn, albeit a > nonlinear one. This is because in the (boolean) Fourier domain its > spectrum consists of a single nonzero coefficient, and functions that > are sparse in that domain are very easy to learn. See N. Linial, Y. > Mansour, and N. Nisan, "Constant depth circuits, Fourier Transform and > learnability", FOCS 1989, or Mansour, Y. (1994). Learning Boolean > Functions via the Fourier Transform. Theoretical Advances in Neural > Computation and Learning, 391?424. doi:10.1007/978-1-4615-2696-4_11 > > --Barak Pearlmutter > From danny.silver at acadiau.ca Mon Jul 18 09:05:57 2022 From: danny.silver at acadiau.ca (Danny Silver) Date: Mon, 18 Jul 2022 13:05:57 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: Dear Connectionists I am hoping that LeCun and Browning meant that symbolic manipulation was needed as an explanatory component of human-like AI. That is, if we wish for the agent to appear human in its intelligence, then it needs to be able to explain (to some extent) how it reached its conclusion. This does not necessitate that reasoning is done with symbols. There is a difference. Animals do complex intelligent things every day, but do not display human-like AI; ie. they cannot explain their decisions with symbols. And humans make great decision every day which they cannot explain fully using symbols; NOR, I suspect, required the use of symbols as fundamental building blocks. My comments are not meant to diminish the importance of symbols in human intelligence ? but rather to distinguish symbols as tools we use to constrain (justify internally) and transfer (communicate externally) human learning and reasoning. I do not see symbols as the foundational components of human learning and reasoning, but they are essential to human-like AI. As I have mentioned earlier in an email. We need to start thinking about architectures that can both (1) learn to classify from a sequence of features extracted from our senses, and (2) learn to explain (and therefore justify) how that classification was done. We have already had some early success in this area ? see ?Learning Arithmetic from Handwritten Images with the Aid of Symbols? ? Danny Silver From: Connectionists on behalf of Gary Marcus Date: Monday, July 18, 2022 at 5:06 AM To: Stephen Jos? Hanson Cc: connectionists at mailman.srv.cs.cmu.edu , AIhub Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton CAUTION: This email comes from outside Acadia. Verify the sender and use caution with any requests, links or attachments. I can?t help but note a profound tension between these two very recent quotes: Hanson, 2022, below: as we go forward, let?s avoid at all costs breaking ?the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet.? LeCun and Browning, 2022: ?everyone working in DL agrees that symbolic manipulation is a necessary feature for creating human-like AI.? https://www.noemamag.com/what-ai-can-tell-us-about-intelligence On Jul 16, 2022, at 12:11 AM, Stephen Jos? Hanson wrote: ? So there are a number of issues here. In the first one, the brain obviously has a specific structure. DL supervised modeling of classification, NOT psychological categorization-- may scale linearly with resources.. however, its clear just adding more hidden layers has diminishing returns, adding various kinds of classification structure (GANs etc) or doing unsupervised learning or reinforcement learning is not going to end up with AGI or really any sort of final AI system and certainly not biological. Humans btw, scale as sub-linear systems.. but with hard limits. We can only recognize about 6000 faces (and the variance is about 4k across individuals), most undergraduates can recognize 100k words again with a large variance., so despite the fact that the brain has trillions of synapses, "loading" in explicit (MTL) memory is complex and somehow localized. Memory systems involve implicit (most DL) or explicit (almost noone), working memory is tiny--like 3 or 4+- 2 (sorry George). Many (Bengio, Le Cun etc) are experimenting with more complex structures, but keeping the system homogeneous, no extra symbolic baggage and forcing the exercise to have to learn the configuration rather than engineer it. In some ways, Grossberg has been out in front in all this, but with his own special and situated neural systems. He has been consistently and persistently creating focused neural modeling, which as an art form has always been a matter of taste. Connectionist models tend toward minimalist forms and Steve and others follow more Baroque or perhaps Victorian forms. What is clear is that the future is going to be blends of these forms and more complex, without at the same time breaking the the Gary Marcus fire alarm glass where symbolic piece parts tumble out of the closet. Steve On 7/15/22 4:01 AM, Dietterich, Thomas wrote: Dear Danko, In my experience, supervised classification scales linearly in the number of classes. Of course it depends to some extent on how subtle the distinctions are between the different categories. The real world is not dominated by generalized XOR problems. --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: Connectionists On Behalf Of Danko Nikolic Sent: Thursday, July 14, 2022 09:17 To: Grossberg, Stephen 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.] Dear Steve, Thank you very much for your message and for the greetings. I will pass them on if an occasion arises. Regarding your question: The key problem I am trying to address and that, to the best of my knowledge, no connectionist system was able to solve so far is that of scaling the system's intelligence. For example, if the system is able to correctly recognize 100 different objects, how many additional resources are needed to double that to 200? All the empirical data show that connectionist systems scale poorly: Some of the best systems we have require 500x more resources in order to increase the intelligence by only 2x. I document this problem in the manuscript and even run some simulations to show that the worst performance is if connectionist systems need to solve a generalized XOR problem. In contrast, the biological brain scales well. This I also quantify in the paper. I will look at the publication that you mentioned. However, so far, I haven't seen a solution that scales well in intelligence. My argument is that transient selection of subnetworks by the help of the mentioned proteins is how intelligence scaling is achieved in biological brains. In short, intelligence scaling is the key problem that concerns me. I describe the intelligence scaling problem in more detail in this book that just came out a few weeks ago and that is written for practitioners in Data Scientist and AI: https://amzn.to/3IBxUpL I hope that this at least partly answers where I see the problems and what I am trying to solve. Greetings from Germany, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ --- A progress usually starts with an insight --- On Thu, Jul 14, 2022 at 3:30 PM Grossberg, Stephen > wrote: Dear Danko, I have just read your new article and would like to comment briefly about it. In your introductory remarks, you write: "However, connectionism did not yet produce a satisfactory explanation of how the mental emerges from the physical. A number of open problems remains ( 5,6,7,8). As a result, the explanatory gap between the mind and the brain remains wide open." I certainly believe that no theoretical explanation in science is ever complete. However, I also believe that "the explanatory gap between the mind and the brain" does not remain "wide open". My Magnum Opus, that was published in 2021, makes that belief clear in its title: Conscious Mind, Resonant Brain: How Each Brain Makes a Mind https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552 The book provides a self-contained and non-technical exposition in a conversational tone of many principled and unifying explanations of psychological and neurobiological data. In particular, it explains roles for the metabotropic glutamate receptors that you mention in your own work. See the text and figures around p. 521. This explanation unifies psychological, anatomical, neurophysiological, biophysical, and biochemical data about the processes under discussion. I have a very old-fashioned view about how to understand scientific theories. I get excited by theories that explain and predict more data than previous theories. Which of the data that I explain in my book, and support with quantitative computer simulations, can you also explain? What data can you explain, in the same quantitative sense, that you do not think the neural models in my book can explain? I would be delighted to discuss these issues further with you. If you are in touch with my old friend and esteemed colleague, Wolf Singer, please send him my warm regards. I cite the superb work that he and various of his collaborators have done in many places in my book. Best, Steve 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 Danko Nikolic > Sent: Thursday, July 14, 2022 6:05 AM To: Gary Marcus > Cc: connectionists at mailman.srv.cs.cmu.edu >; AIhub > Subject: Re: Connectionists: Stephen Hanson in conversation with Geoff Hinton Dear Gary and everyone, I am continuing the discussion from where we left off a few months ago. Back then, some of us agreed that the problem of understanding remains unsolved. As a reminder, the challenge for connectionism was to 1) learn with few examples and 2) apply the knowledge to a broad set of situations. I am happy to announce that I have now finished a draft of a paper in which I propose how the brain is able to achieve that. The manuscript requires a bit of patience for two reasons: one is that the reader may be exposed for the first time to certain aspects of brain physiology. The second reason is that it may take some effort to understand the counterintuitive implications of the new ideas (this requires a different way of thinking than what we are used to based on connectionism). In short, I am suggesting that instead of the connectionist paradigm, we adopt transient selection of subnetworks. The mechanisms that transiently select brain subnetworks are distributed all over the nervous system and, I argue, are our main machinery for thinking/cognition. The surprising outcome is that neural activation, which was central in connectionism, now plays only a supportive role, while the real 'workers' within the brain are the mechanisms for transient selection of subnetworks. I also explain how I think transient selection achieves learning with only a few examples and how the learned knowledge is possible to apply to a broad set of situations. The manuscript is made available to everyone and can be downloaded here: https://bit.ly/3IFs8Ug (I apologize for the neuroscience lingo, which I tried to minimize.) It will likely take a wide effort to implement these concepts as an AI technology, provided my ideas do not have a major flaw in the first place. Does anyone see a flaw? Thanks. Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ On Thu, Feb 3, 2022 at 5:25 PM 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 --- [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 -- Stephen Jose Hanson Professor, Director Psychology Department, RUBIC (Rutgers University Brain Imaging Center) -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Mon Jul 18 09:43:17 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Mon, 18 Jul 2022 15:43:17 +0200 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 Barak, Generalized XOR is defined in the manuscript, in the Supplementary Materials. Here is a snapshot from a Figure that illustrates how to construct it. Also the code for generating such training data is provided. [image: image.png] It is a hard problem to learn for a connectionist network. Perhaps all of the computers in the world if worked together for 100 years could not learn the problem with only 20 bits depth, provided that they used standard deep learning techniques. And yet, it is trivially easy to create a solution for an engineer who uses their human mind to understand(!) the problem. Greetings, Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Mon, Jul 18, 2022 at 1:12 PM Barak A. Pearlmutter wrote: > On Mon, 18 Jul 2022 at 08:28, Danko Nikolic > wrote: > > In short, learning mechanisms cannot discover generalized XOR functions > with simple connectivity -- only with complex connectivity. This problem > results in exponential growth of needed resources as the number of bits in > the generalized XOR increases. > > Assuming that "generalized XOR" means parity, this must rely on some > unusual definitions which you should probably state in order to avoid > confusion. > > Parity is a poster boy for an *easy* function to learn, albeit a > nonlinear one. This is because in the (boolean) Fourier domain its > spectrum consists of a single nonzero coefficient, and functions that > are sparse in that domain are very easy to learn. See N. Linial, Y. > Mansour, and N. Nisan, "Constant depth circuits, Fourier Transform and > learnability", FOCS 1989, or Mansour, Y. (1994). Learning Boolean > Functions via the Fourier Transform. Theoretical Advances in Neural > Computation and Learning, 391?424. doi:10.1007/978-1-4615-2696-4_11 > > --Barak Pearlmutter > -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 71360 bytes Desc: not available URL: From tgd at oregonstate.edu Mon Jul 18 09:49:31 2022 From: tgd at oregonstate.edu (Dietterich, Thomas) Date: Mon, 18 Jul 2022 13:49:31 +0000 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: This depends crucially on the vocabulary (representation). If I look at Fourier components, I can generalize in one way; if I treat each input vector as unique (as in a lookup table), I can't generalize at all. People are able to represent inputs in a wide variety of ways, as shown by their performance on Bongard problems, for example. These can involve relationships over relationships, and other recursive structures. Representation learning still has a ways to go. --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 -----Original Message----- From: Connectionists On Behalf Of Gary Marcus Sent: Monday, July 18, 2022 06:00 To: Barak A. Pearlmutter Cc: Gary Cottrell ; 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.] identity is just as ?easy?. but as I showed long ago (1998 and 2001), what is learned is specific to a space of training examples. there is interpolation in that space, but no reliable extrapolation outside that space. eg if you arrange the problem as training over even numbers represented in binary digits in a standard multi-layer perceptron, the system will not generalize properly to odd numbers. nowadays Bengio and others call this the problem of distribution shift, and you would get the same sort of thing with parity (with a slightly different example), because what is learned and described as ?parity? is fairly superficial, example-based rather than fully abstract. gary > On Jul 18, 2022, at 05:46, Barak A. Pearlmutter wrote: > > ?On Mon, 18 Jul 2022 at 08:28, Danko Nikolic wrote: >> In short, learning mechanisms cannot discover generalized XOR functions with simple connectivity -- only with complex connectivity. This problem results in exponential growth of needed resources as the number of bits in the generalized XOR increases. > > Assuming that "generalized XOR" means parity, this must rely on some > unusual definitions which you should probably state in order to avoid > confusion. > > Parity is a poster boy for an *easy* function to learn, albeit a > nonlinear one. This is because in the (boolean) Fourier domain its > spectrum consists of a single nonzero coefficient, and functions that > are sparse in that domain are very easy to learn. See N. Linial, Y. > Mansour, and N. Nisan, "Constant depth circuits, Fourier Transform and > learnability", FOCS 1989, or Mansour, Y. (1994). Learning Boolean > Functions via the Fourier Transform. Theoretical Advances in Neural > Computation and Learning, 391?424. doi:10.1007/978-1-4615-2696-4_11 > > --Barak Pearlmutter > From barak at pearlmutter.net Mon Jul 18 10:17:41 2022 From: barak at pearlmutter.net (Barak A. Pearlmutter) Date: Mon, 18 Jul 2022 15:17:41 +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: On Mon, 18 Jul 2022 at 14:43, Danko Nikolic wrote: > > [image: image.png] > > It is a hard problem to learn for a connectionist network. > We don't need to invent new terminology, like "inverters problem" or "generalized xor." This is parity. Four (4) bit parity. https://en.wikipedia.org/wiki/Parity_function Parity is *not* a hard function to learn. Even for a connectionist network. It is an interesting function for historic reasons (n-bit parity cannot be loaded by a k-th order perceptron, for k -------------- next part -------------- A non-text attachment was scrubbed... Name: image.png Type: image/png Size: 71360 bytes Desc: not available URL: From gary.marcus at nyu.edu Mon Jul 18 10:18:51 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Mon, 18 Jul 2022 07:18:51 -0700 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: References: Message-ID: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> completely agree on all points > On Jul 18, 2022, at 6:49 AM, Dietterich, Thomas wrote: > > ?This depends crucially on the vocabulary (representation). If I look at Fourier components, I can generalize in one way; if I treat each input vector as unique (as in a lookup table), I can't generalize at all. People are able to represent inputs in a wide variety of ways, as shown by their performance on Bongard problems, for example. These can involve relationships over relationships, and other recursive structures. Representation learning still has a ways to go. > > --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 > > -----Original Message----- > From: Connectionists On Behalf Of Gary Marcus > Sent: Monday, July 18, 2022 06:00 > To: Barak A. Pearlmutter > Cc: Gary Cottrell ; 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.] > > identity is just as ?easy?. but as I showed long ago (1998 and 2001), what is learned is specific to a space of training examples. there is interpolation in that space, but no reliable extrapolation outside that space. eg if you arrange the problem as training over even numbers represented in binary digits in a standard multi-layer perceptron, the system will not generalize properly to odd numbers. > > nowadays Bengio and others call this the problem of distribution shift, and you would get the same sort of thing with parity (with a slightly different example), because what is learned and described as ?parity? is fairly superficial, example-based rather than fully abstract. > > gary > >>> On Jul 18, 2022, at 05:46, Barak A. Pearlmutter wrote: >>> >>> ?On Mon, 18 Jul 2022 at 08:28, Danko Nikolic wrote: >>> In short, learning mechanisms cannot discover generalized XOR functions with simple connectivity -- only with complex connectivity. This problem results in exponential growth of needed resources as the number of bits in the generalized XOR increases. >> >> Assuming that "generalized XOR" means parity, this must rely on some >> unusual definitions which you should probably state in order to avoid >> confusion. >> >> Parity is a poster boy for an *easy* function to learn, albeit a >> nonlinear one. This is because in the (boolean) Fourier domain its >> spectrum consists of a single nonzero coefficient, and functions that >> are sparse in that domain are very easy to learn. See N. Linial, Y. >> Mansour, and N. Nisan, "Constant depth circuits, Fourier Transform and >> learnability", FOCS 1989, or Mansour, Y. (1994). Learning Boolean >> Functions via the Fourier Transform. Theoretical Advances in Neural >> Computation and Learning, 391?424. doi:10.1007/978-1-4615-2696-4_11 >> >> --Barak Pearlmutter >> > From bei.xiao at gmail.com Mon Jul 18 10:18:56 2022 From: bei.xiao at gmail.com (bei.xiao at gmail.com) Date: Mon, 18 Jul 2022 16:18:56 +0200 Subject: Connectionists: The British Machine Vision Conference (BMVC) call for papers, London, November 21-24 Message-ID: The British Machine Vision Conference (BMVC) is one of the major international conferences on computer vision and related areas. It is organised by the British Machine Vision Association (BMVA). The 33rd BMVC will now be a hybrid event from 21st?24th November 2022. Our local in person meeting will be held at The Kia Oval (Home of Surrey County Cricket Club, https://events.kiaoval.com/). Authors are invited to submit full-length high-quality papers in image processing, computer vision, machine learning and related areas for BMVC 2022. Submitted papers will be refereed on their originality, presentation, empirical results, and quality of evaluation. Accepted papers will be included in the conference proceedings published and DOI-indexed by BMVA. Past proceedings can be found online: here . Please note that BMVC is a single-track meeting with oral and poster presentations. The abstract submission deadline is Friday 22nd July 2022 and the paper submission deadline is Friday 29th July 2022 (both 23:59, GMT). Submission instructions are available on the BMVC 2022 website . Submitted papers should not exceed 9 pages (references are excluded, but appendices are included). Topics include, but are not limited to: - 2D object recognition - 3D computer vision - 3D object recognition - Action and behavior recognition - Adversarial learning, adversarial attack and defense methods - Biometrics, face, gesture, body pose - Computational photography - Datasets and evaluation - Efficient training and inference methods for networks - Explainable AI, fairness, accountability, privacy, transparency and ethics in vision - Image and video retrieval - Image and video synthesis - Image classification - Low-level and physics-based vision - Machine learning architectures and formulations - Medical, biological and cell microscopy - Motion and tracking - Optimization and learning methods - Pose estimation - Representation learning - Scene analysis and understanding - Transfer, low-shot, semi- and un- supervised learning - Video analysis and understanding - Vision + language, vision + other modalities - Vision applications and systems, vision for robotics and autonomous vehicles - ?Brave new ideas? Papers submitted under the ?Brave new ideas? subject area are expected to move away from incremental benchmark gains. Proposed ideas should be radically different from the current strand of research or propose a novel problem. Reviewing process BMVC 2022 - Each paper will be reviewed by at least three reviewers. The primary AC will also provide a meta review, summarising the points that need to be addressed during the rebuttal phase. - The authors will have a period to produce a rebuttal to address the reviewers concerns. Due to the tight schedule, there will be no revision of the papers before the final camera ready submission. - The rebuttal will be handled by two ACs, a primary and a secondary, who will facilitate paper discussion and jointly make the recommendations. Conflicts will be jointly managed by the ACs and Program Chairs that will make the final decisions. Please Note: Due the anticipated volume of papers for BMVC 2022 (based on recent year?s experience) there will be NO extension granted to the submission deadline. In keeping with conferences in the field (e.g. NeurIPS , CVPR ) and to cope with the increasing number of submissions, we ask that all authors be prepared to review papers and make use of a compulsory abstract submission deadline a week before the paper submission deadline. The CMT submission site will ask authors to acknowledge this commitment and failure to engage with the reviewing process might be grounds for rejection. Any queries to the Programme Chairs should be sent to pcs at bmvc2022.org. BMVC Organizers https://bmvc2022.org/people/organisers/ -- Bei Xiao, PhD Associate Professor Computer Science & Center for Behavioral Neuroscience American University, Washington DC Homepage: https://sites.google.com/site/beixiao/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From gary.marcus at nyu.edu Mon Jul 18 12:01:49 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Mon, 18 Jul 2022 09:01:49 -0700 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: <2BB8247E-6320-4968-B3F1-5A9ACAD9C49B@nyu.edu> sure, but a person can learn the idea for n-bits from a few examples with a small number of bits, generalizing it to large values of n. most current systems learn it for a certain number of bits and don?t generalize beyond that number of bits. > On Jul 18, 2022, at 7:17 AM, Barak A. Pearlmutter wrote: > > > > On Mon, 18 Jul 2022 at 14:43, Danko Nikolic > wrote: > > > It is a hard problem to learn for a connectionist network. > > We don't need to invent new terminology, like "inverters problem" or "generalized xor." This is parity. Four (4) bit parity. > > https://en.wikipedia.org/wiki/Parity_function > > Parity is *not* a hard function to learn. Even for a connectionist network. > > It is an interesting function for historic reasons (n-bit parity cannot be loaded by a k-th order perceptron, for k > --Barak Pearlmutter. -------------- next part -------------- An HTML attachment was scrubbed... URL: From georg.martius at tuebingen.mpg.de Mon Jul 18 12:01:09 2022 From: georg.martius at tuebingen.mpg.de (Georg Martius) Date: Mon, 18 Jul 2022 18:01:09 +0200 Subject: Connectionists: Real-Robot Challenge 2022 -- offline reinforcement learning for real robots Message-ID: Dear Colleagues, We are happy to announce the Real-Robot Challenge 2022: It is time to test offline reinforcement learning methods in the real world! Take part in our http://real-robot-challenge.com NeurIPS competition. The task is to use pre-recorded data to learn how to manipulate a cube with the TriFinger robot. You will then be able to run the policy on a real robot at the Max Planck Institute for Intelligent Systems in a remote fashion (as easy as submitting a compute job). Please encourage your colleagues to take part and spread the word. https://twitter.com/robo_challenge/status/1549059071332040709?s=20&t=wDuocEfUp6NtUNBSVTGr8A Many thanks and regards, The RRC-Organizers [Georg Martiu, Nico G?rtler, Felix Widmaier, Cansu Sancaktar, Sebastian Blaes, Pavel Kolev, Stefan Bauer, Manuel W?thrich, Markus Wulfmeier, Martin Riedmiller, Arthur Allshire, Annika Buchholz, Thomas Steinbrenner, Vincent Berenz, Bernhard Sch?lkopf] From Menno.VanZaanen at nwu.ac.za Tue Jul 19 02:39:52 2022 From: Menno.VanZaanen at nwu.ac.za (Menno Van Zaanen) Date: Tue, 19 Jul 2022 06:39:52 +0000 Subject: Connectionists: Job opening: Computational linguist Message-ID: <068cccfa061d42782c66b81102c4f05c288accd0.camel@nwu.ac.za> Computational Linguist Purpose of the position: As a Computational Linguist at the South African Centre for Digital Language Resources (SADiLaR) you will have the opportunity to initiate and lead Human Language Technology and Digital Humanities projects stemming from your own research interests. You will work closely with a team of researchers as part of SADiLaR?s extended network, both on your own and commissioned projects. Dissemination of project results at national and international conferences will be encouraged and supported. This position is crucial for research and development in Human Language Technology and Digital Humanities, fields that form the essence of SADiLaR, which is a national Research Infrastructure supported by the Department of Science and Innovation. Minimum Requirements * PhD in one of the following fields: Computational Linguistics, Natural Language Processing, General Linguistics, Human Language Technology, Digital Humanities, Computer Science, Information Technology, Artificial Intelligence or related fields with a focus on computational aspects of linguistics. * Applicable experience in the use of Python (recommended). Other programming languages used within the computational linguistics domain can also be considered. * Experience as a supervisor/co-supervisor of students or playing a mentorship/supervising role for individuals. * Evidence of peer-reviewed academic publications. * Advanced computer literacy. Other competency requirements * Ability to work independently or as part of a team. * Ability to effectively liaise and communicate with public, students, colleagues, and other stakeholders at various levels and from diverse backgrounds. * Demonstration of language proficiency in order to function optimally in the various multilingual environments of SADiLaR. Recommendations: * Experience with writing research reports. * Ability to lead research projects. * Evidence of acquiring research funding. * Experience with using and/or developing computational tools. * Experience related to research within the domain of Language Technology or Digital Humanities. * Experience in the presentation of research-based results at national and international conferences. * Experience related to teaching within the domain of Language Technology or Digital Humanities. * Strong interest in the advancement of under-resourced South African languages. Responsibilities: * Research in the area of Human Language Technology and Digital Humanities. * Teaching in the area of Human Language Technology and Digital Humanities. * Initiating and leading Human Language Technology and Digital Humanities projects. * Mentorship of researchers in the field of Computational Linguistics and Digital Humanities. ENQUIRIES: Prof Menno van Zaanen, menno.vanzaanen at nwu.ac.za CLOSING DATE: 29 July 2022 COMMENCEMENT OF DUTIES: As soon as possible TO APPLY: https://bit.ly/3yQqcnd -- Prof Menno van Zaanen menno.vanzaanen at nwu.ac.za Professor in Digital Humanities South African Centre for Digital Language Resources https://www.sadilar.org ________________________________ NWU CORONA VIRUS: http://www.nwu.ac.za/coronavirus/ NWU PRIVACY STATEMENT: http://www.nwu.ac.za/it/gov-man/disclaimer.html DISCLAIMER: This e-mail message and attachments thereto are intended solely for the recipient(s) and may contain confidential and privileged information. Any unauthorised review, use, disclosure, or distribution is prohibited. If you have received the e-mail by mistake, please contact the sender or reply e-mail and delete the e-mail and its attachments (where appropriate) from your system. ________________________________ From rloosemore at susaro.com Mon Jul 18 16:28:17 2022 From: rloosemore at susaro.com (Richard Loosemore) Date: Mon, 18 Jul 2022 16:28:17 -0400 Subject: Connectionists: If you believe in your work ... In-Reply-To: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> Message-ID: On 7/17/22 11:52 AM, Grossberg, Stephen wrote: > ... if you believe in your work, and the criticisms of it are not valid, do not give up. ... > ... all criticisms by reviewers are valuable and should be taken into account in your revision. > Even if a reviewer's criticisms are, to your mind, wrong-headed, they represent the > viewpoint of a more-than-usually-qualified reader who has given you the privilege > of taking enough time to read your article. Really? 1) I believe in my work, and the criticisms of it are not valid. I did not give up, and the net result of not giving up was ... nothing. 2) No reviewer who has ever commented on my work has shown the slightest sign that they understood anything in it. 3) Good plumbers are more than usually qualified in their field, and if one of those gave you the privilege of taking enough time to read your article and give nonsensical comments, would you pay any attention to their viewpoint? ** - ** I have spent my career fighting against this system, to no avail. I have watched charlatans bamboozle the crowd with pointless mathematics, and get published. I have watched people use teams of subordinates to pump out streams of worthless papers that inflate their prestige. I have written grant proposals that were exquisitely tuned to the stated goal of the grant, and then watched as the grant money went to people whose proposals had only the faintest imaginable connection to the stated goal of the grant. ** - ** The quoted remarks, above, somehow distilled all of that history and left me shaking with rage at the stupidity. I have been a member of the Connectionists mailing list since the early 1990s, and before that I had been working on neural nets since 1980. No more. Best, Richard -- Richard Loosemore Cornell University ... rpl72 at cornell.edu -------------- next part -------------- An HTML attachment was scrubbed... URL: From barak at pearlmutter.net Mon Jul 18 12:19:09 2022 From: barak at pearlmutter.net (Barak A. Pearlmutter) Date: Mon, 18 Jul 2022 17:19:09 +0100 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <2BB8247E-6320-4968-B3F1-5A9ACAD9C49B@nyu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <2BB8247E-6320-4968-B3F1-5A9ACAD9C49B@nyu.edu> Message-ID: On Mon, 18 Jul 2022 at 17:02, Gary Marcus wrote: > > sure, but a person can learn [n-bit parity] from a few examples with a small number of bits, generalizing it to large values of n. most current systems learn it for a certain number of bits and don?t generalize beyond that number of bits. Really? Because I would not think that induction of a two-state DFA over a two-symbol alphabet woud be beyond the current state of the art. --Barak Pearlmutter. From gary.marcus at nyu.edu Mon Jul 18 12:32:25 2022 From: gary.marcus at nyu.edu (Gary Marcus) Date: Mon, 18 Jul 2022 09:32:25 -0700 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> <2BB8247E-6320-4968-B3F1-5A9ACAD9C49B@nyu.edu> Message-ID: <4426B317-BBBD-4E39-AE75-EDAB4EB80D8B@nyu.edu> sure, but this goes back to Tom?s point about representation; i addressed this sort of thing at length in Chapter 3 of The Algebraic Mind. You can solve this one problem in that way but then give up most of the other virtues of Transformers etc if you build a different network and representational scheme for each problem you encounter. > On Jul 18, 2022, at 9:19 AM, Barak A. Pearlmutter wrote: > > On Mon, 18 Jul 2022 at 17:02, Gary Marcus wrote: >> >> sure, but a person can learn [n-bit parity] from a few examples with a small number of bits, generalizing it to large values of n. most current systems learn it for a certain number of bits and don?t generalize beyond that number of bits. > > Really? Because I would not think that induction of a two-state DFA > over a two-symbol alphabet woud be beyond the current state of the > art. > > --Barak Pearlmutter. From Menno.VanZaanen at nwu.ac.za Tue Jul 19 03:07:52 2022 From: Menno.VanZaanen at nwu.ac.za (Menno Van Zaanen) Date: Tue, 19 Jul 2022 07:07:52 +0000 Subject: Connectionists: 2nd CfP Third workshop on Resources for African Indigenous Language (RAIL) Message-ID: Second call for papers Third workshop on Resources for African Indigenous Language (RAIL) https://bit.ly/rail2022 The South African Centre for Digital Language Resources (SADiLaR) is organising the 3rd RAIL workshop in the field of Resources for African Indigenous Languages. This workshop aims to bring together researchers who are interested in showcasing their research and thereby boosting the field of African indigenous languages. This provides an overview of the current state-of-the-art and emphasizes availability of African indigenous language resources, including both data and tools. Additionally, it will allow for information sharing among researchers interested in African indigenous languages and also start discussions on improving the quality and availability of the resources. Many African indigenous languages currently have no or very limited resources available and, additionally, they are often structurally quite different from more well-resourced languages, requiring the development and use of specialized techniques. By bringing together researchers from different fields (e.g., (computational) linguistics, sociolinguistics, language technology) to discuss the development of language resources for African indigenous languages, we hope to boost research in this field. The RAIL workshop is an interdisciplinary platform for researchers working on resources (data collections, tools, etc.) specifically targeted towards African indigenous languages. It aims to create the conditions for the emergence of a scientific community of practice that focuses on data, as well as tools, specifically designed for or applied to indigenous languages found in Africa. Suggested topics include the following: * Digital representations of linguistic structures * Descriptions of corpora or other data sets of African indigenous languages * Building resources for (under resourced) African indigenous languages * Developing and using African indigenous languages in the digital age * Effectiveness of digital technologies for the development of African indigenous languages * Revealing unknown or unpublished existing resources for African indigenous languages * Developing desired resources for African indigenous languages * Improving quality, availability and accessibility of African indigenous language resources The 3rd RAIL workshop 2022 will be co-located with the 10th Southern African Microlinguistics Workshop ( https://sites.google.com/nwulettere.co.za/samwop-10/home). This will be an in-person event located in Potchefstroom, South Africa. Registration will be free. RAIL 2022 submission requirements: * RAIL asks for full papers from 4 pages to 8 pages (plus more pages for references if needed), which must strictly follow the Journal of the Digital Humanities Association of Southern Africa style guide ( https://upjournals.up.ac.za/index.php/dhasa/libraryFiles/downloadPublic/30 ). * Accepted submissions will be published in JDHASA, the Journal of the Digital Humanities Association of Southern Africa ( https://upjournals.up.ac.za/index.php/dhasa/). * Papers will be double blind peer-reviewed and must be submitted through EasyChair (https://easychair.org/my/conference?conf=rail2022). Important dates Submission deadline: 28 August 2022 Date of notification: 30 September 2022 Camera ready copy deadline: 23 October 2022 RAIL: 30 November 2022, North-West University - Potchefstroom SAMWOP: 1 ? 3 December 2021, North-West University - Potchefstroom Organising Committee Jessica Mabaso Rooweither Mabuya Muzi Matfunjwa Mmasibidi Setaka Menno van Zaanen South African Centre for Digital Language Resources (SADiLaR), South Africa -- Prof Menno van Zaanen menno.vanzaanen at nwu.ac.za Professor in Digital Humanities South African Centre for Digital Language Resources https://www.sadilar.org ________________________________ NWU CORONA VIRUS: http://www.nwu.ac.za/coronavirus/ NWU PRIVACY STATEMENT: http://www.nwu.ac.za/it/gov-man/disclaimer.html DISCLAIMER: This e-mail message and attachments thereto are intended solely for the recipient(s) and may contain confidential and privileged information. Any unauthorised review, use, disclosure, or distribution is prohibited. If you have received the e-mail by mistake, please contact the sender or reply e-mail and delete the e-mail and its attachments (where appropriate) from your system. ________________________________ From ioannakoroni at csd.auth.gr Tue Jul 19 05:02:46 2022 From: ioannakoroni at csd.auth.gr (Ioanna Koroni) Date: Tue, 19 Jul 2022 12:02:46 +0300 Subject: Connectionists: Invitation to join the free hybrid "AI Mellontology (futurology) Symposium 2022", 23rd September 2022 References: <2f7c01d89a69$90028f00$b007ad00$@csd.auth.gr> <005001d89a88$9aa9aaf0$cffd00d0$@csd.auth.gr> Message-ID: <319401d89b4e$50d96b10$f28c4130$@csd.auth.gr> Dear AI researcher, scientist, engineer, student, enthusiast, you are welcomed to attend the free hybrid 'AI Mellontology Symposium' on 23/09/2022 to debate futurist topics on AI research and social/industrial impact. Its program, registration and participation links can be found in: http://icarus.csd.auth.gr/ai-mellontology-e-symposium-2022/ It is co-located with International AI Doctoral Academy (AIDA) General Assembly. More details on the panel topic descriptions can be found in the www page. Hybrid (local and remote) participation is foreseen. Remote participation is free. There is a very small registration fee for local participation or for remote participation (if a participation certificate is desired) for people outside the AI4media partners. Zoom link for participation: https://authgr.zoom.us/j/97284234003 Horizon2020 AI flagship R&D projects: AI4Media and ELISE , together with the International AI Doctoral Academy (AIDA) joined forces in sponsoring this event. There will be 6 AI science/technology/society/industry sessions (4 in the morning and 2 in the afternoon). Participants will debate the ?AI future ?, namely hot AI research issues of any form on any scientific, societal, industrial/economic AI aspect. Session panelists will make their short statements during each panel to initiate discussion by the audience. Discussion will be informal, hopefully lively and hot aiming to debate contrasting views (thesis-antithesis) on each issue. You are welcome to a) see panelist and other views and voice your own views on any panel topic, by using the ?Voice your views? area, b) propose your own AI grand challenge and c) actively participate in each panel discussion. The more innovative, futuristic, even provocative your views are, the better. The discussion will be conducted though in a scientific way and in good faith. The entire discussion and contributions will be compiled later on in a public ?AI research white paper? that will be used to guide R&D anywhere in the world, notably in the framework of AIDA and Horizon2020 ICT48 R&D projects. The end of September is a very nice period to visit Thessaloniki and its surroundings (cultural sightseeing, swimming, grape harvesting). An optional cultural trip is foreseen on the 24th of September 2022. For any questions, you can contact the e-symposium secretary Mrs. Ioanna Koroni koroniioanna at csd.auth.gr (for organization matters) or Prof. Ioannis Pitas pitas at csd.auth.gr (for scientific matters). On behalf of the Organizing Committee Prof. Ioannis Pitas AI Mellontology Symposium chair AI Mellontology Symposium Program 23/09/2022 Times are in CEST 8:00-8:05 CEST Opening (Prof. Ioannis Pitas) Morning sessions 1. 8:05-9:00 ?Latent brain?, Organizer: Dr. Patrick van der Smagt (Volkswagen) Panelists: Dr. Patrick van der Smagt, Prof. Henrik J?rntell, Prof. Daniel Gauthier 2. 9:00-10:00 Collaborative AI, Organizer: Prof. N. Cesa Bianchi (U Milan), Prof. John Shawe-Taylor (UCL) Panelists: Prof. N. Cesa Bianchi (U Milan), Prof. John Shawe-Taylor (UCL), Prof. J. Crowley (INRIA), Prof. A. Oulasvirta (Aaalto U) 3. 10:30-11:30 ?Human ? AI Symbiosis?, Organizer: Dr. N. Vretos (CERTH/ITI) Panelists: Prof. Federico ?lvarez Garcia (UPM), Dr. T. Giannakopoulos (NCSR Demokritos) 4. 11:30-12:30 ?AI studies: AI PhD Excellence. AI Curriculum. AI Science and Engineering?? Organizer: Prof. Ioannis Pitas (AUTH), Panel: Prof. A. Geiger (U Tuebingen), Prof. Ioannis Pitas (AUTH), Prof. B. O?Sullivan (UCC), Prof. B. Schoelkopf (to be confirmed). Afternoon sessions 1. 14:00-15:00 ?AI-based recommender systems: social and legal perspectives?, Organizer: L. Dutkiewicz (CiTiP, KUL) Panelists: A. Schj?tt Hansen (UvA), L. Dutkiewicz (CiTiP, KUL) 5. 15:00-16:00 How AI meets human in media houses? (Recommendation systems and Media) Organizer: Dr. P. Kord?k (CVUT) Panelists: Dr. P. Kord?k (CVUT), Dr. T. Lan?ov? (CVUT). 16:30-18:30 AIDA General Assembly Saturday 24/9/2022 (optional): 9:00-15:00 EEST Visit of Royal Macedonian Tombs and Museum (Vergina) and a winery in Naoussa, Central Macedonia, Greece (bus transfer). -------------- next part -------------- An HTML attachment was scrubbed... URL: From dftschool at ini.rub.de Tue Jul 19 05:02:13 2022 From: dftschool at ini.rub.de (DFT Summer School) Date: Tue, 19 Jul 2022 11:02:13 +0200 Subject: Connectionists: Neuronal Dynamics for Embodied Cognition - Virtual Summer School 2022 - Deadline for Workshop closes in 2 weeks Message-ID: <9edd40c7-e690-5ae8-23bd-1c8ff3f340dd@ini.rub.de> Please forward this advertisement to whoever you think might be interested. This year our summer school "Neural Dynamics for Embodied Cognition" will take place in virtual form from the 15th to the 20th of August, 2022. It will consist of two parts: A live-lecture series and a hands-on workshop. Anyone can attend the live-lecture series, but the workshop has a capacity limit, so please apply via our webpage: https://dynamicfieldtheory.org/events/summer_school_2022/ The deadline for workshop applications is *Jul 31, 2022!* Participation in any part of the school is free of charge. Best regards, Raul Grieben and Jan Tek?lve - - - Virtual DFT School 2022 This year our summer school "Neural Dynamics for Embodied Cognition" will take place in virtual form from the 15th to the 20th of August, 2022. Neuronal dynamics provide a powerful theoretical language for the design and modeling of embodied and situated cognitive systems. This school provides a hands-on and practical introduction to neuronal dynamics ideas and enables participants to become productive within this framework. The school is aimed at advanced undergraduate or graduate students, postdocs and faculty members in embodied cognition, cognitive science, and robotics. Topics addressed include neural dynamics, attractor dynamics and instabilities, dynamic field theory, neuronal representations, artificial perception, simple forms of cognition including detection and selection decisions, memory formation, learning, and grounding relational concepts. This virtual edition of our summer school will consist of two parts: A live-lecture series and a hands-on workshop. The lecture series will be held as a video conference and provides a step-by-step introduction to Dynamic Field Theory. The two-and-a-half-day project workshop gives students the opportunity to put to use the newly acquired skills in a concrete hands-on modeling project. Students solve the task in our open-source simulation environment under the guidance of a personal tutor. This year's lectures will be open for everyone, while the one-on-one tutoring limits the number of participants who can take part in the workshop. We also encourage workshop applications by small groups of participants, maybe two or three colleagues who will work together locally on the same project and may share a tutor. Although this format will not retain the appeal of meeting other students in person, we will make this year's experience as interactive as possible. The lecture series will be held from the 15th to the 20th of August and the workshop takes place from the 18th to the 20th of August. Lectures will take place from 3 to 6 p.m. (CET) on each day and personal tutoring will be available on each workshop day. Participation in any part of the school is free of charge. To apply for the lecture series and/or the workshop, please visit our webpage: https://dynamicfieldtheory.org/events/summer_school_2022/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From stdm at zhaw.ch Tue Jul 19 07:25:42 2022 From: stdm at zhaw.ch (Stadelmann Thilo (stdm)) Date: Tue, 19 Jul 2022 11:25:42 +0000 Subject: Connectionists: PhD student for robust deep learning in industrial applications, fully funded, Winterthur/Switzerland Message-ID: I am looking for a new PhD student as of September 2022, working with me on robust applications of deep learning (computer vision, transfer learning, trustworthy AI). Recent MSc graduates from universities of applied sciences are specifically welcome. Details: https://ohws.prospective.ch/public/v1/jobs/d6f0f9bf-539b-427e-92cb-a4c870d4babd?utm_source=jobabo&utm_medium=email More Background: https://www.zhaw.ch/en/about-us/news/news-releases/news-detail/event-news/3-neue-forschungsprojekte-zielen-auf-den-einsatz-vertrauenswuerdiger-ki/ Best, Thilo ------------------------------------------------------------------------------------------------------------- ZHAW School of Engineering Prof. Thilo Stadelmann, Dr. rer. nat., FECLT, SMIEEE Director of Centre for Artificial Intelligence Head of Computer Vision, Perception and Cognition Group Phone: +41 58 934 72 08, fax: +41 58 935 72 08 Email: thilo.stadelmann at zhaw.ch Web: www.zhaw.ch/cai, http://stdm.github.io Office: TN 03.55, Technikumstrasse 71, CH-8400 Winterthur Postal address: ZHAW School of Engineering, Thilo Stadelmann, Postfach, CH-8401 Winterthur, Switzerland Site plan: https://www.zhaw.ch/storage/shared/hochschule/lageplaene/lageplan-winterthur-technikumstrasse.pdf -------------- next part -------------- An HTML attachment was scrubbed... URL: From rloosemore at susaro.com Tue Jul 19 13:39:40 2022 From: rloosemore at susaro.com (Richard Loosemore) Date: Tue, 19 Jul 2022 13:39:40 -0400 Subject: Connectionists: [EXT] If you believe in your work ... In-Reply-To: References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> Message-ID: <9ea52c4c-2b66-e8bd-b494-5dc4ab7ebc8d@susaro.com> Jean-Marc, The problem is systemic; it goes down into the very roots of modern "science". The only solution that $20m could buy would be: 1) An institute run by someone with ethical principles, who would use the money to attract further funding until it could actually take on board researchers with creative ideas and ethical principles, and then free them from the yoke of publish-crap-in-quantity-or-perish. 2) An AI/Cognitive system development tool that would allow people to build and explore complex cognitive systems without being shackled to one particular architecture (like deep learning and its many descendents). A propos of (2) that is one thing I proposed in a (rejected) grant proposal. It would have cost $6.4m. Best, Richard -- Richard Loosemore Cornell University ... rpl72 at cornell.edu On 7/19/22 11:31 AM, Fellous, Jean-Marc - (fellous) wrote: > Assuming there are funders on the list, and funding-related people, > including program officers (and believe or not, there are!): if you > had $20M to invest in the sort of things we do on this list: how would > we make things better? Can we brainstorm an alternative system that > allows for innovating publications and effective funding? > > Jean-Marc > ------------------------------------------------------------------------ > *From:* Connectionists > on behalf of Richard Loosemore > *Sent:* Monday, July 18, 2022 1:28 PM > *To:* connectionists at mailman.srv.cs.cmu.edu > > *Subject:* [EXT]Connectionists: If you believe in your work ... > > *External Email* > > > On 7/17/22 11:52 AM, Grossberg, Stephen wrote: > > > ... if you believe in your work, and the criticisms of it are not > valid, do not give up. ... > > > ... all criticisms by reviewers are valuable and should be taken > into account in your revision. > > > Even if a reviewer's criticisms are, to your mind, wrong-headed, > they represent the > > viewpoint of a more-than-usually-qualified reader who has given you > the privilege > > of taking enough time to read your article. > > Really? > > 1) I believe in my work, and the criticisms of it are not valid.? I > did not give up, and the net result of not giving up was ... nothing. > > 2) No reviewer who has ever commented on my work has shown the > slightest sign that they understood anything in it. > > 3) Good plumbers are more than usually qualified in their field, and > if one of those gave you the privilege of taking enough time to read > your article and give nonsensical comments, would you pay any > attention to their viewpoint? > > ** - ** > > I have spent my career fighting against this system, to no avail. > > I have watched charlatans bamboozle the crowd with pointless > mathematics, and get published. > > I have watched people use teams of subordinates to pump out streams of > worthless papers that inflate their prestige. > > I have written grant proposals that were exquisitely tuned to the > stated goal of the grant, and then watched as the grant money went to > people whose proposals had only the faintest imaginable connection to > the stated goal of the grant. > > ** - ** > > The quoted remarks, above, somehow distilled all of that history and > left me shaking with rage at the stupidity. > > I have been a member of the Connectionists mailing list since the > early 1990s, and before that I had been working on neural nets since 1980. > > No more. > > > Best, > > Richard > > -- > > Richard Loosemore > > Cornell University > > ... > > rpl72 at cornell.edu > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From fellous at arizona.edu Tue Jul 19 11:31:40 2022 From: fellous at arizona.edu (Fellous, Jean-Marc - (fellous)) Date: Tue, 19 Jul 2022 15:31:40 +0000 Subject: Connectionists: [EXT] If you believe in your work ... In-Reply-To: References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> Message-ID: Assuming there are funders on the list, and funding-related people, including program officers (and believe or not, there are!): if you had $20M to invest in the sort of things we do on this list: how would we make things better? Can we brainstorm an alternative system that allows for innovating publications and effective funding? Jean-Marc ________________________________ From: Connectionists on behalf of Richard Loosemore Sent: Monday, July 18, 2022 1:28 PM To: connectionists at mailman.srv.cs.cmu.edu Subject: [EXT]Connectionists: If you believe in your work ... External Email On 7/17/22 11:52 AM, Grossberg, Stephen wrote: > ... if you believe in your work, and the criticisms of it are not valid, do not give up. ... > ... all criticisms by reviewers are valuable and should be taken into account in your revision. > Even if a reviewer's criticisms are, to your mind, wrong-headed, they represent the > viewpoint of a more-than-usually-qualified reader who has given you the privilege > of taking enough time to read your article. Really? 1) I believe in my work, and the criticisms of it are not valid. I did not give up, and the net result of not giving up was ... nothing. 2) No reviewer who has ever commented on my work has shown the slightest sign that they understood anything in it. 3) Good plumbers are more than usually qualified in their field, and if one of those gave you the privilege of taking enough time to read your article and give nonsensical comments, would you pay any attention to their viewpoint? ** - ** I have spent my career fighting against this system, to no avail. I have watched charlatans bamboozle the crowd with pointless mathematics, and get published. I have watched people use teams of subordinates to pump out streams of worthless papers that inflate their prestige. I have written grant proposals that were exquisitely tuned to the stated goal of the grant, and then watched as the grant money went to people whose proposals had only the faintest imaginable connection to the stated goal of the grant. ** - ** The quoted remarks, above, somehow distilled all of that history and left me shaking with rage at the stupidity. I have been a member of the Connectionists mailing list since the early 1990s, and before that I had been working on neural nets since 1980. No more. Best, Richard -- Richard Loosemore Cornell University ... rpl72 at cornell.edu -------------- next part -------------- An HTML attachment was scrubbed... URL: From danko.nikolic at gmail.com Tue Jul 19 17:08:36 2022 From: danko.nikolic at gmail.com (Danko Nikolic) Date: Tue, 19 Jul 2022 23:08:36 +0200 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 Barak, Thank you for the pointers. I have to read the papers. I need to understand then why, if parity is so easy to learn, my deep learning models had such a hard time that it led to an exponential growth in the number of needed parameters with each additional bit added to the input. Strange. I will report back. Best Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Mon, Jul 18, 2022 at 1:12 PM Barak A. Pearlmutter wrote: > On Mon, 18 Jul 2022 at 08:28, Danko Nikolic > wrote: > > In short, learning mechanisms cannot discover generalized XOR functions > with simple connectivity -- only with complex connectivity. This problem > results in exponential growth of needed resources as the number of bits in > the generalized XOR increases. > > Assuming that "generalized XOR" means parity, this must rely on some > unusual definitions which you should probably state in order to avoid > confusion. > > Parity is a poster boy for an *easy* function to learn, albeit a > nonlinear one. This is because in the (boolean) Fourier domain its > spectrum consists of a single nonzero coefficient, and functions that > are sparse in that domain are very easy to learn. See N. Linial, Y. > Mansour, and N. Nisan, "Constant depth circuits, Fourier Transform and > learnability", FOCS 1989, or Mansour, Y. (1994). Learning Boolean > Functions via the Fourier Transform. Theoretical Advances in Neural > Computation and Learning, 391?424. doi:10.1007/978-1-4615-2696-4_11 > > --Barak Pearlmutter > -------------- next part -------------- An HTML attachment was scrubbed... URL: From bhccsymposium2022 at gmail.com Wed Jul 20 04:36:45 2022 From: bhccsymposium2022 at gmail.com (BHCC BHCC) Date: Wed, 20 Jul 2022 10:36:45 +0200 Subject: Connectionists: [CFP] BHCC 2022 - Online - Call For Papers Message-ID: **Apologies for cross-posting** *BHCC 2022 - Call For Papers* The Fourth Symposium On Biases In Human Computing And Crowdsourcing ( https://bhcc2022.org/*) will take place ONLINE, *on *12 - 14 October 2022.* *(FREE OF CHARGE EVENT)* The goal of this symposium is to analyze both existing human biases in hybrid systems, and methods to manage bias via crowdsourcing and human computation. We will discuss different types of biases, measures, and methods to track bias, as well as methodologies to prevent and solve bias. We will provide a framework for discussion among scholars, practitioners, and other interested parties, including industry, crowd workers, requesters, and crowdsourcing platform managers. We expect contributions combining ideas from different disciplines, including computer science, psychology, economics, and social sciences. *Topics Of Interest* Topics of interest include (but are not limited to): - Biases in Human Computation and Crowdsourcing - Human sampling bias - Effect of cultural, gender, and ethnic biases - Effect of human in the loop training and past experiences - Effect of human expertise vs interest - Bias in experts vs. bias in crowdsourcing - Bias in outsourcing vs bias in crowdsourcing - Bias in task selection - Task assignment/recommendation for reducing bias - Effect of human engagement on the bias - Responsibility and ethics in human computation and bias management - Preventing bias in crowdsourcing and human computation - Creating awareness of cognitive biases among human agents - Measuring and addressing ambiguities and biases in human annotation - Human factors in AI - Using Human Computation and Crowdsourcing for Bias Understanding and Management - Biases in Human-in-the-loop systems - Identifying new types of cognitive bias in data or content - Measuring bias in data or content - Removing bias in data or content - Dealing with algorithmic bias - Fake news detection - Diversification of sources by means - Provenance and traceability - Long-term crowd engagement *Submission Guidelines* We welcome the submission of research papers and abstracts which describe original work that has not been submitted or is currently under review, has not been previously published nor accepted for publication elsewhere, in any other journal or conference. Submissions of the research papers must be in English, in PDF format, and be in the current CEUR-WS single-column conference format. We will follow CEUR-WS guidelines, meet their preconditions, and expect to get the proceedings published. However, note that there is no guarantee that our volume will get published at CEUR-WS. We welcome the submission of the following types of contributions: - Full papers should be at most 10 pages in length (including figures, tables, appendices, and references); - Short papers should be at most 5 pages in length (including figures, tables, appendices, and references); - Abstracts should be at most 1 page in length (including figures, tables, appendices, and references), should contain just a title and the abstract, and should detail demos or relevant work or ideas which are under development. They can not contain references. We implement a double-blind review process. Submissions must be anonymous and the submission must be made via EasyChair: https://easychair.org/conferences/?conf=bhcc2022. *Important Dates* - Full, Short, and Abstract papers due: 1 September 2022 AoE (firm deadline) - Notifications: 10 September 2022 - Conference: 12, 13, and 14 October 2022 *Chairs* - General Chair: Lorenzo Bracciale (University of Rome "Tor Vergata") - General Chair: Kevin Roitero (University of Udine) - Proceedings and Website Chair: Michael Soprano (University of Udine) - Social Media Chair: David La Barbera (University of Udine) - Sponsorship Chair: Danula Hettiachchi (RMIT University) *Contact* bhccsymposium2022 at gmail.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From achler at gmail.com Wed Jul 20 04:14:32 2022 From: achler at gmail.com (Tsvi Achler) Date: Wed, 20 Jul 2022 01:14:32 -0700 Subject: Connectionists: [EXT] If you believe in your work ... In-Reply-To: References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> Message-ID: I think the recent trend (starting 70s& 80s) of having articles accepted for publications based on peer review is a problem because it adds more politics to every small decision. Academic governance, which started in the middle ages, is not democratic and is actually rather dysfunctional: think tenure type politics. The less of it the better. It may be better to go back to just having a single editor decide if an article is worthy. That way at least more novel ideas can be presented. Of course the other problem is paid journals which have low costs (in these days of the internet) and free labor and skim lots of money for unclear reasons. Also realize most innovative researchers are on a shoestring budget because of politics. It is an awful combination: journals love the peer review system because it lets them act like they are bringing value by adding more politics and obfuscating the skimming. -Tsvi On Wed, Jul 20, 2022 at 12:01 AM Fellous, Jean-Marc - (fellous) < fellous at arizona.edu> wrote: > Assuming there are funders on the list, and funding-related people, > including program officers (and believe or not, there are!): if you had > $20M to invest in the sort of things we do on this list: how would we make > things better? Can we brainstorm an alternative system that allows for > innovating publications and effective funding? > > Jean-Marc > ------------------------------ > *From:* Connectionists on > behalf of Richard Loosemore > *Sent:* Monday, July 18, 2022 1:28 PM > *To:* connectionists at mailman.srv.cs.cmu.edu < > connectionists at mailman.srv.cs.cmu.edu> > *Subject:* [EXT]Connectionists: If you believe in your work ... > > > *External Email* > > On 7/17/22 11:52 AM, Grossberg, Stephen wrote: > > > ... if you believe in your work, and the criticisms of it are not valid, > do not give up. ... > > > ... all criticisms by reviewers are valuable and should be taken into > account in your revision. > > Even if a reviewer's criticisms are, to your mind, wrong-headed, they > represent the > > viewpoint of a more-than-usually-qualified reader who has given you the > privilege > > of taking enough time to read your article. > > Really? > > 1) I believe in my work, and the criticisms of it are not valid. I did > not give up, and the net result of not giving up was ... nothing. > > 2) No reviewer who has ever commented on my work has shown the slightest > sign that they understood anything in it. > > 3) Good plumbers are more than usually qualified in their field, and if > one of those gave you the privilege of taking enough time to read your > article and give nonsensical comments, would you pay any attention to their > viewpoint? > > ** - ** > > I have spent my career fighting against this system, to no avail. > I have watched charlatans bamboozle the crowd with pointless mathematics, > and get published. > > I have watched people use teams of subordinates to pump out streams of > worthless papers that inflate their prestige. > > I have written grant proposals that were exquisitely tuned to the stated > goal of the grant, and then watched as the grant money went to people whose > proposals had only the faintest imaginable connection to the stated goal of > the grant. > > ** - ** > > The quoted remarks, above, somehow distilled all of that history and left > me shaking with rage at the stupidity. > > I have been a member of the Connectionists mailing list since the early > 1990s, and before that I had been working on neural nets since 1980. > > No more. > > > Best, > > > > Richard > > -- > > Richard Loosemore > > Cornell University > > ... > > rpl72 at cornell.edu > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From achler at gmail.com Wed Jul 20 03:51:59 2022 From: achler at gmail.com (Tsvi Achler) Date: Wed, 20 Jul 2022 00:51:59 -0700 Subject: Connectionists: Stephen Hanson in conversation with Geoff Hinton In-Reply-To: <4426B317-BBBD-4E39-AE75-EDAB4EB80D8B@nyu.edu> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <2BB8247E-6320-4968-B3F1-5A9ACAD9C49B@nyu.edu> <4426B317-BBBD-4E39-AE75-EDAB4EB80D8B@nyu.edu> Message-ID: I believe one way to address this problem that learning with different types of features may give up other virtues (of Transformers etc) is to scale better by: 1) reducing the cost of learning so the same information over more feature types can be learned at once. 2) new features/learning to be able to add modularly to the model (eg avoid catastrophic forgetting) 3) Not making a decision of what features are most important ahead of time 4) taking a shotgun approach and learning with as much features as possible These goals can be better achieved if the networks learning (or at least top layer learning) does not require iid (independent and identically distributed) rehearsal and is super scalable. Feedforward methods (e.g. current neural networks) have issues with 1 & 2 while most other methods such as Bayesian Networks have problems with 1 & scalability. My 2c, -Tsvi On Tue, Jul 19, 2022 at 12:09 AM Gary Marcus wrote: > sure, but this goes back to Tom?s point about representation; i addressed > this sort of thing at length in Chapter 3 of The Algebraic Mind. > > You can solve this one problem in that way but then give up most of the > other virtues of Transformers etc if you build a different network and > representational scheme for each problem you encounter. > > > On Jul 18, 2022, at 9:19 AM, Barak A. Pearlmutter > wrote: > > > > On Mon, 18 Jul 2022 at 17:02, Gary Marcus wrote: > >> > >> sure, but a person can learn [n-bit parity] from a few examples with > a small number of bits, generalizing it to large values of n. most current > systems learn it for a certain number of bits and don?t generalize beyond > that number of bits. > > > > Really? Because I would not think that induction of a two-state DFA > > over a two-symbol alphabet woud be beyond the current state of the > > art. > > > > --Barak Pearlmutter. > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From vaina at bu.edu Wed Jul 20 07:16:11 2022 From: vaina at bu.edu (Vaina, Lucia M) Date: Wed, 20 Jul 2022 11:16:11 +0000 Subject: Connectionists: [EXT] If you believe in your work ... In-Reply-To: References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> Message-ID: Good points Tsvi. Lucia Get Outlook for iOS ________________________________ From: Connectionists on behalf of Tsvi Achler Sent: Wednesday, July 20, 2022 4:14:32 AM To: Fellous, Jean-Marc - (fellous) Cc: connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: [EXT] If you believe in your work ... I think the recent trend (starting 70s& 80s) of having articles accepted for publications based on peer review is a problem because it adds more politics to every small decision. Academic governance, which started in the middle ages, is not democratic and is actually rather dysfunctional: think tenure type politics. The less of it the better. It may be better to go back to just having a single editor decide if an article is worthy. That way at least more novel ideas can be presented. Of course the other problem is paid journals which have low costs (in these days of the internet) and free labor and skim lots of money for unclear reasons. Also realize most innovative researchers are on a shoestring budget because of politics. It is an awful combination: journals love the peer review system because it lets them act like they are bringing value by adding more politics and obfuscating the skimming. -Tsvi On Wed, Jul 20, 2022 at 12:01 AM Fellous, Jean-Marc - (fellous) > wrote: Assuming there are funders on the list, and funding-related people, including program officers (and believe or not, there are!): if you had $20M to invest in the sort of things we do on this list: how would we make things better? Can we brainstorm an alternative system that allows for innovating publications and effective funding? Jean-Marc ________________________________ From: Connectionists > on behalf of Richard Loosemore > Sent: Monday, July 18, 2022 1:28 PM To: connectionists at mailman.srv.cs.cmu.edu > Subject: [EXT]Connectionists: If you believe in your work ... External Email On 7/17/22 11:52 AM, Grossberg, Stephen wrote: > ... if you believe in your work, and the criticisms of it are not valid, do not give up. ... > ... all criticisms by reviewers are valuable and should be taken into account in your revision. > Even if a reviewer's criticisms are, to your mind, wrong-headed, they represent the > viewpoint of a more-than-usually-qualified reader who has given you the privilege > of taking enough time to read your article. Really? 1) I believe in my work, and the criticisms of it are not valid. I did not give up, and the net result of not giving up was ... nothing. 2) No reviewer who has ever commented on my work has shown the slightest sign that they understood anything in it. 3) Good plumbers are more than usually qualified in their field, and if one of those gave you the privilege of taking enough time to read your article and give nonsensical comments, would you pay any attention to their viewpoint? ** - ** I have spent my career fighting against this system, to no avail. I have watched charlatans bamboozle the crowd with pointless mathematics, and get published. I have watched people use teams of subordinates to pump out streams of worthless papers that inflate their prestige. I have written grant proposals that were exquisitely tuned to the stated goal of the grant, and then watched as the grant money went to people whose proposals had only the faintest imaginable connection to the stated goal of the grant. ** - ** The quoted remarks, above, somehow distilled all of that history and left me shaking with rage at the stupidity. I have been a member of the Connectionists mailing list since the early 1990s, and before that I had been working on neural nets since 1980. No more. Best, Richard -- Richard Loosemore Cornell University ... rpl72 at cornell.edu -------------- next part -------------- An HTML attachment was scrubbed... URL: From er.anubajaj at gmail.com Wed Jul 20 07:21:26 2022 From: er.anubajaj at gmail.com (Anu Bajaj) Date: Wed, 20 Jul 2022 16:51:26 +0530 Subject: Connectionists: 2nd CFP: 22nd International Conference on Hybrid Intelligent Systems (HIS'22) - Online - Springer Publication Message-ID: ** Second Call for Papers - please circulate this CFP to your colleagues and networks ** -- The 22nd International Conference on Hybrid Intelligent Systems (HIS'22) -- http://www.mirlabs.net/his22 http://www.mirlabs.org/his22 On the World Wide Web December 13-15, 2022 Proceedings of HIS'22 will be published with Springer Verlag in their Lecture Notes in Networks and Systems (LNNS) series. ( https://www.springer.com/series/15179) (Approval Pending) Proceedings of HIS'21: https://link.springer.com/book/10.1007/978-3-030-96305-7 Indexed by: SCOPUS, INSPEC, WTI Frankfurt eG, zbMATH, SCImago, DBLP, EI Compendex, Japanese Science and Technology Agency (JST), SpringerLink History of HIS series: http://www.mirlabs.net/his22/previous.php **Important Dates** --------------------- Paper submission due: September 30, 2022 Notification of paper acceptance: October 31, 2022 Registration and Final manuscript due: November 15, 2022 Conference: December 13-15, 2022 **About HIS 2022** ------------------ Hybridization of intelligent systems is a promising research field of modern artificial/computational intelligence concerned with the development of the next generations of intelligent systems. A fundamental stimulus to Hybrid Intelligent Systems (HIS) investigations is the awareness in the academic communities about the necessity of the combined approaches to solving the remaining challenging problems in artificial/computational intelligence. Hybrid Intelligent Systems are recently getting popular due to their capabilities in handling several real-world complexities involving imprecision, uncertainty, and vagueness. HIS'22 built on the success of the previous HIS conferences. HIS'22 is the 22nd International conference that brings together researchers, developers, practitioners, and users of soft computing, computational intelligence, agents, logic programming, and several other intelligent computing techniques. It aims to increase the awareness of the research community about the broad spectrum of hybrid approaches. It brings together worldwide AI researchers to present their cutting-edge results, discuss the current trends in HIS research, and develop a collective vision of future opportunities. Thereby helping the researchers and practitioners establish international collaborative opportunities, and as a result, advancing the state-of-the-art of the field. **Topics (not limited to)** --------------------------- Interactions between neural networks and fuzzy inference systems. Artificial neural network optimization using global optimization techniques. Fuzzy clustering algorithms and optimization techniques. Fuzzy inference system optimization using global optimization algorithms. Hybrid computing using neural networks - fuzzy systems - evolutionary algorithms. Hybrid optimization techniques (evolutionary algorithms, simulated annealing, tabu search, GRASP etc.). Hybrid of soft computing and statistical learning techniques. Models using inductive logic programming, logic synthesis, grammatical inference, case-based reasoning etc. Autonomic computing. Hybridizatiion with novel computing paradigms: Qantum computing, DNA computing, membrane computing etc. Hybrid Intelligent Systems: Applications **Submission Guidelines** ------------------------- Submission of paper should be made through the submission page from the conference web page. Please refer to the conference website for guidelines to prepare your manuscript. Paper format templates: https://www.springer.com/de/authors-editors/book-authors-editors/manuscript-preparation/5636#c3324 HIS?22 Submission Link: https://easychair.org/conferences/?conf=his20220 **Plenary Talks** ---------------------------------------- You are welcome to attend 11 Keynote Talks offered by world-renowned professors and industry leaders. The detailed information is available on conference website. Speaker 1: Catarina Silva, University of Coimbra, Portugal Title: Interpretability and Explainability in Intelligent Systems Speaker 2: Chuan-Kang Ting, National Tsing Hua University, Taiwan Title: TBA Speaker 3: Joanna Kolodziej, Cracow University of Technology, Poland Title: TBA Speaker 4: Kaisa Miettinen, Multiobjective Optimization Group, Faculty of Information Technology, University of Jyvaskyla, Finland Title: Some Perspectives to Interactive Evolutionary Multiobjective Optimization Methods. Speaker 5: Kaspar Riesen, Institute of Computer Science, University of Bern, Switzerland University of Applied Sciences and Arts, Switzerland Title: Four Decades of Structural Pattern Recognition ? An Overview of the Three Major Epochs Speaker 6: Katherine MALAN, Department of Decision Sciences, University of South Africa Title: Landscape analysis of optimisation and machine learning search spaces. Speaker 7: Maki Sakamoto, The University of Electro-Communications, Tokyo, Japan Title: Computer Vision for Expressing Texture Using Sound-Symbolic Words Speaker 8: M?rio Antunes, Polytechnic Institute of Leiria, Portugal Title: Cybersecurity: the road ahead Speaker 9: Patricia MELIN, Tijuana Institute of Technology, Tijuana, Mexico Title: Hybrid Intelligent Systems based on Neural Networks, Fuzzy Logic and Bioinspired Optimization Algorithms and their application to Pattern Recognition. Speaker 10: Ren? NATOWICZ, ESIEE-Paris - Universit? Gustave Eiffel, France Title: Machine Learning in Graphs: Where Are We So Far? Speaker 11: Yifei Pu, College of Computer Science, Sichuan University, China Title : Analog Circuit Implementation of Fractional-Order Memristor: Arbitrary-Order Lattice Scaling Fracmemristor. **HIS 2022 Organization** ------------------------- General Chairs Ajith Abraham, Machine Intelligence Research Labs, USA Tzung-Pei Hong, National University of Kaohsiung, Taiwan Art?ras Kaklauskas, Vilnius Gediminas Technical University, Lithuania Program Chairs Ketan Kotecha, Symbiosis International University, India Ganeshsree Selvachandran, UCSI University, Malaysia Publication Chairs Niketa Gandhi, Machine Intelligence Research Labs, USA Kun Ma, University of Jinan, China Special Session Chair Gabriella Casalino, University of Bari, Italy Publicity Chairs Pooja Manghirmalani Mishra, Machine Intelligence Research Labs, India Anu Bajaj, Machine Intelligence Research Labs, USA If you would like to propose a special session, please email Dr. Gabriella Casalino ** For all other Technical Contact ** ------------------------------------- Dr. Ajith Abraham Email: ajith.abraham at ieee.org -------------- next part -------------- An HTML attachment was scrubbed... URL: From frothga at sandia.gov Wed Jul 20 09:39:10 2022 From: frothga at sandia.gov (Rothganger, Fredrick) Date: Wed, 20 Jul 2022 13:39:10 +0000 Subject: Connectionists: Approaches to Publication In-Reply-To: References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> Message-ID: There are some other approaches to publication we could consider. 1) Something like Wikipedia, but which allows original research like Scholarpedia, where people collaboratively build up a knowledge structure about a topic. In this case, credit assignment can come for the revision log and the significance of the edits you make. 2) Something like arXiv, where you make article available without peer-review. In this case, the strength of an article is determined by how many people link to it. Articles can be made visible to search engines rather than a specific editor or group of peer reviewers. Perhaps criticism could be handled via a talk-back section, which would be equivalent to a rebuttal section in some journals. 3) Personal/lab research blogs. This would be a good place to post software and data, and to talk about negative results. All these are enabled by fairly recent (last few decades) technology. Our culture is only slowly catching up. Part of the problem is that we are in constant competition for funding, so credit assignment is a big deal. I dream of a world where competition for basic survival is no longer an issue, where people practice science, or art, or serve in other professions, merely because they love to. That world is technologically possible, but I don't see it happening in my lifetime, due to cultural issues. [http://scholarpedia.org/w/images/thumb/d/d5/SET_book_cover.jpg/90px-SET_book_cover.jpg1zB] Scholarpedia Experimental determination of the CKM matrix. S?bastien Descotes-Genon et al. (2022), Scholarpedia, 17(1):54385. In the development of particle physics describing matter at the smallest distances, it has proved possible not only to understand the structure of the proton and the neutron in nuclei as made of two types... www.scholarpedia.org ? ? ? ________________________________ From: Connectionists on behalf of Tsvi Achler Sent: Wednesday, July 20, 2022 2:14 AM To: Fellous, Jean-Marc - (fellous) Cc: connectionists at mailman.srv.cs.cmu.edu Subject: [EXTERNAL] Re: Connectionists: [EXT] If you believe in your work ... I think the recent trend (starting 70s& 80s) of having articles accepted for publications based on peer review is a problem because it adds more politics to every small decision. Academic governance, which started in the middle ages, is not democratic and is actually rather dysfunctional: think tenure type politics. The less of it the better. It may be better to go back to just having a single editor decide if an article is worthy. That way at least more novel ideas can be presented. Of course the other problem is paid journals which have low costs (in these days of the internet) and free labor and skim lots of money for unclear reasons. Also realize most innovative researchers are on a shoestring budget because of politics. It is an awful combination: journals love the peer review system because it lets them act like they are bringing value by adding more politics and obfuscating the skimming. -Tsvi -------------- next part -------------- An HTML attachment was scrubbed... URL: From h.jaeger at rug.nl Wed Jul 20 17:35:18 2022 From: h.jaeger at rug.nl (Herbert Jaeger) Date: Wed, 20 Jul 2022 23:35:18 +0200 Subject: Connectionists: Three-year lecturing contracts in Machine Learning / AI / Cognitive Science at the University of Groningen (NL) Message-ID: <2d5ba81a-4382-30b7-d428-ad1f836771c6@rug.nl> We have a number of vacancies for 3-year teaching contracts in the Bachelor?s and Master?s programmes in Artificial Intelligence and Computational Cognitive Science, hosted in the AI department of the Bernoulli Institute (https://www.rug.nl/research/bernoulli/) of the University of Groningen. We want to fill these positions as soon as possible. Holders of PhD degrees and also of Master degrees with a relevant background and excellent qualifications are invited to apply. Teaching and campus life are all-English. Groningen University is among the top 100 universities worldwide according to the most important rankings (https://www.rug.nl/about-ug/profile/facts-and-figures/position-international-rankings) and I can happily and emphatically underline that the working conditions and the community spirit at this university are totally inviting. Plus, Groningen is a beautiful city that fulfils every expectation how a dutch city should look like. The official job offer and details can be found at https://www.rug.nl/about-ug/work-with-us/job-opportunities/?details=00347-02S0009DXP - Herbert Jaeger -- Dr. Herbert Jaeger Professor of Computing in Cognitive Materials Rijksuniversiteit Groningen Faculty of Science and Engineering - CogniGron Bernoulliborg Nijenborgh 9, 9747 AG Groningen office: Bernoulliborg 402 phone: +31 (0) 50-36 32473 COVID home office phone: +49 (0) 4209 930403 web: www.ai.rug.nl/minds/ From juergen at idsia.ch Wed Jul 20 10:55:25 2022 From: juergen at idsia.ch (Schmidhuber Juergen) Date: Wed, 20 Jul 2022 14:55:25 +0000 Subject: Connectionists: weight guessing quickly solves n-bit parity In-Reply-To: References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> Message-ID: <38F9FFC3-EA3E-4494-8E2E-324AC81FCC97@supsi.ch> I have never understood the difference between "symbolic" and "sub-symbolic" reasoning. A recurrent neural net (RNN) is a general computer that can do both. n-bit parity for arbitrary n can be solved by a tiny RNN with few connections, sequentially reading bits one by one [1]. The best learning algorithm is NOT gradient descent. Instead keep randomly initializing the RNN weights between -100 and +100 until the RNN solves parity for a few training examples of various large sizes n (this will take just 1000 trials or so). Now the RNN will probably generalize to ANY n. BTW, try that with a Transformer - it will never generalize like that. [1] J. Schmidhuber and S. Hochreiter. Guessing can outperform many long time lag algorithms. Technical Note IDSIA-19-96, IDSIA, 1996 J?rgen On 19 Jul 2022, at 23:08, Danko Nikolic wrote: Dear Barak, Thank you for the pointers. I have to read the papers. I need to understand then why, if parity is so easy to learn, my deep learning models had such a hard time that it led to an exponential growth in the number of needed parameters with each additional bit added to the input. Strange. I will report back. Best Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Mon, Jul 18, 2022 at 1:12 PM Barak A. Pearlmutter wrote: On Mon, 18 Jul 2022 at 08:28, Danko Nikolic wrote: In short, learning mechanisms cannot discover generalized XOR functions with simple connectivity -- only with complex connectivity. This problem results in exponential growth of needed resources as the number of bits in the generalized XOR increases. Assuming that "generalized XOR" means parity, this must rely on some unusual definitions which you should probably state in order to avoid confusion. Parity is a poster boy for an *easy* function to learn, albeit a nonlinear one. This is because in the (boolean) Fourier domain its spectrum consists of a single nonzero coefficient, and functions that are sparse in that domain are very easy to learn. See N. Linial, Y. Mansour, and N. Nisan, "Constant depth circuits, Fourier Transform and learnability", FOCS 1989, or Mansour, Y. (1994). Learning Boolean Functions via the Fourier Transform. Theoretical Advances in Neural Computation and Learning, 391?424. doi:10.1007/978-1-4615-2696-4_11 --Barak Pearlmutter On 18 Jul 2022, at 18:01, Gary Marcus wrote: sure, but a person can learn the idea for n-bits from a few examples with a small number of bits, generalizing it to large values of n. most current systems learn it for a certain number of bits and don?t generalize beyond that number of bits. On Jul 18, 2022, at 7:17 AM, Barak A. Pearlmutter wrote: On Mon, 18 Jul 2022 at 14:43, Danko Nikolic wrote: It is a hard problem to learn for a connectionist network. We don't need to invent new terminology, like "inverters problem" or "generalized xor." This is parity. Four (4) bit parity. https://en.wikipedia.org/wiki/Parity_function Parity is *not* a hard function to learn. Even for a connectionist network. It is an interesting function for historic reasons (n-bit parity cannot be loaded by a k-th order perceptron, for k wrote: Sorry, I can't let this go by: And it is good so because generalize XOR scales worse than power law. It scales exponentially! This a more agressive form of explosion than power law. I'm not sure exactly what you mean by this, but a single-hidden layer network with N inputs and N hidden units can solve N-bit parity. Each unit has an increasing threshold, so, one turns on if there is one unit on in the input, and then turns on the output with a weight of +1. If two units are on in the input, then a second unit comes on and cancels the activation of the first unit via a weight of -1. Etc. g. On Sat, Jul 16, 2022 at 12:03 AM Danko Nikolic wrote: Dear Thomas, Thank you for reading the paper and for the comments. I cite: "In my experience, supervised classification scales linearly in the number of classes." This would be good to quantify as a plot. Maybe a research paper would be a good idea. The reason is that it seems that everyone else who tried to quantify that relation found a power law. At this point, it would be surprising to find a linear relationship. And it would probably make a well read paper. But please do not forget that my argument states that even a linear relationship is not good enough to match bilogical brains. We need something more similar to a power law with exponent zero when it comes to the model size i.e., a constant number of parameters in the model. And we need linear relationship when it comes to learning time: Each newly learned object should needs about as much of learning effort as was needed for each previous object. I cite: "The real world is not dominated by generalized XOR problems." Agreed. And it is good so because generalize XOR scales worse than power law. It scales exponentially! This a more agressive form of explosion than power law. Importantly, a generalized AND operation also scales exponentially (with a smaller exponent, though). I guess we would agree that the real world probably encouners a lot of AND problems. The only logical operaiton that could be learned with a linear increase in the number of parameters was a generalized OR. Finally, I foiund that a mixure of AND and OR resulted in a power law-like scaling of the number of parameters. So, a mixture of AND and OR seemed to scale as good (or as bad) as the real world. I have put this information into Supplementary Materials. The conclusion that I derived from those analyses is: connectionism is not sustainable to reach human (or animal) levels of intelligence. Therefore, I hunted for an alternative pradigm. Greetings, Danko From mrs at fel.cvut.cz Wed Jul 20 13:32:40 2022 From: mrs at fel.cvut.cz (Multi-Robot Systems Group in Prague) Date: Wed, 20 Jul 2022 19:32:40 +0200 Subject: Connectionists: hybrid (online/physical) IEEE Summer School on Multi-Robot Systems (UAVs), August 1-5, 2022, call for online participation Message-ID: <0dc92130-f0ba-e767-cd2d-8d7317d42bc0@fel.cvut.cz> Based on numerous requests from countries affected by COVID with travel restrictions and students with insufficient traveling funding, the IEEE Summer School on Multi-Robot Systems with real UAV experiments will allow both physical and virtual participation.?All lectures, meetings, demos, and practical courses (including the students' experiments in real-world conditions) will be provided also online so that online participants will receive an equivalent amount of?information and knowledge as the students being physically in Prague. Representatives of almost all robotic labs worldwide have registered to the school (so far we have 220 registrations) but only people from Europe and a few other countries can come physically and it would not be fair to exclude students from almost the entire Asia for example. But we believe the online possibility may help also many people from Europe. If you or your students/colleagues consider participation at the school, I?recommend?you to register asap http://mrs.felk.cvut.cz/summer-school-2022/apply.htmlsince the week of the school is coming soon. The Summer School is proper for last year's bachelor?s students, master?s students, and Ph.D. students, but we would also like to invite young scientists from both industry and academia working in the field of Aerial Robotics and Multi Agent Systems. Mainly the planned experiments with a fleet of?real UAVs realized by students (online as well as physical) during the Summer School practicals could be within their interest. You can find more details about the Summer School in the call for participation, below this email. Best Regards, Martin -------------------------------------------------------------------------------------------------- Following the great success of Summer Schools in Singapore, 2016 and in Prague, 2019 & 2020 the MRS RAS technical committee is organising 2022 IEEE RAS Summer School on Multi-Robot Systems http://mrs.felk.cvut.cz/summer-school-2022/?with real-world UAV experiments which will be held again at the campus of Czech Technical University located at the heart of the historic city of Prague. The Summer School aims to promote the newest achievements in multi-robot systems, aerial and swarm robotics to students, academic researchers, and industrial practitioners to enable putting systems of cooperating robots into practice. The?content of the Summer School will be focused but not limited to systems of cooperating aerial vehicles. The topics addressed by well-recognised experts in the field of Multi-Robot Systems are composed to provide the participants?necessary knowledge for understanding the available theory and for realisation real-world experiments with a fleet of autonomous micro aerial vehicles in the last day of the Summer School. This year the school will be focused on?deployment of MRS in real-world conditions being motivated by EU AERIAL-CORE https://aerial-core.eu?project and DARPA SubT challenge www.subtchallenge.com The school is considered for presence participants primarily since a minimal amount of COVID cases and no travel restrictions were experienced in the Czech Republic in July-August in 2020 and 2021. Nevertheless, we will allow also?virtual participation for students from countries with travel restrictions or limitations. ---------------------------------------------------------------------------- 2022 IEEE RAS Summer School on Multi-Robot Systems ---------------------------------------------------------------------------- WHERE: ? ? ? ? ? ? ? ? ? ? ? ? ? ? Prague, Czech Republic WHEN: ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? August 1-5, 2022 HOW TO APPLY: ? ? ? ? ? ? ? Submit CV, personal information, and a brief statement of interest through: https://forms.gle/ehFe6Beesv1LfHKC6 Include title and abstract of your short talk if you are interested in providing a talk on your research (optional) WHO SHOULD ATTEND: ?Graduate students (master or PhD) and students (Bc.) at an advanced stage of their university courses. R&D professionals. An accredited course equivalent to 2 ECTS FEE: ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? Students: 440 ? + 92.4 ? VAT ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? Academic participants: 510 ? + 107.1 ? VAT ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? Industry participants: 590 ? + 123.9 ? VAT * The fee includes all lectures, practicals with real robots, lunches, refreshments, welcome drinks, banquet, farewell drinks, and social programs - guided tour in Prague's Old Town LECTURERS: Konstantinos Alexis. Professor at?Norwegian university of science and technology. Winner of DARPA SubT challenge - motivating Summer School experiments http://mrs.felk.cvut.cz/summer-school-2022/outdoor.html Rachid Alami. Senior Scientist (Directeur de Recherche) at Laboratory for Analysis and Architecture of Systems (LAAS) An?bal Ollero. Professor at the University of Seville. Principal investigator of Aerial Core?-?motivating?Summer School experiments http://mrs.felk.cvut.cz/summer-school-2022/outdoor.html Tom?? Svoboda.?Professor at the?Czech Technical University in Prague.?2nd place at virtual DARPA SubT challenge?-?motivating?Summer School experiments http://mrs.felk.cvut.cz/summer-school-2022/outdoor.html Alyssa Pierson.?Professor at?Boston University Guido de Croon. Professor at Delft University of Technology Vito Trianni. Senior researcher at Istituto di Scienze e Tecnologie della Cognizione Lino Forte Marques.?Professor at?University of Coimbra PRACTICALS: Tom?? B??a, Czech Technical University: Outdoor experiments with swarms of micro aerial vehicles http://mrs.felk.cvut.cz/summer-school-2022/outdoor.html Robert P?ni?ka, Czech Technical University: ROS and Gazebo for verification of multirobot systems ORGANIZERS: Martin Saska, Czech Technical University, Summer School Chair Robert Fitch, The University of Sydney, Technical Committee and Advisory Board Nora Ayanian, University of Southern California, Technical Committee and Advisory Board Antonio Franchi, University of Twente, Technical Committee and Advisory Board Lorenzo Sabatini, University of Modena, Technical Committee and Advisory Board Jen Jen Chung, ETH Z?rich, Technical Committee and Advisory Board Changjoo Nam, ?Inha University, Technical Committee and Advisory Board Javier Alonso-Mora, Delft University of Technology, Technical Committee and Advisory Board Alyssa Pierson, Boston University, Technical Committee and Advisory Board WEB PAGE: http://mrs.felk.cvut.cz/summer-school-2022/ CONTACT: mrs at fel.cvut.cz Best regards on behalf of organizers and IEEE RAS MRS TC, Martin ----------------------- Dr. Martin Saska Head of Multi-robot Systems Group Faculty of Electrical Engineering Czech Technical University in Prague Karlovo n?m?st? 13 CZ 121 35 Prague 2 Office: E120 Voice: +420-776241932 http://mrs.felk.cvut.cz/ http://mrs.felk.cvut.cz/people/martin-saska -------------- next part -------------- An HTML attachment was scrubbed... URL: From frothga at sandia.gov Wed Jul 20 09:13:15 2022 From: frothga at sandia.gov (Rothganger, Fredrick) Date: Wed, 20 Jul 2022 13:13:15 +0000 Subject: Connectionists: Tools for AI In-Reply-To: <9ea52c4c-2b66-e8bd-b494-5dc4ab7ebc8d@susaro.com> References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> <9ea52c4c-2b66-e8bd-b494-5dc4ab7ebc8d@susaro.com> Message-ID: Richard Loosemore said: > 2) An AI/Cognitive system development tool that would allow people to build and explore complex cognitive systems without being shackled to one particular architecture (like deep learning and its many descendents). The model of open-source community software development can address this. We can build some very big things without grant support, if we share a vision of what to build. This of course goes back to the need to gain mind-share among our fellow researchers. Regarding DL tools -- The notion of a computational graph that chains functions together via data-flow, where the primary data type is a tensor, is quite general. It does not restrict you to only DL. Finally, a small advertisement for N2A (https://github.com/frothga/n2a). This is a dynamical-system modeling tool that a couple of neuroscientists and I designed about 10 years ago, and which I actively develop. It's long-term goal is to provide a framework for describing the computation of the entire brain. Of course, that requires getting community mind-share, particularly among neuroscientists, and there are numerous other neuroinformatic and simulation tools out there. The distinction is that N2A is designed from the ground up to integrate very large and complex models. [https://opengraph.githubassets.com/c7d6fb77fcaac8f95f3f7d0b74243e98353f74454e7c40de473c10add7fa1db4/frothga/n2a] GitHub - frothga/n2a: An object-oriented language for modeling large-scale neural systems, along with an IDE for writing and simulating models. An object-oriented language for modeling large-scale neural systems, along with an IDE for writing and simulating models. - GitHub - frothga/n2a: An object-oriented language for modeling large-scale neural systems, along with an IDE for writing and simulating models. github.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From andreas.wichert at tecnico.ulisboa.pt Thu Jul 21 05:07:22 2022 From: andreas.wichert at tecnico.ulisboa.pt (Andrzej Wichert) Date: Thu, 21 Jul 2022 10:07:22 +0100 Subject: Connectionists: weight guessing quickly solves n-bit parity In-Reply-To: <38F9FFC3-EA3E-4494-8E2E-324AC81FCC97@supsi.ch> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <38F9FFC3-EA3E-4494-8E2E-324AC81FCC97@supsi.ch> Message-ID: <76994187-BC20-4B04-A818-85CD025674D8@tecnico.ulisboa.pt> Dear Juergen, Symbols do not, by themselves, represent any utilizable knowledge, they cannot be used for a definition of similarity criteria between themselves. The use of symbols in algorithms which imitate human intelligent behavior led to the famous physical symbol system hypothesis \index{physical symbol system hypothesis} by Newell and Simon (1976) ``The necessary and sufficient condition for a physical system to exhibit intelligence is that it be a physical symbol system.'' Symbols are not present in the world; they are the constructs of a human mind and simplify the process of representation used in communication and problem solving. A Turing machine can simulate any algorithm, the same with RNN, but they do not explain the human problem solving. There is a difference, for example human problem solving can be described by production systems. The most successful model is the SOAR architecture. Such systems were very successful, they learn and can give you an explanation for their doing. in August 1998 Dave Touretzky asked on the connectionistic e-mailing list: ``Is connectionist symbol processing dead??' From ml-connectionists-request$@$mlist-1.sp.cs.cmu.edu Tue Aug 11 17:35:10 1998 From: Dave$\_$Touretzky$@$cs.cmu.edu To: connectionists$@$cs.cmu.edu Subject: Connectionist symbol processing: any progress? Date: Tue, 11 Aug 1998 $03:34:27$ -0400 I'd like to start a debate on the current state of connectionist symbol processing? Is it dead? Or does progress continue? ... People had gotten some interesting effects with localist networks, by doing spreading activation and a simple form of constraint satisfaction.... This approach does not create new structure on the fly, or deal with structured representations or variable binding. Those localist networks that did attempt to implement variable binding did so in a discrete, symbolic way that did not advance the parallel constraint satisfaction/heuristic reasoning agenda of earlier spreading activation research. ... So I concluded that connectionist symbol processing had reached a plateau, and further progress would have to await some revolutionary new insight about representations. ... The problems of structured representations and variable binding have remained unsolved. No one is trying to build distributed connectionist reasoning systems any more, like the connectionist production system I built with Geoff Hinton... 24 years past and not much progress was done. It seems that the progress is only related to pure brute force of computers, but not much insight beside a wishful thinking. The whole DL movements stops the progress in understanding how the brain works. We need some new fresh ideas beside error minimization. I think one of the main problems is the publish or perish altitude, the famous Google impact factor. One does not care what some one is doing, one just checks his Google impact factor.. This is like the saying, eat more shit, one million flies cannot be wrong. Like some mathematician said, computer science is not science at all, but it force us to follow its ideas. Andreas -------------------------------------------------------------------------------------------------- Prof. Auxiliar Andreas Wichert http://web.tecnico.ulisboa.pt/andreas.wichert/ - https://www.amazon.com/author/andreaswichert Instituto Superior T?cnico - Universidade de Lisboa Campus IST-Taguspark Avenida Professor Cavaco Silva Phone: +351 214233231 2744-016 Porto Salvo, Portugal > On 20 Jul 2022, at 15:55, Schmidhuber Juergen wrote: > > > I have never understood the difference between "symbolic" and "sub-symbolic" reasoning. A recurrent neural net (RNN) is a general computer that can do both. n-bit parity for arbitrary n can be solved by a tiny RNN with few connections, sequentially reading bits one by one [1]. The best learning algorithm is NOT gradient descent. Instead keep randomly initializing the RNN weights between -100 and +100 until the RNN solves parity for a few training examples of various large sizes n (this will take just 1000 trials or so). Now the RNN will probably generalize to ANY n. BTW, try that with a Transformer - it will never generalize like that. > > [1] J. Schmidhuber and S. Hochreiter. Guessing can outperform many long time lag algorithms. Technical Note IDSIA-19-96, IDSIA, 1996 > > J?rgen > > > > > On 19 Jul 2022, at 23:08, Danko Nikolic wrote: > > Dear Barak, > > Thank you for the pointers. I have to read the papers. I need to understand then why, if parity is so easy to learn, my deep learning models had such a hard time that it led to an exponential growth in the number of needed parameters with each additional bit added to the input. Strange. > > I will report back. > > Best > > Danko > > Dr. Danko Nikoli? > www.danko-nikolic.com > https://www.linkedin.com/in/danko-nikolic/ > -- I wonder, how is the brain able to generate insight? -- > > > On Mon, Jul 18, 2022 at 1:12 PM Barak A. Pearlmutter wrote: > On Mon, 18 Jul 2022 at 08:28, Danko Nikolic wrote: > In short, learning mechanisms cannot discover generalized XOR functions with simple connectivity -- only with complex connectivity. This problem results in exponential growth of needed resources as the number of bits in the generalized XOR increases. > > Assuming that "generalized XOR" means parity, this must rely on some > unusual definitions which you should probably state in order to avoid > confusion. > > Parity is a poster boy for an *easy* function to learn, albeit a > nonlinear one. This is because in the (boolean) Fourier domain its > spectrum consists of a single nonzero coefficient, and functions that > are sparse in that domain are very easy to learn. See N. Linial, Y. > Mansour, and N. Nisan, "Constant depth circuits, Fourier Transform and > learnability", FOCS 1989, or Mansour, Y. (1994). Learning Boolean > Functions via the Fourier Transform. Theoretical Advances in Neural > Computation and Learning, 391?424. doi:10.1007/978-1-4615-2696-4_11 > > --Barak Pearlmutter > > > > On 18 Jul 2022, at 18:01, Gary Marcus wrote: > > sure, but a person can learn the idea for n-bits from a few examples with a small number of bits, generalizing it to large values of n. most current systems learn it for a certain number of bits and don?t generalize beyond that number of bits. > > On Jul 18, 2022, at 7:17 AM, Barak A. Pearlmutter wrote: > > > > On Mon, 18 Jul 2022 at 14:43, Danko Nikolic wrote: > > > It is a hard problem to learn for a connectionist network. > > We don't need to invent new terminology, like "inverters problem" or "generalized xor." This is parity. Four (4) bit parity. > > https://en.wikipedia.org/wiki/Parity_function > > Parity is *not* a hard function to learn. Even for a connectionist network. > > It is an interesting function for historic reasons (n-bit parity cannot be loaded by a k-th order perceptron, for k > --Barak Pearlmutter. > > > > > On 18 Jul 2022, at 00:58, gary at ucsd.edu wrote: > > Sorry, I can't let this go by: > And it is good so because generalize XOR scales worse than power law. It scales exponentially! This a more agressive form of explosion than power law. > I'm not sure exactly what you mean by this, but a single-hidden layer network with N inputs and N hidden units can solve N-bit parity. Each unit has an increasing threshold, so, one turns on if there is one unit on in the input, and then turns on the output with a weight of +1. If two units are on in the input, then a second unit comes on and cancels the activation of the first unit via a weight of -1. Etc. > > g. > > > On Sat, Jul 16, 2022 at 12:03 AM Danko Nikolic wrote: > Dear Thomas, > > Thank you for reading the paper and for the comments. > > I cite: "In my experience, supervised classification scales linearly in the number of classes." > This would be good to quantify as a plot. Maybe a research paper would be a good idea. The reason is that it seems that everyone else who tried to quantify that relation found a power law. At this point, it would be surprising to find a linear relationship. And it would probably make a well read paper. > > But please do not forget that my argument states that even a linear relationship is not good enough to match bilogical brains. We need something more similar to a power law with exponent zero when it comes to the model size i.e., a constant number of parameters in the model. And we need linear relationship when it comes to learning time: Each newly learned object should needs about as much of learning effort as was needed for each previous object. > > I cite: "The real world is not dominated by generalized XOR problems." > Agreed. And it is good so because generalize XOR scales worse than power law. It scales exponentially! This a more agressive form of explosion than power law. > Importantly, a generalized AND operation also scales exponentially (with a smaller exponent, though). I guess we would agree that the real world probably encouners a lot of AND problems. The only logical operaiton that could be learned with a linear increase in the number of parameters was a generalized OR. Finally, I foiund that a mixure of AND and OR resulted in a power law-like scaling of the number of parameters. So, a mixture of AND and OR seemed to scale as good (or as bad) as the real world. I have put this information into Supplementary Materials. > > The conclusion that I derived from those analyses is: connectionism is not sustainable to reach human (or animal) levels of intelligence. Therefore, I hunted for an alternative pradigm. > > Greetings, > > Danko > -------------- next part -------------- An HTML attachment was scrubbed... URL: From melaahi2 at gmail.com Thu Jul 21 08:07:47 2022 From: melaahi2 at gmail.com (M Elahi) Date: Thu, 21 Jul 2022 16:37:47 +0430 Subject: Connectionists: PhD position in User Modelling, Personalisation & Engagement at MediaFutures Message-ID: There is a vacancy for a PhD position at MediaFutures: Research Centre for Responsible Media Technology & Innovation. The position is for a fixed-term period of 3 years. MediaFutures is a new centre for research-based innovation at the Department for information science and media studies, University of Bergen, and located at Media City Bergen. The centre is a part of the Norwegian Research Council?s Research-based Innovation scheme. MediaFutures is in cooperation with industry partners from news media and media technology research topics like media experiences, recommender technology, content production and analysis, content interaction and availability, and Norwegian language technology. Qualifications and personal qualities: - The applicant must hold a master's degree or the equivalent in one of the relevant disciplines by the starting date of the position: information science, data science, computer science or related fields. - It is a condition for employment that the master has been awarded. - Programming skills and knowledge of a programming language, such as Python or R, is essential. - Data analysis skills in Python, R, or STATA can be an additional asset. - The applicant should be motivated to contribute to the emerging field of recommender systems research. - The requirements are generally a grade B or better on the Master thesis and for the Master degree in total. - As an applicant you should have a considerable work capacity as well as an enthusiasm for research and the ability and interest to work in a team. - A firm basis in quantitative research methods and data analysis is an advantage. - As an applicant you should have an excellent written and spoken command of English. - Experience with recommender systems research is an advantage. Link to apply: https://www.jobbnorge.no/en/available-jobs/job/228463/phd-position-in-user-modelling-personalisation-engagement-at-mediafutures Deadline: *14th August 2022* ??????????????? Mehdi Elahi Associate Professor, University of Bergen, Group: dars.uib.no Project: mediafutures.no Twitter: twitter.com/mehdielaahi LinkedIn: linkedin.com/in/mehdielahi GoogleScholar: scholar.google.com/citations?user=aUWF7LYAAAAJ ??????????????? -------------- next part -------------- An HTML attachment was scrubbed... URL: From malin.sandstrom at incf.org Thu Jul 21 10:42:00 2022 From: malin.sandstrom at incf.org (=?UTF-8?Q?Malin_Sandstr=C3=B6m?=) Date: Thu, 21 Jul 2022 16:42:00 +0200 Subject: Connectionists: INCF IC survey on barriers to neuroscience data sharing and reuse - please forward Message-ID: The INCF Infrastructure Committee has created a brief (~2 min) anonymous survey to chart barriers to data sharing and reuse among neuroscience researchers worldwide. The results will be made public, and will be used by INCF and collaborators to develop strategies and activities for supporting the global neuroscience community. To have real impact, the survey needs to reach the broadest possible range of neuroscience researchers. We would be very grateful for your help in spreading the word to your friends and colleagues - the more responses we get, the more valuable and useful the survey will be! Take the survey *No personal data will be collected within this survey, but please note that the survey uses Google Forms; their privacy policies can be found here . Follow the link to the survey if you agree with these policies. Best regards, Malin Sandstr?m on behalf of the INCF Infrastructure Committee -- Malin Sandstr?m, PhD Community Engagement Officer malin.sandstrom at incf.org ORCID: 0000-0002-8464-2494 International Neuroinformatics Coordinating Facility Karolinska Institutet Nobels v?g 15 A SE-171 77 Stockholm Sweden http://www.incf.org -------------- next part -------------- An HTML attachment was scrubbed... URL: From EmmettRedd at MissouriState.edu Thu Jul 21 11:53:13 2022 From: EmmettRedd at MissouriState.edu (Redd, Emmett R) Date: Thu, 21 Jul 2022 15:53:13 +0000 Subject: Connectionists: weight guessing quickly solves n-bit parity In-Reply-To: <76994187-BC20-4B04-A818-85CD025674D8@tecnico.ulisboa.pt> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <38F9FFC3-EA3E-4494-8E2E-324AC81FCC97@supsi.ch> <76994187-BC20-4B04-A818-85CD025674D8@tecnico.ulisboa.pt> Message-ID: Since a Turing machine cannot "not explain the human problem solving", super-Turing computation should be considered. There are two recent papers exploring computation consistent with super-Turing operation: https://journals.aps.org/prresearch/abstract/10.1103/PhysRevResearch.3.013120 and https://link.springer.com/chapter/10.1007/978-3-030-86380-7_38 . The APS paper is open source. The studies include noise which the brain also has. They simulate a quantal nature. Neuron excitation is also quantal. [https://cdn.journals.aps.org/journals/PRRESEARCH/key_images/10.1103/PhysRevResearch.3.013120.png] Noise optimizes super-Turing computation in recurrent neural networks The authors use noise to restore chaotic behavior and show consistency with super-Turing theory and operation in neural networks. journals.aps.org [https://static-content.springer.com/cover/book/978-3-030-86380-7.jpg] Noise Quality and Super-Turing Computation in Recurrent Neural Networks - SpringerLink Figures 3 and 4 show the number of output sequences which remain consistent with chaos (y-axis) vs. \(log_2\) of the repeat length of the noise (x-axis) for the logistic and H?non maps digital RNNs for both optimum and near-optimum noise magnitudes and the two noise types (MATLAB vs. LFSR). As mentioned in Sec. 2, noise magnitudes of four times LSB is regarded as optimum for the H?non map ... link.springer.com ? Emmett Redd Ph.D. mailto:EmmettRedd at missouristate.edu Professor (417)836-5221 Department of Physics, Astronomy, and Materials Science Missouri State University Fax (417)836-6226 901 SOUTH NATIONAL SPRINGFIELD, MO 65897 USA Dept (417)836-5131 Perfect logic and faultless deduction make a pleasant theoretical structure, but it may be right or wrong; the experimenter is the only one to decide, and he is always right. ?L?on Brillouin, Scientific Uncertainty, and Information (1964). ________________________________ From: Connectionists on behalf of Andrzej Wichert Sent: Thursday, July 21, 2022 4:07 AM To: Schmidhuber Juergen Cc: connectionists at cs.cmu.edu Subject: Re: Connectionists: weight guessing quickly solves n-bit parity CAUTION: External Sender Dear Juergen, Symbols do not, by themselves, represent any utilizable knowledge, they cannot be used for a definition of similarity criteria between themselves. The use of symbols in algorithms which imitate human intelligent behavior led to the famous physical symbol system hypothesis \index{physical symbol system hypothesis} by Newell and Simon (1976) ``The necessary and sufficient condition for a physical system to exhibit intelligence is that it be a physical symbol system.'' Symbols are not present in the world; they are the constructs of a human mind and simplify the process of representation used in communication and problem solving. A Turing machine can simulate any algorithm, the same with RNN, but they do not explain the human problem solving. There is a difference, for example human problem solving can be described by production systems. The most successful model is the SOAR architecture. Such systems were very successful, they learn and can give you an explanation for their doing. in August 1998 Dave Touretzky asked on the connectionistic e-mailing list: ``Is connectionist symbol processing dead??' From ml-connectionists-request$@$mlist-1.sp.cs.cmu.edu Tue Aug 11 17:35:10 1998 From: Dave$\_$Touretzky$@$cs.cmu.edu To: connectionists$@$cs.cmu.edu Subject: Connectionist symbol processing: any progress? Date: Tue, 11 Aug 1998 $03:34:27$ -0400 I'd like to start a debate on the current state of connectionist symbol processing? Is it dead? Or does progress continue? ... People had gotten some interesting effects with localist networks, by doing spreading activation and a simple form of constraint satisfaction.... This approach does not create new structure on the fly, or deal with structured representations or variable binding. Those localist networks that did attempt to implement variable binding did so in a discrete, symbolic way that did not advance the parallel constraint satisfaction/heuristic reasoning agenda of earlier spreading activation research. ... So I concluded that connectionist symbol processing had reached a plateau, and further progress would have to await some revolutionary new insight about representations. ... The problems of structured representations and variable binding have remained unsolved. No one is trying to build distributed connectionist reasoning systems any more, like the connectionist production system I built with Geoff Hinton... 24 years past and not much progress was done. It seems that the progress is only related to pure brute force of computers, but not much insight beside a wishful thinking. The whole DL movements stops the progress in understanding how the brain works. We need some new fresh ideas beside error minimization. I think one of the main problems is the publish or perish altitude, the famous Google impact factor. One does not care what some one is doing, one just checks his Google impact factor.. This is like the saying, eat more shit, one million flies cannot be wrong. Like some mathematician said, computer science is not science at all, but it force us to follow its ideas. Andreas -------------------------------------------------------------------------------------------------- Prof. Auxiliar Andreas Wichert http://web.tecnico.ulisboa.pt/andreas.wichert/ - https://www.amazon.com/author/andreaswichert Instituto Superior T?cnico - Universidade de Lisboa Campus IST-Taguspark Avenida Professor Cavaco Silva Phone: +351 214233231 2744-016 Porto Salvo, Portugal On 20 Jul 2022, at 15:55, Schmidhuber Juergen > wrote: I have never understood the difference between "symbolic" and "sub-symbolic" reasoning. A recurrent neural net (RNN) is a general computer that can do both. n-bit parity for arbitrary n can be solved by a tiny RNN with few connections, sequentially reading bits one by one [1]. The best learning algorithm is NOT gradient descent. Instead keep randomly initializing the RNN weights between -100 and +100 until the RNN solves parity for a few training examples of various large sizes n (this will take just 1000 trials or so). Now the RNN will probably generalize to ANY n. BTW, try that with a Transformer - it will never generalize like that. [1] J. Schmidhuber and S. Hochreiter. Guessing can outperform many long time lag algorithms. Technical Note IDSIA-19-96, IDSIA, 1996 J?rgen On 19 Jul 2022, at 23:08, Danko Nikolic > wrote: Dear Barak, Thank you for the pointers. I have to read the papers. I need to understand then why, if parity is so easy to learn, my deep learning models had such a hard time that it led to an exponential growth in the number of needed parameters with each additional bit added to the input. Strange. I will report back. Best Danko Dr. Danko Nikoli? www.danko-nikolic.com https://www.linkedin.com/in/danko-nikolic/ -- I wonder, how is the brain able to generate insight? -- On Mon, Jul 18, 2022 at 1:12 PM Barak A. Pearlmutter wrote: On Mon, 18 Jul 2022 at 08:28, Danko Nikolic wrote: In short, learning mechanisms cannot discover generalized XOR functions with simple connectivity -- only with complex connectivity. This problem results in exponential growth of needed resources as the number of bits in the generalized XOR increases. Assuming that "generalized XOR" means parity, this must rely on some unusual definitions which you should probably state in order to avoid confusion. Parity is a poster boy for an *easy* function to learn, albeit a nonlinear one. This is because in the (boolean) Fourier domain its spectrum consists of a single nonzero coefficient, and functions that are sparse in that domain are very easy to learn. See N. Linial, Y. Mansour, and N. Nisan, "Constant depth circuits, Fourier Transform and learnability", FOCS 1989, or Mansour, Y. (1994). Learning Boolean Functions via the Fourier Transform. Theoretical Advances in Neural Computation and Learning, 391?424. doi:10.1007/978-1-4615-2696-4_11 --Barak Pearlmutter On 18 Jul 2022, at 18:01, Gary Marcus wrote: sure, but a person can learn the idea for n-bits from a few examples with a small number of bits, generalizing it to large values of n. most current systems learn it for a certain number of bits and don?t generalize beyond that number of bits. On Jul 18, 2022, at 7:17 AM, Barak A. Pearlmutter wrote: On Mon, 18 Jul 2022 at 14:43, Danko Nikolic wrote: It is a hard problem to learn for a connectionist network. We don't need to invent new terminology, like "inverters problem" or "generalized xor." This is parity. Four (4) bit parity. https://en.wikipedia.org/wiki/Parity_function Parity is *not* a hard function to learn. Even for a connectionist network. It is an interesting function for historic reasons (n-bit parity cannot be loaded by a k-th order perceptron, for k wrote: Sorry, I can't let this go by: And it is good so because generalize XOR scales worse than power law. It scales exponentially! This a more agressive form of explosion than power law. I'm not sure exactly what you mean by this, but a single-hidden layer network with N inputs and N hidden units can solve N-bit parity. Each unit has an increasing threshold, so, one turns on if there is one unit on in the input, and then turns on the output with a weight of +1. If two units are on in the input, then a second unit comes on and cancels the activation of the first unit via a weight of -1. Etc. g. On Sat, Jul 16, 2022 at 12:03 AM Danko Nikolic wrote: Dear Thomas, Thank you for reading the paper and for the comments. I cite: "In my experience, supervised classification scales linearly in the number of classes." This would be good to quantify as a plot. Maybe a research paper would be a good idea. The reason is that it seems that everyone else who tried to quantify that relation found a power law. At this point, it would be surprising to find a linear relationship. And it would probably make a well read paper. But please do not forget that my argument states that even a linear relationship is not good enough to match bilogical brains. We need something more similar to a power law with exponent zero when it comes to the model size i.e., a constant number of parameters in the model. And we need linear relationship when it comes to learning time: Each newly learned object should needs about as much of learning effort as was needed for each previous object. I cite: "The real world is not dominated by generalized XOR problems." Agreed. And it is good so because generalize XOR scales worse than power law. It scales exponentially! This a more agressive form of explosion than power law. Importantly, a generalized AND operation also scales exponentially (with a smaller exponent, though). I guess we would agree that the real world probably encouners a lot of AND problems. The only logical operaiton that could be learned with a linear increase in the number of parameters was a generalized OR. Finally, I foiund that a mixure of AND and OR resulted in a power law-like scaling of the number of parameters. So, a mixture of AND and OR seemed to scale as good (or as bad) as the real world. I have put this information into Supplementary Materials. The conclusion that I derived from those analyses is: connectionism is not sustainable to reach human (or animal) levels of intelligence. Therefore, I hunted for an alternative pradigm. Greetings, Danko This message originated outside Missouri State University. Please use caution when opening attachments, clicking links, or replying. -------------- next part -------------- An HTML attachment was scrubbed... URL: From xavier.hinaut at inria.fr Thu Jul 21 14:58:10 2022 From: xavier.hinaut at inria.fr (Xavier Hinaut) Date: Thu, 21 Jul 2022 20:58:10 +0200 Subject: Connectionists: Reminder: SMILES workshop call extended 28th July Message-ID: The SMILES (Sensorimotor Interaction, Language and Embodiment of Symbols) Workshop will take place both on site and virtually at the ICDL 2022 (International Conference on Developmental Learning). * Call for abstracts : - Deadline extended: July 28th - Abstracts call: from 1/2 page to 2 pages (onsite and virtual participation are possible) - Abstract format: same as ICDL conference https://www.ieee.org/conferences/publishing/templates.html - Submissions: smiles.conf at gmail.com + indicate if you will be onsite or online - Workshop dates: September 12, 2022 - Venue onsite: Queen Mary University of London, UK. - Venue online: via Zoom and Discord group. - Join our Discord: https://discord.gg/B8xbemQS Accepted abstract will be asked to make a short video or poster for the workshop. * Workshop Short Description On the one hand, models of sensorimotor interaction are embodied in the environment and in the interaction with other agents. On the other hand, recent Deep Learning development of Natural Language Processing (NLP) models allow to capture increasing language complexity (e.g. compositional representations, word embedding, long term dependencies). However, those NLP models are disembodied in the sense that they are learned from static datasets of text or speech. How can we bridge the gap from low-level sensorimotor interaction to high-level compositional symbolic communication? The SMILES workshop will address this issue through an interdisciplinary approach involving researchers from (but not limited to): - Sensori-motor learning, - Symbol grounding and symbol emergence, - Emergent communication in multi-agent systems, - Chunking of perceptuo-motor gestures (gestures in a general sense: motor, vocal, ...), - Compositional representations for communication and action sequence, - Hierarchical representations of temporal information, - Language processing and language acquisition in brains and machines, - Models of animal communication, - Understanding composition and temporal processing in neural network models, and - Enaction, active perception, perception-action loop. * More info - contact: smiles.conf at gmail.com - organizers: Xavier Hinaut, Cl?ment Moulin-Frier, Silvia Pagliarini, Joni Zhong, Michael Spranger, Tadahiro Taniguchi, Anne Warlaumont. - invited speakers (coming soon) - workshop website (updated regularly): https://sites.google.com/view/smiles-workshop/ - ICDL conference website: https://icdl2022.qmul.ac.uk/ Xavier Hinaut on behalf of the SMILES workshop organisers: - Xavier Hinaut, Inria, Bordeaux, France - Cl?ment Moulin-Frier, Inria and Ensta ParisTech, Bordeaux, France - Anne Warlaumont, UCLA, Los Angeles, USA - Silvia Pagliarini, UCLA, Los Angeles, USA - Michael Spranger, Sony AI and Sony CSL, Tokyo, Japan - Tadahiro Taniguchi, Ritsumeikan University, Kyoto, Japan - Junpei Zhong, Hong Kong Polytechnic University, Hong Kong, China Xavier Hinaut Inria Research Scientist www.xavierhinaut.com -- +33 5 33 51 48 01 Mnemosyne team, Inria, Bordeaux, France -- https://team.inria.fr/mnemosyne & LaBRI, Bordeaux University -- https://www4.labri.fr/en/formal-methods-and-models & IMN (Neurodegeneratives Diseases Institute) -- http://www.imn-bordeaux.org/en --- Our new release of Reservoir Computing library: https://github.com/reservoirpy/reservoirpy -------------- next part -------------- An HTML attachment was scrubbed... URL: From laurenz.wiskott at rub.de Thu Jul 21 12:56:52 2022 From: laurenz.wiskott at rub.de (Laurenz Wiskott) Date: Thu, 21 Jul 2022 18:56:52 +0200 Subject: Connectionists: [jobs] 2 year independent Postdoc position in human centered AI (German required) Message-ID: <20220721165652.GA16776@curry> The Institute for Neural Computation of the Faculty of Computer Science is looking for a Postdoc (m,f,x) for 2 years, full-time, TVL E13 The Institute for Neural Computation is part of the Faculty of Computer Science at the Ruhr-University Bochum, see https://www.ini.rub.de/. It 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. There is an open position for a postdoc in the group of Prof. Dr. Laurenz Wiskott to be filled October 1st. The project is part of the competence center HUMAINE for human-centered artificial intelligence, see https://humaine.info/, and will be developed in close collaboration with colleagues at RUB and application partners in industry, among others. We offer: * An interesting interdisciplinary environment in the field of Artificial Intelligence / Machine Learning. * Opportunity to develop an independent research profile and your own research group. * Opportunity to pursue further scientific qualification. * Diverse and interesting contacts at the university, in transfer, and in industry. * A lot of room for own ideas and initiatives. Your tasks: * Acquiring and maintaining industry contacts. * Scientific advice for companies to help them introduce AI in work processes. * Advising corresponding PhD and student projects. * Coordination of events and teaching of advanced training courses about human-centered AI for developers and decision makers in industry Your profile: * Very good PhD in mathematics, computer science, engineering, or a related field * Solid knowledge in Artificial Intelligence / Machine Learning. * Good programming and mathematical skills * Good command of German. * Good organizational and communicative abilities. * Advantageous: + Experience in industry cooperations and interdisciplinary projects. + Programming skills in Python. Our offerings: * Challenging and varied tasks with a high degree of personal responsibility * Support from and cooperation with competent colleagues * A dynamic environment * Fair and flexible working hours * Support in reconciling family and work/academics (including childcare placement, holiday offers, mobile children's room) Additional information: The load of teaching will be calculated according to ? 3 of Lehrverpflichtungsverordnung (state of North Rhine-Westphalia). Contact details for your application: Prof. Dr. Laurenz Wiskott, Phone: +49234 32 27997 Travel expenses for interviews cannot be refunded. For information on the collection of personal data in the application process see: https://www.ruhr-uni-bochum.de/en/information-collection-personal-data-application-process. We are looking forward to receiving your application with the specification ANR: 800 until 08.08.2022, send by e-mail to the following address: laurenz.wiskott at rub.de extent: full-time duration: temporary beginning: as soon as possible application dateline: 08.08.2022 The Ruhr-Universit?t Bochum is one of Germany?s leading research universities, addressing the whole range of academic disciplines. A highly dynamic setting enables researchers and students to work across the traditional boundaries of academic subjects and faculties. To create knowledge networks within and beyond the university is RUB?s declared aim. The Ruhr-Universit?t Bochum stands for diversity and equal opportunities. For this reason, we favour a working environment composed of heterogeneous teams, and seek to promote the careers of individuals who are underrepresented in our respective professional areas. The Ruhr-Universit?t Bochum expressly requests job applications from women. In areas in which they are underrepresented they will be given preference in the case of equivalent qualifications with male candidates. Applications from individuals with disabilities are most welcome. __________________________________________________________________________ Prof. Dr. Laurenz Wiskott room: NB 3/29 Institut f?r Neuroinformatik phone: +49 234 32-27997 Ruhr Universit?t Bochum fax: +49 234 32-14210 D-44780 Bochum, Germany https://www.ini.rub.de/PEOPLE/wiskott/ laurenz.wiskott at rub.de __________________________________________________________________________ From jonizhong at msn.com Fri Jul 22 00:47:53 2022 From: jonizhong at msn.com (Joni Zhong) Date: Fri, 22 Jul 2022 04:47:53 +0000 Subject: Connectionists: [CFP] IEEE Transactions on Cognitive and Developmental Systems Special Issue on Movement Sciences in Cognitive Systems Message-ID: ================================================================================= Call for Papers: IEEE Transactions on Cognitive and Developmental Systems Special Issue on Movement Sciences in Cognitive Systems ================================================================================= Deadline for submission: Jan 6, 2023 https://cis.ieee.org/images/files/Documents/call-for-special-issues/Movement_Sciences_in_Cognitive_Systems_CFP_2022.pdf Special issue on Movement Sciences in Cognitive Systems - CFP3 [CFP] Special issue on Movement Sciences in Cognitive Systems IEEE Transactions on Cognitive and Developmental Systems Over the past decades, the robotics communities have created computationally efficient mathematical tools for cis.ieee.org Over the past decades, the robotics communities have created computationally efficient mathematical tools for the study, simulation, and optimization of movements of articulated bodies. These techniques are being applied to increasingly complicated mechanical structures such as prosthetics and artificial systems in rehabilitation and sports training. Such formalism for quantitative biological motion analysis and synthesis is enabling broader applications in fields such as medical diagnosis, physical training monitoring and feedback, animation, ergonomic analysis and design, and assistive robots and devices. Learning and adapting to changing environments are challenging for cognitive systems. It is conceivable that the biological principles that underpin these adaptive and learning behaviours in animals or humans will inspire new robotic technology and the design of cognitive systems. To enable more agile and natural behaviours of these systems, it is useful to observe and analyze the mechanisms of the perception and action of biological systems. Complex brain circuitry is responsible for the movement, which ultimately controls muscle contraction. In humans and other mammals, the organization of neural structures including the spinal cord, cerebral cortex, basal ganglia, and cerebellum may be a key role to allow the adaptation and flexibility of movements This special issue on movement sciences in cognitive systems is primarily concerned with the implications of different computational aspects of movement sciences in developing intelligent systems. In cognitive systems, results and methodologies created from human or animal motion analysis in domains such as biomechanics, neuroscience, computer graphics, and computer vision are rapidly being employed in different themes, such as rehabilitation robots, ergonomic designs and assistive robots. This special issue will highlight the scientific findings from movement neuroscience, algorithms developed for the analysis, simulation, and optimization of articulated body movement and their applications in learning and interpreting complex structures or movements in cognitive systems. The primary list of topics (but is not limited to): Movement analysis in cognitive systems; Movement synthesis in cognitive systems; Sensory-motor learning in cognitive systems; Goal-directed movements in cognitive systems; Environmental-driven scaffolding; Movement neurosciences-inspired control; Modelling of human and animal behaviour movements; Reinforcement learning and deep reinforcement learning in movement control; Biomechanical analysis in cognitive systems; Interactive and affective motion design in cognitive systems Submission: Manuscripts should be prepared according to the guidelines in ?Submission Guidelines? of the IEEE Transactions on Cognitive and Developmental Systems in https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=7274989. IEEE Transactions on Cognitive and Developmental Systems | IEEE Xplore IEEE Transactions on Cognitive and Developmental Systems. null | IEEE Xplore ieeexplore.ieee.org Submissions should be done through the journal submission website: https://mc.manuscriptcentral.com/tcds-ieee, by selecting the Manuscript Type of ?Movement Sciences in Cognitive Systems? and clearly marking ?Movement Sciences in Cognitive Systems? in the comments to the Editor-in-Chief. Submitted papers will be reviewed by domain experts. Submission of a manuscript implies that it is the authors? original unpublished work and is not being submitted for possible publication elsewhere. Call for papers: July 2022; Paper submission deadline: 6 January 2023 Notification of acceptance: March 2023; Deadline revised versions: 31st May 2023 Final notification: 30th June 2023; Camera-ready deadline: 31st July 2023 Expected publication date: September 2023 Guest Editors Junpei Zhong, The Hong Kong Polytechnic University, Hong Kong; Ran Dong, Tokyo University of Technology, Japan; Soichiro Ikuno, Tokyo University of Technology, Japan; Yanan Li, University of Sussex, UK Chenguang Yang, University of the West of England, UK Contact: joni.zhong at polyu.edu.hk -------------- next part -------------- An HTML attachment was scrubbed... URL: From chunzhiyi at hit.edu.cn Fri Jul 22 06:15:03 2022 From: chunzhiyi at hit.edu.cn (=?UTF-8?B?6KGj5rez5qSN?=) Date: Fri, 22 Jul 2022 18:15:03 +0800 (GMT+08:00) Subject: Connectionists: [journals] CFP: "On The Bridge of Robotics and Human Motor System" Message-ID: <38938578.1184.18225674e35.Coremail.chunzhiyi@hit.edu.cn> Dear Colleagues, We would like to announce a new article collection for "On The Bridge of Robotics and Human Motor System" in Frontiers in Human Neuroscience. Human motor system executes the motor command from high-order regions, rigorously controls the musculoskeletal system and learns to adapt to novel stimuli of dynamic environments. Due to the highly redundant characteristics of the motor system, understanding the optimal and individualized principles of motor control and motor learning is still a scientific challenge. Robots interacting with or mimicking human motor systems through a completely controlled manner provide a novel perspective for studying the human motor system. On one hand, wearable robots, such as exoskeletons, prostheses and artificial limbs, closely interact with human, alter biomechanical, moto-neuronal, motor cortical and subjective preference aspects, thus fostering new homeostasis of human motor system. The alternation and adaptation of motor systems under such human-robot interactions lead to a well-controlled treatment and can offer novel viewing angles. On the other hand, biomimetic and bio-inspired robots that mimic human motor control or motor learning, from the perspectives of musculoskeletal actuation, exploration and exploitation-balanced motor decision and neuromechanical control of motions, reflect the inner principles of human motor system through comparison studies. Despite the advancement of robotics greatly enabling the potential of such studies, research under this viewpoint is still rare. The topics like inner beliefs, neuromechanical principles and cognitive aspects of motor learning are still left to be understood. The complexity of the human motor system and the rapidly developed robotic technologies and their combined promising usages in neuroscience, diagnostics, rehabilitation and augmentation highlight the necessity of our focus on bridging the gap between robotics and the human motor system study. The Research Topic aims to summarize the recent development of the robot-assisted studies of human motor systems and the motor principle-inspired robotic design. We welcome submissions in the form of original research, systematic reviews, method articles, and perspective articles. Areas of focus include but are not limited to: ? Biomechanical adaptation under human-robot interactions ? Simulation of neuromechanical control ? Computational modelling ? Bio-inspired design of robots ? Motor control and motor learning ? Exploration and exploitation that follow motor decision principles ? Reinforcement learning-based modeling ? Cognitive and cortical alternation under human-robot interactions ? Moto-neuronal aspects of wearable robots Prof. Dr. Feng Jiang Prof. Dr. Seungmin Rho Dr.Hang Su Dr. Chunzhi Yi Guest Editors You can find the website of the SI at https://www.frontiersin.org/research-topics/41086/on-the-bridge-of-robotics-and-human-motor-system Best, Chunzhi ------------------------------ YI Chunzhi Ph.D., Assistant Professor Intelligent Human-Machine Engineering, Department of Biomedical Engineering School of Medicine and Health, Harbin Institute of Technology No.2 Yikuang St. Nangang Dist. Harbin, China? 150001 Homepage: http://homepage.hit.edu.cn/yichunzhi From r.gayler at gmail.com Fri Jul 22 06:26:33 2022 From: r.gayler at gmail.com (Ross Gayler) Date: Fri, 22 Jul 2022 20:26:33 +1000 Subject: Connectionists: weight guessing quickly solves n-bit parity In-Reply-To: <76994187-BC20-4B04-A818-85CD025674D8@tecnico.ulisboa.pt> References: <201B5241-E8FC-49D0-8EE4-88964E0E8B8A@nyu.edu> <200D0D12-795D-4ED7-8F26-3E032E209997@nyu.edu> <38F9FFC3-EA3E-4494-8E2E-324AC81FCC97@supsi.ch> <76994187-BC20-4B04-A818-85CD025674D8@tecnico.ulisboa.pt> Message-ID: Andreas Wichert has mentioned Dave Touretzky's 1998 query to the connectionists mailing list: "Is connectionist symbol processing dead?" The responses to that query were collected and published in Neural Computing Surveys, which doesn't exist anymore. However, the paper is still available from the WaybackMachine: https://web.archive.org/web/20170706013814/ftp://ftp.icsi.berkeley.edu/pub/ai/jagota/vol2_1.pdf Cheers Ross On Thu, 21 Jul 2022 at 20:31, Andrzej Wichert < andreas.wichert at tecnico.ulisboa.pt> wrote: > Dear Juergen, > > Symbols do not, by themselves, represent any utilizable knowledge, they > cannot be used for a definition of similarity criteria between themselves. > The use of symbols in algorithms which imitate human intelligent behavior > led to the famous physical symbol system hypothesis \index{physical symbol > system hypothesis} by Newell and Simon (1976) ``The necessary and > sufficient condition for a physical system to exhibit intelligence is that > it be a physical symbol system.'' Symbols are not present in the world; > they are the constructs of a human mind and simplify the process of > representation used in communication and problem solving. > > A Turing machine can simulate any algorithm, the same with RNN, but they > do not explain the human problem solving. > > There is a difference, for example human problem solving can be described > by production systems. The most successful model is the SOAR architecture. > Such systems were very successful, they learn and can give you an > explanation for their doing. > > in August 1998 Dave Touretzky asked on the connectionistic e-mailing list: > ``Is connectionist symbol processing dead??' > > > From ml-connectionists-request$@$mlist-1.sp.cs.cmu.edu Tue Aug 11 > 17:35:10 1998 > From: Dave$\_$Touretzky$@$cs.cmu.edu > To: connectionists$@$cs.cmu.edu > Subject: Connectionist symbol processing: any progress? > Date: Tue, 11 Aug 1998 $03:34:27$ -0400 > > I'd like to start a debate on the current state of connectionist symbol > processing? Is it dead? Or does progress continue? ... People had > gotten > some interesting effects with localist networks, by doing spreading > activation > and a simple form of constraint satisfaction.... This approach does not > create > new structure on the fly, or deal with structured representations or > variable > binding. Those localist networks that did attempt to implement variable > binding > did so in a discrete, symbolic way that did not advance the parallel > constraint > satisfaction/heuristic reasoning agenda of earlier spreading activation > research. ... So I concluded that connectionist symbol processing had > reached > a plateau, and further progress would have to await some revolutionary new > insight about representations. ... The problems of structured > representations > and variable binding have remained unsolved. No one is trying to build > distributed connectionist reasoning systems any more, like the > connectionist > production system I built with Geoff Hinton... > > > 24 years past and not much progress was done. It seems that the progress > is only related to pure brute force of computers, but not much insight > beside a wishful thinking. The whole DL movements stops the progress in > understanding how the brain works. We need some new fresh ideas beside > error minimization. > > I think one of the main problems is the publish or perish altitude, the > famous Google impact factor. One does not care what some one is doing, one > just checks his Google impact factor.. This is like the saying, eat more > shit, one million flies cannot be wrong. > Like some mathematician said, computer science is not science at all, but > it force us to follow its ideas. > > Andreas > > > > > -------------------------------------------------------------------------------------------------- > Prof. Auxiliar Andreas Wichert > > http://web.tecnico.ulisboa.pt/andreas.wichert/ > - > https://www.amazon.com/author/andreaswichert > > Instituto Superior T?cnico - Universidade de Lisboa > Campus IST-Taguspark > Avenida Professor Cavaco Silva Phone: +351 214233231 > 2744-016 Porto Salvo, Portugal > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From jorgecbalmeida at gmail.com Fri Jul 22 06:48:06 2022 From: jorgecbalmeida at gmail.com (Jorge Almeida) Date: Fri, 22 Jul 2022 11:48:06 +0100 Subject: Connectionists: Post-doctoral position in Portugal in Cognitive/Computational Neuroscience - Application deadline August 3rd Message-ID: Apologies for cross-posting. The Proaction Laboratory (Jorge Almeida?s Lab; proactionlab.fpce.uc.pt) at the University of Coimbra (www.uc.pt), Portugal is looking for 1 Post-Doctoral Researcher (Cognitive/computational Neuroscience with interests in dimensionality reduction, SOMs and object processing) to work on an ERC Starting Grant project (ContentMAP; https://cordis.europa.eu/project/id/802553) on the neural organization of object knowledge. In this project we are exploring how complex information is topographically organized in the brain using fMRI and state of the art analytical techniques, as well as computational approaches, and neuromodulation. We strongly and particularly encourage applications from women, and from underrepresented groups in academia. General Requirements for the positions: 1. Candidates should have a PhD in Psychology, Cognitive Neuroscience, Computer Science, Computational Neuroscience or any other related field as long as their work relates to the specific profiles below. 2. They should already have their diplomas (so that we can start the process of recognition in Portugal, which is a necessary step for hiring ? please go here to start the process of recognizing your diploma). 3. The candidate should be a recent graduate ? the date in their PhD diploma should not be before January 2020. 4. Interest in object recognition and neural representation. 5. Very good English (oral and written) communicative skills are necessary. Specific requirements for each profile: Cognitive Neuroscience/fMRI: 1. Excellent understanding of and experience with fMRI and data analysis, and specifically with MVPA and Representational Similarity Analysis is required. 2. Strong programming skills (matlab, python, etc.) are also a requirement. Mapping/SOMs/Computational Neuroscience: 1. Excellent understanding of Machine Learning especially applied to fMRI 2. Command of computational approaches for mapping and dimensionality reduction especially applied to neuroimaging (e.g., SOM ? self-organizing maps). Salary and duration: The position will start as soon as possible and finish in January 2024. The position involves no formal teaching (unless the candidate wants to). It does involve, however, lab mentoring. The salary is extremely competitive ? 2200 euros per month net value (annual 26400 net value). This value is on par with the average salaries at top American institutions, as well as London, Paris, etc. However, the cost of living in Portugal (and particularly in Coimbra) is much lower. According to Numbeo, 2400 euros would be equivalent to 4850 pounds in London/4800 euros in Paris/6615 USD in Boston/6210 USD in Los Angeles/ 8330 USD in New York. Working conditions: The researcher will work directly with Jorge Almeida in Coimbra. The researcher will also be encouraged to develop her/his own projects and look for additional funding so that the stay can be extended. We have access to 2 3T MRI scanner with a 32-channel coil, to tDCS with neuronavigation, and to a fully set psychophysics lab. We have EEG and eyetracking on site. We also have access, through other collaborations, to a 7T scanner. Finally, 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. How can I apply: Applications are open until August 3rd. The interested candidates should email Jorge Almeida if they have any question. Please send an email (jorgealmeida at fpce.uc.pt) with the subject ?Post-doc position under ERC - ContentMAP?. To apply, please follow this link: https://apply.uc.pt/IT137-22-045 link to the call is here: https://storage.fw.uc.pt/fp/56vu7xubmo2tqff299mh7365kqj7uirt/L822593_Im0757en_Edital_VF.pdf You should send the following documents: 1. The Curriculum Vitae with a list of publications, 2. 2 Reference letters 3. A motivation letter with a short description of your experience in the field and how you fulfill the requirements (fit with the position). -------------- next part -------------- An HTML attachment was scrubbed... URL: From david at irdta.eu Sat Jul 23 10:39:48 2022 From: david at irdta.eu (David Silva - IRDTA) Date: Sat, 23 Jul 2022 16:39:48 +0200 (CEST) Subject: Connectionists: DeepLearn 2023 Winter: early registration August 1st Message-ID: <1532693560.258184.1658587188542@webmail.strato.com> ****************************************************************** 8th INTERNATIONAL SCHOOL ON DEEP LEARNING DeepLearn 2023 Winter Bournemouth, UK January 16-20, 2023 https://irdta.eu/deeplearn/2023wi/ *********** Co-organized by: Department of Computing and Informatics Bournemouth University Institute for Research Development, Training and Advice ? IRDTA Brussels/London ****************************************************************** Early registration: August 1st, 2022 ****************************************************************** SCOPE: DeepLearn 2022 Winter 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, Las Palmas de Gran Canaria and Lule?. Deep learning is a branch of artificial intelligence covering a spectrum of current exciting 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 2023 Winter 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 2023 Winter will take place in Bournemouth, a coastal resort town on the south coast of England. The venue will be: Talbot Campus Bournemouth University https://www.bournemouth.ac.uk/about/contact-us/directions-maps/directions-our-talbot-campus 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: Yi Ma (University of California, Berkeley), CTRL: Closed-Loop Data Transcription via Rate Reduction Daphna Weinshall (Hebrew University of Jerusalem), Curriculum Learning in Deep Networks Eric P. Xing (Carnegie Mellon University), It Is Time for Deep Learning to Understand Its Expense Bills PROFESSORS AND COURSES: Mohammed Bennamoun (University of Western Australia), [intermediate/advanced] Deep Learning for 3D Vision Matias Carrasco Kind (University of Illinois, Urbana-Champaign), [intermediate] Anomaly Detection Nitesh Chawla (University of Notre Dame), [introductory/intermediate] Graph Representation Learning Seungjin Choi (Intellicode), [introductory/intermediate] Bayesian Optimization over Continuous, Discrete, or Hybrid Spaces Sumit Chopra (New York University), [intermediate] Deep Learning in Healthcare Luc De Raedt (KU Leuven), [introductory/intermediate] Statistical Relational and Neurosymbolic AI Marco Duarte (University of Massachusetts, Amherst), [introductory/intermediate] Explainable Machine Learning Jo?o Gama (University of Porto), [introductory] Learning from Data Streams: Challenges, Issues, and Opportunities Claus Horn (Zurich University of Applied Sciences), [intermediate] Deep Learning for Biotechnology Zhiting Hu (University of California, San Diego) & Eric P. Xing (Carnegie Mellon University), [intermediate/advanced] A "Standard Model" for Machine Learning with All Experiences Nathalie Japkowicz (American University), [intermediate/advanced] Learning from Class Imbalances Gregor Kasieczka (University of Hamburg), [introductory/intermediate] Deep Learning Fundamental Physics: Rare Signals, Unsupervised Anomaly Detection, and Generative Models Karen Livescu (Toyota Technological Institute at Chicago), [intermediate/advanced] Speech Processing: Automatic Speech Recognition and beyond David McAllester (Toyota Technological Institute at Chicago), [intermediate/advanced] Information Theory for Deep Learning Dhabaleswar K. Panda (Ohio State University), [intermediate] Exploiting High-performance Computing for Deep Learning: Why and How? Fabio Roli (University of Cagliari), [introductory/intermediate] Adversarial Machine Learning Bracha Shapira (Ben-Gurion University of the Negev), [introductory/intermediate] Recommender Systems Richa Singh (Indian Institute of Technology Jodhpur), [introductory/intermediate] Trusted AI Kunal Talwar (Apple), [introductory/intermediate] Foundations of Differentially Private Learning Tinne Tuytelaars (KU Leuven), [introductory/intermediate] Continual Learning in Deep Neural Networks Lyle Ungar (University of Pennsylvania), [intermediate] Natural Language Processing using Deep Learning Bram van Ginneken (Radboud University Medical Center), [introductory/intermediate] Deep Learning for Medical Image Analysis Yu-Dong Zhang (University of Leicester), [introductory/intermediate] Convolutional Neural Networks and Their Applications to COVID-19 Diagnosis 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 January 8, 2023. 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 January 8, 2023. 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 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 January 8, 2023. ORGANIZING COMMITTEE: Rashid Bakirov (Bournemouth, local co-chair) Marcin Budka (Bournemouth) Vegard Engen (Bournemouth) Nan Jiang (Bournemouth, local co-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/2023wi/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 are available at https://irdta.eu/deeplearn/2023wi/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: Bournemouth University Rovira i Virgili University Institute for Research Development, Training and Advice ? IRDTA, Brussels/London -------------- next part -------------- An HTML attachment was scrubbed... URL: From d.goodman at imperial.ac.uk Fri Jul 22 10:37:47 2022 From: d.goodman at imperial.ac.uk (Goodman, Daniel F M) Date: Fri, 22 Jul 2022 14:37:47 +0000 Subject: Connectionists: Announcement: spiking neural networks workshop Nov 9-10, deadline Sept 28 Message-ID: We will be holding the third incarnation of our annual workshop SNUFA on spiking neural networks. This online workshop brings together hundreds of researchers in the fields of computational neuroscience, machine learning, and neuromorphic engineering to present their work and discuss ways of translating these findings into a better understanding of neural circuits. Topics include artificial and biologically plausible learning algorithms and the dissection of trained spiking circuits toward understanding neural processing. We have a manageable number of talks with ample time for discussions. Key information * The workshop will take place 9-10 November in the afternoons (central europe timezone). * Registration is free but mandatory. Click here to register. * Abstract submission deadline is 28 September. Click here to submit. * Final decisions by 12 October. Invited speakers * Charlotte Frenkel, TU Delft * Priya Panda, Yale * Yiota Poirazi, Institute of Molecular Biology and Biotechnology (IMBB) * Yonghong Tian, Peking University Abstracts will be made publicly available at the end of the abstract submissions deadline for blinded public comments and ratings. We will select the most highly rated abstracts for contributed talks, subject to maintaining a balance between the different fields of, broadly speaking, neuroscience, computer science and neuromorphic engineering. Abstracts not selected for a talk will be presented as posters, and there is an option to submit an abstract directly for a poster and not a talk if you prefer. The workshop will also feature a debate with all speakers invited. The subject of the debate is yet to be decided - please do get in touch if you have a fun idea! To get a feel for the workshop, take a look at our previous incarnations on YouTube for 2020 and 2021. Many thanks, and hope to see you there, The SNUFA executive committee, Dan Goodman Friedemann Zenke Katie Schuman Tim Masquelier -------------- next part -------------- An HTML attachment was scrubbed... URL: From stevensequeira92 at hotmail.com Fri Jul 22 13:39:12 2022 From: stevensequeira92 at hotmail.com (steven gouveia) Date: Fri, 22 Jul 2022 17:39:12 +0000 Subject: Connectionists: =?iso-8859-1?q?Online_Course_n=BA8_-_Consciousnes?= =?iso-8859-1?q?s_=28with_Sir_Roger_Penrose_-_Nobel_Prize_in_Physics_2020?= =?iso-8859-1?q?=2C_David_Chalmers_=26_Christof_Koch=29?= In-Reply-To: References: Message-ID: Dear Colleagues, The slots for Online Course n?8 - Consciousness are officially open, open to the general public (students, researchers, professors) and those curious about the nature of consciousness. The Course will have 3 sessions with 3 guest professors: (1) Orch-OR Theory of Consciousness, with Sir Roger Penrose (Nobel Prize in Physics 2020); (2) Virtual Reality and Consciousness, with David Chalmers; (3) Neural Correlates of Consciousness and Integrated Information Theory, with Christof Koch. All information here: https://stevensgouveia.weebly.com/course-8.html Any questions, I will be at your disposal. Thank you very much, Steven S. Gouveia Ph.D. (University of Minho) PostDoc Research Fellow (University of Ottawa) https://stevensgouveia.weebly.com -------------- next part -------------- An HTML attachment was scrubbed... URL: From fellous at arizona.edu Fri Jul 22 14:07:09 2022 From: fellous at arizona.edu (Fellous, Jean-Marc - (fellous)) Date: Fri, 22 Jul 2022 18:07:09 +0000 Subject: Connectionists: [EXT] If you believe in your work ... In-Reply-To: <9ea52c4c-2b66-e8bd-b494-5dc4ab7ebc8d@susaro.com> References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> <9ea52c4c-2b66-e8bd-b494-5dc4ab7ebc8d@susaro.com> Message-ID: @ Richard, Frederick, Tsvi: Points taken. Regarding funding: Delegating the funding decision to one person is probably dangerous. Note that technically, this is the case at NIH institutes (the director is the only one who makes the final decision, but in practice, that decision is built on the basis of peer review and program officer inputs), and similar at NSF (the PO has more authority in funding, but again, within limits). One other possibility for allowing innovative ideas to push through the ?politics? and ?settled culture? might be to set aside (say) 20% of the fund to randomly fund ?sound? short (2 years?) proposals that were placed far from funding threshold by reviewers. Injecting ?noise? so to speak to get out of local minima? Regarding publications: There are so many journals out there, and the arXiv options. The issue of course is integrity, trust, reproducibility etc. To some extent this is what peer review attempts to address, but there is no guarantee. May be a hybrid system, where researchers can post their work in ArXiv type venue, but with minimal initial ?sanity? checks (i.e. moderated, not reviewed), followed by voluntary or invited anonymous reviews? The quality/strength of the article would be assessed multi-dimensionally by the number of reviews/responses, number of reads, number of downloads, number of citations, availability of code/data, posting of negative results? etc. Can?t help but think we could be on the verge of a paradigm shift in publications and funding models? we need new ideas, and the courage to try them out! Best, Jean-Marc From: Connectionists On Behalf Of Richard Loosemore Sent: Tuesday, July 19, 2022 10:40 AM To: connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: [EXT] If you believe in your work ... External Email Jean-Marc, The problem is systemic; it goes down into the very roots of modern "science". The only solution that $20m could buy would be: 1) An institute run by someone with ethical principles, who would use the money to attract further funding until it could actually take on board researchers with creative ideas and ethical principles, and then free them from the yoke of publish-crap-in-quantity-or-perish. 2) An AI/Cognitive system development tool that would allow people to build and explore complex cognitive systems without being shackled to one particular architecture (like deep learning and its many descendents). A propos of (2) that is one thing I proposed in a (rejected) grant proposal. It would have cost $6.4m. Best, Richard -- Richard Loosemore Cornell University ... rpl72 at cornell.edu On 7/19/22 11:31 AM, Fellous, Jean-Marc - (fellous) wrote: Assuming there are funders on the list, and funding-related people, including program officers (and believe or not, there are!): if you had $20M to invest in the sort of things we do on this list: how would we make things better? Can we brainstorm an alternative system that allows for innovating publications and effective funding? Jean-Marc ________________________________ From: Connectionists on behalf of Richard Loosemore Sent: Monday, July 18, 2022 1:28 PM To: connectionists at mailman.srv.cs.cmu.edu Subject: [EXT]Connectionists: If you believe in your work ... External Email On 7/17/22 11:52 AM, Grossberg, Stephen wrote: > ... if you believe in your work, and the criticisms of it are not valid, do not give up. ... > ... all criticisms by reviewers are valuable and should be taken into account in your revision. > Even if a reviewer's criticisms are, to your mind, wrong-headed, they represent the > viewpoint of a more-than-usually-qualified reader who has given you the privilege > of taking enough time to read your article. Really? 1) I believe in my work, and the criticisms of it are not valid. I did not give up, and the net result of not giving up was ... nothing. 2) No reviewer who has ever commented on my work has shown the slightest sign that they understood anything in it. 3) Good plumbers are more than usually qualified in their field, and if one of those gave you the privilege of taking enough time to read your article and give nonsensical comments, would you pay any attention to their viewpoint? ** - ** I have spent my career fighting against this system, to no avail. I have watched charlatans bamboozle the crowd with pointless mathematics, and get published. I have watched people use teams of subordinates to pump out streams of worthless papers that inflate their prestige. I have written grant proposals that were exquisitely tuned to the stated goal of the grant, and then watched as the grant money went to people whose proposals had only the faintest imaginable connection to the stated goal of the grant. ** - ** The quoted remarks, above, somehow distilled all of that history and left me shaking with rage at the stupidity. I have been a member of the Connectionists mailing list since the early 1990s, and before that I had been working on neural nets since 1980. No more. Best, Richard -- Richard Loosemore Cornell University ... rpl72 at cornell.edu -------------- next part -------------- An HTML attachment was scrubbed... URL: From Donald.Adjeroh at mail.wvu.edu Sun Jul 24 02:22:31 2022 From: Donald.Adjeroh at mail.wvu.edu (Donald Adjeroh) Date: Sun, 24 Jul 2022 06:22:31 +0000 Subject: Connectionists: Last call: Summer School on AI & Smart Health & Mini-workshop series (July 25-27, 2022) In-Reply-To: References: , , , , , , , , , , , , , , Message-ID: Sorry for the multiple posts. This is just a reminder on the summer school on AI & Smart Health, starting tomorrow (Monday). Please visit the website: https://community.wvu.edu/~daadjeroh/projects/cresh/activity/summersch2022/ for more information on the summer school and mini-workshop series, and to register. ________________________________ From: Donald Adjeroh Sent: Tuesday, July 12, 2022 11:27 AM To: DON at CSEE.WVU.EDU Subject: Call for Participation: Summer School on AI & Smart Health & Mini-workshop series (July 25-27, 2022) We apologize if you receive multiple copies. As part of the NSF Track-2 project on "Multi-Scale Integrative Approach to Digital Health: Collaborative Research and Education in Smart Health in West Virginia and Arkansas" (http://community.wvu.edu/~daadjeroh/projects/cresh/ ) and the NSF BridgesDH NRT on "Bridges in Digital Health" (https://community.wvu.edu/~daadjeroh/projects/bridges/ ) we would like to announce a Summer School on AI & Smart Health, and related mini-workshop series. The summer school and mini-workshop series will be from July 25 - 27, 2022. This will be held remotely as a virtual event. The activities are free, with no registration fee. Please visit the website: https://community.wvu.edu/~daadjeroh/projects/cresh/activity/summersch2022/ for more information on the summer school and mini-workshop series, and to register. Please help us to forward this to students, colleagues, and others that may be interested. Regards, Don Adjeroh, PhD Professor and Associate Chair Graduate Coordinator of Computer Science West Virginia University Morgantown, WV 26506 http://community.wvu.edu/~daadjeroh/ Tel: 304-293-9681 Fax: 304-293-8602 -------------- next part -------------- An HTML attachment was scrubbed... URL: From valvilraman at yahoo.co.in Sat Jul 23 14:57:34 2022 From: valvilraman at yahoo.co.in (Anand Ramamoorthy) Date: Sat, 23 Jul 2022 18:57:34 +0000 (UTC) Subject: Connectionists: [EXT] If you believe in your work ... In-Reply-To: References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> <9ea52c4c-2b66-e8bd-b494-5dc4ab7ebc8d@susaro.com> Message-ID: <384795035.1010451.1658602654376@mail.yahoo.com> Hi Jean-Marc, Richard, Tsvi et al.,????????????????????????????????????????????????????????????? Re: publications, The proposed multidimensional assessment method does sound interesting but it inherits some of the problems we encounter with the current model. Looking at things like download volume, "number of reads" (really hard to tell what has been read, and whether it has been read carefully), number of citations can (and I predict will) elicit behaviours aimed at maximising those numbers, diminishing the reliability of the estimate when it comes to ascertaining scientific value. Number of active people in a given field or subfield, cliques that cite each other aggressively, can all inflate the estimate - this is a social problem within scientific research that requires attention. Popularity contests should not coarsen a discourse which is all about the search for truth. Re: paradigm shift in funding models & publications? - long overdue. All the best! ?????? Anand Ramamoorthy On Saturday, 23 July 2022 at 17:39:31 BST, Fellous, Jean-Marc - (fellous) wrote: ? @ Richard, Frederick, Tsvi: ? Points taken. Regarding funding: Delegating the funding decision to one person is probably dangerous. Note that technically, this is the case at NIH institutes (the director is the only one who makes the final decision, but in practice, that decision is built on the basis of peer review and program officer inputs), and similar at NSF (the PO has more authority in funding, but again, within limits). One other possibility for allowing innovative ideas to push through the ?politics? and ?settled culture? might be to set aside (say) 20% of the fund to randomly fund ?sound? short (2 years?) proposals that were placed far from funding threshold by reviewers. Injecting ?noise? so to speak to get out of local minima? Regarding publications: There are so many journals out there, and the arXiv options. The issue of course is integrity, trust, reproducibility etc. To some extent this is what peer review attempts to address, but there is no guarantee. May be a hybrid system, where researchers can post their work in ArXiv type venue, but with minimal initial ?sanity? checks (i.e. moderated, not reviewed), followed by voluntary or invited anonymous reviews? The quality/strength of the article would be assessed multi-dimensionally by the number of reviews/responses, number of reads, number of downloads, number of citations, availability of code/data, posting of negative results? etc. Can?t help but think we could be on the verge of a paradigm shift in publications and funding models? we need new ideas, and the courage to try them out! ? Best, Jean-Marc ? From: Connectionists On Behalf Of Richard Loosemore Sent: Tuesday, July 19, 2022 10:40 AM To: connectionists at mailman.srv.cs.cmu.edu Subject: Re: Connectionists: [EXT] If you believe in your work ... ? External Email Jean-Marc, The problem is systemic; it goes down into the very roots of modern "science". The only solution that $20m could buy would be: 1) An institute run by someone with ethical principles, who would use the money to attract further funding until it could actually take on board researchers with creative ideas and ethical principles, and then free them from the yoke of publish-crap-in-quantity-or-perish. 2) An AI/Cognitive system development tool that would allow people to build and explore complex cognitive systems without being shackled to one particular architecture (like deep learning and its many descendents). A propos of (2) that is one thing I proposed in a (rejected) grant proposal. It would have cost $6.4m. Best, Richard -- Richard Loosemore Cornell University ... rpl72 at cornell.edu On 7/19/22 11:31 AM, Fellous, Jean-Marc - (fellous) wrote: Assuming there are funders on the list, and funding-related people, including program officers (and believe or not, there are!): if you had $20M to invest in the sort of things we do on this list: how would we make things better? Can we brainstorm an alternative system that allows for innovating publications and effective funding? ? Jean-Marc From: Connectionists on behalf of Richard Loosemore Sent: Monday, July 18, 2022 1:28 PM To: connectionists at mailman.srv.cs.cmu.edu Subject: [EXT]Connectionists: If you believe in your work ... ? External Email On 7/17/22 11:52 AM, Grossberg, Stephen wrote: > ... if you believe in your work, and the criticisms of it are not valid, do not give up. ... > ... all criticisms by reviewers are valuable and should be taken into account in your revision. > Even if a reviewer's criticisms are, to your mind, wrong-headed, they represent the > viewpoint of a more-than-usually-qualified reader who has given you the privilege > of taking enough time to read your article. Really? 1) I believe in my work, and the criticisms of it are not valid.? I did not give up, and the net result of not giving up was ... nothing. 2) No reviewer who has ever commented on my work has shown the slightest sign that they understood anything in it. 3) Good plumbers are more than usually qualified in their field, and if one of those gave you the privilege of taking enough time to read your article and give nonsensical comments, would you pay any attention to their viewpoint? ** - ** I have spent my career fighting against this system, to no avail. I have watched charlatans bamboozle the crowd with pointless mathematics, and get published. I have watched people use teams of subordinates to pump out streams of worthless papers that inflate their prestige. I have written grant proposals that were exquisitely tuned to the stated goal of the grant, and then watched as the grant money went to people whose proposals had only the faintest imaginable connection to the stated goal of the grant. ** - ** The quoted remarks, above, somehow distilled all of that history and left me shaking with rage at the stupidity. I have been a member of the Connectionists mailing list since the early 1990s, and before that I had been working on neural nets since 1980. No more. ? Best, ? Richard --? Richard Loosemore Cornell University ... rpl72 at cornell.edu ? ? -------------- next part -------------- An HTML attachment was scrubbed... URL: From mtanveer at iiti.ac.in Mon Jul 25 02:48:32 2022 From: mtanveer at iiti.ac.in (M Tanveer) Date: Mon, 25 Jul 2022 12:18:32 +0530 Subject: Connectionists: ICONIP 2022 | Call for Papers | Deadline: July 28, 2022 In-Reply-To: References: Message-ID: CALL FOR PAPERS *Paper submission deadline extended to July 28, 2022 (Firm deadline)* 29th International Conference on Neural Information Processing (ICONIP 2022) November 22-26, 2022 New Delhi, India https://www.iconip2022.apnns.org/ Dear Colleague, We would like to invite you to submit your paper to the 29th International Conference on Neural Information Processing (ICONIP 2022), New Delhi, India, November 22-26, 2022. Updated flyer of the call for paper is attached with this email. The conference website: https://iconip2022.apnns.org/index.php Submission Page: https://easychair.org/conferences/?conf=iconip2022 Paper Submission: *July 28, 2022* Paper Notification Date: *Sept. 15, 2022* Camera Ready Submission: *Sept. 30, 2022* Thank you so much for your kind support in this conference. Kindly share among your contacts. If you have any queries, please contact us: iconip2022 at gmail.com. Sincerely, General Chairs - ICONIP 2022 ---------------------------------------------------------- Dr. M. Tanveer (General Chair - ICONIP 2022, IEEE CIS SS 2022) Associate Professor and Ramanujan Fellow Department of Mathematics Indian Institute of Technology Indore Email: mtanveer at iiti.ac.in Mobile: +91-9413259268 Homepage: http://iiti.ac.in/people/~mtanveer/ Associate Editor: IEEE TNNLS (IF: 14.25). Action Editor: Neural Networks, Elsevier (IF: 9.65). Associate Editor: Pattern Recognition, Elsevier (IF: 8.52). Associate Editor: Cognitive Computation, Springer (IF: 8.26). Board of Editors: Engineering Applications of AI, Elsevier (IF: 7.80). Associate Editor: Neurocomputing, Elsevier (IF: 5.78). Editorial Board: Applied Soft Computing, Elsevier (IF: 6.72). Associate Editor: International Journal of Machine Learning & Cybernetics (IF: 4.37). -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: CallForPapers-ICONIP 2022v3_final.pdf Type: application/pdf Size: 500290 bytes Desc: not available URL: From bei.xiao at gmail.com Sun Jul 24 14:07:27 2022 From: bei.xiao at gmail.com (bei.xiao at gmail.com) Date: Sun, 24 Jul 2022 20:07:27 +0200 Subject: Connectionists: The British Machine Vision Conference (BMVC) abstract deadline extended to July 25th 23:59 GMT Message-ID: The British Machine Vision Conference (BMVC) is one of the major international conferences on computer vision and related areas. It is organised by the British Machine Vision Association (BMVA). The 33rd BMVC will now be a hybrid event from 21st?24th November 2022. Our local in person meeting will be held at The Kia Oval (Home of Surrey County Cricket Club, https://events.kiaoval.com/). Authors are invited to submit full-length high-quality papers in image processing, computer vision, machine learning and related areas for BMVC 2022. Submitted papers will be refereed on their originality, presentation, empirical results, and quality of evaluation. Accepted papers will be included in the conference proceedings published and DOI-indexed by BMVA. Past proceedings can be found online: here . Please note that BMVC is a single-track meeting with oral and poster presentations. The abstract submission deadline is *extended to Monday July 25th 23:59 GMT (+15min buffer)*and the paper submission deadline is Friday 29th July 2022 (both 23:59, GMT). Submission instructions are available on the BMVC 2022 website . Submitted papers should not exceed 9 pages (references are excluded, but appendices are included). Topics include, but are not limited to: - 2D object recognition - 3D computer vision - 3D object recognition - Action and behavior recognition - Adversarial learning, adversarial attack and defense methods - Biometrics, face, gesture, body pose - Computational photography - Datasets and evaluation - Efficient training and inference methods for networks - Explainable AI, fairness, accountability, privacy, transparency and ethics in vision - Image and video retrieval - Image and video synthesis - Image classification - Low-level and physics-based vision - Machine learning architectures and formulations - Medical, biological and cell microscopy - Motion and tracking - Optimization and learning methods - Pose estimation - Representation learning - Scene analysis and understanding - Transfer, low-shot, semi- and un- supervised learning - Video analysis and understanding - Vision + language, vision + other modalities - Vision applications and systems, vision for robotics and autonomous vehicles - ?Brave new ideas? Papers submitted under the ?Brave new ideas? subject area are expected to move away from incremental benchmark gains. Proposed ideas should be radically different from the current strand of research or propose a novel problem. Reviewing process BMVC 2022 - Each paper will be reviewed by at least three reviewers. The primary AC will also provide a meta review, summarising the points that need to be addressed during the rebuttal phase. - The authors will have a period to produce a rebuttal to address the reviewers concerns. Due to the tight schedule, there will be no revision of the papers before the final camera ready submission. - The rebuttal will be handled by two ACs, a primary and a secondary, who will facilitate paper discussion and jointly make the recommendations. Conflicts will be jointly managed by the ACs and Program Chairs that will make the final decisions. Please Note: Due to the anticipated volume of papers for BMVC 2022 (based on recent year?s experience) there will be NO extension granted to the submission deadline. In keeping with conferences in the field (e.g. NeurIPS , CVPR ) and to cope with the increasing number of submissions, we ask that all authors be prepared to review papers and make use of a compulsory abstract submission deadline a week before the paper submission deadline. The CMT submission site will ask authors to acknowledge this commitment and failure to engage with the reviewing process might be grounds for rejection. Any queries to the Programme Chairs should be sent to pcs at bmvc2022.org. BMVC Organizing team https://bmvc2022.org/people/organisers/ -- Bei Xiao, PhD Associate Professor Computer Science & Center for Behavioral Neuroscience American University, Washington DC Homepage: https://sites.google.com/site/beixiao/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From achler at gmail.com Sun Jul 24 10:57:51 2022 From: achler at gmail.com (Tsvi Achler) Date: Sun, 24 Jul 2022 07:57:51 -0700 Subject: Connectionists: [EXT] If you believe in your work ... In-Reply-To: References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> <9ea52c4c-2b66-e8bd-b494-5dc4ab7ebc8d@susaro.com> Message-ID: Jean-Marc, The idea to let the editor decide was directed more to publications than funding. For funding decisions I think the key is to break the hegemony by primarily funding those who have little say in the system. That's a way to equalize the say and sources of ideas in the system. In that vein, one suggestion is to fund research primarily through junior researchers and give them the power, if a senior researcher needs large money then they to do it by a mechanism setup where they court junior researchers. Another suggestion for funding is let each member of staff in the funding committee have N favorite applications a year that they can choose bypassing the others' scrutiny, Indeed some random funding is a great idea as well. Follow my channel on YouTube "Updating Research" for more ideas: https://www.youtube.com/channel/UCbvTQ3lLVvikKaYnNH3kH3g BTW. this is how bad peer reviewed & politicized research can get: https://www.science.org/content/article/potential-fabrication-research-images-threatens-key-theory-alzheimers-disease -Tsvi On Sat, Jul 23, 2022 at 9:18 AM Fellous, Jean-Marc - (fellous) < fellous at arizona.edu> wrote: > > > @ Richard, Frederick, Tsvi: > > > > Points taken. > > Regarding funding: Delegating the funding decision to one person is > probably dangerous. Note that technically, this is the case at NIH > institutes (the director is the only one who makes the final decision, but > in practice, that decision is built on the basis of peer review and program > officer inputs), and similar at NSF (the PO has more authority in funding, > but again, within limits). One other possibility for allowing innovative > ideas to push through the ?politics? and ?settled culture? might be to set > aside (say) 20% of the fund to randomly fund ?sound? short (2 years?) > proposals that were placed far from funding threshold by reviewers. > Injecting ?noise? so to speak to get out of local minima? > > Regarding publications: There are so many journals out there, and the > arXiv options. The issue of course is integrity, trust, reproducibility > etc. To some extent this is what peer review attempts to address, but there > is no guarantee. May be a hybrid system, where researchers can post their > work in ArXiv type venue, but with minimal initial ?sanity? checks (i.e. > moderated, not reviewed), followed by voluntary or invited anonymous > reviews? The quality/strength of the article would be assessed > multi-dimensionally by the number of reviews/responses, number of reads, > number of downloads, number of citations, availability of code/data, > posting of negative results? etc. > > Can?t help but think we could be on the verge of a paradigm shift in > publications and funding models? we need new ideas, and the courage to try > them out! > > > > Best, > > Jean-Marc > > > > *From:* Connectionists *On > Behalf Of *Richard Loosemore > *Sent:* Tuesday, July 19, 2022 10:40 AM > *To:* connectionists at mailman.srv.cs.cmu.edu > *Subject:* Re: Connectionists: [EXT] If you believe in your work ... > > > > *External Email* > > > Jean-Marc, > > The problem is systemic; it goes down into the very roots of modern > "science". > > The only solution that $20m could buy would be: > > 1) An institute run by someone with ethical principles, who would use the > money to attract further funding until it could actually take on board > researchers with creative ideas and ethical principles, and then free them > from the yoke of publish-crap-in-quantity-or-perish. > > 2) An AI/Cognitive system development tool that would allow people to > build and explore complex cognitive systems without being shackled to one > particular architecture (like deep learning and its many descendents). > > A propos of (2) that is one thing I proposed in a (rejected) grant > proposal. It would have cost $6.4m. > > Best, > > Richard > > -- > Richard Loosemore > Cornell University > ... > rpl72 at cornell.edu > > > > > > > On 7/19/22 11:31 AM, Fellous, Jean-Marc - (fellous) wrote: > > Assuming there are funders on the list, and funding-related people, > including program officers (and believe or not, there are!): if you had > $20M to invest in the sort of things we do on this list: how would we make > things better? Can we brainstorm an alternative system that allows for > innovating publications and effective funding? > > > > Jean-Marc > ------------------------------ > > *From:* Connectionists > on behalf of Richard > Loosemore > *Sent:* Monday, July 18, 2022 1:28 PM > *To:* connectionists at mailman.srv.cs.cmu.edu > > > *Subject:* [EXT]Connectionists: If you believe in your work ... > > > > *External Email* > > > On 7/17/22 11:52 AM, Grossberg, Stephen wrote: > > > ... if you believe in your work, and the criticisms of it are not valid, > do not give up. ... > > > ... all criticisms by reviewers are valuable and should be taken into > account in your revision. > > > Even if a reviewer's criticisms are, to your mind, wrong-headed, they > represent the > > viewpoint of a more-than-usually-qualified reader who has given you the > privilege > > of taking enough time to read your article. > > Really? > > 1) I believe in my work, and the criticisms of it are not valid. I did > not give up, and the net result of not giving up was ... nothing. > > 2) No reviewer who has ever commented on my work has shown the slightest > sign that they understood anything in it. > > 3) Good plumbers are more than usually qualified in their field, and if > one of those gave you the privilege of taking enough time to read your > article and give nonsensical comments, would you pay any attention to their > viewpoint? > > ** - ** > > I have spent my career fighting against this system, to no avail. > > I have watched charlatans bamboozle the crowd with pointless mathematics, > and get published. > > I have watched people use teams of subordinates to pump out streams of > worthless papers that inflate their prestige. > > I have written grant proposals that were exquisitely tuned to the stated > goal of the grant, and then watched as the grant money went to people whose > proposals had only the faintest imaginable connection to the stated goal of > the grant. > > ** - ** > > The quoted remarks, above, somehow distilled all of that history and left > me shaking with rage at the stupidity. > > I have been a member of the Connectionists mailing list since the early > 1990s, and before that I had been working on neural nets since 1980. > > No more. > > > > Best, > > > > Richard > > -- > > Richard Loosemore > > Cornell University > > ... > > rpl72 at cornell.edu > > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: From zk240 at cam.ac.uk Sun Jul 24 19:19:51 2022 From: zk240 at cam.ac.uk (Zoe Kourtzi) Date: Sun, 24 Jul 2022 23:19:51 +0000 Subject: Connectionists: Assistant/Associate Professorship in Neuroinformatics, University of Cambridge, UK In-Reply-To: <909CAF92-27C3-4DAA-B8BA-44E9CFB9FB7C@cam.ac.uk> References: <909CAF92-27C3-4DAA-B8BA-44E9CFB9FB7C@cam.ac.uk> Message-ID: <22A0C0A6-A4F3-4E31-B3B2-B1784142FE93@cam.ac.uk> The Department of Psychology at the University of Cambridge, UK invites applications for a University Assistant/Associate Professorship in Neuroinformatics, focusing on the development and implementation of computational and data analytics approaches for understanding behaviour and cognition across the life span. https://www.jobs.cam.ac.uk/job/35988/ Application by August 29th -------------- next part -------------- An HTML attachment was scrubbed... URL: From evomusart at gmail.com Mon Jul 25 07:11:06 2022 From: evomusart at gmail.com (EvoMUSART) Date: Mon, 25 Jul 2022 12:11:06 +0100 Subject: Connectionists: =?utf-8?q?Call_for_Papers_=E2=80=94_EvoMUSART_202?= =?utf-8?q?3_=2812-14_April_2023=29?= Message-ID: ------------------------------------------------ Call for papers for the 12th International Conference on Artificial Intelligence in Music, Sound, Art and Design (EvoMUSART) ? Please distribute ? Apologies for cross-posting ------------------------------------------------ The 12th International Conference on Artificial Intelligence in Music, Sound, Art and Design (EvoMUSART) will take place on 12-14 April 2023, as part of the evo* event. EvoMUSART webpage: https://www.evostar.org/2023/evomusart Submission deadline: 1 November 2022 Conference: 12 ? 14 April 2023 EvoMUSART is a multidisciplinary conference that brings together researchers who are working on the application of Artificial Neural Networks, Evolutionary Computation, Swarm Intelligence, Cellular Automata, Alife, and other Artificial Intelligence techniques in creative and artist fields such as Visual Art, Music, Architecture, Video, Digital Games, Poetry, or Design. This conference gives researchers in the field the opportunity to promote, present and discuss ongoing work in the area. Submissions must be at most 16 pages long, in Springer LNCS format. Each submission must be anonymised for a double-blind review process. Accepted papers will be presented orally or as posters at the event and included in the EvoMUSART proceedings published by Springer Nature in a dedicated volume of the Lecture Notes in Computer Science series. Indicative topics include but are not limited to: * Systems that create drawings, images, animations, sculptures, poetry, text, designs, webpages, buildings, etc.; * Systems that create musical pieces, sounds, instruments, voices, sound effects, sound analysis, etc.; * Systems that create artefacts such as game content, architecture, furniture, based on aesthetic and/or functional criteria; * Systems that resort to artificial intelligence to perform the analysis of image, music, sound, sculpture, or some other types of artistic objects; * Systems in which artificial intelligence is used to promote the creativity of a human user; * Theories or models of computational aesthetics; * Computational models of emotional response, surprise, novelty; * Representation techniques for images, videos, music, etc.; * Surveys of the current state-of-the-art in the area; * New ways of integrating the user in the process (e.g. improvisation, co-creation, participation). More information on the submission process and the topics of EvoMUSART: https://www.evostar.org/2023/evomusart Flyer of EvoMUSART 2023: https://www.evostar.org/2023/flyers/evomusart Papers published in EvoMUSART: https://evomusart-index.dei.uc.pt We look forward to seeing you in EvoMUSART 2023! The EvoMUSART 2023 organisers Colin Johnson Nereida Rodr?guez-Fern?ndez S?rgio Rebelo (publication chair) -------------- next part -------------- An HTML attachment was scrubbed... URL: From stephan.chalup at newcastle.edu.au Tue Jul 26 00:00:51 2022 From: stephan.chalup at newcastle.edu.au (Stephan Chalup) Date: Tue, 26 Jul 2022 04:00:51 +0000 Subject: Connectionists: [jobs] PhD scholarship (international, Australia): Neural Principles of Dual Sensor Vision Systems in Machine Vision Applications Message-ID: PhD Scholarship: Neural Principles of Dual Sensor Vision Systems in Machine Vision Applications In this project, we will investigate abstract principles of two-stream vision systems in the context of deep neural networks and sensor fusion. The aim is to develop new real-world applications of machine vision using two-stream video or laser sensors. The project will be conducted in collaboration with industry partner 4Tel Pty Ltd in Newcastle, Australia. The PhD candidate will be part of an interdisciplinary team of researchers from computer science and industry. Please find further details at: https://www.newcastle.edu.au/study/research/phd-scholarships/phd-scholarships/neural-principles-of-dual-sensor-vision-systems-in-machine-vision-applications Closing Date: 1 October 2022 Contact: Professor Stephan Chalup School of Information and Physical Sciences The University of Newcastle, Callaghan NSW 2308, Australia Email: Stephan.Chalup at newcastle.edu.au -------------- next part -------------- An HTML attachment was scrubbed... URL: From timofte.radu at gmail.com Mon Jul 25 13:37:09 2022 From: timofte.radu at gmail.com (Radu Timofte) Date: Mon, 25 Jul 2022 19:37:09 +0200 Subject: Connectionists: [Last CFP] ECCV 2022 Advances in Image Manipulation (AIM) workshop and challenges [DEADLINE July 31] Message-ID: Apologies for cross-posting ******************************* CALL FOR PAPERS & CALL FOR PARTICIPANTS IN 8 CHALLENGES AIM: 4th Advances in Image Manipulation workshop and challenges on compressed/image/video super-resolution, learned ISP, reversed ISP, Instagram filter removal, Bokeh effect, depth estimation In conjunction with ECCV 2022, Tel-Aviv, Israel Website: https://data.vision.ee.ethz.ch/cvl/aim22/ Contact: radu.timofte at uni-wuerzburg.de TOPICS Papers addressing topics related to image/video manipulation, restoration and enhancement are invited. The topics include, but are not limited to: ? Image-to-image translation ? Video-to-video translation ? Image/video manipulation ? Perceptual manipulation ? Image/video generation and hallucination ? Image/video quality assessment ? Image/video semantic segmentation ? Saliency and gaze estimation ? Perceptual enhancement ? Multimodal translation ? Depth estimation ? Image/video inpainting ? Image/video deblurring ? Image/video denoising ? Image/video upsampling and super-resolution ? Image/video filtering ? Image/video de-hazing, de-raining, de-snowing, etc. ? Demosaicing ? Image/video compression ? Removal of artifacts, shadows, glare and reflections, etc. ? Image/video enhancement: brightening, color adjustment, sharpening, etc. ? Style transfer ? Hyperspectral imaging ? Underwater imaging ? Aerial and satellite imaging ? Methods robust to changing weather conditions / adverse outdoor conditions ? Image/video manipulation on mobile devices ? Image/video restoration and enhancement on mobile devices ? Studies and applications of the above. SUBMISSION A paper submission has to be in English, in pdf format, and at most 14 pages (excluding references) in single-column, ECCV style. The paper format must follow the same guidelines as for all ECCV 2022 submissions. The review process is double blind. Dual submission is not allowed. Submission site: https://cmt3.research.microsoft.com/AIMWC2022/ WORKSHOP DATES ? Submission Deadline for Early & Regular Papers: *July 31, **2022* (EXTENDED) ? Submission Deadline for Challenge Papers and Papers reviewed elsewhere: *August 10, **2022* ? Decisions: August 15, 2022 ? Camera Ready Deadline: August 22, 2022 AIM 2022 has the following associated challenges (ONGOING!): 1. Compressed Input Super-Resolution 2. Reversed ISP 3. Instagram Filter Removal 4. Video Super-Resolution (Evaluation platform: MediaTek Dimensity APU) - Powered by MediaTek 5. Image Super-Resolution (Eval. platform: Synaptics Dolphin NPU) - Powered by Synaptics 6. Learned Smartphone ISP (Eval. platform: Snapdragon Adreno GPU) - Powered by OPPO 7. Bokeh Effect Rendering (Eval. platform: ARM Mali GPU) - Powered by Huawei 8. Depth Estimation (Eval. platform: Raspberry Pi 4) - Powered by Raspberry Pi PARTICIPATION To learn more about the challenges and to participate: https://data.vision.ee.ethz.ch/cvl/aim22/ CHALLENGES DATES ? Release of train data: May 24, 2022 ? Validation server online: June 1, 2022 ? Competitions end: July 30, 2022 CONTACT Email: radu.timofte at uni-wuerzburg.de Website: https://data.vision.ee.ethz.ch/cvl/aim22/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From pattichi at ucy.ac.cy Tue Jul 26 05:32:05 2022 From: pattichi at ucy.ac.cy (Costas Pattichis) Date: Tue, 26 Jul 2022 09:32:05 +0000 Subject: Connectionists: [CAIP2021] AMALEA - APPLICATIONS OF MACHINE LEARNING, International Workshop, Cetraro, Italy, September 12-16, 2022 Message-ID: AMALEA - APPLICATIONS OF MACHINE LEARNING International Workshop, Cetraro, Italy, September 12-16, 2022 http://amalea.web.rug.nl/index.html With kind regards, Prof. Nicolai Petkov -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- _______________________________________________ CAIP2021 mailing list CAIP2021 at cs.ucy.ac.cy https://listserv.cs.ucy.ac.cy/mailman/listinfo/caip2021 From max.garagnani at gmail.com Tue Jul 26 07:20:57 2022 From: max.garagnani at gmail.com (Max Garagnani) Date: Tue, 26 Jul 2022 11:20:57 +0000 Subject: Connectionists: ** Last few places remaining ** -- MSc in Computational Cognitive Neuroscience, Goldsmiths (London, UK) Message-ID: -- Apologies for cross-posting -- MSc in COMPUTATIONAL COGNITIVE NEUROSCIENCE at Goldsmiths, University of London (UK) ============================================================================= **************************************** ONLY A FEW PLACES REMANING FOR 2022-23 ENTRY **************************************** This established Masters 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). Areas of application range from machine learning to brain-computer interfaces, to experimental and clinical research in computational / cognitive neuroscience. There are only a few places left for 2022-23 entry. 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, applications from 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 (see below). For samples of previous students' MSc projects, visit: https://coconeuro.com/index.php/student-projects/ LINKS WITH INDUSTRY: ==================== The programme benefits from ongoing collaborative partnerships with companies having headquarters in Europe, USA, and Japan. Carrying out your final research project with one of our industry partners will enable you to acquire cutting-edge skills much in demand on the job market, providing a ?fast track? route towards post-Masters internships and employment. Here are some examples of career pathways followed by some of our alumni, along with their feedback: https://coconeuro.com/index.php/alumni/ For any further information (including funding opportunities and fees), please visit: https://www.gold.ac.uk/pg/msc-computational-cognitive-neuroscience/ https://www.gold.ac.uk/pg/fees-funding/ 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 triesch at fias.uni-frankfurt.de Tue Jul 26 09:01:49 2022 From: triesch at fias.uni-frankfurt.de (Jochen Triesch) Date: Tue, 26 Jul 2022 15:01:49 +0200 Subject: Connectionists: Sergey Levine speaking on July 28 in Developing Minds global online lecture series Message-ID: <224806BA-2EA5-4CF1-B082-A8FE9DCEEF86@fias.uni-frankfurt.de> Dear colleagues, On July 28, the Developing Minds global online lecture series will feature Sergey Levine, UC Berkeley, USA: "From Reinforcement Learning to Embodied Learning? https://sites.google.com/view/developing-minds-series/home The live event will take place via zoom at: 09:00 PDT (Pacific Daylight Time) 16:00 UTC (Universal Coordinated Time) 18:00 CEST (Central European Summer Time) 01:00 JST, July 29 (Japan Standard Time) To participate please register here: https://sites.google.com/view/developing-minds-series/home Abstract Reinforcement learning provides one of the most widely studied abstractions for learning-based control. However, while the RL formalism is elegant and concise, real-world embodied learning problems (e.g., in robotics) deviate substantially from the most widely studied RL problem settings. The need to exist in a real physical environment challenges RL methods in terms of generalization, robustness, and capacity for lifelong learning -- all aspects of the RL problem that are often neglected in commonly studied benchmark problems. In this talk, I will discuss how we can devise a framework for learning-based control that is at its core focused on generalization, robustness, and continual adaptation. I will argue that effective utilization of previously collected experience, in combination with multi-task learning, represents one of the most promising paths for tackling these challenges, and present recent research in RL and robotics that studies this perspective. Bio Sergey Levine received a BS and MS in Computer Science from Stanford University in 2009, and a Ph.D. in Computer Science from Stanford University in 2014. He joined the faculty of the Department of Electrical Engineering and Computer Sciences at UC Berkeley in fall 2016. His work focuses on machine learning for decision making and control, with an emphasis on deep learning and reinforcement learning algorithms. Applications of his work include autonomous robots and vehicles, as well as computer vision and graphics. His research includes developing algorithms for end-to-end training of deep neural network policies that combine perception and control, scalable algorithms for inverse reinforcement learning, deep reinforcement learning algorithms, and more. Web: https://people.eecs.berkeley.edu/~svlevine/ The talk will also be recored and the recording made available via the web page: https://sites.google.com/view/developing-minds-series/home Stay healthy, Jochen Triesch -- Prof. Dr. Jochen Triesch Johanna Quandt Chair for Theoretical Life Sciences Frankfurt Institute for Advanced Studies and Goethe University Frankfurt http://fias.uni-frankfurt.de/~triesch/ Tel: +49 (0)69 798-47531 Fax: +49 (0)69 798-47611 From m.plumbley at surrey.ac.uk Tue Jul 26 10:18:52 2022 From: m.plumbley at surrey.ac.uk (Mark Plumbley) Date: Tue, 26 Jul 2022 14:18:52 +0000 Subject: Connectionists: JOB: Research Engineer in Sound Sensing, University of Surrey, UK Message-ID: Dear Connectionists, (with apologies for cross-posting) We are recruiting for a Research Engineer in Sound Sensing, a 6-month post 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 08 August 2022 (23:59 BST) Further details: https://jobs.surrey.ac.uk/025022-R Applications are invited for a Research Engineer (Research Fellow) in Sound Sensing, to work full-time for six months on an EPSRC-funded Fellowship project "AI for Sound" (https://ai4s.surrey.ac.uk/), to start September 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/025022-R Deadline: Monday 08 August 2022 (23:59 BST) For informal inquiries about the position, please contact Prof Mark Plumbley (m.plumbley at surrey.ac.uk). -- Prof Mark D Plumbley EPSRC Fellow in AI for Sound Professor of Signal Processing Centre for Vision, Speech and Signal Processing University of Surrey, Guildford, Surrey, GU2 7XH, UK Email: m.plumbley at surrey.ac.uk From m.biehl at rug.nl Tue Jul 26 11:20:40 2022 From: m.biehl at rug.nl (Michael Biehl) Date: Tue, 26 Jul 2022 17:20:40 +0200 Subject: Connectionists: AMALEA workshop 12-16 September Message-ID: On behalf of Prof. Dr. Nicolai Petkov (www.cs.rug.nl/~petkov) *International workshop on Applications of Machine Learning and Neural Networks* *AMALEA, 12-16 Sept 2022, Cetraro, Calabria, Italy* Dear colleagues and friends, There are a few rooms still available for AMALEA - APPLICATIONS OF MACHINE LEARNING International Workshop, Cetraro, Italy, September 12-16, 2022 http://amalea.web.rug.nl/index.html Venue: Grand Hotel San Michele on the West, Tyrrhenian sea coast of Southern Italy with a surrounding own land estate, golf course, private beach and a conference center. Photo gallery Aim: to provide a discussion forum for scientists from different disciplines where they can inspire and learn from each other, meet new people and form new research alliances. Format: a single track for talks from 10:00-12:30H and 17:00-19:30H in the conference center. Enough time for discussions outside the conference room: during breakfast on the terrace overlooking the Tyrrhenian sea, at lunch in the private beach restaurant, in the afternoon break on the private beach, at dinner on the hotel terrace and after dinner listening to the life piano music in the grand salon. Please visit the workshop website for further information and contact. -- --------------------------------------------------- Prof. Dr. Michael Biehl Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence P.O. Box 407, 9700 AK Groningen The Netherlands https://www.cs.rug.nl/~biehl m.biehl at rug.nl -------------- next part -------------- An HTML attachment was scrubbed... URL: From er.anubajaj at gmail.com Tue Jul 26 14:38:10 2022 From: er.anubajaj at gmail.com (Anu Bajaj) Date: Tue, 26 Jul 2022 14:38:10 -0400 Subject: Connectionists: 2nd CFP: 18th International Conference on Information Assurance and Security (IAS 2022) - Online - Springer Publication Message-ID: ** Second Call for Papers - please circulate this CFP to your colleagues and networks ** -- 18th International Conference on Information Assurance and Security (IAS 2022) -- http://www.mirlabs.org/ias22 http://www.mirlabs.net/ias22 On the World Wide Web December 13-15,2022 Proceedings of IAS'22 will be published with Springer Verlag in their Lecture Notes in Networks and Systems (LNNS) series. ( https://www.springer.com/series/15179) (Approval Pending) Proceedings of IAS'21: https://link.springer.com/book/10.1007/978-3-030-96305-7 Indexed by: SCOPUS, INSPEC, WTI Frankfurt eG, zbMATH, SCImago, DBLP, EI Compendex, Japanese Science and Technology Agency (JST), SpringerLink History of IAS series: http://www.mirlabs.net/ias22/previous.php **Important Dates** --------------------- Paper submission due: September 30, 2022 Notification of paper acceptance: October 31, 2022 Registration and Final manuscript due: November 15, 2022 Conference: December 13-15, 2022 **About IAS 2022** --------------------- Information assurance and security have become important research issues in the networked and distributed information sharing environments. Finding effective ways to protect information systems, networks, and sensitive data within the critical information infrastructure is challenging even with the most advanced technology and trained professionals. The 18th International Conference on Information Assurance and Security (IAS) aims to bring together researchers, practitioners, developers, and policymakers involved in various information assurance and security disciplines to exchange ideas and learn the latest development in this crucial field. **Topics (not limited to)** --------------------------- Information Assurance, Security Mechanisms, Methodologies and Models Authentication and Identity Management Authorization and Access Control Trust Negotiation, Establishment and Management Anonymity and User Privacy Data Integrity and Privacy Network Security Operating System Security Database Security Intrusion Detection Security Attacks Security Oriented System Design Security and Performance trade-off Security Management and Strategy Security Verification, Evaluations and Measurements Secure Software Technologies New Ideas and Paradigms for Security Cryptography Cryptographic Protocols Key Management and Recovery Secure System Architectures and Security Application Image Engineering, Multimedia Signal Processing and Communication Security **Submission Guidelines** ------------------------- Submission of a paper should be made through the submission page from the conference web page. Please refer to the conference website for guidelines to prepare your manuscript. Paper format templates: https://www.springer.com/de/authors-editors/book-authors-editors/manuscript-preparation/5636#c3324 IAS?22 Submission Link: https://easychair.org/conferences/?conf=ias20220 **Plenary Talks** ---------------------------------------- You are welcome to attend 11 Keynote Talks offered by world-renowned professors and industry leaders. The detailed information is available on conference website. Speaker 1: Catarina Silva, University of Coimbra, Portugal Title: Interpretability and Explainability in Intelligent Systems Speaker 2: Chuan-Kang Ting, National Tsing Hua University, Taiwan Title: TBA Speaker 3: Joanna Kolodziej, Cracow University of Technology, Poland Title: TBA Speaker 4: Kaisa Miettinen, Multiobjective Optimization Group, Faculty of Information Technology, University of Jyvaskyla, Finland Title: Some Perspectives to Interactive Evolutionary Multiobjective Optimization Methods. Speaker 5: Kaspar Riesen, Institute of Computer Science, University of Bern, Switzerland University of Applied Sciences and Arts, Switzerland Title: Four Decades of Structural Pattern Recognition ? An Overview of the Three Major Epochs Speaker 6: Katherine MALAN, Department of Decision Sciences, University of South Africa Title: Landscape analysis of optimisation and machine learning search spaces. Speaker 7: Maki Sakamoto, The University of Electro-Communications, Tokyo, Japan Title: Computer Vision for Expressing Texture Using Sound-Symbolic Words Speaker 8: M?rio Antunes, Polytechnic Institute of Leiria, Portugal Title: Cybersecurity: the road ahead Speaker 9: Patricia MELIN, Tijuana Institute of Technology, Tijuana, Mexico Title: Hybrid Intelligent Systems based on Neural Networks, Fuzzy Logic and Bioinspired Optimization Algorithms and their application to Pattern Recognition. Speaker 10: Ren? NATOWICZ, ESIEE-Paris - Universit? Gustave Eiffel, France Title: Machine Learning in Graphs: Where Are We So Far? Speaker 11: Yifei Pu, College of Computer Science, Sichuan University, China Title : Analog Circuit Implementation of Fractional-Order Memristor: Arbitrary-Order Lattice Scaling Fracmemristor. ** IAS 2022 Organization ** ---------------------------- General Chairs Ajith Abraham, Machine Intelligence Research Labs, USA Tzung-Pei Hong, National University of Kaohsiung, Taiwan Art?ras Kaklauskas, Vilnius Gediminas Technical University, Lithuania Program Chairs Ketan Kotecha, Symbiosis International University, India Ganeshsree Selvachandran, UCSI University, Malaysia Publication Chairs Niketa Gandhi, Machine Intelligence Research Labs, USA Kun Ma, University of Jinan, China Special Session Chair Gabriella Casalino, University of Bari, Italy Publicity Chairs Pooja Manghirmalani Mishra, Machine Intelligence Research Labs, India Anu Bajaj, Machine Intelligence Research Labs, USA If you would like to propose a special session, please email Dr. Gabriella Casalino **Technical Contact** --------------------- Dr. Ajith Abraham Email: ajith.abraham at ieee.org -------------- next part -------------- An HTML attachment was scrubbed... URL: From leslie.perez at pucv.cl Tue Jul 26 13:48:24 2022 From: leslie.perez at pucv.cl (Leslie Angelica Perez Caceres) Date: Tue, 26 Jul 2022 13:48:24 -0400 Subject: Connectionists: First Call for Papers: EvoCOP 2023 - The 23nd European Conference on Evolutionary Computation in Combinatorial Optimisation Message-ID: (Apologies for cross-posting) ************************************************************************************* First Call for Papers: EvoCOP 2023 - The 23nd European Conference on Evolutionary Computation in Combinatorial Optimisation http://www.evostar.org/2023/evocop/ April 12 - 14, 2023 held as part of EvoStar (http://www.evostar.org) Venue: Brno, Czech Republic ** EvoCOP is CORE Rank B ** Submission deadline: November 1, 2022 ************************************************************************************* The 23nd European Conference on Evolutionary Computation in Combinatorial Optimisation is a multidisciplinary conference that brings together researchers working on applications and theory of evolutionary computation methods and other metaheuristics for solving difficult combinatorial optimisation problems appearing in various industrial, economic, and scientific domains. Successfully solved problems include, but are not limited to, multi-objective, uncertain, dynamic and stochastic problems in the context of scheduling, timetabling, network design, transportation and distribution, vehicle routing, stringology, graphs, satisfiability, energy optimisation, cutting, packing, planning and search-based software engineering. The EvoCOP 2023 conference will be held somewhere on Earth, together with EuroGP (the 26th European Conference on Genetic Programming), EvoMUSART (the 12th European conference on evolutionary and biologically inspired music, sound, art and design) and EvoApplications (the 26th European Conference on the Applications of Evolutionary Computation), and a special track on Evolutionary Machine Learning in a joint event collectively known as EvoStar (Evo*). Accepted papers will be published by Springer Nature in the Lecture Notes in Computer Science series. (See https://link.springer.com/conference/evocop for previous proceedings.) Download the CFP in PDF format: https://www.evostar.org/2023/wp-content/uploads/2022/07/evo2022-evocop-flyers-prints-05062022-3.pdf The best regular paper presented at EvoCOP 2023 will be distinguished with a Best Paper Award. EvoCOP conference is ranked B in the CORE 2021 ranking: http://portal.core.edu.au/conf-ranks/2195/ **** Areas of Interest and Contributions **** EvoCOP welcomes submissions in all experimental and theoretical aspects of evolutionary computation and other metaheuristics to combinatorial optimisation problems, including (but not limited to) the following areas: * Applications of metaheuristics to combinatorial optimisation problems * Theoretical developments * Neighbourhoods and efficient algorithms for searching them * Variation operators for stochastic search methods * Constraint-handling techniques * Parallelisation and grid computing * Search space and landscape analyses * Comparisons between different (also exact) methods * Automatic algorithm configuration and design Prominent examples of metaheuristics include (but are not limited to): * Evolutionary algorithms * Estimation of distribution algorithms * Swarm intelligence methods such as ant colony and particle swarm optimisation * Artificial immune systems * Local search methods such as simulated annealing, tabu search, variable neighbourhood search, iterated local search, scatter search and path relinking * Hybrid methods such as memetic algorithms * Matheuristics (hybrids of exact and heuristic methods) * Hyper-heuristics and autonomous search * Surrogate-model-based methods Notice that, by tradition, continuous/numerical optimisation is *not* part of the topics of interest of EvoCOP. Interested authors might consider submitting to other EvoStar conferences such as EvoApplications. **** Submission Details **** Paper submissions must be original and not published elsewhere. The submissions will be peer reviewed by members of the program committee. The reviewing process will be double-blind, please omit information about the authors in the submitted paper. Submit your manuscript in Springer LNCS format: http://www.springer.com/computer/lncs?SGWID=0-164-6-793341-0 Page limit: 16 pages Submission link: coming soon The authors of accepted papers will have to improve their paper on the basis of the reviewers? comments and will be asked to send a camera-ready version of their manuscripts. At least one author of each accepted work has to register for the conference, attend the conference and present the work. **** Important Dates **** Submission deadline: November 1, 2022 EvoStar: April 12-14, 2023 **** EvoCOP Programme Chairs **** Thomas St?tzle Universit? Libre de Bruxelles, Belgium stuetzle at ulb.ac.be Leslie P?rez C?ceres Pontificia Universidad Cat?lica de Valpara?so, Chile leslie.perez at pucv.cl -- Leslie P?rez C?ceres Escuela de Ingenier?a Inform?tica Pontificia Universidad Cat?lica de Valpara?so -------------- next part -------------- An HTML attachment was scrubbed... URL: From vasenka at gmail.com Tue Jul 26 14:32:31 2022 From: vasenka at gmail.com (Vasily Vakorin) Date: Tue, 26 Jul 2022 11:32:31 -0700 Subject: Connectionists: Several PhD positions in clinical EEG and MEG at SFU, Vancouver In-Reply-To: References: Message-ID: Dear moderators, Can you please post this ad? Thank you. Several PhD positions in clinical EEG and MEG at SFU, Vancouver Several funded Ph.D. positions are available at Simon Fraser University (Vancouver, Canada), at the Department of Biomedical Physiology and Kinesiology and the School of Engineering Science. Ph.D. candidates will focus on research in neurophysiology, such as electroencephalography (EEG) and magnetoencephalography (MEG), and machine learning to study clinical populations. Two Ph.D. projects will use clinical routine EEG recorded and evaluated in public hospitals in the process of neurological evaluation to predict patients? clinical profiles. The other project will apply MEG to explore cortico-cerebellar brain networks in children with autism. Ph.D. candidates are expected to get experience with digital signal processing, multivariate statistics, predictive analytics (deep learning and classical machine learning) and data analysis with applications in clinical and cognitive neuroscience, and further develop scientific programming skills (Python). The ideal start date is September 2022, but in general it is flexible. For more information, please contact Dr. Sam Doesburg, Dr. Teresa Cheung, and Dr. Vasily Vakorin: https://www.bcni-sfu.net . If you are interested, please send your cover letter and CV to vvakorin at sfu.ca . Please name the attached files according to the format: Lastname_Firstname_cv.pdf and Lastname_Firstname_coverletter.pdf . Thank you, Vasily Vakorin, Ph.D. Scientist & NeuroInformatics Lead - Behavioral & Cognitive Neuroscience Institute Adjunct Professor - Biomedical Physiology & Kinesiology Simon Fraser University, British Columbia, Canada Affiliated Researcher - Fraser Health Authority, British Columbia email: vvakorin at sfu.ca -------------- next part -------------- An HTML attachment was scrubbed... URL: From dwang at cse.ohio-state.edu Wed Jul 27 21:29:41 2022 From: dwang at cse.ohio-state.edu (Wang, Deliang) Date: Thu, 28 Jul 2022 01:29:41 +0000 Subject: Connectionists: NEURAL NETWORKS, Aug. 2022 Message-ID: Neural Networks - Volume 152, August 2022 https://www.journals.elsevier.com/neural-networks Robust kernel principal component analysis with optimal mean Pei Li, Wenlin Zhang, Chengjun Lu, Rui Zhang, Xuelong Li Dynamic image clustering from projected coordinates of deep similarity learning Jui-Hung Chang, Yin-Chung Leung Deep networks may capture biological behavior for shallow, but not deep, empirical characterizations Peter Neri Branching time active inference: Empirical study and complexity class analysis Th?ophile Champion, Howard Bowman, Marek Grzes Social impact and governance of AI and neurotechnologies Kenji Doya, Arisa Ema, Hiroaki Kitano, Masamichi Sakagami, Stuart Russell A neuroscience-inspired spiking neural network for EEG-based auditory spatial attention detection Faramarz Faghihi, Siqi Cai, Ahmed A. Moustafa Deep unsupervised feature selection by discarding nuisance and correlated features Uri Shaham, Ofir Lindenbaum, Jonathan Svirsky, Yuval Kluger Context meta-reinforcement learning via neuromodulation Eseoghene Ben-Iwhiwhu, Jeffery Dick, Nicholas A. Ketz, Praveen K. Pilly, Andrea Soltoggio Multistability analysis of delayed recurrent neural networks with a class of piecewise nonlinear activation functions Yang Liu, Zhen Wang, Qian Ma, Hao Shen Discovering diverse solutions in deep reinforcement learning by maximizing state-action-based mutual information Takayuki Osa, Voot Tangkaratt, Masashi Sugiyama Sparse factorization of square matrices with application to neural attention modeling Ruslan Khalitov, Tong Yu, Lei Cheng, Zhirong Yang Optimistic reinforcement learning by forward Kullback-Leibler divergence optimization Taisuke Kobayashi A new deep learning framework based on blood pressure range constraint for continuous cuffless BP estimation Yongyi Chen, Dan Zhang, Hamid Reza Karimi, Chao Deng, Wutao Yin Distantly Supervised Relation Extraction via Recursive Hierarchy-Interactive Attention and Entity-Order Perception Ridong Han, Tao Peng, Jiayu Han, Hai Cui, Lu Liu Non-linear perceptual multi-scale network for single image super-resolution Aiping Yang, Leilei Li, Jinbin Wang, Zhong Ji, ... Zihao Wei Event-triggered integral reinforcement learning for nonzero-sum games with asymmetric input saturation Shan Xue, Biao Luo, Derong Liu, Ying Gao Attributed graph clustering with multi-task embedding learning Xiaotong Zhang, Han Liu, Xianchao Zhang, Xinyue Liu Deep learning, reinforcement learning, and world models Yutaka Matsuo, Yann LeCun, Maneesh Sahani, Doina Precup, ... Jun Morimoto Multi-level landmark-guided deep network for face super-resolution Cheng Zhuang, Minqi Li, Kaibing Zhang, Zheng Li, Jian Lu Sediment Prediction in the Great Barrier Reef using Vision Transformer with finite element analysis Mohammad Jahanbakht, Wei Xiang, Mostafa Rahimi Azghadi Embedding graphs on Grassmann manifold Bingxin Zhou, Xuebin Zheng, Yu Guang Wang, Ming Li, Junbin Gao Generative convolution layer for image generation Seung Park, Yong-Goo Shin Replacing pooling functions in Convolutional Neural Networks by linear combinations of increasing functions Iosu Rodriguez-Martinez, Julio Lafuente, Regivan H.N. Santiago, Gra?aliz Pereira Dimuro, ... Humberto Bustince LAP: Latency-aware automated pruning with dynamic-based filter selection Zailong Chen, Chubo Liu, Wangdong Yang, Kenli Li, Keqin Li Visual context learning based on textual knowledge for image-text retrieval Yuzhuo Qin, Xiaodong Gu, Zhenshan Tan Semantic consistency learning on manifold for source data-free unsupervised domain adaptation Song Tang, Yan Zou, Zihao Song, Jianzhi Lyu, ... Jianwei Zhang GIU-GANs: Global Information Utilization for Generative Adversarial Networks Yongqi Tian, Xueyuan Gong, Jialin Tang, Binghua Su, ... Xinyuan Zhang A multivariate adaptive gradient algorithm with reduced tuning efforts Samer Saab, Khaled Saab, Shashi Phoha, Minghui Zhu, Asok Ray Human-guided auto-labeling for network traffic data: The GELM approach Meejoung Kim, Inkyu Lee Weighted Incremental-Decremental Support Vector Machines for concept drift with shifting window Honorius Galmeanu, Razvan Andonie Organization of a Latent Space structure in VAE/GAN trained by navigation data Hiroki Kojima, Takashi Ikegami Context-aware dynamic neural computational models for accurate Poly(A) signal prediction Yanbu Guo, Chaoyang Li, Dongming Zhou, Jinde Cao, Hui Liang ExSpliNet: An interpretable and expressive spline-based neural network Daniele Fakhoury, Emanuele Fakhoury, Hendrik Speleers A manifold learning approach for gesture recognition from micro-Doppler radar measurements E.S. Mason, H.N. Mhaskar, Adam Guo Sampled-data synchronization of complex network based on periodic self-triggered intermittent control and its application to image encryption Hui Zhou, Zijiang Liu, Dianhui Chu, Wenxue Li Set-membership filtering for complex networks with constraint communication channels Chang Liu, Lixin Yang, Jie Tao, Yong Xu, Tingwen Huang Risk-based implementation of COLREGs for autonomous surface vehicles using deep reinforcement learning Amalie Heiberg, Thomas Nakken Larsen, Eivind Meyer, Adil Rasheed, ... Damiano Varagnolo Artificial neural networks with conformable transfer function for improving the performance in thermal and environmental processes J.E. Solis-Perez, J.A. Hernandez, A. Parrales, J.F. Gomez-Aguilar, A. Huicochea Tri-view two-photon microscopic image registration and deblurring with convolutional neural networks Sehyung Lee, Hideaki Kume, Hidetoshi Urakubo, Haruo Kasai, Shin Ishii Quantum pulse coupled neural network Zhaobin Wang, Minzhe Xu, Yaonan Zhang DynamicNet: A time-variant ODE network for multi-step wind speed prediction Rui Ye, Xutao Li, Yunming Ye, Baoquan Zhang A dynamical neural network approach for solving stochastic two-player zero-sum games Dawen Wu, Abdel Lisser Spatiotemporal CNN with Pyramid Bottleneck Blocks: Application to eye blinking detection S.E. Bekhouche, I. Kajo, Y. Ruichek, F. Dornaika Multivariate time series forecasting method based on nonlinear spiking neural P systems and non-subsampled shearlet transform Lifan Long, Qian Liu, Hong Peng, Jun Wang, Qian Yang Knowledge-guided multi-task attention network for survival risk prediction using multi-center computed tomography images Liwen Zhang, Lianzhen Zhong, Cong Li, Wenjuan Zhang, ... Jie Tian -------------- next part -------------- An HTML attachment was scrubbed... URL: From dyyeung at cse.ust.hk Wed Jul 27 22:52:51 2022 From: dyyeung at cse.ust.hk (Dit-Yan Yeung) Date: Thu, 28 Jul 2022 02:52:51 +0000 Subject: Connectionists: [Jobs] Research Staff Positions at HKUST in Hong Kong Message-ID: <095D9594-7617-40CC-BE75-F450EA3971E0@contoso.com> We are looking for outstanding candidates to fill multiple research staff positions for several projects in machine learning and its applications in meteorology, earth sciences, and video prediction. The research staff members recruited will be affiliated with the Department of Computer Science and Engineering (https://cse.hkust.edu.hk/) at the Hong Kong University of Science and Technology (HKUST). More information including the online application procedure can be found here: https://cse.hkust.edu.hk/admin/recruitment/dyyeung/ Best regards, Dit-Yan Yeung, Chair Professor Department of Computer Science and Engineering Hong Kong University of Science and Technology Clear Water Bay Hong Kong -------------- next part -------------- An HTML attachment was scrubbed... URL: From iswc.conf at gmail.com Thu Jul 28 11:05:07 2022 From: iswc.conf at gmail.com (International Semantic Web Conference) Date: Thu, 28 Jul 2022 11:05:07 -0400 Subject: Connectionists: [iswc2022] Important information regarding ISWC 2022 conference format change! Message-ID: *Virtual 21st International Semantic Web Conference (ISWC 2022)* Hangzhou, China, October 23-27, 2022 https://iswc2022.semanticweb.org/ We regret to inform you that because of the increasing travel restrictions in China we are forced to *switch ISWC 2022 from a hybrid to a fully virtual* only format. We will be happy to meet all of you remotely. Stay tuned! 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 davy.weissenbacher at gmail.com Thu Jul 28 13:58:48 2022 From: davy.weissenbacher at gmail.com (Davy Weissenbacher) Date: Thu, 28 Jul 2022 10:58:48 -0700 Subject: Connectionists: CFP: #SMM4H'22, 7th Social Media Mining for Health Applications - Workshop at COLING 2022 Message-ID: ============================= Last call for papers, submission deadline is August 17, 2022 ============================= *Apologies if you received multiple copies of this CFP* Location: Gyeongju, Republic of Korea Workshop Date: October 16-17, 2022 Workshop link: https://healthlanguageprocessing.org/smm4h-2022/ Submission link: https://www.softconf.com/coling2022/7thSMM4H/ The workshop will include two components ? a standard workshop and a shared task Workshop The Social Media Mining for Health Applications (#SMM4H) workshop serves as a venue for bringing together researchers interested in automatic methods for the collection, extraction, representation, analysis, and validation of social media data (e.g., Twitter, Reddit, Facebook) for health informatics. The 7th #SMM4H Workshop, co-located at COLING 2022 ( https://coling2022.org/index), invites 4-page paper (unlimited references in standard COLING format) submissions on original, unpublished research in all aspects at the intersection of social media mining and health. Topics of interest include, but are not limited to: Methods for the automatic detection and extraction of health-related concept mentions in social media Mapping of health-related mentions in social media to standardized vocabularies Deriving health-related trends from social media Information retrieval methods for obtaining relevant social media data Geographic or demographic data inference from social media discourse Virus spread monitoring using social media Mining health-related discussions in social media Drug abuse and alcoholism incidence monitoring through social media Disease incidence studies using social media Sentinel event detection using social media Semantic methods in social media analysis Classifying health-related messages in social media Automatic analysis of social media messages for disease surveillance and patient education Methods for validation of social media-derived hypotheses and datasets Shared task The workshop organizers this year are hosting 10 shared tasks i.e. NLP challenges as part of the workshop. Participating teams will be provided with a set of annotated posts for developing systems, followed by a three-day window during which they will run their systems on unlabeled test data and upload it to Codalab for evaluation. For additional details about the tasks and information about registration, data access, paper submissions, and presentations, go to https://healthlanguageprocessing.org/smm4h-2022/ Task 1 ? Classification, detection, and normalization of Adverse Events (AE) mentions in tweets (in English) Task 2 ? Classification of stance and premise in tweets about health mandates related to COVID-19 (in English) Task 3 ? Classification of changes in medication treatments in tweets and WebMD reviews (in English) Task 4 ? Classification of tweets self-reporting exact age (in English) Task 5 ? Classification of tweets containing self-reported COVID-19 symptoms (in Spanish) Task 6 ? Classification of tweets which indicate self-reported COVID-19 vaccination status (in English) Task 7 ? Classification of self-reported intimate partner violence on Twitter (in English) Task 8 ? Classification of self-reported chronic stress on Twitter (in English) Task 9 ? Classification of Reddit posts self-reporting exact age (in English) Task 10 ? Detection of disease mentions in tweets ? SocialDisNER (in Spanish) Organizing Committee Graciela Gonzalez-Hernandez, Cedars-Sinai Medical Center, USA Davy Weissenbacher, Cedars-Sinai Medical Center, USA Arjun Magge, University of Pennsylvania, USA Ari Z. Klein, University of Pennsylvania, USA Ivan Flores, Cedars-Sinai Medical Center, USA Karen O?Connor, University of Pennsylvania, USA Raul Rodriguez-Esteban, Roche Pharmaceuticals, Switzerland Lucia Schmidt, Roche Pharmaceuticals, Switzerland Juan M. Banda, Georgia State University, USA Abeed Sarker, Emory University, USA Yuting Guo, Emory University, USA Yao Ge, Emory University, USA Elena Tutubalina, Insilico Medicine, Hong Kong Luis Gasco, Barcelona Supercomputing Center, Spain Darryl Estrada, Barcelona Supercomputing Center, Spain Martin Krallinger, Barcelona Supercomputing Center, Spain Program Committee Cecilia Arighi, University of Delaware, USA Natalia Grabar, French National Center for Scientific Research, France Thierry Hamon, Paris-Nord University, France Antonio Jimeno Yepes, Royal Melbourne Institute of Technology, Australia Jin-Dong Kim, Database Center for Life Science, Japan Corrado Lanera, University of Padova, Italy Robert Leaman, US National Library of Medicine, USA Kirk Roberts, University of Texas Health Science Center at Houston, USA Yutaka Sasaki, Toyota Technological Institute, Japan Pierre Zweigenbaum, French National Center for Scientific Research, France Contact All questions should be emailed to Davy Weissenbacher ( davy.weissenbacher at cshs.org) -------------- next part -------------- An HTML attachment was scrubbed... URL: From franrruiz87 at gmail.com Fri Jul 29 03:14:09 2022 From: franrruiz87 at gmail.com (=?UTF-8?Q?Francisco_J=2E_Rodr=C3=ADguez_Ruiz?=) Date: Fri, 29 Jul 2022 08:14:09 +0100 Subject: Connectionists: ICBINB Monthly Seminar Series Talk: Benjamin Bloem-Reddy Message-ID: Dear all, We are pleased to announce that the next speaker of the *?I Can?t Believe It?s Not Better!? (**ICBINB)* virtual seminar series will be *Benjamin Bloem-Reddy** (**University of British Columbia**)*. More details about this series and the talk 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: *August 4th, 2022 at 8am PDT / 11am EDT / 5pm CEST (*Note*: The time differs from our usual one.) *Where: *RSVP for the Zoom link here: https://us02web.zoom.us/meeting/register/tZUtceitrzgvHNYgvD02gj57-kxKNahUdTiC *Title:* *From Identifiability to Structured Representation Spaces, and a Case for (Precise) Pragmatism in Machine Learning* *Abstract: **There has been a recent surge in research activity related to identifiability in generative models involving latent variables. Why should we care whether a latent variable model is identifiable? I will give some pragmatic reasons, which differ philosophically from and which have different practical and theoretical implications than, classical views on identifiability, which usually relate to recovering the ?true? distribution or ?true? latent factors of variation. In particular, a pragmatic approach requires us to consider how the structure we are imposing (or not imposing) on the latent space relates to the problems we?re trying to solve. I will highlight how I think a lack of precise pragmatism is holding back modern methods in challenging settings, including how aspects of my own research on identiability has gotten stuck without problem-specific constraints. Elaborating on methods for representation learning more generally, I will discuss some ways we can (and are beginning to) structure our latent spaces to achieve specific goals other than vague appeals to general AI.* *Bio:* *Benjamin Bloem-Reddy is an Assistant Professor of Statistics at the University of British Columbia. He works on problems in statistics and machine learning, with an emphasis on probabilistic approaches. He has a growing interest in causality and its interplay with knowledge and inference and also collaborates with researchers in the sciences on statistical problems arising in their research.* *Bloem-Reddy was a PhD student with Peter Orbanz at Columbia and a postdoc with Yee Whye Teh in the CSML group at the University of Oxford. Before moving to statistics and machine learning, he studied physics at Stanford University and Northwestern University.* 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 h.jaeger at rug.nl Thu Jul 28 18:02:28 2022 From: h.jaeger at rug.nl (Herbert Jaeger) Date: Fri, 29 Jul 2022 00:02:28 +0200 Subject: Connectionists: Preparatory workshop for shaping funded research teams in neuromorphic and unconventional computing (Europe only) Message-ID: <70fe472a-0348-6766-df7a-5a5371b92dab@rug.nl> The Volkswagen Foundation -- Germany's largest private research funding organization -- launches an initiative for the development of neuromorphic and other non-digital approaches to analyse and engineer computing systems. One action within this initiative is to support the formation of INTERDISCIPLINARY, possibly multi-national research teams. In a preparatory "Ideation" workshop, to be held in Einbeck (Germany) in December 5-8, 2022, European researchers interested in non-digital computing are invited to exchange views and ideas and start networking. The target audience is open-minded postdocs and assistant/junior professors with backgrounds in relevant areas - including but not restricted to materials and devices, microchip technology, optical computing, neuroscience, cognitive modeling, theory of computing, machine learning, complex dynamical systems, or philosophy of computing. The aim of the workshop is the nucleation of research teams who subsequently submit funding proposals for large-scale interdisciplinary research projects. More information can be found at https://www.volkswagenstiftung.de/en/funding/our-funding-portfolio-at-a-glance/next-integrating-approaches-to-neuromorphic-computing. -- Dr. Herbert Jaeger Professor of Computing in Cognitive Materials Rijksuniversiteit Groningen Faculty of Science and Engineering - CogniGron Bernoulliborg Nijenborgh 9, 9747 AG Groningen office: Bernoulliborg 402 phone: +31 (0) 50-36 32473 COVID home office phone: +49 (0) 4209 930403 web: www.ai.rug.nl/minds/ From m.biehl at rug.nl Fri Jul 29 14:10:02 2022 From: m.biehl at rug.nl (Michael Biehl) Date: Fri, 29 Jul 2022 20:10:02 +0200 Subject: Connectionists: 2 fully funded PhD positions at the Univ. of Groningen / NL Message-ID: Two fully funded (4 years) PhD positions in Neural Networks / Machine Learning / Deep Learning are available at the University of Groningen / NL within the project: "*Robust Learning of **Sparse Representations: * *Brain-inspired Inhibition and **Statistical Physics* *Analysis*". Main supervisors and sub-projects: George Azzopardi (https://www.cs.rug.nl/~george/) "*Push-pull inhibition in deep, convolutional * *(spiking) neural networks*" Michael Biehl (https://www.cs.rug.nl/~biehl/) "*Statistical physics analysis of learning processes * *in model situations*" For details and your expression of interest please visit https://t.co/VLjgKzPxJp *Apologies for multiple postings.* *Please circulate and distribute further.* --------------------------------------------------- Prof. Dr. Michael Biehl Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence P.O. Box 407, 9700 AK Groningen The Netherlands https://www.cs.rug.nl/~biehl m.biehl at rug.nl -------------- next part -------------- An HTML attachment was scrubbed... URL: From franciscocruzhh at gmail.com Fri Jul 29 07:18:43 2022 From: franciscocruzhh at gmail.com (Francisco Cruz) Date: Fri, 29 Jul 2022 21:18:43 +1000 Subject: Connectionists: PhD position in Cognitive Robotics available at UNSW Sydney Message-ID: A full Ph.D. scholarship at UNSW Sydney is available for a student to work on topics such as social and cognitive robotics. Particular areas of interest include reinforcement learning, interactive machine learning, explainability, and multi-agent systems. UNSW is a research university placed in Sydney, one of the most liveable cities in the world. The School of Computer Science and Engineering this year celebrates 30 years of world-class education, industry-leading research, and global impact. *** Requirements: The minimum requirements to apply are: - A Bachelor's degree with upper second class Honours; or - A Master's degree with a substantial research component; or - An equivalent qualification from a tertiary institution as determined by the Faculty Higher Degree Committee (HDC) Additionally, it is desirable the applicant has relevant background and research experience. *** What we offer: The candidate will be supervised by Dr. Francisco Cruz. We offer a full scholarship for 3.5 years including AUD $28,854 per annum (2022 rate), tuition fees, and overseas student health cover (OSHC). *** Application: If you are interested and have related background please send the following documents to f.cruz (AT) unsw.edu.au by August 31st: - Your CV - Academic transcripts - 2 paragraphs describing your research interests and background - Most relevant publication, if any, or Master thesis (or equivalent) Should you have questions or would like to discuss further details, please get in touch. Dr. Francisco Cruz School of Computer Science and Engineering Faculty of Engineering UNSW Sydney -------------- next part -------------- An HTML attachment was scrubbed... URL: From ivan.ezhov at tum.de Fri Jul 29 07:50:18 2022 From: ivan.ezhov at tum.de (Ezhov, Ivan) Date: Fri, 29 Jul 2022 11:50:18 +0000 Subject: Connectionists: PhD position on radiotherapy personalization for brain tumor patients, TUM, Munich Message-ID: <2afb730543c84ee2925e36f8453abf6a@tum.de> ******** PhD position. A PhD position on image-based personalization of radiotherapy planning for brain tumor patients is currently available at TUM, Munich. The position is supervised by Benedikt Wiestler (TUM) and Bjoern Menze (UZH). The salary is according to TV-L E13 (100%). ******** Project description. Understanding the dynamics of brain tumor progression is essential for optimal treatment planning. Cast in a mathematical formulation, it can be viewed as an evaluation of a system of partial differential equations, wherein the underlying physiological processes that govern the growth of the tumor, such as diffusion and proliferation of tumor cells, are considered. To personalize the model, i.e. find a relevant set of parameters, with respect to the tumor dynamics of a particular patient, the model can be informed from empirical data, e.g., medical images obtained from different diagnostic modalities, such as magnetic-resonance imaging (MRI) or positron-emission tomography (PET). Recently, we performed proof-of-concept studies with deep-learning based personalization techniques [1]. The goal of the PhD project is to pave the way between the existing or potentially newly developed personalization methodologies and clinical practice. ******** Prerequisites. Python (or C++), the ability to understand some math (probability theory, partial differential equations). ******** Keywords. Computational physiology, statistical inference, physics-informed deep learning ******** For further information please feel free to reach out: b.wiestler at tum.de, bjoern.menze at uzh.ch, ivan.ezhov at tum.de References: [1] Learn-Morph-Infer: a new way of solving the inverse problem for brain tumor modeling, https://arxiv.org/abs/2111.04090 -------------- next part -------------- An HTML attachment was scrubbed... URL: From julie.coloigner at gmail.com Fri Jul 29 17:28:12 2022 From: julie.coloigner at gmail.com (Julie Coloigner) Date: Fri, 29 Jul 2022 23:28:12 +0200 Subject: Connectionists: ISBI 2023 Message-ID: Announcing the Call for Papers for the 2023 IEEE International Symposium on Biomedical Imaging (ISBI)! Call for Papers is now open for the 2023 IEEE International Symposium on Biomedical Imaging (ISBI). Submit your papers by 27 October 2022! For its twentieth anniversary, ISBI will take place from 18-21 April 2023 in Cartagena de Indias, Colombia, the first time ever in Latin America. Located on the shores of the Caribbean Sea, Cartagena de Indias brings together the charm of colonial architecture, the excitement of a vivid nightlife, fascinating cultural festivals, and lush landscapes. It is one of the most beautiful, well-preserved cities in Latin America, and has been named a world heritage site by UNESCO. This year?s program will include five Keynotes Talks presented by world-renowned imaging scientists and clinicians. We will also host the fourth edition of Clinical Day, a day-long special session designed to promote the exchange of emerging knowledge of biomedical imaging in the clinical fields. Additionally, we will hold the first ever Industrial Day, in which we will hear about the latest products in our field by industry leaders, learn about the journey of medical imaging startups that have now become successful companies, and have a startup pitch competition. As in previous ISBIs, we also have an open call for Special Sessions, Challenges and Tutorials, and we look forward to your submissions. ISBI 2023 topics of interest include: Image formation and reconstruction Physical, biological, and statistical modeling Computational and statistical image processing and analysis Image segmentation Image quality assessment Machine learning for image analysis Dynamic/ multimodal/ multiplexed/ multiscale imaging Computer-aided diagnosis Integration of imaging and non-imaging biomarkers Imaging informatics Visualization in biomedical imaging, and biomedical applications ISBI 2023 Call for Submissions Paper Submission: All submissions must be original submissions not under concurrent review at any other conference or journal. Accepted 4-page regular papers will be published in the symposium proceedings by IEEE Xplore. For author guidelines and submission instructions, visit: https://2023.biomedicalimaging.org/en/GUIDELINES.html Important Dates: Special Sessions Submission Deadline: 29 September 2022 Notification: 6 October 2022 Tutorials & Challenges Submission Deadline: 13 October 2022 Notification: 3 November 2022 4-Page Papers Submission Deadline: 27 October 2022 Notification: 19 January 2023 Camera Ready: 23 February 2023 1-Page Abstracts Submission Deadline: 26 January 2023 Notification: 9 February 2023 Pitch Competition Submission Deadline: 26 January 2023 Notification: 9 February 2023 Learn More about ISBI 2023 Copyright 2022 IEEE Signal Processing Society, All rights reserved. IEEE Signal Processing Society: 445 Hoes Lane, Piscataway, NJ, 08854, USA Email us |Join the IEEE | Already an IEEE member? Join the IEEE Signal Processing Society | IEEE Privacy Policy Julie Coloigner CNRS Researcher Empenn research team IRISA - CNRS (UMR 6074), Rennes FR35042 https://juliecoloigner.fr/ https://team.inria.fr/empenn/ -------------- next part -------------- An HTML attachment was scrubbed... URL: From achler at gmail.com Sat Jul 30 02:35:48 2022 From: achler at gmail.com (Tsvi Achler) Date: Fri, 29 Jul 2022 23:35:48 -0700 Subject: Connectionists: [EXT] If you believe in your work ... In-Reply-To: References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> <9ea52c4c-2b66-e8bd-b494-5dc4ab7ebc8d@susaro.com> Message-ID: BTW in regards to faulty research (eg the last link in my last message of this thread), certain fMRI data that may support computational modeling can be very faulty and politicised: https://www.wired.com/2009/09/fmrisalmon/ Moreover research supporting feedforward findings and existing feedforward models (e.g. neural networks) in neuroscience is also very politicised. It is clear that the brain has an abundance of top-down feedback connections and likely is overwhelmingly using this type of feedback in computation during recognition. This reality potentially makes the majority of models irrelevant. -Tsvi On Sun, Jul 24, 2022 at 7:57 AM Tsvi Achler wrote: > Jean-Marc, > > The idea to let the editor decide was directed more to publications than > funding. > > For funding decisions I think the key is to break the hegemony by > primarily funding those who have little say in the system. That's a way to > equalize the say and sources of ideas in the system. > In that vein, one suggestion is to fund research primarily through junior > researchers and give them the power, if a senior researcher needs large > money then they to do it by a mechanism setup where they court junior > researchers. > > Another suggestion for funding is let each member of staff in the funding > committee have N favorite applications a year that they can choose > bypassing the others' scrutiny, > > Indeed some random funding is a great idea as well. > > Follow my channel on YouTube "Updating Research" for more ideas: > https://www.youtube.com/channel/UCbvTQ3lLVvikKaYnNH3kH3g > > BTW. this is how bad peer reviewed & politicized research can get: > https://www.science.org/content/article/potential-fabrication-research-images-threatens-key-theory-alzheimers-disease > > -Tsvi > > > > On Sat, Jul 23, 2022 at 9:18 AM Fellous, Jean-Marc - (fellous) < > fellous at arizona.edu> wrote: > >> >> >> @ Richard, Frederick, Tsvi: >> >> >> >> Points taken. >> >> Regarding funding: Delegating the funding decision to one person is >> probably dangerous. Note that technically, this is the case at NIH >> institutes (the director is the only one who makes the final decision, but >> in practice, that decision is built on the basis of peer review and program >> officer inputs), and similar at NSF (the PO has more authority in funding, >> but again, within limits). One other possibility for allowing innovative >> ideas to push through the ?politics? and ?settled culture? might be to set >> aside (say) 20% of the fund to randomly fund ?sound? short (2 years?) >> proposals that were placed far from funding threshold by reviewers. >> Injecting ?noise? so to speak to get out of local minima? >> >> Regarding publications: There are so many journals out there, and the >> arXiv options. The issue of course is integrity, trust, reproducibility >> etc. To some extent this is what peer review attempts to address, but there >> is no guarantee. May be a hybrid system, where researchers can post their >> work in ArXiv type venue, but with minimal initial ?sanity? checks (i.e. >> moderated, not reviewed), followed by voluntary or invited anonymous >> reviews? The quality/strength of the article would be assessed >> multi-dimensionally by the number of reviews/responses, number of reads, >> number of downloads, number of citations, availability of code/data, >> posting of negative results? etc. >> >> Can?t help but think we could be on the verge of a paradigm shift in >> publications and funding models? we need new ideas, and the courage to try >> them out! >> >> >> >> Best, >> >> Jean-Marc >> >> >> >> *From:* Connectionists *On >> Behalf Of *Richard Loosemore >> *Sent:* Tuesday, July 19, 2022 10:40 AM >> *To:* connectionists at mailman.srv.cs.cmu.edu >> *Subject:* Re: Connectionists: [EXT] If you believe in your work ... >> >> >> >> *External Email* >> >> >> Jean-Marc, >> >> The problem is systemic; it goes down into the very roots of modern >> "science". >> >> The only solution that $20m could buy would be: >> >> 1) An institute run by someone with ethical principles, who would use the >> money to attract further funding until it could actually take on board >> researchers with creative ideas and ethical principles, and then free them >> from the yoke of publish-crap-in-quantity-or-perish. >> >> 2) An AI/Cognitive system development tool that would allow people to >> build and explore complex cognitive systems without being shackled to one >> particular architecture (like deep learning and its many descendents). >> >> A propos of (2) that is one thing I proposed in a (rejected) grant >> proposal. It would have cost $6.4m. >> >> Best, >> >> Richard >> >> -- >> Richard Loosemore >> Cornell University >> ... >> rpl72 at cornell.edu >> >> >> >> >> >> >> On 7/19/22 11:31 AM, Fellous, Jean-Marc - (fellous) wrote: >> >> Assuming there are funders on the list, and funding-related people, >> including program officers (and believe or not, there are!): if you had >> $20M to invest in the sort of things we do on this list: how would we make >> things better? Can we brainstorm an alternative system that allows for >> innovating publications and effective funding? >> >> >> >> Jean-Marc >> ------------------------------ >> >> *From:* Connectionists >> on behalf of Richard >> Loosemore >> *Sent:* Monday, July 18, 2022 1:28 PM >> *To:* connectionists at mailman.srv.cs.cmu.edu >> >> >> *Subject:* [EXT]Connectionists: If you believe in your work ... >> >> >> >> *External Email* >> >> >> On 7/17/22 11:52 AM, Grossberg, Stephen wrote: >> >> > ... if you believe in your work, and the criticisms of it are not >> valid, do not give up. ... >> >> > ... all criticisms by reviewers are valuable and should be taken into >> account in your revision. >> >> > Even if a reviewer's criticisms are, to your mind, wrong-headed, they >> represent the >> > viewpoint of a more-than-usually-qualified reader who has given you the >> privilege >> > of taking enough time to read your article. >> >> Really? >> >> 1) I believe in my work, and the criticisms of it are not valid. I did >> not give up, and the net result of not giving up was ... nothing. >> >> 2) No reviewer who has ever commented on my work has shown the slightest >> sign that they understood anything in it. >> >> 3) Good plumbers are more than usually qualified in their field, and if >> one of those gave you the privilege of taking enough time to read your >> article and give nonsensical comments, would you pay any attention to their >> viewpoint? >> >> ** - ** >> >> I have spent my career fighting against this system, to no avail. >> >> I have watched charlatans bamboozle the crowd with pointless mathematics, >> and get published. >> >> I have watched people use teams of subordinates to pump out streams of >> worthless papers that inflate their prestige. >> >> I have written grant proposals that were exquisitely tuned to the stated >> goal of the grant, and then watched as the grant money went to people whose >> proposals had only the faintest imaginable connection to the stated goal of >> the grant. >> >> ** - ** >> >> The quoted remarks, above, somehow distilled all of that history and left >> me shaking with rage at the stupidity. >> >> I have been a member of the Connectionists mailing list since the early >> 1990s, and before that I had been working on neural nets since 1980. >> >> No more. >> >> >> >> Best, >> >> >> >> Richard >> >> -- >> >> Richard Loosemore >> >> Cornell University >> >> ... >> >> rpl72 at cornell.edu >> >> >> >> >> > -------------- next part -------------- An HTML attachment was scrubbed... URL: From ksharma.raj at gmail.com Sat Jul 30 08:10:07 2022 From: ksharma.raj at gmail.com (Raj Sharma) Date: Sat, 30 Jul 2022 17:40:07 +0530 Subject: Connectionists: ACML 2022 --Call for Tutorials [Submission deadline: 2nd September] Message-ID: **apologies if you have received multiple copies of this email* ------------------------------------------------------------------------------ *ACML 2022* The 14th Asian Conference on Machine Learning Hyderabad, India December 12-14, 2022 https://www.acml-conf.org/2022/ ------------------------------------------------------------------------------ *CALL FOR TUTORIALS* We invite proposals for half day (2.5 hour) tutorials on machine learning and related fields. Ideally, the tutorial should attract a wide audience, provide a broad coverage of core research problems in its chosen research area, elucidate technical solutions, discuss their key insights, and stimulate future work. The tutorial should be broad enough to provide a basic introduction to the chosen area, but it should also be deep enough on the most important topics. Presentations that exclusively focus on the presenter's own work or commercial demonstrations are strongly discouraged. Submit your proposal through this form . The deadline for submitting proposals is: *September 2, 2022, 23:59* AoE (Anywhere on Earth) . Accepted proposals will be notified by the end of September. For any questions, please contact the tutorial chairs at acml2022tutorials at gmail.com. -------------- next part -------------- An HTML attachment was scrubbed... URL: From qobi at purdue.edu Sat Jul 30 12:23:38 2022 From: qobi at purdue.edu (Jeffrey Mark Siskind) Date: Sat, 30 Jul 2022 12:23:38 -0400 Subject: Connectionists: [EXT] If you believe in your work ... In-Reply-To: (message from Tsvi Achler on Fri, 29 Jul 2022 23:35:48 -0700) References: <4AFF4095-B8DE-40E7-9BF4-E8CD6BAC54BC@nyu.edu> <9ea52c4c-2b66-e8bd-b494-5dc4ab7ebc8d@susaro.com> Message-ID: BTW in regards to faulty research (eg the last link in my last message of this thread), certain fMRI data that may support computational modeling can be very faulty and politicised: https://www.wired.com/2009/09/fmrisalmon/ For issues in EEG data, first read, in chronological order: https://ieeexplore.ieee.org/document/8099962 CVPR (2017) oral https://ieeexplore.ieee.org/document/8237631 ICCV (2017) https://dl.acm.org/doi/10.1145/3123266.3127907 ACM MM (2017) https://dl.acm.org/doi/10.1145/3240508.3240641 ACM MM (2018) https://ieeexplore.ieee.org/document/9320159 FG (2020) https://ieeexplore.ieee.org/document/9097411 TPAMI (2021) Then read, in chronological order: https://ieeexplore.ieee.org/document/9264220 TPAMI (2021) https://ieeexplore.ieee.org/document/9578178 CVPR (2021) https://ieeexplore.ieee.org/document/9580535 TPAMI (2021) Jeff (http: //engineering.purdue.edu/~qobi) From er.anubajaj at gmail.com Sat Jul 30 14:28:24 2022 From: er.anubajaj at gmail.com (Anu Bajaj) Date: Sat, 30 Jul 2022 11:28:24 -0700 Subject: Connectionists: 2nd CFP: 14th International Conference on Soft Computing and Pattern Recognition (SoCPaR 2022) - Online - Springer Publication Message-ID: ** Second Call for Papers - please circulate this CFP to your colleagues and networks ** -- 14th International Conference on Soft Computing and Pattern Recognition (SoCPaR 2022) -- http://www.mirlabs.org/socpar22 http://www.mirlabs.net/socpar22 On the World Wide Web December 14-16,2022 Proceedings of SoCPaR'22 will be published with Springer Verlag in their Lecture Notes in Networks and Systems (LNNS) series. ( https://www.springer.com/series/15179) (Approval Pending) Proceedings of SoCPaR'21: https://link.springer.com/book/10.1007/978-3-030-96302-6 Indexed by: SCOPUS, INSPEC, WTI Frankfurt eG, zbMATH, SCImago, DBLP, EI Compendex, Japanese Science and Technology Agency (JST), SpringerLink History of SoCPaR series: http://www.mirlabs.net/socpar22/previous.php **Important Dates** --------------------- Paper submission due: September 30, 2022 Notification of paper acceptance: October 31, 2022 Registration and Final manuscript due: November 15, 2022 Conference: December 14-16, 2022 **About SoCPaR 2022** --------------------- After the success of the 13th International Conference on Soft Computing and Pattern Recognition (SoCPaR 2021), SoCPaR 2022 is organized to bring together worldwide leading researchers and practitioners interested in advancing the state-of-the-art in Soft Computing and Pattern Recognition, for exchanging knowledge that encompasses a broad range of disciplines among various distinct communities. It is hoped that researchers and practitioners will bring new prospects for collaboration across disciplines and gain inspiration to facilitate novel breakthroughs. The themes for this conference are thus focused on "Innovating and Inspiring Soft Computing and Intelligent Pattern Recognition". **Topics (not limited to)** --------------------------- Soft Computing and Applications (but not limited to): Evolutionary computing Swarm intelligence Artificial immune systems Fuzzy Sets Uncertainty analysis Fractals Rough Sets Support vector machines Artificial neural networks Case Based Reasoning Wavelets Hybrid intelligent systems Nature inspired computing techniques Machine learning Ambient intelligence Hardware implementations Pattern Recognition and Applications (but not limited to): Information retrieval Data Mining Web Mining Image Processing Computer Vision Bio-informatics Information security Network security Steganography Biometry Remote sensing Medical Informatics E-commerce Signal Processing Control systems **Submission Guidelines** ------------------------- Submission of paper should be made through the submission page from the conference web page. Please refer to the conference website for guidelines to prepare your manuscript. Paper format templates: https://www.springer.com/de/authors-editors/book-authors-editors/manuscript-preparation/5636#c3324 SoCPaR?22 Submission Link: https://easychair.org/conferences/?conf=socpar2022 **Plenary Talks** ---------------------------------------- You are welcome to attend 11 Keynote Talks offered by world-renowned professors and industry leaders. The detailed information is available on conference website. Speaker 1: Catarina Silva, University of Coimbra, Portugal Title: Interpretability and Explainability in Intelligent Systems Speaker 2: Chuan-Kang Ting, National Tsing Hua University, Taiwan Title: TBA Speaker 3: Joanna Kolodziej, Cracow University of Technology, Poland Title: TBA Speaker 4: Kaisa Miettinen, Multiobjective Optimization Group, Faculty of Information Technology, University of Jyvaskyla, Finland Title: Some Perspectives to Interactive Evolutionary Multiobjective Optimization Methods. Speaker 5: Kaspar Riesen, Institute of Computer Science, University of Bern, Switzerland University of Applied Sciences and Arts, Switzerland Title: Four Decades of Structural Pattern Recognition ? An Overview of the Three Major Epochs Speaker 6: Katherine MALAN, Department of Decision Sciences, University of South Africa Title: Landscape analysis of optimisation and machine learning search spaces. Speaker 7: Maki Sakamoto, The University of Electro-Communications, Tokyo, Japan Title: Computer Vision for Expressing Texture Using Sound-Symbolic Words Speaker 8: M?rio Antunes, Polytechnic Institute of Leiria, Portugal Title: Cybersecurity: the road ahead Speaker 9: Patricia MELIN, Tijuana Institute of Technology, Tijuana, Mexico Title: Hybrid Intelligent Systems based on Neural Networks, Fuzzy Logic and Bioinspired Optimization Algorithms and their application to Pattern Recognition. Speaker 10: Ren? NATOWICZ, ESIEE-Paris - Universit? Gustave Eiffel, France Title: Machine Learning in Graphs: Where Are We So Far? Speaker 11: Yifei Pu, College of Computer Science, Sichuan University, China Title : Analog Circuit Implementation of Fractional-Order Memristor: Arbitrary-Order Lattice Scaling Fracmemristor. ** SoCPaR 2022 Organization ** ------------------------------ General Chairs Ajith Abraham, Machine Intelligence Research Labs (MIR Labs), USA Patrick Siarry, Universit? Paris-Est Cr?teil, France Thomas Hanne, University of Applied Sciences and Arts Northwestern Switzerland, Switzerland Program Chairs Isabel Jesus, Institute of Engineering of Porto, Portugal Nazar Zaki, United Arab Emirate University, UAE Publication Chairs Niketa Gandhi, Machine Intelligence Research Labs, USA Kun Ma, University of Jinan, China Special Session Chair Gabriella Casalino, University of Bari, Italy Publicity Chairs Pooja Manghirmalani Mishra, Machine Intelligence Research Labs, India Anu Bajaj, Machine Intelligence Research Labs, USA If you would like to propose a special session, please email Dr. Gabriella Casalino **Technical Contact** --------------------- Dr. Ajith Abraham Email: ajith.abraham at ieee.org -------------- next part -------------- An HTML attachment was scrubbed... URL: From david at irdta.eu Sat Jul 30 16:02:00 2022 From: david at irdta.eu (David Silva - IRDTA) Date: Sat, 30 Jul 2022 22:02:00 +0200 (CEST) Subject: Connectionists: DeepLearn 2022 Autumn: early registration August 15 Message-ID: <1303680238.186751.1659211320199@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: August 15, 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 21 four-hour and a half courses and 2 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: Tommaso Dorigo (Italian National Institute for Nuclear Physics), Deep-Learning-Optimized Design of Experiments: Challenges and Opportunities Elaine O. Nsoesie (Boston University), AI and Health Equity PROFESSORS AND COURSES: Sean Benson (Netherlands Cancer Institute), [intermediate] Deep Learning for a Better Understanding of Cancer 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 Nadya Chernyavskaya (European Organization for Nuclear Research), [intermediate] Graph Networks for Scientific Applications with Examples from Particle Physics S?bastien Fabbro (University of Victoria), [introductory/intermediate] Learning with Astronomical Data Efstratios Gavves (University of Amsterdam), [advanced] Advanced Deep Learning 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 Tin Kam Ho (IBM Thomas J. Watson Research Center), [introductory/intermediate] Deep Learning Applications in Natural Language Understanding Timothy Hospedales (University of Edinburgh), [intermediate/advanced] Deep Meta-Learning 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 Yao Wang (New York University), [introductory/intermediate] Deep Learning for Computer Vision 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: Nosheen Abid (Lule?) Sana Sabah Al-Azzawi (Lule?) Lama Alkhaled (Lule?) Prakash Chandra Chhipa (Lule?) Saleha Javed (Lule?) Marcus Liwicki (Lule?, local chair) Carlos Mart?n-Vide (Tarragona, program chair) Hamam Mokayed (Lule?) Sara Morales (Brussels) Mia Oldenburg (Lule?) Maryam Pahlavan (Lule?) David Silva (London, organization 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 are available 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: