From han.zhao at cs.cmu.edu Thu Jan 10 14:54:48 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Thu, 10 Jan 2019 14:54:48 -0500 Subject: [AI Seminar] Call-for-Talk in AI Seminar, Spring 2019 Message-ID: Dear faculty and students: Happy new year and welcome back to school! The CMU AI Seminar this semester will start in two weeks on Jan. 22nd. Please sign up to contribute a talk! Again, according to our previous experience, the slots will be filled up quickly, so please do it as soon as you can if you would like to give a talk about your great work! Looking forward to seeing all of you at AI seminar soon! -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Jan 20 09:40:48 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 20 Jan 2019 09:40:48 -0500 Subject: [AI Seminar] No seminar on Jan. 22nd Message-ID: Dear faculty and students: This is just a friend notification that due to the incoming ICML deadline on Jan. 23rd, we will not hold seminar next Tuesday on Jan. 22nd. Good luck to everyone! -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Jan 27 15:12:34 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 27 Jan 2019 15:12:34 -0500 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Wen Sun -- Jan. 29th Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Jan. 29th, at noon in *NSH 3305 *for our first AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Wen Sun will give the following talk: *Title: Towards Generalization and Efficiency in Reinforcement Learning* *Abstract*: In classic supervised machine learning, a learning agent behaves as a passive observer: it receives examples from some external environment which it has no control over and then makes predictions. Reinforcement Learning (RL), on the other hand, is fundamentally interactive : an autonomous agent must learn how to behave in an unknown and possibly hostile environment, by actively interacting with the environment to collect useful feedback. One central challenge in RL is how to explore an unknown environment and collect useful feedback efficiently. In recent practical RL success stories, we notice that most of them rely on random exploration which requires large a number of interactions with the environment before it can learn anything useful. The theoretical RL literature has developed more sophisticated algorithms for efficient learning, however, the sample complexity of these algorithms has to scale exponentially with respect to key parameters of underlying systems such as the dimensionality of state vector, which prohibits a direct application of these theoretically elegant RL algorithms to large-scale applications. Without any further assumptions, RL is hard, both in practice and in theory. In this work, we improve generalization and efficiency on RL problems by introducing extra sources of help and additional assumptions. The first contribution of this work comes from improving RL sample efficiency via Imitation Learning (IL). Imitation Learning reduces policy improvement to classic supervised learning. We study in both theory and in practice how one can imitate experts to reduce sample complexity compared to RL approaches. The second contribution of this work comes from exploiting the underlying structures of the RL problems via model-based learning approaches. While there exist efficient model-based RL approaches specialized for specific RL problems (e.g., tabular MDPs, Linear Quadratic Systems), we develop a unified model-based algorithm that generalizes a large number of RL problems that were often studied independently in the literature. We also revisit the long standing debate on whether model-based RL is more efficient than model-free RL from a theoretical perspective, and demonstrate that model-based RL can be exponentially more sample efficient than model-free ones, which to the best of our knowledge, is the first that separates model-based and model-free general approaches. ----------------------------------------------------------------------------------------------------------------------------- P. S. This talk should be of great interest to many people working on ML/RL and other related domains, hence I encourage everyone to attend. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Feb 3 11:40:44 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 3 Feb 2019 11:40:44 -0500 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Graham Yennie Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Feb. 5th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Graham Yennie will give the following talk: *Title: **Machine Learning in Production: A Collection of Lessons Learned from Deploying ML Across Industries* *Abstract*: In this topic, I will present some common themes and high level objects in designing and implementing large ML platforms and workflows. These workflows will include datasets and data ingestion, the train/evaluate cycle, deployment, and application development. In each step of the workflow, I will go over high level considerations and some possible solutions, including current frameworks and libraries that address said challenges. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Feb 10 10:33:54 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 10 Feb 2019 10:33:54 -0500 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Hsiao-Yu Fish Tung Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Feb. 12th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Hsiao-Yu Fish Tung will give the following talk: *Title: **Geometry-Aware Recurrent Networks: A visual system for embodied agents* *Abstract*: In this talk, I will introduce you our new work that integrates two powerful ideas, geometry and deep visual representation learning, into recurrent network architectures for mobile visual scene understanding. The proposed networks learn to "lift" 2D visual features and integrate them over time into latent 3D feature maps of the scene. They are equipped with differentiable geometric operations, such as projection, unprojection, egomotion estimation, and stabilization, in order to compute a geometrically-consistent mapping between the world scene and their 3D latent feature space. We train the proposed architectures to predict novel image views given short frame sequences as input. Their predictions strongly generalize to scenes with a novel number of objects, appearances, and configurations, and greatly outperform predictions of previous works that do not consider egomotion stabilization or a space-aware latent feature space. We train the proposed architectures to detect and segment objects in 3D, using the latent 3D feature map as input--as opposed to any input 2D video frame. The resulting detections are permanent: they continue to exist even when an object gets occluded or leaves the field of view. Our experiments suggest the proposed space-aware latent feature arrangement and egomotion-stabilized convolutions are essential architectural choices for spatial common sense to emerge in artificial embodied visual agents. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Feb 17 12:38:56 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 17 Feb 2019 12:38:56 -0500 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Simon Shaolei Du Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Feb. 19th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Simon Shaolei Du will give the following talk: *Title: Understanding Optimization and Generalization in Deep Learning: A Trajectory-based Analysis* *Abstract*: In this talk, I will present recent progress on understanding deep neural networks by analyzing the trajectory of the gradient descent algorithm. Using this analysis technique, we are able to explain: 1) why gradient descent finds a global minimum of the training loss even though the objective function is highly non-convex, and 2) why a neural network can generalize even the number of parameters in the neural network is more than the number of training data. Based on joint work with Sanjeev Arora, Wei Hu, Jason D. Lee, Haochuan Li, Zhiyuan Li, Barnabas Poczos, Aarti Singh, Liwei Wang, Ruosong Wang, Xiyu Zhai References: https://arxiv.org/abs/1810.02054 https://arxiv.org/abs/1811.03804 https://arxiv.org/abs/1901.08584 Bio: Simon Shaolei Du is a Ph.D. student in the Machine Learning Department at the School of Computer Science, Carnegie Mellon University, advised by Professor Aarti Singh and Professor Barnab?s P?czos. His research interests broadly include topics in theoretical machine learning and statistics, such as deep learning, matrix factorization, convex/non-convex optimization, transfer learning, reinforcement learning, non-parametric statistics, and robust statistics. In 2015, he obtained his B.S. in Engineering Math & Statistics and B.S. in Electrical Engineering & Computer Science from the University of California, Berkeley. He has also spent time working at research labs of Microsoft and Facebook. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Tue Feb 19 11:54:56 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Tue, 19 Feb 2019 11:54:56 -0500 Subject: [AI Seminar] Fwd: AI Seminar sponsored by Apple -- Simon Shaolei Du In-Reply-To: References: Message-ID: A gentle reminder that the following talk is happening soon. ---------- Forwarded message --------- From: Han Zhao Date: 2019?2?17??? ??12:38 Subject: AI Seminar sponsored by Apple -- Simon Shaolei Du To: Dear faculty and students: We look forward to seeing you next Tuesday, Feb. 19th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Simon Shaolei Du will give the following talk: *Title: Understanding Optimization and Generalization in Deep Learning: A Trajectory-based Analysis* *Abstract*: In this talk, I will present recent progress on understanding deep neural networks by analyzing the trajectory of the gradient descent algorithm. Using this analysis technique, we are able to explain: 1) why gradient descent finds a global minimum of the training loss even though the objective function is highly non-convex, and 2) why a neural network can generalize even the number of parameters in the neural network is more than the number of training data. Based on joint work with Sanjeev Arora, Wei Hu, Jason D. Lee, Haochuan Li, Zhiyuan Li, Barnabas Poczos, Aarti Singh, Liwei Wang, Ruosong Wang, Xiyu Zhai References: https://arxiv.org/abs/1810.02054 https://arxiv.org/abs/1811.03804 https://arxiv.org/abs/1901.08584 Bio: Simon Shaolei Du is a Ph.D. student in the Machine Learning Department at the School of Computer Science, Carnegie Mellon University, advised by Professor Aarti Singh and Professor Barnab?s P?czos. His research interests broadly include topics in theoretical machine learning and statistics, such as deep learning, matrix factorization, convex/non-convex optimization, transfer learning, reinforcement learning, non-parametric statistics, and robust statistics. In 2015, he obtained his B.S. in Engineering Math & Statistics and B.S. in Electrical Engineering & Computer Science from the University of California, Berkeley. He has also spent time working at research labs of Microsoft and Facebook. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Feb 24 20:10:23 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 24 Feb 2019 20:10:23 -0500 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Eric Wong Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Feb. 26th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Eric Wong will give the following talk: *Title: Provable defenses against adversarial attacks: from linear programming to dual networks* *Abstract*: In this talk, I will present recent progress on duality-based certified defenses against adversarial attacks for neural networks. Using convex relaxations of network architectures, we are able to: 1) provide a certified bound on the worst case adversarial output of a network over a perturbation region in the input space 2) the bound can be computed as a pass through a "dual network" which has structure similar to the backwards pass of the original architecture 3) training against this bound learns networks which are provably safe against any adversarial attack in the given threat model -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Mar 3 13:56:19 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 3 Mar 2019 13:56:19 -0500 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Anson Kahng Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Mar. 5th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Anson Kahng will give the following talk: *Title: Statistical Foundations of Virtual Democracy* *Abstract*: Virtual democracy is an approach to automating decisions, by learning models of the preferences of individual people, and, at runtime, aggregating the predicted preferences of those people on the dilemma at hand. One of the key questions is which aggregation method ? or voting rule ? to use; we offer a novel statistical viewpoint that provides guidance. Specifically, we seek voting rules that are robust to prediction errors, in that their output on people's true preferences is likely to coincide with their output on noisy estimates thereof. We prove that the classic Borda count rule is robust in this sense, whereas any voting rule belonging to the wide family of pairwise-majority consistent rules is not. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Mon Mar 11 12:06:01 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Mon, 11 Mar 2019 12:06:01 -0400 Subject: [AI Seminar] No AI Seminar this week Message-ID: Dear faculty and students: Due to the spring break, we do not have AI seminar this week. We will resume our weekly seminar from next week. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Mar 17 14:43:19 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 17 Mar 2019 10:43:19 -0800 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Yuexin Wu Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Mar. 19th, at noon in *GHC 8102 (note the unusual location this time) *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Yuexin Wu will give the following talk: *Title: Deep Learning for Epidemiological Predictions* *Abstract*: Predicting new and urgent trends in epidemiological data is an important problem for public health, and has attracted increasing attention in the data mining and machine learning communities. The temporal nature of epidemiology data and the need for real-time prediction by the system makes the problem residing in the category of time-series forecasting or prediction. While traditional autoregressive (AR) methods and Gaussian Process Regression (GPR) have been actively studied for solving this problem, deep learning techniques have not been explored in this domain. In this paper, we develop a deep learning framework, for the first time, to predict epidemiology profiles in the time-series perspective. We adopt Recurrent Neural Networks (RNNs) to capture the long-term correlation in the data and Convolutional Neural Networks (CNNs) to fuse information from data of different sources. A residual structure is also applied to prevent overfitting issues in the training process. We compared our model with the most widely used AR models on USA and Japan datasets. Our approach provides consistently better results than these baseline methods. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Mar 24 16:29:35 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 24 Mar 2019 16:29:35 -0400 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Jeremy Cohen Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Mar. 26th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Jeremy Cohen will give the following talk: *Title: **Certified Adversarial Robustness via Randomized Smoothing* *Abstract*: Extending recent work , we show how to turn any classifier that classifies well under Gaussian noise into a new classifier that is provably robust to perturbations in L2 norm. This method is the only provable adversarial defense that scales to ImageNet. It also outperforms all other provable L2 adversarial defenses on CIFAR-10 by a wide margin. Best of all, the method is extremely simple to implement and to understand. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Mon Apr 1 09:37:46 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Mon, 1 Apr 2019 09:37:46 -0400 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Dr. P. Anandan Message-ID: Dear faculty and students: We look forward to seeing you tomorrow, Apr. 2nd, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website. On Tuesday, Dr. P. Anandan will give the following talk: *Title: **AI Solutions for the Underserved Billons* *Abstract*: While AI has become ubiquitous in the daily lives of most people in developed segments of the world, it?s usage among the billions of poor is practically non-existent. Yet, given that the a primary cause of sustained poverty and hardship is the insufficiency of human expertise and inadequacy of expert human resources, AI may actually provide opportunities to transform the every aspect of the life such as health, agriculture, basic education, infrastructure, and financial inclusion. Wadhwani AI is an independent not-for-profit applied research Institute based in Mumbai founded on this premise and with the goal of developing innovative AI based solutions to address the challenges of the lives of the poor. We are about one year old, and have grown to a team of about 25, working in developing solutions for a few initial use-cases in health and agriculture, specifically in the areas of Maternal and Child health, TB eradication, and Cotton Farming. In this talk, I will described our approach to these problems, including how we select our problems, the technical aspects of our work, and how we propose to see our work have tangible, sustained and scalable impact in the world. I will also use the talk to share some the learnings from the first year of this endeavor, especially regarding some of the key challenges in doing AI in the social sector and how to address them. *Bio: *Dr. P. Anandan is the CEO of Wadhwani Institute of Artificial Intelligence. Previous to this Anandan was VP for Research at the Adobe Research Lab India (2016-2017) and before that a Distinguished Scientist and Managing Director at Microsoft Research (1997-2014). Anandan was the founding director of Microsoft Research India which he ran from 2005-2014. Prior to this, Anandan was researcher at Sarnoff corporation (1991-1997) and an Assistant Professor of Computer Science at Yale University (1987-1991). His primary research area is Computer vision where he is well known for his fundamental and lasting contributions to the problem of visual motion analysis. Anandan received his PhD in Computer Science from University of Massachusetts, Amherst in 1987, a Masters in Computer Science from University of Nebraska, Lincoln in 1979 and his BTech in Electrical Engineering from IIT Madras, India in 1977. He is a distinguished alumnus of IIT Madras, and UMass, Amherst and is on the Nebraska Hall of Computing. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Mon Apr 1 16:56:24 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Mon, 1 Apr 2019 16:56:24 -0400 Subject: [AI Seminar] Fwd: Advertising two talks to AI mailing list? In-Reply-To: References: Message-ID: For those who are interested in the following talks, please come to the talks and sign up on the sheet if you're interested in talking to the speakers. ---------- Forwarded message --------- ???? Fei Fang Date: 2019?4?1??? ??10:24 Subject: Advertising two talks to AI mailing list? To: Han Zhao , Zico Kolter Hi Han and Zico, I am hosting two visitors this week, Prof. Tanya Berger-Wolf and Prof. Kevin Leyton-Brown. They will give seminar talks on Tue and Wed. Attached are the posters of their talks. Also, there are a few open slots in their visiting schedules (Prof Berger-Wolf: https://docs.google.com/spreadsheets/d/1msMdGIQwpOLO68_30nS_cL1bcRoQPB2GDy8rUYQQS28/edit#gid=0 ; Prof. Leyton-Brown: https://docs.google.com/spreadsheets/d/16GCX01B94lwuLUKUzv7ONg-aiGHQoHUnyzHulxYq0pM/edit#gid=0 ). Could you help circulate the information to people who may be interested (e.g., the AI mailing list)? Thanks in advance! Fei Fang Assistant Professor Institute for Software Research, School of Computer Science Carnegie Mellon University http://feifang.info/ -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: Kevin Leyton-Brown Poster.pdf Type: application/pdf Size: 237946 bytes Desc: not available URL: -------------- next part -------------- A non-text attachment was scrubbed... Name: Tanya Berger-Wolf Poster.pdf Type: application/pdf Size: 1231533 bytes Desc: not available URL: From han.zhao at cs.cmu.edu Sun Apr 7 11:51:31 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 7 Apr 2019 11:51:31 -0400 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Omer Ben-Porat Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Apr. 9th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Omer Ben-Porat will give the following talk: *Title: **Regression Equilibrium* *Abstract: *Prediction is a well-studied machine learning task, and prediction algorithms are core ingredients in online products and services. Despite their centrality in the competition between online companies who offer prediction-based products, the strategic use of prediction algorithms remains unexplored. The goal of this presentation is to examine strategic use of prediction algorithms. We introduce a novel game-theoretic setting that is based on the PAC learning framework, where each player (aka a prediction algorithm aimed at competition) seeks to maximize the sum of points for which it produces an accurate prediction and the others do not. We show that algorithms aiming at generalization may wittingly mispredict some points to perform better than others in expectation. We analyze the empirical game, i.e., the game induced on a given sample, prove that it always possesses a pure Nash equilibrium, and show that every better-response learning process converges. Moreover, our learning-theoretic analysis suggests that players can, with high probability, learn an approximate pure Nash equilibrium for the whole population using a small number of samples. Based on a joint work with Moshe Tennenholtz. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Thu Apr 11 10:35:54 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Thu, 11 Apr 2019 10:35:54 -0400 Subject: [AI Seminar] Fwd: Forwarding ICML AI + Climate Change Workshop call for submissions? In-Reply-To: References: Message-ID: FYI. ICML AI + Climate Change Workshop CFP: https://www.climatechange.ai/ ---------- Forwarded message --------- ???? Priya Donti Date: 2019?4?2??? ??2:00 Subject: Forwarding ICML AI + Climate Change Workshop call for submissions? To: Han Zhao , Zico Kolter (CMU) Hi Han and Zico, Would you be willing to forward this call for submissions on to the AI seminar mailing list? Thanks! Priya ---------- Forwarded message --------- From: Priya Donti Subject: ICML AI + Climate Change Workshop - call for submissions *** CALL FOR SUBMISSIONS: ICML workshop ?Climate Change: How Can AI Help?? *** We invite submission of extended abstracts applying machine learning to the problems of climate change. There will be three tracks (Deployed, Research, and Ideas). Date: June 14 or 15, 2019 Location: Long Beach, California, USA Website: www.climatechange.ai Submission deadline: April 30, 11:59 PM Pacific Time Notification: May 15 (early notification possible upon request) Submission website: https://cmt3.research.microsoft.com/CCAI2019 Contact: climatechangeai.icml2019 at gmail.com Summary ------------- Climate change is widely agreed to be one of the greatest challenges facing humanity. We already observe increased incidence and severity of storms, droughts, fires, and flooding, as well as significant changes to global ecosystems, including the natural resources and agriculture on which humanity depends. The 2018 UN report on climate change estimates that the world has only thirty years to eliminate greenhouse emissions completely if we are to avoid catastrophic consequences. Many in the machine learning community want to address climate change but feel their skills are inapplicable. This workshop will showcase the many settings in which machine learning can be applied to reducing greenhouse emissions and helping society adapt to the effects of climate change. Climate change is a complex problem requiring simultaneous action from many directions. While machine learning is not a silver bullet, there is significant potential impact for research and implementation. About ICML ---------------- ICML is one of the premier conferences on machine learning, and includes a wide audience of researchers and practitioners in academia and industry. It is possible to attend the workshop without either presenting or attending the main ICML conference. Those interested should register for the Workshops component of ICML at https://icml.cc/ while tickets last (a number of spots will be reserved for accepted submissions). Call for submissions --------------------------- We invite submission of extended abstracts on machine learning applied to problems in climate mitigation, adaptation, or modeling, including but not limited to the following topics: - Power generation and grids - Transportation - Smart buildings and cities - Industrial optimization - Carbon capture and sequestration - Agriculture, forestry and other land use - Climate modeling - Extreme weather events - Disaster management and relief - Societal adaptation - Ecosystems and natural resources - Data presentation and management - Climate finance Accepted submissions will be invited to give poster presentations at the workshop, of which some will be selected for spotlight talks. Please contact climatechangeai.icml2019 at gmail.com with questions, or if visa considerations make earlier notification important. Dual-submissions are allowed, and the workshop does not record proceedings. Submissions will be reviewed double-blind; do your best to anonymize your submission, and do not include identifying information for authors in the PDF. We encourage, but do not require, use of the ICML style template (please do not use the ?Accepted? format). Submission tracks ------------------------ Extended abstracts are limited to 3 pages for the Deployed and Research tracks, and 2 pages for the Ideas track, in PDF format. An additional page may be used for references. All machine learning techniques are welcome, from kernel methods to deep learning. Each submission should make clear why the application has (or could have) positive impacts regarding climate change. There are three tracks for submissions: DEPLOYED * Work that is already having an impact * Submissions for the Deployed track are intended for machine learning approaches which are impacting climate-relevant problems through consumers or partner institutions. This could include implementations of academic research that have moved beyond the testing phase, as well as results from startups/industry. Details of methodology need not be revealed if they are proprietary, though transparency is encouraged. RESEARCH * Work that will have an impact when deployed * Submissions for the Research track are intended for machine learning research applied to climate-relevant problems. Submissions should provide experimental or theoretical validation of the method proposed, as well as specifying what gap the method fills. Algorithms need not be novel from a machine learning perspective if they are applied in a novel setting. Datasets may be submitted to this track that are designed to permit machine learning research (e.g. formatted with clear benchmarks for evaluation). In this case, baseline experimental results on the dataset are preferred but not required. IDEAS * Future work that could have an impact * Submissions for the Ideas track are intended for proposed applications of machine learning to solve climate-relevant problems. While the least constrained, this track will be subject to a very high standard of review. No results need be demonstrated, but ideas should be justified as extensively as possible, including motivation for the problem being solved, an explanation of why current tools are inadequate, and details of how tools from machine learning are proposed to fill the gap. Organizers --------------- David Rolnick (UPenn) Alexandre Lacoste (ElementAI) Tegan Maharaj (MILA) Jennifer Chayes (Microsoft) Yoshua Bengio (MILA) Karthik Mukkavilli (MILA) Di Wu (MILA) Narmada Balasooriya (ConscientAI) Priya Donti (CMU) Lynn Kaack (CMU) Manvitha Ponnapati (MIT) -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Apr 14 13:18:49 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 14 Apr 2019 13:18:49 -0400 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Otilia Stretcu Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Apr. 16th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Otilia Stretcu will give the following talk: *Title: **Contextual Parameter Generation for Knowledge Graph Link Prediction* *Abstract: *Knowledge graph link prediction is the task of inferring missing relationships between entities in a knowledge graph. It can be seen as a form of question answering, where given a question consisting of an entity and a relation (e.g., ?Shakespeare? and ?bornIn?), we are tasked with predicting the most likely answer entity (e.g., ?England?). Recent methods have focused on inferring answers by learning entity and relation embeddings. For example, the current state-of-the-art, ConvE, stacks the question entity and relation embeddings together and uses a convolutional neural network to predict the answer entity. However, we argue that entities and relations model different kinds of information and thus, integrating them in this way does not represent a meaningful inductive bias for the model. Therefore, we instead propose to treat relations as the ?context? in which question entities are interpreted and transformed to produce answer entities. Concretely, we use relation embeddings to generate the parameters of a model operating over entity embeddings, which in turn produces a distribution over possible answers. This proposed model outperforms all existing methods by a significant margin on several established datasets, thus establishing the new state-of-the-art for this problem, while at the same time reducing training time by up to 98%. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Tue Apr 16 13:21:21 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Tue, 16 Apr 2019 13:21:21 -0400 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Richard Zemel Message-ID: Dear faculty and students: We look forward to seeing you *next Monday, Apr. 22nd*, in the morning *10: 30 am* at *GHC 6115 (note the special time and location) *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Monday, Prof. Richard Zemel will give the following talk: *Title: Controlling the Black Box: Learning Manipulable and Fair Representations* *Abstract: *Machine learning models, and more specifically deep neural networks, are achieving state-of-the-art performance on difficult pattern-recognition tasks such as object recognition, speech recognition, drug discovery, and more. However, deep networks are notoriously difficult to understand, both in how they arrive at and how to affect their responses. As these systems become more prevalent in real-world applications it is essential to allow users to exert more control over the learning system. In particular a wide range of applications can be facilitated by exerting some structure over the learned representations, to enable users to manipulate, interpret, and in some cases obfuscate the representations. In this talk I will discuss recent work that makes some steps towards these goals, allowing users to interact with and control representations. *Bio: *Richard Zemel is a Professor of Computer Science and Industrial Research Chair in Machine Learning at the University of Toronto, and a co-founder and the Research Director at the Vector Institute for Artificial Intelligence. Prior to that he was on the faculty at the University of Arizona, and a Postdoctoral Fellow at the Salk Institute and at CMU. He received the B.Sc. in History & Science from Harvard, and a Ph.D. in Computer Science from the University of Toronto. His awards and honors include a Young Investigator Award from the ONR and a US Presidential Scholar award. He is a Senior Fellow of the Canadian Institute for Advanced Research, an NVIDIA Pioneer of AI, and a member of the NeurIPS Advisory Board. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Apr 28 15:45:44 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 28 Apr 2019 15:45:44 -0400 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Junjie Hu Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday*, Apr. 30th*, in the noon *12: 00 PM* at *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Junjie Hu will give the following talk: *Title: **Cross-Lingual and Cross Domain Transfer for Neural Machine Translation* *Abstract: *Neural machine translation (NMT) models have achieved state-of-the-art performance on many popular benchmark datasets for high-resourced languages (HRL). However, training such deep networks requires a large amount of parallel sentences, and the translation performance in low-resourced languages (LRL) falls far behind that of high-resourced languages. This problem is further exacerbated by domain mismatch, which is unavoidable on many collected datasets. To remedy this problem, we apply transfer learning to adapting NMT models across languages and across domains. In the first part of the talk, we introduce a method of rapidly adapting NMT models to new languages by continuously training on data related to the LRL. In the second part, we propose to adapt NMT models to new domains by generating synthetic data that are closely related to in-domain data. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* Virus-free. www.avast.com <#DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2> -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Tue Sep 10 12:58:36 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Tue, 10 Sep 2019 09:58:36 -0700 Subject: [AI Seminar] [AI-Seminar] Call for Talks in AI Seminar, Fall 2019 Message-ID: Dear all: Welcome back to school and hope you all had a great summer! The CMU AI Seminar will start from Tuesday, Sep. 17th this semester. Please sign up for contributing a talk! Again, based on previous experience, the slots will be filled up soon, so please do it as soon as possible you can by sending an emailing to Han Zhao (han.zhao at cs.cmu.edu) or Aayush Bansal (aayushb at cs.cmu.edu) ! The tentative schedule for this semester is listed in the website . Looking forward to seeing you at the seminars soon! -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Tue Sep 17 19:25:04 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Tue, 17 Sep 2019 19:25:04 -0400 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Aayush Bansal Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Sep. 24th, at noon in *NSH 3305 *for our first AI Seminar this semester, sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Aayush Bansal will give the following talk: *Title: Association and Imagination* *Abstract: *When you see the sunset standing on Seine river in Paris, you can trivially imagine it along the Monongahela in Pittsburgh. When you hear something, you can easily imagine how someone would have said it. When you think of an event from the past, you can relive every bit of it in your imagination. Humans have remarkable abilities to associate different concepts and create visual worlds far beyond what could be seen by a human eye, including inferring the state of unobserved, imagining the unknown, and thinking about diverse possibilities about what lies in the future. These human powers require minimal instructions, and primarily relies on observation and interaction with a dynamic environment. The simple tasks from daily life that are trivial for humans to think and imagine have remained challenging for machine perception and artificial intelligence. The inability to associate and a lack of sense of imagination in machines substantially restricts their applicability. In this talk, I will demonstrate how thinking about association at various levels of abstraction can lead to machine imagination. I will present algorithms that enable association between different domains in an unsupervised manner. This ability to associate allows automatic creation of audio and visual content (images, videos, 4D space-time visualization of dynamic events) that is also user-controllable and interactive. I will show diverse user applications in audio-visual data retargeting, reconstruction, synthesis, and manipulation. These applications are the first steps towards building machines with a powerful audio-visual simulator that will enable them to imagine complex hypothetical situations, and model the aspects of their surroundings that is not easily perceived. *Bio: *Aayush Bansal is a PhD candidate at the Robotics Institute of Carnegie Mellon University. He is a recipient of Uber Presidential Fellowship (2016-17), Qualcomm Fellowship (2017-18), and Snap Fellowship (2019-20). The production houses such as BBC Studios and PBS are using his research work to create documentaries and short movies. Various national and international media such as NBC, CBS, France TV, and The Journalist have extensively covered his work. More details are here . -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Thu Sep 26 16:06:55 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Thu, 26 Sep 2019 16:06:55 -0400 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Gus Guangyu Xia Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Oct. 1st, at noon in *NSH 3305 *for our AI Seminar, sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Gus Guangyu Xia will give the following talk: *Title: Better music representation learning using inductive bias: mind vs. machine* *Abstract: *Gus works on music AI systems for intimate human-computer interaction and co-creation. In this talk, he will present two recent prototype systems: the multimodal flute tutor and the interpretable deep music composer. The former is a hyper instrument which guides human motion and makes music learning more efficient and enjoyable. The latter is a general framework for learning disentangled representation. We will see that for both human learning and machine learning, inductive bias plays an important role. At the end of the talk, Gus will show his vision on merging these two efforts for new music experiences. *Bio: *Gus is an Assistant Professor in Computer Science at NYU Shanghai. He received his Ph.D. in the Machine Learning Department at Carnegie Mellon University in 2016 and was a Neukom Fellow at Dartmouth from 2016 to 2017. Gus is also a professional DI and XIAO (Chinese flute and vertical flute) player. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Thu Oct 3 11:31:35 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Thu, 3 Oct 2019 11:31:35 -0400 Subject: [AI Seminar] AI Seminar sponsored by Apple -- Sarah Scheffler Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Oct. 8th, at noon in *NSH 3305 *for our AI Seminar, sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Sarah Scheffler will give the following talk: *Title: **From Soft Classifiers to Hard Decisions: How fair can we be?* *Abstract: *A popular methodology for building binary decision-making classifiers in the presence of imperfect information is to first construct a non-binary "scoring" classifier that is calibrated over all protected groups, and then to post-process this score to obtain a binary decision. We study the feasibility of achieving various fairness properties by post-processing calibrated scores, and then show that deferring post-processors allow for more fairness conditions to hold on the final decision. Specifically, we show: 1. There does not exist a general way to post-process a calibrated classifier to equalize protected groups' positive or negative predictive value (PPV or NPV). For certain "nice" calibrated classifiers, either PPV or NPV can be equalized when the post-processor uses different thresholds across protected groups, though there exist distributions of calibrated scores for which the two measures cannot be both equalized. When the post-processing consists of a single global threshold across all groups, natural fairness properties, such as equalizing PPV in a nontrivial way, do not hold even for "nice" classifiers. 2. When the post-processing is allowed to `defer' on some decisions (that is, to avoid making a decision by handing off some examples to a separate process), then for the non-deferred decisions, the resulting classifier can be made to equalize PPV, NPV, false positive rate (FPR) and false negative rate (FNR) across the protected groups. This suggests a way to partially evade the impossibility results of Chouldechova and Kleinberg et al., which preclude equalizing all of these measures simultaneously. We also present different deferring strategies and show how they affect the fairness properties of the overall system. We evaluate our post-processing techniques using the COMPAS data set from 2016. Joint work with Ran Canetti, Aloni Cohen, Nishanth Dikkala, Govind Ramnarayan, and Adam Smith. -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sat Oct 12 11:21:20 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sat, 12 Oct 2019 11:21:20 -0400 Subject: [AI Seminar] AI Seminar -- Aditya Grover, Mitigating Bias in Generative Modeling Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Oct. 15th, at noon in *NSH 3305 *for our first AI Seminar this semester, sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Aditya Grover will give the following talk: *Title: *Mitigating Bias in Generative Modeling *Abstract:* In the last few years, there has been remarkable progress in deep generative modeling. However, the learned models are noticeably inaccurate w.r.t. to the underlying data distribution, as evident from downstream metrics that compare statistics of interest across the true and generated data samples. This bias in downstream evaluation can be attributed to imperfections in learning (?model bias?) or be propagated due to the bias in the training dataset itself (?dataset bias?). In this talk, I will present an importance weighting approach for mitigating both these kinds of biases of generative models. Our approach assumes only ?black-box? sample access to a generative model and is broadly applicable to both likelihood-based and likelihood-free generative models. Empirically, we find that our technique consistently improves standard goodness-of-fit metrics for evaluating the sample quality of state-of-the-art deep generative models, suggesting reduced bias. We demonstrate its utility on representative applications in a) data augmentation, b) model-based policy evaluation using off-policy data, and c) permutation-invariant generative modeling of graphs. Finally, I will present some recent work extending these ideas to fair data generation in the presence of biased training datasets. *Bio:* Aditya Grover is a 5th-year Ph.D. candidate in Computer Science at Stanford University advised by Stefano Ermon. His research focuses on probabilistic machine learning, including topics in generative modeling, approximate inference, and deep learning as well as applications in sustainability. His research has been cited widely in academia, deployed into production at major technology companies, and recognized with a best paper award (StarAI), a Lieberman Fellowship, a Data Science Institute Scholarship, and a Microsoft Research Ph.D. Fellowship. He is also a Teaching Fellow at Stanford since 2018, where he co-designed and teaches a new class on Deep Generative Models. Previously, Aditya obtained his bachelors in Computer Science and Engineering from IIT Delhi in 2015, where he received a best undergraduate thesis award. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Thu Oct 17 23:58:04 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Thu, 17 Oct 2019 23:58:04 -0400 Subject: [AI Seminar] AI Seminar -- Kenji Kawaguchi, Special Priorities in Deep Learning and AI Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Oct. 22nd, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Kenji Kawaguchi will give the following talk: *Title: *Special Priorities in Deep Learning and AI *Abstract:* Deep learning and AI have provided high-impact data-driven methods in various applications. However, theoretical guarantees on deep learning and AI tend to provide too pessimistic insights with a gap from practical observations, because of hidden special properties of deep learning and AI problems. Identifying such special properties can provide novel theoretical insights, and is potentially helpful for designing methods and deriving better guarantees. In this talk, I will discuss special properties on non-convex optimization landscapes of deep neural networks and machine learning models, as well as their implications on gradient descent methods and the results on real-world applications based on theoretical insights. *Bio:* Kenji Kawaguchi is a Ph.D. candidate at Massachusetts Institute of Technology (MIT), advised by Prof. Leslie Pack Kaelbling. He received his M.S. in Electrical Engineering and Computer Science from MIT. His research interests span machine learning, deep learning, artificial intelligence, convex/nonconvex optimization and Bayesian optimization. His research has been cited widely in academia and used in classes. He was invited to speak at the 2019 International Congress on Industrial and Applied Mathematics Minisymposium on Theoretical Foundations of Deep Learning. In 2018, he was invited for a summer research visit at Microsoft Research in Redmond. He was awarded the Funai Overseas Scholarship in 2014 and was selected for the Nakajimi Foundation Fellowship in 2013. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From aayushb at cs.cmu.edu Sun Oct 20 22:47:05 2019 From: aayushb at cs.cmu.edu (Aayush Bansal) Date: Sun, 20 Oct 2019 22:47:05 -0400 Subject: [AI Seminar] AI Seminar on Oct 29 -- Yuandong Tian -- Over-parameterization as a Catalyst in Generalization of Deep ReLU networks via Student-Teacher Setting Message-ID: Yuandong Tian, a Research Scientist and Manager at Facebook AI Research, will be giving a seminar on "Over-parameterization as a Catalyst in Generalization of Deep ReLU networks via Student-Teacher Setting" from *12:00 - 01:00 PM* in *Newell Simon Hall (NSH) 3305*. CMU AI Seminar is sponsored by Apple. Lunch will be served. Following are the details of the talk: *Title: *Over-parameterization as a Catalyst in Generalization of Deep ReLU networks via Student-Teacher Setting *Abstract:* In this talk, we study the generalization behaviors of deep networks at interpolation region, where the training error and the gradient at each training data point is small. We use a teacher-student setting: both student and teacher are deep ReLU networks and a student learns from the output of a fixed teacher with SGD. Our conclusion is two-fold. First, with minimal assumptions on the teacher network and the training set, we prove that small gradient at each training data point leads to weight alignment between teacher and student networks at the lowest layer, if both the student and the teacher are. Furthermore, from the proof, over-parameterization makes such alignment more likely to happen. Second, further analysis of the training dynamics shows that student network learns the strong teacher nodes first, leaving weak teacher node unexplained until late stage of the training, and over-parameterization can help cover more weaker node with the same number of iterations. This sheds light on the puzzling phenomena that low training error and over-parameterization could lead to good generalization. *Bio*: Yuandong Tian is a Research Scientist and Manager in Facebook AI Research, working on deep reinforcement learning and its applications, and theoretical analysis of deep models. He is the lead scientist and engineer for ELF OpenGo and DarkForest Go project. Prior to that, he was a researcher and engineer in Google Self-driving Car team in 2013-2014. He received a Ph.D. in Robotics from the Robotics Institute, Carnegie Mellon University in 2013, a Bachelor's and a Master's degree in Computer Science from Shanghai Jiao Tong University. He is the recipient of 2013 ICCV Marr Prize Honorable Mentions. To learn more about the seminar series, please visit the website . -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Nov 3 18:58:11 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 3 Nov 2019 18:58:11 -0500 Subject: [AI Seminar] AI Seminar -- Vaishnavh Nagarajan, Uniform Convergence May Be Unable to Explain Generalization in Deep Learning Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Nov. 5th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Vaishnavh Nagarajan will give the following talk: *Title: *Uniform Convergence May Be Unable to Explain Generalization in Deep Learning *Abstract:* In this talk, I will present our work that casts doubt on the ongoing pursuit of using uniform convergence to explain generalization in deep learning. Over the last couple of years, research in deep learning theory has focused on developing newer and more refined generalization bounds (using Rademacher complexity, covering numbers, PAC-Bayes etc.,) to help us understand why overparameterized deep networks generalize well. Although these bounds are quite different on the surface, essentially, they are 'implementations' of a single learning-theoretic technique called uniform convergence. While it is well-known that many of these existing bounds are numerically large, through a variety of experiments, we first bring to light another crucial and more concerning aspect of these bounds: in practice, these bounds can increase with the dataset size. Guided by these observations, we then present specific scenarios where uniform convergence provably fails to explain generalization in deep learning. That is, in these scenarios, even though a deep network trained by stochastic gradient descent (SGD) generalizes well, any uniform convergence bound would be vacuous, however carefully it is applied. Through our work, we call for going beyond uniform convergence to explain generalization in deep learning. This is joint work with Zico Kolter. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Nov 10 18:39:38 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 10 Nov 2019 18:39:38 -0500 Subject: [AI Seminar] AI Seminar -- Priya Donti, Tackling Climate Change with Machine Learning Message-ID: Dear faculty and students: We look forward to seeing you next Tuesday, Nov. 12th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Priya Donti will give the following talk: *Title: *Tackling Climate Change with Machine Learning *Abstract:* Climate change is one of the greatest challenges facing humanity, and we, as machine learning experts, may wonder how we can help. In this talk, I will describe how machine learning can be a powerful tool in reducing greenhouse gas emissions and helping society adapt to a changing climate. From smart grids to disaster management, I will describe high impact problems where existing gaps can be filled by machine learning, in collaboration with other fields. I will then describe some of my work in this area, which incorporates electricity system domain knowledge into deep network architectures to support low-emissions operation of the electric grid. *Bio: *Priya Donti is a Ph.D. student in Computer Science and Public Policy at Carnegie Mellon University, co-advised by Zico Kolter and Ines Azevedo. Her work lies at the intersection of machine learning, electric power systems, and climate change mitigation. Specifically, she is interested in creating novel machine learning techniques that incorporate domain knowledge (such as power system physics) to reduce greenhouse gas emissions from the electricity sector. Priya is a U.S. Department of Energy Computational Science Graduate Fellow and co-chair of Climate Change AI, a group of volunteers from academia and industry who believe in using machine learning, where it is relevant, to help tackle the climate crisis. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From aayushb at cs.cmu.edu Wed Nov 6 08:46:43 2019 From: aayushb at cs.cmu.edu (Aayush Bansal) Date: Wed, 6 Nov 2019 08:46:43 -0500 Subject: [AI Seminar] Fwd: [graphics] VASC - 11/11/19 -- Madalina Fiterau, an Assistant Professor at UMass Amherst, College of Information & Computer Sciences, presenting "Hybrid Methods for the Integration of Heterogeneous Multimodal Biomedical Data" In-Reply-To: References: <7947fdfcdc144b869338294a27fcb4c9@andrew.cmu.edu> Message-ID: FYI -- this might be of interest to many working on multi-modal data.. Additionally, here are the available meeting slots with the speaker. Feel free to sign-up: https://docs.google.com/spreadsheets/d/13OLRRRRjbA9ivtLi-dECB3xTOje4evN8H27PX7LJ-TY/edit?usp=sharing Aayush ---------- Forwarded message --------- From: Christine A Downey Date: Mon, Nov 4, 2019 at 10:47 AM Subject: [graphics] VASC - 11/11/19 -- Madalina Fiterau, an Assistant Professor at UMass Amherst, College of Information & Computer Sciences, presenting "Hybrid Methods for the Integration of Heterogeneous Multimodal Biomedical Data" To: vasc-seminar at cs.cmu.edu VASC - 11/11/19 -- Madalina Fiterau, an Assistant Professor at UMass Amherst, College of Information & Computer Sciences, will be giving a seminar on "Hybrid Methods for the Integration of Heterogeneous Multimodal Biomedical Data", on 11/11/19, from 3:00-4:00 in *Gates Hillman 6501*. Refreshments will be served. Details are as follows: *Title*: Hybrid Methods for the Integration of Heterogeneous Multimodal Biomedical Data *Abstract**: *The prevalence of smartphones and wearable devices for health monitoring and widespread use of electronic health records have led to a surge in heterogeneous multimodal healthcare data, collected at an unprecedented scale. My research focuses on developing machine learning techniques that learn salient representations of multimodal, heterogeneous data for biomedical predictive models. The first part of the talk describes the construction of hybrid models that combine deep learning with random forests, and the fusing of structured information into temporal representation learning. These methods obviate the need for feature engineering while improving on the state of the art for diverse biomedical applications. Use cases include the prediction of surgical outcomes for children with cerebral palsy, and forecasting the progression of osteoarthritis from subjects' physical activity. The focus of the latter part is on hybrid methods for the integration of images and multi-resolution, irregular time series data for disease trajectory modeling, developed with my students at UMass Amherst. Multi-FIT, a unified model for the construction of flexible temporal representations, is designed to handle missing values and irregularly collected samples in multi-resolution, multivariate time series. Multi-FIT outperforms the state-of-the-art for patient survival prediction on the PhysioNet Challenge 2012 ICU data. FLARe is a model that provides more informative modeling of the temporal relationships between patients' history and the disease trajectory by generating a sequence of latent representations of patients' health status across the time horizon. FLARe improves on the state-of-the-art on forecasting the progression of Alzheimer's disease from brain MRIs and contextual information from the ADNI dataset *Bio: *Ina Fiterau is an Assistant Professor in the College of Information and Computer Sciences at UMass Amherst. She has completed a PhD in Machine Learning from Carnegie Mellon University (Fall 2015), and a Postdoc at Stanford University (Fall 2018). Ina is currently expanding her research on interpretable models, in part by applying deep learning to obtain salient representations from biomedical unstructured data, including time series, text and images. She is the recipient of the Marr Prize for Best Paper at ICCV 2015 and of Star Research Award at the Annual Congress of the Society of Critical Care Medicine 2016. Madalina has co-organized the NeurIPS workshop on Machine Learning in Healthcare. *Homepage:* https://www.cics.umass.edu/people/fiterau-brostean-madalina -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sun Nov 17 12:25:29 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sun, 17 Nov 2019 12:25:29 -0500 Subject: [AI Seminar] AI Seminar Sponsored by Apple -- Chris Liaw, Near-optimal sample complexity bounds for learning mixtures of Gaussians Message-ID: Dear faculty and students: We look forward to seeing you on Tuesday, Nov. 19th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Chris Liaw will give the following talk: *Title: *Near-optimal sample complexity bounds for learning mixtures of Gaussians *Abstract:* Estimating distributions from observed data is a fundamental task in statistics that has been studied for over a century. We consider such a problem where the distribution is a mixture of k Gaussians in R^d. The objective is density estimation: given i.i.d. samples from the (unknown) distribution, produce a distribution whose total variation distance from the unknown distribution is at most epsilon. We prove that Theta(kd^2/epsilon^2) samples are necessary and sufficient for this task, suppressing logarithmic factors. This improves both the known upper bound and lower bound for this problem. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sat Nov 23 10:00:26 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sat, 23 Nov 2019 10:00:26 -0500 Subject: [AI Seminar] AI Seminar Sponsored by Apple -- Shubham Tulsiani, Learning from Geometry and Parsimony in the Visual World Message-ID: Dear faculty and students: We look forward to seeing you on Tuesday, Nov. 26th, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Shubham Tulsiani will give the following talk: *Title: Learning from Geometry and Parsimony in the Visual World* *Abstract:* We live in a structured world, and perceive it in specific ways. In this talk, I?ll argue that approaches that aim to understand the visual world should leverage this in form of inductive biases. I will present a line of work that, by building in the notion that our 2D percepts are projections of an underlying 3D world, can allow us to bypass the need of supervision and learn to infer this 3D, as well as recover 2D to 2D correspondences. I will also show that simply leveraging a prior that complex structures (scenes, objects, etc.) can be thought to have simpler components, we can discover these underlying components, and obtain representations which can more accurately perform downstream tasks e.g. future prediction or robotics manipulation. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sat Nov 30 20:32:36 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sat, 30 Nov 2019 17:32:36 -0800 Subject: [AI Seminar] AI Seminar Sponsored by Apple -- Misha Khodak, Efficient and Adaptive Meta-Learning with Provable Guarantees Message-ID: Dear faculty and students: We look forward to seeing you on Tuesday, Dec. 3rd, at noon in *NSH 3305 *for our AI Seminar sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Misha Khodak will give the following talk: *Title: **Efficient and Adaptive Meta-Learning with Provable Guarantees* *Abstract:* Meta-learning has recently re-emerged as an important direction for developing algorithms for multi-task learning, dynamic environments, and federated settings; however, meta-learning approaches that can scale to deep neural networks are largely heuristic and lack formal guarantees. We build a theoretical framework for designing and understanding practical meta-learning methods that integrates sophisticated formalizations of task-similarity with the extensive literature on online convex optimization and sequential prediction algorithms. Our approach enables the task-similarity to be learned adaptively, provides sharper transfer-risk bounds in the setting of statistical learning-to-learn, and leads to straightforward derivations of average-case regret bounds for efficient algorithms in settings where the task-environment changes dynamically or the tasks share a certain geometric structure. We use our theory to modify several popular meta-learning algorithms and improve performance on standard problems in few-shot and federated learning. Joint work Nina Balcan, Ameet Talwalkar, Jeff Li, and Sebastian Caldas -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From aayushb at cs.cmu.edu Tue Dec 3 12:19:59 2019 From: aayushb at cs.cmu.edu (Aayush Bansal) Date: Tue, 3 Dec 2019 12:19:59 -0500 Subject: [AI Seminar] AI Seminar on Dec 10 (NSH 1305) -- Jason Saragih -- Virtually Indistinguishable Digital Doubles Message-ID: Jason Saragih, a Director Research Scientist at Facebook Reality Labs, will be giving a seminar on "Virtually Indistinguishable Digital Doubles" from *12:00 - 01:00 PM* in *Newell Simon Hall (NSH) 1305*. (NOTE: CHANGE IN LOCATION) CMU AI Seminar is sponsored by Apple. Lunch will be served. Following are the details of the talk: *Title: *Virtually Indistinguishable Digital Doubles *Abstract:* One of the holy grails of AR/VR is a system for telepresence that feels indistinguishable from face to face interactions. A key technology for enabling this is the ability to create digital doubles; representations of humans that are indistinguishable in how they look and move from the real thing. Facebook Reality Labs in Pittsburgh has been working on automating the creation of digital doubles and their animation during social interactions in VR. In this talk, I will give an overview of some of the technology behind our system and outline directions for future work. *Bio*: Jason Saragih is a Director Research Scientist at Facebook Reality Labs (FRL). He works at the intersection of graphics, computer vision, and machine learning, specializing in human modeling. He received his Bachelors in Mechatronics and PhD in Computer Science from the Australian National University in 2004 and 2008 respectively. Prior to joining FRL in 2015, Jason developed computer vision systems for the mobile AR industry. He has also worked as a Post Doc at CMU and as a research scientist at CSIRO where he developed face tracking and modeling technologies. To learn more about the seminar series, please visit the website . -------------- next part -------------- An HTML attachment was scrubbed... URL: From han.zhao at cs.cmu.edu Sat Dec 7 22:27:11 2019 From: han.zhao at cs.cmu.edu (Han Zhao) Date: Sat, 7 Dec 2019 22:27:11 -0500 Subject: [AI Seminar] AI Seminar Sponsored by Apple -- Jason Saragih, Virtually Indistinguishable Digital Doubles Message-ID: Dear faculty and students: We look forward to seeing you on Tuesday, Dec. 10th, at noon in *NSH 1305 (note the different room) *for our last AI Seminar this semester sponsored by Apple. To learn more about the seminar series, please visit the website . On Tuesday, Jason Saragih will give the following talk: *Title: **Virtually Indistinguishable Digital Doubles* *Abstract:* One of the holy grails of AR/VR is a system for telepresence that feels indistinguishable from face to face interactions. A key technology for enabling this is the ability to create digital doubles; representations of humans that are indistinguishable in how they look and move from the real thing. Facebook Reality Labs in Pittsburgh has been working on automating the creation of digital doubles and their animation during social interactions in VR. In this talk, I will give an overview of some of the technology behind our system and outline directions for future work. *Bio: *Jason Saragih is a Director Research Scientist at Facebook Reality Labs (FRL). He works at the intersection of graphics, computer vision, and machine learning, specializing in human modeling. He received his Bachelors in Mechatronics and PhD in Computer Science from the Australian National University in 2004 and 2008 respectively. Prior to joining FRL in 2015, Jason developed computer vision systems for the mobile AR industry. He has also worked as a Post Doc at CMU and as a research scientist at CSIRO where he developed face tracking and modeling technologies. -- *Han ZhaoMachine Learning Department* *School of Computer ScienceCarnegie Mellon UniversityMobile: +1-* *412-652-4404* -------------- next part -------------- An HTML attachment was scrubbed... URL: From aayushb at cs.cmu.edu Tue Dec 10 08:12:04 2019 From: aayushb at cs.cmu.edu (Aayush Bansal) Date: Tue, 10 Dec 2019 08:12:04 -0500 Subject: [AI Seminar] AI Seminar on Dec 10 (NSH 1305) -- Jason Saragih -- Virtually Indistinguishable Digital Doubles In-Reply-To: References: Message-ID: Reminder.. this talk is today at 12:00-01:00 PM in NSH-1305 (NOTE: change in location) On Tue, Dec 3, 2019 at 12:19 PM Aayush Bansal wrote: > Jason Saragih, a Director Research Scientist at Facebook Reality Labs, > will be giving a seminar on "Virtually Indistinguishable Digital Doubles" > from *12:00 - 01:00 PM* in *Newell Simon Hall (NSH) 1305*. (NOTE: CHANGE > IN LOCATION) > > CMU AI Seminar is sponsored by Apple. Lunch will be served. > > Following are the details of the talk: > > *Title: *Virtually Indistinguishable Digital Doubles > > *Abstract:* One of the holy grails of AR/VR is a system for telepresence > that feels indistinguishable from face to face interactions. A key > technology for enabling this is the ability to create digital doubles; > representations of humans that are indistinguishable in how they look and > move from the real thing. Facebook Reality Labs in Pittsburgh has been > working on automating the creation of digital doubles and their animation > during social interactions in VR. In this talk, I will give an overview of > some of the technology behind our system and outline directions for future > work. > > *Bio*: Jason Saragih is a Director Research Scientist at Facebook Reality > Labs (FRL). He works at the intersection of graphics, computer vision, and > machine learning, specializing in human modeling. He received his Bachelors > in Mechatronics and PhD in Computer Science from the Australian National > University in 2004 and 2008 respectively. Prior to joining FRL in 2015, > Jason developed computer vision systems for the mobile AR industry. He has > also worked as a Post Doc at CMU and as a research scientist at CSIRO where > he developed face tracking and modeling technologies. > > To learn more about the seminar series, please visit the website > . > -------------- next part -------------- An HTML attachment was scrubbed... URL: