Bridging Game Theory and Deep Learning

Fei Fang (CMU)

Title:

Integrating Machine Learning with Game Theory for Societal Challenges

Link to video

Abstract:

Real-world problems such as protecting critical infrastructure and cyber networks and protecting wildlife, fishery, and forest often involve multiple decision-makers. While game theory is an established paradigm for such problems, its applicability in practice is often limited by computational intractability in large games, the unavailability of game parameters and the lack of rationality of human players. On the other hand, machine learning has led to huge successes in various domains and can be leveraged to overcome the limitations of the game-theoretic analysis. In this talk, I will introduce our work on integrating machine learning with computational game theory for addressing societal challenges such as security and sustainability, covering the following directions: data-based game-theoretic reasoning, learning-powered strategy computation in large scale games, and end-to-end learning of game parameters.

Short Bio:

Fei Fang is an Assistant Professor at the Institute for Software Research in the School of Computer Science at Carnegie Mellon University. Before joining CMU, she was a Postdoctoral Fellow at the Center for Research on Computation and Society (CRCS) at Harvard University. She received her Ph.D. from the Department of Computer Science at the University of Southern California in June 2016.

Eva Tardos (Cornell University)

Title:

Learning in dynamic multi-agent environments

Link to video

Abstract:

In this talk we will consider on games where players use a form of learning that helps them adapt to a changing environment. We ask if the quantitative guarantees obtained for Nash equilibria for this class of games extend to such out of equilibrium game play, when the game or the population of players is dynamically changing and where participants have to adapt to the dynamic environment.

Short Bio:

Éva Tardos received her Dipl.Math. in 1981 , and her Ph.D. 1984, from Eötvös University , Budapest, Hungary . She joined Cornell in 1989, and was Chair of the Department of Computer Science 2006-2010. She has been elected to the National Academy of Engineering, National Academy of Sciences, and the American Academy of Arts and Sciences, is an external member of the Hungarian Academy of Sciences, and is the recipient of a number of fellowships and awards including the the IEEE John von Neumann Medal, Packard Fellowship, the Gödel Prize, Dantzig Prize, and the Fulkerson Prize. She was editor editor-in-Chief of SIAM Journal of Computing 2004-2009, and is currently editor-in-Chief of the Journal of the ACM, and editor of some other journals includingthe Theory of Computing, and Combinatorica.

David Balduzzi (DeepMind)

Title:

Composition, learning, and games

Link to video

Abstract:

Gradient descent and automatic differentiation provide a powerful framework for composing function-approximators and training them to optimise an objective. However, there are many learning algorithms that do not optimise a single, fixed objective — such as self-play and its generalisations in Go and StarCraft, generative adversarial networks, and adversarial training for robustness. In this talk, I will argue for a new subfield of “differentiable mechanism design”. In support, I will describe extant work from the literature under the unifying themes of meta-games and second-order information.

Short Bio:

David Balduzzi is a researcher at Google DeepMind. He did his PhD in representation theory and algebraic geometry at the University of Chicago. After that he worked on computational neuroscience at UW-Madison and machine learning at the MPI for Intelligent Systems, ETH Zürich and Victoria University Wellington. He now works on game theory and machine learning at DeepMind.

Aryan Mokhtari (UT Austin)

Title:

Understanding the Role of Optimism in Minimax Optimization: A Proximal Point Approach

Link to video

Abstract:

In this talk, we consider solving saddle point problems, and, in particular, we discuss the concept of “optimism” or “negative momentum” - a technique which is observed to have superior empirical performance in training GANs. The goal of this talk is to provide a theoretical understanding on why optimism helps, in particular why the Optimistic Gradient Descent Ascent (OGDA) algorithm performs well in practice. To do so, we first consider the classical Proximal Point algorithm which is an implicit algorithm to solve this problem. We then show that OGDA inherently tries to approximate the proximal point method, and this is the rationale behind the ‘’negative momentum” term in the update of OGDA. This proximal point approximation viewpoint also enables us to provide a much simpler analysis of another well studied algorithm - the Extra-Gradient (EG) method.

Short Bio:

Aryan Mokhtari is an Assistant Professor in the ECE Department of the University of Texas at Austin (UT Austin). Before joining UT Austin, he was a Postdoctoral Associate in the Laboratory for Information and Decision Systems (LIDS) at MIT. Before that, he was a Research Fellow at the Simons Institute for the Theory of Computing at UC Berkeley. His research interests include the areas of optimization, machine learning, and artificial intelligence. His current research focuses on the theory and applications of convex and non-convex optimization in large-scale machine learning and data science problems.

Ioannis Mitliagkas (Mila & University of Montreal)

Ioannis Mitliagkas is an assistant professor in the department of Computer Science and Operations Research (DIRO) at the University of Montréal. Before that, he was a Postdoctoral Scholar with the departments of Statistics and Computer Science at Stanford University. He obtained his Ph.D. from the department of Electrical and Computer Engineering at The University of Texas at Austin. His research includes topics in optimization, statistical learning and inference, and efficient large-scale and distributed algorithms.
He is particularly interested in the dynamics of optimization, like momentum methods, in the presence of system dynamics, adaptivity, and lately, smooth two-player games (ongoing work).

Gauthier Gidel (Mila & University of Montreal)

Gauthier Gidel is a PhD candidate at Mila lab Université de Montréal under the supervision of Simon Lacoste-Julien. Before that, he received the Diplôme de l’École Normale Supérieure in 2017 (ULM MPI2013) and the Master of Science MVA from École Normale supérieur Paris-Saclay in 2016. Gauthier’s PhD thesis topic revolves around saddle point optimization (a.k.a mini-max problems) for machine learning and more generally on understanding the practical and theoretical challenges of differentiable games optimization for multi-agent learning. He is also a recipient of a Graduate Borealis AI fellowship.

Niao He (UIUC)

Niao He is an assistant professor in the Department of Industrial and Enterprise Systems Engineering and Coordinated Science Laboratory at the University of Illinois at Urbana-Champaign. Before joining Illinois, she received her Ph.D. degree in Operations Research from Georgia Institute of Technology in 2015 and B.S. degree in Mathematics from University of Science and Technology of China in 2010. Her research interests are in large-scale optimization and machine learning, with a primary focus in bridging modern optimization theory and algorithms with core machine learning topics, like Bayesian inference, reinforcement learning, and adversarial learning. She is also a recipient of the NSF CISE Research Initiation Initiative (CRII) Award and the NCSA Faculty Fellowship.

Reyhane Askari (Mila & University of Montreal)

Reyhane Askari is a PhD student at Mila lab, Université de Montréal. She works under the supervision of Ioannis Mitliagkas (UdeM) and Nicolas Le Roux (Google Brain). Prior to her PhD, she received her Masters in Computer Science from Université de Montréal and started working as a Machine Learning engineer for two years at Mila. During that time she worked on several open-source software for deep learning such as Theano, Orion and Cortex. She also did her bachelors in Computer Engineering at Amirkabir University of Technology (Tehran Polytechnic).
Her research interests are on understanding accelerated methods in single objective and multi-objective settings using tools from dynamical systems.

Nika Haghtalab

Nika Haghtalab is an Assistant Professor in the Department of Computer Science at Cornell University. She works broadly on the theoretical aspects of machine learning and algorithmic economics. She especially cares about developing a theory for machine learning that accounts for its interactions with people and organizations, and the wide range of social and economic limitations, aspiration, and behavior they demonstrate. Prior to Cornell, she was a postdoctoral researcher at Microsoft Research, New England, in 2018-2019.
She received her Ph.D. from the Computer Science Department of Carnegie Mellon University, where she was co-advised by Avrim Blum and Ariel Procaccia. Her thesis titled Foundation of Machine Learning, by the People, for the People received the CMU School of Computer Science Dissertation Award (2018) and a SIGecom Dissertation Honorable Mention Award (2019).

Simon Lacoste-Julien (Mila & University of Montreal)

Simon Lacoste-Julien is a CIFAR fellow and an assistant professor at Mila and DIRO from Université de Montréal. His research interests are machine learning and applied math, with applications to computer vision and natural language processing. He obtained a B.Sc. in math., physics and computer science from McGill, a PhD in computer science from UC Berkeley and a post-doc from the University of Cambridge. He spent a few years as a research faculty at INRIA and École normale supérieure in Paris before coming back to his roots in Montreal in 2016.
Simon published several papers at the intersection of mathematical programming and machine learning, and in particular for solving min-max games. He is a frequent participant to the NeurIPS OPT workshop series, and co-organized the NeurIPS 2009 workshop on “The Generative & Discriminative Learning Interface” .