Special Seminar in CMS and HSS
Eric Mazumdar is a Ph.D Candidate in Electrical Engineering and Computer Science at UC Berkeley co-advised by Michael Jordan and Shankar Sastry. He is a member of the Berkeley AI Research Lab (BAIR) and the RISELab. His research is on the mathematical foundations of machine learning. He aims to understand when learning algorithms succeed and fail in dynamic environments. In particular, he is interested in understanding the consequences of using machine learning algorithms in economic settings, as well as designing efficient learning algorithms with provable guarantees of convergence in competitive and multi-agent settings. Using tools from dynamical systems and stochastic processes, his research so far has looked at multi-agent learning, min-max optimization, multi-armed bandits, and learning for control.
The ability to learn from data and make decisions in real-time has led to the rapid deployment of machine learning algorithms across many aspects of everyday life. Despite their potential to enable new services and address persistent societal issues, the widespread use of these algorithms has led to unintended consequences like flash crashes in financial markets or price collusion on e-commerce platforms. These consequences are the inevitable result of deploying algorithms--- that were designed to operate in isolation--- in uncertain dynamic environments in which they interact with other autonomous agents, algorithms, and human decision makers.
To address these issues, it is necessary to develop an understanding of the fundamental limits of learning algorithms in societal-scale systems. In this talk, I will give an overview of my work on three aspects of learning and decision-making in societal-scale systems: (i) Understanding why and when learning algorithms fail in game theoretic settings, (ii) Learning expressive models of human decision-making from data, and (iii) Bayesian decision-making in uncertain dynamic environments.