Brian Ziebart “Superhuman Imitation Learning”
Please join us on Tuesday, October 24, 2023 at 12:00pm in Clark Hall Room 110 and on ZOOM for the CIS & MINDS Seminar Series:
Guest: Brian Ziebart
Associate Professor
University of Illinois Chicago
Topic: “Superhuman Imitation Learning”
Join Zoom Meeting: https://wse.zoom.us/j/95386212146
Passcode: cis&minds
If you would like to meet with Brian Ziebart, please sign up at this link: https://docs.google.com/spreadsheets/d/1CeFz-KbXQm0J7ZgauSmsp2qdYwBeAoKE-mlhpG4UHUY/edit#gid=1192317842
Brian Ziebart
Associate Professor
University of Illinois Chicago
“Superhuman Imitation Learning”
Abstract: Significantly outperforming the best humans when driving a car, providing medical diagnosis/treatment decisions, or engaging in conversation represent ambitious, but important motivating goals for artificial intelligence research. One popular approach is to estimate the reward function that motivates demonstrated human behavior, and then employ reinforcement learning (RL) to find a policy that optimizes the estimated reward function better than the demonstrators do. When the demonstrators are inherently less capable than autonomous systems (e.g., due to slower reaction time, motor control imprecision, limited recall), the RL part of this approach becomes easy, but reward estimation becomes error-prone because capability limitations are often difficult to model. Aggressively optimizing the wrong reward function (i.e., value misalignment) can result. This talk presents subdominance minimization, a new objective for imitation learning that improves with larger AI-human capability gaps by seeking policies that are unambiguously better than demonstrations.
Biography: Brian Ziebart is an Associate Professor in the Department of Computer Science at the University of Illinois Chicago. He earned his PhD in Machine Learning from Carnegie Mellon University where he was also a postdoctoral fellow. From 2020-2021, he was a Software Engineer working on autonomous driving at Aurora Innovation, Inc. His interests include imitation learning, distributionally robust optimization methods, and fair machine learning. He has published over 40 articles in leading machine learning, artificial intelligence, and robotics venues, including a Best Paper at the International Conference on Machine Learning.