Abstract: I've been thinking about how robots and AI agents more broadly can optimize for what we actually want as end users for a while now. I'll take the opportunity to reflect on what we've been able to accomplish in this area, as well as what's missing.
(RLHF = Reinforcement Learning from Human Feedback)

Bio: Prof. Anca Dragan is an Associate Professor in the EECS Department at UC Berkeley. Her goal is to enable robots to work with, around, and in support of people. She runs the InterACT Lab, where they focus on algorithms for human-robot interaction -- algorithms that move beyond the robot's function in isolation, and generate robot behavior that coordinates well with people, and is aligned with what they actually want the robot to do and work across different applications, from assistive arms, to quadrotors, to autonomous cars, and draw from optimal control, game theory, reinforcement learning, Bayesian inference, and cognitive science. She also helped found and serve on the steering committee for the Berkeley AI Research (BAIR) Lab, and is a co-PI of the Center for Human-Compatible AI. She has been honored by the Sloan Fellowship, MIT TR35, the Okawa award, an NSF CAREER award, and the PECASE award.

Abstract: Is it a good idea to use machine learning (ML) predictions in education? Would machine learning models treat all students fairly? I will start this talk with our recent analysis on middle school and high school datasets that reveal potential fairness risks of applying vanilla ML on students. To improve fairness in ML for education, there are several practical challenges. First, there are missing values in the datasets that are not evenly distributed across groups (e.g., female and male) which could aggravate the ML model's bias. I will show a fundamental limit of learning with missing values and propose a decision-tree-based algorithm that outperforms state-of-the-art fair ML methods that do not consider missing values. In the second part, we address how to correct bias in a classifier with low-cost post-processing when we have multi-class labels and sensitive attributes. I will introduce the Fair Projection algorithm which utilizes the idea of “information projection” and how it can be applied to a wide range of classifiers while maintaining a competitive fairness-accuracy trade-off.

Bio: Haewon Jeong is an assistant professor of Electrical and Computer Engineering at the University of California Santa Barbara. She received the B.S. degree ('14) in Electrical Engineering from KAIST and the M.Sc. ('16) and Ph.D. ('20) degrees in Electrical and Computer Engineering from Carnegie Mellon University. From 2020 to 2022, she was a postdoctoral scholar at Harvard University. Her research interests include information theory, distributed computing, machine learning, and ethics of AI systems.