Serena Booth
Date:
Speaker
Serena Booth is an Assistant Professor of Computer Science at Brown University. Her research develops methods for enabling AI systems and robots to learn from imperfect human input. She studies reward design and other approaches for inferring and evaluating AI specifications to build systems that better align with human intentions. Serena also studies the governance of AI systems. She served as a AAAS AI Policy Fellow in the U.S. Senate, where she worked on AI, consumer protection, economic policy, and labor, and she currently co-chairs the ACM Technology Policy Subcommittee on AI and Algorithms. She received her Ph.D. from MIT CSAIL in 2023 and her A.B. from Harvard College in 2016.
Speaker Links: Website · Brown CS · Email: serena_booth@brown.edu
Abstract
We often use human feedback to specify the objectives of AI systems and robots, whether through demonstrations, preferences, corrections, or natural language feedback. These signals are frequently used to infer reward functions, but they are often noisy, biased, or influenced by the solicitation process, making it difficult to recover the underlying objectives they are intended to encode. In this talk, I will present computational models of human decision-making for interpreting imperfect human feedback. I will also discuss methods for designing interventions that elicit more informative feedback and more accurately reveal users’ intended reward functions.