Date: Wednesday, February 26th, 2025
9:00 am – 10:00 am Pacific Time
12:00 pm – 1:00 pm Eastern Time
Location: Weekly Seminar, Zoom
Title: Tractable agreement protocols
Abstract:
We present an efficient method to transform any machine learning model into an interactive system that collaborates with another party, such as a human, to improve prediction accuracy and reach a consensus through iterative feedback. Each party operates under computationally and statistically tractable conditions that generalize Bayesian rationality. These conditions apply even in settings without prior knowledge, significantly generalizing Aumann’s foundational “agreement theorem.” In our protocol, the machine learning model begins by making a prediction. The human then responds either by agreeing or by providing feedback, prompting the model to update its prediction. This back-and-forth continues until both parties reach a consensus. In this presentation we focus on the one-dimensional setting, recovering state of the art convergence results under weaker assumptions. Our approach also extends to multi-dimensional predictions and multi-party settings with minimal added complexity. Our protocols are based on simple, efficiently maintainable conditions and result in predictions that are more accurate than any single party’s alone.
Bio:
Natalie Collina is a PhD student in computer science at the University of Pennsylvania, where she is advised by Michael Kearns and Aaron Roth. She works in the intersection of algorithmic game theory and online learning, and is particularly interested in understanding repeated strategic interactions, and the algorithms that agents employ in these settings. Her research is supported by an IBM PhD Fellowship and has previously been supported by AWS AI . She also co-leads the University of Pennsylvania’s weekly Theory Seminar.
Website: https://www.seas.upenn.edu/~ncollina/