Events

MAD Seminar: Participation Dynamics in Learning Systems

Speaker: Sarah Dean

Location: 60 Fifth Avenue, Room 150

Date: Thursday, February 16, 2023

The choice to participate in a data-driven system, often made on the basis of the quality of that system, influences the ability of the system to learn and improve. Participation choices manifest as distribution shifts which are partially endogeneous, i.e. caused by the machine learning system itself. In this talk, I will discuss participation dynamics in the presence of multiple learners. We introduce and study a general class of loss-reducing dynamics, in which learners retrain to improve predictive performance and users shift participation towards better performing learners. We characterize the stable equilibria and discuss the implications in terms of social welfare and fairness. Based on joint work with Mihaela Curmei, Maryam Fazel, Jamie Morgenstern, and Lillian Ratliff.