Date: Wednesday, November 16th, 2022
9:00 am – 10:00 am Pacific Time
12:00 pm – 1:00 pm Eastern Time
Location: Weekly Seminar, Zoom
Title: The many performances of performative prediction
When algorithmic predictions are used to inform social decision-making, these predictions don’t just forecast the world around it: they actively shape it. From online recommender platforms, to financial predictions, machine learning systems are in active feedback with the surrounding environment and have the ability to steer the underlying data distributions towards different targets. While traditionally neglected, these causal forces of prediction have been recently formalized in a new risk minimization framework called performative prediction [PZMH20].
Following a brief overview of the performative prediction framework, in this talk, I will present some work which directly tries to address the distinction between forecasting future outcomes, and steering data distributions towards socially desirable targets. Building upon a new line of research in supervised learning [GKRSW21, GHKRW22], we introduce the idea of performative omnipredictors. These are simple predictive models that simultaneously encode the optimal decision rule with respect to many possibly-competing objectives. As part of our presentation, we will discuss some connections with the outcome indistinguishability literature [DKRRY21] and illustrate how the solution concepts we introduce enable decision-makers to flexibly decide on the goals of prediction in performative settings.
Juan Perdomo is a final-year graduate student at the University of California, Berkeley in EECS where he is advised by Moritz Hardt and Peter Bartlett. His research centers around the foundations of machine learning within social systems. He received his BA in CS and Math from Harvard and was previously supported by an NSF graduate student research fellowship.