Date: September 21, 2026
Title: Omniprediction and Outcome Indistinguishability
Speaker: Michael P. Kim, Assistant Professor, Computer Science, Cornell University
Abstract: Omnipredictors are simple prediction functions that encode loss-minimizing predictions, simultaneously for every loss function within an immense class of losses L. Somewhat surprisingly, recent work demonstrates that this stringent learning guarantee is possible at (essentially) the same cost as loss minimization *for a single loss function*. In this overview talk, I will present the definition of Omniprediction, highlight the state-of-the-art bounds, and present a learning framework, called Outcome Indistinguishability, that plays an important role in establishing the existence of omnipredictors and may be of independent interest.
Based on joint works, including [Dwork,Kim,Reingold,Rothblum] arxiv.org/abs/2011.13426, [Gopalan,Kim,Hu,Reingold,Wieder] arxiv.org/abs/2210.08649, [Okoroafor,Kleinberg,Kim] arxiv.org/abs/2501.17205