A Unifying Theory of Distance to Calibration

Explore a comprehensive Google TechTalk presented by Jarosław Błasiok on the fundamental question of defining and measuring the distance from calibration for probabilistic predictors. Delve into a rigorous framework for analyzing calibration measures, inspired by property testing literature. Discover the proposed ground-truth notion of distance from calibration and learn about three consistent calibration measures: smooth calibration, interval calibration, and Laplace kernel calibration. Examine the information-theoretic optimal quadratic approximations to the ground truth distance and understand the fundamental lower and upper bounds on measuring distance to calibration. Gain insights into the theoretical justification for preferring certain metrics in practice, based on joint work with Parikshit Gopalan, Lunjia Hu, and Preetum Nakkiran. Explore the speaker's background in Theoretical Computer Science, including his work in streaming algorithms, error-correcting codes, machine learning, differential privacy, and compressed sensing.

A Unifying Theory of Distance to Calibration