High-Dimensional Prediction for Sequential Decision Making

Explore high-dimensional prediction techniques for sequential decision making in this Google TechTalk presented by Georgy Noarov. Dive into efficient algorithms for making online predictions of adversarially chosen high-dimensional states, tailored to downstream decision makers. Discover applications including no-regret guarantees over large action spaces, high-dimensional best-in-class results, fairness guarantees, and a novel framework for uncertainty quantification in multiclass settings. Learn about tractable algorithms offering optimal swap regret for multiple decision makers, and efficient no-regret algorithms for online combinatorial optimization. Examine how these methods can be applied to online routing problems and extensive-form games. Investigate new approaches to uncertainty quantification in machine learning, including multiclass probability vector estimation with fairness guarantees and best-in-class predictions. Gain insights into alternatives to conformal and top-K prediction paradigms, and understand the implications for high-dimensional omniprediction.

High-Dimensional Prediction for Sequential Decision Making