A New Paradigm for Learning with Distribution Shift

Watch a 48-minute lecture from the Simons Institute where Adam Klivans from the University of Texas at Austin presents groundbreaking research on learning with distribution shift. Explore a novel approach called TDS learning that addresses the challenges of training AI systems when test and training data distributions differ. Learn how this paradigm allows for detecting distribution shifts and guaranteeing low test error without making assumptions about test distributions. Discover how this innovative method combines concepts from pseudorandomness, property testing, and sum of squares proofs to create efficient learning algorithms for well-studied function classes. Understand the limitations of traditional generalization bounds based on distribution distances and see how this new framework overcomes previous computational barriers in the field.