Toward Understanding the Extrapolation of Nonlinear Models to Unseen Domains

Watch a 42-minute lecture from Stanford University professor Tengyu Ma at the Simons Institute exploring the complex challenges and theoretical frameworks surrounding how nonlinear models perform when applied to previously unseen domains or when attempting to generalize across an entire domain space. Delve into key concepts of domain adaptation and examine the mathematical foundations that govern how machine learning models extrapolate beyond their training data. Learn about the limitations, possibilities, and emerging research directions in understanding model behavior when faced with out-of-distribution scenarios or when required to make predictions across the complete domain spectrum.