Seminar, Hosted by Kempner Institute, Harvard University
Some Very Old and Very New Problems in Learning Theory
Weizmann Institute of Science
Abstract
Two recent theoretical results on the limits of neural-network learning. The first part examines model collapse: what happens to models that are repeatedly retrained on data containing synthetic samples generated by earlier models. The second part asks when and why standard gradient-based learning fails, and presents a framework for proving computational lower bounds for neural networks trained with standard stochastic gradient descent.