59 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Generalization and overfitting in two-layer neural networks
Recording
Abstract
Implicit bias, benign overfitting, and the emergence of useful features can depend on how an overparameterized network changes during training. Using dynamical mean-field theory from statistical physics, the speaker analyzes nonlinear training dynamics in large two-layer networks and examines when generalization and overfitting arise.