1 h 2 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Benefits of Early Stopping in Gradient Descent for Overparameterized Logistic Regression
UC Berkeley
Recording
Abstract
In high-dimensional logistic regression, training until convergence can yield poor calibration even when the direction of gradient descent approaches a maximum-margin solution. The talk contrasts the excess logistic and classification risks of early-stopped and interpolating predictors, derives sample requirements, and relates early stopping to explicit squared-norm regularization.