1 h
Seminar, Hosted by Simons Institute for the Theory of Computing
On continual learning with gradient descent for neural networks
UC San Diego
Recording
Abstract
Arya Mazumdar studies forgetting in one-hidden-layer quadratic networks trained by gradient descent on sequential noisy XOR classification tasks with orthogonal cluster means. The analysis bounds train- and test-time forgetting as the sample size, update count, width and number of tasks change. Numerical experiments probe how the relationships carry over beyond the analyzed setting.