35 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Principles of learning in distributed neural networks
University College London
Recording
Abstract
Andrew Saxe uses solvable deep-network models to study how successive stages of processing influence the dynamics of learning. He relates the models to rodent experiments in which mice change strategies during weeks of visual detection training. The comparison examines how environmental statistics and nonlinear learning dynamics shape both representations and behavior.