Title: An entropy formula for the deep linear network
Speaker: Tianmin Yu
Speaker Info: Brown University
Brief Description: An entropy formula for the deep linear network
Special Note:
Abstract:
Deep Linear Network (DLN) is a simplified model of deep neural network with all active functions being identity. In recent work (Bah et al., 2020), a Riemannian metric structure in the gradient flow of learning DLN is observed. In this talk I will introduce a geometric framework for finding entropy formulas based on theories of Brownian motion on Riemannian manifolds, and explain how this framework applies to DLN and helps us understand overparameterization problems.Date: Tuesday, May 14, 2024