1 paper
Cheng Shi, Liming Pan, Ivan DokmaniÄ
Feature-learning deep nets progressively collapse data to a regular low-dimensional geometry. How this emerges from the collective action of nonlinearity, noise, learning rate, and…