1 paper
Liu Ziyin, Botao Li, Tomer Galanti +1
Characterizing and understanding the dynamics of stochastic gradient descent (SGD) around saddle points remains an open problem. We first show that saddle points in neural networks…