Convergence of gradient descent for deep neural networks

Sourav Chatterjee

We give a simple local Polyak-Lojasiewicz (PL) criterion that guarantees linear (exponential) convergence of gradient flow and gradient descent to a zero-loss solution of a nonnega…