The Convex Geometry of Backpropagation: Neural Network Gradient Flows Converge to Extreme Points of the Dual Convex Program.

Yifei Wang, Mert Pilanci

Browse the full ICLR paper archive.