Gradient Descent on Neural Networks Typically Occurs at the Edge of Stability.

Jeremy Cohen, Simran Kaur, Yuanzhi Li, J. Zico Kolter, Ameet Talwalkar

Browse the full ICLR paper archive.