Edge of Stability

Unstable Power: How Sharpness Drives Deep Network Learning

A larger learning rate leaves the NTK better aligned with the labels at the Edge of Stability: Kernel Target Alignment jumps exactly when sharpness drops.