M-F AC: Efficient Matrix-Free Approximations of Second-Order Information

Efficiently approximating local curvature information of the loss function is a key tool for optimization and compression of deep neural networks.