Identifying and attacking the saddle point problem in high-dimensional non-convex optimization

W e apply this algorithm to deep or recurrent neural network training, and provide numerical evidence for its superior optimization performance.