Saddle-free Hessian-free Optimization

Martin Arjovsky

Nonconvex optimization problems such as the ones in training deep neural networks suffer from a phenomenon called saddle point proliferation. This means that there are a vast numbe…