Identifying and Attacking the Saddle Point Problem in High-dimensional Non-convex Optimization