Stochastic Gradient Descent Escapes Saddle Points Efficiently