Why Deep Learning Works: Implicit Self-Regularization in Deep Neural Networks