The paper observes the similarity between the stochastic optimal control over discrete dynamical systems and the lear ning multilayer neural networks. It focuses on contemporary deep networks with nonconvex nonsmooth loss ...