The paper observes the similarity between the stochastic optimal control of discrete dynamical systems and the training multilayer neural networks. It focuses on contemporary deep networks with nonconvex nonsmooth loss and activation functions. In the paper, the machine learning problems are treated as nonconvex nonsmooth stochastic optimization problems. As a model of nonsmooth nonconvex dependences, the so-called generalized differentiable functions are used. A method for calculating stochastic generalized gradients of the learning quality functional for such systems is substantiated basing on Hamilton-Pontryagin formalism. This method extends a well-known “backpropagation” machine learning technique to nonconvex nonsmooth networks. Stochastic generalized gradient learning algorithms are extended for training nonconvex nonsmooth neural networks.
Preprint, September 2019. V.M.Glushkov Institute of Cybernetics of the National Academy of Sciences of Ukraine (NASU), Kyiv. To appear in Dopovidi NASU.