MIT Center for Brains, Minds and Machines researchers provide one of the first theoretical analyses covering optimization, generalization, and approximation in deep networks and offers new insights into the properties that emerge during training.