On the regularized risk of distributionally robust learning over deep neural networks