Machine Learning's Dropout Training is Distributionally Robust Optimal