Beyond cross-entropy: learning highly separable feature distributions for robust and accurate classification.