Preventing Arbitrarily High Confidence on Far-Away Data in Point-Estimated Discriminative Neural Networks