Mixup: Beyond Empirical Risk Minimization

A Stabilizing data augmentation technique for image classifiers and GANs. Create inputs as a linear blend of two training images, with a matching linear blend of their labels as the target.