There are no existing augmentation techniques that can correct adataset that has very poor diversity with respect to the testing data. All these augmenta-tion algorithms perform best under the assumption that the training data and testingdata are both drawn from the same distribution.

suggests that it is best to initially train with the original data only and then finish trainingwith the original and augmented data,

A robust classifier is thus defined as having alow variance in predictions across augmentations.

By taking a test image and augmenting it in thesame way as the training images, a more robust prediction can be derived.