Supplemental Material: On Interaction Between Augmentations and Corruptions in Natural Corruption Robustness

Here we describe an alternative experiment that shows how the introduction of dissimilar augmentations affects corruption error. We expect that this should be the case, since neural networks are often good at memorizing rare examples. Here we show the effect using a real augmentation scheme. Even though training on AugMix with our training parameters would normally would produce 5 million uniquely sampled augmentations, only around 1000 are needed to achieve equivalent performance. This work completed as part of the Facebook AI residency program.