Drawing Multiple Augmentation Samples Per Image During Training Efficiently Decreases Test Error

In computer vision, it is standard practice to draw a single sample from the\ndata augmentation procedure for each unique image in the mini-batch. However\nrecent work has suggested drawing multiple samples can achieve higher test\naccuracies. In this work, we provide a detailed empirical evaluation of how the\nnumber of augmentation samples per unique image influences model performance on\nheld out data when training deep ResNets. We demonstrate drawing multiple\nsamples per image consistently enhances the test accuracy achieved for both\nsmall and large batch training. Crucially, this benefit arises even if\ndifferent numbers of augmentations per image perform the same number of\nparameter updates and gradient evaluations (requiring the same total compute).\nAlthough prior work has found variance in the gradient estimate arising from\nsubsampling the dataset has an implicit regularization benefit, our experiments\nsuggest variance which arises from the data augmentation process harms\ngeneralization. We apply these insights to the highly performant NFNet-F5,\nachieving 86.8$\\%$ top-1 w/o extra data on ImageNet.\n