MetaAugment: Sample-Aware Data Augmentation Policy Learning

Automated data augmentation has shown superior performance in image recognition. Existing works search for dataset-level policies without considering individual sample variations, which are likely to be sub-optimal. On the other hand, learning different samples naively could greatly increase computing cost. In this paper, we learn a sample-aware policy efficiently by formulating it as reweighting problem. Specifically, an network takes transformation and corresponding augmented inputs, outputs weight adjust loss computed task network. At training stage, minimizes weighted losses of images, while on validation set via meta-learning. We theoretically prove convergence procedure further derive exact rate. Superior is achieved widely-used benchmarks including CIFAR-10/100, Omniglot, ImageNet.