Contrastive Learning with Latent Augmentation

Contrastive learning with well-designed data augmentation has recently greatly improved the quality of self-supervised representation learning. However, existing works only concentrate on various forms of hand-crafted data augmentation on raw image pixels. We explore data augmentation in feature space and find that it can achieve a similar effect with augmentation on image pixels in contrastive learning. Based on this finding, we present Contrastive Learning with Latent Augmentation (CLLA), which is a new automatic data augmentation strategy that augments intermediate features for contrastive learning. Specifically, aiming at automatic data augmentation, CLLA generates adversarial examples of intermediate features and then leverages them as feature augmentations. Moreover, we extend CLLA to the multi-level data augmentation framework M-CLLA which regards data augmentation on image pixels as a special case. Extensive experiments demonstrate that CLLA and M-CLLA achieve better performance over strong baselines on multiple visual benchmarks.