Regularized stochastic representation for reliable contrastive learning

Abstract

Contrastive learning, which has recently received a lot of attention, exploits multi-views of instances (images) to learn view-invariant representation. Conventionally, the multi-views are generated by composition of multiple stochastic augmentations. Since the contrastive learning methods so far implicitly assumed that the generated multi-views are always an appropriate positive pair, their representations were made to be recklessly close in the representation space. However, in the case of complex images, which is common in real-world, inappropriate view pairs are likely to be generated. It likely results in learning problematic representation by encouraging representations to be closer even though the pair is not appropriate. To alleviate this problem, we propose to use a regularized stochastic representation considering the adequacy of the given view pair. With this, we devised a novel method that the model can attenuate the effect from inappropriate pairs. The proposed method consistently outperforms baselines for various downstream tasks (image classification, object detection) on various benchmark datasets (CIFAR-100, ImageNet-100, COCO).