Contrastive Attraction and Contrastive Repulsion for Representation Learning

Contrastive learning (CL) methods effectively learn data representations without label supervision, where the encoder contrasts each positive sample over multiple negative samples via a one-vs-many softmax cross-entropy loss. By leveraging large amounts of unlabeled image data, recent CL methods have achieved promising results when pre-trained on ImageNet, a well-curated data set with balanced image classes. However, they tend to yield worse performance when pre-trained on images in the wild. In this paper, to further improve the performance of CL and enhance its robustness on uncurated data sets, we propose a doubly CL strategy that contrasts the positive (negative) samples of a query within themselves before deciding how strongly to pull (push) them. We realize this strategy with contrastive attraction and contrastive repulsion (CACR), which makes the query not only exert a greater force to attract more distant positive samples but also do so to repel closer negative samples. Theoretical analysis reveals that CACR generalizes CL's behavior by taking into consideration the differences between the distributions of the positive/negative samples, which in general are sampled independently of the query, and their true conditional distributions given the query. We demonstrate this unique intra-positive attraction and intra-negative repulsion mechanism, which helps remove the need to assume uniform prior distributions on both the data and their latent representation, is particularly beneficial when data sets are less curated. Extensive large-scale experiments on a number of standard vision tasks show that CACR not only consistently outperforms existing CL methods on benchmark data sets in representation learning, but also shows better robustness when generalized to pre-training on imbalanced image data sets.