Self Supervised Contrastive Learning Combining Equivariance and Invariance

Current self-supervised representation learning methods are mainly based on contrastive learning and proxy tasks. These methods acquire semantically rich features by contrasting samples with invariant transformations (positive pairs) against other samples (negative pairs), and simply discard transformations that degrade performance when used as invariances. However, using only invariant transformations often leads to an over-reliance on invariant transformations, which affects the generalisation ability and robustness of the model, while the large number of negative sample pairs in contrast learning imposes a huge computational overhead. In order to address these issues, we reduce the dependence on invariant transformations by transforming the discarded invariant transformations into equivariant transformations. In contrast learning, we reduce the computational overhead by using only positive pairs to obtain semantically rich features. Specifically, we enhance feature semantic quality by encouraging certain transformations to exhibit non-trivial equivariance on samples of invariant transformations in the form of a proxy task, while preserving original transformation invariance. The model learns the invariant transformations further by learning equivariance at the same time, and our approach can improve the accuracy of the model without changing the structure of the original model. Experimental results show that significant improvements are obtained on several benchmark datasets.