Multi-network contrastive learning of visual representations

摘要

Contrastive learning, as an important way of self-supervised learning, has achieved great success in visual representations, which significantly reduces its gap with supervised learning. The essential strategy is to maximize the similarities between two augmented views of the same image (positive pairs) and to make such image easily distinguishable from other images of different types (negative pairs). Previous methods rely heavily on a large number of negative samples, such as SimCLR and MoCo. However, some recently proposed methods, such as BYOL and SimSiam, discard negative samples by introducing asymmetric structures. This paper proposes a multi-network contrastive learning methodology for visual representations (MNCLR), which integrates the end-to-end and the momentum encoder mechanisms to introduce more negative samples under a multi-network framework. The classification results on three benchmark image datasets demonstrate that the proposed MNCLR algorithm outperforms some classic contrastive learning methods.