Supervised Asymmetric Contrastive Learning of Visual Representations

Contrastive learning, as an unsupervised representation learning method, learns representations by exploring intrinsic relationships between sample pairs, achieving notable success in many visual tasks. However, the lack of class labels in contrastive learning means that different instances from the same class may be considered negative pairs, disrupting representation learning and weakening the model’s discrimination ability. The introduction of SupCon partially addresses this issue, but its base model, SimCLR, requires the construction of numerous negative pairs to avoid model collapse, demanding large batch sizes and imposing stringent hardware requirements.In this paper, we propose a novel supervised asymmetric contrastive learning model SupAsCon, based on the self-supervised contrastive learning framework BYOL. By combining asymmetric structure with supervised learning, SupAsCon prevents model collapse while enhancing class recognition abilities, especially excelling in small-batch training. Experimental results demonstrate that SupAsCon outperforms SupCon in terms of performance.