Decoupled Contrastive Learning

Contrastive learning is one of the most successful paradigms for self-supervised learning (SSL). In a principled way, it considers two augmented views of the same image as positive to be pulled closer, and all other images as negative to be pushed further apart.