Supervised Contrastive Learning


  • generalize contrastive learning used in self-supervised learning to the fully supervised setting
  • Clusters of points belonging to the same class are pulled together in embedding space, while simultaneously pushing apart clusters of samples from different classes.
  • Achieves top-1 accuracy of 81/4% on ResNet-200
  • self-supervised contrastive loss contrasts a single positive against a set of negatives