supervised-contrastive-learning

This paper shows that cross-entropy and supervised contrastive learning are both forms of prototype learning on the hypersphere and proposes normalized losses (NTCE and NONL) that achieve Neural Collapse by design, outperforming standard methods.