Definition
Contrastive loss trains representations by pulling matched pairs together and pushing unmatched pairs apart in embedding space. It’s the workhorse of self-supervised representation learning — CLIP, SimCLR, sentence embeddings — and lives or dies on how you mine negatives.