Contrastive Learning

Contrastive learning trains encoders so that related pairs (augmentations, text–image matches) are close in embedding space and unrelated pairs are far apart.

InfoNCE-style losses underpin CLIP-like models and many embedding models used for retrieval.