Contrastive Learning

What Contrastive Learning is

Contrastive objectives learn what makes two things similar without needing class labels, which is how text-image models learn to connect captions with pictures.

How it works

Positive pairs come from augmentations of the same item or naturally paired data such as image and caption; negatives come from other items in the batch. A temperature-scaled similarity loss shapes the space.

Why it matters

It produced the embedding models behind semantic and cross-modal search, and remains a core self-supervised technique.

Common uses

• →Text and image embedding models
• →Cross-modal retrieval
• →Representation pretraining