Review Contrastive Learning

Contrastive training is a method where a model learns by pulling related examples closer together and pushing unrelated examples farther apart.

Contrastive learning teaches a model by comparing examples. Representations of related examples are pulled closer together, while representations of unrelated examples are pushed apart.

The method can learn useful representations without a manually written label for every example. Its behavior still depends on how positive and negative pairs are created.

Two altered views of the same photograph are treated as a positive pair, while views from different photographs are treated as negatives.

Contrastive learning is not automatically unbiased or label-free. The pairing rules act as supervision and can encode assumptions about what should count as similar.

Bring related examples together; separate the rest.

* [[self-supervised-learning|Self-Supervised Learning]] * [[embedding|Embedding]] * [[latent-representation|Latent Representation]] * [[training-data|Training Data]]

* [Ting Chen and colleagues, “A Simple Framework for Contrastive Learning of Visual Representations”](https://arxiv.org/abs/2002.05709) — demonstrates a contrastive objective that aligns two augmented views of an example.

* Compare [[self-supervised-learning|Self-Supervised Learning]], [[embedding|Embedding]], [[latent-representation|Latent Representation]], and [[training-data|Training Data]] to place Contrastive Learning in the surrounding system. * Follow the cited source material above for the method, standard, or research context behind this draft.