Contrastive self-supervision is an approach in the field of self-supervised learning where a model learns to generate representations that bring similar data points closer together and push different ones further apart.
This is usually done by using so-called positive and negative pairs, where positive pairs are different views or transformations of the same example.
Typical techniques involve the use of contrastive loss functions such as the InfoNCE loss, which instruct the model to maximize the similarity between positive pairs and minimize it between negative pairs.
This method enables learning useful features that can be used for various tasks such as classification or clustering without requiring explicit labels.