Contrastive learning is a learning method aimed at producing a representation of data where similar data points lie closer together in the feature space than dissimilar ones.
This is achieved by using pairs or triplets of examples, where positive pairs (similar examples) are brought together and negative pairs (dissimilar examples) are pushed apart.
Typical approaches use loss functions such as triplet loss or InfoNCE loss to optimize distances in the feature space.
Contrastive learning is particularly effective in self-supervised learning, as it works without explicit labels and produces robust features for further tasks such as classification or clustering.
Well-known applications can be found in image processing, speech processing, and in the pretraining of neural networks.