Contrastive coding is a learning method in which a model is trained to map representations of similar data points closer together and dissimilar ones further apart in the feature space.
Typically, pairs of examples are used: positive pairs (similar) and negative pairs (dissimilar).
The model optimizes a loss function, such as the contrastive loss or the InfoNCE loss, to enforce this structure.
This technique is often applied in self-supervised learning approaches to generate robust and meaningful feature vectors that can later be used for classification or other tasks.
An example is image or text representations, where different views or transformations of the same object are considered positive.