Supervised Contrastive Learning

The method uses class labels to define positive and negative example pairs, training an encoder to produce representations that help distinguish categories, often alongside a classification loss. Common approaches include supervised contrastive loss, prototype and angular objectives, and weighted or hard-negative pair selection; extensions apply these ideas across modalities and network layers. Researchers use it for image and text classification, speech recognition, medical imaging, and long-tailed settings, with benchmarks such as ImageNet-LT, IEMOCAP, and MIMIC-CXR. A central challenge is choosing pairs and loss geometry that improve class separation without discarding useful within-class variation, especially when classes are imbalanced or labels are scarce.