Unsupervised Representation Learning

Unsupervised representation learning studies how to turn unlabeled data into features that reveal useful structure for tasks such as clustering, retrieval, anomaly detection, and prediction. Common approaches include autoencoders, contrastive learning, clustering and prototype methods, graph and spectral embeddings, and models that use symmetry or temporal structure. Methods are tested across settings including images, graphs, biomedical data, and time series, often by measuring both representation quality and performance on downstream tasks. A central challenge is balancing compact, invariant features against preserving information, while ensuring that learned structure generalizes and does not encode unwanted biases.