Disentangled Representation Learning

Disentangled representation learning studies how to encode separate factors—such as content and style, identity and viewpoint, or disease and acquisition conditions—in distinct parts of a representation. Common approaches include variational autoencoders, contrastive objectives, sparse or orthogonal subspaces, causal models, and symmetry-based methods; many applications use the resulting factors for editing, robust prediction, or transfer across domains. Work now spans images, speech, graphs, time series, medicine, and robotics, with increasing attention to interventions and practical controllability. A central tension remains: factors are generally not identifiable from unlabelled data alone without assumptions or useful supervision, and standard disentanglement metrics can fail when their assumptions do not hold.