Domain-Adversarial Learning
Overview
Domain-adversarial learning studies how to make models work across domains, such as different hospitals, devices, users, imaging modalities, or time periods. A common approach uses a domain discriminator and gradient reversal to encourage domain-invariant features, often combined with feature alignment, contrastive learning, class prototypes, or pseudo-labeling. These methods are used in supervised and unsupervised adaptation, including settings where target-domain labels are scarce or unavailable. A central tension is that removing domain information can also discard useful task information: alignment may fail under label shift or harm representations that already encode relevant domain knowledge, so invariance does not guarantee better transfer.
Latest papers
35 papers6 papers in the last four weeks, up 100% on the four weeks before. 0.0% of all new papers.