Unsupervised Domain Adaptation

Unsupervised domain adaptation studies how to transfer a model trained on labeled source data to a target setting where the data distribution differs and labels are unavailable. Common approaches align source and target features with adversarial losses such as DANN, discrepancy measures such as MMD or CORAL, and self-training with pseudo-labels and consistency regularization; newer work also uses foundation models, synthetic data, and parameter-efficient or test-time updates. Methods are tested across shifts in imaging devices, weather, sensors, languages, and graph structure, often on benchmarks such as Office-Home, DomainNet, and Cityscapes. A central tension is that alignment can blur class distinctions or reinforce incorrect pseudo-labels, while choosing a reliable model without target labels remains difficult, especially under label shift or large real-world changes.