Cross-Domain Transfer Learning

Cross-domain transfer learning studies how a model trained in one setting can help in another, especially when target labels are scarce and data differ across hospitals, sensors, cities, or tasks. Common approaches include fine-tuning and parameter-efficient adapters, feature or distribution alignment, shared and domain-specific representations, and selective transfer using retrieval, prototypes, or expert routing. Work spans vision, medical prediction, graphs, language models, and reinforcement learning, with evaluations increasingly testing unseen domains, few-shot adaptation, calibration, and transfer in both directions. A central tension is that shared knowledge can improve accuracy and data efficiency but can also cause negative or asymmetric transfer, so in-domain results and simple measures of domain similarity do not reliably predict deployment performance.