Domain Generalization

Domain generalization studies how models trained on one or more source environments can make reliable predictions when data shift across hospitals, cameras, sensors, users, locations, or time. Common approaches include learning invariant or disentangled features, causal modeling, adversarial and style-based augmentation, and adapting pretrained models with prompts, adapters, or mixtures of experts. Evaluation often uses leave-one-domain-out and cross-dataset tests, including benchmarks such as PACS and OfficeHome, alongside domain-specific measures of calibration and robustness. A central tension is that enforcing invariance can discard useful information or cause negative transfer, while many proposed methods offer only modest gains over empirical risk minimization and may not reflect deployment conditions.