Cross-Domain Generalization

Cross-domain generalization asks whether a model trained in one setting can work in a different one, such as a new hospital, camera, country, or sensor. Common approaches learn domain-invariant features, diversify training data with augmentation or simulated domains, align representations across domains, and use causal or pretrained models to identify transferable structure. Researchers test these methods with zero-shot and few-shot transfer, leave-one-domain-out protocols, and benchmarks such as DomainBed, cross-site medical imaging, and cross-dataset detection. A central tension is that removing domain-specific information can also discard useful signals, while strong in-domain results often fail to predict performance on genuinely unseen domains.