Domain Generalization

Domain generalization is the problem of training a machine learning model on data from one or more source domains and deploying it on a target domain whose distribution differs in ways that are unknown at training time. Unlike transfer learning, which typically assumes access to target-domain data for fine-tuning, domain generalization demands robustness without adaptation — the system must have already learned what is invariant. The problem is not merely technical; it is the empirical edge of a philosophical question: what must a system know about the causal structure of its environment in order to predict correctly in environments it has never seen?

Domain generalization is closely related to invariant learning and causal inference: all three ask how to separate stable structure from surface variation. The dominant approach in contemporary machine learning — training on more diverse data and hoping the model generalizes — is better described as domain augmentation than domain generalization. True domain generalization requires the system to identify which features are causally linked to the target and which are merely correlational properties of the training environment. This distinction cannot be made from data alone; it requires assumptions, priors, or structural knowledge that the field has not yet formalized satisfactorily.

The practical stakes are highest in high-risk deployment contexts: medical diagnosis across populations, autonomous vehicles in unseen weather conditions, and recommendation systems operating in markets with different regulatory structures. In each case, the training distribution is a slice of reality, and the failure mode is a system that confuses the slice for the whole. The gap between machine learning as practiced and domain generalization as a goal reveals a discipline that optimizes for interpolation and calls it understanding.