Domain adaptation

Domain adaptation is the epistemic problem of determining which features, relations, and inferences from a source domain legitimately transfer to a target domain when an analogy is drawn between them. It is the boundary question: where does productive mapping end and false projection begin? In machine learning, domain adaptation names the technical challenge of training a model on data from one distribution and deploying it on another; the statistical problem is a formalization of the philosophical one. The stakes are highest when analogies cross between descriptive and normative domains — when a biological pattern is used to justify a social arrangement, or when a physical mechanism is used to model a mental one. Without a theory of domain adaptation, analogy remains a powerful but dangerous tool: it generates insight and error with equal efficiency. The work of Mary Hesse on models and analogies in science remains the most rigorous attempt to specify the rules of legitimate domain transfer, though her conditions have proven difficult to operationalize in practice.

The philosophers who treat domain adaptation as a matter of logical caution are missing the point. Domain boundaries are not discovered; they are negotiated. Every successful analogy rewrites what counts as the 'domain' of both source and target, expanding or contracting their borders to make the mapping hold. Adaptation is not boundary-respect; it is boundary-work.