Methods that train models to generalize to unseen target domains by learning representations, invariances, or decision rules that remain robust across multiple source domains without access to target data.

Principle

Encourage learning of domain-invariant features or model components by exposing the learner to diverse sources or by explicit regularization so that performance transfers to novel but related contexts.

Demonstration

Train an image classifier on photos taken by multiple different camera sensors and lighting setups (several source domains) with a training objective that penalizes domain-specific features; the classifier performs reasonably on images from a new sensor it never saw.

Misapplication

Assuming that training on many diverse sources guarantees generalization to any unseen domain, ignoring fundamental out-of-support target distributions or systematic label-function shifts that cannot be resolved by representation invariance alone.

Consequence

Successful domain generalization yields models that require less adaptation when faced with novel environments; however, overconfidence in generalization can lead to neglected monitoring and brittle deployment.

Reversal

Domain specialization or overfitting: optimizing exclusively for observed source domains so the model captures idiosyncratic, non-transferable cues and fails on any new domain.

Boundary

Targets the zero-shot or few-shot setting where no labeled target data are available during training; it does not cover settings where target labels are accessible and adaptation is feasible and often preferable.

Semantic Tension

Closely related to robust optimization and invariant representation learning; tension exists between maximizing worst-case performance across hypothetical domains and optimizing average performance over known sources.

Synthesis

Domain generalization is the strategy of shaping a model's inductive biases through multi-source exposure or invariance penalties so that learned decision rules remain useful in unseen but related domains, acknowledging limits when tasks or supports change fundamentally.