outofdomain

Out-of-domain, often written as out-of-domain or out-of-distribution, refers to data inputs or situations that lie outside the distribution of data on which a model or system was trained. In machine learning and artificial intelligence, this concept helps indicate when a model’s predictions may be unreliable because the input does not resemble the training examples.

Examples of out-of-domain conditions include images from unseen classes, text with novel vocabulary, data corrupted by noise, or adversarial inputs designed to mislead a model. Out-of-domain scenarios can arise from distribution shift, changes in data collection processes, or deployment in new environments. In some contexts, domain can also refer to a network, application, or trust boundary, and inputs outside that boundary are considered out-of-domain, requiring filtering or access control.

Detection and handling of out-of-domain inputs are active areas of research and practice. Approaches include calibrating model confidence, applying thresholds, and using dedicated OOD detectors. Techniques such as temperature scaling, ODIN, density-based methods (e.g., kernel density estimation), Mahalanobis distance, deep ensembles, and autoencoder-based anomaly detection are employed to flag or manage OOD inputs. Systems may reject OOD inputs, route them to specialized models, or defer to human review. The overarching goal is to improve reliability and safety under distribution shift and to support open-set recognition.

See also: out-of-distribution, open-set recognition, anomaly detection, distribution shift.