outofdistribution

Out-of-distribution (OOD) refers to inputs that do not lie within the distribution of data used to train a machine learning model. In supervised learning, the training data are assumed to be representative of the situations the model will encounter at deployment; OOD inputs violate this assumption, potentially causing unreliable predictions, miscalibrated uncertainty estimates, and degraded performance. OOD phenomena arise from distributional shifts such as covariate shift, concept drift, domain shift, or the appearance of novel classes not seen during training.

Detection and mitigation strategies aim to identify OOD samples and reduce their impact. Common approaches include calibrating predicted probabilities (for example with temperature scaling), applying confidence thresholds or entropy-based rejection, and using distance- or density-based methods in feature space (such as k-nearest neighbors, Gaussian or flow-based density estimation). Open-set or novelty detection techniques, one-class classifiers, and ensemble methods are also employed. In some settings, models are trained to be robust to OOD data or to abstain from predicting when uncertainty is high, sometimes through specific training objectives or exposure to outliers.

Evaluation of OOD performance typically involves testing on in-distribution data alongside separate out-of-distribution datasets, using metrics such as AUROC, false-positive rate at a fixed true-positive rate, or detection error. Benchmarks and methods vary across domains, including computer vision, natural language processing, and speech processing.

Challenges persist in defining what constitutes OOD in a given context, balancing detection with maintaining in-distribution accuracy, and reliably calibrating uncertainty in high-dimensional spaces. Robust OOD handling remains an active area of research with practical importance for safety and reliability.