Typicality-based point OOD detection with contrastive learning

Typicality-based inference methods for OOD detection find a typical value (often the mean value) of a model statistic from the training data and then flag test points as anomalous if the model statistic of the test data point deviates significantly from the typical value. These methods are effective for detecting a group of OOD data points when OOD data points are labeled into groups, but ineffective for the detection of individual OOD data points. In this paper, we extend typicality-based inference to be effective for point OOD detection by utilizing latent features learned from contrastive learning and then obtaining the nearest neighbors of a test data point to provide additional context used for point OOD detection. The typicality-based inference approach is shown to improve point OOD detection relative to several benchmarks.