Neighborhood Transformation Marginalization forOOD Detection

Abstract

Out-of-distribution (OOD) detection is an important part of enabling the real world deployment of machine learning models. Many recent methods developed to perform OOD detection rely on calculating a score function on a given test point then thresholding the value to classify the point as in-distribution (ID) or OOD. However, calculating a score function on a single example may give biased or inaccurate estimates, especially as examples are sampled further and further OOD. In this paper we propose TraM: Transformation Neighborhood Marginalization, a method to improve the estimation of score functions used for OOD detection by calculating their expectation over a transformation neighborhood. TraM demonstrates improvements on a subset of commonly used OOD score functions in the OpenOOD benchmark, improving a baseline ODIN score function by up to 6 AUROC. However, it is not found to improve other baseline metrics signficantly, indicating the need for further research on this topic.