A View on Out-of-Distribution Identification from a Statistical Testing Theory Perspective

TL;DR

This paper analyzes out-of-distribution detection using statistical testing theory, proposing a framework based on Wasserstein distance with convergence guarantees and empirical evaluation.

Contribution

It reformulates OOD detection as a statistical testing problem, establishing conditions for identifiability and analyzing convergence guarantees of Wasserstein-based tests.

Findings

Wasserstein distance-based test shows promising convergence properties.

The proposed framework clarifies conditions for reliable OOD detection.

Empirical evaluation supports theoretical insights.

Abstract

We study the problem of efficiently detecting Out-of-Distribution (OOD) samples at test time in supervised and unsupervised learning contexts. While ML models are typically trained under the assumption that training and test data stem from the same distribution, this is often not the case in realistic settings, thus reliably detecting distribution shifts is crucial at deployment. We re-formulate the OOD problem under the lenses of statistical testing and then discuss conditions that render the OOD problem identifiable in statistical terms. Building on this framework, we study convergence guarantees of an OOD test based on the Wasserstein distance, and provide a simple empirical evaluation.