A Simple Fix to Mahalanobis Distance for Improving Near-OOD Detection

TL;DR

This paper identifies limitations of Mahalanobis distance in near-OOD detection and introduces a simple, more robust modification called relative Mahalanobis distance, significantly enhancing detection performance across diverse benchmarks.

Contribution

The paper proposes a simple fix called relative Mahalanobis distance that improves near-OOD detection and robustness over the standard Mahalanobis distance method.

Findings

RMD outperforms MD by up to 15% AUROC on genomics OOD benchmarks.

RMD is more robust to hyperparameter choices.

The method improves OOD detection across vision, language, and biology datasets.

Abstract

Mahalanobis distance (MD) is a simple and popular post-processing method for detecting out-of-distribution (OOD) inputs in neural networks. We analyze its failure modes for near-OOD detection and propose a simple fix called relative Mahalanobis distance (RMD) which improves performance and is more robust to hyperparameter choice. On a wide selection of challenging vision, language, and biology OOD benchmarks (CIFAR-100 vs CIFAR-10, CLINC OOD intent detection, Genomics OOD), we show that RMD meaningfully improves upon MD performance (by up to 15% AUROC on genomics OOD).