Out-of-Distribution Detection with Deep Nearest Neighbors

TL;DR

This paper proposes a non-parametric nearest-neighbor approach for out-of-distribution detection that does not rely on distributional assumptions, demonstrating superior performance on benchmarks compared to parametric methods.

Contribution

It introduces a flexible, assumption-free nearest-neighbor method for OOD detection, outperforming existing parametric approaches like Mahalanobis distance.

Findings

Reduces false positive rate by 24.77% on ImageNet-1k

Outperforms Mahalanobis-based baseline in OOD detection

Effective across multiple benchmark datasets

Abstract

Out-of-distribution (OOD) detection is a critical task for deploying machine learning models in the open world. Distance-based methods have demonstrated promise, where testing samples are detected as OOD if they are relatively far away from in-distribution (ID) data. However, prior methods impose a strong distributional assumption of the underlying feature space, which may not always hold. In this paper, we explore the efficacy of non-parametric nearest-neighbor distance for OOD detection, which has been largely overlooked in the literature. Unlike prior works, our method does not impose any distributional assumption, hence providing stronger flexibility and generality. We demonstrate the effectiveness of nearest-neighbor-based OOD detection on several benchmarks and establish superior performance. Under the same model trained on ImageNet-1k, our method substantially reduces the false positive rate (FPR@TPR95) by 24.77% compared to a strong baseline SSD+, which uses a parametric approach Mahalanobis distance in detection. Code is available: https://github.com/deeplearning-wisc/knn-ood.