RankFeat: Rank-1 Feature Removal for Out-of-distribution Detection

Song Yue, Nicu Sebe, Wei WangZenodo 2023 · DOI · OpenAlex

Abstract

The task of out-of-distribution (OOD) detection is crucial for deploying machine<br> learning models in real-world settings. In this paper, we observe that the singular<br> value distributions of the in-distribution (ID) and OOD features are quite different:<br> the OOD feature matrix tends to have a larger dominant singular value than the<br> ID feature, and the class predictions of OOD samples are largely determined by it.<br> This observation motivates us to propose RankFeat, a simple yet effective post<br> hoc approach for OOD detection by removing the rank-1 matrix composed of<br> the largest singular value and the associated singular vectors from the high-level<br> feature. RankFeat achieves state-of-the-art performance and reduces the average<br> false positive rate (FPR95) by 17.90% compared with the previous best method.<br> Extensive ablation studies and comprehensive theoretical analyses are presented to<br> support the empirical results.
BibTeX
@inproceedings{yue2023rankfeat,
  author = {Song Yue and Nicu Sebe and Wei Wang},
  title = {{RankFeat: Rank-1 Feature Removal for Out-of-distribution Detection}},
  booktitle = {Zenodo},
  year = {2023},
  doi = {10.5281/zenodo.7566083},
}

Loading related content...

Figure