Rectifying Gradient-based OOD Detection via Fisher Information Matrix

Yufan Hu, Qizhou Wang, Xinyu Liu, Bo Han, Yixuan Yuan

Keywords: OOD Detection

Abstract:

Out-of-distribution (OOD) detection is an anomaly-handling mechanism, for which classification systems should detect outliers with true labels outside the label space, distinguishing them from normal in-distribution (ID) data.
Advanced works suggest that gradient information preserve sufficient cues to indicate the confidence of being OOD.
However, we discover previous gradient-based detection methods suffer from limited effectiveness mainly due to over-parameterization. As gradient-based OOD scores derive from the over-parameterized weight space, a widely recognized cause for the suboptimal OOD detection performance, there are also some gradient components which lack necessary informativeness for OOD detection, thereby impair the performance.
This observation motivates us to propose gradient rectification (GradRect), using fisher information matrix to correct gradients in directions that are uninformative to discern the distribution change. Moreover, we connect GradRect with classical theories in identifying influential observations, verifying that model fine-tuning with outlier exposure can further improve GradRect. We conduct extensive experiments on various OOD detection setups, revealing the power of GradRect against state-of-the-art counterparts.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6875