Out-of-Distribution Detection with Hyperspherical Energy

Jirayu Burapacheep, Yixuan Li

Primary Area: general machine learning (i.e., none of the above)

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.

Keywords: Hyperspherical energy, model reliability

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: The ability to detect if inputs are out-of-distribution (OOD) is essential to guarantee the reliability and safety of machine learning models that are deployed in an open environment. Recent studies have shown that an energy-based score is effective. However, unconstrained energy scores from a model trained with cross-entropy loss may not necessarily reflect the log-likelihood. To address this limitation, we introduce a novel hyperspherical energy score that connects energy with hyperspherical representations. By modeling hyperspherical representations using von Mises-Fisher distribution, our method provides a more accurate interpretation from a log-likelihood perspective, making it an efficient OOD detection indicator. Our method consistently achieves competitive performance on popular OOD detection benchmarks. On the large-scale ImageNet-1k benchmark, our method is more than 10 times faster than the KNN-based score, while simultaneously reducing the average FPR95 by 11.85%.

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: 2102