ROMA: Regularization for Out-of-distribution Detection with Masked Autoencoders

Xiaochen Feng, Yuan Jiang, Hao Sha, Yongbing Zhang

Keywords: out-of-distribution detection, masked autoencoders, regularization, masked image modeling

TL;DR: We propose a novel regularization framework based on masked autoencoders, addressing the over-fitting of ID representation that is harmful to some challenging OOD detection.

Abstract: Existing out-of-distribution (OOD) detection methods without outlier exposure learn effective in-distribution (ID) representations distinguishable for OOD samples, which have shown promising performance on many OOD detection tasks. However, we find a performance degradation in some challenging OOD detection, where pre-trained networks tend to perform worse during the fine-tuning process, exhibiting the over-fitting of ID representations. Motivated by this observation, we propose a critical task of hidden OOD detection, wherein ID representations provide limited or even counterproductive assistance in identifying hidden OOD data. To address this issue, we introduce a novel Regularization framework for OOD detection with Masked Autoencoders (ROMA), which utilizes the masked image modeling task to regularize the network. With distribution-agnostic auxiliary data exposure, ROMA notably surpasses previous OOD detection methods in hidden OOD detection. Moreover, the robustness of ROMA is further evidenced by its state-of-the-art performance on benchmarks for other challenging OOD detection tasks.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

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: 9628