Breaking the Detection-Generalization Paradox on Out-Of-Distribution Data

Xuan Li, Zhanke Zhou, YiXiao ZHENG, Jiangchao Yao, Lu Zhang, Bo Han

Keywords: Trustworthy Machine Learning; Out of distribution data

Abstract:

This work studies the trade-off between out-of-distribution (OOD) detection and generalization. We identify the Detection-Generalization Paradox in OOD data, where optimizing one objective can degrade the other. We investigate this paradox by analyzing the behaviors of models trained under different paradigms, focusing on representation, logits, and loss across in-distribution, covariate-shift, and semantic-shift data. Based on our findings, we propose Distribution-Robust Sharpness-Aware Minimization (DR-SAM), an optimization framework that balances OOD detection and generalization. Extensive experiments demonstrate the method's effectiveness, offering a clear, empirically validated approach for improving detection and generalizationability in different benchmarks.

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.

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: 10299