A Graph-Theoretic Framework for Joint OOD Generalization and Detection

Han Wang, 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: OOD generalization, OOD detection

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

Abstract: In the context of modern machine learning, models deployed in real-world scenarios often encounter various forms of data shifts, leading to challenges in both out-of-distribution (OOD) generalization and detection. While these two aspects have received significant attention individually, they lack a unified framework for theoretical understanding and practical usage. This paper bridges the gap by formalizing a graph-theoretical framework tailored for both OOD generalization and detection. In particular, based on our graph formulation, we introduce spectral learning with wild data (SLW) and show the equivalence of minimizing the objective and performing spectral decomposition on the graph. This equivalence allows us to derive provable error quantifying OOD generalization and detection performance. Empirically, SLW demonstrates competitive performance against existing baselines, aligning with the theoretical insight.

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: 2386