SODA: Stream Out-of-Distribution Adaptation
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: out-of-distribution detection, distribution shift, machine learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: In open-context environments, machine learning models require out-of-distribution (OOD) awareness to ensure safe operation. However, existing OOD detection approaches have primarily focused on the offline setting, where OOD detectors remain static and fixed after deployment. This limits their ability to perform in real-world environments with unknown and ever-shifting out-of-distribution data. To address this limitation, we propose a novel online OOD detection framework that allows for continuous adaptation of the OOD detector. Our framework updates the ID classifier and OOD detector sequentially, based on samples observed from the deployed environment, and minimizes the risk of incorrect OOD predictions at each timestep. Unlike traditional offline OOD detection methods, our online framework provides the adaptivity and practicality needed for real-world environments. Theoretical analysis demonstrates that our algorithm provably achieves sub-linear regret and converges to the optimal OOD detector over time. Empirical evaluation in various environments shows that our online OOD detector significantly outperforms offline methods, highlighting the superiority of our framework for real-world applications of OOD detection.
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: 3012
Loading