Guaranteed Out-Of-Distribution Detection with Diverse Auxiliary Set
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
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Our theory and experiments demonstrate that training with diverse auxiliary outliers enhances OOD detection performance.
Abstract: Out-of-distribution (OOD) detection is crucial for ensuring reliable deployment of machine learning models in real-world scenarios. Recent advancements leverage auxiliary outliers to represent the unknown OOD data to regularize model during training, showing promising performance. However, detectors face challenges in effectively identifying OOD data that significantly deviate from the distribution of the auxiliary outliers, limiting their generalization capacity. In this work, we thoroughly examine this problem from the generalization perspective and demonstrate that a more diverse set of auxiliary outliers improves OOD detection. Constrained by limited access to auxiliary outliers and the high cost of data collection, we propose Provable Mixup Outlier (ProMix), a simple yet practical approach that utilizes mixup to enhance auxiliary outlier diversity. By training with these diverse outliers, our method achieves superior OOD detection. We also provide insightful theoretical analysis to verify that our method achieves better performance than prior works. Furthermore, we evaluate ProMix on standard benchmarks and demonstrate significant relative improvements of 14.2\% and 31.5\% (FPR95) on CIFAR-10 and CIFAR-100, respectively, compared to state-of-the-art methods. Our findings emphasize the importance of incorporating diverse auxiliary outliers during training and highlight ProMix as a promising solution to enhance model security in real-world applications. Compared with other methods, the proposed method achieves excellent performance on different metrics in almost all datasets.
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: 2473
Loading