Correcting the Bias of Normalizing Flows by Synthetic Outliers for Improving Out-of-Distribution Detection
Keywords: OOD Detection, Normalizing Flow
TL;DR: We propose leveraging synthetic outliers alongside a specialized training objective to enhance the OOD detection ability of normalizing flows for both images and texts.
Abstract: Out-of-distribution (OOD) detection is critical for ensuring the reliability and robustness of deep learning models in real-world applications. While normalizing flows have demonstrated impressive performance for various task of image OOD detection, recent findings suggest that they still encounter limitations and severe biases when applied to datasets with different statistics. Specifically, it has been observed that normalizing flow models tend to assign higher likelihoods to OOD samples with low complexity, which undermines the effectiveness of likelihood based OOD detection methods. In this paper, we explore the bias related to data complexity linked to normalizing flow models in OOD detection. We propose a novel method for bias correction by incorporating synthetic outliers during training, guiding the model to assign lower likelihoods to OOD samples. Additionally, we introduce a specialized training objective that leverages the softplus function for OOD data, ensuring a smooth and effective training process. Extensive experiments on benchmark and high-dimensional real-world datasets, including both images and texts, confirm that our proposed approach significantly enhances OOD detection accuracy, achieving performance comparable to models trained with a limited number of real outliers. Moreover, our method increases the Lipschitz constant, supporting the hypothesis presented in related literature.
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.
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: 1532
Loading