Conditional Density Ratio Score for Post Hoc Deep Outlier Detection
Keywords: Deep Learning, Outlier Detection, Out-Of-Distribution Detection
Abstract: The ability to accurately identify out-of-distribution (OOD) samples is essential not only as a stand-alone machine learning task but also for maintaining the reliability and safety of machine learning systems. Within this domain, post hoc density estimators like the energy score are popular ways for detecting OOD samples. However, most of the existing post hoc density estimation have mainly focused on marginalizing the conditional distributions over all possible classes. In this paper, we introduce the Conditional Density Ratio (CDR) score, a principled post hoc density estimator that leverages both a class-conditional generative model in the latent space and a discriminative classifier model, allowing us to estimate the marginal densities of the latent representation without marginalization. We demonstrate that a key component to the success of the CDR score lies in correctly calibrating the two models and propose a simple yet effective method to automatically tune the temperature parameter without the need for out-of-distribution samples. We illustrate the general compatibility of the proposed method with two popular density estimators, the kernel density estimator and the Mahalanobis estimator. Through experiments on a wide range of OOD benchmark tasks, we verify the effectiveness of the proposed method and advocate it as an easy-to-implement baseline that can achieve competitive performance in most tested scenarios.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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: 7268
Loading