Pseudo-Labels are All You Need for Out-Of-Distribution Detection

Atik Garg, Yu-Shuen Wang

Keywords: Pseudo-Labels, Unsupervised Out-of-Distribution Detection

TL;DR: We bridge supervised and unsupervised out-of-distribution detection methods by leveraging pseudo-labels, achieving state-of-the-art performance.

Abstract: Detecting out-of-distribution (OOD) samples is a significant challenge in real-world deep-learning applications, such as medical imaging and autonomous driving. Traditional machine learning models, primarily trained on in-distribution (ID) data, often struggle when encountering OOD instances, resulting in unreliable predictions. While supervised OOD detection methods generally outperform unsupervised approaches due to the availability of labeled data, our research uncovers a crucial insight: their success is not necessarily due to recognizing the actual object categories in the images; instead, these methods rely on a specific classification strategy that may not correspond to real-world understanding. Essentially, supervised methods detect OOD samples by identifying the difficulties in classifying unfamiliar data. This challenge is similar to what unsupervised OOD detection methods face, as they also depend on the failure to reconstruct OOD data due to the lack of prior exposure. In this study, we bridge the gap between supervised and unsupervised OOD detection by introducing a novel approach that trains models to classify data into pseudo-categories. We employ self-supervised learning (SSL) to convert raw data into representations, which are then clustered to generate pseudo-labels. These pseudo-labels are subsequently used to train a classifier, enabling its OOD detection capabilities. Experimental results show that our approach surpasses state-of-the-art techniques. Furthermore, by training models on different sets of pseudo-labels derived from the dataset, we enhance the robustness and reliability of our OOD detection method.

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.

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: 4042