Exploring distribution-based approaches for out-of-distribution detection in deep learning models

Detecting unknown samples is a crucial task for deep learning applications, especially when considering open-set problems such as autonomous driving or disease classification. To improve DL models’ robustness in identifying unseen classes, out-of-distribution (OOD) methods are utilized to distinguis...

Full description

Saved in:

Bibliographic Details
Published in:	Neural computing & applications Vol. 37; no. 17; pp. 10807 - 10822
Main Authors:	Carvalho, Thiago, Vellasco, Marley, Amaral, José Franco
Format:	Journal Article
Language:	English
Published:	London Springer London 01.06.2025 Springer Nature B.V
Subjects:	Artificial Intelligence Computational Biology/Bioinformatics Computational Science and Engineering Computer Science Data Mining and Knowledge Discovery Deep learning Image Processing and Computer Vision Probabilistic models Probability and Statistics in Computer Science S.I.: Timely Advances of Deep Learning with applications and Data Driven Modeling Special Issue on Timely Advances of Deep Learning with applications and Data Driven Modeling Out-of-distribution detection Gaussian distribution Dirichlet Process Mixture Deep learning application
ISSN:	0941-0643, 1433-3058
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Detecting unknown samples is a crucial task for deep learning applications, especially when considering open-set problems such as autonomous driving or disease classification. To improve DL models’ robustness in identifying unseen classes, out-of-distribution (OOD) methods are utilized to distinguish between in-distribution (ID) and OOD samples using distinct patterns from the model’s output space. While utilizing the output space for OOD detection is common due to its practicality and effectiveness, relying solely on model scores may lead to issues such as overconfidence. In this study, we propose leveraging the logit distribution, with a focus on the Dirichlet Gaussian Mixture Models, to identify OOD samples. These approaches consider not only the score associated with the highest class but also those assigned to other classes. We evaluate different distributional assumptions and analyze the advantages and disadvantages of using Gaussian-based models for logit distribution in OOD detection. Based on the experiments of this work, the Dirichlet Gaussian Mixture approach obtained up to 27.5% improvement if compared to the best baseline strategy using the output space in the AUROC metric and 13.6% for the FPR95 metric.
Bibliography:	ObjectType-Article-1 SourceType-Scholarly Journals-1 ObjectType-Feature-2 content type line 14
ISSN:	0941-0643 1433-3058
DOI:	10.1007/s00521-024-10912-w