Browse by author
Lookup NU author(s): Dr Shidong WangORCiD
Full text for this publication is not currently held within this repository. Alternative links are provided below where available.
© 2025 Elsevier LtdTo address the issue of domain shift between training and testing phases, Domain Generalization (DG) has been proposed to enable models to maintain accuracy in unseen domains. However, during the testing phase, not only domain shift but also label space shift occurs. Traditional DG methods fail to classify samples from unknown classes. To solve these problems, Open Set Domain Generalization (OSDG) has been proposed, which aims to overcome domain shift while also providing the model with the ability to recognize unknown samples. Meta-learning based methods are common used in DG, where the source domains are divided into meta-training and meta-testing domains, and joint training is performed to prevent the model from over-fitting to the source domains. Meta-learning in DG often takes cross-entropy loss from supervised learning, which can be understood as enabling the model to learn “how to learn to classify” under domain shift. This paper proposes a meta-learning method based on Optimal Transport (OT), Learning to Learn to Optimal Transport (L2OT), which splits both between domains and classes, allowing the model to simulate how to transport the distribution from source domain to the target domain during the meta-training/testing phase. Unlike other methods, L2OT aims to let the model learn “how to learn to transport”, i.e., how to optimally transport the new class mass distribution in the presence of domain shift. Experimental results show that L2OT achieves leading results on multiple datasets. Code is available at: https://github.com/Ashengl/L2OT.
Author(s): Li C, Wang S, Long Y, Zhang H
Publication type: Article
Publication status: Published
Journal: Pattern Recognition
Year: 2026
Volume: 174
Print publication date: 01/06/2026
Online publication date: 26/12/2025
Acceptance date: 22/12/2025
ISSN (electronic): 0031-3203
Publisher: Elsevier Ltd
URL: https://doi.org/10.1016/j.patcog.2025.112988
DOI: 10.1016/j.patcog.2025.112988
Altmetrics provided by Altmetric
