Understanding Domain Generalization: A View of Necessity and Sufficiency
Keywords: Domain Generalization
Abstract: Despite the rapid advancements in domain generalization (DG), the majority of DG studies center on establishing theoretical guarantee for generalization under the assumption of sufficient, diverse or even infinite domains. This assumption however is unrealistic, thus there remains no conclusive evidence as to whether the existing DG algorithms can truly generalize in practical settings where domains are limited. This paper aims to elucidate this matter. We first study the conditions for the existence and learnability of an optimal hypothesis. As the sufficient conditions are non-verifiable, our identified two necessary conditions become critical to guaranteeing the chance of finding the global optimal hypothesis in finite domain settings. In light of the theoretical insights, we provide a comprehensive review of DG algorithms explaining to what extent they can generalize effectively. We finally introduce a practical approach that leverages the joint effect of the two sets of conditions to boost generalization. Our proposed method demonstrates superior performance on well-established DG benchmarks.
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: 6866
Loading