Learning invariances via correlation and marginal alignments for out-of-distribution generalization

doi:10.1016/j.neucom.2026.132876

CSpace

	Learning invariances via correlation and marginal alignments for out-of-distribution generalization
	Guo, Zong 1,2; Ma, Bingpeng 2
	2026-04-14
发表期刊	NEUROCOMPUTING
ISSN	0925-2312
卷号	674 页码:11
摘要	Out-of-distribution generalization is a crucial challenge in machine learning, especially when spurious correla tions exist. To this end, Invariant Risk Minimization (IRM) proposes a promising paradigm by learning invariances from multiple training environments. However, IRM-based methods can be ineffective in certain situations. In this paper, we first analyze two main defects in IRM: the necessity of linear representation function assumption, and undesired solutions induced by misalignment of marginal distributions. Then, we propose a novel method to address these defects. We realize that the necessity of linear representation function assumption stems from not specifying the optimal classifier for each environment. Therefore, we define the optimal classifier to be the Canonical Correlation Analysis (CCA) classifier. This enables our method to be solved without linear representa tion function assumption. Furthermore, since the undesired solutions are attributed to misalignment of marginal distributions, we instead align representation-label correlation and marginal distribution of the representation. While slightly stronger, this is sufficient for matching the optimal CCA classifier. With the alignment of marginal distribution of the representation, the problem of undesired solutions can be mitigated. Experiments on five vision benchmarks and one regression task demonstrate the effectiveness of our proposed method. Our code is available at https://github.com/ZongGuo1995/CMA.
关键词	Out-of-distribution generalization Spurious correlation Invariant risk minimization Correlation alignment Marginal distribution alignment
DOI	10.1016/j.neucom.2026.132876
收录类别	SCI
语种	英语
WOS研究方向	Computer Science
WOS类目	Computer Science, Artificial Intelligence
WOS记录号	WOS:001684145500002
出版者	ELSEVIER
引用统计
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/42820
专题	中国科学院计算技术研究所
通讯作者	Ma, Bingpeng
作者单位	1.Chinese Acad Sci, Inst Comp Technol, Key Lab Intelligent Informat Proc, Beijing 100190, Peoples R China 2.Univ Chinese Acad Sci, Sch Comp Sci & Technol, Beijing 100049, Peoples R China
推荐引用方式 GB/T 7714	Guo, Zong,Ma, Bingpeng. Learning invariances via correlation and marginal alignments for out-of-distribution generalization[J]. NEUROCOMPUTING,2026,674:11.
APA	Guo, Zong,&Ma, Bingpeng.(2026).Learning invariances via correlation and marginal alignments for out-of-distribution generalization.NEUROCOMPUTING,674,11.
MLA	Guo, Zong,et al."Learning invariances via correlation and marginal alignments for out-of-distribution generalization".NEUROCOMPUTING 674(2026):11.