Meta-Domain Variant Neural Architecture Search For Robust Cross-Domain Generalization Using Meta-Learning

Abstract

Neural Architecture Search (NAS) has demonstrated remarkable success in automatically designing deep neural networks. However, conventional NAS methods assume that training and deployment data originate from similar distributions, resulting in significant performance degradation under domain shifts. To address this limitation, this paper proposes Meta-Domain Variant Neural Architecture Search (MDV-NAS), a novel framework that integrates meta-learning with domain-aware architecture optimization. The proposed method treats each domain as an individual meta-task and learns architecture parameters that generalize across multiple source domains. A Domain Variance Estimator (DVE) is introduced to quantify distribution discrepancies among domains and guide architecture selection toward domain-invariant representations. During the search process, architectures are trained on sampled source domains and validated on unseen domains through episodic meta-learning. Experimental evaluations are conducted on PACS, Office-Home, DomainNet, and VLCS benchmarks which demonstrated superior robustness and generalization compared with conventional NAS and domain generalization methods. Results produced indicate that MDV-NAS achieves higher unseen-domain accuracy while maintaining computational efficiency comparing with other frameworks, making it suitable for real-world deployment in heterogeneous environments.