Leveraging Meta-Learning To Improve Unsupervised Domain Adaptation

Unsupervised Domain Adaptation (UDA) techniques in real-world scenarios often encounter limitations due to their reliance on reducing distribution dissimilarity between source and target domains, assuming it leads to effective adaptation. However, they overlook the intricate factors causing domain s...

Full description

Saved in:
Bibliographic Details
Published inComputer journal Vol. 67; no. 5; pp. 1838 - 1850
Main Authors Farhadi, Amirfarhad, Sharifi, Arash
Format Journal Article
LanguageEnglish
Published Oxford University Press 22.06.2024
Subjects
Takagi-Sugeno
Unsupervised Domain adaptation (UDA)
Image recognition
Weight Optimization
Meta-learning
Machine learning
Online AccessGet full text

Cover

More Information
Summary:Unsupervised Domain Adaptation (UDA) techniques in real-world scenarios often encounter limitations due to their reliance on reducing distribution dissimilarity between source and target domains, assuming it leads to effective adaptation. However, they overlook the intricate factors causing domain shifts, including data distribution variations, domain-specific features and nonlinear relationships, thereby hindering robust performance in challenging UDA tasks. The Neuro-Fuzzy Meta-Learning (NF-ML) approach overcomes traditional UDA limitations with its flexible framework that adapts to intricate, nonlinear domain gaps without rigid assumptions. NF-ML enhances domain adaptation by selecting a UDA subset and optimizing their weights via a neuro-fuzzy system, utilizing meta-learning to efficiently adapt models to new domains using previously acquired knowledge. This approach mitigates domain adaptation challenges and bolsters traditional UDA methods’ performance by harnessing the strengths of multiple UDA methods to enhance overall model generalization. The proposed approach shows potential in advancing domain adaptation research by providing a robust and efficient solution for real-world domain shifts. Experiments on three standard image datasets confirm the proposed approach’s superiority over state-of-the-art UDA methods, validating the effectiveness of meta-learning. Remarkably, the Office+Caltech 10, ImageCLEF-DA and combined digit datasets exhibit substantial accuracy gains of 30.9%, 6.8% and 10.9%, respectively, compared with the best-second baseline UDA approach.
ISSN:0010-4620
1460-2067
DOI:10.1093/comjnl/bxad104