Generalized Meta-FDMixup: Cross-Domain Few-Shot Learning Guided by Labeled Target Data

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Few-Shot Learning (FSL) typically assumes source and target classes share the same domain.
• Cross-Domain Few-Shot Learning (CD-FSL) addresses the challenge of significant domain shifts between source and target datasets.
• Existing CD-FSL methods often overlook the explicit utilization of limited labeled target data during training.

Purpose of the Study:

• To propose a more practical training scenario for CD-FSL that leverages labeled target data.
• To introduce a novel network, GMeta-FDMixup, to effectively utilize labeled target data for bridging domain gaps.
• To enhance the generalization ability of models in CD-FSL tasks.

Main Methods:

• Developed two mixup modules (mixup-P and mixup-M) to facilitate the use of unbalanced and disjoint datasets, enabling diverse image generation for source domain training.
• Introduced a feature disentanglement module to decouple domain-irrelevant and domain-specific features, mitigating domain inductive bias.
• Incorporated a contrastive learning module (ConL) to prevent over-reliance on category-specific features and improve generalization.

Main Results:

• The proposed GMeta-FDMixup network effectively utilizes labeled target data to bridge the domain gap in CD-FSL.
• Experimental results on two benchmarks demonstrate the superiority of the proposed training setting and method.
• The feature disentanglement and contrastive learning components contribute to improved model generalization.

Conclusions:

• Leveraging limited labeled target data is crucial for effective CD-FSL, especially when facing significant domain shifts.
• The GMeta-FDMixup network offers a novel and effective approach to address CD-FSL challenges.
• The proposed method shows significant potential for improving few-shot classification performance in cross-domain scenarios.