[Paper Review] Meta-FDMixup

Yuqian Fu, Yanwei Fu+1 more|arXiv (Cornell University)|Oct 17, 2021

Domain Adaptation and Few-Shot Learning23 references1 citations

TL;DR

This paper proposes Meta-FDMixup, a novel meta-learning framework for Cross-Domain Few-Shot Learning (CD-FSL) that leverages limited labeled target data to bridge domain gaps. By integrating a re-imagined mixup module with a disentangled feature extractor and domain classifier, the method improves generalization to unseen target domains, achieving state-of-the-art performance on benchmark datasets.

ABSTRACT

A recent study finds that existing few-shot learning methods, trained on the source domain, fail to generalize to the novel target domain when a domain gap is observed. This motivates the task of Cross-Domain Few-Shot Learning (CD-FSL). In this paper, we realize that the labeled target data in CD-FSL has not been leveraged in any way to help the learning process. Thus, we advocate utilizing few labeled target data to guide the model learning. Technically, a novel meta-FDMixup network is proposed. We tackle this problem mainly from two aspects. Firstly, to utilize the source and the newly introduced target data of two different class sets, a mixup module is re-proposed and integrated into the meta-learning mechanism. Secondly, a novel disentangle module together with a domain classifier is proposed to extract the disentangled domain-irrelevant and domain-specific features. These two modules together enable our model to narrow the domain gap thus generalizing well to the target datasets. Additionally, a detailed feasibility and pilot study is conducted to reflect the intuitive understanding of CD-FSL under our new setting. Experimental results show the effectiveness of our new setting and the proposed method. Codes and models are available at https://github.com/lovelyqian/Meta-FDMixup.

Motivation & Objective

• To address the generalization failure of few-shot learning models when a domain gap exists between source and target domains.
• To explore the untapped potential of using limited labeled target data in CD-FSL to improve model adaptation.
• To develop a meta-learning framework that effectively combines source and target data from different class sets.
• To disentangle domain-invariant and domain-specific features to reduce domain shift and improve robustness.
• To validate the feasibility and effectiveness of the proposed CD-FSL setting with empirical studies.

Proposed method

• A novel mixup module is re-designed to combine support sets from both source and target domains, even when they belong to different class sets.
• The mixup module is integrated into a meta-learning paradigm to enable few-shot adaptation using both source and target labeled data.
• A disentangle module is introduced to separate features into domain-invariant and domain-specific components.
• A domain classifier is trained jointly with the disentangle module to encourage the model to learn domain-robust representations.
• The model is trained end-to-end using a meta-learning objective that optimizes for few-shot generalization on the target domain.
• The framework enables effective knowledge transfer from source to target domains by minimizing domain shift through feature disentanglement.

Experimental results

Research questions

• RQ1 Can limited labeled target data improve few-shot generalization in cross-domain settings?
• RQ2 How can mixup be adapted to combine support sets from source and target domains with disjoint class sets?
• RQ3 To what extent can disentangled feature learning reduce domain shift in CD-FSL?
• RQ4 Does the integration of target data into meta-learning improve performance on unseen target datasets?
• RQ5 How does the proposed method compare to existing FSL methods under domain shift?

Key findings

• The proposed Meta-FDMixup framework achieves state-of-the-art performance on standard CD-FSL benchmarks, demonstrating superior generalization under domain shift.
• The integration of labeled target data significantly improves model robustness and accuracy on novel target domains.
• The disentangled feature learning component effectively separates domain-invariant and domain-specific features, reducing domain shift.
• The ablation study confirms that both the mixup and disentangle modules contribute meaningfully to performance gains.
• The pilot study validates the feasibility of the proposed CD-FSL setting and highlights the importance of leveraging target data.