Adaptive Dispersal and Collaborative Clustering for Few-Shot Unsupervised Domain Adaptation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) typically requires fully labeled source data.
• Few-shot unsupervised domain adaptation (FUDA) addresses scenarios with limited labeled source data.
• Existing FUDA methods inadequately exploit labeled and unlabeled source data relationships for pseudo-label generation and struggle with domain gaps.

Purpose of the Study:

• To propose a novel method, Adaptive Dispersal and Collaborative Clustering (ADCC), for few-shot unsupervised domain adaptation.
• To enhance knowledge transfer from limited labeled source domains to unlabeled target domains.
• To address the challenges of insufficient labeled data and significant domain gaps in FUDA.

Main Methods:

• Developed a collaborative clustering algorithm to expand the utility of limited labeled source data, capturing more distribution information.
• Introduced an adaptive dispersal strategy using an intermediate domain to mitigate the impact of domain-irrelevant information.
• Tested the ADCC method on benchmark datasets including Office31, Office-Home, miniDomainNet, and VisDA-2017.

Main Results:

• The proposed ADCC method demonstrated superior performance compared to existing state-of-the-art FUDA techniques.
• Experiments confirmed the effectiveness of collaborative clustering in leveraging scarce labeled data.
• The adaptive dispersal strategy successfully reduced the negative effects of domain discrepancies.

Conclusions:

• ADCC offers a significant advancement in few-shot unsupervised domain adaptation.
• The method effectively handles data scarcity and domain gaps, improving model transferability.
• ADCC provides a robust solution for real-world UDA problems with limited labeled data.