Generation, division and training: A promising method for source-free unsupervised domain adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Conventional unsupervised domain adaptation (UDA) requires labeled source data, which is often unavailable due to privacy concerns in source-free UDA (SFUDA).
• Existing SFUDA methods using self-supervised learning struggle with inaccurate pseudo-labels, hindering target model performance.

Purpose of the Study:

• To propose a novel SFUDA method, Generation, Division and Training (GDT), that improves pseudo-label reliability and enhances feature representation.
• To address the limitations of existing SFUDA approaches by refining pseudo-labels and employing a dual strategy for reliable and unreliable samples.

Main Methods:

• The GDT method refines target pseudo-labels using deep clustering and divides samples into reliable and unreliable sets.
• Reliable samples are trained using self-supervised learning and information maximization.
• Unreliable samples undergo contrastive learning to attract similar features and repel dissimilar ones, promoting efficient feature clustering.

Main Results:

• The proposed GDT method significantly improves the performance of target models in SFUDA tasks.
• Experiments on three benchmark datasets demonstrate the effectiveness and superiority of the GDT approach.
• The method successfully enhances pseudo-label reliability and leads to more discriminative feature representations.

Conclusions:

• GDT offers a robust solution for source-free unsupervised domain adaptation by improving pseudo-label quality.
• The combination of deep clustering, self-supervised learning, and contrastive learning in GDT effectively addresses key challenges in SFUDA.
• The proposed approach demonstrates strong potential for advancing domain adaptation techniques in privacy-sensitive scenarios.