Universal and Scalable Weakly-Supervised Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Domain adaptation aims to transfer knowledge from labeled source domains to unlabeled target domains.
• Real-world data often contains noise, necessitating weakly-supervised domain adaptation (WSDA) methods.
• Existing WSDA methods require prior knowledge of the noise rate, limiting practical application.

Purpose of the Study:

• To develop a universal and scalable WSDA method that does not require prior knowledge of the noise rate.
• To address the limitations of single-source domain adaptation in multi-source scenarios.
• To improve the robustness and effectiveness of domain adaptation in the presence of noisy data.

Main Methods:

• Proposes PDCAS with two stages: progressive distillation and domain alignment.
• Progressive distillation iteratively refines noisy source samples without supervision.
• Domain alignment uses Class-Aligned Sampling to balance source and target domain samples and global feature distributions.

Main Results:

• PDCAS effectively handles label and feature noises in the source domain.
• The proposed MSPDCAS demonstrates scalability for multi-source noisy domain adaptation.
• Experiments on Office-31 and Office-Home datasets show superior performance compared to state-of-the-art methods.

Conclusions:

• PDCAS offers a robust and generalizable solution for weakly-supervised domain adaptation.
• The framework's scalability is validated through its extension to multi-source settings.
• The method significantly enhances domain adaptation accuracy and robustness in noisy, real-world conditions.