Bridging the gap with grad: Integrating active learning into semi-supervised domain generalization.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Domain generalization (DG) typically requires extensive labeled source data, which is often impractical.
• Semi-supervised learning (SSL) has been adapted for domain generalization (SSDG) to leverage unlabeled data, but comparisons are inconsistent.
• Existing SSDG methods suffer from unstable training due to random initial labeling of unlabeled data.

Purpose of the Study:

• To address the challenges of data labeling in domain generalization.
• To propose a standardized training paradigm for semi-supervised domain generalization.
• To introduce Active Semi-supervised Domain Generalization (ASSDG) by integrating active learning (AL) with SSDG.

Main Methods:

• A unified framework, Gradient-Similarity-based Sample Filtering and Sorting (GSSFS), is proposed.
• GSSFS iteratively trains both SSDG and AL components.
• Gradient similarity is employed to select informative unlabeled source samples for both SSDG and AL.

Main Results:

• The proposed methods achieve state-of-the-art results on DG datasets, particularly in low-data regimes.
• The GSSFS framework demonstrates efficiency and simplicity.
• The approach enhances training stability and reliability compared to random labeling.

Conclusions:

• ASSDG offers a more effective approach to domain generalization with limited labeled data.
• The GSSFS framework provides a robust and efficient solution for active semi-supervised domain generalization.
• This work establishes a standardized paradigm for SSDG research and highlights the benefits of integrating AL.