FGPLFA: Fine-Grained Pseudo-Labeling and Feature Alignment for Source-Free Unsupervised Domain Adaptation

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Source-free unsupervised domain adaptation (SFUDA) is vital for privacy-preserving machine learning.
• Current SFUDA methods often use pseudo-labeling, which can be noisy and degrade performance.
• Existing techniques overlook fine-grained data features, limiting adaptation effectiveness.

Purpose of the Study:

• To develop a novel method, FGPLFA, to enhance SFUDA performance.
• To address limitations of noisy pseudo-labels in existing SFUDA approaches.
• To improve model adaptability in target domains without source data access.

Main Methods:

• Introduced a gradient-based metric integrating model knowledge and data features for reliable sample assessment.
• Developed the fine-grained pseudo-labeling (FGPL) module for sample-level data clustering and category/domain-specific pseudo-labeling.
• Implemented mean-covariance adjustment feature alignment (MCAFA) for sequential feature alignment across subsets.

Main Results:

• FGPLFA significantly reduces noisy pseudo-labels through multilevel granularity.
• The method enhances feature alignment, improving model adaptability.
• Experimental validation across multiple datasets confirms FGPLFA's superior performance.

Conclusions:

• FGPLFA offers a robust solution for SFUDA, outperforming existing methods.
• The proposed FGPL and MCAFA modules effectively address key challenges in domain adaptation.
• This work advances SFUDA by enabling more accurate and adaptable models under data constraints.