Multi-Source Unsupervised Domain Adaptation via Pseudo Target Domain.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Multi-source domain adaptation (MDA) addresses knowledge transfer from multiple sources to an unlabeled target domain.
• Significant domain shift exists between target and source domains, and among diverse sources, posing a major challenge.
• Existing MDA methods often estimate mixed source distributions or combine single-source models, neglecting inter-source relevant information.

Purpose of the Study:

• To propose a novel MDA approach, Pseudo Target for MDA (PTMDA), that effectively utilizes relevant information among diverse source domains.
• To improve the performance of knowledge transfer in scenarios with significant domain shifts.
• To enhance the transferability of deep neural networks (DNNs) in MDA tasks.

Main Methods:

• PTMDA employs adversarial learning with a metric constraint to map source and target domains into group-specific subspaces.
• It constructs pseudo target domains and aligns remaining source domains within these subspaces.
• A matching normalization layer replaces traditional batch normalization to enforce latent layer alignments in DNNs.

Main Results:

• PTMDA exploits additional structured source information through pseudo target domain training, improving real target domain performance.
• Theoretical analysis indicates PTMDA reduces the target error bound and better approximates target risk.
• Extensive experiments show PTMDA outperforms state-of-the-art methods on MDA tasks.

Conclusions:

• PTMDA offers an effective solution for multi-source domain adaptation by better leveraging inter-source information.
• The proposed method, including matching normalization, enhances DNN transferability and overall MDA performance.
• PTMDA demonstrates significant improvements over existing methods in various experimental settings.