Leveraging Mixture Alignment for Multi-Source Domain Adaptation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Conventional Domain Adaptation (DA) typically involves one source and one target domain.
• Real-world applications often require knowledge transfer from multiple sources to a single target, a problem known as Multi-source Domain Adaptation (MDA).
• Existing adversarial-based MDA methods face limitations, including the need for classifier-aware metrics and insufficient consideration of target sample consistency and semantic information.

Purpose of the Study:

• To propose a novel adversarial learning algorithm, MDAMA, for addressing the challenges in Multi-source Domain Adaptation (MDA).
• To enhance knowledge transfer by aligning the target domain with a mixture distribution of multiple source domains.
• To improve target domain performance through enhanced domain alignment, target sample consistency, and semantic information transfer.

Main Methods:

• MDAMA utilizes an adversarial learning framework for knowledge transfer from multiple sources to a target domain.
• The algorithm employs margin-based discrepancy and augmented intermediate distributions for effective domain alignment.
• Target sample consistency is achieved through confidence thresholding, and semantic information is transferred from source to augmented target domains.

Main Results:

• MDAMA demonstrates effective alignment of the target domain with a mixture of source domains.
• The proposed method successfully incorporates target sample consistency and semantic information transfer.
• Extensive experiments on benchmark datasets (OfficeHome, Office31, PACS, Office-Caltech, DomainNet) show top performance.

Conclusions:

• MDAMA offers a robust solution for Multi-source Domain Adaptation (MDA) by addressing limitations of previous adversarial methods.
• The algorithm effectively transfers knowledge from multiple sources to a target domain, enhancing performance.
• MDAMA achieves state-of-the-art results on widely used MDA benchmark datasets.