Attention-Based Multi-Source Domain Adaptation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Multi-source domain adaptation (MSDA) transfers knowledge from multiple source domains to a single target domain.
• Current MSDA methods align target and source domains, but this can be detrimental when source domains are dissimilar.
• Harmful effects from dissimilar domains can negatively impact representation learning in MSDA.

Purpose of the Study:

• To propose an attention-based approach for MSDA that mitigates negative impacts from dissimilar source domains.
• To enhance the positive influence of similar domains in the knowledge transfer process.
• To improve the effectiveness of MSDA by considering domain correlations.

Main Methods:

• Attention-Based Multi-Source Domain Adaptation (ABMSDA) utilizes an attention mechanism to weigh domain importance.
• A domain recognition model calculates target-source domain correlations.
• Weighted Moment Distance (WMD) and Attentive Classification Loss (ACL) are proposed for better representation learning.

Main Results:

• The proposed ABMSDA model effectively alleviates negative effects from dissimilar domains.
• WMD prioritizes source domains with higher similarity to the target domain.
• ACL ensures the generation of aligned and discriminative visual representations.

Conclusions:

• ABMSDA demonstrates superior performance in multi-source domain adaptation tasks.
• The attention mechanism is crucial for selectively leveraging knowledge from relevant source domains.
• The method shows significant effectiveness, achieving an average of 6.1% improvement on the DomainNet dataset.