Dual Alignment for Partial Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Partial Domain Adaptation (PDA) addresses knowledge transfer challenges from label-rich to label-scarce domains.
• A key difficulty is maximizing positive transfer within shared label spaces while minimizing negative transfer due to label space mismatches.

Purpose of the Study:

• To propose a novel Dual Alignment Approach for Partial Domain Adaptation (DAPDA).
• To enhance positive transfer and mitigate negative transfer in PDA scenarios.

Main Methods:

• Utilized a Siamese network for feature extraction from source and target domains.
• Implemented a reweighting network to generate class-level (hard) and instance-level (soft) weights.
• Employed a dual alignment network for intra-domain compactness and inter-domain distribution alignment.

Main Results:

• DAPDA effectively minimizes intraclass variances for improved compactness.
• Inter-domain alignment reduces discrepancies through domain-wise and class-wise adaptations.
• The method demonstrated superior performance over state-of-the-art PDA techniques on benchmark datasets.

Conclusions:

• The proposed DAPDA method successfully addresses the challenges of partial domain adaptation.
• By selectively weighting source features and aligning distributions, DAPDA enhances knowledge transfer efficiency.
• The approach offers a significant improvement for scenarios with label space mismatches.