Informative pairs mining based adaptive metric learning for adversarial domain adaptation.

Area of Science:

• Computer Science
• Machine Learning

Background:

• Adversarial domain adaptation aims to improve feature transferability across domains.
• A key challenge is the degradation of feature discrimination during this process.

Purpose of the Study:

• To propose a novel method, informative pairs mining based adaptive metric learning (IPM-AML), to enhance feature discrimination.
• To integrate IPM-AML with Conditional Domain Adversarial Network (CDAN) for improved feature representation.

Main Methods:

• Developed a two-triplet-sampling strategy to identify informative positive and negative pairs.
• Utilized a weighted metric loss to focus on informative pairs, improving discrimination.
• Incorporated IPM-AML into CDAN, creating IPM-AML-CDAN.
• Implemented a threshold strategy for reliable pseudo target label selection with theoretical validation.

Main Results:

• IPM-AML-CDAN learns feature representations that are both transferable and discriminative.
• The method achieves competitive results on four cross-domain benchmarks.
• Validated the effectiveness of the informative pairs mining and adaptive metric learning approach.

Conclusions:

• The proposed IPM-AML-CDAN effectively addresses the feature discrimination degradation in adversarial domain adaptation.
• This approach leads to superior performance in cross-domain tasks.
• The informative pairs mining strategy is crucial for enhancing discriminative power.