Adaptive multi-source domain collaborative fine-tuning for transfer learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Transfer learning is crucial for tasks with limited data.
• Single-source domain fine-tuning struggles with large data distribution differences.
• Effective feature extraction is key for successful transfer learning.

Purpose of the Study:

• To propose a novel transfer learning framework, adaptive multi-source domain collaborative fine-tuning (AMCF).
• To address the limitations of single-source fine-tuning in scenarios with significant domain shifts.
• To enhance feature extraction capabilities for target tasks using multiple source domains.

Main Methods:

• AMCF utilizes multiple source domain models for collaborative fine-tuning.
• An adaptive multi-source domain layer selection strategy customizes fine-tuning schemes.
• A multi-source domain collaborative loss function ensures precise feature extraction and minimizes output discrepancies.

Main Results:

• AMCF was validated on seven public visual classification datasets.
• Experimental results show enhanced accuracy in feature extraction compared to single-source methods.
• The framework provides precise layer fine-tuning schemes, significantly improving overall performance.

Conclusions:

• AMCF effectively overcomes the challenges of large domain distribution differences in transfer learning.
• The proposed method significantly improves fine-tuning performance by enhancing feature extraction.
• AMCF offers a robust solution for improving model adaptability and accuracy in cross-domain tasks.