Multisource Deep Transfer Learning Based on Balanced Distribution Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional unsupervised transfer learning often relies on single-source domains, which is insufficient for many practical scenarios.
• Multisource unsupervised transfer learning aims to align data in a common feature space to minimize distribution differences between source and target domains.
• Existing methods often treat marginal and conditional distributions equally, leading to suboptimal performance, and balanced distribution algorithms are typically single-source.

Purpose of the Study:

• To develop a novel multisource transfer learning algorithm that addresses the limitations of existing methods.
• To improve the performance of transfer learning by effectively adapting distributions from multiple sources.
• To introduce a method that balances the adaptation of marginal and conditional distributions.

Main Methods:

• Proposes a multisource transfer learning algorithm named MTLBDA (Multisource Transfer Learning based on Distribution Adaptation).
• Employs a strategy that adjusts the weights of marginal and conditional distributions for effective domain adaptation.
• Focuses on aligning data in a common feature space while managing distribution differences.

Main Results:

• MTLBDA demonstrates significant improvements in performance on popular image classification datasets.
• The method achieves notable results on the Office-31 dataset, validating its effectiveness.
• Experimental results indicate superior performance compared to existing approaches.

Conclusions:

• The proposed MTLBDA algorithm effectively handles multisource unsupervised transfer learning challenges.
• Adjusting distribution weights is a viable strategy for enhancing domain adaptation in multisource scenarios.
• The method offers a promising solution for practical applications requiring data from multiple domains.