Multi-Source Deep Transfer Neural Network Algorithm.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Transfer learning leverages source domain knowledge for target domain tasks, especially with limited data.
• Multi-source transfer learning offers improved performance but faces challenges from probability distribution mismatch.
• Deep learning methods show promise in extracting robust features to mitigate domain mismatch.

Purpose of the Study:

• To propose a novel multi-source deep transfer neural network algorithm, MultiDTNN.
• To address the performance degradation caused by probability distribution mismatch in multi-source transfer learning.
• To enhance feature transferability and classification accuracy in target domains with insufficient data.

Main Methods:

• Developed MultiDTNN algorithm integrating convolutional neural networks and multi-source transfer learning.
• Employed joint probability distribution adaptation (JPDA) to reduce source-target domain distribution mismatch.
• Trained convolutional neural networks on source and target datasets to generate multiple classifiers.
• Implemented a selection strategy to choose the best-performing classifier for the final MultiDTNN framework.

Main Results:

• MultiDTNN demonstrated effectiveness in enhancing classification performance.
• The algorithm successfully reduced the mismatching between source and target domains.
• Comparative analysis showed superior performance against state-of-the-art deep transfer learning methods on three datasets.

Conclusions:

• MultiDTNN offers a robust solution for multi-source deep transfer learning.
• The proposed JPDA and classifier selection strategy effectively improve target domain classification.
• This work advances deep transfer learning techniques for data-scarce scenarios.