Deep Neural Network with Joint Distribution Matching for Cross-Subject Motor Imagery Brain-Computer Interfaces

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Motor imagery brain-computer interfaces (BCIs) show promise but require lengthy calibration due to nonstationary signals.
• Current methods struggle with distribution shifts between source and target subjects, hindering real-world application.

Purpose of the Study:

• To develop a zero-training approach for motor imagery BCIs by addressing the distribution shift problem.
• To enable effective calibration transfer from source to target subjects without user-specific training.

Main Methods:

• Proposed a novel measure for joint distribution discrepancy (JDD) to align source and target data distributions.
• Developed a deep neural network incorporating joint distribution matching for zero-training BCI.
• Explored both marginal and joint distribution adaptation techniques.

Main Results:

• The proposed JDD measure effectively aligns source and target data, showing a direct correlation with classification accuracy.
• The deep neural network with joint distribution matching achieved effective feature generalization in an aligned common space.
• Experimental results on two datasets demonstrated superior performance compared to existing methods.

Conclusions:

• Joint distribution adaptation is crucial for successful transfer learning in motor imagery BCIs.
• The developed zero-training BCI system significantly alleviates distribution discrepancy across subjects.
• This approach enhances the practicality and accessibility of motor imagery BCIs for real-world use.