Deep and Wide Transfer Learning with Kernel Matching for Pooling Data from Electroencephalography and Psychological...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) is crucial for motor learning and brain-computer interface (BCI) systems using electroencephalogram (EEG) decoding.
• Inefficiency in MI self-regulation necessitates longer training periods for BCI users.
• Improving BCI performance for individuals with poor MI skills is a significant challenge.

Purpose of the Study:

• To develop a parameter-based cross-subject transfer learning approach to enhance MI-BCI performance.
• To improve the accuracy of BCI systems for individuals with MI inefficiency.
• To integrate EEG data with psychological questionnaire data for improved BCI performance.

Main Methods:

• Implemented a Deep and Wide neural network for MI classification and pre-training.
• Utilized a fine-tuning procedure to transfer network parameters to a target network.
• Employed stepwise kernel-matching via Gaussian-embedding for data fusion of categorical and real-valued features.
• Applied inefficiency-based clustering for subject selection and evaluation of source-target sets.

Main Results:

• The proposed Deep and Wide neural network demonstrated competitive accuracy in MI classification.
• Transfer learning effectively improved performance for individuals with MI inefficiency.
• Integration of questionnaire data alongside EEG data did not compromise accuracy.

Conclusions:

• The parameter-based cross-subject transfer learning approach is effective in enhancing MI-BCI performance.
• This method offers a promising solution for users with MI inefficiency.
• The integration of diverse data sources, including psychological data, can improve BCI system robustness.