Weighted Transfer Learning of Dynamic Time Warped Data for Motor Imagery based Brain Computer Interfaces

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signals are crucial for brain-computer interfaces (BCI).
• The non-stationary nature of EEG data necessitates large calibration datasets for effective BCI model training.
• Existing transfer learning methods often struggle with temporal variations in EEG data.

Purpose of the Study:

• To develop a novel weighted transfer learning algorithm for EEG-BCI.
• To reduce the requirement for extensive subject-specific calibration data.
• To improve BCI model performance using data from other subjects.

Main Methods:

• Dynamic Time Warping (DTW) for aligning temporal variations between subject data and external datasets.
• Kullback-Leibler (KL) divergence to measure similarity between aligned data.
• A weighted transfer learning approach, where external data is weighted based on KL similarity.
• Training subject-specific BCI models using weighted external data.

Main Results:

• The proposed algorithm demonstrated an average improvement of 9% over subject-specific models trained with limited data (4 trials).
• A 10% average performance improvement was observed compared to naive transfer learning methods.
• The method effectively alleviates the need for large calibration datasets in EEG-BCI.

Conclusions:

• The DTW-aligned KL weighted transfer learning algorithm shows significant promise for EEG-BCI.
• This approach enables efficient BCI model training with minimal subject-specific calibration data.
• The findings contribute to more accessible and practical BCI applications.