Cross-dataset transfer learning for motor imagery signal classification via multi-task learning and pre-training

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning (DL) models excel at decoding motor imagery (MI) from Electroencephalogram (EEG) data.
• However, DL models require extensive training data, which is challenging to acquire for EEG.
• Cross-dataset transfer learning offers a solution, but transferring knowledge between datasets with different MI tasks is difficult.

Purpose of the Study:

• To propose a novel pre-training-based cross-dataset transfer learning method for EEG-based MI decoding.
• To address the challenge of transferring knowledge across datasets with distinct MI paradigms.
• To enhance the practicality and user-friendliness of Brain-Computer Interfaces (BCIs).

Main Methods:

• A pre-training approach inspired by Hard Parameter Sharing in multi-task learning was developed.
• Distinct MI datasets were treated as separate tasks within a unified model featuring shared feature extraction and task-specific layers.
• Knowledge transfer was achieved through pre-training and fine-tuning, with four fine-tuning schemes explored.

Main Results:

• Pre-trained models showed a maximum accuracy increase of 7.76% compared to non-pre-trained models.
• With limited data, pre-training improved DL model accuracy by up to 27.34%.
• Pre-trained models demonstrated faster convergence, reduced training time per subject (up to 102.83s), and improved robustness (variance decreased by 75.22%).

Conclusions:

• This study presents the first comprehensive investigation of cross-dataset transfer learning for differing MI tasks.
• The proposed pre-training method significantly reduces the need for fine-tuning data when applying DL models to new MI paradigms.
• This advancement makes Brain-Computer Interfaces more accessible and user-friendly.