Semi-supervised generative and discriminative adversarial learning for motor imagery-based brain-computer interface.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) are used for feature extraction but face challenges in Brain-Computer Interfaces (BCIs).
• Key issues include small training datasets and lack of interpretability in deep learning models.
• Existing methods struggle with effectively utilizing limited and unlabeled electroencephalography (EEG) data.

Purpose of the Study:

• To develop a semi-supervised generative and discriminative learning framework for BCIs.
• To address limitations of small sample sizes and improve interpretability in EEG-based BCIs.
• To discover class-discriminative features by effectively using synthesized and real EEG samples.

Main Methods:

• A semi-supervised framework combining generative and discriminative learning models was employed.
• A generative model learned EEG signal distributions in an embedding space.
• Synthesized and real EEG data were used to train models for class-discriminative spatio-temporal feature representation.

Main Results:

• The proposed framework demonstrated statistically significant performance improvements over conventional linear models.
• Experiments on three public datasets showed superior classification accuracy.
• Analysis of activation maps and generated samples supported the model's stability and neurophysiological plausibility.

Conclusions:

• The developed framework effectively enhances feature extraction for BCIs using semi-supervised learning.
• It successfully leverages unlabeled EEG data to improve discriminative feature discovery.
• The method offers a promising approach for robust and interpretable EEG-based BCIs.