BrainWave-Scattering Net: a lightweight network for EEG-based motor imagery recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) leverage electroencephalography (EEG) signals for direct brain-world communication.
• Convolutional Neural Networks (CNNs) are prevalent in EEG tasks but lack interpretability and struggle with limited data.
• Existing deep learning models for EEG analysis often require substantial training data and computational resources.

Purpose of the Study:

• To develop a novel, lightweight, and interpretable neural network architecture for EEG-based BCIs.
• To address the limitations of current deep learning models in terms of data requirements and interpretability.
• To enable efficient personalized BCI models even with limited training data.

Main Methods:

• Introduction of a fully-learnable neural network architecture utilizing Gabor filters.
• Scattering decomposition of EEG signals along frequency and temporal modulation paths.
• Evaluation in both generic (cross-subject) and personalized (within-subject) classification settings.

Main Results:

• The proposed model achieves high performance across multiple datasets (public and in-house).
• Demonstrated superior efficiency with significantly fewer trainable parameters and reduced training time compared to state-of-the-art deep architectures.
• The network exhibits enhanced interpretability, particularly in its temporal filtering operations.

Conclusions:

• The novel Gabor filter-based neural network offers a promising, efficient, and interpretable solution for EEG-based BCIs.
• This architecture effectively handles limited training data, facilitating personalized BCI development.
• The findings suggest a new direction for designing more accessible and understandable BCI systems.