SPD-CNN

SPD-CNN applies symmetric positive definite (SPD) matrix transformation, convolutional neural networks (CNNs), and meta-transfer-learning (MTL) to perform cross-subject electroencephalography (EEG) classification for Brain-Computer Interface (BCI) applications.

Key Features:

• Symmetric Positive Definite (SPD) matrix transformation: Transforms raw EEG signals into SPD matrices to capture intrinsic geometric properties of the data and reduce sensitivity to EEG noise and artifacts.
• Convolutional Neural Networks (CNNs): Processes SPD matrices with CNN architectures for automatic feature extraction and classification of spatial hierarchies in EEG data.
• Meta-Transfer Learning (MTL): Employs meta-transfer-learning to enable generalization across subjects and reduce the need for subject-specific calibration.

Scientific Applications:

• Cross-subject EEG classification: Improves classification accuracy across different individuals by combining SPD representations, CNNs, and MTL to address domain shifts and inter-subject variability.
• Brain-Computer Interfaces (BCIs): Supports development of calibration-reduced BCI systems by enabling cross-subject model transfer for cognitive and motor tasks.
• Evaluation on motor-imagery datasets: Validated using three public motor-imagery datasets to assess cross-subject classification performance.

Methodology:

Transform raw EEG signals into SPD matrices, input SPD matrices to a CNN architecture for feature extraction and classification, and apply meta-transfer-learning to refine model generalization across subjects.