[Motor imagery classification based on dynamic multi-scale convolution and multi-head temporal attention].

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Context:

• Motor imagery (MI) electroencephalogram (EEG) signals are crucial for brain-computer interfaces.
• Convolutional neural networks (CNNs) are widely used for MI-EEG classification but struggle with individual signal variability.
• Existing methods often fail to capture complex spatiotemporal dynamics effectively.

Purpose:

• To propose a novel dynamic multi-scale CNN and multi-head temporal attention (DMSCMHTA) model for robust MI-EEG signal classification.
• To address the challenge of strong inter-individual variability in MI-EEG data.
• To enhance the extraction of discriminative spatiotemporal features from MI-EEG signals.

Summary:

• The DMSCMHTA model employs multi-band filtering, dynamic multi-scale CNN for temporal feature extraction, and spatial convolution for spatiotemporal feature extraction.
• A multi-head attention mechanism is utilized to optimize temporal correlations and reduce dimensionality, generating more discriminative features.
• Classification is achieved using cross-entropy and center loss, with experimental validation on BCI Competition IV datasets.

Impact:

• The DMSCMHTA model achieves high average accuracies of 80.32% and 90.81% on BCI Competition IV datasets 2a and 2b, respectively.
• Demonstrates superior performance compared to current mainstream methods in MI-EEG classification.
• Highlights the model's capability for adaptive, personalized spatiotemporal feature extraction, paving the way for improved brain-computer interfaces.