An Efficient Motor Imagery BCI Classification Architecture Using Non-linear Network Tuning suitable for Cloud-Edge Environments

This paper proposes a novel deep learning architecture for EEG-based Brain-Computer Interface (BCI) systems, optimized for real-world deployment. BCI is fundamentally a communication problem, yet many works focus on lab-based implementations. Our approach addresses practical implementation considerations crucial for use beyond controlled environments. By incorporating a non-linear Network Tuning Block (NTB) and applying a two-phase training process, our model enhances classification accuracy while minimizing computational demands. Tested on the BCI Competition IV-2b dataset, our architecture achieves a $\mathbf{9 3. 2 \%}$ accuracy, surpassing state-of-the-art methods with fewer parameters, underscoring its potential for practical, real-world applications.