Deep Learning Architectures for Code-Modulated Visual Evoked Potentials Detection.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Non-invasive Brain-Computer Interfaces (BCIs) using Code-Modulated Visual Evoked Potentials (C-VEPs) face challenges with temporal variability and noise in EEG signals.
• Robust decoding methods are crucial for reliable C-VEP-based BCI performance.

Purpose of the Study:

• To propose and evaluate deep learning architectures for robust decoding of C-VEPs.
• To compare deep learning models against traditional methods like Canonical Correlation Analysis (CCA).

Main Methods:

• Developed and tested Convolutional Neural Networks (CNNs) for m-sequence reconstruction and classification.
• Implemented Siamese networks for similarity-based decoding.
• Utilized Earth Mover's Distance (EMD) for distance-based decoding and temporal data augmentation.

Main Results:

• Deep learning models significantly outperformed traditional approaches.
• Distance-based decoding with EMD demonstrated superior robustness to latency variations.
• Temporal data augmentation enhanced generalization across sessions.
• The multi-class Siamese network achieved the highest accuracy at 96.89% for single-trial C-VEP decoding.

Conclusions:

• Data-driven deep learning architectures show significant potential for reliable single-trial C-VEP decoding.
• Siamese networks offer a promising approach for adaptive non-invasive BCI systems.
• Advanced decoding methods are essential for overcoming EEG signal variability in BCIs.