POC-CSP: a novel parameterised and orthogonally-constrained neural network layer for learning common spatial patterns...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Common Spatial Patterns (CSP) is a key feature extraction technique for EEG.
• Traditional CSP methods are sensitive to noise and have fixed weights.
• Integrating CSP into trainable neural networks is challenging.

Purpose of the Study:

• To develop a trainable CSP model that learns from data and integrates into end-to-end networks.
• To overcome the limitations of conventional CSP, such as noise sensitivity and weight rigidity.
• To enhance EEG signal processing for brain-computer interfaces.

Main Methods:

• Developed a novel parameterised and orthogonally-constrained neural network layer for learning CSPs (POC-CSP).
• Utilized Lie Group theory for parameterization, converting constrained optimization to unconstrained optimization.
• Integrated the POC-CSP layer into standard neural network training methods.

Main Results:

• POC-CSP demonstrated superior performance over conventional CSP and existing neural network approaches in subject-specific tasks.
• Achieved superior generalization in a novel multi-subject paradigm.
• Attained 0.95 average accuracy across subjects when fine-tuned with only 50% of new subject data.

Conclusions:

• Combining CSP's effectiveness with neural network flexibility significantly enhances EEG signal processing.
• POC-CSP offers improved generalization across subjects and high accuracy with minimal subject-specific data.
• This approach is highly valuable for practical brain-computer interface applications.