Interpretable deep neural networks for single-trial EEG classification

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Deep neural networks (DNNs) show promise for complex classification but lack interpretability.
• The 'black box' nature of DNNs hinders understanding of underlying neurophysiological processes.
• Layer-wise relevance propagation (LRP) offers a method to explain individual DNN decisions.

Purpose of the Study:

• To apply DNNs with LRP for the first time to electroencephalography (EEG) data analysis.
• To transform single-trial DNN decisions into relevance heatmaps for neurophysiological insight.
• To compare DNN-LRP performance against established methods in motor-imagery Brain-Computer Interfaces (BCIs).

Main Methods:

• Utilized deep neural networks (DNNs) for EEG data classification.
• Applied layer-wise relevance propagation (LRP) to generate relevance heatmaps from DNN decisions.
• Compared classification accuracy of DNNs with LRP against CSP-LDA on motor-imagery BCI datasets.

Main Results:

• DNNs achieved classification accuracies comparable to CSP-LDA.
• Subject-to-subject transfer of DNNs improved performance in low-performing subjects.
• Single-trial LRP heatmaps displayed neurophysiologically plausible patterns, similar to CSP scalp maps but with higher temporal specificity.

Conclusions:

• DNNs are powerful non-linear tools for EEG analysis.
• LRP provides high-resolution assessment of neural activity, enhancing DNN interpretability in neuroscience.
• LRP's specificity opens new research avenues for neural processes in perception and decision-making.