New avenues for understanding what deep networks learn from EEG

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning models are increasingly used for electroencephalography (EEG) decoding.
• Understanding the specific features learned by these models remains a significant challenge.
• Previous interpretability studies often focused on limited aspects, not the complete network.

Purpose of the Study:

• To uncover and visualize the features learned by complete deep learning networks for EEG decoding.
• To develop novel interpretability methods for analyzing neural network functions in EEG analysis.
• To compare learned features against established EEG knowledge and identify novel patterns.

Main Methods:

• Introduced two complementary network architectures with dedicated visualization techniques.
• Utilized invertible networks to generate class-specific prototypical EEG signals.
• Developed a compact, fully visualizable network for EEG decoding.
• Applied these methods to binary classification of nonpathological and pathological EEG data.

Main Results:

• Identified expected features, such as increased delta and theta band oscillations in pathological EEG.
• Discovered unexpected differences in sub-delta frequencies (<0.5 Hz) between healthy and pathological EEG.
• Revealed higher spectral amplitudes in sub-delta frequencies at frontal sensors for the healthy class, a novel finding.

Conclusions:

• Visualization techniques can effectively reveal features learned by complete deep learning networks for EEG decoding.
• The study identified both anticipated and novel EEG features, enhancing understanding of pathological conditions.
• This approach offers a powerful tool for interpreting complex neural network behavior in biomedical signal analysis.