Explainable AI Insights Into EEG Classification and Its Alignment to Neural Correlates

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning significantly enhances electroencephalography (EEG) analysis but lacks interpretability.
• Understanding what deep learning models learn from neural data is crucial for neuroscientific insights.

Purpose of the Study:

• To develop a comprehensive interpretability framework for deep learning models of neural data.
• To analyze abstract concepts encoded by neurons and filters in EEG models.
• To bridge the gap between deep learning features and neuroscientific concepts.

Main Methods:

• Concept Relevance Propagation (CRP), an extension of layer-wise relevance propagation, was developed.
• CRP was applied to filters of EEGNet models trained via leave-one-out cross-validation.
• Relevance maximization, UMAP dimensionality reduction, and density-based clustering identified filter concepts.
• A virtual inspection layer projected explanations into the frequency domain for spatial, temporal, and spectral analysis.

Main Results:

• The framework revealed interpretable, task-relevant neural patterns across auditory attention, attention, and motor imagery tasks.
• Learned features generalized across participants, demonstrating robustness.
• The approach provided insights into task-specific neuroscience without model retraining.

Conclusions:

• The developed framework enhances the interpretability of deep learning models in EEG analysis.
• It offers a pathway to understand model learning and gain neuroscientific insights.
• This work facilitates a deeper connection between artificial intelligence and brain function studies.