Rethinking Saliency Map: A Context-Aware Perturbation Method to Explain EEG-Based Deep Learning Model.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning models are increasingly used for electroencephalogram (EEG) signal decoding.
• Explaining these EEG-based deep learning models remains a significant challenge due to data characteristics like non-stationarity and inter-subject variability.

Approach:

• This paper reviews existing methods for explaining EEG-based models and identifies their limitations.
• A novel context-aware perturbation method is proposed, extending instance-level saliency identification by incorporating representative context information.
• The method also leverages context to suppress artifacts in EEG deep learning models and includes an optional area limitation strategy for simplified explanations.

Key Points:

• Existing explanation methods struggle with the inherent complexities of EEG data.
• The proposed context-aware perturbation method enhances saliency map estimation by capturing relevant context.
• Contextual information plays a dual role: improving explanation accuracy and mitigating model artifacts.

Conclusions:

• The developed context-aware perturbation method offers a more effective approach to explaining EEG-based deep learning models.
• Experimental results on the DEAP dataset demonstrate the advantages of the proposed method over existing techniques.
• The method's ability to handle EEG data characteristics and provide user-friendly explanations is highlighted.