Related Experiment Video
Updated: Jan 9, 2026
08:51
Statistical Modelling of Cortical Connectivity Using Non-invasive Electroencephalograms
Published on: November 1, 2019
6.0K
What does my network learn? Assessing interpretability of deep learning for EEG
Pinar Göktepe-Kavis1,2, Florence M Aellen1,2, Sigurd L Alnes1,2
1Institute of Computer Science, University of Bern, Bern, Switzerland.
Imaging Neuroscience (Cambridge, Mass.)
|December 5, 2025
Summary
Deep learning models like EEGNet and ResNet show promise for analyzing electroencephalography (EEG) data. Careful selection of network architecture and visualization methods is crucial for interpreting these complex models.
Area of Science:
- Neuroscience
- Machine Learning
- Signal Processing
Background:
- Electrophysiological studies increasingly use multivariate pattern analysis.
- Traditional machine learning methods often assume consistent response latencies, limiting their application.
- Deep learning offers high performance in EEG analysis but faces challenges in feature interpretability.
Purpose of the Study:
- To evaluate how preprocessing, network architecture, and feature extraction/visualization impact deep learning interpretability for electroencephalography (EEG) data.
- To compare the interpretability of EEGNet and ResNet convolutional neural networks (CNNs) for EEG decoding.
- To assess the influence of gradient-based visualization techniques (saliency, GradCam) on feature interpretability.
Main Methods:
- Trained two CNNs, EEGNet and ResNet, to decode single-trial EEG responses to visual and auditory stimuli.
- Extracted and visualized learned features using saliency and Gradient-weighted Activation Maps (GradCam).
- Compared the performance and feature interpretability of the two CNN architectures and visualization techniques.
Main Results:
- EEGNet and ResNet demonstrated comparable decoding performance.
- Different network architectures learned distinct data features; EEGNet features showed higher similarity to raw EEG data.
- GradCam visualization yielded features more aligned with EEG data compared to saliency mapping.
- Visualization techniques influenced the perceived latency and distribution of important electrodes.
Conclusions:
- Network architecture and feature visualization significantly impact the interpretability of deep learning models for EEG.
- EEGNet may offer better interpretability than ResNet due to feature similarity to EEG data.
- GradCam is a more effective visualization technique for EEG interpretability than saliency.
- Further research into architecture and visualization is essential for advancing deep learning in EEG research.
