Grad-TopoCAM: EEG Brain Region Visual Interpretability via Gradient-Based Topographic Class Activation Map
Keywords: Electroencephalogram, Class Activation Map, Deep Learning, Visualization, Interpretability
Abstract: The visualization and interpretability of electroencephalogram (EEG) decoding significantly contribute to brain-computer interfaces (BCI) and cognitive neuroscience. Although some existing research has attempted to map EEG features to specific brain regions, these approaches fail to fully utilize raw signals and lack extensibility to other Deep Learning (DL) models. In this work, Grad-TopoCAM (Gradient-Based Topographic Class Activation Map) is proposed, which enhances interpretability in DL models for EEG decoding adaptively. Grad-TopoCAM calculates the gradient of feature maps for the target class at the target layer. The weights of the feature maps are obtained through global average pooling of the gradients. The class activation map is generated by performing a linear combination of weights and feature maps, which is subsequently mapped to different brain regions. Grad-TopoCAM is validated across eight DL models on four public datasets. Experimental results indicate that Grad-TopoCAM effectively identifies and visualizes brain regions that significantly influence decoding outcomes, while also facilitating channel selection for different decoding tasks. The code and data are open-source.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10439
Loading