Explaining Convolutional Neural Networks for EEG-based Brain-Computer Interface Using Influence Functions

Although the high performance of the convolutional neural networks (CNNs) for brain-computer interface (BCI) tasks based on raw electroencephalography (EEG) signals, the explanation of the prediction result remains challenging owing to their complex structure and numerous parameters. We propose a novel framework for explaining CNNs for EEG-based BCI tasks by using the perturbation-based influence scores. The method supports the interpretation of CNN classification for EEG signals at both the example-level and the feature-level. The experiments on the BCIC III-IVa dataset demonstrate that the proposed method is effective for not only the interpretation of the predictive models, but also for the improvement of the classification accuracy.