EEG Signal Classification Based on Neural Network with Depthwise Convolution

Abstract

Brain-Computer Interface (BCI) have attracted a lot of attention recently as it enables people to communicate between the human and the machine by electroencephalography (EEG) signals decoding. EEG signals classification is important for BCI system. Convolution neural network (CNN) can automatically extract features and enhance the classification accuracy. However, limited EEG data easily leads to over-fitting of neural network. In this study, we proposed two novel neural networks (DCNN and DSCNN) based on depthwise convolution for EEG signals classification. The proposed models used depthwise convolution to learn a spatial filter and fuse channels information. And we performed multiple experiments to evaluate the EEG signal classification performance of the proposed CNN models on multiple sets of public datasets. The proposed DCNN method achieved 88.33% and 78.00% on the dataset III and the dataset IV from BCI competition II. The proposed DSCNN method could also achieve very competitive accuracy. In addition, the proposed models exhibited a significant ability in reducing model parameters and mitigating over-fitting.