Multi-band CNN with Band-dependent Kernels and Amalgamated Cross Entropy Loss for Motor Imagery Classification

Motor imagery (MI) electroencephalography (EEG) signal has been widely used as control commands in Brain-computer interface (BCI). However, the limited performance of MI classification makes it difficult to apply BCI to everyday life. Recently, deep-learning, specifically convolutional neural network (CNN), based approaches have been proposed to improve classification performance, but they suffer from subject dependency issue due to the kernel size optimization problem. In this paper, we present a novel MI classification method based on multi-band CNN with band-dependent kernel sizes, named MBK-CNN, to improve classification performance. The proposed structure exploits the frequency diversity of the EEG signals and resolves the subject dependent kernel size issue at the same time. EEG signal is decomposed into overlapping multi-band and then passed through multiple CNNs (termed 'branch-CNNs') equipped with different kernel sizes to generate frequency dependent feature vectors. The features are then combined by a simple weighted sum. In contrast to the existing works where single-band multi-branch CNNs with different kernel sizes are used to resolve the subject dependency issue, a unique kernel size per frequency band is used. To prevent possible overfitting induced by a weighted sum, each branch-CNN is additionally trained by tentative cross entropy loss while overall network is optimized with respect to the end-to-end cross entropy loss, which is named amalgamated cross entropy loss. In addition, we further propose multi-band CNN with enhanced spatial diversity, named MBK-LR-CNN, by replacing each branch-CNN with several sub branch-CNNs applied for channel subsets (termed 'local region') to further improve the classification performance. We evaluated the performance of the proposed methods, MBK-CNN and MBK-LR-CNN, on publicly available datasets, BCI Competition IV dataset 2a and High Gamma Dataset. The experimental results confirm the performance improvement of the proposed methods in comparison with the currently studied MI classification methods.