Deep Learning Solutions for Motor Imagery Classification: A Comparison Study

Motor imagery classification has been widely applied in constructing brain computer interface to control the outside equipment as an alternative neural muscular pathway. EEG as the most popular non-invasive brain signal suffers from low signal to noise ratio and unpredictable pattern variation even for the same subject. To improve the classification accuracy of EEG based motor imageries, many deep learning based solutions have been developed, mainly including convolutional neural network (CNN) based methods and recurrent neural network (RNN) based methods. There is no unanimous acknowledgement of the most appropriate deep learning solution for motor imagery classification. In order to evaluate the performance of different deep learning solutions for motor imagery classification, a comprehensive comparison study has been conducted in this paper. CNN based method, RNN based method, temporal convolution network (TCN) based method, paralleled combination of CNN and SRU (Simple Recurrent Unit), cascaded combination of CNN and SRU have been constructed and compared based on extensive experiments. The experiments have been conducted on a fair basis with the same dataset, the same preprocessing of the data, and the same platform. Experiments have shown that TCN based method has obtained the best performance and the paralleled combination of CNN and RNN has obtained the second best performance, which inspired us to explore the spatial-temporal feature learning deep network solutions for further improvement.