Cross-Subject EEG Signal Classification with Deep Neural Networks Applied to Motor Imagery

The Brain-Computer Interface (BCI) is a system able to serve as a mean of communication between machine and human where the brainwaves are the control signals acquired by electroencephalography (EEG). One of the most used brainwaves is the sensorimotor rhythm (SMR) which appears for real or imagined motor movement. In general, EEG signals need feature extraction methods and classification algorithms to interpret the raw signals. Deep learning approaches; however, permit the processing of the raw data without any transformation. In this paper, we present a deep learning neural network architecture to classify SMR signals due to its success for some previous works and to visualize the learned features. The architecture is composed of three parts. The first part contains a temporal convolution operation followed by a spatial convolution one. The second part contains recurrent layers. Finally, we use a dense layer to assign the signal to its class. The model is trained with Adam optimizer algorithm. Also, we use various regularization techniques such as dropout to prevent learning problem like overfitting. To evaluate the performance of the proposed architecture, the well known Dataset IIa of the BCI Competition IV is used. As a result, we get equivalent results to those ones of EEGNet.