A Temporal Convolution Network Solution for EEG Motor Imagery Classification

Abstract

EEG motor imagery recognition based brain computer interface has been an import scheme to construct an alternative pathway of the brain to the outside world. EEG signal is usually buried in noise and has very low signal to noise ratio (SNR), which has presented great challenge for efficient motor imagery classification. In addition, the large intra-subject and inter-subject signal variance toward one specific motor imagery also brings difficulty for accurate classification. In recent years, some deep learning solutions based on AutoEncoder, Restricted Boltzmann Machine, CNN and RNN have been proposed for EEG motor imagery classification which have well improved the motor imagery classification accuracy. However, the multi-subject and multi-task motor imagery classification problem remains a challenge. The high computational cost of the existing deep learning solutions is another serious issue to be addressed. In this paper, a new motor imagery classification solution based on Temporal Convolutional Network (TCN) is developed. The dilated causal convolution within TCN could well incorporate the temporal information in a parallel way with much higher computational efficiency than the traditional RNNs. Time stacked spatial EEG signal has been employed as the input to the TCN. Based on which, both the spatial distribution information and temporal variation of the brain signal have been considered. Extensive experiments have shown that the proposed TCN solution has obtained state of the art performance on multi-subject and multi-task motor imagery classification. A high classification accuracy as 97.89% on 20 subjects and 5 tasks has been reached.