Residual Learning Attention CNN for Motion Intention Recognition Based on EEG Data

Abstract

Brain computer interfaces (BCIs) recognize users' intentions to directly control the external environment or devices. Thus, accuracy and robustness of decoding algorithms have an essential effect on the performance of BCIs. Traditional algorithms depend on manually extracting features and updating parameters in various experimental conditions, which limit BCI paradigm application in reality. Nowadays, inspired by the success of deep learning in computer vision and speech recognition applications, scientific researchers take advantages of convolutional networks to recognize EEG motion intention with superior performance. Although the accuracy of methods based on convolutional networks can reach or exceed that of traditional machine learning methods, it is still not high enough to guarantee widespread application, especially in complex tasks, like multi-dimension motion intention recognition. In this paper, we propose a novel motor imagery EEG decoding method based on Residual Learning Attention Convolutional Neural Network (RLA-CNN). First, RLA-CNN architecture is demonstrated based on the formulated motion intention recognition problem. The basic idea of RLA-CNN is to learn attention weights of EEG channel, time points, and feature maps automatically, which can simplify the decoding algorithms and working electrodes in BCIs. Second, Compared with previous work in four, three, and two classes tasks on the same dataset, the average validation accuracy of the proposed method is 75.85%, 84.81%, and 84.69% respectively, which outperforms the state-of-the-art by 5.65%, 10.06%, and 7.41%. Finally, the attention weights learned by RLA-CNN are visualized to explain the effectiveness of RLA-CNN.