A Lightweight End-to-End Neural Networks for Decoding of Motor Imagery Brain Signal

Abstract

In motor imagery-based brain-computer interface (MI-BCI), the variants of convolutional neural networks (CNNs) have been increasingly received attention due to relatively outstanding decoding performance. However, the growing network size for high decoding performance and the inefficient procedures of BCI systems lead to limited availability in real-life MI-BCI systems. To tackle these issues, we propose an end-to-end neural network named lightweight EEG-inception squeeze-and-excitation network (LiteEEG-ISENet). The architecture is built to remedy the two parts: 1) depthwise convolution is adopted to reduce the computational complexity of the network and train the intrinsic features for each channel of the MI dataset; 2) In addition to the previous motivation, the squeeze-and-excitation (SE) blocks are employed to recalibrate channel-wise feature response adaptively. The experimental results on the public dataset widely used in the MI-BCI study demonstrate that the proposed method outperforms the existing method in terms of decoding performance and neural network memory efficiency.