CS-Net: An End-to-end Network for Motor Imagery Brain-Machine Interface with Adaptive Channel Selection and Compressed Sensing

Motor imagery (MI) based brain-machine interface (BMI) typically relies on multi-channel electroencephalogram (EEG) data and high-complexity algorithms, leading to huge computational burden. This paper proposes an end-to-end neural network for MI classification tasks, namely CS-Net, which integrates a Channel Selection layer and a Compressed Sensing layer to select the optimal channel subset while compressing data automatically. The parameters across the entire network are jointly optimized and the most compact network structure is explored by the Neural Architecture Search (NAS)-based framework. Evaluated on two public MI datasets (the High Gamma Dataset and BCI Competition IV-2a Dataset), the proposed network ensures competitive accuracies with minimal computational complexity. Compared to the full-channel and uncompressed situation, multiply-accumulate operations (MACs) are reduced by 90.90% and 60.72%, while the accuracies only drop by 2.14% and 1.37%, respectively. Compared to other state-of-the-art (SOTA) works, the proposed network achieves a comparable classification accuracy using fewer channels and better generalization ability without the subject-specific method.