Constructing lightweight and efficient spiking neural networks for EEG-based motor imagery classification

Spiking neural networks (SNNs) are gaining attention across various fields, including EEG-based brain–computer interfaces. SNNs utilize spike sequences to characterize and convey information, offering a more bio-interpretable approach and consuming less energy than artificial neural networks (ANNs). However, the binary nature of spike sequences makes it challenging for SNNs to be trained directly using the backpropagation method like ANNs. A prevalent SNN training method involves converting ANNs to SNNs, while the converted SNN often suffers severe performance degradation problems. We theoretically derive the ANN to SNN conversion conditions. Based on this, we comprehensively analyze the sources of conversion errors, and it seems that excessive ANN parameters may result in accuracy loss during conversion. Building upon this analysis, we propose a lightweight and efficient neural network (LENet) for motor imagery (MI) classification. Specifically, LENet can effectively capture spatiotemporal features while replacing the traditional fully connected layer (FCL) with the classification convolution block (CCB), which can reduce the model parameters compared to FCLs. Through experiments on two MI datasets (BCI IV-2a and IV-2b), LENet not only effectively reduces conversion errors but also adeptly balances parameters and performance. The results show that LENet outperforms state-of-the-art classification methods. Meanwhile, our proposed CCB effectively improves the performance of baseline methods compared to FCL. Furthermore, through an analysis of energy consumption, we observe a substantial reduction in power consumption for SNNs compared to ANNs. This work introduces novel perspectives for future SNN application scenarios for the non-stationary MI signals.