Towards neuromorphic brain-computer interfaces: Model and circuit Co-design of the spiking EEGNet

Abstract

The compact end-to-end deep neural network model, EEGNet, scalable to various brain-computer interface (BCI) paradigms, has been realized in hardware but with low efficiency. Neuromorphic circuit with high efficiency has not yet been applied to BCI. In this paper, the activation-quantization-based (AQB) method that converts EEGNet to spiking neural network is proposed and optimized on the standard motion imagery dataset BCIC-IV-2a with an accuracy of 63.35%. The proposed circuit of the AQB EEGNet is implemented in FPGA, and shows an energy efficiency of 318μJ/inference, which is comparable to that of a low power MCU at 337μJ/inference. Estimation of the energy efficiency in neuromorphic hardware shows the high efficiency of 34μJ/inference, which is a huge improvement compared to previous works, and benefits the BCI for long-term use.