A soma‐synapses neuron model and FPGA implementation

Abstract

SummaryThe neuron model serves as the foundation for building a neural network. The goal of neuron modeling is to shoot a tradeoff between the biological meaningful and the implementation cost, so as to build a bridge between brain science knowledge and the brain‐like neuromorphic computing. Unlike previous neuron models with linear static synapses, the focus of this research is to model neurons with relatively detailed nonlinear dynamic synapses. First, a universal soma‐synapses neuron (SSN) is proposed. It contains a soma represented by a leaky integrate‐and‐fire neuron and multiple excitatory and inhibitory synapses based on ion channels dynamics. Short‐term plasticity and spike‐timing‐dependent plasticity linked to biological microscopic mechanisms are also presented in the synaptic models. Then, SSN is implemented on field‐programmable gate array (FPGA). The performance of each component in SSN is analyzed and evaluated. Finally, a neural network SSNN composed of SSNs is deployed on FPGA and used for testing. Experimental results show that the stimulus‐response characteristics of SSN are consistent with the electrophysiological test findings of biological neurons, and the activities of SSNN exhibit a promising prospect. We provide a prototype for embedded neuromorphic computing with a small number of relatively detailed neuron models.