An FPGA-based Implementation Method For Networks Of Piecewise Linear Spiking Neurons

Spiking neural networks that composed of biologically plausible spiking neurons are usually known as the third generation of artificial neural networks,which are shown to be suitable tools for the processing of spatio-temporal information.Combing the reprogrammable feature of Field Programmable Gate Arrays(FPGA)and the information processing model of spiking neural networks,the hardware implementation of independent learning ability for large scale spiking liquid state machines is a significant and challenging task in the artificial intelligence research area.However,in the current researches on the hardware implementation of spiking neural networks,not only the performance of the neuron model is relatively simple,but also the design of the network hardware architecture is unreasonable enough,which results in the lack of biological physiological activities and neurodynamic properties.This paper analyzes the current research status of neuromorphic systems,and then compares the differences between digital systems,analog circuits,and hybrid implementations.According to the two-dimensional piecewise linear spiking neuron model,the hardware architecture of spiking neural network and coupled random networks with synaptic connections,the spatio-temporal property of spiking neural network communication system is designed and implemented based on FPGA.The main work is as follows:(1)We propose the circuit design and implementation of neuron model based on FPGA by introducing the two-dimensional piecewise linear spiking neuron model.It is discretized with the Euler method for the original model equations,and moreover,a digital implementation is used to design the schematic diagram of the parallel arithmetic unit for neuron and the digital calculation pipeline of the membrane potential V and the recovery variable U,so that the arithmetic tree can maximize the parallelism of time and space.In the experiment,we analog 20 neural computational properties and 6 biological cortical neurons firing patterns,and also verify the parallel computing performance and biologically interpretable of digital circuits.(2)We propose a hardware architecture of communication system for spiking neural networks,which combines synapse interaction mechanism and coupled random networks to design and perform digital circuits.The real-time simulation of spike trains firing and processing is performed to make it possess the spatio-temporal property and biological reality.The proposed communication system is evaluated through experiments,we simulate the spatio-temporal for spike trains of 2 different coupled neural networks,and also verify the neurodynamic property of the communication system.Finally,through experimental analysis and comparison,the proposed works can fire and process spike trains which include spiking neuron digital circuit and spiking neural network communication system,verifying the properties of the expressed neuromorphic dynamics.Furthermore,by analyzing the performance data of communication system simulation,it shows a good resource utilization and power loss.