Research On Key Technologies Of Large Scale Neuromorphological System Based On FPGA

Since the development of brain science,people have been committed to recreating a computing system with the scale and computing power of a human brain,that is,brain-inspired computing based on computational neuroscience.It is difficult to realize brain-inspired intelligence on traditional Von Neumann computers,with high energy consumption and low efficiency.Neuromorphic system has an information processing mechanism close to the human brain,with excellent parallelism,fault tolerance and low power consumption,it's helpful to the realize brain-inspired computing.The FPGA's hardware programmable,highly parallel computing,and low power consumption characteristics can give full play to the advantages of parallel execution of neuromorphic computing,could provide a new way for the realization of neuromorphic computing.This paper put forward a large-scale neuromorphic system,completed the design,verification,compilation,and test the related functional applications.This paper focuses on four aspects of research,including neuron circuit design,neuromorphic system architecture design,system internal modules design,functional verification and application deployment.The main research work of this paper is as follows:(1)Design and verify the LIF neuron circuit.The basic unit of the Spiking Neural Network is the spiking neuron.After comparing various spiking neuron models,the LIF neuron is finally selected as the basic of the system.The mathematical model of the LIF neuron is simplified to obtain a model suitable for the realization of digital circuits.After the introduction of parameter sharing technology,the digital circuit design of a single neuron is completed.Verilog language is used to complete the coding,and after the coding completed,a UVM verification platform is built for simulation verification,and finally a LIF neuron module with simple structure,strong reliability and parameter sharing function is obtained.(2)Complete the design of neuromorphic system architecture.Based on LIF neurons,the neuromorphic system is constructed according to the requirements of scale,function,performance and other indicators.The system design includes function division,interconnection communication design and top-level port logic design.The system is divided into top controller,cores,and transmission modules according to functions;the interconnection communication design determines the data interaction method;the top-level port logic realizes the interaction function between the system and the host computer,and specifies the data transmission timing.After finishing the architecture design,the basic framework of neuromorphological system with good scalability and off chip storage is obtained.(3)Complete the design of the internal modules of the neuromorphic system.Complete the design of the top controller,cores,and transmission module.The cores undertake the neuromorphic computing function,adopt parameter sharing technology and core multiplexing technology,which greatly saves hardware overhead;the top controller is used to control the operation of the system;the transmission module is used to complete the transmission of computing data and parameters.After integrating the above modules,a complete large-scale neuromorphic system is obtained.The system contains 128 physical neurons and a maximum of 1048576 virtual neurons.The scale of the neural network can be selected freely,the parameters of the neuron can be freely configured,and the computing cores can be connected freely.It has strong compatibility and can be used as a hardware platform to achieve various functions.(4)System verification and functional application research.The UVM verification platform is built to complete the simulation verification of the entire system,and the resource overhead is obtained after compiling with Quartus and DC respectively.Discuss the compatibility of the system to different functional applications.After completing the parameter deployment,the system is burned to the FPGA development board for testing,then achieve the handwritten digit recognition function.A PC is used as the host to participate the recognition of 1,056 pictures,and the final accuracy rate is97.92%.After completing the research on Spiking Neural Network and spiking neuron,neuron circuit design,neuromorphic system architecture design,system internal modules design,functional verification and application deployment,I finally realized a large-scale neuromorphic system with high degree of freedom,high compatibility,and low resource overhead.