Desgin And Implementation Of A Parameterized Neuron Model Generation System In Brain Simulation Environment

In recent years,researchers have studied the operating mechanism and working principle of biological neural systems through biological brain simulation.To this end,many different spiking neuron models and spiking neural network structures have been proposed,and a brain simulation platform has been built to run pulse neural networks.Running different biological brain simulation tasks in a brain simulation environment requires constructing different pulse neural networks,which typically use different neuron models.Providing different neural models and reducing the consumption generated during the calculation process of neural models for brain simulation tasks is a valuable issue.This thesis mainly studies the design and implementation of a "parameterized neuron model generation system".The main work of this thesis is as follows:1)In order to solve the problem of neuron parallel computing of pulse neural network in brain simulation platform,the vectorization scheme of neuron computing is studied based on the primitive language of brain like chip provided by brain simulation platform.In this scheme,this thesis designs the implementation scheme of vectorization division operation,the implementation scheme of vectorization conditional judgment statement,and the static memory allocation scheme of neuron model data supporting vectorization calculation.2)In order to generate a customizable neural model assembly language program file for a certain brain simulation platform,this thesis designs and implements a neural model generation module in the system.In the neuron model generation module,based on the chip primitive library function in the simulation platform,the spiking neuron model high-level programming language program supporting neuron parallel computing is coded and implemented.The chip primitive library function can generate the chip assembly instruction code corresponding to this function during execution;After the neuron model generation module receives the parameter settings such as the number of neurons set by the user,neuron parameter values such as neuron model particle conductance,and initial values of state variables such as neuron model membrane potential,it executes the pulse neuron model high-level programming language program in the module to obtain the parameterized neuron model assembly language program file that can be run on a brain simulation platform.3)In order to support users in studying spiking neurons with different temporal dynamics models,this thesis designs and implements a custom neuron model generation module in the system.This module accepts the spiking neuron model defined by the four arithmetic operations and basic conditional judgment statements in NESTML language.When the spiking neuron model defined by the user can use the Euler method to solve its temporal dynamics model,this module can generate the Python code file of neuron model that can run in the PyTorch library environment.The custom neuron model generation module first uses NESTML’s grammar processing tool to parse the user input custom neuron model text,obtaining the abstract syntax tree of the custom neuron model.Then,key information such as expressions and variables are extracted from the abstract syntax tree.Finally,a set of specified Jinja2 template files are used to generate the Python code file of the neuron model.This thesis discusses the requirements analysis,design,and implementation of a parameterized neural model generation system.Finally,the system testing analysis is completed,and the work of the thesis is summarized.