Design And Implementation Of Deep Spiking Neural Network Framework

Deep learning has achieved significant breakthroughs in the field of intelligent applications by drawing lessons from the hierarchical information processing structure of the brain at a macroscopic scale.However,deep learning typically requires a lot of computing resources and power consumption,making it difficult to deploy on resource-constrained devices.Inspired by the biological brain,spiking neural networks simulate the information encoding and processing mechanisms of biological neurons at a microscopic scale,with the potential for ultra-low power consumption and latency.Therefore,combining the advantages of deep learning and spiking neural networks,the research of deep spiking neural network models is expected to achieve a new paradigm of brain-like intelligence with high intelligence,low power consumption,and low latency.However,the current applications of deep spiking neural networks are not as widespread as deep learning models,partly because their discrete spiking mechanism makes network training very difficult,and partly because their more complex model operation mechanisms increase the difficulty of programming.Existing frameworks mostly focus on simulating the characteristics of spiking neurons,but lack tools for constructing and training various deep spiking neural network models.To address these issues,the thesis designs and implements a programming framework based on deep spiking neural networks.At the same time,a new temporal coding based deep spiking neural network learning algorithm is integrated into the programming framework.The specific research contents include:(1)The thesis designs and implements a programming framework for deep spiking neural networks,which reduces the programming complexity of deep spiking neural networks.The framework is based on the Pytorch framework and includes five modules:dataset,information encoding,synapse,neuron,and training algorithm modules,which fully support the commonly used neuron models and their corresponding coding methods and learning algorithms for deep spiking neural networks.This can help researchers in various fields to easily implement the entire process from spiking input information processing to model construction and then to model training.(2)The thesis proposes a deep spiking neural network conversion learning algorithm based on temporal coding and integrates it into the framework,making the framework more complete for training deep spiking neural networks.Firstly,a synchronous spiking neural network model is proposed to solve the problem of temporal dynamic characteristics,and a corresponding surrogate artificial neural network model is derived based on this model.Since these two models have high equivalence to ensure that the converted spiking neural network model still has low conversion loss even in the case of a single spike emission.Experimental results show that this algorithm achieves good performance while ensuring network sparsity.(3)The thesis verifies the framework in two pattern recognition tasks respectively.By using the modules in the framework to load and process the dataset,build and train the neural network model,thesis demonstrates the framework’s programming simplicity in building deep spiking neural network applications and the high performance of the trained models.