Research On Algorithms And Implementation Of Spiking Neural Network

Deep learning approaches have achieved exceptional results on the tasks of image recognition,speech recognition,and natural language processing,but deep learning uses continuous values for the communication of neurons,and the weights between neurons are learned by backpropagation,which is biologically implausible and energyconsuming.In contrast,spiking neural network(SNN)models are inspired by brain,in which the neurons communicate using spikes,and they have shown great potential to become alternatives to deep learning.However,the development of SNNs is still in a primary stage: existing SNN models have bad performances,and the implementation of SNNs lacks a mature programming framework.Therefore,this thesis explores the algorithms and implementation of SNNs,and proposes a flexible and fast method to implement SNNs,and a novel SNN model targeting speech recognition tasks.First,this thesis introduces the current research status of the SNN area,existing SNN models and related experiments.Then,different programming frameworks of spiking neural network are compared and reviewed by experiment.A new method of implementing spiking neural networks with NumPy and PyCUDA is proposed.Finally,this thesis presents a convolutional neural network,which utilizes the synaptic plasticity as the learning rule to complete the task of speech recognition,and achieves high performance which is comparable to traditional methods.