Font Size: a A A

Research Of Pattern Recognition Based On Spiking Neural Networks

Posted on:2018-11-28Degree:MasterType:Thesis
Country:ChinaCandidate:G M GaoFull Text:PDF
GTID:2348330512484811Subject:Engineering
Abstract/Summary:PDF Full Text Request
Spiking Neural Networks(SNNs)are known as “the third generation neural network”,are the latest achievements of neural network.Compared with the first two neural networks generations,SNNs simulate the neuron more close to our reality,and they have more powerful computing ability,hence,they are widely used in fields of machine learning,image processing and visual information processing.SNNs have strong nonlinear processing ability.In addition,their encoding and synaptic learning mechanism have distinct differences with first two neural networks generations.Therefore,it is important to deeply research SNNs because they are more close to biological neuron.Recently,pattern recognition is becoming more and more popular in realms of research and application,and has obtained various excellent results.To improve the performance of neural network in pattern recognition,researching and applying SNNs are extremely significant.The content of this thesis mainly includes introduction of basic theory of neural network,research of Spiking Neural Model and learning algorithm of SNNs.Moreover,this thesis combines SNNs and CNN(Convolutional Neural Network),propose an innovative model based on spiking convolution.In addition,this thesis improves existing learning algorithm,and propose a new algorithm based on Tempotron.The main points of this topic are as follows:1.Briefly introducing the theoretical knowledge about biological neurons and Spiking Neural Networks.Then it also describes Spiking neuron model,SNNs' learning algorithm and encoding methods of neuron.2.Proposing an image edge character encoding schemes based on Spiking convolution.This thesis applies the Convolutional Neural Network and image edge extraction into Spiking timing encoding,and then proposes the encoding schemes,which is used to encode the input information.3.Porposing the R-Tempotron learning algorithm,which improves the disadvantage of Tempotron learning algorithm,applying the noise threshold into training process.It is improving the robustness of Tempotron learning algorithm.4.Proposing a pattern recognition model of Spiking convolution based on RTempotorn utilizing Spiking convolution and R-Tempotorn algorithm.And then we tests it on MNIST dataset.The experiments successfully apply the Spiking convolution to image edge detection,the effect of extracted edge is well,which can effectively distinguish different categories after encoding.Meanwhile,the integration model can correctly identify most of the MNIST test dataset.In noise environment,the R-Tempotron algorithm can have a higher robustness than the Tempotron.This thesis provides a new idea for Spiking neural encoding and the research of pattern recognition based on Spiking Neural Networks.
Keywords/Search Tags:Spiking Neural Networks, Pattern Recognition, Spiking convolution, R-Tempotron
PDF Full Text Request
Related items