The Model Optimization And Application Of Spiking Neural Network

The idea of creating a computational model by simulating the human brain has been promulgated in the middle of the 20 th century,but was not used in practice before the late 80 s.This model,named the ‘artificial neural network',had maintained popularity only for a few short years before being replaced by simpler artificial intelligence models in the following years.Advances in the neural networks field in the early 2000 s have sparked interest once again,as a plethora of improvements has brought forth richer neuron models and more biologically realistic network designs.Spiking Neural Networks(SNNs)represent a special class of artificial neural networks,in which neuron models communicate by sending spikes(action potentials).SNNs model the dynamics and learning capabilities of the brain in a more biologically inspired way than previous generations of neural networks.However,training these networks is still problematic due to over-training and weight instabilities and issues regarding the synaptic weight dependency on the firing history of a neuron are still unresolved.Based on the existing research results,this paper focuses on the optimization of convergence rate,classification and prediction accuracy,and network stability of SNNs supervised learning algorithm.And the research on emotion recognition of EEG by using SNNs.The main works of the thesis include:1.The research background of this project is introduced firstly.Next the basic knowledge of the traditional artificial neural network and the SNN is described,and the advantages of the SNN are given by comparison.Then the development and basic working principles of SNNs,and the encoding ways of information in SNNs are introduce.The variety of supervised and unsupervised learning algorithms are also listed in the thesis.Finally,the existing problems of SNNs are summarized,and the optimization direction of the algorithm is proposed.2.This thesis presents two methods of using the dynamic momentum and learning rate adaption,to improve learning performance in spiking neural networks where neurons are modelled as spiking multiple times.3.Different methods are employed to enhance the ability of noise tolerance.The comparison of noise SNN and anti-noise SNN by using the various test tasks shows that the anti-noise SNN can commendably improve the noise tolerance.The anti-noise SNN causes less concussion during the training period and obtains reliable output.4.This thesis proposes a method by using the spiking neural network and the electroencephalograph(EEG)processing techniques to recognize emotion states.For the emotion state classification,three methods including variance,fast Fourier transform and discrete wavelet transform are employed to process EEG signals.The SNN is then applied to the processed EEG data for the emotion classification.In the meantime,this work achieves a better emotion classification performance than the benchmarking approaches,which demonstrates the advantages of using SNN for the emotion state classifications.