Research On The Application Of Memristor-based Spiking Neural Network In Image Recognition

Spiking neural network,also known as the third-generation artificial neural network,is one of the most advanced neural morphological networks at present.This neural network uses biologically realistic neuron models for information encoding and computation to take full advantage of the efficiency of the neural network.Spiking neural network is a highly brain-like and highly biological computing network,which is usually trained by the biological plausibility algorithm.Spiking neurons interact with each other through sparse pulse sequences that propagate across the synapses.And spiking neural network transmits information in the form of binarization and does not need to transmit analog values between neurons.Therefore,spiking neural network can be well combined with digital circuit design.In the thesis,in which the memristor is combined with the spiking neural network,mainly carried out the following two aspects of work: in the terms of neuron model,the traditional Leaky Integrated and Fire(LIF)model is improved to simulate the spikes that more closely matches the true biological form.In the terms of network training algorithm,the Spiking Timing Dependent Plasticity(STDP)unsupervised learning rules is used to complete the entire network construction and network synaptic training on the pure digital platform,and combined with the biological mechanism of lateral inhibition.The memristor-based spiking neural network designed in the thesis has completed the functional verification,which can realize spikes sending and training memristor weight update.Clock constraint is added to verify the correctness of the timing analysis of the design.The RTL code has passed the logic synthesis and downloaded to the FPGA,and finally completed the prototype verification,which proves the function of the memristorbased spiking neural network is correct.Assisted with the software environment,the task of black and white handwritten digital image recognition is completed,and the classification results are analyzed.Unlike the traditional deep learning neural networks,the structure of the spiking neural network designed in the thesis is more simple,which only adds input and output layer,and it does not need to learn from convolutional neural networks to build,thus it leaves out the convolutional layer,pooling layer and full connection layer.The addition of STDP unsupervised learning algorithm also makes the network more biologically plausible,and the network can have a fast speed while reducing the design complexity.