Memristive Implementation And Application Of Homeostatic Plasticity Of Neural Network

The memristor is a new-type electronic component,which can vary its resistance according to applied voltage or current signal,and is non-volatile.Since the memristor is resistance changeable,non-volatile,power-efficient and high-density integration friendly,it is very promising in areas like storage,logic computation and neuromorphic computing.Homeostatic plasticity originates from the nervous system,which is an important mechanism to maintain the stability of the nervous system,and has been widely used in the field of computer science.In this thesis,homeostatic plasticity of neural networks is taken as the focus of research.Aiming at the bottleneck problems and current research gaps in its circuit implementation,the following aspects are carried out:The basic concept of memristor is introduced,and three fingerprints of memristor are expounded by taking Hewlett Packard memristor model as an example.The model of VTEAM memristor which has been accepted by researchers is introduced in detail,and the memristor based on AIST material is modeled by VTEAM memristor modeling method.Combining the rationality of physical materials and the discovery of biological memristor,the homeostatic plasticity is induced into the memristor mathematical model,so that the memristor threshold can be adjusted adaptively according to the mechanism of homeostatic plasticity.The general mathematical model is introduced into three kinds of memristor models: bipolar memristor model,titanium dioxide memristor model and AIST memristor model.The corresponding SPICE model is also proposed to facilitate circuit simulation.The basic neuron circuit with homeostatic plasticity is constructed by using titanium dioxide memristor model and CMOS devices.The proposed neuron circuit not only exhibits the basic characteristics of the cumulative excitation discharge,but also can adjust the firing frequency adaptively to keep it within the inherent frequency range of neurons.On this basis,the basic neuron circuit is improved by combining with the memristor model with homeostatic threshold variations,which makes the neuron circuit more compact and less power consumption.In addition,inspired by the prototype of biological neurons,a neuron circuit with double regulation mode for threshold is designed,which improves the basic neuron circuit proposed above again.The neuron circuit is analyzed,and the correctness and effectiveness of the proposed circuit are verified on the PSPICE platform.Based on the above improved neuron circuit,spiking neural network circuit is further constructed.The input neurons,output neurons,the circuit implementation of the weight matrix and the learning rules of the network are designed.The spiking neural network is applied to a pattern recognition task.Since the output neuron circuit in the network is equipped with homeostatic plasticity mechanism,the proposed network can effectively prevent the problems of over-training and under-training in the network.All the simulations are carried on Pspice platform,and the effectiveness of the network is verified.