Research On Competitive Neural Network Based On Memristor

With the AlphaGo sweeping across the four sides in the circle of weiqi,artificial intelligence is gradually rising,and AI technology,such as speech recognition and image recognition,has been used in daily life.As the third generation of artificial neural network,pulse neural network is gradually being studied because of its significant advantages in classification,image recognition and other fields.The memristor is a two-terminal device with the same resistance as a dimension.The memristor has the characteristic of recalling resistance: its resistance is related to the amount of charge that is passed at a given time.Because memristor has the advantages of high speed and low power consumption,and is compatible with CMOS technology,it has wide application prospects in the field of memory and neural network.This paper first introduces the basic characteristics of memristor and the basic concept of pulse neural network,and further explores two kinds of models of impulse neuron: Hodgkin-Huxley(HH)model and Leaky-Integrate-And-Fire(LIF)model.Based on this,this paper designs two LIF neuron models,one based on the board level circuit model and one based on the CMOS process.The pulse rate of LIF neuron circuit is linearly related to the injection current based on the board level circuit.Meanwhile,the neuron circuit also has the characteristics of the refractory period,which can regulate the refractory period by adjusting the bias voltage.In the CMOS model,the DPI module is used as the membrane potential integration and leakage of the neuron.Then the concept of synapses is summarized and the neural network is simplified as an explanation of synapses and neurons.Then a synapse model based on CMOS circuit is designed,and the plasticity of synaptic circuit is verified by simulation.When the pulse signal is different,synapse will show different activity.The basic concept of competitive neural network is introduced,and a competitive neural network based on CMOS circuit is built by using CMOS neuron circuit and CMOS synapse circuit.The competitive neural network is based on pulse neural network,and it is a hardware level model.It is verified by simulation: neurons will compete differently because of the difference between synaptic weights and initial excitation currents.Finally,the designed LIF neuron board level circuit is tested to verify the characteristics of the neuron membrane potential and the pulse rate.A memristor based synapse circuit is designed based on memristor resistance and synaptic power.The synaptic plasticity is verified by simulation and test.A competitive neural network is constructed by using a neuron board level circuit and a memristor based synaptic circuit.The characteristics of neurons in individual competition,group competition and inter group competition are tested,and the correctness of the competitive network model is verified.