NbO_x Local Active Memristor And Its Application In Neural Network

Niobium oxide(NbO_x)memristor is an integral passive and locally-active device.Its quiescent voltage and current characteristic curve includes a negative differential resistance region.With locally-active characteristics and nanoscale advantages,NbO_xmemristors can directly simulate biological neurons at the physical level,and further build neural networks for pattern recognition.At present,in view of the lack of systematic theoretical design methods for the memristor and its application,the research of memristor neural network is still in its infancy.To this end,the basic characteristics of NbO_xlocally-active memristors and their applications in oscillators and neuromorphic calculations are studied,and the main research works are as follows:(1)Based on small signal analysis method,locally-active and chaotic edge theory,the electrical characteristics of NbO_xlocally-active memristor are quantitatively analyzed.The second order oscillator circuit is constructed by connecting the DC bias circuit and the capacitor element,where the DC bias can be provided by the series resistance of the current source or the voltage source.Based on Hopf bifurcation theory,the oscillation condition of the circuit is obtained,that is,the value range of the capacitance.The phase difference between two identical second-order memristor oscillators coupled by resistance and capacitance is quantitatively analyzed by numerical method.The phase difference between the two oscillators varies with the initial condition and the coupling element.Based on the second-order oscillating circuit,a third-order neuron circuit is designed,and the simulation results show that the circuit can simulate seven neuromorphic behaviors.(2)NbO_xlocally-active memristor was used to construct neurons,passive memristor to realize synaptic circuits,and a 25×10 Spiking Neural Network was further constructed.According to the quantitative relationship between the frequency of neuronal spike signal and the intensity of excitation current,combined with Hebbian learning rules,the synaptic weight was calculated.The voltage amplitude of the input black and white pixels is encoded,and the frequency coding(RC)and time coding(TC)methods are used to complete the effective identification of ten different modes of numbers0–9.(3)NbO_xmemristor neurons are used to construct a coupled oscillating neural network combined with memristor memrization capacitance coupling elements.The input pixel is encoded using the phase of the signal,which is reflected in the initial phase of the excitation signal.According to Hebbian learning rules,calculate the coupling strength(resistance and capacitance)between different neurons in the neural network.The influence of different coupling elements on the recognition accuracy and fault tolerance of neural network is compared.The simulation results show that the phase difference of the memristor and capacitance coupled oscillating neural network is stable at 160°,and the phase difference of the pure resistance coupled oscillating neural network is stable at 86°.The proposed coupled oscillatory neural network can realize pattern recognition effectively.