Nb₂O₅ memristive neurons-based unsupervised learning network

Memristors exhibit potential applications in neuromorphic computing, because of their nanoscale and low power. Non-volatile passive memristors usually behave as electronic synapses, while volatile locally active memristors can be used to construct artificial neurons. In this paper, we apply an Nb2O5 locally active memristor with the parasitic capacitance as a LIF neuron and analyze the possibility of generating spiking oscillations by the neuron through small signal equivalent circuits and the Hopf bifurcation method. By combining Nb2O5 memristive neurons with the voltage-controlled non-volatile memristive synapses, we construct an unsupervised learning network and classify 5 x 3 letter images and 5 x 5 number images. In particular, before building the hardware circuit, we predict the training time, recognition time, and recognition accuracy of the pattern recognition network through theoretical analysis, which guides the actual circuit experiment. Specifically, the training time of the network is related to the synaptic memristor resistance change rate, the recognition time of the network is related to the oscillation period of the Nb2O5 memristor, and whether the network can work properly is related to the parameters of Nb2O5 memristor and NMOS. The LTspice simulation results manifest that the proposed circuit can recognize different patterns and can be applied to the neural morphological system of pattern recognition.