Abstract— In this article, we design a memristive competitiveneural network circuit based on the winner-take-all (WTA)mechanism and the online Hebbian learning rule. Each synapseof the network contains two memristors whose terminals of signalinputs are opposite. However, only one memristor participates inthe calculation each time, and that one is determined by theoriginal input signal. The competitive neural network circuitincludes two parts: forward calculation and weight update.In this article, the forward calculation part of the circuit isdesigned based on the WTA mechanism. The combination ofthe leaky-integrate-and-fifire (LIF) model and pMOS realizesthe lateral inhibition of neurons. The design of the weightupdating part is based on Hebbian learning rules. In eachcycle, only synaptic memristors connected to the winner outputneuron in forward calculation can be adjusted. The voltage usedfor synaptic memristor adjustment comes from the membranevoltage of the winner output neuron. The whole neural networkcircuit does not need the participation of a central processingunit (CPU) or a fifield-programmable gate array (FPGA) andreally realizes parallel calculation, the saving of area, powerconsumption, and a certain extent computing-in-memory. Basedon the circuit designed in PSPICE, we simulated the classifificationof 5×3 pixel pictures. The changing trend of weights inthe training phase and the high recognition accuracy in therecognition phase prove that the network can learn and recognizedifferent patterns. The competitive neural network can be appliedto the neuromorphic system of visual pattern recognition.

Index Terms— Circuit implementation, competitive neural network, memristor, pattern recognition, unsupervised learning,winner-take-all (WTA).