Memristive neural network circuit with fault tolerance for character recognition

Abstract Memristor-based neural networks are one of the most promising approaches for the hardware implementation of artificial neural networks. In this paper, a memristor-based neural network circuit based on a one-memristor–one-resistor (1M1R) synaptic array structure is designed for character recognition. Compared with other memristive synaptic arrays, the 1M1R structure can reduce the number of memristors used. However, memristors may malfunction due to fabrication defects and the influence of external factors, resulting in a decrease in the accuracy of the circuit’s character recognition, and a suitable solution needs to be found to improve the stability and durability of the circuit. Therefore, in this paper, a fault-tolerant module with feedback adjustment capability is designed in the memristive neural network circuit that can re-adjust the weights of the memristors through in-situ training to solve multiple faults in the memristive neural network. The effect of fault tolerance is verified by character recognition. The experimental results show that the designed memristive neural network circuit can accurately realize character recognition, and the designed fault-tolerant circuit can well tolerate multiple faults, ensuring stable operation of the circuit under fault conditions.