Research on Fault-Tolerant Algorithm for Memristor Neural Networks Based on On-Chip Training

Neural networks are widely used in fields such as image recognition and natural language processing due to their powerful learning ability. Traditional neural network hardware devices suffer from issues such as volatility and high power consumption. In recent years, memristors have become a research hotspot in neural network hardware due to their outstanding characteristics such as scalability, low power consumption, and non volatility. However, due to current process limitations, memristors have Stuck At Fault (SAF) and conductance variation defects, which result in a decrease in the accuracy of the memristor neural network. In this paper, a fault tolerance approach for SAF is proposed using a mask matrix, and fault tolerance for conduction variation is achieved through cosine regularization. The results demonstrate that the proposed fault tolerance algorithms can effectively restore the accuracy of memristor neural networks with defects.