Abstract: Resistive random access memory (RRAM) exhibits enormous potential in accelerating neural network computations due to its characteristics such as non-volatility and low latency. It can efficiently implement vector-matrix multiplication operations while avoiding massive data transmission. However, stuck-at faults (SAFs) can lead to a significant degradation in the inference accuracy of RRAM-based neural networks. This paper proposes a fault-tolerant scheme for SAFs, which includes methods such as weight mapping adjustment, weight range modification, and loss function regularization, aiming to minimize the weight deviations introduced by SAFs. Comprehensive evaluations through applying image recognition tasks on different neural networks show that the proposed fault-tolerant scheme can effectively recover the accuracy loss caused by SAFs. Even under the condition of 10% SAFs, the average accuracy loss does not exceed 1.5%.

Key words: memristor, neural network, stuck-at fault, fault-tolerant computing