Robust Fault-Tolerant Design Based on Checksum and On-Line Testing for Memristor Neural Network

Abstract

The matrix-vector product is the most essential operation in the weight calculation of deep learning, and greatly impacts the calculation speed and power consumption of neural network circuits. A memristor is one of the most promising components used to efficiently develop matrix-vector products. However, it has been pointed out that memristors have a severely low write endurance limitation and large variation during operation owing to their manufacturing immaturity. While an algorithm-based fault tolerance method has thus far been proposed to enhance the reliability by applying checksum function and online testing, the effectiveness of such the function remains limited because it can apply only the forward propagation and multiple hard faults cannot be repaired. This paper proposes an extension of the conventional method to achieve a more robust fault-tolerant method for memristor-based neural network circuits. Numerical experiments using the Hopfield network and three-layered neural network demonstrate that the proposed method achieves 5.25% and 1.88% higher classification accuracies compared with a conventional fault-tolerant method, respectively.