Approximated Triple Modular Redundancy of Convolutional Neural Networks Based on Residual Quantization

Triple modular redundancy (TMR) is a fault-tolerant system that detects faults and allows processing to continue by employing three identical modules. However, TMR introduces challenges such as tripled circuit area and increased power consumption. In this paper, we propose an approximated TMR (ATMR), which consists of the original module and two smaller modules achieved through quantization, specifically tailored for convolutional neural networks (CNNs). For evaluation, we conduct experiments using ResNet-20 to simulate fault recovery through linear and residual quantization when the original module fails. The results demonstrate that ATMR maintains high inference accuracy across many layers, even when employing low-bit linear or residual quantization for recovery.