An efficient structure to improve the reliability of deep neural networks on ARMs

Abstract

As Deep Neural Networks (DNNs) become more pervasive in safety-critical embedded systems, improving the soft error resilience of DNNs will grow increasingly important. This paper proposes a Distribution-based Error Detector (DED) to improve DNN's reliability. We compare the proposed approach with the regularization method and the typical Symptom-based Error Detector (SED). From the perspective of the bit error resilience, DED provides the highest fault coverage. Our results show that DED DNNs' Silent Data Corruption rates are less than 0.02 even if error bit rates are up to 1. Further, regarding Architecture Vulnerability Factor (AVF) results, we observe that the regularization method and the SED cannot improve the error resilience of register files for quantized DNNs. On the contrary, DED can reduce the SDC AVF by order of magnitude. In addition, DED can increase Mean Work To Failure (MWTF) by more than 19×, while the regularization method and the SED only increase MWTF by less than 2×.