Analyzing the Impact of Soft Errors in Deep Neural Networks on GPUsfrom Instruction Level

Abstract

­ Deep Neural Networks (DNNs) used in safety­critical systems cannot compromise their performance due to reliability issues. In particular, soft errors are the worst. Selective software­based protection solutions are among the best techniques to improve the reliability of DNNs efficiently. However, their most significant challenge is precisely hardening portions of the DNN model to avoid performance degradation. In this work, we propose a comprehensive methodology to analyze the reliability of object detection and classification algorithms run on GPUs from the lowest (instruction) evaluation level. The ultimate goal is to avoid the performance penalty of full instruction duplication by confidently identifying the vulnerable instructions. For this purpose, we propose a technique, Instruction Vulnerability Factor (IVF). By applying our methodology on ResNet and YOLO models, we demonstrate that both models’ most vulnerable instructions can be precisely determined. Moreover, we show that YOLO is more sensitive to the changes caused by soft errors than ResNet. Also, ResNet depends on the input image in its reliability, while YOLO tends to be independent.