Soft Error Resilience of Deep Residual Networks for Object Recognition

Abstract

Convolutional Neural Networks (CNNs) have truly gained attention in object recognition and object classification in particular. When being implemented on Graphics Processing Units (GPUs), deeper networks are more accurate than shallow ones. Residual Networks (ResNets) are one of the deepest CNN architectures used in various fields including safety-critical ones. GPUs have proven to be the major accelerator for CNN models. However, modern GPUs are prone to radiation-induced soft errors, which is a serious issue in safety-compliant systems. In this work, we analyze and propose an approach to address the reliability of ResNet on GPUs. We firstly analyze three popular ResNet models, explicitly, ResNet-50, ResNet-101, and ResNet-152 through NVIDIA's fault injector, SASSIFI. We perform an in-depth analysis of the model from the perspective of layer and kernel vulnerability. Then, we experimentally show the vulnerability of ResNet models and identify the most vulnerable portions. Finally, we validate our solution, which is a selective-hardening technique, through hardening the worth-hardening kernels to avoid unnecessary overheads. Our strategy is demonstrated to mask up to 93.38% of the injected errors with performance overhead less than 5.35%. Furthermore, the percentage of the errors causing misclassifications can be reduced from 4.2% to 0.104%, thereby significantly improving the model's reliability.