Analyzing the Reliability of Alternative Convolution Implementations for Deep Learning Applications

Convolution represents the core of Deep Learning (DL) applications, enabling the automatic extraction of features from raw input data. Several implementations of the convolution have been proposed. The impact of these different implementations on the performance of DL applications has been studied. However, no specific reliability-related analysis has been carried out. In this paper, we apply the CLASSES cross-layer reliability analysis methodology for an in-depth study aimed at: i) analyzing and characterizing the effects of Single Event Upsets occurring in Graphics Processing Units while executing the convolution operators; and ii) identifying whether a convolution implementation is more robust than others. The outcomes can then be exploited to tailor better hardening schemes for DL applications to improve reliability and reduce overhead.