Understanding the Impact of Quantization, Accuracy, and Radiation on the Reliability of Convolutional Neural Networks on FPGAs

Convolutional neural networks are quickly becoming viable solutions for self-driving vehicles, military, and aerospace applications. At the same time, due to their high level of design flexibility, reprogrammable capability, low power consumption, and relatively low cost, the field-programmable gate arrays (FPGAs) are very good candidates to implement the neural networks. Unfortunately, the radiation-induced errors are known to be an issue in static random-access memory (SRAM)-based FPGAs. More specifically, we have seen that particles can change the content of the FPGA's configuration memory, consequently corrupting the implemented circuit and generating the observable errors at the output. Through extensive fault injection, we determine the reliability impact of applying binary quantization to the convolutional layers of neural networks on FPGAs, by analyzing the relationships between model accuracy, resource utilization, performance, error criticality, and radiation cross section. We were able to find that a design with quantized convolutional layers can be 39% less sensitive to radiation, whereas the portion of errors that are considered critical (misclassifications) in the network is increased by 12%. Moreover, we also derive generic equations that consider both accuracy and radiation in order to model the overall failure rate of neural networks.