FKeras: A Sensitivity Analysis Tool for Edge Neural Networks

Abstract

Edge computation often requires robustness to faults, e.g., to reduce the effects of transient errors and to function correctly in high radiation environments. In these cases, the edge device must be designed with fault tolerance as a primary objective. FKeras is a tool that helps design fault-tolerant edge neural networks (NNs) that run entirely on chip to meet strict latency and resource requirements. FKeras provides metrics that give a bit-level ranking of NN weights with respect to their sensitivity to faults. FKeras includes these sensitivity metrics to guide efficient fault injection campaigns to help evaluate the robustness of an NN architecture. We show how to use FKeras in the codesign of edge NNs trained on the high-granularity endcap calorimeter dataset, which represents high energy physics data, as well as the CIFAR-10 dataset. We use FKeras to analyze an NN’s fault tolerance to consider alongside its accuracy, performance, and resource consumption. The results show that the different NN architectures have vastly differing resilience to faults. FKeras can also determine how to protect NN weights best, e.g., by selectively using triple modular redundancy on only the most sensitive weights, which reduces area without affecting accuracy.