Kernel and layer vulnerability factor to evaluate object detection reliability in GPUs

Abstract

Video recognition applications running on Graphics Processing Unit are composed of heterogeneous software portions, such as kernels or layers for neural networks. The authors propose the concepts of kernel vulnerability factor (KVF) and layer vulnerability factor (LVF), which indicate the probability of faults in a kernel or layer to affect the computation. KVF and LVF indicate the high-level portions of code that are more likely, if corrupted, to impact the application's output. KVF and LVF restrict the architecture/program vulnerability factor analysis to specific portions of the algorithm, easing the criticality analysis and the implementation of selective hardening. We apply the proposed metrics to two Histogram of Oriented Gradients (HOG), and You Only Look Once (YOLO) benchmarks. We measure the KVF for HOG by using fault-injection at both the architectural level and high level. We propose for HOG an efficient selective hardening technique able to detect 85% of critical errors with an overhead in performance as low as 11.8%. For YOLO, we study the LVF with architectural-level fault-injection. We qualify the observed corrupted outputs, distinguishing between tolerable and critical errors. Then, we proposed a smart layer duplication that detects more than 90% of errors, with an overhead lower than 60%.

Usability-based Cross-Layer Reliability Evaluation of Image Processing Applications

“…Due to the criticality of the downstream control application that detects the moving cars, the oracle has been defined as in [12] to verify that all identified bounding boxes overlap the one in the golden output by a Jaccard index measure larger than 50%.…”

Approximation-based Fault Tolerance in Image Processing Applications

“…The Oracle classifies the usability by checking the number and the position of the bounding boxes based on the Jaccard index as in [35]. To build T RS, V S and T ES, we employed 1,000 600x300 images downloaded from Italian highway webcams.…”

An Approximation-based Fault Detection Scheme for Image Processing Applications

“…The idea of analysing the usability of the overall result of an image processing application has been considered in recent publications to assess the robustness of machine learning applications for image processing [11], [12]. In particular, SDCs affecting pedestrian detection applications are analyzed and classified as critical/not critical based on the fact that the corrupted output is still usable or not.…”