Research papers · Posted 2021.08.16 · Publication / acceptance: author-reported
Towards a Safety Case for Hardware Fault Tolerance in Convolutional Neural Networks Using Activation Range Supervision
Towards a Safety Case for Hardware Fault Tolerance in Convolutional Neural Networks Using Activation Range Supervision
Overview
The study builds a prototype CNN safety case around activation-range supervision for hardware bit flips. It examines nonuniform range restrictions and demonstrates fault detection and mitigation in a ResNet-50 vehicle-classification application.
Publication and version information
Publication, acceptance or venue information is author-reported on arXiv. The exact source wording is preserved below.
Author-supplied source information
Author-supplied information on arXiv.
Proceedings of the Workshop on Artificial Intelligence Safety 2021
8 pages, 7 figures
Version: arXiv v1 — manuscript revision