Efficient Fault-Criticality Analysis for AI Accelerators using a Neural Twin | IEEE Conference Publication | IEEE Xplore