Pruning of Deep Neural Networks for Fault-Tolerant Memristor-based Accelerators
Hosted in Virtual Platform
Manufacturing Test and Reliability
DescriptionCritical hardware faults (CFs) lead to misclassification of deep neural networks (DNNs) mapped to memristor crossbars. We present a machine learning technique to identify these CFs with over 98% accuracy and is 20x faster than a baseline using random fault injection. Next, we present a fault-tolerance technique that iteratively prunes a DNN by targeting weights mapped to CFs in the crossbars. Our results show that the proposed pruning technique eliminates up to 95\% of the CFs with negligible DNN accuracy loss, leading to a 99% savings in the hardware redundancy required for fault tolerance.