Preliminary Experiments on Fault-Tolerance of a Small Convolutional Neural Network

Fault-tolerance is an important property of safety critical systems, and there exist many established fault-tolerant technologies for conventional computing devices. As preliminary for fault-tolerant design of complex neural networks, this paper presents some simulation results of fault-tolerance of a small convolutional neural network for digit classification, and we show relations between error probability of synaptic weights and classification accuracy.