Fig. 4: Comparative simulation results of hardware fault tolerance schemes on the MNIST handwritten digit classification task.
From: Layer ensemble averaging for fault tolerance in memristive neural networks
a Overview of the MNIST dataset and employed architecture of the 2-layer perceptron network used for image classification. b Network test accuracies and mapping errors averaged across network layers at increasing levels of device stuck-at faults and the redundancy parameter α for each fault tolerance scheme (see Supplementary Table 1 and corresponding discussion for details). Bar heights correspond to averages and error bars correspond to standard deviations across 10 independent cycles of the entire simulation process. The proposed layer ensemble averaging scheme boasts a higher degree of fault tolerance compared to other hardware fault tolerance schemes, evident by lower mapping errors and higher test accuracies at similar levels of stuck-at faults and redundancy.
