Fault Pruning: Robust Training of Neural Networks with Memristive Weights

Neural networks with memristive memory for weights have been proposed as an energy-efficient solution scaling up of neural network implementations. However, training such is still challenging due to various memristor imperfections and faulty elements. Such faults are becoming increasingly severe the density arrays increases in order scale weight memory. We propose fault pruning, a robust scheme based on idea identify behavior fly during prune corresponding connections. test this algorithm simulations using both feed-forward convolutional architectures standard object recognition data sets. show its ability mitigate detrimental effect training.