Title: Nonuniform magnetic domain-wall synapses enabled by population coding

Abstract: Traditional artificial intelligence implemented in software is usually executed on accurate digital computers. Nevertheless, the nanoscale devices for the implementation of neuromorphic computing may not be ideally identical, and the performance is reduced by nonuniform devices. In biological brains, information is usually encoded by a cluster of neurons such that the variability of nerve cells does not influence the accuracy of human cognition and movement. Here, we introduce the population encoding strategy in neuromorphic computing and demonstrate that this strategy can overcome the problems caused by nonuniform devices. Using magnetic memristor device based on current-induced domain-wall motion as an example, we show that imperfect storage devices can be applied in a hardware network to perform principal component analysis (PCA), and the accuracy of unsupervised classification is comparable to that of conventional PCA using ideally accurate synaptic weights. Our results pave the way for hardware implementation of neuromorphic computing and lower the criteria for the uniformity of nanoscale devices.