Аннотация:Inspired by the architecture and principle of the human brain, neuromorphic computing has attracted enormous research interest due to its potential for massively parallel and energy-efficient computing, where nanoscale memristors are considered as perfect building blocks for hardware neural networks, serving as compact, analog synapses. However, the inherent variation in memristors has been regarded as a major obstacle to their practical application in neuromorphic computing. Here, for the first time, we demonstrate that this long-standing issue can be addressed by introducing fuzziness into the neural networks. We found that the cycle-to-cycle and device-to-device conductance variations in the on and off states of Pt/TaOx/Ta memristors statistically follow Gaussian distributions, and using an experimentally verified compact synapse model based on the electrical characteristics of Pt/TaOx/Ta devices, a fuzzy restricted Boltzmann machine (FRBM) network was constructed where all the weight states were fuzzified to accommodate device stochasticity. The FRBM network has shown significantly improved tolerance to device variation, as confirmed by increased accuracy in the benchmark test of MNIST handwritten digit classifications. This study thus provides a new route towards highly robust neuromorphic computing, even if the computing elements can be stochastic and inhomogeneous.
