Binary Resistive-Switching-Device-Based Electronic Synapse with Spike-Rate-Dependent Plasticity for Online Learning

Neuromorphic computing is promising to become a future information processing technology with much less energy consumption, higher efficiency, and larger volume. A key element to build an electronic neuromorphic computing system is the synapse, which should be implemented with a simple device that can be easily integrated into large systems because of its vast amount. Here, inspired by a stochastic-forgetting learning concept, we experimentally demonstrate the spike-rate-dependent plasticity (SRDP) synaptic behavior with a one-transistor–one-resistive-switching-device structure and its application in unsupervised online learning. Furthermore, a convolutional neural network-SRDP two-layer network is proposed to recognize the full MNIST data set with 92% accuracy, which is a significant improvement compared with the 56% accuracy achieved by the same scale neural network using a back-propagation algorithm. This work may introduce new possibilities into the implementation of a high-performance artificial neural system for applications with none or very few human-labeled data.