Realizing spike-timing dependent plasticity learning rule in Pt/Cu:ZnO/Nb:STO memristors for implementing single spike based denoising autoencoder

Abstract

In this work, the Cu:ZnO based memristors were fabricated and modelled and its biological synaptic characteristics were realized. Phenomenon similar to long-term potentiation and long-term depression were observed in the proposed devices and spike timing dependent plasticity learning rule was established by engineering appropriate input voltage spikes, making the devices suitable for use in artificial neural networks. In order to demonstrate learning mechanism, a denoising autoencoder network was developed by incorporating the synaptic characteristics of the device along with the concept of rank coding. To evaluate the feasibility and performance of the network, images from the MNIST database for handwritten digits were employed. The training of the proposed network was accomplished by incorporating noisy images, and it was validated with images corrupted with Gaussian, Salt & Pepper and Speckle noises. Surprisingly, the obtained results demonstrated that the shape of the digits was recovered to a great extent and almost all noise in the background were removed. The accuracy of the denoising was found to be more than 90% for most cases. The proposed network shows the efficacy of ferroelectric Cu:ZnO memristors as artificial synapses in spiking neural networks, which opens up a new path towards developing future generation biologically compatible neuromorphic systems.