A Compact Memristor-based GAN Architecture with a Case Study on Single Image Super-Resolution

Abstract

The generative adversarial network (GAN) is a kind of unsupervised learning approach with the capacity of dealing with the challenge of limited labeled data in the supervised learning world. However, the limited data bandwidth, as well as the performance gap between processing units and memory of the conventional computing platform becomes a major obstacle in the GAN based applications. In this paper, a memristor-based convolution neural network (CNN) unit synthesized with the spintronic memristor crossbar circuit and a general sigmoid activation function circuit is designed for the implementation of convolutional calculation. Notably, multiple memristor-based CNN units can be utilized for the construction of the generator and discriminator respectively. Based on this, a compact GAN architecture composed of the generator block and discriminator block is presented. For verification, the presented memristor-based GAN is applied to the single image super-resolution (SR). The experimental results demonstrate the validity and effectiveness of the entire scheme.