Experimental Quantum Generative Adversarial Networks for Image Generation

Abstract

Quantum machine learning is expected to be one of the first practical\napplications of near-term quantum devices. Pioneer theoretical works suggest\nthat quantum generative adversarial networks (GANs) may exhibit a potential\nexponential advantage over classical GANs, thus attracting widespread\nattention. However, it remains elusive whether quantum GANs implemented on\nnear-term quantum devices can actually solve real-world learning tasks. Here,\nwe devise a flexible quantum GAN scheme to narrow this knowledge gap, which\ncould accomplish image generation with arbitrarily high-dimensional features,\nand could also take advantage of quantum superposition to train multiple\nexamples in parallel. For the first time, we experimentally achieve the\nlearning and generation of real-world hand-written digit images on a\nsuperconducting quantum processor. Moreover, we utilize a gray-scale bar\ndataset to exhibit the competitive performance between quantum GANs and the\nclassical GANs based on multilayer perceptron and convolutional neural network\narchitectures, respectively, benchmarked by the Fr\\'echet Distance score. Our\nwork provides guidance for developing advanced quantum generative models on\nnear-term quantum devices and opens up an avenue for exploring quantum\nadvantages in various GAN-related learning tasks.\n