OrganiQ: Mitigating Classical Resource Bottlenecks of Quantum Generative Adversarial Networks on NISQ-Era Machines

TL;DR

OrganiQ is introduced, the first quantum GAN capable of producing high-quality images without using classical neural networks, and it is claimed to be the first quantum GAN capable of producing high-quality images without using classical neural networks.

Abstract

Driven by swift progress in hardware capabilities, quantum machine learning has emerged as a research area of interest. Recently, quantum image generation has produced promising results. However, prior quantum image generation techniques rely on classical neural networks, limiting their quantum potential and image quality. To overcome this, we introduce OrganiQ, the first quantum GAN capable of producing high-quality images without using classical neural networks.

Figures (14)

• Figure 1: Image quality on the MNIST digit 3 after 50 iterations of training using the QCPatch technique. Higher ratio indicates classical discriminator needs to have more parameters than quantum generator to match quantum generator’s computing capability.
• Figure 2: The Baseline is a GAN architecture ported onto quantum machines. The quantum generator produces an image that is measured classically and tested on the quantum discriminator with amplitude embedding. The discriminator also receives real (non-generated) images separately, which it learns to differentiate from the generated images.
• Figure 3: This figure demonstrates the three different passes as part of OrganiQ's training procedure. In A, the discriminator is trained on real images from the initial dataset, where amplitude embedding places the classical data on the quantum circuit. This is followed by a randomized unitary matrix which introduces phases to the classical image. Then, the discriminator is trained to update the tunable RX gates through gradient descent. In B, the discriminator is trained on synthetic data by placing the generator on the discriminator circuit. The generator weights are greyed out to indicate the gradients do not optimize the generator. In C, the generator is trained, where the discriminator's weights are greyed out to indicate that only the generator's weights are optimized in this step.
• Figure 4: Inference procedure for OrganiQ. Noise is sent through the quantum generator, then scaled to the domain of PCA components, before PCA inverse is performed using the eigenvalues of the original dataset.
• Figure 5: MNIST images generated by the Baseline, QCPatch, and OrganiQ. OrganiQ generates more easily recognizable digits with less blur than other methods.