Data augmentation experiments with style-based quantum generative adversarial networks on trapped-ion and superconducting-qubit technologies

TL;DR

This paper demonstrates the implementation and effectiveness of style-based quantum GANs for data augmentation on two different quantum hardware platforms, highlighting hardware-specific performance and optimization strategies.

Contribution

It is the first to show successful implementation of style-based qGANs on both superconducting and trapped-ion quantum computers, with analysis of performance and parallelization techniques.

Findings

Comparable quality results on both hardware types

Aria-1 provides more accurate results than IBM Torino

Parallelization reduces runtime and job submissions

Abstract

In the current noisy intermediate scale quantum computing era, and after the significant progress of the quantum hardware we have seen in the past few years, it is of high importance to understand how different quantum algorithms behave on different types of hardware. This includes whether or not they can be implemented at all and, if so, what the quality of the results is. This work quantitatively demonstrates, for the first time, how the quantum generator architecture for the style-based quantum generative adversarial network (qGAN) can not only be implemented but also yield good results on two very different types of hardware for data augmentation: the IBM bm_torino quantum computer based on the Heron chip using superconducting transmon qubits and the aria-1 IonQ quantum computer based on trapped-ion qubits. The style-based qGAN, proposed in 2022, generalizes the state of the art for qGANs and allows for shallow-depth networks. The results obtained on both devices are of comparable quality, with the aria-1 device delivering somewhat more accurate results than the ibm_torino device, while the runtime on ibm_torino is significantly shorter than on aria-1. Parallelization of the circuits, using up to 48 qubits on IBM quantum systems and up to 24 qubits on the IonQ system, is also presented, reducing the number of submitted jobs and allowing for a substantial reduction of the runtime on the quantum processor to generate the total number of samples.