Hybrid Quantum-Classical Normalizing Flow

TL;DR

This paper introduces a hybrid quantum-classical normalizing flow model that leverages parameterized quantum circuits for image generation, demonstrating improved efficiency and quality over existing quantum and classical generative models.

Contribution

The paper proposes a novel hybrid quantum-classical normalizing flow model using parameterized quantum circuits, tailored for NISQ devices, and demonstrates its effectiveness in image generation tasks.

Findings

Achieves lower FID scores than quantum GANs.

Generates high-quality images with fewer parameters than classical models.

Proves the advantage of hybrid quantum-classical approach in generative tasks.

Abstract

With the rapid development of quantum computing technology, we have entered the era of noisy intermediate-scale quantum (NISQ) computers. Therefore, designing quantum algorithms that adapt to the hardware conditions of current NISQ devices and can preliminarily solve some practical problems has become the focus of researchers. In this paper, we focus on quantum generative models in the field of quantum machine learning, and propose a hybrid quantum-classical normalizing flow (HQCNF) model based on parameterized quantum circuits. Based on the ideas of classical normalizing flow models and the characteristics of parameterized quantum circuits, we cleverly design the form of the ansatz and the hybrid method of quantum and classical computing, and derive the form of the loss function in the case that quantum computing is involved. We test our model on the image generation problem. Experimental results show that our model is capable of generating images of good quality. Compared with other quantum generative models, such as quantum generative adversarial networks (QGAN), our model achieves lower (better) Fr\'echet inception distance (FID) score, and compared with classical generative models, we can complete the image generation task with significantly fewer parameters. These results prove the advantage of our proposed model.