A Hybrid Quantum-Classical Generative Adversarial Network for Near-Term Quantum Processors

TL;DR

This paper introduces a hybrid quantum-classical GAN designed for near-term quantum processors, utilizing quantum neural networks for both generator and discriminator, optimized for circuit depth and trained with classical algorithms.

Contribution

It presents a modular design for quantum neural networks in GANs, enabling control over circuit complexity and integrating gradient computation without auxiliary qubits.

Findings

Achieved KL divergence score of 0.39

Achieved JS divergence score of 0.52

Implemented on a two-qubit system with 63 gates

Abstract

In this article, we present a hybrid quantum-classical generative adversarial network (GAN) for near-term quantum processors. The hybrid GAN comprises a generator and a discriminator quantum neural network (QNN). The generator network is realized using an angle encoding quantum circuit and a variational quantum ansatz. The discriminator network is realized using multi-stage trainable encoding quantum circuits. A modular design approach is proposed for the QNNs which enables control on their depth to compromise between accuracy and circuit complexity. Gradient of the loss functions for the generator and discriminator networks are derived using the same quantum circuits used for their implementation. This prevents the need for extra quantum circuits or auxiliary qubits. The quantum simulations are performed using the IBM Qiskit open-source software development kit (SDK), while the training of the hybrid quantum-classical GAN is conducted using the mini-batch stochastic gradient descent (SGD) optimization on a classic computer. The hybrid quantum-classical GAN is implemented using a two-qubit system with different discriminator network structures. The hybrid GAN realized using a five-stage discriminator network, comprises 63 quantum gates and 31 trainable parameters, and achieves the Kullback-Leibler (KL) and the Jensen-Shannon (JS) divergence scores of 0.39 and 0.52, respectively, for similarity between the real and generated data distributions.