<p>We propose a hybrid generative adversarial network (GAN) by combining the classical and quantum approaches with input noise samples drawn from the dependent and nonidentical distribution learned from the training data. Leveraging quantum computing our approach is built on styleGAN for image generation. Present day generative models are still dependent on classical computation method for adversarial learning. They also use independent and identically distributed (IID) noise without considering the data statistics. In this work, we use a hybrid approach by combining quantum computing with the classical way of learning. Unlike StyleGAN, our work is implemented by training intermediate independent generative adversarial networks (GANs) in parallel where we make use of the available training images. The trained parameters initialize the combined architecture to refine the parameters in the final training. A variational quantum circuit (VQC) is embedded in the generators during the training process. The use of parallel setup as well as combining the quantum method with classical significantly reduce the training time. The non-IID samples as input which are generated using the training data statistics enhance the accuracy as well as contribute to training time reduction. The efficacy of the proposed method is tested on Flickr-Faces High-Quality (FFHQ) dataset and Celeba High-Quality (CelebaHQ) dataset. We have conducted experiments with both variants of GANs, namely the classical generator and the hybrid generator, referred to as the proposed classical approach and the proposed quantum approach. A hybrid generator, also referred to as a quantum generator, combines a quantum circuit with a classical GAN. An ablation study is also performed to check performance of VQC by changing the rotation gates and type of entanglements. We compare our results with the classical styleGAN with IID noise as well as other state of the art methods qualitatively as well as using the quantitative metrics such as Fréchet inception distance (FID), density, and coverage. Experiments show that the proposed method reduces the computational complexity as well as generates better quality images. The proposed method of non-IID, hybrid and parallel GAN improves the FID score up to 75<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7495_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>.</p>

Hybrid and parallel GAN architecture with non-IID noise input

期刊论文
发表日期：

2025年6月10日

查看全文

Prashant Gohel,
Manjunath Joshi

摘要

We propose a hybrid generative adversarial network (GAN) by combining the classical and quantum approaches with input noise samples drawn from the dependent and nonidentical distribution learned from the training data. Leveraging quantum computing our approach is built on styleGAN for image generation. Present day generative models are still dependent on classical computation method for adversarial learning. They also use independent and identically distributed (IID) noise without considering the data statistics. In this work, we use a hybrid approach by combining quantum computing with the classical way of learning. Unlike StyleGAN, our work is implemented by training intermediate independent generative adversarial networks (GANs) in parallel where we make use of the available training images. The trained parameters initialize the combined architecture to refine the parameters in the final training. A variational quantum circuit (VQC) is embedded in the generators during the training process. The use of parallel setup as well as combining the quantum method with classical significantly reduce the training time. The non-IID samples as input which are generated using the training data statistics enhance the accuracy as well as contribute to training time reduction. The efficacy of the proposed method is tested on Flickr-Faces High-Quality (FFHQ) dataset and Celeba High-Quality (CelebaHQ) dataset. We have conducted experiments with both variants of GANs, namely the classical generator and the hybrid generator, referred to as the proposed classical approach and the proposed quantum approach. A hybrid generator, also referred to as a quantum generator, combines a quantum circuit with a classical GAN. An ablation study is also performed to check performance of VQC by changing the rotation gates and type of entanglements. We compare our results with the classical styleGAN with IID noise as well as other state of the art methods qualitatively as well as using the quantitative metrics such as Fréchet inception distance (FID), density, and coverage. Experiments show that the proposed method reduces the computational complexity as well as generates better quality images. The proposed method of non-IID, hybrid and parallel GAN improves the FID score up to 75 $\%$ $%$ .