QGAN-based data augmentation for hybrid quantum–classical neural networks

Abstract

Hybrid Quantum–Classical Neural Networks (HQNNs) have the potential to achieve faster training convergence and higher accuracy than classical neural networks in complex feature spaces. However, data augmentation for quantum machine learning is underexplored due to challenges in generating quantum data and worsening data scarcity. To address this, we integrate Quantum Generative Adversarial Networks (QGANs) with HQNNs, and propose two strategies: a general approach to improve data processing and classification, and a customized strategy that generates samples based on HQNNs’ performance on specific classes. Simulation experiments on the MNIST dataset demonstrate that QGAN outperforms conventional data augmentation methods(random rotation, translation, contrast adjustment) and classical GANs. Compared to deep convolutional GAN, QGAN achieves similar performance with 50% fewer trainable parameters, balancing efficiency and effectiveness. These results highlight the advantages of quantum data augmentation techniques, offering potential solutions for real-world applications such as rare disease diagnosis and endangered species classification.