Mutual information maximizing quantum generative adversarial networks

TL;DR

Numerical simulations demonstrate that InfoQGAN effectively mitigates mode collapse while achieving robust feature disentanglement in the quantum generator, highlighting the potential of InfoQGAN as a foundational approach for advancing quantum generative modeling in the NISQ era.

Abstract

One of the most promising applications in the era of Noisy Intermediate-Scale Quantum (NISQ) computing is quantum generative adversarial networks (QGANs), which offer significant quantum advantages over classical machine learning in various domains. However, QGANs suffer from mode collapse and lack explicit control over the features of generated outputs. To overcome these limitations, we propose InfoQGAN, a novel quantum-classical hybrid generative adversarial network that integrates the principles of InfoGAN with a QGAN architecture. Our approach employs a variational quantum circuit for data generation, a classical discriminator, and a Mutual Information Neural Estimator (MINE) to explicitly optimize the mutual information between latent codes and generated samples. Numerical simulations on synthetic 2D distributions and Iris dataset augmentation demonstrate that InfoQGAN effectively mitigates mode collapse while achieving robust feature disentanglement in the quantum generator. By leveraging these advantages, InfoQGAN not only enhances training stability but also improves data augmentation performance through controlled feature generation. These results highlight the potential of InfoQGAN as a foundational approach for advancing quantum generative modeling in the NISQ era.

Figures (7)

• Figure 1: The basic structure of a quantum generative adversarial networks (QGAN).
• Figure 2: Generator ansatz for a 4-qubit quantum circuit with $l$ layers.
• Figure 3: Comparison of latent code disentanglement. Samples were generated from the epoch with the highest $p$-value during training. InfoQGAN and InfoGAN exhibit superior disentanglement capabilities, as evidenced by the clear color separation of latent codes in the generated samples. InfoQGAN also demonstrates feature disentanglement in a noisy environment; However, due to the noise-induced variance reduction, it produces a smaller-shaped distribution.
• Figure 4: Comparison between the target distributions and those generated by QGAN and InfoQGAN. This experiment demonstrates how feature disentanglement can be utilized. The same can be done with InfoGAN, but not with QGAN and GAN. Figure \ref{['main:fig:2d_custom_InfoQGAN_Q']} shows an example where InfoQGAN mapped the target distribution to code space when creating the heart.
• Figure 5: Comparison of t-SNE visualizations. The legend labels 1, 2, and 3 represent data generated by the generator, corresponding to $c_1$ values of -1.0, 0, and 1.0, respectively.