Generative adversarial network based on chaotic time series

Makoto Naruse, Takashi Matsubara, Nicolas Chauvet, Kazutaka Kanno, Tianyu Yang, Atsushi Uchida

Abstract

While Generative Adversarial Networks (GANs) typically rely on software-based pseudorandom number generators (PRNGs) for latent space sampling, this study explores the integration of physical chaos from semiconductor lasers. By utilizing chaotic time series as the source of randomness, the researchers demonstrate that the physical properties of the chaos—specifically its temporal and spatial correlations—are inherited by the generated images. This approach enhances the robustness of the generative model without compromising output diversity, offering a novel intersection between nonlinear physical dynamics and deep learning.

Introduction

In the architecture of Generative Adversarial Networks (GANs), the latent space serves as the foundational "seed" from which synthetic data is grown. Traditionally, this space is populated using software-based pseudorandom number generators (PRNGs), such as the Mersenne Twister. However, these mathematical algorithms lack the complex, multi-scale structural properties found in physical systems.

A recent study by Makoto Naruse and colleagues investigates the replacement of standard PRNGs with chaotic signals generated by a semiconductor laser. The research asks whether the raw, physical unpredictability of a chaotic system can be reflected or transformed into generated data, potentially improving the quality and robustness of the generative process.

The Role of Chaos in Latent Variables

Chaos exists in a unique state between total randomness and deterministic order. The authors hypothesize that the internal structures of a chaotic system—its nonlinear dynamics—can be mapped onto the manifold of a GAN. The study focuses on two primary, often competing, performance metrics:

  1. Similarity in Proximity (Robustness): The degree to which small perturbations in the latent vector result in consistent, logical changes in the output image (e.g., a small tweak to the noise results in a slight, logical change in a face's expression).
  2. Diversity: The breadth of the distribution of generated images, ensuring the model can produce a wide variety of outputs without suffering from mode collapse.

Experimental Methodology

The researchers employed a Deep Convolutional GAN (DCGAN) architecture trained on the CelebA dataset, a large-scale face attributes dataset containing over 200,000 images. The images were resized to $64 \times 64$ pixels for the training phase.

The core innovation lies in the construction of the latent variable vector $Z$. The source of randomness was a semiconductor laser with delayed optical feedback, producing chaotic oscillations. These signals were sampled at 100 GSample/s, and the resulting 8-bit values were concatenated to form 32-bit floating-point numbers, normalized between -1 and 1.

The study compared several experimental configurations:

  • Time-domain chaos: Sequences where temporal correlations from the laser are preserved across successive samples.
  • Space-domain chaos: Sequences where correlations are embedded within a single latent vector.
  • Baselines: Standard uniform (RAND) and normal (RANDN) distributions generated by software.

Inheritance of Physical Signatures

A significant discovery of this research is that the "signature" of the physical laser persists in the digital output. The laser used in the experiment exhibits a characteristic negative autocorrelation at a 50 ps time lag. When the GAN was fed chaos sampled at this specific interval, the similarity in proximity of the generated faces also peaked. This suggests a direct mapping from physical nonlinear dynamics to the geometric properties of the generative model's manifold.

Furthermore, the "peaky" distribution of chaotic signals—which often cluster near the edges of the distribution (-1 and 1)—proved more effective for latent space exploration than a flat uniform distribution. This resulted in superior image retrieval performance, characterized by a higher average minimum Hamming distance between latent variables (~10 for chaos vs. ~3 for standard RAND).

Balancing Robustness and Diversity

Typically, increasing a model's robustness (ensuring neighbors look similar) leads to a reduction in overall diversity. However, the chaotic time series appears to mitigate this trade-off.

  • Space-domain chaos achieved the highest similarity in proximity compared to all other methods, including standard PRNGs.
  • Time-domain chaos provided the most effective balance, ranking second in robustness while maintaining high diversity (nearly as high as standard random numbers).

Conclusion and Future Work

This research demonstrates that the source of randomness in GANs is not merely a neutral utility but a structural component that can be optimized. By leveraging the physical properties of laser chaos, the authors have shown a path toward more robust and efficient generative models.

While this study utilized the DCGAN architecture, future work remains to test these chaotic sources on more advanced models like StyleGAN or Diffusion Models. Additionally, the persistence of the chaotic signature in the output images suggests potential applications in GAN security, such as watermarking generated content or identifying the provenance of AI-generated media. The ultimate goal remains the direct integration of photonic chaotic hardware into the deep learning signal processing pipeline.

Summary updated

Abstract

Generative adversarial networks (GANs) are becoming increasingly important in the artificial construction of natural images and related functionalities, wherein two types of networks called generators and discriminators evolve through adversarial mechanisms. Using deep convolutional neural networks and related techniques, high-resolution and highly realistic scenes, human faces, etc. have been generated. GANs generally require large amounts of genuine training data sets, as well as vast amounts of pseudorandom numbers. In this study, we utilized chaotic time series generated experimentally by semiconductor lasers for the latent variables of a GAN, whereby the inherent nature of chaos could be reflected or transformed into the generated output data. We show that the similarity in proximity, which describes the robustness of the generated images with respect to minute changes in the input latent variables, is enhanced, while the versatility overall is not severely degraded. Furthermore, we demonstrate that the surrogate chaos time series eliminates the signature of the generated images that is originally observed corresponding to the negative autocorrelation inherent in the chaos sequence. We also address the effects of utilizing chaotic time series to retrieve images from the trained generator.
BibTeX
@inproceedings{naruse2019generative,
  author = {Makoto Naruse and Takashi Matsubara and Nicolas Chauvet and Kazutaka Kanno and Tianyu Yang and Atsushi Uchida},
  title = {{Generative adversarial network based on chaotic time series}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-019-49397-2},
}
Figure