Coulomb GANs: Provably Optimal Nash Equilibria via Potential Fields

Thomas Unterthiner, Bernhard Nessler, Günter Klambauer, Martin Heusel, Hubert Ramsauer, Sepp Hochreiter

Abstract

Generative adversarial networks (GANs) evolved into one of the most successful unsupervised techniques for generating realistic images. Even though it has recently been shown that GAN training converges, GAN models often end up in local Nash equilibria that are associated with mode collapse or otherwise fail to model the target distribution. We introduce Coulomb GANs, which pose the GAN learning problem as a potential field of charged particles, where generated samples are attracted to training set samples but repel each other. The discriminator learns a potential field while the generator decreases the energy by moving its samples along the vector (force) field determined by the gradient of the potential field. Through decreasing the energy, the GAN model learns to generate samples according to the whole target distribution and does not only cover some of its modes. We prove that Coulomb GANs possess only one Nash equilibrium which is optimal in the sense that the model distribution equals the target distribution. We show the efficacy of Coulomb GANs on a variety of image datasets. On LSUN and celebA, Coulomb GANs set a new state of the art and produce a previously unseen variety of different samples.

Introduction

Generative Adversarial Networks (GANs) have revolutionized the field of unsupervised learning, yet they remain notoriously difficult to train. Two primary obstacles hinder their performance: local Nash equilibria and mode collapse. Standard GAN discriminators act as binary classifiers, providing only local signals about whether a specific point is real or fake. Consequently, the generator often fails to learn the full diversity of the training set, instead "collapsing" into a subset of the data that successfully fools the discriminator. Coulomb GANs address this by redefining the adversarial game through the lens of classical physics, ensuring the generator receives a global signal that guides it toward the true data distribution.

The Coulomb GAN Framework

Coulomb GANs treat data samples as electrically charged particles interacting within a potential field. In this analogy, training samples act as "sinks" (attractive charges), while generated samples act as "sources" (repulsive charges).

Instead of a binary classifier, the discriminator in a Coulomb GAN learns to approximate a potential field $\Phi(a)$. This field maps the "energy" of the entire space based on the positions of all real and generated samples. The generator's objective is to minimize the total energy of the system by moving its samples along the resulting "force field"—the gradient of the potential—toward the training samples while simultaneously being pushed away from other generated samples to maintain diversity.

Visualization of the potential field and force vectors guiding the generator. Red dots represent training samples (sinks) and blue dots represent generated samples (sources). The streamlines indicate the force field that guides the generator.

A critical component of this framework is the use of a Plummer kernel. Unlike the Gaussian kernels often used in Maximum Mean Discrepancy (MMD) approaches, the Plummer kernel is mathematically proven to lead to a single, global Nash equilibrium where the model distribution exactly matches the target distribution. This "global awareness" ensures that even if a generated sample is far from a training mode, it still "feels" the pull toward it, preventing the model from ignoring parts of the data.

Experimental Results

The efficacy of Coulomb GANs was evaluated across several benchmarks, ranging from toy datasets to high-dimensional image sets:

  • Toy Datasets: In a 2D mixture of 25 Gaussians, standard GANs and WGANs frequently suffered from mode collapse. In contrast, Coulomb GANs captured all 25 modes perfectly, demonstrating their resistance to the "one-trick pony" problem.
  • Image Generation: On the CelebA dataset, Coulomb GANs achieved a state-of-the-art Fréchet Inception Distance (FID) of 9.3, significantly outperforming competitors like BEGAN (28.5) and DCGAN (12.5). The model demonstrated a support size of approximately 1 million faces, indicating a vast and continuous learned distribution rather than simple memorization.
  • Text Generation: The model also proved flexible enough for discrete data, performing on par with WGAN-GP on the Google Billion Word dataset.

Discussion and Limitations

The primary strength of Coulomb GANs lies in their provable optimality. By ensuring only one Nash equilibrium exists, the researchers provide a theoretical guarantee that convergence leads to the true data distribution. However, practical implementation presents challenges. Approximating a complex, non-smooth potential field in high dimensions requires significant neural network capacity. If the discriminator is insufficiently powerful, the model may produce "non-sensical interpolations"—blurry artifacts in regions where the potential field is poorly mapped. Additionally, the generator must move slowly to allow the discriminator to accurately track the field, which can result in longer training times compared to standard architectures.

Despite these challenges, Coulomb GANs represent a significant step forward in stabilizing GAN training and ensuring global coverage of complex data distributions.

Summary updated

Abstract

Generative adversarial networks (GANs) evolved into one of the most successful unsupervised techniques for generating realistic images. Even though it has recently been shown that GAN training converges, GAN models often end up in local Nash equilibria that are associated with mode collapse or otherwise fail to model the target distribution. We introduce Coulomb GANs, which pose the GAN learning problem as a potential field of charged particles, where generated samples are attracted to training set samples but repel each other. The discriminator learns a potential field while the generator decreases the energy by moving its samples along the vector (force) field determined by the gradient of the potential field. Through decreasing the energy, the GAN model learns to generate samples according to the whole target distribution and does not only cover some of its modes. We prove that Coulomb GANs possess only one Nash equilibrium which is optimal in the sense that the model distribution equals the target distribution. We show the efficacy of Coulomb GANs on a variety of image datasets. On LSUN and celebA, Coulomb GANs set a new state of the art and produce a previously unseen variety of different samples.
BibTeX
@inproceedings{unterthiner2018coulomb,
  author = {Thomas Unterthiner and Bernhard Nessler and Günter Klambauer and Martin Heusel and Hubert Ramsauer and Sepp Hochreiter},
  title = {{Coulomb GANs: Provably Optimal Nash Equilibria via Potential Fields}},
  booktitle = {International Conference on Learning Representations},
  year = {2018},
  url = {https://openreview.net/forum?id=SkVqXOxCb},
}
Figure