Energy-based Generative Adversarial Network

TL;DR

The paper proposes EBGAN, a novel GAN variant that models the discriminator as an energy function, enabling more flexible architectures and improved training stability, demonstrated with auto-encoder based EBGANs producing high-resolution images.

Contribution

Introduces the EBGAN framework, allowing diverse discriminator architectures and demonstrating stability and high-resolution image generation.

Findings

EBGAN exhibits more stable training than traditional GANs.

Auto-encoder based EBGAN can generate high-resolution images.

Flexible energy-based discriminator architectures are feasible.

Abstract

We introduce the "Energy-based Generative Adversarial Network" model (EBGAN) which views the discriminator as an energy function that attributes low energies to the regions near the data manifold and higher energies to other regions. Similar to the probabilistic GANs, a generator is seen as being trained to produce contrastive samples with minimal energies, while the discriminator is trained to assign high energies to these generated samples. Viewing the discriminator as an energy function allows to use a wide variety of architectures and loss functionals in addition to the usual binary classifier with logistic output. Among them, we show one instantiation of EBGAN framework as using an auto-encoder architecture, with the energy being the reconstruction error, in place of the discriminator. We show that this form of EBGAN exhibits more stable behavior than regular GANs during training. We also show that a single-scale architecture can be trained to generate high-resolution images.