Energy-based Generative Adversarial Network

We introduce the "Energy-based Generative Adversarial Network" (EBGAN) model which views the discriminator in GAN framework as an energy function that associates low energies with the regions near the data manifold and higher energies everywhere else. Similar to the probabilistic GANs, a generator is trained to produce contrastive samples with minimal energies, while the energy function is trained to assign high energies to those 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 discriminant network. Among them, an instantiation of EBGANs is to use an auto-encoder architecture alongside the energy being the reconstruction error. 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.