EvolGAN Boosts Image Quality for Small or Difficult Datasets

Original article was published by Synced on Artificial Intelligence on Medium

EvolGAN Boosts Image Quality for Small or Difficult Datasets

Generative Adversarial Networks (GANs) are a model architecture for automatically discovering and learning the regularities or patterns in input data and using the learned patterns to generate or output new examples that plausibly could have been drawn from the original dataset. GANs are the current SOTA generative models in many domains — most notably image synthesis and translation tasks.

GAN models however require massive amounts of training data to reach decent performance. In an effort to make GANs more effective and reliable when only small, difficult, or multimodal datasets are available, a group of researchers from Facebook AI, University of the Littoral Opal Coast, University of Grenoble and University of Konstanz have proposed Evolutionary Generative Adversarial Networks (EvolGAN).

The novel model uses a quality estimator and evolutionary optimization methods to search the latent space of generative adversarial networks trained on small or difficult datasets. Instead of randomly generating a latent vector z as classical GANS do, EvolGAN performs an evolutionary optimization process, with z as decision variables. And unlike previous methods, EvolGAN performs its evolutionary optimization without modifying the training phase.

General EvolGAN approach (left) and optimization loop (right)

The researchers say their approach is “simple, generic, easy to implement, and fast.” With a quality estimator for the outputs of the GAN, it can be used as a drop-in replacement for classical GAN.

The team presents applications of EvolGAN on three different GAN models: StyleGAN2 for faces, cats, horses and artworks; PokeGAN for mountains and Pokemons; and PGAN from Pytorch GAN Zoo for FashionGen.

In experiments, the proposed method generated “significantly higher quality images while preserving the original generator’s diversity.” The human evaluators preferred EvolGAN-generated images with a probability of 83.7 percent for Cats, 74 percent for FashionGen, 70.4 percent for Horses, and 69.2 percent for Artworks.

The researchers say the EvolGAN approach applies to any quality scorer and GAN generators.

The paper EvolGAN: Evolutionary Generative Adversarial Networks is on arXiv.