Introducing EvolGAN: An Evolutionary Method To Generate Higher Quality Images From Small or Difficult Datasets

Generative Adversarial Networks (GANs) are a model architecture that automatically discovers and learn the regularities or patterns in input data. GANs use these learned patterns to generate new examples that could have been drawn from the original dataset.

Generative Adversarial Networks are the SOTA generative models in many domains, mostly like image synthesis and translation tasks. To reach up to a mark performance, GAN models require a massive amount of training data. To make GANs more effective and dedicated efforts are to be made, either small, complicated, or multimodal datasets are available.

Facebook’s AI Researchers group from the University of the Littoral Opal Coast, the University of Konstanz, and the University of Grenoble have proposed Evolutionary Generative Adversarial Networks (EvolGAN).

Using a quality estimator and evolutionary optimization methods, the novel model searches 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. EvolGAN performs its evolutionary optimization without modifying the training phase, unlike the previous versions.

Researchers say that their approach is generic, simple, easy to implement, and fast. A quality estimator for the outputs of the Generative Adversarial Networks can be used as a drop-in replacement for classical GAN.

Three different Generative Adversarial Networks models offered by the application of EvoIGAN are:

  1. PokeGAN – mountains and Pokemons;
  2. StyleGAN2 – faces, cats, horses, and artwork
  3. PGAN from Pytorch GAN Zoo – FashionGen. In experiments, the proposed methods generated significantly higher quality images while the original generator’s diversity was preserved.

The human evaluators preferred the image generated by EvolGAN with a probability of 83.7 % for the Cats, 74 % for the FashionGen, 70.4 % for the Horses, and 69.2 % for the Artworks.

Researchers say the EvolGAN approach applies to any quality scorer and Generative Adversarial Networks generators.



Consultant Intern: He is Currently pursuing his Third year of B.Tech in Mechanical field from Indian Institute of Technology(IIT), Goa. He is motivated by his vision to bring remarkable changes in the society by his knowledge and experience. Being a ML enthusiast with keen interest in Robotics, he tries to be up to date with the latest advancements in Artificial Intelligence and deep learning.