Generative Adversarial Networks (GANs)

“(GANs), and the variations that are now being proposed is the most interesting idea in the last 10 years in ML, in my opinion.”
Yann LeCun, Director of AI Research, Facebook


  • Neural networks have made great progresses in the fields of image recognition…
  • But tasks where human creativity is needed (e.g. generating a new image) are far from being solved;
  • GAN are making some of these tasks possible.

What is a GAN:

A GAN can be seen as the relation between a forger and an investigator.

  • The forger objective is to create fraudulent imitations of original paintings by famous artists;
  • The investigator (which knows the properties which sets the original artist apart) must identify the fake paintings;
  • Both the forger and the investigator will try to become more and more capable than the other part as the time goes on.

A GAN is composed by two main components :

  • Generator;
  • Discriminator;
  • The basic idea of GANs is to set up a game between Generator and Discriminator.
Basic GAN structure.

GANs Examples

We use gan to generate fashion images, to perform style tranfer and we try to modify their architectures to improve the generation.


MetalGAN: a Cluster-based Adaptive Training for Few-Shot Adversarial Colorization – ICIAP 2019 (under review)

Tomaso Fontanini is the supervisor for this project.