Introduction

A Generative Adversarial Network (GAN) is a generative network architecture, capable of generating new content, such as images and audio, made to look real. They can be used to generate special effects in movies, new data for small datasets and speed up photo editing.

Generative networks existed before GANs but the “adversarial” part actually adds a lot of value and new perspectives. The novelty comes from the fact that in this framework we train two models at the same time: the Generative and the Discriminative ones. The Generative model tries to capture the data distribution, while the Discriminative one estimates a probability that some content was generated by the Generative model, and did not come from the actual data. The Generative model then tries to fool the Discriminative model, by maximising the probability that it makes a mistake.

This competition between these two models is what makes GANs so powerful. It’s like training a hacker at the same time you develop your IT security: the more the hacker gets skilled, the more your IT security develops to protect against atacks, and the more the IT security develops, the better the hacker has to get to overcome it.

Now let’s go through 5 articles you should read in order to better understand GANs and how to use them.

#data-science #data #research #science #machine-learning