GANs

Generative Adversarial Networks, or GANs, are unsupervised learning algorithms that make it possible to generate artificial data with a high degree of realism.

Here’s how it works. A GAN’s architecture is made up of two neural networks set-up in competition. During the learning process, the first network, called the “generator”, creates a sample of data that resembles the training data as much as possible. The second network, called the “discriminator”, attempts to detect if the sample is real or if it is the work of its “opponent”. The first network is therefore trying to trick the second.

Just like the forger and the detective, this process enables both networks to improve, each in their respective task, the aim being to obtain the most realistic fake possible.

Last year, electronic chip manufacturer Nvidia thus managed to create fake photos of celebrities that were particularly credible! But GANs can also be very useful in a wide range of areas: finding the best strategy for a particular problem (to learn to play chess… or still to improve network performance), creating music or paintings inspired by a particular artist or style, etc.

GANs can thus be used in different areas, from video games to particle physics, through art (referred to as “GANism”) and pharmaceutical research.

They constitute a powerful tool for artificial intelligence. Unlike other methods of machine learning that require huge amounts of learning data, these algorithms can work with relatively little data because they produce it themselves. This has enabled a giant leap forward in unsupervised learning.

Since their invention in 2014 by Ian Goodfellow, GANs have attracted a lot of interest. Yann LeCun, the boss of AI research at Facebook, has for example presented them as “the most interesting idea in the last ten years in machine learning”. In effect, GANs are pushing back the boundaries of artificial intelligence, which is no longer simply learning to recognise data, but is becoming creative.

Such progress does however entail risks: a machine that is trained to create imitations that are as realistic as possible is the ideal weapon for fake news-mongers wishing to put words in the mouths of politicians, for example.