What is a GAN and why should we be aware of it?

04/06/2019

Take a look at the faces below? Do you notice anything odd about them, or do they look perfectly normal?

You may find it hard to believe, but none of the people above actually exist - they have all been digitally generated using Artificial Intelligence (AI)!

Generative Adversarial Network - or GAN - is type of deep-network AI, built to learn how to create. The creators started by showing the AI faces in a pixelated 'noisy' form that increases in clarity so that the program is able to learn what a face typically consists of. GAN works by splitting the AI’s work up into two separate algorithms: the first is tasked with creating and is known as a 'generator', and the second is known as a 'judge' and tries to determine whether what was created by the generator is a real image or a product of AI. Programmers then tweak the algorithms until it produces consistently realistic images, at which point the AI is smart enough run by itself! The output data can even be altered so that only people with certain characteristics, such an a certain skin, eye, or hair type are generated.

You can try this out for yourself here, just press refresh on your browser and a newly generated face will appear each time; pretty crazy!

The same principle has been applied to fashion models too. DataGrid showed their GAN thousands of photos of fashion models until it learnt the pattern in the images, which it could then use to create its own made up versions using the same type of algorithm as above. This application could eventually be used by companies to reduce the need for photo shoots to show off their products.

Other GANs have been taught to produce fake photos of food, Airbnb accommodation, to turn photos into the style of Monet or Van Gough, and to even alter videos of horses trotting about in a field into zebras! In some cases it’s being used in the field of medicine to spot abnormal cells, which would save time waiting for a trained person to look over the images and could mean you would get results from tests back quicker. Text-to-image GAN is also being developed, where you can describe something like a flower in text format and it will generate an image based on your description!

Samsung AI have devloped a deep learing machine that can turn portraits into moving, talking images known as living portraits! This AI studies a large collection of videos of faces talking and moving and learns where the main areas of the face are and how they move. Once this is applied to a portrait, it allows the photo to look like it is speaking or moving in a realistic way. You can view some of the examples of this here, which range from everyday people and celebrities to famous portraits such as the Mona Lisa.

There is also a slightly darker side to GAN, which could be used for more nefarious reasons. You can show the AI videos of someone speaking and it will learn their facial expressions. These can then be used to create convincing footage of someone saying something, or even potentially doing something they’ve never actually said or done before. This example someone created of Obama speaking shows you just how realistic it can be!

GAN is still in its early stages of development, but could grow to become a very powerful piece of software, for better or for worse, this tech is growing! Either way, it is interesting to see how AI is evolving and what else it could be used for in the future!