In 2018, a new kind of AI will show off its ability to produce artworks that can not only imitate old masters but which can take off in startling new creative directions. Generative adversarial networks (GANs) bring a new level of sophistication to graphics. Not only can they produce totally convincing artificial images on demand ("Donald Trump on a skateboard being chased by a polar bear"), they can tweak existing images in subtle ways ("make it look like the Sun is shining"). A GAN involves two separate neural networks, a generator and a discriminator. The generator produces images and the discriminator rates them. For example, the generator might be fed a large database of images of dogs and attempt to produce its own imitation dog picture. The discriminator then tries to tell the difference between the fake dog and real ones, and feeds back to the generator. The generator rapidly gets better at producing dogs, and the discriminator becomes better at spotting fakes.
GANs have exploded since Ian Goodfellow from Google Brain published a paper on them in 2014, though the basic idea is much older. The results can be impressive. A team that included art historian Marian Mazzone and members of Facebook's AI lab in California used a GAN working with a database of 80,000 paintings to produce pictures in the style of great artists.
An Amazon team based in San Francisco is developing a GAN fashion designer. This learns about a particular style from images, then generates more clothes in the same style. US and Japanese researchers have used a system called DRAGAN to generate unique, high-quality animé characters based on a given set of characteristics. Others use GANs to generate computer- game scenery and clean up blurry video.
InfoGAN, developed by Peter Chen at UC Berkeley, has the ability to manipulate images of faces. It can change the pose, lighting or camera angle while keeping the face the same. It shows InfoGAN has worked out for itself the features that make up the face, and how to change the image while keeping these intact.
All of these applications use GAN's power not just to process images but understand their content. By grasping the essential "Trumpiness" of Donald Trump, a GAN could turn out convincing images of the president in situations that never occurred.
There will be a huge increase in original content. Netflix plans to spend $7 billion on programming in 2018, and wants half of its library to be self-produced. Facebook and Apple are also reportedly prepared to spend $1 billion each.
Then there’s DABUS, which stands for Device for the Autonomous Bootstrapping of Unified Sentience. Developed by Stephen Thaler, CEO of Imagination Engines, Thaler has been working on what he terms Creativity Machines for more than two decades, noting that “GANs are at best a subset of Creativity Machines.”
“Creativity Machines are brainstorming neural nets selectively reinforcing ideas that work, while weakening those that don’t,” says Thaler. He believes that DABUS’ architecture represents the best prospect for an AI able to approach genius levels of creativity.
DABUS does not just tweak the style of a picture, but can create a whole new composition out of elements it finds aesthetically pleasing. One image started with a classic car, morphed it into a shark and then depicted the car/shark devouring a victim. DABUS entitled the result ‘Road Shark Kill.’ It does not just shuffle pixels, but manipulates ideas. However, Thaler says DABUS’ artworks are just an appetiser.
“The art is meant to be an introduction to the system,” says Thaler. “Like Einstein entertaining us with his violin before telling us about relativity.”
Thaler says DABUS can produce innovative concepts in fields as diverse as corporate strategy, economic prediction, and scientific theorising. Thaler is not revealing specifics or who he is working with, but mentions that DABUS has been working on stock market prediction.
“Some important players would like to use the DABUS system for predicting the future on an unprecedented scale,” is all Thaler will say.
We can expect to see GANs producing surreal artworks, perfectly retouched photographs and cool clothes in 2018. And these may be the least of their achievements.
This article was originally published by WIRED UK
