The images featured on thispersondoesnotexist.com may seem indistinguishable from those of regular people that you come into contact with every day. The shocking twist of the website is that each photo on the site was created by artificial intelligence. The artificial artist behind the images is an artificial intelligence algorithm known as a generative adversarial network (GAN). The site, created by former software engineer for Uber Phillip Wang, produces a new, completely fake picture of a nonexistent person’s face each time it is refreshed. This site is a great representative example of what GANs are capable of. In this article, Haval Dosky discusses GANs and some of their most interesting implications for future innovation.
The principle behind GANs was first discussed by computer scientist Ian Goodfellow in 2014. Goodfellow designed a system where two AI systems (a “generator” and a “discriminator”) are pitted against each other in order to improve the quality of the results. The programs compete against each other millions of times to refine their imagine generation abilities until they are practiced enough to create pictures that are indistinguishable from fakes. Up until late 2017, as Haval Dosky notes, researchers were unable to create high-quality images using this method, but Nvidia finally cracked the code using a technique explained in its famous ProGAN paper.
With Facebook, Google, Nvidia and a host of other tech companies making continued developments with their own versions of this A.I. technique, Haval Dosky wonders what the future of the technology will hold. It is widely believed that the end goal of GANs is to create entire virtual worlds using automated methods as opposed to relying on hard coding. This could be a huge development for the virtual reality industry as it could point towards quicker development and more life-like designs, as hard code for fully-rendered, immersive worlds is already massively labor-intensive and time-consuming in the world of video game development – VR worlds would requires an entirely different level of immersion, making GANs a potential key to VR development.
While using GANs for VR development may still be a far-flung goal, experts in key industries are already finding ways to utilize GANs to produce revenue now. For example, brands have been making use of computer-generated characters to advertise their products, some of which have amassed millions of social media followers. As GANs make these computer-generated models more and more realistic, they will continue to reduce the labor involved in producing them. Haval Dosky also notes that, in a similar sense, GANs could be utilized in order to revolutionize image editing on real images, enabling editors to perform tasks such as applying complex variations to real image and quickly removing textures like rain or even blur from images.