A chocolate brown dog looks down the barrel of the camera lens. A spider is suspended above grass by its web. They are gorgeous photographs, but they don鈥檛 actually exist 鈥� artificial intelligence dreamt them up.
Andrew Brock of Heriot-Watt University and colleagues at Google鈥檚 DeepMind created a generative adversarial network (GAN), a type of algorithm that pits two聽AIs聽against each other, to produce new images.
The GAN, known as BigGAN, was first trained on thousands of images linked to particular words, such as dog or butterfly. Normally, systems like this only get around 64 images per term, but due to its massive processing power, BigGAN can cope with 2,000. Once trained, BigGAN can then create its own images related to each word.
Advertisement
鈥淎 lot of the images are photorealistic, and the algorithm has managed to do hundreds of different items,鈥� says Janelle Shane, a research scientist who has previously used AI to create April Fool鈥檚 pranks.
However, BigGAN still struggles with some parts of nature. For example, the AI trips up over the vast number of legs on a spider, giving them 20 or more in some instances, or adding too many eyes to frogs to make mutants.
Last year computer chipmaker Nvidia trained a GAN to generate realistic fake celebrity faces, but the range of images produced by BigGAN are significantly more varied.
They鈥檙e also good enough to hoodwink its creator. During the project, Brock came across a strip of images of jaguars and bears that BigGAN had generated and mistook it for a Google Image search result.
Whilst the world is concerned with fake news, hyper-realistic AI images could seem a problem 鈥� but BigGAN鈥檚 current best work is generating dogs. 鈥淚t鈥檚 pretty good, but the internet is already flooded with dog pictures,鈥� says Shane.
搁别蹿别谤别苍肠别:听
