This AI can paint real worlds completely from memory

Google may have taught its AI to dream up filthy images based off of real-world examples, but one researcher has built an AI capable of painting entirely new worlds completely from scratch.

The image at the top of this story isn’t real, it’s a fake world painted by an AI. Created by Qifeng Chen from Stanford University and Intel, the AI essentially paints by numbers what it believes a street view should look like. By using thousands of reference images, learning where a “car”, “road” or “sign” should be placed, it can then paint a depiction of the real world – one that technically doesn’t exist at all.

Trained with 3,000 images of German streets, the AI is able to recognise what a “car” is, drawing one from memory. “We want the network to memorise what it’s seen in the data,” explains Chen.

As neural networks go, Chen’s managed to build something currently unheard of in the world of AI. Current generations of a realistic-looking fake world aren’t very convincing, but Chen’s are incredibly detailed for a computer-created scene.

Chen understands that, while his creation is quite the achievement, it still has quite some way to go. The images it currently produces feel dream-like: some things are slightly off and everything looks like oil paints on a canvas than a photo-realistic depiction of the world. The AI has trouble filling in the details we’d expect of an image, and so Chen has already begun work on a larger system that should be capable of creating accurate artificial scenes.

His ultimate goal is to create an AI that allows for the building of complete virtual worlds, not just flat paintings. “Using deep learning to render video games could be the future,” said Chen, who has already used his algorithm to replace parts of Grand Theft Auto V.

To become a creator of virtual worlds Chen’s AI also needs to branch out from simply looking at German roads. That’s easier said than done, especially as each image has to be labelled by hand for the AI to understand what each component of a picture is, but if Chen were to combine his neural network with an image- recognition one, perhaps the task of labelling up another 3,000-plus photographs wouldn’t be so labour intensive.

The possibilities of Chen’s AI could mean being able to create new VR worlds ripe for exploration from little more than a description spoken aloud. It could become a digital creator, allowing you to explore dream-like incarnations of the real-world in VR. It may well be the perfect setting for a dystopian world, one where an AI has enslaved us all but keeps us happy inside a network of its own creation – The Matrix, anyone? – but thankfully we’re a long way off from that if that’s the case.

Intel and Chen will be showcasing his neural network at this year’s International Conference on Computer Vision in Venice, Italy this October.

Source: arXiv:1707.09405v1

Leave a Reply

Your email address will not be published. Required fields are marked *

Disclaimer: Some pages on this site may include an affiliate link. This does not effect our editorial in any way.