Artificial Intelligence is developing at an incredible rate, but it may not always be immediately apparent to a layman.
Where you can immediately see improvement by leaps and bounds however, is in image generation. And boy, are we good that that by now.
Images courtesy: Nvidia
Get ready to have your mind blown, because none of the people in those images up there are real. They've all been generated by an AI tool developed by Nvidia, using a technique called generative adversarial network (GAN).
This is a system where the AI is split into two parts, where one generates and the other checks for authenticity, and they essentially compete against each other. This way, the AI is working against its most formidable opponent (itself) and can improve exponentially in a very short amount of time.
And to give you an idea of how far AI has come along, take a look at the photo below.
That's a set of faces generated by an AI back in 2014, as part of the published paper that first introduced the concept of GAN to researchers. As you can see, there's a world of difference between the two outputs.
What's great is that the Nvidia images aren't just ?incredibly accurate, they're also customisable. Nvidia's researchers included a tool called neural style transfer, which blends the characteristics of one or more images to the image its fed. It's like that formerly popular app Prisma, which could make your selfie look like a Van Gogh piece.
Also Read:?Chinese State News Agency Has AI Anchor To Read English News Bulletins And 'He' Looks Real!
In this case, style transfer allows researchers to blend the major features of someone's face to a large degree. In the grid below, the top row are the source images of real people being manipulated. The left hand column has styles of other faces blended in, to varying degrees, with the central block being the output. What you get at the end, is an entirely new, yet believable, person.
This AI isn't limited to just faces either. It can also recreate interior decoration photos...
... and images of cars.
But talking about how cool all of this is, brings us to how terrifying it can be. Experts have for the past few years already been raising these questions. For instance being able to realistically generate fake faces can have serious consequences for a lot of sectors. News media, police, the justice system, all will have their jobs made harder. And in the Trump era, it also raises the question of how much more damaging this can make propaganda imagery.
Consider deepfakes, the AI tool that's able to fairly accurately create non-consensual porn. It has positive uses too, but the majority of content it's used to output is nothing like that.
Of course, it took Nvidia a week of training their AI on eight Tesla GPUs, which isn't cheap. So there's some time before your everyday amateur AI enthusiast can generate these kinds of results. However, it's the kind of problem AI researchers, in collaboration with lawmakers, need to consider before we get to that point.