Self-driving cars, genuinely useful smart AI assistants, and impressive machine translation tools are just a few of the things experts once predicted would be impossible for a computer to carry, but which are now very much a part of our world.
Something computers still can’t quite nail, though? Creating computer-generated faces convincing enough to fool the brain into thinking we’re looking at something real. Nicknamed the “uncanny valley” by robotics professor Masahiro Mori in 1970, the results can be unconvincing at best — and, at worst, pretty darn creepy.
A new project carried out by Google researcher Mike Tyka (on his own time, and not an official Google project) sets out to take AI-generated faces to the next level. Called “Portraits of Imaginary People,” it utilizes neural networks to create faces that look impressively real — albeit stylized.
“It uses a neural network technique called ‘generative adversarial networks’ invented by Ian Goodfellow,” Tyka told Digital Trends. “GANs work by using two artificial neural networks which are playing an adversarial game. One — the ‘Generator’ — tries to generate increasingly convincing output, while the second — the ‘Critic’ — tries to learn to distinguish real photos from generated ones. With time, the generated output becomes increasingly realistic, as both adversaries try to outwit each other. As with all machine learning technology what’s cool is that you can create an algorithm simply by feeding through a large number of examples rather than having to hand construct the rules that govern the algorithm.”
If you think that’s cool, however, then you should check out an online demo by web developer AlteredQualia, which has taken Tyka’s work and combined it with yet another neural network called DeepWarp to create a demo in which Tyka’s faces follow your mouse as it moves around the screen.
We may not yet be at the point where AIs can create faces that completely fool our brains into thinking they’re real. However, as this work shows, we’re not a million miles away, either.