Halloween Horror: AI Learns How to Terrify Us
Just in time for Halloween, MIT researchers have taught machines how to scare us.
The U.S. Capitol drenched in an eerie green slime of toxic waste. The Tower Bridge relocated from London to hell. And the fairytale Neuschwanstein Castle — the model Disney used for Cinderella’s palace — now horrifyingly haunted under a webbed, grey sky.
Be afraid: The Nightmare Machine lives up to its name. Using GPUs and deep learning, it turns ordinary images of places into spooky spectacles and those of people into ghoulish avatars.
“Nowadays, nothing seems to frighten humanity more than runaway intelligent machines,” said Iyad Rahwan, a professor of Media Arts & Sciences in the MIT Media Lab. “The rapid progress in artificial intelligence has people worried about everything from mass unemployment to the annihilation of the human race at the hand of evil robots.”
AI Turns Movie Stars into Zombies
No one and nothing is spared Nightmare Machine’s menace. Its deep learning algorithms turn Marilyn Monroe, Brad Pitt and even cuddly Kermit the Frog into zombies and other monsters. Superman is transformed from hero to haunted villain. And if Jaws wasn’t frightening enough, the Nightmare Machine version of the man-eating shark shows it is definitely not safe to go back in the water.
To generate the gruesome faces, the team trained its deep learning algorithm on 200,000 celebrity images. Researchers generated faces using a method published early this year, and “scarified” them with an algorithm that learns the artistic style of a given image, and transfers this style onto another picture.
The researchers converted treasured landmarks into landscapes of doom by training the algorithm to turn plain images into horror styles such as haunted house, slaughterhouse, inferno and toxic city.
GPUs Speed Deep Learning
To accelerate the deep learning, they used our GeForce TITAN X GPUs, along with cuDNN, to train and infer the scary imagery.
“Without the GPUs, we wouldn’t have been able to launch this project in such a short time — just two weeks,” said Yanardag, who is a post-doctoral researcher in the MIT Media Lab.
To help MIT’s Nightmare Machine be even more hair-raising, the team is asking people to rate the faces as “scary” or “not scary.” Users have submitted more than 400,000 evaluations so far.
“Maybe in the future we can generate ‘personalized’ horror images,” said Cebrian, a principal at the Commonwealth Scientific and Industrial Research Organisation in Melbourne, Australia.
An even more terrifying Nightmare Machine. Now that’s the stuff of nightmares.