How MIT Is Teaching AI to Scare Us
Help an algorithm learn how to scare the sh*t out of you.
Taj Mahal. Image: MIT
Halloween's approaching, and MIT wants to use AI to scare the daylights out of you.
MIT's Nightmare Machine is an experiment to see if machines can learn to scare us. Through a series of algorithms designed to create creepy images, it aims to identify what makes something horrific, and then apply that information to freaky faces and portraits of famous places around the world, such as the Taj Mahal or the Golden Gate Bridge.
Users look through the images on a website or on Instagram and vote for what freaked them out. "Do they scare you?" MIT's spooky project, asks the audience about its Haunted Faces. "Help our algorithm learn!"
The researchers trained deep neural network algorithms to recognize exact features from a scary image, such as a haunted house, Iyad Rahwan, associate professor at the MIT Media Lab, told Motherboard. Then the algorithms would build their own internal representations of those features responsible for making the image "scary," and apply them to new images.
Google Deep Dream, a computer vision program that enhances images to a kind of hallucinogenic appearance, was also vital in turning images of well-known tourist sites into Halloweeny death traps in the style of "haunted house," "slaughterhouse," "toxic city," "ghost town," "fright night," "alien invasion," "inferno," and "tentacle monster." The program uses neural networks and feedback loops to enhance and repeat image patterns in increasingly trippy or abstract styles
"Quantifying what is scary is not an easy problem, let alone describing it to an algorithm," Rahwan said. "This is precisely why machine learning is useful here, because we can show the algorithm a picture of something scary, and ask it to learn those scary features on its own."
But the process isn't perfect, and not all the AI images turned out to be scary, he admitted. That's why are asking people to vote on what they consider scary or not. So far, they've collected over 100,000 individual evaluations.
"Initial tallies reveal that humans quickly converge on finding some of them very scary, and others not so much," said Manuel Cebrian, another member of the MIT team.
To the computer, however, these faces make no difference, they're all "equally creepy," team member Pinar Yanardag added. "So that reveals that there is extra information in how humans perceive horror that can be exploited to make even scarier faces, or even personalized horror images, were we to tailor the generation process to the individual data."
For now, the Nightmare Machine is just a fun experiment for Halloween, but ultimately, according to Rahwan, "the goal is to understand the barriers between human and machine cooperation."
Reconciling the psychological perceptions of what makes humans and machines tick is important for cooperation between the two, he said. "This project tries to shed some light on that front, of course in a goofy Halloween manner, in the Tradition of MIT hacks!"
Get six of our favorite Motherboard stories every day by signing up for our newsletter.