The Nightmare Machine: artificial intelligence gets spooky

By Sophie Schmidt

28 October 2016

Kermit the Zombie before/after. Image - Nightmare Machine

Kermit the Zombie before/after. Image – Nightmare Machine.

One of the biological side effects of being a human is the will to live. Luckily for us, one of the ways in which our brain gives the heads up to inform us of potentially dangerous situations is by invoking that little old survival instinct called “fear”.

Have you ever been stuck sitting next to someone in a cinema, completely unfazed by a horror movie, while you diverted your attention to the closest escape door? Everyone gets spooked out by different stimuli – whether rational or irrational – clowns, gigantic spiders, or even marshmallows.

Since we know that stimuli can evoke varying psychological responses, one group of researchers from our team at Data61 and MIT Media lab, set out to find what unites us in our phobia and terrifies us on a universal scale. And so they created a horrifying algorithm to generate scary imagery, designed to spook the living daylights out of us mere civilians.

Welcome to the Nightmare Machine

A monster creation just in time for Halloween, which transforms an idyllic scene – whether it be an Ikea catalogue or the Taj Mahal – into an slaughterhouse or inferno, using cutting-edge Artificial Intelligence.

The spooky formula is powered by deep learning algorithms, and a secret ingredient that our principal research scientist at Data61, Manuel Cebrian, claims as “evil spirits!”. The team used two deep learning algorithms: one for extracting artistic styles from one image to apply it to another, and a second algorithm that generates “imagined” faces from trained data.

Do these faces scare you? You can help the algorithm learn with your answers. Image - Nightmare Machine

Which faces are scarier? You can help the algorithm learn with your answers. Image – Nightmare Machine

So far, the team has collected over 200,000 individual evaluations of computer-generated images, using a website form that presents a spooky image and asks the user to rate its scariness factor. The algorithm grew hungrier and hungrier for more user data, until it was able to think and feel on its own.

“We started little by little, experimenting with what we call the “nightmarifying” process. We use deep learning algorithms to learn first how haunted houses, then ghost towns, and more recently toxic cities look. Then, we apply the learned style to famous landmarks. It’s surprising how well the algorithm can extract the element from the “scary” templates and plant it into the landmarks,” Manuel said.

From an initial look at the data, Manuel said that the “tallies reveal that humans quickly converge on finding some of them very scary, and others not so much”.

The group’s primary goal is to understand the barriers between human and machine co-operation – psychological perceptions of what makes people tick and what make computers tick are an important barrier for such cooperation to emerge.

If you’re trying to decide what’s more terrifying– the images, or the idea that an intelligent machine is capable of generating them – Manuel assures us that the team is interested in testing this experimentally to find out – so keep your eyes peeled for more updates from the Nightmare Machine.

You can find out more about Data61’s (less spooky) work here.