It's almost Halloween, and spooky things happen in October. Just in time, MIT scientists have succeeded in creating an AI that makes ordinary images look scary.
The AI makes images look like they are filtered, but indeed they are more than that.
A "scary team" of scientists (Pinar Yanardag, Pinar Yanardag and Iyad Rahwan) created an algorithm called Nightmare Machine to find the roots of horror. What it does is generating Halloween-inspired images from either people's faces (Haunted Faces) or places (Haunted Places).
AI in creating strange images isn't new, previously, Google's had come up with machine learning AI that dreams in 2015.
As a matter of fact, the team did used Google's DeepDream method to create Nightmare Machine's ghastly portraits and locations. The interpretation is similar, but the AI is dedicated for it.
Like earlier said, just in time for Halloween.
The algorithm is still a work in progress. It still needs to learn, and participants are allowed to teach it by voting whether or not different sets of faces are scary. This will give the AI more data to learn about what scary means for humans.
The AI learns by everything it has seen before, combining that with what it knows about "scary" or "not scary". Each vote participants give will pull the best fit line to slightly alter the images in some direction: more teeth, paler skin, darker background and so forth.
With enough information, it should understand the essence of horror to humans, thus theoretically makes it able to scare us even more.
To accomplish this feat, the team uses two main techniques in the project: style transfer and generative adversarial networks. These two were published in papers in 2015. Nthe technology is implemented in a novelty project by the MIT's team that only consists of three scientists.
At MIT's website, nightmare.mit.edu was created to publish images created by this horror-loving AI.