MIT’s latest project toes the line between “could” and “should.”
A series of algorithms dubbed the Nightmare Machine is an effort to find the root of horror by generating ghoulish faces, and then relying on user feedback to see which approach makes the freakiest images. MIT also used Google’s DeepDream method to create ghastly portraits of famous locations around the world, just in time for Halloween.
It’s easiest to think of the fear-generating AI as a complex black box that draws a best fit line. When given the task of creating something, it generates an image based on everything it has seen before: in this case, scary faces. Each “scary” or “not scary” vote in MIT’s game pulls the best fit line slightly in some direction: more teeth, paler skin, darker background. With enough information, the AI could theoretically generate the sum of human fears.
While it’s a fun game around Halloween, the project also shows how quickly AI research is progressing. The two main techniques used in the project, style transfer and generative adversarial networks, were published in papers only last year. Now the technology is easy enough to implement in a novelty project made by just three computer scientists.