Computer programs such as FaceTracker can find human faces in photos and video feeds, but it seems particularly human to find faces in animate objects such as wall outlets or car fronts. There's even a word for the human tendency: pareidolia. Yet FaceTracker has a kind of pareidolia of its own. Its algorithms and training can lead it to find faces that don't exist, as Greg Borenstein, who studies interactive technology at New York University, blogged in January. He posted an example of a face FaceTracker found in a shadowed window in a studio in Philadelphia.
In this case, the program's pareidolia is very different from human pareidolia; people wouldn't "see" the shadows in the window as a face. "Something in that window looks like a face to the algorithm even though we can't see it," Borenstein wrote in an email to InnovationNewsDaily. He wondered how often machine pareidolia differed from human pareidolia, so he wrote a program that applied FaceTracker to 681 photos from a Flickr group called " Hello Little Fella! ", which is dedicated to photos of faces people see in animate objects.
FaceTracker found a face in 50 of the photos, which Borenstein posted on Flickr. In seven of those, the program saw the same faces people saw, as in the cheerful countenance of this key. The program's judgment seemed most human when there was a defined outline to the face, Borenstein wrote.
Eleven photos were a little off, according to our judgment at InnovationNewsDaily. In this photo, for example, the program saw the same eyes people saw, but disagreed about where the mouth was. In 32 photos, FaceTracker's algorithms found a face in a totally different part of the photo than a person would:
It's this last scenario that Borenstein found eeriest. "It made me wonder what I was missing," he blogged. "It's a feeling akin to having a conversation with someone who's gradually losing interest in what you're saying and starting to scan the room over your shoulder."
How does this happen? FaceTracker "learns" to see faces after being fed images of people with their eyes, nose and mouth labeled, Borenstein explained over email. But what the program learns from those examples is not always the same as what people learn from their own innate impulses and from seeing faces around them from birth. From our numbers above, it appears that what the program learns is actually mostly not the same.
Yet FaceTracker works well with real human faces. Borenstein previously demonstrated that it can recognize hand-drawn faces. The program also drew some attention in 2011 when programmer Arturo Castro used it to make a program that can meld and morph users' faces with other faces, such as those of celebrities.
Facial recognition programs might improve by thinking in more human ways and thus seeing more human pareidolia. But Borenstein, like some other designers, sees the impulse of the future in another way, he wrote in his blog. Instead of making programs think more like people, people may try to understand how programs "see," so they can design things to look more recognizable to robots.
Follow InnovationNewsDaily on Twitter @ News_Innovation, or on Facebook.