This article was taken from the April 2014 issue of Wired magazine. Be the first to read Wired's articles in print before they're posted online, and get your hands on loads of additional content by subscribing online.
If you've ever spotted faces or animals in floating clouds, you've experienced pareidolia -- one of the brain's inexplicable quirks. But it's not unique to you -- machines do it too. "Machine vision is processed in a similar way to ours. They have the same bias to spot patterns," says London-based web designer and coder Henry Cooke. "So do they spot these faces, too?"
To explore this, he created a Tumblr to post images where machines have drawn in faces they have "spotted", but don't actually exist. Cooke's algorithm pulls random images from Flickr, Instagram and FFFFound, and puts each one through an open-source face-scanner that draws rectangles around possible faces, giving each a probability score. "If the score comes back too high, I know it's definitely a face, so I reject it.
Similarly, if it's too low, it gets rejected," Cooke explains.
If the probability score is somewhere in the middle, it is selected. Next, a more specialised face-tracker called offline_face draws in wire-frame features -- eyes, nose and a mouth -- where it believes they are. "The best ones are when it's a bit creepy," says Cooke. "Like a face drawn into a window or a tree.
It's just a face hanging there, but it gives you a shiver. Even though the process is completely mechanical, the spark of recognition you get is rooted in human perceptions."
MV
This article was originally published by WIRED UK
