If you haven’t already, go check out What Dog Dot Net (what-dog.net), a website that has been collectively humping the internet’s leg since Thursday afternoon. From the same Microsoft minds who brought us that website that incorrectly predicts your age, this program sometimes correctly guesses the breed of your dog.
The real joy is not using the way the program is expressly laid out, however, but forcing the algorithm to choke down garbage and spit it back up. People have forced presidential candidates in the system. And college football coaches. The computer tries to do its job, we laugh at its attempts to frame the universe as a canine, and hilarity (?) ensues:
That the Microsoft Garage app sees all the world a dog isn’t quite a malfunction, but it does highlight what happens when object recognition gets hyper-specific. (It also fails biology in its claim that visuals “show a percentage of the closest breed” for mutts or unknown dogs — to reveal accurate information about a dog’s ancestry we really need genetic analysis.)
Microsoft researchers say that the AI powering the app does know when there are only inanimate objects in the frame — though it can’t tell a dog from a cat — and that the person-as-pooch is a baked-in feature. It knows what a dog is supposed to look like, but it couldn’t tell you what a dog is.
But don’t let this app lull you into the sense that all recognition software is doomed to make silly mistakes. Though computers that analyze images from security cameras and other in-the-wild images still have a ways to go, online recognition systems, like the one Facebook uses, are increasingly powerful. We all get a number based on how far our eyes are from our nose and ears, and Facebook also relies on clothes and hairdos as identifying cues.
Not all uses of image recognition are as frivolous as tagging friends or turning Han Solo into a basset hound — the app Finding Rover, for instance, wants to bring lost dogs home through the eyes of a computer.