In April researchers at Microsoft released How-Old.net. Using machine learning, the software guessed the age of people in photos. The results were, ahem, not very accurate, but people loved submitting photos of themselves and their friends. Now a new project through Microsoft's experimental program Garage is doing a similar thing for identifying dog breeds. And it's terrible in the best way.
You can use Fetch! through a website, What-Dog.net, or an iOS app. When the service evaluates a dog photo and settles on a breed, it provides a rating of how strong it thinks the match is, and also lists runner-up breeds that it considered. Then it lists characteristics of the breed it assigned, like disposition and size.
Lily Hay Newman
With How-Old.net, Microsoft researchers seemed a bit out of touch with how Internet users would interact with the service. (They expected most people to try the service out on stock photos, not personal photos). Fetch! is much sleeker, but like with How-Old.net, the true joy of using it doesn't really come across in the description, which says: "This is the kind of app you’re going to take out when you’re with your friends. You’ll make fun of each other, comparing which breeds you look like, and posting the tagged photos."
The Fetch! mode that allows people to be assigned a dog breed is kind of funny, sure, but I was much more interested in throwing other animals and inanimate objects at the A.I. What would it think about a photo of a plate of bagels?! (To the service's credit, it did realize that the bagels were not a dog.)
When I wasn't trying to trick it, the service did make some impressive guesses. But it also made a lot of mistakes. Maybe Fetch! will get better as more people feed it data to learn from. For now it's pretty delightful as is and a great excuse to look at dog photos.