Google's 'Quick, Draw!' AI Guesses Your Scribbles

Advertisement
Advertisement

Have a look at the scribbles below. Drawn on a laptop's touch screen with a somewhat errant index finger, they're not exactly fit to be displayed in The Louvre. Artificial intelligence is still working on things like driving and playing Go, so one would think that it would probably take a human to identify some of these drawings, but that's not actually the case. Thanks to the magic of machine learning and neural networking, Google's new AI, called "Quick, Draw!", is able to decipher all but the very scribbliest of scribbles. It's presented through a web app that anybody can access, and the more people use it, the better it gets at identifying drawn objects.

The way it works is pretty simple, if a bit limited in its scope. Users hop on and are given 20 seconds on the clock to draw an object of the AI's choosing. This can be done with the mouse, a drawing tablet, a touch screen, or just about any other apparatus that can be hooked up to a computer and draw a picture. Once the user is finished with their drawing or time is up, the AI flips up the curtain and it's on to the next picture. A user gets six chances to draw something, then they are shown a gallery of their drawings, which they can click on to check out a quick showing of whether the AI recognized the drawing, what it thought the drawing looked like, and drawings that are used as examples of the object at hand.

Google may be one of the best in the biz when it comes to artificial intelligence, machine learning, and neural networking, but they didn't go it alone this time around. The nifty project was part of a huge joint AI venture with Google's bosom buddy, Snap Inc. Specifically, their head of research, Jia Li, who is now a full member of Google staff. Li serves under cloud head Diane Greene, and is jointly in charge of a bevy of new projects alongside Greene and Fei-Fei Li, Stanford's old AI director. The news came amid announcements of new cloud products, and that Google would be making even more heavy use of machine learning going forward than they already do.

Advertisement