Can a machine learn to recognize bad doodles? And even get better at recognizing them the more it sees?
The answer is apparently yes, as this impressive-but-potentially-frightening A.I. experiment is showing us.
How does it work? It picks six things for you to draw on your screen. You get 20 seconds to draw each thing, and then the program (a "neural network") works to guess what you're drawing.
For example, look at these six awful attempts (though in my defense, scribbling with a mousepad is hard – I'd recommend going to the site on a phone/tablet).
Quick, Draw! managed to know every one, including that thing that was on its way to being a lamp post – after I'd drawn just two lines.
The idea is that it'll get better and better at knowing the object, and be able to guess faster, the more people draw. PC Mag says it wasn't very good at guessing when it first launched (which appears to be a couple days ago). But now it does really well.
The developers said they trained it in a few hundred concepts, and want to add more. And "the more you play with it, the more it will learn."
Here's a video from Google Developers, which explains how it's not just about what you draw, but how you draw it – the order of the strokes, for example.
So there you go: Whip up a few doodles for Google right now, and in 20 years when robots are able to walk the Earth and recognize every individual's face, we'll know we only have ourselves to blame.