This is what I think a passport looks like:
I drew this—more of an abstract half-eaten apple inside a frame—to add to the 120,000 crudely drawn passports collected by Google. Now you can explore every single hastily doodled one.
Fifty million doodles have been created in Quick, Draw! a game built by machine learning researchers, in collaboration with Google’s Creative Lab, since its release in November. The doodles come in categories like “giraffe,” “moustache,” and “the Mona Lisa.”
Yesterday the game’s creators made all of those drawings and their metadata open-source, for anyone to explore, download, and analyze. An interactive site makes it easy to look at how drawers interpreted different subjects. Here’s what a few Quick, Draw! doodlers think a parrot looks like, for example:
The purpose of Quick, Draw! is to train a neural network about how humans draw things. Human users are given a cue like “owl,” and then, with a mouse or finger, must draw the prompt in 20 seconds.
Like in Pictionary, Google’s algorithm shouts out guesses: “I see bread…pool…purse…” until it matches with the right answer. The guess is based not only on what you draw, but how you draw it, a promo video says, taking into account the order and direction of the strokes.
The game has produced some pretty cute animal doodles, like these camels:
…as well as some pretty basic looking body parts, like these knees:
Blackberries were clearly a challenge for many would-be artists:
As was the Great Wall of China:
Since the point of the game is to help a computer guess what you’re drawing, Quick, Draw! users tend to draw the simplest and most traditional versions of their subjects. We don’t find, for example, many “campfires” from an aerial view. And as far as I can see, no one drew baseball mitt-shaped ice pops for the prompt “popsicle,” either.
Last month another team used collected drawings from Quick, Draw! to create a second neural network that draws as humans do.