Last April, Google’s machine learning crew revealed AutoDraw, a fun little demo of all that neural network theory. In a nutshell, the web app tries to guess what your scribble looks like and identify it. Now Google’s researchers are taking that idea one step further. Called Sketch-RNN, this “recurrent neural network” model does for doodles and drawing that autocomplete does for words and sentences, trying to predict how to best complete your indistinguishable scribble.
Unlike AutoDraw, Sketch-RNN isn’t exactly trying to guess the object you’ve drawn. It already knows before hand whether you are trying to draw a cat, a spider, or the Mona Lisa. You yourself get to choose that before the drawing session begins. That, however, doesn’t make what it does any less impressive.
What the neural network basically does is to try and fill in the blanks you’ve left. For example, when drawing a cat, you start with a circle for the head. After a brief pause, Sketch-RNN analyzes what you’ve drawn and tries to predict whether it’s a head or, say, the body. It then draws other parts of the cat, like ears, whiskers, or legs, taking over where you left off. You can continue to refine your own drawing and Sketch-RNN will change its own predictions accordingly.
It’s far from perfect, of course, and the neural network will sometimes place body or facial parts in ridiculous places. But the very fact that it can even do that is enough of a testament to the complexity of this neural network. It had to learn these almost like how we learn how to identify objects. And it had a lot of help, thanks to the Quick Draw game from last April.
If anything, Sketch-RNN is fun to play with and might even help make it better. Google recommends using a desktop browser, but even a mobile browser would do. However, smartphone users on data connections should be aware that each model, that is, each type of object being drawn, is a 5 MB download that gets downloaded again and again every time you switch models.