Pictionary. That cursed game if ever there was one. I'm terrible at guessing, and particularly horrible at drawing. Between the laughter of my rivals and the puzzled “what the heck is that?” from my team, a game of Pictionary can become very frustrating—but what if you could train with a neural network? In fact, that works for both sides: Google's new project Quick, Draw! doesn't just force you to create convincing drawings in twenty seconds or less, but in the process you make its neural network smarter.
“Yoga.” How can I draw something in less than twenty seconds with a trackball that represents “yoga”? Solution: exploit the absence of rules. I drew the word “yoga” and added an arrow. “Oh, it’s yoga!” exclaimed the neural network. Mission accomplished. I laughed for five minutes, and I didn’t repeat the strategy with “watermelon” because I wasn’t going to draw a watermelon with a trackball... but then I felt chills. This thing is much smarter than it appears. 4 out of 6 doesn’t sound bad at all for a technology that’s under training, and that must interpret those horrors that only a pious person would call drawings. That’s the key. A couple of years ago, that 4 out of 6 was 0 out of 6, and if we go further back, it was 0 out of 500. Machines couldn’t do this. The code couldn’t do this. Now it can, and it adopts names like Quick, Draw!, my new partner in digital Pictionary.
Technically, Quick, Draw! isn't an isolated project. The folks at Google launched a portal called A.I. Experiments that serves as a showcase for the general public to appreciate the workings of artificial intelligences, neural networks, and the concept of “machine learning”. There are eight experiments available, though not all work online. Giorgio Cam creates music from a photograph. Bird Sounds classifies the sounds of certain birds, and The Infinite Drum Machine does something similar with everyday noises, and has a much more pronounced musical profile. But I keep coming back to Quick, Draw! If you click on one of the drawings, you'll see what the neural network visualized, and what the approximate results were.
I must warn you: Quick, Draw! is a little more addictive than it should be. Between your own drawings and the strange interpretations the neural network makes, there’s material for a while. I imagine Google will add other experiments in the future, including that Thing Translator that looks very good.
