The game uses doodles to teach computers to understand images
Researchers at the Allen Institute for Artificial Intelligence (AI2) invented a children’s game to teach computers common sense. The game, Iconary (think Pictionary with icons instead of doodles), can train AI to think more like humans.
In previous experiments, we learned that AI can master chess and beat a human playing Go. Those tasks are impressive, but fall short of complex thought or understanding complex images. Unlike human children, AI does not intuitively understand that a drawing of a house and three stick figures represents a family. Or that a dog and a house equal doghouse.
To play chess or Go, AI simply needs enough data to understand the rules and consequences of each move. Iconary demands more. It builds on deep learning, or self-education for computers. The computer has to deduce what the icons could stand for based on previous examples. That is much closer to how humans think naturally. But it is difficult for computers to master.
AllenAI does well on the easy level with simple phrases like “ride a bicycle” (stick figure on a bicycle). But it struggles with more complicated tasks. For instance, when trying to communicate “measuring a hand”, AllenAI selected icons showing a hand and two scales.
Iconary is online and available to anyone. To play, simply guess what AllenAI draws or have the computer guess what you are drawing. The project is collaborative: the more you draw, the more AllenAI learns.