For decades, scientists have sought to give computers common sense—a basic understanding of the world that lets humans navigate everything from conversation to city traffic. Now, researchers have come up with a new approach: They’ve designed an artificial intelligence (AI) that can abstract knowledge and generalize it to play the surprisingly subtle drawing game Pictionary.
“This is a first step toward exploiting common sense,” says Aniruddha Kembhavi, a computer scientist on the project at the Allen Institute for Artificial Intelligence (AI2), a nonprofit lab in Seattle, Washington. Angeliki Lazaridou, a computer scientist at DeepMind in London who tried the game, agrees. She says the AI learns how people understand basic concepts by finding the minimum elements required to convey them.
Previously, computer scientists aimed for common sense by programming AIs with the laws of physics or uploading lists of facts. That works for pool or trivia night. But Pictionary is far more complex: It asks players to guess words or phrases based on the sketches of a partner. That requires abstraction, reasoning, communication, and collaboration.
The researchers created a new game, Iconary, where players select from 1200 icons (images that depict trees or arrows, say) and arrange them to convey a randomly generated phrase. A partner then guesses the phrase until they get it right—or asks for new sketch.
They also created an AI, made public today on their website, that anyone can play Iconary with. It learned by watching 100,000 games between human players. The AI relies on neural networks, software that emulates the brain by learning from experience, and a vast database of numeric codes used in translation software. The codes represent the meaning of words: “Chair” and “couch,” for example, are closer in value than “chair” and “dog.” The AI translates words into codes—using the entire phrase as context—but instead of translating the codes into another language, it translates them into icons.
“We wanted to build an AI system that can collaborate with human beings, and at the same time is learning about how humans think, how they act,” says Ali Farhadi, an AI2 computer scientist on the project. Going forward, he adds, it will learn by playing against people.
AI experts have mixed reactions to the new algorithm’s importance. Catherine Havasi, a computer scientist at the Massachusetts Institute of Technology in Cambridge, appreciates its ability to generalize, given the inflexibility of many machine learning algorithms. “There’s a depth here, and the ability to take things it learns from one phrase or one instance and generalize to others,” she says. But Ernest Davis, a computer scientist at New York University in New York City, was more skeptical of the project’s link to common sense. Much of the task is simply matching words to icons. “That’s a very limited form of commonsense knowledge,” he says.
A key element of the new system is that it can resketch ideas based on a partner’s guesses. Farhadi senses a real collaboration when playing: “I actually kind of feel that this system is connecting to me deep in my thoughts.”