Skip to Content
Uncategorized

An AI is playing Pictionary to figure out how the world works

Forget Go or StarCraft—guessing the phrase behind a drawing will require machines to gain some understanding of the way concepts fit together in the real world.
February 5, 2019
ICONARYIconary

It might be a frivolous after-dinner game to you, but Pictionary could perhaps give AI programs a deeper understanding of the world.

AI’s lack of common sense is one of the main obstacles to the development of chatbots and voice assistants that are genuinely useful. What’s more, while AI programs can trounce the best human players of many games, including chess, Go, and (more recently) StarCraft, mastering them offers only a narrow measure of artificial intelligence. Learning to play chess, for instance, does nothing to help a computer play Sudoku.

Researchers at the Allen Institute for AI (Ai2) believe that Pictionary could push machine intelligence beyond its current limits. To that end, they have devised an online version of the game that pairs a human player with an AI program.

In case you’ve never played it before, Pictionary involves trying to draw an image that conveys a written word or phrase for your teammates to guess. This tests a person’s drawing skills but also the ability to convey complex meaning using simple concepts. Given the phrase “wedding ring,” for example, a player might try to draw the object itself but also a bride and groom or a wedding ceremony.

“You’ve got to use a lot of sophisticated reasoning,” says Ali Farhadi, the lead research on the Ai2 project. “It actually teaches common sense.”

That makes it the perfect vehicle to help teach machines. The team developed an online version of the game, called Iconary, that pairs a user with an AI bot called AllenAI. Both take turns as the artist and the guesser. Playing as artist, a user is given a phrase and then has to sketch things to convey it. The sketches are first turned into clip-art icons using computer vision; then the computer program tries to guess the phrase using a database of words and concepts and the relationship between them. If the program gets only part of the phrase, it will ask for another image to clarify.

Given the phrase “turning a page,” for instance, a player could try drawing a book, a hand, and a curved arrow. In that particular case, the AllenAI program guesses correctly after just a couple of tries.

The AI program uses a combination of AI techniques to draw and guess. Over time, by playing against enough people, AllenAI should learn from their common-sense understanding of how concepts (like “books” and “pages”) go together in everyday life, Fahadi says. It will also help the researchers explore ways for humans and machines to communicate and collaborate more effectively.

In total there are 1,200 icons, 75,000 possible phrases, and a vocabulary of 20,000 words. And there are two modes: easy and hard. The plan is to create a leaderboard to bring out people’s competitive spirit—and help accelerate the AI's learning.

Fahadi says Iconary is a better measure of AI smarts than Go or chess because it requires much broader intelligence. “Games have been a successful platform for testing AI,” he says, “but intelligence is more than pattern matching.”

Eventually, he says, it could be used as a new kind of Turing test—players could try to guess whether their teammate is a human or a machine.

Keep Reading

Most Popular

Large language models can do jaw-dropping things. But nobody knows exactly why.

And that's a problem. Figuring it out is one of the biggest scientific puzzles of our time and a crucial step towards controlling more powerful future models.

OpenAI teases an amazing new generative video model called Sora

The firm is sharing Sora with a small group of safety testers but the rest of us will have to wait to learn more.

Google’s Gemini is now in everything. Here’s how you can try it out.

Gmail, Docs, and more will now come with Gemini baked in. But Europeans will have to wait before they can download the app.

This baby with a head camera helped teach an AI how kids learn language

A neural network trained on the experiences of a single young child managed to learn one of the core components of language: how to match words to the objects they represent.

Stay connected

Illustration by Rose Wong

Get the latest updates from
MIT Technology Review

Discover special offers, top stories, upcoming events, and more.

Thank you for submitting your email!

Explore more newsletters

It looks like something went wrong.

We’re having trouble saving your preferences. Try refreshing this page and updating them one more time. If you continue to get this message, reach out to us at customer-service@technologyreview.com with a list of newsletters you’d like to receive.