Researchers are using a Pictionary-style game to teach AI 'common sense': 4 notes

Jessica Kim Cohen - Print  | 

Researchers from the Allen Institute for Artificial Intelligence — a Seattle-based research institute funded by Microsoft co-founder Paul Allen — created a Pictionary-style game to teach AI "common sense," according to MIT Technology Review.

Four notes:

1. The online game, which pairs human players against an AI program, is called Iconary. Ali Farhadi, PhD, lead researcher on the project, argues Iconary is a useful measure of an AI's knowledge base because it requires broad intelligence.

2. During Pictionary — and now Iconary — a player is asked to draw an image that conveys a written word or phrase, which their opponent attempts to guess. However, players typically don't only draw the phrase itself — they also use context clues. For example, when given the phrase "wedding ring," a player might draw a bride and groom in addition to the ring itself.

"You've got to use a lot of sophisticated reasoning," Dr. Farhadi told MIT Technology Review. "It actually teaches common sense."

3. Today, AI's lack of common sense is one of the ongoing obstacles developers face when creating lifelike chatbots and voice assistants. The Allen Institute researchers hope that, through playing enough games of Iconary, their AI program will develop common-sense understandings of how different concepts — such as "books" and "pages," for example — work together in real life.

4. To play Iconary, human players and the AllenAI program take turns as the drawer and the guesser. As the drawer, the user is provided one of more than 75,000 possible phrases, and then given the opportunity to sketch a drawing that conveys it. These sketches are turned into clip-art icons, from which the AI program tries to guess the phrase. AllenAI might request another image to clarify.

To play Iconary on the Allen Institute for AI's website, click here.

More articles on artificial intelligence:
Northwell Health adds AI for readmissions to its EMR
New algorithm from MIT might be able to 'de-bias' AI: 3 notes
Dr. Eric Topol: 10 potential AI applications for clinicians, hospitals

© Copyright ASC COMMUNICATIONS 2019. Interested in LINKING to or REPRINTING this content? View our policies by clicking here.

To receive the latest hospital and health system business and legal news and analysis from Becker's Hospital Review, sign-up for the free Becker's Hospital Review E-weekly by clicking here.