What’s the latest game artificially intelligent software can play, you ask? Well, it’s Pictionary.
AI researchers at the Allen Institute for Artificial Intelligence (AI2), in Seattle, USA, have built Iconary, an “AI-powered drawing and guessing game” inspired by Pictionary. You can play against a bot, AllenAI, with one of you being the drawer and the other the guesser. And yes, that’s Allen as in the late Microsoft cofounder Paul Allen, who funded the institute.
As the drawer, you’re given phrases like “chasing a deer,” and have to draw the scenario for AllenAI to guess. It’s pretty straight forward. Two stick figures running to look like they’re chasing each other and a deer.
Unlike a real game of Pictionary, in Iconary you draw each section of the scene at a time, and each section is analyzed before moving on to the next part. In the case of “catching a ball,” first you draw a person with their arm out. The software makes a few suggestions from your sketch, and hopefully “woman with arm out” or “man with arm out” is listed. Pick one, and place its icon on the screen. Then draw a ball. The software should suggest a ball out of list of other round objects. Place that icon as well near the outstretched hand. Then AllenAI guesses something like “throwing a ball” then “catching a ball,” and the game’s complete.
And as the guesser, the bot takes the role of the drawer instead. What is comes up with is rather abstract, so be warned.
At first glance, compared to other games that AI can play such as Go, Chess or Dota 2, Iconary is pretty simplistic and unimpressive. It doesn’t require any complicated strategic thinking or lighting fast reactions. AI2, however, claims the game is more useful for teaching machines common sense.
“There has been a long history of the AI community choosing games like Chess, Go, or Atari as a sandbox to develop and test AI algorithms,” Ani Kembhavi, lead research scientist of the project, told The Register. “While the AI breakthroughs achieved on these games has been remarkable, these games are quite different from situations and scenarios that might be posed to an AI system in the real world.”
Iconary is an exercise getting computers to try and understand both images and text, he explained.
“It requires players to use every day common sense knowledge in their game play. For instance, drawing the word ‘family’ requires the system to understand that the word refers to a set of living things, such as adults, children, pets like dogs and cats, etc. And the word ‘father’ might be drawn with a large male icon standing next to smaller male and female icons, denoting children.”
OK, smarty pants AI. You can beat us humans at video games. But how about real-world puzzles like Jenga? Oh, oh no
AllenAI was trained by observing roughly 100,000 games between humans. There are 1,200 possible icons that can be used to represent 75,000 different scenarios, containing about 12,000 unique words. As the bot watches the games, it learns to associate the icons to the corresponding words in the phrases. The researchers held back roughly 2,000 phrases for testing.
Iconary is just a first step, Kembhavi said. Researchers plan to introduce more challenging concepts into the game such as Roommate or President.
“Our eventual goal is to do large scale Visual Turing Test by having AllenAI play with a large and diverse set of people,” he said. “This would entail people being paired either with AllenAI or with another person, and at the end of the game, they would be required to guess if their partner was a human being or AllenAI.
“The metric would then be defined as the number of times people could differentiate the two. This would be a mature of the system’s ability to exhibit intelligent behavior and be effective in collaborating with human partners.” ®