Researchers have taught a computer to do a better-than-expected job of predicting what characters on TV shows will do, just by forcing the machine to study 600 hours’ worth of YouTube videos.
The experiment could serve as a commentary on the state of research into artificial intelligence, or on the predictability of sitcom plots. It also calls to mind the scenes from countless science-fiction movies where the alien gets up to speed on human culture just by watching TV.
MIT’s Carl Vondrick and his colleagues are due to present the results of their experiment next week at the International Conference on Computer Vision and Pattern Recognition in Las Vegas.
The researchers developed predictive-vision software that uses machine learning to anticipate what actions should follow a given set of video frames. They grabbed thousands of videos showing humans greeting each other, and fed those videos into the algorithm.
To test how much the machine was learning about human behavior, the researchers presented the computer with single frames that showed meet-ups between characters on TV sitcoms it had never seen, including “The Big Bang Theory,” “Desperate Housewives” and “The Office.” Then they asked whether the characters would be hugging, kissing, shaking hands or exchanging high-fives one second afterward.
The computer’s success rate was 43 percent. That doesn’t match a human’s predictive ability (72 percent), but it’s way better than random (25 percent) as well as the researchers’ benchmark predictive-vision programs (30 to 36 percent).
“Still a long way to go,” the team says in its own YouTube video.
The point of the research is to create robots that do a better job of anticipating what humans will do.
“It could help a robot move more fluidly through your living space,” Vondrick told The Associated Press. “The robot won’t want to start pouring milk if it thinks you’re about to pull the glass away.”
That predictive capability would also come in handy for robot caregivers. For example, if it looked as if someone was about to fall down, the robots could anticipate that and intervene, Vondrick said.
And who knows? Maybe they’ll be better able to rescue falling robots as well.