Resarchers at MIT use must-see-TV to train computers to predict human behavior - CNET

There's a lot that artificial intelligence can do, but understanding human behavior isn't one of the strong suits. Now, a team at MIT's Computer Science and Artificial Intelligence Laboratory wants to try and help change that. The magic bullet? TV shows.

Researchers essentially tued computers into couch potatoes, feeding them hundreds of hours of footage from popular TV shows like The Office, Scrubs, and Desperate Housewives. Each clip ends with one of four actions: a hug, a kiss, a high five, or a handshake. The computer's challenge? Predict which one is about to happen.

With the help of a leaing algorithm, the artificially intelligent test subjects were able to predict the correct action 43 percent of the time, compared to 71 percent from human test subjects. Researchers hope to see that number increase as the computers consume more and more video examples and lea to pick up pattes from them.

The long-term goal is to train artificial intelligence to recognize things like danger, injury, or crime as they're happening or even about to happen. Breakthroughs like that are still likely a long way off, but as one of the most successful studies in the field to date, MIT's researchers are optimistic that they can move us closer. At the very least, here's hoping that computers pick up the nuanced complexities of Michael Scott's "that's what she said" jokes.