Watching your favorite TV shows might be relaxing for you, but for a robot, it’s hard work.
Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory have discovered that they could teach artificially intelligent systems about human interactions – as well as how to predict them – by simply feeding them with TV shows and videos.
MIT researchers have created a new algorithm that can predict human reactions based on six hundred hours of The Big Bang Theory, The Office, and other sitcoms. So far, the artificial intelligence system (AI) can not only identify kisses, high-fives, hugs, and handshakes but also recognize the moments leading up to these behaviors.
After watching all those shows for training, the algorithm was showed a single frame of a brand new video and was requested to predict what type of interaction will take place next. In 43 percent of the cases, the AI got it right.
By comparison, a human answered correctly 71 percent of the time, but the researchers are nevertheless impressed with the algorithm’s performance, given its basic training.
Humans have much more data to compare the videos with – previous situations and personal experience – but the AI could be fed even more TV shows and become better at recognizing human interactions.
According to Carl Vondrick, the lead author of the project, an algorithm has a very different way of understanding what happens in a particular frame of a video. Humans understand social cues, while and AI system understands pixels and patterns.
After analyzing the composition of each frame, the resulting patterns help the system predict what will happen next. “It draws its own conclusions in terms of correlations between the visuals and the eventual action,” explained Vondrick.
Vondrick was joined by two other researchers in his endeavor which has taken place over the past three years. Even though it seems completely useless to have a robot watch reruns as an experiment, there could be more applications for this technology.
An improved version of the AI could be installed in hospitals or in places where it could help prevent injuries. Similarly, smart cameras equipped with the algorithm could analyze video feeds and alert authorities should they discover an emergency.
Image Source: Open the Fridge