For Facebook AI, machine vision is just the beginning

Yesterday, at Facebook’s day-one F8 keynote, CTO Mike Schroepfer talked about the company’s progress in teaching computers to recognize elements in still images and videos. At the day-two keynote today, he returned to the topic of AI—but emphasized that machine vision is only one aspect.

Schroepfer divided AI into four challenges: perception, understanding and learning, prediction, and planning. He spoke of work Facebook is doing to auto-compose photo captions for the vision impaired, such as labeling a photo of a soccer goalie with “image may contain one or more people playing sports outdoors.”

But then he emphasized how much computers don’t yet understand by flashing a photo of a pepperoni pizza on-screen and asking the audience whether it was a vegetarian pizza. For humans, that’s an easy question, because we know what pepperoni and vegetarianism are. But “computers are confounded by these questions because they don’t understand the world around us,” Schroepfer said.