Despite recent progress in artificial intelligence, humans and animals vastly surpass machine agents in their ability to quickly learn about their environment. While humans generalize to new concepts from small numbers of examples, state-of-the-art artificial neural networks still require huge amounts of supervision. We hypothesize that humans benefit from such data-efficiency because their internal representations support a much wider set tasks (such as planning and decision-making) which often require making predictions about future events. Using the curvature of natural videos as a measure of predictability, we find that human perceptual representations are indeed more predictable than their inputs, whereas current deep neural networks are not. Conversely, by optimizing neural networks for an information-theoretic measure of predictability, we arrive at artificial classifiers whose data-efficiency greatly surpasses that of purely supervised ones. Learning predictable representations may therefore enable artificial systems to perceive the world in a manner that is closer to biological ones.