“Garbage in, garbage out.” According to the old adage from computer science, what you get from a computer is no better than what you give it. And it would seem to imply that because computers can’t think for themselves, they can never do anything more sophisticated than what they’ve been explicitly instructed to.
But that last part appears to be no longer true. Neural networks—computing architectures, inspired by the human brain, in which signals are passed among nodes called artificial neurons—have, in recent years, been producing wave after wave of stunning results. (See, for example, page 17 of this issue.) Individual artificial neurons perform only the most elementary of computations. But when brought together in large enough numbers, and when fed on enough training data, they acquire capabilities uncannily reminiscent of human intelligence, seemingly out of nowhere.
Physicists are no strangers to the idea of unexpected phenomena emerging from simpler...
