Computers Already Learn From Us. But Can They Teach Themselves? - The New York Times

Computers Already Learn From Us. But Can They Teach Themselves? - The New York Times

Computer, teach thyself

Further progress in artificial intelligence will require having machines teach themselves, computer scientists say. Machine learning has made impressive breakthroughs in the last decade, in large part through ‘supervised’ training: algorithms go through vast reams of data that have been previously labelled by humans (for example, about pictures being of cats rather than dogs). But ‘self-supervised’ programs (such as AlphaGo Zero) and ‘unsupervised’ techniques can sidestep human labelling, and can perform better at tasks such as discovering new laws of nature or coordinating motion. After all, babies learn how to walk by stumbling around, not by being told which moves are correct. “Humans don’t need that much supervision,” says machine-learning pioneer Yoshua Bengio.

New York Times | 6 min read