Information / AI & Robotics

Backpropagation

A method by which multilayer networks learn from their mistakes: the mathematical basis of today’s AI training.

The idea of distributing errors backward through a network with the chain rule appeared with Linnainmaa (1970) and Werbos (1974) and became widely known in 1986 through Rumelhart, Hinton and Williams (Nature). From 1989 Yann LeCun used it to train networks that read ZIP codes. Too little data and computing power kept the method small for almost two decades.