Backpropagation

Backpropagation made multilayer training computationally practical: output error was successively converted into updates for every layer. Versions of the method existed earlier, but the 1986 work established it as a general way to learn internal representations rather than only tune an output layer.