Backpropagation
Algorithm that trains neural networks by passing the output error backwards through the layers and computing how much each weight contributed to it; the basis of training by gradient descent.
Backpropagation is a training algorithm for neural networks that passes the output error backwards through the layers and computes how much each weight contributed to that error. It combines a forward pass, an error measure (typically half the sum of squared output deviations), and a backward pass using the chain rule so weights move by gradient descent.
The procedure was described in 1986 by David Rumelhart, Ronald Williams and Geoffrey Hinton in Nature, building on earlier work by others and showing that networks with hidden layers can learn tasks a single perceptron cannot. It later fine-tuned pretrained stacks in deep belief networks (2006). See the article and timeline entry.