Backpropagation

David Rumelhart, Geoffrey Hinton, and Ronald Williams publish “Learning Representations by Back-propagating Errors” in Nature, demonstrating that multi-layer neural networks can be trained by propagating error signals backward through the network – adjusting each weight in proportion to its contribution to the output error.

The technique is not entirely new, but the paper’s clear formulation and empirical results make it foundational. It becomes the standard training algorithm for neural networks. The full impact on practical AI systems doesn’t arrive until the 2010s, when increased compute and data make it possible to train the deep architectures that backpropagation enables.

Source: David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams, “Learning Representations by Back-Propagating Errors,” Nature, 323(6088), 1986, 533–536.