In 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams published in Nature on learning representations via error backpropagation. The idea had precedents, but their exposition catalyzed a research wave.
At last, multilayer networks could be trained systematically for non-linearly separable functions. Applications emerged in character recognition, control, and prediction.
Compute and data were still scarce; the mass boom would wait for GPUs and ImageNet. Even so, 1986 marks the theoretical restart of modern deep learning.
