The Connectionist Revival

David Rumelhart

Cognitive scientist who, with Geoffrey Hinton and Ronald Williams, popularized backpropagation in 1986 — the algorithm that made multi-layer neural networks trainable.

David Rumelhart (1942–2011) was a cognitive psychologist, not an engineer by training, which shaped how he approached neural networks: as models of how minds might actually process information, not just curve fitters. With Geoffrey Hinton and Ronald Williams, he co-authored the 1986 paper "Learning Representations by Back-propagating Errors" — the paper that made backpropagation the standard way to train multi-layer networks.

The algorithm itself had been derived independently by others earlier (including a 1970 formulation by Seppo Linnainmaa and applications by Paul Werbos in the 1970s), but Rumelhart, Hinton, and Williams's paper is what made it land: clear derivation, working experiments, and a direct answer to the objection Marvin Minsky and Seymour Papert had raised in 1969 — that single-layer perceptrons couldn't represent functions like XOR. Backpropagation showed that adding hidden layers, and training all of them jointly by propagating error gradients backward through the network, fixed exactly that limitation.

Rumelhart also co-edited the two-volume Parallel Distributed Processing (1986), the book that gave the resurgent connectionist movement of the 1980s its name and its intellectual foundation.

See also: Geoffrey Hinton, Frank Rosenblatt

Learn more: Neural Networks & Backprop · Wikipedia: David Rumelhart

Mentioned in

Lessons where this comes up in context.