# David Rumelhart > 1942–2011 · Cognitive Scientist > > **Recorded contribution:** Popularized error backpropagation for learning internal representations with Geoffrey Hinton and Ronald Williams, following earlier derivations; parallel distributed processing ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation David Rumelhart (1942–2011) was a cognitive scientist who developed influential models of reading, schemas, and parallel distributed processing. With Geoffrey Hinton and Ronald Williams, he published the 1986 paper that demonstrated error backpropagation as an effective method for learning internal representations in multilayer networks. Backpropagation had important earlier derivations and applications, including work by Seppo Linnainmaa, Paul Werbos, and others; the trio popularized and experimentally legitimized it within connectionist AI rather than inventing it from nothing. Rumelhart's broader contribution was treating cognition as distributed activation and learned representations instead of a sequence of explicit symbolic rules. ## 2. The problem inherited Single-layer perceptrons could not learn nonlinear features such as XOR, while multilayer networks lacked a widely understood efficient rule for assigning output error to hidden weights. ## 3. The central contribution Rumelhart, Hinton, and Williams showed how reverse-mode differentiation trains multilayer networks end to end and used it to learn distributed hidden representations. ## 4. Reconstruct the mechanism 1. Run an input forward through weighted layers and differentiable nonlinearities to produce a prediction and scalar loss. 2. Differentiate the loss at the output and propagate gradients backward through each operation using the chain rule. 3. Compute every weight's contribution from its incoming activation and downstream error signal. 4. Update weights opposite the gradient and repeat across examples until representations and predictions improve. ## 5. What changed downstream - Backpropagation became the central training mechanism for multilayer neural networks. - Parallel Distributed Processing revitalized connectionist cognitive modeling. - Learned hidden features became an alternative to fully hand-engineered symbolic representations. ## 6. Attribution, limits, and uncertainty - Hinton and Williams are coauthors, and earlier chain-rule and reverse-mode work makes 'invented backprop' inaccurate. - Training can encounter poor conditioning, vanishing gradients, overfitting, shortcut learning, and data bias. - A network's predictive success does not establish that its hidden representations reproduce human cognitive mechanisms. ## 7. Reconstruction lab Compute one full forward and backward pass for a two-input, two-hidden-unit XOR network by hand. Check each gradient with finite differences, then train it and inspect how different random initializations produce different hidden representations. ## 8. Evidence trail - [Learning Representations by Back-Propagating Errors](https://doi.org/10.1038/323533a0) — Nature - [David Rumelhart](https://en.wikipedia.org/wiki/David_Rumelhart) — Wikimedia Foundation --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*