Werbos's dissertation
Era 2 · AI is born (1943–1979) › Neural networks and learning
- Lineage
- Neurons and networks
- Kind
- Theory
Who
Harvard University
What happened
At Harvard, Paul Werbos submits the doctoral dissertation "Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences," on computing derivatives efficiently in multistage models.
Why it mattered
It brings backward computation of derivatives to prediction models and proposes using it to train neural networks. The explicit application to neural networks is published in 1982.
The full story
Werbos wanted to build prediction models for the social sciences and was drawn to the idea of a network that learns the way the brain does. The dissertation developed what Werbos called ordered derivatives: a systematic way of computing how each parameter of a many-stage system influences its final result, by working backward through the system.
The dissertation went almost unnoticed. It arrived when connectionism was at its lowest ebb and few researchers worked on neural networks. In 1982 Werbos published a paper on nonlinear sensitivity analysis that applies the method to neural networks the way it is used today.
Priority is disputed. Werbos and much of the literature place the first description of training networks by backpropagating errors in the 1974 dissertation, and in 1995 the IEEE gave Werbos its Neural Network Pioneer Award for it. Schmidhuber argues that the application to neural networks appears in 1982 and not yet in the dissertation.
What is not disputed is that the method did not spread until 1986, when Rumelhart, Hinton, and Williams showed that a network trained this way learned useful representations. In 1994 Werbos collected the dissertation and other writings in The Roots of Backpropagation.
Videos
- Paul Werbos (opens in a new tab) (IEEE CIS History)
Leads to
Backpropagation by Rumelhart, Hinton, and Williams (October 1986)
Independent convergence
Rumelhart arrived at the method independently around 1982, and the 1986 paper does not cite the earlier work
Sources
Scholarship and context
Further reading
How to cite
“Werbos's dissertation.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/werbos-thesis
This entry was last reviewed on November 13, 2026.