Textbook of AI

1947–, Engineer, neural-network theorist

Also known as: Paul J. Werbos

Paul John Werbos is an American engineer whose 1974 Harvard PhD thesis Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences contained the first derivation of the backpropagation algorithm for training multilayer neural networks, twelve years before its independent rediscovery and popularisation by Rumelhart, Hinton and Williams (1986).

The thesis was largely overlooked at the time, partly because the field was in its post-Perceptrons slump and partly because Werbos's framing was in terms of dynamic feedback systems and generalisations of regression rather than neural networks specifically. He spent his career at the National Science Foundation as a programme director and continued contributions to recurrent networks, adaptive critics in reinforcement learning, and theoretical foundations of intelligence.