Paul John Werbos submitted his PhD thesis, “Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences,” to Harvard University’s Committee on Applied Mathematics in August 1974. The verified scan of the original thesis title page confirms the author, title, the Doctor of Philosophy degree in the subject of statistics, and the date of August 1974 at Harvard in Cambridge, Massachusetts.
The thesis introduced what Werbos called a method for efficiently computing how the error of a model changes with respect to each of its parameters, working backwards through a chain of calculations. This is the mathematical idea that the neural network community later named backpropagation, the workhorse algorithm for training multilayer networks.
The work received little attention at the time, in part because it appeared during the AI winter when neural network research was out of favor. Its significance was only widely recognized after the 1986 Nature paper by Rumelhart, Hinton, and Williams popularized backpropagation. Werbos is now credited as one of the originators of the technique that underpins essentially all modern deep learning.