Paul Werbos

• 1. Introduction

  ›

  Paul Werbos completed a doctoral dissertation in 1974 that contained the first formal description of training artificial neural networks through backpropagation of errors. For years, the work attracted little attention. He was not a computer scientist. He was an American social scientist, approaching one of computing's central problems from an unusual angle. What drew him to that question? What did he work on as the years passed before recognition arrived? And how did a researcher with interests stretching from neural networks to quantum mechanics and human consciousness come to be seen as a founding figure in machine learning?

• 2. A Social Scientist's Reach

  ›

  Paul John Werbos was born on the 4th of September, 1947. His route into machine learning ran through social science, a starting point that set him apart from most researchers working on artificial intelligence in the early 1970s. His doctoral thesis established one line of pioneering work. A second followed: Werbos was also a pioneer of recurrent neural networks, a separate branch of the discipline. Those two bodies of work placed him at the foundation of multiple areas of machine learning research. His contributions across both areas would eventually factor into how the field formally acknowledged his career.

• 3. IEEE Recognition, 1995

  ›

  In 1995, the Institute of Electrical and Electronics Engineers gave Werbos its Neural Network Pioneer Award. The citation acknowledged the discoveries that had first established him as a machine learning pioneer and his work on a framework called Adaptive Dynamic Programming, one of the other basic neural network learning approaches he had developed. Werbos was also one of the architects of the field's institutional life. He served as one of just three founding presidents of the International Neural Network Society, known as INNS, with each of the three holding a two-year term. Beyond the neural network community, Werbos had also been developing interests in questions that formal scientific institutions rarely address directly, including the nature of consciousness and the interpretation of quantum mechanics.

• 4. Beyond Machine Learning

  ›

  Werbos wrote on quantum mechanics and other areas of physics, engaging with fields well beyond the one that brought him recognition. His interests extended to consciousness, the foundations of physics, and what he described as human potential. These subjects sit at a considerable distance from the technical work of training neural networks. He served as a program director at the National Science Foundation for several years, holding that role until 2015. In the foundations of physics, Werbos found questions as open and contested as anything in machine learning, and he treated both as worthy of the same serious inquiry.