Paul Utgoff (1985-2008)

Paul Utgoff


Paul E. Utgoff, professor of Computer Science, died Oct. 11, 2008 due to complications from surgery that he elected to undergo as part of his battle against appendiceal cancer. He was 57.

He joined the department as an assistant professor in 1985 and was promoted to associate professor in 1991 and to professor last month.

"Paul was a genuine scholar who never lost sight of the deepest problems in his field, and he was an essential force in helping the department make important decisions. I cannot express our sadness at this loss of a truly valued colleague and friend," said Andrew Barto, the chair of the department at the time of Paul's death.

Utgoff was recognized internationally as a pioneer and leader in the area of machine learning. Among the topics he addressed is the longstanding fundamental problem of how intelligent systems can acquire features, terms and other representational structures that are prerequisite to further learning. He started his work in this area as a Ph.D. student at Rutgers where he developed a system that could modify its own representation language by adding terms to repair faulty generalizations, an accomplishment now recognized as a seminal contribution to the field. His subsequent research demonstrated his skill in selecting key problems and his persistence in pursuing elegant solutions. Three areas of his research are particularly notable for their impact on the field: his work on incremental decision tree induction, multivariate decision trees and problem representation.

In 1993, he designed the Incremental Tree Inducer (ITI) program and a new algorithm called Direct-Metric Tree Inducer (DMTI) made feasible by the ability to restructure trees efficiently. He made his code available online in October 1994 and maintained it ever since. It has been downloaded 6,530 times. His work remains the state of the art in incremental decision tree induction. Returning to his interest in problem representation, his recent research focused on "many-layered learning," following the principle that all learning is simple if the prerequisites are in place.

Utgoff recently wrote, "...the fascination I have had in representation induction since my doctoral dissertation remains at the heart of my research interests. The great majority of my work has been motivated by ... my desire to see machines learn difficult concepts and skills in a sustained manner. ... I have been working on representation repair and modification, term construction, feature construction of various kinds, goal regression to identify essential sets, learning deeply layered knowledge from data or from textbooks, learning from expert choices, hybrid representations, and incremental decision tree induction."

In other recent work, Utgoff was a co-principal investigator on the National Science Foundation-ITR collaborative project "Interactive Software Systems for Expert-Assisted Image Analysis and Classification of Aquatic Particles." In this project, he worked with his department's Computer Vision group and Bigelow Laboratories building systems to perform various classifications of phytoplankton and zooplankton from digital images obtained from ocean studies.

Utgoff served as an action editor for the journal Machine Learning from 1991-95, with membership on the editorial board through 1997. He also served in many organizational roles for the International Conference on Machine Learning (ICML). He chaired and hosted the Tenth ICML held on campus and served as an area chair for the conference in 2000 and 2002. Most recently, he was the area chair of the 2007 European Conference on Machine Learning. In addition, he has held many leadership positions on University and departmental committees, including chairing the Research Council from 2005-07. He is also credited with essential contributions to the overhaul of the undergraduate curriculum that the department expects to unveil next semester.

He earned a bachelor of music in pipe organ performance from the Oberlin College Conservatory of Music in 1979 and a Ph.D. in computer science from Rutgers University in 1984. Following two years as a research scientist at Siemens Research and Technology Laboratories, he joined the Computer Science Department. His lifelong passion for music manifested itself in recent work on how artificial intelligence meets various topics in music, such as music perception, cognition, and understanding, as well as his work as music director at Saint Philip's Episcopal Church in Easthampton. He was an avid backgammon player and as a young man, piloted light airplanes, qualifying as an instrument flight instructor.

He leaves his wife, Karen, and children Naomi Utgoff, Emily VanHassel, and Ariel Utgoff.