David Marr (1945-1980), theoretical neurophysiologist and cognitive scientist, integrated neurophysiological and psychophysical studies with the computational methods of artificial intelligence (AI) to found a new, more powerful approach to understanding biological information-processing systems. The approach has come to redefine the standard for achieving a suitable comprehension of brain structure and function.
Marr was born in Essex, England on 19 January 1945 to Douglas and Madge Marr, and died at age thirty-five of leukemia. After attending Rugby, he entered Trinity College, Cambridge in 1963, obtaining the B.S. degree in mathematics in 1966 with first-class honors. Shortly thereafter, under the guidance of Giles Brindley, he began an intensive year of study on all aspects of brain function, with the intent of focusing on the neural implementation of efficient associative memories. By the end of 1968 he had submitted a dissertation for a title, a fellowship at Trinity College, and was elected. He received two advanced degrees from Trinity in 1971: an M.S. in mathematics and a Ph.D. in theoretical neurophysiology. The first part of his thesis was a theoretical analysis of the cerebellar cortex, published in the Journal of Physiology in 1969. This work was the first detailed theory on any really complex piece of neural machinery, with very specific predictions concerning the input-output relations and details of synaptic modifiabilities during the learning of new motor movements. The essence of this theory is still viable today, and continues to be the benchmark for further advances in understanding cerebellar cortex. Two other papers also appeared before 1971: one on the archicortex, the other on the neocortex, both of which remain landmarks in theoretical neurophysiology.
After obtaining his Ph.D., Marr accepted an appointment at the MRC Laboratory of Molecular Biology under Sydney Brenner and Francis Crick, and he retained an affiliation with MRC until 1976. The thrust of Marr's work changed rather dramatically in 1972, however, following an interdisciplinary workshop on the brain where he met Marvin Minsky and Seymour Papert, who extended an invitation to visit the MIT Artificial Intelligence Laboratory. This visit reinforced Marr's growing conviction that a complete theory of any brain structure must go beyond interpreting anatomical and physiological facts to include an analysis of the task being performed, or more specifically, an understanding of the problem that the information-processing device was "solving." Equally important, he recognized the weakness of theories that appeared explanatory but were not demonstrative. However, demonstrative theories of brain function required large and flexible computing resources such as those available at MIT. Consequently, Marr's initial three-month visit to the AI lab oratory in 1973 became extended, extended again, and then by 1976 became permanent. In 1977 he was appointed to the faculty of the (current) Department of Brain and Cognitive Sciences, becoming full professor at MIT in 1980, while continuing to hold his AI Lab appointment.
Marr's years at the AI laboratory were incredibly productive. The goal was to understand both the competence as well as the performance of a biological information-processing system. Although some studies of movement continued, the primary thrust was understanding the mammalian visual system. Although small, Marr's remarkably talented group included Tomaso Poggio and Shimon Ullman. They shared Marr's conviction that explanations of a complex system are found at several levels. A complete study should address at least three levels: an analysis of the competence, the design of an algorithm and choice of representation, and an implementation. In order to provide a coherent framework for organizing and attacking visual problems -- more properly, problems of "seeing" -- Marr proposed separating perceptual processing tasks into three main stages: a primal sketch, where properties of the image are made explicit; a 2-D sketch, which is a viewer-centered representation of the surface geometry and surface properties; and lastly a 3-D model representation, which is object centered rather than viewer based. His 1982 book, Vision: A Computational Investigation into the Human Representation and Processing of Visual Information, published posthumously, summarizes these ideas, as well as the contributions to image processing, grouping, color, stereopsis, motion, surface geometry, TEXTURE, SHAPE PERCEPTION, and OBJECT RECOGNITION. Many of these contributions had appeared by the late 1970s, and in recognition of this work, Marr received in 1979 the Computers and Thought Award from the International Joint Conference on Artificial Intelligence. More recently, "best paper" awards in the name of David Marr have been created by the International Conference on Computer Vision and by the Cognitive Science Society.
As a person, David Marr was charismatic and inspiring. He was both fun and brilliant, enjoying the adventures of understanding brain structure and function, as well as life itself. He communicated his pleasure in clear and compelling ways, not only in personal exchanges and in his writing but also in music. He was an accomplished clarinetist. His early death was a great loss but even in his short life, he was able to bring together two previously diverse disciplines, set a higher standard for explanatory understanding, and open new doors to unraveling the mysteries of mind and brain.
Marr, D. (1982). Vision: A Computational Investigation into the Human Representation and Processing of Visual Information. San Francisco: W. H. Freeman.
Vaina, L. (1991). From Retina to Cortex: Selected Papers of David Marr. Boston: Birkhäuser.
Copyright © 1999 Massachusetts Institute of Technology