S. Geman; "Four Arguments in Support of Hierarchy in Computer Vision"

Faculty of Engineering and Natural Sciences
FENS SEMINARS

Four Arguments in Support of Hierarchy in Computer Vision
Prof. Stuart Geman, Brown University

I will give evidence from the neural sciences, the cognitive sciences, and the mathematical sciences that supports hierarchies of reusable parts as the correct framework for artificial intelligence, in general, and computer vision, in particular.  The neural science evidence mostly concerns the anatomy and physiology of the ventral visual pathway.  The cognitive science evidence is mostly from computational linguistics. And the mathematical evidence is in the form of two theorems that address, respectively, the computation of scene interpretations, and ROC performance in the face of the highly structured "noise" that constitutes "background" in natural scenes.

Prof. Geman has a BS with Highest Honors in Physics from University of Michigan,
National Boards I from Dartmouth Medical School, Masters in Physiology from Dartmouth College, and a  Ph.D. in Applied Mathematics from Massachusetts Institute of Technology.

He is currently a professor in the Division of Applied Mathematics, Brown University.

His honors include: Presidential Young Investigator Award 1984-1989, Fellow of Institute of Mathematical Statistics, Elected member in International Statistical Institute, 1997 Rietz Lecture, Institute of Mathematical Statistics, Named Chair: James Manning Professor, Division of Applied Mathematics, Brown University, 1997-2001 Hotelling Memorial Lectures, University of North Carolina and Philip J. Bray Award for Excellence in Teaching in the Physical Sciences, Brown University in 2001

His Work in Progress include: * Compositional Vision. Development of theory and algorithms for a "composition machine," a scene interpretation system based upon a hierarchy of composite structures. On-going experiments with face recognition and license plate reading.
* Neural Representations. Compositionality makes strong predictions about the nature of the neural code. In particular, it would appear that there must be mechanisms for rapidly and reversibly binding otherwise-uncorrelated spatio-temporal patterns of neural activities.
A compelling candidate is the fine-temporal (millisecond time windows) structure of neural
discharges. Statistical methods are being devised for identifying this structure in stable
multi-unit recordings. Phsyiological experiments are being performed to test for the
predicted discharge patterns.

May 5, 2005, 15:00, G032