Junmei Zhu
Assistant Professor, Department of Computer Science
University of Memphis
http://www.cs.memphis.edu/~jzhu/
Thursday, April 6
3 :10-4:00 p.m.
1065 Kemper - refreshments to follow in 1131 Kemper Hall
Invariant visual recognition in light of brain organization
Computers have transformed our lives, yet they have achieved only limited progress in visual perception, a task that humans perform effortlessly. A theory of general organizing principles is the key to understanding how the brain works as well as building intelligent machines. In this talk, I present my work on brain state organization in the context of visual recognition with invariance to transformation (position, scale, orientation) and deformation. At the heart of this process is the correspondence problem which establishes a set of links connecting neural units projected from the same scene points. I will describe a self-organizing dynamical system of the links that is based on direct and transformation-specific link interactions, mediated by high-order links called maplets. The fast convergence and recognition performance of this system have made it a foundation on which several projects on neural implementation of invariant recognition are built. I will also describe a more natural implementation of maplets which models task-adapted link interactions and their learning directly, leading to a system that can be analyzed more easily. I will discuss the essential aspects of brain organization emerged from these model systems.