In this talk, we will entertain the proposal that the visual cortex can be conceptualized as a hierarchical Bayesian network. Such a hierarchical network can provide the needed flexibility and power to model with the complexity of the natural world and to generate predictions during learning and inference. I will discuss neurophysiological evidence and constraints that we have learned about the visual cortex from this perspective and, in particular, the functional role of recurrent feedback in visual computation. I will present findings that provide insight as to how the brain is shaped by the statistical structures of the natural environments and how the real neural networks perform computation using such priors to solve 3D perception problem. These findings argue that a class of graphical models provides a viable theoretical framework for us to conceptualize the computation in the visual cortex.
讲演者简介
Dr. Tai Sing Lee is on the faculty of the Computer Science Department and the Center for the Neural Basis of Cognition at Carnegie Mellon University, Pittsburgh, U.S.A. He was born in Guangzhou and grew up in Hong Kong. He obtained his S.B., M.S., and Ph.D. all from Harvard University. He was also trained in medical sciences, medical physics and medical engineering in the Harvard-MIT Division of Health Sciences and Technology during his doctoral training. He did his postdoctoral training in primate neurophysiology in MIT’s Brain and Cognitive Science before being recruited to CMU to start an interdisciplinary research laboratory to study the computational and neural basis of visual perception. He is well known for his works with David Mumford on the role of recurrent feedback in the hierarchical visual system, and has received the ICCV Helmholtz award for his work with Song Chun Zhu and Alan Yuille, and the CAREER award from the National Science Foundation (USA).
