Mumford 9615444 The problem of vision can be studied from many points of view: from an engineering, system-building problem to a problem in biology and psychology. The investigator seeks to understand vision from the perspective of Bayesian statistics, as one of modeling and learning the `priors' of visual signals and of finding algorithms for efficiently sampling the posterior probability distribution on the unobserved 3D world variables. More specifically, he constructs a larger, more inclusive, probability model for vision based on a pyramid architecture that combines low-level texture segmentation with high-level object recognition. He seeks more effective algorithms for sampling the posterior or making `near-MAP (maximum a posteriori probability)' estimates by combining region growing and genetic algorithm ideas. He also seeks to model the neurophysiology of the primate visual system, using the insights gained by this statistical analysis. He continues the analysis of cortical functioning based on single/multiple-cell recordings, especially with models of cortical areas V1 and M1, based on information multiplexed via synfire-like chains. The investigator believes that logic-based methods for solving problems in artificial intelligence are reaching their limits and that Bayesian statistics is the best tool for modeling how we think. The ability to see, to transform raw measurements of light in a camera or on the retina into a symbolic description of the world being seen, is an extremely challenging problem for these methods. Vision is the perfect test case, especially, for massively parallel computation because of the huge size of its data. In the last 5 years, the technology for processing the data streams produced by video cameras in real time has become available and the challenge of building general purpose computer vision systems may now be within reach. On the theoretical side, the understanding of human and animal vision systems has complemented the engineering of artificial vision systems in a remarkable synergy. In neurophysiology, the possibility of massively parallel neural recordings is now available and may lead to understanding how mammalian cortex uses parallelism so successfully. This project seeks to attack the problem of vision from these two perspectives.
