Thursday, April 29, 2004

12:00 noon

Location: BioX/Clark Center, Room S360, 318 Campus Drive, Stanford University

 

 

Title: "Inferential Processes and the Architecture of Visual Cortex"

 

 

Malcolm Young

Provost of Science, Agriculture & Engineering

University of Newcastle

Newcastle upon Tyne, England

 

Abstact:

A striking aspect of vision is that observers do not see what their eyes tell their brains (e.g. Young 1994). This mismatch is clearly evident in the fact that apparently detailed coloured texture is seen in extra-foveal vision, where such information is very modest or absent in the input from the retina, and in other cases, where information at one location and time influences the likelihood that information will be present at a different location, and at a later time. These considerations, and detailed observations of what neuroanatomy can tell us of the processing architecture in visual cortex, suggest that the visual system may undertake an inferential style of computation, rather than analysing bottom-up information from the eyes, which latter model is in the ascendant in visual neurophysiology.

 

We have been exploring the different predictions that the inference and traditional models make, when the prior probabilities of stimuli are manipulated, as they are during normal vision. In one paradigm, we use spatiotemporal sequences of oriented bars to provide the visual system with "reason to believe" that a bar of particular orientation will be presented at a particular time. We then vary the congruence, incongruence and presence of the bar in the RF in relation to these prior expectations. The results so far are striking: V1 neurons are often more modulated by these prior and distant events - in many cases more than 400ms, and 6-10 degrees away from their classical RF - than they are by events within their RF. Similar results relate to an experimental paradigm that tests whether V1 neurons really suffer from the "aperture problem". The interaction between priors and likelihood functions is well fitted by a Bayesian model, and only poorly fitted by the traditional view that V1 cells are local filters of one kind or another. Similarly, Bayesian models fit better than mismatch models, such as that of Rao and Ballard. These results suggest that V1 neurons are signaling the posterior probability of visual stimuli, on the basis both of prior knowledge about the world and information from the eye, effectively imputing feature constellations to the visual world, rather than simply analysing local contours.