アブストラクト

Most perceptual and motor tasks performed by the central nervous system are essentially probabilistic. They infer a few important but ``hidden'' properties, such as object shapes, time to collision, or motor commands, from many noisy, local and ambiguous cues. We show that leaky integrate and fire neurons with adaption can be interpreted as Bayesian integrators accumulating evidence about events in the external world or the body, and communicating to other neurons their certainties about these events. Networks of these neurons can perform exact and approximate bayesian inference, using standard models of synaptic transmission. We explore the implications of this model for dendritic computations, the origin and meaning of irregular inter-spike intervals and neural synchrony, population coding, neural integration and decision making and, at a more global level, modularity as a tractability/accuracy tradoff.

I will present a theoretical proposal that primary visual cortex, V1, generates a bottom up saliency map to direct visual attention. The physiological and psychophysical data that have motivated this theoretical proposal will be described. This proposal is demonstrated by showing that V1's output, as simulated in a biologically based model, can explain or predict the ease of visual search and segmentation tasks. Experimentally testable predictions are provided by this proposal, and some of them have already been tested and confirmed. Psychophysical experiments to test the difference between this and alternative proposals of visual saliency maps will be presented.

Part of the materials of this presentation can be seen in the article;
"A saliency map in primary visual cortex", Zhaoping Li, Trends in Cognitive Sciences, Vol 6, p.9-16, 2002,

＊Downloadable from: http://www.gatsby.ucl.ac.uk/￣zhaoping/prints/Fortics.pdf