Neural computation of statistical structure

Abstract

The present thesis comprises four distinct approaches to the problem of how the visual system encodes complex visual structure within the cerebral cortex. In the present context complexity refers to visual structure defined by spatial relationships (correlations) between multiple points (greater than 3) within an image (higher-order spatial correlations). Such spatial structure, for example corners or intersections, constitutes the visually meaningful elements of an image. While the encoding of these spatial correlations is of central concern, some findings may point to the types of computations performed within cortical (pyramidal) neurons in general. The four approaches fundamentally differ in the scales at which visual encoding has been explored. In Chapter 2 encoding has been studied at the behavioural (psychophysical) level. Findings in Chapter 2 suggest that the range of complex spatial relationships that we are capable of detecting is highly restricted relative to the set of all possible spatial relationships. Further, such detection may be accounted for by relatively few statistically independent driving mechanisms. Chapter 3 then considers the above results within the context of information theory, exploiting measures of image information to identify both the spatial and computational limits of structure encoding mechanisms. These findings suggest that correlations up to 9th order may be locally computed over highly restricted spatial domains. Chapter 4 then studies encoding at the level of neurotransmitters, their receptors and subcellular structures such as dendrites and dendritic spines. The results in this section demonstrate how such subcellular components confer individual neurons (reconstructed from morphological data) with the computational power to detect relationships between the activities of three or more synapses. The detection of these higher-order correlations by individual neurons suggests that these fundamental units of cognition are capable of performing computations significantly more complex than is generally supposed. Such computations confer neurons with the ability to detect meaningful complex relationships within patterns of synaptic input. Finally, by approaching the problem of encoding from the scale of human visual ecology, Chapter 5 attempts to place the above results within a broader theoretical context. Work in this section further exploits information theory, suggesting that the encoding properties of some visual neurons are adapted to efficiently encode visual structure at the centre of gaze. Show more