Off-campus UMass Amherst users: To download dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.
Non-UMass Amherst users, please click the view more button below to purchase a copy of this dissertation from Proquest.
(Some titles may also be available free of charge in our Open Access Dissertation Collection, so please check there first.)
PAN: A distributed memory model of cortical knowledge representation in the mammalian primary visual cortex
Abstract
The role of the single visual-cortical neuron in vision is not well understood. Three proposed functions for these cells are line-and-edge feature detectors, spatial frequency filters and matching-to-experience templates. Each of these proposals entail specific predictions about the spatial structure of the receptive fields that single neurons develop. The Primitive Acquisition Network (PAN) was implemented in an effort to account for empirical data collected in support of these three putative functions. PAN was designed in a Hebbian framework that favors memory models but makes no assumptions about the spatial structure of the receptive fields it develops. Any structure that PAN 'neurons' develop results from the patterns of input the network experiences and structural structural parameters that determine the maximal range of connections within the network. After training with sinusoidal gratings, random noise, or natural forest scenes PAN 'neurons' show orientation and spatial frequency selectivity. Such training also causes these networks to self-organize bands or clusters of adjacent 'neurons' into groups with the same or similar response selectivities. Classification of these 'neurons' based on the spatial structure of their receptive fields differs from that based on the selectivity of their response and is also organized into bands or clusters. The simulation results demonstrate that the model neuron implemented in PAN is adequate to describe the major empirical observations of cortical simple cell properties and columnar organization and that a memory model of cortex is capable of accounting for these observations. Investigations of the information processing capabilities of PAN networks in figure-ground separation and image classification tasks show that the representations they develop store and make available both local and global information extracted from the image on the network retina. Parallel studies of the spatial structure (at the scale of physiological RFs) of the natural images used to train the networks reveal complex micro-feature from which PAN's representation could be constructed. Taken together these results suggest that the representation of the visual world in primary visual cortex can contain parallel representations using complex primitives derived from natural scenes alongside orientation and spatial frequency representations.
Subject Area
Neurosciences|Computer science|Psychobiology
Recommended Citation
Grasso, Frank William, "PAN: A distributed memory model of cortical knowledge representation in the mammalian primary visual cortex" (1994). Doctoral Dissertations Available from Proquest. AAI9420631.
https://scholarworks.umass.edu/dissertations/AAI9420631