In this thesis we begin to articulate a computational theory that provides a neuronal-leveldescription of the macro-level function of the visual system, and we propose a particularinstantiation of it that is based on a biologically plausible large scale integration of dynamicalsystems. Inspired by the columnar organization of the visual cortex and the brain's immenseassociative power, we suggest a framework composed of a very large number of small autoassociativeneural networks that create network assemblies by hetero-associating their internalstates. The constituent networks, as abstractions of cortical columns that become selective tolocal receptive fields, form associative memories of the respective image patches. At the sametime, these networks are linked with each other and form lateral and hierarchical assembliesthat become selective to complex configurations of whole visual percepts. We believe that thiscognitive fusion of the dynamical states at various scales, is the basis for high level visualrecognition. In order to face the immense computational demands of our approach, we modelthe proposed framework as a parallel distributed system and we conduct our experiments with ahigh performance computing cluster. Experimental results demonstrate that the proposedframework can model a variety of behavioral data related to visual cognition in humans. Theinherent use of dynamics, in combination with the versatile associativity, make it very suitablefor modeling the temporal aspects of visual recognition and provide testable predictionsregarding the interplay of time with the neuronal connectivity.
"Visual Recognition with a Large Scale Network of Dynamical Systems"
Cognitive Sciences Theses and Dissertations.
Brown Digital Repository. Brown University Library.