A situated cortical model exhibiting attention, learning ane memory : implications for cognition

Stratton, Peter Gregory (2001). A situated cortical model exhibiting attention, learning ane memory : implications for cognition PhD Thesis, School of Information Technology and Electrical Engineering, The University of Queensland.

       
Attached Files (Some files may be inaccessible until you login with your UQ eSpace credentials)
Name Description MIMEType Size Downloads
THE16332.pdf Full text application/pdf 10.79MB 2
Author Stratton, Peter Gregory
Thesis Title A situated cortical model exhibiting attention, learning ane memory : implications for cognition
School, Centre or Institute School of Information Technology and Electrical Engineering
Institution The University of Queensland
Publication date 2001
Thesis type PhD Thesis
Supervisor Downs, Tom
Wyeth, Gordon
Total pages 159
Collection year 2002
Language eng
Subjects L
280212 Neural Networks, Genetic Alogrithms and Fuzzy Logic
700199 Computer software and services not elsewhere classified
Formatted abstract The brain's capacity for learning, adaptation and intelligent behaviour is unmatched by any artificial system. By modelling what we know of brain function we hope to gain further understanding of how the brain accomplishes this, as well as to build new systems which exhibit brain-like abilities. There is no doubt that the facilitator of intelligent adaptive behaviour within the brain is the cortex. Despite its myriad functional roles, the cortex of the brain is remarkably uniform in overall structure, even regarding frontal (largely concerned with motor control and planning) and posterior (largely concerned with perception) cortices. This suggests that similar learning processes are at work across all cortical functional domains. This work presents a working cortical model called Cognac - the 'Cognition in Action' neural network - which models the macroscopic structure of the mammalian cortex, including perceptual (posterior) and motor (anterior) processing hierarchies and the many feedforward and feedback intra- and interconnections. Cognac utilises two novel, biologically-plausible learning algorithms, the first for training of the feedforward connections (and called here Biologically-Plausible Hebbian Learning or BP-Hebb), and the second for training of the feedback connections (called here Expectation Learning). BP-Hebb performs self-organised hierarchical feature extraction for the perceptual cortex, and with slight modifications also performs what is called here 'motor feature abstraction' for the motor cortex. Motor feature abstraction is a self-organised hierarchical temporal pattern recognition scheme which is capable of 'chunking' - the recognition of identical repeated subsequences within two or more differing supersequences - the ability for which has never before been demonstrated using any unsupervised learning algorithm. Expectation Learning uses the feedback connections to predict incoming stimuli from top-down representations. Features on the input which have the largest discrepancy between the top-down expectation and bottom-up incoming stimulus equate to those features which are least-well represented by the perceptual hierarchy, and as such are used to guide further learning by the network. Cognac exhibits interesting behaviour as it learns, with no external control, to look over simple images of digitised hand-written numerals. Some thoughts further exploring the mechanisms of cognition are presented in the context of this cortical model.
Keyword Cognition

 
Citation counts: Google Scholar Search Google Scholar
Access Statistics: 78 Abstract Views, 2 File Downloads  -  Detailed Statistics
Created: Fri, 24 Aug 2007, 17:53:32 EST