The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #94115

Title:Extension of Lau'S 1992 Simulation of Marr'S Theory of the Neocortex
Date: 1994
Abstract:In Marr's classification theory of the neocortex [Marr 70], it is proposed that a two-layer self organizing net can learn to classify events presented to it by reducing the spatial redundancy in the information received. The central idea of the theory is that modifiable synapses between cells in the hidden layer (the so-called codon layer) and the output cell come to represent the conditional probability of pre- and post- synaptic activity co-occuring. Marr proposes that this is a biologically reasonable calculation for cells in the neocortex. The synapses of the lower layer are unmodifiable, except in an initial "setting up" stage, and thus the second layer can legitimately be studied in isolation. Lau's work [Lau 92] showed that the most biologically plausible update rules that could be derived from Marr's ideas perform poorly on classification tasks. In this project some of Lau's findings for the net's second layer were duplicated and an alternative iterative strategy for weight and threshold update implemented. This is based on update rules proposed by Minsky and Papert [Minsky and Papert 69]. The strategy is biologically more plausible than that of Marr's, in part because it does not depend on the computing and retaining of large numbers, and its classification performance is very good, for the event spaces on which it was tested.

[Search These Pages] [DAI Home Page] [Comment]