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

MSc Thesis #9540

Title:Properties of Convergence Zone Associative Memory: an Experimental Investigation
Date: 1995
Abstract:Moll et al (1990) propose an implementation of a type of associative memory that was inspired by the work of Damasio (1989), who has put forward a new type of theoretical cognitive model that attempts to solve the binding problem, and account for biological data that other more traditional networks (such as the Hopfield net and back-propagation nets) fail to incorporate. this model relies on the notions of feature maps, convergence zones and feedback linking, omits traditional - and non-biologically plausible - features such as dynamic or recurrent activity, and shows promise as a realistic model of human brain processes. Moll et al take an idealised and simplified version of the model, and show analytically that with biologically plausible dimensions it has human-like memory capacity. In this study we have implemented Moll et al's network for relatively small network sizes, and performed experiments to gather performance data (by recall of random patterns with error measurement) over various parameter changes. Our results show that the network has a predictable behaviour, so that performance can be predicted as parameter changes are made, but that it has a poor response to noisy patterns, and a "catastrophic" response to certain parameter changes. Overall the result suggest that further work would be worth-while, particularly in experiments with larger networks, and in the development of lower-level and more biologically plausible implementation for connectivity.

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