Thesis on Adaptive Neural Growth

[Sam Joseph]

Dear Connectionists,

For your information, the phd thesis entitled:

"Theories of Adaptive Neural Growth"

by Sam Joseph, Edinburgh University 1998

is available to download in gzipped postscript format at:

http://www.cns.ed.ac.uk/students/sam/work.html

On the same web page is a shorter paper entitled "Feature Generation
Mechanism" that summarises some of the main results from the phd thesis,
regarding the construction of boolean neural networks that takes
inspiration from the relationship between electrical activity and outgrowth
in biological neurons.

All comments gratefully received, particularly on the shorter (draft)
paper, which is designed to be in a more digestable format.  Full abstract
of thesis follows at end of this mail.

Sincerely

Sam Joseph
email to: sam at cns.ed.ac.uk 


****************

Abstract of "Theories of Adaptive Neural Growth"

When interpreting the results of experiments that investigate
biological development, one is faced with a wealth of data.  Producing
a model of such development must always involve some degree of
abstraction.  The appropriate level of abstraction and the importance
of particular experimental evidence is determined by one's modelling
objective.  Models may potentially be motivated by one of two
complementary aims:

1. To understand how biological neurons achieve their mature
interconnectivity.

2. To improve the learning ability of artificial neural nets (ANNs)
by taking inspiration from the growth of biological nervous systems.
 
These aims are exemplified in the thesis by two simulated neural
models.  The first is a model of neuromuscular development that places
an emphasis on achieving biological plausibility.  The second is a
platform for modifying the connectivity of artificial neural networks.
This feature generation mechanism (FGM) platform supports a variety of
growth procedures that are inspired by evidence from biological
development.

The model of development at the mammalian neuromuscular junction (NMJ)
focusses on the achievement of single innervation, and is an extension of
the existing dual constraint model (DCM).  The identities of the molecules
involved in the DCM are proposed and the framework is
adjusted accordingly.  This extension allows a variety of developmental
phenomena to be replicated, including the presence of both
activity-dependent \& independent competition between terminals. A further
framework is established that provides a potential explanation for the
paradoxical results of synaptic interaction under focal blockade
conditions.

The FGM model concerns feed-forward ANNs and attempts to improve their
unsupervised pattern recognition ability.  Different FGMs consist of
functions that in the right combination produce connectivity patterns
that maximise the average Shannon information provided by the output
of individual nodes.  They also allow the network to construct partial
input features which form an any-of-N representation of a given input
pattern.  FGM networks are shown to outperform other straightforward
unsupervised ANNs in trials on simple data sets.  More demanding tests
are performed indicating that an FGM net with Boolean weights
outperforms a competitive network using continuous weights.  The slight
superiority of the FGM performance is achieved with a third of the
free parameters of the competitive net.  The nature of the FGM induced
partial connectivity implies that these networks would scale up to
larger problems more easily their fully connected counterparts.