A modular neural network tech rep

[Eric Ronco]

Dear all,

We completed recently a technical report about modular neural networks.
This one is actualy in the web page of our Centre for System and Control in
the University of Glasgow (see the abstract below):

http://www.mech.gla.ac.uk/Control/reports.html

Ronco, E. and Peter Gawthrop, 1995. Modular Neural Networks: a state of the
art. Tech. rep. CSC-95026, Centre for Systems and Control, faculty of
engineering, Glasgow University.


Any comments about this work would be well come.


Regards,

Eric Ronco


Abstract:

Title: Modular Neural Networks: a state of the art
Author: Eric Ronco and Peter Gawthrop
Keywords: Neural networks; Modularity; Global computation; Local computation; 
Clustering; Function approximation


The use of ``global neural networks'' (as the back propagation neural
network) and ``clustering neural networks'' (as the radial basis
function neural network) leads each other to different advantages and
inconvenients. The combination of the desirable features ot those two
neural ways of computation is achieved by the use of Modular Neural
Networks (MNN). In addition, a considerable advantage can emerge from
the use of such a MNN: an interpreatable and relevant neural
representation about the plant's behaviour. This very desirable
feature for function approximation and especially for control
problems, is what lake other neural models. This feature is so
important that we introduce it as a way to differenciate MNN between
other local computation models.

However, to enable a systematic use of MNN three steps have to be
achieved. First of all, the task has to be decomposed into subtasks,
then the neural modules have to be properly organised considering the
subtasks and finally a way of communication inter-modules has to be
integrated in the whole architecture. We achieved a study of the main
modular applications according to those steps. This study leads to the
main fact that a systematic use of MNN depends on the type of task
considered. The clustering networks and especially the Local Model
Networks can be seen as MNN in the frame of classification or
recognition problems. The Euclidean distance criterion that they apply
to cluster the input space leads to a relevant decomposition according
to the properties of those tasks. But, it is irrelevant to apply such
a criteria in case of function approximation problems. As spatial
clustering seems to be the only existing decomposing method,
therefore, an ``ad hoc'' decomposition and organisation of the
architecture is achieved in case of function approximation. So, to
improve the systematic use of MNN in the framework of function
approximation it is now essential to conceive a method of relevant
task decomposition.