Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!pipex!uknet!cf-cm!C.M.Sully
From: C.M.Sully@cm.cf.ac.uk (Chris Sully)
Subject: Refs wanted: knowledge augmentation, composition of data sets
Message-ID: <1994Dec5.103916.770@cm.cf.ac.uk>
Sender: C.M.Sully@cm.cf.ac.uk (Chris Sully)
Nntp-Posting-Host: topaz.cm.cf.ac.uk
Organization: University of Wales College of Cardiff, Cardiff, WALES, UK.
Date: Mon, 5 Dec 1994 10:39:14 +0000
Lines: 38

I have been training ANNs (using BP) to predict birthweight of children. My
approach thus far has been simple: use all the data available (split into equal
training, validation and testing sets), throwing it all at the network and 
seeing how well it can cope. I have undertaken limited optimisation of network
parameters though performance increases have been limited. The results are
unacceptable and thus I am looking for other ways of improving predictive
performance. Two areas identified:

Knowledge augmentation: there exists some statistically derived information 
concerning, for example, the relationships between the input variables. I'd
like to encorporate this information within training.

Data set manipulation: a segmental analysis based on birthweight reveals
performance degrades at the extremes. Purely due to the frerquency of cases?
The current intent is to manipulate the data to improve overall performance.
A further idea currently being tried is splitting the data into three sets, to
provide three ANN models for low, high, and the majority of birthweights. But
shouldn't the network be able to handle the model as a whole?

Any comments on any of the above, particularly from a statistical viewpoint.

Also pointers to references on any of the above areas would be most welcome.

Cheers.

Chris.

========================================================
Christopher Sully
Ph.D. Student (Neural Networks)  
Department of Computing Mathematics (Room M1.36), 
University of Wales College of Cardiff,   
PO Box 916, CARDIFF CF2 4YN, Wales, UK.   
E-Mail: C.M.Sully@cm.cf.ac.uk
Phone:  +44 (0)222 874000 x6070
Fax:	+44 (0)222 666182
Home:	+44 (0)222 484494
=========================================================
