DM: Another Delayed Introduction

[Tom Osborn]

Hi all,

I joined the list a couple of months ago. My background includes 
applied
maths/modelling, statistical mechanics, neural networks (theory and 
application),
evolution (we 'invented' Genetic Logic Programming here), some 
machine learning 
and trad AI (UTS was Ross Quinlan's "home" from '83 till about '88), 
and many 
kinds of information theoretic issues to do with processing data sets 
(eg, 
signal processing and financial time series), including use of rank 
statistics.

Some recent consulting involved learning a medical tissue 
classification problem 
for 83 dim input to 12 classes, and was fraught with problems 
(smallish training 
set, some classes were composed of disjoint subclasses, class sizes 
varied a lot,
noise, instrument drift, an imposed utility (cost) on classification, 
and some 
classes were both poorly identified and over dispersed). 
[Significantly, our
model outperformed 18 person months of statisticians' work]

All the data was quantitative, so we ultimately used neural approach 
with lots
of pruning, and took a residual approach to learn some superclass 
separation as 
a filter for subsequent classifications. Our model was later used as 
a guide
to fit a slightly more restricted non-linear regression.

If we go to another stage, there will be a LOT of statistical 
pre-processing
built in.

Tom.
-- 
Dr Tom Osborn,                        "I told the elephants to forget 
it... 
Head - Adaptive Methods Group,                         ...but they 
couldn't",
School of Computing Sciences,
University of Technology, Sydney,                         Zazoo to 
Mufasa.   
PO Box 123 Broadway 2007,  AUSTRALIA.