Newsgroups: comp.ai.neural-nets
From: David@longley.demon.co.uk (David Longley)
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!gatech!swrinde!pipex!peernews.demon.co.uk!news.demon.co.uk!longley.demon.co.uk!David
Subject: NN or Logistic Regression or Multiple Regression
Organization: Myorganisation
Reply-To: David@longley.demon.co.uk
X-Newsreader: Demon Internet Simple News v1.29
Lines: 25
X-Posting-Host: longley.demon.co.uk
Date: Fri, 28 Apr 1995 08:00:52 +0000
Message-ID: <799056052snz@longley.demon.co.uk>
Sender: usenet@demon.co.uk

Over the years there have been some helpful papers on the relation between
neural network technology and  traditional  multiple regression,  logistic
regression and discriminant  function analysis. The advantage of  the more
traditional techniques is  that you get each of your 'input' variables out
as weighted  coefficients,  and your  non-linearities can be coded  in  as
interactions. Furthermore, as the traditional systems  are based  on sound
statistical theory, one can develop prediction (good fitting) models which
don't overfit the data (as indicated by minimal shrinkage  from the const-
ruction sample space to the validation sample space).

Having been a great fan of NN research, I wonder whether it offers *that*
much more *if anything* over the traditional classification &  prediction
prediction technology listed above?  NN's may be good at modelling how we
make judgements about the world (cf. my series 'Fragments  of  Behaviour'
in sci.pstchology, sci.philosophy.tech, sci.stat.edu, etc), but  for  an
AI system worth trust, do we want such systems ultimately?

I think the whole GOFAI vs NN debate has been very fruitful in making us
think about exactly what we are trying to do in Cognitive Science  &  AI
but think that the value of AI lies beyond imitating  human  information
processing. I'd be interested to see a discussion develop on this issue. 

Can we coax MM to put up his views?
-- 
David Longley
