Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!gatech!concert!sas!mozart.unx.sas.com!saswss
From: saswss@hotellng.unx.sas.com (Warren Sarle)
Subject: Re: Refs wanted: knowledge augmentation, composition of data sets
Originator: saswss@hotellng.unx.sas.com
Sender: news@unx.sas.com (Noter of Newsworthy Events)
Message-ID: <D0M8HJ.HED@unx.sas.com>
Date: Sat, 10 Dec 1994 22:15:19 GMT
References:  <1994Dec5.103916.770@cm.cf.ac.uk>
Nntp-Posting-Host: hotellng.unx.sas.com
Organization: SAS Institute Inc.
Lines: 31


In article <1994Dec5.103916.770@cm.cf.ac.uk>, C.M.Sully@cm.cf.ac.uk (Chris Sully) writes:
|> 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. ...
|>
|> ... a segmental analysis based on birthweight reveals
|> performance degrades at the extremes. Purely due to the frerquency of cases?

Generalization always degrades as you move away from the bulk of the
data. This happens even if the training cases are uniformly distributed
over the region of interest. For linear models, the proof is simple
and can be found in any textbook on regression. For nonlinear models,
the effect is similar but more complicated.

|> 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

But if you are trying to predict birthweight, how would you know which
set to put a case in?

|> shouldn't the network be able to handle the model as a whole?

Yes, if you have enough hidden units.

-- 

Warren S. Sarle       SAS Institute Inc.   The opinions expressed here
saswss@unx.sas.com    SAS Campus Drive     are mine and not necessarily
(919) 677-8000        Cary, NC 27513, USA  those of SAS Institute.
