Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!EU.net!nova.puug.pt!news.rccn.net!master.di.fc.ul.pt!news.inesc.pt!animal.inescn.pt!bart.inescn.pt!jcardoso
From: jcardoso@bart.inescn.pt (Joao Cardoso)
Subject: New Q: How does training with noise effect ... ?
Message-ID: <D7wK65.Gto@animal.inescn.pt>
Sender: news@animal.inescn.pt (USENET News System)
Nntp-Posting-Host: bart.inescn.pt
Organization: INESC-Porto, Portugal
References: <3nllru$oqu@jaws.cs.hmc.edu> <3nnk5g$5rg@oslo.uni-paderborn.de> <3norg8$buq@jaws.cs.hmc.edu> <3nseg1$si7@news.iastate.edu>
Date: Mon, 1 May 1995 13:58:52 GMT
Lines: 27


Hi,

In real-world problems, training data come from measurements, which have
an associated uncertainty, either due to instrumentation, inherent phenomena noise, etc.
Usually one considers the uncertainty to have a normal distribution about the
'real' true value, and one makes several measurements around the same 'working point', as to
diminish the measuring error, or at least to diminish the uncertainty.

If we know the measuring error of each measurement, can't we add noise to each pattern
with an equal variance as the know measuring error?

If we have  not enough examples to train a ANN, we can get more patterns with this
technique; the question is if there is any advantage with this. Of course, the input
space is not more covered than before, but at least we have more _representative_
samples.


Joao Cardoso
INESC
R. Jose Falcao 110
4000 PORTO
PORTUGAL

e-mail: jcardoso@bart.inescn.pt
tel:	+ 351 2 2094300
fax:	+ 351 2 2008487	
