Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!godot.cc.duq.edu!newsgate.duke.edu!interpath!news.interpath.net!sas!newshost.unx.sas.com!saswss
From: saswss@hotellng.unx.sas.com (Warren Sarle)
Subject: Re: Adding noise to inputs
Originator: saswss@hotellng.unx.sas.com
Sender: news@unx.sas.com (Noter of Newsworthy Events)
Message-ID: <Dv637u.FC2@unx.sas.com>
Date: Fri, 26 Jul 1996 20:21:29 GMT
X-Nntp-Posting-Host: hotellng.unx.sas.com
References: <4svoie$7a6@news1.panix.com> <4t9a6f$nqj@kelly.iaonline.com>
Organization: SAS Institute Inc.
Lines: 27


In article <4t9a6f$nqj@kelly.iaonline.com>, dmce@iaonline.com (David McEldowney) writes:
|> Jeffrey O. Katz <off@panix.com> wrote:
|> >I am trying to predict financial time series using NNs, and am
|> >curious as to what experiences others have had when adding
|> >noise to the inputs during training.  
|> 
|> >Does noise yield better generalization? 
|> 
|> It has been my experience that financial time series already contain
|> plenty of noise -- providing you have enough cases.

Before the dash, I thought this reply was a joke. After the dash, I'm
not so sure. Anyhow, it is vitally important not to confuse the effects
of jitter (artificial noise added to inputs) with the effects of noise
in the actual training data--the latter does _not_ improve
generalization! See:
   "How does noise affect generalization?"
   "What is jitter? (Training with noise)"
in the Neural Network FAQ, part 3 of 7: Generalization, 
ftp://ftp.sas.com/pub/neural/FAQ3.html

-- 

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.
