Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!cornellcs!newsstand.cit.cornell.edu!news.kei.com!newsfeed.internetmci.com!EU.net!Germany.EU.net!news.dfn.de!news.dlr.de!smagt
From: smagt@df.op.dlr.de (Dr. Patrick van der Smagt)
Subject: Re: Guidelines to no. of neurons
X-Nntp-Posting-Host: saar.df.op.dlr.de
Message-ID: <rlogin.818144297@saar>
Sender: news@news.dlr.de
Organization: DLR
References: <49h9mm$t3j@nuscc.nus.sg> <49upf6$kdt@fstgal00.tu-graz.ac.at>
Date: Tue, 5 Dec 1995 06:18:17 GMT
Lines: 40


GILLETTE@JOANNEUM.ADA.AT (Karine Gillette) writes:

>In article <49h9mm$t3j@nuscc.nus.sg>, eng20047@leonis.nus.sg says...
>>
>>        Just like to find out whether there are any general rules, to 
>>determine the no. of neurons within a layer.
>>
>>Liang Moh

>I also tried to answer this question and I definitively can say that I never 
>found any helpful documentation about sizes for neural networks.

>The only thing you can do is to test your network for different sizes and 
>take the better one.

>You won't find more information in papers (though some seem tobe 
>interessant).

Mistake, for sure there are methodologies to pick # neurons.
Pick this up:
   http://www.op.dlr.de/FF-DR/dr_rs/STAFF/patrick_van_der_smagt/home/thesis/
and get chapter 2, "Learning with neural networks".  Read section 2.2,
and then get the papers by Vysniauskas et al, some of which can be obtained
from
	ftp://ftp.fwi.uva.nl/pub/computer_systems/aut_sys/reports/
(the reports *before* 1995 discuss this topic).

In effect, a model of the approximation error is created, as a function
of the number of hidden units and learning samples.  This model can be
used to determine the optimal # for a particular problem.  It requires
some work, but gives a good result (for the problems I tested it on.
See also chapter 5, "Visual feedback in motion".)

Patrick van der Smagt
--
{\let\c=\catcode\c`\{= 12 \c`\}= 12 \c`\@=11 \c`\[=1 \c`]=2 \c`\|=0 \c`\\=12 
 |global|long|def|@comment#1\end{comment}[|end[comment]]
 |global|def|comment[|let|do|@makeother|dospecials|@comment]]
\begin{comment}
