Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news4.ner.bbnplanet.net!cpk-news-feed2.bbnplanet.com!cam-news-hub1.bbnplanet.com!howland.erols.net!agate!news.Stanford.EDU!kithrup.com!cygnus.com!sunnyside.com!nntp-hub3.barrnet.net!voder!nsc!taux01!usenet
From: Marcelo Krygier <marcelo@taux01.nsc.com>
Subject: [Q] control over unsupervised learning ?
X-Nntp-Posting-Host: tasu24
Content-Type: text/plain; charset=us-ascii
Message-ID: <3234084A.41C67EA6@taux01.nsc.com>
Sender: usenet@taux01.nsc.com (Usenet news account)
Content-Transfer-Encoding: 7bit
Organization: National Semiconductor I.C.
Mime-Version: 1.0
Date: Mon, 9 Sep 1996 12:06:34 GMT
X-Mailer: Mozilla 3.0 (X11; I; SunOS 4.1.3 sun4m)
Lines: 34

First of all, let me say : I've already read the FAQ regarding
unsupervised learning.

Still I have some questions, mostly derived from my inexperience
in the field. May be some experienced ANN developers/investigators
already came to conclusions on the following:


1) Given unsupervised nets classify by themselves the input,
   how can I know if I am distracting the net with superfluous
   information ?
   ( I don't really want to classify all that is classifiable
     in the input, but, on the other hand, If I knew how to
     separate the info I want from the junk info, I wouldn't
     be using ANNs. So, I have to give the net all the input
     I've got. )
   ( I assume superfluous information can be divided in two:
     - Unclassifiable superfluous information
       Is the net task to get rid of this kind of info
     - Classifiable superfluous information
       Could this kind of info cause the net to diverge ? )

2) Are ther any means to control the way unsupervised learning nets
   do their classification ?
   I assume the net topology is one method, but, I'd like to know
   what is the exact influence it has on the final classification
   

   
-- 
---------------------------------------------------------------------
 Marcelo Krygier                      @email : marcelo@taux01.nsc.com
                                      Tel    : (972) 9 594210
---------------------------------------------------------------------
