Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!ceylon!wizard.pn.com!Germany.EU.net!howland.reston.ans.net!gatech!concert!sas!mozart.unx.sas.com!saswss
From: saswss@hotellng.unx.sas.com (Warren Sarle)
Subject: Re: normalization in competitive learning
Originator: saswss@hotellng.unx.sas.com
Sender: news@unx.sas.com (Noter of Newsworthy Events)
Message-ID: <CzH4uI.4AE@unx.sas.com>
Date: Fri, 18 Nov 1994 17:35:06 GMT
References:  <3acsu8$ng6@lorne.stir.ac.uk>
Nntp-Posting-Host: hotellng.unx.sas.com
Organization: SAS Institute Inc.
Keywords: competetive learning, Kohonen
Lines: 35


In article <3acsu8$ng6@lorne.stir.ac.uk>, psi@cs.stir.ac.uk (Peter Sincak (Guest)) writes:
|> 1. COncerning a simple competiteve NN - the inputs should be normalized.
|> Well - it is note very clear to me why - but anyway - lets normalize it.
|> Whats gonna happen if I get on input 2 vectors with diffrent size but
|> same directions ? Like (1,2) and (40,80) - these are absolutely in
|> different places in feature space !!

You would not have these problems if you used ordinary clustering
methods such as various k-means methods available in numerous
statistical software products.

|> 2. Concerning Kohonen - have I normalize input to a K - NN ? I was told that no .
|> Why - as far as I percent Kohonen - it is only generalized competitive
|> learning and a diference is in neiborhood funcion ? So why - I do not
|> need to normalize ? The results of Kohonen processing are hidden in
|> weight vectors - I mean centers of clusters - is that right ?

When talking about Kohonen networks, it would be a big help if people
would specify which type of Kohonen network they mean. There are at
least three:

 1. Plain vanilla (not a technical term--is there one?) Kohonen network
    that does simple clustering/AVQ

 2. Self-organizing maps, which are a form of dimensionality reduction

 3. Learning vector quantization, which is a supervised (unlike the
    above two) classification method

-- 

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.
