Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!scramble.lm.com!news.math.psu.edu!news.cse.psu.edu!uwm.edu!math.ohio-state.edu!howland.reston.ans.net!swrinde!newsfeed.internetmci.com!in2.uu.net!news.interpath.net!sas!newshost.unx.sas.com!saswss
From: saswss@hotellng.unx.sas.com (Warren Sarle)
Subject: Re: [Q]Comparison between ART and Kohonen's Net?
Originator: saswss@hotellng.unx.sas.com
Sender: news@unx.sas.com (Noter of Newsworthy Events)
Message-ID: <Dr9958.3uG@unx.sas.com>
Date: Sat, 11 May 1996 19:07:56 GMT
X-Nntp-Posting-Host: hotellng.unx.sas.com
References: <4mgpb3$2nr@news.eecs.nwu.edu> <Dr1xu2.L3@unx.sas.com> <4moqd1$2ai@news.eecs.nwu.edu> <4mvkle$nb7@bignews.shef.ac.uk>
Organization: SAS Institute Inc.
Lines: 42


In article <4mvkle$nb7@bignews.shef.ac.uk>, COP95AJW@shef.ac.uk (COP95AJW) writes:
|> ...
|> First of all Kohonen nets are topological whereas ART nets are not. Consider a 
|> pattern which would normaly be attributed to one particular category but which 
|> actually gets attributed to another due to corruption of the data. In Kohonen 
|> nets this alternative category will be topologically close to the ideal one. 
|> As a result the classification produced is similar to the ideal. 

That is true only if the SOM grid is not highly curved. In general,
a misclassified case can show up anywhere on the grid.

|> ...
|> However, Kohonen nets are not gauranteed to produce a stable categorisation, 
|> the vectors representing cluster centres can continue adapting indefinitely. 
|> The usual way of overcomming this is to reduce the degree to which cluster 
|> vectors can adapt as time progresses. As a result, the network becomes unable 
|> to learn any new data. A major feature of ART is its ability to remain 
|> adaptive and yet protect the information already learnt. 

Another way of viewing this is that if the data processed early cause
ART to learn a poor classification, ART will never be able to correct 
itself. Which implies that ART networks do not have the property of
statistical consistency.

|> This is essentially 
|> achieved by performing a test on the category which the network selects to 
|> represent the input. If the input and the vector representing the category are 
|> similar then learning is allowed to take place. If it is not then the current 
|> category is not modified and a new category is searched for. If no existing 
|> category is a good match then a new one is dynamically allocated. This is 
|> another advantage of ART, you do not need to specify the size of the network, 
|> it grows as required.

But you do have to specify the vigilance instead of the number of
clusters. And what rational basis is there for specifying the vigilance?

-- 

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.
