Path: utzoo!attcan!uunet!wuarchive!gem.mps.ohio-state.edu!uakari.primate.wisc.edu!polyslo!unmvax!ariel!wayback.unm.edu!bill
From: bill@wayback.unm.edu (william horne)
Newsgroups: comp.ai.neural-nets
Subject: Re: Representational NN
Message-ID: <708@ariel.unm.edu>
Date: 6 Oct 89 22:53:44 GMT
References: <6403@watdcsu.waterloo.edu> <485@spinifex.eecs.unsw.oz> <1678@cs.yale.edu>
Sender: news@ariel.unm.edu
Reply-To: bill@wayback.unm.edu (william horne)
Distribution: comp
Organization: University of New Mexico, Albuquerque
Lines: 35

In article <1678@cs.yale.edu> zador-anthony@CS.YALE.EDU (anthony zador) writes:
>In article <485@spinifex.eecs.unsw.oz> hughc@spinifex.eecs.unsw.oz (Hugh Clapin) writes:
>>I'm presently getting my act together for my honours thesis (in philosophy),
>>and want to explore the comparisons between representation in NNs and in
>>traditional symbol systems. Also of interest is a comparsion between how
>>NNs use the knowledge they represent and how a Von Neumann computer uses
>
>Connections and Symbols, ed. S. Pinker ~1988
>

These are generally critical articles of the NN approach.  The best
pro- article I saw was in Brain and Behavioral Sciences 1988 by
Paul Smolensky, called "On the Proper Treatment of Connectionism", 11, 1-74.
BBS is an "Open Peer" journal.  For this particular article there
were at least thirty responses.  This is a very good article, although it
is pretty difficult to get through (its long, and packed with information).

Basically, Smolensky is trying to characterize connectionist models as
using "subsymbolic" representations, which are quite different from
symbolic representations in traditional symbol systems.  These subsymbols
somehow describe micro-features of cognition as opposed to macro-features.
In essence, it is a reductionist approach.  One of the issues he brings
up (actually in his response to the peer responses) is the classification
of various approaches to cognition researchers have takens with
connectionist systems:  Eliminativist, Limitivist, Revisionist, or
Implementationalist.  Which are basically opinions as to whether
connectionism is just as doomed as traditional AI, something completely
new and different, or just the same old concepts in a different
architecture.  I think that there are new concepts being offered that
are unique to connectionism, and that the consequences (if connectionism
is successful) is that we would see a paradigm shift away from logic
towards pattern recognition as the dominant mechanism of cognition.
What do people out there think?  

-bill