Newsgroups: comp.ai.philosophy
From: Lupton@luptonpj.demon.co.uk (Peter Lupton)
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!udel!gatech!swrinde!pipex!demon!luptonpj.demon.co.uk!Lupton
Subject: Re: Strong AI and consciousness
References: <1994Dec2.143356.8747@oracorp.com>
Distribution: world
Organization: No Organisation
Reply-To: Lupton@luptonpj.demon.co.uk
X-Newsreader: Newswin Alpha 0.6
Lines:  67
Date: Tue, 6 Dec 1994 00:39:44 +0000
Message-ID: <135458154wnr@luptonpj.demon.co.uk>
Sender: usenet@demon.co.uk

In article: <1994Dec2.143356.8747@oracorp.com>  daryl@oracorp.com (Daryl McCullough) writes:
> 
> Lupton@luptonpj.demon.co.uk (Peter Lupton) writes:
> 
> > AC provides an account of how classification can come about which has
> > no dependency upon any logicist notions. Nor does AC depend on
> > propositions, content or any such notion. AC is a computational, not a
> > logical notion.
> 
> I agree with your point that classifications can be judged by their
> data-compressing qualities. However, I don't agree that algorithmic
> complexity is the right way to form classifications. Algorithmic
> complexity as you've defined makes the complexity of data dependent on
> the shortest program that can be used to encode the data. The problem
> with using the shortest program is that for any finite amount of data,
> there will always be silly, spurious correlations. 

First of all, I am not sure what you have in mind by 'silly, spurious 
correlations'. I would think there are cases where we do form silly, 
spurious, correlations.

That said, I think you *might* be pre-judging the issue. The structure 
of a short program which is able to reproduce a large given stream of 
data is really going to be very elaborate. Contexts and sub-contexts 
will arise - the degree to which a certain sequence of bits can be 
compressed will come to depend crucially upon the context in which
it arises. Identical data arising in quite distinct contexts might
well turn out to be compressed in very different ways - it will 
all depend on what is shortest-in-context.

It may well be that the sort of spurious relationships you
propose will arise initially, when there is little data.
But it may be that as the quantity of data increases and, 
particularly, as the sophistication of the short program increases, 
such spurious relationships may tend to be eliminated. I believe this 
to be a consequence of AC.

> Therefore, if you
> use AC to give rise to classifications, there will be correspondingly
> spurious classifications. Of course, in the limit as the amount of
> data being compressed goes to infinity, presumably all spurious
> correlations will eventually disappear, but I thought that your main
> point about the benefit of AC was its applicability to finite data.

The notion of a spurious classification is interesting in its own
right. What, presumably, is meant, is that the classification may
well be made and then revised. This does not require infinite data,
and the trend of AC to force revision has already been mentioned.
If your point is that AC will tend to have spurious classifications
which will be revised, so be it. I am sure that I do have such
classifications right now - what will force me to revise them will
either be new information (new data) or more thought (additional
computation).

> I don't think that the *only* purpose in classification is the
> compression of data.

I agree. I could, for example, right now, classify in order to
make a point. The classification could be blatantly contrary to
the purpose of data compression. What is at issue, I think,
is whether one could be in a position to classify for *other*
purposes unless one already had the ability to classify for
data compression purposes.

Cheers,
Pete Lupton
