Philosophical Foundations

[Robert Toews]

Gerry

This thread has been very interesting, however, I think that more effort
should be spent on actually building workable thinking machines than trying
to classify the existing ones (Although I do realize there is usefulness in
that.)

It  is often said that most new discoveries stem from the merging of
technologies.  In that vein, I suggest that there is promise in the merging
of your compression theories with the knowledge representation theories
developed by Dr. Tim Lethbridge.  See Chapter 3 of his PhD thesis at:

http://www.csi.uottawa.ca/~tcl/thesis_html/thesis.html

For what it's worth, I think that an adaptation of your compression
algorithms may fit well as a front-end for Dr. Tim Lethbridge's knowledge
representation model.  I'm currently working on that approach, however, N
heads are better than one <grin>.

Bob Toews

-----Original Message-----
From: Gerry Wolff <gerry at sees.bangor.ac.uk>
To: Computing as Compression Mail List <casc at sanna.com>
Date: Thursday, December 24, 1998 5:19 AM
Subject: Re: Philosophical Foundations


>Freddy Offenga wrote:
>...
>>
>> If it's true that compressionism underlies both symbolic and
>> connectionist models, it might be useful to step to a lower level
>> (compressionism). But this should only be done if the other
>> approaches (symbolic and connectionistic) get stuck. In other
>> words, the brain should be explained at the right level. It's
>> not useful to talk about neural firings, since this level is too
>> low for a functional explanation (I don't say examining the neural
>> level isn't useful for cognitive science).
>>
>> But maybe there's another argumentation. If compressionism is
>> done by a symbol system, wouldn't it mean that the symbolic system
>> is the "lower level"?
>> I get a little confused by this. One time it's claimed that
>> compressionism has at least the same power as a Turing machine.
>> Another time it's stated that the SP system can be executed on a
>> Turing machine (or conventional computer). This is why I asked
>> the question; is compressionism a _sub-class_ of symbolicism.
>> Since Gerry tells me this isn't the case, I wonder what the real
>> differces are. If the only difference is the aspect of "massive
>> parallelism", I don't see why this can't be done with the "old"
>> symbolic approach.
>
>Regarding "symbolic" and "connectionist" approaches to AI etc, I feel
>that much of the confusion arises because neither term is very precise.
>In the work I am doing I have chosen to define a 'symbol' as any kind of
>arbitrary 'mark' that can be discriminated from other symbols in a
>yes/no manner. This includes 'symbols' like the binary symbols '0' and
>'1' which are the currency at the lowest levels of all digital
>computers. If this idea of a symbol is adopted, then one could argue
>that 'symbolic' approaches to AI and cognition can underly connectionist
>approaches - because neural nets can be and often are simulated on
>digital computers! It is also possible to argue that connectionism
>(possibly) underlies symbolic mechanisms - because any Turing machine
>can be simulated by a Post Canonical System and because it is probable
>that a PCS can be simulated by a neural net (I don't know whether anyone
>has tried it).
>
>I have indeed argued that the SP system can model a Turing machine (the
>relevant article needs some tightening up but I think the argument is
>sound). But, since current versions of the SP system are running as
>software simulations on a conventional computer (which is similar to a
>Turing machine), we do seem to have a paradox here.
>
>My view of the paradox is, at present, something like this:
>
>* An SP system may be characterised roughly as a system for finding
>'good' multiple alignments whose chief feature is a relatively
>sophisticated search mechanism.
>
>* It is possible to simulate such a system using a conventional computer
>(approximately, a Turing machine) but WITH ADDED PROGRAMMING. In other
>words, the simulation is not achieved using the basic machine. There
>must be certain additional information on the 'tape' of the machine.
>
>* I suspect that the current means of simulating the SP system (a
>program on a conventional computer) will not, in the long run, be the
>best way of setting it up. It should be possible to design new kinds of
>hardware dedicated to the processes in the SP system. It is likely that
>high levels of parallel processing will be used. In any such realization
>of the SP system, it may be hard to detect within it anything that
>resembles a Turing machine.
>
>* A neural net might very well provide a foundation for an SP machine.
>This could be an interesting and fruitful exercise which might help to
>remedy some of the weaknesses of the current generation of neural nets
>in performing some kinds of 'symbolic' task.
>
>The critical point seems to be: In asking questions like "Do systems of
>type X 'underly' systems of type Y?" or "Can systems of type Y be
>simulated by systems of type X?" we need to be very precise about
>whether we are talking about X or Y in 'basic' form or whether we might
>really be talking about "X (or Y) plus some programming".
>
>I agree that these kinds of issues can be quite confusing. Questions
>like "which kind of system is 'lower'?" may not always make sense. In a
>conventional program we can say that function B is at a 'lower level'
>than function A if it is 'called' from within function A. But what do we
>say about functions A and B if A calls B in one context but B calls A in
>another? There may be a recursive loop with A calling B and B calling A
>through many 'levels'. What do we say about 'levels' if a function calls
>itself (which, of course, is commonly done)?
>
>> Special thanks to the following people:
>> Sergio, Remy, Phil, Nate, Detlef, Chuck and Gerry.
>
>And thanks very much to you, Freddy, for raising these interesting
>questions. Happy Christmas to you and everyone on the list!
>
>Gerry
>