NetNews Usenet Archive 1992 #26/NN_1992_26.iso/spool/bit/listserv/csgl/1409

Internet Message Format | 1992-11-09 | 3.5 KB

Path: sparky!uunet!noc.near.net!hri.com!ukma!psuvax1!psuvm!auvm!FAC.ANU.EDU.AU!ANDALING
From: [e-mail] ((Avery Andrews))
Newsgroups: bit.listserv.csg-l
Subject: language, discreteness
Message-ID: <[e-mail]>
Date: 9 Nov 92 14:23:40 GMT
Sender: "Control Systems Group Network (CSGnet)" <[e-mail]>
Lines: 58
Comments: Gated by [e-mail]

[Avery Andrews (921108.1244)]
 (Bill Powers (921107.0800))

> I have some questions about the "knowledge base." I gather that what
> you mean by a knowledge base is a set of statements describing
> perceptions, rather than the perceptions themselves. I can see how it
> would be possible to mark a list of statements as "said" or "not
> said," but how would you mark the perceptions to which they refer?

The perceptions being described are typically long-gone in these
situations, so I don't see that there's any sense in which they
can be marked at at all.  It's my impression, though I don't know
the area, is that there is a certain amount of evidence that memory
really is symbolic to a considerable degree, and so, when one is
talking on the basis of what one remembers, it is quite appropriate
to mark statements as `said' or `not said', at least as a first
approximation.

As for modelling, what you are sensing missing is any idea of how
linguistic tasks could be done with brain-like-as-we-know-it hardware.
The tricky bit, I think, is how to deal with the fact that we seem
to be able to build up networks in which arbitrarily large numbers
of individuals are classified and related with a finite number
of properties and relationships.  E.g. we can learn of any
number of women that they are daughters of, say, Adeline, any number
of men that they are (current or former) husbands of these women,
etc. & people can learn this kind of stuff very quickly--much faster
than a connectionist network could be trained.

I don't think anyone has a clue as to how to do this with
neurologically realistic hardware (wetware?), just like nobody had
any idea in the nineteenth century about how to reduce chemistry to
physics, though people seemed to have assumed that it had to be able to
be done somehow.  Maybe this means that PCT should ignore language entirely
until this problem is solved, but I think that the idea of closing loops
through the environment has enough power to be worth trying to apply to
language even without neurologically concrete modelling.

And, it seems to me that trying to build models with the emphasis on
interaction & minimal representation might turn up some constraints that
would make a neurologically realistic model easier to attain.  At
any rate, it's the best we can do at the moment.

The point of my little piece on discreteness is not that it is
not necessarily problematic for PCT, but that a different collection
of things might be important in a regime where disturbances are
highly limited (e.g. once something has been said, nothing can
make it unsaid), and where reference levels can be attained
*exactly*.   In particular, it is not at all clear to me that the
usual equations have much significance--what makes equations
interesting is the possibility of solving them non-trivially,
but I don't see how to do this in this sort of domain:  the perception
that cancels the error signal is to hear somebody saying what is
to be said, & that's a lot less exciting that the notion of the equilibrium
state of a feedback system, at least to me (since the solution is a
trivial restatement of the problem).

[e-mail]