NetNews Usenet Archive 1992 #26/NN_1992_26.iso/spool/bit/listserv/csgl/1439

Text File | 1992-11-10 | 8.6 KB | 169 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!stanford.edu!bcm!convex!darwin.sura.net!paladin.american.edu!auvm!VAXF.COLORADO.EDU!POWERS_W
X-Envelope-to: [e-mail]
X-VMS-To: @CSG
MIME-version: 1.0
Content-transfer-encoding: 7BIT
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Tue, 10 Nov 1992 13:52:01 -0700
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         "William T. Powers" <[e-mail]>
Subject:      Linguistics, misc
Lines: 154

[From Bill Powers (921110.1200)]

Penni Sibun (921109.1300) --

>well, i actually don't believe in ``meaning'' very much.  i think >it's
a systematic error that people tend to think actions that >involve
language have meanings and those that don't don't.

PCT and I agree with at least your second sentence. It isn't actions
that have meanings, whether they're verbal or other motor outputs. As
Bruce Nevin pointed out, in PCT (and elaborations thereon) meanings are
perceptions evoked by other perceptions. If you perceive the word
"green", its meaning is the color that comes to mind. If you perceive
the color green on a traffic light, its meaning is a perception of your
car starting to move (in that context). This translation between one
perception and another associated one is our proposition for how the
folk concept of meaning works.

As to doing without meaning entirely, I think that would pretty much put
an end to communication.

>a higher order system looks for the completeness criterion to be
>satisfied, and then salix *can* stop--but doesn't *have* to.  can
>you have that kind of variable in pct?

Sure. It's what we call a "single ended" or "one-way" control system --
one that considers a departure of a perception from the reference signal
in only one direction to be an error. This is like controlling for how
close to the edge of a cliff you like to perceive yourself. When you're
closer than the reference distance, there's a big error and you pull
back. But if you step back so you're farther from the edge than that
amount, the control system doesn't act to bring you closer again. One-
way control.

Actually all simple neural control systems have to be one-way systems,
as I model them in my head. The reason is that comparators are made of
neurons with one inhibitory and one excitatory input, the excitatory one
being the reference signal. You get an error signal only when the
perceptual signal is less than the reference signal, leaving some
effects of the reference signal uninhibited. You increase your action
until the perceptual signal completely inhibits the effect of the
reference signal, leading to zero error signal (the output of the
neuron). When the perceptual signal is more than large enough to
completely inhibit the reference signal, you still get zero error.
Neural frequencies can't go negative -- less then zero impulses per
second.

To get a two-way neural control system you need two comparators, one of
which has inhibition and excitation swapped relative to the other. This
implies that control actions that pass through a zero or a neutral state
will not, except by accident, be completely symmetrical. If you measure
the parameters of control actions, separating the actions for positive
errors from those for negative errors, you should find a slight
difference in the parameters. We haven't tried that yet.
>... what i'd like you to try to do is to assume--for just
>a few minutes--that i have an interesting and intelligent proposal
>(viz., what has been said and what there is to say codetermine what
>can be said now), and work with me to cast that model in pct terms.

For this you must propose a model (my meaning). The model would be a
proposal as to HOW what has been said and what is to be said do this
codetermining. You may actually have built some of this model into your
program, without identifying the process implied by the program
manipulations in PCT terms.

Suppose that in the knowledge base there are elements called John and
Penni. Whether you mean the person or the name of the person is
immaterial. In PCT terms we would say that the program perceives things,
names them, and stores the names in memory structures. The perceptual
signal would be whatever is used in the program to stand for the memory
structure -- a 32-bit address, for example, although in a closed-loop
system any arbitrary but unique symbol would do (it won't be necessary
in a closed-loop system to work backward from the symbol to the thing it
indicates).

Also in your knowledge base you want some relationship perceptions to
exist -- for example (father-of John Penni). This implies a general
perceptual function of the form (father-of x y) where x and y are any
two inputs from the lower level elements. I don't know how you'd design
such a perceptual function; presumably it would rely on multiple sources
of evidence other than just the names John and Penni. There would have
to be something about John and something about Penni and something about
their relationships that would reveal that the first element is the
father of the second element -- or is not.

Of course you could let this process occur behind the scenes or under
the table, and just deposit the specific statement (father-of John
Penni) in the knowledge base. In that case you would be relying on
processes in the programmer that identify the father-of relationship,
but not specifically representing those processes in the model.

The same could be done at the level of the individual words. Instead of
proposing some process by which a set of sensory inputs is identified as
"John" or "Penni," you could just put those terms in the knowledge base,
letting the processes of perceiving and naming take place outside the
model.

This takes us only the first step toward a PCT model of how what has
been said and what is to be said codetermine what is to be said next. We
have been trying to determine how something can be known to be said. We
have yet to determine how the system can know what remains to be said,
so more parts of the model can be constructed. Then we will have to make
some proposals as to what "codetermination" is to mean. And finally, we
will have to close the loop, so that what is finally said can be made to
satisfy all the conditions we have set up in the earlier parts of the
model.

I would be very surprised if the answers to many of these questions were
not already embodied in the steps of your program. To unravel the PCTequivalent
 of the model that you are already using (if such exists), it
would be necessary to analyze the program itself to see how one kind of
information is extracted from other information, to see how comparisons
are made, and to see how the results of comparisons lead to actions. I
won't dispute that the outcome of all these processes will be a
codetermination of what is to be said next by what has been said and
what remains to be said. But we won't have a model to account for that
phenomenon until we have taken your program apart and seen just how this
codetermination is brought about.

The best person to do that, I suggest, is you. If you don't want to take
the time off to learn the fundamentals of PCT, then it's not likely to
get done. It's up to you to judge whether the result would be worth the
effort.
------------------------------------------------------------------
Dag Forssell and Hank Folson --

PCT has been shown to make very accurate predictions for some very
simple tasks. I get extremely uncomfortable when I see public claims
that the PCT model is 99.999% accurate. That is not just misleading,
it's a lie. Saying things like that will make people think that PCT can
predict what people will do with that kind of accuracy in all
situations. This is simply not true. There are years and years of
experimentation and model-building to do before we can know how well PCT
will apply in areas such as language, social interactions, learning, and
so forth -- all the fields where PCT has not yet been systematically
applied or tested.

If you want to sell PCT that's fine. If it has engaged your enthusiasm,
wonderful. But please tone down the hype. That doesn't further the
cause. Sorry about the bludgeon, but I want to get the donkey's
attention.
------------------------------------------------------------------
Jixhuan Hu (921110) --

Welcome to CSGnet! I think that before you start making comparisons
between our uses of control theory and cybernetics, you should read some
of the introductory material. It takes most people more than a week to
realize what we're talking about. You won't find much in the kind of
cybernetics you were mentioning that is like what we do.
-------------------------------------------------------------------
Best to all,


Bill P.