NetNews Usenet Archive 1992 #27/NN_1992_27.iso/spool/bit/listserv/csgl/1543

Text File | 1992-11-18 | 10.0 KB | 198 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!news.gtech.com!noc.near.net!mars.caps.maine.edu!maine.maine.edu!cunyvm!psuvm!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:         Wed, 18 Nov 1992 09:19:21 -0700
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         "William T. Powers" <[e-mail]>
Subject:      words & images;logical perceptions; intro
Lines: 183

[From Bill Powers (921118.0730)]

Avery Andrews (921118.1430) --

  >>When I hear "large tree at the top of a hill next to the path" I
  >>construct something like an imagined top of a hill with an
  >>imagined big tree next to the path.

>This won't work, because images have the wrong kind of information
>in them. A classical example from  philosophy:  an image of a
>triangle has to either be isoceles or not isoceles, but the concept
>triangle is neither.

You must understand that when I say "image" I mean a perceptual signal
derived from inputs and indicating that something is present in those
input signals, as per PCT. If I tell you "draw a triangle" you just draw
a triangle -- it doesn't matter whether it's equilateral or isosceles or
obtuse, or whether the end of the last line exactly touches the start of
the first without crossing it. Yet you always draw a SPECIFIC triangle.
Any triangle will give rise to a perceptual signal indicating that
triangleness is present. If you want a special kind of triangle, you
have to specify TWO reference signals: triangleness and equilateralness,
for example. Then you are asking for perceptions of both triangleness
and equilateralness, which are independent dimensions of variation and
can be controlled (more or less) independently. If I ask you to draw an
equilateral figure, you might draw a triangle, a square, a hexagon, an
octagon, and so forth. You might make some of the vertices in more
complex figures indented. In the uncontrolled dimensions, anything will
do. All that matters in controlling for equilateralness is that all the
sides be perceived as equal.

The same is true for the Ronald Reagan image. To control for such an
image, all that matters is that some aspect of RRness be present,
sufficient for a perceptual function to wake up and signal "I see my
guy." Cartoonists know how to wake up such perceptual functions by
picking out a sufficient set of inputs and exaggerating them to make up
for distortions of other inputs. That's called a caricature.

>It is indeed something of a mystery how anything can translate >between
images and the `concepts' that we like to represent as >propositions in
logic, but that mystery has got to be solved >anyway, & once it is, this
solution could be used to build a >concept-to-image comparator.

In HPCT there are no such things as concepts. We break down perceptions
in a different way into lower level and higher level perceptions of
specific hierarchically-related classes (maybe not the right classes,
but it's a start). In the informal language of pre-HPCT philosophy there
are only a few categories of perception and they are poorly defined: you
have concrete and abstract, or sensations-perceptions-concepts, and
other such vague notions.

Propositions are generally composed of symbols related by implicit
operations. In "Socrates is mortal" the operation is that of equality or
interchangeability or class membership, depending on how you think of
it. "Socrates" is the name of a configuration, a person. "is" is thename of a
 relationship or a class membership. "Mortal" is the name of a
rather fuzzy collection of perceptions having to do with dying (which
for oneself can only be imagined), time duration, and perhaps some
principles like "Life ends."

But in dealing with "Socrates is mortal" as a proposition, the actual
meanings of the terms make no difference; the words are just place-
holders. The proposition is "x is a member of class y." It is then put
together with other similar propositions at the logic level, or the
ninth level in current HPCT, to yield a perception of TRUE or FALSE. If
a conclusion is TRUE, and the reference level is TRUE, then the logical
control process is satisfied and the logical process stops. The output
proposition is then in the state that causes the input symbols to be in
states that make the truth value of the perceived proposition match the
reference value.

Logic has its uses, and its abuses. It helps in trying to figure out
complex relationships. But it can be used to do logical manipulations of
symbols where the symbols stand for perceptions that don't actually
follow the logical rules. You can start by saying "All men are bicycles;
Socrates is a man; therefore Socrates is a bicycle" and nothing at the
logic level will object. Logical and other rule-driven processes may be
powerful and complex, but they are stupid. They are certainly not the
highest level of function in the brain. If they were, we would see
nothing wrong with concluding that Socrates is a bicycle.

My chief objection to overmathematicized approaches to modeling is that
mathematics doesn't need to make sense. There has to be a higher
perception of the appropriateness of the mathematics. The judgement as
to whether or not to believe the mathematical conclusions is not itself
a mathematical process. It involves principles and system concepts --
the aspects of perception and action that we call, in older fuzzier
terms, "wisdom."
------------------------------------------------------------------
John Gabriel (921117.2131) --

See my last paragraph to Avery.

>My baseline concept is a Kanerva database of truth values for a set
>of N Aristotelian propositions that together discriminate within an
>area of discourse.  The truth values are bit strings N long, and
>can be thought of as being in a space of N dimensions, containing
>2**N points each representing a possible object in the area of
>discourse. Typical values of N are in the range of 100-200, and the
>space is very sparsely populated.

As no reference to meaning has been made so far, this is a strictly
logic-level proposition. The universe of discourse is being DEFINED to
have the desired properties: the rules of the game are being laid out.

>New propositions can be introduced by refining old ones. For >example,
if we both understand "green", we can qualify "green" into
>subcategories of the taxonomy by
     green.1   Green like grass.
     green.2   Green like waterweed.
In real perceptions, there are also
     green.3 Green like about 2/3 of the way from grass to
             waterweed.
     green.4 Green like about 4/7 of the way from grass to
             waterweed.
     ....

By assigning meanings to symbols as you propose, you're introducing an
abitrary grid into the world of lower-level perceptions, such that
perceptions can symbolized accurately only at the intersections of the
grid lines, and jump from one intersection to another instantaneously
(as, for example, the grass dries out). In the world of symbols there
are no states of green between green.1 and green.2. Therefore if one
person is perceiving such an intermediate shade of green, it will be
impossible (save by luck) to convey that shade of green to the other
person because the only terms available for description refer to fixed
points in the continuum.

This means that the world of discourse represents the experienced world
by forcing it into discrete forms. This limits the meanings that can be
conveyed, and that can be perceived to have been conveyed. The fewer the
agreed-on Aristotelian propositions that are used, the more stereotyped
the discourse becomes. Even in the limit, allowing the maximum
conceivable number of agreed-on propositions, the world of direct
experience is always far more detailed than what we can say about it.
-------------------------------------------------------------------
Dag Forssell (921117.2200) --

Finally I'm getting around to your intro to CSGnet.

I'm of two minds about this document. I know that if I were a stranger
to the net, I would be put off by it. I don't need to be told why I
might want to read CSGnet, particularly not when the telling comes
across as a sales pitch. You say I'll be interested "If you're
interested in how people and organisms work... If you're dissatisfied
with explanations ... if you insist on thinking things through for
yourself...". Of course that means that if I don't want to read CSGnet I
must be some sort of stupid dolt who isn't interested in people or
organisms, who is satisfied with nonrigorous explanation, and who
doesn't like to think for himself. This sounds like an insurance
salesman who says "Do you care what happens to your children after you
die?"

I wrote my intro thing at your house and was being a salesman. I'm
uncomfortable now when I read it.

Mary's stuff looks fine, but I have to say that, don't I? Let others
judge.

In general, I draw back from the air of the expert aiding the neophyte
who is wandering through CSGnet in a state of mild confusion. Most of
the people who look in on CSGnet will be smarter than we are, and have
long experience with approaching new ideas and learning about them
efficiently. I don't think they need to be told how to ask effectivequestions. I
 think a description of the available materials should be
enough -- nobody is going to consult them just because you tell them to,
and most people who browse around networks know enough to read the
source material before starting to express opinions.

I think you're doing us a service by working up this document, but I
also hope it can be revised some more to eliminate the salesmanship and
condescension.
-------------------------------------------------------------------
Bruce Nevin (921118) --

The analogue neural net circuit sounds interesting. The transistor count
per neuron is about the same as for an operational amplifier. Any idea
where I should write to get the specs?
---------------------------------------------------------------
Best to all,


Bill P.