O tell me where is fancy bred

[bnhpct]

From Bruce Nevin (970930.1420)

Bill Powers (970929.1954 MDT) --

Are you the only person in the world who can't make sense of what I'm
saying here? Or am I the only one who can?

No, it makes perfect sense to me. It seems to me (1) that it gets into
trouble on the supposed correspondence of category perceptions to words and
word constructions, and (2) that it is not required, that a sufficient
mechanism is memory -- whatever that is -- which must be there with or
without a category level.

OK. I'm certainly not going to solve any problems in linguistics. I'll
stick to what I know.

Are you saying that you know there is a category level in the human
perceptual control hierarchy? That you know that the experiences you cite
are evidence for a category level of perceptual control, and that they are
not results of a different mechanism involving memory? Or are you saying
only that the work of making and testing models should proceed at lower
levels where the issues are better understood?

This is not a problem in linguistics. It may just be my problem, my limited
grasp of how the explicit neural connections work in a control hierarchy,
and I should go away and study these things (talk to you in a decade or so).

It seemed to me there is not enough plasticity for reorganization to account
for how quickly category membership can be modified by analogy or in the
invention of nonce-categories, as quickly abandoned after use, as when
sorting fallen leaves for a collection of fall foliage; I don't see how, in
a model with gazillion hierarchical neural connections up to a category
control system, to explain the evidence that categories are organized by
similarity or analogy to exemplars (rather than set membership); nor the ad
hoc appearance of the categories that programs seem often work with. And it
seemed to me that we have to have some account of memory, association, and
analogy anyway; and that, however those processes can be modelled (my
weakness here, I haven't a clue), our experience of categories, or of
categorizing, might well fall out as an effect.

In B:CP there is no Category level. I think I recall your saying it was
brought in to account for abstractions and symbols. The example of a mundane
domestic program "looking for my glasses" in B:CP does not obviously involve
abstractions or symbols. It seems to me the program is

For each place
      { Remembering where I have been with my glasses
        Imagining where I might have been with my glasses
      }
      Go to that place.
      Remember or imagine leaving them.
      Control the perception of picking the glasses up.

In each place, including the starting place, controlling the perception of
picking the glasses up involves another loop something like
      { Imagining where they might have fallen from these places.
        Imagining what might have been placed on top of them or next to them
          in these places.
      }

What determines the places to look, and the sequence of them? Ah yes, how
does associative memory work? Is it easier to remember or imagine leaving
the glasses some places than others, or does the memory or imagined
experience come to awareness more quickly for some places than others?

The program itself is not limited to glasses, anything loseable will do.
(Where did I park the car?) Hence the apparent need for symbols. But
"loseable thing" seems a very ad hoc category, invented for the program
"finding things" and perhaps a few others. It's difficult to say that
programs determine what categories there are, if the category level is prior.

From an exemplar (with which category-members are associated--whatever that

means) to a symbol is a short step of conventionalization. Exemplars and
symbols can be perceptions input to the program level. The suggestion is
that the link from the particular (my glasses, my car) to the abstraction (a
losable thing) is associative. I reach for my glasses, they're not there.
This (the finding-things routine) is what I do when my control of a
perception of something being present occasions error. Is it possible that
there are effects of error (perhaps chemical) that trigger processes of
associative memory?

What control systems might be involved when a thing is lost whose presence
is necessary for successful control? Depends on what the thing is, what
perceptions in that CEV are controlled, what's going on in the environment.
It seems to me that the set of control systems that might be involved in the
experience of something being lost are too many and too varied for there to
be a neural connection from each such control system to the program for
finding lost things, or from each such control system to a category
perception "things that are lost" and thence to the finding-lost-things
program. The implausibility of that really tilts me toward an alternative
explanation.

It seems to me that with error begins imagination and memory. I reach for my
glasses to read the paper. They're not there. I imagine picking them up.
Picking them up is associated with previously putting them down. I imagine
putting them down. Not here. I remember putting them down various places. I
remember where I have been recently, and imagine putting them down. I
imagine where I might have been, and imagine putting them down. Controlling
a perception of picking them up where I previously put them down, I go to
each place in order of immediacy (in some sense or senses -- interesting
problems there -- Nasrudin searchng for the key under the streetlamp because
there's more light there).

In this case, maybe the appearance of a program is an effect of something
more primitive. But surely there are programs that cannot be explained as
effects of reducing error in memory and imagination, then trying to reduce
error in fact (in the environment) in the remembered and imagined ways. For
those, exemplars and conventional or institutionalized symbols could be
available by means other than by neural connections to a category level of
perceptual control.

I have to stop here and do some work.

        Bruce Nevin

[Bill_Powers1]

[From Bill Powers (971001.0840 MDT)]

Bruce Nevin (970930.1420)--

Are you saying that you know there is a category level in the human
perceptual control hierarchy? That you know that the experiences you cite
are evidence for a category level of perceptual control, and that they are
not results of a different mechanism involving memory? Or are you saying
only that the work of making and testing models should proceed at lower
levels where the issues are better understood?

I think we're getting confused here between what a level is supposed to do
and how it does it. The first concerns a phenomenon; the second proposes a
model for how the phenomenon works.

The phenomenon I'm trying to describe, that I call category perception, is
the ability to perceive things that are actually different (at lower
levels) as being the same thing. So Fido, Spot, Rover, and Nigel are all
perceived as what we call "dogs." When we look at pictures of different
animals, we can pick out the "dogs" even without having to say to
ourselves, "dog, dog, dog ...". There's a set of different configurations
that we put into one category, while other sets of configurations, also
different from each other, are called "cats" and are not seen as dogs. We
can't know whether to use the term "dog" or "cat" until we have already
perceived membership in the appropriate category.

We use words to "stand for" other perceptions, just as in algebra we say
"let x be the width of the field" or of the road, or box, or river. Once we
have made this substitution, we can then manipulate relationships among the
words, the symbols, instead of manipulating actual relationships among the
perceptions being symbolized. In doing this, we naturally generalize and
lose the lower-level details. We say "dogs chase cats," when in fact an
individual dog-thing may not chase an individual cat-thing -- Fido may
never in fact chase Tabby. We can say that the area of two fields is the
same because the products of width and length are the same, even though the
referent of "area", the ground enclosed by the boundaries of the two
fields, is occupied by grass in one field and trees in another, and one
field is round while the other is square.

This is what I'm talking about when I use the term "category." It has
nothing to do with memory, or association, or flip-flops, or "or" gates, or
any other model you might propose to explain how such perceptions can
exist. I'm just trying to describe one of the ways the world presents
itself to us in human perception. I've never seriously proposed a model to
explain how such perceptions are generated, beyond pointing out that for
every model someone proposes, another model can be proposed that does the
same thing in a different way. If you want to propose a model in which such
perceptions are generated by memory, or association, or any other
mechanism, it's incumbent on you to show how this model works and that it
is correct. But we don't have to do any of that to agree that we do
perceive in categories, at some level.

When I'm reminded, I am very conscious of thinking in categories. It's
usually helpful to know I'm doing so, because thinking in categories, while
useful, can be a mistake -- as when one perceives a "behaviorist" as
someone who has a stereotyped set of characteristics. If we can't help
thinking in categories, it's at least helpful, as Madam Blavatski said, to
be able to "think in different categories." It's also helpful, as Korzybski
reminded us, to be conscious of abstracting (which is what categories do)
and to remember that the categorical map is not the lower-level territory.
The weather map says "scattered showers" while I am looking out my window
at a clear blue sky.

This is not a problem in linguistics. It may just be my problem, my limited
grasp of how the explicit neural connections work in a control hierarchy,
and I should go away and study these things (talk to you in a decade or so).

That's a matter of explaining the phenomenon, not of describing it.

I don't see how, in
a model with gazillion hierarchical neural connections up to a category
control system, to explain the evidence that categories are organized by
similarity or analogy to exemplars (rather than set membership);

But what is "analogy to exemplars" but presenting the disparate members of
a given category and hoping that the listener can perceive them as examples
of the same thing? Perceiving common elements in a set of exemplars does
not explain a category; in fact, it reduces a category to a lower-level
perception, whatever it is that remains the same over exemplars. But
categories don't have to be composed of perceptions that have something in
common at a lower level. If I empty my pockets, I can give the name "mine"
to every item dropped on the table. They don't have to have anything in
common in themselves. In fact, by perceiving them as "mine", I create a
category of things I can dispose of as I wish.

Remember, there isn't "a" category control system in the HPCT model; there
is a category level, which includes many control systems which are set up
to produce certain category perceptions, as well as many input functions
that recognize certain categories and pass the result upward, but with no
control system attached. We have as many categorical input functions as
there are categories that we are prepared to recognize. This is not an
infinite number, but in an adult it is large, at least the size of one's
vocabulary.

nor the ad
hoc appearance of the categories that programs seem often work with.

I don't think the categories are nearly as ad hoc as they may seem. We can
say "let x be the width of the field," but "x", "width", and "field" are
not ad-hoc categories. What is ad-hoc is the structure that organizes these
categories into a process, which is what higher systems do.

In B:CP there is no Category level. I think I recall your saying it was
brought in to account for abstractions and symbols. The example of a mundane
domestic program "looking for my glasses" in B:CP does not obviously involve
abstractions or symbols.

First, look at pages 156 to 158. I was talking about category perceptions
there, very explicitly, and did't realize I was doing it.

Second, "looking for my glasses" most certainly does involve abstractions
and symbols. "Looking" is a symbol for a certain category of activities,
and "glasses" the name of a large category of objects. To say "My glasses"
intersects two categories: the category of glasses and the category of
things that belong to me. In my case that narrows the objects for which I'm
searching down to about 5, three of which are in a bureau drawer (old and
spare glasses), and two of which I consistently get mixed up when I'm in a
hurry; my bifocals and my computer glasses. Sometimes, thinking too
categorically, I dash downstairs to get my glasses to watch a TV program,
and come upstairs with the wrong ones. I come up with glasses, in the right
category but not the right individuaol items.

It seems to me the program is

For each place
     { Remembering where I have been with my glasses
       Imagining where I might have been with my glasses
     }
     Go to that place.
     Remember or imagine leaving them.
     Control the perception of picking the glasses up.

In each place, including the starting place, controlling the perception of
picking the glasses up involves another loop something like
     { Imagining where they might have fallen from these places.
       Imagining what might have been placed on top of them or next to them
         in these places.
     }

All the program can do is to manipulate the symbol, "glasses", and symbols
for "places" and so forth. Once it comes up with the name of a place, that
name has to be converted into an actual perception in the named category.
You might realize that you left them "beside" a "chair", so now you have to
translate "beside" into a perceived spatial relationship, and "chair" into
the right example of a chair. Only then can you actually go to the right
place, and pick up the actual glasses instead of the name, "glasses." You
can't watch TV through the name of the object, "glasses". But all the
program level can come up with is the symbol for a category (or the names
of lower-level perceptions, which most often turn out to be names of
_categories_ of perceptions, not specific ones).

I think that trying to _model_ this kind of category perception can be
postponed until we agree that there is a phenomenon to be modeled.

Best,

Bill P.

[bnhpct]

From Bruce Nevin (971001.1859)

Bill Powers (971001.0840 MDT) --

Thank you, Bill, for your extended reply.

I think we're in agreement that a phenomenon exists that we can call
category perception.

I've never seriously proposed a model to
explain how such perceptions are generated

You have described a model in which many perceptions (tail wagging, barking,
four legs, the word "dog") all input to a single category control system--as
I said, "gazillion hierarchical neural connections up to a category control
system". I thought this was a seriously proposed model. No?

... beyond pointing out that for
every model someone proposes, another model can be proposed that does the
same thing in a different way. If you want to propose a model in which such
perceptions are generated by memory, or association, or any other
mechanism, it's incumbent on you to show how this model works and that it
is correct.

I agree that this is the challenge, and I have confessed (do confess) that I
am not up to it, not real modelling.

But we don't have to do any of that to agree that we do
perceive in categories, at some level.

I think we do have some disagreement about the character of category
perception, the phenomenon. The assignment of words and symbols as names of
categories lends to them a tidiness that may be deceptive. It is easy to
substitute our usage of words and symbols in place of our more primitive
perception of categories. This is hard to get clear about, and hard to
communicate when I think I have, and I will have to work on it more than I
can for this note.

A couple of misunderstandings to clear up:

Remember, there isn't "a" category control system in the HPCT model; there
is a category level ...

I know this. Guess I said something misleading?

the evidence that categories are organized by
similarity or analogy to exemplars (rather than set membership)

... what is "analogy to exemplars" but presenting the disparate members of
a given category and hoping that the listener can perceive them as examples
of the same thing? Perceiving common elements in a set of exemplars does
not explain a category; in fact, it reduces a category to a lower-level
perception, whatever it is that remains the same over exemplars. ...

An exemplar is a single instance,perhaps idealized, like the
hyperarticulated phoneme. "Examples" are members of the category by analogy
to the exemplar, perhaps because they are perceived as more analogous to
that exemplar than they are to the exemplar or idealized core member of any
other category. Maybe the exemplar would be highest-strength signals from
all the "gazillion inputs" in your model, and analogy would work by
diminishing the strength of some signals.

"looking for my glasses" most certainly does involve abstractions
and symbols. "Looking" is a symbol for a certain category of activities,
and "glasses" the name of a large category of objects. To say "My glasses"
intersects two categories: the category of glasses and the category of
things that belong to me. In my case that narrows the objects for which I'm
searching down to about 5 ...

This is one way of modelling the program. I tried to describe another which
so far as I can tell does not use symbols. In fact, it's not a program,
really, it's a side effect of controlling the perception of picking up my
glasses, first in real time, then in memory and imagination, and for a
succession of remembered or imagined scenarios of putting them down
controlling the perception of picking them up in the remembered or imagined
place, which requires going there. Surely not all program perceptions can be
reduced in this way, but by golly it is suggestive.

All the program can do is to manipulate the symbol, "glasses", and symbols
for "places" and so forth. Once it comes up with the name of a place, that
name has to be converted into an actual perception in the named category.
You might realize that you left them "beside" a "chair", so now you have to
translate "beside" into a perceived spatial relationship, and "chair" into
the right example of a chair. Only then can you actually go to the right
place, and pick up the actual glasses instead of the name, "glasses." You
can't watch TV through the name of the object, "glasses". But all the
program level can come up with is the symbol for a category (or the names
of lower-level perceptions, which most often turn out to be names of
_categories_ of perceptions, not specific ones).

Here, you have simply ignored my description and substituted your own.

This hurried post is the best I can do now, and your considered reply
deserves better. I will try to give more thought to this this weekend.

        Bruce Nevin

[Bill_Powers1]

[From Bill Powers (971001.2011 MDT)]

Bruce Nevin (971001.1859)

You have described a model in which many perceptions (tail wagging, barking,
four legs, the word "dog") all input to a single category control system--as
I said, "gazillion hierarchical neural connections up to a category control
system". I thought this was a seriously proposed model. No?

That's not a model, only a possible start toward one. It simply expresses
the idea that a category is a signal that can be evoked by any of the
lower-level perceptual signals that are being treated as members of the
same category.

I agree that this is the challenge, and I have confessed (do confess) that I
am not up to it, not real modelling.

Neither am I. It's much more of a problem that it seems on the surface.

I think we do have some disagreement about the character of category
perception, the phenomenon. The assignment of words and symbols as names of
categories lends to them a tidiness that may be deceptive.

Remember that I'm not proposing that categories _are_ words. They're
perceptions; we can recognize them silently. We can refer to them using
words and symbols, with all the abiguities involved in referring to
anything with words or symbols.

An exemplar is a single instance,perhaps idealized, like the
hyperarticulated phoneme. "Examples" are members of the category by analogy
to the exemplar, perhaps because they are perceived as more analogous to
that exemplar than they are to the exemplar or idealized core member of any
other category.

Maybe the exemplar would be highest-strength signals from
all the "gazillion inputs" in your model, and analogy would work by
diminishing the strength of some signals.

Now you have to explain how we perceive analogies.

"looking for my glasses" most certainly does involve abstractions
and symbols. "Looking" is a symbol for a certain category of activities,
and "glasses" the name of a large category of objects. To say "My glasses"
intersects two categories: the category of glasses and the category of
things that belong to me. In my case that narrows the objects for which I'm
searching down to about 5 ...

This is one way of modelling the program. I tried to describe another which
so far as I can tell does not use symbols. In fact, it's not a program,
really, it's a side effect of controlling the perception of picking up my
glasses, first in real time, then in memory and imagination, and for a
succession of remembered or imagined scenarios of putting them down
controlling the perception of picking them up in the remembered or imagined
place, which requires going there. Surely not all program perceptions can be
reduced in this way, but by golly it is suggestive.

If the glasses you're looking for are a specific pair, as implied by "my
glasses," then no category perception need be involved in the search.
However, the program level, I think, probably always works in terms of
symbols, not directly with lower-level perceptions. Considering how many
words _do_ refer to categories instead of unique individual perceptions, I
think we also have to consider examples of the form I proposed.

All the program can do is to manipulate the symbol, "glasses", and symbols
for "places" and so forth. Once it comes up with the name of a place, that
name has to be converted into an actual perception in the named category.
You might realize that you left them "beside" a "chair", so now you have to
translate "beside" into a perceived spatial relationship, and "chair" into
the right example of a chair. Only then can you actually go to the right
place, and pick up the actual glasses instead of the name, "glasses." You
can't watch TV through the name of the object, "glasses". But all the
program level can come up with is the symbol for a category (or the names
of lower-level perceptions, which most often turn out to be names of
_categories_ of perceptions, not specific ones).

Here, you have simply ignored my description and substituted your own.

All right, you're right. But consider the case in which you're searching
not for a specific item, but simply for examples of a category. Shopping
for a shirt would be an example.

Best,

Bill P.