Is there a category level?

[From Bill Powers (931022.1500 MDT)]

Bruce Nevin (931022.1547) --

The only strong examples of category perceptions that I recall
being forwarded are those to which we refer with words.

Yes, we refer to them with words, but we refer to many other
perceptions with words -- for example we refer to a certain range
of color-perceptions with the word "green." But the word green is
not the perceived color-sensation.

We can also refer to them without naming them:"Would you hand me
the color chip that matches this?"

The same goes for categories. "I need something to pry with" is
an implicit request that can't be fulfilled unless you can
perceive items as being in that category. Unless you have already
perceived membership in that category, there's no way to know
whether the words apply to that item.

To me, the hallmark of a category perception is that when you
look at an example of the category, you see not that particular
example but a likeness to a general kind of thing. That is, you
see "Mississippi" and you perceive a rivery sort of thing. Or you
look at a stack of umbrellas and you see which one is "yours." Or
you see not "that big umbrella, that small umbrella, that broken
umbrella" but what you refer to as "umbrellas."

I suggest that the discreteness and referentiality that we
attribute to those category perceptions could actually
properties of the word perceptions.

What properties could a word perception have? If there is a
perceptual signal indicating that a word has occurred, what other
information could that same signal carry? How can a signal carry
out any kind of function?

I don't understand "referentiality." Do you mean some property of
being capable of referring to something else, with no outside
help? If a word "refers" to something, it's not the word doing
the referring, is it? The word is just a signal; it can't DO
things to other things. I can understand a process that uses a
word-signal to refer to (in place of) some other perception, but
the process is neither the signal nor the other perception: it is
something that performs the function of referring, whatever that
means.

If phonemes are categories, they are unlike other
(or many other?) categories in this exhaustiveness of a well
defined and familiar subdomain of perceptions.

I'm not willing to say that phonemes are categories. They're
perceptions, yes, but why at that level? The mere fact that
different sets of sensations can produce the same phoneme-signal
doesn't imply that category perception is involved. That's just
how all many-to-one perceptual functions work, whatever the kind
of perception. Any why assume that categories exhaustively define
a domain of perception? As far as I know, the number of possible
categories is unlimited and there is no need for them to be
mutually exclusive. A doughnut and three dimes can be one
category, and two dimes and two doughnuts another. If you have
some reason to make such categories.

Any perceptual signal exhausts the domain of values that that
perceptual signal can attain. Three perceptual signals combined
in three orthogonal ways exhaust the space defined by the three
signals. This is true of any perceptual function. There's nothing
unique to any one level there.

···

===================================
I don't deny that there are perceptions of the sort that we call
categories. I am only proposing that they are not located on a
category level.

                      > reference (r)
                      >
       perception (p) V error (e)
        ---------->comparator-----------
       > >
    perceptual output ORGANISM
      input (PIF) function
     function |
       > >
       ><---------IMAGINATION-----------|
       > (copy of r) |
       : :
       : : [lower levels
omitted]
       : :
       : :
=======^================================V=====
       > >
       > > WORLD
        <------------CEV<--------------- [multiple loops from
lowest ECSs]
                      ^
                      > disturbance
                      >

In the first place, the arrow you have drawn is not a copy of r,
but of the output signal, which is a function of r - p (the error
signal), not r alone. The imagination connection is a copy of the
reference signal sent to a system at the level BELOW, acting like
the lower-level perceptual signal that would be generated if
control were perfect at the lower level. And this is only one
INPUT to the higher-level system, not the perceptual signal. For
total imagination, EACH reference signal sent to a lower system
would have to be routed back to the input function of the higher
system.

In the second place, in PCT nothing is perceived unless it is in
the form of a perceptual signal. The arrow crossing from output
to input is not a perceptual signal, but a pathway carrying
information to an input function that generates a perceptual
signal. The whole structure of the HPCT model is predicated on
the assumption that ONLY perceptual signals are experiencable. If
other signals were experiencable, there would be no need for the
imagination connection. You'd just perceive the output signal.
But then, of course, you'd need appropriate perceiving functions
for each possible source of experience. Also, you'd have arrows
running from every part of a control system to higher levels,
instead of just copies of the perceptual signal (which is
anatomically correct, as far as we know).

If a copy of r can be used in the imagination loop within this
ECS, then it can equally feasibly be used as a category
perception that categorizes all of the possible combinations of
lower-level signals into the PIF of this ECS such that p = r in
this ECS.

Your assumption that the imagination signal represents r as drawn
is incorrect; all the imagination signal represents is one of the
perceptions that a lower system would produce if it controlled
perfectly. But beyond that, YOU are recognizing that some signal
could be used to stand for a category, but where in the diagram
is the capability of having that recognition? You're standing off
to one side, using your own ability to recognize categories, but
you haven't said what it is about the diagrammed system that
would be capable of reporting "that is category X." In fact there
is nothing in the diagrammed system that can do that. You need a
perceptual function that is capable of taking in information and
emitting a signal indicating that a category is exemplified in
the information. I claim that you're USING such a perceptual
function, but it's not in the system you're pointing to. If all
perceptions are the outputs of perceptual functions, then
perception of categoriness of any kind is the output of a
perceptual function. If you don't think that's true, then you're
starting a new model from scratch.

I don't know how to get this across. I've actually had people
tell me essentially the same thing about every higher level in
the model. It all depended on where they liked to spend their
time being aware from. People who like to operate from the
principle/generalization level have told me that principles are
simply the basic laws of nature, not a level of perception.
People who like to be logical have told me that logic is just
inherent in the way things interact with each other. People who
like to study relationships assure me that they study
relationships, they don't make them up. Now you, who like to
categorize, are telling me that categories aren't a level of
perception. What am I supposed to say? What would you say if
someone said to you that colors aren't perceptions, they're
attributes of objects? If I start backing down here, I'll soon
have to admit that relationships, configurations, sensations,
events, principles, yahdahyahdahyahdah, are really there and
somehow we just sort of KNOW about them, without the aid of
perceptual functions to construct them. The whole thing would
collapse, wouldn't it?
--------------------------------------------------------------
Best,
Bill P.

[From: Bruce Nevin (Mon 931025 13:16:25 EDT)]

Bill Powers (931022.1500 MDT)--

Haste and consequent carelessness again obscure what I am saying.
Yes, the diagram I sent shows a copy of e within the current ECS.
But that same signal e contributes to r in lower-level ECSs, hence the
label r.

If lower-level ECSs are controlling perfectly, the perceptual signal
that comes up to the current ECS (superordinate to them) reflects the
conseqence of each matching its reference signal r.

But r in each ECS on the lower level is not simply the error signal e
output by the current ECS. It is a product of combining error signals
from various higher-level ECSs in the reference input function of each
lower-level ECS. So it would appear that I am not justified in the
simplified diagram that I sent.

Let me try to explain the intuition.

Start with the (nonoccurring) case where there is just one ECS on the current
level n and just one ECS below it on level m:

          RIF.n This case does not occur because it omits that
           > which makes p.n and p.m perceptual signals of
           > r.n different order. It is the particular organization
           > of signals of order p.m in the PIF of the
           V n-level ECS (PIF.n) that makes p.n a signal of
    p.n +->C-+ e.n different order from p.m. (Conversely r.m is a
        > > signal of different order from e.n because of the
       PIF.n OF.n organization of various e.n signals into one r.m
        > > signal for each ECS.m.)
        > RIF.m
        > > But allow me this oversimplification for a moment.
        > >r.m If ECS.m is controlling perfectly, then p.m = r.m.
        > V But since p.m = p.n, then a copy of r.m
    p.m +--->C-+ e.m serves to replace p.m for the imagination loop.
        > >
       PIF.m OF.m From the point of view of ECS.n, whatever complexity
        > > of ECSs may exist on level m below it might
=======^======V===== just as well be a single ECS.m, it has no way of
        > > knowing how the signal e.n is copied and what
        > > WORLD is done with it. It is just an error signal emitted
        +-CEV<-+ into its environment for the effect of making p.n
           ^ = r.n. The aggregate effect of the various ECSs
           > disturbance on level m that actually receive e.n (however
           > combined with other error signals in their RIF.m
                            reference-signal input functions) is equivalent,
from the point of view of ECS.n, to a single ECS.m below it.

Indeed, from the point of view of ECS.n, it is equivalent to the simplest
case, a single ECS with effectors and perceptual input functions:

             > From the point of view of ECS.n, its error signal
            RIF.n e.n goes across the organism-environment boundary
             > and affects a CEV in the world, whence returns
             >r.n its perceptual input signal p.n across the
             V organism-envirnoment boundary.
    p.n +--->C-+ e.n
        > > From the point of view of ECS.n, all the lower
       PIF.n OF.n parts of the perceptual control hierarchy are
        > > within the CEV in the environment.
=======^======V=====
        > > In this simplest case, the imagination loop
        > > WORLD returns a copy of e.n to PIF.n, where
        +-CEV<-+ it becomes p.n.
           ^
           > disturbance But the quantity e.n is not the quantity p.n
           > nor is the output of OF.n equal the input of PIF.n
                             (when ECS.n is controlling perfectly).

> For total imagination, EACH reference signal sent to a lower system
> would have to be routed back to the input function of the higher
> system.

Going back to the multi-level case, if k copies of e.n are made in the
output function OF.n and sent to k ECSs on level m, then k copies of e.n
are also sent to level m, one to the PIF.m of each of those ECSs on level
m. Then each PIF.m combines its copy with other signals to produce its
p.m passed up to its comparator, with a copy to ECS.n. For total
imagination, each other signal entering each PIF.m would also have to be
imagined, that is, would have to be a copy of e.n from some other ECS.n,
as appropriate. But for one ECS.n (as shown here) to imagine, only the k
copies of its e.n are sent down to the PIF of each ECS.m to which it
concurrently sends a copy of e.n as a contribution to its reference
signal r.m.

So the diagram was oversimplified and wrong. I am glad to have
correction. But the diagram is an attempt to communicate an intuition
that does not depend upon the diagram. That is, I did not come up with
the intuition from pondering the diagram, I came up with the diagram as a
way of communicating and (I hoped) underwriting the intuition.

So here it is again. In some way, a reference signal can substitute for
perceptual input in the imagination loop. I may not have understood the
mechanism correctly. My intuitive leap says a reference signal that can
substitute for perceptual input from a CEV in the world can also be used
as perceptual input to a program-level ECS. It is in the PIF to the
program-level ECS that it becomes a category perception. It is "cut off"
from boss reality in just the same way as control in imagination is.

So far as I can see, the only justification for an ECS that recognizes a
category is to serve as PIF for higher-level "symbol-manipulating" ECSs.
Categories are not (so far as I can see) perceptions; categorizing is
something that appears to take place as I carry out programs, etc., in
practice or in imagination.

> "I need something to pry with" is
> an implicit request that can't be fulfilled unless you can
> perceive items as being in that category. Unless you have already
> perceived membership in that category, there's no way to know
> whether the words apply to that item.

With each thing that I perceive (in fact or in imagination) I imagine
prying with it, or I imagine you prying with it in your present task. No
need for a category perception. The process of prying with it (in
imagination) does the categorization. The category and a supposed
category-level ECS for "things one can pry with", need not pre-exist. I
make it up on the spot, or rather the process of prying (carried out in
imagination) makes it up. Similarly for seeing which in the stack of
umbrellas is mine (memory--you call that a category?) or perceiving them
as all umbrellas without distinction. I see one and can imagine opening
and using it, or wanting it if it's raining, or looking for my own
remembered one; and I see repetition of the same familiar configuration
in all the others in the pile. I don't look at them attentively, having
identified the one and having identified the others as the same
configuration, hence the perception of "not that particular example but a
likeness to a general kind of thing." (Your response to the word
"Mississippi" is a word-association thing, which brings in many
extraneous issues. Let's stick to non-verbal examples. I'll drop the
stuff about discreteness and referentiality for the same reason.)

> why assume that categories exhaustively define
> a domain of perception?

I didn't. I was emphasizinb that phonemes are different from categories
in this respect.

> As far as I know, the number of possible
> categories is unlimited and there is no need for them to be
> mutually exclusive. A doughnut and three dimes can be one
> category, and two dimes and two doughnuts another. If you have
> some reason to make such categories.

Yes, the appearance of categories is an ad-hoc byproduct of control at
higher levels, as in what we are calling program perceptions. Some
categories are conventionalized for the sake of social cooperation at
those levels. That's where language comes in.

> YOU are recognizing that some signal
> could be used to stand for a category, but where in the diagram
> is the capability of having that recognition?

On a higher level, e.g. program. It is not shown in the diagram because
the diagram shows (or was meant to show) only imagination. The claim is
that programs manipulate imagined perceptions; they can be applied to the
environment precisely because (and when) perceptual input from the
environment is controlled rather than copies of reference signals.
Thence the abstractness of programs and the categories they appear to
manipulate, when considered apart from one's actual use of them (e.g. to
get a beer from the fridge, in Avery's favorite example).

> Now you, who like to
> categorize, are telling me that categories aren't a level of
> perception.

What evidence have you for categorizing me as such a person?

Note that in doing so you are merely using language. The frame "one who
likes to __________" can be used to "categorize" as many ways as you can
think up words to put in the blanks. Is there one ECS for each? Aside
from the ECSs postulated for recognizing the word or phrase or short story
in the blank.

What evidence is there for a category level other than words? The fact
that you need them for programs, etc. doesn't mean much to me, PIFs at
that level will accomplish what you need for that.

     Bruce
     bn@bbn.com (but now officially employee of LightStream Inc.)