Categories, levels, decisions

[From Bill Powers (931101.0640 MST)]

Martin Taylor (931031.1010) --

I think you're letting yourself slip into a view that categories
are naturally existing things which we learn to perceive. While
that may be true, it may also not be true. To establish the truth
of that proposition, we would need some method of observing
"actual" categories to see if our perceived categories correctly
reflect them, without in the process using our category
perceivers. And that method we do not have.

Phil Runkel was telling me, a few years ago, that philosophers
were all excited about "natural kinds." It seemed to me then, and
still does, that this idea is an attempt to circumvent the basic
problem of perception -- in fact, to beg the question. All
evidence we have about "natural kinds" is perceptual, just as all
evidence we have about anything whatsoever is perceptual.

It may seem that the world forbids certain categories to coexist,
but people keep coming up with clever counterexamples. For
example, if we consider a two-dimensional surface with some
finite thickness, we find that we can control an object's
distance from that surface in a smooth way, maintaining any
particular distance normal to that surface, positive or negative,
with no restriction -- except for the apparent restriction that
we can't make a point be both "above" and "below" that surface at
the same time. As soon as we make that assertion, up pops Mr.
Moebius or Mr. Klein. Above and below, inside and outside, are
categories which rest on assumptions about the world.

I once took a car to a small repair shop to get a water pump
replaced. To get the water pump off, a "screwdriver" was needed.
However, there was no room to wield an ordinary screwdriver. A
right-angle screwdriver was required. But the only right-angle
screwdriver the mechanic had was bent too far from the tip, and
wouldn't fit in the space.

No problem. Chris Kotsiopolis lit his torch, heated the
screwdriver red-hot, and moved the bend closer to the bit, at the
same time making the angle of twist more convenient.

I confess that I was thinking of going to a hardware store and
trying to buy a right-angle screwdriver more suited to the space.
I was thinking in terms of a category; either something belongs
to a category or it doesn't. Right-angle screwdrivers are a
"natural kind," and in fact a small set of natural kinds. We
didn't have the right natural kind of screwdriver, so we were
stuck.

To Chris, however, a right-angle screwdriver was just one
configuration on a continuum of configurations, and by using the
right output operations he could alter the configuration in any
way needed to do the job.

Another incident along the same lines: I worked on a hay ranch
one summer in my youth. A 60-foot steel truss boom was used to
lift the hay, and it was too long to get across a bridge and make
the turn to avoid a chain-link fence. No problem. Bob Tomingus,
the same ranch hand who made a steam engine out of a brake drum,
got out his torch and cut the boom in half. After the pieces were
moved, he welded them back together. That was easier than taking
the other road and dragging the boom an extra 10 miles. To Bob
Tomingus, the category "boom" was optional, as was "brake drum."

I think that categories are strictly inventions, constructed for
a purpose. They simplify experience. But by simplifying
experience, they also limit it because they make the world more
rigid, reducing its degrees of freedom and the number of states
in which a perception can be. What we get for this is the ability
to grab "a screwdriver" without the need to set up a specific
reference signal for just one specific screwdriver-shape, so we
can get something in the right category without having to search
through all the screwdrivers for an exact fit. Any specific shape
that is in the category will do -- until the situation arises
where no item fits a predefined category. Then we pay for this
convenience by getting unnecessarily stuck (if we can't get our
awareness out of the category level).

Categories also permit us to do logic, which sometimes helps us
to be smart and sometimes helps us to be stupid. Some control
problems are so complex that only rule-driven symbol manipulation
can lead to a solution. But some are so simple that rule-driven
symbol-manipulation leads us deeper and deeper into thickets of
complexity that keep us from seeing the simple answer.

ยทยทยท

-------------------------------

Bill P. suggests that the scheme I proposed might be modified
by taking the cross-links between copies of the perceptual
signals representing the contrasting category members, rather
than between the originals. He says that one could otherwise
not simultaneously walk and perceive "I am walking." This
brings up a real issue, because I don't think Bill's suggestion
works, either. The important point is not that there is cross-
connection between the contrasting perceptual SIGNALS, but that
each of the perceptual signals is fed to the input of the other
perceptual FUNCTION.

That is what I meant: the signals cross-connect to the other
functions. How else would a flip-flop work? Signals don't affect
signals except through functions.

The perceptual signal coming out of a flip-flop function can be
in only one of two states. In either of those states, it isn't
variable (or it's nowhere near as variable as without the cross-
connection). Therefore the signal can't be used for smooth analog
control. If you want smooth analog control AND category control,
then you have to leave the original signal smoothly variable, so
it can follow a reference signal to any value within its limits.
To get the sharp-edged category, you have to take a copy of that
signal and feed it into a function whose output is cross-
connected to a function handling another analog signal, so that
the output of the new function is properly binary.
--------------------------------------------------------------

If one works with copies, the copies must be of the entire PIFs
of the ECSs concerned.

Why?
-------------------------------------------------------------

Let's consider walking while controlling for perceiving
"walking." From somewhere, a reference signal is provided that
sets a high level for the intended perception of "walking,"
which is contrasted to "standing," "running," "sitting," and
the like. All these others have low values for their reference
signals.

Fine. Now consider the ways in which one can walk. You start out
sauntering slowly, increase speed to a leisurely stroll, increase
more to a purposeful walk, then a hurried walk, then a heel-and-
toe sinuous flurry of hips and elbows. You walk smoothly or with
a bounce; you walk seriously or in a silly way. You walk in a
crouch or with your legs stiff.

If your only choices were between walking and standing, running,
or sitting, then you would perceive either "walking" or something
else that is not walking. You would perceive the category
"walking" or not perceive it. There could be no variations in
walking. The only meaningful reference signals would be 1 (walk)
or 0 (don't walk). You could then say "I'm walking" or "I'm not
walking", but you couldn't control your way of walking at the
same time unless the analog signals were also there, still being
controlled through all those infinite variations.
---------------------

The same is true for speech, but in a more complex way. A word
can be a pattern of sounds, references for which may be
provided to event-level ECSs by a word-level category system.
But it can also be a sequence of phoneme categories, each of
which can provide references to the same event-level ECSs.

And at the same time, the word can vary in smooth analog ways:
YES? (speak up!); Yeeesss? (skeptically); YES! (a goal by our
side); yes... (a reluctant admission); YEHhhss (now I'm beginning
to see). At the category level it's the same word with the same
categorical meaning of acquiescence, but at the analog levels its
meanings are infinitely variable. If you flip-flop-connect the
lower systems, you lose that infinite variability of the
perceptual signals.
-----------------------

The picture might look something like this (skipping many
analogue levels for ease of presentation):

                   c

    principles---a
     > > > t---\ -------------
     events------e \ | logical | other
    > > > > g------| |-----logical
configurations--o / | expressions | levels
   > > > > > r---/ -------------
  intensities----y

It could also look like this:

     Higher logical levels
               >
  C a t e g o r i e s
    > > > >
--- | | |

------- | |

----------- |

--------------

--principles
   > > >

--- events
   > > > >

  -configurations

   > > > > >

-- intensities

... which, save for specific levels, is the current arrangement
in HPCT. All you've done is to switch some levels around without
affecting the connections.

The big problem is putting principles below logic, and prior to
categorization. I don't see how a principle could exist prior to
categorization. But maybe you do.
----------------------------------------------------------------

An ECS within the logical hierarchy generates output if the
current perception does not correspond with its reference
level, and not if it does correspond. In that sense, a
"decision" can be seen as a choice of whether or not to
act.

Suppose the logic says, "If A=B, do C, else don't do C". This is
a "decision" as you define it. But what decision is necessary,
given B and A? I prefer to call this an (elementary) algorithm
with an outcome that is completely determined, given the
reference level B, as soon as the state of the perception A
becomes known. While you can define this as decision-making if
you like, this conception of a decision removes the problematic
aspect of decision-making. If one has algorithms already set up
which convert specific logical perceptions automatically into
specifications for action, the decision is made continuously and
immediately, changing automatically as the variables change
values.

My associations with the word "decision" imply conflict, not
immediate action. A decision (or a choice) is needed precisely
where existing algorithms lead to contradictory specifications
for action. When an algorithm exists, there is only one possible
outcome given the perceptual and reference inputs; there is no
choice between alternatives because there are no alternatives.
Decision or choice implies for me _reorganization_, required when
there is no existing basis for going one way or the other.

If, whenever it's raining outside, you ALWAYS "decide" to carry
your umbrella, is any decision needed? I should think that (what
I call) a decision would be required only if you had just one
umbrella and thought it would be polite to give the umbrella to
your guest who is leaving 20 minutes after you have to leave. Be
polite, or stay dry? That's the decision you have to make.
--------------------------

If Bill's argument applies, then no matter where the sensory
inputs to the category level come from, the outputs should go
only to the highest analogue level. This causes a conceptual
problem.

Yes, it's a problem. One solution, although awkward, might be
that in level-skipping cases, the intermediate levels are simply
turned off. This would prevent the conflict due to downward
level-skipping. Suppose one squeezes a fist in a rhythmic (event-
like) way. Then the event level could directly alter sensation
reference signals and interpret sensations in terms of a temporal
pattern, as long as no other system wanted to produce any
particular configurations or transitions using those same
muscles. I'm not terribly happy with that.

I think we need to build up some experience with what seem to be
higher levels of control. There are just too many possibilities
and complexities for us to guess correctly from our current state
of knowledge. If we actually try some experiments, we'll find out
what does and doesn't happen. If we just keep guessing there'll
never be any resolution of these problems.

That's enough speculation for now, I think.

You took the words out of my mouth.
-------------------------------------------------------------
Best,

Bill P.

[Martin Taylor 931101 13:50]
(Bill Powers 931101.0640)

Martin Taylor (931031.1010) --

I think you're letting yourself slip into a view that categories
are naturally existing things which we learn to perceive.

I tried to avoid that trap. Perhaps I was not explicit enough. What
I was distinguishing was contrastive categories--those that are of
the same kind, such as red and blue, or above and below, but that nature
does not allow to co-occur under normal conditions--versus constructed
categories, such as "something to pry with" or your right-angled
screwdriver.

What I argue is naturally occurring is the contrast. That is, that in
a variety of perceptual contexts, element A may occur or element B
may occur, but you never see both at once. A and B are of the same "kind"
because they tend to occur in similar contexts, not because they are
naturally occurring categories. Their ECSs presumably are at the same
perceptual level, though this is not necessary.

I had some trepidation in using the term "natural kind," much as I imagine
you did originally in using "perception" to cover arbitrary single-valued
functions of sensory input. It may have misled the rest of your interpretation
of my posting.

All evidence we have about "natural kinds" is perceptual, just as all
evidence we have about anything whatsoever is perceptual.

Yes.

It may seem that the world forbids certain categories to coexist,
but people keep coming up with clever counterexamples. For
example, if we consider a two-dimensional surface with some
finite thickness, we find that we can control an object's
distance from that surface in a smooth way, maintaining any
particular distance normal to that surface, positive or negative,
with no restriction -- except for the apparent restriction that
we can't make a point be both "above" and "below" that surface at
the same time. As soon as we make that assertion, up pops Mr.
Moebius or Mr. Klein. Above and below, inside and outside, are
categories which rest on assumptions about the world.

I only ask that the non co-occurrence be usual, not that it be mandatory,
and that it be not normally possible for actions by the HCS to generate
the two perceptions at once. Special cases, even rare situations in
which the two perceptions can co-occur, are permitted, provided they
are sufficiently rare in nature that it is advantageous to the organism
to discard them.

The underlying assumption is that the cross-linking happens in much the
same way as any other reorganization, by the improvement in control that
it permits. If things tend not to co-occur, it is likely to be better
at some stage to determine that the state is "A" rather than "B." To use
your example, the object is (in)edible. Control is easier if distinctly
different actions follow from one perception than from the other, given
a higher reference perception: Ref=thing to eat, Perc=thing not to eat,
Act=no eat.

Right-angle screwdrivers are a
"natural kind," and in fact a small set of natural kinds. We
didn't have the right natural kind of screwdriver, so we were
stuck.

I wouldn't call "right-angle screwdrivers" a natural kind. For one thing,
it is a function of two categories: right-angle, and screwdriver. Is
either of those a "natural kind?" I'm not sure. It depends on what they
contrast with. "Screwdriver" can be smoothly modified into "chisel" but
not, I think, "hammer." Maybe there is a screwdriver-chisel contrast that
would qualify each as a natural kind.

Maybe a better approach than to use the term "natural kind" would be to
say "member of a naturally contrasting group." The "chisel-screwdriver"
example reinforces the point that it is the affordances of the world that
matter--how the perceptions can be controlled. An object is a screwdriver
because it can be a part of the actions involved in controlling an event
perception of a screw going into a piece of wood, whereas it is a chisel
because it can be part of the action involved in a different event perception.
Some objects can serve both purposes, and these would be perceived as
being on one side of the contrast boundary or the other depending on context.

I think that categories are strictly inventions, constructed for
a purpose. They simplify experience. But by simplifying
experience, they also limit it because they make the world more
rigid, reducing its degrees of freedom and the number of states
in which a perception can be.

Don't you think I agree? The only shift I would make is of emphasis.
The word "constructed" can apply to logical expressions, or to the
reorganized construction of the relationship of PIFs. A "logical"
construction is labile, whereas a reorganized one is quasi-permanent
(and an evolved one even more permanent--if they exist).

If one works with copies, the copies must be of the entire PIFs
of the ECSs concerned.

Why?

Maybe I'm missing something. You agree that:

The important point is not that there is cross-
connection between the contrasting perceptual SIGNALS, but that
each of the perceptual signals is fed to the input of the other
perceptual FUNCTION.

That is what I meant: the signals cross-connect to the other
functions. How else would a flip-flop work? Signals don't affect
signals except through functions.

If the cross-connection is through the PIFs of a pair of ECSsthat control in
a smooth analogue way, they immediately stop their analogue behaviour, as
you say:

The perceptual signal coming out of a flip-flop function can be
in only one of two states. In either of those states, it isn't
variable (or it's nowhere near as variable as without the cross-
connection). Therefore the signal can't be used for smooth analog
control.

So, continuing the quote:

If you want smooth analog control AND category control,
then you have to leave the original signal smoothly variable, so
it can follow a reference signal to any value within its limits.

How can you do this, unless the analogue signal is generated by a copy
of the function that generates the signal that is cross-linked? You
can't just take a copy of the perceptual signal "A" and feed it into the
original PIF "B" without affecting perceptual signal "B." You would
have to feed it into a copy of PIF "B."

If your only choices were between walking and standing, running,
or sitting, then you would perceive either "walking" or something
else that is not walking. You would perceive the category
"walking" or not perceive it. There could be no variations in
walking.

Your key words there are "your only choices." Reference signals don't
come from only one higher level, do they? I assume that reference signals
derived in part from the output of a category ECS (or several) will also
have contributions from other ECSs, some within the analogue hierarchy.
Why not? At any rate, that's what I have always assumed to be the case
with the emotional modulation of speech (independent of this speculation
about flip-flop connections), and I see no reason it should be different
for walking, or making ice-cream.

    Higher logical levels
              >
C a t e g o r i e s
   > > > >
--- | | |

------- | |

----------- |

--------------

--principles
   > > >

--- events
   > > > >

-configurations

   > > > > >

-- intensities

... which, save for specific levels, is the current arrangement
in HPCT. All you've done is to switch some levels around without
affecting the connections.

The big problem is putting principles below logic, and prior to
categorization. I don't see how a principle could exist prior to
categorization. But maybe you do.

Non-committal on the last point; I haven't thought about it except to
say that I see no a priori reason to deny the possibility of a principles
level within the analogue hierarchy.

I don't think that your repositioning of the diagram changes anything.
In your way or mine, category is not "above" pronciples, although a category
of principles does depend on there being a perception of "principle" type.
I don't really like the term "above," because it can mean "immediately
above" or "anywhere above." Most discussion of standard HPCT uses the
"immediately above" connotation, because level-skipping is forbidden on
output in a well reorganized hierarchy, and is rare if not forbidden on
the perceptual side. The difference here is that there is no difference
in how far category is "above" intensity, "above" configuration, "above"
event, or "above" analogue principle (if such exists). So I think it
worthwhile not to think of the "above" relation in the same way when
dealing with categories. From category, we can talk about "above" within
the logical hierarchy in the standard way, but I see the logical hierarchy
as a side view on the analogue one (and perhaps vice-versa, to be even
more speculative).

Remember, I propose this structure as a straw man, not as a position to
be maintained. I will defend it or attack it, as I see the need. But
for it to be worth attacking, I have to show that the straw has some
bonding cement around it. Speculation cannot be avoided, I'm afraid.
Perhaps the discussion will show where it is worthless speculation, and
where there might be something worth following up.

On "decision," for example, there may be a difference in connotation.
You say:

My associations with the word "decision" imply conflict, not
immediate action. A decision (or a choice) is needed precisely
where existing algorithms lead to contradictory specifications
for action. When an algorithm exists, there is only one possible
outcome given the perceptual and reference inputs; there is no
choice between alternatives because there are no alternatives.
Decision or choice implies for me _reorganization_, required when
there is no existing basis for going one way or the other.

For me, "decision" implies that the program level is involved. A
program may have a conditional "If P " where "P" is a logical expression.
The "decision" is whether "P" is true.

As I read your connotation for "decision," it involves the construction
of a program-level ECS (PIF?), not its activation.

If, whenever it's raining outside, you ALWAYS "decide" to carry
your umbrella, is any decision needed?

Yes--Is it or will it be raining hard enough to make the trouble
of carrying the umbrella worthwhile?

I should think that (what
I call) a decision would be required only if you had just one
umbrella and thought it would be polite to give the umbrella to
your guest who is leaving 20 minutes after you have to leave. Be
polite, or stay dry? That's the decision you have to make.

You probably have a lot of sub-expressions involved in "If P" within
the program "If P then give umbrella." As I read the example, the
problem is in the perception of how much you want to please the guest,
how you perceive social convention to apply, and so forth. You probably
have principles that supply appropriate references (and I agree that
they are probably largely or totally in the logical hierarchy). I
should not have thought reorganization comes into it much, and even if
it did on the first occasion this situation occurred, would reorganization
again happen the next time? The decision would still be there to make.

I think we need to build up some experience with what seem to be
higher levels of control. There are just too many possibilities
and complexities for us to guess correctly from our current state
of knowledge. If we actually try some experiments, we'll find out
what does and doesn't happen. If we just keep guessing there'll
never be any resolution of these problems.

I'm taking home a new version of Allan Randall's Control Builder for the
Mac, and intend to try out a few experiments on the flip-flop mechanism.
Not much, but it's a start.

That's enough speculation for now, I think.

You took the words out of my mouth.

Speculation is a tool. Misused, it is damaging. Properly used, it is
constructive. I hope I'm being constructive, though at the end of that
posting, I began to feel that things were getting away a bit. I have no
(speculative) solution to the level-skipping problem of the side-looking
category interface.

Martin