Category perceptions

[From Bill Powers (940224.1730 MST)]

Martin Taylor (930424.1140) --

In the standard hierarchy, there is a "category" level. I
observed that categories can apply to almost any analogue
perception where the world seems usually to have a logical
exclusion, such as that a blue object is not at the same time a
red object. Using this observation, I proposed that the
category level, which is already an interface level between
analogue and symbolic control levels, should be conceived as
being connected simultaneously to all lower levels, rather than
just to the event (?) level. This places the symbolic levels
beside rather than above the analogue levels.

The fact that a signal indicating "blue" is not the same as a
signal indicating "red" doesn't have anything to do with
categories (necessarily). Different perceptions are carried on
different pathways; they are in different places. "Not blue" does
not mean "red," and vice versa. Nor is there anything to prevent
a particular color from appearing both blue and red (to different
systems): for example, purple will excite both blue-detectors and
red-detectors at the sensation level. Some of the names we give
to colors are there strictly to handle cases that lie between
whatever we have learned to call "pure" colors (which is hardly
standardized).

In your definition of a category level, it seems to me that you
are focusing on an almost-irrelevant side-issue (mutual
exclusivity) and failing to answer the main question, which is
_what kind of operation generates a category perception_?

I proposed that the *kind* of perception involved in the
category level was analogous to the operation of a flip-flop,
in that the perception of one of a set of mutually exclusive
categories actively inhibited the possible perception of any of
the others.

This proposal assumes that we already have two systems which
respond to inputs by generating signals that have a categorical
meaning. This skirts the primary question: what kind of operation
will provide a signal when one or more members of a category is
present? Whether or not the signals are made mutually exclusive,
we must state how perception of either an apple or a grape can
lead to the category-perception "fruit." This is quite
independent of whether we go further to propose that "fruit" and
"vegetable" are to be mutually exclusive, a result that could be
achieved by a positive-feedback cross-connection between the two
category-recognizers, or simply by a logical convention. In
principle, this mutual exclusivity could be established between
any two categories: "house" and "automobile," for example -- and
furthermore, people could argue as to whether an object could be
both at the same time (a motor home).

When we get into the realm of logical argument about whether two
categories are or are not, or should or should not be, mutually
exclusive, then we are no longer talking about basic capacities
of the brain. We are just using brains to manufacture arguments:
pushing words around in a logical framework.

The essence of the category level is that a number of disparate
input signals, meaning quite different lower-level perceptions,
come to evoke _one single signal_. A wallet, a house, a computer,
a car, all evoke in me a sense that they are "mine", and that any
one of them or any combination of them is "mine." This has
nothing to do with objects that are not mine -- the essential
problem is to say what kind of perceptual operation would create
this relationship between lower-level perceptions and a category-
level perception. The simplest kind of operation I can think of
is the "OR". A category signal is produced by or-ing together all
the signals representing the individual members of the category.
Nothing else seems required.

Mutual exclusivity can be added at the logic level; it is not
needed, and indeed is not called for, at the category level. Two
categories can be mutually exclusive in one context yet can
coexist in another. If I say that no daughter of mine will have a
child out of wedlock, I am saying that a person can't
simultaneously be my daughter and an unmarried mother, so if my
daughter turns up pregnant I must either force her to have an
abortion or disown her to keep the exclusive-or true. Yet in
another context, I would have to admit that this person is in
fact both an unmarried mother and my daughter (this is not an
autobiographical example). The mutual exclusivity depends on my
logic, not on the categories "unmarried mother" and "daughter."

All this indicates to me that mutual exclusivity is not a
property of the category level, but of the way the logic level
chooses to deal with category signals. I am shown a purple
object, and am asked "Is it blue?" I answer "Yes, somewhat." If I
am asked "Is it red?" I also answer "Yes, somewhat." Only if I
feel that objects must _logically_ either be one color or the
other must I choose one word -- and then I would have to say that
since it is not exclusively blue or exclusively red, it is
neither, and it must be purple and only purple. But that is not a
categorical consideration; it is a logical consideration. And one
person's logic is not necessarily another's.

At the stage in which our model-building exists now, I see no
merit in trying to add complexities to it. Rather, we should be
trying to boil it down to the simplest form possible that still
covers the main phenomena. We need to define levels that are
simple enough to test. Treating every proposal in the basic model
as a springboard for leaping to other possibilities, especially
irrelevant ones, does not further our progress.

···

---------------------------------------------------------------
Best,

Bill P.

<Martin Taylor 940224 21:00>

Bill Powers (940224.1730)

"Not blue" does not mean "red," and vice versa.

But "Blue" does mean "not red."

Nor is there anything to prevent
a particular color from appearing both blue and red (to different
systems): for example, purple will excite both blue-detectors and
red-detectors at the sensation level.

Exactly! At the sensation level it excites red detectors and blue detectors.
So does a green wavelength. But the colour category you describe is
purple, therefore "not red" and "not blue."

In your definition of a category level, it seems to me that you
are focusing on an almost-irrelevant side-issue (mutual
exclusivity) and failing to answer the main question, which is
_what kind of operation generates a category perception_?

I thought it was quite the opposite. I concentrated on the operation
that generates a category, and mutual exclusivity within the class is
the side-effect.

Let me ask you--what operation did you think generated a category perception
when you proposed the existence of a category level below sequence and
program, but above event?

I proposed that the *kind* of perception involved in the
category level was analogous to the operation of a flip-flop,
in that the perception of one of a set of mutually exclusive
categories actively inhibited the possible perception of any of
the others.

This proposal assumes that we already have two systems which
respond to inputs by generating signals that have a categorical
meaning.

Huh?!? This comment passes me by utterly. Is it the case that if
we have a sensation perception we presuppose that we have a perceptual
function that responds to that sensation? I presume it is. So what
is the point of the comment? Does it say more than that if we have
a perceptual signal that represents a category we have a way of generating
that signal? And where did that gratuitous word "two" spring from?
In my proposal, there are N perceptual signals that the world tends
to produce in a more or less common context, but seldom at the same
time as one another. Out of those N signals, our perceptual systems
have organized so as to take advantage of this exclusivity, and to allow
us to act so that a desired member of the class comes to be perceived.

Whether or not the signals are made mutually exclusive,
we must state how perception of either an apple or a grape can
lead to the category-perception "fruit."

You forget that the proposal included a distinction between a logical level
at which such combinations can be made arbitrarily, and a (lower) category
level that provides the basic categories out of which the combinations
can be made. You can do anything with logic, once you have the symbolic
basis for it. The category perceptions provide that basis, in my proposal.
And as I understood it, they did so in your original proposal as well.

So far as I can see, the only difference between your original proposal
and mine is that your category level receives input and provides output
to exactly one other level of the hierarchy, whereas mine connects to
all of them. (And I have suggested an explicit operation to generate
the category perceptual signals that may be different from yours, as I
don't remember what yours was).

The essence of the category level is that a number of disparate
input signals, meaning quite different lower-level perceptions,
come to evoke _one single signal_.

I take that to be the essence of a logical combination level that is
not explicit in your hierarchy. Its functions seem to be split between
your sequence and program levels. Maybe that's right, but I think that
a separate logical level that is neither sequence nor program feels right
to me. In the category level as I see it, the signals that come to
form one category are topological neighbours--in other words, you can
change one into another with no discontinuous changes. This is not the
case with the categories you discuss, which I called "derived" categories
in the earlier discussion.

the essential
problem is to say what kind of perceptual operation would create
this relationship between lower-level perceptions and a category-
level perception. The simplest kind of operation I can think of
is the "OR". A category signal is produced by or-ing together all
the signals representing the individual members of the category.

But you can't "OR" analogue signals. You can add them, divide them,
multiply them, correlate them... but to "OR" them, you have to turn
them into logical signals, a job that I assign to the category level,
and had misunderstood you also to do.

Mutual exclusivity can be added at the logic level; it is not
needed, and indeed is not called for, at the category level. Two
categories can be mutually exclusive in one context yet can
coexist in another.

I think it is quite the other way round. Logically derived categories
can be mutually exclusive or not, depending on context. For the most
part, categories based on observations of nature cannot. I say "for the
most part" since our observations may have been a poor sample of what
nature is prepared to provide us with, something we never can know. So
we may reorganize to develop category perceptions from our observations
and success in control that are unwarranted as we observe further. But
all we can deal with is what we have perceived, and the best reorganizing
system in the world can't prepare us for a world of a kind different from
the one we have so far encountered.

...
All this indicates to me that mutual exclusivity is not a
property of the category level, but of the way the logic level
chooses to deal with category signals.

The examples you choose are ones I would have assigned to the logic level,
regardless of how the category level perceptions are generated.

I am shown a purple
object, and am asked "Is it blue?" I answer "Yes, somewhat." If I
am asked "Is it red?" I also answer "Yes, somewhat." Only if I
feel that objects must _logically_ either be one color or the
other must I choose one word -- and then I would have to say that
since it is not exclusively blue or exclusively red, it is
neither, and it must be purple and only purple. But that is not a
categorical consideration; it is a logical consideration.

Quite so. In fact, you are verging on a very interesting issue here,
but I think it is another tarbaby issue that is inappropriate for discussion
at this stage. It relates to the control of gain and the appearance of
the (trendy science, Rick) cusp catastrophe whose end point is exactly
the appearance of the category perception when a higher reference signal
demands that there be one. You are asserting the perceptual correlate
of a mechanism previously postulated. I'll be happy to interact privately
with you on this topic, but I fear being dragged into fruitless (neither
apple nor grape) discussions if we go into it publicly.

At the stage in which our model-building exists now, I see no
merit in trying to add complexities to it. Rather, we should be
trying to boil it down to the simplest form possible that still
covers the main phenomena.

I think that's always true. It's a restatement of Occam's razor.
But there are indeed phenomena relating to categories and logic,
especially including language, which is where all this discussion
proceeded from the last time it came up.

I might ask again--if it was not worthwhile dealing with such topics,
why did you include them in BCP?

We need to define levels that are
simple enough to test.

Pretty hard, when even Rick proposes a two level model for combining
pursuit with compensatory tracking, and neither of them is a "classic"
level.

Treating every proposal in the basic model
as a springboard for leaping to other possibilities, especially
irrelevant ones, does not further our progress.

Progress, I think, requires a good balance between keeping the feet on
the ground here and now, and scouting out possible paths up the mountain
we want to climb later. To put it in another way, the lower reference levels
are set from the errors at higher levels, and control is needed both
high and low.

I think it is fruitful to look at the implications of mechanisms. In
fact, that has been my whole approach to PCT, and the wealth of results
it produces are the source of my sometimes excessive enthusiasm. In the
case of categories, I looked at the question "what would happen if the
perceptual signals at some particular level were fed back as sensory
inputs at the same level rather than only to a higher level?" The answer
is that flip-flop-like operation would occur (and has been modelled
in our control-builder program).

The result is the development of classes of mutually exclusive
perceptual signals. Mutual exclusion is a property of contrastive
categories, and so, perceptually, is the hysteresis that typically
is part of category perception. Similar hysteresis is a property of
the cross-linked perceptual functions. The properties of the model
match the properties of a particular perceptual type, at least to first
order. So it seems plausible to me that this "contrast" process is what
produces the basic categorical perceptions that we do have, and that
these are the raw materials on which we can base the kind of logical
categories you discussed in your posting.

There's no proof that it is so. But at least the model works as intuition
says it should. Try it in Simcon (remembering that the perceptual functions
must saturate--it won't work with infinitely linear perceptual functions.)

Martin