[From Bill Powers (920623.1100)][Delayed]
Martin Taylor (920622.1930)]
You and Bruce are right that the Test requires anthropomorphizing. But
Sagan and Drayer did it wrong in two ways.
First, they were crossing species. Instead of using the lowest common
denominator of description, they used the highest: they described the
(potentially controlled) perceptions in words appropriate for human beings.
I would pick terms that imply the least sophistication of perception and
control, not the most, in making my guesses about another species.
Second, the description of the social interactions employs stereotypes that
probably don't reveal what even human participants are controlling for in
such situations. There's a bit of cynicism in the description: I'll bet
Carl Sagan never stood in a crowd and reached out to touch a charismatic
leader like George Bush, or that he ever bowed down in awe before any other
person. That's just his (their) stereotype of how the common masses behave.
If S&D could get inside the heads of most "common" participants in rituals
and ceremonies, they might be shocked at finding how much of the same sort
of cynicism is there. If Sagan could hear people reacting to him when he
strikes an imposing pose on television, he might be furious, as well.
Most people, I think, behave in ritual situations as they do not because
they have some corresponding inner conviction, but because they think
everyone else does. You take off your hat, stand, and put your hand on your
heart when the anthem is played and the flag is raised at Yankee Stadium,
because a show of patriotism is what's expected -- even if you're Canadian
or an anarchist. The main pressure is concerned with controlling for
appearances, not beliefs. If you act acceptably, you can believe anything
you want. Your real goals are your own business.
A great many of the silly and boring things that people do become
understandable if you know that lots of people are controlling for
stereotyped actions that they see others carrying out, simply to conform
with what's expected in the situation. I don't think there are nearly as
many True Believers as there seem to be.
When we find out what people are really controlling for, I don't think we
will find much relationship to commonly-held stereotypes like those that
Sagan and Drayer described (or imagined). The most commonality will be
found in the overt behaviors that other people can see, and by which
individuals are judged in social circumstances. This could even be true in
a chimp society. The gestures, if produced simply because they're expected,
don't have much to do with anything except the penalties for nonconformity.
Inner and outer conformity, however, are not the same thing.
···
-----------------------------------------------------------------------
I don't find CSG revolting, and I don't want to. Jolting is OK, but
revolution seldom has good results, politically or scientifically.
Seldom isn't never. There have been some pretty good revolutions among the
bad ones.
There's lots of good food for thought out there among the garbage. >Some
just needs to be made a bit more tasty by being taken with a grain >of
salt.
Salted garbage is still garbage. I don't want to sort through garbage for
something that looks edible, even if it's there. When things get messy
enough, it's time to go to the grocery store and start over, even though,
in principle, you could sort out the peanuts from the coffee grounds.
-----------------------------------------------------------------------
RE: perception of multiple Xs
Bill points out that in principle one can only [track]the objects if >the
targets maintain differences from the distractors in at least one >of
position, velocity and acceleration. That's quite true, and
experimentally verified. But even after two have been confused, the
observer can still track about four (depending, I think, as Bill says, >on
the motion statistics of the ensemble). One of the four may be, >from the
experimenter's viewpoint, wrong, but the observer sees the one >that best
fits whatever tracking method is used by the ECSs that are >employed.
I don't understand the "But...". In the "gating" model I proposed, the
tracking systems don't give up when they confuse two objects. They go right
on tracking the n objects, as if the original ones were still being
tracked. The problem is that the systems are NOT confused. If four objects
were being tracked, four continue to be tracked. They are not necessarily
the original set, but the tracking systems don't know that. If a given
tracker knew that there was more than one item in the gated area, it
wouldn't confuse two items. Such gated tracking systems simply center the
centroid of whatever is in the gated area, on the ASSUMPTION that it's a
single object.
As I see it, the key problem is more like: third, how does the brain
identify that there exists N objects of a certain kind rather than one
strong exemplar of the kind. I've forgotten how Bruce originally put >it,
but perhaps I can paraphrase. If there is an ECS controlling for >the
perception of an X, it should satisfy its reference if an X is in >its
input. What distinguishes thes existence of exactly three Xs in >its input
from the existence of one X.
If the ECS is controlling for the perception of "an X", then any number of
Xs greater than zero at its input will result in zero error. If it is
controlling for "exactly n Xs", then there are two ECSs involved: one
elementary control system can handle only one variable. The second ECS is
controlling for number of objects. Another level is required to handle the
"exactly" part: the number must be neither less than nor greater than n,
AND for each item taken alone, the item must yield the perception of X
rather than Y. This is a logical condition -- at least a relationship (the
AND relationship).
In the pandemonium model, the presence of an X anywhere in the visual
field, with any size or orientation, is enough to produce perception of X-
ness. I proposed several posts ago a model in which each instance
contributed to the total X-ness. That was wrong -- that would mean that X-
ness is not independent of the attribute of position. All that can change
the amount of perceived X-ness is that some or all items, while still
recognizeably X, differ from the canonical X to some degree. The number of
items in the visual field brings in a new dimension of perception.
If you're presented with a visual field containing dozens of Xs and Ys, and
are asked to find 3 Xs, you will be able to do so, but only by serially
searching among the items and discarding those that aren't Xs. If there is
no contiguous group of 3 Xs, you probably won't be able to construct a
perception of 3 Xs, easily or perhaps at all. You'll have to handle it by
counting at the symbolic representation levels, putting gates around
qualifying objects, and so on.
If all the items are Xs, then the X-ness perception is satisfied wherever
you look, and the 3-ness perception is satisfied by any configuration of 3
objects. If explicit perception of position is introduced (it doesn't have
to be), we could also require that "there-ness" be perceived - i.e., a
specific location or set of locations, either predesignated or specified by
relationship ("not the same").
For any of these ideas to work, I think we have to assume that judgements
are made within some limited region around the center of vision, or
attention, or both.
While this still doesn't handle all problems, it does help us to remember
that an ECS perceives ONLY ONE PERCEPTION. In describing any experimental
situation in words, it's easy to pack multiple dimensions of perception
into unitary-seeming sentences or phrases, like "three Xs." n-ness and X-
ness are independent dimensions of perception, and can vary independently.
To see "n objects of a given kind" is to see (1) n objects, regardless of
kind, and (2) a given kind of object. If the additional condition is set
that (3) no OTHER kind of object is to be perceived, another dimension of
perception is introduced. If the condition is that (4) n and only n objects
shall be perceived, we have still another condition to satisfy and
perceive, which would be different from the condition (4a) at least n
objects, or (4b) at most n objects. Also, "see n objects and see X-ness" is
a reference condition different from "see n perceptions such that each is
an object AND is an X."
When you have to analyze a perceptual situation down to individual and
independent perceptions, a lot of ambiguities are usually revealed. This is
one of my beefs with psychological experiments: most of them are ambiguous
in ways their authors overlooked because of taking informal verbal
descriptions for granted. For the same reason, most such experiments
involve phenomena that are immensely more complex than the authors
recognize -- this is one reason for the lousy correlations that are
published.
----------------------------------------------------------------------
Bill said in some posting shortly after I joined this group that >whenever
he had thought of a level-jumping control, it turned out not >to be
(appropriate/correct/necessary/simple?). Maybe so. In the >neural network
business, multilayer perceptrons often work well, but >there are other
architectures that are more appropriate for complex >problems, including
specifically gated modular architectures in which >smallish modules solve
sub-problems, and gating structures determine >which subnetworks present
their solutions to higher modules.
See Gene Boggess' excellent post from 920622. If all you want to do is
solve a problem, there are many ways to do it. But if you want the solution
to be consistent with neural architecture and function, behavioral facts,
and subjective experience, the possible choices are greatly reduced.
In BCP and elsewhere I said that level-jumping perceptions seem to be
permissible, but jumping levels in the downward direction leads to a
problem. The reason in general is that a control system receives its net
reference signal from multiple sources. If a system of level n+2
contributes a reference signal to a system of level n, the result will be a
disturbance of all systems of level n+1 that are using the same level-n
system to achieve their own goals. Those systems will react by cancelling
the effect of the level n+2 output. Furthermore, the level n+2 system does
not perceive in the right terms for attaining a compromise with the
competing level n+1 systems: it would be like looking for a simultaneous
solution of a set of equations in which the variables are x,y,z, and
"consistency".
This problem doesn't exist in every possible case; it is obviated when
intervening levels don't come into play -- when, for example, there are
transitions under control, but no specific events. Then you can control
relationships among transitions.
Perhaps control nets might work better on complex problems if conflicts
can be resolved by gating structures that permits some modules but not
others to exercise control?
This is already part of the HPCT model, although not all MODES of control
are "official" yet. Particularly at level 9, logic and programs, it's
possible for control systems to work by sending reference signals to some
lower systems but not others: select one sequence rather than another, or
where sequence is not controlled, one category of action rather than
another. And of course even within the program level it's possible to have
nested subroutines -- why not? That's just a matter of what program is
running, not of basic organization. Anything that can be programmed in a
neural computer can run at the program level. There's nothing about any
particular program that characterises human nature. I wouldn't be surprised
if ALL propositions at this level can be found exemplified in SOMEONE'S
head -- if only the head of the guy who thought up the proposition.
-----------------------------------------------------------------
Simplicity is to be preserved where possible in science, and HPCT is a
simple structure. If it can solve problems as basic to perception as
"there's an X and there's another" or "I see a lot of Xs among the Ys"
without the introduction of new structures or concepts, so much the
better. At present I don't see the answer.
HPCT is expressed in terms too general to provide specific answers to
specific questions -- particularly when they are questions about the
content, rather than the form, of behavior. HPCT is really a toolkit. The
claim is that by using the elements of this model, by making specific
propositions that employ only these elements, one can construct
explanations of all behaviors and organizations of behavior. The only way
to test this claim is to see whether the design principles and parts list
in the HPCT model will suffice, given enough creative effort. Of course if
the effort involved is greater than the effort needed when using some other
kit of ideas, we would become suspicious of the HPCT model and either
modify it or abandon it.
This is why I can't answer a lot of questions, like "What does HPCT have to
say about cognitive dissonance?" It doesn't have anything to say until you
go through the effort of creating something to say using the concepts of
HPCT instead of whatever concepts underly "cognitive" and "dissonance" (if
any). All the problems of behavior that come out of other approaches are
defined in terms of a different conceptual toolkit and a different assumed
parts list, usually irrelevant to the concepts and components of HPCT.
Tho most difficult questions to answer are those that contain implicit
theoretical assertions, or unconscious assumptions, or taken-for-granted
interpretations. I've been asked, for example, "What causes anxiety?" Such
a question really stumps me at first, because not only can't I answer it as
stated, but I doubt whether it HAS an answer as stated.
Under HPCT we don't deal in things that "cause" things in a simple 1-2
manner. Terms like "anxiety" don't refer to any objective state of a
system, but to how a person perceives an internal state. An internal state
of what? Evaluated with respect to what reference level? And of course in
asking for a "what" that causes anxiety, one is asking for a force that,
acting on a system, makes it respond in a specific way -- and in the HPCT
model, that never happens. The real answer to a question like "What causes
anxiety" is "Nothing. There's no thing called anxiety, and there's no cause
for things that don't exist, and anyway 'causation' is the wrong concept."
And then, of course, to handle the PHENOMENON with which this person is
concerned, one would have to start from scratch and study it as a
phenomenon of control. It's a common predicament that for me to answer
certain classes of questions, I'd have to do 10 years of research in HPCT
terms (which the asker wants me to report on instead of doing it himself or
herself).
A basic premise of HPCT, that an elementary control system handles only one
variable in one dimension of variation, forces us to analyze phenomena into
single variables of just a few classes. Doing this usually reveals that we
have posed more than one question, thinking it was a single question. It
doesn't really matter whether the premise here is absolutely true all of
the time; just thinking that it's true is often enough to lead to
reformulation of a problem in a way that suggests multiple solutions.
--------------------------------------------------------------------
Gene Boggess (920622) --
Unfortunately, not much AI is actually based on ANY model of human
intelligence; "whatever works" seems to be the watchword. So I don't
know that HPCT will have an overwhelming impact on the field; but it
could, and it should. It is certainly affecting how I think about my >own
AI work.
A lovely post -- so THAT's what's been going on! I think a lot of people
would be interested in your description of just what changes in your
approach to AI took place. I expect that many people on this net have
suffered through a paradigm shift like yours, and will have a deep
understanding of the struggle and the cost involved.
--------------------------------------------------------------------
Best to all,
Bill P.