[From Bill Powers (950526.1100 MDT)]
Bruce Abbott (950526.0930 EST) --
As seen from the outside, most behavior consists of sequences such
as I described for the fly. Consider the directions for making
coffee: Remove the filter basket and check it for contents. If it
is full of old grounds, empty it. If it is empty, place a new
filter-cup in the basket. If it contains a new filter-cup, add
three measuring spoons-full of fresh ground coffee to the cup.
Replace the basket into the coffee maker. . . . and so on. Each
step describes the relevant sensory conditions and the behaviors
that should occur under each condition.
The behaviors are sensory conditions, too. In fact these instructions
say nothing about the motor outputs that are to be produced. The motor
outputs that will actually occur depend on the happenstance state of the
environment as each prescribed perceptual result is brought about. The
motor acts involved in removing the filter basket depend on whether it's
stuck and whether it contains old grounds and the angle of your body
relative to the basket and the distance your shoulders happen to be from
it. Emptying the basket, if needed, will involve motor acts that depend
on what is already in the garbage can, where the garbage can is relative
to your body and the basket, whether the grounds are wet or dry, and
whether the wet filter tears or stays intact. Placing a new filter into
the basket involves motor acts that depend on whether two filters stick
together, how far down in the plastic bag they are, exactly where on the
filter your fingers happen to grasp it, where the basket is and how it
is oriented, and again your bodily configurations relative to the
filters and the basket. And so forth.
When we name behaviors we are hardly ever actually naming outputs. We
are naming perceptual consequences that we are to bring about by means
of whatever output will accomplish them in the environment that exists
at a given moment. This is why control systems work as a model of
behavior, and open-loop systems don't.
The sequences are there; what matters is how you explain them: as
an S-R chain or as, for example, a program-level perceptual control
system.
If you were to watch flies landing on the ceiling over and over, and
make quantitative measurements instead of simply classifying similiar-
looking patterns as if they were identical, you would discover endless
variations in the approach and landing pattern. Because of these endless
variations in position, velocity, and acceleration, both linear and
angular, the motor outputs involved must also be endlessly varying to
keep the pattern converging toward the same final result. If the motor
patterns were not varying from instance to instance, their consequences
would be even less similar. The movements of the fly that you describe
are not actions, but consequences of actions. To keep these consequences
even moderately similar from instance to instance, the fly must be
producing very different motor outputs each time.
The problem lies in naming behaviors and then assuming that all
behaviors with the same name must be produced by the same motor actions.
The process of naming creates similarities and obscures differences. If
a fly "leaps" off a surface ten times, its motor actions might be quite
different all ten times, depending on the positions of the legs, the
adhesiveness of the surface, the state of the wings, and other processes
that were in progress at the time the leap interrupted them. Yet we make
all the instances seem identical by calling each of them a "leap." If
indeed the leaps are quite similar, we can only conclude that the
actions which produced them must have been quite different, because
careful observation would show us that the initial conditions were quite
different.
Your way of describing the sequence of events is the customary way, the
way that behavior is most commonly described. But this level of
description is actually a barrier to understanding what is going on. By
making differences disappear through classification, you cut out the
variations that would actually explain how the outcomes can be _nearly
the same_ when the actions are _markedly different_.
···
--------------------------------------
RE: Lobsters
I was more interested in the demonstration that different output
functions could be carried out by the same neural system as a
function of input state than in the neural oscillator function
itself.
From what I got out of the article, it seemed that the individual
oscillators just kept working the same way; what changed was their
synchronization. or whether some were turned on or off. This is more
like higher-level control achieved through actions on lower systems of
fixed organization.
The idea of using a set of lower-order components to achieve different
functions when driven by higher-order systems is surely nothing new to
us, is it? I use the same hand to scratch that I use to write. What's
the big deal?
--------------------------------------
What I meant by saying this sort of neurological analysis is a crock was
not that a competent analysis is a crock, but that an analysis that
simply invokes a new causal signal every time a new effect is needed is
a crock. To do a competent analysis of a neural system, you have to know
the input-output laws that govern each component, and derive the
behavior of the whole system from the interactions of the components.
Same problem as deducing what an electronic circuit does from reading
the circuit diagram. Unless you know the laws governing resistors,
capacitors, inductors, transistors, transducers, transformers, and so
forth, you are extremely unlikely to guess what the circuit really does,
even if you can identify input and output signals.
And as Rick pointed out, even if we did understand the neural circuit in
this sense, we still would have to know the rest of the loop to make any
predictions about what such a system would actually do.
In B:CP I laid out a few basic components and discussed what their
properties would be. Such components could be hooked together in many
different ways (as with electronic components) to achieve complex
functions. Of course it would be even better to have a set of components
which have been derived from the actual properties of neurons.
-----------------------------------------------------------------------
Best,
Bill P.