[From Bruce Nevin (2003.06.13 20:54 EDT)]
Bill Powers (2003.06.08.1258 MDT)–
What you propose is a class of controlled
variable
to be added to to the model, or if not added, at least identified
within
available features of the model.
I conceded to Rick that dependences are relationships, so there’s no new
class of variables.
What I propose is a more extensive and more active role for associative
memory in the model.
We know we have to start
with simple things and then add details; just diving into the deep end
of
the pool before paddling around in shallow water for a while is
dangerous.
A memory bit flipping on and
off is sufficient to indicate that something has happened, and serves
as
a memory of an occurrance when all you need to know is whether it
has
happened.
When that is all that we need to know, then this suffices.
When we need to know more, a more elaborate model of memory is necessary.
If that model can also account for this simple requirement, then would
the above still suffice? (That is, redundantly.)
Any lasting trace can be called
“memory”, but not all memory is the
equivalent of a full-blown tape recorder.
The following is not an example of this.
The retina, for example, retains
the effects of a flash of light for something like a tenth of a second,
so
we continue to see the light after the photons are no longer
arriving.
Should we call that a “memory?” I don’t think so; asserting
membership in
that abstract class tells us nothing we don’t already know.
The point of the example is rather that not all retention of effects
should be called memory. (Although such metaphors are common in studies
of materials.)
Time-delays are
involved in perceiving rates of change, but I doubt that they share
any
important properties with other processes we call
memory.
The basis for this doubt is not clear. In a canonical control loop, there
is only the present moment, the delta of r-p and its transforms around
the loop. Loop delay has no bearing on perception of rates of change. I
suppose it is possible that rate perception is a class of controlled
variables to be added to the model, but that hasn’t been proposed so far
as I know. I had assumed that change, and rate of change, were derived
from relationship between memory and present input, but my basis for that
assumption is weak indeed. (Does it require a perception of the age of a
memory, begging the question?) Rate of visually perceived motion could be
a derivative of effort required to maintain apparent motionlessness in
the visual field, some kind of cousin of Rick’s ball catcher.
Then there is the fact that some perceptions
… take time to form, and an equal time to change state.
Yes, in the model the time it takes for continuous control, however
time-consuming, is not a function of memory storage and retrieval.
See? You’re designing with neural currents
already.
I anticipated that someone might think that my point is to reject logic
design and so it seemed best to demonstrate understanding at least.
Note
that the state of A’ indicates one or more occurrences of
A.
In general, I should think a sequence ABC isn’t expansible as
AABBCC*
(that is, with Kleene star notation, AABC, AAABC, ABBC, etc.), so
this
design is departing from the phenomenon.
OK, if you want to say that ringing the doorbell to get someone to open
the
door won’t work if you ring more than once.
So when can a step be repeated? Sounds like a do-until loop in a program,
not a sequence.
But this is an inapt example. Suppose getting someone to open the door is
a step in a sequence. It is controlled by various means, starting with
ringing the doorbell. If someone sees you coming, and you see through the
window that they’re getting up and coming to the door, you don’t ring the
doorbell. If you ring the doorbell and no one comes, maybe you ring it
again, maybe you knock, maybe you peer through the window, maybe you go
around to the back door. These are not in any sequence, and the openness
of choice about them indicates that no one of them is “the”
current step in that sequence; rather, they are alternative means of
getting someone to come and talk to you. The ringing of the doorbell
might be repeated, but is iteration a sequence? “Hammer until nail
is driven and set” might be an iterative loop, interruptible by a
bent nail or a bruised thumb. But ringing the doorbell is not iterated in
a mindless loop “ring bell until door opens” except perhaps by
some small children (not all!) doing what they have been told is supposed
to work. (Setting aside bell-ringing as provocation, where the iteration
itself is controlled as such.)
Here’s the confusion of ends and means again:
I don’t think you want to say
that. In cases where multiple occurrences of A are not equivalent to
the
sense that “A has occurred,”
If I am typing the sequence ABCDE, multiple occurrences of C as in
ABCCCCCCCCCCCCCDE definitely constitute an error, even ABCCDE. The only
reason for iteration is if to accomplish the current step the pressing of
the C key has some effect and the first try didn’t complete the step by
having that effect. So if the typewriter key is bent and the one for the
letter C bounces off the guides and doesn’t reach the paper through the
ribbon (remember those?), or if the keyboard is defective and a C doesn’t
appear on the screen, maybe I press the C again and again until it does.
Or maybe I interrupt the sequence to fix the mechanism. Or press the
space bar and write it in with a pen later. Different means to complete
that step and move on to the next.
you need to state in what way the second
occurrence changes the situation, and build that change into the
model.
Obviously, the effect of the “second” A can’t occur until a
previous A has
occurred, so we’re still talking about a related problem. We need an
indication that A has occurred at least once.
A provision for temporal delay and
possible interruption between terms of a sequence probably would as
a
byproduct limit each term to one occurrence.
If that special rule, for some reason, is really what is required of the
model.
What sequence cannot be interrupted?
Now
consider the statement, “B cannot occur unless A has occurred.”
This
statement as it stands does not say that if A does occur, B will
occur.
Yes. It only says if you’re controlling B you can’t do so unless you
are
already controlling A (or the result or product of controlling A). I
can’t
drive to the store without first getting into the car. I can’t replace
the
swing-latch on the fence without first cutting a piece of wood of
appropriate size.
If you’re controlling B you can’t do so? Do you mean that if you try
to
control B, you will fail unless you’ve already controlled A? I think
there’s some confusion here between states (which are persistent) and
state
transitions (which are transient occurrances).
I’m controlling getting enough sleep, being rather short of that lately,
but I can’t actually go to sleep right now as I have a meeting at 3:30.
So I plan and expect not to get up at 6:00 tomorrow because I don’t have
to get my daughter off to school on Saturday. If someone were to disturb
that CV with a need for me to do something early tomorrow I would
immediately resist the disturbance. That demonstrates that I am
controlling that perception, as means of getting enough sleep, even
though I can’t get enough sleep right now at this moment. There is a
dependency between my staying in bed past 6:00 and my getting enough
sleep.
In each case, what is the
“flip-flop” that indicates that the previous
event has occurred? In the first case, the event of getting into
the car
is indicated as having occurred by being seated in the car, isn’t it?
You
can’t drive until you’re seated in the car. In the second case,
suppose
you’re interrupted with an emergency that takes you away for a couple
of
hours, so when you get back to the workshop, how do you know whether
you
finished the event of cutting the piece of wood before the phone
rang?
I perceive the state of the piece of wood. I might not go look at it
until I had perceived the state of the gate and ‘remembered’ (leaving the
question of what that means unanswered) my control of it being fixed, the
sequence as means of control, the steps of that sequence that have been
done, and the step that was not yet done when I was interrupted. But even
in that degree of forgetfulness and reminder, I know that I hadn’t
finished cutting the piece of wood by perceiving the state of the wood,
either in memory or directly in the environment when I go look at
it.
I
will assume that the occurrence (change of state) of B depends not only
on A’ being on, but on some other event taking place: D (a control
action, I suppose). If D occurs while A’ is on, B will occur. If D
occurs with A’ in the off state, B cannot occur.
You don’t need D etc. if when I say “A occurs” I am referring
to control of
A, as above. My use of the word “occur” has confused
things.
Has it? What happens if you try to control B before A has occurred?
I can control waking up later than 6:00 tomorrow before I go to sleep
tonight. I can control the length of the piece of wood being 7
inches (say) before selecting the wood and putting it on the saw table.
This is of course a confusion of ends and means. The B (waking up after
6:00, the wood being 7 inches long) comes temporally after the A (going
to sleep tonight, selecting and cutting the wood) but the control
relationship between them is not sequential, it is contingent: the
subordination of means to ends. Here, different means may be substituted;
in a sequence ABCDE nothing else, such as the letter F, may be
substituted for C. ABFDE is not the sequence ABCDE. But if I find an
otherwise appropriate piece of wood that happens to be 7 inches long I
don’t have to run it past the saw blade anyway so that I can move on to
the next step.
But my objection to your introducing A’, D, and so on was that they are
not necessary. I don’t need some program-level CV in order to perceive
that the gate is not fixed (which originally occasioned my perceiving in
imagination the steps to fix it and starting to carry them out, and which
now evokes those perceptions from memory) or that the wood hasn’t yet
been cut (which reminds me “where I was” when I was
interrupted).
You can
apply an output to B, I suppose, but under the rules,
Under the rules? In these examples, it’s under the environmental
conditions.
B will not occur
because A has not occurred. You can try to start the car, but you
can’t
because you you haven’t got into it yet. “Trying to control B”
is the sort
of thing I meant by “D”.
An unanswered
question is whether these logical relationships are bein
imposed on perceptions, or whether they express properties of the
environment.
Both, so long as the control loop is closed
through the environment.
“Imposed on” was not a good expression. I was asking whether
the condition
that you can’t do B until A has occurred is a property of the
environment,
or an arbitrary rule you have adopted (you can’t have dessert until
you’ve
eaten your vegetables).
Socially mandated rules are arbitrary but they are part of the
environment. They are controlled as means to an end. Social ends in the
environment constituted of social relationships.
I assumed you used
the word “logical” because this is an exercise in logic
design, and not because a logic level of the hierarchy is involved. But
now
you take that as an obvious step.
Either way, it’s pretty clear that a
logic-level control
system is needed to provide the required variables D and E in the
proper
order to allow C to occur, and to perceive the states of A’, B’, maybe
C’,
D, and E.
That’s quite a leap, and an unnecessary one. If “A occurs” etc.
refers to
control of A etc. then there is no D etc.
OK, you’re saying that if you control A, you WILL immediately be
controlling B. I don’t think so. I think that you can try to control B,
but
it won’t be possible unless A has occurred. If you try to control B
by
applying the appropriate output, you will fail unless A has
occurred.
“Applying the appropriate output” is D.
I don’t follow. If I try to cut the wood before picking it up I will
fail, not because of a problem of logic, but because the wood is not
brought against the saw blade. This is an environmental contingency. Is
the example inappropriate?
We had the surmise that memory provides
reference signals. This ran afoul
of the origination of reference signals in higher-level error. No way
has
been specified for error output to address memory yielding
reference
signals that vary appropriately.
Why get hung up on that? There are many ways to convert a signal of
variable magnitude into an address that picks one discrete item from a
list
of them.
And at lower levels then converts that discrete item back into a smoothly
varying reference “45 degrees up and to the left of the mark and 1
inch away”?
This is especially easy if we suppose that
this mode of
control-through-memory is confined to the higher levels where all
signals
are symbolic rather than analog. For that kind of control system it
would
be simple to write a computer program that would convert each error
signal
into a choice of a remembered lower reference condition that would,
if
achieved, tend to reduce that error because it did so in the
past.
So the proposal only works at higher levels? If there is a different
mechanism at lower levels, why not use it throughout? So the fact that
this is easier at postulated higher levels accomplishes nothing for your
argument.
But let’s get the trotter back in front of the
sulky. I don’t insist that
reference signals come from memory. What I do insist on is that the
model
be able to reproduce a perception that occurred in the past, the
process
called “doing the same thing again.” I insist that the model be
able to
create signals standing for perceptions that have actually occurred in
the
past, and to make sure those signals are sent into the perceptual
pathways
where that kind of signal will be correctly understood. I insist that it
be
able to create perceptions in imagination as well as by acting on
the
environment. I insist that we be able to examine potential reference
signals to see whether making perceptions match them would produce
the
result we want, and to do this without putting those reference signals
into
active service until we want to. I insist because these seem to be
phenomena that require explanation. My memory proposal was a fairly
simple
start on an explanation.
We are of course in violent agreement.
Got something better?
Is this belligerent phrase suppose to mean that I should say nothing
until I do? That the search for something possibly better should be
silent and private? Please be more explicit.
When perceptual
inputs that originate in memory are controlled, how is the
control loop closed? Not through the environment. Through the
imagination?
switch? That would have to be imagination switches, plural, many, down
to
whatever level of detail at whatever levels of perception are evoked in
the
memory and affected by control.
As many as are required to explain what we experience, no more and no
fewer.
I liked your example of working with mooring your boat, but am not
sure
what point you were making amid all the details.
It’s too easy to diddle facile abstractions without highly specific
examples on which to test them. So this was a fairly specific example.
And I couldn’t see how the imagination switch could account for it.
One question was where do remembered perceptions get into the loop. I was
starting from the notion that they can’t get in as reference perceptions
by interposing memory between error output and reference input because
the reference input is continuously variable. (As in: “perceptions
are carried by neural signals that indicate only how much of a given
type of perception is present: how much intensity, how good a fit to a
given form, how much like a given relationship, and so on.” (Bill
Powers 2003.06.13.0750 MDT).) If we say that the variability itself - the
rate and direction of change, and the rate and direction of acceleration,
etc. - is stored in memory then we have stored plans for behavioral
outputs.
They seem to get into the loop as though perceived in the environment. Of
course this could be an effect of references being looped back to input.
But as I say I don’t see how they can come in as reference input. So
maybe indeed they do come in as perceptual input. Then we control them
with our usual references. (Indeed, we must, in order to try things out
in imagination.)
In the standard view, input signals can be references looped back to
input by the imagination switch. If the imagination switch is closed
“in as many places as necessary” (a cute evasion), then
references could come in as a consequence of controlling a few perceptual
inputs that come from memory. It’s the same commodious vicus of
recirculation (sorry - obscure literary reference).
From memory come perceptions of the eye in the end of the line and the
cleat on the bow deck of the boat. Suppose these enter the hierarchy from
memory as perceptual input signals going up the hierarchy. They go up to
a “loop over cleat” relationship controller, which in turn
sends signals to a system for which that relationship perception is means
of controlling the boat being fast to the mooring. That system starts
controlling the “loop-over-cleat” relationship perception. But
now the input from memory has the loop separate from the cleat. (I
witness this happening. I don’t know what occasions the change.) To the
“loop-over-cleat” system, and consequently to the “make
the boat fast” system, these inputs are disturbed relative to the
reference. Perhaps the signal is an “amount” of that
loop-over-cleat relationship, and the amount is lower than the reference,
so the relationship controller outputs an error signal. This is mapped to
a reference value for the location of the loop and a reference value for
the location of the cleat.
Why do those perceptions arise from memory? I think maybe the higher
systems start controlling, and in the absence of sensory input they evoke
input from memory.
If this is so, then remembering and imagining are really
indistinguishable. In either case, perceptions arise from memory because
a control system is controlling in the absence of sensory input. If the
process of controlling seems to us to change the perceptions, we call it
imagining, and if it doesn’t seem to we call it memory.
I agree withn these observations. It’s not
necessarily done through
associative memory; in fact I’m uncomfortable with settling on any
particular mechanisms here. We have a long way to go before guessing
in
that much detail will get us anywhere.
Saying that these signals come from memory really says nothing at all
about mechanism. It only says that they don’t come through sensors from
the environment.
So where does this leave us? Sort of
inconclusive.
If memory is local, as your RNA guess suggests, then the evoked input
signal starts out vague, completely unspecified at lower levels. That
fits my subjective impression. By persisting with control - and that
requires a degree of discipline in continuing to control the same
signal, not skipping around - we enable reference signals to propagate
down to lower systems. As we control the inputs of more systems at more
levels, the subjective impression is of a clearer, more specific sensory
experience.
Here of course we touch on the basketball players improving their foul
shots with no ball in their hands, Maxwell Maltz & Co. and the
varieties of ‘creative visualization’ discipline generally.
Here, too, we may find that we have a plausible account of hypnotic
suggestion and kindred phenomena.
For now, this reply has been delayed long enough, or perhaps too long,
and the window for making it is closing again.
/Bruce
Nevin
···
At 05:37 PM 6/8/2003, Bill Powers wrote: