Learning References

[From Rick Marken (960105.1010)]

Shannon Williams (960105) --

You cannot think in terms of making a perception match a reference until
you have a reference. Do you see why I am saying this?

In a negative feedback system there is always a reference, even if it is
implicit, at zero.

Me:

So my preferred definition of an "effective output" is "an output that
allows control of perception".

You:

Does this definition mean anything before a reference exists?

No. There must be a reference. But if the system exists in a negative
feedback relationship with respect to it's input, there is always at least an
implicit reference for input level.

I am only sure that, before you build a reference, you must have an error.

How can you have error before you "build" a reference? Error is the
difference between what is wanted (the reference for what is to be perceived)
and what is (the actual state of the perceptual variable). You can't build a
reference based on error unless you have a reference with which to compute
error (e = r-p; error = reference-perception). Fortunately, even if there is
no explicit reference signal (r) entering the comparator of the control
system, there is an implicit reference (r=0) when e =-p. That is, if the
comparator is just a signal inverter, then there is an implicit reference of
0 (e=0-p).

In PCT, the "intrinsic" reference signals are not necessarily explicit
reference signals; they may simply be the zero level of inherited perceptual
signal inverters. The implicit zero reference level of one such inverter
might specify the genetically determined concentration of O2 that must be
in the blood, for example.

So I don't think our learning system has to learn reference signals; but it
might well have to learn how to change the reference specifications going to
some control systems as a function of error in other control systems.

Best

Rick

[From Shannon Williams (960105)]

Rick Marken (960105.1010)--

You cannot think in terms of making a perception match a reference until
you have a reference. Do you see why I am saying this?

In a negative feedback system there is always a reference, even if it is
implicit, at zero.

Make it happen. Try to make your system evolve as a negative feedback
system.

Negative feedback occurs/just happens/crystalizes BECAUSE certain input
elicits certain output.

look at it this way:

  input = 13, output = move right.
  input = 12, output = move right.
  input = 11, output = move right.
  input = 10, output = 0.
  input = 9, output = move left.
  input = 8, output = move left.
  input = 7, output = move left.

The accumulative effect of these input/output is to center you at 10. 10
becomes a reference and the system begins to exhibit the characteristics of
negative feedback. (But only for certain recognizable input).

How can you have error before you "build" a reference?

In the beginning we have error from innate references. Just like you have
always said. But these references are not learned. Signals enter our brains
as pure error. All the brain 'knows' is that it is in an error condition.
At first, it does not have a method to counteract the error.

Error is the difference between what is wanted (the reference for what is
to be perceived) and what is (the actual state of the perceptual variable).

When you observe an organism generating output you say that he is in error.
You watch his environment change, and then you see him stop generating
output. At this point you say he is not in error. Whenever you change his
environment, he puts it back. So you call this his reference.

This is ALL you know.

Just because you have named one thing 'error' and another thing 'reference'
does not mean that you understand what they are. You are in no position to
demand that 'error' be defined as you do above.

You can't build a reference based on error unless you have a reference with
which to compute error (e = r-p; error = reference-perception).

Maybe error is not computed the way that you think it is computed.

So I don't think our learning system has to learn reference signals; but it
might well have to learn how to change the reference specifications going to
some control systems as a function of error in other control systems.

But all you have done is recite the Bible to me. Offer explanations from
the world, not from your PCT manual.

-Shannon

[Martin Taylor 950105 14:40]

Shannon Williams (960105)

Rick Marken (960105.1010)--

You cannot think in terms of making a perception match a reference until
you have a reference. Do you see why I am saying this?

In a negative feedback system there is always a reference, even if it is
implicit, at zero.

Make it happen. Try to make your system evolve as a negative feedback
system.

Well, at least we got that far with no problem in the Little Baby project.

There's really no trick at all to doing this with a "conventional"
reorganization approach, either Hebbian or discrete. "Reorganization"
happens withe higher probability when there is large error or rapidly
increasing error (at least in the LB). If an output affects its own input
at all, then if it affects it wrongly (positive feedback), a switch of
sign will occur soon enough, leading to negative feedback. If an output
doesn't affect its input, random shifts of action linkages will at some
point cause actions that do affect the input, and then you have the first
situation again. After that, all you need is Hebbian learning, perhaps
through the e-coli technique, perhaps by some other method.

Negative feedback occurs/just happens/crystalizes BECAUSE certain input
elicits certain output.

look at it this way:

input = 13, output = move right.
input = 12, output = move right.
input = 11, output = move right.
input = 10, output = 0.
input = 9, output = move left.
input = 8, output = move left.
input = 7, output = move left.

The accumulative effect of these input/output is to center you at 10. 10
becomes a reference and the system begins to exhibit the characteristics of
negative feedback.

You haven't said how "move right" and "move left" affect the value of the
input. If "move right" reduces the input and "move left" increases it, then
you already have a negative feedback system with a reference value of 10.
What's to learn? How did it get to be that way?

How can you have error before you "build" a reference?

In the beginning we have error from innate references.

What does that mean? I think you are using words rather differently from
their usual meanings within PCT. Error exists because something is different
from what it "should" be. "Reference" is a word meaning a "should be" value
for something; "perception" is a word for something that has a value. If
that perception is the thing for which a particular reference is the value,
then the deviation between the value of the perception and the value of
the reference is an error. If not, it isn't. How can you get an "error"
from a "reference" alone?

Signals enter our brain as pure error.

Huh? Is this Greek, or what? Let's think about what the words might mean.
A "signal" should be a temporal pattern of the value of some variable, for
which the changes are induced by the values of one or more variables at
another location or locations. If you mean something else, then you should
define your terms appropriately. The "signal path" is the means whereby the
variables at the distant locations affect the variable in question. So,
signals can "enter the brain" by way of the senses, which is what I suspect
you to mean. But then the result is a pattern of values of many variables in
the brain--about a million, if the "signal" is the set of outputs of the
optic nerve fibre from one eye.

"Error" is a difference between something and something else. It is a
relationship, whether you read the "PCT Bible" or not. You can't have error
without the possibility of not having error--of being "correct." So, for
signals to "enter the brain as pure error" there must be something that
these signals should be--the "correct" version of the same signals. What
is that?

All the brain 'knows' is that it is in an error condition.

All the brain "knows" is a set of values of many variables.

At first, it does not have a method to counteract the error.

Disallowing evolution, and assuming that we emerge into the world as a
complete _tabula rasa_, you would be right. For simulations, it's good
to make that assumption, and we do so in the Little Baby. I think Bill P
does in the Artificial Cerebellum, too; he certainly has done so in many
discussions of reorganization. But in saying this, we are talking about
"error" in the PCT sense, not whatever it is you mean by the word. We
assume that there are intrinsic reference values and perceptual signal
paths associated with them, to allow the concept of "error" to mean
something.

You can't build a reference based on error unless you have a reference with
which to compute error (e = r-p; error = reference-perception).

Maybe error is not computed the way that you think it is computed.

Maybe not. There are all sorts of possibilities. But they all come down
to a comparison between a "should be" and an "is thus." When one thinks of
the "signals" as scalar variables, the "-" sign is easily identified with
an algebraic subtraction, but it doesn't have to be. And when the "signals"
are more complicated variables, the operator "-" has to be taken as some
kind of a more generalized difference operator, such as a vector difference,
for example.

But all you have done is recite the Bible to me. Offer explanations from
the world, not from your PCT manual.

It's nice that you are thinking from first principles. Keep at it, but if
you are going to use concepts that are non-standard in PCT, try to identify
them with words that don't have standard usages in PCT, or else mark them
with asterisks or something so that we don't read your messages in the
wrong way. You are clearly using "error" in a way that differs both from
PCT and from everyday usage, so a different word would be appropriate.
I don't know what that word would be, because I don't know what concept
you are trying to convey.

Sorry.

Martin