[Martin Taylor 991104]
[From Bill Powers (991025.0410 MDT)]
Martin Taylor 991022 10:19--
This is a reprise of a notion that I expounded a couple of years ago,
about the "category level" and those above it. I am inspired to rewrite
it by Bill's self-doubts expressed in his message to Bruce Abbott,
because I think it resolves the issues that cause him problems.Bill:
As I think further about this, I'm inclined to change my mind about the
idea that an input function of a given level can recieve signals from _any_
lower level of perceptions. I can't explain what has swayed me -- there's
just a sort of general sense of difficulties associated with that
assumption. ..... I'm almost persuaded that
sequences belong below categories, and that the category level is where
symbols are created.Martin:
My proposal agrees with that.
Bill:
I recall your saying that the category level should be able to receive from
all lower levels, but not your saying that sequences belong below
categories.
I didn't say they do. I said that there would be an analogue level of
sequence perceptions and a discrete (post-categorical?) level of sequence
perceptions. I don't think of the category interface as being "above"
or "below" _any_ analogue level, since for every level of analogue
perceptions there could be a corresponding level of discrete (categorized,
symbolic) perceptions.
I don't see any principled way to decide which kind of perception is of the
higher level. I can think of examples that lead either way.
Exactly why I brought up the notion of the "category interface" as opposed
to the "category level." You don't need to decide which perception is at
the "higher level" They are both at the same level, though the discrete
perceptions at a level are, of course, derived from the analogue ones, so
you could, I suppose, designate "configuration discrete" as a level
above "configuration analogue" if you wanted. I wouldn't want.
Bill:
So: If anything, the program level would control a perception that is the
name of, or a symbol for, a category-level perception.Martin:
And with that.
That works for the way the program level would use symbols if the
categorizing level were just below programs,
or anywhere else between "programs" and the sensory-muscular interface to
the outer world.
but does not resolve the
problem of whether that is true. So we can't deal with the program level
properly until we settle the issue of whether categorizing lies above
sequence.
The proposal says it both does and doesn't. There's a sequence category
perception and a perception of sequences of categories.
And of course we are still faced with the problem that "program"
is a name for what the program level does, implying that categories lie
above programs. You're agreeing with contradictory statements.
I see no contradictions whatever. And I think that by conflating an
analyst's perception of an element of the hierarchy (the name of the
"category level") with the functioning of the element in the system
analyzed, you add little to the discussion.
Bill:
I agree. I'm not trying to dismiss the problems, only to express how I have
attempted to sort them out. Chances are that there are more levels than I
have detected in this region of the hierarchy. Or it could be that some
other architecture takes over at the high levels.Martin:
That is what I proposed and now re-describe.
Bill:
It's hard to see what you're saying you proposed, since I'm offering
different ideas as alternatives and "That" isn't a sufficient reference to
say which you're claiming to have proposed. You've done this for several
paragraphs above -- what "that" are you agreeing with, and what are you
ignoring?
I'm agreeing with all the alternatives you set up as mutually incompatible.
You find evidence in your thinking for each alternative to be correct.
I try to show how they are neither mutually exclusive nor contradictory,
and your subjective evidence for each alternative can be considered
independently.
Martin:
Simple statement (i.e., abstract or executive summary)
Bill:
I thought the "executive summary" was supposed to be the dumbed-down one!
It was! I simplified too much, it now seems.
Martin:
Analogue levels are organized as described in the "standard model of HPCT".
The "category level" produces symbols, but is not a level in the sense
that say "event" is a level.Bill:
How does it produce a symbol like "John" to refer to a person-object?
In my proposal, the category level produces symbols, not references
to labels. The _association_ between symbols generates references. One
symbol is the what you perceive as the person John (note: not 90% John,
but John or someone else). The other symbol is the word "John" (note:
not 90% "John") that is sensorily derived from an acoustic waveform, a
string of letters, or, often, memory/imagination.
Also, the reverse phenomenon exists. Perception of the category "John"
from an acoustic configuration is likely to result in the perception of
the image of the person referenced.
I
can't see it actually emitting that word!
Of course not. But it does make the perception of that word more probable
than if the person you saw were the one you label "Jane". And of seeing
the person viewed as the one labelled "John" if you hear the acoustic
configuration categrized as the word "John."
What I have been thinking is that
the output of a category-perceiver would be just a neural signal, as usual,
its magnitude representing the presence of the category detected by the
category-generating input function.
Same here.
When the signal is present, the
category is present. In some cases the category-signal could be
continuously variable, and in others it might flip between two states, or
even show hysteresis. Your "super flip-flop" with varying degrees of
positive feedback is one circuit design for accomplishing this, but I don't
really think it's time to start guessing about circuitry at that level of
complexity.
That's why my original posting (991022 10:19) dissociated the mechanism
for perceiving categories from the architecture of the hierarchy around
the category interface. Above, I referred to the _association_ between
categories (symbols) such as persons and words. One needs a mechanism for
association as well as one for category generation. The "grand flip flop"
produces both, so I didn't try to reprise it. How categorization and
association are done--the circuitry--is a separate matter from whether
they are done and how the results are used.
We have to establish the phenomenon first. And there are
alternative means of achieving the same result -- why settle on any one of
them?
Exactly. But the properties of the different possibilities should be
investigated before any are dismissed out of hand.
My biggest gap in understanding is how we can use word-configurations (or
spoken word-events) as the names of categories.
I think it's the same issue as for any other associations. If there exists
a mechanism whereby the perception of _any_ one category is made more
probable in the presence of the perception of _any_ other category, then
the association of a "name" category (a discrete word-configuration) to
any symbol (categorized perception) can use the same mechanism.
Martin:
I prefer to call it the "category interface"
since its job is to interface between analogue levels and digital levels.Bill:
Well, you're _proposing_ that that's what it's job is -- you shouldn't
refer to that proposal as if it's an established fact.
I don't think I did. I posted very tentatively, and labelled it a "notion"
right up front in the first line of my posting, and as a "proposal" several
times thereafter. However, _on the assumption_ that the proposal is
correct, then the system works in particular ways. And in my proposal,
the job of the category interface _IS_ to interface between digital and
analogue levels, one-to-one. Even in your classic hierarchy, the job of
the category level is to "interface between analogue levels and digital
levels", so I really don't know what you are complaining about, here.
Martin:
The assumption is that perceptual signals flow upward through the
analogue levels, being non-linearly transformed level-by-level, but
that the category interface accepts perceptual signals from all the
analogue levels.Bill:
A side-issue: saying that signals "flow through" levels has already proven
to be misleading (private communication with another person). I'd prefer to
avoid that image. My concept is that each level creates a new kind of
abstraction in which lower-level signals are not individually discernible.
The world in which one level exists is composed of systems of the next
level down.
Didn't I say exactly that in first part of the section you just quoted?
The specific proposal in the second part is that the category
level/interface is different precisely in having direct access to the
perceptual signals from all levels, untransformed.
The category interface accepts inputs from all analogue perceptual
functions at all levels, and produces symbols to which linguistic names
may sometimes be attached, but often are not. If the analogue perceptual
control hierarchy is visualised as a vertical array of levels, the
category interface can be visualised as a skin lying along its whole
height, like this:category
<-------> interface
>>> d | | analogue
>>> i | | levels
>>> g |symbols| -------------
>>> i |<-- -->| level N
>>> t | s | -------------
>>> a |<--y-->| level N-1
>>> l | m | -------------
>>> ><--b--> .......
>>> l | o | -------------
>>> e |<--l-->| level 2
>>> v | s | -------------
>>> e |<-- -->| level 1
>>> l |symbols| -------------
>>> s | ^ ^ | |
<--------> | | V V
> > > >
======^==^====V==V=========
external environmentThe effect of conceiving it this way is that the arguments against level
jumping in the analogue levels are fully applicable.Bill:
It seems to me they are fully inapplicable. You show a two-way interaction
between this category "skin" and each level, implying that the category
level can alter reference signals, and hence perceptions, at any level in
the analog hierarchy.
I guess that's a problem with ASCII diagrams. The "category level" doesn't
alter reference signals or perceptions at _any_ level of the analogue
hierarchy. In this diagram (but not in the alternative I drew just below
it in my original posting) the analogue perceptions cross from, say,
"analogue event" to become "discrete event" and the outputs from any
"discrete event" control unit go to the same places in the analogue
hierarchy as would the outputs from an "analogue event" control unit.
The interface itself works across units. There's no level skipping on
the analogue side.
Likewise there's no level-skipping on the disgital side, because the base
level (corresponding to sensation on the analogue side) is "symbol" or
"category". I like "symbol" as the word for it, because that's what
logicians use as variables in logical functions, and logical functions
are what the base digital level perceptions feed into. A symbol is a symbol
regardless of what analogue level it originally came from. And the output
of the lowest level control units on the discrete side are symbols. But
just as the perceptual symbols do each come from some specific level of
the analogue hierarchy, so do the output symbols feed to the reference
inputs at their own specific levels of the analogue hierarchy. Just
like any other signal in the system, a symbol is only a value of a neural
signal.
But doing that would disturb perceptions in any
higher levels of the analog hierarchy, so the higher levels would react as
to any disturbance, altering their own effects on lower-level reference
signals and cancelling any effects from the category-skin. So you've
ignored the condition that is my very reason for concluding that
level-skipping must not happen.
You see now how this is not correct?
All symbols are just that--symbols--within the discrete part of the
system. But in relation to the analogue part of the system, symbols
are of different kinds. Symbols for events are different from symbols for
configurations simply because they connect into different parts of the
analogue hierarchy. An "event" symbol output from the discrete part
to the analogue part is seen simply as would be a "high" value output
from an analogue "event" control unit.
You can allow the category-skin to receive copies of perceptual signals
from any lower system, but you can't allow its outputs to affect reference
signals in any level but the highest one in the analog systems. So it may
as well just be one of the levels.
I hope I've suggested (not shown) how that is not correct--however, you
may have noted also that I did show a top-level-only output to the analogue
system as an alternate possibility in my original posting.
Symbols may be
produced by the category interface from any analogue perception, but
not all analogue perceptions are categorized into symbolic form.What is "symbolic form?" And how does the category interface produce a
symbol "from" an analog perception?
"Symbolic form" means that the perception is there or it isn't. It has
been categorized. (Actually, that's an overstatement, in that my mechanism
doesn't produce strictly one-zero values for symbols. Their values can
vary somewhat, but tend to be either near maximum or near minimum).
You're making a huge assumption here,
or leaving a huge hole in your reasoning. How does the continuous flow of
sound-changes we know as the word-symbol "John" get produced by the
category level? That can't be at all how it works. You're skipping over a
whole lot of problems.
Two problems with this comment: (1) Who says that the association between
the symbol for the person and the symbol for a sound (or letter)
configuration must be produced by the category level? (2) In point of
fact, the mechanism I proposed some years ago (the grand flip-flop)
has as a side-effect of producing categories the production of just such
associations. I didn't want to bring it up, as I said above, because the
mechanism is a separate issue. One is enough to discuss at a time,
especially when it seems so hard to get this one very simple idea across.
Or is the comment intended to ask how speech recognition works? The
automatic speech recognition people have done a lot of work on that--which
is not to say that the machines work the same way as people do. Anyway,
if that's the question, I think it's inappropriate to introduce here,
though interesting in its own right.
I show a two-way arrow between symbols and the digital levels on the one
hand and the analogue levels on the other. This is too simple.No, it's hiding enormous complexities and problems. Are these digital
signals, having only two values? What happens when a digital signal enters
an analog system, especially a low-level analog system? How can a neural
signal have two forms, one being symbolic and the other being -- what, a
train of impulses?
Both are of the same form--values of a neural signal, which you choose to
think of as frequencies of neural impulses. The "high" value of a symbol
(an analyst/observer might see "symbol present") is a large value of the
neural signal; a "low" value is a small value of the neural signal. Where's
the enormous complexity and problem in that?
My model
allows symbols as both input and output to both sides, so that a symbol
can be input to an analogue OR a digital perceptual function (it's only
a value of a signal, not an entity like a word)What, isn't a word a signal? Aren't all experienced entities signals?
A word is NOT a signal (the map is not the territory). A word is an
analyst's concept. The perception of a word (or any other entity) is
a signal within the system being analyzed.
Martin:
Either way, I think this architecture can resolve some of Bill's misgivings,
and at the same time answer the question of how an apparently low-level
perception can seem to set references for program-level actions.Sorry, Martin, but your "architecture" is just too full of unresolved
issues to strike me as useful. It does nothing to resolve my misgivings and
uncertainties.
I hope that you are now a little closer to seeing what the proposal _is_.
If that is so, and you still don't see how it resolves any of your
misgivings about the "classical" hierarchy, we can discuss why that might
be, and what might be a proposal better than either. But until you start to
talk about the actual proposal, it's hard to discuss it technically.
I can't see any unifying principle in it.
Well, here's one. Any kind of perception can be categorized, and is likely
to be if so doing improves control of intrinsic variables. The presence
or absence of a perceived category would be the reference for any control
system using such symbolic (categorized) input. Categorized values are
amenable to logical analysis and control, and there is no apparent reason
why the nature of the original (analogue) perception should matter to the
logical analysis. The outputs of logical-valued control units are likely
to be logical, and there is no apparent reason why the output value should
relate to the same kind of analogue perceptual level as the input.
You _say_ that
certain things are acccomplished by it, but I see no reasoning at all
behind such statements; however you reached your conclusions, it wasn't by
any public process that I could follow.
Try again. Have I made it public enough yet? I've reduced the simplification
a little, but I do acknowledge that the current re-description is still
vastly over simplified.
And please don't mix up the mechanism proposed for categorization and
association with the architectures that use those functions. Both the
"classic" hierarchy and my proposal need both mechanisms. But the nature
of those mechanisms can be argued and experimented separately.
Martin