NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/bit/listserv/csgl/600

Text File | 1992-07-21 | 11.0 KB | 202 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!darwin.sura.net!ukma!psuvax1!psuvm!auvm!BEN.DCIEM.DND.CA!MMT
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Tue, 21 Jul 1992 18:33:21 EDT
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         [e-mail]
Subject: Re: More catching up
Lines: 191

[Martin Taylor 920721 17:30]
(Bill Powers 920719.0800)

We were disconnected from the Internet for a couple of days, and have just
been reconnected.  Nothing I posted since (including) Saturday 18 July has
shown up here, and I don't know whether it got out.  I suspect it did, from
some comments in what Bill and Rick wrote.  If so, it probably bounced back
to the Mail Server.  I hope it returns some day, for my HyperCard stacks, if
nothing else.
=====================

>To lump the CNS hierarchy with the biochemical one is to ignore time-scales
>and relationships to the immediate world of the senses. It is the CNS
>hierarchy that produces the overt behavior of organisms and our own ability
>to observe and make sense of that behavior.

Well, this gives me a perfect intro to a paper I was going to mention anyway:
Behavioral Hypothermia and Survival of Hypoxic Protozoans Paramecium
caudatum, by Gary M. Malvin and Stephen C. Wood, Science, 13 March 1992, 255,
pp1423-1425.

Malvin and Wood start from the proposition that hypothermia is believed to
be a protection agains reduced oxygen conditions (hypoxia), and that many
species there is either a behavioural response or a physiological response
that cools the organism if it gets hypoxic.  "In mammals, hypoxia induces
hypothermia by decreasing heat production and increasing heat loss.  In
ectotherms [I think that means lizards, insects, and the like. MMT], hypoxia
induces behavioral selection of a lower ambient temperature."
(Right there is a suggestion that control can transfer between CNS and non-CNS
modes over an evolutionary time scale.)

M&W wanted to see whether an organism without a nervous system would control
its temperature differently as a function of depressed oxygen levels. They
chose a single-celled protozoan.  "Two specific hypotheses were tested (i)
Hypoxia causes paramecia to select a lower temperature in a thermal gradient,
(ii) survival of hypoxic paramecia is increased at lower temperatures."
Both hypotheses proved true.

I do not see where your reference level comes from for segregating effects
that act solely through the CNS from effects that occur by other media.
CNS events can affect physiological states directly, and vice-versa.  Why
should the systems have to be considered as separate.  Applying your own
criteria of faithfulness to what is observed in life, I would have thought
that separating CNS and biochemistry into two systems would have been difficult
for you.  The Paramecium exerts behavioural control of an intrinsic variable
without a CNS.  Pre-mammals presumably used behavioural regulation of
temperature, as lizards now do. Mammals developed physiological means to
do the same thing.  The controlled percept is the same.  Isn't it easier
to imagine that we added a physiological ability to a bahavioral ability
that we continue to use, within a single hierarchy rather than to imagine that
we developed a new ability in a separate hierarchy to duplicate an ability
we had anyway?


>When you say that no controlled variables have meaning to the behaving
>systems, you commit a peculiar reductionistic contortion, using knowledge
>and meaning to deny knowledge and meaning. Isn't meaning one of the
>phenomena we're trying to explain in terms of the organization of a complex
>system? Don't higher systems perceive, and even think about, information
>contained in the signals arising from lower-level systems, thus giving
>meaning to these lower-level processes?

I suspect we are talking at cross-purposes here.  I think of an ECS as dealing
only with neural currents or similar measurable variables.  Within the ECS
they don't have meaning..."An ECS has to do what an ECS has gotta do."  Whether
the perceptual input functions of high-level ECSs themselves "perceive, and
even think about, information contained in the signals arising from lower-level
systems" is, I think irrelevant.  Maybe they do, maybe they don't.  What would
it mean to the ECS if they did?  It wouldn't change the result that a perceptual
value must be compared to a reference value and the resulting error
transformed into an output that is distributed the ECS knows not where.

> How is it that an intrinsic
>variable, which represents the internal state of the organism, can lead to
>effects that alter the details of the way the organism controls variables
>external to the organism? This, not control of local conditions by local
>action inside the organism, is the heart of the problem of reorganization.

I agree, except that I am not convinced that one has to go beyond control
of local conditions to obtain the systemic result.

> I want to explain how this kind of
>reorganization can work on any level in the hierarchy, regardless of the
>cause of the error, so that a child can learn to add 2 and 2 as a way of
>alleviating chronic physical discomfort, or can learn to exercise and build
>up muscles as a way of doing the same thing.

Yep, that's what I am getting at.  I can see how my system would do that, though
I haven't modelled it, and my experience with the time-bomb in the hierarchy
leads me not to trust intuition too far.  But I can't see how yours would do
it for the reasons already discussed. Any way you get out of the degrees of
freedom problem (at least so far proposed) strikes me as Ptolemaic.

>I resist the pressure, which comes not just from you but from large
>segments of modern science, to search for global generalizations that will
>wrap up the grand principles of behavior as Einstein expressed the
>principles of General Relativity in four simple (-looking) equations.

Maybe so, but I think you failed in this resistance some 40 years ago.  I
think that you have indeed produced a "grand principle of behaviour," perhaps
the principal principle.  All else is symmetry-breaking, and the specification
of the specific symmetries that have been broken.

> True generalizations, I believe, arise from considering
>details, not generalities. And even the best of them eventually fail and
>must be modified.

Probably true.  All scientific laws are provisional descriptions.  But they
are helpful descriptions unless they guide us away from even more helpful ones.
Failure is relative.

==================
On my commentary with respect to Allan Randall's posting...

>>I suspect Allan was getting more at the question of whether a group of
>>ECSs at a level can act as a population vector ...
>
>Not in a way that specifically controls the vector rather than one variable
>contributing to it (see my post of yesterday).
>
>>... "given that we have sensors only for THIS red, THIS green, and THIS
>>blue, how is it that we can control for a pretty close match to a wide
>>range of blends."
>
>The obvious answer to this penetrating question is not very satisfactory to
>me. In simplistic terms, the target color has to be remembered (all three
>components, in different control systems), and then each remembered value
>must be reproduced by varying the individual color dimensions. My biggest
>problem with this explanation is that I don't experience colors in terms of
>their trichromatic components, but simply as whatever colors they are.
>Another problem is that people can compare two colors and pronounce them
>similar or different (hearking back to a previous conversation).

(I don't quote your "more pregnant" solution, because it goes off in a different
direction that is unaffected by the following comment.)

The core of my question, and of my interpretation of Allan's, has to do with
the concept of "coarse coding."  Coarse coding is a very important principle
that we seem to use a lot, all the way from the retina (which uses it both in
space and in colour) to the motor systems.  The colour system is probably the
simplest to describe, but the concept applies in multidimensional spaces
equally well.  We have in our retinas three types of cone, all of which respond
to light over a wide range of frequencies.  Call them R, G, and B.  Forget
B for the moment.  The sensitivities of R and G vary across frequency, R
being most sensitive to a colour we might call orange, and G to a yellow-green.
But we can make exquisite discriminations over a wide range of frequencies
because anywhere between green and red the ratio of the R cone outputs to the
nearby G cone outputs depends strongly on the frequency. We perceive some
function of this ratio as a colour.  For bluer colours, and for non-spectral
hues, the B cones come into play, and we use the ratios among all three--or
more correctly, between the B and the sum of R and G.

In general, coarse coding depends on there being a set of detectors each of
which is most sensitive to some specific values of some features that can have
a continuum of values. These would be the perceptual outputs of some ECSs
(i.e. the outputs of their perceptual input functions).  The ranges of values
to which these detectors are appreciably sensitive overlap substantially,
and a subsequent system can obtain precise information about the location of
some entity in the feature space by deriving ratios among the outputs of the
detectors.  The resolution that can be obtained by this procedure can be
much greater than the grain of the detector centres in the feature space.

Coarse coding is very robust against loss of a detector (even in the colour
system, a person missing one of the ayatems still can work reasonably well
with the other two), as well as giving fine possibilities for control.  It
is at the heart of my porposal that coordination be considered an emergent
of the hierarchy--that no single ECS actually controls a specific percept
at any level, though each acts as if it does.

This is different from the notion of a vector representing the multiple
features of a percept, since in coarse coding the detectors all represent
different central values of the same features.  I have tried to distinguish
three ways ECSs can relate within a level, based on their connections at
lower levels: conflict, cooperation, and coordination. (I'm not sure I have
done it on this mailing list, but it is in the current draft of the Paris
paper, which is still not finished.)

>I can't visualize "recurrent reference connections." What does that mean?

I was thinking within a level, if the output of ECS A contributed to the
reference level of ECS B (even indirectly) and vice-versa.  Such connections
do not exist within your scheme, but there are all these little niggly bits
of data that suggest it might happen among percepts.  If it did, then it
should be expected among references, too.  All I wanted to say was that we
can't rule it out, even if we don't want to think about it yet.

Best for the Durango meeting.  Durango was featured on the front page of the
Travel section of the Toronto Globe and Mail on Saturday.  Sounds like a
nice place, if somewhat touristy.  Wish I could come.

Martin