[Martin Taylor 2019.01.09.17.14]
[Rick Marken 2019-01-09_10:39:27]
I think I now understand the source of our misunderstanding, and
I can quite sympathize with what you say at the end of your
message. So I requote the last bit of your message here, before I
try to clarify a little according to my understanding of the free
energy principle.
RM:
The thing I liked about Powers theoretical work from the
start was that it was always tied to experimental
demonstrations of the phenomena to be explained. There was
no hand-waving or attempts to impress with mathematics
heavy laden with Greek symbols. It was always clear how
the math related to the phenomena to be explained and the
math itself was always presented clearly, like Bill
actually wanted you to understand it.
I do so agree. But I think you are a bit unfair to Friston when
you say…
Friston's
approach would have impressed me when I was a grad
student. But I can now tell when math is being used to
make an impression rather than provide understanding. And
I am not impressed by Friston’s free energy principle and
lovely mathematics.
I don't think that's what Friston does. The problem, I think, is
that different people have their perceptual functions differently
developed. If you are used to controlling perceptions of
equations, it is natural to communicate through equations. If you
are not, you do not see the equation. You see a bunch of symbols,
some of them Greek, some representing functions that mean nothing
to you, but you don’t see any higher-level variable. If you see a
bunch of Arabic or Tamil text, you don’t see any words, and you
don’t get any sense out of the sequencing of the funny symbols,
but if you see English text written in Roman alphabetic symbols,
you don’t really see the symbols or perhaps even the words. You
see the meaning of what is written.
When I was an undergraduate Engineering Physicist and for several
years thereafter, I could look at an expression for, say, a
Fourier Transform or a Laplace Transform and see in my mind a
dynamic visual picture of a scene in which properties changed in a
coordinated fashion according to the equation. I can’t do that
anymore, because it is a long time since I ever controlled any
such perceptions and the perceptual functions seem to have
corroded away with age. But at the time, it was as natural to
communicate to myself or to others through the language of
equations and time-space transforms as through English text. I
don’t read Chinese, and you don’t read maths. But then I don’t
disparage those who communicate in Chinese.
Now let’s see if I can help with the rest of your comments.
[Martin Taylor 2019.01.08.23.09]
MT: Friston's insight doesn't ignore a fact
at the core of PCT, rather, it uses a fact
that PCT ignores.
RM: What is Friston's insight? What is the fact
that Friston’s insight uses that PCT ignores? What
data does Friston’s theory explain that PCT
doesn’t?Â
MT: That has been explained in the WIRED thread, and if
that is insufficient for you, you can look it up in
Wikipedia or in Friston’s writings.
RM: The closest I get to anything that sounds like PCT from
the wiki is that the free energy principle says that a system
“t ries to
minimize the difference between its model of the world and the perception  of its
sensors”. But things kind of goes off the rails when it says
that the difference between perception and world model is " minimized by
constantly updating the world model". This seems analogous
to PCT saying that an error signal is reduced by changing
the reference for the perception.
I don't interpret it that way. I understand this version of
“world model” to to be represented not only in the current values
of perceptions but also in the perceptual and output functions in
the hierarchy. I don’t see it as applying to reference values,
except inasmuch as disturbances to perceptions at level N are
opposed by altering reference values sent down to level N-1
control units.
This is
a way to “control” the error, keeping it at 0, but it
certainly isn’t a way to control perception. Perception
would actually just be causing the world-model to match it.
Doesn’t seem very PCT like to me.
No, your interpretation doesn't seem at all like PCT. Mine does.
If you interpret “updating the world model” as “reorganizing the
functions and connecting weights in the control hierarchy” as well
as reducing current error values, then it sounds more like PCT. I
don’t think that’s a stretch at all. Something that long bothered
me about reorganization was time-scale, ever since you and Bill
did that experiment in which you switched the sign of the
environmental feedback function, and used the word
“reorganization” to describe the rapid switch the subject made to
keep tracking effectively.
Reorganization, or at least its time scale and its modularity,
have never been satisfactorily described. It was a problem Bill
was addressing right to the end. In Friston’s approach, it happens
concurrently with what we would call on-line control. There is no
“now I am reorganizing” phase that alternates with a “now I am
controlling” phase. In both ways of looking at what I believe to
be the same phenomenon, reorganization happens along with control.
Well, so does any version of reorganization that I have
encountered. I think reorganization happens early and quickly at
low levels in both approaches, to produce stabilities (things
unlikely to create surprise) on which further stabilities can rely
– building ever higher levels in either hierarchy.
RM: The
only thing I see that might be considered a “fact” that PCT
ignores is that biological systems " maintain their
order ( non-equilibrium
steady-state ) by restricting themselves to a
limited number of states." I can’t find anything in the
article that gives concrete examples of a
non-equilibrium steady state and how you measure it.
Ah, for that you have to go back to a quite different
intellectual foundation. But here’s an example–the temperature
inside your fridge. It is maintained in a non-equilibrium steady
state (colder than its environment) by exporting entropy in an
energy flow powered by its motor. But if you recall, I did not say
that PCT ignored a “fact”. I said that the free energy approach
was based on an “insight” that is ignored in PCT, and vice-versa.
I don't know how far back the concept of self-organizing
steady-state structures goes, but at least to 1927 in the work of
a meteorologist called Lewis Fry Richardson, who was interested in
how the atmospheric patterns interacted with each other at all
scales from hemispheric to sub-micron. The relevant concept is
captured in a little ditty of his: * Big whirls have little
whirls that feed on their velocity/Little whirls have littler
whirls and so on to viscosity* . But the real work began, I
think, with Ilya Prigogine in about 1947, who recognized that the
Second Law of Thermodynamics (about entropy always increasing)
holds only in a closed system, and that such systems cannot, in
principle, be observed in Nature. Living things most certainly are
not closed systems, but they are structures of interacting
non-equilibrium steady states conceptually related to Richardson’s
whirls.
Control is one way in which these non-equilibrium steady states
continue to exist, at many levels of complexity and dynamism. To
maintain each controlled perception or environmental variable
takes work – energy – to avoid them being disturbed and even
disrupted by their environment. Kent McClelland has a lovely
passage about that in LCS IV. The business of exporting entropy in
a through energy flow goes back a lot further. One site that I
found by googling for “first refrigerator invented” goes back to
the 11th century for a discovery of how to do it, and has been
more or less continuously developed since around 1750. The cold
refrigerator “has” a lot of free energy in a warm environment, in
the sense that one could run a motor or something using the heat
difference between the inside and the outside.
So I
don’t know how the free energy principle accounts for that
“fact”.
I'm not sure which "fact" you refer to. But I think the
unfortunately animistic wording in the sentence you first quoted
– “t ries to
minimize” – might be seriously misleading. It’s like saying
that the refrigerator tries to get its inside warm. If you
unplug the refrigerator, its inside will indeed get warm over
time, but is it “trying to minimize” the difference between its
inside and its outside? Does a control loop “try to minimize”
the difference between its reference and perceptual values? No,
but as a consequence of basic physics, minimizing that
difference is what happens. The principle is the same.
I think a
better way of putting it might be that the self-organizing
properties of feedback in a through
energy flow producing stabilities (low entropy regions)
accounts for Friston’s formulation of the free energy principle
(if he was the inventor). The “limited number of states”
intuitively comes from the fact that if you have an eddy centred
“here”, you can’t have a similarly sized one centred on a point
too close to “here” or the two eddies will merge into one.
Taking the eddies as analogues of perceptual functions (in
either PCT or Friston formulation), it says that you can’t have
two controlled perceptions that are too close to one another,
because if you did, whatever influences one will similarly
influence the other.
If you look
closely enough into it, the same self-organizing property that
is introduced by any asymmetry in a sufficient energy flow can
account in exactly the same way for the “Minimize surprise”
free-energy hierarchy described by Seth and Friston in the paper
I linked and for the equivalent perceptual control hierarchy.
Even though Powers and Friston arrived at this structure from
very different intuitive starting points, constraints imposed by
the underlying physics is probably the reason they both came up
with structures that function identically. Convergent conceptual
evolution!
Despite that,
it does seem to me that thinking of perceptual control and of
the purposes of organisms is likely to be more productive in the
long run. Maybe that’s just my bias, based on my relative
current fluency in the different languages of description.
By the way,
the “insight” that I said PCT missed was that the laws of
non-equilibrium thermodynamics lead to what we see as the
control hierarchy, and the “insight” that Friston seems to miss
is that control of perception leads to what he sees as a
“surprise-reduction” hierarchy.
Indeed, I don't see how the free energy principle connects
to any measurable aspect of the behavior of living things.
No, that's a problem when you don't understand the language. I'm
not sure that my explanation will help you, because I may have
pitched it at too simple a level for the connections to be clear,
just as the use of equations and Greek symbols pitches it in a
language that is unintelligible to you at its root. I hope that
some glimmering might get through, though.
Martin
···
RM:
The thing I liked about Powers theoretical work from the
start was that it was always tied to experimental
demonstrations of the phenomena to be explained. There was
no hand-waving or attempts to impress with mathematics
heavy laden with Greek symbols. It was always clear how
the math related to the phenomena to be explained and the
math itself was always presented clearly, like Bill
actually wanted you to understand it. Friston’s approach
would have impressed me when I was a grad student. But I
can now tell when math is being used to make an impression
rather than provide understanding. And I am not impressed
by Friston’s free energy principle and lovely mathematics.
rsm
–
Richard S. MarkenÂ
"Perfection
is achieved not when you have
nothing more to add, but when you
have
nothing left to take away.�
   Â
            --Antoine de
Saint-Exupery