AI, Penrose, and BBS (Cariani)

(From Peter Cariani)
(To Bill Powers)
Thanks for the BBS references (Shastri & Ajjanagadde; Penrose (Sept 93)) --
I'll try to obtain the issue. Again, I apologize for my response time constant

···

--
I realize it's longer than most in this forum ....
   I very much agree that much of the connectionist program is really not as
different from that of symbolic AI as they make it out to be (or as they might
like it to be). When you look closely at the mechanisms that are involved, you
see that they are all discrete, and their conception of knowledge not being
explicitly encoded is that instead of one symbol per object (or property) there
is some
more complex set of symbols and relations in place. They still are
"computationalists" in the most restricted sense of the term (i.e. discrete
operations on unambiguous tokens; formal procedures). [The more casual use of
the word "computation" as in "neural computation" is usually so vague and
unrestricted that it has all but rendered the concept meaningless.]
    I had analogous discussions and debates within the artificial life
community over the genotype-phenotype distinction, which is really a digital-
analog distinction. The computationalists wanted to say that a more complicated
(yet still syntactic) relationship between genotype states and phenotypes in
their
simulations was what made these relationships "implicit" and "unpredictable"
rather than "explicit" and "predictable". Underlying these beliefs was a general
feeling that mere computational complexity confers upon a given system
mysterious powers ("unpredictability") and qualitatively different
structure ("life").
        The BIG difference between the connectionists and symbolic AI,
I think, should be their attitudes towards learning from past performance.
The connectionists at least embrace some form of adaptivity and learning,
using measures of past performance to change the structure of their devices.
Old ways of thinking die hard, however, and probably even a larger segment of
the connectionist community associates neural nets with particular, distributed
strategies of computing rather than adaptivity within an external (real,
unpredictable, ill-defined) environment. Thus scores of applied
mathematicians and computer scientists have steered neural net
models from the intellectual milieu of cybernetics and theoretical
neuroscience into one of optimization theory.
        So Munsat's and your alternative (analog) accounts seem much more reason
able,
allowing for all of the perceptions related to an event to have some influence
on the
form of its memory representation (instead of encoding many little sensory atoms
and
their discrete connections).
        I would be utterly surprised, however, if Penrose is indeed on this same
track. He may be a fine mathematician, but he's a REALLY TERRIBLE philosopher
(that BBS would seriously consider his muddled platonic fantasies (which are not
testable)
and not perceptual control theory (which is testable) speaks reams about BBS,
their
priorities, and their intellectual values).
        Godel's Theorems apply only to potentially infinite systems of symbols,
which are physically not realizable, therefore they say NOTHING about the symbol
systems (implementable formal systems, computer programs and simulations,
information-processing operations embedded in biological organisms) that are
actually used in the material world (so Roger, where are the infinite tapes?).
It is even doubtful that the common (semantic) interpretations
of Godel's Theorems are reasonable even for potentially-infinite symbol systems
(see discussion of Wittgenstein's critique of Godel in S.G. Shanker's
"Wittgenstein and the Turning Point in the Philosophy of Mathematics",
SUNY Press, 1987).
        I think when Penrose says "... such as the fact that the actual meaning
of
the symbol [inverted A] is to be 'for all natural numbers'" he REALLY believes
1) that the (actually infinite) set of all natural numbers
exists independently of him and anyone else (i.e. Boss Platonic Reality)
and 2) that there is some logical relationship between the symbol
[inverted A] and this Platonic entity. Far from being grounded in some sort
of neural representation or even in some conception of analog processes,
or even in some conception of himself as an epistemic agent perceiving/acting
in the world, Penrose is really out there in la-la land.
He is trying to apply a set of concepts, which were meant to
deal with infinite sets of discrete symbol strings, to the brain,
which is neither infinite nor made up of discrete symbol strings.
I thought it was exceedingly embarassing that the New
York Times Book Review called "The Emperor's New Mind" "the most important book
in
40 years" (which speaks reams about the science editors at New York Times Book
Review). It really seemed to me to be more evidence of an ongoing general
decline
in our scientific culture (this coupled with the current blathering in the
media of some theoretical physicists about God and the Big Bang). This is what
happens when we get out of the habit of demanding that scientific theories be
tested (in whatever way) by observations of the world (however obtained)--
science and religion quickly get blurred together.
        But maybe we are turning a corner. There is a very good contemporary
critique of infinity by (former mathematician) Brian Rotman (I forget the exact
title, something like "From Zero to Infinity: On and on and on....",
Stanford University Press, 1993, $13 paper) and an equally incisive
critique of platonically-based "final theories" in physics
("The End of Physics", Basic Books, 1993, $25 hard,
I forget the author's name -- he's an editor of Science).

Over & out,

-- Peter Cariani (9/28/93)