RE: A New Kind of Science

From: Ben Goertzel (ben@goertzel.org)
Date: Wed May 22 2002 - 19:41:42 MDT


One thing I should add. His comments on AI and cognition, which are
speculative and not grounded at all directly in any of the numerous
computational experiments he reports in the book, lean much more in the A2I2
direction than the Novamente or DGI or Cyc directions.

He forcibly states his intuition that

a) cognition is a set of emergent behaviors that come out of very simple
rules

b) CA-ish rules will be able to give rise to these emergent behaviors.

(Note that NN models with local learning rules are easily castable in
"CA-ish" form ... especially NN's embedded in an ambient dimensional space,
e.g. a 3D space)

I agree with his philosophical position but I think that, due to lack of
time spent pondering and analyzing the details of cognition and experiential
learning, he overestimates the simplicity of the rules underlying emergent
cognitive behavior (although I'm sure these rules can be made very very
simple compared to the emergent behaviors, I don't think they can be made as
simple as he does)

-- Ben G

> -----Original Message-----
> From: Ben Goertzel [mailto:ben@goertzel.org]
> Sent: Wednesday, May 22, 2002 5:59 PM
> To: sl4@sysopmind.com
> Subject: A New Kind of Science
>
>
>
> Hi all,
>
> I got the book yesterday. I skimmed it and read a few chapters
> carefully (Chapters 1 and 2, and the chapters on Physics and
> Perception/Analysis). Over the next few weeks I'll read the
> whole thing carefully. Now I just have a few preliminary observations:
>
> 1) There seemed to be nothing in the book fundamentally
> conceptually new to me. However, there is much in the book that
> has been conceptually treated only in various "marginal" places
> rather than in the scientific mainstream.
>
> 2) What he has done, from my point of view, is to go through a
> LOT of ideas from the complex systems literature and place them
> all within a common framework of cellular-automaton-ish systems.
> (Not all his dynamical systems are strictly CA's, but most of
> them are either CA's or very similar). This allows a single book
> to discuss what otherwise would take 10 different books, each
> involving a different notation and mathematical formalism.
>
> 3) His physics approach is cool, but in my view no better (or
> worse) than some other discrete physics approaches, such as Tony
> Smith's (even fringier) "Vodou Physics" model [There is a lot of
> cool stuff on Tony's website, see
> http://www.innerx.net/personal/tsmith/TShome.html ]
>
> 4) His treatment of perception and analysis is conceptually fine,
> but very limited and does not even come close to being
> pragmatically useful for AI
>
> 5) His discussion of biological pattern-formation is a repeat, in
> somewhat more elegant form, of a lot of old ideas from the
> systems theory literature. For instance, A. Lima de Faria's old
> book "Evolution without Selection" makes the same conceptual
> points (along with other more controversial ones) but without the
> numerous CA and quasi-CA examples.
>
>
> Is it progress? Yeah. It's nice to have a common formalism
> (CA's and quasi-CA's) for dealing with all these diverse
> complex-systems phenomena.
>
> Is it a new kind of science? Maybe. But it's not *that* new --
> it's "complex systems science" which has been gathering steam for a while.
>
> MAYBE his CA-ish formalism will prove powerful enough to make
> complex systems science into a "real science", but this book does
> not demonstrate that, because it doesn't really come to any
> dramatic and definitive scientific conclusions in any particular
> branch of science.
>
> He admits that the truly revolutionary nature of his approach
> will only be seen in 10-20 years. Maybe so.
>
> I think that maybe the reason he doesn't agree that most of his
> *concepts* are old ones, is that he thinks he's gotten them
> "right". Whereas to me it just seems like he's said familiar
> things a little *differently*.
>
> The big problem is, given a complex system you want to build, or
> a complex phenomenon you want to analyze, how do you go about
> constructing a CA-ish model of it? (By intuition based on other
> examples, is his method, it seems.) And if the CA-ish model you
> find just *qualitatively* but not exactly models the phenomenon
> or realizes the behavior, how useful is it?
>
> Finding a CA-ish model to *exactly* model a given collection of
> data about a given system or set of systems, is a problem of
> function learning approachable via GP-type methods, but he
> doesn't go there...
>
> I certainly look forward to reading the book more fully, as there
> is much to be learned from the details in the other chapters I
> haven't yet read carefully I'm sure...
>
> -- Ben G
>
>
>
>



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT