> I was referring to the time when I read Elaine Rich,
> Hofstadter and Minsky, and all kinds of weird AI journals, and believed
> their approach was not sterile. This was about 15 years ago. Then I got
> a bad case of cellular automata, complexity, resulting in processing of
> lots of dead tree labeled with Fredkin, Toffioli, Holland, Koza, Kauffman,
> Wolfram, and their illustrous ilk. I still haven't recovered yet.
> I'm looking for other infections material, but so far haven't found
> much. Maybe I've become immune, that would be a pity.
The GOFAI approach based on pure logic has been pretty much proven sterile,
although there have been a few successes, such as ("statistical") machine
learning, constraint logic programming.
Although we express our ideas in a serial, 'logical' way, we don't internally
think this way. Our thinking is fuzzy, continuous. But to communicate, it's
necessary to provide a baseline that most can understand and that can be
comunicated easily and reliably.
On the other hand, evolutionary programming approaches currently suffer because
of a variant of the "gravel-and-bare-hands" problem. GA and SA are appopriate
for optimizing and tweaking an existing approximate solution, but not for
solving a hard problem from scratch. For successful hill-climbing evolution, you
need a smooth yet uneven "fitness" surface. For a course of development to be
taken by evolution, almost immediate fitness rewards are needed. Dawkins'
Climbing Mount Improbable book has a most beautiful depiction of how gradually
evolution works in nature, and in what ways it is limited.
Basically, the "invention" of a photo-sensitive cell is very hard (it only
happened a few times in the past billion years, regardless of the trillion-fold
parallelism), the tuning of the optimal number of photosensitive cells is very
simple. Evolving an oscillator is simple, but evolving a TV set is very hard. We
don't know how difficult is to evolve planning, association, deliberative
thought --- but we could provide many hints that would skip an enormous number
of evolutionary steps. Also, better search methods optimize far more efficiently
than evolution: if we want to find roots of a polynomial, we would not use
evolution. Evolution is only appropriate when no search algorithm is known.
As far as CA's go - I agree that they could well be a vital paradigm - I can't
imagine perception and spatial thinking without massive fine-grained
parallelism. And they're more flexible than NN's with predefined rigid
topologies (which can be easily simulated on CA's). We still have to (let them
:) figure out what else could they be useful for.
However, for the concept-level cognition, coarse-grained paralellism is
sufficient -- with individual processors as basic units of computation. Ben's
nodes are quite a persuasive paradigm. A WMish network might be in many ways
preferable to a rigid hierarchical ontology (and ontologies are a holy grail for
Semantic Web adherents, who will soon be forced to show results --maybe they
could show them with WM better than they would with, say CYC).
> But we don't have kLoCs inside us, nor sequential threads of control
> (funny that consciousness should seem to absurdly sequential to
> introspection),
I believe the sequentiallity of introspection is due to human languages being
sequential. Of course, consciousness (along with introspection) and language
coevolve and cannot be considered separately.
Best regards,
Aleks
This archive was generated by hypermail 2b30 : Mon May 28 2001 - 10:00:01 MDT