From: Ben Goertzel (firstname.lastname@example.org)
Date: Fri Mar 18 2005 - 14:58:55 MST
I don't claim to have proven that Bayesian inference is not a
computationally feasible foundation for AGI. The probability that I'm
wrong is, IMO, significantly greater than zero. But nevertheless this is my
best guess.... I don't think this is an OBVIOUS conclusion; I have come to
this conclusion only after a lot of hard thinking about both probability
theory and AGI. And I realize that I haven't given an adequate
justification for the conclusion here; to do so in a reasonably brief email
would be quite hard....
As a teaser, though, I'll note that I did a fair amount of work with MAXENT
a while back. It's powerful but yet the maximum-entropy assumption about
priors isn't really good enough. Yet the Solomonoff-Levin prior is
intractable.... Is there a "middle ground" (not quite the right expression,
I know) that works for AGI yet is tractable? IMO this "middle ground" is
going to involve a heck of a lot of stuff that has nothing directly to do
with Bayes theorem but a lot to do with cognitive science..
----- Original Message -----
From: "J. Andrew Rogers" <email@example.com>
Sent: Friday, March 18, 2005 1:37 PM
Subject: RE: basic BayesCraft training?
> Ben wrote:
>> I believe what Marc really wants to say here is NOT
>> that Bayes theorem is "broken" (clearly it's correct
>> math), but rather that explicitly applying
>> Bayesian inference is not a computationally feasible
>> strategy in most cases. So it's the idea that
>> "intelligence should be achieved primarily via
>> explicit application of Bayes Theorem" that is broken.
> Marc said what Marc said. If he wanted to say something else, then he
> should be a little more thoughtful before pressing the "Send" button.
> Even if he had written exactly what you had written above, I would still
> disagree with the reasoning. Just because some implementations of Bayes
> may be intractable does not mean that it is *necessarily* intractable in
> this domain. Even you use the "most cases" weasel words, which makes
> the idea that "Bayes theorem is broken" a real stretch even if Marc
> merely meant "infeasible".
> It is not obvious to me that some fairly pervasive application of Bayes
> is always going to be intractable for this application. It would seem
> to me to be making some implicit assumptions in the design,
> implementation, and problem space that are not warranted. I'll simply
> make the observation that people don't implement mathematics on
> computers, they implement finite approximations that often have
> different properties than the pure math description would suggest. As
> an obvious example, useful lossless data compression algorithms do a
> pretty fine job on ordinary computers even though a perfect
> mathematically pure implementation of data compression would be
> generally intractable in this universe.
> j. andrew rogers
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:50 MDT