From: Ben Goertzel (email@example.com)
Date: Fri Mar 24 2006 - 08:49:51 MST
> My claim is that I have just set you an impossible task, because
> Bayesian reasoning does not play a significant role in most of the
> things that cognitive systems do, and even when it does kick in, what is
> often attributed to it is partially the work of the peripheral processes.
I don't agree with you on this point. I think that probabilistic
inference does play a major role in most of the things that cognitive
I cannot prove that I'm right, but I just want to point out that you
have not proved you're right either!
You have listed a bunch of tasks that we humans carry out without
consciously, reflectively feeling like we're carrying out
probabilistic inference. But, so what? Analogously: We don't feel
like we're carrying out windowed Fourier or wavelet transforms when we
see, yet neuroscience tells us that (from an objective, scientific
standpoint) we actually are.
My hypothesis is that the neurodynamics underlying most cognition is
in effect carrying out a set of crude approximations to probabilistic
inference. For instance Hebbian learning on cell assemblies can
easily be shown to constitute a crude approximation to probabilistic
deduction. This is a speculative theory of neurodynamics but it's not
one that can be dismissed out of hand....
I suggest that, by building AI systems explicitly founded on
probabilistic inference, we are making systems that do in a more exact
and effective way what the human brain sloppily tries to approximate!
But, this does not mean that complex self-organization plays no role
in human or AI cognition, of course. Because the rules of inference
don't tell you much, directly, about what hypotheses to generate or
what inference trajectories to follow. These things can in principle
be carried out based on probabilistic inference, but I conjecture that
in human brains (and tractable near-term AGI's) they happen in other
ways, which do involve the complexity-oriented concepts you are
frequently talking about.
-- Ben G
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT