This discussion has been skirting very close to what I said in my
AGIRI talk and book...

evolution invested massive computation in getting the KR right. Yes,the KR
is built for (3+1)-D and a lot more-- it's not just a list of facts,
or some database where you enter logical statements that
are then processed by modus ponens or some such;
it's procedures that can be invoked by code and guide when they should
be and exploit the structure of the world in deep ways.
Object permanence is not learned-- its essentially programmed. Yes,
it requires sensory input to compile, but that's just the way it naturally
evolved, because the genome programs chemical processes that take
place in the brain, in the context of sensory inputs flowing through,
and so naturally evolved to exploit them. 
Yes, general learning is always going to be inherently exponential, which is
why we are not general intelligences. We are great at building on what
we understand, which are the problems we already almost know how to
solve, and happily include most of the problems we care about. 
But weak methods will not produce human-level results in any feasible 
amount of computation.

A likely failure mode is inability to reproduce this underlying
structure, because we don't have the resources evolution did, and its 
computationally hard and complex. Its not a given that we will
fail-- but we have to recognize the problem and focus efforts,
because success won't be easy.

Eric Baum
http://whatisthought.com

-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/[EMAIL PROTECTED]

Reply via email to