A computational operation is a kind of generalization.  It can be
extended to be used with a great range of different specific numbers by an
algorithm.  So even though my computer does not have 256 bit
multiplication, I could take 64 bit multiplication and extend it
algorithmically to create 256 bit multiplication.  I can take a
generalization of an arithmetic operation and extend the specifications of
that generalization through a (relatively) simple algorithm.

However, it does not look like this works with General AI.  So does that
mean that AGI is going to be really slow?  It looks that way right now.
There is no simple computational algorithm to apply a generalization to a
range of specifications in AGI.

But, suppose that an AGI program was able to learn about numerous related
specifications.  Then it might derive generalizations from the
specifications.  But what good is this?  It will never be mathemagical.
Well, maybe not but the derived generalizations then might be used as a key
index to quickly look up those specifications that are related to the
problem under consideration. And the generalizations might serve as
organizing principles of thought that can be used on problems that are less
well understood.

But how could an AGI program verify the theories that it makes?  One way is
to rely on structural integration.  Structural integrations are various
ways for theoretical principles to be applied in such a way as to create a
structure that explains a great many things.  Or, what might you might call
a generalization.  However, as I just mentioned we know that in general
intelligence there is no mathemagic that can extend an application of a
general operation to a range of specifications like there are in
computational arithmetic.  So this means that our structural integrations
will not be (relatively) simple generalizations but will instead be
problematic because the application of the generalization will need to rely
on all sorts of specifications.  Now if a learned generalization is based
on learned specifications, then that small subsets of possible
generalizations that would represent effective structural integrations
would be feasible (because many of the specifications would have already
been registered with the generalization.)  The effectiveness of a
key structural integration process would not be based on the nearly
impossible task of finding all the specifications that would be needed to
test it out because it would at some point already be populated with the
key specifications that would be needed.
Jim Bromer



-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com

Reply via email to