I think the problem of confirmation is one of the main problems.
Complexity itself may be the real problem but it may just be that
complexity makes the comprehension of what is going on or being said
unreliable because the methods of confirmation are too crude to be
used reliably in complicated situations.   Because confirmation of
some sort has to be used to build basic knowledge, the lack of robust
methods to confirm theories means that general learning is quite
difficult. The problem is that many little detailed pieces of
knowledge are fundamental to even elementary theories about what is
going on.  If human beings could specify what those fundamental pieces
of knowledge consisted of and how they help more substantial concepts
interrelate then hand-crafted AI programs might stand as strong basis
for further AGI design.  But the lack of robust methods of
confirmation means that many fundamental pieces of knowledge are still
elusive.  I could refer to these fundamental pieces of knowledge as
structural knowledge but that might be a little misleading because my
definition of structural knowledge goes beyond that.

The use of imagination to develop theories and to then develop
theories to confirm or disconfirm those theories will obviously lead
to circular (or loopy) theories.  This could serve to explain how
human beings become stuck in ineffective strategies and can even
become delusional but the problem is that our AGI efforts have not
even reached the stage where meaningful delusion can be achieved.
(Perhaps I could make the achievement of loopy delusional theories as
an initial test of my AGI program – as long as the meaningful
delusions are achieved in a system that used true general learning to
reach that plain.)

So, one of the major problems in a true AGI model, even one designed
for limited feasibility tests, is that reliable methods of
confirmation are elusive.  But since true AGI has not really been
achieved with supervised learning methods perhaps this problem might
be better expressed as a problem of discovering the tiny theories that
have to fill the interstices between the more well-illuminated
theories that hold broader meaning.  Perhaps it was a mistake to start
by focusing on the broadest kind of knowledge that holds the strongest
clues to comprehension and meaning.  Maybe the secret formulas have
been hidden in the trivial details all along.  Perhaps there is a
syntax of the glue of thought that explains how broader concepts can
be held together.

My feeling is that structural concepts hold the most promise to
explain how knowledge may be advanced in an AGI program.  If there is
a glue to hold concepts together then that would definitely be called
structural.  Concepts play different roles when interacting with other
concepts, just as some words are verbs some are nouns and so on.  For
me, a structural insight is one that is discovered when you realize
that a combination of concepts (about words or events) can be
understood by realizing how different concepts might be combined to
make sense of the situation.

The idea of conceptual structure has been used before my use of the
idea to refer to the abstract rules that lay behind the application of
some learned knowledge.  For example, a logical deduction might be
called structural learning.  Or the programming behind a decision
process might be called structural.  But since I (and I assume a lot
of other people) envision an AGI program where abstractions could be
derived and learned by experience or by education that means that my
idea of structural learning refers to concepts that can be learned or
otherwise acquired as well.  So then my idea of structural-concepts
comes back to concepts that can hold a bunch of other concepts
together in a conceptual complex by explaining the kinds of roles the
concepts play in the complex.  This understanding is definitely
necessary for understanding.

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