Hi Richard,
I think you are precisely correct to say that one needs "a sufficiently
and appropriately flexible KR format (which is then really more of a
meta-format)" but I would object that when you go on to say that "a
probabilistic weighted, labeled hypergraph [etc]..." is a good way to
get that flexible KR format, you are underestimating the level at which
the blowback is going to happen.
What I mean by that is that the hypergraph idea is already locking down
many KR assumptions: the nodes are not open to multiple choices for
internal active structure, they interact with other nodes in one
particular choice of interaction space, relationships between nodes are
encoded with relatively simple probabilistic clusters that have direct,
high level semantics (IIRC), and so on. As far as flexible formats are
concerned, this is a thoroughly collapsed wave function. The remaining
flexibility is minimal.
My strong feeling is that the neural net structure in the brain is
ALSO locking down many KR assumptions.... I think you are vastly
overestimating the amount of flexibility present in the brain's
implicit approach to KR...
But, since none of us knows how the brain does KR, we can't really do
much besides opine here...
I would note that if the brain is doing anything like Hebbian learning
btw neural clusters, then it is roughly as constrained as Novamente's
probabilistic representation ... a Hebbian link btw neural clusters is
not semantically all **that** unlike a probabilistic inheritance link
btw two Novamente nodes....
But, I can't rule out the possbility that the brain is doing some
other kind of wild metalearning voodoo that is totally obscure to us
at the moment.. though I really doubt it...
If there were any proof-of-concept systems out there that showed pick up
of even halfway sophisticated concepts purely as a result of learning
mechanisms and real world sensory data, using the class of KRs into
which the hypergraph fits, I would be less pessimistic about it. As it
is, the hypergraphs are as much a leap of faith as anything else.
Please define what you mean by a "halfway sophisticated concepts", and
I can tell you if we're there yet.
And if we're not, maybe we can add your suggestions to our list of
short-term research goals, as this is something we're working on at
the moment (behavior and concept learning based on coupling to a
simulation world, not the physical worls...)
I think it is possible to define classes of cognitive systems that make
relatively few assumptions, and that seem consistent with what we know
of the human system, and then go on, without prejudging them, to
investigate their developmental behavior to *see* what kind of KR
formats they like to develop. Then, once we see what they develop, use
it (a last, and fairly trivial step in the process).
This was what I was saying in my AGIRI workshop presentation.
Yep, but your idea was expressed (both in your presentation and in
this email) at such a vague level of abstraction that I can't really
assess what it means... I look forward to seeing more details at some
point..
Ben
-----
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]
