To me this point seems only partially valid. 1M hand coded rules seems
excessive, but there should be some number (100? 1000?) of hand-coded
rules (not unchangeable!) that it can start from. An absolute minimum
would seem to be "everything in 'Fun with Dick and Jane' through 'My
Little White House'". That's probably not sufficient, but you need to
at least cover those patterns. Most (though not all) of the later
patterns are, or can be, built out of the earlier ones via miscellaneous
forms of composition and elision. This gives context within which other
patterns can be learned.
Note that this is extremely much simpler that starting your learning
from a clean slate.
Vladimir Nesov wrote:
Linas,
Yes, you probably can code all the patterns you need. But it's only
the tip of the iceberg: problem is that for those 1M rules there are
also thousands that are being constantly generated, assessed and
discarded. Knowledge formation happens all the time and adapts those
1M rules to gazillion of real-world situations. You can consider those
additional rules 'inference', but then again, if you can do your
inference that good, you can do without 1M of hand-coded rules,
allowing system to learn them from the ground up. If your inference is
not good enough, it's not clear how many rules you'd need to code in
manually, it may be 10^6 or 10^12, or 10^30, because you'd also need
to code _potential_ rules which are not normally stored in human
brain, but generated on the fly.
I plan to support recent context through a combination of stable
activation patterns (which is analogous to constant reciting of
certain phrase, only on lower level) and temporary induction (roughly,
cooccurrence of concepts in near past leads to them activating each
other in the present, and similarly there are temporary concepts being
formed all the time, of which only those which get repeatedly used in
their short lifetime are retained for longer and longer).
On 11/2/07, Linas Vepstas <[EMAIL PROTECTED]> wrote:
On Thu, Nov 01, 2007 at 02:58:07PM -0700, Matt Mahoney wrote:
--- Linas Vepstas <[EMAIL PROTECTED]> wrote:
Thus, I find that my interests are now turning to representing
conversational state. How does novamente deal with it? What
about Pei Wang's NARS? It seems that NARS is a reasoning system;
great; but what is holding me back right now is not an ability
to reason per-se, but the ability to maintain a conversational
state.
If the matrix is sparse it can be compressed using singular value
decomposition (SVD).
[...]
This is not a model you can tack onto a structured knowledge base.
Why not? Bot NARS and novamente have real-number-valued associative
deduction abilities. I see no reason why simple matrix or nerual net
algo's couldn't be layered on top of it.
Your approach has been tried
hundreds of times.
Yes, I figured as much. I haven't yet seen a cogent explanation of
why folks gave up. For shrdlu, sure .. compute power was limited.
There's discussion about grounding, and folks wander off into the weeds.
There is a great temptation to insert knowledge directly,
but the result is always the same. Natural language is a complicated beast.
You cannot hand code all the language rules. After 23 years of developing the
Cyc database, Doug Lenat guesses it is between 0.1% and 10% finished.
And hasn't stopped trying. We also have Wordnet, and assorted ontology
projects.
How many english words are there? About 250K, but this hasn't stopped
classical dictionary authors, nor wordnet, nor Lenat.
How many sentence parse patterns are there? 10K? 100K? 1M? Its not
infinite, even though it can feel that way sometimes. Just because
you personally don't feel like trying to hand-build an association
matrix between sentence parse patterns and a semantic "current topic
of conversation" dataset doesn't mean its unbuildable.
I didn't claim the approach I described as being "good"; its not;
and I can see its limitations already. I did claim that its "practical",
and after half-a-dozen weekends coding, I have a demo. I'm trying to
understand just how far the approach can be pushed. I get the impression
that it hasn't been pushed very far at all, before people give up.
--linas
-----
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/?&;
-----
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/?member_id=8660244&id_secret=60477015-d0bfa3
