From: Ben Goertzel (ben@goertzel.org)
Date: Sun Jun 23 2002 - 13:28:23 MDT
james wrote:
> For example, I seem to be much better at solving abstract design problems
> and intuitively understanding complex systems than most of the people I
> have worked with. Even if this could be factually proven, neither I nor
> anyone I know would be able to explain why this is the case.
...
> Except, I'm not so certain that we can "transmit [this] to AI minds
> implicitly though interaction..." Maybe we can, but without
> understanding
> the nature of it we can't (or at least shouldn't) make any
> assumptions.
I've thought about this a lot in the domain of math & programming. I think
we can go a long way by collaboratively solving small math and programming
problems with it, so it can pick up patterns in what we're doing in various
cases and generalize from what it sees.... This is actually how most people
work with automated theorem proving tools right now... interactively,
guiding the system when it gets stuck. But, current automated theorem
provers don't significantly learn inductively from the guidance they get ;-(
>We
> do, of course, need to try and implement this in the AI regardless if it
> takes programming or teaching...
You raise a very good point.
A lot of thought has gone into the question: Do we need to program
high-level cognitive schema into Novamente?
I have gone back and forth on this, actually. You may well be right...
though my current brain-state hopes not ;>
The way to do that in Novamente, technically, would be to create compound
schema embodying "ways of organizing ideas and approaching problems." For
this purpose one would want to actually implement the Sasha
Novamente-scripting language we've designed, or something like it. The ways
of organizing and approaching are then Sasha programs loaded into the
mind...
My hope is that the Sasha scripts we end up needing to write will be more in
the spirit of "nudging the system into context-appropriate emergent
responses" rather than say, "encoding explicit inference control schemes."
But I look forward to getting to the stage where we can discover such things
empirically!
In the context of language processing, for instance, I created a whole bunch
of such "schema-scripts" for things like word-sense disambiguation, parsing,
etc. But I'd rather not have to use them.... So we'll see!
The Novamente architecture is really flexible, which is because I know
there's a lot left to discover, and I'd like to be able to incorporate what
we learn along the way without scrapping the system and starting over
(though I won't be surprised if a total code rewrite is needed before we get
to human-level AI, I'm still hoping ... ;)
-- ben g
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT