RE: project COSA + Novamente & Cyc

From: Ben Goertzel (ben@goertzel.org)
Date: Sat Aug 10 2002 - 07:53:28 MDT


hi,

> I'm have some unanswered questions about the relationship between
> Novamente and Cyc.
>
> I understand Cyc to be primarily a KB, but with some reasoning abilities
> added on. Is this a correct assessment?

In my opinion, it's *almost* correct, though the Cyc team might disagree ;)

They have definitely put a lot of man-hours into their reasoning system as
well as their knowledge base, and they are recently extending their work
into other areas of AI such as language processing. Stephen Reed, who works
there, recently gave a talk there about how to extend it into the domain of
perception and action based on James Albus's Reference Model work.

So they are slowly moving in the direction of trying to make an AGI centered
around their knowledge DB, but the knowledge DB remains the center of their
work, yes.

> It seems that Cyc currently has a very well populated KB, with
> rudimentary tools to allow that population to increase at a growing
> rate.

It depends what you mean by "well-populated."

Compared to other common-sense DB's it is well-populated. (Though
Teknowledge's SUMO DB, which is smaller, is the biggest one available for
free online... most of Cyc's DB is still proprietary.)

Compared to the total sum of important common sense knowledge, in my view,
Cyc is still pretty bloody small, but the Cyc-ers might disagree. Really
this is a matter of conjecture.

>IIRC, Cyc is reading on its own now, although it still asks
> programmers many questions about what it reads, the time to answer
> questions being the primary governor on it's self-learning speed. Is
> this a correct assessment?

I don't think Cyc is "reading" in any ordinary sense of the word at this
point.

They *might* have some human-guided NLP information extraction tools set up,
but I actually doubt it.

The problem is that existing NLP information extraction tools are only able
to extract very simple information from text (with mediocre reliability),
not complex nested quantifier relationships such as are typical in the CycL
language,

> The combination of Cyc's formidable KB with Novamente's strong
> reasoning, modelling and goal seeking abilities seems a good match. Is
> there any chance of a pairing of the two, or is much duplicated effort
> inevitable?

There will be no duplicated effort in any narrow sense, because

a) Novamente will not make a knowledge base

b) Novamente will not make a predicate-logic-based reasoning engine

c) All tools Cyc will build will be centered around their knowledge base and
predicate-logic-based reasoning engine

Novamente is based on a philosophy of AI in which knowledge is learned
through interaction with the environment and with other minds in the context
of the environment. Very different from Cyc's philosophy, which is based on
knowledge encoding.

Conceivably, a Novamente mind could benefit from Cyc's database, as a kind
of "childhood dictionary infusion into the brain", but not as a primary
means of gains of knowledge.

I have more faith in the Novamente approach, but I also consider it possible
that both of these very different approaches to AGI could ultimately
succeed.

> In an unrelated vein, what clustering/grid/process-distribution software
> will Novamente use? Have you seen OpenMosix?

At the moment our plan is to do what we did with Webmind AI Engine, our last
AI system -- i.e., cook our own.

The requirements for Novamente process distribution are very specialized.
It's not very much code (or headache) to write our own mechanisms for
messaging, load balancing, etc. among the different Novamente instances in a
cluster.

We haven't done this for Novamente yet, but since we did it for Webmind AI
Engine, we know how to do it pretty well.... The current system is written
so as to be "distributed processing ready" tho it's not yet
distributed-processing-enabled.

-- Ben G



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:40 MDT