| Date: Fri Nov 02 2001 - 08:28:15 PST | 04 00067 61 01110201 |
| From: | Eric Armstrong |
|
eric.armstrong@eng.sun.com |
| To: | unrev-II@yahoogroups.com |
| Subject: | Self-healing Systems, Cyc |
|
Personal Ontologies Teaching |
Eugene Eric Kim wrote:
The Deep Blue approach strikes me as being similar to Doug Lenat's Cyc approach, which is to give the computer as much information as possible so that it can make "intelligent" decisions. The problems in both systems is that the information must be structured in a highly-constrained manner.
The question is, how do we take terabytes of mostly unstructured information, and structure it so that high-powered computers can do intelligent things with them? Having a small group of people spend decades manually structuring that information, like Lenat's team has been doing, is not a very scaleable solution.
Lenat's exciting assertion (the truth of which remains to be proven) is that there is a certain threshhold at which the system is smart enough to begin filling in gaps in its knowledge on its own -- to recognize what it needs to know, and either ask for it or find it from sources at its disposal.
He stated that Cyc had just now "crossed that threshhold", and held out the promise that knowledge acquisition was now at the "take off" point in the exponential curve.
I think the brute-force approach compensates for the kind of "conceptual chunking" that creates structure by grouping together similar things, and then building on that partial ordering to create taller structures.
It is fascinating to contemplate a cyc-style version of Deep Blue. It should start "overlooking things", the same way a person does, by slighly inaccurate categorization, after which analysis should lead to an improved categorizing strategy that rectifies the oversight.
Basically, the answer to "how do we structure terabytes of information", is with cyc, topic maps, or RDF -- with meta-information that builds a network of interrelationships, and applications which can deal with that meta-information to:
In such a system, the teaching process would be crafted to suit the individual, building on familiar things as much as possible to introduce new ones.
For example, to teach Java to a C programmer, a lot of the constructs are an exact match. But teaching the object oriented aspects of the language requires an appeal to analogies taken from life (like cars), since the C language doesn't have a whole lot of similar constructs.
But suppose the person had built a routine that functioned like an object! By inspection of the person's personal knowledge base, the system could recognize the object-orientedness of that example, and use it to give the person new terms for concepts they had already intuited.
Or the system might discover very little familiarity with cars, and instead build examples based on totally different kinds of examples.
I think we're a long way from being able to do those things, but those are the kinds of things that I suspect will become possible in time.
Sincerely,
Eric Armstrong
eric.armstrong@eng.sun.com