Re: Some references about ontology and analogy
Rich,
In addition to your note and Danny's notes, I also received
a couple of offline notes. I'll try to give a combined answer.
RC> Do you have any specific wish-list items for an approach
> that would be more likely to result in child-level AI?
First, child-level AI is much harder (or at least no easier)
than adult-level AI. Back in the 1970s, many AI researchers
tried to analyze children's stories in the hope that they might
be easier than adult stories. But their hopes were very quickly
dashed.
The main problem is that children's stories tend to have many
dependencies on visual imagery and lots of action. The text
in those stories is very difficult to analyze without having
complex geometrical modeling, including models of the human
body and its movements (running, jumping, skipping, etc.).
Some research is now being done on using such models to analyze
verbs and the relationships to the corresponding actions. But
there is a lot more work to do.
Project Halo, which was the topic of the previous note, used
chemistry because all the relevant knowledge could be expressed
in chemical formulas and rules that could be expressed in logic
and other representations. They deliberately avoided physics
because many problems in physics require some sort of diagram or
other spatial representation. Even that restriction was not
enough to make it really practical, at least not with today's
technology.
Re moving the coding offshore (i.e., to India, China, etc.):
That might reduce costs somewhat, but even a factor of 10 would
still cost $1000 per page or about half a million dollars for a
typical textbook. Something much more radical is needed.
Re commercial success: The definition of what makes any project
commercially successful has not changed for as long as humans
engaged in commerce: it makes a profit -- i.e., customers are
willing to pay more for the product than it costs to produce.
Cyc has not reached that level, and any similar technology that
costs $10,000 or even $1,000 per page to encode the knowledge
will not make a profit except for small, highly specialized
projects carried out for businesses or governments with very
important needs and very deep pockets to fund them.
Offline> What would be sufficient to be successful?
As I said in the slides (challenge.pdf), current systems of
deduction are already at the level where they can compete with
(or sometimes surpass) the ability of the average human. Just
adding more formal axioms and definitions and more efficient
theorem provers won't do anything for the problem of knowledge
acquisition.
The areas where current systems are woefully inadequate compared
to the average human are on the left side of the cognitive cycle.
Those are the areas related to learning -- which Peirce called
induction and abduction. Current systems of data mining, for
example, are better than humans in finding statistical patterns
in a relational database, but the really hard work is analyzing
the real-world situation in order to determine what to put in
a relational database. No learning systems today can even come
close to doing anything like that.
In those slides I mentioned our current work on the VivoMind
Analogy Engine. I believe that's part of the solution, but
there is a lot more work needed to address the "challenge of
knowledge soup": analyze the "blooming buzzing confusion",
as William James called the sensory overload on an infant
(or an adult for that matter) and determine how to organize it
effectively.
People call natural language texts "unstructured", but those
texts have an enormous amount of structure built into them
in comparison to the total sensory input that impinges on the
human (or any animal) body. The most difficult problem is to
analyze input of that kind (or input from a video camera, for
that matter) and to relate it to words in any natural language.
But as anyone who has been working with NLP knows, there is a
lot of hard work needed even after that step has been done.
Unlike Roger Schank, who believes that logic is totally
irrelevant, I believe that logic is relevant. But I agree
with Roger that the most important work is on the learning
processes that enable an infant to learn language and enable
an adult to analyze and organize any real-world input.
John Sowa