David> I think that humans intertwine their thoughts with their
David> language.  

There is some anecdotal evidence that the
language facility is not necessary to thought,
for example a Monk who would periodically undergo localized 
epileptic seizures
in which he would report (after the fact) that his comprehension
and production of language vanished, while his cognitive abilities
and general abilities to function
reportedly were otherwise unimpaired. Of course, the removal
of language  could be
just at a conscious level, perhaps his modules still gabbed away
internally unbeknownst to him. I would also argue that other
animals think pretty effectively-- but maybe they use internally
pidgin English as you think we do.

Good automatic language translation has been stumped
David> by the semantic problems etc.  Context is also a huge problem
David> for people working on computer based language problems.  I
David> think AGI will be found by using models (I use this term is the
David> most general way) that communicate in English at a high level
David> where the language and knowledge is interpreted within the
David> model.  Any model could call any other model as needed and many
David> models could be called even at the top level.  Detail
David> information would be known only by the specialist models but
David> abstraction and patterns would be generated at every level to
David> encourage analogies and finding appropriate models to use.  The
David> top most level would also be a model for determining the
David> results from lower level models.

David> In Schanks CD (conceptual dependency) language model, a
David> relatively small set of primitive actions are necessary to
David> represent the semantics of sentences.  Although I don't
David> necessarily advocate only using this set of techniques, I do
David> believe that most language can be translated quite quickly into
David> some useable semantic form.  It could take a lot more effort to
David> get fool proof language out of the AGI if you were intent on
David> tricking the AGI.  Many humans can be easily ticked by other
David> humans as well.  A naive AGI is better than no AGI at all, I
David> believe.

David> The interesting thing about a model versus entering a huge
David> number of rules can be shown by the following analogy.  If I
David> have a function that takes a parameter and produces a result
David> based on a linear algorithm, I only need to store the Y
David> intercept and the slope to produce an infinite set of answers.
David> If I am given some number of numerical pairs, I could create a
David> best fit linear line but I wouldn't know that it was the best
David> because of the small data set or that it was linear instead of
David> geometric etc.  Knowledge is like the line.  Small differences
David> in the input still produce a pretty good answer and this
David> information can be stored very efficiently.  Entering a bunch
David> of rules into an inference engine is like the numerical pairs.
David> The system still has to guess at the function to generate any
David> useful information unless you can just look up the answer
David> directly from previous experience.

My book
What is Thought? studies how this picture extends more generally to
thought. It explains how understanding, semantics, language etc
arises from a generalized version of Occam's razor, in which if you
find a compact enough program behaving well enough on enough data,
it is so constrained as to exploit underlying structure and generalize
to solve new problems. The Occam program is argued to be primarily
located in the genome. The book surveys and extends a computer science
literature exploring these kinds of ideas, as well as data and ideas
from a number of other fields.

David> If you want consistency, then try to enter statements that
David> don't contradict each other.  It would probably be easy for a
David> while but eventually it would be almost impossible to find the
David> knowledge holes that need plugged (CYC).  You would also
David> probably find out just how inconsistent humans really are.  The
David> idea is to teach the AGI knowledge and not just meaningless
David> symbols.  This can be done using models which use data
David> representations and algorithms that are appropriate to the
David> domain the model was created for.  Context can be had by making
David> models that encapsulate language and other models for different
David> contexts.  This means no single dictionary with the meaning of
David> every word.

David> How quickly would humans learn if the teacher could reach right
David> into their heads and place an appropriate analogy, algorithm
David> and exceptions right into their brain structures?  Instead, we
David> use English to encourage a model to be created in the persons
David> head while using repetition to make a deep enough groove for
David> the memory to stick.  Over time this model normally has to be
David> thrown away and replaced to make way for more sophisticated
David> information.  In Math alone, how many times was your internal
David> model thrown out and started over from kindergarten to grade
David> 12?  My estimate is at least 5 times.  I think a combination of
David> programming models directly, programming models that program
David> other models and AGI created models from English language
David> teaching will end up being the quickest way to AGI.  Even given
David> the genetic hardware that people have, it doesn't create an
David> intelligent creature without extensive teaching from other
David> humans.  If spontaneous intelligence doesn't work for humans,
David> why would we think we can create an AGI this way?

According to my theory, "Spontaneous intelligence" is a miasma because
finding Occam code is a hard computational problem. Our intelligence
emerged through an astoundingly huge evolutionary program. When I am
taught a new concept, say by you, I have to build new code. Your code is not
identical to mine, so it would be a hard problem for you to reach
in and simply supply code-- even if you could you'd have to interface
it with my code. Instead, you provide through language, program
sketches and examples of subconcepts so that I can build the code
very rapidly. And in fact, I build it remarkably rapidly-- learning
is incredibly fast if you consider the complexity of the problem.
(If you had to reach in and specify synapse values and connections,
good luck, it would take you a lot longer to teach me mathematics
so that I understood it and could go out and prove new theorems! ;^)

David> I think that fuzzy logic and best guess given "experience and
David> to work with insufficient knowledge" (NARS) is a useful
David> technique but hardly the technique to use for everything.  Many
David> things are known.  The name of the town I live in is not up for
David> debate.  There is only one answer and you know it or you don't.
David> Many patterns also exist or not.  Are close or not.
David> Probabilities have a place but are not the whole answer.

>> However, I don't understand how smaller modules within the brain or
>> mind could communicate like this, in English. The module that deals
>> with the word ``word" for example, in order to deal with a sentence
>> including lots of other words, would have to invoke the other
>> modules themselves. This is discussed at more length in my book
>> What is Thought?, if memory serves in Ch. 13. If you can propose a
>> solution to this, I would be most interested.

David> Sorry, I haven't read your book!

I highly recommend it to you, because you seem to be thinking along similar
lines, and I suspect it will extend your thoughts in a number of directions
you haven't yet encountered.

David> "Word" doesn't have to be contained in only 1 model.  Many
David> jokes are made because the meaning of words are so context
David> sensitive in our brains that we are surprised when other
David> (legitimate in other contexts) meanings are later used instead.
David> Context would be contained in a model that would contain the
David> language and relations appropriate to that domain.  The model
David> could use stored experience for some results and use other
David> techniques if there was significant changes or more detail
David> necessary.  I don't propose that a sentence would be
David> syntactically parsed and then the models for each word called.
David> I think the whole sentence would go to a context model and the
David> semantic meaning of the sentence extracted using local and
David> global tools (more models) as necessary.  Previous sentences
David> and other sources could be included in determining the semantic
David> meaning of the sentence and adding that information to the
David> model to be used further.

David> The "English" communication at some levels could be like
David> <Command word> <optional parameters> and not full sentences.
David> Some models could be called that have access to the context
David> model or other higher levels so that their output could change
David> depending on how they were created.


>> (3) Cassimatis has another interesting proposal. He proposes that
>> all modules (at some high level of granularity) must support a
>> stipulated interlingua.

David> This is exactly what I propose.  I think this interlingua can
David> be a subset of normal English but more likely a group of
David> English subsets depending on the level of interaction.  The
David> highest levels could probably communicate in normal English
David> while at the lowest of some levels it could be a matrix of
David> numbers or <Command> <parameter> like I described above.

I don't know that we disagree on language as you specify your beliefs 
more. You don't actually seem to be proposing that internal modules
speak full English to each other, and I wouldn't assert they don't
have some simplified communication scheme. I'd also recommend looking
up Cassimatis-- he has working models that achieve interesting demos.

David> -- David Clark

-----
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/?list_id=303

Reply via email to