project three

xyz wrote:

> ========
>
> People have been working on this since before computers were invented.
> However, you are right that there are no "architectures" that
> encourage cooperation and team development.  What is needed is a
> definition of a "AI Virtual Machine" where everyone could experiment,
> participate, and combine knowledge bases and procedures.
>
> It seems that some of the capabilities needed of such a AIVM would be:
> 1) observation.  The AIVM should be capabe of making observations and
> recording facts.  The observations could be pictures, text, speech,
> etc. This could generate enormous databases.

Yes.

> 2) generalization.  The AIVM could spend a lot of time extracting
> features, comparing, and evaluating the enormous databases to find
> common elements.  This would enable the AIVM to reduce the amount of
> data that would need to be stored.

Yes, but I think there is much to talk about in this step. 
I think the storage structure should be capabel of storing 
both instance and generalization.  Any time we generalize, 
there is a bias.  How do you propose to deal with that?


> 3) abstraction.  This is similar to generalization in that it involves
> evaluating features to produce rules that can reduce the amount of
> data stored or allow reaching conclusions that would have statistical
> validity.


Could you expound on the distinction between 2 and 3 ?


> 4) interaction. The AIVM should be capable of interacting with a user
> to answer questions or to gather additional facts.

Definitely.

> 5) prioritization.  The AIVM would have to know what is more important
> when.  If someone asks a question, providing an answer should be a
> first priority and the generalization and abstraction processes would
> get the major emphasis when no observations are being made and no
> questions are being asked (the sleep cycle).

Good idea.


> To achieve this kind of system you need to establish some primitive
> functions that can be used to build procedures. (and which could be
> generated by the abstraction mechanism so that the system can program
> itself).  This self-programming should also be part of the
> interaction.  When a task is given to the AIVM, a procedure should be
> built and then executed.  This self-programing is what would make this
> an AI project instead of an information retrieval project.


Yes ... identifying the primitive functions and defining 
them in terms of the virtual machine is, i think, where we 
need to start.  Perhaps we should start by making some 
initial proposals and seeing if they are at the right level.


> How will the data be stored?  This also needs to be established.  How
> should text be stored ASCII, Unicode?  How about speech or music?
> Photographs?  The definition of the primitive functions to be
> performed determines what might be best, but there should be
> provisions for improvements in technology.


> The system could be implemented in steps.  Start with only text, for
> example.  Pictures and speech can come later. Define primitive
> functions for morphological analysis, boolean logic, statistical
> analysis, etc. Establish procedure-building conventions that can be
> used to achieve some desired functions, e.g., data mining, citation
> analysis, etc.

Yes, I agree.  I think the virtual machine is defined in 
layers.  Perhaps we are ready to start sketching the bottom 
layer.  I'll be proposing some of my thoughts here in a 
separate post.  Perhaps you could do the same.

Could you expound on what you call "procedure-building 
conventions"  ?

> This kind of project requires a lot of resources (grad student
> workhorses) and never losing sight of the goal which is to have a
> system that can grow and interface with other subcomponents.

Yes definitely !  The interoperability of subcomponents is 
paramont.

Thanks for taking the time to make this post.  We need this 
kind of thinking in the project ... lots of it :)


Patty