I have to talk about some of the mechanisms.  I can't help myself. I
would expect the program, if I got it to some level of fundamental
feasibility, to handle numerous kinds of situations as long as the
knowledge it had built up was useable for those situations. The
question is how could it be able to support the kind of reasoning that
I think should be possible? If I was able to teach the program
something about a simple world model it should be able to subsequently
answer something about that model. And I also should be able to use
generalizations and figures of speech that could be applied to that
simple model but which could potentially be applied to situations of
greater complexity as well.  But the problem is (of course) that as it
learns more the number of possibilities should increase sufficiently
to eventually slow it down and befuddle it.

I am hoping to get back to working on a text-based program. However,
if I was able to get it to work I think it would be simple to one day
expand it to include some kind of visual processing as well. The
combination of imagery and text would be interesting.

Although I will program it to initially look for superficial relations
in the text and to recombine them in different ways, I want it to be
able to derive concepts through trial and error. From there it has to
build further knowledge partly based on the way the user (me) reacts
to the program. So it would have a slight tendency to draw conclusions
about the basic relations between words (and other parts of text) by
the way the user responds to its expression of how it combines them.
(The use of fundamental kinds of linguistic behavior to indicate how
words might be related may need to be learned.)

I believe that a simple piece of information, like a simple concept,
has to be associated with hundreds or thousands of other simple pieces
of information. I also believe that the analysis of some input has to
be matched against an imaginative projection (including the projection
of previously learned knowledge) in order to build a better foundation
of what the meaning of the input is and how it should be responded to.
This is a complexity problem so I also believe that extensive indexing
also has to be developed for the acquired knowledge. The indexing
might, for example, be based on generalizations derived from the
knowledge that it had acquired.

Ben's example of a child learning about a pet is a good one. Of course
a text only AI/AGI program is not going to have the experiences a
child can have with a pet. However, the program can be exposed to a
lot of information about pets. I think this extensive knowledge,
combined with trial and error interactions with a user-teacher should
make the program capable of good concept formation even though it will
be different from a child's.

Human beings often seem to deal with opposing and contradictory
theories about the world with little bother. It is only when a
contradictory theory leads directly to some obstacle or the study of a
situation starts to highlight the conflict in theories does it become
a problem. So I think this is a situation that can be described best
with conceptual relativism. Even when we discover a contradiction we
usually first explain it away as a variation that can occur. It takes
some hard headedness to assume that an unexpected variation might
represent a contradiction in theories.

I believe that reason-based reasoning is also important. So a pet-like
object might be visually noticed in a room based on its features and
actions.  If the animal or object is seen frequently and it stands out
against the background, a concept about it will be developed using
concepts about the features and actions of other pets.

Finally, let me add one more thing. Concepts may represent or refer to
objects but they can also play functional roles. So while a conceptual
function prototype might be sufficient to potentially represent any
kind of conceptual relation, I believe it is more to the point to say
that that the program must be capable of deriving conceptual function
prototypes in response to the events it observes in the IO data
environment. Let me draw a parallel. The argument can be made that any
program is a system of yes-no questions and responses. But that
doesn't mean that programmers could effectively use a programming
language that was designed solely on that principle. Similarly, I
believe that an AGI program has to be designed to implement the
eventual formation of conceptual function prototypes and to be
prepared to handle their application and development. Even if I am
unable to figure out how the program could soundly derive functional
prototypes (dynamically) I can use the idea in imaginative
projections. The reason dynamic functional prototypes is so important
is because if concepts become structurally (or abstractly)
specialized, which is part of my theory, then there will probably be a
need to new kinds of conceptual relations to generalize across them. I
think this makes sense and this kind of reasoning comes almost
directly from speculation about the consequences of conceptual
relativism as I see it.



Jim Bromer


On Sun, Jan 4, 2015 at 2:25 PM, Peter Voss <[email protected]> wrote:
> I would find it useful if you could provide one or two specific examples of 
> concepts being derived using existing concepts -- not the mechanics, but 
> situations.
>
> Best,
>
> Peter
>
> -----Original Message-----
> From: Jim Bromer via AGI [mailto:[email protected]]
> Sent: Sunday, January 04, 2015 10:52 AM
> ...
> I was asked if the differences of my theories from the mainstream theories 
> and the theories behind the AI / AGI Frameworks that are being devised are 
> just a matter of semantics. I don't think they are....
>
> A true AGI program will need to derive concepts about its interactions with 
> the IO data environment that it is exposed to.
> It is going to take other concepts to interpret a concept....
>


-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com

Reply via email to