Re: What is Needed to Teach a Computer to Read?

[Peter Olcott]

On 9/1/2012 4:56 PM, Mok-Kong Shen wrote:
> Apparently the researchers don't yet fully know what is needed.
> BTW, there is a group comp.ai.nat-lang.
>
> M. K. Shen
I will add them to the list of cross-postings.

Here is a recap of the prior points:
If we are to design an ontology that knows how to learn where do we start?

We would have to start with the meta-knowledge of {how to learn
syntax}.** We would aim to encode this meta-knowledge directly within
the ontology. It would be encoded within the ontology such this
knowledge within the ontology can control the process of {learning syntax}.

Some representation of syntax would also be required. This
{representation of syntax} would be entirely encoded as data, such that
changes to the data will change the grammar.

** Why focus on syntax first?
Syntax and logical entailment would seem to form all of the connections
between every language element therefore form the basis of all
Compositionality.

[Peter Olcott]

We would have two different kinds of syntax:
1) Natural Language syntax would be required to process ordinary text.
2) The (much simpler) syntax of the internal semantic representation
must be specified.
A mapping from 1) to 2) would also be required.

[Peter Olcott]

On 9/1/2012 5:06 PM, casey wrote:

> On Sep 2, 7:55 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
>> Apparently the researchers don't yet fully know what is needed.

> A program that learns the way we do.
>
The key aspect of {learning the way we do} is knowing enough about
natural language to be able to populate an ontology from reading
ordinary text. We do not need a fully populated ontology for this, we
only need a BootStrap degree of natural language understanding. Since
all of knowledge is constructed on the basis of connections between
language elements (Linguistic Principle of Compositionality) we should
start with a way for the ontology to learn the structure of these
connections.

It seems to me that the set of these connections can be exhaustively
divided into two categories:
1) Syntax
2) Logical Entailment

http://plato.stanford.edu/entries/compositionality/
If we define Compositionality to include every source of meaning
{lexical, contextual} (excluding speaker intent) and we define syntax
broadly to include every connection between elements of language besides
logical entailment, then (at least) most of the arguments against
Compositionality would cease to exist.

[Burkart Venzke]

Am 02.09.2012 15:26, schrieb Peter Olcott:
> On 9/1/2012 5:06 PM, casey wrote:
>> On Sep 2, 7:55 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
>>> Apparently the researchers don't yet fully know what is needed.
>> A program that learns the way we do.
>>
> The key aspect of {learning the way we do} is knowing enough about
> natural language to be able to populate an ontology from reading
> ordinary text. We do not need a fully populated ontology for this, we
> only need a BootStrap degree of natural language understanding.

Only something for a boot strap for natural language understanding, ah,
if this is all... ;)

I think it is not good possible to learn natural language without
understanding (or better: learning) something about the world, too.

> Since
> all of knowledge is constructed on the basis of connections between
> language elements (Linguistic Principle of Compositionality) we should
> start with a way for the ontology to learn the structure of these
> connections.

What are the atoms of this language structure? Words without any
understanding of them? Good luck... it will be never perfect (relatively
to our human language understanding).

Burkart

[casey]

On Sep 2, 11:26 pm, Peter Olcott <OCR4Screen> wrote:
> On 9/1/2012 5:06 PM, casey wrote:> On Sep 2, 7:55 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
> >> Apparently the researchers don't yet fully know what is needed.
> > A program that learns the way we do.
>
> The key aspect of {learning the way we do} is
> knowing enough about natural language to be
> able to populate an ontology from reading
> ordinary text.

We do not simply learn by reading ordinary text.

Our concepts first must be extracted from examples
before they can be codified and used to understand
new things via language, written or spoken, and
that means understanding how those concepts are
formed, represented and used in the brain.

A computer program may learn via reading:

"Jim has brown hair and blue eyes."
"Jill had red hair and brown eyes."

And we may ask a question.

"Does Jill have blue eyes?"

A program can decode the sentence and reply.

"No, Jill does not have blue eyes."

Programs can search and do logic on information in
a data base *but it is not knowing as we know*.

We may give it category knowledge.
Jill is a person.
Blue is a color.
Hair is part of a human.

But we learn those things via our senses using
real examples, not sentences, and it is not
stored in lists but as a rich interconnection
with everything else we know in a way we do
not yet understand.


jc

[Peter Olcott]

On 9/2/2012 4:13 PM, Burkart Venzke wrote:
> Am 02.09.2012 15:26, schrieb Peter Olcott:
>> On 9/1/2012 5:06 PM, casey wrote:
>>> On Sep 2, 7:55 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
>>>> Apparently the researchers don't yet fully know what is needed.
>>> A program that learns the way we do.
>>>
>> The key aspect of {learning the way we do} is knowing enough about
>> natural language to be able to populate an ontology from reading
>> ordinary text. We do not need a fully populated ontology for this, we
>> only need a BootStrap degree of natural language understanding.
>
> Only something for a boot strap for natural language understanding,
> ah, if this is all... ;)

This is really all there is to it, yet to make BootStrap Natural
Language is a difficult problem.

>
> I think it is not good possible to learn natural language without
> understanding (or better: learning) something about the world, too.

Since we can learn about the world by reading about the world, a machine
can too, iff (if and only if) is has the minimum prerequisite
meta-knowledge. The exact nature of this missing meta-knowledge can be
reverse-engineered.

>
>> Since
>> all of knowledge is constructed on the basis of connections between
>> language elements (Linguistic Principle of Compositionality) we should
>> start with a way for the ontology to learn the structure of these
>> connections.
>
> What are the atoms of this language structure? Words without any
> understanding of them? Good luck... it will be never perfect
> (relatively to our human language understanding).
>

There will eventually be a meaning postulate for every piece of human
knowledge. It would seem that the only {atoms} of this language
(everything is essentially defined in terms of everything else) would be
the hierarchy of connections between {units of meaning}. Everything is
either a relation or a property of something else. Even the concept of
existence itself is broken down into its properties and relations.

> Burkart
>

[Peter Olcott]

On 9/2/2012 4:46 PM, casey wrote:
> On Sep 2, 11:26 pm, Peter Olcott <OCR4Screen> wrote:
>> On 9/1/2012 5:06 PM, casey wrote:> On Sep 2, 7:55 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
>>>> Apparently the researchers don't yet fully know what is needed.
>>> A program that learns the way we do.
>> The key aspect of {learning the way we do} is
>> knowing enough about natural language to be
>> able to populate an ontology from reading
>> ordinary text.
> We do not simply learn by reading ordinary text.
>
> Our concepts first must be extracted from examples
> before they can be codified and used to understand
> new things via language, written or spoken, and
> that means understanding how those concepts are
> formed, represented and used in the brain.

Which is essentially the process of encoding meaning postulates within
an ontology.

> A computer program may learn via reading:
>
> "Jim has brown hair and blue eyes."
> "Jill had red hair and brown eyes."
>
> And we may ask a question.
>
> "Does Jill have blue eyes?"
>
> A program can decode the sentence and reply.
>
> "No, Jill does not have blue eyes."
>
> Programs can search and do logic on information in
> a data base *but it is not knowing as we know*.
>
> We may give it category knowledge.
> Jill is a person.
> Blue is a color.
> Hair is part of a human.
>
> But we learn those things via our senses using
> real examples, not sentences, and it is not
> stored in lists but as a rich interconnection
> with everything else we know in a way we do
> not yet understand.

Right that is the key missing piece, we must understand exactly how we
encode the {rich interconnection with everything else}. That is the part
that I have begun to work out here in this forum. The answer to this can
be reverse-engineered.


>
> jc

[Mok-Kong Shen]

Am 03.09.2012 04:38, schrieb Peter Olcott:

> Since we can learn about the world by reading about the world, a machine
> can too, iff (if and only if) is has the minimum prerequisite
> meta-knowledge. The exact nature of this missing meta-knowledge can be
> reverse-engineered.

As layman I surmise that a machine couldn't have the concept of
meta-knowldege to begin with.

M. K. Shen

[Peter Olcott]

Of course not, a machine lacking an ontology (when measured against its
comprehension of the meaning of words from natural language) begins with
an effective IQ of zero . Some BootStrap meta-knowledge would have to
be created manually. If we focus on finding this minimal set of
meta-knowledge then it will be much easier to find than if we are not
focusing on finding this set.

The *only* thing that is required within the BootStrap ontology is the
meta-knowledge of how to learn more knowledge (including learning more
meta-knowledge).

[António Marques]

Is it workable, though? For instance, an empty Word document doesn't start
off with 0 bytes plus Word's knowledge of hoW to expand that. It starts with
some *data* structures already in place, a template if you will. This is
actually very common. IOW, the software doesn't know how to generate its
data from scratch, though it knows how to use it from a certain core, so the
'how to' is not the only part needed, there's also a little bit of bootstrap
*data* needed. In a way like that cake or whatever that they have in England
where they use a little bit of the previous year's cake.

[Peter Olcott]

This seems to be saying the same thing that I said using different words.
I envision the system that comprehends the meaning of words to be data
driven on the basis of encoded {Meaning Postulates} .

[Mok-Kong Shen]

Am 03.09.2012 15:02, schrieb Peter Olcott:

> Of course not, a machine lacking an ontology (when measured against its
> comprehension of the meaning of words from natural language) begins with
> an effective IQ of zero . Some BootStrap meta-knowledge would have to
> be created manually. If we focus on finding this minimal set of
> meta-knowledge then it will be much easier to find than if we are not
> focusing on finding this set.

Sorry for a layman's view: Wouldn't the machine that should "understand"
that minimal meta-knowlege need a second level minimal meta-knowledge,
and so forth and so forth?

M. K. Shen

[Burkart Venzke]

Am 03.09.2012 04:38, schrieb Peter Olcott:

> On 9/2/2012 4:13 PM, Burkart Venzke wrote:
>> Am 02.09.2012 15:26, schrieb Peter Olcott:
>>> On 9/1/2012 5:06 PM, casey wrote:
>>>> On Sep 2, 7:55 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
>>>>> Apparently the researchers don't yet fully know what is needed.
>>>> A program that learns the way we do.
>>>>
>>> The key aspect of {learning the way we do} is knowing enough about
>>> natural language to be able to populate an ontology from reading
>>> ordinary text. We do not need a fully populated ontology for this, we
>>> only need a BootStrap degree of natural language understanding.
>>
>> Only something for a boot strap for natural language understanding,
>> ah, if this is all... ;)
> This is really all there is to it, yet to make BootStrap Natural
> Language is a difficult problem.
>>
>> I think it is not good possible to learn natural language without
>> understanding (or better: learning) something about the world, too.
> Since we can learn about the world by reading about the world,

Who is "we"? Little children? Aborigines? Most of the people 500 years ago?
Individually, children first have to learn to read and to understand,
this is not innate. And we have been living of thousands of years
without the ability of reading, it cannot be trivial.

> a machine
> can too, iff (if and only if) is has the minimum prerequisite
> meta-knowledge.

Theoretically, you are right...

> The exact nature of this missing meta-knowledge can be
> reverse-engineered.

...but how do you want to manage this reverse-engineering?

>>> Since
>>> all of knowledge is constructed on the basis of connections between
>>> language elements (Linguistic Principle of Compositionality) we should
>>> start with a way for the ontology to learn the structure of these
>>> connections.
>>
>> What are the atoms of this language structure? Words without any
>> understanding of them? Good luck... it will be never perfect
>> (relatively to our human language understanding).
>>
> There will eventually be a meaning postulate for every piece of human
> knowledge. It would seem that the only {atoms} of this language
> (everything is essentially defined in terms of everything else) would be
> the hierarchy of connections between {units of meaning}. Everything is
> either a relation or a property of something else. Even the concept of
> existence itself is broken down into its properties and relations.

What are your first, foundamental knowledge items? Only "empty"
words/notions?
How can an AI (machine) know or comprehend something?

Burkart

[Burkart Venzke]

Nice idea but it is not mandatory if the machine can understand (and
learn later) something about the world without (further) meta-knowledge.
Basics might be a kernel of comprehension, some senses and communication
abilities, some inate aims to reach (or something like ethical values)...

Burkart

[Burkart Venzke]

What kind of "Meaning Postulates"? How shall they work? Are they fix or
can they be expanded by learning?

Burkart

[Peter Olcott]

On 9/3/2012 4:51 PM, Burkart Venzke wrote:
> Am 03.09.2012 04:38, schrieb Peter Olcott:
>> On 9/2/2012 4:13 PM, Burkart Venzke wrote:
>>> Am 02.09.2012 15:26, schrieb Peter Olcott:
>>>> On 9/1/2012 5:06 PM, casey wrote:
>>>>> On Sep 2, 7:55 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
>>>>>> Apparently the researchers don't yet fully know what is needed.
>>>>> A program that learns the way we do.
>>>>>
>>>> The key aspect of {learning the way we do} is knowing enough about
>>>> natural language to be able to populate an ontology from reading
>>>> ordinary text. We do not need a fully populated ontology for this, we
>>>> only need a BootStrap degree of natural language understanding.
>>>
>>> Only something for a boot strap for natural language understanding,
>>> ah, if this is all... ;)
>> This is really all there is to it, yet to make BootStrap Natural
>> Language is a difficult problem.
>>>
>>> I think it is not good possible to learn natural language without
>>> understanding (or better: learning) something about the world, too.
>> Since we can learn about the world by reading about the world,
>
> Who is "we"? Little children? Aborigines? Most of the people 500 years
> ago?
> Individually, children first have to learn to read and to understand,
> this is not innate. And we have been living of thousands of years
> without the ability of reading, it cannot be trivial.

It is not trivial, yet all humans have this capability.

>
>> a machine
>> can too, iff (if and only if) is has the minimum prerequisite
>> meta-knowledge.
>
> Theoretically, you are right...
>
>> The exact nature of this missing meta-knowledge can be
>> reverse-engineered.
>
> ...but how do you want to manage this reverse-engineering?

Proceeding from the broad goal (machine comprehension of natural
language), and work backwards to what this requires.

>
>>>> Since
>>>> all of knowledge is constructed on the basis of connections between
>>>> language elements (Linguistic Principle of Compositionality) we should
>>>> start with a way for the ontology to learn the structure of these
>>>> connections.
>>>
>>> What are the atoms of this language structure? Words without any
>>> understanding of them? Good luck... it will be never perfect
>>> (relatively to our human language understanding).
>>>
>> There will eventually be a meaning postulate for every piece of human
>> knowledge. It would seem that the only {atoms} of this language
>> (everything is essentially defined in terms of everything else) would be
>> the hierarchy of connections between {units of meaning}. Everything is
>> either a relation or a property of something else. Even the concept of
>> existence itself is broken down into its properties and relations.
>
> What are your first, foundamental knowledge items? Only "empty"
> words/notions?
> How can an AI (machine) know or comprehend something?

The foundational knowledge items are the (differing types of)

connections between {units of meaning}.

A Machine can comprehend anything on the basis of the {Meaning
Postulates} within its ontology.

>
> Burkart

[Peter Olcott]

I envision something like Predicate Logic as the basis for the
specification of {Meaning Postulates}.

I would venture to guess that the {Meaning Postulates} themselves may
not change whereas learning would mostly be the case of adding more
{Meaning Postulates}.

[DKleinecke]

On Sep 3, 4:41 pm, Peter Olcott <OCR4Screen> wrote:
> On 9/3/2012 4:51 PM, Burkart Venzke wrote:

> > Individually, children first have to learn to read and to understand,
> > this is not innate. And we have been living of thousands of years
> > without the ability of reading, it cannot be trivial.
>
> It is not trivial, yet all humans have this capability.

Sadly that is not true.

There are people who do not read and there are people who do not talk.
We cope with their existence by called them handicapped and thereafter
ignoring their existence. I think this population might be very
insightful to work with - and maybe someone has - but I know of no
linguistic study of mute people (the ones I have met seem to
understand but not respond) or people with poor speech skills.

[Peter Olcott]

On 9/3/2012 8:44 PM, DKleinecke wrote:
> On Sep 3, 4:41 pm, Peter Olcott <OCR4Screen> wrote:
>> On 9/3/2012 4:51 PM, Burkart Venzke wrote:
>>> Individually, children first have to learn to read and to understand,
>>> this is not innate. And we have been living of thousands of years
>>> without the ability of reading, it cannot be trivial.
>> It is not trivial, yet all humans have this capability.
> Sadly that is not true.

All humans normal humans have the innate ability to learn to read.

[Peter T. Daniels]

On Sep 3, 9:49 pm, Peter Olcott <OCR4Screen> wrote:
> On 9/3/2012 8:44 PM, DKleinecke wrote:> On Sep 3, 4:41 pm, Peter Olcott <OCR4Screen> wrote:
> >> On 9/3/2012 4:51 PM, Burkart Venzke wrote:
> >>> Individually, children first have to learn to read and to understand,
> >>> this is not innate. And we have been living of thousands of years
> >>> without the ability of reading, it cannot be trivial.
> >> It is not trivial, yet all humans have this capability.
> > Sadly that is not true.
>
> All humans normal humans have the innate ability to learn to read.

Some people can only learn to read a very little bit. There are many
kinds of reading disorders ("dyslexia" per se is not a single
syndrome, just as there's no one thing that is "cancer"), and some are
apparently insuperable.

Reading is unlike speaking, however, in that there does not seem to be
any "critical period" for learning to read as there is for acquiring
language -- as far as the psychologists can tell me, people who learn
to read as adults do not show deficiences by comparison with those who
learned to read as children.

[Mok-Kong Shen]

I notice however your "if". A "kernel of comprehension" has a meaning
for a subject that "can" somehow "comprehend" in the first place,
isn't?

M. K. Shen

[Burkart Venzke]

The idea of using Predicate Logic is already some decades old...

> for the specification of {Meaning Postulates}.

...the main problem is the real specification of your "Meaning Postulates".
Which are they? What are they able to describe or to fulfill?

[Burkart Venzke]

All? But this point is not important for me...

>>> a machine
>>> can too, iff (if and only if) is has the minimum prerequisite
>>> meta-knowledge.
>>
>> Theoretically, you are right...
>>
>>> The exact nature of this missing meta-knowledge can be
>>> reverse-engineered.
>>
>> ...but how do you want to manage this reverse-engineering?
> Proceeding from the broad goal (machine comprehension of natural
> language), and work backwards to what this requires.

This is your goal, right. But how do you really want to manage it?
Try to explain it concretely and you will realize the problems.

>>>>> Since
>>>>> all of knowledge is constructed on the basis of connections between
>>>>> language elements (Linguistic Principle of Compositionality) we should
>>>>> start with a way for the ontology to learn the structure of these
>>>>> connections.
>>>>
>>>> What are the atoms of this language structure? Words without any
>>>> understanding of them? Good luck... it will be never perfect
>>>> (relatively to our human language understanding).
>>>>
>>> There will eventually be a meaning postulate for every piece of human
>>> knowledge. It would seem that the only {atoms} of this language
>>> (everything is essentially defined in terms of everything else) would be
>>> the hierarchy of connections between {units of meaning}. Everything is
>>> either a relation or a property of something else. Even the concept of
>>> existence itself is broken down into its properties and relations.
>>
>> What are your first, foundamental knowledge items? Only "empty"
>> words/notions?
>> How can an AI (machine) know or comprehend something?
> The foundational knowledge items are the (differing types of)
> connections between {units of meaning}.

Where does the different types come from? Who defines them or do you
think they are "natural", however?

> A Machine can comprehend anything on the basis of the {Meaning
> Postulates} within its ontology.

"Comprehend" or "deduce"? If you only/primarily think of comprehension:
In which sense is it really more than deduction?

[Burkart Venzke]

I am not sure what you mean by 'that "can" somehow "comprehend" in the
first place'. Do you want to express that "can" (ability) is necessary
to "comprehend"? Or the other way around?

When we think of little children we know that the "can" (see, listen,
touch, feel, move...) and that they can relate input of different senses
which lead to comprehension.

Burkart

[Peter Olcott]

Right and arithmetic is much older than that, yet will not be replaced
by anything better.
There are many details to this {something like} predicate logic that
have not yet been elaborated.

>
>> for the specification of {Meaning Postulates}.
>
> ...the main problem is the real specification of your "Meaning
> Postulates".
> Which are they? What are they able to describe or to fulfill?

They are defined by Montague, in his Montague Grammar (more aptly
Montague Semantics).

The best that I can tell some extension of Montague's work can fully
specify every subtle detail of anything that can every be expressed in
words. Also it expresses these concepts using enormously simpler**
syntax, and the syntax can be organized to facilitate reasoning.

** Simpler than natural language.

[Peter Olcott]

Even Chimps can be taught to read and write.

>
>>>> a machine
>>>> can too, iff (if and only if) is has the minimum prerequisite
>>>> meta-knowledge.
>>>
>>> Theoretically, you are right...
>>>
>>>> The exact nature of this missing meta-knowledge can be
>>>> reverse-engineered.
>>>
>>> ...but how do you want to manage this reverse-engineering?
>> Proceeding from the broad goal (machine comprehension of natural
>> language), and work backwards to what this requires.
>
> This is your goal, right. But how do you really want to manage it?
> Try to explain it concretely and you will realize the problems.

Any purely analytical problem no matter how complex can always be broken
down into a hierarchy of simple steps that all fit together seamlessly
to solve the enormous problem.

The most important thing is to see how humans organize the set of all
knowledge. We don't have to know anything about the physiology, we only
need to know how everything {units of meaning} is connected together.
The second most important thing is to find the most efficient way to
exhaustively and correctly document natural language syntax.

>
>>>>>> Since
>>>>>> all of knowledge is constructed on the basis of connections between
>>>>>> language elements (Linguistic Principle of Compositionality) we
>>>>>> should
>>>>>> start with a way for the ontology to learn the structure of these
>>>>>> connections.
>>>>>
>>>>> What are the atoms of this language structure? Words without any
>>>>> understanding of them? Good luck... it will be never perfect
>>>>> (relatively to our human language understanding).
>>>>>
>>>> There will eventually be a meaning postulate for every piece of human
>>>> knowledge. It would seem that the only {atoms} of this language
>>>> (everything is essentially defined in terms of everything else)
>>>> would be
>>>> the hierarchy of connections between {units of meaning}. Everything is
>>>> either a relation or a property of something else. Even the concept of
>>>> existence itself is broken down into its properties and relations.
>>>
>>> What are your first, foundamental knowledge items? Only "empty"
>>> words/notions?
>>> How can an AI (machine) know or comprehend something?
>> The foundational knowledge items are the (differing types of)
>> connections between {units of meaning}.
>
> Where does the different types come from? Who defines them or do you
> think they are "natural", however?

Here is one type of connection between {units of meaning}
The concept of {plural} combined with the concept of an instance of a
{physical thing} derives a set of more than one {physical things}.

>
>> A Machine can comprehend anything on the basis of the {Meaning
>> Postulates} within its ontology.
>
> "Comprehend" or "deduce"? If you only/primarily think of
> comprehension: In which sense is it really more than deduction?

One does not deduce that liquid water is wet, one refers to the internal
meaning postulate for liquid water, and finds one of its properties is
{wet}.

[Burkart Venzke]

Could you post them here?

> The best that I can tell some extension of Montague's work can fully
> specify every subtle detail of anything that can every be expressed in
> words. Also it expresses these concepts using enormously simpler**
> syntax, and the syntax can be organized to facilitate reasoning.
>
> ** Simpler than natural language.

Where is the connection between a single word (isn't it an atom of the
Grammar?) and an object in the world? Can a machine/AI learn such a
world from being told by referencing to the object?

[Burkart Venzke]

Am 05.09.2012 00:39, schrieb Peter Olcott:
> On 9/4/2012 3:13 PM, Burkart Venzke wrote:
>> Am 04.09.2012 01:41, schrieb Peter Olcott:
>>> On 9/3/2012 4:51 PM, Burkart Venzke wrote:
>>>> Am 03.09.2012 04:38, schrieb Peter Olcott:
>>>>> a machine
>>>>> can too, iff (if and only if) is has the minimum prerequisite
>>>>> meta-knowledge.
>>>>
>>>> Theoretically, you are right...
>>>>
>>>>> The exact nature of this missing meta-knowledge can be
>>>>> reverse-engineered.
>>>>
>>>> ...but how do you want to manage this reverse-engineering?
>>> Proceeding from the broad goal (machine comprehension of natural
>>> language), and work backwards to what this requires.
>>
>> This is your goal, right. But how do you really want to manage it?
>> Try to explain it concretely and you will realize the problems.
> Any purely analytical problem no matter how complex can always be broken
> down into a hierarchy of simple steps that all fit together seamlessly
> to solve the enormous problem.

Yes, but how? There are a lot of possible break downs.
You need a concrete algorithm, not only the idea that is may be possible.

> The most important thing is to see how humans organize the set of all
> knowledge. We don't have to know anything about the physiology, we only
> need to know how everything {units of meaning} is connected together.

Do you think that there is a unique way how everything is connected
together?
Do you think that knowledge itself is always clear (and not unsure etc.)?

Is the concept of plural defined or learned? If learned: How?

>>> A Machine can comprehend anything on the basis of the {Meaning
>>> Postulates} within its ontology.
>>
>> "Comprehend" or "deduce"? If you only/primarily think of
>> comprehension: In which sense is it really more than deduction?
> One does not deduce that liquid water is wet, one refers to the internal
> meaning postulate for liquid water, and finds one of its properties is
> {wet}.

Where does the internal meaning postulate for liquid water initially
comes from? Definition?

[Peter Olcott]

The concept is best build up within the mind within the context of his
whole framework:
http://www.amazon.com/Formal-Semantics-Introduction-Cambridge-Linguistics/dp/0521376106/ref=sr_1_sc_2?s=books&ie=UTF8&qid=1346799399&sr=1-2-spell&keywords=formal+sematnics
The above was the first book that I read about Montague Semantics.

>
>> The best that I can tell some extension of Montague's work can fully
>> specify every subtle detail of anything that can every be expressed in
>> words. Also it expresses these concepts using enormously simpler**
>> syntax, and the syntax can be organized to facilitate reasoning.
>>
>> ** Simpler than natural language.
>
> Where is the connection between a single word (isn't it an atom of the
> Grammar?) and an object in the world? Can a machine/AI learn such a
> world from being told by referencing to the object?

The atoms of the system that I am envisioning are {units of meaning} and
the connections between them. {units of meaning} do not perfectly map to
word boundaries.

All of those things that can only be known by the first-hand direct
experience of the actual physical sensation, the machine that I am
envisioning, will never know. Everything that can be expressed within
words can be fully understood by this machine.

[Peter Olcott]

On 9/4/2012 5:56 PM, Burkart Venzke wrote:
> Am 05.09.2012 00:39, schrieb Peter Olcott:
>> On 9/4/2012 3:13 PM, Burkart Venzke wrote:
>>> Am 04.09.2012 01:41, schrieb Peter Olcott:
>>>> On 9/3/2012 4:51 PM, Burkart Venzke wrote:
>>>>> Am 03.09.2012 04:38, schrieb Peter Olcott:
>>>>>> a machine
>>>>>> can too, iff (if and only if) is has the minimum prerequisite
>>>>>> meta-knowledge.
>>>>>
>>>>> Theoretically, you are right...
>>>>>
>>>>>> The exact nature of this missing meta-knowledge can be
>>>>>> reverse-engineered.
>>>>>
>>>>> ...but how do you want to manage this reverse-engineering?
>>>> Proceeding from the broad goal (machine comprehension of natural
>>>> language), and work backwards to what this requires.
>>>
>>> This is your goal, right. But how do you really want to manage it?
>>> Try to explain it concretely and you will realize the problems.
>> Any purely analytical problem no matter how complex can always be broken
>> down into a hierarchy of simple steps that all fit together seamlessly
>> to solve the enormous problem.
>
> Yes, but how? There are a lot of possible break downs.
> You need a concrete algorithm, not only the idea that is may be possible.

It is entirely incorrect to even begin to look for an algorithm at this
early stage of design.
I have already begun to divide the problem space into its simpler
components.
1) Determine the structure of knowledge.
2) Determine the details of natural language syntax.

It is learned based on having its definition applied to concrete instances.

>
>>>> A Machine can comprehend anything on the basis of the {Meaning
>>>> Postulates} within its ontology.
>>>
>>> "Comprehend" or "deduce"? If you only/primarily think of
>>> comprehension: In which sense is it really more than deduction?
>> One does not deduce that liquid water is wet, one refers to the internal
>> meaning postulate for liquid water, and finds one of its properties is
>> {wet}.
>
> Where does the internal meaning postulate for liquid water initially
> comes from? Definition?

Sets of meanings are assigned to arbitrary word labels.

[DKleinecke]

On Sep 4, 4:25 pm, Peter Olcott <OCR4Screen> wrote:

> 2) Determine the details of natural language syntax.

Isn't that exactly what linguists have been trying to do ever since
syntax was identified as a field of research? [It turns out I have no
idea when syntax was recognized as different than morphology.] Do you
have an attitude toward any of Chomsky's proposals ? Optimality Theory?

[Peter Olcott]

I have not read read very much. Most of my ideas are ideas that I came
up with beginning thirty years ago. I finally have a means of expressing
these ideas using (Montague Semantics) . Montague Semantics and its
possible extensions have been the focus of my recent investigation.

It seems that the key to solving both of the above two goals

1) Determine the structure of knowledge.

2) Determine the details of natural language syntax.

Can be simultaneously fulfilled by the complete elaboration of
Compositionality.

Fully elaborating natural language syntax should not be that difficult.
There has already been great process towards this goal using statistical
methods.
I have ideas here too.

[Peter T. Daniels]

On Sep 4, 6:39 pm, Peter Olcott <OCR4Screen> wrote:

> Even Chimps can be taught to read and write.

?

[Mok-Kong Shen]

I mean that "can" of "can see" is somehow given in nature, we don't
know though how, sadly. A manine registers a picture in pixels and
does some computations, of course, but that's all. It does not "feel"
the content of the picture. That's that problem. It seems we can't
create that capability in a machine.

M. K. Shen

[Peter Olcott]

http://en.wikipedia.org/wiki/Yerkish

[Peter T. Daniels]

What does that have to do with anything like reading and writing? None
of the 1970s nonhuman primate experiments yielded anything like
language in chimpanzees, gorillas, etc.

[Burkart Venzke]

I see, you think about image processing. The usual way (at least when I
studied 20 years ago) of analyizing the pixels to find edges, zones etc.
bottom up is not the natural (human) way in my mind.

> It does not "feel"
> the content of the picture. That's that problem. It seems we can't
> create that capability in a machine.

One difference is that the machine usually cannot learn to see, humans
can and do. To "feel" content needs to have much experience -> learning.

Burkart

[Burkart Venzke]

Aren't there some central points you can write here? I don't want to
read more than 300 pages first.

> The above was the first book that I read about Montague Semantics.
>
>>
>>> The best that I can tell some extension of Montague's work can fully
>>> specify every subtle detail of anything that can every be expressed in
>>> words. Also it expresses these concepts using enormously simpler**
>>> syntax, and the syntax can be organized to facilitate reasoning.
>>>
>>> ** Simpler than natural language.
>>
>> Where is the connection between a single word (isn't it an atom of the
>> Grammar?) and an object in the world? Can a machine/AI learn such a
>> world from being told by referencing to the object?
> The atoms of the system that I am envisioning are {units of meaning} and
> the connections between them. {units of meaning} do not perfectly map to
> word boundaries.

Do you have (an) example(s) for "units of meaning"?

> All of those things that can only be known by the first-hand direct
> experience of the actual physical sensation, the machine that I am
> envisioning, will never know.

Hm? You don't want the machine to have experience in the real world?

> Everything that can be expressed within
> words can be fully understood by this machine.

If a machine really have no contact to the real world it will never
understand anything like a (theoretical) human without any senses.

[Peter Olcott]

Whatever.

[Burkart Venzke]

OK, you are not so far for more concrete steps.
I am thinking of strong AI since my studies 25 years ago. Without real
algorithms it seems to me that strong AI cannot be very difficult. But
when implementing a real learning algorithm I can realize the real
problems (currently that (abstract) "safe facts" do not really exists
when learning from the world (inductively), only ideas/theories which
may be correct but also may be wrong (partially)).

So you want to define all the concepts first like "plural"?

>>>>> A Machine can comprehend anything on the basis of the {Meaning
>>>>> Postulates} within its ontology.
>>>>
>>>> "Comprehend" or "deduce"? If you only/primarily think of
>>>> comprehension: In which sense is it really more than deduction?
>>> One does not deduce that liquid water is wet, one refers to the internal
>>> meaning postulate for liquid water, and finds one of its properties is
>>> {wet}.
>>
>> Where does the internal meaning postulate for liquid water initially
>> comes from? Definition?

> Sets of meanings are assigned to arbitrary word labels.

As above: Do you define first the whole set of meanings?
If so: Do you think that humans think and learn in that way?

[Peter Olcott]

I do not have that much initiative.
Basically every concept is defined using something like predicate logic,
and if it is not defined it does not exist.
Here are some key terms {truth conditional propositions} {model
theoretic} {possible worlds}.

>
>> The above was the first book that I read about Montague Semantics.
>>
>>>
>>>> The best that I can tell some extension of Montague's work can fully
>>>> specify every subtle detail of anything that can every be expressed in
>>>> words. Also it expresses these concepts using enormously simpler**
>>>> syntax, and the syntax can be organized to facilitate reasoning.
>>>>
>>>> ** Simpler than natural language.
>>>
>>> Where is the connection between a single word (isn't it an atom of the
>>> Grammar?) and an object in the world? Can a machine/AI learn such a
>>> world from being told by referencing to the object?
>> The atoms of the system that I am envisioning are {units of meaning} and
>> the connections between them. {units of meaning} do not perfectly map to
>> word boundaries.
>
> Do you have (an) example(s) for "units of meaning"?

The concept of {Plural} when combined with {Physically_Existing_Thing}
each of these are a {unit of meaning} and so is the result of the
combination.

>
>> All of those things that can only be known by the first-hand direct
>> experience of the actual physical sensation, the machine that I am
>> envisioning, will never know.
>
> Hm? You don't want the machine to have experience in the real world?

It will never be able to actually appreciate the beauty of a rainbow,
but, it will be able to talk as if it does.

>
>> Everything that can be expressed within
>> words can be fully understood by this machine.
>
> If a machine really have no contact to the real world it will never
> understand anything like a (theoretical) human without any senses.

All of its contact will be through words, and these will be enough.

[Peter Olcott]

Ah very good we are on about the same wave length.

I propose that there exists a means of self-sustaining machine learning
that is not stochastic, and is capable of learning everything on its own
such that it will eventually have the sum total of all human knowledge.
This machine must begin with a kernel of knowledge, {BootStrap Natural
Language Understanding} that would provide the basis for the {knowledge
of gaining more knowledge}.

The would be the mission of the first one of these two paths:

1) Determine the structure of knowledge.

Exhaustively specify all Compositionality of logical entailment.

The other path that would be further developed:

2) Determine the details of natural language syntax.

Exhaustively specify all Compositionality (besides logical entailment}

>
>>>>>> A Machine can comprehend anything on the basis of the {Meaning
>>>>>> Postulates} within its ontology.
>>>>>
>>>>> "Comprehend" or "deduce"? If you only/primarily think of
>>>>> comprehension: In which sense is it really more than deduction?
>>>> One does not deduce that liquid water is wet, one refers to the
>>>> internal
>>>> meaning postulate for liquid water, and finds one of its properties is
>>>> {wet}.
>>>
>>> Where does the internal meaning postulate for liquid water initially
>>> comes from? Definition?
>
>> Sets of meanings are assigned to arbitrary word labels.
>
> As above: Do you define first the whole set of meanings?

The order that knowledge is specified to the ontology will aim to
minimize the effort to get to self-sustained learning.
Besides that the order does not matter.

> If so: Do you think that humans think and learn in that way?

We can skip how humans learn and goes straight to the organization of
{units of meaning} that is equivalent to human comprehension. We can
think of this as differing types of connections between {units of
meaning}. All it seems to break down to individuals (including sets),
properties, and relations.

[Peter Olcott]

> That would be the mission of the first of these two paths:

> 1) Determine the structure of knowledge.
> Exhaustively specify all Compositionality of logical entailment.
>
> The other path that would be further developed:
> 2) Determine the details of natural language syntax.
> Exhaustively specify all Compositionality (besides logical entailment}
>

> Determining the minimum set of prerequisite concepts that must be
> manually specified would be another sub-task of the project. It might
> be able to learn something as basic as plural, on its own.

>>
>>>>>>> A Machine can comprehend anything on the basis of the {Meaning
>>>>>>> Postulates} within its ontology.
>>>>>>
>>>>>> "Comprehend" or "deduce"? If you only/primarily think of
>>>>>> comprehension: In which sense is it really more than deduction?
>>>>> One does not deduce that liquid water is wet, one refers to the
>>>>> internal
>>>>> meaning postulate for liquid water, and finds one of its
>>>>> properties is
>>>>> {wet}.
>>>>
>>>> Where does the internal meaning postulate for liquid water initially
>>>> comes from? Definition?
>>
>>> Sets of meanings are assigned to arbitrary word labels.
>>
>> As above: Do you define first the whole set of meanings?
> The order that knowledge is specified to the ontology will aim to

> minimize the effort to get to achieve self-sustained learning.

> Besides that the order does not matter.
>
>> If so: Do you think that humans think and learn in that way?

> We can skip how humans learn and go straight to determining the
> organization of {units of meaning} equivalent to human comprehension.

> We can think of this as differing types of connections between {units

> of meaning}. It all seems to break down to individuals (including
> sets), properties, and relations.
>

[Peter T. Daniels]

> >>http://www.amazon.com/Formal-Semantics-Introduction-Cambridge-Linguis...

Mr. Olcott, meet Mr. Turing.

> >> Everything that can be expressed within
> >> words can be fully understood by this machine.
>
> > If a machine really have no contact to the real world it will never
> > understand anything like a (theoretical) human without any senses.
>
> All of its contact will be through words, and these will be enough.

How does it encounter "words"? I thought computers only know "on" and
"off."

[Mok-Kong Shen]

To accumulate experiences in "feeling" it has to be able to feel somehow
basically/primitively in the first place. And one is then back to
square one, IMHO.

M. K. Shen

[casey]

A computer *program* can know as much about its binary switches as we
know about our neurons.

Computers can only do ONE thing, carry out a series of steps in a
program.

What a program can do, including "knowing" something, is still an open
question.

jc

[Dissitra]

find me xxx

[Peter Olcott]

Computers will know words through the connections within an ontology.

[Dissitra]

like some leek she shall certainly know their farmer labourer

and homsexuals shat certaily be aware of their own geometricies-ess


still no money in that crap?

try an abortion

[Burkart Venzke]

OK, my idea of your "feel" was the "intuitive" comprehension of the
content of the picture, "intuitive" in the sense of using much knowledge
seemingly unconsciously.

On the other hand, we cannot feel the content of a picture as we can
feel the burning sun or the wind. But we can associate the colors, forms
or interpretated elements of the picture with more or less directly with
our natural feelings. This is a point machines as non-biological
creatures are not able to do directly (except we implement them feelings
explicitely).

However, I don't think that we want perfect copies of humans; artifical
intelligence without (real) emotions should be enough. So, an
unemotional analysis of a picture based on learned criteria may be enough.

Burkart

[Burkart Venzke]

I see, this is a (set of) model(s) without any inductive aspects.

>>> The above was the first book that I read about Montague Semantics.
>>>
>>>>
>>>>> The best that I can tell some extension of Montague's work can fully
>>>>> specify every subtle detail of anything that can every be expressed in
>>>>> words. Also it expresses these concepts using enormously simpler**
>>>>> syntax, and the syntax can be organized to facilitate reasoning.
>>>>>
>>>>> ** Simpler than natural language.
>>>>
>>>> Where is the connection between a single word (isn't it an atom of the
>>>> Grammar?) and an object in the world? Can a machine/AI learn such a
>>>> world from being told by referencing to the object?
>>> The atoms of the system that I am envisioning are {units of meaning} and
>>> the connections between them. {units of meaning} do not perfectly map to
>>> word boundaries.
>>
>> Do you have (an) example(s) for "units of meaning"?
> The concept of {Plural} when combined with {Physically_Existing_Thing}
> each of these are a {unit of meaning} and so is the result of the
> combination.
>>
>>> All of those things that can only be known by the first-hand direct
>>> experience of the actual physical sensation, the machine that I am
>>> envisioning, will never know.
>>
>> Hm? You don't want the machine to have experience in the real world?
> It will never be able to actually appreciate the beauty of a rainbow,
> but, it will be able to talk as if it does.

It cannot appreciate the beauty of a rainbow, not even it cann see it
for itself. So it can only talk about it what it "heard" (was being
told/learned/implemented) from others (for example the programmer).

>>> Everything that can be expressed within
>>> words can be fully understood by this machine.
>>
>> If a machine really have no contact to the real world it will never
>> understand anything like a (theoretical) human without any senses.
> All of its contact will be through words, and these will be enough.

For me, words (as only abstract items) are not enough.

[Burkart Venzke]

Am 06.09.2012 21:40, schrieb casey:
> On Sep 6, 1:39 pm, "Peter T. Daniels"<gramma...@verizon.net> wrote:
>> On Sep 5, 6:34 pm, Peter Olcott<OCR4Screen> wrote:
>>
>>
>>
>>
>>
>>> On 9/5/2012 4:55 PM, Burkart Venzke wrote:
>>
>>>> Am 05.09.2012 01:02, schrieb Peter Olcott:
>>>>> On 9/4/2012 5:41 PM, Burkart Venzke wrote:
>>>>>> Am 05.09.2012 00:26, schrieb Peter Olcott:
>>>>>>> On 9/4/2012 3:02 PM, Burkart Venzke wrote:
>>>>>>>> Am 04.09.2012 01:45, schrieb Peter Olcott:
>>>>>>>>> On 9/3/2012 5:10 PM, Burkart Venzke wrote:
>>>>>>>>>> Am 03.09.2012 19:10, schrieb Peter Olcott:

Right.

> Computers can only do ONE thing, carry out a series of steps in a
> program.

...as humans carry out their series of steps determined by their brain
and body.

> What a program can do, including "knowing" something, is still an open
> question.

Additionally, programs can learn so that their series of steps become
unpredictable.

"Knowing" is really difficult, even philosophical seen. What can we
really know? Or how precisely and/or surely?

Burkart

[Peter Olcott]

Yet that is all that we have between us: You and I.

[casey]

On Sep 7, 7:55 am, Burkart Venzke <b...@gmx.de> wrote:
> [...]

> "Knowing" is really difficult, even philosophical seen. What can we
> really know? Or how precisely and/or surely?

Does a light seeking robot know where the light is?

When does an act of reacting to something become knowing that
something?

jc

>
> Burkart
>
>

[Burkart Venzke]

It might be possible. But a good teacher (as for children) will save a
lot of time (the time of human evolution more or less).

>> that is not stochastic,

I have been hoping so, too, but because I think of inductive learning as
the real basis, it is not really possible. For example, you can learn
that Paris is the capital city of France, but if you do so you have to
believe your teacher and he might be wrong, even might lie...

>> and is capable of learning everything
>> on its own such that it will eventually have the sum total of all
>> human knowledge. This machine must begin with a kernel of knowledge,
>> {BootStrap Natural Language Understanding} that would provide the
>> basis for the {knowledge of gaining more knowledge}.

Something like that, that's true. We should not forget to guide the
machine to "good knowledge", for example to help humans instead of
fighting against them.

I fear that an explicite, finally structure is not perfect and/or
flexible enough.

>> The other path that would be further developed:
>> 2) Determine the details of natural language syntax.
>> Exhaustively specify all Compositionality (besides logical entailment}
>>
>> Determining the minimum set of prerequisite concepts that must be
>> manually specified would be another sub-task of the project. It might
>> be able to learn something as basic as plural, on its own.

This sounds better. I don't know exactly what "natural language syntax"
is but even language should be learned on a small communication basis by
interaction with a speaking (communicating) human who additionally uses
world items in combination with words so that words can really be
connect to the world.

>>>>>>>> A Machine can comprehend anything on the basis of the {Meaning
>>>>>>>> Postulates} within its ontology.
>>>>>>>
>>>>>>> "Comprehend" or "deduce"? If you only/primarily think of
>>>>>>> comprehension: In which sense is it really more than deduction?
>>>>>> One does not deduce that liquid water is wet, one refers to the
>>>>>> internal
>>>>>> meaning postulate for liquid water, and finds one of its
>>>>>> properties is
>>>>>> {wet}.
>>>>>
>>>>> Where does the internal meaning postulate for liquid water initially
>>>>> comes from? Definition?
>>>
>>>> Sets of meanings are assigned to arbitrary word labels.
>>>
>>> As above: Do you define first the whole set of meanings?
>> The order that knowledge is specified to the ontology will aim to
>> minimize the effort to get to achieve self-sustained learning.
>> Besides that the order does not matter.
>>
>>> If so: Do you think that humans think and learn in that way?
>> We can skip how humans learn

We could but I think that we still can learn a lot of human's learning.

>> and go straight to determining the
>> organization of {units of meaning} equivalent to human comprehension.
>> We can think of this as differing types of connections between {units
>> of meaning}. It all seems to break down to individuals (including
>> sets), properties, and relations.

As far as you do not define/explain here what "units of meaning" etc
are, it is too abstract/unimaginative for me.

[Peter Olcott]

When people acquire knowledge this knowledge is almost never stochastic.
I will use your own example:
Do most people feel that there is an extremely good chance that Paris is
the capitol of France?
or Do most people simply comprehend that Paris is defined as the capitol
of France?

>
>>> and is capable of learning everything
>>> on its own such that it will eventually have the sum total of all
>>> human knowledge. This machine must begin with a kernel of knowledge,
>>> {BootStrap Natural Language Understanding} that would provide the
>>> basis for the {knowledge of gaining more knowledge}.
>
> Something like that, that's true. We should not forget to guide the
> machine to "good knowledge", for example to help humans instead of
> fighting against them.

Because of the Chinese room thought experiment, this machine will never
have a will of its own, so we must infuse it with something. Wde have to
be careful when we do this or we could possibly get the result of
Asimov's I Robot.

It already has its own perfect structure, we only need to discover this,
to reverse-engineer it.

>
>>> The other path that would be further developed:
>>> 2) Determine the details of natural language syntax.
>>> Exhaustively specify all Compositionality (besides logical entailment}
>>>
>>> Determining the minimum set of prerequisite concepts that must be
>>> manually specified would be another sub-task of the project. It might
>>> be able to learn something as basic as plural, on its own.
>
> This sounds better. I don't know exactly what "natural language syntax" is

a specific example would be: English Grammar

> but even language should be learned on a small communication basis by
> interaction with a speaking (communicating) human who additionally
> uses world items in combination with words so that words can really be
> connect to the world.

I was envisioning that most of its knowledge would be from books, and
articles. It would also have the ability to ask questions.

>
>>>>>>>>> A Machine can comprehend anything on the basis of the {Meaning
>>>>>>>>> Postulates} within its ontology.
>>>>>>>>
>>>>>>>> "Comprehend" or "deduce"? If you only/primarily think of
>>>>>>>> comprehension: In which sense is it really more than deduction?
>>>>>>> One does not deduce that liquid water is wet, one refers to the
>>>>>>> internal
>>>>>>> meaning postulate for liquid water, and finds one of its
>>>>>>> properties is
>>>>>>> {wet}.
>>>>>>
>>>>>> Where does the internal meaning postulate for liquid water initially
>>>>>> comes from? Definition?
>>>>
>>>>> Sets of meanings are assigned to arbitrary word labels.
>>>>
>>>> As above: Do you define first the whole set of meanings?
>>> The order that knowledge is specified to the ontology will aim to
>>> minimize the effort to get to achieve self-sustained learning.
>>> Besides that the order does not matter.
>>>
>>>> If so: Do you think that humans think and learn in that way?
>>> We can skip how humans learn
>
> We could but I think that we still can learn a lot of human's learning.

The reasoning that is required to fully integrate the many dimensions of
a piece of knowledge is required.
Rote memorization is entirely moot, an machine needs never forget any
detail even if only encountered once.

>
>>> and go straight to determining the
>>> organization of {units of meaning} equivalent to human comprehension.
>>> We can think of this as differing types of connections between {units
>>> of meaning}. It all seems to break down to individuals (including
>>> sets), properties, and relations.
>
> As far as you do not define/explain here what "units of meaning" etc
> are, it is too abstract/unimaginative for me.

Pretty much the sense meanings of morphemes.
http://en.wikipedia.org/wiki/Morpheme

[casey]

On Sep 7, 8:09 am, Peter Olcott <OCR4Screen> wrote:
> On 9/6/2012 4:45 PM, Burkart Venzke wrote:

[...]

> >>> If a machine really have no contact to the real world it will never
> >>> understand anything like a (theoretical) human without any senses.
> >> All of its contact will be through words, and these will be enough.
>
> > For me, words (as only abstract items) are not enough.
>
> Yet that is all that we have between us: You and I.

But you both share the same world unlike a machine without contact to
a real world.

jc

[casey]

On Sep 7, 8:53 am, Peter Olcott <OCR4Screen> wrote:
> [...]

>
> Because of the Chinese room thought experiment, this machine will never

> have a will of its own, so we must infuse it with something. We have to

> be careful when we do this or we could possibly get the result of
> Asimov's I Robot.

This is about how the machine values a set of possible moves.

All machines that make decisions (like a chess playing program)
must be able to value its moves.

The idea of a "will" is really about what move you have decided to
make.

A lot of these words are colored by being conscious or unconscious.

My fingers *know* where the keys are when I touch type but the
conscious
part of me does not. All I *know* is what I want to type and the
subconscious
parts of the brain carry it out.

The only difference between me knowing and the finger control circuits
knowing is the former is conscious knowing and I can talk about it.

> I was envisioning that most of its knowledge would be from books, and
> articles. It would also have the ability to ask questions.

And how is this knowledge to be represented in the machine?

> Rote memorization is entirely moot, an machine needs never forget any
> detail even if only encountered once.

Even if everything was memorized (stored) how would it be recalled?

I get the feeling from this thread you guys haven't actually
started coding any of these thoughts?

jc

[Dissitra]

I wonder when the Chinese experiment was done. Was this after the that
Shanghai revolution or before in normal native popululance of Chinese,
he?

[Peter Olcott]

That only means that it has to be told everything until it knows enough
to go out on its own surfing the web, reading books, and articles. It
will also have the ability to ask questions, and/or request reading
material.

[Peter Olcott]

On 9/6/2012 6:55 PM, casey wrote:
> On Sep 7, 8:53 am, Peter Olcott <OCR4Screen> wrote:
>> [...]
>>
>> Because of the Chinese room thought experiment, this machine will never
>> have a will of its own, so we must infuse it with something. We have to
>> be careful when we do this or we could possibly get the result of
>> Asimov's I Robot.
> This is about how the machine values a set of possible moves.
>
> All machines that make decisions (like a chess playing program)
> must be able to value its moves.
>
> The idea of a "will" is really about what move you have decided to
> make.

Yes that is exactly it. It will follow a hierarchy of abstract goals,
probably including a pinnacle goal at the top of the hierarchy. I think
that the goal for humanity would be to maximize the duration, and depth
of joy.

> A lot of these words are colored by being conscious or unconscious.

What I am learning by studying Linguistics is that there are two aspects
to denotation:
a) A reference to a thing in the world.
b) A set of properties that selects a set of things in the world.
What the concept of {consciousness} is depends on the set of properties
that are assigned to the term "consciousness".

> My fingers *know* where the keys are when I touch type but the
> conscious
> part of me does not. All I *know* is what I want to type and the
> subconscious
> parts of the brain carry it out.
>
> The only difference between me knowing and the finger control circuits
> knowing is the former is conscious knowing and I can talk about it.

I think that these aspects may be moot to ontology design.

>
>> I was envisioning that most of its knowledge would be from books, and
>> articles. It would also have the ability to ask questions.
> And how is this knowledge to be represented in the machine?

Different types of connections between nodes of a di-graph.

>
>> Rote memorization is entirely moot, an machine needs never forget any
>> detail even if only encountered once.
> Even if everything was memorized (stored) how would it be recalled?

Traversals through the knowledgebase nodes. There are a pair on ontologies:
a) Knowledge Base
b) Specific Situation.

The Specific Situation ontology is optimized for reasoning, and is
constructed on the basis the the Knowledge Base and the Discourse.

> I get the feeling from this thread you guys haven't actually
> started coding any of these thoughts?
>
> jc
>
>
>
>

I am working on the architectural design of self-sustaining machine
learning from natural language words, right here right now as I answer
these posts. Although I am doing this for enjoyment, I do expect to make
substantial progress. I have been thinking about these things since at
least 1986.

[Mok-Kong Shen]

Am 06.09.2012 23:35, schrieb Burkart Venzke:

> However, I don't think that we want perfect copies of humans; artifical
> intelligence without (real) emotions should be enough. So, an
> unemotional analysis of a picture based on learned criteria may be enough.

But that casts IMHO essential doubt on whether we could build machines
that can "understand" and hence acquire languages in the sense as we do.

M. K. Shen

[António Marques]

PO thinks knowledge can be acquired exclusively from language plus a certain
kernel of it. Not only the jury is out on that, we haven't got anywhere on
the matter of how to encode the required kernel, or even if it's encodable
at all. PO thinks we communicate through language only and forgets that we
have an awful lot in common and language may simply be a catalyzer.

[Peter Olcott]

Language is the structure of thought.

Conceptual comprehension is essentially nothing more than knowing the
meaning of words.
Every possible conceptual meaning that can ever be expressed can be
expressed within an ontology.

A machine will be able to demonstrate the functional equivalent of human
comprehension as soon as its ontology is sufficiently populated, and is
organized to facilitate every type of reasoning that humans are capable of.

The required kernel of meta-knowledge needed to facilitate fully
self-sustained LearningByReading can be reverse-engineered.

[Mok-Kong Shen]

Am 07.09.2012 16:34, schrieb Peter Olcott:

> The required kernel of meta-knowledge needed to facilitate fully
> self-sustained LearningByReading can be reverse-engineered.

Reverse-engineered from what? From the neural network of the
human brain??

M. K. Shen

[Peter Olcott]

No those paths are dead-ends. Reverse-engineered on the basis of the
fundamental structure of the universal set of conceptual knowledge.

[casey]

On Sep 8, 12:34 am, Peter Olcott <OCR4Screen> wrote:
> [...]

> Language is the structure of thought.

Is it really? Are animals unable to think in any
way at all? We may talk to ourselves when we think
but I also feel there is a wordless basis to how
we think. Words may order our thoughts and allow
us to communicate those thoughts but is language
really the mechanism of thought? Or did it evolve
to allow us to communicate our thoughts and then
become a means to talk to ourselves?

> Conceptual comprehension is essentially nothing
> more than knowing the meaning of words.

And what neurologically does "knowing the meanings
of words" mean? Do you think animals cannot have
concepts such as that will eat me, that looks bad,
that looks good, that is something I can climb and
so on simply because they cannot express those
thoughts in words?

> Every possible conceptual meaning that can ever
> be expressed can be expressed within an ontology.
>
>
> A machine will be able to demonstrate the functional
> equivalent of human comprehension as soon as its
> ontology is sufficiently populated, and is organized
> to facilitate every type of reasoning that humans
> are capable of.

And at point is a child's ontology sufficiently
populated in order to reason?

> The required kernel of meta-knowledge needed to
> facilitate fully self-sustained LearningByReading
> can be reverse-engineered.

When is it a mechanism required to learn vs. the
knowledge needed to learn. We have the knowledge
to add numbers. A computer has the mechanism to
add numbers. The "knowledge" for adding is embodied
in the connections between the logic gates.

Do we have the knowledge to see the world or do
we have the mechanisms that allow us to see a
world which we can then have knowledge about?
Do we need language to know anything about the
world? What is it an animal without language may
be said to know?

jc

[casey]

And the brain is a working example of a mechanism that can reverse
engineer the Universe so it seems to me that is exactly what you want
to build and that would be possible if you first reverse engineer (or
duplicate in some way) what the brain does.

jc

[Peter Olcott]

Yes we reverse-engineer the functional end-result of comprehension.

[Peter Olcott]

We can not assume that animals do not have internal language.

[Franz Gnaedinger]

On Sep 7, 4:34 pm, Peter Olcott <OCR4Screen> wrote:
>
> A machine will be able to demonstrate the functional equivalent of human
> comprehension as soon as its ontology is sufficiently populated, and is
> organized to facilitate every type of reasoning that humans are capable of.

I will save the world, as soon as someone can give me the phone number
and e-mail address of God.

[Burkart Venzke]

Am 07.09.2012 00:13, schrieb casey:
> On Sep 7, 7:55 am, Burkart Venzke<b...@gmx.de> wrote:
>> [...]
>> "Knowing" is really difficult, even philosophical seen. What can we
>> really know? Or how precisely and/or surely?
>
> Does a light seeking robot know where the light is?

It think that "knowing" is a fuzzy notion.
Perhaps we often really should use it in the way your robot does.

> When does an act of reacting to something become knowing that
> something?

It's fuzzy, isn't it?
Perhaps something like a degree or hierarchy of knowing would be helpful.

Burkart

[Burkart Venzke]

We have different models and certainly different aims we want to reach.

I hope that AI systems will learn as much as possible to be helpful in
artifical and in real life as robots.

[Burkart Venzke]

For me, the possibility of understanding every single word is missing.
The only chance is to implement the understanding in your meanings.

[Peter Olcott]

The understanding that the machine has is represented within its second
ontology, the ontology of the specific situation. It builds this on the
basis of its big ontology combined with the discourse.

[Curt Welch]

Burkart Venzke <b...@gmx.de> wrote:
> Am 07.09.2012 00:09, schrieb Peter Olcott:
> > On 9/6/2012 4:45 PM, Burkart Venzke wrote:
> >> For me, words (as only abstract items) are not enough.
> > Yet that is all that we have between us: You and I.

Actually, what we have between us, is a good many computers. Interesting
how the computers become so invisible to us.

> We have different models and certainly different aims we want to reach.
>
> I hope that AI systems will learn as much as possible to be helpful in
> artifical and in real life as robots.

--
Curt Welch http://CurtWelch.Com/
cu...@kcwc.com http://NewsReader.Com/

[Burkart Venzke]

Hm, where does the data for your "ontology of the specific situation"
come from? Let's say for example we watch a boat coming down a river...

[Mok-Kong Shen]

Am 07.09.2012 19:45, schrieb Peter Olcott:

> Yes we reverse-engineer the functional end-result of comprehension.

But IMHO we don't know where/what/how "really" "exactly" comprehension
or consciousness are.

M. K. Shen

[Peter Olcott]

The purpose of the machine is to answer questions from people. All input
and output is in natural language.

[Peter Olcott]

We do not need to know this for the purpose of duplicating its
functional end-result.

[Burkart Venzke]

Then good luck with such a machine. My machine/AI should also have
contact with a common (object) world, I also want to talk with the AI
about the world.

[Mok-Kong Shen]

Am 09.09.2012 15:04, schrieb Peter Olcott:
> On 9/9/2012 4:07 AM, Mok-Kong Shen wrote:
>> Am 07.09.2012 19:45, schrieb Peter Olcott:
>>
>>> Yes we reverse-engineer the functional end-result of comprehension.
>>
>> But IMHO we don't know where/what/how "really" "exactly" comprehension
>> or consciousness are.
>>

> We do not need to know this for the purpose of duplicating its
> functional end-result.

I am a layman and could well err from what I read about a little bit.
But wasn't there a certain general argument against this kind of
thinking termed "Chinese room"?

M. K. Shen

[Peter Olcott]

No you are right that you are wrong.
The Chinese room proves that if a machine does duplicate the functional
end-result of consciousness that it is not real.
I say forget this get it to duplicate the functional end-result of
comprehension, and then it can provide medical advice for $1.50.

[Mok-Kong Shen]

Am 11.09.2012 12:00, schrieb Peter Olcott:

> The Chinese room proves that if a machine does duplicate the functional
> end-result of consciousness that it is not real.
> I say forget this get it to duplicate the functional end-result of
> comprehension, and then it can provide medical advice for $1.50.

I doubt to have rightly captured what you intend to achieve. Let
me use an example (or only a more or less far analogy) to illustrate
my point: If one has a database and a software to let a machine answer
questions of the kind posed to a travelling agency, one could certainly
make efforts to refine the software for better performance. But has the
machine "learned" something that way? It is IMHO the human (the
programmer) that has made the higher performance of the machine
available. That's of course fine for the practice. But the machine has
"learned" nothing thereby in my view. (The programmer might have
learned something though.)

An�way I like to quote from a paper of H. M�hlenbein, Computational
Intelligence: The Legacy of Alan Turing and John von Neumann, in
C. K. Mumford, L. C. Jain (Eds.), Computational Intelligence,
Berlin, 2009. The author concludes with:

Up to now computational intelligence was successful in specialized
applications only, automatic passing the Turing test or understanding
languages are not yet in sight.

M. K. Shen

[PeteOlcott]

On Sep 11, 9:31 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
> Am 11.09.2012 12:00, schrieb Peter Olcott:
>
> > The Chinese room proves that if a machine does duplicate the functional
> > end-result of consciousness that it is not real.
> > I say forget this get it to duplicate the functional end-result of
> > comprehension, and then it can provide medical advice for $1.50.
>
> I doubt to have rightly captured what you intend to achieve. Let
> me use an example (or only a more or less far analogy) to illustrate
> my point: If one has a database and a software to let a machine answer
> questions of the kind posed to a travelling agency, one could certainly
> make efforts to refine the software for better performance. But has the
> machine "learned" something that way? It is IMHO the human (the
> programmer) that has made the higher performance of the machine
> available. That's of course fine for the practice. But the machine has
> "learned" nothing thereby in my view. (The programmer might have
> learned something though.)
>

> Anýway I like to quote from a paper of H. Mühlenbein, Computational

> Intelligence: The Legacy of Alan Turing and John von Neumann, in
> C. K. Mumford, L. C. Jain (Eds.), Computational Intelligence,
> Berlin, 2009. The author concludes with:
>
>     Up to now computational intelligence was successful in specialized
>     applications only, automatic passing the Turing test or understanding
>     languages are not yet in sight.
>
> M. K. Shen

The functional equivalent to human comprehension by a machine can be
mostly achieved by a sufficiently populated ontology that is organized
to facilitate reasoning.

[Doc O'Leary]

In article <k2mv2s$h1c$1...@news.albasani.net>,

There are all sorts of flawed arguments against AI.

--
iPhone apps that matter: http://appstore.subsume.com/
My personal UDP list: 127.0.0.1, localhost, googlegroups.com, theremailer.net,
and probably your server, too.

[Mok-Kong Shen]

Am 11.09.2012 16:40, schrieb PeteOlcott:
> On Sep 11, 9:31 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
>> Am 11.09.2012 12:00, schrieb Peter Olcott:
>>
>>> The Chinese room proves that if a machine does duplicate the functional
>>> end-result of consciousness that it is not real.
>>> I say forget this get it to duplicate the functional end-result of
>>> comprehension, and then it can provide medical advice for $1.50.
>>
>> I doubt to have rightly captured what you intend to achieve. Let
>> me use an example (or only a more or less far analogy) to illustrate
>> my point: If one has a database and a software to let a machine answer
>> questions of the kind posed to a travelling agency, one could certainly
>> make efforts to refine the software for better performance. But has the
>> machine "learned" something that way? It is IMHO the human (the
>> programmer) that has made the higher performance of the machine
>> available. That's of course fine for the practice. But the machine has
>> "learned" nothing thereby in my view. (The programmer might have
>> learned something though.)
>>

>> An�way I like to quote from a paper of H. M�hlenbein, Computational

>> Intelligence: The Legacy of Alan Turing and John von Neumann, in
>> C. K. Mumford, L. C. Jain (Eds.), Computational Intelligence,
>> Berlin, 2009. The author concludes with:
>>
>> Up to now computational intelligence was successful in specialized
>> applications only, automatic passing the Turing test or understanding
>> languages are not yet in sight.
>>
>> M. K. Shen
>
> The functional equivalent to human comprehension by a machine can be
> mostly achieved by a sufficiently populated ontology that is organized
> to facilitate reasoning.

Could you kindly illustrate with something concrete to let one
understand what's that "onlotogy" stuff that you want to put into
a machine?

M. K. Shen

[Mok-Kong Shen]

Am 11.09.2012 16:41, schrieb Doc O'Leary:
> In article <k2mv2s$h1c$1...@news.albasani.net>,
> Mok-Kong Shen <mok-ko...@t-online.de> wrote:
>
>> Am 09.09.2012 15:04, schrieb Peter Olcott:
>>> On 9/9/2012 4:07 AM, Mok-Kong Shen wrote:
>>>> Am 07.09.2012 19:45, schrieb Peter Olcott:
>>>>
>>>>> Yes we reverse-engineer the functional end-result of comprehension.
>>>>
>>>> But IMHO we don't know where/what/how "really" "exactly" comprehension
>>>> or consciousness are.
>>>>
>>> We do not need to know this for the purpose of duplicating its
>>> functional end-result.
>>
>> I am a layman and could well err from what I read about a little bit.
>> But wasn't there a certain general argument against this kind of
>> thinking termed "Chinese room"?
>
> There are all sorts of flawed arguments against AI.

As far as I could remember, the person who raised that argument
was a not-unknown scientist.

M. K. Shen

[PeteOlcott]

On Sep 11, 10:26 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
> Am 11.09.2012 16:40, schrieb PeteOlcott:
>
>
>
>
>
> > On Sep 11, 9:31 am, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
> >> Am 11.09.2012 12:00, schrieb Peter Olcott:
>
> >>> The Chinese room proves that if a machine does duplicate the functional
> >>> end-result of consciousness that it is not real.
> >>> I say forget this get it to duplicate the functional end-result of
> >>> comprehension, and then it can provide medical advice for $1.50.
>
> >> I doubt to have rightly captured what you intend to achieve. Let
> >> me use an example (or only a more or less far analogy) to illustrate
> >> my point: If one has a database and a software to let a machine answer
> >> questions of the kind posed to a travelling agency, one could certainly
> >> make efforts to refine the software for better performance. But has the
> >> machine "learned" something that way? It is IMHO the human (the
> >> programmer) that has made the higher performance of the machine
> >> available. That's of course fine for the practice. But the machine has
> >> "learned" nothing thereby in my view. (The programmer might have
> >> learned something though.)
>

> >> Anýway I like to quote from a paper of H. Mühlenbein, Computational

> >> Intelligence: The Legacy of Alan Turing and John von Neumann, in
> >> C. K. Mumford, L. C. Jain (Eds.), Computational Intelligence,
> >> Berlin, 2009. The author concludes with:
>
> >>      Up to now computational intelligence was successful in specialized
> >>      applications only, automatic passing the Turing test or understanding
> >>      languages are not yet in sight.
>
> >> M. K. Shen
>
> > The functional equivalent to human comprehension by a machine can be
> > mostly achieved by a sufficiently populated ontology that is organized
> > to facilitate reasoning.
>
> Could you kindly illustrate with something concrete to let one
> understand what's that "onlotogy" stuff that you want to put into
> a machine?
>

> M. K. Shen- Hide quoted text -
>
> - Show quoted text -

http://cyc.com/cyc/technology/whatiscyc_dir/whatdoescycknow

[Mok-Kong Shen]

Am 11.09.2012 20:15, schrieb PeteOlcott:

> http://cyc.com/cyc/technology/whatiscyc_dir/whatdoescycknow

I see. I remember now having even read a news about CYC but long long
ago. Now what's the intention of your proposal? Are you thinking of
getting a better result than CYC, which after about 3 decades and
presumably a big sum of expenses doesn't yet reach a success (at least
according to the article I cited)?

M. K. Shen

[PeteOlcott]

They skipped the design step and went straight to coding. I am
attempting to fill in this gap.

[Curt Welch]

Peter Olcott <OCR4Screen> wrote:
> On 9/11/2012 4:08 AM, Mok-Kong Shen wrote:
> > Am 09.09.2012 15:04, schrieb Peter Olcott:
> >> On 9/9/2012 4:07 AM, Mok-Kong Shen wrote:
> >>> Am 07.09.2012 19:45, schrieb Peter Olcott:
> >>>
> >>>> Yes we reverse-engineer the functional end-result of comprehension.
> >>>
> >>> But IMHO we don't know where/what/how "really" "exactly"
> >>> comprehension or consciousness are.
> >>>
> >> We do not need to know this for the purpose of duplicating its
> >> functional end-result.
> >
> > I am a layman and could well err from what I read about a little bit.
> > But wasn't there a certain general argument against this kind of
> > thinking termed "Chinese room"?
> >
> > M. K. Shen
> >
> No you are right that you are wrong.
> The Chinese room proves that if a machine does duplicate the functional
> end-result of consciousness that it is not real.

Ha ha ha ha ha.

The Chinese Room proves nothing. It's just a thought experiment. It works
the same way as a Rorschach inkblot test does. People see what they were
already predisposed to see in it. You can use it as a psychological test
to uncover someone's beliefs on duality and the mind/body problem.

It was not created to be a test, it was created to be a proof. But it's not
a proof.

> I say forget this get it to duplicate the functional end-result of
> comprehension, and then it can provide medical advice for $1.50.

Yes, that's a perfectly good approach.

[Curt Welch]

PeteOlcott <peteo...@gmail.com> wrote:

:)

[Mok-Kong Shen]

Am 11.09.2012 21:18, schrieb PeteOlcott:
> On Sep 11, 1:31 pm, Mok-Kong Shen <mok-kong.s...@t-online.de> wrote:
>> Am 11.09.2012 20:15, schrieb PeteOlcott:
>>
>>> http://cyc.com/cyc/technology/whatiscyc_dir/whatdoescycknow
>>
>> I see. I remember now having even read a news about CYC but long long
>> ago. Now what's the intention of your proposal? Are you thinking of
>> getting a better result than CYC, which after about 3 decades and
>> presumably a big sum of expenses doesn't yet reach a success (at least
>> according to the article I cited)?

> They skipped the design step and went straight to coding. I am
> attempting to fill in this gap.

Sorry for "plainly" expressing my humble personal opinions: You
seem to like to do some "philosophical" work to achieve that grand
goal, presumably lasting at least a few decades, which the CYC people
have already spent till now.

M. K. Shen

[Doc O'Leary]

In article <k2nldf$voq$2...@news.albasani.net>,

And? Appeal to authority is a logical fallacy. The argument itself is
painfully flawed.

[Mok-Kong Shen]

Am 12.09.2012 16:36, schrieb Doc O'Leary:
> In article <k2nldf$voq$2...@news.albasani.net>,
> Mok-Kong Shen <mok-ko...@t-online.de> wrote:
>
>> Am 11.09.2012 16:41, schrieb Doc O'Leary:
>>>
>>> There are all sorts of flawed arguments against AI.
>>
>> As far as I could remember, the person who raised that argument
>> was a not-unknown scientist.
>
> And? Appeal to authority is a logical fallacy. The argument itself is
> painfully flawed.

Then it would be fine always to give some reasons for that or else
refer to literature instead of simply claiming it's flawed, IMHO.

M. K. Shen

[Peter Olcott]

It is not philosophical, it is system design.
I do not expect to take many decades, yet, it is common knowledge that
any sufficiently complex system can take as much as an infinite amount
of time, if built prior to sufficient design.

[Burkart Venzke]

Very optimistic - let's see when you really have a running system!
In theory, a lot of things are (seem to be) quite easy...

Perhaps it is better to make a first concrete step (a concrete system
design, not only vague theory).

[Peter Olcott]

I have already made some progress at this.
I have broken down the goal of Learning by Reading (LbR) into
two key elements that must be further elaborated:
(1) Complete Elaboration of Natural Language syntax.
This would begin with the findings of the current statistical
methods, and would not be limited to conventional parts-of-speech.

(2) Complete Elaboration of the inherent structure of knowledge.
It looks like predicate logic (including higher order predicate
logic) would form the sufficient expressive basis.

Also I have found that a key concept of Linguistics Compositionality can
be greatly simplified:
All conceptual meanings are comprised of connections between units of
meaning, and these
connections can be broken down into two essential types:
(1) Logical Entailment (semantics)
(2) Everything else (syntax) (many sub-types, including thematic
relations/roles, and sub-categorization)
and that the {units of meaning} are can come from semantics and pragmatics.

[Mok-Kong Shen]

As Venzke remarked, you should better post after one year your first
beta-version of your software and let people discuss.

M. K. Shen