A philosophy of AI based on Epistemology
Will Pearson
the mind and whatever that makes of the environment.
So what does this have to do with Epistemology? Well the question of
how information is derived information from the environment is
Epistemology (I shall be using it in the non-normative sense).
I am interested in making AI, primarily to make computers easier to
use. So it will be useful for them to be able to understand humans and
their irrationality (I don't want my computer to not be able to answer
questions about fiction for example).
So what can epistemology tell us about AI, I think the most important
one is Fallibilism, derived from the fact that we don't know when our
senses are reflecting the world. That is every concept about the world
should be allowed to be discarded (including ideas about learning,
what ideas should be created etc..).
So we obviously cannot start with a tabula rasa with no concepts, else
nothing would be done. However whatever concepts we start with, should
be able to change. Now assuming a computational basis of mind, I shall
translate the concepts to functional programs. By functional programs,
I mean a program that interacts with the outside world or has a
complex affect on another functional program.
My current research area is an ALife like system with many programs in
trying to overwrite each other, and being rewarded for performing a
task. The programs are turing complete so could calculate any
function. I am currently interested in discovering how to create
programs that change there offspring in a reasonable fashion so that
different ideas can be found in the system.
In brief my philosophy of mind is a fallibilistic mind that has ideas
(as programs) creating other ideas creating new ways to create ideas,
cooperating with other ideas, competing with others all to try and
achieve a goal.
Will Pearson
Neil W Rickert
>So what can epistemology tell us about AI, I think the most important
>one is Fallibilism, derived from the fact that we don't know when our
>senses are reflecting the world. That is every concept about the world
>should be allowed to be discarded (including ideas about learning,
>what ideas should be created etc..).
Quite a bit of AI work is already based on epistemology. I'm
not sure what you plan that is different.
Personally, I think epistemology is a pile of horse manure.
nucleus
The "right" term is horseshit (tm).
David Longley
writes
I take it that horseshit is your trademark nucleus <g>?
--
David Longley
David Longley
Pearson <whaz...@yahoo.com> writes
It is a sad fact that some academic disciplines are largely ignorant of
what other disciplines are all about. Given the size of the literature
in any one domain, this is understandable. The recent fusion of
philosophy, computer science, cognitive psychology and linguistics is an
attempt to address this problem, but it often fails. If you have a look
at the following (and the Rescorla-Wagner model really is just a grain
of sand here), you should get some idea of how vast this is and how it
is being implemented, and has been for over the past half a century or
so.
Va=alpha.beta(lambda-Vx)
(see [3] and its developments)
[1] Quine W.V.O (1951) "Two Dogmas of Empiricism" in, "From a
Logical Point of View", Harvard Press
1961
[2] Quine W.V.O (1969) "Epistemology Naturalized"
in "Ontological Relativity and Other
Essays" Columbia Press
[3] Rescorla R & Wagner A (1972) "A Theory of Pavlovian Conditioning:
variations in the effectiveness of
reinforcement and non-reinforcement.
in Clasical Conditioning II: Current
Theory and Research (A Black & W
Prokasy Eds) pp 64-99 Appleton
Century Crofts, NY
[4] Rescorla R A (1988) "Pavlovian Conditioning: It's not you
think it is" American Psychologist,
43, 151-160.
--
David Longley
Eray Ozkural exa
> In brief my philosophy of mind is a fallibilistic mind that has ideas
> (as programs) creating other ideas creating new ways to create ideas,
> cooperating with other ideas, competing with others all to try and
> achieve a goal.
Whilst a noble goal, metaphysics can help you only this much perhaps,
ie. in terms of formulating your problem. No work on epistemology
gives insight to the mechanisms required for giving birth to concepts,
let alone describe what a concept is computationally.
My suggestion is to approach the problem of knowledge from a
scientific point of view and leave behind as much metaphysical luggage
as possible. As you well know, we have an information theory in
science. Unfortunately, we don't really have a theory of knowledge.
(And I'm beginning to think we might not have one) The closest thing
that comes to a theory of knowledge is computational ontology,
specification of a system of "concepts" (although ontology is normally
concerned with what exists). Those works derive directly from ontology
in metaphysics, however you will see that they are severely limited in
their
a) means of description: mostly logical, opaque, inflexible
b) view of the world: object-based, simplistic, too high-level,
possibly not-grounded
Furthermore as evidenced by some people around here, a
misunderstanding of ontology/epistemology or subscribing to a
particular theory can hamper one's ideas about intelligence. (I will
not elaborate on that)
A small observation: you can't just "implement" a philosophical paper.
;)
Now about your idea that "we obviously cannot start with a tabula rasa
with no concepts", I don't know where that comes from. This claim
seems to depend on the use of concept itself which is vague enough.
You can probably enlarge it to contain anything on an AI's mind, and
yes, truly an AI will have to have a mind to learn new concepts.
However, I must point out that the "no tabula rasa" argument is flawed
from a computational point of view. Obviously unsupervised learning
methods *can* discover new concepts on their own starting with no
concepts. (But if you count the processes that generate concepts
concepts themselves, you would be right)
In summary, I don't think philosophical ontology and epistemology
research can aid much in the way of building minds, but it's nice to
know about them.
Let me say something slightly useful about your research. It is well
possible that a strong AI can be created by a genetic programming
experiment. Computational view of mind does not exclude any particular
approach to AI programming.
Good luck in your research,
Regards,
__
Eray Ozkural <er...@cs.bilkent.edu.tr>
CS Dept., Bilkent University
Acme Debugging
Depends on whether it's in a pile or in your face.
Larry
Acme Debugging
<snip>
>
> My current research area is an ALife like system with many programs in
> trying to overwrite each other, and being rewarded for performing a
> task. The programs are turing complete so could calculate any
> function. I am currently interested in discovering how to create
> programs that change there offspring in a reasonable fashion so that
> different ideas can be found in the system.
>
<snip>
>
> Will Pearson
Hi Will. This above part seems interesting and to stand by itself
without the philosopy. Was that to be useful for introducing
creativity into the offspring process? I wonder what your ideas are to
accomplish that.
Larry
Trewth Seeker
> the mind and whatever that makes of the environment.
You start out wrong and go downhill from there. Look into basic
evolutionary biology and understand the difference between
genotype and phenotype.
> In brief my philosophy of mind is a fallibilistic mind that has ideas
> (as programs) creating other ideas creating new ways to create ideas,
> cooperating with other ideas, competing with others all to try and
> achieve a goal.
Yeah, so? Why are you telling us?
Aatu Koskensilta
> I am interested in making AI, primarily to make computers easier to
> use. So it will be useful for them to be able to understand humans and
> their irrationality (I don't want my computer to not be able to answer
> questions about fiction for example).
How is fiction irrational?
--
Aatu Koskensilta (aatu.kos...@xortec.fi)
"Wovon man nicht sprechen kann, daruber muss man schweigen"
- Ludwig Wittgenstein, Tractatus Logico-Philosophicus
David Longley
<aatu.kos...@xortec.fi> writes
>Will Pearson wrote:
>
>> I am interested in making AI, primarily to make computers easier to
>> use. So it will be useful for them to be able to understand humans and
>> their irrationality (I don't want my computer to not be able to answer
>> questions about fiction for example).
>
>How is fiction irrational?
>
See threads in intension, referential opacity and pretending etc.
--
David Longley
Will Pearson
Hi Larry
I try and and introduce things depending upon what I am thinking about
and what I think my audience is. At the moment I am interested in
epistemology and there seemed to be some discussion going on here.
To be truthful, I don't quite know how to do the creativity. Too
"creative" and the programs destroy their offspring and get bred out
of the system by those safer programs that are not creative. My
current thinking is to try and store some meta information about the
programs within a data section, indicating information about which
sections are safe to alter. This meta-information might also
information about for loops and the like. It could also be written to
by the programs and the way it interperated change over time as the
program that reads it change.
I think that I will get hung up on the "too creative" programs for a
while as the programs at the moment are very basic.
Will
Will Pearson
> concerned with what exists). Those works derive directly from ontology
> in metaphysics, however you will see that they are severely limited in
> their
> a) means of description: mostly logical, opaque, inflexible
> b) view of the world: object-based, simplistic, too high-level,
> possibly not-grounded
>
> Furthermore as evidenced by some people around here, a
> misunderstanding of ontology/epistemology or subscribing to a
> particular theory can hamper one's ideas about intelligence. (I will
> not elaborate on that)
To me a concept has to be a program of sorts. For example logical
deduction I hope you will agree is a concept. Yet it can also be used
as a program (or series of instructions) to create new concepts. And
since a concept is a program, the creation of a concept is a form of
programming. So yes a tabula rasa from my perspective is a Strawman
as it would have no programs that could create other programs.
> A small observation: you can't just "implement" a philosophical paper.
> ;)
True it will need lots of work. But without the correct philosophical
grounding, you may head off in the wrong direction.
> Let me say something slightly useful about your research. It is well
> possible that a strong AI can be created by a genetic programming
> experiment. Computational view of mind does not exclude any particular
> approach to AI programming.
Do you have any other philosophical ideas about AI, apart from a
computational view, I would be interested to know. Also I am not sure
my system would be properly described as GP. Tierra, the sort of
system that mine is similar to, never got that epithet and with GP
there is certain set things that people expect such as fixed
generations, no interaction between programs and a closely controlled
selection process. However there seems to be no other name apart from
Tierra-like system, and that was based on a completly different
philosophy, that of trying to make the individual programs
intelligent, whereas I am trying to make the system as a whole
Will Pearson
Will Pearson
> whaz...@yahoo.com (Will Pearson) wrote in message news:<623a9073.03071...@posting.google.com>...
> > Now the mind is roughly made of two things. Whatever genetics gives
> > the mind and whatever that makes of the environment.
>
> You start out wrong and go downhill from there. Look into basic
> evolutionary biology and understand the difference between
> genotype and phenotype.
So I was ignoring the interaction between the environment and
phenotype in its development. So? Still, my basic point remains, the
information that determines the mind comes from two sources the
genotype and the environment that it develops in.
> > In brief my philosophy of mind is a fallibilistic mind that has ideas
> > (as programs) creating other ideas creating new ways to create ideas,
> > cooperating with other ideas, competing with others all to try and
> > achieve a goal.
>
> Yeah, so? Why are you telling us?
Philosophy? Ye know debate, conversation, discussion. What is your
philosophy of the brain?
Acme Debugging
This part about storing meta info in a data section sounds similar to
a part of Curt Welch's project where he is building an "interpreter"
to create algorithms. Perhaps you might mention this to him in one of
his threads.
>
> I think that I will get hung up on the "too creative" programs for a
> while as the programs at the moment are very basic.
>
> Will
It sounds like an interesting project.
Larry
Trewth Seeker
> trewth...@yahoo.com (Trewth Seeker) wrote in message news:<d690a633.03071...@posting.google.com>...
> > whaz...@yahoo.com (Will Pearson) wrote in message news:<623a9073.03071...@posting.google.com>...
> > > Now the mind is roughly made of two things. Whatever genetics gives
> > > the mind and whatever that makes of the environment.
> >
> > You start out wrong and go downhill from there. Look into basic
> > evolutionary biology and understand the difference between
> > genotype and phenotype.
>
> So I was ignoring the interaction between the environment and
> phenotype in its development. So? Still, my basic point remains, the
> information that determines the mind comes from two sources the
> genotype and the environment that it develops in.
That's a "point"? If you want to say something beyond pointless
tautologies, you might at least try to be accurate. Surely the
acutral ontogeny of the mind is relevant to understanding it,
of understanding what it is "made of".
> > > In brief my philosophy of mind is a fallibilistic mind that has ideas
> > > (as programs) creating other ideas creating new ways to create ideas,
> > > cooperating with other ideas, competing with others all to try and
> > > achieve a goal.
> >
> > Yeah, so? Why are you telling us?
>
> Philosophy? Ye know debate, conversation, discussion.
Why would anyone want to waste their time debating *your* "philosophy",
especially when it is as unoriginally vague as the above?
Ever heard of memes?
> What is your
> philosophy of the brain?
Zen.
Trewth Seeker
Non sequitur. Dealing in fiction (counterfactuals) is in fact
a distinguishing (from "lower" animals) feature of human rationality.
Will Pearson
> whaz...@yahoo.com (Will Pearson) wrote in message
> >
> > So I was ignoring the interaction between the environment and
> > phenotype in its development. So? Still, my basic point remains, the
> > information that determines the mind comes from two sources the
> > genotype and the environment that it develops in.
>
> That's a "point"? If you want to say something beyond pointless
> tautologies, you might at least try to be accurate. Surely the
> acutral ontogeny of the mind is relevant to understanding it,
> of understanding what it is "made of".
It one that is often ignored, especially in computer science, people
tend to program their computers with information that obviously the
brain is not born with in an attempt to make the computer intelligent.
I do try to take account of the growing of the brain into my theories
(or what little I know of it), for example the fact that sections of
the brain that are normally visual can be used for other processing[1]
if the eyes are not sending information suggests to me either a
centralised control (which I do not like for structural reasons and it
does not often happen in biology) or competitive sections that can
take over unused areas.
> > > > In brief my philosophy of mind is a fallibilistic mind that has ideas
> > > > (as programs) creating other ideas creating new ways to create ideas,
> > > > cooperating with other ideas, competing with others all to try and
> > > > achieve a goal.
> > >
> > > Yeah, so? Why are you telling us?
> >
> > Philosophy? Ye know debate, conversation, discussion.
>
> Why would anyone want to waste their time debating *your* "philosophy",
> especially when it is as unoriginally vague as the above?
> Ever heard of memes?
Yep, but memes are mainly used to when discussing replication between
brains, not within brains, dawkins was specific on this point in the
Selfish Gene, he believed that there was something else to human
intelligence. Also there is AFAIK no theory about how they are mutated
or created apart from trying to fit it into a darwinistic theory,
which is a little bit different to my description as things are only
darwinistic at the lowest layer and neo-lamarckian at higher levels.
It was vague because I didn't think people would appreciate me going
into details about the permissions each program can set on their
memory, the different reward schemes that could be used to make sure
the good programs to survive, how the utility is encoded into the
program and sundry other details. It is also vague, because I am not
trying to match human intelligence but I want to make something to
make computers program themselves and have backup slightly different
programs if one of them crashed. But the general ideas would be the
same, if I was trying to replicate humans, but I would need to think
more about the utility function actually was, rather than relying on
humans to choose the correct behaviour.
> > What is your
> > philosophy of the brain?
>
> Zen.
How enlightening. It is hard to debate a single word. Does Zen
philosophy take into consideration how the brain develops, if so I
have not seen any of it.
David Longley
Empiricism".
The following is an online link to the paper itself.
http://www.ditext.com/quine/quine.html
A second paper, "Epistemology Naturalized" can be found in Quine's
"Ontological Relativity & Other Essays" 1969, Columbia Press.
As I have said elsewhere, the traditional, (Romantic) objective of AI
seem something of a paradox, or worse still, a muddle. If a) the model
is to be living creatures capable of learning, this would mean that any
systems created in their likeness would be prone to all of the errors
which animals are prone to. If on the other hand, the model is b) that
of best practice engineering drawing on science, then "AI" applications
would seem to be co-extensive with almost all human technological
developments. Furthermore, given the nature of science and engineering,
the products would only seem "intelligent" to the extent that the
observers were ignorant of how the systems worked.
Perhaps, as has been said by many others before, this just highlights
the problems we encounter when moving from Natural Language to Science
and technology. In their own place and for their limited everyday
purposes, maybe the terms of folk psychology suffice. They just don't
travel well outside that domain.
This is why some urge us to begin with an Experimental Analysis of
Behavior.
--
David Longley
Bill Modlin
> As I have said elsewhere, the traditional, (Romantic) objective of AI
> seem something of a paradox, or worse still, a muddle. If a) the model
> is to be living creatures capable of learning, this would mean that any
> systems created in their likeness would be prone to all of the errors
> which animals are prone to. If on the other hand, the model is b) that
> of best practice engineering drawing on science, then "AI" applications
> would seem to be co-extensive with almost all human technological
> developments. Furthermore, given the nature of science and engineering,
> the products would only seem "intelligent" to the extent that the
> observers were ignorant of how the systems worked.
I concur with the objective analysis almost hidden in the rhetoric above,
and applaud Longley for his recognition (made explicit in other posts) that
the essence of traditional AI (GOFAI) is equivalent to the (b) alternative.
However, Longley seems to think that the error-proneness of animal learning
processes renders them unworthy of serious consideration, that there is no
point in an AI based on model (a). I disagree with this subjective
judgment, and claim that Longley simply misses the point. There is no
paradox, and while there may be a muddle it is due to incomplete
understanding of the (a) model and its role in the functioning of
intelligence.
Errors are unavoidable when dealing with real-world situations. Information
is never complete, and infallibility is a fools goal.
Logic, no matter how carefully applied, cannot yield a useful result when
applied to irrelevant premises, and it is precisely those fallible
heuristics which Longley dismisses with such disdain which allow us to
select relevant premises from the welter of possibilities presented by
ongoing experience.
Formal statistical tools are demonstrably superior to heuristic judgment, if
and only if heuristic judgment has first selected appropriate inputs to
those tools.
Type (b) AI achieves precision at the expense of generality, type (a)
achieves generality at the expense of precision.
Type (b) AI is domain-specific and brittle. We have many working examples
of this.
Type (a) AI, if we achieve it, will be more general in application, more
ductile and adaptable. So far I don't know of any good working examples of
this kind of AI, so we have to keep working on it. I have my own notions of
how best to proceed, others have different ideas. Time will tell.
However, it seems clear that if we succeed in producing type (a) AI we will
have a machine which can learn to use scientific and engineering procedures
where they are available and appropriate, while still having its primary
heuristic intelligence as a guide for the many situations in they are not.
Type (a) subsumes type (b).
There is no equivalent symmetrical hope for type (b): no finite collection
of formal procedures can substitute for the generality of heuristic
intelligence.
Type (a) AI is intelligent.
Type (b) AI is a collection of tools which intelligence may learn to use.
Bill Modlin
P.S.
Several years ago I posted a rather long and rambling note to this
newsgroup, sketching my position on the nature and definition of
intelligence. I recently came across a copy of it, and uploaded it to the
free website space made available by my ISP. It is at:
http://www.metrocast.net/~modlin1
Today I would express similar thoughts somewhat differently, and I've come
to realize that the problem of creating an effective type (a) AI is more
complex than I had in mind at the time. However the essence of my
philosophy on these matters has not changed and I would welcome any comment
on that earlier paper.
Neil W Rickert
>Several years ago I posted a rather long and rambling note to this
>newsgroup, sketching my position on the nature and definition of
>intelligence. I recently came across a copy of it, and uploaded it to the
>free website space made available by my ISP. It is at:
>http://www.metrocast.net/~modlin1
>Today I would express similar thoughts somewhat differently, and I've come
>to realize that the problem of creating an effective type (a) AI is more
>complex than I had in mind at the time. However the essence of my
>philosophy on these matters has not changed and I would welcome any comment
>on that earlier paper.
Thanks for the link.
At one time, I held a view very similar to what you present.
However, I have come to recognize that view as faulty.
Your view is that the world is a highly patterned place, and intelligence
involves the discovery of these patterns.
What I have concluded, is that the world is a disorderly place
without clear patterns or recurrent events.
I certainly grant that the world appears to be a patterned, orderer
place. But in some sense, that is an illusion. It is not that the
world is patterned. Rather, it is that what we see of the world is
patterned.
The heart of intelligence is in perception. It is our perceptual
systems that order the world as we see it. In effect, our perceptual
systems impose some sort of organization on the world as we see it.
The patterns then become apparent, relative to the system of organization.
To achieve AI, you would need a system that could organize the world
in the process of developing a perceptual system. It would have to
develop organization systems that work, in the sense of fitting well
with the actual world. An arbitrarily imposed system of organization
would probably work quite poorly. Thus there is an element of trial
and error experimentation, in developing suitable systems of
organization, and there is an element of pragmatic judgement involved
in evaluating the outcomes of this experimentation.
Incidently, it is also my conclusion that computationalism cannot
work. For it can only make truth-based judgements. It cannot make
the needed pragmatic judgements. Biological systems, by contrast, are
able to make pragmatic judgements. Evolution itself is a
pragmatically based system.
Curt Welch
"Bill Modlin" <mod...@metrocast.net> wrote:
> So intelligence is the ability to find patterns.
Bill and I have had this discussion in e-mail, but I thought I'd put a few
ideas out here to see where it might lead.
The single line above from the paper should be its subject. The rest of
the paper simply supports this claim.
I agree with almost everything Bill says in the paper, and the approach is
very consistent with my own, however, I think if falls short of defining
"full AI".
A machine which spends its life finding patterns would not be considered by
many to be intelligent. It has to do something with that knowledge to be
intelligent. This issue introduces what I consider to be the other half of
the problem, and the half which I see as the "real" purpose of AI.
Behavior.
Given a huge amount of knowledge about the universe from a life time of
pattern extraction, what does a machine have to do to be "intelligent"?
This is a question I started looking at long ago. If a robot has an arm,
why would it want to lift the arm instead of letting it hang by it's side?
Why would one action be "intelligent" and the other "stupid"? How does a
huge database of knowldege about the universe help the AI know if it should
lift the arm, or let it hang?
The only answer I've seen to this is that it must have a purpose. If it
has a purpose, then its ability to use the knowledge to achieve its purpose
would be intelligence. How well it performed this task would be a measure
of it's intelligence.
I see this as the other half of AI. And the more important of the two.
Because the purpose of AI is never to just acquire knowledge, it's to use
knoweldge. If you have not defined the AI's purpose, then how do you know
the knowledge the "pattern extraction" system builds is helpful to its true
purpose?
So if we need a purpose, what should that purpose be?
You could say the purpose is to be intelligent. But that self-refrential
statement does not help us define intelligence which is what we are after.
So it is pointless to say that.
You could say the purpose is to survive. That seems to be a big part of
our purpose. But how does the robot know if it should lift its arm or let
it hang? How does it know which is better for "survivial"? What can the
code in our AI brain do to answer that question? The answer is that it
can't know the answer to that without dying. That's a fine technique for
evolution, but it's not AI. AI is the ability to answer these questions
between the time you are born and die, not after you die.
You could pick a purpose like "to serve mankind". That would be nice if we
could build intelligent machines who's purpose was to server us. But how
do you build that idea into the hardware? How does the robot know if
lifting its arm is a better way to serve mankind than to let it hang?
All the machines we build has a purpose of serving us. But they know "what
to do", because we tell them. We build their behvior into them. We
specify what they do, and when, and how. And this clearly isn't AI.
The problem with most "purpose statements", is that you have to be
intelligent to understand your purpose. But if AI can't understand it's
purpose, how is it going to learn to be "intelligent" using that purpose?
We have a real boot-straping problem when we try to talk about any
high-level abstract "purpose" that requires full intelligence to
understand. It really is no better than just saying that the purpose of
the machine is to "be intelligent".
To solve this purpose problem, you must find a way to define the purpose
which a "dumb" machine can respond to, or which we can hard-wired into the
machine without requiring us to create AI first.
If you look at humans and animals with the qustion of how do they decide if
they should raise their arm or keep it lowered, you find answers to this
propose question. You find that we all learn "what to do" using the same
basic systems. It's the one behavior research has been looking at for many
years. It just the basic conditioning system at work.
We do not raise are arm "to survive" or keep it hanging "to survive". We
do it becuse we have been conditioned to do it based on our training. We
do it to get a reward, or to avoid a punishment. We do it to get food, or
prevent pain.
This points us to a simple and obvious way to give AI the purpose it needs,
in a way that the machine can deal with. We just give it a reward and
punishment system that tells it when something "bad" is happening and when
something "good" is happening.
With that simple defintion of "purpose", we can no look at what it takes to
build a machine that can decide for itself if it's better to raise it's
arm, or keep it hanging. The purpose of the hardware which we have to
hard-code into the machine is to do whatever it can to reduce the
probability of future punishment, and increase the probability of future
reward.
That gives us something we can actually build, unlike the "be intelligent"
purpose, which gives us nothing to build.
Now, when we look at the "find patterns" idea, we have something to guide
us. We aren't just looking for any pattern, we are looking for patterns
that help us achieve our purpose. To achieve the purpose of minimsing
punishment and maximize reward, there is only one thing we need to do well,
and that's to predict how much reward, or punishment, any choice we make
now is likely to create.
This of course is not a simple problem. It's like the chess game problem
of "what is the best move now?". The better you can predict the outcome
of the entire rest of the game, the better you can answer that question.
The problem of AI is not to just prevent short-term pain, but to minimize
it for the entire rest of time. So doing something like raising an arm,
which is known to create more pain that keeping it hanging is not wrong, if
the AI belives it will reduce pain in the long term.
The only tool we have for predicting the future is to study the past. And
this is where the "find patterns" ideas comes into play.
So now we not only need to "find patterns", we know why we need to find
them. We need to find them so we can best predict what actions to take
next.
So to me, intelligence is the ability to act in ways that minimise all
future punishment and maximize all future rewards.
Finding patterns by studying the past is the obvious tool it's going to use
to do this.
And to be a bit more specific, there are two obvious things the machine
will analize. The first is to predict the probability of pain/pleasure
which any action is likely to create. If raising the arm always causes
more punishment than keeping it hanging, then the obvious choice for the AI
is to keep it hanging.
But the second thing that is important is to try and find patterns that
identify what "situation" the AI is in, so that it can make predictions
based on the current situation. So maybe raising your arm in some
situation is good, and in others it is bad. So, what we need to do is
create a machine which is able to define "what" situation we are in, so the
anayisis can be based on that information.
Working on a machine which can do these things I believe is how we will
solve AI.
--
Curt Welch http://CurtWelch.Com/
cu...@kcwc.com Webmaster for http://NewsReader.Com/
Acme Debugging
> "David Longley" <Da...@longley.demon.co.uk> wrote in message
> news:7D2MiJAEylG$Ew...@longley.demon.co.uk...
<snip>
>Formal statistical tools are demonstrably superior to heuristic
judgment, if
>and only if heuristic judgment has first selected appropriate inputs
to
>those tools.
Did you intend to leave "sometimes" out?
If so then please explain, completely within the methodology of
statistics, why statistical prediction of commodity prices (i.e.
so-called "behavior" of hedgers and speculators) doesn't work when the
input (prices) is perfectly quantifiable, and the methodology
satisfies every last criteria of multiple regression with correlation
well beyond significant levels.
Due to the above example and others, I maintain that the reliability
of statistics with human personalities as the samples is unpredictable
until confirmed in real-time tests or tests in the fields as measured
by good objective tests independent of the predicting methodology;
that each such case has its own unique logic, and even then in
real-world application future tests are ultimately unpredictable
because all other things never remain equal, a well-known requirement
of statistical prediction, and one cannot predict what new or
previously overlooked logic might come to light.
Larry
Curt Welch
> "Bill Modlin" <mod...@metrocast.net> wrote in message
> news:<eemdnZkKZ6E...@metrocast.net>...
> > "David Longley" <Da...@longley.demon.co.uk> wrote in message
> > news:7D2MiJAEylG$Ew...@longley.demon.co.uk...
>
> <snip>
>
> >Formal statistical tools are demonstrably superior to heuristic
> judgment, if
> >and only if heuristic judgment has first selected appropriate inputs
> to
> >those tools.
>
> Did you intend to leave "sometimes" out?
>
> If so then please explain, completely within the methodology of
> statistics, why statistical prediction of commodity prices (i.e.
> so-called "behavior" of hedgers and speculators) doesn't work when the
> input (prices) is perfectly quantifiable, and the methodology
> satisfies every last criteria of multiple regression with correlation
> well beyond significant levels.
Larry,
Are you honestly trying to claim that heuristic judgement outperforms
statistical tools for market prediction when the judgment is based only on
the same quantifiable input data?
Neil W Rickert
>L.F...@lycos.co.uk (Acme Debugging) wrote:
>> "Bill Modlin" <mod...@metrocast.net> wrote in message
>> news:<eemdnZkKZ6E...@metrocast.net>...
>> >Formal statistical tools are demonstrably superior to heuristic
>> judgment, if
>> >and only if heuristic judgment has first selected appropriate inputs
>> to
>> >those tools.
>> Did you intend to leave "sometimes" out?
>> If so then please explain, completely within the methodology of
>> statistics, why statistical prediction of commodity prices (i.e.
>> so-called "behavior" of hedgers and speculators) doesn't work when the
>> input (prices) is perfectly quantifiable, and the methodology
>> satisfies every last criteria of multiple regression with correlation
>> well beyond significant levels.
>Larry,
>Are you honestly trying to claim that heuristic judgement outperforms
>statistical tools for market prediction when the judgment is based only on
>the same quantifiable input data?
I can't speak for Larry. However, in many cases human judgement
makes use of information not available to the statistical tools.
Neil W Rickert
[big snip leading to the need for a purpose]
I agree that purpose is important. I'm not sure we have much
agreement beyond that.
>So if we need a purpose, what should that purpose be?
>You could say the purpose is to be intelligent. But that self-refrential
>statement does not help us define intelligence which is what we are after.
>So it is pointless to say that.
>You could say the purpose is to survive. That seems to be a big part of
>our purpose. But how does the robot know if it should lift its arm or let
>it hang? How does it know which is better for "survivial"? What can the
>code in our AI brain do to answer that question? The answer is that it
>can't know the answer to that without dying. That's a fine technique for
>evolution, but it's not AI. AI is the ability to answer these questions
>between the time you are born and die, not after you die.
That seems simplistic. Choosing whether to raise an arm is rarely a
matter of life and death. I realize you are using the example as a
metaphor. But I think it is too simplistic for that role.
>You could pick a purpose like "to serve mankind". That would be nice if we
Which "mankind"? Should it server George Bush's mankind? Or Saddam
Hussein's mankind? How is it to know which is which?
>All the machines we build has a purpose of serving us. But they know "what
>to do", because we tell them. We build their behvior into them. We
>specify what they do, and when, and how. And this clearly isn't AI.
Note that many AI people think this is AI, so the claim may not
be as clear as you think. (Still, I agree with your assessment that
it is not the route to AI).
>The problem with most "purpose statements", is that you have to be
>intelligent to understand your purpose. But if AI can't understand it's
>purpose, how is it going to learn to be "intelligent" using that purpose?
>We have a real boot-straping problem when we try to talk about any
>high-level abstract "purpose" that requires full intelligence to
>understand. It really is no better than just saying that the purpose of
>the machine is to "be intelligent".
>To solve this purpose problem, you must find a way to define the purpose
>which a "dumb" machine can respond to, or which we can hard-wired into the
>machine without requiring us to create AI first.
>If you look at humans and animals with the qustion of how do they decide if
>they should raise their arm or keep it lowered, you find answers to this
>propose question. You find that we all learn "what to do" using the same
>basic systems. It's the one behavior research has been looking at for many
>years. It just the basic conditioning system at work.
I would suggest that you are overrating conditioning.
What was it that conditioned Beethoven to write his symphonies,
instead of just replaying symphonies that already were available and
were known to be audience pleasers? What was it that conditioned
Darwin to decide on his theory of evolution? What was it that
conditioned Einstein to decide on his relativity theories? Or what
conditioned von Neumann to come up with his design for the computer?
The same question could be asked in countless other ways.
Conditioning is too simple an explanation to account for the
diversity and creativity of human activity.
>We do not raise are arm "to survive" or keep it hanging "to survive". We
>do it becuse we have been conditioned to do it based on our training. We
>do it to get a reward, or to avoid a punishment. We do it to get food, or
>prevent pain.
How do you account for acts of altruism?
>This points us to a simple and obvious way to give AI the purpose it needs,
>in a way that the machine can deal with. We just give it a reward and
>punishment system that tells it when something "bad" is happening and when
>something "good" is happening.
When designing the AI appliance, how will you figure out what to
count as good and what to count as bad in your programming?
And how would that be different from what you were criticizing
above when you wrote:
But they know "what to do", because we tell them. We build
their behvior into them. We specify what they do, and when,
and how. And this clearly isn't AI.
>With that simple defintion of "purpose", we can no look at what it takes to
>build a machine that can decide for itself if it's better to raise it's
>arm, or keep it hanging. The purpose of the hardware which we have to
>hard-code into the machine is to do whatever it can to reduce the
>probability of future punishment, and increase the probability of future
>reward.
But then you have to design the machine so that it can estimate these
probabilities of future reward and punishment. It seems to me that
this would require a lot more than conditioning.
>That gives us something we can actually build, unlike the "be intelligent"
>purpose, which gives us nothing to build.
>Now, when we look at the "find patterns" idea, we have something to guide
>us. We aren't just looking for any pattern, we are looking for patterns
>that help us achieve our purpose. To achieve the purpose of minimsing
>punishment and maximize reward, there is only one thing we need to do well,
>and that's to predict how much reward, or punishment, any choice we make
>now is likely to create.
>This of course is not a simple problem. It's like the chess game problem
>of "what is the best move now?". The better you can predict the outcome
>of the entire rest of the game, the better you can answer that question.
In the case of chess, prediction is based on the rules of the game.
If you want the analogy to be relevant, it would seem that you would
have to preprogram your AI system with the rules of the game of
minimizing punishment and maximizing reward. In other words, you
would have to pre-program intensive knowledge. Such a view is
consistent with the No Free Lunch theorems, but it seems inconsistent
with your goal of a learning system.
>The problem of AI is not to just prevent short-term pain, but to minimize
>it for the entire rest of time. So doing something like raising an arm,
>which is known to create more pain that keeping it hanging is not wrong, if
>the AI belives it will reduce pain in the long term.
How will the AI come to believe that?
>The only tool we have for predicting the future is to study the past. And
>this is where the "find patterns" ideas comes into play.
The past is a poor predictor of the future. From a past of computers
too big and too expensive for anything other than large business and
research projects, how would one predict today's world? It would
seem that an AI that underwent conditioning in 1950 would have been
useless by 1970 and even more useless today. Yet there are plenty of
people around in 1950 who are still doing useful things today.
>So now we not only need to "find patterns", we know why we need to find
>them. We need to find them so we can best predict what actions to take
>next.
>So to me, intelligence is the ability to act in ways that minimise all
>future punishment and maximize all future rewards.
By the way -- it isn't quite the same thing -- but you might find it
useful to read William T. Powers, "Behavior: the control of
perception".
>Finding patterns by studying the past is the obvious tool it's going to use
>to do this.
>And to be a bit more specific, there are two obvious things the machine
>will analize. The first is to predict the probability of pain/pleasure
>which any action is likely to create. If raising the arm always causes
>more punishment than keeping it hanging, then the obvious choice for the AI
>is to keep it hanging.
>But the second thing that is important is to try and find patterns that
>identify what "situation" the AI is in, so that it can make predictions
>based on the current situation. So maybe raising your arm in some
>situation is good, and in others it is bad. So, what we need to do is
>create a machine which is able to define "what" situation we are in, so the
>anayisis can be based on that information.
Well now you may be getting somewhere. But you will need a clear
idea of what is to count as a "situation" before you can make much
progress with this.
Acme Debugging
> "Bill Modlin" <mod...@metrocast.net> writes:
>
<snip>
>>
>>Incidently, it is also my conclusion that computationalism cannot
>>work. For it can only make truth-based judgements. It cannot make
>>the needed pragmatic judgements. Biological systems, by contrast,
are
>>able to make pragmatic judgements. Evolution itself is a
>>pragmatically based system.
I agree it does seem impossible, but has it been proven that the
pragmatic judgements cannot be generalized? Do you have any logic
besides intuition or common sense to convince you that it could not be
generalized sufficiently to serve wide-application, if not
universally?
Larry
Neil W Rickert
>Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bff1mo$q4j$1...@husk.cso.niu.edu>...
><snip>
>>>Incidently, it is also my conclusion that computationalism cannot
>>>work. For it can only make truth-based judgements. It cannot make
>>>the needed pragmatic judgements. Biological systems, by contrast,
>are
>>>able to make pragmatic judgements. Evolution itself is a
>>>pragmatically based system.
>I agree it does seem impossible, but has it been proven that the
>pragmatic judgements cannot be generalized? Do you have any logic
>besides intuition or common sense to convince you that it could not be
>generalized sufficiently to serve wide-application, if not
>universally?
Logic? No. These questions are outside of logic. Logic is good for
deriving conclusions from premises. But it can never get anything
that is outside the system. That's its limitation. But then most
decisions we make in life are made outside of logic. For the
constraints of logic are too severe.
The Sherlock Holmes stories are sometimes given as examples of
logic. They are not. Sherlock Holmes actually had to go out into
the field and look for those cigarette butts (or whatever else was
the needed evidence). Going out and collecting new evidence is not
part of logic.
Sure, you can describe it as if it were logic. The mistake often
made is to assume that if you can describe it as if it were logic,
then it really was logic.
I was recently on a jury in a local court case. When instructing the
jury, the judge was very clear that we were not expected to make a
logical decision based only on the testimony given. Rather, we were
expected to call upon our life experiences in reaching our
conclusions. (In fact, that's the whole point of a jury of one's
peers).
The intelligent human has access to vastly more information than is
available within the limited input channels of a computer. For the
important decisions of every day life, logic applied to the defined
input usually cannot match the ability of the human, made possible by
the broad array of additional information that is available.
No, I cannot give a logical proof of the above. It is one of the
limitations of logic that no such proof is possible.
----------
Addendum.
It is often argued that a pragmatic decision is really a truth based
decision in disguise. X was a pragmatically good decision means "It
was true that X was a pragmatically good decision". But if an AI
system is to use truth criteria (i.e. logic) to make pragmatic
decisions, then it must be preprogrammed with comprehensive rules as
to which decisions are to count as pragmatically good. This amount
to the same problem that TS raised about Curt's proposals.
My current view is that intelligence is all about pragmatic decision
making, as distinct from truth based decision making.
Bill Modlin
>
> >Several years ago I posted a rather long and rambling note to this
> >newsgroup, sketching my position on the nature and definition of
> >intelligence. I recently came across a copy of it, and uploaded it to
the
> >free website space made available by my ISP. It is at:
>
> >http://www.metrocast.net/~modlin1
>
> >Today I would express similar thoughts somewhat differently, and I've
come
> >to realize that the problem of creating an effective type (a) AI is more
> >complex than I had in mind at the time. However the essence of my
> >philosophy on these matters has not changed and I would welcome any
comment
> >on that earlier paper.
>
> Thanks for the link.
You're welcome, thanks for reading it... Hi Neil :-)
> At one time, I held a view very similar to what you present.
> However, I have come to recognize that view as faulty.
>
> Your view is that the world is a highly patterned place, and intelligence
> involves the discovery of these patterns.
>
> What I have concluded, is that the world is a disorderly place
> without clear patterns or recurrent events.
Certainly there is plenty of real or apparant disorder to be found... even
ignoring quantum indeterminancy, if we measure any quantity precisely enough
we find it varying randomly by small amounts from measurement to
measurement, a result of what has come to be called thermal noise.
But if we ignore variations below the noise level we can often get
measurements that are reasonably well behaved and non-random in quite
definite ways.
For example, we often find that measurements exhibit autocorrelation such
that future values can be predicted to some degree as a function of recent
past values of the same measurement at the same place.
For another example, we often find that measurements exhibit correlation
with other measurements taken from different places nearby.
So far as recurrent events are concerned, we get to define what we wish to
treat as a recurrence. At this lowest level of individual measurements, it
is reasonable to define recurrent events as repetitions of a specified range
of values in a series of samples, or (more common in biological sensory
measurements) the reaching of a threshold value in the integral of a
continously monitored quantity. It's true that nothing ever repeats exactly
in the real world, so we have to ignore some differences to treat a category
of things as "the same". To get started, we have to impose some rather
arbitrary categorization at the sensory level, basically just ignoring
differences that are small in the measurement metric. As we learn more we
will choose categories that improve the correlations we find... but for now
just pick something.
> I certainly grant that the world appears to be a patterned, orderer
> place. But in some sense, that is an illusion. It is not that the
> world is patterned. Rather, it is that what we see of the world is
> patterned.
>
> The heart of intelligence is in perception. It is our perceptual
> systems that order the world as we see it. In effect, our perceptual
> systems impose some sort of organization on the world as we see it.
> The patterns then become apparent, relative to the system of organization.
We do not and cannot usefully impose arbitrary patterns on data. No matter
how I choose to organize random data it will remain random: I can discover a
recurring predictable pattern only to the extent that it exists in the data.
In those primitive notions of patterning of simple measurements that I
mentioned about, there is no way to find a pattern in thermal noise. That's
why we design our sensing/measuring systems to ignore noise-level
variations. If we ignore this principle and measure the noise, there will
be no additional correlations to detect, no patterns beyond those detectable
in the signal quantitized at a level which ignores the noise.
You may of course try to force a pattern. Given a training set of many
sequential samples from some source you could for example assert that there
is an alternation between two sets of "patterns" of a dozen samples in a
row. You could start tabulating the patterns in each set... put the first
dozen samples as an example pattern in the first set, the second dozen in
the second set, the third goes back in the first set, and so on.
Testing over the same sample will show a perfect pattern: sure enough, the
dozen-sample-chunks alternate between the two sets perfectly, because that's
the way you listed them.
But it won't work on any new data from the same source, unless there really
was such a pattern inherent in the data. This is called "overfitting", and
is a waste of time. You cannot impose a pattern that is not there and
expect it to be of any use.
It's fine if you can find some non-imposed criteria for grouping chunks of
samples, and possible that it will turn out that your groupings alternate in
the data. But you have to find rather than impose the groupings of
combinations, or you have no reason to expect the pattern to extend beyond
the data on which it was imposed.
If you have previously discovered that such a pattern of alternation existed
for a time in the data, you may set up your perceptual system to look for
it, and notice it quickly when it arises again rather than having to
discover it from scratch. Once you have learned to see this pattern, you
can see a single chunk that you recognize as a member of one of these sets
and perhaps expect to see a chunk from the alternating set next, depending
on how often you see this kind of chunk outside the alternating pattern.
Perceptual systems can and do organize the incoming data so that we can see
the patterns we are prepared to see. The organization can prevent us from
seeing other patterns that we might see under a different organization. But
it cannot impose patterns that are not there, it can only allow us to see
patterns that are actually present. It is incumbent on the mechanisms which
guide development of our perceptual systems to provide organizations
corresponding to the patterns in the data: we don't get to invent them.
> To achieve AI, you would need a system that could organize the world
> in the process of developing a perceptual system. It would have to
> develop organization systems that work, in the sense of fitting well
> with the actual world. An arbitrarily imposed system of organization
> would probably work quite poorly. Thus there is an element of trial
> and error experimentation, in developing suitable systems of
> organization, and there is an element of pragmatic judgement involved
> in evaluating the outcomes of this experimentation.
An arbitrarily imposed system would be more likely not to work at all than
to work poorly.
Trial and error... yes, there is some of that, when you have no clues to
guide you. It's a valid fallback strategy to get you out of a dead end,
and given enough time it might even find a solution. After all, we assume
evolution came up with us as a result of a long series of trial and error
experiments. But I would not call it a particularly intelligent strategy.
To me intelligence is a matter of being able to find patterns much faster
than you could stumble on them by trial and error, a matter of heuristically
guided search.
In the end we are left with the same conflict with which we started: you
see patterns as imposed, I see them as discovered. I know they cannot be
imposed, you think you know they cannot be readily discovered. :-)
It probably isn't profitable to debate the point further, since it is
unlikely that either of us will be convinced to change.
But if you do want to continue, could you address the simple case of a
single signal source (continuous or discrete, doesn't matter), which may or
may not be "patterned"... which is to say, somehow predictable? Can you
describe a way in which you could impose a pattern on this signal without
discovering a pattern that actually exists?
For my side of it, the signal could be patterned in any number of ways
without my being able to discover it. I'm not claiming that I can discover
any arbitrary pattern. I think I know useful ways to search out the kinds
of patterns that are evident to humans, which is enough to build a
human-style intelligence. But I can't for the life of me see how you or
anyone can force that signal to be patterned if it was not that way from the
start.
??
Bill Modlin
> Incidently, it is also my conclusion that computationalism cannot
> work. For it can only make truth-based judgements. It cannot make
> the needed pragmatic judgements. Biological systems, by contrast, are
> able to make pragmatic judgements. Evolution itself is a
> pragmatically based system.
Oops. Forgot about this paragraph. I have no idea what you mean here... I
don't know what a "truth-based" judgment is, or how any system computational
or otherwise could know what is really "true" or "false" except in the
pragmatic sense that some things work better than others. Is my
hypothetical computer program that implements AI
"computational"? What defines "computationalism"?
Acme Debugging
> L.F...@lycos.co.uk (Acme Debugging) wrote:
> > "Bill Modlin" <mod...@metrocast.net> wrote in message
> > news:<eemdnZkKZ6E...@metrocast.net>...
>> <snip>
>>
>> >Formal statistical tools are demonstrably superior to heuristic
>> judgment, if
>> >and only if heuristic judgment has first selected appropriate
inputs
>> to
>> >those tools.
>>
>> Did you intend to leave "sometimes" out?
>>
>> If so then please explain, completely within the methodology of
>> statistics, why statistical prediction of commodity prices (i.e.
>> so-called "behavior" of hedgers and speculators) doesn't work when
the
>> input (prices) is perfectly quantifiable, and the methodology
>> satisfies every last criteria of multiple regression with
correlation
>> well beyond significant levels.
>
>Larry,
>
>Are you honestly trying to claim that heuristic judgement outperforms
>statistical tools for market prediction when the judgment is based
only on
>the same quantifiable input data?
No, but I can infer that and also claim that it is trivial to the
example. At any given time, enough traders will be using statistical
prediction techniques to fully discount and randomize the results, in
fact due to the way the commodities market works, randomized almost
right up to the moment of a trade. I can then infer that *the best*
human judgment will always beat the flip of a coin even if it is only
eyeballing trend lines, etc. If anyone disagrees, fine, I don't need
it, it's specific to the example, it's not the point.
But (and I know you didn't ask this) I judiciously did not limit
judgement to prices in the example, only for the statistics, and
that's the way it always seems when human personalities are the
samples. The prices are the "perfect" aggregate measurement of
"behavior" (and where have I heard about aggregate measurements
before?). The example holds because there is a world of data
available to human judgment that cannot be sufficiently quantified,
for instances news reports of an impending strike in a gold-producing
country, insider trading, Micheal Jackson concerts, trivial Miss World
asides that cause mass rioting and death, and just about anything
else you can think of that affects the so-called behavior of traders
or hedgers, including some well-known examples of market hysteria.
Successful traders however become very good at interpreting these
factors in real-time (by definition of course).
Note that this "logic" such as "discounting" behind the statistics is
real-world, it is outside of the methodology, and is unique to the
case. Also that when human personalities are the samples, additional
unquantifiable data is (in every case I can think of) always available
to human judgment (but one case satisfies "sometimes" anyway). Samples
aside, it is also theory intended to be generalized into the
real-world which is another unsolved problem discussed in this group.
I know this doesn't mean that your project will fail on these points
because we can't know if it will develop informed wisdom aka heuristic
judgement on it's own, and besides I've learned not to say what a
programmer cannot do (except to motivate the hell out of them <g>). It
will be interesting to find out. It is also interesting that, the
first AI researcher to be able to systematically quantify news
reports, insider trading, etc. affecting commodity prices so that this
can't be discounted by other traders may get rich rather quickly.
It's much easier just to say that in real-life all the marbles are not
the same. (One of my "dogmatic" messages <g>)
Thanks for the question. I don't know anything for sure.
Larry
Bill Modlin
"Acme Debugging" <L.F...@lycos.co.uk> wrote in message
news:35fae540.03072...@posting.google.com...
Lol... In my mind the word "appropriate" carried the burden for
"sometimes"... in the case of commodity price prediction, so far nobody has
come up with an appropriate set of quantifiable inputs such that regression
produces a particularly useful output, if indeed such a set exists.
Remember the context for the statement. I'm responding to Longley, who has
frequently posted examples of cases in which regression works better than
human judgment, and argues that therefore regression is a desirable
substitute for human intelligence. I wish to cede that there are such
cases, but dispute the "therefore".
I'm not pushing the reliability of regression, just saying its purported
reliability does not make it intelligent.
I probably should have included "sometimes" anyway. :-)
Any comment on the actual issues discussed in the post?
Bill
Neil W Rickert
>"Neil W Rickert" <ricke...@cs.niu.edu> wrote in message
>> Thanks for the link.
>You're welcome, thanks for reading it... Hi Neil :-)
Note that I am snipping heavily, since you have indicated your
disdain for this kind of discussion.
>> At one time, I held a view very similar to what you present.
>> However, I have come to recognize that view as faulty.
>> Your view is that the world is a highly patterned place, and intelligence
>> involves the discovery of these patterns.
>> What I have concluded, is that the world is a disorderly place
>> without clear patterns or recurrent events.
>Certainly there is plenty of real or apparant disorder to be found... even
>ignoring quantum indeterminancy, if we measure any quantity precisely enough
>we find it varying randomly by small amounts from measurement to
>measurement, a result of what has come to be called thermal noise.
>But if we ignore variations below the noise level we can often get
>measurements that are reasonably well behaved and non-random in quite
>definite ways.
You are falling for the illusion.
Measurements are not any part of the natural world. Measurements are
the application of an invention (the measuring system). The
invention of suitable measuring systems part of how we organize the
world.
>So far as recurrent events are concerned, we get to define what we wish to
>treat as a recurrence.
Quite. Events are, in effect, our invention. A recurrent event is
really two different things that we have chosen to consider a
repetition of the one event. This is part of how we organize the
world.
>We do not and cannot usefully impose arbitrary patterns on data.
I don't know why "arbitrary" keeps coming up in discussions of this
sort of topic. I said that we organize the world. I did not say
that we *disorganize* the world.
What has "arbitrary" to do with anything I suggested?
> No matter
>how I choose to organize random data it will remain random: I can discover a
>recurring predictable pattern only to the extent that it exists in the data.
Data are not a natural part of the world. They are human artifacts.
If you were to put a simple sensor somewhere -- say a light sensor,
or an air pressure sensor -- and monitor the signals it receives,
they would not be highly patterned. There may be some weak statistical
trends, but nothing strong enough that you would call a pattern.
Data results from organized ways of dealing with the world. It does
not exist without organization.
>In those primitive notions of patterning of simple measurements that I
>mentioned about, there is no way to find a pattern in thermal noise. That's
>why we design our sensing/measuring systems to ignore noise-level
>variations. If we ignore this principle and measure the noise, there will
>be no additional correlations to detect, no patterns beyond those detectable
>in the signal quantitized at a level which ignores the noise.
>You may of course try to force a pattern.
And we do. You just described one of the many ways we do it in the
above paragraph:
we design our sensing/measuring systems to ignore noise-level
variations.
> Given a training set of many
>sequential samples from some source you could for example assert that there
>is an alternation between two sets of "patterns" of a dozen samples in a
>row. You could start tabulating the patterns in each set... put the first
>dozen samples as an example pattern in the first set, the second dozen in
>the second set, the third goes back in the first set, and so on.
But I am not suggesting anything of the kind.
>> To achieve AI, you would need a system that could organize the world
>> in the process of developing a perceptual system. It would have to
>> develop organization systems that work, in the sense of fitting well
>> with the actual world. An arbitrarily imposed system of organization
>> would probably work quite poorly. Thus there is an element of trial
>> and error experimentation, in developing suitable systems of
>> organization, and there is an element of pragmatic judgement involved
>> in evaluating the outcomes of this experimentation.
>An arbitrarily imposed system would be more likely not to work at all than
>to work poorly.
>Trial and error... yes, there is some of that, when you have no clues to
>guide you.
You decide. Are you looking for a learning system, or are you
considering a system with preprogrammed innate knowledge?
I always thought you were looking for learning systems. In such
systems you are often faced with a situation with no clues to guide
you.
>In the end we are left with the same conflict with which we started: you
>see patterns as imposed, I see them as discovered. I know they cannot be
>imposed, you think you know they cannot be readily discovered. :-)
This is because you have fallen for the illusion. You seem unable to
contemplate a world with no imposed measuring systems, no imposed
ways of collecting data. You fail to recognize that all of the data
we use in science comes from methodologies imposed on the world. You
take what we have imposed on the world for granted. And therefore it
is completely invisible to you.
>It probably isn't profitable to debate the point further, since it is
>unlikely that either of us will be convinced to change.
>But if you do want to continue, could you address the simple case of a
>single signal source (continuous or discrete, doesn't matter), which may or
>may not be "patterned"... which is to say, somehow predictable? Can you
>describe a way in which you could impose a pattern on this signal without
>discovering a pattern that actually exists?
>For my side of it, the signal could be patterned in any number of ways
>without my being able to discover it.
A pattern is a property of a representation. It is not a property of
the natural world, except to the extent that the natural world
includes organisms that form representations. Any method of forming
representations depends on imposing that method on the world.
A focussed lens form more highly patterned representations on the
film than does an unfocussed lens, or a camera without any lens.
>Bill Modlin
>> Incidently, it is also my conclusion that computationalism cannot
>> work. For it can only make truth-based judgements. It cannot make
>> the needed pragmatic judgements. Biological systems, by contrast, are
>> able to make pragmatic judgements. Evolution itself is a
>> pragmatically based system.
>Oops. Forgot about this paragraph. I have no idea what you mean here... I
Let's postpone discussion, until you have read my response above.
Mr Michael Bibby
i like the idea of recursive programes that re-write over old axioms so as to
compute new functions. if you are interested in the epistemology of
system-relative representations, then i highly recommend you check out the
second order cybernetics approach to systems research. this persective argues
against Brooks notion that 'the world is its own model' and says that as far as
an 'operationally closed' systesm is concerned, its system-relative
representations *are* the environment instead of being anthropomorphically
defined enitities (Riegler calls this the pacman syndrom). in this orientation,
the job of the agent (or artifact) is to be viable in its context- where models
(representations) produce the desired outcome, then the representations are
viable and retained untile they are thwarted. this approach is based on the
radical constructivist orientation to human knowing (epistemology) which state
that "'to know' is not to possess true descriptions of reality but rather to
possess means and ways of acting that allows one to attain the goals one happens
to have chosen". if your interested in these ideas, i suggest you go to the
"radical constructivism homepage" where you will find a bunch of papers that
relate directly or inderectly to this approach to AI. perhaps the most seminal
text i have come accross, available at this site, is called "does representation
need reality" - it addressess the problem of epistemology in AI.
mickeyd
Mr Michael Bibby
>In: <a href="http://www.metrocast.net/~modlin1">http://www.metrocast.net/~modlin1</a>,
or, one could just as easily say "intelligence is the ability to *create*
patterns"
pattern recognition can also be studied from a 'pattern construction'
perspective which overcomes many of the problems inherrint in the alternate
view, namely, the problem of finding correlations in artifical neural nets or
hypercolums and the environment
mickeyd
Acme Debugging
> L.F...@lycos.co.uk (Acme Debugging) writes:
> >Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bff1mo$q4j$1...@husk.cso.niu.edu>...
>
>><snip>
>
>>>>Incidently, it is also my conclusion that computationalism cannot
>>>>work. For it can only make truth-based judgements. It cannot make
>>>>the needed pragmatic judgements. Biological systems, by contrast,
>>are
>>>>able to make pragmatic judgements. Evolution itself is a
>>>>pragmatically based system.
>
>>I agree it does seem impossible, but has it been proven that the
>>pragmatic judgements cannot be generalized? Do you have any logic
>>besides intuition or common sense to convince you that it could not be
>>generalized sufficiently to serve wide-application, if not
>>universally?
>
>Logic? No. These questions are outside of logic. Logic is good for
>deriving conclusions from premises. But it can never get anything
>that is outside the system. That's its limitation.
>
>
>But then most
>decisions we make in life are made outside of logic. For the
>constraints of logic are too severe.
Well all the major decisions in my life were like "If I major in
Astronomy then I will *most likely* wind up working all night on the
top of some mountain in the middle of nowhere. If we move to Indiana,
then we'll never run out of corn." Etc., times 10,000. But this has
been questioned so many times in this group I'm beginning to think I'm
the only one. How does everybody else make major decisions?" I even
think, "If there's an *80%* chance of rain, I'll take an umbrella.
Otherwise there's a *good chance* I can buy one at K-Mart."
(*emphasizing probability*). Now that I think about it, when discussing
the map with my wife, she'll say "If we take route 43 then we'll avoid
the rush hour"...etc., etc., so I'm not the only one.
>The Sherlock Holmes stories are sometimes given as examples of
>logic. They are not. Sherlock Holmes actually had to go out into
>the field and look for those cigarette butts (or whatever else was
>the needed evidence). Going out and collecting new evidence is not
>part of logic.
Agreed. Logic is rarely constructive. Dark Ages, etc. But then
programming is the really neat exception. But the major budget in most
commercial projects is debugging and beta-testing, and that's mostly
science.
>Sure, you can describe it as if it were logic. The mistake often
>made is to assume that if you can describe it as if it were logic,
>then it really was logic.
Yeah, it is usually contrived, like on Star Trek. If you're ever
arguing with an AI bomb trying to convince it not to go off, and it
tells you you are just a fallible human, reply "An idea is valid
regardless of the source" and it will shut down. (Well according
to the author of Dark Star anyway...) Or just tell it you are it's
creator and you are imperfect (Kirk to Nomad).
>I was recently on a jury in a local court case. When instructing the
>jury, the judge was very clear that we were not expected to make a
>logical decision based only on the testimony given. Rather, we were
>expected to call upon our life experiences in reaching our
>conclusions. (In fact, that's the whole point of a jury of one's
>peers).
Couldn't agree more. Life experiences: so empirical. Juries
are a great example. Also for the importance of the decision.
>The intelligent human has access to vastly more information than is
>available within the limited input channels of a computer. For the
>important decisions of every day life, logic applied to the defined
>input usually cannot match the ability of the human, made possible by
>the broad array of additional information that is available.
Agree.
>
>No, I cannot give a logical proof of the above. It is one of the
>limitations of logic that no such proof is possible.
>
Well, I just thought there was an off-chance you might have something
like that, and I think it was you that mentioned Nancy Cartwright. I
have a burning interest lately. There are logical proofs of things, but
most are real-world trivial.
And while on the subject, in "informal logic" and "convincing" (read
persuasive argumentation) terms both you and the other guy used, there
is no burden of proof so you were both right. In formal logic it is on
the person making the assertion, so that guy was right too, IMHO.
Surprised nobody made that distinction because it was all over the
place in the terminology. OOPS! Wrong group! Nevermind. :-)
Thanks,
Larry
Curt Welch
> cu...@kcwc.com (Curt Welch) writes:
>
> [big snip leading to the need for a purpose]
>
> I agree that purpose is important. I'm not sure we have much
> agreement beyond that.
:)
> That seems simplistic. Choosing whether to raise an arm is rarely a
> matter of life and death. I realize you are using the example as a
> metaphor. But I think it is too simplistic for that role.
It wasn't a metaphor actually. It was simplistic to make a very important
point. Long before an AI can decide if it's going to go read messages on
usenet, or go watch TV, it has to first figure out the simple things in
life, - like whether it should raise it's arm, or let it hang. Or whether
it should open it's eyes or leave them shut. Or turn it's head to the
left, or to the right.
Honstly, these are first things the AI has to decide. How will it decide
what to do?
If it doesn't have a way of deciding these most simple of things, how is it
ever going to learn later in life to make the complex decisions?
Are we the creators required to make every decision for it, like "when it
should raise it's arm"? If it is intelligent and will later in life be
able to make far more complex decision on it's own, why can't it start it's
life figuring the arm question out?
As creators of AI, if we think we know how to create an AI that can make
the complex decisions, don't you think it would be trivial for us to know
how it's going to make the simple decisions first? If on the other hand,
we don't understand how it's going to make these decisions, how can we hope
to understand how the machine is going to make the complex ones later?
Much of what we think about in terms of "decision" making starts to look
very different when we ask the simple questions of "to rise, or not to
raise", instead of the the more complex ones like, "fight or flight".
> >You could pick a purpose like "to serve mankind". That would be nice if
> >we
>
> Which "mankind"? Should it server George Bush's mankind? Or Saddam
> Hussein's mankind? How is it to know which is which?
Yeah, that's my point. Any purpose that high-level and complex will
require full AI to understnad. It has to be smart enough to ask the type
of questions you just asked if given "to server mankind" as it's purpose.
And since we are trying to give it the purpose to answer the question of
how we create AI, giving it a purpose which still requires we create AI
befor it can use the purpose isn't helping us.
> I would suggest that you are overrating conditioning.
Yeah, and you are not alone.
> What was it that conditioned Beethoven to write his symphonies,
> instead of just replaying symphonies that already were available and
> were known to be audience pleasers? What was it that conditioned
> Darwin to decide on his theory of evolution? What was it that
> conditioned Einstein to decide on his relativity theories? Or what
> conditioned von Neumann to come up with his design for the computer?
> The same question could be asked in countless other ways.
> Conditioning is too simple an explanation to account for the
> diversity and creativity of human activity.
Do you know any of those people well? Why would you expect to understand
what makes them do the things they do?
Instead of asking these questions about people you have never meet, try
asking them about people you know well. Ask it about yourself. Why do you
do all the things you do? Do you not do the things you do because past
experience has shown that the behavior makes things better for you?
Have you ever had a bad experience which did not condition you to try and
aviod the experience in the future?
Have you ever had a good experience which did not condition you to try and
repeat the experience in the future?
I think most people that study behavior would not be so quick to say
"conditioning is overrated".
But, even firm belivers in the power of conditioning are not so quick to
say "it's everything" to inetlligence.
Keep in mind however that I'm not defining intelligence to be "human
behavior". In this approach, I've redefined what intelligence is. I've
defined it to be "learning skill". A machine can be very intelligent, and
learn a lot about some limited domain, and still act totaly un-human.
To pass a Turing test, you have to not only be intelligent, but you must
have learned a huge base of "human" behavior. I do not argue that to get
all that human behavior not going to be easy. My "strong learning"
algorthm won't just be born "acting human". You either need to give it a
very human-like life to live (not very easy), or you have to help train it
"act human" like an actor learns to act like many different types of
people. And that's years of work.
Sometimes when people debate me, I think they total miss this point. They
_define_ intellignece to be human behavior, and then argue that my learning
approach doesn't explain human behavior.
This of course is a non argument because I never said my approach would
naturally create human behavior.
> >We do not raise are arm "to survive" or keep it hanging "to survive".
> >We do it becuse we have been conditioned to do it based on our training.
> >We do it to get a reward, or to avoid a punishment. We do it to get
> >food, or prevent pain.
>
> How do you account for acts of altruism?
The easy out is to just call all such acts innate.
But the truth is, it's trivial to condition people to do anything,
including perform acts of altruism. The military is very good at that type
of conditioning. Society encourages a lot of of that type of behavior as
well.
> >This points us to a simple and obvious way to give AI the purpose it
> >needs, in a way that the machine can deal with. We just give it a
> >reward and punishment system that tells it when something "bad" is
> >happening and when something "good" is happening.
>
> When designing the AI appliance, how will you figure out what to
> count as good and what to count as bad in your programming?
> And how would that be different from what you were criticizing
> above when you wrote:
>
> But they know "what to do", because we tell them. We build
> their behvior into them. We specify what they do, and when,
> and how. And this clearly isn't AI.
When my AI Module is "punished", I am not telling it what it did wrong.
This is a very important concept you can't overlook. I am not telling it
that "you lifted your arm when you should have left it still".
In other words, I am _not_ telling it _what_ to do, or _when_ to do it.
I'm only telling it "life for you sucks now more than it did a second ago".
And when I reward it, I am telling it "life is better for you now".
What it does with this information is up to it.
I for example can motivate it to keep it's battery charged just by giving
it increasing amounts of pain as the battery runs down. It has to figure
out on it's own what it has to do to make life better for itself.
Giving it a motivation is a very different problem than telling it what,
and how, and when, to do everything.
> >With that simple defintion of "purpose", we can no look at what it takes
> >to build a machine that can decide for itself if it's better to raise
> >it's arm, or keep it hanging. The purpose of the hardware which we have
> >to hard-code into the machine is to do whatever it can to reduce the
> >probability of future punishment, and increase the probability of future
> >reward.
>
> But then you have to design the machine so that it can estimate these
> probabilities of future reward and punishment. It seems to me that
> this would require a lot more than conditioning.
That's really just a mater of what you think "conditioning" is. I see it
as a system that can predict future reward and punishment. The better it
is at doing this, the better choice it can make, and the more intelligent
it is. I've never said that doing a good job at this was going to be easy.
I've only said that this is what we need to work on to build AI.
However, I don't think the human-level AI is going to be all that complex
either. I do think it's just basically a good statistic system at work.
I'm quite sure that human behavior is many orders of magnature more complex
than AI is.
> >That gives us something we can actually build, unlike the "be
> >intelligent" purpose, which gives us nothing to build.
>
> >Now, when we look at the "find patterns" idea, we have something to
> >guide us. We aren't just looking for any pattern, we are looking for
> >patterns that help us achieve our purpose. To achieve the purpose of
> >minimsing punishment and maximize reward, there is only one thing we
> >need to do well, and that's to predict how much reward, or punishment,
> >any choice we make now is likely to create.
>
> >This of course is not a simple problem. It's like the chess game
> >problem of "what is the best move now?". The better you can predict
> >the outcome of the entire rest of the game, the better you can answer
> >that question.
>
> In the case of chess, prediction is based on the rules of the game.
> If you want the analogy to be relevant, it would seem that you would
> have to preprogram your AI system with the rules of the game of
> minimizing punishment and maximizing reward. In other words, you
> would have to pre-program intensive knowledge. Such a view is
> consistent with the No Free Lunch theorems, but it seems inconsistent
> with your goal of a learning system.
Why can't it learn the "rules"?
Chess programs predict future "reward and punishment" by performing a huge
search very quickly. We don't. We predict future "reward and punishment"
in a chess game by heuristics which we learn by plaing many games. The
more you play, the better your heuristics of "predicting reward and
punishment" become, and the better your playing ability.
Do you not think that a chess player just "knows" how dangerous a position
is by looking at it to a degree that normally far excceds a normal
chess-playing computers ability? Do you not see this as his brain using
this same strong ability to predict future reward and punishment based on
past experience (i.e., all the games the guy has won and lost in the past).
> >The problem of AI is not to just prevent short-term pain, but to
> >minimize it for the entire rest of time. So doing something like
> >raising an arm, which is known to create more pain that keeping it
> >hanging is not wrong, if the AI belives it will reduce pain in the long
> >term.
>
> How will the AI come to believe that?
Because life is not so simple that raising your arm _always_ causes you
pain.
You can for example learn that rasing your arm is what you do to to call
for help if someone that knows you can see you. And you have done this
many times in your life and the help is almost always able to "get you out
of pain".
And you can learn that under a different condition (when a "bully" is
attacking you), raising your arm will cause you pain. You have learned
this because you have been running into this bully for years.
But now the bully is there, and so is a few people that would come to help
you. And you belive that if you don't call for help, this bully will
continue to abuse you for the next 30 minutes. And if you raise your arm,
he's likely to hit it with his stick. But the pain of getting hit seems
much less than the pain of being abused for the next 30 minutes, so you
raise the arm.
> >The only tool we have for predicting the future is to study the past.
> >And this is where the "find patterns" ideas comes into play.
>
> The past is a poor predictor of the future.
So what do you suggest we use? It's all we have got.
> From a past of computers
> too big and too expensive for anything other than large business and
> research projects, how would one predict today's world? It would
> seem that an AI that underwent conditioning in 1950 would have been
> useless by 1970 and even more useless today. Yet there are plenty of
> people around in 1950 who are still doing useful things today.
"conditioning" never stops. It's active 24x7 for your entire life.
> >So now we not only need to "find patterns", we know why we need to find
> >them. We need to find them so we can best predict what actions to take
> >next.
>
> >So to me, intelligence is the ability to act in ways that minimise all
> >future punishment and maximize all future rewards.
>
> By the way -- it isn't quite the same thing -- but you might find it
> useful to read William T. Powers, "Behavior: the control of
> perception".
It sounds like I might from the title.
> >Finding patterns by studying the past is the obvious tool it's going to
> >use to do this.
>
> >And to be a bit more specific, there are two obvious things the machine
> >will analize. The first is to predict the probability of pain/pleasure
> >which any action is likely to create. If raising the arm always causes
> >more punishment than keeping it hanging, then the obvious choice for the
> >AI is to keep it hanging.
>
> >But the second thing that is important is to try and find patterns that
> >identify what "situation" the AI is in, so that it can make predictions
> >based on the current situation. So maybe raising your arm in some
> >situation is good, and in others it is bad. So, what we need to do is
> >create a machine which is able to define "what" situation we are in, so
> >the anayisis can be based on that information.
>
> Well now you may be getting somewhere. But you will need a clear
> idea of what is to count as a "situation" before you can make much
> progress with this.
Not only do I have a clear idea of how I feel we need to define a
"situation", I've got a working network doing it. This is the purpose of
my inet - to define the current "situation" for the onet. But half of what
it uses for it's input is the output of my onet, which in itself, is an
important part of the defintion of "situation", and I don't have that
completely answered yet. That's what I'm working towards now.
A lot of Bill's work is very closely tied to what my onet still needs to do
in this area of "situation prediction" which is why I'm interested in what
he is doing. So far, I'm not sure if his network designs can fit into my
structure, but what he's trying to make his networks do is very much the
same stuff I need mine to do, so there's good potential for sharing
understanding of the general problem area.
Curt Welch
> >In: <a
> > href="http://www.metrocast.net/~modlin1">http://www.metrocast.net/~mod
> >"Bill Modlin" <mod...@metrocast.net> wrote:
> >
> >> So intelligence is the ability to find patterns.
>
> or, one could just as easily say "intelligence is the ability to *create*
> patterns"
>
> pattern recognition can also be studied from a 'pattern construction'
> perspective which overcomes many of the problems inherrint in the
> alternate view, namely, the problem of finding correlations in artifical
> neural nets or hypercolums and the environment
This also brings up another point I've had in playing with various "patern
recognition" system. That's the question of "what is a pattern"?
What I have found is that the processes can easily define the pattern, and
that depending on what definition of "pattern" your system is built on, it
can always build information about that type of pattern in a dataset. But
if you change the defintion of pattern, you can build a completely
different "understanding" of the meaning of the same dataset.
As a simple example, think of a pattern being a sequence of two letters in
a text stream. You can build a pattern extraction system to identify two
character patterns. And then, becase you give those patterns a new "name",
you can look for pairs of repetitions of those patterns. Continuing to do
this gives you a hierachy of understanding about your data set based on
this type of pattern.
Now do the same thing, with the same dataset, but use tuples instead of
pairs. The heiarchy of understaning you create will be very different -
yet, because it's based on the same dataset, will of course have a lot in
common.
So if you are going to do unsupervised pattern extraction with the intent
of "automatic understanding" how do you know that any defintion of
"pattern" your code uses is a useful or a "correct" one?
To answer this, I think you must look elsewhere. You must have other
reaons for why you are doing the pattern extraction or for judging why one
type of pattern would be better than another (like for effecience or
functional reasons).
I don't have all the answers for this, but I find this question an
intersting one.
David Longley
<ricke...@cs.niu.edu> writes
Sure - and you are a master at that if nothing else!
--
David Longley
Neil W Rickert
>Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bffjv5$7fr$1...@husk.cso.niu.edu>...
>> L.F...@lycos.co.uk (Acme Debugging) writes:
>>>I agree it does seem impossible, but has it been proven that the
>>>pragmatic judgements cannot be generalized? Do you have any logic
>>>besides intuition or common sense to convince you that it could not be
>>>generalized sufficiently to serve wide-application, if not
>>>universally?
>>Logic? No. These questions are outside of logic. Logic is good for
>>deriving conclusions from premises. But it can never get anything
>>that is outside the system. That's its limitation.
>Ok.
>>But then most
>>decisions we make in life are made outside of logic. For the
>>constraints of logic are too severe.
>Well all the major decisions in my life were like "If I major in
>Astronomy then I will *most likely* wind up working all night on the
>top of some mountain in the middle of nowhere. If we move to Indiana,
>then we'll never run out of corn." Etc., times 10,000.
Quite so. And if you had majored in astronomy, that would not have
been wrong. It merely would have directed your life in a different
direction from the one you took via computer science.
These are pragmatic decisions. Often, any choice could work. It
isn't a true/false dichotomy that is being tested. Sometimes it is a
better or worse pragmatic choice. At other times, you will never
really be sure. Maybe astonomy would have been a better choice, but
there is no way to tell since you took a different fork of the road.
That our life is chock full of such decisions -- which fork in the
road to take, when you have no particular destination -- is the basis
for our ideas about having free will.
Once you have made your choice -- say for computer science, you can
then invent arguments that have the apparent form of logic, in order
to explain your choice. But such arguments are rationalizations.
They don't accurately account for how you made the decision.
> But this has
>been questioned so many times in this group I'm beginning to think I'm
>the only one. How does everybody else make major decisions?"
They probably make them the same way you do. But then they invent
after the fact rationalizations, so that they can maintain the
illusion that they use logic for their decision making.
To a large extent, both academic philosophy and AI are the religions
of idolatorous worship of the false god of logic. Every good
decision is to be credited to the deity (logic). Every bad decision
is to be blamed on the satan (fallacy).
[I have been known to say that analytic philosophers are called
"analytic" for the same reason that Little John was called
"little" -- because they aren't.]
dan michaels
> In: http://www.metrocast.net/~modlin1,
> "Bill Modlin" <mod...@metrocast.net> wrote:
>
> > So intelligence is the ability to find patterns.
>
Hi Curt, I have extracted a few of the lines from Bill's paper:
==========================
We might call an animal intelligent, but that was relative to other
animals. Only people could do complicated thinking... no dog could
converse, or play tic-tac-toe. We tied real intelligence to such
complex things, and felt we had no serious competition for the label
"intelligent".
...............
To me, the central part of intelligence isn't the skills and behaviors
or the particular knowledge which it allows us to learn, but the basic
ability to understand how the things we observe fit together. I
suggest that what we call intelligence is mostly an ability to
discover underlying patterns in a variety of events.
.....................
But all this behavior depends on finding the patterns in the first
place, and to me that's the interesting, and the hard, part of the
job. The rest is "mere mechanics", playing out the information
entailed by the pattern.
So intelligence is the ability to find patterns.
=======================
Looks like Bill tried to come up with the minimalist definition of
what constitutes intelligence ...... but there are a couple of
internal inconsistencies here, plus a hint at what's missing.
First, dogs and other animals all do pattern extraction of some sort,
so they would be labelled as intelligent by the last line - as would
most neural networks [which do pattern extraction better than most
anything else].
In paragraph 2, Bill says: "... but the basic ability to understand
how the things we observe fit together ...". This points at what's
missing from his last statement, which is, the ability of intelligent
life to put the information it extracts from the environment into the
context of its past experiences or some inherent knowledge. [note - it
doesn't necessairly need to do this consciously, so consciousness is a
different issue].
At a minimum, you need both pattern extraction and context. As I
mentioned elsewhere, referring to how the brain works, Rosenfield
said: "... perception, recognition, and memory are not separate
processes, but an integral procedure ... there are no symbols in the
brain; there are patterns of activity that acquire different meanings
in different contexts".
The context can come from built-in knowledge, as the case of
instinctual knowledge in lower animals, or from prior learning. Either
way, the new patterns extracted are only useful in relation to some
context. Higher intelligences have a much wider context for the
purposes of comparison. This is the reasoning behind giving Cyc an
encyclopedic knowledge - so the system has something to compare new
input against. Cyc is missing a few other things, but it has this
part.
just a couple of 0.02,
- dan
========================
Neil W Rickert
>Neil W Rickert <ricke...@cs.niu.edu> wrote:
>> cu...@kcwc.com (Curt Welch) writes:
>> [big snip leading to the need for a purpose]
>> I agree that purpose is important. I'm not sure we have much
>> agreement beyond that.
>:)
>> That seems simplistic. Choosing whether to raise an arm is rarely a
>> matter of life and death. I realize you are using the example as a
>> metaphor. But I think it is too simplistic for that role.
>It wasn't a metaphor actually. It was simplistic to make a very important
>point. Long before an AI can decide if it's going to go read messages on
>usenet, or go watch TV, it has to first figure out the simple things in
>life, - like whether it should raise it's arm, or let it hang. Or whether
>it should open it's eyes or leave them shut. Or turn it's head to the
>left, or to the right.
I expect that a human infant has worked out how and when to move its
arms or open its eyes, long before it has discovered that it has arms
and eyes.
>Honstly, these are first things the AI has to decide. How will it decide
>what to do?
My point stands. If often doesn't matter whether you raise your
arms.
>If it doesn't have a way of deciding these most simple of things, how is it
>ever going to learn later in life to make the complex decisions?
>Are we the creators required to make every decision for it, like "when it
>should raise it's arm"? If it is intelligent and will later in life be
>able to make far more complex decision on it's own, why can't it start it's
>life figuring the arm question out?
>As creators of AI, if we think we know how to create an AI that can make
>the complex decisions, don't you think it would be trivial for us to know
>how it's going to make the simple decisions first?
You said in an earlier post, that your AI system would have free
will. If it will have free will, then you cannot know how it will
make a decision to raise its arm. That becomes a decision for the
AI, and one that is not up to you.
> If on the other hand,
>we don't understand how it's going to make these decisions, how can we hope
>to understand how the machine is going to make the complex ones later?
If your system is going to be intelligent and have free will, then
you cannot know how it will make decisions.
>> I would suggest that you are overrating conditioning.
>Yeah, and you are not alone.
>> What was it that conditioned Beethoven to write his symphonies,
>> instead of just replaying symphonies that already were available and
>> were known to be audience pleasers? What was it that conditioned
>> Darwin to decide on his theory of evolution? What was it that
>> conditioned Einstein to decide on his relativity theories? Or what
>> conditioned von Neumann to come up with his design for the computer?
>> The same question could be asked in countless other ways.
>> Conditioning is too simple an explanation to account for the
>> diversity and creativity of human activity.
>Do you know any of those people well? Why would you expect to understand
>what makes them do the things they do?
I don't expect to understand that. But I do understand that
conditioning cannot account for the creativity we see in the cases
mentioned.
>Instead of asking these questions about people you have never meet, try
>asking them about people you know well. Ask it about yourself. Why do you
>do all the things you do? Do you not do the things you do because past
>experience has shown that the behavior makes things better for you?
It does not seem to be explainable as conditioning.
>Have you ever had a bad experience which did not condition you to try and
>aviod the experience in the future?
I had bad experiences when I argued with Longley a few years back. I
don't seem to be trying to avoid it now.
>Have you ever had a good experience which did not condition you to try and
>repeat the experience in the future?
Many good experiences are not repeatable. They are singular events.
>> >We do not raise are arm "to survive" or keep it hanging "to survive".
>> >We do it becuse we have been conditioned to do it based on our training.
>> >We do it to get a reward, or to avoid a punishment. We do it to get
>> >food, or prevent pain.
>> How do you account for acts of altruism?
>The easy out is to just call all such acts innate.
>But the truth is, it's trivial to condition people to do anything,
>including perform acts of altruism. The military is very good at that type
>of conditioning. Society encourages a lot of of that type of behavior as
>well.
Conditioning does not adequately explain it, in my view.
>> >This points us to a simple and obvious way to give AI the purpose it
>> >needs, in a way that the machine can deal with. We just give it a
>> >reward and punishment system that tells it when something "bad" is
>> >happening and when something "good" is happening.
>> When designing the AI appliance, how will you figure out what to
>> count as good and what to count as bad in your programming?
>> And how would that be different from what you were criticizing
>> above when you wrote:
>> But they know "what to do", because we tell them. We build
>> their behvior into them. We specify what they do, and when,
>> and how. And this clearly isn't AI.
>When my AI Module is "punished", I am not telling it what it did wrong.
Sure you are.
>> >This of course is not a simple problem. It's like the chess game
>> >problem of "what is the best move now?". The better you can predict
>> >the outcome of the entire rest of the game, the better you can answer
>> >that question.
>> In the case of chess, prediction is based on the rules of the game.
>> If you want the analogy to be relevant, it would seem that you would
>> have to preprogram your AI system with the rules of the game of
>> minimizing punishment and maximizing reward. In other words, you
>> would have to pre-program intensive knowledge. Such a view is
>> consistent with the No Free Lunch theorems, but it seems inconsistent
>> with your goal of a learning system.
>Why can't it learn the "rules"?
The rules are too complex.
>Chess programs predict future "reward and punishment" by performing a huge
>search very quickly. We don't. We predict future "reward and punishment"
>in a chess game by heuristics which we learn by plaing many games.
That's questionable. I think it would be more accurate to say that
we describe it as "heuristics" to avoid having to admit that we don't
know how it is done.
>Do you not think that a chess player just "knows" how dangerous a position
>is by looking at it to a degree that normally far excceds a normal
>chess-playing computers ability?
Sure. But a computer does not look at it at all. A computer has no
perception, and no ability to look.
> Do you not see this as his brain using
>this same strong ability to predict future reward and punishment based on
>past experience (i.e., all the games the guy has won and lost in the past).
I don't believe that is a correct explanation.
>> >The only tool we have for predicting the future is to study the past.
>> >And this is where the "find patterns" ideas comes into play.
>> The past is a poor predictor of the future.
>So what do you suggest we use? It's all we have got.
We do not need to predict the future. We need only to direct our own
actions as best we can.
David Longley
<ricke...@cs.niu.edu> writes
>L.F...@lycos.co.uk (Acme Debugging) writes:
>>Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bffjv5$7fr$1@husk
So when we find you posting obnoxious, ignorant, self-opinionated drivel
here each day, it's really no more than a "rationalization" of your ill-
informed-decision making then? (all of which, is sadly beyond your
control). No wonder you crave "discussion" - you delude yourself to such
an extent that even you don't realise that what you are up to amounts to
nothing less than plagiarism.
>
>> But this has
>>been questioned so many times in this group I'm beginning to think I'm
>>the only one. How does everybody else make major decisions?"
>
>They probably make them the same way you do. But then they invent
>after the fact rationalizations, so that they can maintain the
>illusion that they use logic for their decision making.
>
>To a large extent, both academic philosophy and AI are the religions
>of idolatorous worship of the false god of logic. Every good
>decision is to be credited to the deity (logic). Every bad decision
>is to be blamed on the satan (fallacy).
>
>[I have been known to say that analytic philosophers are called
>"analytic" for the same reason that Little John was called
>"little" -- because they aren't.]
>
You've been known to say an awful lot of obnoxious, ignorant drivel in
your time Neil. The only time you don't write drivel is when you are
plagiarising someone else's work, often having dismissed their whole
field out of hand.
You are an intellectual parasite - and what's worse - you're hopelessly
irrational with it.
--
David Longley
Acme Debugging
> "Acme Debugging" <L.F...@lycos.co.uk> wrote in message
> news:35fae540.03072...@posting.google.com...
> > "Bill Modlin" <mod...@metrocast.net> wrote in message
> news:<eemdnZkKZ6E...@metrocast.net>...
> >
>>
>> >Formal statistical tools are demonstrably superior to heuristic
>> judgment, if
>> >and only if heuristic judgment has first selected appropriate
inputs
>> to
>> >those tools.
>>
>> Did you intend to leave "sometimes" out?
>>
>> If so then please explain, completely within the methodology of
>> statistics, why statistical prediction of commodity prices (i.e.
<snip>
>Lol... In my mind
>
"In my mind" - great phrase. I like how you avoid unnecessary
argument.
>
>the word "appropriate" carried the burden for
>"sometimes"... in the case of commodity price prediction, so far
nobody has
>come up with an appropriate set of quantifiable inputs such that
regression
>produces a particularly useful output, if indeed such a set exists.
>
"Appropriate" probably worked. I was just looking for a place to post
the commodities example...
I was always able to get good correlation even just using price
variables, albeit fancily concocted. Just didn't work in real-time.
>
>Remember the context for the statement. I'm responding to Longley,
who has
>frequently posted examples of cases in which regression works better
than
>human judgment, and argues that therefore regression is a desirable
>substitute for human intelligence. I wish to cede that there are
such
>cases, but dispute the "therefore".
>
I understood that more-or-less, and felt somewhat ambiguous on that
score. I emphasize that, from everything I've read about his prison
work, his methodology was "best" (Not implying that you don't agree).
Who wouldn't want to systematize (benevolent) behavior improvement of
criminals and scientific measurements thereof? (Not implying that's
all he did). I think he must have been up against some "uninformed
wisdom" and/or politics, and I don't see how one could avoid some
negative conviction about both. But he said one time "Maybe someone
higher up knew something I didn't" (I doubt anything legitimate) and
that says "objective" to me. I've been in some political situations
and it really is a load of crap to deal with.
>I'm not pushing the reliability of regression, just saying its
purported
>reliability does not make it intelligent.
It's been posted and I don't think anyone disagreed, not me anyway.
>
>I probably should have included "sometimes" anyway. :-)
Hah - Like I tell Longley I appreciate adults and hope to be one
someday.
>
>Any comment on the actual issues discussed in the post?
>
No because you are doing such a professional job and you know what
you're talking about. I know little about behaviorism or CS. Of course
however old anyone is, that's how long they've been studying and at
least attempting to modify human behavior. Some are obviously better
at it than others, so there's a curve, and some will be on the far
right edge of that curve. I think David is usually talking about the
rest of the curve, legitimately so in his context.
Larry
David Longley
<ricke...@cs.niu.edu> writes
>cu...@kcwc.com (Curt Welch) writes:
>
>>Instead of asking these questions about people you have never meet, try
>>asking them about people you know well. Ask it about yourself. Why do you
>>do all the things you do? Do you not do the things you do because past
>>experience has shown that the behavior makes things better for you?
>
>It does not seem to be explainable as conditioning.
<possibly because you know so little about it?>
>
>>Have you ever had a bad experience which did not condition you to try and
>>aviod the experience in the future?
>
>I had bad experiences when I argued with Longley a few years back. I
>don't seem to be trying to avoid it now.
>
>
Possibly because you are a parasitic psychopath who also benefits form
the exchanges?
>>Have you ever had a good experience which did not condition you to try and
>>repeat the experience in the future?
>
>Many good experiences are not repeatable. They are singular events.
Ever had sex Neil? Got any favourite foods?
You do like arguing!!!! (and you do it to the point of pathology)
--
David Longley
David Longley
michaels <d...@oricomtech.com> writes
>cu...@kcwc.com (Curt Welch) wrote in message news:<20030721044857.809$Dm@newsread
>er.com>...
>> "Mr Michael Bibby" <s403...@student.uq.edu.au> wrote:
>> > >In: <a
>> > > href="http://www.metrocast.net/~modlin1">http://www.metrocast.net/~mod
>> > > li n1</a>,
>> > >"Bill Modlin" <mod...@metrocast.net> wrote:
>> > >
>> > >> So intelligence is the ability to find patterns.
>> >
>> > or, one could just as easily say "intelligence is the ability to *create*
>> > patterns"
>> >
>
>> So if you are going to do unsupervised pattern extraction with the intent
>> of "automatic understanding" how do you know that any defintion of
>> "pattern" your code uses is a useful or a "correct" one?
>>
>> To answer this, I think you must look elsewhere. You must have other
>> reaons for why you are doing the pattern extraction or for judging why one
>> type of pattern would be better than another (like for effecience or
>> functional reasons).
>>
>
>
>compare the new pattern - be it an innate or learned context. I think
>you will have this is your concept of inet-onet, but this part is
>missing from Bill's direct definition in his paper.
>
>Living animals [like humans] that learn context, learn it from
>interactions with their environment at a young age. Babies have the
>ability to babble built-in by genetics, but they learn language by
>listening to and making noises back at humans, until they get it right
>- a process which takes a huge amount of trial-and-error and feedback
>over 4-5 years.
Glen calls these operants.
>
>Your inet-onet will learn things from its environment, and then by
>playing those back to the environment, and seeing how the environment
>responds in return, it will learn if it is doing things correctly.
>Real-world learning is really supervised. Just like a child, if
>inet-onet were in a vacuum, it would learn ... next to nothing. The
>child's initial motivations come from comfort [ie, dry-bottom and
>not-pain] as well as food [ie, tastes-good], so not-pain and
>tastes-good are no doubt the "contexts" pre-programmed in by genetics,
>and later on, learning vectors off of that, one bit at a time.
So where's the argument with Glen? Surely it takes little to say "sorry,
I've got it now"? He might even reward you with a few nice words and/or
the withdrawal of his hostility... <g>
>
>
>a few more 0.02,
>- dan
>===========================
--
David Longley
Neil W Rickert
>So when we find you posting obnoxious, ignorant, self-opinionated drivel
>here each day, it's really no more than a "rationalization" of your ill-
The Longley bot needs to spend time in a Skinner box, so as to correct
its improper use of ""rationalization".
>informed-decision making then? (all of which, is sadly beyond your
>control). No wonder you crave "discussion" - you delude yourself to such
>an extent that even you don't realise that what you are up to amounts to
>nothing less than plagiarism.
The bot also needs retraining on appropriate use of "plagiarism".
David Longley
<ricke...@cs.niu.edu> writes
Amazing isn't it - when will you learn and learn that you're learning?
--
David Longley
Curt Welch
> "Neil W Rickert" <ricke...@cs.niu.edu> wrote in message
> news:bff1mo$q4j$1...@husk.cso.niu.edu...
> > Your view is that the world is a highly patterned place, and
> > intelligence involves the discovery of these patterns.
> >
> > What I have concluded, is that the world is a disorderly place
> > without clear patterns or recurrent events.
It is mostly disorder, there is no doubt about that. But there are plenty
of patterns as well and that is what the brain finds, and makes use of.
> Certainly there is plenty of real or apparant disorder to be found...
> even ignoring quantum indeterminancy, if we measure any quantity
> precisely enough we find it varying randomly by small amounts from
> measurement to measurement, a result of what has come to be called
> thermal noise.
>
> But if we ignore variations below the noise level we can often get
> measurements that are reasonably well behaved and non-random in quite
> definite ways.
Ah, but be careful how you say that.
It's trivial to extract data hidden below the noise floor as well. You
would never want to "ignore" variations below the noise level.
For example, I can create a signal stream (A) like this:
for (t = 0; ; t < 100; t++) {
X = sin(t*3.1415926/100);
A = X + (random()%10000)/100.0 - 50.0;
}
Where the signal (sin(t) with a value of +- 1) is well below the noise
floor (+-50).
It produces numbers like this:
T X A
0 0.000 43.830
1 0.031 10.451
2 0.063 -46.337
3 0.094 41.264
4 0.125 30.355
5 0.156 17.766
6 0.187 -49.003
7 0.218 13.308
Yet, by looking for correlations between T and A, it is trivial to
reconstruct X simply by avergering the value of A over time relative to t
like this:
for (t = 0; ; t = t++ % 100) {
X = sin(t*3.1415926/100);
A = X + (random()%10000)/100.0 - 50.0;
e[t] += (e[t] - A) * 0.00001;
}
Which produces results like this if you let it run for one million cycles:
t x e[t]
0 0.000 0.038
1 0.031 0.058
2 0.063 0.067
3 0.094 0.122
4 0.125 0.063
5 0.156 0.195
6 0.187 0.122
7 0.218 0.102
8 0.249 0.234
9 0.279 0.194
10 0.309 0.358
11 0.339 0.215
12 0.368 0.351
13 0.397 0.457
14 0.426 0.399
15 0.454 0.448
The longer it runs, and the smaller the 0.00001 factor is, the better the
results. e[t] is calcuated purely from A, and not from X, yet it is able
extract the value of X from A because of it's correlation with t.
e[t] is a measure of the correlation of A with respect to t. It shows the
causeality relationship between t and A. (which turns out to be sin(t) in
this exmaple).
So the above shows how it's quite trivial to extract correlation data even
when the data is over an order of magnatude below the noise floor in a
signal. This is exactly what the brain is constantly doing to extract
useful "patterns" from the noise of the enviroment even when the data is
all hidden below the noise floor. This is why the world can look like 100%
noise, yet still contain a lot of very useful "data" to allow the brain to
make predictions with.
This is exactly what my net is doing to predict reward and punishment.
The level you quantize at changes your ability to extract signal, but even
if you take my data above, and quantize the value at +-1, you can still
extract the sin singal which is below the quantization level. So here's
the data with the A value quntized to +-1 using a floor() function (always
round down):
t x A
0 0.000 43.000
1 0.063 -42.000
2 0.125 -23.000
3 0.187 19.000
4 0.249 28.000
5 0.309 33.000
6 0.368 4.000
7 0.426 -45.000
8 0.482 16.000
9 0.536 -36.000
10 0.588 -26.000
And here is the extracted signal from the noise:
6500000 Cycles:
t x e[t]
0 0.000 -0.504
1 0.063 -0.496
2 0.125 -0.423
3 0.187 -0.333
4 0.249 -0.272
5 0.309 -0.216
6 0.368 -0.135
7 0.426 -0.080
8 0.482 0.006
9 0.536 0.009
10 0.588 0.087
Notice that e[t] is now .5 lower than it was before. This is because the
floor() function on the average will remove .5 from every sample.
I think prgramatic judgements can be defined in terms of truth-based
judgements so there's no problem at all here. A decision based on the
extracted data in my little example is the heart of what a pragmatic
judegement is all about.
Here is the C code I wrote to create the above numbers in case anyone wants
to play with it (I was running it on a FreeBSD system):
#include <math.h>
main()
{
int t;
double x;
double a;
double d[100];
int cycles;
for (t = 0; t < 100; t++)
d[t] = 0.0;
for (cycles = 0; cycles < 10000000; cycles++) {
for (t = 0; t < 100; t++) {
x = sin(t*2.0*3.1415926/100.0);
a = x + (random()%10000)/100.0 - 50.0;
a = floor(a);
d[t] += (a - d[t]) * 0.000001;
printf("%4d %10.3f %10.3f\n", t, x, a);
}
if (cycles%100000 == 0) {
printf("%d Cycles:\n", cycles);
for (t = 0; t < 100; t++) {
x = sin(t*2.0*3.1415926/100.0);
printf("%4d %10.3f %10.3f\n", t, x, d[t]);
}
printf("\n");
printf("%d Cycles:\n", cycles);
printf("\n");
Bill Modlin
"dan michaels" <d...@oricomtech.com> wrote in message
news:4b4b6093.03072...@posting.google.com...
> cu...@kcwc.com (Curt Welch) wrote in message
news:<20030720191503.053$gQ...@newsreader.com>...
> > In: http://www.metrocast.net/~modlin1,
> > "Bill Modlin" <mod...@metrocast.net> wrote:
> >
> > > So intelligence is the ability to find patterns.
> >
>
>
> Hi Curt, I have extracted a few of the lines from Bill's paper:
>
> ==========================
> We might call an animal intelligent, but that was relative to other
> animals. Only people could do complicated thinking... no dog could
> converse, or play tic-tac-toe. We tied real intelligence to such
> complex things, and felt we had no serious competition for the label
> "intelligent".
Note that this was meant to depict how people may have thought about
intelligence in an earlier time, it is in no way my personal view, and
indeed the thrust of the paper is that this is the wrong way to look at it.
...............
>
> To me, the central part of intelligence isn't the skills and behaviors
> or the particular knowledge which it allows us to learn, but the basic
> ability to understand how the things we observe fit together. I
> suggest that what we call intelligence is mostly an ability to
> discover underlying patterns in a variety of events.
> .....................
>
> But all this behavior depends on finding the patterns in the first
> place, and to me that's the interesting, and the hard, part of the
> job. The rest is "mere mechanics", playing out the information
> entailed by the pattern.
>
> So intelligence is the ability to find patterns.
> =======================
>
>
> Looks like Bill tried to come up with the minimalist definition of
> what constitutes intelligence ...... but there are a couple of
> internal inconsistencies here, plus a hint at what's missing.
>
> First, dogs and other animals all do pattern extraction of some sort,
> so they would be labeled as intelligent by the last line - as would
> most neural networks [which do pattern extraction better than most
> anything else].
In my view dogs and other animals are indeed intelligent. Yes. We can both
broader and deeper patterns and are therefore more intelligent, but they
have intelligence of the same kind as we do.
Obviously biological neural networks, like the ones in brains, can do the
job. I don't see any of the current NN models as really discovering
patterns in the sense I'm talking about... they do some of it, but so far as
I can see are a bit too simple, missing some mechanisms that would be
required for stable operation in a very large recurrent sparsely connected
configuration under an almost completely unsupervised continuous learning
regime. I'd have no problem calling a functioning NN model intelligent if
it really could find patterns on the scale of a dog brain, and indeed my
notion of an AI is itself just another NN with those extra mechanisms
included. I'd be glad to talk more about this, but so far there is still a
pretty large glob of handwaving involved.
> In paragraph 2, Bill says: "... but the basic ability to understand
> how the things we observe fit together ...". This points at what's
> missing from his last statement, which is, the ability of intelligent
> life to put the information it extracts from the environment into the
> context of its past experiences or some inherent knowledge. [note - it
> doesn't necessarily need to do this consciously, so consciousness is a
> different issue].
>
> At a minimum, you need both pattern extraction and context. As I
> mentioned elsewhere, referring to how the brain works, Rosenfield
> said: "... perception, recognition, and memory are not separate
> processes, but an integral procedure ... there are no symbols in the
> brain; there are patterns of activity that acquire different meanings
> in different contexts".
>
> The context can come from built-in knowledge, as the case of
> instinctual knowledge in lower animals, or from prior learning. Either
> way, the new patterns extracted are only useful in relation to some
> context. Higher intelligences have a much wider context for the
> purposes of comparison. This is the reasoning behind giving Cyc an
> encyclopedic knowledge - so the system has something to compare new
> input against. Cyc is missing a few other things, but it has this
> part.
But but but... how did you imagine that patterns can be found and focused
without context in the first place?
Finding out "how the things we observe fit together" includes fitting
content within context as content within broader context extending
recursively to incorporate everything in one all encompassing pattern
spanning everything we have experienced and everything we have inferred from
that experience.
It's a never-ending exercise, and at any stage of the game there may be any
number of discontinuities and discrepancies and incompletely assimilated
sections and dead-ends that don't yet connect anywhere. But local adaptive
processes keep nibbling at the problem and smoothing out the glitches and
noticing new bits of the pattern to make new connections.
New input flows into the net and triggers excitation of nodes recognizing
surface patterns and the patterns of those trigger deeper patterns and so
on, so that the new input is compared to and classified with previous
experience at many levels. These excitation flows evoke associated flows
so
that the current pattern is filled in and projected forward and backward in
time to include all manner of context and predictions from the
contextually-placed new instances of the pattern... it's all in there. :-)
Lol... Did I mention handwaving?
>
> just a couple of 0.02,
> - dan
> ========================
Just a few more points. I can't imagine how you would embed knowledge in
this network There is no representation language or equivalent, I can't
"build in" cyc-like knowledge. I have to let it learn for itself.
The pattern recognition network is passively self-organizing to represent
its entire input stream as compactly as possible. It's a big adaptive
lossy compression algorithm. It has no effectors, nor does it have any
goals or purposes... it just does its thing with the data. The way it does
its compression is to build a model of the causal processes inferred to
exist as an explanation of the inputs it sees.
If you want it to actually "do something", to "behave", put it in a robot,
and give the robot some starter set of innate behaviors. Design the
robotic behavioral controls with a network of nodes capable of associative
conditioning. Cross connect the behavior generation network with the
modeling network, or just use some of the nodes of the modeling network to
implement a behavioral subsystem, initialized with connections and weights
to implement the desired innate behavior. The modeling network gets
inputs from the behaving network including any sensory inputs and effectors
controls and intermediate stages. So the modeling network models the
behavior system, and the behavior system learns connections from the model
to its own corresponding parts through associative conditioning, and the
model eventually will take over control of behavior, which can then diverge
from its initial innate definition to include responses to projected and
remembered situations arising deeper in the model. If the modeling system
is good enough it will spin off a model of itself as an agent in its own
modeled world, patterning itself as wanting the things its innate behaviors
suggested and so on.
If the inputs include people talking in context with general sensory input
from the situations they are talking about, a submodel will develop of
talking, associated with things discussed, and the system will be able to
associate words with situations and ideas, and if the behavior system
includes an ability to generate speech sounds and a little innate babbling
to bootstrap from the system will learn to talk.
Similarly for reading and writing, and pretty soon it will be smarter than I
am and will post a withering note telling me to shut up and let it explain
for itself.
So I will.
Bill
Neil W Rickert
>In article <bfh7na$ab3$1...@husk.cso.niu.edu>, Neil W Rickert
><ricke...@cs.niu.edu> writes
>>cu...@kcwc.com (Curt Welch) writes:
>>>Have you ever had a bad experience which did not condition you to try and
>>>aviod the experience in the future?
>>I had bad experiences when I argued with Longley a few years back. I
>>don't seem to be trying to avoid it now.
>Possibly because you are a parasitic psychopath who also benefits form
>the exchanges?
These radical behaviorist morons have a pseudo-explanation for
everything.
Curt Welch
> "dan michaels" <d...@oricomtech.com> wrote in message
> news:4b4b6093.03072...@posting.google.com...
> > In paragraph 2, Bill says: "... but the basic ability to understand
> > how the things we observe fit together ...". This points at what's
> > missing from his last statement, which is, the ability of intelligent
> > life to put the information it extracts from the environment into the
> > context of its past experiences or some inherent knowledge. [note - it
> > doesn't necessarily need to do this consciously, so consciousness is a
> > different issue].
> >
> > At a minimum, you need both pattern extraction and context. As I
> > mentioned elsewhere, referring to how the brain works, Rosenfield
> > said: "... perception, recognition, and memory are not separate
> > processes, but an integral procedure ... there are no symbols in the
> > brain; there are patterns of activity that acquire different meanings
> > in different contexts".
There is some real confusion in those words...
> Finding out "how the things we observe fit together" includes fitting
> content within context as content within broader context extending
> recursively to incorporate everything in one all encompassing pattern
> spanning everything we have experienced and everything we have inferred
> from that experience.
And the result of it all is always, by definition, the networks
understanding of it's current "context". This is why we do this in the
first place - to define context to a higher degree of resolution.
If you have an input stream that looks like this:
ABAABBAABABABAAABBBABABABAAABBBABABABAAABBBAAA
The only "context" you know about is that the current input is A, or it is
B. You have only have two types of "context" on which to base all your
actions.
But if the pattern extraction system spots the pattern AAABBB and calls it
Y, then you have this to define our context from:
ABAABBAABABABAAABBBABABABAAABBBABABABAAABBBAAA
Y Y Y
So now we know about the Y context and not just the A and B context. And
we can try to use this new understanding of what "context" or "state" we
are in to make better decisions about what action to take based on our
understanding of our current "context".
And like Bill was saying, the pattern system works using all the prior
information as well. So the pattern spotting system uses the fact that we
are in the Y context as it contines to do it's job of defing more "context"
information for us.
Patterns are not "understood" in context, their existence _defines_ the
context.
Daryl McCullough
>>>> So intelligence is the ability to find patterns.
I think that that's in the right ballpark, but it isn't
completely right.
First of all, in humans we can distinguish between *automatic*
pattern-recognition and patterns that require hard work to
recognize.
We recognize visual patterns (faces, circles, lines, etc.)
automatically (pre-consciously). On the other hand, if we
are given a sequence of numbers (perhaps the outcomes of
experiments) and we have to figure out the pattern, it is
a lot of hard work. Basically, what we have to do in the
latter case is to guess a rule, and then generate a sequence
according to that rule and see if the sequences match.
I think that the automatic, "easy" pattern recognition
(recognizing that the person standing in the corner is your old
friend Neil) works so well because it is (at least partly)
hardwired into our brains. I believe that there are definite
limitations to what we can come to recognize automatically.
On the other hand, the non-automatic, "hard" pattern recognition
is completely open-ended, but it is also much harder, and much
less reliable. Some perfectly intelligent humans may never come
to suspect that a sequence of numbers was generated using the
integer part of the Riemann zeta function (or whatever).
When we claim that humans are so much better than other
beings at pattern recognition, we need to worry about whether
we are being almost tautological: humans are better at recognizing
those patterns that humans consider important (or interesting).
We aren't necessarily better at recognizing *all* patterns.
What makes a pattern important? A pattern is
important or not depending on what we intend to do with it --- if
recognizing the pattern serves our purposes, then it is important,
and otherwise it is trivia. When we judge certain phenomena to be
"noise" rather than "signal" it doesn't really mean that it lacks
a pattern---it just means that we believe that the pattern is
irrelevant for us. A super-pattern-recognizer who goes around
discovering faces in clouds or prime numbers in social security
numbers would not be considered intelligent, he would be considered
an eccentric or a crackpot.
To me, the most difficult part of human intelligence to mimick
by a computer is judgement: judgement about what avenues are
worth exploring and which ones are likely to be a waste of time.
There's a chicken-or-egg problem inherent in such decisions: we
don't really *know* whether something is a waste of time until
after we've done it, in which case it is too late. It is therefore
wrong to think of decision-making as something that *follows*
thinking about alternatives---in fact, we have to make decisions
about which alternatives to think about, as well.
--
Daryl McCullough
Ithaca, NY
Acme Debugging
> L.F...@lycos.co.uk (Acme Debugging) writes:
> >Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bffjv5$7fr$1...@husk.cso.niu.edu>...
>
>>>decisions we make in life are made outside of logic. For the
>>>constraints of logic are too severe.
>
>>Well all the major decisions in my life were like "If I major in
>>Astronomy then I will *most likely* wind up working all night on the
>>top of some mountain in the middle of nowhere. If we move to
Indiana,
>>then we'll never run out of corn." Etc., times 10,000.
>
>Quite so. And if you had majored in astronomy, that would not have
>been wrong. It merely would have directed your life in a different
>direction from the one you took via computer science.
Agree. We'd probably have comet Larry by now. ~ ~ ~~:-)
>These are pragmatic decisions. Often, any choice could work. It
Agree. I gave that sermon just yesterday (to email victims).
>isn't a true/false dichotomy that is being tested.
I'm rarely talking about true/false (except in programming). That is
academic logic. I'm usually talking about logic/probability, i.e.
real-world application. The logic is simple, as is the probability.
There is the problem of generalizing sometimes tricky logic behind the
probability unique to the case. But the most intractible problem, also
the end-point of reasoning in humans, is the issue of objectivity.
>Sometimes it is a
>better or worse pragmatic choice. At other times, you will never
>really be sure. Maybe astonomy would have been a better choice, but
>there is no way to tell since you took a different fork of the road.
Agree.
>That our life is chock full of such decisions -- which fork in the
>road to take, when you have no particular destination -- is the basis
>for our ideas about having free will.
Agree. But fair to ask which fork has the highest chance of success.
In many cases, that can be tested, so one generalizes to cases which
can't be tested. Obviously some things are quite random and
unpredictable, others are quite predictable, with all things in
between. Elsewhere in this thread we're beating that to death.
>
>Once you have made your choice -- say for computer science, you can
>then invent arguments that have the apparent form of logic, in order
>to explain your choice. But such arguments are rationalizations.
>They don't accurately account for how you made the decision.
>
Agree. I've said/thought it 100 times. You experience (science or
whatever), and then you find the logic. I also claim that proving an
assertion in which you are invested simply doesn't work due to the
investment in the result. You will always find the logic and it will
always be airtight (in your mind). Then you post it and guess what?
Some jerk disagrees!
The safest way to do "logic" is to perform experiments and see what
happens. But there are other ways to use logic constructively, more or
less. There are ways to force a jury to use logic regardless of
instructions. The most forceful device is telling them two or three
things they already know, but in a different order. It's done all the
time.
>> But this has
>>been questioned so many times in this group I'm beginning to think
I'm
>>the only one. How does everybody else make major decisions?"
>
>They probably make them the same way you do. But then they invent
>after the fact rationalizations, so that they can maintain the
>illusion that they use logic for their decision making.
Ditto.
>
>To a large extent, both academic philosophy and AI are the religions
>of idolatorous worship of the false god of logic. Every good
>decision is to be credited to the deity (logic). Every bad decision
>is to be blamed on the satan (fallacy).
The "To a large extent" quantifier makes all the difference and makes
it work for me.
>
>[I have been known to say that analytic philosophers are called
>"analytic" for the same reason that Little John was called
>"little" -- because they aren't.]
>
My analytic philosophy has always been "Define the problem. Everything
else accounts for 2%." I'm trying to define the problem of how to be
more analyticitistical. (sarcasm)
Logic works. Humans just aren't very good at accuracy. But it gives
you the list of things to try, including the list of things to try to
do it better. High on that list is AI, specifically AI reasoning.
Larry
dan michaels
> > > At a minimum, you need both pattern extraction and context. As I
> > > mentioned elsewhere, referring to how the brain works, Rosenfield
> > > said: "... perception, recognition, and memory are not separate
> > > processes, but an integral procedure ... there are no symbols in the
> > > brain; there are patterns of activity that acquire different meanings
> > > in different contexts".
>
> There is some real confusion in those words...
>
Yeah, it's part of a 190 page book.
====================
.................
> And like Bill was saying, the pattern system works using all the prior
> information as well. So the pattern spotting system uses the fact that we
> are in the Y context as it contines to do it's job of defing more "context"
> information for us.
>
> Patterns are not "understood" in context, their existence _defines_ the
> context.
The context Rosenfield and I were talking about is in the broader
sense, not just the temporally-local sense. IE, long-term memory LTM
as opposed to short-term STM. You're basically talking about strings
here, and STM, we were talking about more life-like situations, and
LTM context.
If you say to your sister, "I took Lucy to the zoo", your sister will
know this means your daughter Lucy and not something else, because
Lucy means something to both of you. Also, you will both probably know
where the zoo is without having to say the one at 9th and Cedar.
When the patterns were "initially" stored, at that point in time they
define the STM context. However, later on, when you recall them in
order to compare them with new incoming data, then the new patterns
are understood in terms of the stored LTM context. Lucy is stored as
daughter in memory, so when you say Lucy, both you and your sister
know who Lucy is, without having to say so directly. We always talk in
terms of stored context [memories]. "I'm going downtown". Where
downtown is depends upon where you live. Stored context.
In order to make any real sense of new incoming pattern data, it has
to be compared against old pattern data stored in LTM. This is what we
do, starting as babies.
Bill Modlin
right. :-)
I don't know about Curt's model, but LTM context would be evoked in mine, as
a recursive pattern completion effect... if I ever got to the point that I
understood what I am doing well enough to actually build it. :-( For now
my theories have too many gaps.
Bill
Curt Welch
> >It wasn't a metaphor actually. It was simplistic to make a very
> >important point. Long before an AI can decide if it's going to go read
> >messages on usenet, or go watch TV, it has to first figure out the
> >simple things in life, - like whether it should raise it's arm, or let
> >it hang. Or whether it should open it's eyes or leave them shut. Or
> >turn it's head to the left, or to the right.
>
> I expect that a human infant has worked out how and when to move its
> arms or open its eyes, long before it has discovered that it has arms
> and eyes.
Yes, that is very true. But I wasn't trying to talk about that.
Sorry, my wording is making my concepts more than a little hard to follow.
It's just so damn hard for me to talk about this stuff without causing
confusion because I like to personify the hardware when I talk about it's
function yet I am not trying to talk about the conscious mind we are trying
to create.
When I said "the AI has to first figure out the simple things of life", I
was talking about the function of the AI hardware has to have some
algorithm which determins if it will lift the arm or not. I'm asking, what
will the function of that hardware be? Why does the baby AI hardware raise
the arm? What software to I write and what does my software base the
decision on wether to lift the arm or not?
> >Honstly, these are first things the AI has to decide. How will it
> >decide what to do?
>
> My point stands. If often doesn't matter whether you raise your
> arms.
Yes, but a real baby does. What should AI do? Why? These are that
questions that lead me to the idea of needing the motivation system.
> >If it doesn't have a way of deciding these most simple of things, how is
> >it ever going to learn later in life to make the complex decisions?
>
> >Are we the creators required to make every decision for it, like "when
> >it should raise it's arm"? If it is intelligent and will later in life
> >be able to make far more complex decision on it's own, why can't it
> >start it's life figuring the arm question out?
>
> >As creators of AI, if we think we know how to create an AI that can make
> >the complex decisions, don't you think it would be trivial for us to
> >know how it's going to make the simple decisions first?
>
> You said in an earlier post, that your AI system would have free
> will. If it will have free will, then you cannot know how it will
> make a decision to raise its arm. That becomes a decision for the
> AI, and one that is not up to you.
But I have to hard-code that free will into the hardware. I have to build
the decision system that guides its actions. If I build a robot which has
an arm under the control of the computer, I have to write software that
determines if the arm moves, or stays still. How do you write this
software, yet give the robot "Free will" at the same time?
The answer comes with the motivation system. You can't answer it any other
way that I've ever seen. This allows me to write software to solve the
"motivation problem". I've still written the software that controls the
arm, but the robot will end up doing things that I could never predict.
And it's not because it's moving "randomly". It's because it learns what
type of arm movements help solve the motivation issues and which don't, and
it uses what's work. And I can't predict what works because it has to
learn that for itself.
This is the only answer I've every seen that seems to fit the requirements
of what intelligent behavior is all about.
> > If on the other
> >hand, we don't understand how it's going to make these decisions, how
> >can we hope to understand how the machine is going to make the complex
> >ones later?
>
> If your system is going to be intelligent and have free will, then
> you cannot know how it will make decisions.
If I can not know how it works, I could never build it, and AI is
impossible. I wouldn't be working on that if I thought that was true.
The truth is that I can know how it works, without knowing what decision it
will make, because I don't have access to its experience. It will make
decisions based on what it's experience tells it works best for it. I can
build hardware to do this (I alaredy have software doing this), yet I will
not have any clue what it will do next (any more so than a human say).
:)
> >Have you ever had a good experience which did not condition you to try
> >and repeat the experience in the future?
>
> Many good experiences are not repeatable. They are singular events.
>
> >> >We do not raise are arm "to survive" or keep it hanging "to survive".
> >> >We do it becuse we have been conditioned to do it based on our
> >> >training. We do it to get a reward, or to avoid a punishment. We do
> >> >it to get food, or prevent pain.
>
> >> How do you account for acts of altruism?
>
> >The easy out is to just call all such acts innate.
>
> >But the truth is, it's trivial to condition people to do anything,
> >including perform acts of altruism. The military is very good at that
> >type of conditioning. Society encourages a lot of of that type of
> >behavior as well.
>
> Conditioning does not adequately explain it, in my view.
When it comes down to, I do not know how far "conditioning" will go. It's
obvious something that clearly works at the low level, and has effects all
the way up into highest levels of our behavior, but can "strong learning"
really create AI? I don't know.
What I feel strong about however is that this is the right foundation to
build AI on. No other foundation I belive can explain intelligent
behavior.
After I have hardware doing as good of a job as I can at "leaning behavior"
though conditioning, I will then see just how "smart" this thing seems to
be and how dumb it seems to be, and I will then try to tell, like with all
other AI approaches, is this the right path to AI or not? Does this show
signs of intelligtence or is it just another AI toy which something
important still missing?
It's a big leap from learning not to stick your hand in a fire because it
hurts all the way up to full intelligent and creative behavior. I of
course think this is the correct path to get there. It does not surprise
me that others think the path is wrong, or "not enough" to explain
intelligence.
> >> >This points us to a simple and obvious way to give AI the purpose it
> >> >needs, in a way that the machine can deal with. We just give it a
> >> >reward and punishment system that tells it when something "bad" is
> >> >happening and when something "good" is happening.
>
> >> When designing the AI appliance, how will you figure out what to
> >> count as good and what to count as bad in your programming?
> >> And how would that be different from what you were criticizing
> >> above when you wrote:
>
> >> But they know "what to do", because we tell them. We build
> >> their behvior into them. We specify what they do, and when,
> >> and how. And this clearly isn't AI.
>
> >When my AI Module is "punished", I am not telling it what it did wrong.
>
> Sure you are.
I don't think you understand what these networks must do with "punishement"
and "reward" data to learn correctly. It's a statistical correlation
problem. A single reward or punishment tells the AI almost nothing. It's
only after millions of reward and punishment signals that the networks
learn what is really "good" and what is "bad".
> >> >This of course is not a simple problem. It's like the chess game
> >> >problem of "what is the best move now?". The better you can predict
> >> >the outcome of the entire rest of the game, the better you can answer
> >> >that question.
>
> >> In the case of chess, prediction is based on the rules of the game.
> >> If you want the analogy to be relevant, it would seem that you would
> >> have to preprogram your AI system with the rules of the game of
> >> minimizing punishment and maximizing reward. In other words, you
> >> would have to pre-program intensive knowledge. Such a view is
> >> consistent with the No Free Lunch theorems, but it seems inconsistent
> >> with your goal of a learning system.
>
> >Why can't it learn the "rules"?
>
> The rules are too complex.
I think these type of algorithms have more powers to understand "rules"
than you reallize.
> >Chess programs predict future "reward and punishment" by performing a
> >huge search very quickly. We don't. We predict future "reward and
> >punishment" in a chess game by heuristics which we learn by plaing many
> >games.
>
> That's questionable. I think it would be more accurate to say that
> we describe it as "heuristics" to avoid having to admit that we don't
> know how it is done.
>
> >Do you not think that a chess player just "knows" how dangerous a
> >position is by looking at it to a degree that normally far excceds a
> >normal chess-playing computers ability?
>
> Sure. But a computer does not look at it at all. A computer has no
> perception, and no ability to look.
>
> > Do you not see this as his brain using
> >this same strong ability to predict future reward and punishment based
> >on past experience (i.e., all the games the guy has won and lost in the
> >past).
>
> I don't believe that is a correct explanation.
>
> >> >The only tool we have for predicting the future is to study the past.
> >> >And this is where the "find patterns" ideas comes into play.
>
> >> The past is a poor predictor of the future.
>
> >So what do you suggest we use? It's all we have got.
>
> We do not need to predict the future. We need only to direct our own
> actions as best we can.
If you can believe that, you clearly look at "intelligence" every
differently than I do. It's unlikely you will ever see any hope in my
approach if that is how you think about intelligence and behavior.
If you don't understand how central the concept of "predicting the future"
is to everything we do, you have no hope of building AI or understand work
such as mine.
I do appreciate you taking the time to debate these points with and to try
and understand what I'm doing.
Neil W Rickert
>Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bfh4a5$8a2$1...@husk.cso.niu.edu>...
>>That our life is chock full of such decisions -- which fork in the
>>road to take, when you have no particular destination -- is the basis
>>for our ideas about having free will.
>Agree. But fair to ask which fork has the highest chance of success.
>In many cases, that can be tested, so one generalizes to cases which
>can't be tested. Obviously some things are quite random and
>unpredictable, others are quite predictable, with all things in
>between. Elsewhere in this thread we're beating that to death.
Keep in mind that 20 or so years back, many people chose to work in
the steel mills because of the highest chance of success. Most of
them fell on hard times, while other parts of the economy prospered.
>>Once you have made your choice -- say for computer science, you can
>>then invent arguments that have the apparent form of logic, in order
>>to explain your choice. But such arguments are rationalizations.
>>They don't accurately account for how you made the decision.
>Agree. I've said/thought it 100 times. You experience (science or
>whatever), and then you find the logic. I also claim that proving an
>assertion in which you are invested simply doesn't work due to the
>investment in the result. You will always find the logic and it will
>always be airtight (in your mind). Then you post it and guess what?
>Some jerk disagrees!
Most of the time, disagreements over logic are really disagreements
over the premises assumed in the logic argument.
It seems to follow that if you want an AI system, it is not
sufficient to have one that does good logic. You also need a system
with the ability to generate good premises.
>My analytic philosophy has always been "Define the problem. Everything
>else accounts for 2%." I'm trying to define the problem of how to be
>more analyticitistical. (sarcasm)
That's about right. The logic is what comes after you define
the problem. The intelligence is what is used to define the
problem in the first place.
>Logic works. Humans just aren't very good at accuracy. But it gives
>you the list of things to try, including the list of things to try to
>do it better. High on that list is AI, specifically AI reasoning.
Logic is fine for what it does. The problem isn't the logic -- it's
the people who mischaracterize what they are doing.
Curt Welch
> the ability of intelligent
> life to put the information it extracts from the environment into the
> context of its past experiences or some inherent knowledge. [note - it
> doesn't necessairly need to do this consciously, so consciousness is a
> different issue].
To me, this is somewhat backwards, because I think it's impossible to
extract "information" without the contex. It doesn't extract and then
"analize in context". The network it builds is the context which has been
created through experience. The network which defines the data that is
extracted is the context.
If you send raw pixel data in, and out comes "fox", this happens because
the network is the context which is used to transform the raw pixel data
into "fox".
It's impossible to do pattern extraction without context. The pattern
extraction network or software or alogorthm is the context which defines
the meaning of what comes out.
It makes no sense to talk about "pattern extraction" and "context" as two
separate things. You can not separate them. They are one and the same.
> At a minimum, you need both pattern extraction and context.
I agree with the idea, but I say pattern extraction is context, and context
is pattern extraction, so saying you need "both" is misleading because it
implies you need two things when there is only one.
> The context can come from built-in knowledge, as the case of
> instinctual knowledge in lower animals,
Yes, if there is a hard-wired pattern extraction system, then it is
hard-wired context defintion learned by the processes of evolution from
past experience.
Curt Welch
> In order to make any real sense of new incoming pattern data, it has
> to be compared against old pattern data stored in LTM. This is what we
> do, starting as babies.
Yeah, I agree with that. But I think it happens as the data goes though
the network. I think LTM is the network. STM is just part of the system
which defines the current "context" of the environment. All new data
comming into the network is cobined with the current STM context and out of
it all comes a decision by the hardware about how it's going to change the
STM context.
When you are talking with your sister, part of the current STM context is
the fact you are talking with your sister. When you hear "How did Lucy
like the zoo", the network takes this "Lucy" input, plus the rest of the
current STM context as input, and produces an output which which activates
your STM context of "your dauther Lucy".
So "STM-YourSister + STM-SheSpeaks + NewInput-Lucy passes though your LTM
network and produces the "behavior" of activating STM-DaughterLucy.
This is how my simple inet onet can hope to do all this. It's because
changes to the short term memory is simply another type of "learned
behavior" to my network.
David Longley
>X...@longley.demon.co.uk>...
>
>
>> >Yes, as mentioned on the other thread, you need a context in which to
>> >compare the new pattern - be it an innate or learned context. I think
>> >you will have this is your concept of inet-onet, but this part is
>> >missing from Bill's direct definition in his paper.
>> >
>> >Living animals [like humans] that learn context, learn it from
>> >interactions with their environment at a young age. Babies have the
>> >ability to babble built-in by genetics, but they learn language by
>> >listening to and making noises back at humans, until they get it right
>> >- a process which takes a huge amount of trial-and-error and feedback
>> >over 4-5 years.
>>
>> Glen calls these operants.
>>
>
>invented.
>====================
>
And whatever the universe was doing it was doing it before Einstein was
"invented".
>
>> >
>> >Your inet-onet will learn things from its environment, and then by
>> >playing those back to the environment, and seeing how the environment
>> >responds in return, it will learn if it is doing things correctly.
>> >Real-world learning is really supervised. Just like a child, if
>> >inet-onet were in a vacuum, it would learn ... next to nothing. The
>> >child's initial motivations come from comfort [ie, dry-bottom and
>> >not-pain] as well as food [ie, tastes-good], so not-pain and
>> >tastes-good are no doubt the "contexts" pre-programmed in by genetics,
>> >and later on, learning vectors off of that, one bit at a time.
>>
>> So where's the argument with Glen? Surely it takes little to say "sorry,
>> I've got it now"? He might even reward you with a few nice words and/or
>> the withdrawal of his hostility... <g>
>>
>
>
>I have no argument with G. He comes and argues with me. I have never
>once sought him out myself just to argue. It's just not worth the
>trouble. If you look back at years worth of posts by G, you will find
>he does the same thing with everyone who doesn't say what he wants to
>hear - and then he gets pissed and starts namecalling <-- this is his
>trademark on many forums.
>
He is fairly representing a much misunderstood and falsely represented
line of work which is 1) very useful and 2) does have promise. He'd be
rather irresponsible NOT to fight his corner.
You *have* got a number of things wrong Dan, and it's important to get
them right. You can then reject the whole approach as not of interest or
not useful to you, or whatever. I'm not asking for more than an accurate
understanding and portrayal from you. I suspect Glen isn't either.
Is that too much to ask?
>I read very few of his posts, but of those I have read, I have seen
>nothing new - except for a change in "terminology".
>
>Sorry, but the stuff I wrote above was not invented by G. He has NO
>claim to it. "I've got it now", and I've had it a LONG LONG time too,
>and it has nothing whatsoever to do with G.
It's not got much to do with ownership. Science doesn't work like that,
you KNOW that. It's a matter of pursuit of true statements. We *do* need
to change how we talk about things, hopefully some of my recent posts
will, in the end, persuade you of this.
--
David Longley
Pierre-Normand Houle
> >> At one time, I held a view very similar to what you present.
> >> However, I have come to recognize that view as faulty.
>
> >> Your view is that the world is a highly patterned place, and intelligence
> >> involves the discovery of these patterns.
>
> >> What I have concluded, is that the world is a disorderly place
> >> without clear patterns or recurrent events.
>
> >Certainly there is plenty of real or apparant disorder to be found... even
> >ignoring quantum indeterminancy, if we measure any quantity precisely enough
> >we find it varying randomly by small amounts from measurement to
> >measurement, a result of what has come to be called thermal noise.
>
> >But if we ignore variations below the noise level we can often get
> >measurements that are reasonably well behaved and non-random in quite
> >definite ways.
>
> You are falling for the illusion.
>
> Measurements are not any part of the natural world. Measurements are
> the application of an invention (the measuring system). The
> invention of suitable measuring systems part of how we organize the
> world.
>
> >So far as recurrent events are concerned, we get to define what we wish to
> >treat as a recurrence.
>
> Quite. Events are, in effect, our invention. A recurrent event is
> really two different things that we have chosen to consider a
> repetition of the one event. This is part of how we organize the
> world.
I don't understand this. Events to be recognized as such must be conceptualized in
some manner and the concepts, no doubt, are our inventions. The concepts we operate
with reflect our needs, our interests, the contingent feature of our perceptual
organs, etc. But does it follow that the events themselves are our inventions? Why
can't we say that things or events in the world can exhibit orderly patterns when
they show up for us through the exercise of our conceptual capabilities?
> > No matter
> >how I choose to organize random data it will remain random: I can discover a
> >recurring predictable pattern only to the extent that it exists in the data.
>
> Data are not a natural part of the world. They are human artifacts.
>
> If you were to put a simple sensor somewhere -- say a light sensor,
> or an air pressure sensor -- and monitor the signals it receives,
> they would not be highly patterned. There may be some weak statistical
> trends, but nothing strong enough that you would call a pattern.
>
> Data results from organized ways of dealing with the world. It does
> not exist without organization.
It does not seem that what you say contradicts Modlin's claim. He says that there is
a pattern in the data and that this pattern is not of our making. What you say is
that the pattern can not show up for us unless we have developed an suitable sensor
to extract genuine data from what are otherwise meaningless signals. I understand
your grounds for disagreeing with Modlin that the starting point for "organizing" the
data consists in establishing correlations within meaningless signals. But I don't
understand your apparent issue with the end point -- when an organism is equipped
with some working perceptual system (after it has developed a way to organize the
data) -- why couldn't it be said to be able to detect objective patterns in the data
he is then enabled to collect?
David Longley
STM, LTM, SEMANTIC, EPISODIC, SPATIAL, ICONIC and it goes on and on -
are all theoretical concepts which have become very popular because we
have computer systems which have "memories". But as Glen and I have
pointed out, and as Ashby and others since has said, it may not be this
way. We resort to models of memory when we describe a system from a
perspective which gives us a limited window on all of the systems states
and transitions. The promise of Cognitivism has always been that the
models it comes up with allow better prediction and control of the
output (R) given a specific input (S) - what Glen and I have been saying
is that it's not really all that clear that this has happened - in fact,
what seems to have happened is very much what Skinner warned us would
happen -
The Brooks approach, provided he keeps a good eye on what
neurobiologists mapping real nets of simple creatures (like insects),
seems worthy of attention - although I wonder to what extent he may
just end up playing second fiddle as the modelling technology in
neuroscience advances.......(as I have said elsewhere).
--
David Longley
Glen M. Sizemore
DM: If you look back at years worth of posts by G,
you will find he does the same thing with everyone who
doesn't say what he wants to hear - and then he gets
pissed and starts namecalling <-- this is his trademark
on many forums.
GS: This is a lie. The way things usually go is something
like this:
Someone will say something like, "We can be
interested in subjective experience, unlike the
behaviorists who ignored it."
And I will say something like, "Sorry, but you are
wrong. Radical behaviorism is defined by its stance on
‘subjective experience' which is, briefly, that what is
‘felt' when one introspects is a one's own behavior,
and we come to be aware of this behavior in the same
fashion in which we come to be aware of the rest of the
world."
The reply to this can take a variety of forms and is
frequently insulting right off the bat. Sometimes it takes
a post or two. Eventually the other person's reply will
be something like, "Any idiot knows that behaviorism
cannot account for complex behavior. Perhaps you
need to take an intro. psych course and find out what
someone who THINKS says about psychology..." and
so on.
I'm supposed to take shit like that? I don't think so. I
suggest you take a look back at some of your posts of
a couple weeks ago. Pretty insulting stuff there, Danny.
And as to that being "my" trademark, I suggest that you
read posts by Rickert, Eray, and a host of others.....oh
and did I mention your posts?
d...@oricomtech.com (dan michaels) wrote in message
David Longley
<mod...@metrocast.net> writes
know).
--
David Longley
rick++
even though A.I. researchers tend to work on just one or two at a time.
dan michaels
> >> So where's the argument with Glen? Surely it takes little to say "sorry,
> >> I've got it now"? He might even reward you with a few nice words and/or
> >> the withdrawal of his hostility... <g>
> >>
> >
> >
> >I have no argument with G. He comes and argues with me. I have never
> >once sought him out myself just to argue. It's just not worth the
> >trouble. If you look back at years worth of posts by G, you will find
> >he does the same thing with everyone who doesn't say what he wants to
> >hear - and then he gets pissed and starts namecalling <-- this is his
> >trademark on many forums.
> >
>
> He is fairly representing a much misunderstood and falsely represented
> line of work which is 1) very useful and 2) does have promise. He'd be
> rather irresponsible NOT to fight his corner.
>
> You *have* got a number of things wrong Dan, and it's important to get
> them right. You can then reject the whole approach as not of interest or
> not useful to you, or whatever. I'm not asking for more than an accurate
> understanding and portrayal from you. I suspect Glen isn't either.
>
> Is that too much to ask?
>
OK .... Glen is an abusive person. He has a long history of this type
of behavior on the internet with many people. There is no reason for
anyone to have to deal with someone like that - in real life or in
cyberspace.
"Choice is the luxury of the free" - Sappho
David Longley
>a...@longley.demon.co.uk>...
But, and this really is not an excuse - there comes a point where
misunderstanding and misrepresentation just becomes insulting, even
though the people doing the former don't necessarily see it that way.
I'm sure you can think of examples where you are 100% sure of something
you are familiar with, say the spec sheet on a particular chip. There
comes a point where after you have explained and documented over and
over again, you just want to SHOUT in the hope it will shock the
recipient into seeing that what they are doing is wrong.
I am not suggesting for a second that you become a radical behaviorist.
There may be many reasons why you'd prefer not to - e.g. it might
require a lot of time and re-thinking, none of which you might feel has
any real payoff for you. Fine, that's reasonable. But all I think is
being asked of you is to see it accurately from the radical
behaviourists perspective, even if it's just an overview. Curt's done
that admirably well. I'm sure he'd be the first to say that there may be
lots to learn but he'll probably say he'll learn that if and when he
needs to.
You don't need to work with rats or pigeons to appreciate the
principles, you can see it from a robotics perspective and I'm sure
there will be literature out there apart from Curt's posts which may be
of interest to you given what I've seen of your posts to date.
--
David Longley
dan michaels
Hi Curt, in reply to both Bill and your esteemed self, yes, we humans
would suffer different kinds of cognitive deficits if missing either
STM or LTM. STM allows us to see the consistencies [or changes] in the
environment from second to second, so we can deal with them
effectively. LTM allows us to put everything into the perspective of
our past experience, and among other things, helps us not to make the
same mistakes too many times in a row :). A good AI should have both,
if it's gonna deal with the real world.
I can see how both can be part of [or make use of] the same mechanisms
- at least from the sensory end of things. You continually abstract
incoming data in real-time, and compare it with what came just
beforehand. STM. Then, automatically, this new abstraction also
elicits stored LTM of similar content, and compares against that, to
see how the new data fits into the experiences of your life in
general.
Personally, I see a whole chain of modules here, as the easiest way to
deal with all of the goings on - input processing, abstraction, STM
comparisons, LTM comparisons, actions, etc - but this is moot, and
exactly where to put STM with respect to LTM is an open question. I
guess it's conceivably that STM can go either before or after the
abstraction level, while LTM will almost certainly be after the
abstraction level, but .... ??
The abstractions I am thinking of here pull out the salient features
of the incoming data, for the purposes of generalization. This way, as
we've discussed previously, you don't need to take a snapshot
representation of everything you experience, but rather you can put
things into categories. LTM almost certainly works on abstracted data,
so it can generalize. However, whether STM might work on the new data
directly, or on abstracted data, is an open question. Since everything
is being done automatically, STM can well be downstream too.
In the brain there is massive feedback from higher centers to lower
centers everywhere along the way - and it's a little unclear how
feedback would work from a higher center holding abstracted data back
to a center earlier in the processing chain prior to where the
abstraction occurs, but something like this does appear to be taking
place. One can only wonder .... ??????????
- dan
=================
dan michaels
> STM, LTM, SEMANTIC, EPISODIC, SPATIAL, ICONIC and it goes on and on -
> are all theoretical concepts which have become very popular because we
> have computer systems which have "memories".
..............
I suppose a "complete" model would have to include all of those. Maybe
one day ....
===============
> The Brooks approach, provided he keeps a good eye on what
> neurobiologists mapping real nets of simple creatures (like insects),
> seems worthy of attention - although I wonder to what extent he may
> just end up playing second fiddle as the modelling technology in
> neuroscience advances.......(as I have said elsewhere).
Since Brooks' basic model has ZERO memory, other than built-in
hard-wired reflexes - you'll have to decide for yourself where it fits
into the overall fray.
Bill Modlin
"Daryl McCullough" <da...@atc-nycorp.com> wrote in message
news:bfhsl...@drn.newsguy.com...
> Bill Modlin says...
>
> >>>> So intelligence is the ability to find patterns.
>
> I think that that's in the right ballpark, but it isn't
> completely right.
>
> First of all, in humans we can distinguish between *automatic*
> pattern-recognition and patterns that require hard work to
> recognize.
>
> We recognize visual patterns (faces, circles, lines, etc.)
> automatically (pre-consciously). On the other hand, if we
> are given a sequence of numbers (perhaps the outcomes of
> experiments) and we have to figure out the pattern, it is
> a lot of hard work. Basically, what we have to do in the
> latter case is to guess a rule, and then generate a sequence
> according to that rule and see if the sequences match.
>
> I think that the automatic, "easy" pattern recognition
> (recognizing that the person standing in the corner is your old
> friend Neil) works so well because it is (at least partly)
> hardwired into our brains. I believe that there are definite
> limitations to what we can come to recognize automatically.
>
> On the other hand, the non-automatic, "hard" pattern recognition
> is completely open-ended, but it is also much harder, and much
> less reliable. Some perfectly intelligent humans may never come
> to suspect that a sequence of numbers was generated using the
> integer part of the Riemann zeta function (or whatever).
>
Hi Daryl.
Some good points.
Before trying to respond, let me introduce a quibble. You say "pattern
recognition", while I'm talking about finding or discovering pattern
classes.
You may mean the same thing, but in many contexts "pattern recognition"
is taken as recognition of a set of variables as belonging to some already
known class of patterns. That's subtly different from discovering a new
class of patterns. Of course once the new class is discovered one must
be prepared to recognize instances of it... anyway, I hope there is no
actual disagreement or misunderstanding reflected in your choice of words.
I am not much interested in what you call "hard" patterns, I think they have
more to do with education than intelligence. Indirectly, more intelligence
might make it easier, but I don't think the rule generate-and-test algorithm
is a core ingredient in intelligence.
As to the automatic recognition we find easy: true, it is easy because
these are the patterns that our brains have evolved to recognize, and so in
a sense the ability to recognize them is "hardwired".
However, I'd like to distinguish two kinds of hardwiring.
There may be hardwired network specialization to provide improved
discrimination of patterns following some specific templates. There are
almost certainly innate "feature detectors" honed for recognition of
particular low level patterns, and we probably have an innate tendency to
recognize and discriminate finely among human faces. There could be others.
I don't consider this sort of thing crucial to general intelligence.
But there is a very broad class of patterns which can be discovered by the
sort of networks we have in our brains with individual cells running in
unsupervised auto-associative mode, where "the sort of networks we have in
our brains" corresponds to a loose set of constraints on interconnectivity
and local interaction through intermediary cells and chemical gradients.
It is this broad class of discoverable patterns that we can learn to
recognize automatically and which we use to partition experience into
discriminable elements. And it is the innate ability to build deep
structures of such patterns that I identify with intelligence.
> When we claim that humans are so much better than other
> beings at pattern recognition, we need to worry about whether
> we are being almost tautological: humans are better at recognizing
> those patterns that humans consider important (or interesting).
> We aren't necessarily better at recognizing *all* patterns.
Talk of *all* patterns is scary... it puts us into NFL territory where
nothing can do better than random search. We recognize a very much
constrained subset of all possible patterns.
However I'd say it isn't about what we consider important or interesting,
it is about what sort of patterns can actually be discovered by heuristic
search.
In abstract there can be patterns involving combinations of arbitrarily many
distinct elements scattered over arbitrary separations in space and time.
But combinatorial growth renders exhaustive search of such a huge pattern
space impractical. Pragmatically we are restricted to examination of
second-order statistics, looking for pairwise correlations among elements
available for observation. We cannot realistically examine more than an
insignificant percentage of cases involving higher-order statistics.
Even with the limitation to pair-wise statistical criteria, if we consider
a domain containing millions of individual elements there are too many pairs
for practical examination of all of them individually.
If we are to find patterns involving relationships between combinations of
dozens or thousands or millions of elements spanning a lifetime of
experience, neither we nor any other intelligence could look for them
directly. No physical system even with universe-wide resources and the
life of the universe to work with can scratch the surface of such a number
of combinations.
If such patterns are to be discoverable, there has to be a way to work up to
them in stages, each stage involving no more than pair-wise statistics and a
practical number of individual elements. That "practical number" depends
on the resources available to the searching system, but for either brains or
computers it is small compared to the total number of potentially-correlated
signals in the brain. Any practical physical implementation of
intelligence is limited to construction out of "building blocks" able to
find only small patterns ... patterns recognizable in the pairwise
statistical behavior of some moderate number of elements at a time.
We can cascade the process to find larger patterns using the
first-recognized patterns as elements, and repeat this as many times as
desired.
But this allows us to find only a limited class of patterns: those
decomposable into smaller independently discoverable patterns. We cannot
discover just any arbitrary high-order pattern. Out of all the uncountably
many possible complex patterns we can only discover a vanishingly small
percentage, those which are decomposable in such a fashion.
Since this is all we or any conceivable physical implementation of
intelligence can manage, it has to be sufficient. It is all we have to work
with. It is enough, at least, to give us the intelligence we do have.
Oh. Forgot to explain that word "heuristic" and relate it to the above.
The point is that for a decomposable pattern, the recognizable structure at
each level is related to the structure at the next higher level: it
provides the clues necessary to recognize the next level. This is not a
characteristic of patterns in general, it is a defining characteristic of
decomposable patterns that we can use the structure at one level as a
heuristic guide to finding the next level.
> What makes a pattern important? A pattern is
> important or not depending on what we intend to do with it --- if
> recognizing the pattern serves our purposes, then it is important,
> and otherwise it is trivia. When we judge certain phenomena to be
> "noise" rather than "signal" it doesn't really mean that it lacks
> a pattern---it just means that we believe that the pattern is
> irrelevant for us. A super-pattern-recognizer who goes around
> discovering faces in clouds or prime numbers in social security
> numbers would not be considered intelligent, he would be considered
> an eccentric or a crackpot.
For the background automatic processes, I don't think "importance" or
"purpose" enter into it. As our cells discover new patterns they are
incorporated into our model of the world. Period. The conscious result is
that we understand something better than we did before, we "see" or "know"
how some discriminable elements of experience connect to some others. If
we were not focused on that part of the model we probably don't even notice
that anything has happened.
The super-pattern-recognizer in the last sentence is engaging in spurious
pattern matching, not pattern-class discovery, when he finds a face in
the clouds. I don't know how finding a prime number is related to pattern
recognition or discovery.
> To me, the most difficult part of human intelligence to mimick
> by a computer is judgement: judgement about what avenues are
> worth exploring and which ones are likely to be a waste of time.
> There's a chicken-or-egg problem inherent in such decisions: we
> don't really *know* whether something is a waste of time until
> after we've done it, in which case it is too late. It is therefore
> wrong to think of decision-making as something that *follows*
> thinking about alternatives---in fact, we have to make decisions
> about which alternatives to think about, as well.
Well. To me most of this goes on under the covers as a parallel heuristic
exploration of the alternatives... once the automatic processes decide what
we are going to do we may become conscious that we have made a decision.
The most that conscious processes have to do with it is that they can
circulate of stream of words or images or other token-packets to stimulate
automatic activity in sections of the model previously decided to be
important. Eventually a new decision pops into consciousness, often well
after we have acted on it. Sequential conscious thinking is a useful
mechanism for keeping the under-cover analysis focused on a longer chain of
connections than it would normally pursue, but decisions aren't made at the
conscious level... we can only think those thoughts that the network has
decided might get it where it is trying to go.
Hmm.
> --
> Daryl McCullough
> Ithaca, NY
Bill Modlin
Neil W Rickert
>"Neil W Rickert" <ricke...@cs.niu.edu> wrote in message
>news:bffvf6$f3n$1...@husk.cso.niu.edu...
>> >So far as recurrent events are concerned, we get to define what we wish to
>> >treat as a recurrence.
>> Quite. Events are, in effect, our invention. A recurrent event is
>> really two different things that we have chosen to consider a
>> repetition of the one event. This is part of how we organize the
>> world.
>I don't understand this. Events to be recognized as such must be conceptualized in
>some manner and the concepts, no doubt, are our inventions. The concepts we operate
>with reflect our needs, our interests, the contingent feature of our perceptual
>organs, etc. But does it follow that the events themselves are our inventions? Why
>can't we say that things or events in the world can exhibit orderly patterns when
>they show up for us through the exercise of our conceptual capabilities?
No doubt the atoms move around much the same whether we are watching
them or not. But an events are socially constructed in ways that
motion of atoms are not.
Consider:
A: the things that happened between 1914 and 1918
B: the things that happened between 1939 and 1945
C: the things that happened between 1959 and 1965
We think of A as an event (World War I), and of B as
an event (World War II), but we don't consider C
to be an event at all -- perhaps a conglomeration
of lots of little events, but not an event in its
own right.
Or consider the cars driving down my street between 5 pm and 5:30 pm
today. We don't think of that as an event. But suppose that
tommorrow a local TV station will be recording it on camera, and
reporting it on their late news. Then suddenly that becomes an
event, socially constructed by virtue of the decision of the local TV
station to select it out.
>> > No matter
>> >how I choose to organize random data it will remain random: I can discover a
>> >recurring predictable pattern only to the extent that it exists in the data.
>> Data are not a natural part of the world. They are human artifacts.
>> If you were to put a simple sensor somewhere -- say a light sensor,
>> or an air pressure sensor -- and monitor the signals it receives,
>> they would not be highly patterned. There may be some weak statistical
>> trends, but nothing strong enough that you would call a pattern.
>> Data results from organized ways of dealing with the world. It does
>> not exist without organization.
>It does not seem that what you say contradicts Modlin's claim. He says that there is
>a pattern in the data and that this pattern is not of our making. What you say is
>that the pattern can not show up for us unless we have developed an suitable sensor
>to extract genuine data from what are otherwise meaningless signals. I understand
No, that's not what I am saying. I am denying that "genuine data" has any real
meaning, other than as a human construct.
A sensor does not "extract data". There isn't any data for the
sensor to extract.
>your grounds for disagreeing with Modlin that the starting point for "organizing" the
>data consists in establishing correlations within meaningless signals. But I don't
>understand your apparent issue with the end point -- when an organism is equipped
>with some working perceptual system (after it has developed a way to organize the
>data) -- why couldn't it be said to be able to detect objective patterns in the data
>he is then enabled to collect?
I could argue with the details of what you say. But I'll skip that.
More importantly, you have missed my real disagreement with Bill.
There are two things needed:
(a) Some sort of system to collect data in a suitable organized
manner.
(b) Making use of that data in one form or another.
No doubt (b) is important. We use the data to control our activity.
Here is the disagreement. For Bill (and for most philosophers,
psychologists, AI folk, etc), everything important happens
in (b). So for Bill, learning is pattern discovery in (b). What
happens for (a) is just assumed and taken for granted as existing
a priori.
What I am arguing, is that learning, consciousness, free will,
intensionality -- are all about (a).
People who see (b) as where the action is, will for ever be offering
proofs that we have no free will, that learning must be mainly an
illusion or the outcome of a selection process from innate knowledge,
and that consciousness must be no more than an epiphenomenon since it
has no apparent role.
Neil W Rickert
>Neil W Rickert <ricke...@cs.niu.edu> wrote:
>> cu...@kcwc.com (Curt Welch) writes:
>> >It wasn't a metaphor actually. It was simplistic to make a very
>> >important point. Long before an AI can decide if it's going to go read
>> >messages on usenet, or go watch TV, it has to first figure out the
>> >simple things in life, - like whether it should raise it's arm, or let
>> >it hang. Or whether it should open it's eyes or leave them shut. Or
>> >turn it's head to the left, or to the right.
>> I expect that a human infant has worked out how and when to move its
>> arms or open its eyes, long before it has discovered that it has arms
>> and eyes.
>Yes, that is very true. But I wasn't trying to talk about that.
>Sorry, my wording is making my concepts more than a little hard to follow.
>It's just so damn hard for me to talk about this stuff without causing
>confusion because I like to personify the hardware when I talk about it's
>function yet I am not trying to talk about the conscious mind we are trying
>to create.
I'll agree that it is hard to find clear ways of talking about this.
>When I said "the AI has to first figure out the simple things of life", I
>was talking about the function of the AI hardware has to have some
>algorithm which determins if it will lift the arm or not. I'm asking, what
>will the function of that hardware be? Why does the baby AI hardware raise
>the arm? What software to I write and what does my software base the
>decision on wether to lift the arm or not?
My answer to "Why does a human baby raise its arm" would be "because
it wants to". I don't believe that there is any algorithmic answer
to be given. You could look more closely, and see that nothing is
happening that violates any scientific principles. But, I suggest,
you would never identify an algorithmic answer to the question on arm
raising.
For sure, there need to be control mechanisms available. "It raised
the arm because the control mechanism was activated". But if I
understand what you are asking, then that merely moves the question
to "why was the control mechanism activated". I don't believe that
there can be algorithmic answers to such questions, if you want the
AI system to have anything we would consider to be free will.
I do see the problem. You are wanting to design a computation based
AI system. And unless there is an algorithmic answer, the arm never
does get raised. I see that as a limitation of computationalism.
I should add that I am not trying to dissuade you from trying to
find answers.
>> You said in an earlier post, that your AI system would have free
>> will. If it will have free will, then you cannot know how it will
>> make a decision to raise its arm. That becomes a decision for the
>> AI, and one that is not up to you.
>But I have to hard-code that free will into the hardware.
>But I have to hard-code that free will into the hardware. I have to build
>the decision system that guides its actions. If I build a robot which has
>an arm under the control of the computer, I have to write software that
>determines if the arm moves, or stays still. How do you write this
>software, yet give the robot "Free will" at the same time?
Personally, I think it cannot be done. I would want to start with
something other than logic chips (more like biological cells).
But you can look at something simpler. Consider the clock in your
computer or your AI system. Ask the question, why will the clock
emit the next tick?
We can give sort-of answers, that involve the physics. We can
provide something of a bell curve for the probability distribution on
when the next tick will occur. But the actual emitting of the next
tick is not controlled by any algorithm.
[snippage]
>> >When my AI Module is "punished", I am not telling it what it did wrong.
>> Sure you are.
>I don't think you understand what these networks must do with "punishement"
>and "reward" data to learn correctly. It's a statistical correlation
>problem. A single reward or punishment tells the AI almost nothing. It's
>only after millions of reward and punishment signals that the networks
>learn what is really "good" and what is "bad".
Instead of directly programming what is good and what is bad, you
have designed a system that will gradually converge to your idea of
what is good and what is bad. But you are still programming it.
That you have designed to be a very slow learner does not alter the
fact that it is deriving its standards of good and bad from what you
have set as the reward/punishment system.
>> >> >This of course is not a simple problem. It's like the chess game
>> >> >problem of "what is the best move now?". The better you can predict
>> >> >the outcome of the entire rest of the game, the better you can answer
>> >> >that question.
>> >> In the case of chess, prediction is based on the rules of the game.
>> >> If you want the analogy to be relevant, it would seem that you would
>> >> have to preprogram your AI system with the rules of the game of
>> >> minimizing punishment and maximizing reward. In other words, you
>> >> would have to pre-program intensive knowledge. Such a view is
>> >> consistent with the No Free Lunch theorems, but it seems inconsistent
>> >> with your goal of a learning system.
>> >Why can't it learn the "rules"?
>> The rules are too complex.
>I think these type of algorithms have more powers to understand "rules"
>than you reallize.
Wittgenstein has written something about the impossibility of
following a rule. Chomsky's "poverty of stimulus" argument is to the
effect that the rules of language cannot be learned. E. M. Gold,
"Language identification in the limit" (Information and Control,
1967) provides some sort of proof supporting Chomsky's conclusion.
J. Stephen Judd, "Neural network design and the complexity of
learning" shows that the learning problem for neural networks is
computationally intractable (NP complete). Vladimir Vapnik, "The
Nature of Statistical Learning Theory" shows that learning is usually
intractable. The "No Free Lunch" theorems show something similar.
Roughly speaking, the type of rule learning you would need is only
possible if you suitably greased the skids in your design.
>> >> The past is a poor predictor of the future.
>> >So what do you suggest we use? It's all we have got.
>> We do not need to predict the future. We need only to direct our own
>> actions as best we can.
>If you can believe that, you clearly look at "intelligence" every
>differently than I do. It's unlikely you will ever see any hope in my
>approach if that is how you think about intelligence and behavior.
Let's look at two examples:
(a) I predict that it will rain tomorrow.
(b) I write a note "We need some more milk". I predict that if I
look at that note when I am next at the grocery, it will
still say "We need some more milk".
While predictions of type (a) are useful, they are not essential. If we
get them wrong, well c'est la vie. We have to take life as it comes.
Predictions of type (b) are far more important. But what that requires,
is that when we invent technology (such as writing in this case), then
we should make sure that the technology works.
In some sense, intelligence is all about inventing. It may be small
inventions that never make it to the patent office. But they are
inventions none the less.
>If you don't understand how central the concept of "predicting the future"
>is to everything we do, you have no hope of building AI or understand work
>such as mine.
I understand what you are trying to do. Count me as skeptical as to
whether it can succeed.
If I were to try to build an AI system, it would not be computation
based. But at present the needed hardware components are not readily
available, except as parts of biological organisms. And we would
have difficulty assembling them as efficiently as do biological
systems.
gmb
From a freshman's point of view, consciousness is not important
consideration for AI. Software + hardware builds robots that
do what is expected from software and hardware: perform tasks,
autonomous control of lunar vehicles, housecleaning machines,
etc. But once these machines begin communicating besides
standard training mechanism issues, once these machines get
involved with people, then questions about consciousness may
kick in. Even less sophisticated robots could raise eyebrows.
People love chatting with simple psychological chatbots.
It is programmed to respond to simple keywords:
Person: I just broke up with my girlfriend.
Artificial Psychologist: Tell me more about your girlfriend.
Sounds intelligent, right? The program does not interpret
"I just broke up with", it responds simply "my <object>",
provoking further talk.
George
Eray Ozkural exa
> Incidently, it is also my conclusion that computationalism cannot
> work. For it can only make truth-based judgements. It cannot make
> the needed pragmatic judgements. Biological systems, by contrast, are
> able to make pragmatic judgements. Evolution itself is a
> pragmatically based system.
I think this wrong view of the "sufficient" conditions for a mind
stems from an ineffective philosophical understanding of what
computation is.
That is not your fault, but you could recognize that there is nothing
"unbiological" or "biological" about computation.
You think the universe is not an orderly place. Fine. How did you
measure it?
I have no idea what you mean by "pragmatic judgement", but I don't
think it's instrumental in this erroneous view.
I sometimes think I will have to rewrite much of philosophy :((
Regards,
__
Eray
gmb
I took this idea back in the early 90's and built up a complex
chat bot, responding to a gazillion keywords and entertaining
folks. It was a fun project.
George
Eray Ozkural exa
> To me a concept has to be a program of sorts. For example logical
> deduction I hope you will agree is a concept. Yet it can also be used
> as a program (or series of instructions) to create new concepts. And
> since a concept is a program, the creation of a concept is a form of
> programming. So yes a tabula rasa from my perspective is a Strawman
> as it would have no programs that could create other programs.
>
That's what I think, too.
In fact I think it *is* an algorithm, and nothing else.
Programming and debugging are therefore the most basic reflective
thinking skills that lead to consciousness.
> > A small observation: you can't just "implement" a philosophical paper.
> > ;)
>
> True it will need lots of work. But without the correct philosophical
> grounding, you may head off in the wrong direction.
>
Yes, but with unneeded metaphysical baggage you don't get anywhere.
That's why studying ontology and epistemology is almost guaranteed to
waste your time.
> > Let me say something slightly useful about your research. It is well
> > possible that a strong AI can be created by a genetic programming
> > experiment. Computational view of mind does not exclude any particular
> > approach to AI programming.
>
> Do you have any other philosophical ideas about AI, apart from a
> computational view, I would be interested to know.
If you mean "do you subscribe to a view that is wholly apart from the
computational view?", my answer will be no. But if you are asking
about particulars, yes, I have ideas about
i) perception
ii) logical reasoning
iii) reflective thinking
iv) what basic cognitive tasks are
among others. I post here from time to time.
I also have a lot of ideas about where the AI research should be
going. ;)
I'm currently working on a parallel data mining algorithm.
Application-wise a friend of mine and I are working on (profitable!)
stock market prediction and intelligent web search. I wish to extend
my research to new methodologies and techniques in machine learning,
possibly designing a scalable parallel algorithm and a few experiments
:)
> Also I am not sure
> my system would be properly described as GP. Tierra, the sort of
> system that mine is similar to, never got that epithet and with GP
> there is certain set things that people expect such as fixed
> generations, no interaction between programs and a closely controlled
> selection process. However there seems to be no other name apart from
> Tierra-like system, and that was based on a completly different
> philosophy, that of trying to make the individual programs
> intelligent, whereas I am trying to make the system as a whole
The system is a program. However, a change in architecture could make
all the difference. I think we should be actively designing parallel
systems, not outdated serial symbolic engines.
What do you name your system if it isn't GP? I'm not too familiar with
the terminology in that field.
Regards,
__
Eray Ozkural
Glen M. Sizemore
arm" would be "because it wants to".
GS: And God said "Let there be arm movement!"
Neil W Rickert <ricke...@cs.niu.edu> wrote in message
gmb
Why does Bush want Iraq's oil? To do somethiung intelligent
with it. Let there be more US control of oil, and more
money in the mafia's pocket. Why does a baby want to move
his/her arm? To do something intelligent with it. Why does
a monkey want to move its arm? A monkey arm makes a monkey
behave like a monkey.
george
Neil W Rickert
>Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bff1mo$q4j$1...@husk.cso.niu.edu>...
>> Incidently, it is also my conclusion that computationalism cannot
>> work. For it can only make truth-based judgements. It cannot make
>> the needed pragmatic judgements. Biological systems, by contrast, are
>> able to make pragmatic judgements. Evolution itself is a
>> pragmatically based system.
>I think this wrong view of the "sufficient" conditions for a mind
>stems from an ineffective philosophical understanding of what
>computation is.
I expect people to disagree with me. But when you say it is "wrong",
you should provide some sort of cogent argument.
>That is not your fault, but you could recognize that there is nothing
>"unbiological" or "biological" about computation.
The point is not relevant. Biological systems do things other than
computation. I am suggesting that some of those other things are
needed.
>You think the universe is not an orderly place. Fine. How did you
>measure it?
You are seriously missing the point. It is measurement, and our
measurement systems, that make the world appear somewhat orderly to
us.
My primary reason for rejecting computationalism, is that I see
measurement processes as essential components of intelligence.
A Turing machine has no ability to measure anything.
dan michaels
> d...@oricomtech.com (dan michaels) wrote:
> > the ability of intelligent
> > life to put the information it extracts from the environment into the
> > context of its past experiences or some inherent knowledge. [note - it
> > doesn't necessairly need to do this consciously, so consciousness is a
> > different issue].
>
> To me, this is somewhat backwards, because I think it's impossible to
> extract "information" without the contex. It doesn't extract and then
> "analize in context". The network it builds is the context which has been
> created through experience. The network which defines the data that is
> extracted is the context.
Well, I was more or less laying out the various tasks. I think the
classical approach of AI-computer vision uses the process/extract,
then analyze/compare paradigm, but some of the brain modelers think
this sequential approach is wrong, and try to do things in a more
integrated fashion. [like you too :)].
OTOH, it seems fairly clear that the 30+ visual areas in the brain
perform various unitary operations [orientation, edges, color blobs,
etc] on the signals, but it's still unclear how these fit into the
overall scheme of things, esp considering the massive amounts of
interconnections between regions.
==================
>
> If you send raw pixel data in, and out comes "fox", this happens because
> the network is the context which is used to transform the raw pixel data
> into "fox".
>
> It's impossible to do pattern extraction without context. The pattern
> extraction network or software or alogorthm is the context which defines
> the meaning of what comes out.
>
> It makes no sense to talk about "pattern extraction" and "context" as two
> separate things. You can not separate them. They are one and the same.
>
This is basically what Rosenfield was saying in his book, because of
all the interconnections between brain areas, however, in the model he
shows at the end of the book - [which is Edelman's Darwin II
processor] - there are 2 parallel pathways with 2 levels each, and the
first levels do unitary feature extraction, and the 2nd levels do
associations and comparisons, using interconnections between the level
2 regions. So, it's not really lall that integrated.
==================
> > At a minimum, you need both pattern extraction and context.
>
> I agree with the idea, but I say pattern extraction is context, and context
> is pattern extraction, so saying you need "both" is misleading because it
> implies you need two things when there is only one.
>
ok - as noted above, you can apparently attack the problem either way.
======================
> > The context can come from built-in knowledge, as the case of
> > instinctual knowledge in lower animals,
>
> Yes, if there is a hard-wired pattern extraction system, then it is
> hard-wired context defintion learned by the processes of evolution from
> past experience.
My concept of what is going on in the brain - based upon presence of
retina, geniculate, and 30+ visual cortical areas - is that the
functions of the lowest levels are mostly hard-wired [retina,
geniculate], as are basically the 1st few cortical levels [although
there is some developmental plasticity presence in infants], but then
once you get past some [uncertain] cortical level, then that is where
most of the learning takes place.
Related to this is that, as you move up the levels, feature extraction
of many different types occurs in the more or less hard-wired regions.
IE, the signals are abstracted in different ways, and these
abstractions are used by the higher levels, which are less hard-wired
and where most of the learning occurs, and where generalization into
categories occurs. So, in this case, you have about 30+LGN+retina
areas doing the extraction - and another who knows how many areas
during storage and association.
OTOH, because of all the interconnection pathways at every level, it's
certainly not clear how everything fits together in the end. Eg, V4,
the color area, connects directly to ~20 of the other areas of the 30.
Why so much cross-traffic at that level?
dan michaels
> > the ability of intelligent
> > life to put the information it extracts from the environment into the
> > context of its past experiences or some inherent knowledge. [note - it
> > doesn't necessairly need to do this consciously, so consciousness is a
> > different issue].
>
> To me, this is somewhat backwards, because I think it's impossible to
> extract "information" without the contex. It doesn't extract and then
> "analize in context". The network it builds is the context which has been
> created through experience. The network which defines the data that is
> extracted is the context.
Well, I was more or less laying out the various tasks. I think the
classical approach of AI-computer vision uses the process/extract,
then analyze/compare paradigm, but some of the brain modelers think
this sequential approach is wrong, and try to do things in a more
integrated fashion. [like you too :)].
OTOH, it seems fairly clear that the 30+ visual areas in the brain
perform various unitary operations [orientation, edges, color blobs,
etc] on the signals, but it's still unclear how these fit into the
overall scheme of things, esp considering the massive amounts of
interconnections between regions.
==================
>
> If you send raw pixel data in, and out comes "fox", this happens because
> the network is the context which is used to transform the raw pixel data
> into "fox".
>
> It's impossible to do pattern extraction without context. The pattern
> extraction network or software or alogorthm is the context which defines
> the meaning of what comes out.
>
> It makes no sense to talk about "pattern extraction" and "context" as two
> separate things. You can not separate them. They are one and the same.
>
This is basically what Rosenfield was saying in his book, because of
all the interconnections between brain areas, however, in the model he
shows at the end of the book - [which is Edelman's Darwin II
processor] - there are 2 parallel pathways with 2 levels each, and the
first levels do unitary feature extraction, and the 2nd levels do
associations and comparisons, using interconnections between the level
2 regions. So, it's not really lall that integrated.
==================
> > At a minimum, you need both pattern extraction and context.
>
> I agree with the idea, but I say pattern extraction is context, and context
> is pattern extraction, so saying you need "both" is misleading because it
> implies you need two things when there is only one.
>
ok - as noted above, you can apparently attack the problem either way.
======================
> > The context can come from built-in knowledge, as the case of
> > instinctual knowledge in lower animals,
>
> Yes, if there is a hard-wired pattern extraction system, then it is
> hard-wired context defintion learned by the processes of evolution from
> past experience.
Peter F
> In the brain there is massive feedback from higher centers to lower
> centers everywhere along the way - and it's a little unclear how
> feedback would work from a higher center holding abstracted data back
> to a center earlier in the processing chain prior to where the
> abstraction occurs, but something like this does appear to be taking
> place. One can only wonder .... ??????????
It seems to me that AI designers don't have to worry about building-in (from
the bottom up) any "selectively consciousness canceling" mechanisms other
than some very basic, electrical overload averting, safety switch.
In contrast, such selectively consciousness canceling situations are
invaluable to biological individuals evolving in biological environments,
since such enviroments are replete with "selective Hibernation" imploring
type situations (feel free to acronymize of this subcategory of
life-situational and naturally selective "adversity (super-) type"
evolutionary pressures %-}) that, to make matters worse, normally cause (or
usually inevitably tend to "condition in") a type (kind or category) of
memories that can suitably be called "CURSES".
----
By the way -- here is a patch of my (EPT) philosophy, thrown in:
An individual's brain, and its individual 'consciousness' ["''" since I mean
it both more and less simply than people normally do] may be 'metaphorically
seen' as a corrupt socialist democracy with voting/competitions going on all
the time between unequally important (vote-weighted) mutually incompatible
policy proposals/proposers. Closer to reality (or somewhat less
metaphorically), these "policy proposals" are neurophysioanatomically
concrete potentials, or "actention (sensorimotor/behaviour) modules", and
they are "vote-weighted" via phylogeny and by prior and current
environmental influences; and they tend to compete (partly by means of the
general principle of lateral inhibition (or center/surround
excitation/inhibition, as per Paul Bush's brain page
http://www.keck.ucsf.edu/~paul/brain.html) towards a transient predominance
within the a brain's repertoire of "actention modules" (each of these is
physically available to become transient "focuses of actention" at the
active exclusion of other actention modules, depending on the individual's
"total situations" - a very wide concept). The actention modules that
miss-out in this ongoing, environmentally "cheered/booed-on", competition do
however as if 'actively abide' in subliminal (or transiently sublimated)
states within the "actention selecting (both brain and body) system".
----
Cheers,
P
Acme Debugging
> cu...@kcwc.com (Curt Welch) writes:
> >Neil W Rickert <ricke...@cs.niu.edu> wrote:
> >> cu...@kcwc.com (Curt Welch) writes:
>
important. I and others have been thinking, accumulating, on the broad
subject for quite some time, mentioned in numerous posts in numerous
ways, and I think you are well-informed in this area. This has to do
with the difference between what can be scientifically predicted and
what brains can predict but on a subconsious level (not "informed
wisdom") with the ultimate goal of gaining insight to the problem of
generalizing "rules" (I call it "logic") of various problem solving.
Introducing "free will" and "real-world complexity" in terms of Curt's
project seem to lead to interesting places.
I wonder if you would clarify a couple of points so I can understand
your thinking better. I've edited to the bare minimum of what I see as
relevant to the broad issue. Apologies in advance if you find that
troubling. Feel free to put anything back.
I don't mean to exclude Curt. I pretty much understand his points.
>
>My answer to "Why does a human baby raise its arm" would be "because
>it wants to". I don't believe that there is any algorithmic answer
>to be given. You could look more closely, and see that nothing is
>happening that violates any scientific principles. But, I suggest,
>you would never identify an algorithmic answer to the question on arm
>raising [...edit raising mechanics...] if you want the AI system to
have anything
>we would consider to be free will.
>
>>> You said in an earlier post, that your AI system would have free
>>> will. If it will have free will, then you cannot know how it will
>>> make a decision to raise its arm. That becomes a decision for the
>>> AI, and one that is not up to you.
>
>>But I have to hard-code that free will into the hardware. I have to
build
>>the decision system that guides its actions. If I build a robot
which has
>>an arm under the control of the computer, I have to write software
that
>>determines if the arm moves, or stays still. How do you write this
>>software, yet give the robot "Free will" at the same time?
Need some time to think about this. Like where might
probably-not-entirely randomness come from in a computational model in
context of free will. It's interesting.
>
><snip>
>
>>> >> In the case of chess, prediction is based on the rules of the
game.
>>> >> If you want the analogy to be relevant, it would seem that you
would
>>> >> have to preprogram your AI system with the rules of the game of
>>> >> minimizing punishment and maximizing reward. In other words,
you
>>> >> would have to pre-program intensive knowledge. Such a view is
>>> >> consistent with the No Free Lunch theorems, but it seems
inconsistent
>>> >> with your goal of a learning system.
>
>>> >Why can't it learn the "rules"?
>
>>> The rules are too complex.
>
>>I think these type of algorithms have more powers to understand
"rules"
>>than you reallize.
An example of real-world complexity would be great here. Something
where humans seem to be able to come up with some meaningful
probability that science can't, but subconsciously. I am reluctant to
repeat my ET example yet again. I think we're looking for something
really elusive and possibly residual, so examples aren't that easy.
But one with empirical tests, not ESP, philosophy, etc. Maybe the
question "How does a dog tell the difference between someone tripping
over it and someone kicking it?" helps to find an example. Mark Twain
said, "When you pick up a cat by the tail you learn something you
can't learn in any other way." Hey, nothing worthwhile is easy. The ET
example took years...
>
>Roughly speaking, the type of rule learning you would need is only
>possible if you suitably greased the skids in your design.
>
I don't understand the "greased the skids" analogy. Is this the
difference between the theoretical probability of a coin-toss v. real
world measurement of it? Like "grease" the coin until neither side has
unique attributes?
>>> >> The past is a poor predictor of the future.
Do you mean the past is "sometimes" a poor predictor of the future?
(Plenty of real-world, real-time statistical tests with significant
correlation, from gambling to all kinds of business stuff,
achievement/aptitude tests, not to mention simple physical experiments
like bullet trajectories.)
>>> We do not need to predict the future. We need only to direct our
own
>>> actions as best we can.
>
Is this the same as "We need to guess the likelihood of future
events?" What's the "direct our actions" distinction? (If one
predicted that pilot training would improve the odds of avoiding a
future crash, would that not be confirmable for aggregates of pilots?)
>
>>If you don't understand how central the concept of "predicting the
future"
>>is to everything we do, you have no hope of building AI or
understand work
>>such as mine.
I think Neil understands it, but is trying to make an important point
about it. (And it would really be nice if we could keep away from
butterflies affecting weather or chaotic v. orderly universe. At least
I'm not ready to resolve *that* issue today!)
>
>If I were to try to build an AI system, it would not be computation
>based. But at present the needed hardware components are not readily
>available, except as parts of biological organisms. And we would
>have difficulty assembling them as efficiently as do biological
>systems.
That makes a lot of sense, I've posted as much. But if the brain has
some extra facility for estimating the future that scientific
prediction does not have (and I don't necessarily mean better overall,
but just any facility), we could try to list some attributes. Until
then, I don't think we can decide if it is computational or not, or
even whether it exists. If it turns out to be some mechanism for
calculating logic/probabilty, then computation is supported. But if
there is some component of unconscious perception, or conscious
perception with unrecognized significance, then it may not be
computational. Etc.
Bonus quote: Einstein said examples are not one way of teaching, they
are the only way of teaching. You're Einstein and Larry is your dumb
student. I also think that would be good for learning about this
problem and help keep it empirical.
Larry
boris.k
It's nice to see some meaningful definitions on this ng for a change!
But I think it's not quite right: the purpose of intelligence is
prediction. It IS done by projecting patterns, but assymmetrically:
the value of a pattern is its projection in the future, not in the
past. Talking about value, I think pattern can only be defined as a
record that compresses a set of inputs by replacing them (complexity
theory), & value is the degree of compression. You need to quantify
things.:)
Boris.
David Longley
But you try telling Rogerian Client-Centred Counsellors that what they
are doing (all due reference to ELIZA) that what they are doing is
fundamentally behavioural, allowing the client to hear what they are
saying a bit better.... like so many folk in this newsgroup they think
there is something more, "understanding". They fail to see that the very
language they use is a public system of behaviour.
Almost the defining feature of "cognition" is that it "goes beyond the
information given" (see Bruner 1957). This is often cited as a good
thing, and it *is* what psychologists spend a lot of time doing, with
some justification as their province is "psychology". But science of the
real world includes behaviour not just the theories which folk come up
with to make sense of the world... and folk psychology is very probably,
just such a theory.
--
David Longley
David Longley
Ozkural exa <er...@bilkent.edu.tr> writes
>I sometimes think I will have to rewrite much of philosophy :((
>
Or re-evaluate your understanding of what you have read at least ;-/
--
David Longley
dan michaels
>
> >Neil W Rickert <ricke...@cs.niu.edu> wrote:
> >> cu...@kcwc.com (Curt Welch) writes:
>
> >> >It wasn't a metaphor actually. It was simplistic to make a very
> >> >important point. Long before an AI can decide if it's going to go read
> >> >messages on usenet, or go watch TV, it has to first figure out the
> >> >simple things in life, - like whether it should raise it's arm, or let
> >> >it hang. Or whether it should open it's eyes or leave them shut. Or
> >> >turn it's head to the left, or to the right.
>
> >> I expect that a human infant has worked out how and when to move its
> >> arms or open its eyes, long before it has discovered that it has arms
> >> and eyes.
>
> >When I said "the AI has to first figure out the simple things of life", I
> >was talking about the function of the AI hardware has to have some
> >algorithm which determins if it will lift the arm or not. I'm asking, what
> >will the function of that hardware be? Why does the baby AI hardware raise
> >the arm? What software to I write and what does my software base the
> >decision on wether to lift the arm or not?
>
> My answer to "Why does a human baby raise its arm" would be "because
> it wants to". I don't believe that there is any algorithmic answer
> to be given. You could look more closely, and see that nothing is
> happening that violates any scientific principles. But, I suggest,
> you would never identify an algorithmic answer to the question on arm
> raising.
>
How about ... "because it has to" ... [if it ever wants to find out
who and what it is ... see below].
=================
> For sure, there need to be control mechanisms available. "It raised
> the arm because the control mechanism was activated". But if I
> understand what you are asking, then that merely moves the question
> to "why was the control mechanism activated". I don't believe that
> there can be algorithmic answers to such questions, if you want the
> AI system to have anything we would consider to be free will.
>
> I do see the problem. You are wanting to design a computation based
> AI system. And unless there is an algorithmic answer, the arm never
> does get raised. I see that as a limitation of computationalism.
>
How about ....
1 - raising its arm is the first step on the road to intelligence, and
2 - the act of arm raising should be hard-wired to occur immediately
upon powering up the naiive AI for the first time. [or soon
thereafter]
===================
> I should add that I am not trying to dissuade you from trying to
> find answers.
>
> >> You said in an earlier post, that your AI system would have free
> >> will. If it will have free will, then you cannot know how it will
> >> make a decision to raise its arm. That becomes a decision for the
> >> AI, and one that is not up to you.
>
Ah ha, good questions here ... but the designer should make the
decision initially .... because:
Due to some of these threads, I have started reading up on infant
development during the first 3-4 years. Haven't gotten very far along
yet, but I already have some opinions ;-).
And the first one is .... when it somes to training a naiive AI, maybe
one should investigate early development in infants and use them as a
role-model.
The answer as to why a baby raises its arms turns out to be very
simple. Reflex. At birth, the baby has many reflexes built-in by
genetics. I guess these are pre-wired in order to get things going -
so the brain doesn't have to make any real decisions before its ready.
Built-in are how to suck, how to breathe, basic movements, etc. As
soon as it pops out, those little wigglies start wiggling. It has no
choice in the matter. It also has a very strong grasping reflex. Touch
it's palm with your finger, and it'll latch on.
Now this may all sound pretty innocuous, but I don't think so. For one
thing, the baby has to get all of its pathways wired up and the
various systems coordinated. I imagine that practice in using the
pathways has a lot to do with this. Practice, practice, practice.
More importantly, and I don't have a good feel for this mechanism yet,
one of the major things the baby has to accomplish during its first
year is to learn the difference between "me" and "not-me". Initially,
its brain doesn't know what a "me" is - or anything else for that
matter. It learns a lot from visual and auditory input, but I have a
hunch that the major way it distinquishes the difference between me
and not-me is tactile, by using its hands and feet, and touching and
grasping and exploring. I haven't seen this specifically written, but
to me, the me-not-me problem is "the" major cognitive problem the baby
needs to solve in its first year or so - and I have a gut feeling that
vision and audition are more like magic to the little kid, but that
touch is the basic process by which this problem is solved. This is is
just a working hypothesis, but I got a couple of books from the
library to follow it up.
At any rate, getting back to the issue .... initially, arm and leg
movements are reflex actions, but as the kid develops, they come more
and more under cerebral control. The kid explores things by touch and
manipulation, and by age one year, the kid is touching, moving, and
biting everything it can get its little mitts on. The point is, this
entire thing is an active-process, and not a passive process. If it
only had vision and audition, which are really more passive, it
couldn't really interact with its world directly like it can using
motion+touch. It does of course start to learn language, and there is
the hearing+speaking process, but this still isn't as intimate for the
kid as motion+touch.
Regards the me-not-me problem, I think this is directly related to why
the baby chooses to move its arms, once past the auto-reflex phase.
Initially, when the arm whizzes by the baby's line of sight, it
probably doesn't mean anything at all to the kid, but as time goes on,
the baby begins to make a connection between the fact that it itself
is controlling that arm. IOW, it realizes "I" moved it, and "I" see
it, and hey that arm thingie is "me" - and not just something that
simply happened. It's just a thought, but I have a feeling that the
brain --> arm --> eye --> brain loop is the critical element in early
cognitive development of the kid. I am sure this is all written
somewhere by Piaget.
Now, that being said ..... think about how all of this can apply to
the process of early learning in a naiive infant AI, to how it learns
about its environment, and eventually to how it attains an
appreciation for me-not-me [ok, the last is a little whimsical maybe].
Active-feedback loops between the AI and the world are a key element.
Basically Brooks' idea about physical grounding - except here we are
also going put in some memory and learning. This is how the naiive AI
learns the simple things in life.
some more 0.02,
- dan
=================
dan michaels
>
> >Neil W Rickert <ricke...@cs.niu.edu> wrote:
> >> cu...@kcwc.com (Curt Welch) writes:
>
> >> >It wasn't a metaphor actually. It was simplistic to make a very
> >> >important point. Long before an AI can decide if it's going to go read
> >> >messages on usenet, or go watch TV, it has to first figure out the
> >> >simple things in life, - like whether it should raise it's arm, or let
> >> >it hang. Or whether it should open it's eyes or leave them shut. Or
> >> >turn it's head to the left, or to the right.
>
> >> I expect that a human infant has worked out how and when to move its
> >> arms or open its eyes, long before it has discovered that it has arms
> >> and eyes.
>
> >When I said "the AI has to first figure out the simple things of life", I
> >was talking about the function of the AI hardware has to have some
> >algorithm which determins if it will lift the arm or not. I'm asking, what
> >will the function of that hardware be? Why does the baby AI hardware raise
> >the arm? What software to I write and what does my software base the
> >decision on wether to lift the arm or not?
>
> My answer to "Why does a human baby raise its arm" would be "because
> it wants to". I don't believe that there is any algorithmic answer
> to be given. You could look more closely, and see that nothing is
> happening that violates any scientific principles. But, I suggest,
> you would never identify an algorithmic answer to the question on arm
> raising.
>
How about ... "because it has to" ... [if it ever wants to find out
who and what it is ... see below].
=================
> For sure, there need to be control mechanisms available. "It raised
> the arm because the control mechanism was activated". But if I
> understand what you are asking, then that merely moves the question
> to "why was the control mechanism activated". I don't believe that
> there can be algorithmic answers to such questions, if you want the
> AI system to have anything we would consider to be free will.
>
> I do see the problem. You are wanting to design a computation based
> AI system. And unless there is an algorithmic answer, the arm never
> does get raised. I see that as a limitation of computationalism.
>
How about ....
1 - raising its arm is the first step on the road to intelligence, and
2 - the act of arm raising should be hard-wired to occur immediately
upon powering up the naiive AI for the first time. [or soon
thereafter]
===================
> I should add that I am not trying to dissuade you from trying to
> find answers.
>
> >> You said in an earlier post, that your AI system would have free
> >> will. If it will have free will, then you cannot know how it will
> >> make a decision to raise its arm. That becomes a decision for the
> >> AI, and one that is not up to you.
>
Eray Ozkural exa
> My primary reason for rejecting computationalism, is that I see
> measurement processes as essential components of intelligence.
> A Turing machine has no ability to measure anything.
That is an interesting view. But you see a Turing machine is simply a
formalization! The "real" computation takes place in the physical
world and is fully grounded.
Each set of quanta participates in a number of computations. Every
such computation necessarily involves physical measurement. Please
argue against this.
I wouldn't expect your argument to be refuted in two sentences so I
would like you to clarify what you think measurement is, if it isn't
physical measurement which is typically seen as nothing more than
physical interaction (ie. photons deflected from a surface)
Regards,
__
Eray Ozkural
Eray Ozkural exa
metrics.
Typically the system makes a number of transformations (like
generating patterns) in the direction of decreasing system DL.
bk...@earthlink.net (boris.k) wrote in message news:<248f6ba2.03072...@posting.google.com>...
Eray Ozkural exa
> OTOH, because of all the interconnection pathways at every level, it's
> certainly not clear how everything fits together in the end. Eg, V4,
> the color area, connects directly to ~20 of the other areas of the 30.
> Why so much cross-traffic at that level?
Isn't it obvious? :)
It's because the relations among features are used by the learning
algorithm. It is the synthesis of perception.
There is only so much one can infer from analyzing each feature
independently. You could say that the brain looks for associations
more than anything else.
Michael Feldhake
I think Neil is right. I believe we could label the brain as an
Instrument, not a Computer. The AI industry labeled it that way only
because they were Computer Science people. Granted it may produce
some results that can be demonstrated using mathmatics, but it does
not operate that way. Just look at your basic Neuron.
Mike
dan michaels
Yes, mistated by me. The why question is obvious ;-). The how it all
works question is the tough one.
Neil W Rickert
>> My primary reason for rejecting computationalism, is that I see
>> measurement processes as essential components of intelligence.
>> A Turing machine has no ability to measure anything.
I am breaking up the next paragraph for commenting.
>That is an interesting view.
Ok.
> But you see a Turing machine is simply a
>formalization!
I fully agree with that.
> The "real" computation takes place in the physical
>world and is fully grounded.
I don't agree with that.
>Each set of quanta participates in a number of computations. Every
>such computation necessarily involves physical measurement. Please
>argue against this.
The physical implementatation details for the computation are
irrelevant. We could compute in a different physical manner. That
does not change what is computed.
Computation is inherently abstract, even though physically
implemented.
>I wouldn't expect your argument to be refuted in two sentences so I
>would like you to clarify what you think measurement is, if it isn't
>physical measurement which is typically seen as nothing more than
>physical interaction (ie. photons deflected from a surface)
I am certainly talking about physical measurement. The idea that it
is "seen as nothing more than physical interaction" is absurd.
Neil W Rickert
>Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bfkbpn$b3m$1...@husk.cso.niu.edu>...
>> My answer to "Why does a human baby raise its arm" would be "because
>> it wants to". I don't believe that there is any algorithmic answer
>> to be given. You could look more closely, and see that nothing is
>> happening that violates any scientific principles. But, I suggest,
>> you would never identify an algorithmic answer to the question on arm
>> raising.
>How about ... "because it has to" ... [if it ever wants to find out
>who and what it is ... see below].
>=================
Usually "because it has to" implies some kind of compulsion. If you
wanted to just say "in order to explore the world and to explore
itself" then I could agree.
>How about ....
>1 - raising its arm is the first step on the road to intelligence, and
I don't agree with that. There are babies born without arms. As far
as I can tell, then does not prevent them from becoming intelligent.
>> I should add that I am not trying to dissuade you from trying to
>> find answers.
>> >> You said in an earlier post, that your AI system would have free
>> >> will. If it will have free will, then you cannot know how it will
>> >> make a decision to raise its arm. That becomes a decision for the
>> >> AI, and one that is not up to you.
>Ah ha, good questions here ... but the designer should make the
>decision initially .... because:
>Due to some of these threads, I have started reading up on infant
>development during the first 3-4 years. Haven't gotten very far along
>yet, but I already have some opinions ;-).
>And the first one is .... when it somes to training a naiive AI, maybe
>one should investigate early development in infants and use them as a
>role-model.
>The answer as to why a baby raises its arms turns out to be very
>simple. Reflex.
Maybe that's what behaviorists say. But it is a pseudo-answer.
It does not explain anything.
David Longley
<ricke...@cs.niu.edu> writes
>d...@oricomtech.com (dan michaels) writes:
>>Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bfkbpn$b3m$1@husk
And what if the behaviourist says it's an operant?
--
David Longley
Neil W Rickert
>Neil W Rickert <ricke...@cs.niu.edu> wrote in message news:<bfkbpn$b3m$1...@husk.cso.niu.edu>...
>Hi Neil. I think the below conversation is heading somewhere
>important. I and others have been thinking, accumulating, on the broad
>subject for quite some time, mentioned in numerous posts in numerous
>ways, and I think you are well-informed in this area. This has to do
>with the difference between what can be scientifically predicted and
>what brains can predict but on a subconsious level (not "informed
>wisdom") with the ultimate goal of gaining insight to the problem of
>generalizing "rules" (I call it "logic") of various problem solving.
>Introducing "free will" and "real-world complexity" in terms of Curt's
>project seem to lead to interesting places.
>I wonder if you would clarify a couple of points so I can understand
>your thinking better. I've edited to the bare minimum of what I see as
>relevant to the broad issue. Apologies in advance if you find that
>troubling. Feel free to put anything back.
>I don't mean to exclude Curt. I pretty much understand his points.
This is usenet. Nobody is excluded.
>>My answer to "Why does a human baby raise its arm" would be "because
>>it wants to". I don't believe that there is any algorithmic answer
>>to be given. You could look more closely, and see that nothing is
>>happening that violates any scientific principles. But, I suggest,
>>you would never identify an algorithmic answer to the question on arm
>>raising [...edit raising mechanics...] if you want the AI system to
>have anything
>>we would consider to be free will.
>>>> You said in an earlier post, that your AI system would have free
>>>> will. If it will have free will, then you cannot know how it will
>>>> make a decision to raise its arm. That becomes a decision for the
>>>> AI, and one that is not up to you.
>>>But I have to hard-code that free will into the hardware. I have to
>build
>>>the decision system that guides its actions. If I build a robot
>which has
>>>an arm under the control of the computer, I have to write software
>that
>>>determines if the arm moves, or stays still. How do you write this
>>>software, yet give the robot "Free will" at the same time?
>Need some time to think about this. Like where might
>probably-not-entirely randomness come from in a computational model in
>context of free will. It's interesting.
"Free will" is one of those topics that generates more heat than
light. I only mentioned it because Curt does.
I personally would not identify free will with randomness, although
some people suggest a connection.
>><snip>
>>>> >Why can't it learn the "rules"?
>>>> The rules are too complex.
>>>I think these type of algorithms have more powers to understand
>"rules"
>>>than you reallize.
>An example of real-world complexity would be great here. Something
>where humans seem to be able to come up with some meaningful
>probability that science can't, but subconsciously.
We don't know what people come up with subconsciously.
Here's a different example.
The distance between two small coffee stains on my desk is 0.6 inches. I
just measured it.
Sure, a scientist could come up with that -- if he happened to think
about doing it.
The usual AI and philosophy discussions assume something of a fixed
data input. Then one is supposed to find rules or patterns in
that data.
I am suggesting that, instead, we can go out and collect a new type
of data, perhaps something that nobody had considered collecting
before. We do better, not because we have fancy rules (discovered
patterns in the data), but because we are collecting better data.
Here's a thought experiment (or two thought experiments).
Consider weather predictions. We (or the forecasters) are not
perfect, but it is a lot better today than it was 30 years ago. You
can imagine that it wasn't very good at all 200 years ago. Yet the
concepts used (temperature, humidity, barometric pressure, etc) were
mostly all available 200 years ago.
Thought experiment 1:
We take our present equations, and atmospheric models for
weather prediction. We load them onto a time machine. We
throw in some computers powered by solar cells. We take them
all back to 1800, and give them to the scientists of that
time. (The computers are for running the atmospheric models,
and the solar cells are needed since they didn't exactly have
a modern electricity grid at that time).
Question: Will these scientists of 1800, now armed with superior
knowledge about patterns in the weather system, be able to make
better predictions?
My answer is that no, they will not.
Thought experiment 2:
In this experiment, we also take a time machine back to 1800.
However, we skip all of our theoretical knowledge of meteorology.
We allow the scientists of that era to get by with their own
knowledge. However, while there, we do launch weather satellites,
and provide weather balloons, and a network of instrumentation to
gather the sort of data that is used by meteorologists today.
Question: Will these scientists of 1800, now provided with vastly more
information, be able to make better weather predictions than they
previously could.
My answer is yes, they will. Without the computers and atmospheric models,
they won't do as well as we do today. But they will still be able to do
far better then they previously could.
My point: The major advance are not that we have discovered fancy
schmancy rules about how things work. The major advances are because
we are collecting more information and better quality information.
My skepticism about computationalism with respect to AI is just
that. No amount of computation will increase the information that is
collected, or improve the quality of information that is collected.
You have to get your hands dirty and go out and be physically
involved in the world to do that.
As I suggested on my answer to Pierre-Normand Houle, I see most
important parts of intelligence and learning having to do with how we
collect information, rather than with how we process information that
has been collected. In terms of learning, I am referring to
perceptual learning (as, for example, in Eleanor J. Gibson,
"Principles of Perceptual Learning and Development").
>probability that science can't, but subconsciously. I am reluctant to
>repeat my ET example yet again. I think we're looking for something
>really elusive and possibly residual, so examples aren't that easy.
>But one with empirical tests, not ESP, philosophy, etc. Maybe the
>question "How does a dog tell the difference between someone tripping
>over it and someone kicking it?" helps to find an example. Mark Twain
>said, "When you pick up a cat by the tail you learn something you
>can't learn in any other way." Hey, nothing worthwhile is easy. The ET
>example took years...
>>Roughly speaking, the type of rule learning you would need is only
>>possible if you suitably greased the skids in your design.
>I don't understand the "greased the skids" analogy. Is this the
>difference between the theoretical probability of a coin-toss v. real
>world measurement of it? Like "grease" the coin until neither side has
>unique attributes?
Around here "greased the skids" is used in political discussions,
such as "The mayor greased the skids so that the XXX company would
get the contract". In terms of neural networks, I'm suggesting
that if you want a neural network to be able to learn a rule,
then you had best design your neural network programming in such
a way as to make finding that particular rule easy.
>>>> >> The past is a poor predictor of the future.
>Do you mean the past is "sometimes" a poor predictor of the future?
I allow that there are specific areas where predicting does well.
>Is this the same as "We need to guess the likelihood of future
>events?" What's the "direct our actions" distinction? (If one
>predicted that pilot training would improve the odds of avoiding a
>future crash, would that not be confirmable for aggregates of pilots?)
You are looking at this wrongly.
The human species evolved in very different circumstances from
those of modern life. Think of a band of hunter-gatherers on
the African veldt, a few thousand years ago. There weren't too many
predictions that they could make that would help.
They could predict that if they didn't go find some food today, they
would be pretty hungry tomorrow. But just because they found an
antelope they could kill and eat yesterday, it does not follow that
they could predict they could do the same thing at the same place
today.
Yet intelligence evolved so as to serve the needs of just such a
band. What they most needed was a keen perceptions, so that they
could readily observe likely sources of food.
Incidently, when I suggested that the world is a disordely place, and
that we attempt to organize it (or create order), you might think of
the world of those ancient hunter-gatherers. I think you would
probably see why I call it disorderly.
>>>If you don't understand how central the concept of "predicting the
>future"
>>>is to everything we do, you have no hope of building AI or
>understand work
>>>such as mine.
>I think Neil understands it, but is trying to make an important point
>about it. (And it would really be nice if we could keep away from
>butterflies affecting weather or chaotic v. orderly universe. At least
>I'm not ready to resolve *that* issue today!)
The places where we best predict, involve our technology, rather than
the natural world before we tampered with it.
Newton's laws would have meant nothing to that ancient band of
hunter-gatherers. Knowing Newton's laws would not have improved
their lives one iota. But once we had people shooting cannon balls,
then Newton's laws became more valuable. In a sense, they are not
so much laws of nature, they are laws of technology, and how to
control that technology.
>>If I were to try to build an AI system, it would not be computation
>>based. But at present the needed hardware components are not readily
>>available, except as parts of biological organisms. And we would
>>have difficulty assembling them as efficiently as do biological
>>systems.
>That makes a lot of sense, I've posted as much. But if the brain has
>some extra facility for estimating the future that scientific
>prediction does not have (and I don't necessarily mean better overall,
>but just any facility), we could try to list some attributes.
I have suggested an answer in the thought experiments above. The
human does better than science, when the human has at his disposal
information that is not being used by science. Better information
leads to better predictions. Much of our information input is of an
informal kind, hard to describe and hard to document.
There is the old story of the retired school janitor.
The school board found that they were having problems
with the heating system (a hot water system). So they
asked the former janitor if he would help them. He agreed,
providing he would be paid his $100 consulting fee.
The janitor comes in. He picks up a block of wood, and
hits the pipe 3 times. The heating system starts to work.
Then he asks for his $100 consulting fee.
He was told he would have to submit an itemized bill.
His bill:
For hitting the pipe $5
For knowing that this
would solve the problem $95
>Bonus quote: Einstein said examples are not one way of teaching, they
>are the only way of teaching. You're Einstein and Larry is your dumb
>student. I also think that would be good for learning about this
>problem and help keep it empirical.
Here's another example -- really an analogy.
In the usual way of looking at things, data is input. The data
is often referred to as a representation (it represents something
about the state of the world).
What most people in AI want to do, and what most epistemologists
say we should do, is find patterns in the data. This is sometimes
called rule induction, sometimes it is called statistical analysis.
I guess "data mining" is the hot button term of the moment.
Here is the analogy. One way of forming representations is to use a
camera. The photographs (markings on paper) form the representations
or data.
According to conventional wisdom from AI and philosophy, the way to
get improved results is to do a statistical analysis of the data (the
markings on paper).
According to the Rickert view, you will do far better by improving
the optics of the lens, the grain of the photographic emulsion, or by
using infra-red or uv photography. These methods bring in new
information that was never available before, whereas the statistical
methods merely try to make better use of the same old data.
Glen M. Sizemore
DM: How about ....
GS: Sorry, Dan, wrong again. Or at least, mostly
wrong. A great deal of the movements of babies are,
like babbling, not reflexive.
DM: At birth, the baby has many reflexes built-in by
genetics.
GS: This is true, but reflexes are not the sort of thing
that "can be used to explore the environment." Instead,
such behavior is rigidly to eliciting stimuli. Behavior that
"explores the environment" is not "committed" to
particular eliciting stimulus classes, and is said to be
emitted, rather than elicited.
DM: I guess these are pre-wired in order to get things
going -
so the brain doesn't have to make any real decisions
before its ready.
GS: Brains don't make decisions, people do.
DM: Built-in are how to suck, how to breathe, basic
movements, etc. As
soon as it pops out, those little wigglies start wiggling. It
has no
choice in the matter. It also has a very strong grasping
reflex. Touch
it's palm with your finger, and it'll latch on.
GS: Most of what you are referring to as "basic
movements" are not elicited. Indeed, coordinated
movements begin to occur in utero before the sensory
nervous system has made functional connections.
DM: Now this may all sound pretty innocuous, but I
don't think so. For one
thing, the baby has to get all of its pathways wired up
and the
various systems coordinated. I imagine that practice in
using the
pathways has a lot to do with this. Practice, practice,
practice.
GS: The infant isn't "practicing" anything in the sense
that a person might "practice a drop shot." It has
nothing to do with "practice" and has everything to do
with the consequences of responses. Reinforcement,
reinforcement, reinforcement.
DM: More importantly, and I don't have a good feel for
this mechanism yet,[...]
GS: Oh really?
DM: [...]one of the major things the baby has to
accomplish during its first
year is to learn the difference between "me" and "not-
me".
GS: How do you know this, and how does one know
when this is finished? Is this "imperative" hard-wired?
And, if so, wasn't this "knowledge" simply hard-wired?
My guess is that the first thing the baby has to learn is
how to converge its eyes on an object, and how to
touch it.
DM: Initially,
its brain doesn't know what a "me" is[...]
GS: Brains don't know things. People know things.
DM: [...]- or anything else for that
matter. It learns a lot from visual and auditory input,
GS: No, it learns a lot because certain responses have
visual and "auditory input," or alters such "input." It is
probably true that some operant behavior (like
convergence) is partially acquired when stimuli are
merely presented, but it is likely that such learning
occurs much more rapidly when the stimuli are part of
some reinforcement contingency.
DM: but I have a
hunch that the major way it distinquishes the difference
between me
and not-me[...]
GS: What would one observe in order to say that the
child did distinguish between "me" and "not-me?"
DM: is tactile, by using its hands and feet, and touching and
grasping and exploring. I haven't seen this specifically
written, but
to me, the me-not-me problem is "the" major cognitive
problem the baby
needs to solve in its first year or so[...]
GS: Why do you say this?
DM: [...]- and I have a gut feeling that
vision and audition are more like magic to the little kid,
but that
touch is the basic process by which this problem is
solved. This is is
just a working hypothesis, but I got a couple of books
from the
library to follow it up.
GS: Make sure to keep us posted. But do tell us how
one would know that the infant "learns this."
DM: At any rate, getting back to the issue .... initially,
arm and leg
movements are reflex actions,[...]
GS: Nope.
DM: [...]but as the kid develops, they come more
and more under cerebral control.
GS: That must be where the little executive sits.
DM: The kid explores things by touch and
manipulation, and by age one year, the kid is touching,
moving, and
biting everything it can get its little mitts on. The point is,
this
entire thing is an active-process, and not a passive
process.
GS: Who argues that it is a passive
process?.......actually, you do. If all the infant,s
behavior were a reflex then everything would be
passive response to "input."
DM: If it
only had vision and audition, which are really more
passive,[...]
GS: Well, the stimuli "come to you" so to speak. But
vision and audition are not passive. Indeed, to "see,"
"hear," "feel," etc. is to behave actively. This has
caused much difficulty in psychology.
DM: [...] it couldn't really interact with its world directly
like it can using
motion+touch. It does of course start to learn language,
and there is
the hearing+speaking process, but this still isn't as
intimate for the
kid as motion+touch.
GS: This strikes me as somewhat silly. Indeed, it stems
from 17th and 18th century philosophy......oh, but you
probably think that there is no sense in which this is
true.
DM: Regards the me-not-me problem, I think this is
directly related to why
the baby chooses to move its arms, once past the auto-
reflex phase.
GS: This is mostly dead-end thinking. You are focusing
on the alleged choosing instead of focusing on the
consequences of the spontaneous behavior. The baby
emits behavior, these responses have consequences,
and the response classes become come to be emitted
more frequently. To see the "choosing" as an act in-
and-of-itself (yes, that is an implication) is misleading;
indeed, "choosing" in the sense of engaging in behavior
that affects other behavior and makes prepotent on of 2
or more alternatives IS an act, but it is not what is
happening every time an operant is acquired. But, of
course, "choose" has several meanings - you have
selected a trivial one; you have made every non-reflex
action a "choice."
DM: Initially, when the arm whizzes by the baby's line
of sight, it
probably doesn't mean anything at all to the kid, but as
time goes on,
the baby begins to make a connection between the fact
that it itself
is controlling that arm.
GS: That is, moving its own arm has visual (as well is
proprioceptive etc.) consequences. And my guess is
that such novel events are quite important...they are
probably unconditioned reinforcers. The notion that it
"makes a connection" is animism, and it eventually
causes many problems - like mainstream psychology.
DM: IOW, it realizes "I" moved it, and "I" see
it, and hey that arm thingie is "me" - and not just
something that
simply happened. It's just a thought, but I have a feeling
that the
brain --> arm --> eye --> brain loop is the critical
element in early
cognitive development of the kid. I am sure this is all written
somewhere by Piaget.
GS: The translation of the facts into "realizations" and
"choices" is exactly what is wrong with mainstream
psychology. It leads your beloved neurobiologists to
look for things that simply don't exist. It doesn't matter
that these terms are useful in everyday discourse (in
what David and others call folk-psychology); a
realization is not a thing or event in the brain; it is a thing
"in" behavior.
DM: Now, that being said ..... think about how all of
this can apply to
the process of early learning in a naiive infant AI, to
how it learns
about its environment, and eventually to how it attains an
appreciation for me-not-me [ok, the last is a little
whimsical maybe].
Active-feedback loops between the AI and the world
are a key element.
GS: You may have thought this up yourself, but the idea
is very, very, old. In addition, it is the subject matter of
the experimental analysis of behavior and its discussion
of theses issues are orders of magnitude more
sophisticated than yours.
DM: Basically Brooks' idea about physical grounding -
except here we are
also going put in some memory and learning. This is
how the naiive AI
learns the simple things in life.
GS: Here is what you said, but said efficiently and
without animism: babies emit behavior and that
behavior has consequences, some of these
consequences make responses that have this effect
more probable. Babies also have reflexes. It is
important to note that many important consequences
are simply "sensory," as when our own movements
produce changes in visual, auditory and proprioceptive
stimuli.
"Realizations" and "choices" etc. etc. etc. etc. etc. etc.,
like behavior, are "things" that happen in the world.
Indeed, "realizations" etc. are behavior. They are not
things that precede or accompany or cause behavior. In
this important sense, to search for these things in the
brain is to search in vain. The relevant events in the
brain that mediate behavior are not themselves
behavior, and they are not "realizations." And they are
not "mental."
Most cordially,
Glen