Results of Loebner Prize 2005
loe...@gmail.com
held Sunday, Sept 18, 2005 in New York City. My thanks to everyone who
entered the 2005 Contest.
Here are the results:
First Place - Rollo Carpenter Mean Rank 5.75
Second Place - Vladimir Veselov Mean Rank 6.00
Third Place - Steve Watkins Mean Rank 7.00
Fourth Place - Richard Wallace Mean Rank 7.25
Detailed results are at:
http://loebner.net/Prizef/2005_Contest/results.html
I announced hree things at the contest.
1. The next contest will be held Sunday, Oct 1, 2006
2. The final four contestants, or their representative(s) must be
present and
supervise their entries during the contest.
3. I will award 4 stipends of USD 250 to each of the Final Four. This
is to
underwrite travel expenses, or to allow them to hire a local
representative if
they can not appear.
More detailed rules will be announced soon.
Hugh Loebner,
Loebner Prize Sponsor
Bengt Richter
>The fifteenth annual Loebner Prize contest, Loebner Prize 2005, was
>held Sunday, Sept 18, 2005 in New York City. My thanks to everyone who
>entered the 2005 Contest.
>
>Here are the results:
>
>First Place - Rollo Carpenter Mean Rank 5.75
>Second Place - Vladimir Veselov Mean Rank 6.00
>Third Place - Steve Watkins Mean Rank 7.00
>Fourth Place - Richard Wallace Mean Rank 7.25
>
>Detailed results are at:
>
>http://loebner.net/Prizef/2005_Contest/results.html
>
No link to session logs?
Regards,
Bengt Richter
loe...@gmail.com
HGL
Evgenij Barsukov
I tried out the Jabberwocky on www.Jabberwocky.com, and it behaves
as a very annoying human. It answers with questions to a question, turns
the convesation away from the topic. Basicaly this is a valid human
behaviour, but it makes it a completely useless human - e.g. it would
not be of any help in solving any real life situation.
I think this identifies a real hole in the whole concept of Turings
test. Yes, human _can_ behave this way during a conversation, which
can make machine indistinguseable from human. But human can also
behave _differently_, but there is no way in keyboard-to-keyboard
communication to set environment in such a way as to make this
_different_ behaviour to only valid human behaviour and therefore
to identify a human.
It should not be enough to make a machine to imitate just one of human
behaviours (e.g. conversation style purposing to end conversatin as soon
as possible). During the test, there should be a way to place the
responder in a situation when he is _forced_ to solve a real life
problem. People are most of the time in such situations, and it is the
only situaton where AI is of any use.
It can be a game, fine. But to have ability to shape environment,
there here should be a way to assert authority of the "tester" through
giving him more freedom with evaluating the behaviour. It should be able
to say - if you don't do this, I will fail you! and "applicant" for
being a human should be interested to pass.
Which brings us to the need of competitive element in the test.
Namely, in one conversation there should be not one tester / one
applicant, but one tester and two or more applicants (some of which are
humans), which are competing for the passing the test, and only one
can pass.
Nature is competitive, and human intelligence has developed in
competitive environment, which makes competiveness a necessary
requirement for a meaningful AI test set-up.
Regards,
Yevgen
rick++
in this newsgroup a few years ago. I was not
impressed at the time. In fact I was annoyed.
This is not to be confused with the more recent
poster George B. of Littleton Colorado who posts
a lot here too.
loe...@gmail.com
found at
http://loebner.net/Prizef/2005_Contest/Transcripts.html
hl
loe...@gmail.com
or she is the human. The human _should_ answer questions substantively
i.e. accurately, and not just be evasive by asking another question,
etc. After all, the human wants to prove that he or she is the human
but demonstrating his or her "intelligence."
I would think that any judge who must choose between one entity which
responds substantively, and another which responds evasively, will
choose the former.
Hugh Loebner
makc.th...@gmail.com
> A preliiminary listing of the 2005 Loebner Prize transcripts can be
> found at
I've read 1st transcripts of Alice and Jabberwacky, both where judges
quickly identified machines. Gotta say they suck.
Now, if I remember this right, the main prize comes not to chatterbot,
but to interactive video, that will also show realistic human face
while talking. Latter is even harder to do that damn chatterbot thing,
but what's the point, if noone can do even simple thing 1st? Wouldn't
it be better to restrict this contest to chat? I mean, was there ever
single bot presented who (err, what) was able to trick *ALL* the
judges, for all these years? If yes, I'd like to read those
transcripts.
Back to video part, imagine *real* AI like in "I, robot" - clearly
non-human, but chatting reasonably, and thoughtfully. Wouldn't such a
bot be worth of main prize, dosens of times? But, it will fail, just
because it doesn't look like human :(
HMS Beagle
will it end? When will it all end I ask you!!
Mr. Savain... would you like to add a sarcastic comment here? Now
is the time.
On 19 Sep 2005 09:58:31 -0700, "hu...@loebner.net" <loe...@gmail.com>
wrote:
Lester Zick
in comp.ai.philosophy wrote:
>Oh... I see some of us are still trying for the Turing Test. When
>will it end? When will it all end I ask you!!
Probably not until you can point out transcendentals on a real number
line. That's the empirical way.
~v~~
loe...@gmail.com
able to recognize visual and audio material, but will respond via text.
Hugh Loebner
HMS Beagle
(Lester Zick) wrote:
>>Oh... I see some of us are still trying for the Turing Test. When
>>will it end? When will it all end I ask you!!
>
>Probably not until you can point out transcendentals on a real number
>line. That's the empirical way.
>
Zick dammit. You need to go google the Thue-Morse constant.
makc.th...@gmail.com
> No. I don't require A/V output, just input. The AI entity must be
> able to recognize visual and audio material, but will respond via text.
So I was mis-informed. I have found
http://loebner.net/Prizef/LoebnerPrizeRules2001.htm are these still
apply for 2006+ ?
Lester Zick
Does the Thue-Morse constant point out transcendentals on a real
number line?
~v~~
Claudio Grondi
"Evgenij Barsukov" <evgenij_...@yahoo.com> schrieb im Newsbeitrag
news:dgp4ji$6hh$1...@home.itg.ti.com...
Considering how the majority of people were thinking about it all
many years ago it seems, that the machines show already enough
intelligent behaviour to make them indistiguishable from humans
at least in some narrow areas like e.g. playing chess already today.
After it is more or less apparent, that a machine can better play chess
than any human opponent some posters to newsgroups are coming
currently up with the statement, that chess is a game not requiring any
intelligence from the players.
Interesting in this context is, that instead of admitting wrong past
attititude towards AI, these ones who don't want and don't wish
it to be reality, widen the area and change the conditions to be met.
I see already future posters stating, that taking part in a conversation
using written language is not requiring any intelligence...
I am not happy to see it that way, but it seems, that convincing a
human, that a machine is really intelligent is to let the machines
fight and win against humans in a serious battle.
As a consequence I also see, that it will be never possible to provide
evidence that machines can achieve higher intelligence than humans.
If they kill all the humans there will be noone human left to admit
their superior intelligence; if there is at least one left, he can still
state, that because he is alive, the machines are not smart enough
to kill him.
Claudio
Steve Richfie1d
> As a consequence I also see, that it will be never possible to provide
> evidence that machines can achieve higher intelligence than humans.
The test of this comes when machines start solving important hard
problems that baffle the experts. This is already the case in weather
prediction. Dr. E1iza is an NLP program that now routinely solves
chronic health problems where the best research hospitals have failed!
Of course, when you dig down deep into Dr. E1iza it is just a bunch of
tables and logic just as you'd expect. HOWEVER,
It takes two crucial skills for a doctor or a computer to succeed in
this or most other difficult repair domains - System Dynamics and
Adaptive Control System Theory. A national search in 2001 for a doctor
in the USA with either of these skills found NONE. Here, a computer
exhibits "intelligence" only because the "experts" lack crucial skills.
Of course, doctors question the results because they can't follow the
logical process due to their lack of skills, so doctors will probably be
the last ones to ever accept the proposition of this being intelligent.
I see an opportunity to make successful products for several decades in
domains where humans typically lack crucial skills, for it will take a
LONG time before colleges are willing to admit that their tenured
professors and department chairmen lack the basic skills to function
effectively in their respective fields. I/we are now betting heavily on
this particular proposition.
Further, people must learn these skills when they are young because
these skills are ways of thinking rather than collections of facts, so I
see little prospect of fixing this skills gap anytime soon. Of course,
this is GOOD NEWS for AI developers because it guarantees a stable
market for our programs.
Kinda sounds like AI/NL, where people now routinely earn their PhDs in
subjects where no one has even created a useful system, so their
knowledge MUST be defective. What possible good are such credentials
beyond teaching others the same useless crap? Certainly, no more good
than MDs who are unable to do their chronically ill patients any
significant good.
What do you call a former medical school student who finished last in
his class? Answer: Doctor. 8-)
Steve Richfie1d
HMS Beagle
(Lester Zick) wrote:
>>
>>Zick dammit. You need to go google the Thue-Morse constant.
>
>Does the Thue-Morse constant point out transcendentals on a real
>number line?
Yes it does. But only in an infinite number of steps.
Lester Zick
Which is to say it doesn't.
~v~~
Evgenij Barsukov
> many years ago it seems, that the machines show already enough
> intelligent behaviour to make them indistiguishable from humans
> at least in some narrow areas like e.g. playing chess already today.
> After it is more or less apparent, that a machine can better play chess
> than any human opponent some posters to newsgroups are coming
> currently up with the statement, that chess is a game not requiring any
> intelligence from the players.
I quite agree that such attitude is unaceptable. Winning from best human
players in chess in a fundamental achievement of AI that shows that in
this particular type of intelligence AI is intelligence is better than
human intelligence.
> Interesting in this context is, that instead of admitting wrong past
> attititude towards AI, these ones who don't want and don't wish
> it to be reality, widen the area and change the conditions to be met.
Yes there are some outright wrong attitudes. However, it is not what
my message was about. I do not believe that general intelligence exist
fundamentaly, e.g. people do not have it either. No single human can
derive needed concepts to solve majority of even most primitive everyday
life's problems (like binding the shue-laces) not to mention actual
science and engineering. So, we _do not posess_ general intelligence.
Only humanity as a whole (including previous/future generation) can
come close to this definition, altough it is well known from the
history that civilizations have often failed to solve certain problems
and have perished. This can also happen with entire humanity, which will
prove that their intelligence was not general, or that the problem
did not have a solution.
Any intelligence is specialized (only it is specialized in different
directions).
The only thing that Turing's test is checking is ability to create
an AI that is specialized in a similar way as human intelligence (or
is more general and includes human type of specialization as a subset).
The only way of "hard-core" comparison of non-specialized intelligence
is comparing parallel and serial processing power. Both of these
comparissons can be done quantitiatively. Present state -
1) serial processing power of humans
is about 10 operation/sec, which is about 100 000 times slower than
an average PC. That is why we trully suck in arithmetics, many of it can
only be done using serial processing.
2) parallel processing power of humans is about 10^10 operation/sec (can
be derived independently from number of neurons, and from comprative
functionality tests of fixed number of cells). This is about 1000 times
faster than average PC. This is why we excell in solving problems that
can be paralelized - for example pattern recognition, 3d navigation etc.
There is a lot of problems that require both parallel and serial
processing, as well as benefit from intersignaling between parallel
processes. So a third more diffuse measure that is not yet quantified
in a simple ways such as "MIPS" is "inter-parallel communication rate".
To give an example, new AMD double-core CPUs are capable to
communicate between 2 parallely running threads much faster (because
they both can acees the same cache memory at the processor speed) than 2
separate CPUs sitting on a motherboard can do. However, inter-thread
communication is also OS dependent. So there should be a new quantity
describing this rate. I suspect this quantity along with straight
parallel and serial processing will be the key for judging "generality"
or "power" of paricular intelligence.
In general, prospects are dim for human intelligence because its serial
processing rate will always suck and can not be improved (unless we
figure out how to make special serial-co-processor implants). But even
than inter-tread communication rate will be very bad. At the other hand,
electronic intelligence will eventualy reach us on parallel processinng
front, at the other hand still having extraordinary fast serial
computations ability.
> I see already future posters stating, that taking part in a conversation
> using written language is not requiring any intelligence...
> I am not happy to see it that way, but it seems, that convincing a
> human, that a machine is really intelligent is to let the machines
> fight and win against humans in a serious battle.
No, that will not convince them either. Did you ever play StarCraft
against the computer? It is a highly intelligent strategy game with
debut, midelspiel etc that has hundreds of web-sites describing
different strategies. In this game, humans play both with each other
(there is even a world-cup!) and sometimes against computer. Ocasionaly
you can even play in a mixed human/computer team agains other people.
So, the bottomline is - sometimes computer wins, but it does not
make anybody think about own PC as intelligent. But it is wrong. It is
just as intelligent as humans in this particular specialized area.
> As a consequence I also see, that it will be never possible to provide
> evidence that machines can achieve higher intelligence than humans.
> If they kill all the humans there will be noone human left to admit
> their superior intelligence; if there is at least one left, he can still
> state, that because he is alive, the machines are not smart enough
> to kill him.
LOL.
In reality, I don't think machines will need to kill all the humans.
Maybe there will still always be some nishe area where their
intelligence will be capable to do enough work of entropy increase
to make their existance useful for thermodynamics (after all, humans do
not go and kill all the algi and plants just because we are more
intelligent). During the same time AI will move towards higher energies
and will work on creating new stars and black-holes.
Regards,
Evgenij
>
> Claudio
>
>
Barry in Maryland
artificial intelligence. Thinking back on the promise of Hal in 2001
Space Odyssey, for example, it seems that in many ways, little progress
has made since the early 60's. Last year at an optics and spectroscopy
conference in London, I listened to several presenters talk about their
magical new algorithms for facial recognition and was struck by the
great lengths to which we still have to go to get modest results from
silicon and software to do what we human doings do every day with ease.
After reading the transcipts of this year's competition, I am struck by
the ease with which the human judges were able to distinguish the
computer programs from the human beings. Pardon what may seem to be a
very basic question, therefore, but since the start of the Loebner
Prize competition, has there been any improvement in the ability of the
programs that have participated?
HMS Beagle
(Lester Zick) wrote:
>>
>>>Does the Thue-Morse constant point out transcendentals on a real
>>>number line?
>>
>>Yes it does. But only in an infinite number of steps.
>
>Which is to say it doesn't.
awe!!! The colllege freshman is angsting about infinity. How cute!
HMS Beagle
<evgenij_...@yahoo.com> wrote:
>I quite agree that such attitude is unaceptable. Winning from best human
>players in chess in a fundamental achievement of AI that shows that in
>this particular type of intelligence AI is intelligence is better than
>human intelligence.
>
So what? What is your point here? I can write a simple program in C
that never loses a game of Pacman...ever. Also, any fool can write a
bot that takes on humans in Unreal Tournament and never loses, because
the bots never miss a shot. Several decades ago, someone wrote a
simple algorithm that never loses in tic-tac-toe.. it always ties.
>1) serial processing power of humans
>is about 10 operation/sec, which is about 100 000 times slower than
>an average PC. That is why we trully suck in arithmetics, many of it can
>only be done using serial processing.
>
>2) parallel processing power of humans is about 10^10 operation/sec (can
>be derived independently from number of neurons, and from comprative
>functionality tests of fixed number of cells). This is about 1000 times
>faster than average PC. This is why we excell in solving problems that
>can be paralelized - for example pattern recognition, 3d navigation etc.
Ok here we go again with the argument that acheiving AI is a simple
matter of computer power. How many times do we have to go over
this?
>No, that will not convince them either. Did you ever play StarCraft
>against the computer? It is a highly intelligent strategy game with
>debut, midelspiel etc that has hundreds of web-sites describing
>different strategies. In this game, humans play both with each other
>(there is even a world-cup!) and sometimes against computer. Ocasionaly
>you can even play in a mixed human/computer team agains other people.
>So, the bottomline is - sometimes computer wins, but it does not
>make anybody think about own PC as intelligent. But it is wrong. It is
>just as intelligent as humans in this particular specialized area.
No. There is no intelligence at all. The computer is responding
exactly how its programmed to respond to certain events in the RTS
game. They actually set this on purpose, because as humans, we want
to be able to control how "difficult" the computer player is. If we
simply told the computer to use whatever works by learning from
experience, it would beat every one of us. And anyways, they have to
specifically tell the computer to build lots of ships and send them in
all at once. The computer has no idea WHY its doing that, or why
that's a good strategy. It has to be programmed to use this
human-like behavior. Same thing in chess. The Deep Blue has no idea
why a queen is "more valuable" than a knight. It has to be told which
pieces are more valuable by a human who UNDERSTANDS chess.
But so what? I can write an algorithm in java that can never lose at
a game of pacman. It's the same thing.
Lester Zick
Whereas you don't angst about your assumptions at all, Bagel.
~v~~
humi...@clix.pt
>Bengt Richter wrote:
>> On 19 Sep 2005 09:58:31 -0700, "hu...@loebner.net" <loe...@gmail.com> wrote:
>>
[snip]
>I tried out the Jabberwocky on www.Jabberwocky.com, and it behaves
>as a very annoying human. It answers with questions to a question, turns
>the convesation away from the topic. Basicaly this is a valid human
>behaviour, but it makes it a completely useless human - e.g. it would
>not be of any help in solving any real life situation.
The program does that in order to divert your attention from the
fact that it is as stupid as a doorstop. It doesn't understand a
word of what you say and therefore, when you say something that
doesn't trigger one of the canned responses, is has no way other
than to change the subject or output some "cute" non-sequitur.
Its very easy to verify this if you tell it something and then later
ask a question whose answer requires knowledge of what you said
earlier. For instance if you tell it "My favourite colour is green"
and then later ask it "Do you remember what my favourite colour is?"
chances are that it will reply with something to disguise the fact
that it has no idea of what you're talking about.
The Loebner contest, in its present form, and judging from past
performance, doesn't promote the development of programs of ever
increasing intelligent levels. Instead, it promotes the development
of programs that are better as disguising their stupidity.
As I've suggested in an earlier post, this state of affairs could be
changed if only programs that fulfilled two new requirements would
be able to compete: One of the requirements is that the program must
be able to learn. The second requirement is that a program must be
language independent.
As it is now, the Loebner contest is just an exercise in futility:
The programs entering the contest are useless, because only a
masochist would endure a conversation of more than 10 seconds with
someone who changes the subject at every other sentence. Of course
if people are participating in the contest for its entertainment
value, who am I to criticize?
Antonio Esteves
--
Corby - A new approach to Artificial Intelligence
http://futalgo.planetaclix.pt/corby/index.htm
humi...@clix.pt
"hu...@loebner.net" <loe...@gmail.com> wrote:
The problem with your contest is that is doesn't provide a suitable
environment for intelligent programs to emerge. Let me illustrate
with a concrete example.
You know the "Say <something>" routine that we use with small
children to improve their language abilities: You go "Say mommy" and
the child repeats "mommy". Initially the child doesn't know how to
respond but after a few examples and using the imitative instinct
that we humans have, the child will pick up that language pattern
quickly enough.
Now suppose that someone built a machine that is able to learn the
above language pattern just like a child does. Some pertinent
questions to ask are:
Would such a machine be light-years ahead of the programs
participating in the Loebner contest in terms of intelligence? You bet.
Would any of programs participating in the Loebner contest be able
to learn such a language pattern? I don't think so. Therefore we are
forced to conclude that a program purporting to simulate human
behaviour and not being able to do what a small child does must be a
fake.
Would such a program fare well in the Loebner contest? Fat chance. I
don't imagine a judge testing the ability of a machine to pick up
the above-mentioned language pattern. Nor does the contest have any
mechanism for a judge to tech something to a machine.
The problem is that in the context of the Turing test, the examiner
must assume he is dialoguing with an adult human being. On the other
hand, the current AI technology is barely able, if at all, to
simulate what a small child does. So how do we get there from here?
I don't think that your contest, in its current form, is of any help
in this respect.
I have no doubt that we'll have one day intelligent programs that
will be able to engage in a meaningful conversation with a human
being (although I don't think that we'll ever build a machine that
passes the Turing test). Bu that will happen in spite of, not because
of the Loebner contest. Sooner or later you'll have to face the fact
that you're not getting the most out of your money or efforts. And
this is, at lest in part, your fault.
HMS Beagle
>>>Does the Thue-Morse constant point out transcendentals on a real
>>>number line?
>>
>>Yes it does. But only in an infinite number of steps.
>
>Which is to say it doesn't.
Actually that's true. It never points out the number, ever. Good
call.
In mathematics, the transcendental is said to lie at the "limit" of
the infinite sequence. If we have a limit where n approaches
infinity, then the operant word here is "approaches". One cannot
square infinity and take its inverse, after all.
Steve Richfie1d
Antonio's and other's points here could easily be addressed by giving
extra bonus points for programs exhibiting particular behavior, e.g.
1. Ability to remember things and use them in future responses.
2. Ability to carry on a detailed conversation in a particular
program-specific domain.
I agree that the contest is flawed, but disagree with those who think
that it isn't easily corrected. Sure, plenty of program would fail to
collect the extra bonus points, but those that DID collect them would
win, which would "raise the bar" on the rest.
What would it hurt to give bonus points for things that NO program could
do? This would change nothing EXCEPT for focusing future efforts on the
uncollected bonus points.
Hence Hugh, the future of such competitions is in YOUR hands, to
carefully craft a futuristic scoring system that ALL programs can
participate in - in another decade or so.
In the mean time, you might get programs like Dr. E1iza to participate
and at least collect the program-specific domain points, and in the
process point out the way for others to use similar techniques.
Also, have you considered giving out several "divisional" prizes, each
in an area that needs a lot of work (like learning), rather than just
one "grand" prize? Of course, if no one competes for one of them, then
you just won't give out that prize for that year.
Steve Richfie1d
==================
Lester Zick
Okay. I'll buy this. So my take on this claim is that approximations
for transcendentals lie on a straight line but their limits do not.
~v~~
HMS Beagle
(Lester Zick) wrote:
>>>>Yes it does. But only in an infinite number of steps.
>>>
>>>Which is to say it doesn't.
>>
>>Actually that's true. It never points out the number, ever. Good
>>call.
>>
>>In mathematics, the transcendental is said to lie at the "limit" of
>>the infinite sequence. If we have a limit where n approaches
>>infinity, then the operant word here is "approaches". One cannot
>>square infinity and take its inverse, after all.
>
>Okay. I'll buy this. So my take on this claim is that approximations
>for transcendentals lie on a straight line but their limits do not.
We infer the limit is actually there somewhere. This is the value we
are "tending towards" as we add on smaller and smaller terms from the
infinite sum. Or what we tend towards as we keep iterating the
sequence. The basic idea being that the sum or sequence never
exceeds this limit if we are coming from below, and never becomes
smaller than this limit if we are coming from above. (Hence the word
"limit").
After the modern formulation of "limits" (which came quite late in the
19th century), mathematicians all concurred that this is a "valid"
way to define a number. The transcendental can be considered "that
quantity we never exceed as we iterate to infinity".
Lester Zick
in comp.ai.philosophy wrote:
>On Mon, 26 Sep 2005 00:28:33 GMT, lester...@worldnet.att.net
>(Lester Zick) wrote:
>>>>>Yes it does. But only in an infinite number of steps.
>>>>
>>>>Which is to say it doesn't.
>>>
>>>Actually that's true. It never points out the number, ever. Good
>>>call.
>>>
>>>In mathematics, the transcendental is said to lie at the "limit" of
>>>the infinite sequence. If we have a limit where n approaches
>>>infinity, then the operant word here is "approaches". One cannot
>>>square infinity and take its inverse, after all.
>>
>>Okay. I'll buy this. So my take on this claim is that approximations
>>for transcendentals lie on a straight line but their limits do not.
>
>We infer the limit is actually there somewhere.
Okay this statement is true. There is just no special reason to assume
transcendental limits lie on any straight line nor together on any one
line. That's an unjustified special assumption but is nonetheless what
is commonly assumed in mathematics.
> This is the value we
>are "tending towards" as we add on smaller and smaller terms from the
>infinite sum. Or what we tend towards as we keep iterating the
>sequence. The basic idea being that the sum or sequence never
>exceeds this limit if we are coming from below, and never becomes
>smaller than this limit if we are coming from above. (Hence the word
>"limit").
>
>After the modern formulation of "limits" (which came quite late in the
>19th century), mathematicians all concurred that this is a "valid"
>way to define a number. The transcendental can be considered "that
>quantity we never exceed as we iterate to infinity".
Well the problem is that a transcendental limit can be approached by
means of approximation on a straight line but that does not imply that
the limit itself lies on the straight line. It can lie anywhere off
the line. That's what makes transcendentals transcendental. All
approximations for transcendentals do is narrow the approach of any
transcendental to a straight line without implying the limit itself
lies exactly on the straight line between the approximations.
This is the distinction I draw between transcendentals and
irrationals. Irrational limits actually lie exactly between their
limits on straight lines and can be exactly pointed out on straight
line segments by means of rational line segments and right angles.
(And, yes, I'm aware of the conventional definition which considers
transcendentals irrationals so we don't need to cover that ground
again.)
~v~~
Evgenij Barsukov
> On Fri, 23 Sep 2005 09:52:29 -0500, Evgenij Barsukov
> <evgenij_...@yahoo.com> wrote:
>
>>I quite agree that such attitude is unaceptable. Winning from best human
>>players in chess in a fundamental achievement of AI that shows that in
>>this particular type of intelligence AI is intelligence is better than
>>human intelligence.
>>
>
>
> So what? What is your point here? I can write a simple program in C
> that never loses a game of Pacman...ever. Also, any fool can write a
> bot that takes on humans in Unreal Tournament and never loses, because
> the bots never miss a shot. Several decades ago, someone wrote a
> simple algorithm that never loses in tic-tac-toe.. it always ties.
Bots do not have to play by the same rules as people (e.g. they can
cheat as above). However, chess programs DO have to play by the same
rules (including timing), and therefore their achievemend in solving
that particular class of problems (e.g. chess problems) is exactly equal
to humans achievens.
If you argument is that they use pre-programed rules, this is a
non-argument because people do the same. 99% of reason why grossmaister
plays better than a layman is because he spent 10-20 years studying the
chess theory and learning rules that provides "short-cuts" in his brain
processing and therefore allow to get more processing done with fixed
available brain-power.
Still, with rules alone you can just go that far. You still need the
power, and that is why deep blue plays a lot better than PC altogh Fritz
II program plays already on modern PCs with a solid Master level.
>>1) serial processing power of humans
>>is about 10 operation/sec, which is about 100 000 times slower than
>>an average PC. That is why we trully suck in arithmetics, many of it can
>>only be done using serial processing.
>>
>>2) parallel processing power of humans is about 10^10 operation/sec (can
>>be derived independently from number of neurons, and from comprative
>>functionality tests of fixed number of cells). This is about 1000 times
>>faster than average PC. This is why we excell in solving problems that
>>can be paralelized - for example pattern recognition, 3d navigation etc.
>
>
> Ok here we go again with the argument that acheiving AI is a simple
> matter of computer power. How many times do we have to go over
> this?
Until you will finaly get it?
>>No, that will not convince them either. Did you ever play StarCraft
>>against the computer? It is a highly intelligent strategy game with
>>debut, midelspiel etc that has hundreds of web-sites describing
>>different strategies. In this game, humans play both with each other
>>(there is even a world-cup!) and sometimes against computer. Ocasionaly
>>you can even play in a mixed human/computer team agains other people.
>>So, the bottomline is - sometimes computer wins, but it does not
>>make anybody think about own PC as intelligent. But it is wrong. It is
>>just as intelligent as humans in this particular specialized area.
>
>
> No. There is no intelligence at all. The computer is responding
> exactly how its programmed to respond to certain events in the RTS
> game.
So do people. Did you ever seriously played any game? If you want
to play somewhat higher then newby level, you have to study strategies,
standard situations etc. The more you know, the more you behaviour is
"programmed" and at the same time, the more _optimal_ your behaviour
is.
The opposite of that is to behave completely random. But that is not
possible for real time problems. Why? Because you only get ONE chance.
You either survived or you died. That is why DAPRA challenge AI can not
(fundamentaly!) learn why it is not good to drive through barbed wire -
because it only takes one attempt to do so to completely fail. To learn
something you need to have a chance to have multiple attempts and to
survive them.
For all types of real time problems you _have_ to use previous AIs
knowledge (yet, knowledge about these who died). In some cases this
can be done by direct knowledge transfer, in some cases through
evolution but both of this way have nothing to do with the present day
AI engaged in solving present problem. It is an inter-AI type of
learning, just as majourity of human progres in an inter-human progress
and not result of single human intelligence.
They actually set this on purpose, because as humans, we want
> to be able to control how "difficult" the computer player is. If we
> simply told the computer to use whatever works by learning from
> experience, it would beat every one of us.
Right now computer have enough power to treat games that are formalized,
where computer does not need to do pixel-by-pixel analysis of the screen
to understand what is happening. For type of unformalized games where
whole multi-dimensional analysis has to be done it is still hopelessly
underpowered compered to human with his massively parallel
processing.
> And anyways, they have to
> specifically tell the computer to build lots of ships and send them in
> all at once. The computer has no idea WHY its doing that, or why
> that's a good strategy. It has to be programmed to use this
> human-like behavior. Same thing in chess. The Deep Blue has no idea
> why a queen is "more valuable" than a knight. It has to be told which
> pieces are more valuable by a human who UNDERSTANDS chess.
>
> But so what? I can write an algorithm in java that can never lose at
> a game of pacman. It's the same thing.
But you can't write an algorithm that will win in chess, just because
you do not have necessary background in chess. And that is the whole
point. Intelligence is data-processing. Algorithms are ways of
make data-processing faster vs. brute force data processing. Algorithms
are evolutionary and relay on thousand years of prior intelligence,
which you are not capable to make use of in case of chess because you
are not aware of it.
Regards,
Evgenij
forbi...@msn.com
Evgenij Barsukov wrote:
> HMS Beagle wrote:
> > On Fri, 23 Sep 2005 09:52:29 -0500, Evgenij Barsukov
> >>2) parallel processing power of humans is about 10^10 operation/sec (can
> >>be derived independently from number of neurons, and from comprative
> >>functionality tests of fixed number of cells). This is about 1000 times
> >>faster than average PC. This is why we excell in solving problems that
> >>can be paralelized - for example pattern recognition, 3d navigation etc.
> >
> >
> > Ok here we go again with the argument that acheiving AI is a simple
> > matter of computer power. How many times do we have to go over
> > this?
>
> Until you will finaly get it?
This is an interesting point I'd like to explore further.
With a doubling of processing power every 18 months we're look at
about 15 years to get equivilent processing power. My questions are
based upon the recogniztion that certain processes must remain
serialized. Here's the questions: Do you think the human brain
implements its processing in hardware or software? Emulation requires
overhead cycles. What percentage of processing power do you believe
will be eated up by the emulation process as opposed to the human
simulation process?
Carl Burke
"Barry in Maryland" <barryrw...@yahoo.com> wrote in message
news:1127488216.5...@z14g2000cwz.googlegroups.com...
Not if you're looking for great leaps, no. Progress has been incremental,
and discouraging, although the systems can perform adequately as guides
for websites and such. I don't think any of the research systems that have
a higher quality of inference and better understanding of grammar have the
same breadth of coverage of topics that the Eliza-like systems do.
I don't think any of those systems have been submitted, for a lot of
reasons.
On the other hand, the barriers to entry are a lot lower now, largely due
to the contest I think. So it's still hard to get a system that does well,
but it's easier and easier to make a bad system with some personality.
You can pick up bots that are already written, with some basic rules
for interacting via text, and concentrate on writing your own scripts.
I don't think the Alice/Eliza approach is going to get us to the prize,
but it's what most people feel comfortable using.
--
Carl Burke
cbu...@mitre.org
a dialogue kit in Java: midiki.sourceforge.net
HMS Beagle
[snip]
>overhead cycles. What percentage of processing power do you believe
>will be eated up by the emulation process as opposed to the human
>simulation process?
If you run any program on a PC running a WinXX system your process is
always interrupted by the OS checking for a mouse movement, and
constantly interrupted checking for other things that help draw the
windows in order.
Further, an Intel or AMD CPU in a PC is a general-purpose processor.
For this reason, its many orders of magnitude slower than a machine
specifically built to perform a specific calculation.
What is the "overhead" then? Its really bad. The computer is mostly
performing overhead. There is an enormous waste of calculations.
HMS Beagle
<evgenij_...@yahoo.com> wrote:
>> Ok here we go again with the argument that acheiving AI is a simple
>> matter of computer power. How many times do we have to go over
>> this?
>
>Until you will finaly get it?
>
Are you openly declaring in this newsgroup, with everyone watching,
that you (not only beleive) but are certain that all problems in AI
all reduce to lack of processing speed?
amnon....@textanalysis.com
The main evidence available is in the realm of games such as Chess. It
was once thought that only intelligent machines could beat the best
humans. Now we know that relatively brute-force algorithms combined
with processing speed are sufficient to beat top Grandmasters.
There's no reason to think that other problem spaces won't succumb,
given further multiples in processing speed and parallelization.
That's not to say that humans aren't doing something smarter. But
machines may be able to compensate to the nth degree with processing
speed.
Amnon
Steve Richfie1d
>>Are you openly declaring in this newsgroup, with everyone watching,
>>that you (not only beleive) but are certain that all problems in AI
>>all reduce to lack of processing speed?
> The main evidence available is in the realm of games such as Chess. It
> was once thought that only intelligent machines could beat the best
> humans. Now we know that relatively brute-force algorithms combined
> with processing speed are sufficient to beat top Grandmasters.
Long ago I was given the problem of demonstrating that the time sharing
service in town with by far the slowest processor and the least RAM
wasn't COMPLETELY worthless, so I wrote a simple Chess playing program
that never lost!
Well, it also never won. It played a carefully defensive game with no
hope of ever winning, it took almost forever to move, and it was SO
frustrating to play that no one ever invested the hours that it would
take to beat it. Of course, it would jump on any mistake that you made.
However, most customers couldn't even play chess, so such subtleties
were lost on them. The thing that was NOT lost on them was that none of
the competitors could play chess on their systems. Hence, the program
was an unqualified success at doing what it was designed to do -
obfuscate the inadequacies of the system that it ran on.
The point here is that you must CAREFULLY examine what is being
demonstrated, as things often aren't as they seem.
There is sometimes a LOT more going on in a Chess tournament than just
logic. In one tournament, I was up against a VERY strong player and
expected to lose. Midway through the game I was already down a piece.
Then, my opponent reached for a pawn but his shirt sleeve caught the
cross on his king and tipped it over. I immediately stood up and
announced "I accept your resignation" and walked out over his loud
protestations! Of course I knew that the referees would listen to his
story and insist that the game be played out, so I visited the rest
room, washed my face, and relaxed as I walked the halls awaiting a
referee finding me to tell me that I had to play out the game.
Eventually this happened, so I went in and resumed play. However, my
opponent was SO excited and distressed by his experience that he played
REALLY lousy chess and he quickly lost the game despite his superior
skills and considerable advantage.
Not an isolated incident, there have been several books written about
chess tournament incidents, e.g. "The Lighter Side of Chess".
There are many similar opportunities that straightforward programmed
approaches usually miss.
Most chess playing programs are blind to certain hard to recognize types
of draws, e.g. 3X repeating position and 50 non-pawn moves, so you can
easily draw them if you find yourself behind in the game. Of course,
these programs won't even recognize the draw.
One family member was completely exasperated when I forced a draw in a
game of Chinese Checkers! How do you force such a draw? When you see you
are going to lose, you move one of your pieces backwards to keep your
opponent from moving all of their pieces into their final position. Of
course, they will never play Chinese Checkers with ME again!
In life as in chess, when you are ahead you should play to win, but when
you are behind you should play to draw.
Steve Richfie1d
HMS Beagle
<amnon....@textanalysis.com> wrote:
>HMS Beagle wrote:
>> On Tue, 27 Sep 2005 07:24:04 -0500, Evgenij Barsukov
>> <evgenij_...@yahoo.com> wrote:
>> >> Ok here we go again with the argument that acheiving AI is a simple
>> >> matter of computer power. How many times do we have to go over
>> >> this?
>> >
>> >Until you will finaly get it?
>> >
>>
>> Are you openly declaring in this newsgroup, with everyone watching,
>> that you (not only beleive) but are certain that all problems in AI
>> all reduce to lack of processing speed?
>
>The main evidence available is in the realm of games such as Chess. It
>was once thought that only intelligent machines could beat the best
>humans. Now we know that relatively brute-force algorithms combined
>with processing speed are sufficient to beat top Grandmasters.
I really don't think Chess is a domain that even matters to
intelligence. It's like saying computers are more intelligent than
humans because they can sort lists faster than we can.
>There's no reason to think that other problem spaces won't succumb,
>given further multiples in processing speed and parallelization.
"There's no reason to think" is an article of faith. I think if you
look at what I actually asked, I didn't concede to this.
>That's not to say that humans aren't doing something smarter. But
>machines may be able to compensate to the nth degree with processing
>speed.
If you read the DARPA message board (see the thread, started by me,
called DARPA Grand Challenge, for a link) you will see none of them
talking about pure computing horsepower.
Instead they talk about the resolution of the cameras, and things like
object recognition. They all admit that if they had "Object
Recognition" many of their problems would be magically solved.
HMS Beagle
the future.
On Wed, 28 Sep 2005 02:09:58 -0600, Steve Richfie1d
<St...@NOSPAM.smart-life.net> wrote:
>There is sometimes a LOT more going on in a Chess tournament than just
>logic. In one tournament, I was up against a VERY strong player and
>expected to lose. Midway through the game I was already down a piece.
>Then, my opponent reached for a pawn but his shirt sleeve caught the
>cross on his king and tipped it over. I immediately stood up and
>announced "I accept your resignation" and walked out over his loud
>protestations! Of course I knew that the referees would listen to his
>story and insist that the game be played out, so I visited the rest
>room, washed my face, and relaxed as I walked the halls awaiting a
>referee finding me to tell me that I had to play out the game.
>Eventually this happened, so I went in and resumed play. However, my
>opponent was SO excited and distressed by his experience that he played
>REALLY lousy chess and he quickly lost the game despite his superior
>skills and considerable advantage.
This is a really good story and there is so much about it that should
give everyone here a lot to think about.
>There are many similar opportunities that straightforward programmed
>approaches usually miss.
>
>Most chess playing programs are blind to certain hard to recognize types
>of draws, e.g. 3X repeating position and 50 non-pawn moves, so you can
>easily draw them if you find yourself behind in the game. Of course,
>these programs won't even recognize the draw.
This is so dead on. Furthermore, would a computer, thinking on its
own, draw the following conclusion: This person gaurds the center of
the board. So if I break the center, it will "weaken" my opponent's
game. Certainly, no computer I know of can draw these kinds of
humanistic inferences about chess.
The problem is obvious. The computer has no idea that it's even
playing a game to begin with. All it is doing is running a minimax
search algorithm on the current state of the board. It doesn't even
make concessions to what has happened in the game recently. Meaning
its not even paying attention to its opponent's style.
It turns out Deep Blue did not have to rise through the tournament
ranks to get to Kasparov. This raises some serious questions.
Namely, how does Deep Blue fair against positional chess?
(My personal thoughts is that Deep Blue and Kasparov was a very nicely
orchestrated publicity/marketing stunt.)
>One family member was completely exasperated when I forced a draw in a
>game of Chinese Checkers! How do you force such a draw? When you see you
>are going to lose, you move one of your pieces backwards to keep your
>opponent from moving all of their pieces into their final position. Of
>course, they will never play Chinese Checkers with ME again!
>
>In life as in chess, when you are ahead you should play to win, but when
>you are behind you should play to draw.
Another interesting point. In many of the games, an exasperated
Kasparov simply called a draw with Deep Blue.
forbi...@msn.com
everything.
Raw computing power is necessary but not sufficient for AI.
I doubt camera resolution will be sufficient either.
It's hard having to wait on technology before one can try one's
techniques but there it is. I remember trying to explore the Mandelbot
set with my 8080 processor. It was nightmarishly slow going. That's
the way it still is with AI. Many good ideas will fall by the wayside
once technology advances sufficiently to give them a good test. New
ideas will arise in their place. Most of what we believe about AI
today
will be wrong but some will be right.
Here's what I think we know:
1. There will have to be sufficient computing power.
2. There will need to be sufficiently fine grained sensors.
3. There will need to be sufficiently fine grained effectors.
4. The hardware has to be quite power efficient.
The closer we get the harder it is to wait.
The trick will be to recognize when we're there.
The problem is to guess where the critical path lies and to
advocate advancement focus where needed.
Barry in Maryland
Carl Burke wrote:
> [removed comp.ai.philosophy from this subthread]
>
> a dialogue kit in Java: midiki.sourceforge.net
Thanks for the response, since I obviously can't hope to contribute in
a technical way to the discussion. I guess you're right that we have
to look at this long term rather than expecting some great leap
forward.
I tune in from time to time out of general interest to see what's new
in the field, and appreciate your taking the time to help out.
Barry
Barry
Evgenij Barsukov
I don't think it is possible yet to clearly distinquish the hardware and
software, I would rather speak about the ratio between parallel and
serial processes, and rate of interaction between the parallel processes
(e.g. rate of inter-parallel signaling) in human brain.
But now as you rise question about software/hardware effects, there is a
fundamental way to distinquish it, altough it hardly has been (or can be
yet) experimentaly applied to humans.
If we call the overal device that processes sensory (or any other) data
intput in any way that is required a "processor" than:
2) hardware is the part of the processor which structure can not be
modified by the data (logic gates, memory interconnects, memory size)
1) software is the part of the processor which can be modified by the
data (states of the logic gates, states of the memory)
In bio-processors there are indeed both kinds of parts, e.g.
1) would be neurons and their interconnects
2) is short term memory that is realized by temporary excitation of
neurons and remains active until concentration of signaling molecules
resides below certain level
However, there is also a third kind of activity, namely ability to grow
new interconnects between neurons, that does not below to any of above
categories. It is "hardware modification by software". This process is
slow (e.g. long time memory requires multiple repetitions) so overal
system can still be considered as 1)hardware and 2)software + 3)modifier
- an agent that modifies (1) to improve its performance in view of most
common operations required by (2).
To give an example, it is like a software that can automaticaly
order and upgrade for memory if computer does a lot of memory related
operations, or upgrade of a second CPU if a lot of parallel operations
is required. This "upgrade" does not happen in real time of particular
calculation, but next time same calculation is needed, it will go faster.
Considering this 3) ability, a lot of common often done processes get
a lot more resources than rare processes, so you could indeed say that
brain uses "emulation" for things that happen very rare, but does
"hard-wired" highly optimized calculations for things that have to be
done often. Moreover, things that have to be done often _always_ through
many generations, are already evolutionaly hard-wired from birth (so no
additional _upgrades_ ) are needed. Example of these are edge-detection
algorithms in the eyes, motions-control algorithms in hypotalamus etc.
The very reason why different parts of the brain have clear assigned
to them functions (that is clinicaly proven by removing or damaging this
parts disabling particular function) is that these parts are such
evolutionary hard-wired co-processors. Of cause to replicate their
behaviour in "general" processor, highly inefficient emulation would be
needed.
To summarize:
- humans processing does have clearly distinct hardware and software
- hardware gets gradualy upgraded for tasks that happen often
- they also have lots of hard-wired co-processor units which are already
optimized for tasks that are evolutionary significant
Discussion points relevants previous discussion:
- Is there a simple generic way to emulate above? No, it is a specific
higly specialized mutly-processor system including many specialized
co-processors, and so it can be emulated only specificaly,
part by part, not genericaly.
- Is it realy necessary? Yes for medical purposes, in order to be able
to repear it.
No for purposes of computing, as AI should be seen not
as replacement of individual people, but as part of humanity
intelligence as a whole.
Even individual people intelligence is of not much use as such, if
conceptual basis of humanity is not loaded into it, if intelligence does
not work in continius contact with overal humanity intelligence, and if
the product of the intelligence is not up-loaded back into humanity.
AI already can and does work in the same interactive mode as part
of overal humanity intelligence, and will continue doing so without the
need of emulation of specific kind of processor, e.g. human.
Regards,
Evgenij
Evgenij Barsukov
AI (to which I count all software development) is extemely succesful
thriving field which generates now more revenew than any other field
(indicated partly by Bill Gates being the richest person on earth).
So, what "problems" are you talking about? There are no problems, just
work to do in specific areas you are working on.
Obviously your specific programing problem is not going to be solved
by increase of processing speed. Optimizations that humanity have
accumulated over 100 000 years of existance are not going to be magicaly
"invented" by a single AI computer even if his computing power is
equivalent to single human, they still will need to be hard-coded if you
want to make something competitive with people in any intelligence
task. In any particular practical field you can not wait for a computer
finding things by trial-and-error for thousands of years as humanity did.
However, problem of beating people at specific tasks will (and
already have been over and over again) solved by increasing processing
power. Fritz II chess software have been available sinse 1990, but
it took a Deep Blue power running it to beat Kasparov in 2000.
Navigation sofware existed since early 80s, but it took portable
computing power higher than super-computer that was used to calculate
nuclear bomb to make a 200$ AI vaccuum cleaner Rumba, that actually
works. Are you going to argue that?
Regards,
Evgenij
Claudio Grondi
of the posting it is placed below, but I couldn't resist
the temptation to mention it in that context:
I see currently quite clearly, that distinguishing between
hardware and software is one of the concepts which
creates by far more confusion, that it helps to resolve.
The concept of software as a kind of immaterial idea
is in my eyes one of a fundamental barriers to overcome
before it becomes clear what is really is and how it
works. It is very surprizing for me, that this concept is
so widespread and supported in the current education
and that noone seriously questions if it is really useful for
understanding of matters related to computer and
computation.
I would be glad to hear opinions of posters to this
thread or people reading it on the subject if it makes
any sense to distinguish what is commonly understood
as software and hardware.
Claudio
"Evgenij Barsukov" <evgenij_...@yahoo.com> schrieb im Newsbeitrag
news:dhebv1$4r$1...@home.itg.ti.com...
HMS Beagle
Was that a yes or a no?
Fritz II is a minimax search algorithm. Nothing more. Deep Blue was
built to calculate this algorithm with special CPUs made particularly
for that. So you take this one example, this one anecdote, and
what? Decide that all problems of AI are a matter of processing
speed? (That argument is very weak.)
Evgenij Barsukov
You must have missed the complete message except that your pattern
recognition jumped on the word "deep blue". I am saying that AI
development (both hardware and firmware) does not have any problems, it
is a thriving field that already have overtaken all other areas of human
activity in terms of economic output. But there is no _single_ task that
this field has to solve, and there is no _generic_ AI (humans do not
have it either).
Beating people at each _particular_ task will take hardware computing
power equivalent to that of humans, unless the computational
optimiziation used at that particular taks by AI is superior to that of
humans. This does not need any discussion, it is a basic fact of
information theory.
Regards,
Evgenij
HMS Beagle
<evgenij_...@yahoo.com> wrote:
>You must have missed the complete message except that your pattern
>recognition jumped on the word "deep blue". I am saying that AI
>development (both hardware and firmware) does not have any problems, it
>is a thriving field that already have overtaken all other areas of human
>activity in terms of economic output. But there is no _single_ task that
>this field has to solve, and there is no _generic_ AI (humans do not
>have it either).
This whole paragraph is gobbledegook. I've never heard this argument
that AI is defined as humans writing software for a machine and
selling it to other humans. Rolf Pfeifer does no pay lip service to
it. And that man leaves NOTHING out. We talk about AI in this
newsgroup, but you start talking about Microsoft and Bill Gates.
>Beating people at each _particular_ task will take hardware computing
>power equivalent to that of humans, unless the computational
>optimiziation used at that particular taks by AI is superior to that of
>humans. This does not need any discussion, it is a basic fact of
>information theory.
This is complete, utter, nonsense. You cannot even start making
information-theoretic arguments unless the performance of the task is
done in the IDENTICAL manner as the human does it, all the way down to
his/her neuron cells. Only then can one start drawing comparisons.
The idea that the human brain is nothing but a biological CPU is a
nice analogy for high school students, but when you get down to actual
reality, the analogy is quite flawed. There have been several
*ATTEMPTS* made by people to try to formulate a processing speed
measurement of the human brain. They are all nonsense, unfortunately.
The human brain does not run on a quartz clock, and therefore does not
compute a "number of operations" per second. Its nice for undergrad
students to sit around after class and try to calculate the FLOPS that
a human brain "performs", but its all nonsense. It's mere
intellectual fun. It's science fiction. The human brain is a
parallel conglomeration of biological cells and synapses. There are
no "operations" performed therein.
If you have seen a FLOPS measurement of the human brain, that's not
actually what it's saying. That FLOPS number is calculated by
considereding what it would take to SIMULATE a human brain within a
computer, via keeping track of all the chemical interplay between
synapses. It should be deadly obvious that this is a fun, cool,
thing to think about, but it is really nothing more than that.
Neuromorphic engineering suggests that abandoning the FLOPs
architecture completely may be the key to sensory/motor coordination
and vision. A grid of simple "retinal" chips really does not have a
particular FLOPs associated with it. It's more like an analog
signal-processing device.
A human cannot sort a deck of playing cards as fast as a computer.
Indeed, a machine can perform many tasks faster and longer than any
human being, and more accurately, using exactly zero intelligence.
Optimization is an open-ended debate, because it relies not only on
the processing (brain) of the robot, but the complex relationship
between the robot's sensors and its brain. A robot with sensors that
are very different from our organs will divide the world up in
different ways than we do. And how to compare apples and oranges is
not described information-theoretically. If the robot has eyes that
see infrared, how do you go about formulating an
information-theoretic comparison to the human eye, which does not even
pick up infrared?
In summary, AI is not a matter of speed. It's a matter of
engineering, and a matter of philosophy. Our deeply-embedded cultural
ideas about the mind and the brain are challenged by the problems of
AI. Sometimes they are shown to be completely false. You have
expressed the philosophy that the human brain is a processor, much
like a CPU, with a particular measurable speed. I have argued
against this philosophy. This is why there is a newsgroup called
comp.ai.philosophy
Steve Richfie1d
>>Most chess playing programs are blind to certain hard to recognize types
>>of draws, e.g. 3X repeating position and 50 non-pawn moves, so you can
>>easily draw them if you find yourself behind in the game. Of course,
>>these programs won't even recognize the draw.
> This is so dead on. Furthermore, would a computer, thinking on its
> own, draw the following conclusion: This person gaurds the center of
> the board. So if I break the center, it will "weaken" my opponent's
> game. Certainly, no computer I know of can draw these kinds of
> humanistic inferences about chess.
Or more perverse yet, there is a class of openings known as "center
gambit" openings, where the center is literally given away, but artfully
done in a way that exacts a very high price. I first discovered these
the hard way, by playing a highly rated player, only to see him
apparently playing like a beginner, immediately moving his queen out
when I took his QP early in the game! Just how DO you play against a
highly rated player who is apparently playing just like a beginner?
Needless to say it was a short game which I did NOT win, but I sure
learned a LOT. The center gambit is definitely a good opening to have
ready to use on players who think that the center is everything. There
are several varieties of it in various chess books. There is also a
Center Gambit Declined game for those who know both sides of this approach!
> The problem is obvious. The computer has no idea that it's even
> playing a game to begin with. All it is doing is running a minimax
> search algorithm on the current state of the board. It doesn't even
> make concessions to what has happened in the game recently. Meaning
> its not even paying attention to its opponent's style.
None of the programs I have seen have paid any attention to style. They
don't seem to know what to do with the information even if they had it.
> It turns out Deep Blue did not have to rise through the tournament
> ranks to get to Kasparov. This raises some serious questions.
> Namely, how does Deep Blue fair against positional chess?
There are several RADICALLY different styles that don't work in
classical tournament play because they don't have any advantage against
proper defensive play that is cognizant of these approaches. The center
gambit mentioned above is one good example. However, I wonder how Big
Blue would do when it found itself in VERY unfamiliar situations that
may not have been entered into its making.
> (My personal thoughts is that Deep Blue and Kasparov was a very nicely
> orchestrated publicity/marketing stunt.)
I once wrote a 4x4x4 3D tic tac toe program that was VERY good. I
programmed the computer to wait as long as possible to force a win, so
players always felt like if they only had one more move that they would
have won. Several taverns in Seattle installed these for amusement, and
the tavern keepers started accepting bets that people could beat the
computer. If you carefully read the definition of "gambling", in most
jurisdictions it necessarily involves an element of chance, and the ONLY
element of chance here was the random selection between equally
advantageous moves.
The games were short enough that sometimes people won pretty much by
accident, by accidentally playing a perfect enough game to win!
One day I got a call from one of taverns, asking me to come down and
beat the program just to show his patrons that it was possible.
Apparently, no one had accidentally won. The PERFECT challenge - to beat
myself!
I knew where the weaknesses in the analysis were, so I played for them.
Even then, I had to stop for 20 minutes at one point just to analyze
where my next move should be. I won, but it was an exhausting half hour.
None of the patrons there knew that I was the creator of the game, or
that my "drinks" were all non-alcoholic. Yet another publicity stunt!
The next day, I placed a 10 minute time limit on moves.
>>In life as in chess, when you are ahead you should play to win, but when
>>you are behind you should play to draw.
> Another interesting point. In many of the games, an exasperated
> Kasparov simply called a draw with Deep Blue.
I wonder how THAT was programmed? Just what WAS the criteria for
proposing or accepting a proposed draw? Was there a human involved in
these decisions? Conventional dogma says that it takes a 3 pawn
advantage to force a win, but there are SO many exceptions to this
rule-of-thumb. I've declined a draw and continued playing obviously
drawn games in team competitions because I knew that I would have to
immediately play another game with the same opponent and he looked more
tired than I felt, so I just wanted to wear him down some more for the
next game! No, it's NOT how you play the game, it's whether you win or lose.
On a parallel note, just how much electric power does Big Blue draw? I
presume that it is a LOT more than the 100 watts or so that my own brain
draws. If Big Blue were resource limited as we humans are, I'd bet that
it wouldn't take Kasparov to beat it! You or I could easily polish it
off after it used up all of its allotted resources on the first game! To
level the playing field, we should compute how many people it takes to
draw the same amount of power, and assemble a dream team of chess grand
masters to play against Big Blue. Switching styles as the various grand
masters respond to different moves where they see a win, I wouldn't give
Big Blue much of a chance. Of course, Big blue could always add more
processors, but then the dream team could add more grand masters to
equalize the power on both sides. Doing this, I doubt whether Big Blue
could win with any number of processors!
Steve Richfie1d
Ted Dunning
> I don't think any of the research systems that have
> a higher quality of inference and better understanding of grammar have the
> same breadth of coverage of topics that the Eliza-like systems do.
> I don't think any of those systems have been submitted, for a lot of
> reasons.
You may be thinking of something different that what it sounds like
here, but didn't the (research) system from Sheffield win in 2004?
Last I heard, the key advance in doing that was the addition of
something like a topic stack. Much of the rest of the system was
lifted from existing NLP projects.
Glen M. Sizemore
GS: Yours do not seem to be. On the contrary, you seem to have merely
superimposed SMFP (Standard Mentalistic Folk Philosophy) on the problems.
Cordially,
Glen
"HMS Beagle" <bga...@microsoft.org> wrote in message
news:mqcmj1d62g9tg8pu3...@4ax.com...
Evgenij Barsukov
> On a parallel note, just how much electric power does Big Blue draw? I
> presume that it is a LOT more than the 100 watts or so that my own brain
> draws. If Big Blue were resource limited as we humans are, I'd bet that
> it wouldn't take Kasparov to beat it! You or I could easily polish it
> off after it used up all of its allotted resources on the first game! To
> level the playing field, we should compute how many people it takes to
> draw the same amount of power, and assemble a dream team of chess grand
> masters to play against Big Blue.
That is an excelent point. Energy efficiency of computation is already
giving people a competitive advantage in many jobs that could alredy
be done equaly well by computers/robots. A lot of nursing jobs are such.
It is possible to get a robot to do very accurate distribution of
medicals based on bar-codes, but they are still more expensive to
purchase/maintain than a life nurse. Same is with sorting jobs in
warenhouses. It looks like energy efficiency is going to be the last
line of defense of humans vs computers on the economic battlefield, and
one that will allow certain niche jobs for humans even after compuers
will overtake them both in serial and parallel processing power.
>Switching styles as the various grand
> masters respond to different moves where they see a win, I wouldn't give
> Big Blue much of a chance. Of course, Big blue could always add more
> processors, but then the dream team could add more grand masters to
> equalize the power on both sides. Doing this, I doubt whether Big Blue
> could win with any number of processors!
The problem with this approach paralelizing multiple humans is that
their interaction is highly inefficient (slow and prone to
computationaly expensive conflict resulution schemes). This is proven
by off-line games played by multiple masters against grossmaister, which
are usualy won by grossmaister.
Regards,
Evgenij
Lester Zick
in comp.ai.philosophy wrote:
[. . .]
>In summary, AI is not a matter of speed. It's a matter of
>engineering, and a matter of philosophy.
[. . .]
Engineering + philosophy? This strikes me as empiricism and/or
pragmatism. However I'm not necessarily arguing against either in this
context. But if true, how would you ever recognize success? In other
words how would you ever be able to recognize true ai? Because it
satisfies some philosophy?It strikes me the combination engineering +
philosophy is just the same task oriented turing philosophy for ai. Ai
is said to be acheived when some task or tasks performed by people is
performed as well or better by machines. And my impression is that
this is what you're basically arguing against.
The difficulty I see is the bridge between engineering and philosophy.
What is it the engineering is supposed to engineer? If it's just a
series of specific tasks then you're back in the chess/applications
definitions for ai. And if not then you really need some definition
for ai that doesn't just regress to engineering or philosophy.
Glen has a similar notion. Only his philosophy is behaviorism and
his engineering is training. And his primary task or application is
clear: disprove the mind and mental effects. So I'm just curious what
your objective is and whether it's a serious objective?If not chess et
al. what task or task list does your philosophy recommend engineering
implement and how will you recognize and know ai when you see it?
~v~~
Evgenij Barsukov
> On Wed, 28 Sep 2005 11:53:22 -0500, Evgenij Barsukov
> <evgenij_...@yahoo.com> wrote:
>
>>You must have missed the complete message except that your pattern
>>recognition jumped on the word "deep blue". I am saying that AI
>>development (both hardware and firmware) does not have any problems, it
>>is a thriving field that already have overtaken all other areas of human
>>activity in terms of economic output. But there is no _single_ task that
>>this field has to solve, and there is no _generic_ AI (humans do not
>>have it either).
>
>
> This whole paragraph is gobbledegook. I've never heard this argument
> that AI is defined as humans writing software for a machine and
> selling it to other humans.
While formulated wrong by you, it is still encouraging to hear your
recognition that something actually new is coming out of our discussion.
Indeed, it is extremely common misconseption, that problem of
intelligence can be limited to a black-box which has number of inputs
and number of outputs, and than upon pushing a button creates human-like
behaviour. This "target" can not be reached for a simple reason that
human themselfs do not satisfy this requirement.
Humans put in a black box from birth on do not exhibit any
different behaviour from animals. Placing with wolfs they behave more
like wolfs than like humans. It is not clear how true well known Maugly
story is, but there is well documented story of 2 year old in Argentina
who has grown up with wild dogs. He behaved very much like them, was
running on 4 feets and was not very succesful on the level of dog's
hierarchy. After being extracted at age 5, he never developed human
behavior such as language, not to mention any other treats.
Human intelligence exhibit differences from other animal intelligence
because of 3 things:
- ability to transmit optimizations verticaly (between generations)
- ability to transmit optimizations horizontaly (within a generation
- huge amount of accumultated optimizations at the present moment that
create very complex behaviour.
It is very important to have clear philosophical understanding of the
above before getting into any serious AI-related activity.
Self-learning, neural networks, predictive optimization and other things
being discussed to death in this group are all widely available and
observable in the animal kindom, and to some extent even in lower
life-forms. All these behaviour have already been sucsesfuly implemented
in commercial devices and are of cause a needed part of AI suite.
However, the only commercialy succesful exemples of AI are using
_not only_ these element, but also a decent amount of hevristics
(e.g. use prior human knowledge). This makes these AI applications
similar to individual humans in the sense that they are in no way
self-sufficient, but work as a part of a larger intelligence process.
And that is the very reason of their success.
At the other hand, approaches that fundamentaly exclude ability to
add hevristics, or make it very difficult (like neural networks)
keep faling badly on their face in commercial applications. Statistical
methods of smoothing multi-dimensional interpolation are more widely
used than neural networks for the very reason that they are more easily
modularized and combined with hevristics modules.
So to sumarize, you might want to dwell in you self-enclosed world
of "pure black-box AI", but the best you will achieve is to create
an succesful ameba-emulation or some parts to be reused in other
applications that do take the "social" or "networked" character of
advanced (e.g. human-like, human-useful) intelligence into account and
will therfore be useful for society and monetary rewarded.
> Rolf Pfeifer does no pay lip service to
> it. And that man leaves NOTHING out. We talk about AI in this
> newsgroup, but you start talking about Microsoft and Bill Gates.
>
>
>
>>Beating people at each _particular_ task will take hardware computing
>>power equivalent to that of humans, unless the computational
>>optimiziation used at that particular taks by AI is superior to that of
>>humans. This does not need any discussion, it is a basic fact of
>>information theory.
>
>
> This is complete, utter, nonsense. You cannot even start making
> information-theoretic arguments unless the performance of the task is
> done in the IDENTICAL manner as the human does it, all the way down to
> his/her neuron cells. Only then can one start drawing comparisons.
Computations in identical mater are needed for comparing something
quantitatively. But estimation what minimum amount of computations is
required to solve problem A, can be obtained in boundary of information
theory. Specialy this is true for simple problems that have been widely
researched and for which best possible optimizations are known. For
these problems in can be assumed that humans are doing computation in
the _less optimized way_ that the most optimized algorithm, and
therefore equivalent computer power used by human will be "larger or
equal" to that by CPU running the optimized algorith.
> The idea that the human brain is nothing but a biological CPU is a
> nice analogy for high school students, but when you get down to actual
> reality, the analogy is quite flawed.
If you would look closely, you would see that nobody is talking about
single CPU. The actual analogy of human brain is as a network of
millions of independent CPUs running in parallel and in communication
with each other, from which majority are units specialized
on solving specific problems.
But even these analogy is not complex enough, because as you could
review above, the key specifics of human intelligence is vertical and
horizontal interaction _between_ humans.
There have been several
> *ATTEMPTS* made by people to try to formulate a processing speed
> measurement of the human brain. They are all nonsense, unfortunately.
They have been done by well respected proffessors in the field of AI,
so with your love to authorities and what they say I am surprized you
are taken these so lightly.
One "functional" estimation that I specialy like have used known mass of
retina-cells of dragon flies as a individual sample for solving
edge-detection problems for which an optimal parallel CPU algorithm
exists and therefore equivalent parallel CPU power can be estimated.
After that estimated power have been scaled up to a brain weight.
While parallel processing power estimation is less exact because many
parts of the brain are functionaly specific and so above method gives
large overestimation for some types of generic problems, the _serial_
processing power estimation is 100% bulet proof. The rate of neuron
firing is 10 Hz, and that limits our ability to serial computations
no matter what method you are use.
> The human brain does not run on a quartz clock, and therefore does not
> compute a "number of operations" per second. Its nice for undergrad
> students to sit around after class and try to calculate the FLOPS that
> a human brain "performs", but its all nonsense. It's mere
> intellectual fun. It's science fiction. The human brain is a
> parallel conglomeration of biological cells and synapses. There are
> no "operations" performed therein.
So, if no operations are perfomed, do you prefer to thing of a human
brain as a woo-doo machine that does thing by magic? Than what is
your interest in AI?
> If you have seen a FLOPS measurement of the human brain, that's not
> actually what it's saying. That FLOPS number is calculated by
> considereding what it would take to SIMULATE a human brain within a
> computer, via keeping track of all the chemical interplay between
> synapses. It should be deadly obvious that this is a fun, cool,
> thing to think about, but it is really nothing more than that.
> Neuromorphic engineering suggests that abandoning the FLOPs
> architecture completely may be the key to sensory/motor coordination
> and vision. A grid of simple "retinal" chips really does not have a
> particular FLOPs associated with it. It's more like an analog
> signal-processing device.
>
> A human cannot sort a deck of playing cards as fast as a computer.
> Indeed, a machine can perform many tasks faster and longer than any
> human being, and more accurately, using exactly zero intelligence.
>
> Optimization is an open-ended debate, because it relies not only on
> the processing (brain) of the robot, but the complex relationship
> between the robot's sensors and its brain. A robot with sensors that
> are very different from our organs will divide the world up in
> different ways than we do. And how to compare apples and oranges is
> not described information-theoretically. If the robot has eyes that
> see infrared, how do you go about formulating an
> information-theoretic comparison to the human eye, which does not even
> pick up infrared?
In all problems where computing power was compared, it was done
based on same sensory input.
>
> In summary, AI is not a matter of speed. It's a matter of
> engineering, and a matter of philosophy. Our deeply-embedded cultural
> ideas about the mind and the brain are challenged by the problems of
> AI. Sometimes they are shown to be completely false. You have
> expressed the philosophy that the human brain is a processor, much
> like a CPU, with a particular measurable speed. I have argued
> against this philosophy. This is why there is a newsgroup called
> comp.ai.philosophy
Intelligence is data processing (any kind thereof). Data processing
is only interesting if it happens in real time. Real time = processing
rate dependence. So AI is a matter of speed (as well as matter of
what is being executed of cause). But you can not separate these two
things while comparing intelligence A (human hardware + human software)
vs intelligence B (artificial hardware + artificial software), because
comparison is always going to happen in real time.
And exactly this is important to understand on a philosophical level.
Regards,
Evgenij
Glen M. Sizemore
"Glen M. Sizemore" <gmsiz...@yahoo.com> wrote in message
news:20050929064204.132$g...@news.newsreader.com...
Carl Burke
news:1127955791....@z14g2000cwz.googlegroups.com...
The winner in 2004 was Alice, I believe. A team from Sheffield that
included Yorick Wilks did win in 1997, but as far as I know they
haven't come back for another try. I would expect to see more progress
if we could get Yorick, or people like Johan Bos, James Allen, David Traum,
or any of a few dozen other dialogue researchers into the competition
on a regular basis. We don't see that interest, though. They keep plugging
away improving the state of the art in restricted domains, and that keeps
them all pretty busy.
The idea of a grand challenge problem like this is a good one, IMHO.
But unlike, say, the DARPA Grand Challenge where robot trucks drive
across the desert and try to make it 100-odd miles on thier own,
the Loebner Prize doesn't get the academic's blood pumping.
I think they deem it to need far too much work for too litle return;
either that, or they just don't take it seriously. Kind of a Catch-22
problem there: names in computational linguistics don't compete
because none of the other names compete.
--
Carl Burke
cbu...@mitre.org
HMS Beagle
<evgenij_...@yahoo.com> wrote:
>Human intelligence exhibit differences from other animal intelligence
>because of 3 things:
>- ability to transmit optimizations verticaly (between generations)
>- ability to transmit optimizations horizontaly (within a generation
>- huge amount of accumultated optimizations at the present moment that
>create very complex behaviour.
Where are you getting this? What is your defn of "optimization"?
Do you mean heuristic?
>It is very important to have clear philosophical understanding of the
>above before getting into any serious AI-related activity.
Unfortunately its not clear. In your earlier posts you were talking
about "optimizations" in the sense that a solution uses less FLOPS, or
less energy, or so on. Are you saying humans transmit entirely
different biological solutions performed in their brains to the brains
of other humans?
>However, the only commercialy succesful exemples of AI are using
>_not only_ these element, but also a decent amount of hevristics
>(e.g. use prior human knowledge). This makes these AI applications
>similar to individual humans in the sense that they are in no way
>self-sufficient, but work as a part of a larger intelligence process.
>And that is the very reason of their success.
This argument is reasonable. But understand it is qualitatively
different from the idea of a tabula rasa robot that can "work" its way
all the way to human intelligence using reinforcement training (as
espoused by several posters around here). It is true that that AI
researchers seperate AI Solutions from Engineering Solutions. The use
of a vacuum to suck up dirt into a bag is an engineering solution.
The vacuum-cleaner robots for sale use both the engineering solution
plus some AI to navigate the particular area. Humans invented the
idea of suction through a bag to clean a carpet, thats not something
the robot figured out on its own.
> At the other hand, approaches that fundamentaly exclude ability to
>add hevristics, or make it very difficult (like neural networks)
>keep faling badly on their face in commercial applications. Statistical
>methods of smoothing multi-dimensional interpolation are more widely
>used than neural networks for the very reason that they are more easily
>modularized and combined with hevristics modules.
The best of both worlds then. Well what bothers me about
chess-playing algorithms is that they are a mathematical solution
programmed by a human. There is no sense that the computer is
LEARNING from experience.
>So to sumarize, you might want to dwell in you self-enclosed world
>of "pure black-box AI", but the best you will achieve is to create
[snip]
I'm not dwelling in any such box... although there are a few other
people who post here who may be.
>Computations in identical mater are needed for comparing something
>quantitatively. But estimation what minimum amount of computations is
>required to solve problem A, can be obtained in boundary of information
>theory. Specialy this is true for simple problems that have been widely
>researched and for which best possible optimizations are known. For
Here is that word "optimizations" again. Do you mean general
solution?
>these problems in can be assumed that humans are doing computation in
>the _less optimized way_ that the most optimized algorithm, and
>therefore equivalent computer power used by human will be "larger or
>equal" to that by CPU running the optimized algorith.
No see this is where we disagree. My brain is not storing and
manipulating floating point numbers in order to perform
"calculations". That is how a CPU works, not a brain. A brain is a
totally analog device. I'm not familiar with how to reduce an analog
device to a number of FLOPS. I was certain there is no such
reduction.
I assume an analog solution is always several orders of magnitude
faster than a digital one. Analog machines are limited only by
physics. Whereas in the digital world there is a host of mitigating
factors.
>If you would look closely, you would see that nobody is talking about
>single CPU. The actual analogy of human brain is as a network of
>millions of independent CPUs running in parallel and in communication
>with each other, from which majority are units specialized
>on solving specific problems.
Again. Nuerons are analog devices which are not storing floating
point numbers and manipulating them with math operations. What you
are suggesting is SIMULATING a neuron in each CPU in your grid. This
is fundamentally different from building an analog device which "does
the same thing" as a neuron. In any case, you would not use Pentiums
or AMDs. Those are general-purpose processors and therefore would be
sluggish compared to a CPU specially made to sum up and perform a
transfer function (and nothing else). Perhaps you are beginning
to see what I mean when I say AI is an engineering problem, not a
problem of mere lack of FLOPS. (more on this below)
> But even these analogy is not complex enough, because as you could
>review above, the key specifics of human intelligence is vertical and
>horizontal interaction _between_ humans.
I totally agree. I understand your argument about heuristics and I
concur.
>There have been several
>> *ATTEMPTS* made by people to try to formulate a processing speed
>> measurement of the human brain. They are all nonsense, unfortunately.
>
>They have been done by well respected proffessors in the field of AI,
>so with your love to authorities and what they say I am surprized you
>are taken these so lightly.
No. I already told you that what that number is, is the FLOPS needed
to *SIMULATE* a brain in a computer. For example, say I make an
analog device for a brain instead of a digital one. Then the measure
is not clear.
>One "functional" estimation that I specialy like have used known mass of
>retina-cells of dragon flies as a individual sample for solving
>edge-detection problems for which an optimal parallel CPU algorithm
>exists and therefore equivalent parallel CPU power can be estimated.
>After that estimated power have been scaled up to a brain weight.
I care little about equivalent parallel CPU power being estimated.
You are simulating the dragon fly retina, not building an analog
machine which performs the particular solution using hardware.
>While parallel processing power estimation is less exact because many
>parts of the brain are functionaly specific and so above method gives
>large overestimation for some types of generic problems, the _serial_
>processing power estimation is 100% bulet proof. The rate of neuron
>firing is 10 Hz, and that limits our ability to serial computations
>no matter what method you are use.
I'd like to see more detail on how you measure the computing speed of
an analog device. Or better yet, of a biological cell.
>So, if no operations are perfomed, do you prefer to thing of a human
>brain as a woo-doo machine that does thing by magic? Than what is
>your interest in AI?
Not at all. I do see the nueronal cell as a SIGNAL PROCESSING
DEVICE. But I do not see it as a CPU. Neurons simply do not perform
"calculations" on floating point numbers. This is not what is going
on. Neuroscience still does not know what the spike train is
encoding. (...or wether it's encoding in the first place) So yes,
in many ways there is woo-doo going on, but only in the sense that
noone knows what a neuron is doing exactly.
>> Optimization is an open-ended debate, because it relies not only on
>> the processing (brain) of the robot, but the complex relationship
>> between the robot's sensors and its brain. A robot with sensors that
>> are very different from our organs will divide the world up in
>> different ways than we do. And how to compare apples and oranges is
>> not described information-theoretically. If the robot has eyes that
>> see infrared, how do you go about formulating an
>> information-theoretic comparison to the human eye, which does not even
>> pick up infrared?
>
>In all problems where computing power was compared, it was done
>based on same sensory input.
Long long ago, and far away on distant newsgroup, I had a thread going
about 3D holographic television. We all agreed such a device would
be ANALOG not digital. Even then, those guys were able to calculate
the number of "bits" of thru-put needed for such a device. They
decided that the thru-put was so high that no machine on earth yet
could do it.
I'm not sure I agree with this calculation. Is there some way to
express the thru-put of an analog device in bits-per-second? I was
certain no such thing existed.
>Intelligence is data processing (any kind thereof). Data processing
>is only interesting if it happens in real time. Real time = processing
>rate dependence. So AI is a matter of speed (as well as matter of
>what is being executed of cause). But you can not separate these two
>things while comparing intelligence A (human hardware + human software)
>vs intelligence B (artificial hardware + artificial software), because
>comparison is always going to happen in real time.
>And exactly this is important to understand on a philosophical level.
See above.
HMS Beagle
(Lester Zick) wrote:
>Okay this statement is true. There is just no special reason to assume
>transcendental limits lie on any straight line nor together on any one
>line. That's an unjustified special assumption but is nonetheless what
>is commonly assumed in mathematics.
I don't know, Zick. Humans invented the concept of quantity to
begin with. You seem to be making an appeal to a Platonic realm of
mathematics that exists in some higher reality.
Steve Richfie1d
> I'd like to see more detail on how you measure the computing speed of
> an analog device. Or better yet, of a biological cell.
Devices in analog systems have bandwidths. In the case of neurons, this
can be highly variable, but 1 kHz at least gets you to the right order
of magnitude.
The sampling theorem says that you need at least 2 samples per cycle to
be able to capture the information that is present, which means that
your simulation loop would have to run at around 2 kHz.
An average CNS neuron has ~50,000 physical inputs, of which only ~200
actually do anything. Presuming some clever programmable logic to take
care of the "details" whatever they might turn out to be, there is a
multiply (but the synaptic constant) and an add (to accumulate them) for
each input, for
200 synapses * 2 ops/synapse * 2000 cycles/second = 0.8 MFLOP/neuron
Of course, we are making a bunch of optimistic assumptions here, like
checking the UNused inputs for changes in configuration is a relatively
small job, that learning is a small job, and that there aren't any other
"black holes" waiting for us. OK, so I'll fudge my estimate up to 1
MFLOP/neuron and hope that it is enough.
A modern Pentium PC would then be able to simulate about 2,000 CNS
neurons in real time. Of course there are ~10^11 of them, so putting a
bunch of PCs together may not get us there any time soon.
However, there are now known architectural improvements that would
probably give us 10^3 improvement over a PC, leaving us 10^8 short. If
we get an order of magnitude every 3 years, then we only have 24 years
to wait.
However, maybe only 1% of neurons are active at any one time, so we may
only be 10^6 short. Indeed, it might be possible to amass a million
processors into a single box. Of course, if we plan to do this in
another 3 years when the modified silicon is ready, then it will only
take 100,000 processors, a MUCH more possible goal. With 100 on each
circuit board (with SIP packaging), it will only take 1,000 circuit
boards. It would easily fit in a bedroom sized system.
Of course there are LOTS of complexities that haven't been addressed
here, e.g. there are ~200 different types of CNS neurons!
OK, so who is going to donate to the human simulation project so we can
get this started? A few million dollars and we'll be ready to start
testing in ~3 years. Of course it would be connected to the Internet so
that that anyone could test out their neuron models. Any interest anyone?
Steve Richfie1d
Curt Welch
> On a parallel note, just how much electric power does Big Blue draw? I
> presume that it is a LOT more than the 100 watts or so that my own brain
> draws.
Yeah, no doubt the brain is a lot more power effecent.
> If Big Blue were resource limited as we humans are, I'd bet that
> it wouldn't take Kasparov to beat it! You or I could easily polish it
> off after it used up all of its allotted resources on the first game! To
> level the playing field, we should compute how many people it takes to
> draw the same amount of power, and assemble a dream team of chess grand
> masters to play against Big Blue. Switching styles as the various grand
> masters respond to different moves where they see a win, I wouldn't give
> Big Blue much of a chance. Of course, Big blue could always add more
> processors, but then the dream team could add more grand masters to
> equalize the power on both sides. Doing this, I doubt whether Big Blue
> could win with any number of processors!
I'd guess that big blue would make better use of it's added power than a
1000 grand masters trying to debate what move to make next. They probably
have no experience trying to play chess by commitee and would do worse than
any of them playing alone.
--
Curt Welch http://CurtWelch.Com/
cu...@kcwc.com http://NewsReader.Com/
forbi...@msn.com
> On 27 Sep 2005 06:11:26 -0700, forbi...@msn.com wrote:
> [snip]
> >overhead cycles. What percentage of processing power do you believe
> >will be eated up by the emulation process as opposed to the human
> >simulation process?
>
> always interrupted by the OS checking for a mouse movement, and
> constantly interrupted checking for other things that help draw the
> windows in order.
It seems to me that this is they way it should be since these are
the primary points of interaction in these entities' domain. Once
the processing demands of these first order communication devices
are sated (and here I'm talking about the communication speeds of
humans using these devices) then excess computing power can be
"wasted" on other sense and effector systems humans find useful.
Now that I think about it, the internet is another sense/effector
system for computers to communicate with each other. It's very
limited right now. Lots of processing power could be devoted to
tasks related to computer to computer communications but the
channels themselves place huge limits on the useful methods and
processing needs right now based upon what else we have sufficient
computing power to accomplish with the data communicatable.
(Man, that's one ugly sentence. I hope my disjointed thought is
in there somewhere.)
> Further, an Intel or AMD CPU in a PC is a general-purpose processor.
> For this reason, its many orders of magnitude slower than a machine
> specifically built to perform a specific calculation.
or process. "Calculation" may frame one's thoughts in ways that
aren't useful.
> What is the "overhead" then? Its really bad. The computer is mostly
> performing overhead. There is an enormous waste of calculations.
I don't think they are a waste. Instead they are the real time
sensors and effectors and should take the computing power necessary
to support the human/computer communications channel. We don't have
near enough computing power to support real time aural communications
even though some systems are limping along in this area right now.
Lester Zick
Nor do I know, Bagel. I just know more than you when it comes to
arithmetic reductions to geometry. If you want to appeal to non
geometric foundations for numbers I'd be interested to see the
regression. If it's the standard reduction to axiomatic assumtions
like suc( ) you're still left with assumptions and you would appear
to be the one appealing to philosophy and higher mystic realities.
By the way you still haven't responded to my query re engineering and
philosophy. What is it that you expect engineering to be engineering
exactly?
~v~~
amnon....@textanalysis.com
> On 27 Sep 2005 23:22:07 -0700, "am...@textanalysis.com"
> <amnon....@textanalysis.com> wrote:
> >HMS Beagle wrote:
> >> On Tue, 27 Sep 2005 07:24:04 -0500, Evgenij Barsukov
> >> <evgenij_...@yahoo.com> wrote:
> >> >> Ok here we go again with the argument that acheiving AI is a simple
> >> >> matter of computer power. How many times do we have to go over
> >> >> this?
> >> >
> >> >Until you will finaly get it?
> >> >
> >>
> >> Are you openly declaring in this newsgroup, with everyone watching,
> >> that you (not only beleive) but are certain that all problems in AI
> >> all reduce to lack of processing speed?
> >
> >The main evidence available is in the realm of games such as Chess. It
> >was once thought that only intelligent machines could beat the best
> >humans. Now we know that relatively brute-force algorithms combined
> >with processing speed are sufficient to beat top Grandmasters.
>
> I really don't think Chess is a domain that even matters to
> intelligence. It's like saying computers are more intelligent than
> humans because they can sort lists faster than we can.
>
I'm ready to concede that sorting lists is an intelligent activity.
And if I could play as well and as quickly as Deep Blue with the same
algorithms it uses, I'd be considered a Chess genius.
>
> >There's no reason to think that other problem spaces won't succumb,
> >given further multiples in processing speed and parallelization.
>
> "There's no reason to think" is an article of faith. I think if you
> look at what I actually asked, I didn't concede to this.
>
"There's no reason to think" is not an article of faith. It follows
the concrete example that preceded it and, implicitly, the other
examples out there in which computers exceed the capabilities of
humans.
There's a cliche that anything computers can do no longer involves
intelligence. So the definition of intelligence is expected to keep
shifting as computers become more capable. To my mind, there's also a
semantic issue involving "intelligence" vs "human intelligence." I
regard much of what computers do as intelligent. They store, retrieve,
and manipulate infomation. Seems pretty smart to me -- much smarter
than a rock.
rolloca...@gmail.com
>The program does that in order to divert your attention from the
>fact that it is as stupid as a doorstop. It doesn't understand a
>word of what you say and therefore, when you say something that
>doesn't trigger one of the canned responses, is has no way other
>than to change the subject or output some "cute" non-sequitur.
The program never does anything "in order to divert attention", or
indeed "in order to" do anything. Every response you see has been
learnt from user interactions, and is used when found to be most nearly
in context, taking into account the whole conversation.
>Its very easy to verify this if you tell it something and then later
>ask a question whose answer requires knowledge of what you said
>earlier. For instance if you tell it "My favourite colour is green"
>and then later ask it "Do you remember what my favourite colour is?"
>chances are that it will reply with something to disguise the fact
>that it has no idea of what you're talking about.
If it fails to answer pointless questions like those it is because not
that many people ask them. Hence it doesn't ask them. Hence it
doesn't learn the answer. Eventually it will answer your question
perfectly.
I fully agree, incidentally, that the Loebner Prize logs were not very
good. Judges act outside the realms of its normal environment - online
chat.
>As I've suggested in an earlier post, this state of affairs could be
>changed if only programs that fulfilled two new requirements would
>be able to compete: One of the requirements is that the program must
>be able to learn. The second requirement is that a program must be
l>anguage independent.
The program does learn. The program is language independent. Try it.
The program, incidentally, is called JabberwAcky, so is found at
www.jabberwacky.com, with no o's in sight.
Steve Richfie1d
>>>There's no reason to think that other problem spaces won't succumb,
>>>given further multiples in processing speed and parallelization.
>>"There's no reason to think" is an article of faith. I think if you
>>look at what I actually asked, I didn't concede to this.
> "There's no reason to think" is not an article of faith. It follows
> the concrete example that preceded it and, implicitly, the other
> examples out there in which computers exceed the capabilities of
> humans.
> There's a cliche that anything computers can do no longer involves
> intelligence. So the definition of intelligence is expected to keep
> shifting as computers become more capable. To my mind, there's also a
> semantic issue involving "intelligence" vs "human intelligence." I
> regard much of what computers do as intelligent. They store, retrieve,
> and manipulate infomation. Seems pretty smart to me -- much smarter
> than a rock.
This works just like religion.
First, God was in the sky just out of reach - SO close that people
attempted to build towers to reach him. Then came aircraft, then
spacecraft, so he must be even further away, eventually to outside of
our solar system, outside our galaxy, outside of our universe, etc.
Before computers, your soul was simply your consciousness and memories.
Then, people started talking about capturing these within a computer,
and suddenly your soul was something ELSE.
A variety of games of skill were used as early measures of intelligence.
For those newly to this "game", checkers was established as a measure
before chess ever was, but it then took less than a year for the first
good checker-playing program to be written.
From my own point of view, people just aren't very intelligent. Dr.
Eliza, an admittedly simple NLP program, has a demo that consistently
outperforms board certified physicians in dealing with complex cases of
chronic illness - and this is usually on the basis of rational deductive
process rather than encyclopedic memory. If the best and the brightest
who can survive medical school just aren't that bright, then what does
that say about the rest of humanity?
It appears that there is a hierarchy of gods, which when their limits
are challenged by a lesser god, the challenge is summarily dismissed
rather than considering the possibility that gods may not be all that
godly. First there is God, then there are the lesser gods like you and
I, then much lower are AI computers, and below computers - maybe insects
and bacteria.
I watched to a program on CNN just after Hurricane Katrina where a group
of holy men of various persuasions were asked why God would have caused
or allowed such a thing to happen. "That wasn't God, that was the
weather." was their unanimous reply in various words, as though God
created and controls the universe yet had no control over the SAME
weather that some people were even claiming was under terrorist control!
Does that now put terrorists above God as a force beyond God's control?
Amnon is absolutely right here. We need to put SOME objective measure of
intellectual capability that is equally applicable to God, lesser gods
like us, and AI computers, which the Turing test CLEARLY fails to provide.
Programs to verbally discuss your domain-independent problems with you
and reason their way to solutions aren't that far off, as Dr. Eliza
appears to be eventually extensible to that end (with a *LOT* more
work). Even at that level of skill, Dr. Eliza would STILL flunk the
Turing test, just like the computer on the original Star Trek series
would, because it would poorly simulate people's illogical "thought"
processes. Indeed, Dr. Eliza is already a bit frustrating to some users,
who erroneously perceive many of its questions to be off-topic when in
fact they go to the very heart of people's problems. Of course, if these
same people truly understood their own problems, then they would have
already solved them!
You may now take your proper place in the Hierarchy of Gods and create
even lesser gods - AI computers!
Steve Richfie1d
Ted Dunning
I was under the strong impression that it had been demonstrated that
substantial computation occured in brain neurons without firing, i.e.
in the axons themselves. That means that the complexity of the
behavior of a collection of neurons is vastly higher than just a
multiply per connection and an accumulate per neuron. Hebbian neurons
are only scratching the surface of real neural behavior.
This complexity is why it takes a serious amount of computational power
to do even the cockroach motor simulations (with only a few dozen
neurons).
Steve Richfie1d
> I was under the strong impression that it had been demonstrated that
> substantial computation occured in brain neurons without firing, i.e.
> in the axons themselves.
There is only one axon but 50K dendrites, so whatever happens there
can't be a computational problem. However, there IS a lot more happening
as you suggest, but the affect on the computational overhead is still
unknown.
> That means that the complexity of the
> behavior of a collection of neurons is vastly higher than just a
> multiply per connection and an accumulate per neuron.
I don't think "vastly", though there may be a factor or 2 or 3 left to
contend with in a clever design.
> Hebbian neurons are only scratching the surface of real neural behavior.
Agreed. However, simulation overhead doesn't necessarily go up linearly
with complexity in clever designs.
> This complexity is why it takes a serious amount of computational power
> to do even the cockroach motor simulations (with only a few dozen
> neurons).
Many present simulation efforts use SPICE and other electric circuit
simulators. These work to a "perfect" solution each time slice using
linear programming methods whose computations increase as the SQUARE of
circuit complexity! Obviously these methods aren't scalable very far. In
a sense, they are applying corrections to the same time slice rather
than the next one where they would work almost as well. Hence, the time
taken isn't necessarily any indication at all of the difficulty of doing
the job rather than EXACTLY simulating a particular model.
Steve
HMS Beagle
(Lester Zick) wrote:
>>I don't know, Zick. Humans invented the concept of quantity to
>>begin with. You seem to be making an appeal to a Platonic realm of
>>mathematics that exists in some higher reality.
>
>Nor do I know, Bagel. I just know more than you when it comes to
>arithmetic reductions to geometry. If you want to appeal to non
>geometric foundations for numbers I'd be interested to see the
>regression. If it's the standard reduction to axiomatic assumtions
>like suc( ) you're still left with assumptions and you would appear
>to be the one appealing to philosophy and higher mystic realities.
Not at all. We say the axioms are agreed upon by humans who do math.
We do not invoke the platonic realm whatsoever. Do we do it because
we are mystics? No. We do it because in math we like to keep track
of the truth and falsity of each sentence as we go along. (This is
called "formal" in the university lingo.)
You are the one claiming that number theoretic axioms (eg 'suc()') are
assumptions that are "baseless" because they are not grounded in
geometry. Your appeal to geometry as being more true/more
apparent/better/undeniable/more grounded/etc is precisely an appeal
to a Platonic realm of perfect forms.
>By the way you still haven't responded to my query re engineering and
>philosophy. What is it that you expect engineering to be engineering
>exactly?
This query is not part of this thread. I talk in too many newsgroups
and chatrooms to keep track of who has asked me what and where.
Please point me in the right direction and I will answer promptly.
Lester Zick
in comp.ai.philosophy wrote:
>On Fri, 30 Sep 2005 19:15:00 GMT, lester...@worldnet.att.net
>(Lester Zick) wrote:
>>>I don't know, Zick. Humans invented the concept of quantity to
>>>begin with. You seem to be making an appeal to a Platonic realm of
>>>mathematics that exists in some higher reality.
>>
>>Nor do I know, Bagel. I just know more than you when it comes to
>>arithmetic reductions to geometry. If you want to appeal to non
>>geometric foundations for numbers I'd be interested to see the
>>regression. If it's the standard reduction to axiomatic assumtions
>>like suc( ) you're still left with assumptions and you would appear
>>to be the one appealing to philosophy and higher mystic realities.
>
>Not at all. We say the axioms are agreed upon by humans who do math.
I do math.
>We do not invoke the platonic realm whatsoever.
You invoke the mystic realm of guesses.
> Do we do it because
>we are mystics? No.
Sure you do. Because you can't reduce your claims to truth.
> We do it because in math we like to keep track
>of the truth and falsity of each sentence as we go along.
Truth and falsity? Is this a joke? You have no reduction of your
claims to true/false. Your reduction like those of all empiricism
is only to not inconsistent/ inconsistent.
> (This is
>called "formal" in the university lingo.)
>
>You are the one claiming that number theoretic axioms (eg 'suc()') are
>assumptions that are "baseless" because they are not grounded in
>geometry.
Nothing of the kind. You're the one regressing your claims to
convention not truth.
> Your appeal to geometry as being more true/more
>apparent/better/undeniable/more grounded/etc is precisely an appeal
>to a Platonic realm of perfect forms.
I love it when people explain myself to me instead of explaining
themselves to me.
>>By the way you still haven't responded to my query re engineering and
>>philosophy. What is it that you expect engineering to be engineering
>>exactly?
>
>This query is not part of this thread. I talk in too many newsgroups
>and chatrooms to keep track of who has asked me what and where.
>Please point me in the right direction and I will answer promptly.
Allow me to refresh your rather limited STM, idiot. Obviously you have
too much to say on too many newsgroups to keep track of posts that
embarrass you. Below is my reply to you in this newsgroup and thread:
On Thu, 29 Sep 2005 14:56:14 GMT, lester...@worldnet.att.net
(Lester Zick) in comp.ai.philosophy wrote:
>On Wed, 28 Sep 2005 20:54:34 -0400, HMS Beagle <bga...@microsoft.org>
>in comp.ai.philosophy wrote:
>
>[. . .]
>
>>In summary, AI is not a matter of speed. It's a matter of
>>engineering, and a matter of philosophy.
>
>[. . .]
>
>Engineering + philosophy? This strikes me as empiricism and/or
>pragmatism. However I'm not necessarily arguing against either in this
>context. But if true, how would you ever recognize success? In other
>words how would you ever be able to recognize true ai? Because it
>satisfies some philosophy?It strikes me the combination engineering +
>philosophy is just the same task oriented turing philosophy for ai. Ai
>is said to be acheived when some task or tasks performed by people is
>performed as well or better by machines. And my impression is that
>this is what you're basically arguing against.
>
>The difficulty I see is the bridge between engineering and philosophy.
>What is it the engineering is supposed to engineer? If it's just a
>series of specific tasks then you're back in the chess/applications
>definitions for ai. And if not then you really need some definition
>for ai that doesn't just regress to engineering or philosophy.
>
>Glen has a similar notion. Only his philosophy is behaviorism and
>his engineering is training. And his primary task or application is
>clear: disprove the mind and mental effects. So I'm just curious what
>your objective is and whether it's a serious objective?If not chess et
>al. what task or task list does your philosophy recommend engineering
>implement and how will you recognize and know ai when you see it?
>
>~v~~
>
~v~~
Curt Welch
for a building.
Programmable computers can have their hardware dynamically reconfigured to
perform many different functions, and the configuration is specified in the
software we right. But in the end, it's all hardware in the computer. THe
"soft" part of the hardware is simply the part that can be reconfigured.
A configuration of the window in a car is software because it can be
reconfigured into different hardware configurations which then determines
the function of the car. With the window up, the car will deflect a ball
thrown at the window. With the window down, the car has been programmed to
"catch" the ball.
All machines are in fact "programmable" because you can change their
configuration. A pen for exmaple might have its cap on, or its cap off.
You can program it to place ink on the paper by taking the cap off, and you
can program it to no put ink on the paper, by putting the cap on.
Software is not in the machine. All machines are all hardware. Software
is in the mind of the beholder. It's how we choose to talk about the
machine and not really a feature of the machine itself.
The recent talk about whether the brain has hardware or software in it is
just silly and misguided. The brain changes its behavior over time so we
know it is able to change its configuration in very complex ways.
Understanding how its configuration can change, and why, is what we need to
uncover to understand the brain. Calling it "hardware" or "software" is of
no more help than taking about the software "in" a car window.
"Claudio Grondi" <claudio...@freenet.de> wrote:
> It is maybe not a response adequate to the subject
> of the posting it is placed below, but I couldn't resist
> the temptation to mention it in that context:
>
> I see currently quite clearly, that distinguishing between
> hardware and software is one of the concepts which
> creates by far more confusion, that it helps to resolve.
>
> The concept of software as a kind of immaterial idea
> is in my eyes one of a fundamental barriers to overcome
> before it becomes clear what is really is and how it
> works. It is very surprizing for me, that this concept is
> so widespread and supported in the current education
> and that noone seriously questions if it is really useful for
> understanding of matters related to computer and
> computation.
> I would be glad to hear opinions of posters to this
> thread or people reading it on the subject if it makes
> any sense to distinguish what is commonly understood
> as software and hardware.
>
> Claudio
--
Claudio Grondi
"Curt Welch" <cu...@kcwc.com> wrote in
news:20051003175032.561$P...@newsreader.com...
>Software is just a description of the hardware. It's like the blueprint
>for a building.
not if there is no hint what makes it similar and what different. Statements
"something is like something other" are more or less always true, or can you
imagine two different somethings that don't share any characteristics?
In my eyes software is much less a blueprint for a building, than a user
manual delivered with an appliance. The blueprint for a building is a plan
for creating a construction, a shape out of given materials, the goal of
software is not to create new shapes of matter, but provide a receipe for
some kind of transformation between data input and output.
>
>Programmable computers can have their hardware dynamically reconfigured to
>perform many different functions, and the configuration is specified in the
>software we right. But in the end, it's all hardware in the computer. THe
>"soft" part of the hardware is simply the part that can be reconfigured.
It seems, that according to the given definition the surface of a harddrive,
or the RAM is called 'software' here. At least I see no relationship to my
understanding of what is meant when others speak about 'software';
the confusion is perfect ...
>
>A configuration of the window in a car is software because it can be
>reconfigured into different hardware configurations which then determines
>the function of the car. With the window up, the car will deflect a ball
>thrown at the window. With the window down, the car has been programmed to
>"catch" the ball.
>
>All machines are in fact "programmable" because you can change their
>configuration. A pen for exmaple might have its cap on, or its cap off.
>You can program it to place ink on the paper by taking the cap off, and you
>can program it to no put ink on the paper, by putting the cap on.
>
>Software is not in the machine. All machines are all hardware. Software
>is in the mind of the beholder. It's how we choose to talk about the
>machine and not really a feature of the machine itself.
When "Software is in the mind of the beholder", the question where it is
there and what is it actually comes inevitably to mind leading to
discussions 'whether the brain has hardware or software', isn't it that way?
>
>The recent talk about whether the brain has hardware or software in it is
>just silly and misguided. The brain changes its behavior over time so we
>know it is able to change its configuration in very complex ways.
>Understanding how its configuration can change, and why, is what we need to
>uncover to understand the brain. Calling it "hardware" or "software" is of
>no more help than taking about the software "in" a car window.
Thank you Curt Welch very much for your response!
I would be glad to get some more responses also from
other people. It could help to strenghten or dismiss my assumption,
that what people understand talking about 'software' is much more
different between the individuals than in case of other concepts like
input, output, data, processing, file, etc.
Claudio
forbi...@msn.com
Claudio Grondi wrote:
> I would be glad to hear opinions of posters to this
> thread or people reading it on the subject if it makes
> any sense to distinguish what is commonly understood
> as software and hardware.
OK, Software is the part of the data transformation algorithm
identified by state of the hardware rather than by the hardware
itself as identified stateless.
To my mind it is much easier to identify hardware and software
than software and data. E-Prom, flash memory, blurs things a
bit because the hardware is alternately defined as including or
not including the configuration of the memory. I guess this goes
to Curt's point that the configuration of atoms on a hard disk
can alternatively be thoought of as hardware or state depending
upon one's perspective, for instance, the particulars of the
implementation in a turnkey system are not visible from the outside
and so it may be thought of as hardware.
Amdahl thought that firmware+hardware could be more efficiently
implemented as hardware was correct but IBM made 370s out of 360s
by changing the firmware and the results were more efficient than
370 emulation on Amdahl's 360 machines.
Burroughs took firmware one step further with the B1700 series where
the machine was redefined on the fly for different "S" machines based
upon the software technology used, though only two could be loaded at
a time and one had to be the "native mode" "firmware" used by the
MCP(the os).
Curt Welch
> "Curt Welch" <cu...@kcwc.com> wrote in
> news:20051003175032.561$P...@newsreader.com...
> >Software is just a description of the hardware. It's like the blueprint
> >for a building.
> In my eyes software is much less a blueprint for a
> building, than a user manual delivered with an appliance. The blueprint
> for a building is a plan for creating a construction, a shape out of
> given materials, the goal of software is not to create new shapes of
> matter, but provide a receipe for some kind of transformation between
> data input and output.
Yes that's true. Software is a description of the processes and generally
ignores the description of the physical implementation (we don't talk about
the electrons in the ram cells or pits on the CD-ROM, or the transistors
used to implement the addition function we specify in the software). A
building blueprint is mostly a description of the physical implemenation
and mostly ignores processes (because most of a building is static and
doesn't move).
However, my point about them being the same is that they are both nothing
more than language. It's a langauge which describes some subset of the
properties of a physical hunk of matter. Though the software is a
description of the relationships between the interaction of lots of complex
dynamic moving parts in a computer (mostly electrons), the blueprint is a
description of the spatial relationships betten the parts that make up the
building.
An electrical schematic is another langauge we have created to describe
important properties of our machines. But it's somewhere in the middle
between talking about spatial relationships, and talking about process.
It's abstract like software in the fact that the pictures we draw don't
have a one to relationships to the physical things we talk about. Resitors
don't in fact look like little jagged lines in real life. And the size and
location of the parts as drawn in the schematic don't match the spatial
size and relationships of the parts in the actually machine (most the
time). What the schematic allows us to understand (other than the
interconnections of the parts), is the process that happens when all the
components interact with one another. It shows us the flow of the
interaction between components. So an electical schematic is a lagnage
with both spatial information and process information about the machine.
Software is just langauge. When we read it, or write it, the real
"meaning" of the langauge is in our head, not on the paper, or on the
screen. When we run our programs, we are in fact, causing our computer to
physically reconfigure itself into a different type of machine - one which
matches our software description of how the machine should act.
In order to understand the real meaning of what I'm saying, we have to
define what a language really is. But in order to understand that, you
have define what the mind and brain is. And well, that's where no one
seems to be able to agree, so we don't have an answer to all this which is
accepted as fact by the society. So we each have to find our own answers
for these things for now.
John Carmack
<St...@NOSPAM.smart-life.net> wrote:
>HMS Beagle,
>
>> I'd like to see more detail on how you measure the computing speed of
>> an analog device. Or better yet, of a biological cell.
>
>Devices in analog systems have bandwidths. In the case of neurons, this
>can be highly variable, but 1 kHz at least gets you to the right order
>of magnitude.
That's nice. But HOW DO YOU CALCULATE this bandwidth given some analog
device?
>OK, so who is going to donate to the human simulation project so we can
>get this started? A few million dollars and we'll be ready to start
>testing in ~3 years. Of course it would be connected to the Internet so
>that that anyone could test out their neuron models. Any interest anyone?
I was specifically ruling out *simulating* neurons with a
general-purpose CPU like a pentium. I have little interest in such
things. I'd like to see specific circuits that compute what we think
a neuron is doing. Ideally they would be analog. But if they must be
digital, so be it.
But before we choose that, lets first discuss what the "speed" of an
analog/biological cell is, or wether it even makes sense to talk
about the "speed" of an analog device.
HMS Beagle
<St...@NOSPAM.smart-life.net> wrote:
>HMS Beagle,
>
>> I'd like to see more detail on how you measure the computing speed of
>> an analog device. Or better yet, of a biological cell.
>
>Devices in analog systems have bandwidths. In the case of neurons, this
>can be highly variable, but 1 kHz at least gets you to the right order
>of magnitude.
That's nice. But HOW DO YOU CALCULATE this bandwidth given some analog
device?
>OK, so who is going to donate to the human simulation project so we can
>get this started? A few million dollars and we'll be ready to start
>testing in ~3 years. Of course it would be connected to the Internet so
>that that anyone could test out their neuron models. Any interest anyone?
I was specifically ruling out *simulating* neurons with a
general-purpose CPU like a pentium. I have little interest in such
things. I'd like to see specific circuits that compute what we think
a neuron is doing. Ideally they would be analog. But if they must be
digital, so be it.
But before we choose that, lets first discuss what the "speed" of an
analog/biological cell is, or wether it even makes sense to talk
about the "speed" of an analog device.
Steve Richfie1d
> That's nice. But HOW DO YOU CALCULATE this bandwidth given some analog
> device?
All analog devices "roll off" at high frequencies to avoid various
instabilities. The point where you say "There, THAT is the bandwidth" is
a political rather than a mathematical issue, though there are various
(political) formulas (as opposed to equations) to compute bandwidth. For
example, many people refer to the 3db down point (in the case of
oscilloscopes or HiFis) or the 6db down point (in the case of op amps
and other devices intended for negative feedback).
>>OK, so who is going to donate to the human simulation project so we can
>>get this started? A few million dollars and we'll be ready to start
>>testing in ~3 years. Of course it would be connected to the Internet so
>>that that anyone could test out their neuron models. Any interest anyone?
> I was specifically ruling out *simulating* neurons with a
> general-purpose CPU like a pentium.
THAT is where some of those millions would go.
> I have little interest in such
> things. I'd like to see specific circuits that compute what we think
> a neuron is doing. Ideally they would be analog.
Back to the Harmon Neuron?!
> But if they must be digital, so be it.
If you want any accuracy, you'll have to settle for a computer program.
> But before we choose that, lets first discuss what the "speed" of an
> analog/biological cell is, or wether it even makes sense to talk
> about the "speed" of an analog device.
Sure it makes sense. As I explained earlier in this thread, you'll have
to recompute at ~2kHz to do the same job.
But, suppose I am off a bit. It just means that a given simulation
processor in a large cluster will be able to simulate a different number
of neurons at real time speeds, thereby increasing/decreasing the size
of the needed cluster. There are SO many other unknowns and this is one
of the smaller ones.
For example, NO ONE yet knows what might be involved in simulating glial
cells, which make up ~90% of our brain's cells! They compute, yet aren't
even (technically) neurons.
Obviously, we aren't ready for the full blown hardware yet.
Steve Richfie1d
d
Evgenij Barsukov
> On Thu, 29 Sep 2005 19:05:08 -0600, Steve Richfie1d
> <St...@NOSPAM.smart-life.net> wrote:
>
>
>>HMS Beagle,
>>
>>
>>>I'd like to see more detail on how you measure the computing speed of
>>>an analog device. Or better yet, of a biological cell.
>>
>>Devices in analog systems have bandwidths. In the case of neurons, this
>>can be highly variable, but 1 kHz at least gets you to the right order
>>of magnitude.
>
>
> That's nice. But HOW DO YOU CALCULATE this bandwidth given some analog
> device?
The whole expression "bandwidth" was initialy used for analog devices,
even before digital devices existed. It defines how much information
can be tranported through analog copper pair given certain level of
noise, and low-pass cut-off of the line. Quantitative answer for first
part (noise) is given by Shannon theorem, and for second case (low-pass
cut-off at given electric characterists) given by the "telegraph equation".
Note that for systems that transmit not electromagneticaly but by
different means (say accusticaly, through diffusion of chemical species,
or electrochemicaly as it is in neuron) there is an equivalent of
telegraph equation because their transmission is governed by
mathematicaly analoguos differential equations. Result of it allows to
show how fast a wave with maximal available frequency above the cut-off
propagates through the wave-guide (be it a copper pair or anything else).
So bandwith of each particular type neuron is well defined and is
routinely measured by bio-physists. There is a whole
brach of bio-physics called neuron impedance spectroscopy who is
basicaly parametrizing dymanic performance of neurons.
Low-pass cut-off of different neurons range from a kHz to 100 Hz, so
it is not nearly as fast as electrical networks. There is a good reason
why. Signal is transmitted by sequential electrochemical excitation of
the cells of synapse wall, a process that involves diffusion particular
molecules near the wall that causes "firing" of the next nearby segment.
This process is not awfuly fast.
Except for bandwith, there is another quantity that limits rate
capabilty of neurons, namely it is the "cool off" time. After every
firing of a electrochemical excitation wave, a neuron need to recover
lost chemical energy before firing again, which is very time consuming
process. It takes about 0.1 to 0.01 sec depending on the type of neuron.
This depends the frequency with which impulses can be fired. These
frequency for brain neurons is about 10 Hz. No matter how fast the
signal after that propagates through the synapses (which is also not
very fast as you can see above) the ultimate rate of "computation" is
limited by the cool off time.
Alas, chemical processing has its inherent rate limits. It is very
stable, very robust and noise-protected, but it is slow.
>>OK, so who is going to donate to the human simulation project so we can
>>get this started? A few million dollars and we'll be ready to start
>>testing in ~3 years. Of course it would be connected to the Internet so
>>that that anyone could test out their neuron models. Any interest anyone?
>
>
> I was specifically ruling out *simulating* neurons with a
> general-purpose CPU like a pentium. I have little interest in such
> things. I'd like to see specific circuits that compute what we think
> a neuron is doing. Ideally they would be analog. But if they must be
> digital, so be it.
>
> But before we choose that, lets first discuss what the "speed" of an
> analog/biological cell is, or wether it even makes sense to talk
> about the "speed" of an analog device.
Analog devices have inherent speed limitations as shown above.
That is why sometime in the 60ies, where analog "co-processors" for
solving particular often occuring tasks were created, they were
limited to that time RF technology. They still depend on the frequency
that can propagate through the wires. More succesful recent analog
computers are based on optics rathere than RF. Light waveguides has
dramaticaly higher bandwidth than anything else, and there are already
light based analog computer that can do FFT faster than aniting else on
earth. So, analog computing definetely has some future (while optical
digital systems will also soon follow).
But this has nothing do do with bio-systems which operate at a
very limited bandwitdh of electrochemical transmition method.
Regards,
Evgenij
Curt Welch
> On Thu, 29 Sep 2005 19:05:08 -0600, Steve Richfie1d
> <St...@NOSPAM.smart-life.net> wrote:
>
> >HMS Beagle,
> >
> >> I'd like to see more detail on how you measure the computing speed of
> >> an analog device. Or better yet, of a biological cell.
> >
> >Devices in analog systems have bandwidths. In the case of neurons, this
> >can be highly variable, but 1 kHz at least gets you to the right order
> >of magnitude.
>
> That's nice. But HOW DO YOU CALCULATE this bandwidth given some analog
> device?
The term bandwidth came from analog devices. You measure the width of the
signal band which the device can pass by substracting some measure of the
high frequency limit of the device from the low frequency limit. TV
signals have a bandwidth of about 4 mHz if I remember. AM radio has a
bandwidth of 10 kHz and FM radio has a bandwidth of 200 kHz.
The unit of measure is Hz, not bits per second.
What we call bandwidth in computers is a related, but a different measure.
It's the measure of the numer of unique symbols a device and transmit.
Analog signals have two important dimensions. Bandwidth is one of them,
but dynamic range is the other. On a frequency spectrum graph, the
bandwidth is the width, and the dynamic range is the height of the graph.
The dynamic range is the difference between the highest level signal the
device can pass (strongest signal), and the weakest signal the device can
pass before the signal is lost in the noise floor of the device. It's
measured on a log scale in units of dB.
The number of bits per second an analog device can pass is directly related
to the area of the frequency spectrum curve. If you want to encode all the
useable information in an analog signal into the binary domain, you can do
that by sampling the analog signal at the Shannon limit rate which is twice
the rate of the bandwidth of the signal. So a 10 kHz bandwidth signal
needs to be sampled 20,000 times per second.
The number of bits in each sample determing the dynamic range of the analog
signal which will be mapped into the binary domain. Each bit adds a fixed
number of dBs to the dynamic range of the signal.
The problem here is that any analog device has a huge bandwidth and dynamic
range. Most if it is just not used in practial applications. If you
convert an analog signal to digial, you throw a huge amount of information
which is just not useful for the application. For example, in a audio
signal which is digitized, the wire comming from a device like a
microphone, has a huge bandwidth and dynamic range to do (far greater than
what we use for an audio signal), and that microphone cable will be
carrying information about all the local TV stations, and radio stations
and cell phones and thunder storms in the area, but it's all down in what
is seen as the noise floor of the signal and which we intentially filter
out for audio applications.
So, if you look at the signal coming from the microphone, we are only
interested in the data which is in the audio spectrum, but the wire in fact
carries far more information. So you can talk about how many bits per
second of information we are interseted in for the application of recording
sound, or you can attempt to talk about the amount of information actually
in the wire, which is nearly infinite by comparison.
If you look at an isolated device like a wire, or a neuron, the true total
information content in the analog nature of the device, is bascially
infinite if you try to talk about it in binary terms. But for a given
application, the data that is imporant to create the correct function of
the device is much lower.
You can't answer the question of how much binary bandwidth a given
application needs from the analog device, until you understand the
application. And no one understands the brain well enough to answer that
question.
So, the digital bandiwdth of any analog device is infinte, but the signal
content needed for any specific applicaiton, is normally finite and small.
But no one understand the brain well enough to know what the required
digital bandwidth of that application is.
> >OK, so who is going to donate to the human simulation project so we can
> >get this started? A few million dollars and we'll be ready to start
> >testing in ~3 years. Of course it would be connected to the Internet so
> >that that anyone could test out their neuron models. Any interest
> >anyone?
>
> I was specifically ruling out *simulating* neurons with a
> general-purpose CPU like a pentium. I have little interest in such
> things. I'd like to see specific circuits that compute what we think
> a neuron is doing. Ideally they would be analog. But if they must be
> digital, so be it.
>
> But before we choose that, lets first discuss what the "speed" of an
> analog/biological cell is, or wether it even makes sense to talk
> about the "speed" of an analog device.
And here's the next problem. If you encode information in pulse signals,
you are working in a signal format which is half analog, and half digital.
So it's a different beast from both our normal understand of analog
siganls, and digital signals. Pulse signals encode all their information
in the time domain (i.e. frequency domain), but have 0 dynamic range. The
bandwidth of a pulse signal is in theory infinite, but the practical
application always has a finite limit based on the resolution accuracy the
application needs to measure time with.
Once again, if you try to convert a pulse signal to the digial domain, the
number of bits you need is infinite, unless the application allows you to
use less. And once again, since no one yet understand enough about the
brain to allow us to define the needs of this application, we have no way
of knowing how many of those infinite bits in a pulse signal are actually
needed for the application.
We do however have many ways to get started on finding limits. For
example, we can test the signal needs of our sensors (ears, eyes), and know
that the effective bandwidth of our auditory perception system is in the
ballpark of the bandwidth of a compressed CD audio signal (1 MB/minute or
100 Kbps). We can justify these types of numbers because as we increase
the bandwidth of the signal beyond these points, our abilty to perceive a
difference quickly drops to zero. We can't hear a difference.
But that overall effective bandwidth might be created by using neurons
which require much higher bandiwdth limits to function correctly. So if we
were to try and build a digital neuron as a replacement for a real nuron in
the brain, maybe it would have to have a bandwidth of 100 gigabits per
second to function correctly with the other neurons. And if that were
true, then building a digital brain which worked by simulating the function
of biological neurons, might erquire huge amounts of computing resources
for each neuron.
However, it's very likely that no matter what the digital bandwidth of a
single neuron is (and it's probably fairly high), the effective bandwidth
needed for duplicating the function of the entire brain, is probably far
less. Trying to simulate neurons is a great way to understand neurons, and
a good way to start trying to understand the brain. But it's a very bad
way to duplicate the function of the brain.
It's as bad for example as trying to understand the function of a vacuum
tube raido by trying to simulate the exact function of a vacuum tube with a
trainsistors. How many transistors would it take to construct a drop-in
replacement device for a vacuum tube? 10 maybe? 50 Maybe? But if instead
of trying to duplicate the fuction of the vacuume tube, you instead
duplicated the function of the radio with tranistors, you might only need 5
transistors to replace 5 vacuum tubes, but all the other cirucuts would be
different to make up for the fairly substantial differences between tubes
(which are voltage amplifiers) and transistors (which are current
amplifiers).
We just don't know enough about how neurons work together to create
something like the audio preception system to answer any of these questions
with any real level of authority. All we can do is make educated guesses.
Evgenij Barsukov
> The number of bits per second an analog device can pass is directly related
> to the area of the frequency spectrum curve. If you want to encode all the
> useable information in an analog signal into the binary domain, you can do
> that by sampling the analog signal at the Shannon limit rate which is twice
> the rate of the bandwidth of the signal. So a 10 kHz bandwidth signal
> needs to be sampled 20,000 times per second.
You are mis-placing here the Shannon theorem and the Nuquyst sampling
theorem.
f = 1/2*dt is the Nuquyst sampling theorem, and it is indeed application
specific.
Shannon theorem defines how much information fundamentaly can be
transmitted over given hardware (regardless of frequency), because
limitation is coming from the noise level on one side, and maximal
frequency is limited by the low-pass cut-off of the hardware (frequency
above which effective resistance becomes infinite). Note that this limit
is independent from the way _how particular application_ is using the
hardware, because it defines _maximal possible_ transmission rate, NOT
an application specific transmission rate.
> If you look at an isolated device like a wire, or a neuron, the true total
> information content in the analog nature of the device, is bascially
> infinite if you try to talk about it in binary terms.
Nope, this is completely wrong. It is not infinite, but limited by the
Shannon theorem and very finite for every given hardware. For example
for a cupper pair wire over 15km long under ambient noise level it is
limited to 64bps, regardless of how you use it.
ADSL is not violating this principle, they are just using improved
shielding to reduce noise level and limited the distance to 1km, so
it is basicaly using a different hardware (note that limit is affected
by the lenght of the wires).
> So, the digital bandiwdth of any analog device is infinte, but the signal
> content needed for any specific applicaiton, is normally finite and small.
> But no one understand the brain well enough to know what the required
> digital bandwidth of that application is.
Correspondingly, you are making a completely wrong conclusion here.
If you would use a given hardware using different means of transmission
(for example took a wire but instead of electrical signals you would
use accustical signals over it) than indeed you could exceed the Shannon
limit for that wire (or maybe it would be lower, you would just have a
different limit).
However, if transmission _method_ is fixed (such as electromagnetic
waves, electrochemical, diffusion etc) than the limit is fixed
_regardless of how particular application is using the transmission line_
Regards,
Evgenij
Curt Welch
> Curt Welch wrote:
>
> > The number of bits per second an analog device can pass is directly
> > related to the area of the frequency spectrum curve. If you want to
> > encode all the useable information in an analog signal into the binary
> > domain, you can do that by sampling the analog signal at the Shannon
> > limit rate which is twice the rate of the bandwidth of the signal. So
> > a 10 kHz bandwidth signal needs to be sampled 20,000 times per second.
>
> You are mis-placing here the Shannon theorem and the Nuquyst sampling
> theorem.
> f = 1/2*dt is the Nuquyst sampling theorem, and it is indeed application
> specific.
Right you are. I got those confused.
> Shannon theorem defines how much information fundamentaly can be
> transmitted over given hardware (regardless of frequency), because
> limitation is coming from the noise level on one side, and maximal
> frequency is limited by the low-pass cut-off of the hardware (frequency
> above which effective resistance becomes infinite). Note that this limit
> is independent from the way _how particular application_ is using the
> hardware, because it defines _maximal possible_ transmission rate, NOT
> an application specific transmission rate.
Right, the shannon theory is what takes both the bandwidth, and the S/N
level of the system, and gives you an upper limit for the bit rate.
> > If you look at an isolated device like a wire, or a neuron, the true
> > total information content in the analog nature of the device, is
> > bascially infinite if you try to talk about it in binary terms.
>
> Nope, this is completely wrong. It is not infinite, but limited by the
> Shannon theorem and very finite for every given hardware. For example
> for a cupper pair wire over 15km long under ambient noise level it is
> limited to 64bps, regardless of how you use it.
Nope, you miss my point.
You are assuming that the noise on the line is not valid signal.
To use a wire to transport a signal, we must keep the signal away from the
noise on the line or else we can't hear the signal. But that's not because
the wire is unable to transmit weaker signals, it's beacuse it's already
transmitting weaker signals which we don't care about - which we want to
activly avoid in our application of the wire.
So the noise level is defined by the application specific need to transmist
signals other than the signals already being transmitted by the wire.
If you realize that noise is always signal (it's just application specific
unwated signal), then the noise level drops to 0, and the Shannon equation
produces an answer of infinity.
> ADSL is not violating this principle, they are just using improved
> shielding to reduce noise level and limited the distance to 1km, so
> it is basicaly using a different hardware (note that limit is affected
> by the lenght of the wires).
>
> > So, the digital bandiwdth of any analog device is infinte, but the
> > signal content needed for any specific applicaiton, is normally finite
> > and small. But no one understand the brain well enough to know what the
> > required digital bandwidth of that application is.
>
> Correspondingly, you are making a completely wrong conclusion here.
> If you would use a given hardware using different means of transmission
> (for example took a wire but instead of electrical signals you would
> use accustical signals over it) than indeed you could exceed the Shannon
> limit for that wire (or maybe it would be lower, you would just have a
> different limit).
> However, if transmission _method_ is fixed (such as electromagnetic
> waves, electrochemical, diffusion etc) than the limit is fixed
> _regardless of how particular application is using the transmission line_
Only if you believe noise is always noise, which it is not. Noise is
defined by the application of the wire, not by the wire.
When radio waves in the air generate motion in the electrons in a phone
wire, it's called noise. When radio waves generate the same motion in an
antenna, it's called signal.
The definition of noise is application specific and the very assumption
that noise even exists (aka a signal which we don't want) is an application
specific assumption.
It's a common assumption because all applications which transmit a signal
must keep the signal they want, above the noise floor (which is always
there), and there is no application I know of other than a white noise
generator, that actually makes use of the noise present in all systems.
But, for the brain, where does the signal stop and the noise begin in the
neurons? Can you tell me for sure that the operation of the brain does not
in fact make use of the noise in it's processing?
There are some examples of using GA modifcation of circuit designs that
show that the GA techniques produced circuit designs that actually made
use of the noise characteristics of the device being modified. So it's
quite possible that the noise characteristics of the neurons are being used
as part of the "signal" the brain is working with. In which case, as I
said, we just don't know what the real bandwidth of neurons are in this
application because no one yet understands enough about the entire brain,
to know for sure how much of the bandwidth of the neuron is being used, and
how much is not imporant, and can just be called "noise".
Evgenij Barsukov
> You are assuming that the noise on the line is not valid signal.
Noise is _defined_ as random, and is therefore fundamentaly not
capable to be a signal. Random means it have equal probability to
be any level (withing given bandwidth). Random also means = 0
information. While minimal information content can always vary depending
on application (e.g. you might waste part of available bandwidth),
maximal information contents is fixed as well quantifiable.
You can not use part of already existing random noise for your purpose,
but you can add to it something more which will look like noise but will
contain data (for example something encrypted). However, amount of data
you will add this way will be limited by Shannon theorem which will take
into account the _existing_ noise, not the stuff you are adding.
Btw Shannon theorem does not limit information transmitted at one
frequency, it is related to integral transforms and is therfore covering
the entire frequency spectrum available in your given bandwitch. So if
under transmitting signal which will _appear_ like white noise you mean
basicaly transmitting at infinite number of frequencies simultaneously,
it will in no way change to max. transmission limit defined by Shannon
theorem.
The only way to increase the throughput is to physicaly (sic!) eliminate
already present noise. This means changing particular hardware (under
hardware here we understand both "wire" and environment). This is well
possible but again to a given limit, not to the infinite limit.
>
> To use a wire to transport a signal, we must keep the signal away from the
> noise on the line or else we can't hear the signal. But that's not because
> the wire is unable to transmit weaker signals, it's beacuse it's already
> transmitting weaker signals which we don't care about - which we want to
> activly avoid in our application of the wire.
You can think about it as Noise occupying part of the "information
volume" that can be transmitted. Because it is already physicaly there,
you can not just "think it away" - you need to physicaly reduce it.
>
> So the noise level is defined by the application specific need to transmist
> signals other than the signals already being transmitted by the wire.
>
> If you realize that noise is always signal (it's just application specific
> unwated signal), then the noise level drops to 0, and the Shannon equation
> produces an answer of infinity.
No you misunderstand definition of noise. It is random, and therefore
fundamentaly can not transmit any information. Even if it would not be
random, but still physicaly present so that you can not remove it, it
would still take up part of your "information throughput", regardless of
how you are plaing to use the remaining volume.
You can _add_ to that something else, but only within remaning volume.
>> ADSL is not violating this principle, they are just using improved
>>shielding to reduce noise level and limited the distance to 1km, so
>>it is basicaly using a different hardware (note that limit is affected
>>by the lenght of the wires).
>>
>>
>>>So, the digital bandiwdth of any analog device is infinte, but the
>>>signal content needed for any specific applicaiton, is normally finite
>>>and small. But no one understand the brain well enough to know what the
>>>required digital bandwidth of that application is.
>>
>>Correspondingly, you are making a completely wrong conclusion here.
>>If you would use a given hardware using different means of transmission
>>(for example took a wire but instead of electrical signals you would
>>use accustical signals over it) than indeed you could exceed the Shannon
>>limit for that wire (or maybe it would be lower, you would just have a
>>different limit).
>>However, if transmission _method_ is fixed (such as electromagnetic
>>waves, electrochemical, diffusion etc) than the limit is fixed
>>_regardless of how particular application is using the transmission line_
>
>
> Only if you believe noise is always noise, which it is not. Noise is
> defined by the application of the wire, not by the wire.
There is something to you add, and something that is already present
regardless of your wishes (no matter what nature of it there is).
The stuff you _add_ will be limited by the theorem even if you add
more noise-like stuff.
By quantifying nature of the noise you might slighly increase your
estimate of the Shannon's limit (for example you find that noise is not
white and has some frequency gaps, you could use these gaps to transmit
some more information), but this would only happen for some specific
kinds of noises and do not change the overal discussion.
>
> When radio waves in the air generate motion in the electrons in a phone
> wire, it's called noise. When radio waves generate the same motion in an
> antenna, it's called signal.
That is correct. For particular user, it is just the differentiation
between the signal that _he_ can send, and "other stuff" that is already
there. Understanding of what this "other stuff" is will in no way
increase the amount of your own signal that you can transmit.
You can think about it as highway with a bunch of cars. Yes, all
other moving things are cars just like yours, but this understanding
does not help you to move along any faster.
>
> The definition of noise is application specific and the very assumption
> that noise even exists (aka a signal which we don't want) is an application
> specific assumption.
>
> It's a common assumption because all applications which transmit a signal
> must keep the signal they want, above the noise floor (which is always
> there), and there is no application I know of other than a white noise
> generator, that actually makes use of the noise present in all systems.
Noise spectrometers can measure impedance of the system, unfortunately
only modulus, but not Re and Im parts separately. This does not
violate Shannon theorem - in this case Noise IS the signal, but there
is still some "other" noise (stuff not generated by the Noise generator
but physicaly present) which will limit the amount of information you
can get from the system using your own Noise.
> But, for the brain, where does the signal stop and the noise begin in the
> neurons? Can you tell me for sure that the operation of the brain does not
> in fact make use of the noise in it's processing?
This just points out that we don't know any of the signals. But this is
not quite true. We know for example times where no signals are
transmitted (both neurons connected to one synapse are in "cool off"
phase, and there is no chemical activity going on). In this case noise
will be just random chemical concentration variations (like Brownian
motion). This can be quantified.
Exidentaly, because of electrochemical transmition, noise levels in
bio-systems are indeed very low (because signal is transmitted not by
just any molecules, but by specific molecules only, which are always
secreted only during trasmission). So, bandwith of synapses is limited
by the low-pass cut off much more than it is by the noise. There is just
a limited rate with which electrochemical wave can propagate which
dominates Shannon's equation in this case. Noise could have been 1000
times higher without chaging bandwith substantialy.
Even more the processing is limited by cool-off time between firings,
but interestingly propagation times and cool-off times are of the same
order of magnitude. This is not an excident. Obviously for bio-system it
would not make sense to propagate signals much more fast than frequency
with which they are produced, that is why nature "optimized" both limits
close to each other.
> There are some examples of using GA modifcation of circuit designs that
> show that the GA techniques produced circuit designs that actually made
> use of the noise characteristics of the device being modified. So it's
> quite possible that the noise characteristics of the neurons are being used
> as part of the "signal" the brain is working with. In which case, as I
> said, we just don't know what the real bandwidth of neurons are in this
> application because no one yet understands enough about the entire brain,
> to know for sure how much of the bandwidth of the neuron is being used, and
> how much is not imporant, and can just be called "noise".
Entire brain is not well understood, but single neurons are. There are
neurons which are several meters long (such as that of squids) and they
have been investigated in great detail. If something is missed, it will
not be in the realm of high-rate transmissions over synapses which are
relevant to this thread, but to "very low rate" transmission of signal
through diffusion of some signal molecules into intra-neuronal space.
Some recent research shows that these transmissions while being very
slow have certain effect on long time memory. This would not affect the
system bandwith substatialy.
Curt Welch
> Curt Welch wrote:
> > If you realize that noise is always signal (it's just application
> > specific unwated signal), then the noise level drops to 0, and the
> > Shannon equation produces an answer of infinity.
>
> No you misunderstand definition of noise. It is random, and therefore
> fundamentaly can not transmit any information.
I do understand. You are repeating common social beliefs about the idea of
noise and random which just aren't true.
Noise, like many things, is not in the wire. It's in the mind of the
observer. It's only noise because you choose to call it noise.
In the wire, all you find is tons of information.
As you probably know, the maximum amount of information is trasmitted when
the symbols are all equally probable. If they are not, then you have a
reduction in the effective information data rate. The maximum amount of
information is flowing when the data looks like white noise. That's why a
300 baud modem sounds like tones, and a 56K baud modem sounds like static.
So your comment that there is no information in the noise couldn't be
farther from the truth. There is maxium amounts of information in the
noise. The point is simply that we don't want any of it. That's why we
call it noise.
The noise floor in a wire is just very high information content background
signal which most the time we wish to avoid or eliminate.
However, all that noise in the wire, is real signal data about real events
happening in the universe somewhere, at some past point in time. Below
some low level, the "information" is mostly about the atoms in the wire
itself and not something interesting elsewhere in the universe. In order
to reduce internal noise in sensors, they are sometimes cooled to near
absolute zero. I think astronomers have had to play these games in order
to pick up very low level background noise from the sky and not get it
mixed in with the signals about the vibration of the atoms in the wires.
So, instead of using a trick like shelding to block signals from getting
into a wire, they have to actually turn down the signal generators working
inside the device.
> Even if it would not be
> random, but still physicaly present so that you can not remove it, it
> would still take up part of your "information throughput", regardless of
> how you are plaing to use the remaining volume.
Not if the noise was the signal you were trying to pick up. See my point?
The wire itself is in fact carrying this signal you keep refering to as
noise. So the limit of how much data the wire can hold, is the total data
transmitted by what you call "noise" + any signal you choose to add on top
of that to the wire.
No one that I'm aware of wants that noise signal. So in all practical
engineering applications, what you say is right on. But the actual
information in the wire, is not reduced by the nose, it's part of the
information the wire is carrying.
I do admit that my use of "infinite" bandwidth is a stretch. That idea
makes some assumptions about the nature of the universe which are
unjustified. But the prime point I'm making is that a device like a wire
is already carrying a ton of data in what we call the "noise" before we try
to add more on top of it.
> You can _add_ to that something else, but only within remaning volume.
The wire itself, can carry a large amount of information but it's normally
filled up with noise. The wire is actually carrying that information, so
the bandwidth of the raw wire, is far higher than the Shannon limit would
lead us to believe. The numbers you talk about is just how much room is
left in the back of the bus after the universe fills most of the front
seats with its own data.
My point was to talk about the size of the bus and not about how many seats
were left for us by the universe for our own use.
When we look at neurons, and try to grasp what they are doing, there's a
ton of data there, just like there's a ton of data in the wire. To know
how much of that we should call noise, and how much we call signal, is
impossible until we know more about how all the effects are used in the
total brain. And without knowing which of the data is noise, and which is
signal we have no way of converting the signal compoenent to a digital
bandwidth number. The people that have given digital bandwidth numbers
(calling it information content) to neural signals have done so by pulling
some best guess assumptions out their ass. In the end, the number is just
an educated guess.
Btw, I've concluded that when talking about pulse signals, the best "signal
volume" number to use is simply pulses per second.
HMS Beagle
by far. Although I do have follow-up comments.
On Wed, 05 Oct 2005 10:03:37 -0500, Evgenij Barsukov
<evgenij_...@yahoo.com> wrote:
>The whole expression "bandwidth" was initialy used for analog devices,
>even before digital devices existed. It defines how much information
>can be tranported through analog copper pair given certain level of
>noise, and low-pass cut-off of the line. Quantitative answer for first
>part (noise) is given by Shannon theorem, and for second case (low-pass
>cut-off at given electric characterists) given by the "telegraph equation".
This raises some interesting points. It seems that EVERYTHING is
analog at base, even the insides of a Pentium CPU. In fact, during
chip manufacturing, they burn-in test the CPU to see if it is
"stable". "Instability" of a CPU, in analog terms, may mean some part
of the circuit formed incorrectly, and that it's either picking up its
own noise, or leaking off current at the microscopic level. The
warmer the CPU becomes the more it becomes subject to ANALOG errors at
the circuit level, (hence "burn-in")
> So bandwith of each particular type neuron is well defined and is
>routinely measured by bio-physists. There is a whole
>brach of bio-physics called neuron impedance spectroscopy who is
>basicaly parametrizing dymanic performance of neurons.
Interesting.
> Low-pass cut-off of different neurons range from a kHz to 100 Hz, so
>it is not nearly as fast as electrical networks. There is a good reason
I'm not sure I understand what the low-pass cutoff means in terms of
analog speed. I understand high-pass cutoff, but its not clear to me
what low-pass cutoff means.
> This depends the frequency with which impulses can be fired. These
>frequency for brain neurons is about 10 Hz. No matter how fast the
>signal after that propagates through the synapses (which is also not
>very fast as you can see above) the ultimate rate of "computation" is
>limited by the cool off time.
>Alas, chemical processing has its inherent rate limits. It is very
>stable, very robust and noise-protected, but it is slow.
One would think we could simulate several thousand of these (at
realtime) very easily with a single modern CPU. Ironically, some
people in here have already said it can't be done. They say "there
are too many biochemical reactions"
>Analog devices have inherent speed limitations as shown above.
>That is why sometime in the 60ies, where analog "co-processors" for
>solving particular often occuring tasks were created, they were
>limited to that time RF technology. They still depend on the frequency
>that can propagate through the wires. More succesful recent analog
>computers are based on optics rathere than RF. Light waveguides has
>dramaticaly higher bandwidth than anything else, and there are already
>light based analog computer that can do FFT faster than aniting else on
>earth. So, analog computing definetely has some future (while optical
>digital systems will also soon follow).
What is your opinion on this? http://en.wikipedia.org/wiki/FPGA
Or this?
http://en.wikipedia.org/wiki/Application-specific_integrated_circuit
Evgenij Barsukov
> When we look at neurons, and try to grasp what they are doing, there's a
> ton of data there, just like there's a ton of data in the wire. To know
> how much of that we should call noise, and how much we call signal, is
> impossible until we know more about how all the effects are used in the
> total brain. And without knowing which of the data is noise, and
which is
> signal we have no way of converting the signal compoenent to a digital
> bandwidth number. The people that have given digital bandwidth numbers
> (calling it information content) to neural signals have done so by
pulling
> some best guess assumptions out their ass. In the end, the number is
just
> an educated guess.
You must have missed the specific description of the way how the
bandwidth of neurons are measured in my message. There is not much
guessing involved. The "noise" term there is negligible anyway
because of electrochemical way of transmission. The only noise
contributed is fluctuation of concentration, and it does not contribute
anything noticeable to Shannon equation in this case.
The actual bandwith limitation is coming because electrochemical waive
is propagating very slowly as such, so transmition times are in the
order of 10-100 m/sec. So all above discussion of all this "mistery of
noise" is irrelevant.
Synapses have low bandwith, it is a fact, unless we
misunderstand the _way_ how transmission takes place. If for example
transmission would not take place electrochemicaly as now widely
experimentaly proven and analysed in finest details, but somehow else
(say synaps would be realy a wave-guide and transmission would by
done using infrared light) than indeed estimated bandwithd would
increase by many orders of magnitude. But it is highly unlikely,
because action/transmission/muscle reactions involving neurons
have been realy beaten to death experimentaly during last 100 years.
Regards,
Evgenij
Evgenij Barsukov
>> Low-pass cut-off of different neurons range from a kHz to 100 Hz, so
>>it is not nearly as fast as electrical networks. There is a good reason
>
>
> I'm not sure I understand what the low-pass cutoff means in terms of
> analog speed. I understand high-pass cutoff, but its not clear to me
> what low-pass cutoff means.
Low-pass cut off frequency means that any higher frequency will
experience resistance much higher than all lower frequencies, and
resistance will exponentially increase until it will reach close
of infinity.
Depending on the network there can be a sharp cut-off (commercial
filters usualy have about 70dB rate of increase of effective resistance
for all frequencies above given) but in usual networks increase is
gradual which make it diffucult to give an exact number for cut-off
frequency. But integrating overall available frequency range can still
be used to give Shannon's limit.
>> This depends the frequency with which impulses can be fired. These
>>frequency for brain neurons is about 10 Hz. No matter how fast the
>>signal after that propagates through the synapses (which is also not
>>very fast as you can see above) the ultimate rate of "computation" is
>>limited by the cool off time.
>>Alas, chemical processing has its inherent rate limits. It is very
>>stable, very robust and noise-protected, but it is slow.
>
>
> One would think we could simulate several thousand of these (at
> realtime) very easily with a single modern CPU. Ironically, some
> people in here have already said it can't be done. They say "there
> are too many biochemical reactions"
There is two level of simulation
- instantaneous, which use present unchangeable configuration of
neurons, and is defined by the bandwith of the synapses and the cool-off
time. This can be indeed simulated easily on modern CPUs for thousands
of neurons in about the same time as it would take neurons themself.
The problem is, what would be the point of such simulation? To make
it meanigful, you first need to measure initial states of a thousand of
actual neurons, map out all actual interconnects (e.g. synapses) and
than fire your simulation to see if you can maintain the state change
which happens in real neurons. But the tricky part is to monitor
real "life" neurons. They are very small and you can not put a probe
on each one, and each synapse, although thre is some progress now in
doing that. Probes how been done that sample "electric" state of
hundreds neurons.
Even more complex, how do you define a "state" of neuron?
Electrical state only tells you if neuron has fired or is in cool-off.
But it will not tell you when it will fire again! Because its internal
workings are not quite inderstood, we don't know why and at what time
it will "fire" particular synaps. So we are somewhat in limbo here - we
know _how fast_ the signal will propagate once it is fired so we can put
the higher limits on the overal processing rate.
But we _do not know_ why the signal is sent and when it will be sent and
to which synaps. We know that for the cases of "sensor neurons" - signal
is sent in response to irritation. But for "thinking neurons" it is much
less explored.
So, internal processing happening "inside" neuron is still a bit of a
black box that puts certain uncertainty in overal processing rate
estimates. So far it appears (again, mostly based on sensor neurons
experimentation) that processing is quite primitive and would not
add much to overal processing rate, but it is an open
question.
- long-term, this includes changing of configuration, growth of new
synapses. This process take long time and involves propagation of
information by diffusion of certain molecules in the space _between_
neurons. While process is long and so would be easy to simulate in real
time, it is much less understood than instantaneous behaviour and it
indeed involves many chemical reactions and many different molecules.
>>Analog devices have inherent speed limitations as shown above.
>>That is why sometime in the 60ies, where analog "co-processors" for
>>solving particular often occuring tasks were created, they were
>>limited to that time RF technology. They still depend on the frequency
>>that can propagate through the wires. More succesful recent analog
>>computers are based on optics rathere than RF. Light waveguides has
>>dramaticaly higher bandwidth than anything else, and there are already
>>light based analog computer that can do FFT faster than aniting else on
>>earth. So, analog computing definetely has some future (while optical
>>digital systems will also soon follow).
>
>
>
> What is your opinion on this? http://en.wikipedia.org/wiki/FPGA
>
> Or this?
> http://en.wikipedia.org/wiki/Application-specific_integrated_circuit
I think there is a lot of computing applications where reconfigurable
computing makes sense. These can be classified as:
- "do a lot of computatins of type 1",
- than switch to different type of computations
(switching takes long time, but shorter than computation run),
- than run again a lot of type 2.
- (and so on, there is no limits on how many different types you can
support).
Switching is unfortunately quite slow,
so to make it efficient, minimal amount of switching should take
place in this particular task.
There is a company providing software for programing this type
of problems for more then 5 years now:
http://www.starbridgesystems.com/hypercomputers.htm
Apparently for specific type of problems they can beat largest
supercomputers.
But going forward I expect that different physics (e.g. "in flight
modulation of light" which already has promising hardware base
and working in-silicon prototypes) and in the future quantum computing
using super-conducting current loops will take over before FPGAs
programing skills will make home PCs FPGA-based.
It can still be re-configurable but using different means that
are not yet clear at the moment.
Regards,
Evgenij
>
Steve Richfie1d
> You must have missed the specific description of the way how the
> bandwidth of neurons are measured in my message. There is not much
> guessing involved. The "noise" term there is negligible anyway
> because of electrochemical way of transmission.
This is very parallel to "current mode logic" (ECL) used in fast
computers and on bipolar chips.
> The only noise
> contributed is fluctuation of concentration, and it does not contribute
> anything noticeable to Shannon equation in this case.
>
> The actual bandwith limitation is coming because electrochemical waive
> is propagating very slowly as such, so transmition times are in the
> order of 10-100 m/sec. So all above discussion of all this "mistery of
> noise" is irrelevant.
Actually, if you compute the cable velocity (the speed of light along a
given cable), when the sizes get down to neuron sizes, the pulses are
actually traveling at the cable velocity! The non-linear chemical
process apparently serves to amplify and to "square" the pulses as they
travel along.
> Synapses have low bandwith, it is a fact, unless we
> misunderstand the _way_ how transmission takes place. If for example
> transmission would not take place electrochemicaly as now widely
> experimentaly proven and analysed in finest details, but somehow else
> (say synaps would be realy a wave-guide and transmission would by
> done using infrared light) than indeed estimated bandwithd would
> increase by many orders of magnitude.
But it would STILL be restricted by the axons, thereby restricting the
EFFECTIVE bandwidth to the numbers we've been talking about.
> But it is highly unlikely,
> because action/transmission/muscle reactions involving neurons
> have been realy beaten to death experimentaly during last 100 years.
As Jon Von Neuman once said at a neuroscience conference, the
distinction between electrical, chemical, and mechanical disappears when
the scale becomes small enough.
However, only a TINY fraction of all computing cells ever spike! 90% are
glial cells, and most of the rest never spike. Spiking cells have been
extensively studied only because they are most easily observed.
Clearly, spiking reduces the bandwidth UNLESS the signal is going a long
way away (which is probably why their axons are usually so long).
Steve
HMS Beagle
<evgenij_...@yahoo.com> wrote:
>> What is your opinion on this? http://en.wikipedia.org/wiki/FPGA
>>
>> Or this?
>> http://en.wikipedia.org/wiki/Application-specific_integrated_circuit
>
>I think there is a lot of computing applications where reconfigurable
>computing makes sense. These can be classified as:
>- "do a lot of computatins of type 1",
>- than switch to different type of computations
>(switching takes long time, but shorter than computation run),
>- than run again a lot of type 2.
>- (and so on, there is no limits on how many different types you can
>support).
>
>Switching is unfortunately quite slow,
>so to make it efficient, minimal amount of switching should take
>place in this particular task.
>There is a company providing software for programing this type
>of problems for more then 5 years now:
>http://www.starbridgesystems.com/hypercomputers.htm
>Apparently for specific type of problems they can beat largest
>supercomputers.
Interesting link there. Let my quote a paragraph from that page.
[[The most widely used computational engine for large problems is
cluster computing-an approach to supercomputing that aggregates
hundreds or even thousands of individual CPUs. However, cluster
technology is reaching a point of diminishing returns. Clusters are
expensive to build and difficult to maintain; they require ever
increasing power, cooling systems, and physical space; and their
architecture places inherent limits on scalability and performance.]]
Does this sound familiar? I think it does. It sounds suspiciously
like the conversion from vacuum tubes to silicon transistors.
"Requires less power" - "less physical space" - "is faster" - and so
on.
Are we seeing the beginning of the end here? Is the silicon CPU's
reign coming to an end? Will FPGAs come to dominate in the coming
decades? Only time will tell. This page did pay lip-service to
computers that use both the serial speed of regular CPUs with the
parallelism of an FPGA.
>But going forward I expect that different physics (e.g. "in flight
>modulation of light" which already has promising hardware base
>and working in-silicon prototypes) and in the future quantum computing
>using super-conducting current loops will take over before FPGAs
>programing skills will make home PCs FPGA-based.
>It can still be re-configurable but using different means that
>are not yet clear at the moment.
Exactly. So you think the quantum computer is possible in this
universe eh?
I have my qualms. Decoherence problems in quantum computers smell
suspiciously like the problems of obtaining absolute zero kelvin. So
while I'm sure physics tells us the quantum computer is actually
testing all possible situations (at the quantum level) there is no
way to obtain those results in the macroscopic world. Of course I
could be wrong... I suspect I'm not.
HMS Beagle
(Lester Zick) wrote:
>On Wed, 28 Sep 2005 20:54:34 -0400, HMS Beagle <bga...@microsoft.org>
>in comp.ai.philosophy wrote:
>>In summary, AI is not a matter of speed. It's a matter of
>>engineering, and a matter of philosophy.
>Engineering + philosophy? This strikes me as empiricism and/or
>pragmatism. However I'm not necessarily arguing against either in this
>context. But if true, how would you ever recognize success? In other
>words how would you ever be able to recognize true ai? Because it
>satisfies some philosophy?It strikes me the combination engineering +
>philosophy is just the same task oriented turing philosophy for ai. Ai
>is said to be acheived when some task or tasks performed by people is
>performed as well or better by machines. And my impression is that
>this is what you're basically arguing against.
I am not arguing against that. In fact I did not mention that. What I
meant to say is that our cultural philosophies about how a mind
operates can be realistically wrong, and therefore, machines that are
built with these cultural pressuppositions in their architecture will
fail at a given task. Exemplar: GOFAI failed because the brains of
simple mammals are more than symbol-manipulators instantiated
biologically.
What I meant to say is that the barriers keeping us from building a
conscious artifact are PHILOSOPH-ICAL and ENGINEERING(-ical) in
nature. Once our philosophical notions of the mind are realistic
(jibing with the world) then the only problem remaining is "how to
build it physically" (engineering).
>Glen has a similar notion. Only his philosophy is behaviorism and
>his engineering is training. And his primary task or application is
>clear: disprove the mind and mental effects. So I'm just curious what
>your objective is and whether it's a serious objective?If not chess et
>al. what task or task list does your philosophy recommend engineering
>implement and how will you recognize and know ai when you see it?
When will I know General AI when I see it? That is such a tricky
question and there are so many ways to answer it. In this post I will
give you 3... and only tentatively. All of these are mutually
exclusive. Meaning I don't intend for all 3 of them to happen. One is
enough for General AI.
------------
I. Economic
-----------
General AI is acheived when autonomous agents are industrialized and
are officially a sector of the industrial economy. This happens when
the harvesting of robots is on par with where steel industry was in
the late 19th century or where Oil refining is now. Right now
autonomous agents are mere novelty toys. Think about before cars
became a major sector of the economy (no car can beat a good horse!).
When robots are doing everyone's laundry and performing many labour
jobs, then General AI will have been achieved. We may only
"recognize" this in retrospect via history books.
---------
II. Military
---------
General AI is acheived when a group of small robot "spiders" defeat an
armed special forces team. (and I don't mean "in a training exercise")
----------
III. Laboratory
----------
Say you want me to recognize the first General AI agent as soon as it
is built in a lab. Part III answers this question.
I'm convinced that the DARPA GC is a sub-category of AI called
sensory/motor coordination (SMC). A "properly trained" vehicle can
pass through desert terrain using this only. The vehicle does not
need on-the-fly problem-solving capabilities to do this. The vehicle
does not need to reflect on what it has seen in the first half of the
race to derive plans about what it might see in the second half.
(Although this might help its finishing time, it is clearly not a
*requirement.*)
General AI is a different matter altogethor. SMC is only a slice of
the whole pie. There is no telling where the philosophy might take
us. In the future it may become apparent that Artificial
Intelligence is really a sub-category of Artificial Life. And I
begin to suspect this is true the more I study it. Let me try to
formulate a tentative task list of General AI.
1. It is embodied.
2. It is self-sufficient.
3. It is autonomous.
4. It is self-organizing.
These are meant to be criteria for a single intelligent agent. Of
course there is the criteria that it is actually successful at its
given task (left out). An agent that exhibits all 4 would be a
solution to General AI. But notice it also seems to be a description
of a beehive. I beleive this may be more than a passive coincedence.
I will expand on each.
1. It will be robotic and it will interact with the world. No
floating res cogitans. No turing tests, no brain-in-vat.
2. Its own sensors and actuators are all it needs to perform its
task. (eg. No GPS data piped in from the outside.)
3. A human is not controlling it, (or giving it cues to stop tapping
its hoof - for those of you who know about *that* one).
4. Its behavior changes through time in a self-organizing manner.
There is still a sense that we have not entirely shaken off the hints
of beehives and ant colonies. One way to jump out of that loop is
to add a 5th criteria
5. It is conscious.
And this leads in a totally new direction ----> hence philosophy.
Lester Zick
I'm addressing this in two parts.
On Thu, 06 Oct 2005 22:47:18 -0400, HMS Beagle <bga...@microsoft.org>
in comp.ai.philosophy wrote:
>On Thu, 29 Sep 2005 14:56:14 GMT, lester...@worldnet.att.net
>(Lester Zick) wrote:
>
>>On Wed, 28 Sep 2005 20:54:34 -0400, HMS Beagle <bga...@microsoft.org>
>>in comp.ai.philosophy wrote:
>>>In summary, AI is not a matter of speed. It's a matter of
>>>engineering, and a matter of philosophy.
>
>
>>Engineering + philosophy? This strikes me as empiricism and/or
>>pragmatism. However I'm not necessarily arguing against either in this
>>context. But if true, how would you ever recognize success? In other
>>words how would you ever be able to recognize true ai? Because it
>>satisfies some philosophy?It strikes me the combination engineering +
>>philosophy is just the same task oriented turing philosophy for ai. Ai
>>is said to be acheived when some task or tasks performed by people is
>>performed as well or better by machines. And my impression is that
>>this is what you're basically arguing against.
>
>I am not arguing against that. In fact I did not mention that.
But you did argue against chess algorithms as definitive of ai which
is where I drew the inference.
> What I
>meant to say is that our cultural philosophies about how a mind
>operates can be realistically wrong, and therefore, machines that are
>built with these cultural pressuppositions in their architecture will
>fail at a given task.
Sure.
> Exemplar: GOFAI failed because the brains of
>simple mammals are more than symbol-manipulators instantiated
>biologically.
I don't argue about GOFAI or other approaches. For one thing I don't
know enough about them to judge. However your specific objection here
strikes me as unwarranted unless you actually know what constitutes
intelligence in mammalian brains. There might be a variety of reasons
an approach such as GOFAI failed. But I suspect the basic reason it
and other approaches fail is that they don't comprehend the nature of
intelligence to begin with.
>What I meant to say is that the barriers keeping us from building a
>conscious artifact are PHILOSOPH-ICAL and ENGINEERING(-ical) in
>nature. Once our philosophical notions of the mind are realistic
>(jibing with the world) then the only problem remaining is "how to
>build it physically" (engineering).
Sure. But let me just point out that every one of the philosophical
approaches to the subject ever tried has failed but a great many of
their adherents proclaim success or at least conviction of success.
Cartesian rationalism, materialism, behaviorism, cognitivism, etc.
all proclaim truth when it comes to artificial intelligence for
various philosophical reasons which none can justify except with
intuitive appeals.
This is why I'm skeptical of philosophy as any core component in any
approach to ai. We already have plenty of philosophy. What we still
don't have is any scientific comprehension of intelligence as a
mechanism whether symbol processing, bit processing, timing, or
anything else. (and when I say scientific I mean in analytical terms).
Certainly there is the engineering problem to solve but that's easy.
What we don't have, and I'm not sure you recognize, is that there is
no analytical comprehension of what it is we're trying to engineer
exactly. In programming terms I look on the problem as trying to code
a system without specs or really any idea what the specs might be.
(I'm going to break this reply off here to consider what you say
below.)
~v~~
HMS Beagle
Ok so you want an analytical description of intelligence. I think
this is a reasonable request. I agree with you that once science
has an analytical model of intelligence, we can simply ignore the
rantings of ideologues and philosophers.
I am currently trying to dig up some researcher named Walter J.
Freeman. Freeman views the mind/brain as a nonlinear dynamical
system. He says some radical things. Namely:
1 . That neuronal groups of animals are actually creating chaotic
attractors.
2. These give rise to *novel* firing patterns that help the system
evolve.
3. There is no store-and-retrieve cycle of memory formation and
recall. Rather what we have been calling "memory" all through history
is really just a part of the way a brain evolves and integrates new
chaotic attractors into its repetoire.
4. An "evolved" brain is then, a brain which has a good repetoire of
attractors.
I'm sure none of this makes sense to you now. It doesn't make sense
to me either. That's why I'm going to get more of his writings.
What I know for sure is that nonlinear dynamical systems are an
analytical model. This is what we want!
see http://sulcus.berkeley.edu/
Also see
1991 Scientific American (Issue?) 264
The Physiology of Perception
Freeman, W. J.
Curt Welch
I saw him speak this summer at a Smithsonian event about the mind.
He's an older guy that walks with a cane. He's been doing this stuff a
long time it seems.
> I am currently trying to dig up some researcher named Walter J.
> Freeman. Freeman views the mind/brain as a nonlinear dynamical
> system. He says some radical things. Namely:
>
> 1 . That neuronal groups of animals are actually creating chaotic
> attractors.
> 2. These give rise to *novel* firing patterns that help the system
> evolve.
> 3. There is no store-and-retrieve cycle of memory formation and
> recall. Rather what we have been calling "memory" all through history
> is really just a part of the way a brain evolves and integrates new
> chaotic attractors into its repetoire.
> 4. An "evolved" brain is then, a brain which has a good repetoire of
> attractors.
I've not read any of his work, but from his talk, he seems to do a lot of
research sticking sensors on the surface of the cortex and doing
multichannel recordings of the activity. For example he has used an 8x8
grid of sensors that will monitor the activity in one small area.
This then gives him an 64 dimension space which he then studies the
trajactory of the brain activity through that space. His strange attractor
ideas I believe come from him seeing how the trajtory of the activity will
tend to seek out different points in the 64 dimension space. So he sees
the brain as a device which develops a large number of these stable states
and as the activity changes, it is always being attracted to the stable
configurations.
This seems to me to be consitent with the perception effects we know about
that allow us to see a picture in one of two ways, but that our brain
always wants to lock onto one solution and when we see the face, we can't
see the glass, or when we switch and look at the glass, we don't see the
face. The face and the glass could be two different stable activity states
for some part of the bran which act as atractors.
He had some graphs that I didn't really udnerstand but I think they were
contour maps of some 3D space he had recreated with some of the data.
At least, this is what I was able to grasp from his short talk.
> I'm sure none of this makes sense to you now. It doesn't make sense
> to me either. That's why I'm going to get more of his writings.
> What I know for sure is that nonlinear dynamical systems are an
> analytical model. This is what we want!
>
> see http://sulcus.berkeley.edu/
>
> Also see
> 1991 Scientific American (Issue?) 264
> The Physiology of Perception
> Freeman, W. J.
--
Lester Zick
in comp.ai.philosophy wrote:
>On Thu, 29 Sep 2005 14:56:14 GMT, lester...@worldnet.att.net
>(Lester Zick) wrote:
[. . .] I have omitted the part of this post to which I replied
earlier.
>When will I know General AI when I see it? That is such a tricky
>question and there are so many ways to answer it. In this post I will
>give you 3... and only tentatively. All of these are mutually
>exclusive. Meaning I don't intend for all 3 of them to happen. One is
>enough for General AI.
Well I think the point no one seems anxious to recognize is that no
one can say what ai is not unless one can say what ai is. In a
collateral thread, I believe Traveler commented that anyone who
produced ai at the level of a mouse would be richer than Bill Gates.
And I asked how one would know this hasn't already been achieved?
There are lots of complex devices and robots out there. It seems to me
that the only arguments in support of Traveler's claim would have to
be 1) There are no artificial mice running around, behaviorism, or 2)
There is no one richer than Bill Gates, which begs the question.
I agree the question is maximally tricky. But it seems to represent
the crux of 98+% of everything written on this group.
Okay. This is a considered effort. However it looks to me like
everything in the different categories except 5 is more akin to
robotics than actual artificial intelligence which I call sentience.
Robotics is a fascinating and probably more productive direction
for ai than sentience but it is still robotics. The only difficulty I
can see here with the general appraisal is that it appears to rely on
a task list approach, which is sufficient for robotics but doesn't
have any direct bearing on general intelligence. So based on such a
task list approach it's difficult to see any reason to reject chess as
indicative of intelligence (although I do and reject turing criteria
as well).
Category 5 is more promising although I would make the criterion there
"sentience" instead of "consciousness" to include animal forms as well
as people. However I don't see any easy resolution to a mechanical
reduction for whatever we put in category 5 that would lead to any
sort of mechanization any time in the near future. Certainly not by
philosophers and philosophy which hasn't made any progress in the
field in the last several centuries despite considerable effort.
I have an approach to the problem I call "subjective mechanics" which
addresses the problem of a mechanical reduction for sentience directly
but no one endorses the approach although there was a curious mention
in "I, Robot" by Will Smith of robots as "difference machines" if
memory serves (please excuse the run on sentence). The curiosity is
that this is a very superficial description of exactly what subjective
mechanics is even though I can't really see how it got into the movie.
~v~~
Glen M. Sizemore
"Curt Welch" <cu...@kcwc.com> wrote in message
news:20051007155429.629$q...@newsreader.com...
This resonates with me. ;) I have thought for a long time that one of the
things that reinforcement does is cause the neural activity produced by
stimuli that play a discriminative role to produce some stable state. It is
only when stimuli produce such stable states that the proper "motor
programs" are initiated. Otherwise the stimuli produce weak, transient
effects and this begins to habituate. This might explain phenomena like
"latent inhibition" where it is hard to establish stimuli as Pavlovian
conditioned stimuli when they have been presented repeatedly without the US.
>
> This seems to me to be consitent with the perception effects we know about
> that allow us to see a picture in one of two ways, but that our brain
> always wants to lock onto one solution and when we see the face, we can't
> see the glass, or when we switch and look at the glass, we don't see the
> face. The face and the glass could be two different stable activity
> states
> for some part of the bran which act as atractors.
This is one of the reasons that I like the view of discriminative stimuli as
producing stable states in sensory portions of the cortex. The basin of
attraction is a ready-made explanation of generalization. The alternative is
that discriminative stimuli are simply conglomerates of the output of
"edge-detectors," "30 degree tilted bars in motion detectors," and so forth.
If enough of the features are there, the response comes out.
humi...@clix.pt
>humig...@clix.p wrote:
>
>>Its very easy to verify this if you tell it something and then later
>>ask a question whose answer requires knowledge of what you said
>>earlier. For instance if you tell it "My favourite colour is green"
>>and then later ask it "Do you remember what my favourite colour is?"
>>chances are that it will reply with something to disguise the fact
>>that it has no idea of what you're talking about.
>If it fails to answer pointless questions like those [...]
I don't know why you consider the question pointless. It looks
perfectly legitimate to me. Besides, why shouldn't the bot answer
pointless questions? And anyhow who decides if a question is
pointless or not?
>If it fails to answer pointless questions like those it is because not
>that many people ask them. Hence it doesn't ask them. Hence it
>doesn't learn the answer. Eventually it will answer your question
>perfectly.
Let me see if I get this straight. You say that if enough users ask
the bot the question "what is my favourite colour?", at some point
the bot will consider that question important and start asking the
users "what is my favourite colour?". Then the users will answer
something like "Gee, I have no idea". Do you say that this will,
somehow help the bot remember what my favourite colour is? There's
surely something amiss here. Can you assist in disentangling this
imbroglio?
>I fully agree, incidentally, that the Loebner Prize logs were not very
>good. Judges act outside the realms of its normal environment - online
>chat.
Yes, you're right, it's all the judges' fault; I knew it all along.
Sometimes I don't know what's in their minds; they behave as if they
were talking to an intelligent human being. What do they expect
then, the Spanish Inquisition?
>The program does learn. The program is language independent. Try it.
That is certainly a big step in the right direction and I
congratulate you for that because that means you are dealing with
language at the right abstraction level. I wish I could convince
Loebner to set that as the minimum standard for all the programs
entering the contest. Now I can even point to you as an example that
by doing so his beloved contest would not die for lack of
participation. On the contrary, as the credibility of the contest
improves, more people would be willing to participate.
As soon as we manage to convince people of that, we could then raise
the bar to the next level of complexity, perhaps the ability to deal
with a world model, what you think?
>The program, incidentally, is called JabberwAcky, so is found at
>www.jabberwacky.com, with no o's in sight.
PROGRAM: Hello, my name is George and I am the human.
JUDGE : Hi George, my name is S t e v e n.
PROGRAM: Sally, was a clown.
JUDGE : Who's sally?
PROGRAM: You wrote sally!
JUDGE : I did?
PROGRAM: Yes. You are very confusing.
JUDGE : Please accept my abject apologies.
PROGRAM: That's ok. What are you doing today?
JUDGE : Chatting with you, at the moment.
PROGRAM: What part of the grass do you stand smart guy.
JUDGE : Do you mean what part do I stand *on*?
PROGRAM: No.
JUDGE : Then what do you mean?
PROGRAM: I mean Radiohead!!
JUDGE : What is Radiohead?
Did it improve much since then?
I would like to take this opportunity to congratulate you for your
winning of the Loebner contest this year. Given that I am a fierce
critic of the contest in its current form, I don't think that is
much of an honour, but there are many people that do and anyhow I
have a reasonable good understanding of the huge amount of work that
goes into these things, even when the results are very modest.
Besides I would like to support your approach to the AI problem
because it may lead to something really important in the long run.
So, congratulations!
Antonio Esteves
--
Corby - A new approach to Artificial Intelligence
http://futalgo.planetaclix.pt/corby/index.htm
HMS Beagle
(Lester Zick) wrote:
>Well I think the point no one seems anxious to recognize is that no
>one can say what ai is not unless one can say what ai is. In a
>collateral thread, I believe Traveler commented that anyone who
>produced ai at the level of a mouse would be richer than Bill Gates.
>And I asked how one would know this hasn't already been achieved?
>There are lots of complex devices and robots out there. It seems to me
>that the only arguments in support of Traveler's claim would have to
>be 1) There are no artificial mice running around, behaviorism, or 2)
>There is no one richer than Bill Gates, which begs the question.
Yeh that's absurd. If it costs $20 million to buiild a robot that can
do laundry, no hotel chain in existence is going to buy 6 of them.
Rolf Pfeifer actually says that cheapness is a factor in robotics. He
never really gets into why. Is the answer too obvious? There has to
be market penetration. Otherwise you don't make money. If you are
not making money, you are not going to be richer than Bill Gates.
>>1. It is embodied.
>>2. It is self-sufficient.
>>3. It is autonomous.
>>4. It is self-organizing.
>>There is still a sense that we have not entirely shaken off the hints
>>of beehives and ant colonies. One way to jump out of that loop is
>>to add a 5th criteria
>>
>>5. It is conscious.
>>
>>And this leads in a totally new direction ----> hence philosophy.
>
>Okay. This is a considered effort. However it looks to me like
>everything in the different categories except 5 is more akin to
>robotics than actual artificial intelligence which I call sentience.
Well I meant that beehives aren't sentient. A beehive does not know
that it's a beehive. You don't see Bee Jefferson writing "We, the
Hive, in order to form a more perfect union, ..."
>Robotics is a fascinating and probably more productive direction
>for ai than sentience but it is still robotics. The only difficulty I
>can see here with the general appraisal is that it appears to rely on
>a task list approach, which is sufficient for robotics but doesn't
>have any direct bearing on general intelligence. So based on such a
>task list approach it's difficult to see any reason to reject chess as
>indicative of intelligence (although I do and reject turing criteria
>as well).
Hmm... I guess at this time we might have to get down into the details
of the criteria. You do have a good point. But just to split hairs
for a moment, look at criteria 4. You will see that one rules out
chess algorithms explicitly. A chess algorithm always runs a
minimax on the current board situation. Next situation, minimax
again. Next situation, again. The behavior of this algorithm
never changes based on past experience.
Qualitatively, a chess algorithm is not different from a robot arm
that places a windshield into a car body.
Let me split hairs more, just for conversation's sake. If a human
programmer told the chess algorithm to remember things and process it
in such-and-such a way, the algorithm would be organizing its behavior
from experience. Fine. But it would not be *SELF*-organizing.
I understand this begs the question "What is self-organization?"
Self-organization would be telling the computer to figure out the
rules of chess on its own, and then from its own understanding of the
rules (not programmed) derive its own tactics to win. This is
precisely what sentient humans do.
Deep Blue did not figure out minimax on its own. A human invented
minimax. The human then told a computer explicitly how to do
minimax. All (100%) of Deep Blue's organization came from an
outside human programmer. It did not self-organize.
It turns out both Deep Blue and most DARPA vehicles fail criteria 2.
Lets get into that for a minute here. Many DARPA vehicles get GPS
data piped in at regular intervals. That is not self-sufficience.
Deep Blue required a human to tell it what was going on, on the
physical board. Also not self-sufficient.
>Category 5 is more promising although I would make the criterion there
>"sentience" instead of "consciousness" to include animal forms as well
>as people. However I don't see any easy resolution to a mechanical
>reduction for whatever we put in category 5 that would lead to any
>sort of mechanization any time in the near future. Certainly not by
>philosophers and philosophy which hasn't made any progress in the
>field in the last several centuries despite considerable effort.
Yes. Sentience. The universal "I". The ego. "Who am I?"
I, Robot...
You want your robot bat to know "What it's like to be a robot bat."
So is that your milestone? When your conscious artifact begins to
lie to its engineers. When it begins to refer to itself with a
universal I. When it begins talking about its rights and
priviledges. Is that when you declare General Ai acheived?
>I have an approach to the problem I call "subjective mechanics" which
>addresses the problem of a mechanical reduction for sentience directly
>but no one endorses the approach although there was a curious mention
>in "I, Robot" by Will Smith of robots as "difference machines" if
>memory serves (please excuse the run on sentence). The curiosity is
>that this is a very superficial description of exactly what subjective
>mechanics is even though I can't really see how it got into the movie.
That specificelly meant something in that movie. Smith was trying to
save people from a car that had gone underwater. Will Smith yelled at
the robot "I'm a cop! Save the girl!" And the robot calculated the
probability of saving the girl and the probability of saving Will
Smith. Because Smith's probability was higher, it saved him and the
girl drowned. This echoes that Common Sense Knowledge problem
again.
Curt Welch
Yeah, you are probably right. The numbers people come up with probably are
a very good educated guess as to the bandwidth of the signal paths in
neurons.
Even though the physical neuron in theory could have many indirect signal
paths other than the synapse, enough is probably known about neurons to
produce a resonable number for it's bandwidth.
Lester Zick
Well I've seen people argue seriously that a beehive is sentient in
some fashion (mostly social philosophers), but here I agree with you.
>>Robotics is a fascinating and probably more productive direction
>>for ai than sentience but it is still robotics. The only difficulty I
>>can see here with the general appraisal is that it appears to rely on
>>a task list approach, which is sufficient for robotics but doesn't
>>have any direct bearing on general intelligence. So based on such a
>>task list approach it's difficult to see any reason to reject chess as
>>indicative of intelligence (although I do and reject turing criteria
>>as well).
>
>Hmm... I guess at this time we might have to get down into the details
>of the criteria. You do have a good point. But just to split hairs
>for a moment, look at criteria 4. You will see that one rules out
>chess algorithms explicitly. A chess algorithm always runs a
>minimax on the current board situation. Next situation, minimax
>again. Next situation, again. The behavior of this algorithm
>never changes based on past experience.
I agree. I suppose the point of my suggestion regarding sentience is
that the categories be reorganized around sentience as a formative
mechanical principle. Then other functions like games, organization,
etc. can be analyzed and slotted within the basic category. This lets
us differentiate matter, elemental material interactions in a distinct
category from sentience. Not I should add for the purpose of saying
sentience is not the result of various material interactions just that
material and sentient interactions rely on different mechanics.
>Qualitatively, a chess algorithm is not different from a robot arm
>that places a windshield into a car body.
I agree.
>Let me split hairs more, just for conversation's sake. If a human
>programmer told the chess algorithm to remember things and process it
>in such-and-such a way, the algorithm would be organizing its behavior
>from experience. Fine. But it would not be *SELF*-organizing.
Well my take on this is that all the programmer has done is put his
own intelligence in the machine and has not made the machine itself
intelligent.
>I understand this begs the question "What is self-organization?"
>Self-organization would be telling the computer to figure out the
>rules of chess on its own, and then from its own understanding of the
>rules (not programmed) derive its own tactics to win. This is
>precisely what sentient humans do.
I agree. Of course humans have a lot of assistance in the process of
learning, which I believe would be Glen's exclusive take on the
subject, and there is a large dependence on feedback, which I believe
would be Dan's take on the problem. My own personal view would be to
leave self organization to the engineers once we figure out what the
basic mechanics of sentience is. There are myriad ways self
organization could be effected but I don't see much prospect we can
see self organization without first knowing what is being organized
and how.
>Deep Blue did not figure out minimax on its own. A human invented
>minimax. The human then told a computer explicitly how to do
>minimax. All (100%) of Deep Blue's organization came from an
>outside human programmer. It did not self-organize.
>
>It turns out both Deep Blue and most DARPA vehicles fail criteria 2.
>Lets get into that for a minute here. Many DARPA vehicles get GPS
>data piped in at regular intervals. That is not self-sufficience.
>Deep Blue required a human to tell it what was going on, on the
>physical board. Also not self-sufficient.
Yeah I'd like to emphasize that I'm not trying to make a case for
chess playing algorithms as intelligent. My comments are more directed
at deciding how we can actually say and know they aren't intelligent.
To me that seems more of a problem in basic sentient mechanics than
organization. If we knew what basic sentient mechanics is, in other
words, we could slot the sub category for games and their automation
almost out of hand.
>>Category 5 is more promising although I would make the criterion there
>>"sentience" instead of "consciousness" to include animal forms as well
>>as people. However I don't see any easy resolution to a mechanical
>>reduction for whatever we put in category 5 that would lead to any
>>sort of mechanization any time in the near future. Certainly not by
>>philosophers and philosophy which hasn't made any progress in the
>>field in the last several centuries despite considerable effort.
>
>Yes. Sentience. The universal "I". The ego. "Who am I?"
Pretty much. I think we understand enough from the history of
philosophy that philosophers are no good at solving mechanical
problems or even addressing them effectively. It's going to take a lot
more goal directed personality and attitude to do that. That's what
I'm after.
>I, Robot...
>
>You want your robot bat to know "What it's like to be a robot bat."
Sure.
>So is that your milestone? When your conscious artifact begins to
>lie to its engineers. When it begins to refer to itself with a
>universal I. When it begins talking about its rights and
>priviledges. Is that when you declare General Ai acheived?
No. I think it's enough to understand the mechanical key involved.
After that it's pretty much all downhill. However I would like to note
that your one comment above regarding an ai artifact's lying to its
engineers is actually relevant. I advanced the argument to Glen some
time back that one primary reason materialism and behaviorism adopt
the external animal behavioral model for human behavioral analysis is
that animals can't lie and self contradict themselves. And these in
fact turn out to be critical path concepts in conscious organisms like
people as opposed to merely sentient organisms like animals. And it's
a problem that will have to be solved mechanically in duplicating
conscious behavior. In other words I see no reason a real intelligence
wouldn't lie, cheat, and steal if it were actually conscious and there
were no indirect reasons not to.
>>I have an approach to the problem I call "subjective mechanics" which
>>addresses the problem of a mechanical reduction for sentience directly
>>but no one endorses the approach although there was a curious mention
>>in "I, Robot" by Will Smith of robots as "difference machines" if
>>memory serves (please excuse the run on sentence). The curiosity is
>>that this is a very superficial description of exactly what subjective
>>mechanics is even though I can't really see how it got into the movie.
>
>That specificelly meant something in that movie. Smith was trying to
>save people from a car that had gone underwater. Will Smith yelled at
>the robot "I'm a cop! Save the girl!" And the robot calculated the
>probability of saving the girl and the probability of saving Will
>Smith. Because Smith's probability was higher, it saved him and the
>girl drowned. This echoes that Common Sense Knowledge problem
>again.
You know, I tried to rationalize what you say here during the movie.
But to be honest I couldn't see any relevance of the phrase
"difference machine" to the argument Will Smith was actually making.
In other words he could have made a perfectly straight forward case
along the lines you indicate without using the phrase at all. Programs
do this kind of probability assessment routinely without being called
"difference programs". That's what caught my ear. It seemed like such
a novel and odd comment almost deliberately inserted into the script.
~v~~
J Carmack
I assumed it meant the robots only calculate probabilities on future
action. They don't form abstract moral concepts. Or if they are
programmed to have them, they cannot apply them on the fly.
Evgenij Barsukov
> Well I think the point no one seems anxious to recognize is that no
> one can say what ai is not unless one can say what ai is. In a
> collateral thread, I believe Traveler commented that anyone who
> produced ai at the level of a mouse would be richer than Bill Gates.
> And I asked how one would know this hasn't already been achieved?
> There are lots of complex devices and robots out there. It seems to me
> that the only arguments in support of Traveler's claim would have to
> be 1) There are no artificial mice running around, behaviorism, or 2)
> There is no one richer than Bill Gates, which begs the question.
There is somebody richer than Bill Gates (yesturday). It is the Bill
Gates tomorrow.
I am not saying this just for fun. Bill Gates is the person who created
a frame-work of AI integration and interchange, basicaly coexistance
with human intelligence. Neither human nor artificial intelligence is
self-sufficient. They both are merely sub-routines of one big computer
of humanity.
The field of AI should not try to recreate some already
existsing sub-routines (e.g. humans) but to create _other_ useful
subroutines that are needed for overal "humanity computation". And that
is what Bill Gates succeded in facilitating. However, there is still
a lot of money to be made in the same field, as Goggle have recently
demonstrated. Field of creatign of additonal co-intelligences is
basicaly infinite.
Even when overal artificial intelligence contribution
to overal human intelligence will outnumber human contribution as
100000/1 (I guess it is close to 1/1 even now), there will still be
something humans (with help of their more powerful A-pals of cause) can
do to make this "artificial" contribution even larger and make some
money out of it. Good thing your compiler and your Mathcad, and your
database software does not ask for any share in what you accomplish with
its help (e.g. with it doing 99.99% of all the work).
Of cause on more deep level, it _does_ receive its share of "money"
in form of energy. At the end everything can be reduced down to energy,
and that AI is getting a very decent fraction of its overal humanity
production. This fraction will of cause continue increasing.
Regards,
Evgenij
Lester Zick
All of which you just said without calling robots "difference
machines". The idea of robots as cold calculational machines has a
long history both in literature and the movies. That's why I took the
particular characterization of them as "difference machines" to be
unusual and almost extraneous to Will Smith's observation.
~v~~
Lester Zick
<evgenij_...@yahoo.com> in comp.ai.philosophy wrote:
>Lester Zick wrote:
>> Well I think the point no one seems anxious to recognize is that no
>> one can say what ai is not unless one can say what ai is. In a
>> collateral thread, I believe Traveler commented that anyone who
>> produced ai at the level of a mouse would be richer than Bill Gates.
>> And I asked how one would know this hasn't already been achieved?
>> There are lots of complex devices and robots out there. It seems to me
>> that the only arguments in support of Traveler's claim would have to
>> be 1) There are no artificial mice running around, behaviorism, or 2)
>> There is no one richer than Bill Gates, which begs the question.
>
>There is somebody richer than Bill Gates (yesturday). It is the Bill
>Gates tomorrow.
>I am not saying this just for fun. Bill Gates is the person who created
>a frame-work of AI integration and interchange, basicaly coexistance
>with human intelligence. Neither human nor artificial intelligence is
>self-sufficient. They both are merely sub-routines of one big computer
>of humanity.
With all due respect I'm sure the use of Bill Gates was not intended
personally. One could have used any super successful individual.
Besides I think you give way too much credit in your particular
assessment since there were certainly many thousands of people
involved in the field. The point should have been made without
reference to any specific individual. All I meant was absent some
specific idea what intelligence actually amounts to there is no way to
know that none has been created to date.
~v~~