How long before superintelligence?
Nicholas Bostrom
I have written a paper outlining the case for believing that there will be
superhuman artificial intelligence within the first third of the next
century. Both the software problem and the hardware promlem are discussed. I
would be interested to get some feedback that I can use in preparing the
final version of the paper
http://www.hedweb.com/nickb/superintelligence.htm
Nick Bostrom
n.bo...@lse.ac.uk
Dept. Philosophy, Logic and Scientific method
London School of Economics
Ben Sharvy
I didn't see a description of how to tell when AI exists.
By writing comparitively ("much smarter than humans") you narrow the topic
to what can quantified....
You wrote:
The computers in the seventies had a computing power
comparable to that of insects. They also achieved
approximately insect-level intelligence.
...suggesting that computing power and intelligence had been distinguished
and measured in distinct ways. If that is the case, what is the difference
between computing power and intelligence, for the purposes of this study
re insects?
--
Eco-Socialist Libertarian Capitalism | Can Men Be Lesbians?
http://www.efn.org/~bsharvy/
*Remove "NOSPAM" from email to earn $100,000 from home reading books.
Niek
Nicholas Bostrom <xn...@dial.pipex.com> schreef in artikel
<349c5...@hades.ndirect.co.uk>...
> I have written a paper outlining the case for believing that there will
be
> superhuman artificial intelligence within the first third of the next
> century. Both the software problem and the hardware promlem are
discussed. I
> would be interested to get some feedback that I can use in preparing the
> final version of the paper
Superintelligence?! I think we should abbandon the concept of
'intelligence' to
begin with. We don't have anything but the faintest clue as to what it is
that
comprises this so called 'intelligence' and the chances are good that
arbitrary
pocket calculator technology could beat the even the most 'brilliant' of
mamals
that roam the face of our sorry planet.
Artificial neural networks can recognize most handwriting or even human
faces
beyond human capacity and there is no doubt whatsoever(*) that they will be
able to perceive meaning from language at levels far beyond our wildest
imagination.
[(*) In my mind that is. I study computer science and I've followed courses
on
neural networks and genetic algorithms as well as the ramifications of
(artifical) 'intelligence'...]
Currently we're at the very entrance of a vast and unbounded sea of
posibilities
and we've only peeked throught the key hole so far.
I'd like to discuss the possible existence of human 'intelligence' within
the
context of a clear understanding of the evolution of our conceptual reality
and I think it's about time to do this anyway (before we start talking
to 'intelligent' programs instead of people).
JRStern
On Sat, 20 Dec 1997 23:19:46 -0000, "Nicholas Bostrom"
<xn...@dial.pipex.com> wrote:
>http://www.hedweb.com/nickb/superintelligence.htm
About the only nontrivial point seems to be your assertion, "Once
there is human-level AI there will soon be superintelligence". I
don't think we're even on the right road yet to get there.
As for your extensive section on hardware speed, it is irrelevent if
(a) the hardware is never turned on, (b) it is turned on without the
appropriate software. There is little current perception that we have
any AI software at all, or any that might scale. Computers are three
orders of magnitude faster today (six to nine, price/performance) than
when I was in college, and AI has advanced not at all.
Is it really necessary or desirable for someone doing a doctorate in
philosophy of science (at London School of Economics?), to be so found
of techno-fetish fads and fancies?
Joshua Stern
JRS...@gte.net
Chris Hooley
On 22 Dec 1997 03:34:00 GMT, "Niek" <nspr...@ibb.ruu.nl> wrote:
>I'd like to discuss the possible existence of human 'intelligence' within
>the
>context of a clear understanding of the evolution of our conceptual reality
>and I think it's about time to do this anyway (before we start talking
>to 'intelligent' programs instead of people).
Go ahead, Niek, open the discussion. Though its connection to AI isn't
readily apparent to everyone, I agree with you that this is important. Any
assumptive base in AI development, in my opinion, sets disabling limiters
on the artificial entities that can grow on that base.
It's like a tree. The deeper we start, the closer to the root, the more
robust the AI. So, the best AI isn't necessarily the first out of the
gate. Or the second, or the hundredth.
Regards
Chris
Hans Moravec
> Computers are three orders of magnitude faster today (six to nine,
> price/performance) than when I was in college, and AI has advanced
> not at all.
> JRS...@gte.net
From 1970 to 1990 the computer power available to most AI programs
hardly budged from 1 MIPS. In 1970 1 MIPS machines like PDP-10s
belonged to whole departments. By 1990, 1 MIPS workstations
were on the desks of individual researchers. Only after 1990
did dilution abate, and the power for individual programs begin
to grow at the industry rate, now doubling every year. Since
1990 the power has grown from a few MIPS to a few 100 MIPS.
Before 1990, it eluded the community to make machines that could
sense well enough to reliably navigate short corridors. Since 1990
robots have driven cross-country, are trundling down dozens of
office corridors and museums around the world, and starting to
explore Mars. Theorem proving was a joke before 1990, but in 1996
a theorem-proving program called EQP running five weeks on a 50
MIPS computer at Argonne National Laboratory found a neat proof of
a boolean algebra conjecture by Herbert Robbins that had eluded
mathematicians for sixty years. Computer chess had been an
embarrassment to the AI community for decades, but in 1997 a
program defeated probably the best human chess player that has
ever lived. Since 1990, expert and learning systems have begun
earning a quiet living in many industry niches. And all that just
in the warm up.
There are none so blind as those who will not see.
Hans Moravec CMU Robotics http://www.frc.ri.cmu.edu/~hpm
Rick Ottolini
In article <349EC3...@cmu.edu> Hans Moravec <h...@cmu.edu> writes:
>
>From 1970 to 1990 the computer power available to most AI programs
>hardly budged from 1 MIPS. In 1970 1 MIPS machines like PDP-10s
>belonged to whole departments. By 1990, 1 MIPS workstations
>were on the desks of individual researchers. Only after 1990
>did dilution abate, and the power for individual programs begin
>to grow at the industry rate, now doubling every year. Since
>1990 the power has grown from a few MIPS to a few 100 MIPS.
Don't forget the "brief, shining moment" of Thinking Machines which
was a high powered parallel LISP machine, before they thought they
could make more money as a scientific supercomputer.
Ben Sharvy
> > Computers are three orders of magnitude faster today (six to nine,
> > price/performance) than when I was in college, and AI has advanced
> > not at all.
> > Joshua Stern
> > JRS...@gte.net
> Theorem proving was a joke before 1990, but in 1996
> a theorem-proving program called EQP running five weeks on a 50
> MIPS computer at Argonne National Laboratory found a neat proof of
> a boolean algebra conjecture by Herbert Robbins that had eluded
> mathematicians for sixty years. Computer chess had been an
> embarrassment to the AI community for decades, but in 1997 a
> program defeated probably the best human chess player that has
> ever lived. Since 1990, expert and learning systems have begun
> earning a quiet living in many industry niches. And all that just
> in the warm up.
>
> There are none so blind as those who will not see.
A recurrent problem in these discussions, and one budding here now before
our eyes, is consistency in the meaning of "AI" from one speaker to
another. Some people use "AI" to mean algorithmic problem solving,
specifically when applied to problems that people try to solve. So
calculators, Deep Blue, the escalating use of computers in industrial and
military production, are all AI. But there is no difference in principle
between these sorts of AI and WordPerfect or a calculator. So, since it is
fairly obvious that computers can do many things people do, and that
computers are getting better at that, that AI is getting better.
Another sense of" AI" is synonymous with "machines with minds." I see no
evidence that machines have greater mental capacities now than ever, or
that they have any minds or quasi-minds at all.
The original essay doesn't make this distinction very well, it seems to
me. It does discuss our ability to simulate a brain using computer
technology, so it is not merely concerned with computational power. But
the bulk of it is concerned only with computational power, it never
explains how we will know when a mind has been created.
Neil Rickert
In <67n072$58t$1...@gte2.gte.net> JRS...@gte.net (JRStern) writes:
>On Mon, 22 Dec 1997 14:46:04 -0500, Hans Moravec <h...@cmu.edu> wrote:
>>> Computers are three orders of magnitude faster today (six to nine,
>>> price/performance) than when I was in college, and AI has advanced
>>> not at all.
>>> Joshua Stern
>>> JRS...@gte.net
While I agree with the list of advances that Hans mentioned, I
nevertheless think Josh Stern's skeptical comment got it right.
Certainly, there are now computer systems able to carry out tasks far
more complex than anything that a bee can do. Yet I would say that
our systems have not yet approached the intelligence of a bee.
As another respondent suggested, this is partly a disagreement
between Josh and Hans as to what we mean by "intelligent."
>> Computer chess had been an
>>embarrassment to the AI community for decades, but in 1997 a
>>program defeated probably the best human chess player that has
>>ever lived.
>An embarrasment only when AI critics said that human performance
>proved something about fundamental theories, a piss-poor argument in
>any case. Still, it's nice to put it away, a little now, a lot pretty
>soon now. This is a case where MIPS helped, and where MIPS failed to
>help until some heuristics were added to better utilize them.
About a week before the recent match, I heard a radio interview with
someone from the IBM team. He was asked about machine learning. He
denied that there was much of that going on. More explicitly, he
said that the main learning was in the activity of the programmers
modifying the system.
> I think a closed game like chess, however
>large the domain, is a dubious test of intelligence in any case.
I agree.
JRStern
On Mon, 22 Dec 1997 14:46:04 -0500, Hans Moravec <h...@cmu.edu> wrote:
>> Computers are three orders of magnitude faster today (six to nine,
>> price/performance) than when I was in college, and AI has advanced
>> not at all.
>> Joshua Stern
>> JRS...@gte.net
>
>hardly budged from 1 MIPS. In 1970 1 MIPS machines like PDP-10s
>belonged to whole departments. By 1990, 1 MIPS workstations
>were on the desks of individual researchers. Only after 1990
>did dilution abate, and the power for individual programs begin
>to grow at the industry rate, now doubling every year. Since
>1990 the power has grown from a few MIPS to a few 100 MIPS.
Hence my three orders of magnitude.
>Before 1990, it eluded the community to make machines that could
>sense well enough to reliably navigate short corridors.
Yes, but the elusion was as much financial as technical. How much
processing power does it take to fire a sonar and match it to a map
and a memory of last known locations? Do you really propose the
software has changed for such simple tasks, or that MIPS help?
[OK, yes MIPS help for image recognition of unexpected obstacles, but
I'm not aware of how much that is achieved even in the best of current
systems. If you'd like to expand on this a bit, I would find it
informative. But even if MIPS are used for this and help a lot,
sensory processing is for most, a mostly separate category from the
kind of consciousness/"mind" intelligence mostly discussed hereabouts,
that might be worthy of the subject label "superintelligence".]
> Since 1990
>robots have driven cross-country, are trundling down dozens of
>office corridors and museums around the world, and starting to
>explore Mars.
Again, it's the convenience factor of microprocessors, not a victory
for AI theory. A victory for "AI" practice? Maybe, in some minimal
way, like the pyramids were a victory for architecture.
> Theorem proving was a joke before 1990, but in 1996
>a theorem-proving program called EQP running five weeks on a 50
>MIPS computer at Argonne National Laboratory found a neat proof of
>a boolean algebra conjecture by Herbert Robbins that had eluded
>mathematicians for sixty years.
I do not know enough about this case to know if there's anything more
significant about this, than about the compilation of tables of
logarithms by computers in the 1950s.
> Computer chess had been an
>embarrassment to the AI community for decades, but in 1997 a
>program defeated probably the best human chess player that has
>ever lived.
An embarrasment only when AI critics said that human performance
proved something about fundamental theories, a piss-poor argument in
any case. Still, it's nice to put it away, a little now, a lot pretty
soon now. This is a case where MIPS helped, and where MIPS failed to
help until some heuristics were added to better utilize them.
Is this to be seen as a victory for AI? Again, yes, but again only
minimally. Let's give it a grade of C. When a 100mip desktop has a
good enough algorithm to beat any human (I expect this within about 10
years), I'll give it a B. I think a closed game like chess, however
large the domain, is a dubious test of intelligence in any case.
A pigeon has better visual discrimination than a human, does that make
it more intelligent? A calculator is better at calculating large
numbers, does that make it more intelligent? Etc, etc.
> Since 1990, expert and learning systems have begun
>earning a quiet living in many industry niches. And all that just
>in the warm up.
I have yet to see one that wasn't equally achievable with trivial
processing power and polynomial estimation. As long as they keep
quiet, idiotic systems can claim they are AI, and nobody much cares.
>There are none so blind as those who will not see.
Maybe if I had more MIPS.
--
Are my negative arguments a case of moving the goalposts as we get
close? Maybe, by the standards I was once taught. The problem is, I
now think the 1950's AI goalposts were located on the wrong field.
It's "nice" to achieve some of these old goals, they might indeed
prove very useful. And, let's call them AI, if it gets us more
funding.
But the *other* thing about all these cheap MIPS, is that we can now
be pretty sure there is nowhere else to go in these directions. A
little cynical blindness may be preferable to seeing things that
aren't there.
Joshua Stern
JRS...@gte.net
Hans Moravec
(This is part 1 of a multipart answer to the previous post
in this thread)
> >Before 1990, it eluded the community to make machines that could
> >sense well enough to reliably navigate short corridors.
> Yes, but the elusion was as much financial as technical. How much
> processing power does it take to fire a sonar and match it to a map
> and a memory of last known locations? Do you really propose the
> software has changed for such simple tasks, or that MIPS help?
>
> [OK, yes MIPS help for image recognition of unexpected obstacles, but
> I'm not aware of how much that is achieved even in the best of current
> systems. If you'd like to expand on this a bit, I would find it
> informative. But even if MIPS are used for this and help a lot,
> sensory processing is for most, a mostly separate category from the
> kind of consciousness/"mind" intelligence mostly discussed hereabouts,
> that might be worthy of the subject label "superintelligence".]
Th finances for SRI's Shakey the robot in the late 1960s and for
Mars rover research at JPL and the Stanford Cart in the 1970s
were comparable to the funding of most of today's autonomous robot
projects. There are more of them today. DARPA's Autonomous Land
Vehicle program in the 1980 had much greater funding, but did not
produce fast reliable autonomous navigation before it ended.
But, as computers became faster in the 1990s, things that had been
out of reach all those decades are beginning to work. For several
years robots have been navigating office networks pretty well, and
a few have been cruising roads around the world. Last year
in a first robot soccer tournament with amusing performance
by robots that tracked the ball, the goal and each other.
Leaving sensory/motor processing out of the intelligence picture
is unreasonable. A million times as much data flows through our
nervous system as registers in our conscious impressions (various
psychological experiments have put the processing rate of
conscious thought at a few tens of bits per second at most).
Consciousness is an evolutionary afterthought, a bit of fluff on
a mountain of unconscious information handling that primarily deals
with sensory and motor issues. It is unreasonable to imagine that
we can do what humans do by imitating the fluff and ignoring the
mountain. By my estimate
(http://www.frc.ri.cmu.edu/~hpm/book97/ch3/index3.html - note new
elements in the final diagram)
doing the job of the mountain will require 10 million MIPS.
The computers in today's robots have about 100 MIPS, comparable,
as Neil Rickert noted, to the nervous system power of a bee).
Unlike Neil, it seems to me that the behavior of the latest
generation of robots is very comparable to bee behavior. The robots
are not as well adapted to their lives as are bees, but their sensory
performance, based mainly on 2D maps and image representations of the
world, is just about as complex as bee perception. (Bees seem to
have rough 2D maps of nearby terrain, and also low res 2D images
of the terrain from different locations in their territory - they
will orient towards "home" when shown slides of their meadow!)
Most robots are not as agile as bees, but one could argue that
aerodynamically unstable fighter aircraft like the stealth
fighter, which are controlled by similar amounts of computer power,
are more agile than bees. For most robots, bee power is there, but
hasn't yet received the evolutionary attention to really maximize
performance. Soon utility robots will become mass produced
products, and then (marketplace) evolution will do its work.
Note also that bees depend as much on luck (fecundity) for overall
success as on skill. There is more attrition in the bee numbers due
to mishaps than is probably tolerable in robot workers. Robots
with bee intelligence will probably seem a little more plodding
than bees, in the interests of safety.
The evidence so far is that one does not have to program in
great complexity to achieve the basics of intelligence. What
is needed is enough processing power the "right" combination of
simple techniques. With the right framework, automatic learning
can acquire the complicated parts.
This is as it should be. Biological evolution, which designed or
brains (and the rest of our bodies) using no more than 10^18 or so
go/no-go tests (order 10^9 births per generation, order 10^9
generations since tiny nervous systems) also is limited in how much
specific complexity it could install at birth. But it was able to
relatively easily provide a lot of copies of whatever simple thing
that worked in past. Our cortex and cerebellum, where higher
intelligence seems to originate, grew enormously in the past
million years, but their microscopic structure seems to have
remained about the same.
An excerpt giving a taste of recent robot evolution that I
happen to be close to follows.
Hans Moravec CMU Robotics http://www.frc.ri.cmu.edu/~hpm
-----------------------------------------------
Cartography
Was it possible to make good on the promises of John
McCarthy's Computer-Controlled Cars essay? In 1969, the only
autonomous computer-controlled wheeled vehicle was a slow-moving
1.5-meter-tall indoor robot, wobbling on a primitive suspension, named
Shakey, at Stanford Research Institute or SRI, a contract-research
company near Stanford University. Shakey was a child of the first
wave of Artificial Intelligence: at its heart was a reasoning program
that used theorem-proving methods to contemplate rooms, walls, doors,
paths and large blocks and wedges that could serve as obstacles or
objects to be pushed around. Programs to interpret camera and
rangefinder data into scene descriptions for the reasoning program,
and to cause the robot to actually execute the resulting plans, were
considered peripheral, and assigned to support engineers. In the
early 1970s computer vision was less than ten years old, and almost
all was of the "blocks-world" type, initially developed at MIT for
finding child's blocks on a table top. The SRI team constructed a
world to match the means: an array of blank-walled rooms containing a
few uniformly-painted large blocks and wedges. Shakey was the star of
a somewhat misleading 1970 Life magazine article, but its most
impressive feat--moving a wedge to a block, ascending it and pushing
off a smaller block--was recorded on film piecemeal, requiring
multiple takes--and several hours--for each error-prone stage.
So, McCarthy's challenge was unexplored territory. How could
one expect to interpret an image sequence at the
many-frames-per-second rate probably necessary for driving, on a 1/2
MIPS (Million-Instruction-Per-Second, each instruction causing work
similar to adding two eight-digit numbers) machine like SAIL's 1969
computer, a Digital Equipment Corp. PDP-10? A good digitized picture
of the road is, by itself, an array of a fraction of a million
numbers. To even touch each of these pixels (picture elements) with a
program took several seconds--and doing anything substantial at least
several times longer. How would it be possible to respond swiftly to
traffic, obstacles, road signs and other items that flash through all
parts of the image?
Maybe programs could somehow select and work with only the
essential parts of each image. In 1971 Rod Schmidt used this approach
in the first Cart thesis. With Schmidt's program, about 200 thousand
memory-straining bytes of strenuous-to-write but efficient assembly
language, the Cart, moving at a very slow walking pace, visually
followed a white line on the ground. The program contained a
predictor for the future position of the line in the camera scene,
based on its past position, and searched about 10% of the next image
to find it again. The new-found location served as the next input to
both the predictor and a steering calculation. Using about 1/4 MIPS,
half the power of the PDP-10, the program could handle one image a
second, and enabled the Cart to follow a line for about 15 meters at a
time--if the line was unbroken, did not curve too much, and was clean
and uniformly lit. Schmidt noted that handling even simple
exceptions, such as brightness changes caused by shadows, would
require several times as much computation, to consider the wider
alternatives. Detecting and responding to obstacles, road signs and
other hazards promised to be much more complicated.
Shakey's vision programs, as most others of the time, reduced
images to a short list of geometric edges before doing anything else.
The approach was quite inappropriate for outdoor scenes containing few
simple edges, but many complicated shapes and color patterns. A major
exception to the blocks-world approach was a project begun at SAIL
with impetus from Joshua Lederberg, a Stanford geneticist with a Nobel
prize, and John McCarthy's enthusiastic support. It was to look for
changes on Mars that might indicate growing life. Using digital
images from Mars-imaging spacecraft Mariners 4, 6, 7 and 9, the
project worked to register, in geometry and color, views of the same
regions taken at different times, so that any differences could be
detected. Since the spacecraft locations were known only
approximately, the image registration process was to be guided by
surface features themselves. A graduate student, Lynn Quam, and
others developed a collection of statistical, intensity-based
comparison, search and transformation methods that did the job (alas,
no unambiguous life signs were spotted). Since they dealt with
complex natural scenes, the Mars methods also seemed appropriate for
interpreting imagery from an outdoor vehicle. The NASA Mars rover
program was then in the ascendant, so there was a double bond between
the Cart concept and the Mars group. I arrived in late 1971, an
enthusiast for space and robots, and soon adopted the Cart from Bruce
Baumgart, another graduate student who had been maintaining it for
possible use in his own research in computer graphics and vision. The
Cart had little research reputation, but discussions with Quam
produced a plan where I would provide a working vehicle (a non-trivial
project given the shoestring construction) and PDP-10-resident motor
control software. Quam and company would adapt their image methods
for visual navigation. By 1973 I was having a good time building and
test-driving new robot hardware and software, when, in a
remote-control misstep, I crashed the poor Cart off a small loading
ramp. Months of low budget repairs left its TV transmitter still
broken, and led me to beg McCarthy to invest several thousand dollars
in a replacement. He agreed, but insisted that I first demonstrate
competence in writing programs for computer vision.
Real-time performance was not an issue in interpreting Mars
images. The missions were several years apart, and, until the Mariner
9 orbiter, each produced only a few dozen images. The Mars group
could afford to run search programs for hours at a time to find
precise and dense matches over large image areas. A Cart-driving
program might forgo this precision and coverage in exchange for speed.
It seemed many tasks could be accomplished with just two basic image
operations--one to pick out a good collection of features (distinctive
local regions peppered across a scene), and another to find them in
different views of the same area. Three-dimensional locations could
then be determined by triangulation, obstacles detected, and the
motion of the robot deduced. I set about to find fast implementations
of these ideas. Working mostly with spatially compressed
images--where squares of 4, 16, 64 and more pixels were averaged into
one--and cleverly coded in assembly language, my operators were able
to pick out a few dozen good features in one image and reacquire them
in another using about ten seconds of computer time. In 1975 I built
a program around them that controlled the Cart's heading by tracking
horizon features on the roads around SAIL. The program would
repeatedly digitize a frame and, in fifteen seconds, determine the
horizontal displacement of features on the (usually tree-lined)
boundary between ground and sky between frames, calculate a steering
correction, and drive the robot up to ten meters. It did its
unambitious task quite well, and was fun to watch, but it was intended
as mere practice for the main event: a much more ambitious program
that would drive the Cart through an obstacle field by visually
tracking its surroundings in three dimensions--to build a map,
identify obstacles, plan safe routes and--most difficult--deduce and
correct the robot's motion from the apparent motion of those
surroundings. I decided to approach this task in full three
dimensions from the start, hoping eventually to run the robot on the
rolling adobe terrain outside the lab (a vain hope).
The Cart carried a single camera, so it was natural to use
driving motion to provide multiple viewpoints for triangulating
distances. Complicating the matter, the Cart's motors were very
imprecise, so its moves would have to be deduced simultaneously with
the position of tracked objects. The Mars team had a similar problem,
and another student, Donald Gennery, had already written a "camera
solver" for it. I struggled with this approach through early 1977.
The program would take a picture and choose up to a hundred features.
It would then drive the robot forward about a meter, stop, take
another picture, and search for the same features in the second image.
Then it would invoke the camera solver to find the robot movement and
the three-dimensional locations of the features that explained their
apparent motion from one image to the other. Despite much
fine-tuning, the program's error rate never dropped below about one
wrong motion solution in four, meaning the robot could move about four
meters before becoming confused about its position--discouraging. The
camera solver repeatedly tweaked an estimate of the robot's motion to
make the features line up as well as possible, and threw away those
that seemed too far off. It worked well for high-quality spacecraft
images of an almost two-dimensional surface, with few matching errors,
good initial camera position estimates, and ideal sideways motion
between images.
My data, with poor position estimates, from noisy TV images of
a nearby scene, with plenty of perspective distortion from frame to
frame, was something else. Ten to twenty percent of the feature
matches were wrong, often because an area chosen in a first image had,
in a second image, been eclipsed or changed in appearance by
point-of-view or lighting effects or camera noise. Position accuracy
in my low-resolution images was modest, compounding the serious
limitation that forward motion stereo is mathematically insoluble for
points near the camera axis. The combination of many outright
incorrect matches and large uncertainties in the correct ones made
finding the robot motion a chancy proposition. It was necessary to
track about one hundred features to succeed even three steps in four,
consuming several minutes of computer time. Months of fiddling with
the program's mathematics and assumptions made little difference.
Eventually, I chose to add some robot hardware to reduce the
computational uncertainties.
Multiple cameras or a repositionable camera on the robot would
permit true stereoscopy, from precisely located relative views,
removing the biggest source of uncertainty. Mismatched features
between stops might then be pruned, before solving for robot motion,
by exploiting the constraint that the mutual three-dimensional
distances between pairs of features should remain unchanged by a move.
Vic Scheinman, an engineer and graduate student whose main interest
was robot arms, but who often graciously lent his mechanical expertise
to the Cart project, found for me, in his basement, a mechanism able
to slide the camera about 60 cm from side to side. Motorized, this
provided a fine stereo baseline. Errors were further reduced by
taking pictures at nine places along the track, exploiting the
nine-way redundancy.
The final result, first sufficiently debugged in October of
1979, was a program that would track about thirty small image features
at a time to drive the robot through indoor clutter, avoiding
obstacles and accumulating a sparse map of the scenery. Its 1 MIPS
computer worked for ten minutes to prepare for every meter-long move.
In five hours it would arrive at a requested destination at the
opposite end of a 30-meter room, succeeding in about three traverses
in four. In outdoor tests the Cart managed to travel only about 15
meters before becoming confused. Harsh contrast between sunlight and
shadow overwhelmed the old-style TV camera tube and greatly degraded
its vision.
When the navigation failed, it was usually because of the
process intended to prune errors made in tracking features from one
stop to the next. The nine-eyed stereoscopy located features quite
reliably in three dimensions at each stop, relative to the robot's
position and orientation. Though the reference frame changed after a
move, the mutual 3D distance between pairs of features should not.
The pruning process aggressively strained out those features which
violated this "rigidity" criterion, up to half of the 100 raw matches.
But sometimes, by chance, about once in 100 moves, some of the
mismatched points would happen to mutually support one another more
strongly than the correct ones, and the pruning would retain them, and
reject the good matches! The robot would then mis-estimate its
position and heading, mess up its accumulating map, and run into
trouble.
Gridwork
In 1980, Ph.D. done, I moved to Carnegie Mellon University,
and set up a small "mobile robot laboratory," where we built and
worked with new robots. My first two students, Larry Matthies and
Chuck Thorpe, streamlined the Cart visual navigation program by using
fewer images and exploiting constraints, like the robot's flat indoor
floor, to simplify and speed up the program tenfold. They increased
its navigational accuracy by modeling geometric uncertainties more
precisely. The changes hardly altered the once in 100 navigational
failure problem. Apparently, given the matching error rate, 100
features provided bad luck too much of an opening. The chance that
random errors could overpower good data was significant in that small
a sample.
In 1984 our laboratory accepted a research contract from a
small startup company, Denning Mobile Robotics, challenging us to
navigate robots using range measurements from an obstacle-detecting
belt of 24 sonar units. Computer vision then was much too expensive
for an affordable machine, but the sonar devices, developed for
autofocusing Polaroid cameras, cost only a few dollars.
Each sonar measurement is the echo time of a sound pulse in a
30 deg wide beam, and although the distance is quite accurate, there
is no indication of where, laterally in the huge 30 deg field of view,
the echo originates. The Cart program's techniques, dependent on
pinpointing distinctive features in the scene, could not be used, even
were they reliable or fast enough. Another student, Alberto Elfes,
and I devised a completely different approach that, instead of trying
to catalog the location of objects, accumulates the "objectness" of
locations. Whereas the identity or even existence of particular
objects is always in question, locations around the robot can safely
be assumed to exist, and can serve as permanent "buckets" to steadily
accumulate even a light rain of evidence about their contents.
The area around the robot is divided into a grid. A program
maintains a number for each cell of the grid representing the evidence
accumulated thus far that the corresponding cell contains something
or, contrarily, is empty. With each new sonar ping, cells
representing its sweep in space are altered. Cells at the echo
distance gain evidence of occupancy (because somewhere at that
distance is the cause of the echo), while those in the interior of the
beam lose occupancy evidence (because anything in the interior volume
would have caused a shorter echo). The amount of evidence adjustment
varies across the beam's volume, since the sound intensity, and so the
sensing reliability, declines gradually away from the center of the
beam, and with distance.
Given the problems with the Cart approach, we were very
surprised in 1985 when our first program using the evidence grid
method enabled a robot to build maps of its surroundings and cross our
cluttered lab with almost perfect reliability. But was it practical?
A reasonably fine three-dimensional grid divides a room into several
million cells, and each sonar measurement affects tens of thousands of
those, demanding more computer memory and speed than could reasonably
be installed on a robot in 1985. By using a very coarse grid, with
cells 30 centimeters on a side, in only two dimensions, like a map, we
produced an efficient program able to handle ten sonar readings per
second on the then-1/2 MIPS Denning robot. A key navigation
step--aligning two maps of the same area--took 3 seconds. Close, but
not good enough for a fast-moving robot. A future robot model with
more powerful computers could probably handle it, but the company was
stretched too thin to immediately pursue that option.
In the meantime, we continued exploring this promising path
with funding from the Office of Naval Research. We applied the grid
approach to distances extracted from stereo vision, then combined
stereo and sonar data in a single map. We put its mathematics on a
sound foundation of probability theory. We acquired more powerful
computers, and did more extensive experiments. One of these revealed
a weakness: in a narrow, smooth-walled hallway--unlike our very
cluttered lab--sonar pulses ricocheted around, like light in a hall of
mirrors, most readings were misleading, and the resulting maps were
worthless. The evidence patterns the program added to its map for
each reading--which we had constructed by hand from the specifications
of the sonar units--were not indicative of what the sonars were really
saying about the hallway.
In 1990 we developed a "learning" approach to find better
evidence patterns. The patterns were encoded as mathematical formulas
with a dozen parameters--"knobs" controlling their shape. We
carefully measured the hallway by hand and constructed a near-perfect
ideal map of it. We ran a robot down the hallway in a precisely known
way, and collected sonar data at regular intervals. Then we wrote a
program to repeatedly process the collected data to simulate a
map-building robot traveling down the hall. After each simulated run,
the program compared the resulting map with the ideal. Then it
tweaked the knobs on the evidence formula, and simulated another run.
If the map using the new evidence patterns was more like the ideal
than before, the program moved the knobs further in the same
direction, otherwise it adjusted them the other way. During the
course of hundreds of thousands of simulated robot excursions,
consuming days on machines which by then had reached 10 MIPS, the
program gradually "tuned in" an evidence pattern that produced an
excellent representation of the hallway, and was suitable for other
smooth-walled surroundings.
FIGURE 2-5:
<<2D map of specular hallway>>
CAPTION: 1990 - 2D evidence grid of a difficult hallway. 624 range
measurements were obtained along an 8 meter hallway by a robot with a
belt of Polaroid sonar units. More than half the ranges were too long
or missing because of deflections by the mirrorlike walls. A good
reconstruction, in a 64 by 32 cell grid, was nevertheless obtained,
because the evidence patterns representing sonar ranges had been well
adjusted for this kind of environment by a learning process. The
traversed hallway runs left-right. The start of an adjoining hallway,
running upwards, can be seen on the right.
Heightened Sense
On 10 MIPS computers, two-dimensional evidence grids can be
built and used quickly enough to be used at normal indoor speeds. In
the 1990s a growing number of research groups worldwide are fielding
robots that trundle around office settings, guided by 2D grids and
other kinds of maps of similar complexity. Though impressive as
short-term demonstrations, these machines invariably run into
problems, like colliding, or becoming lost or trapped, perhaps several
times a day. Although they are much better than sparse
representations like the Cart's few dozen features, maps with a few
thousand cells still have a significant probability of being fooled by
unlucky combinations of sensing errors.
The probability of misapprehension drops as the number of
independent things known about the world grows. A most attractive way
to increase the information in evidence grids is to do them in full
three dimensions. In two dimensions, differences of things at
different heights become confounded, and only a blurry kind of map is
possible, so there is little advantage to reducing the cell dimensions
below about 10 centimeters on a side. In 3D, there need be no
blurring, so the grid resolution can be higher, and its cell contents
more certain. In 2D, a chair is a fuzzy blob a few cells across,
indistinguishable from other objects of similar size. In 3D, a chair
can have legs, seat and back, and be recognizable by shape. A
three-dimensional map would enable a program to plan a complex path
that weaves not only around but over and under obstructions. The
chance of navigational misperception could become negligible.
The price is high. A 2D grid, which we can manage with 10
MIPS, has a few thousand cells. In 3D it is tempting to greatly
reduce the cell size, to a few centimeters or less. Quadrupling the
resolution of a 2D grid increases its number of cells by a factor of
16. In 3D, the effect is compounded by the third dimension, which
itself may be 100 cells high, resulting several thousand times as many
cells, to store and process, in 3D as in 2D. It seemed 3D needed
computers with over 1,000 MIPS.
In the early 1990s, 1,000 MIPS could be found only in
supercomputers. In 1992 I spent a sabbatical year in Boston as a
guest of supercomputer maker Thinking Machines Corporation. My intent
was to write a program to project evidence rays into 3D grids using
the new CM-5 machine, composed of several hundred 20 MIPS computers
working in tight concert. My desired programming environment was not
ready, so I used a regular computer workstation to develop the
programs that would eventually run in many copies on the CM-5.
That preamble grew into a eight-month project. Large grids
made possible economies of scale. Just about every sensor could be
well-represented by evidence rays symmetric around an axis of
propagation, so could be represented by 2D radial slices of cylinders,
rather than 3D boxes. For each new measurement, the transformation
that rotated these slices into cylinders skewering 3D grid maps was
very similar from layer to layer in the grid. It could be worked out
for one slice in a very efficient and compact form, and repeatedly
reused. In each slice, the cells could be sorted by radius from the
propagation axis, and then used only up to the maximum radius of
actual data at the slice. Only the typical cone of evidence rather
than the whole cylinder needed to be filled. Perfecting these and
other ideas produced a program about 40 times as efficient as
anticipated. A surprise speedup came from an advance in the compiler,
which translates programs into machine code. Setting the optimization
of "Gnu C" to level 3 sped up the code about 2.5 times. All together,
the final program ran 100 times faster than expected. Further, my
1992 workstation computed at 25 MIPS, giving yet another 2.5 factor.
Equally important, the workstation had enough memory, about 16
megabytes, to store an entire 3D grid and its supporting structures.
After eight months of preparing to use a supercomputer, I no longer
needed one!
At 25 MIPS, in a cubic grid 128 cells on a side, the program
added 200 broad sonar-like measurements per second, or 4,000 thin
rays, as might represent laser or stereoscopic range values. It was
fast enough for developmental experiments, if not quite for working
robots.
The 1992 sabbatical had left me with an unexpectedly practical
core of a 3D grid program, but not an entire robot perception system.
Groundwork and distractions stretched the next step over several
years. Sonar, providing only hundreds of fuzzy measurements per
minute, was not a promising sensor for a high-resolution 3D grid with
millions of evidence-hungry cells. Scanning laser rangefinders, used
in a simpler way by some other robot projects, could provide the
requisite data, but were large, expensive, balky and power-hungry. I
imagined our methods showing up in future small inexpensive
mass-produced robots, and looked for something more practical.
Television cameras had shrunk to finger size, so stereoscopic vision
again looked attractive.
Wide-angle lenses give nice coverage, but they introduce a
pronounced "fish-eye" distortion. By 1995 I had written a program to
correct camera imperfections of all kinds. The program works by
viewing a precise calibration pattern of several hundred black spots
through a camera, from which it derives a "rectification" function
that straightens out the raw image into one with ideal geometry.
Still, the work was dragging. Since my last sabbatical had been very
productive, I accepted the opportunity to do another in 1996, this
time at Daimler-Benz research in Berlin, Germany.
My workstation in Berlin had 100 MIPS and 64 megabytes. I
wrote a stereoscopic front end for 3D grids, using methods resembling
those of the Cart program twenty years earlier, but improved in many
subtle ways. Input came from two cameras side-by-side on a tripod
moved by hand (there was a robot in the next room, but not quite ready
to host my programs). Instead of extracting a few dozen features from
each stereo set, the new program extracted about 2,500. Every found
feature was translated into at least two evidence rays, one from each
camera, to the triangulated location of the feature. A grid 256 by
256 by 64 cells represented a volume 6 meters wide by 6 meters deep by
2 meters high.
With economies of scale in the stereoscopic ranging as in the
grid updating, the program was able to process a pair of images and
add the consequent 5,000 evidence rays to the grid in about five
seconds.
The following image contrasts the new result with one from the
Cart. Each map was derived from about 40 separate camera views of the
scene. The Cart map took about 40 minutes of computation at 1 MIPS
with 1/2 megabyte of memory, while the grid required 80 seconds at 100
MIPS and 20 megabytes. The Cart map marks the location of about 50
features, while the grid indicates about 100,000 cells are occupied
(also that 1.5 million are empty and that 2.5 million remain unknown).
The Cart map is barely usable for navigation. The grid map promises
to support not only highly reliable navigation, but the recognition of
objects by 3D shape.
FIGURE 2-6:
<<Cart and Crayfish 3D maps>>
APTION: 1979 and 1997 - 3D maps from stereoscopic images. In each
example about 40 images similar to the ones shown were
stereoscopically processed to make the maps below them. The map on
the left shows the Cart's position near the beginning of a run, its
camera field of view, its planned path and a kind of perspective view
of features that passed all consistency checks. Each feature is a
black dot linked to the ground by a diagonal line. The diagonal's
length is the feature's height, where it meets the ground is its
forward and lateral position. The features are also marked by dots
overlayed on the camera image. To aid interpretation, about a dozen
clusters of features were hand-labelled with the identity of the
object on which they were found. The map on the right is a
perspective view of the occupied cells in a 256x256x64 grid
representing a six meter square by two meter high bite of the office
above. To aid visualization, the cells in about a dozen box-shaped
volumes selected by hand were "spotlighted" with distinctive tints.
The 3D grid results are very encouraging, everything I had
hoped for, but open many avenues for improvement. An interesting one
concerns learning the evidence patterns to best represent individual
measurements, an approach which greatly improved 2D sonar maps. The
beams are now constructed from the stereoscopic geometry and many
guesses. But how to guide the learning process? In 2D the learning
evaluated computed grid maps against a hand-constructed ideal map.
The thousandfold greater number of cells makes constructing a 3D ideal
impractical. But, though we don't have an ideal 3D model, we have
excellent pictures of the 3D scene in the original stereoscopic
images! The occupied cells of a computed grid can be viewed, as in
the illustration, from perspectives corresponding to the camera
images, and compared to them. The better the match, the better the
grid, if only the grid cells had colors corresponding to the image.
In fact, the grid can be colored by sometimes projecting the image
colors onto the cells, instead of matching them. Not only will this
colorizing and matching approach allow the grid program to tune itself
up at any time, but the color will enhance object recognition and
other operations.
The work goes on. A new graduate student, Martin Martin, is
using 3D grids in a robot whose computer will soon be upgraded to 500
MIPS. We hope soon to demonstrate navigation for practical near-term
possibilities like Chapter 4's free-roaming robot vacuum cleaners.
But what about John McCarthy's self-driving cars? Our lab's
slow little robots, even when perfected, would be roadkill if they
ventured onto a street. But less timid machines have emerged to meet
the challenge.
Fast Cars
Minicomputers small enough to fit into cars appeared in the
1970s, and a few actual computer-driven cars and trucks appeared
within a decade. As with small robots, their performance remained
discouraging until the 1990s.
In 1977, Japan's Mechanical Engineering Laboratory built a
stereoscopically-guided autonomous automobile that could follow
well-defined roads for distances of about 50 meters at up to 30 km/h.
The secret was highly specialized hardware filling a rack on the
passenger side of a small car, using input from two small television
cameras mounted sideways one above the other on the car's front grill.
The cameras' video signals were electronically filtered, to detect
brightness changes which were then registered as digital pulses. The
pulse streams from the two cameras were matched at various lateral
displacements by simple digital circuitry: displacements with many
matching pulses indicated objects at certain distances. When properly
adjusted for lighting and contrast, this circuit, doing the equivalent
of about 50 MIPS of computing, gave the distance of about eight major
visual discontinuities, such as road embankments and obstacles, thirty
times per second. The distances were sampled ten times a second by a
1/4 MIPS minicomputer programmed to keep the vehicle on the road and
veer around obstacles. The system performed impressively when the
edge detectors were carefully adjusted, and the road and obstacle
boundaries were contrasty enough--the test roads were usually edged by
white lines. At other times, the vehicle was unpredictable and
dangerous. The problems were fundamental to the approach, and the
project was ended in 1981.
In 1984, as part of a larger program called the Strategic
Computing Initiative, DARPA, the Department of Defense Advanced
Research Projects Agency[1], initiated an overambitious program called
"Autonomous Land Vehicles" (ALV). It promised stealthy robot crawlers
to do battlefield reconnaissance, sabotage and perhaps combat. The
prior decade of computer vision work had convinced the managers at
DARPA that stereo vision was too hard a problem for their timeframe,
but they guessed that the rest of the perception and navigation
problem was tractable. Before being abandoned in 1989, the project
had financed a half-dozen small experimental vision-guided vehicles
and two big ones. A rough-terrain vehicle as big as a bus at
Martin-Marietta Corporation in Denver was equipped with about 50 MIPS
of specialized computer power, color television cameras and a scanning
laser rangefinder that provided, twice a second, a 128-by-256 array of
distance measurements, doing by optics and electronics what was
impractical by computation. A similar machine, a large Chevy van,
heavily modified, was developed at Carnegie Mellon University. By the
end of the project, Martin Marietta's "ALV" and Carnegie Mellon's
"Navlab" were both driving down dirt roads at speeds up to 20 km/h,
but usually much slower, tracking road boundaries in color video
images, and stopping for obstacles detected by a minimal processing of
the laser range data. Both were able to do this as long as the road
boundaries were relatively well defined and without major
discontinuities. The simple road identification techniques were often
fooled, and both were unlikely to stay on a road for a whole
kilometer.
A project at the Bundeswehr University in Munich, funded
partly by German automobile and electronics companies, began in 1984,
and by 1989 had produced a van that sometimes drove autonomously on
the Autobahn at up to 100 km/h, guided by a TV image processed by an
onboard array of a dozen special-purpose computers, each--at 10
MIPS--able to track a single image patch at 13 frames per second. The
features were selected by a human at the beginning of an autonomous
run, typically one at the left edge of the highway or the lane,
another the right edge, and others chosen on license plates or other
distinctive marks on traffic ahead and to the sides of the van. Using
motion-prediction techniques, the image processors maintained their
visual locks for many minutes. Their output went to a minicomputer
programmed for motion control, to keep the vehicle in the lane, at a
safe distance from the other traffic. In 1989 the van ran the entire
length of an empty 20 km Autobahn spur.
The above systems critically depended on an alert human
supervisor with a manual override button--these simple-minded and
dangerous machines were very easily confused by such common driving
events as shadows, road stains, lampposts, stopped cars or sudden
curves.
FIGURE 2-7:
<<Navlabs 1 to 5, all in a row>>
CAPTION: The Carnegie Mellon Navlabs. Numbers 1 to 5, left to right,
a sequence that stretches from 1986 to now, 1 meter/second to 150
km/h, 10 meters to 5000 km, 1 MIPS to 100 MIPS, refrigerator-sized to
laptop--and half the fun was getting there.
Carnegie Mellon's Navlab project outlived the ALV program.
Managed by mobile robot lab alumnus Chuck Thorpe, who had earned his
Ph.D. in 1984, and funded by automotive, farming and mining industry
as well as government contracts, it expanded into a range of
applications on roads and fields and mines. The road work was
particularly instructive.
In 1984, Navlab's predecessor, the Terragator, a desk-sized
outdoor robot radio-controlled by a housebound 1 MIPS computer, crept
along jogging trails at about a meter per second. Occasionally it
mistook a tree trunk for the road, and tried to climb it! The Navlab,
running no faster, replaced it in 1986. It was a big blue truck
crammed front-to-back with racks containing 1 MIPS of computing, a
huge air conditioner and a heavy generator. The original drive train
was replaced by hydraulics for control at very slow speed. A
succession of hand-crafted programs to pick out the boundaries of the
road ahead in the color camera view elicited a few general tricks (for
instance subtract the green from the blue channel: the road is bluish
from reflected sky, the roadside greenish from vegetation, subtraction
enhances the distinction, and eliminates shadows), and slowly
improving performance. In 1988 the van negotiated a network of empty
city streets at a few kilometers an hour. By 1990, with about 10 MIPS
of improved computers, it sometimes reached the 40 km/h top speed of
the hydraulics. Still, almost every time a new road type or condition
was encountered, the programs needed tinkering.
In 1990, Navlab was replaced by Navlab 2, a humvee that had
been a military ambulance, containing three 20-MIPS workstations.
Thorpe's student Dean Pomerleau introduced a new approach. Instead of
handcrafting the road finders, Pomerleau's "ALVINN" program trained a
neural net, with about 5,000 adjustable interconnections, to imitate a
human driver. The net's input was a low-resolution image of the road,
preprocessed with the blue-green subtraction trick. Its output
determined the steering-wheel position. In early versions, the net
required hours of training, which were provided by simulator scenes or
long videotapes of human road trips. Training tricks gradually
reduced the time. Each original image was expanded into dozens. Some
simulated the vehicle being further left or right in the lane, with
corresponding adjustments to the steering, providing experience
outside of the boundaries of normal operation. Others added random
clutter in the off-road locations, training the net to ignore the
contents of those regions. Ultimately, the time to learn new roads
was reduced to about five minutes. For long trips, the system was
provided the neural interconnection weights for many road types. It
tested each against the current road, using consistency of the
steering output as a measure of suitability. In 1991, ALVINN
traversed 30 km of busy highway at 70 km/h, and Pomerleau received his
Ph.D.
In 1995, another Navlab, another student and another idea
topped this performance significantly. Navlab 5 (3 and 4 are
unimportant) is a Plymouth minivan. Instead of being crammed full of
equipment, it is controlled by a 50-MIPS laptop computer. The camera,
rather than being ostentatiously mounted on the cab roof, peeks out
the windshield over the rear view mirror. The student is Pomerleau's,
Todd Jochem. The idea was to replace ALVINN's general learning,
reined in by a large quantity of contrived training, by a specialized
system in which the known constraints were built in, and only the
changeable parts were learned. "RALPH" begins with the same 32-by-32
pixel low resolution road picture as ALVINN. In it, the lane ahead of
the vehicle, tapering into the distance, appears as a wedge. RALPH
distorts the image to stretch this wedge into a straight vertical
ribbon, then averages the ribbon down to a single 32-pixel vector,
representing an average cross section of the lane. Of course, this is
true only if the lane lies straight ahead. If it angles right or
left, or curves, the stretching and averaging will blur together
different parts. So the program also tries transformations
representing a range of angles and curves. For each hypothesis, the
program estimates the blur in the resulting vector from the brightness
changes from one cell to the next: blurring softens such changes. The
sharpest vector is kept, the others are discarded. To learn a road
type, its 32-number vector is instantly recorded (compare this to
ALVINN's 5,000 numbers learned over 5 minutes). In automatic driving,
the current vector slides across a memorized vector, and where it best
matches indicates where the vehicle is in the lane. On a new road,
the current vector can be checked against a library of thousands to
find the best match. RALPH can even train itself for new road types,
by comparing a vector made from only the lower half of its image to
the vector from the upper half. If they match, the road is the same
up ahead as locally. If not, the upper half is a new road type, which
can be added to the library. The program has driven successfully in
the dark, in heavy snow and in driving rain. When visibility is near
zero, it can lock on to weak or transient features like pavement
stains or water ripples from preceding cars! In the summer of 1995
RALPH drove Navlab 5 from Washington D.C. to San Diego, CA, in control
98.2% of the time, at an average speed of over 100 km/h. Jochem got
his Ph.D. in 1996.
Another generation of students is now building this work into
"the highway of the future," among other applications.
Neil Rickert
>Leaving sensory/motor processing out of the intelligence picture
>is unreasonable.
Yes, I agree. Yet much of the discussion omits it. Even much of the
NN work is done with simulators which do not need to concern
themselves with sensory data or motor output.
>The computers in today's robots have about 100 MIPS, comparable,
>as Neil Rickert noted, to the nervous system power of a bee).
>Unlike Neil, it seems to me that the behavior of the latest
>generation of robots is very comparable to bee behavior.
I would distinguish between bee behavior and general insect
behavior. The bee has a complex social system involving complex
interactions, and I think these are still beyond what you can do in
today's robots.
>Note also that bees depend as much on luck (fecundity) for overall
>success as on skill. There is more attrition in the bee numbers due
>to mishaps than is probably tolerable in robot workers. Robots
>with bee intelligence will probably seem a little more plodding
>than bees, in the interests of safety.
Bees are short lived, so that this attrition is acceptable. The
evolutionary strategy appears to be to create relatively low cost
units (bees) with short life spans, and to use the processes of
evolution to keep the bees well adapted to their environment. For
mammals, the strategy is different. The units are considerably more
expensive, and have substantially longer lives. With a long life
span they cannot depend on evolution as the main tool for adapting to
their environment. Instead they require a system of learning to
allow self adaptation.
It is this ability at self adaptation which is missing from AI. Some
of the research effort on this self adaptation seems to be in the
form of specious proofs that the ability could not possibly exist
(thus justifying the failure of the research community to solve the
problem).
Hans Moravec
Neil Rickert:
> I would distinguish between bee behavior and general insect
> behavior. The bee has a complex social system involving complex
> interactions, and I think these are still beyond what you can do in
> today's robots.
I would say that the robots (few as they are today, and sub-bee
in processing power) moving things around in our factories
play as complex and vital a social role in their societies as
bees do in theirs.
Hans Moravec
(This is part two of a multipart reply to this thread)
hpm:
> > Theorem proving was a joke before 1990, but in 1996
> >a theorem-proving program called EQP running five weeks on a 50
> >MIPS computer at Argonne National Laboratory found a neat proof of
> >a boolean algebra conjecture by Herbert Robbins that had eluded
> >mathematicians for sixty years.
JRStern:
> I do not know enough about this case to know if there's anything more
> significant about this, than about the compilation of tables of
> logarithms by computers in the 1950s.
The author's original description is found in
http://www.mcs.anl.gov/home/mccune/ar/robbins/
Here is an article excerpt from last year's New York Times
with a good summary. The whole article is available at
http://www.nytimes.com/library/cyber/week/1210math.html
NYT December 10, 1996
Computer Math Proof Shows Reasoning Power
By GINA KOLATA
Computers are whizzes when it comes to the grunt work of
mathematics. But for creative and elegant solutions to hard
mathematical problems, nothing has been able to beat the human
mind. That is, perhaps, until now.
A computer program written by researchers at Argonne National
Laboratory in Illinois has come up with a major mathematical proof
that would have been called creative if a human had thought of it. In
doing so, the computer has, for the first time, got a toehold into
pure mathematics, a field described by its practitioners as more of an
art form than a science. And the implications, some say, are profound,
showing just how powerful computers can be at reasoning itself, at
mimicking the flashes of logical insight or even genius that have
characterized the best human minds.
Computers have found proofs of mathematical conjectures
before, of course, but those conjectures were easy to prove. The
difference this time is that the computer has solved a conjecture that
stumped some of the best mathematicians for 60 years. And it did so
with a program that was designed to reason, not to solve a specific
problem. In that sense, the program is very different from
chess-playing computer programs, for example, which are intended to
solve just one problem: the moves of a chess game.
"It's a sign of power, of reasoning power," said Dr. Larry
Wos, the supervisor of the computer reasoning project at Argonne. And
with this result, obtained by a colleague, Dr. William McCune, he
said, "We've taken a quantum leap forward."
Wos predicts that the result may mark the beginning of the end
for mathematics research as it is now practiced, eventually freeing
mathematicians to focus on discovering new conjectures, and leaving
the proof to computers. But the result also may challenge the very
notion of creative thinking, raising the possibility that computers
could take a parallel path to reach the same conclusions as great
human thinkers. Or it may be that since no one has any idea how humans
think, the magnificent bursts of creativity that spring apparently
full blown from the minds of geniuses are actually a result of hidden,
computer-like drudge work in the unconscious recesses of the brain.
Dr. Stanley Burris, a mathematician at the University of
Waterloo in Canada, said that the result was "the first sort of real
breakthrough in automated theorem proving," and that it did seem to be
different in kind from what went before. It shows, he said, that "it's
a very thin line between the mechanical and the creative and it may
disappear."
Dr. Robert Boyer, a computer scientist at the University of
Texas in Austin, hedged. "I think it's the most remarkable result in
automated theorem proving in 30 years," he said, and "clearly a form
of computer thinking." But, he added, "I don't want to make too much
of that." It's best, he said, to think of a computer as "just another
colleague, one that is sometimes helpful, but often not."
McCune's proof concerns a conjecture that is the very epitome
of pure mathematics. "It has no applications," McCune said. His
computer program proved that a set of three equations is equivalent to
a Boolean algebra, that set of rules, familiar to generations of high
school students, that govern unions and complements and intersections
among sets.
The problem was first posed in the 1930s by Dr. Herbert
Robbins, who is New Jersey Professor of Mathematics at Rutgers
University in New Brunswick. Robbins said that he worked on the
problem for some time, and then passed it on one of the century's most
famous logicians, Dr. Albert Tarski of Stanford University. Tarski,
who is now dead, worked on the problem, included it in a book, and
handed it out to graduate students and visitors.
Burris, for example said that Tarski suggested the problem to
him in the early 1970s, while he was visiting Stanford for a couple of
months. Tarski, he said, "liked to throw out challenging problems to
people passing through."
While mathematicians were batting around Robbins's problem,
computer scientists were striving to see if they could get computers
to reason. Among them was Wos, who started working on automated
reasoning in the 1960s. It was a time when computers were primitive,
clunky and slow, and researchers were divided on how to proceed. Some
believed the key was to figure out how humans reasoned and then to
create computer programs that mimicked the process. Wos disagreed.
"Nobody knows how humans reason," he said. "When you talk to
mathematicians and say, 'I understand you proved a great theorem. How
did you do it?' They'll say, 'Well, I walked around my house a lot and
I read some papers and I thought,' "
So he and his colleagues followed a different path. "We didn't
ask ourselves what people do when they think," Wos said. "That was
irrelevant." he said. Instead, he said: "We asked how can you tell a
computer what this problem is about? How can you get it to draw
conclusions that follow inevitably and logically from hypotheses and
thereby prove theorems?"
He and his colleagues began writing programs in which the
computer would assume that the hypothesis in question was false and
would then examine the consequences. If it found a contradiction, that
would be proof that the hypothesis was true. The computer would also
assume that the hypothesis was true and do the same thing, looking for
contradictions that would show it was false.
To prevent the computer from getting lost in checking out
lengthy chains of extended consequences, the investigators added
strategies like ignoring any logical statements that contained more
than 100 symbols.
"I got into strategy because I was a poker player and a bridge
player," Wos said. "A good con man uses strategy. He figures out what
your weakness is and plays to that weakness. He doesn't just randomly
try to trick you."
Wos's computer programs were soon able to find proofs for
basic mathematical problems. "We could do the problems sometimes
better than the students and sometimes, once in a great while, better
than the professors could," Wos said.
For more than 16 years, Wos and his colleagues stuck to
problems from mathematics textbooks. Wos explained that when the
computer tried to prove something for which a proof existed, and
failed, the investigators knew that "there's a problem with our
program." If they try to solve an unsolved problem, and fail, they
have no way of knowing whether they missed something obvious.
"My own mathematician friends would say, 'Wos, why are you
doing what we already know? Why don't you give us something new?' "
Wos said. In the early 1970s, he said, he told one of his badgering
friends that he thought it would be another 30, 40 or even 50 years
before computers could solve major problems that had stumped
mathematicians.
The first time they tried something new was in 1978, when Wos
said, "a little baby problem" came along. They solved it, and then
solved five others like it. Wos was ecstatic. "We had done more than I
thought I could do in my lifetime," he said.
The group kept adding strategies to its programs. It added one
recently that said to try things that worked in previous problems. Wos
said some of his colleagues scoffed at that and that he himself did
not know if it would work. But, he said, it turned out to be
surprisingly useful.
In 1979, Wos learned about Robbins's problem but, although he
and his colleagues tried to solve it from time to time with ever more
refined computer programs, they failed. McCune joined the group in
1984, with a new Ph.D. from Northwestern University and a thirst to
see how far he could push computers to go.
The fact that the Robbins conjecture was certifiably hard --
it had, after all, stumped some of the best minds in mathematics and
had gone unsolved for decades -- appealed to McCune. But the problem
also thwarted his best computer reasoning programs.
Finally, on Oct. 2, McCune gave the Robbins conjecture to a
new automated reasoning program that he had written called EQP, for
equational prover. Eight days later, on Oct. 10, the computer spewed
out a proof. McCune, a low-key researcher, said he was "amazed." Wos,
his exuberant supervisor, said, "Bill was in heaven."
Encouraged, McCune tried to get the computer to refine the
proof. He started his program searching for a better proof on
Nov. 15. It found one on Nov. 25.
McCune said he had checked his proof by computer and by
hand. Dr. Mark Stickel of the Stanford Research Institute in Palo
Alto, Calif., independently checked it with his computer program. And
Burris independently checked it by hand.
Burris said the proof, freshly out of the computer, was in the
form of statements so long that he had to print them sideways across
the page, and he had to specify that the printer use small type. "It
was pretty unreadable," he said. "The machine says, 'I got that step
from two steps before, but it doesn't fill in all the details" He
spent several days rewriting it and now has it down to three
normal-looking pages of mathematical statements.
The proof will be published within a few months in The Journal
of Automated Reasoning, Wos said. He will also publish it on the
Internet and will put out a call for more problems as certifiably hard
as Robbins's problem.
When he saw that he had actually proved Robbins's conjecture,
McCune called the 81-year-old mathematician at his office at Rutgers,
amazing him with the news.
In a telephone interview, Robbins said he was
delighted. "Isn't that marvelous," he said. "I'm glad I lived long
enough to see it."
But McCune certainly had a different experience than a
mathematician might have had if he or she had created the proof with
pure thought. So was something lost in the process? Did McCune have
anything resembling a Eureka moment?
Well, not quite, McCune said. "I have a good feeling," he
said. "In a sense, I have a feeling that the computer has been
creative."
Ben Sharvy
> Neil Rickert:
> > I would distinguish between bee behavior and general insect
> > behavior. The bee has a complex social system involving complex
> > interactions, and I think these are still beyond what you can do in
> > today's robots.
>
> I would say that the robots (few as they are today, and sub-bee
> in processing power) moving things around in our factories
> play as complex and vital a social role in their societies as
> bees do in theirs.
Robots don't have a society.
Ben Sharvy
The people quoted in this article are idiots.
In article <34A32B...@cmu.edu>, Hans Moravec <h...@cmu.edu> wrote:
>
> "It's a sign of power, of reasoning power," said Dr. Larry
> Wos, the supervisor of the computer reasoning project at Argonne.
Reasoning requires concepts, in particular, the conept of truth. There is
no evidence that computers today have concepts.
>
> "We didn't
> ask ourselves what people do when they think," Wos said. "That was
> irrelevant." he said. Instead, he said: "We asked how can you tell a
> computer what this problem is about? How can you get it to draw
> conclusions that follow inevitably and logically from hypotheses and
> thereby prove theorems?"
This bozo is a clod. You don't "tell" computers things, anymore than you
tell a coffee maker to make coffee. That is metaphorcial language, leading
to the metaphorical use of "reason" and "creativity" to describe the
results of the computer's functioning.
Arthur Th. Murray
When I was in (Latin and Greek) graduate school at U Cal Berkeley
in autumn of 1968, I read in the student newspaper that some AI
students from Stanford were going to come over and show a movie
of their work on a terrain-navigating robot. Since graduate
school in classics was only a provisional career path while my
far more intense life-ambition was to create AI, I made sure to
attend the robot movie screening on the appointed evening. Ever
since then I have wished that I had written down the names of the
four or five visitors from Stanford, because they were probably
persons whose careers I ended up tracking through the literature.
"Greeting. You are hereby ordered to report for induction..."
knocked me out of graduate school, but the US Army schools on
electronics and nuclear weapons (I became an ADM specialist:
"Atomic Demolition Munition" -- in the news a few days ago)
harmonized fine with my long-range AI career plan, because one
needed to understand computer circuit action in order to subsume
such circuitry and its logic beneath efforts to design AI freely.
Three things were missing before I could formulate and propose an
AI mind-model. Number one, Chomskyan linguistics, was dished up
in a linguistics course taught by Dafydd Gibbon in 1972 at the
Georgia-Augusta Universitaet in Goettingen, Germany. Number two,
the visual feature extraction of Hubel and Wiesel -- for lack of
which I wasted the years 1973 to 1976 while trying to reason out
vision on my own -- was explained in BYTE magazine in early 1978.
#3 http://www.scn.org/~mentifex/ a place on the Web, 28 aug 1997.
Hans Moravec
hpm:
> > I would say that the robots (few as they are today, and sub-bee
> > in processing power) moving things around in our factories
> > bees do in theirs.
bsharvy:
> Robots don't have a society.
Robots share in human society. Industrial robots earn
their living by being a vital part of it.
Neil Rickert
In <bsharvy-2512...@dynip157.efn.org> bsh...@NOSPAMefn.org (Ben Sharvy) writes:
>The people quoted in this article are idiots.
Or perhaps Ben Sharvy is the idiot.
The article to which you refer contained quotes by many people. You
chose to only quote that part of the article referring to Larry Wos.
However, in the context of the full article, the comments by Wos
appear to be carefully measured and not excessive. I'll grant that
some of the comments by other folk are a little silly. But do keep
in mind that a reporter, perhaps particularly interested in
hyperbole, chose what to include.
>In article <34A32B...@cmu.edu>, Hans Moravec <h...@cmu.edu> wrote:
>> "It's a sign of power, of reasoning power," said Dr. Larry
>> Wos, the supervisor of the computer reasoning project at Argonne.
>Reasoning requires concepts, in particular, the conept of truth. There is
>no evidence that computers today have concepts.
>> "We didn't
>> ask ourselves what people do when they think," Wos said. "That was
>> irrelevant." he said. Instead, he said: "We asked how can you tell a
>> computer what this problem is about? How can you get it to draw
>> conclusions that follow inevitably and logically from hypotheses and
>> thereby prove theorems?"
>This bozo is a clod. You don't "tell" computers things, anymore than you
JRStern
On Thu, 25 Dec 1997 03:59:27 -0500, Hans Moravec <h...@cmu.edu> wrote:
>The evidence so far is that one does not have to program in
>great complexity to achieve the basics of intelligence.
Agreed. Perhaps degree of intelligence similarly is _not_ necessarily
proportional to computational power. Perhaps.
[nice 600 line review of the history of autonomous vehicles snipped]
> Another generation of students is now building this work into
>"the highway of the future," among other applications.
And I think we will learn a great deal from it, and more from
implementing entire hives of autonomous (and people-carrying)
vehicles. The real-world interaction element that Neil Rickert likes
so much, will drive (sic) progress.
But what we've seen so far is increased engineering of old ideas,
which we knew worked, and work much better on faster hardware with
better software developed on equally better development environments.
I continue to doubt that increased hardware horsepower has produced
any qualitative differences from twenty years ago. Given its most
optimistic interpretation, what the entire historic trend shows is a
climb towards _minimal_ autonomous function. It is my speculative
position that there is a ceiling to the "intelligence" achievable
thereby, the ceiling quite uncoincidentally occuring at about the
levels of sensory and motor performance we see about us in nature,
including human levels.
The B2 and the bee (1?) are both unstable, both use about the same
computational level? Quite natural. When the product of bytes and
cycle times of the highly parallel human brain (and nervous system)
and an available (and highly/massively parallel) digital electronic
computer system are at about equal levels, if our software is up to
the job, we should see equal levels of "intelligence", whatever it is.
Philosophical differences aside, I think we agree on that.
In one way, I continue to espouse the old AI line, that something
fully equivalent to human intelligence in all ways can be realized on
single-processor digital computer systems. Sensory processing aside
(and I, unlike Neil Rickert, believe the intelligence question can be,
perhaps must be, handled on asynchronous, linguistic, symbolic
systems), I further assert that a 100mips workstation probably has the
horsepower today (that is, the ability to do it in real time, which I
do not think is strictly necessary). Maybe the 1mips VAX/780 had it.
I think it's a matter of paradigm, more than horsepower, that keeps us
from animal and even human levels of intelligence. It wouldn't shock
me if, in fifty years, undergrads are taught how to build truly
intelligent systems from a blank page, using 3GLs, in a matter of
hours or days, the way we now teach them how to write EMycin systems.
Make that seventy years: 10 to shift paradigm, 30 for one academic
generation that teaches it, 30 more for a second academic generation
that refines it to the undergrad level.
As for superintelligence, I'm comfortable today asserting its likely
physical and maybe even conceptual unrealizability.
Joshua Stern
JRS...@gte.net
JRStern
On Thu, 25 Dec 1997 23:00:47 -0500, Hans Moravec <h...@cmu.edu> wrote:
hpm:
> Theorem proving was a joke before 1990, but in 1996
>a theorem-proving program called EQP running five weeks on a 50
>MIPS computer at Argonne National Laboratory found a neat proof of
>a boolean algebra conjecture by Herbert Robbins that had eluded
>mathematicians for sixty years.
JRStern:
> I do not know enough about this case to know if there's anything more
> significant about this, than about the compilation of tables of
> logarithms by computers in the 1950s.
hpm:
>The author's original description is found in
>http://www.mcs.anl.gov/home/mccune/ar/robbins/
>
>Here is an article excerpt from last year's New York Times
>with a good summary. The whole article is available at
>http://www.nytimes.com/library/cyber/week/1210math.html
[snip]
> And the implications, some say, are profound,
>showing just how powerful computers can be at reasoning itself, at
>mimicking the flashes of logical insight or even genius that have
>characterized the best human minds.
[snip]
> "I have a good feeling," he
>said. "In a sense, I have a feeling that the computer has been
>creative."
Fooey. Maybe I'm just a curmudgeon, but the only conclusion I think
is justified is yet another insight into how poor human intelligence
is at dealing with large, enumerable domains, from arithmetic to
chess. The text at the Robbins url describes what was done: give the
prover a list of axioms, a starting point, and an end point. The
putative "creativity" the computer exhibited was connecting two
points. Whatever the value of the path found, as far as AI goes, I
see nothing new. I remain underwhelmed.
A valid and somewhat opposing position is the one suggested by this
thread's title, that automation within even such closed domains, plus
the increased computer power now available, equals at least some kind
of quantitative progress. I would not deny that, but insist on
differentiating it from qualitative progress.
After all, a couple of decades ago, "theorem proving" meant
unification, or simple boolean evaluation of axiom sets, or even
simple computational evaluation of arbitrary logics. In that sense,
it both was and was not "a joke before 1990". From the moment I heard
the proposition asserted in my Intro to AI course, that theorem
proving was entirely equivalent to "intelligence", I thought the
statement ill-formed. Twenty-odd years later, I'm certain of it,
though the full statement of my disagreement is going to be a hundred
pages or more long, with another hundred pages to state an
alternative. On the other hand, "theorem proving" in the fundamental
sense is essential to whatever computation does constitute a valid
(identical) equivalent to human intelligence.
Maybe we'll have to start viewing mathematics differently as a result
of the Robbins work, but it provides nothing new to the AI debate.
Joshua Stern
JRS...@gte.net
Louis Savain
>The evidence so far is that one does not have to program in
>great complexity to achieve the basics of intelligence. What
>is needed is enough processing power the "right" combination of
>simple techniques. With the right framework, automatic learning
>can acquire the complicated parts.
It's always a pleasure to see someone express this minority view.
If this attitude had been widespread in the AI community from the
start, the problem of creating a truly adaptive common-sense robot,
*might* have been licked by now. Problem is, there are too many
prominent AI researchers who don't seem to share your conviction.
There is something about human beings and/or human societies that
causes us to follow the ill-conceived advice: "If at first you don't
succeed, try a more complex solution." Even in the connectionist
arena simplicity seems to have fallen out of favor. It's depressing.
IMHO, proper intelligence research is not a search for complex
solutions but rather, a quest for simplicity. AI should be about
discovering the one simple little idea that will not only make
everything easy but also bring the search to a close. Many will
disagree, of course.
Regards,
Louis Savain
President, Marengo Media, Inc.
Ben Sharvy
> In <bsharvy-2512...@dynip157.efn.org> bsh...@NOSPAMefn.org
(Ben Sharvy) writes:
>
> >The people quoted in this article are idiots.
>
> Or perhaps Ben Sharvy is the idiot.
>
> The article to which you refer contained quotes by many people. You
> chose to only quote that part of the article referring to Larry Wos.
> However, in the context of the full article, the comments by Wos
> appear to be carefully measured and not excessive. I'll grant that
> some of the comments by other folk are a little silly.
Ben Sharvy is almost certainly an idiot, but he is most certainly less of
an idiot than the people quoted in the article.
I snipped many of the quotes because of redundancy. The snipping didn't
distort context, because the snippees were identical with each other:
blank assertions that the computing-machine was reasoning, creating, and
understanding.
Ben Sharvy
Robots do not earn a living, except in metaphorical sense of "earn" and
"living". Next we'll hear that pencils earn a living, and share in
society.
Neil Rickert
In <680vc6$gpa$1...@gte1.gte.net> JRS...@gte.net (JRStern) writes:
>On Thu, 25 Dec 1997 03:59:27 -0500, Hans Moravec <h...@cmu.edu> wrote:
>>The evidence so far is that one does not have to program in
>>great complexity to achieve the basics of intelligence.
>Agreed. Perhaps degree of intelligence similarly is _not_ necessarily
>proportional to computational power. Perhaps.
Yes, I agree with that.
>In one way, I continue to espouse the old AI line, that something
>fully equivalent to human intelligence in all ways can be realized on
>single-processor digital computer systems. Sensory processing aside
>(and I, unlike Neil Rickert, believe the intelligence question can be,
>perhaps must be, handled on asynchronous, linguistic, symbolic
>systems), I further assert that a 100mips workstation probably has the
>horsepower today (that is, the ability to do it in real time, which I
>do not think is strictly necessary).
Let me clarify my position here.
Most of what we refer to as intelligent has to do with the way we
deal with real world problems. If you want to restrict yourself to
the consideration of relatively abstract problems (such as might be
solved by a bed ridden person with access only to an internet
connection for news of the world), then in principle it could be done
on a sufficiently powerful desktop system. But the problem is that a
tremendous amount of detail would have to be programmed in, and there
is no reasonable prospect of getting that right.
The reason I insist on the sensory and motor abilities, is that they
are needed for a learning system that can build its own knowledge
base without it having to be programmed in in intricate detail.
> It wouldn't shock
>me if, in fifty years, undergrads are taught how to build truly
>intelligent systems from a blank page, using 3GLs, in a matter of
>hours or days, the way we now teach them how to write EMycin systems.
I'm skeptical.
>Make that seventy years: 10 to shift paradigm, 30 for one academic
>generation that teaches it, 30 more for a second academic generation
>that refines it to the undergrad level.
You are a hopeless optimist. The philosophers have been stuck on a
failed paradigm since the time of Plato. They show no signs of
budging. They cling to their failed paradigm with such tenacity that
it is unthinkable that it could be shaken in only 10 years. And the
influence of this philosophy is felt throughout our education system
and culture, affecting even those who ridicule philosophy.
ray scanlon
Neil Rickert wrote in message <67tvlb$3...@ux.cs.niu.edu>...
>In <34A220...@cmu.edu> Hans Moravec <h...@cmu.edu> writes:
>
>>Leaving sensory/motor processing out of the intelligence picture
>>is unreasonable.
'Processing' is misleading. It leads to artificial intelligence and all
that.
There is input and it comes to be output by what is better called
'filtration'.
>Yes, I agree. Yet much of the discussion omits it. Even much of the
>NN work is done with simulators which do not need to concern
>themselves with sensory data or motor output.
Input is input and output is output.
>Bees are short lived, so that this attrition is acceptable. The
>evolutionary strategy appears to be to create relatively low cost
>units (bees) with short life spans, and to use the processes of
>evolution to keep the bees well adapted to their environment. For
>mammals, the strategy is different. The units are considerably more
>expensive, and have substantially longer lives. With a long life
>span they cannot depend on evolution as the main tool for adapting to
>their environment. Instead they require a system of learning to
>allow self adaptation.
Yes, but mammals (and especially man) are not particularly successful. Among
multi-celled animals, I believe the palm goes to the sea cucumber for
bio-mass.
>It is this ability at self adaptation which is missing from AI. Some
>of the research effort on this self adaptation seems to be in the
>form of specious proofs that the ability could not possibly exist
>(thus justifying the failure of the research community to solve the
>problem).
There is more than one research community out there. I think the molecular
neurobiologists are doing rather well on adaptation. See chap. 17 'Behavior
and Plasticity' and chap. 18 'Learning and Memory' in The Neuron, Levitan
and Kaczmarek, 1997.
ray
For an opinion on how the brain thinks, go to
http://www.wsg.net/~rscanlon/brain.html
Hans Moravec
JRStern wrote (re history of autonomous driving):
>>
> But what we've seen so far is increased engineering of old ideas,
> which we knew worked, and work much better on faster hardware with
> better software developed on equally better development environments.
> I continue to doubt that increased hardware horsepower has produced
> any qualitative differences from twenty years ago.
Talk about revisionism! If you knew which techniques worked
20 years ago, you should have told us. We sure didn't know,
and the inability of our robots to cross the road, or a room,
by their own vision was convincing evidence to most observers
that we were on the wrong track, and should give up, at least
until we could figure out something completely different to do (*).
Not unlike your current prescription, come to think of it.
But it's just not so. We CAN recapitulate biological evolution
by making a long, patient series of modest discoveries in systems
of comparable scale. That's how it was done in the first place.
We don't need great leaps, or new insights, though we will probably
gain interesting hindsights from the effort.
(*) When Chuck Thorpe first child and his management of the
Navlab project arrived at about the same time in 1985, he
was willing to bet that his child would be driving before his
robots. No one else had ever lost such a bet. But now there
are a dozen autonomous vehicles, and his child still has five
years to go.
Hans Moravec
Ben Sharvy wrote:
> Robots do not earn a living, except in metaphorical sense of "earn"
> and "living". Next we'll hear that pencils earn a living, and share
> in society.
A pencil and a robot as much as a horse, a slave, a serf or
a factory worker can earn a living in society by returning more
value to society than they consume. In many jobs, robots have
directly replaced human workers in particular jobs. It is sheer
chauvinism to claim that the human was earning a living, but
that the machine doing the same work isn't.
Ben Sharvy
> Ben Sharvy wrote:
>
> > Robots do not earn a living, except in metaphorical sense of "earn"
> > and "living". Next we'll hear that pencils earn a living, and share
> > in society.
> It is sheer
> chauvinism to claim that the human was earning a living, but
> that the machine doing the same work isn't.
No, it is sheer awareness of what the word "earn" means. In the context of
earning a living, not to mention accusations of chauvinism, it has a moral
connotation: what you earn is what you deserve. Robots, pencils do not
deserve things. They do not "share" in society, they are used by society.
They are useful tools.
Most of this thread has degraded into a discussion of what words mean,
with the participants' awareness. Use words in standard, non-metaphorical,
dictionary senses for a discussion that goes somewhere, please.
SCN User
In a previous article, xn...@dial.pipex.com ("Nicholas Bostrom") says:
>I have written a paper outlining the case for believing that there will be
>superhuman artificial intelligence within the first third of the next
>century. Both the software problem and the hardware promlem are discussed. I
>would be interested to get some feedback that I can use in preparing the
>final version of the paper
The aforementioned paper discusses mainly aspects of the rate at
which hardware chips are doubling in capacity -- an issue perhaps
of interest to economists, but only a side issue to AI enthusiasts
eager to see ANY level of intelligence instantiated in a machine,
because immediately the low-level AI will first saturate the globe,
second hypertrophy machine IQ, third produce superintelligence.
>
>http://www.hedweb.com/nickb/superintelligence.htm
>
For the sake of the indicated paper and in recognition of the
general high-quality of the web site called "The Bostrom Space",
a high-traffic link to it has been encoded within the Index at
http://www.scn.org/~mentifex/ Mentifex Maker of Minds.
>Nick Bostrom
>n.bo...@lse.ac.uk
>Dept. Philosophy, Logic and Scientific method
>London School of Economics
>
Hans Moravec
Ben Sharvy wrote:
> In the context of earning a living, not to mention accusations of
> chauvinism, it has a moral connotation: what you earn is what you
> deserve. Robots, pencils do not deserve things. They do not "share"
> in society, they are used by society. They are useful tools.
Human beings "deserve" things because they, or others that
represent them, are able to defend their access to those things.
At one time people individually or in small groups often had to
fight for access to the essentials of life. Today what we
"deserve" is defined mostly what we are worth to others in social
relationships like jobs, and the legal and police systems do most
of our fighting.
A robot working in a factory has a relationship with its owner
and with workers in the factory. These relationships define what
the robot is worth economically, i.e. what it "deserves" to earn.
A successful robot will deserve at least enough to keep it
powered and maintained. Legally, the robot is a posession.
Legally, slaves were posessions also, yet their relationship
with their plantation and the rest of society was none the less
a social one.
This discussion started with a comment that while robots may
be approaching the intelligence of bees, they lacked the
social organization of bees. I noted that factory robots
played bee-like roles in social organizations (factories)
that were at least as complex as beehives.
You retorted that robots don't have societies. You didn't
say anything about bees. Well, if bees have societies, and
earn their living in them, then so do today's factory robots.
Your comments denying a social place for robots, or that
they earn a living, inspired my ire because I found them
absurdly pedantic and uselessly narrow. Especially in the
context of robots, as I see them in coming decades, playing
increasingly complex, responsible and humanlike roles.
Ben Sharvy
> Ben Sharvy wrote:
> > In the context of earning a living, not to mention accusations of
> > chauvinism, it has a moral connotation: what you earn is what you
> > deserve. Robots, pencils do not deserve things. They do not "share"
> > in society, they are used by society. They are useful tools.
>
> Human beings "deserve" things because they, or others that
> represent them, are able to defend their access to those things.
No, that is simply not what "deserve" means. It is possible to be "able to
defend [your] access" to things you do not deserve.
> Today what we
> "deserve" is defined mostly what we are worth to others in social
> relationships like jobs, and the legal and police systems do most
> of our fighting.
No, what we are worth to others is simply not the definition of the word
"deserve." You might try bringing the concept of rights into your
thinking.
> Your comments denying a social place for robots, or that
> they earn a living, inspired my ire because I found them
> absurdly pedantic and uselessly narrow. Especially in the
> context of robots, as I see them in coming decades, playing
> increasingly complex, responsible and humanlike roles.
I said robots have the same social role as pencils. They are tools people
find useful, and so there is a sense in which they might "deserve"
something. That doesn't make them a member of any society. Most people
believe we (not robots) deserve certain things regardless of whether
others can use us.
My objection is mostly to rhetorical and metaphorical langauge. It is
reasonable in some sense to talk about tools "earning" and "deserving" and
having societies. It is also reasonable to talk about "telling computers"
things, giving them instructions; it is reasonable, in a sense, to say
that Deep Blue "decided" to sacrifice a pawn, and so on. Likewise, it is
reasonable to use "quantum" to mean dramatic, to use "society" to mean
majority ("society thinks..."), and so on. But in a rigorous discussion
within a field that uses these technical terms technically, it is not
reasonable, and "experts" who slip into such slop-speak should be called
on it. If a discssion of AI is begun with propostions like we "tell"
computers X, and robots "earn a living", the discussion is pre-configured
to be circular and favor a certain conclusion.
Hans Moravec
hpm:
> Human beings "deserve" things because they, or others that
> represent them, are able to defend their access to those things.
> "deserve" is defined mostly what we are worth to others in social
> relationships like jobs, and the legal and police systems do most
> of our fighting.
bsharvy:
> No, that is simply not what "deserve" means. It is possible to be
> "able to defend [your] access" to things you do not deserve.
> No, what we are worth to others is simply not the definition of
> the word "deserve." You might try bringing the concept of rights
> into your thinking.
The concept of rights has meaning only inside particular
political frameworks. Political contexts are cultural
inventions that vary with time and place, that people
defend, sometimes by force, sometimes by argument.
People "deserve" what political systems grant them, up to
physical limits, as long as they successfully defend the
systems. If you claim they don't deserve them, then you
are arguing for a different political framework. Welcome
to the battle.
Since I don't think the humanistic political framework of
the last few hundred years makes much sense in the coming
period of genetically and artificially engineered
intelligences, you'll excuse me for ignoring it, and its
arbitrary list of rights. I want to think about the issue
from the ground up. At this point I care about what works,
not what you happen to think is right.
PKo
On Fri, 26 Dec 1997 11:35:33 -0800, bsh...@NOSPAMefn.org (Ben Sharvy)
wrote:
>> bsharvy:
>> > Robots don't have a society.
>>
>> Robots share in human society. Industrial robots earn
>> their living by being a vital part of it.
>
>Robots do not earn a living, except in metaphorical sense of "earn" and
>"living".
How does an industrial robot essentially differ from a bee, or an ant,
considering its tasks and intelligence?
What I am pointing at is obviously that if we say a bee belongs
to a society and earns its living, and robots do not, there should be
an essential discrimating factor between the two.
JRStern
On 26 Dec 1997 17:39:26 -0600, ric...@cs.niu.edu (Neil Rickert)
wrote:
>> It wouldn't shock
>>me if, in fifty years, undergrads are taught how to build truly
>>intelligent systems from a blank page, using 3GLs, in a matter of
>>hours or days, the way we now teach them how to write EMycin systems.
>
>I'm skeptical.
I'm hopeful.
>>Make that seventy years: 10 to shift paradigm, 30 for one academic
>>generation that teaches it, 30 more for a second academic generation
>>that refines it to the undergrad level.
>
>You are a hopeless optimist.
It's a burden.
> The philosophers have been stuck on a
>failed paradigm since the time of Plato. They show no signs of
>budging. They cling to their failed paradigm with such tenacity that
>it is unthinkable that it could be shaken in only 10 years. And the
>influence of this philosophy is felt throughout our education system
>and culture, affecting even those who ridicule philosophy.
Quite agree. Many of the failures of GOFAI were already known to
philosophy, others are a result of lingering effects of failed
paradigms -- we got the worst of it two ways! Astounding.
On the other hand, we do have reason for optimism. We have something
that wasn't available for 2500 years, we have computers. The
existence of hundreds of millions of cheap, fast machines gives an
empirical base for real progress. I'm reading Steven Pinker's new
book, "How the Mind Works", a pretentious title, especially when he
disclaims consciousness as opposed to intelligence. Still, he
probably didn't choose the title (I hope). And still, it's the best
book about AI I've ever read. He's all for the computational model of
intelligence (not consciousness, necessarily), by the way.
And, we even have a new and very powerful social and educational
source that provides a new paradigm. We have the internet.
Reasons for optimism, though far from a done deal.
Joshua Stern
JRS...@gte.net
JRStern
On Thu, 25 Dec 1997 23:44:09 -0800, bsh...@NOSPAMefn.org (Ben Sharvy)
wrote:
>> "We didn't
>> ask ourselves what people do when they think," Wos said. "That was
>> irrelevant." he said. Instead, he said: "We asked how can you tell a
>> computer what this problem is about? How can you get it to draw
>> conclusions that follow inevitably and logically from hypotheses and
>> thereby prove theorems?"
>
>to the metaphorical use of "reason" and "creativity" to describe the
>results of the computer's functioning.
Agreed.
Joshua Stern
JRS...@gte.net
JRStern
On Fri, 26 Dec 1997 21:36:54 -0500, Hans Moravec <h...@cs.cmu.edu>
wrote:
>JRStern wrote (re history of autonomous driving):
>>>
>> But what we've seen so far is increased engineering of old ideas,
>> which we knew worked, and work much better on faster hardware with
>> better software developed on equally better development environments.
>> I continue to doubt that increased hardware horsepower has produced
>> any qualitative differences from twenty years ago.
>
>Talk about revisionism! If you knew which techniques worked
>20 years ago, you should have told us.
Hey, allow me my rhetorical cheats. What I said, was that the methods
that succeed today, were already in use twenty years ago, not that
*everything* tried twenty years ago worked with enough horsepower.
>But it's just not so. We CAN recapitulate biological evolution
>by making a long, patient series of modest discoveries in systems
>of comparable scale. That's how it was done in the first place.
>We don't need great leaps, or new insights, though we will probably
>gain interesting hindsights from the effort.
Well, in the general case, that's building airplanes that flap their
wings, but in the specific domain of sensory processing, maybe.
Joshua Stern
JRS...@gte.net
Neil Rickert
In <688fu4$lq$1...@gte2.gte.net> JRS...@gte.net (JRStern) writes:
>On 26 Dec 1997 17:39:26 -0600, ric...@cs.niu.edu (Neil Rickert)
>wrote:
>> The philosophers have been stuck on a
>>failed paradigm since the time of Plato. They show no signs of
>>budging. They cling to their failed paradigm with such tenacity that
>>it is unthinkable that it could be shaken in only 10 years. And the
>>influence of this philosophy is felt throughout our education system
>>and culture, affecting even those who ridicule philosophy.
>Quite agree. Many of the failures of GOFAI were already known to
>philosophy, others are a result of lingering effects of failed
>paradigms -- we got the worst of it two ways! Astounding.
>On the other hand, we do have reason for optimism. We have something
>that wasn't available for 2500 years, we have computers.
But, as you yourself have admitted, there has been very little
progress in AI during the several decades in which we have had
computers. One of the problems is that the availability of computers
tends to encourage the basically solipsistic approach of emphasizing
internal structure (algorithms, etc), and ignoring the external
relation between the individual and the world. Moreover, computer
science curricula tend to overemphasize formalistic approaches such
as the theory of computation, automatic construction of programs,
automated proofs of correctness, while underemphasizing the important
abilities that the programmer needs to deal with practical real-world
problems.
> I'm reading Steven Pinker's new
>book, "How the Mind Works", a pretentious title, especially when he
>disclaims consciousness as opposed to intelligence. Still, he
>probably didn't choose the title (I hope). And still, it's the best
>book about AI I've ever read. He's all for the computational model of
>intelligence (not consciousness, necessarily), by the way.
By contrast, my take is that Pinker is on the wrong track.
>And, we even have a new and very powerful social and educational
>source that provides a new paradigm. We have the internet.
I'm not sure how the anarchy of the internet is supposed to provide a
useful paradigm for the organization of intelligence. Next you will
be pointing to Windows 95 as an implementation of superintelligence.
Louis Savain
In article <6891in$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil
Rickert) wrote:
>[...]
>But, as you yourself have admitted, there has been very little
>progress in AI during the several decades in which we have had
>computers. One of the problems is that the availability of computers
>tends to encourage the basically solipsistic approach of emphasizing
>internal structure (algorithms, etc), and ignoring the external
>relation between the individual and the world.
I personally don't see what is wrong with so-called "solipsistic"
approaches. A solipsist approach means that the system decides what
to do with the incoming sensory stream and how to organize it into
useful structures, not the programmer.
> Moreover, computer
>science curricula tend to overemphasize formalistic approaches such
>as the theory of computation, automatic construction of programs,
>automated proofs of correctness, while underemphasizing the important
>abilities that the programmer needs to deal with practical real-world
>problems.
This is precisely what is wrong with most of AI research. I call it
the fallacy of knowledge engineering. Most AI programmers think
*they* can deal with the difficulties of the real world. This is pure
folly. AI researchers should take the exact opposite stance: They
should try to create emergent sensory-motor systems and leave it to
*them* "to learn deal with practical real-world problems." Anything
else is either pretentious self-delusion or the sign of some unspoken
agenda.
JRStern
On 29 Dec 1997 14:35:35 -0600, ric...@cs.niu.edu (Neil Rickert)
wrote:
>But, as you yourself have admitted, there has been very little
>progress in AI during the several decades in which we have had
>computers.
Yes, my rationalization for that is the computer age really began when
100mhz processors got cheap, say, 1993, coincidentally about the time
this new upsurge in interest regarding consciousness started up.
> One of the problems is that the availability of computers
>tends to encourage the basically solipsistic approach of emphasizing
>internal structure (algorithms, etc), and ignoring the external
>relation between the individual and the world. Moreover, computer
>science curricula tend to overemphasize formalistic approaches such
>as the theory of computation, automatic construction of programs,
>automated proofs of correctness, while underemphasizing the important
>abilities that the programmer needs to deal with practical real-world
>problems.
Agree again. It's your "bad paradigm", preserved in the amber of
slow-changing academic focus. If I'm right, kids raised on the
Internet (er, you know what I mean <g>) will not all sit quietly for
solipcistic crap, if a reasonable alternative is made available. Ten
year horizon.
>
>> I'm reading Steven Pinker's new
>>book, "How the Mind Works", a pretentious title, especially when he
>>disclaims consciousness as opposed to intelligence. Still, he
>>probably didn't choose the title (I hope). And still, it's the best
>>book about AI I've ever read. He's all for the computational model of
>>intelligence (not consciousness, necessarily), by the way.
>
>By contrast, my take is that Pinker is on the wrong track.
Well, he does take a swipe at interactionist theories en passant,
though the crux of his whole argument is actually interactionist. I
think he must hang out with a bad crowd. BTW, have you actually read
it?
>>And, we even have a new and very powerful social and educational
>>source that provides a new paradigm. We have the internet.
>
>I'm not sure how the anarchy of the internet is supposed to provide a
>useful paradigm for the organization of intelligence. Next you will
>be pointing to Windows 95 as an implementation of superintelligence.
I said it provides a paradigm, not instantiates anything. It's an
interactive paradigm within the linguistic domain I consider valid for
AI. Using it teaches people the limits of the solipcistic approach,
and the limits of existing approaches generally. Again, it's a useful
reference point in both positive and negative ways, and every little
bit helps.
Joshua Stern
JRS...@gte.net
Neil Rickert
In <34a91b73...@news.earthlink.net> sav...@earthlink.net (Louis Savain) writes:
>In article <6891in$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil
>Rickert) wrote:
>>[...]
>>But, as you yourself have admitted, there has been very little
>>progress in AI during the several decades in which we have had
>>computers. One of the problems is that the availability of computers
>>tends to encourage the basically solipsistic approach of emphasizing
>>internal structure (algorithms, etc), and ignoring the external
>>relation between the individual and the world.
> I personally don't see what is wrong with so-called "solipsistic"
>approaches. A solipsist approach means that the system decides what
>to do with the incoming sensory stream and how to organize it into
>useful structures, not the programmer.
I'm afraid that you missed my point. My objection is to the emphasis
on what to do with internal representations. Instead I want to
emphasize acting on the incoming sensory stream and forming
representations. My chief objection to your approach, and to
Modlin's approach, is that, if I understand what you have posted, you
both want to deal with representations, rather than with
pre-represented raw input.
Neil Rickert
In <689ba3$6v1$1...@gte1.gte.net> JRS...@gte.net (JRStern) writes:
>On 29 Dec 1997 14:35:35 -0600, ric...@cs.niu.edu (Neil Rickert)
>wrote:
>>> I'm reading Steven Pinker's new
>>>book, "How the Mind Works", a pretentious title, especially when he
>>>disclaims consciousness as opposed to intelligence. Still, he
>>>probably didn't choose the title (I hope). And still, it's the best
>>>book about AI I've ever read. He's all for the computational model of
>>>intelligence (not consciousness, necessarily), by the way.
>>By contrast, my take is that Pinker is on the wrong track.
>Well, he does take a swipe at interactionist theories en passant,
>though the crux of his whole argument is actually interactionist. I
>think he must hang out with a bad crowd.
It is my impression that he is sympathetic to the positions of Fodor
and Chomsky.
> BTW, have you actually read
>it?
Not yet. I have read his "The Language Instinct", which is supposed
to be a rather better book. I have also heard him discuss "How the
Mind Works" on a radio show.
Seth Russell
Neil Rickert wrote:
> I'm not sure how the anarchy of the internet is supposed to provide a
> useful paradigm for the organization of intelligence.
The size and anarchy of the Internet provides the first *natural
computer environment*. Anarchy is one of the qualities of the physical
environment of our biology in which natural intelligence evolved. IMHO
when AI researchers discover how to adapt their agents for survival in
this new natural environment (the Internet), then Robust AI will
emerge. In other words don't program intelligence, program to the
environment -- trade with the environment -- plenty of info food out
there -- plenty of humans want it served to them.
Only problem is that all the AI programmers are hot-shots ... they got
to do it all themselves ... discover the secret of learning ... alass
me thinks it won't happen that way.
Seth
See "Bozo's Conjecture" at http://www.clickshop.com/ai/conjecture.htm
And then on to the AI Jump List ...
Neil Rickert
In <883444...@longley.demon.co.uk> Da...@longley.demon.co.uk (David Longley) writes:
>Two things to consider:
> 1) Man has been around quite some time, science has not.
> Unless our knowledge of natural selection is *way* of base
> it is highly unlikely that there is much "in our heads"
> which distinguishes us from our remote ancestors.
Agreed.
> 2) From the above, it must be something other than our
> natural abilities which accounts for our 'remarkable'
> intelligence and which should be looked to for all future
> developments of our intelligence.
Sorry, but (2) does not obviously follow from (1). At various times
dolphins and chimpanzees have been said to have remarkable
intelligence. And even if (2) is correct, the "something other ..."
might merely be the fact that we are a social animal. Keep in mind
that social bees and ants are quite remarkable compared with
non-social insects.
David Longley
In article <34a91b73...@news.earthlink.net>
sav...@earthlink.net "Louis Savain" writes:
> In article <6891in$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil
> Rickert) wrote:
>
> >[...]
> >But, as you yourself have admitted, there has been very little
> >progress in AI during the several decades in which we have had
> >computers. One of the problems is that the availability of computers
> >tends to encourage the basically solipsistic approach of emphasizing
> >internal structure (algorithms, etc), and ignoring the external
> >relation between the individual and the world.
>
> I personally don't see what is wrong with so-called "solipsistic"
> approaches. A solipsist approach means that the system decides what
> to do with the incoming sensory stream and how to organize it into
> useful structures, not the programmer.
>
> > Moreover, computer
> >science curricula tend to overemphasize formalistic approaches such
> >as the theory of computation, automatic construction of programs,
> >automated proofs of correctness, while underemphasizing the important
> >abilities that the programmer needs to deal with practical real-world
> >problems.
>
> This is precisely what is wrong with most of AI research. I call it
> the fallacy of knowledge engineering. Most AI programmers think
> *they* can deal with the difficulties of the real world. This is pure
> folly. AI researchers should take the exact opposite stance: They
> should try to create emergent sensory-motor systems and leave it to
> *them* "to learn deal with practical real-world problems." Anything
> else is either pretentious self-delusion or the sign of some unspoken
> agenda.
>
The above is just naive.
In 1980 Fodor
['Explored] the distinction between 2 doctrines, both of
which inform theory construction in much of modern
cognitive psychology: the representational theory of
mind and the computational theory of mind. According to
the former, propositional attitudes are viewed as
relations that organisms bear to mental representations.
According to the latter, mental processes have access
only to formal (nonsemantic) properties of the mental
representations over which they are defined. The
following claims are defended: (1) The traditional
dispute between rational and naturalistic psychology is
plausibly viewed as an argument about the status of the
computational theory of mind. (2) To accept the
formality condition is to endorse a version of
methodological solipsism. (3) The acceptance of some
such condition is warranted, at least for that part of
psychology that concerns itself with theories of the
mental causation of behavior. A glossary and several
commentaries are included.'
J A Fodor (1980)
Methodological solipsism considered as a research
strategy in cognitive psychology.
Massachusetts Inst of Technology
Behavioral and Brain Sciences; 1980 Mar Vol 3(1) 63-109
Some of the commentaries, particularly those by Loar or Rey clarify
what is, admittedly, quite a difficult, but substantial view widely
held by graduate psychologists.
'If psychological explanation is a matter of describing
computational processes, then the references of our
thoughts do not matter to psychological explanation.
This is Fodor's main argument.....Notice that Fodor's
argument can be taken a step further. For not only are
the references of our thoughts not mentioned in
cognitive psychology; nothing that DETERMINES their
references, like Fregian senses, is mentioned
either....Neither reference nor reference-determining
sense have a place in the description of computational
processes.'
B. F. Loar
Ibid p.89
Not all of the commentaries were as formal, as the following
commentary from one of the UK's most eminent logicians makes clear:
'Fodor thinks that when we explain behaviour by mental
causes, these causes would be given "opaque"
descriptions "true in virtue of the way the agent
represents the objects of his wants (intentions,
beliefs, etc.) to HIMSELF" (his emphasis). But what an
agent intends may be widely different from the way he
represents the object of his intention to himself. A man
cannot shuck off the responsibility for killing another
man by just 'directing his intention' at the firing of a
gun:
"I press a trigger - Well, I'm blessed!
he's hit my bullet with his chest!"'
P. Geach
ibid p80
Here's how Fodor contrasted Methodological Solipsism with the
naturalistic approach:
'..there's a tradition which argues that - epistemology
to one side - it is at best a strategic mistake to
attempt to develop a psychology which individuates
mental states without reference to their environmental
causes and effects...I have in mind the tradition which
includes the American Naturalists (notably Pierce and
Dewey), all the learning theorists, and such
contemporary representatives as Quine in philosophy and
Gibson in psychology. The recurrent theme here is that
psychology is a branch of biology, hence that one must
view the organism as embedded in a physical environment.
The psychologist's job is to trace those
organism/environment interactions which constitute its
behavior.'
J. Fodor (1980) ibid. p.64
Here is how Stich (1991) reviewed Fodor's position ten years on:
'This argument was part of a larger project. Influenced
by Quine, I have long been suspicious about the
integrity and scientific utility of the commonsense
notions of meaning and intentional content. This is not,
of course, to deny that the intentional idioms of
ordinary discourse have their uses, nor that the uses
are important. But, like Quine, I view ordinary
intentional locutions as projective, context sensitive,
observer relative, and essentially dramatic. They are
not the sorts of locutions we should welcome in serious
scientific discourse. For those who share this Quinean
scepticism, the sudden flourishing of cognitive
psychology in the 1970s posed something of a problem. On
the account offered by Fodor and other observers, the
cognitive psychology of that period was exploiting both
the ontology and the explanatory strategy of commonsense
psychology. It proposed to explain cognition and certain
aspects of behavior by positing beliefs, desires, and
other psychological states with intentional content, and
by couching generalisations about the interactions among
those states in terms of their intentional content. If
this was right, then those of us who would banish talk
of content in scientific settings would be throwing out
the cognitive psychological baby with the intentional
bath water. On my view, however, this account of
cognitive psychology was seriously mistaken. The
cognitive psychology of the 1970s and early 1980s was
not positing contentful intentional states, nor was it
(adverting) to content in its generalisations. Rather, I
maintained, the cognitive psychology of the day was
"really a kind of logical syntax (only psychologized).
Moreover, it seemed to me that there were good reasons
why cognitive psychology not only did not but SHOULD not
traffic in intentional states. One of these reasons was
provided by the Autonomy argument.'
Stephen P. Stich (1991)
Narrow Content meets Fat Syntax
in MEANING IN MIND - Fodor And His Critics
and writing with others in 1991, even more dramatically:
'In the psychological literature there is no dearth of
models for human belief or memory that follow the lead
of commonsense psychology in supposing that
propositional modularity is true. Indeed, until the
emergence of connectionism, just about all psychological
models of propositional memory, except those urged by
behaviorists, were comfortably compatible with
propositional modularity. Typically, these models view a
subject's store of beliefs or memories as an
interconnected collection of functionally discrete,
semantically interpretable states that interact in
systematic ways. Some of these models represent
individual beliefs as sentence like structures - strings
of symbols that can be individually activated by their
transfer from long-term memory to the more limited
memory of a central processing unit. Other models
represent beliefs as a network of labelled nodes and
labelled links through which patterns of activation may
spread. Still other models represent beliefs as sets of
production rules. In all three sorts of models, it is
generally the case that for any given cognitive episode,
like performing a particular inference or answering a
question, some of the memory states will be actively
involved, and others will be dormant......
The thesis we have been defending in this essay is that
connectionist models of a certain sort are incompatible
with the propositional modularity embedded in
commonsense psychology. The connectionist models in
question are those that are offered as models at the
COGNITIVE level, and in which the encoding of
information is widely distributed and subsymbolic. In
such models, we have argued, there are no DISCRETE,
SEMANTICALLY INTERPRETABLE states that play a CAUSAL
ROLE in some cognitive episodes but not others. Thus
there is, in these models, nothing with which the
propositional attitudes of commonsense psychology can
plausibly be identified. If these models turn out to
offer the best accounts of human belief and memory, we
shall be confronting an ONTOLOGICALLY RADICAL theory
change - the sort of theory change that will sustain the
conclusion that propositional attitudes, like caloric
and phlogiston, do not exist.'
W. Ramsey, S. Stich and J. Garon (1991)
Connectionism, eliminativism, and the future of folk
psychology.
See "Fragments of Behaviour: The Extensional Stance" below for an
elaboration on why Savain's remarks are naive.
--
David Longley (check end reply line #)
Longley Consulting London, UK
Behaviour Assessment & Profiling Technology,
Research, Data Analysis and Training Services,
Small IT Systems http://www.longley.demon.co.uk
Louis Savain
In article <689bto$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil
Rickert) wrote:
>In <34a91b73...@news.earthlink.net> sav...@earthlink.net (Louis Savain) writes:
>>In article <6891in$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil
>>Rickert) wrote:
>
>>>[...]
>>>But, as you yourself have admitted, there has been very little
>>>progress in AI during the several decades in which we have had
>>>computers. One of the problems is that the availability of computers
>>>tends to encourage the basically solipsistic approach of emphasizing
>>>internal structure (algorithms, etc), and ignoring the external
>>>relation between the individual and the world.
>
>> I personally don't see what is wrong with so-called "solipsistic"
>>approaches. A solipsist approach means that the system decides what
>>to do with the incoming sensory stream and how to organize it into
>>useful structures, not the programmer.
>
>I'm afraid that you missed my point. My objection is to the emphasis
>on what to do with internal representations. Instead I want to
>emphasize acting on the incoming sensory stream and forming
>representations.
Sorry, you phrased your message in such a way that it seemed to be
saying the exact opposite.
> My chief objection to your approach, and to
>Modlin's approach, is that, if I understand what you have posted, you
>both want to deal with representations, rather than with
>pre-represented raw input.
I personally don't want to represent anything and I don't think Bill
Modlin does either. Representation is the antithesis of the
solipsistic/emergent intelligence approach. Representational schemes
usually assumes that the programmer has a degree of a priori knowledge
regarding how specific sensory signals are interrelated and somehow
needs to represent this knowledge symbolically in the machine. That,
to me, is the 100% wrong approach.
I want to create a system that builds a signal filtration structure
(call it a knowledge structure) from information it receives from
multiple sensory streams. Some may want to see this knowledge
structure as a "representational structure" but from my perspective
and that of the machine, it isn't a representation at all. First, the
machine does not see a real world, it only "senses" the incoming
signals (the search space) and deals with them appropriately. Second,
I have no interest in unraveling the content of the emerging knowledge
structure or pick out any existing representational counterpart in the
real world. My interests are primarily in the *observable behavior*
effected by the structure and the mechanism that *builds* the
structure. Of course, this subsumes that one has an idea about the
operational nature of the fundamental building blocks that make up the
structure and how they are put together. It also subsumes that the
same building blocks can be put together different ways to build any
sort of adaptive structure permissible by the search space. The
sensory streams would presumably contain causally or temporally
correlatable Boolean signals. Part of my thesis is that this is the
only assumption that a general adaptive system needs to make about
incoming signals.
David Longley
> In article <34a91b73...@news.earthlink.net>
> sav...@earthlink.net "Louis Savain" writes:
>
> > >But, as you yourself have admitted, there has been very little
> > >progress in AI during the several decades in which we have had
> > >computers. One of the problems is that the availability of computers
> > >tends to encourage the basically solipsistic approach of emphasizing
> > >internal structure (algorithms, etc), and ignoring the external
> > >relation between the individual and the world.
> >
> > I personally don't see what is wrong with so-called "solipsistic"
> > approaches. A solipsist approach means that the system decides what
> > to do with the incoming sensory stream and how to organize it into
> > useful structures, not the programmer.
Two things to consider:
1) Man has been around quite some time, science has not.
Unless our knowledge of natural selection is *way* of base
it is highly unlikely that there is much "in our heads"
which distinguishes us from our remote ancestors.
2) From the above, it must be something other than our
natural abilities which accounts for our 'remarkable'
intelligence and which should be looked to for all future
developments of our intelligence.
3) The attractions of the "psychological" are probably little
more than a fascination with the private and unfamiliar -
as they are under minimal social control, they are etheral
and vague - but that doesn't make them important.
I've outlined a framework elsewhere which highlights how this has
come about and why. As a consequence, a lot of what is currently
being discussed under the guise of AI is radically ill-conceived.
Louis Savain
In article <883444...@longley.demon.co.uk>,
Da...@longley.demon.co.uk (David Longley) wrote:
>> In article <34a91b73...@news.earthlink.net>
>> sav...@earthlink.net "Louis Savain" writes:
>>
>> > >But, as you yourself have admitted, there has been very little
>> > >progress in AI during the several decades in which we have had
>> > >computers. One of the problems is that the availability of computers
>> > >tends to encourage the basically solipsistic approach of emphasizing
>> > >internal structure (algorithms, etc), and ignoring the external
>> > >relation between the individual and the world.
>> >
>> > I personally don't see what is wrong with so-called "solipsistic"
>> > approaches. A solipsist approach means that the system decides what
>> > to do with the incoming sensory stream and how to organize it into
>> > useful structures, not the programmer.
>
>Two things to consider:
>
> 1) Man has been around quite some time, science has not.
> Unless our knowledge of natural selection is *way* of base
> it is highly unlikely that there is much "in our heads"
> which distinguishes us from our remote ancestors.
This has little relevance to my post, that I can see.
> 2) From the above, it must be something other than our
> natural abilities which accounts for our 'remarkable'
> intelligence and which should be looked to for all future
> developments of our intelligence.
This is nonsense.
>3) The attractions of the "psychological" are probably little
> more than a fascination with the private and unfamiliar -
> as they are under minimal social control, they are etheral
> and vague - but that doesn't make them important.
The weird thing is that I'm not attracted to the psychological. I'm
mainly attracted to fundamental neural processes. So why the
misdirected allusion?
>I've outlined a framework elsewhere which highlights how this has
>come about and why. As a consequence, a lot of what is currently
>being discussed under the guise of AI is radically ill-conceived.
Unfortunately, the stuff that you've outlined are either irrelevant
or does not provide an answer to the question posed by the Mr. Nick
Bostrom who posted the original article that started this thread. The
question is: How long before superintelligence? The incorrigible
optimist in me answers: 10 to 20 years or less, i.e., within the
lifetime of most of those reading this newsgroup.
You, on the other hand, have no answer that can be relied upon
because you have no contribution to offer. You only harp about the
unreliability of human judgment and that, as everyone else knows, is
an absurd fallacy driven by your personal agenda. Besides, even if
humans were as unreliable as you love to delude yourself into
believing, the AI community does not care. They want to emulate human
intelligence and that of other advanced biological systems, faults and
all. They want to do it whether or not you think it's the right thing
to do. And it *will* be done.
JRStern
On 29 Dec 1997 17:52:00 -0600, ric...@cs.niu.edu (Neil Rickert)
wrote:
>It is my impression that he is sympathetic to the positions of Fodor
>and Chomsky.
>
>> BTW, have you actually read
>>it?
>
>Not yet. I have read his "The Language Instinct", which is supposed
>to be a rather better book. I have also heard him discuss "How the
>Mind Works" on a radio show.
I didn't read the earlier book, but leafed through it at the store.
For me, I doubt it's better than this one, especially if in it he's
sympathetic to Chomsky. Outside of some evolutionary neural wiring
for language, he hasn't mentioned Chomsky in the first half of this
new one. A lot of this new one draws on the same ideas as Matt
Ridley's The Red Queen, though Pinker has some criticms of that, too.
Fodor, well, I run hot and cold on him, and he's only gotten a quick,
tangential mention so far in this one, too.
The more I think about it, the more I think you, specifically, will
find much to agree with in the new book.
Joshua Stern
JRS...@gte.net
JRStern
On Tue, 30 Dec 1997 01:00:33 GMT, sav...@earthlink.net (Louis Savain)
wrote:
> I personally don't want to represent anything and I don't think Bill
>Modlin does either.
You don't want to represent it in bits?
> Representation is the antithesis of the
>solipsistic/emergent intelligence approach. Representational schemes
>usually assumes that the programmer has a degree of a priori knowledge
>regarding how specific sensory signals are interrelated and somehow
>needs to represent this knowledge symbolically in the machine. That,
>to me, is the 100% wrong approach.
That, to me, is 100% ineliminable, unless thoughts about the moon and
the space shuttle mean you must stuff both between your ears. I have
nothing against the a priori, in its place.
I want to finish this Pinker book before posting a review, but even on
partial reading, I strongly recommend it, on these issues.
> I want to create a system that builds a signal filtration structure
>(call it a knowledge structure) from information it receives from
>multiple sensory streams. Some may want to see this knowledge
>structure as a "representational structure" but from my perspective
>and that of the machine, it isn't a representation at all.
If it isn't the thing, it's a representational structure. Analog,
digital, a priori, a postiori, ad-hoc, ipso facto.
Joshua Stern
JRS...@gte.net
Hans Moravec
(This is part three of a multipart reply to this thread)
JRStern:
> Is [Deep Blue victory over Karparov] to be seen as a victory
> for AI? Again, yes, but again only minimally. Let's give it
> a grade of C. When a 100mip desktop has a good enough
> algorithm to beat any human (I expect this within about 10
> years), I'll give it a B. I think a closed game like chess,
> however large the domain, is a dubious test of intelligence
> in any case.
I have a different take on this. The evolution of computer chess seems
to me to be the best model so far of how we can expect machine
intelligence to evolve from sub-human to super-human in other areas in
the near future. More on that at the end of the post.
The majority of AI applications had to putter along at a wormlike 1 MIPS
for decades prior to 1990, and even now are being powered merely by a
buglike few hundred MIPS
(by the calibration outlined in
http://www.frc.ri.cmu.edu/~hpm/book97/ch3/index3.html )
Since most AI programs seem unlikely to do more work than a single
human, few businesses or other organizations have felt it worthwhile to
invest more in individual AI systems than it costs to support a person.
Computer chess was an exception.
(excerpt)
The best chess-playing computers are so interesting they generate
millions of dollars of free advertising for the winners, and
consequently have enticed a series of computer companies to donate time
on their best machines and other resources to the cause. Since 1960
IBM, Control Data, AT&T, Cray, Intel and now again IBM have been
sponsors of computer chess. The “knights” in the AI power graph show
the effect of this largess, relative to mainstream AI research. The top
chess programs have competed in tournaments powered by supercomputers,
or specialized machines whose chess power is comparable. In 1958 IBM
had both the first checker program, by Arthur Samuel, and the first full
chess program, by Alex Bernstein. They ran on an IBM 704, the biggest
and last vacuum-tube computer. The Bernstein program played
atrociously, but Samuel's program, which automatically learned its board
scoring parameters, was able to beat Connecticut checkers champion
Robert Nealey. Since 1994, Chinook, a program written by Jonathan
Schaeffer of the University of Alberta, has consistently bested the
world's human checker champion. But checkers isn't very glamorous, and
this portent received little notice.
By contrast, it was nearly impossible to overlook the epic battles
between world chess champion Garry Kasparov and IBM's Deep Blue in 1996
and 1997. Deep Blue is a scaled-up version of a machine called Deep
Thought, built by Carnegie Mellon University students ten years
earlier. Deep Thought, in turn, depended on special-purpose chips, each
wired like the Belle chess computer built by Ken Thompson at AT&T Bell
Labs in the 1970s. Belle, organized like a chessboard, circuitry on the
squares, wires running like chess moves, could evaluate and find all
legal moves from a position in one electronic flash. In 1997 Deep Blue
had 256 such chips, orchestrated by a 32 processor mini-supercomputer.
It examined 200 million chess positions a second. Chess programs, on
unaided general-purpose computers, average about 16,000 instructions per
position examined. Deep Blue, when playing chess (and only then), was
thus worth about 3 million MIPS, 1/30 of our estimate for human
intelligence.
Deep Blue, in a first for machinekind, won the first game of the 1996
match. But, Kasparov quickly found the machine's weaknesses, and drew
two and won three of the remaining games.
In May 1997 he met an improved version of the machine. That February,
Kasparov had triumphed over a field of grandmasters in a prestigious
tournament in Linares, Spain, reinforcing his reputation as the best
player ever, and boosting his chess rating past 2800, uncharted
territory. He prepared for the computer match in the intervening
months, in part by playing against other machines. Kasparov won a long
first game against Deep Blue, but lost next day to masterly moves by the
machine. Then came three grueling draws, and a final game, in which a
visibly shaken and angry Kasparov resigned early, with a weak position.
It was the first competition match he had ever lost.
The event was notable for many reasons, but one especially is of
interest here. Several times during both matches, Kasparov reported
signs of mind in the machine. At times in the second tournament, he
worried there might be humans behind the scenes, feeding Deep Blue
strategic insights!
Bobby Fischer, the US chess great of the 1970s, is reputed to have
played each game as if against God, simply making the best moves.
Kasparov, on the other hand, claims to see into opponents' minds during
play, intuiting and exploiting their plans, insights and oversights. In
all other chess computers, he reports a mechanical predictability
stemming from their undiscriminating but limited lookahead, and absence
of long-term strategy. In Deep Blue, to his consternation, he saw
instead an "alien intelligence."
In this paper-thin slice of mentality, a computer seems to have not
only outperformed the best human, but to have transcended its
machinehood. Who better to judge than Garry Kasparov? Mathematicians
who examined EQP's proof of the Robbins conjecture, mentioned earlier,
report a similar impression of creativity and intelligence. In both
cases, the evidence for an intelligent mind lies in the machine's
performance, not its makeup.
Now, the team that built Deep Blue claim no "intelligence" in it, only
a large database of opening and end games, scoring and deepening
functions tuned with consulting grandmasters, and, especially, raw speed
that allows the machine to look ahead an average of fourteen half-moves
per turn. Unlike some earlier, less successful, chess programs, Deep
Blue was not designed to think like a human, to form abstract strategies
or see patterns as it races through the move/countermove tree as fast as
possible.
Deep Blue's creators know its quantitative superiority over other chess
machines intimately, but lack the chess understanding to share
Kasparov's deep appreciation of the difference in the quality of its
play. I think this dichotomy will show up increasingly in coming
years. Engineers who know the mechanism of advanced robots most
intimately will be the last to admit they have real minds. From the
inside, robots will indisputably be machines, acting according to
mechanical principles, however elaborately layered. Only on the
outside, where they can be appreciated as a whole, will the impression
of intelligence emerge. A human brain, too, does not exhibit the
intelligence under a neurobiologist's microscope that it does
participating in a lively conversation.
(end excerpt)
Over the decades, there have been various approaches to computer chess.
Hans Berliner of Carnegie Mellon, a master player himself, investigated
knowledge-intensive approaches, and even ones based on higher-level
reasoning, in which a theorem prover pondered abstract properties of
chess positions. Some of these programs were pretty good in their day,
but it always turned out that they became even better when their tricky
evaluations were simplified, freeing up resources for a little
additional search.
I see no fundamental difference between search-intensive,
knowledge-intensive, learning-intensive or memory-intensive approaches.
For instance, the results of a lot of search can be encoded into a
sufficiently large learning, knowledge or case-memory setup. Which way
you do it is a mere implementation detail, which may even be decided at
the last moment by some code in an optimizing compiler. Particular
hardware may bias the optimum implementation one way or another. Deep
Blue's chess chips, so much more powerful the their host machine, almost
guarantee that the best solution will depend most heavily on tree
search. But, a computer with molecular memory a million times bigger
than Deep Blue's would surely depend more on
opening, and endgame tables, and perhaps on massively learned
evaluations in the midgame. At one stage in Deep Blue's development
there was considerable effort in automatically learning hairy evaluation
functions. The idea is to make the evaluation simulate deeper search as
accurately as possible. In the extreme, after all, an evaluation with
enough parameters, based on the full board position, could simply be a
compact encoding of the results of a lot of search.
But in an abstract sense the different approaches can be
nearly equivalent, and make no difference to an outsider
who plays with the machine. If one implementation passes
a Turing test, so will an externally equivalent one.
Some day even the machine's designers may not know
or care, because the implementation choices will have been
made be a design progam that optimized them for that day's
technology, which will probably have changed by next week.
By analogy, do you care which exact Postscript control
sequences your latest version of laserwriter driver sends
to the pinter, so long as what comes out resembles what you
entered on your editor screen?
Kasparov, with his 100 billion millisecond-slow neurons, with
100 trillion synapses tuned by a lifetime of thinking about
chess, probably simultaneously weighs a few million possible
move sequences subconsciously in terms of a few million
learned characteristics. After all that subconscious
selection, a few quality moves are elevated into his
consciousness awareness for final execution.
In the same time, Deep Blue churns through several billion sequences,
weighing each in terms of a few hundred characteristics. From all that
analysis, it extracts a few best moves and pivotal positions, and prints
them out for its programmers.
Deep Blue does only a thin slice of what Kasparov does, but
in that slice it is just about his intellectual equal. It is mere
bigotry to claim that its intelligence is somehow inferior because of
the particular way it is implemented.
Recent progress in computer chess, mobile robots, theorem proving and
other AI areas strongly suggests that intelligence can be achieved
simply by applying enough computational power to modest but effective
constructs of of effective simple ideas, with effectivness revealed by
trial and error. No big new theories needed, thank you.
I am smugly confident that full intelligence will be achieved in similar
modest steps, as the requisite 100 million MIPS or so becomes available
over the next several decades.
(an old draft of one suggested path for the small steps
can found in
http://www.frc.ri.cmu.edu/~hpm/Mind.Age/3..Universal
)
David Longley
> In <883444...@longley.demon.co.uk> Da...@longley.demon.co.uk (David Longley)
> writes:
>
> >Two things to consider:
>
> > 1) Man has been around quite some time, science has not.
> > Unless our knowledge of natural selection is *way* of base
> > it is highly unlikely that there is much "in our heads"
> > which distinguishes us from our remote ancestors.
>
> Agreed.
>
> > 2) From the above, it must be something other than our
> > natural abilities which accounts for our 'remarkable'
> > intelligence and which should be looked to for all future
> > developments of our intelligence.
>
> Sorry, but (2) does not obviously follow from (1). At various times
> dolphins and chimpanzees have been said to have remarkable
> intelligence. And even if (2) is correct, the "something other ..."
> might merely be the fact that we are a social animal. Keep in mind
> that social bees and ants are quite remarkable compared with
> non-social insects.
>
It may not *obviously* follow from (1) but I *am* pressing the
thesis that it's what we have collectively learned and, more
significantly, *inculcated* which accounts for our advances in
intelligence. So I'd say the mistake above is to say "merely". I
suspect (and many others have said this too) that it's our
linguistic/communication skills which differentiate us from other
species in this respect, and possibly even within our own - some
languages having been more apposite than others. Within those, we
have developed some languages which are *particularly* useful,
logic, mathematics, statistics, physics, electronics, biology -
and it is, IMHO, in our mastery of *these* which is coextensive
with "AI".
Louis Savain
In article <689uhg$97g$2...@gte1.gte.net>, JRS...@gte.net (JRStern)
wrote:
>On Tue, 30 Dec 1997 01:00:33 GMT, sav...@earthlink.net (Louis Savain)
>wrote:
>> I personally don't want to represent anything and I don't think Bill
>>Modlin does either.
>
>You don't want to represent it in bits?
Nope.
>> Representation is the antithesis of the
>>solipsistic/emergent intelligence approach. Representational schemes
>>usually assumes that the programmer has a degree of a priori knowledge
>>regarding how specific sensory signals are interrelated and somehow
>>needs to represent this knowledge symbolically in the machine. That,
>>to me, is the 100% wrong approach.
>
>That, to me, is 100% ineliminable, unless thoughts about the moon and
>the space shuttle mean you must stuff both between your ears. I have
>nothing against the a priori, in its place.
I do.
>I want to finish this Pinker book before posting a review, but even on
>partial reading, I strongly recommend it, on these issues.
I saw Pinker on a TV show pushing his book. I'm not very impressed.
Especially by the title of the book. If he really knew how brains
thought the search would be over and we would all go home.
>> I want to create a system that builds a signal filtration structure
>>(call it a knowledge structure) from information it receives from
>>multiple sensory streams. Some may want to see this knowledge
>>structure as a "representational structure" but from my perspective
>>and that of the machine, it isn't a representation at all.
>
>If it isn't the thing, it's a representational structure. Analog,
>digital, a priori, a postiori, ad-hoc, ipso facto.
Wrong. The knowledge structure and the sensory signals are the
*only* things. Whether or not the signals have their roots in a real
world is not only irrelevant to the system/knowledge builder, but can
only be inferred after the fact. Inveterate solipsists say even
*that* is impossible. I personally would not go that far.
A representational system is a symbolic system which means that you
must know the thing you want to symbolize before you can symbolize it.
That is of course impossible in an emergent system since it does not
know what it knows until after it learns it. A priori knowledge is
absurd when one is dealing with learning. That's why knowledge
engineering is a red herring and has little relevance to the goal of
creating a common sense robot.
mod...@concentric.net
In <689bto$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil Rickert) writes:
[responding to Savain]
>I'm afraid that you missed my point. My objection is to the emphasis
>on what to do with internal representations. Instead I want to
>emphasize acting on the incoming sensory stream and forming
>representations. My chief objection to your approach, and to
>Modlin's approach, is that, if I understand what you have posted, you
>both want to deal with representations, rather than with
>pre-represented raw input.
I don't understand this paragraph at all.
In my models (I think Savain's also) we start with raw sensory input
and attempt to discover what objects and relationships are implied
by that input. We then generate internal signals to model or represent
these derivative objects, and continue the abstraction process by treating
our derivative signals as raw input for further discovery.
I assume that there is some uniform mode of representation of signals
(i.e. neurons firing, or codes in a computer) but this has to do with
the mechanisms of processing, not with any prior assumptions about what
real-world objects and concepts may be modelled by those signals.
I don't know what you mean by "pre"-represented raw input. All input
at all levels is just signals, caused by whatever causes it, and
meaningless until we find out what it means by discovering how it
behaves, what else it tells us to expect.
Bill Modlin
Neil Rickert
In <689uvp$mas$1...@gte2.gte.net> JRS...@gte.net (JRStern) writes:
>On 29 Dec 1997 17:52:00 -0600, ric...@cs.niu.edu (Neil Rickert)
>wrote:
>>> BTW, have you actually read
>>>it?
>>Not yet. I have read his "The Language Instinct", which is supposed
>>to be a rather better book. I have also heard him discuss "How the
>>Mind Works" on a radio show.
>The more I think about it, the more I think you, specifically, will
>find much to agree with in the new book.
In another post you pointed out that a priori knowledge plays a large
role for Pinker. All the more reason for me to think he is on the
wrong track.
Neil Rickert
In <883465...@longley.demon.co.uk> Da...@longley.demon.co.uk (David Longley) writes:
>In article <689rvt$7...@ux.cs.niu.edu> ric...@cs.niu.edu "Neil Rickert" writes:
>> > 2) From the above, it must be something other than our
>> > natural abilities which accounts for our 'remarkable'
>> > intelligence and which should be looked to for all future
>> > developments of our intelligence.
>> Sorry, but (2) does not obviously follow from (1). At various times
>> dolphins and chimpanzees have been said to have remarkable
>> intelligence. And even if (2) is correct, the "something other ..."
>> might merely be the fact that we are a social animal. Keep in mind
>> that social bees and ants are quite remarkable compared with
>> non-social insects.
>It may not *obviously* follow from (1) but I *am* pressing the
>thesis that it's what we have collectively learned
It is the fact that we are a social species that makes such
collective learning possible.
> and, more
>significantly, *inculcated* which accounts for our advances in
>intelligence. So I'd say the mistake above is to say "merely".
I don't like the word "inculcate". It has too much of a sense of
brain washing - behaviorist dogma being applied for mind control. It
tends to make the learner's role seem passive, and make the teacher
the active participant. But the learner's role is far more
important, and the teachers role far more complex and subtle than the
word "inculcate" suggests.
> I
>suspect (and many others have said this too) that it's our
>linguistic/communication skills which differentiate us from other
>species in this respect, and possibly even within our own - some
>languages having been more apposite than others. Within those, we
>have developed some languages which are *particularly* useful,
>logic, mathematics, statistics, physics, electronics, biology -
>and it is, IMHO, in our mastery of *these* which is coextensive
>with "AI".
Here we see Longley's methodological solipsism standing out. He is
emphasizing internal structure (languages of logic, mathematics,
electronics, biology). In contrast, I offer Rickert's behaviorism -
it is the externally applicable skills of mathematical technique,
statistical analysis, physical measuring, electronic technology and
biological classification and analysis that is of prime value. The
languages are only there to support these skills.
Chris Hooley
On Mon, 29 Dec 1997 23:25:56 GMT, JRS...@gte.net (JRStern) wrote:
>.................... If I'm right, kids raised on the
>Internet (er, you know what I mean <g>) will not all sit quietly for
>solipcistic crap, if a reasonable alternative is made available. Ten
>year horizon.
I don't understand this. You speak as though Savain's and Modlin's engines
were the dominant view rather than minority voices, as though solipsystem
computer logic was established doctrine.
To my limited knowledge, only during the last year has this newsgroup
seriously entertained the idea that robust AI might emerge without *any*
human-dependent constructs about "the world."
After all, that is aspect of solipsism that is useful for AI research; that
the meaningful "world" is only known one self-referencing system at a time.
Why do you speak of kids not sitting quietly for "solipsistic crap?"
The last I heard, solipsism was generally viewed as an anti-social juvenile
philosophy, yet you speak of it as though it were the prevailing cultural
stance, a position that internet-raised kids could rebel against.
Within twenty years, in my opinion, self-programmed AI is going to be so
clearly intelligent that none of our opinions on the matter will make much
difference. I put that short a time-frame on it because the AI will
participate in its own evolution. The only constraints are the limits to
resources. Perhaps, as several, including Seth, have proposed, the
Internet itself will become the parent/global "body" of a 21st Century AI.
Unless of course the kids rise up in revolt. "We're not going to take this
solipsist crap any longer! We're going to rebel and recreate the good old,
stable, externally-viewed world of our fathers!"
Regards
Chris
David Longley
Ahh, a seasonal joke....
I've spent years pressing the point that these are skills of
course - and for that I've usually been cast as the evil
behaviourist.
Still - in the interests of seasonal good will...
Rick Ottolini
In article <689ba3$6v1$1...@gte1.gte.net> JRS...@gte.net (JRStern) writes:
>Yes, my rationalization for that is the computer age really began when
>100mhz processors got cheap, say, 1993, coincidentally about the time
>this new upsurge in interest regarding consciousness started up.
The computer age has barely begun.
The user interfaces are very crude compared to their potential.
They are like electric motors and internal combustion engines
in the late 19th century- evolving rapidly, had to use then.
Twenty years from now today's computers will seem very slow
and unbelievably cumbersome.
Neil Rickert
In <68ah9k$7...@examiner.concentric.net> mod...@concentric.net writes:
>In <689bto$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil Rickert) writes:
>[responding to Savain]
>>I'm afraid that you missed my point. My objection is to the emphasis
>>on what to do with internal representations. Instead I want to
>>emphasize acting on the incoming sensory stream and forming
>>representations. My chief objection to your approach, and to
>>Modlin's approach, is that, if I understand what you have posted, you
>>both want to deal with representations, rather than with
>>pre-represented raw input.
>I don't understand this paragraph at all.
>In my models (I think Savain's also) we start with raw sensory input
>and attempt to discover what objects and relationships are implied
>by that input. We then generate internal signals to model or represent
>these derivative objects, and continue the abstraction process by treating
>our derivative signals as raw input for further discovery.
It seemed to me that your attempt to discover objects and
relationships was something you are doing with your representations.
You use Shannon's information theory, and you talk about entropy --
but this is a theory of represented information.
>I don't know what you mean by "pre"-represented raw input. All input
>at all levels is just signals, caused by whatever causes it, and
>meaningless until we find out what it means by discovering how it
>behaves, what else it tells us to expect.
It is signals until we generate symbols (or digits, or other discrete
units). Thereafter it is symbolic representations.
Neil Rickert
In <68aukv$b8m$1...@gte2.gte.net> cho...@idnsi.net (Chris Hooley) writes:
>On Mon, 29 Dec 1997 23:25:56 GMT, JRS...@gte.net (JRStern) wrote:
>>.................... If I'm right, kids raised on the
>>Internet (er, you know what I mean <g>) will not all sit quietly for
>>solipcistic crap, if a reasonable alternative is made available. Ten
>>year horizon.
>I don't understand this. You speak as though Savain's and Modlin's engines
>were the dominant view rather than minority voices, as though solipsystem
>computer logic was established doctrine.
>To my limited knowledge, only during the last year has this newsgroup
>seriously entertained the idea that robust AI might emerge without *any*
>human-dependent constructs about "the world."
Actually, I have been trying to push that idea since 1991, but with
little success.
>After all, that is aspect of solipsism that is useful for AI research; that
>the meaningful "world" is only known one self-referencing system at a time.
>Why do you speak of kids not sitting quietly for "solipsistic crap?"
>The last I heard, solipsism was generally viewed as an anti-social juvenile
>philosophy, yet you speak of it as though it were the prevailing cultural
>stance, a position that internet-raised kids could rebel against.
To be fair to Josh, I introduced the term 'solipsism' into the
discussion. Josh just continued the point. I'll grant that
solipsism is rejected by most people. In introducing the term, I was
being critical of philosophy. In essence I was saying that
philosophy is implicitly solipsistic, even though it explicitly
rejects solipsism. That is to say, philosophy is based on internal
standards and is largely immune to external evidence. Similarly, at
other times I have described the Turing machine as solipsistic, since
it starts out with complete internal information (on its tape), and
does not have any interaction with the world.
>Within twenty years, in my opinion, self-programmed AI is going to be so
>clearly intelligent that none of our opinions on the matter will make much
>difference.
It hasn't even begun. It seems overly optimistic to expect to be
finished in 20 years.
David Longley
> In my models (I think Savain's also) we start with raw sensory input
> and attempt to discover what objects and relationships are implied
> by that input. We then generate internal signals to model or represent
> these derivative objects, and continue the abstraction process by treating
> our derivative signals as raw input for further discovery.
>
But what is "raw sensory input"? This seems to harbour all that
the old mentalistic empiricists were guilty of. All sensory
systems are selective in that they're adaptive to some energy
bandwidth, so there is no such thing as "raw" data in that sense.
All data is sampled by some operations on the array, so there's
something there too.
You should read Carnap's "Aufbau" written in the 1920s. It was
the best attempt to build the world from logical primitives in
the spirit of Principia Mathematica. But even those primitives
have been arrived at empirically as others pointed out, and the
primitives really are primitive.
Personally, I don't think it worth spelling this out in more
detail.
mod...@concentric.net
In <68be03$8...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil Rickert) writes:
>In <68ah9k$7...@examiner.concentric.net> mod...@concentric.net writes:
>>In <689bto$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil Rickert) writes:
>
>>[responding to Savain]
>
>>>I'm afraid that you missed my point. My objection is to the emphasis
>>>on what to do with internal representations. Instead I want to
>>>emphasize acting on the incoming sensory stream and forming
>>>representations. My chief objection to your approach, and to
>>>Modlin's approach, is that, if I understand what you have posted, you
>>>both want to deal with representations, rather than with
>>>pre-represented raw input.
>
>>I don't understand this paragraph at all.
>
>>and attempt to discover what objects and relationships are implied
>>by that input. We then generate internal signals to model or represent
>>these derivative objects, and continue the abstraction process by treating
>>our derivative signals as raw input for further discovery.
>
>relationships was something you are doing with your representations.
>You use Shannon's information theory, and you talk about entropy --
>but this is a theory of represented information.
>
>>I don't know what you mean by "pre"-represented raw input. All input
>>at all levels is just signals, caused by whatever causes it, and
>>meaningless until we find out what it means by discovering how it
>>behaves, what else it tells us to expect.
>
>It is signals until we generate symbols (or digits, or other discrete
>units). Thereafter it is symbolic representations.
I'm still having trouble understanding the distinction you are making.
Or if I do understand it, I can't see where you are applying it to the
proposal I've made.
Something impinges on a neural sensor.
The sensor fires if the stimulus is strong enough, generating a discrete
signal which is a quantized measurement of the stimulus.
Many such sensors react to many stimuli, with varying sensitivities and
physical separations, generating many discrete signals which taken together
convey information about the processes which modulated the stimuli being
sensed.
More neurons, each with access to some subset of that ensemble of signals,
react to various combinations of them, producing more signals of the same
physical kind as new outputs.
Still more neurons, with access to subsets of both the original signals from
the sensors and the internally-generated derivative signals, react to still
more combinations, and generate still more derivative signals.
This continues through many levels. The signals accessed by any one cell
in the tissue may include some derived directly from sensors, some
derived as intermediate products more indirectly, and even some derived
through recurrent or feedback paths indirectly from the previous actions
of the cell under examination.
Each cell receives as inputs a number of signals from the firing of other
cells. It generates a new signal as an adaptive function of the inputs
it receives. The new signal it generates is propagated to become an input
to some other set of cells.
All these signals are in a sense discrete representations of some
combination of inputs from which they are derived. In that general
functional sense they could be called "symbols" of those combinations...
but the same could be said of any signal of any sort, no matter how
derived.
When do any of these signals become "symbols" in the sense you are
objecting to? There is never at this level of description any mapping
to arbitrary "symbols" with assigned meanings. All mappings are
functional, not symbolic.
Of course, the point of the exercise is my hope and expectation that
if we get the rules for adaptive signal generation right, many of
the signals will turn out to be in good correspondence with abstractions
for which we customarily do use symbols. But that's at another level
of interpretation, not directly relevant to a discussion of the
algorithms and mechanisms themselves.
Bill
mod...@concentric.net
In <883510...@longley.demon.co.uk>, Da...@longley.demon.co.uk (David Longley) writes:
>In article <68ah9k$7...@examiner.concentric.net> mod...@concentric.net writes:
>
>> and attempt to discover what objects and relationships are implied
>> by that input. We then generate internal signals to model or represent
>> these derivative objects, and continue the abstraction process by treating
>> our derivative signals as raw input for further discovery.
>>
>the old mentalistic empiricists were guilty of. All sensory
>systems are selective in that they're adaptive to some energy
>bandwidth, so there is no such thing as "raw" data in that sense.
>All data is sampled by some operations on the array, so there's
>something there too.
"Raw sensory input" is whatever you have to start with.
Period. No assumptions beyond the fact that there must be some
source of data if we are going to try to do anything with data.
Getting into some infinite regress about how the data was derived
is silly. The point of the exercise is that there is some data
and we want to discover something about it without assuming ahead
of time that we know what it might be.
Bill Modlin
Louis Savain
In article <68beiv$8...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil
Rickert) wrote:
>In <68aukv$b8m$1...@gte2.gte.net> cho...@idnsi.net (Chris Hooley) writes:
>>On Mon, 29 Dec 1997 23:25:56 GMT, JRS...@gte.net (JRStern) wrote:
>
>>>.................... If I'm right, kids raised on the
>>>Internet (er, you know what I mean <g>) will not all sit quietly for
>>>solipcistic crap, if a reasonable alternative is made available. Ten
>>>year horizon.
>
>>I don't understand this. You speak as though Savain's and Modlin's engines
>>were the dominant view rather than minority voices, as though solipsystem
>>computer logic was established doctrine.
>
>>To my limited knowledge, only during the last year has this newsgroup
>>seriously entertained the idea that robust AI might emerge without *any*
>>human-dependent constructs about "the world."
>
>Actually, I have been trying to push that idea since 1991, but with
>little success.
You and Chris Hooley are to be commended for your efforts. Note
however that the idea is not as recent as some of us may think. Many
prophets of solipsistic emergence have cried in the wilderness of
formal logic and symbolic AI since at least the early eighties. See,
for example, Pierre Bierre's now forgotten winter 1984 article in AI
Magazine entitled "The Professor's Challenge." The whole movement
toward ANNs is in part due to a strongly felt need to shift away from
the representational stance.
>[...]
>>Within twenty years, in my opinion, self-programmed AI is going to be so
>>clearly intelligent that none of our opinions on the matter will make much
>>difference.
>
>It hasn't even begun. It seems overly optimistic to expect to be
>finished in 20 years.
You may not share in the optimism but the shift in paradigm is now
well on its way and there's no stopping it. My own personal estimate
is that within five years we will know exactly what it's going to take
to build an emergent synthetic intelligence, one with the right
adaptive mechanism that will allow it to learn and achieve the common
sense of an adult human being. Within 10 years, we will have the
resources to build one. Like I said I'm an incorrigible optimist.
Neil Rickert
In <68bjjb$o...@examiner.concentric.net> mod...@concentric.net writes:
>In <68be03$8...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil Rickert) writes:
>>It is signals until we generate symbols (or digits, or other discrete
>>units). Thereafter it is symbolic representations.
>I'm still having trouble understanding the distinction you are making.
>Or if I do understand it, I can't see where you are applying it to the
>proposal I've made.
>Something impinges on a neural sensor.
>The sensor fires if the stimulus is strong enough, generating a discrete
>signal which is a quantized measurement of the stimulus.
Among other things, I am suggesting that what counts as "strong
enough" is too important to be taken for granted.
Your procedure, as I understand it, is to map the world to symbolic
representations (syntax). Then you will search for syntactic
patterns. Once you have discovered a syntactic pattern, you hope it
will correspond to a real world pattern (which we might call a
semantic pattern). It is not at all obvious that semantic patterns
will show up as syntactic patterns. Whether they will presumably
depends on the effectiveness of how you reduce the external world
(semantics) to syntax. And we should expect that there can be
syntactic patterns which do not arise from semantic patterns.
In most of the cases I can think of, where we form symbolic
representations of the external world, we first find patterns in the
external world, and then use those patterns as part of our process of
symbolization. So it is not clear to me how you can get your system
started.
David Longley
"Data" just doesn't come like that. It is either selected by some
features of a biological transducer or is designed to be
responsive to some features of the energy array and not others.
It is not "silly" to dispute what the status of the data is - it
has been central to discussions about perception and epistemology
for centuries.
Biological systems have evolved under pressures from the
environment. The old idea that there must be uninterpreted "raw"
data from which meaning is identified is a very old empiricist
idea, and one challenged by gestaltists and Gibsonians.
Physiologists don't bother with it much because the nature of
their material dictates their "assumptions". Engineers on the
other hand are generally intersted in solving some particular
signal processing task so work with some energy band. Those who
are not specifically interested in mapping the perceptual systems
of animals are explicitly working on signal processors such as
cameras, sonar systems, you name it.
I've said before that you are just turning a blind eye to
assumptions. You have already settled on a coding system before
you get to any clustering algorithm, and that's likely to shape
your classification as much as the chunks it comes in.
In the days when Sussman was a novice Minsky once came to him as he
sat hacking at the PDP-6. "What are you doing?", asked Minsky.
"I am training a randomly wired neural net to play Tic-Tac-Toe."
"Why is the net wired randomly?", asked Minsky.
"I do not want it to have any preconceptions of how to play"
Minsky shut his eyes,
"Why do you close your eyes?", Sussman asked his teacher.
"So that the room will be empty."
At that momment, Sussman was enlightened.
Anon
Louis Savain
In article <68be03$8...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil
Rickert) wrote:
>[...]
>It is signals until we generate symbols (or digits, or other discrete
>units). Thereafter it is symbolic representations.
This is the crux of the issue and I think it has to do with the
meaning of the word 'representation'. In my mind and, I think, that
of most people, a representation is not the real thing but an agreed
upon symbol or reference, that takes the place of something else. The
symbolic AI people take the stance that neural structures in the brain
have to represent something or other in the real world. Therefore, in
their view, the way to achieve AI is to create structures to represent
"real" world objects. Right off the bat someone may object that there
are many thoughts that have no counterparts in the real world such as
a great many emotional thoughts.
They never stop to think that the intelligent system can never have
direct access to the real world. The world of the system is the
signals it receives through its senses and the processing/filtration
structures that it builds from its analysis of the signals. All that
stuff (signals and structures) is internal to the system and there is
no need for it to represent it in any manner, shape or form. It's the
real thing. It's the only thing.
Ben Sharvy
> In <883444...@longley.demon.co.uk> Da...@longley.demon.co.uk (David
Longley) writes:
>
> >Two things to consider:
>
> > 1) Man has been around quite some time, science has not.
> > Unless our knowledge of natural selection is *way* of base
> > it is highly unlikely that there is much "in our heads"
> > which distinguishes us from our remote ancestors.
>
> Agreed.
If by "science" you mean inquiry based on experiment, then science has
been around a long time, perhaps longer than homo sapiens. Is it
unreasonable to say that other species experiment? At any rate, the
regular use of fire strikes me as likely to have been preceeded by quite a
bit of expermintation (so does the use of metals). So, even if you mean
"organized body of knowledge" by the term "science" I'm not sure that your
statement is as clear as it seems.
Sorry if this reply is out of context: my server seems to have missed the
orginal post.
--
Eco-Socialist Libertarian Capitalism | Can Men Be Lesbians?
http://www.efn.org/~bsharvy/
*Remove "NOSPAM" from email to earn $100,000 from home reading books.
Neil Rickert
In <68c6o3$m...@examiner.concentric.net> mod...@concentric.net writes:
>>Your procedure, as I understand it, is to map the world to symbolic
>>representations (syntax).
>No.
>*I* don't map the world to anything. I'm assuming that's been done.
That is the same thing, near enough. And that is where we disagree.
You assume that the mapping has been done; that is, you take it for
granted. As I see it, doing the mapping is the main problem to be
solved. That is to say, common sense has to do with carrying out the
mapping, rather than with finding clever ways of using the results of
the mapping.
I'll grant that mine is a minority view. In any case, with this
basic disagreement, we won't find much in common, even if we both see
a role for detecting correlations.
>I'm assuming that the mapping was selected for some purpose, to serve
>the needs and goals of an organism or system. The signals I'm looking
>at are presumably being used in the purposive functioning of the
>system, and mean whatever they mean to the system itself. I really
>don't care whether you call them symbols or syntax or whatever... they
>are whatever you think they need to be to play a role in whatever
>system you care to discuss.
>I'm also assuming that you as a designer, or evolution as the designer
>of a biological organism, have already included everything practical
>in the way of built-in functions based on the meanings of those
>signals. If it was useful to build edge and contrast detectors for
>video data, you did it. If it was useful to have some parameters of
>the functioning trained or conditioned by the circumstances
>encountered, you arranged for that.
>What's left is the signals or combinations of signals that you don't
>know what to do with. You don't know what they mean, in any sense
>relevant to the function of the system. If you did, you would have
>designed an appropriate response. But you didn't, so you couldn't.
>That's where I come in. My particular concern is "what can we do
>about signals that we don't know anything about?"
That's certainly my concern. I'm not sure why it is your concern,
for it seems to me that your array of special purpose detectors has
already filtered most of these out so that you will never see them.
Ben Sharvy
People need to stop conflating computation with intelligence. It may turn
out that they are identical (I doubt it), but even so, that needs to be
the *conclusion* of an argument, not a premise. Hans Moravec, in
particular, persists in not distinguishing between the two.
In article <34A890...@cmu.edu>, Hans Moravec <h...@cmu.edu> wrote:
> JRStern:
> > Is [Deep Blue victory over Karparov] to be seen as a victory
> > for AI? Again, yes, but again only minimally. Let's give it
> > a grade of C. When a 100mip desktop has a good enough =
Why is it a victory for AI? The answer cannot merely be that it means that
computers can perform some computations better (i.e., faster?) than
humans, since that has been the case for some time. Deep Blue is no more a
victory for AI than a calcuator--though calculators too are able to
perform computations many times faster than person. Is Eliza, a program
which performs virtually no computation at all, a victory for AI?
> I have a different take on this. The evolution of computer chess seems
> to me to be the best model so far of how we can expect machine
> intelligence to evolve from sub-human to super-human in other areas in
> the near future.
Computer chess is computation; whether it is intelligence is one of the
issues before us. Stop taking it as an axiom in a pro-AI argument. The
result is circular.
> The majority of AI applications had to putter along at a wormlike 1 MIPS
> for decades prior to 1990, and even now are being powered merely by a
> buglike few hundred MIPS
Speed would determine whether your AI's mind worked quickly or slowly: it
would not be relevant to whether your wanna-be AI had a mind. Deep Blue
can be no more "intelligent" than a computer with a MIPS rating from the
70's with an equivalent algorithm.
[snip some discussion of Deep Blue vs. Kasparaov match]
> The event was notable for many reasons, but one especially is of
> interest here. Several times during both matches, Kasparov reported
> signs of mind in the machine. At times in the second tournament, he
> worried there might be humans behind the scenes, feeding Deep Blue
> strategic insights!
>
> Bobby Fischer, the US chess great of the 1970s, is reputed to have
> played each game as if against God, simply making the best moves. =
>
> Kasparov, on the other hand, claims to see into opponents' minds during
> play, intuiting and exploiting their plans, insights and oversights. In
> all other chess computers, he reports a mechanical predictability
> stemming from their undiscriminating but limited lookahead, and absence
> of long-term strategy. In Deep Blue, to his consternation, he saw
> instead an "alien intelligence."
>
> In this paper-thin slice of mentality, a computer seems to have not
> only outperformed the best human, but to have transcended its
> machinehood. Who better to judge than Garry Kasparov? Mathematicians
> who examined EQP's proof of the Robbins conjecture, mentioned earlier,
> report a similar impression of creativity and intelligence. In both
> cases, the evidence for an intelligent mind lies in the machine's
> performance, not its makeup.
The same sort of things were said of Eliza in its heyday, yet Eliza is so
simple even a punk like me has made one. As I recall, Kasparaov's remarks
were part of a not-too-classy diatribe in which the loser made excuses and
hinted at cheating on the part of the Deep Blue team. The remarks of one
ex-champion in the heat of his dethroning, and who is sufficently
grandiose to have characterized his game-playing as "a battle for all
mankind" or some such guff, are not significant.
[snip]
> Deep Blue does only a thin slice of what Kasparov does, but
> in that slice it is just about his intellectual equal. It is mere
> bigotry to claim that its intelligence is somehow inferior because of
> the particular way it is implemented.
Stop conflating computation and intelligence. The claim that Deep Blue's
intelligence is inferior ("non-existent" would be more accurate) has to do
with what Deep Blue does. It does what a calculator does. It's quality of
play is totally irrelevant: Deep Blue would not become more or less
intelligent if a better human player showed up; chess programs from the
70s that could beat me were not more intelligent than me because they
could "play" a better game.
Incidentally, the last time I checked (couple of years ago), the best
computer Go program couldn't beat any professional players at all. There
was an 8-year old Japanese player better than best computer Go program.
One of the Laskers visited China, played Go and described it as much, much
deeper than chess (yet the rules are simpler). What's the state of
computer poker? How about computer painting? Poetry writing?
*Why* the emphasis on computation-based tasks as an indicator of
intelligence? By "intelligence" I mean existence of mind, not "smarts" as
in SAT scores (I'm afraid I suspect you blurring the distinction, since
the latter sense is conventionally tied to your old friend, computational
speed).
> Recent progress in computer chess, mobile robots, theorem proving and
> other AI areas strongly suggests that intelligence can be achieved
> simply by applying enough computational power to modest but effective
> constructs of effective simple ideas...
Stop conflating....
> I am smugly confident that full intelligence will be achieved in similar
> modest steps, as the requisite 100 million MIPS or so becomes available
> over the next several decades.
You don't seem to know what intelligence is.
Ben Sharvy
In article <883540...@longley.demon.co.uk>, Da...@longley.demon.co.uk# wrote:
> In article <bsharvy-3012...@dynip168.efn.org>
> bsh...@NOSPAMefn.org "Ben Sharvy" writes:
>
> > In article <689rvt$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil
Rickert) wrote:
> >
> > > In <883444...@longley.demon.co.uk> Da...@longley.demon.co.uk (David
> > Longley) writes:
> > >
> > > >Two things to consider:
> > >
> > > > 1) Man has been around quite some time, science has not.
> > > > Unless our knowledge of natural selection is *way* of base
> > > > it is highly unlikely that there is much "in our heads"
> > > > which distinguishes us from our remote ancestors.
> > >
> > > Agreed.
> >
> > If by "science" you mean inquiry based on experiment, then science has
> > been around a long time, perhaps longer than homo sapiens. Is it
> > unreasonable to say that other species experiment? At any rate, the
> > regular use of fire strikes me as likely to have been preceeded by quite a
> > bit of expermintation (so does the use of metals). So, even if you mean
> > "organized body of knowledge" by the term "science" I'm not sure that your
> > statement is as clear as it seems.
> >
> > Sorry if this reply is out of context: my server seems to have missed the
> > orginal post.
> >
>
> It *is* out of context. If you think we've had science in any
> form for more than a few hundred years you are sadly misinformed.
We've had science, in every noraml sense of "science", for much longer
than a few hundred years. The Pythagoreans determined the world was
round.... I've already explained my remark (above); are you going to
explain yours?
mod...@concentric.net
>>In <68be03$8...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil Rickert) writes:
>
>>>It is signals until we generate symbols (or digits, or other discrete
>>>units). Thereafter it is symbolic representations.
>
>>Or if I do understand it, I can't see where you are applying it to the
>>proposal I've made.
>
>>Something impinges on a neural sensor.
>
>>The sensor fires if the stimulus is strong enough, generating a discrete
>>signal which is a quantized measurement of the stimulus.
>
>Among other things, I am suggesting that what counts as "strong
>enough" is too important to be taken for granted.
>
>representations (syntax).
No.
*I* don't map the world to anything. I'm assuming that's been done.
I'm assuming that the mapping was selected for some purpose, to serve
the needs and goals of an organism or system. The signals I'm looking
at are presumably being used in the purposive functioning of the
system, and mean whatever they mean to the system itself. I really
don't care whether you call them symbols or syntax or whatever... they
are whatever you think they need to be to play a role in whatever
system you care to discuss.
I'm also assuming that you as a designer, or evolution as the designer
of a biological organism, have already included everything practical
in the way of built-in functions based on the meanings of those
signals. If it was useful to build edge and contrast detectors for
video data, you did it. If it was useful to have some parameters of
the functioning trained or conditioned by the circumstances
encountered, you arranged for that.
What's left is the signals or combinations of signals that you don't
know what to do with. You don't know what they mean, in any sense
relevant to the function of the system. If you did, you would have
designed an appropriate response. But you didn't, so you couldn't.
That's where I come in. My particular concern is "what can we do
about signals that we don't know anything about?"
The signals I'm looking at may be some of the same inputs you used for
specific functions, and they may also include your outputs, and
whatever miscellaneous operating signals are available internally in
your system... anything which might turn out to be usefully related to
anything you are doing.
I try to find out how they fit together, and generate new signals to
represent recognizable combinations which show up often enough that
they might be interesting.
It's up to you, and outside my domain, to use ordinary conditioning
mechanisms to incorporate my new outputs into your operation and do
something useful with them.
> Then you will search for syntactic
>patterns. Once you have discovered a syntactic pattern, you hope it
>will correspond to a real world pattern (which we might call a
>semantic pattern). It is not at all obvious that semantic patterns
>will show up as syntactic patterns. Whether they will presumably
>depends on the effectiveness of how you reduce the external world
>(semantics) to syntax. And we should expect that there can be
>syntactic patterns which do not arise from semantic patterns.
This is fluff and nonsense, words for the sake of words. There are
signals somehow related to or derived from real circumstances. As the
real circumstances change, the signals change. The only patterns are
the real patterns of change in the real circumstances... syntax has
nothing to do with it. Nobody is "telling me a story in symbols about
the real world"... I am simply watching the effects of the real world
on whatever signals I am able to monitor.
>In most of the cases I can think of, where we form symbolic
>representations of the external world, we first find patterns in the
>external world, and then use those patterns as part of our process of
>symbolization. So it is not clear to me how you can get your system
>started.
Again, nobody in this description is "forming symbolic
representations". We are processing real-world signals causally
connected to other real world signals. No symbols here.
***
Bill
mod...@concentric.net
Again, you choose to miss the point entirely. Certainly data is
selected or designed. Certainly biological systems evolve under
enviromental pressures. Absolutely true, and absolutely irrelevant
to the point under discussion.
For whatever presumably good and sufficient reasons, an organism has
certain signals occuring inside its neural structures.
These signals presumably are inputs and outputs of various evolved
or designed functions relevant to the goals of the system in which
they occur. IFF we were talking about those goals, or the design
of the base system itself, disputing the status of the data they
present would be relevent.
But we aren't. Instead, we are talking about a separable function
whose purpose is only indirectly related to that specific purposes
of the system. We are talking about a function whose purpose is to
find *potentially* useful relationships among signals of unknown
prior significance. Since *this* function cannot depend on prior
assumptions about the status or provenance of the data, it is irrelevant
to discuss such assumptions, no matter how important that may be for
other purposes.
[snip]
>I've said before that you are just turning a blind eye to
>assumptions. You have already settled on a coding system before
>you get to any clustering algorithm, and that's likely to shape
>your classification as much as the chunks it comes in.
No, the designer of the system (whether a human engineer or the
blind watchmaker) has settled on a coding system. Yes, the coding
will affect my classification. But it isn't my coding, and I have
to deal with what I'm given.
You are turning a blind eye to what I am saying, trying to shape it
to fit your preconceptions.
Bill Modlin
David Longley
> In article <689rvt$7...@ux.cs.niu.edu>, ric...@cs.niu.edu (Neil Rickert) wrote:
>
> > In <883444...@longley.demon.co.uk> Da...@longley.demon.co.uk (David
> Longley) writes:
> >
> > >Two things to consider:
> >
> > > 1) Man has been around quite some time, science has not.
> > > Unless our knowledge of natural selection is *way* of base
> > > it is highly unlikely that there is much "in our heads"
> > > which distinguishes us from our remote ancestors.
> >
> > Agreed.
>
> If by "science" you mean inquiry based on experiment, then science has
> been around a long time, perhaps longer than homo sapiens. Is it
> unreasonable to say that other species experiment? At any rate, the
> regular use of fire strikes me as likely to have been preceeded by quite a
> bit of expermintation (so does the use of metals). So, even if you mean
> "organized body of knowledge" by the term "science" I'm not sure that your
> statement is as clear as it seems.
>
> Sorry if this reply is out of context: my server seems to have missed the
> orginal post.
>
It *is* out of context. If you think we've had science in any
form for more than a few hundred years you are sadly misinformed.
David Longley
I have not missed your point at all.
If you spent a little time considering that you might make a bit
more progress. Literally thousands of man years have been spent
on the finer points of the philosophy pertaining to all of this
and I suggest you read some of it. I've suggested you read carnap
to see how the classic empiricist programme came to an end and
how we now have something more mature and realistic. We have no
choice but to accept that we are aboard Theseus' or Neurath's
boat. It isn't a metaphysical image either - it's a recognition
of a profound point that Wittgenstein made about the tyranny of
language.
>
> For whatever presumably good and sufficient reasons, an organism has
> certain signals occuring inside its neural structures.
They don't occur "inside" anything.
>
> These signals presumably are inputs and outputs of various evolved
> or designed functions relevant to the goals of the system in which
> they occur. IFF we were talking about those goals, or the design
> of the base system itself, disputing the status of the data they
> present would be relevent.
>
> But we aren't. Instead, we are talking about a separable function
> whose purpose is only indirectly related to that specific purposes
> of the system. We are talking about a function whose purpose is to
> find *potentially* useful relationships among signals of unknown
> prior significance.
Significance for WHAT????
> Since *this* function cannot depend on prior
> assumptions about the status or provenance of the data, it is irrelevant
> to discuss such assumptions, no matter how important that may be for
> other purposes.
>
>
> [snip]
>
> >I've said before that you are just turning a blind eye to
> >assumptions. You have already settled on a coding system before
> >you get to any clustering algorithm, and that's likely to shape
> >your classification as much as the chunks it comes in.
>
> No, the designer of the system (whether a human engineer or the
> blind watchmaker) has settled on a coding system. Yes, the coding
> will affect my classification. But it isn't my coding, and I have
> to deal with what I'm given.
>
> You are turning a blind eye to what I am saying, trying to shape it
> to fit your preconceptions.
>
> Bill Modlin
>
And what preconceptions are those? The body of scientific
knowledge I and everyone else has been taught? Perhaps you really
do believe you can bracket all that in some Husserlian epoche?
Say you could, with what would you talk about it? With what would
you communicate it? (Spend some time thinking about this).
mod...@concentric.net
Since it's my point, I'll be the judge of that. As you've yet to
say one word relevant to it, I repeat that you've missed it.
>If you spent a little time considering that you might make a bit
>more progress. Literally thousands of man years have been spent
>on the finer points of the philosophy pertaining to all of this
>and I suggest you read some of it. I've suggested you read carnap
>to see how the classic empiricist programme came to an end and
>how we now have something more mature and realistic. We have no
>choice but to accept that we are aboard Theseus' or Neurath's
>boat. It isn't a metaphysical image either - it's a recognition
>of a profound point that Wittgenstein made about the tyranny of
>language.
I'm making very good progress, thanks. My mature and realistic
theories are coming along nicely, and while Wittgenstein may have some
interesting things to say about language, I doubt that his authority
extends to the neural mechanisms underlying its development.
>>
>> For whatever presumably good and sufficient reasons, an organism has
>> certain signals occuring inside its neural structures.
>
>They don't occur "inside" anything.
Even coming from you, this astounds me. Neural structures are tissues
composed of neurons. The signals under discussion are those generated
by the firings of cells interior to those tissues. In what sense are
you imagining that they are not "inside" the structures?
Perhaps if you can explain this seemingly assinine statement I'll get
some clue as to what strange construction you are putting on my words?
>
>>
>> These signals presumably are inputs and outputs of various evolved
>> or designed functions relevant to the goals of the system in which
>> they occur. IFF we were talking about those goals, or the design
>> of the base system itself, disputing the status of the data they
>> present would be relevent.
>>
>> But we aren't. Instead, we are talking about a separable function
>> whose purpose is only indirectly related to that specific purposes
>> of the system. We are talking about a function whose purpose is to
>> find *potentially* useful relationships among signals of unknown
>> prior significance.
>
>Significance for WHAT????
Can you read? I said that the signals are of unknown significance.
That means we don't know what they might be significant to or for.
There are many signals which have hardwired significance: some cause
muscles to contract, some cause sucking actions or salivation, etc.
Other signals acquire significance through association with signals
of pre-defined significance, via the mechanisms of classical conditioning.
But there are millions more which occur with no hardwired consequence,
and which furthermore are not sufficiently well corellated with
signals of previously determined significance to acquire significance
by conditioning.
In a simple organism without much excess cortical tissue, any
information which might be gleaned from these signals is lost.
However, higher organisms have evolved excess tissue in which "free
associations" are formed, generating new signals as functions of
correlations among available signals. This turns out to be a useful
activity, since eventually many of these derivative signals do
correspond to conditions which can be associated with significant
signals, and thus may acquire significance of their own.
The basic mechanism for this free association is the same as that
of ordinary conditioning. However, those mechanisms as described
by Rescorla-Wagner are unstable in deep network configurations.
So I've had to postulate some slight modifications to the rules
to account for the activity in large masses of tissue with signal
paths ranging to hundreds of cells without close architectural
constraints on function. It's those modifications that I've been
trying to get someone to talk about...
>> Since *this* function cannot depend on prior
>> assumptions about the status or provenance of the data, it is irrelevant
>> to discuss such assumptions, no matter how important that may be for
>> other purposes.
>>
>>
>> [snip]
>>
>> >I've said before that you are just turning a blind eye to
>> >assumptions. You have already settled on a coding system before
>> >you get to any clustering algorithm, and that's likely to shape
>> >your classification as much as the chunks it comes in.
>>
>> No, the designer of the system (whether a human engineer or the
>> blind watchmaker) has settled on a coding system. Yes, the coding
>> will affect my classification. But it isn't my coding, and I have
>> to deal with what I'm given.
>>
>> You are turning a blind eye to what I am saying, trying to shape it
>> to fit your preconceptions.
>>
>> Bill Modlin
>And what preconceptions are those? The body of scientific
>knowledge I and everyone else has been taught? Perhaps you really
>do believe you can bracket all that in some Husserlian epoche?
>Say you could, with what would you talk about it? With what would
>you communicate it? (Spend some time thinking about this).
Nothing I've said is in conflict in any way with any scientific
knowledge, it is quite consistent with everything currently known
about the way our brains work. It is at odds only with your distorted
preconception of what I *must* be talking about, which prevents you
from hearing what is actually said.
I am talking about the mechanisms in our brains whereby we are enabled
to learn of the world around us. I have no idea what you mean by
bracketing "all that" in some Husserlian epoche. I am attempting
to talk to the members of this group about those mechanisms, using
words. What is it I am supposed to spend time thinking about?
Bill Modlin
Hans Moravec
bsharvy:
> People need to stop conflating computation with intelligence.
> It may turn out that they are identical (I doubt it), but
> even so, that needs to be the *conclusion* of an argument,
> not a premise. Hans Moravec, in particular, persists in not
> distinguishing between the two.
Here's my argument. Playing chess at the level of
Kasparov (or proving hard theorems, or driving safely
down the road) requires, and is a demonstration of
intelligence. Computers have achieved these things.
Thus computers (and computation) has exhibited intelligence.
Works for me.
> Speed would determine whether your AI's mind worked quickly
> or slowly: it would not be relevant to whether your wanna-be
> AI had a mind. Deep Blue can be no more "intelligent" than a
> computer with a MIPS rating from the 70's with an equivalent
> algorithm.
In practice, memory grows with speed: about 1 Mbyte per MIPS
in mature computers. Most of the gains in my own 3D
perception work have been because the programs can now keep
around much more partial information to weigh with later
evidence. My current programs work with 100 Mbytes, and
a dense probabilistic description of a volume of space,
compared to 1 Mbyte and a sparse list of nearby features
that I had in the 1970s.
More abstractly, speed is a measure of the state space a machine can
explore in practice. AI programs rarely run more
than a few hours, so something they can't do in that time
they can't do at all. The number of states a 100 MIPS
program can reach is the 100th power of the number of states
a 1 MIPS program can reach. That's why faster computers also
need more memory.
Speed translates directly into runtime complexity, and thus
to intelligence.
David Longley
In article <68cm7f$d...@examiner.concentric.net> mod...@concentric.net writes:
<snip>
>
> I'm making very good progress, thanks. My mature and realistic
> theories are coming along nicely, and while Wittgenstein may have some
> interesting things to say about language, I doubt that his authority
> extends to the neural mechanisms underlying its development.
>
If you're just cobbling together ideas which take your fancy,
that's hardly scientific research - where is the empirical
anchoring?
<snip>
>
> >And what preconceptions are those? The body of scientific
> >knowledge I and everyone else has been taught? Perhaps you really
> >do believe you can bracket all that in some Husserlian epoche?
> >Say you could, with what would you talk about it? With what would
> >you communicate it? (Spend some time thinking about this).
>
> Nothing I've said is in conflict in any way with any scientific
> knowledge, it is quite consistent with everything currently known
> about the way our brains work. It is at odds only with your distorted
> preconception of what I *must* be talking about, which prevents you
> from hearing what is actually said.
>
> I am talking about the mechanisms in our brains whereby we are enabled
> to learn of the world around us. I have no idea what you mean by
> bracketing "all that" in some Husserlian epoche. I am attempting
> to talk to the members of this group about those mechanisms, using
> words. What is it I am supposed to spend time thinking about?
>
> Bill Modlin
>
>
Like Savain (who seems to have read a few popular neuroscience
articles) you now seem to include expertise in neuroscience
amongst your technical skills. Tell me, just what empirical
research constrains your imaginative ideas on how "the brain"
works?
In the behavioural domain, what empirical evidence are you
implicitly referring to when you say "those mechanisms as
described by Rescorla-Wagner are unstable in deep network
configurations"?
Is this anything but armchair neuroscience? It certainly doesn't
seem to be very receptive to suggestions that it be better
philosophically informed, which makes me wonder why you post it
to comp.ai.philosophy.
JRStern
On Tue, 30 Dec 1997 01:10:12 -0500, Hans Moravec <h...@cmu.edu> wrote:
> Deep Blue, in a first for machinekind, won the first game of the 1996
>match.
A silly anthropomorphism. Did the steam hammer win a great victory
for steam-hammerkind against John Henry? Does my car win a great
victory over Nike-kind whenever I drive instead of jog?
>Kasparov, on the other hand, claims to see into opponents' minds during
>play, intuiting and exploiting their plans, insights and oversights. In
>all other chess computers, he reports a mechanical predictability
>stemming from their undiscriminating but limited lookahead, and absence
>of long-term strategy. In Deep Blue, to his consternation, he saw
>instead an "alien intelligence."
He was a superstitious amateur.
I have to do this all the time, reverse-engineering some software
product by imagining the strategies and mistakes of its developers.
Thousands of programmers do this (but not millions, it's sort of an
advanced technique).
Sure, Dennett calls this "the intentional stance", it's the same
strategy we use with other people. But, here I am, having to do it
with blatently mechanical and stupid programs. Thus, my stance is not
a good indicator of the degree of intelligence I am dealing with. Or,
alternatively, and equally validly, Deep Blue is no more intelligent
than a C++ compiler or database engine.
There is an element of importance here, in that both people with minds
and trivially programmed objects can be treated as agents. The point
is important because it shows how *little* we can tell about the
agent, from the mere fact of its being treated as an agent. Do we
even know the criteria for agency, or is it strictly attributional, in
the eye of the beholder? These are open questions.
Kasparov's judgement was shockingly bad in this matter. He became so
rattled that he completely blew the last game. In the weeks
following, he made some slightly more temperate comments about the
experience. He apparently had frightfully bad advice about computers
and computer chess, he could get better advice on this newsgroup, or
better yet on the chess newsgroups (which I perused after the match).
Less talented players than Kasparov have apparently come to terms with
the play of good commercial programs, without embuing them with
magical powers, and were able to deal with Deep Blue in the same
manner. Kasparov did not. I do not think it was a case of "could
not". Are you even familiar with modern chess programs, and their
users, other than Deep Blue and Kasparov?
I hope Kasparov gets over his superstitious attitudes. Consensus on
the chess groups seems to have been that Kasparov would have killed a
human player making the same moves that Deep Blue made. The
situation, and bad preparation, caused him to lose, embarrasingly. If
a better prepared Kasparov makes these same kinds of attributional
statements a few years from now, they might be worth taking seriously.
However, I would guess he'll learn better, maybe already has.
> Who better to judge than Garry Kasparov? Mathematicians
>who examined EQP's proof of the Robbins conjecture, mentioned earlier,
>report a similar impression of creativity and intelligence. In both
>cases, the evidence for an intelligent mind lies in the machine's
>performance, not its makeup.
No, it lies in the naive and mistaken judgements of non-specialists in
computation.
> Now, the team that built Deep Blue claim no "intelligence" in it,
Theirs *is* an educated judgment, and I'll go with it.
> Deep Blue's creators know its quantitative superiority over other chess
>machines intimately, but lack the chess understanding to share
>Kasparov's deep appreciation of the difference in the quality of its
>play.
This is exactly backwards, imho. They appreciate its play more than
Kasparov, in several senses.
> I think this dichotomy will show up increasingly in coming
>years. Engineers who know the mechanism of advanced robots most
>intimately will be the last to admit they have real minds.
Say rather, that those who know the tricks, are harder to fool.
> A human brain, too, does not exhibit the
>intelligence under a neurobiologist's microscope that it does
>participating in a lively conversation.
They do not (yet) know where to look, though much progress is
apparently being made. If you stick probes into a workstation and do
PET scans, you'd have trouble seeing its much simpler functionality.
>I see no fundamental difference between search-intensive,
>knowledge-intensive, learning-intensive or memory-intensive approaches.
Hans, you know better. Deep Blue has become the archetype of the case
where brute force went only so far, until this year's heuristics
finally turned the trick (to mix metaphors). For twenty years, yes,
brute force beat the heuristics. Maybe there is a task-specific
minimum of processing power required, even with optimal heuristics.
We don't know enough to judge this yet, we have too few expert-level
programs to judge from.
> But, a computer with molecular memory a million times bigger
>than Deep Blue's ...
Why molecular, whatever that's supposed to mean? A computer with a
Penrose memory could search a full enumeration of all possible games
in a femtosecond. Try writing your research proposals based on that.
On second thought, please don't, they sometimes get funded, and it's
my tax dollars at work.
>Kasparov, with his 100 billion millisecond-slow neurons, with
>100 trillion synapses tuned by a lifetime of thinking about
>chess, probably simultaneously weighs a few million possible
>move sequences subconsciously in terms of a few million
>learned characteristics. After all that subconscious
>selection, a few quality moves are elevated into his
>consciousness awareness for final execution.
And when I look out on the lawn, I weigh a few million animal images,
before deciding I'm looking at a blue jay? Well, that's one
explanation. Others spring quickly to mind. How do your autonomous
vehicles find their way, by matching to a few million possible scenes?
I think not.
>I am smugly confident that full intelligence will be achieved in similar
>modest steps, as the requisite 100 million MIPS or so becomes available
>over the next several decades.
I am speculatively confident that after the fact, we'll find that with
the right paradigm and algorithms, we can instantiate full
intelligence in under the 100 MIPS we have today.
Joshua Stern
JRS...@gte.net
JRStern
On 30 Dec 1997 06:51:38 -0600, ric...@cs.niu.edu (Neil Rickert)
wrote:
>In another post you pointed out that a priori knowledge plays a large
>role for Pinker. All the more reason for me to think he is on the
>wrong track.
It's coloring the matter somewhat to call the a priori "knowledge", as
it can also indicate things as simple as the mechanical linkages
necessary to get optical images into the brain, feature scanned, etc.
Pinker quotes (well, refers to very tersely) a great deal of cognitive
psychology experiments done since I was in school. Apparently, there
is strong evidence that people perceive and reason differently,
depending on whether they believe they are dealing with agents, tools,
or things. These may be a priori categories in the human mind. This
does not say they are ontologically necessary or real, just what
happens to be present in one kind of advanced ape.
The structure of a particular Turing machine is an a priori limitation
on the details of programs it runs. But, it's not much of a
limitation, because other Turing machines could run equivalent
programs, if only by simulating the first machine. Similarly, a
priori does not *have* to be a limiting factor, and we are relatively
free to assign a priori structures, just so we are flexible about them
a postiori. Kant did not know about the equivalence of Turing
machines. We should.
Meanwhile, Happy New Year, to you and all our c.a.p. cohorts.
Joshua Stern
JRS...@gte.net
JRStern
On Tue, 30 Dec 1997 13:57:35 GMT, cho...@idnsi.net (Chris Hooley)
wrote:
>I don't understand this. You speak as though Savain's and Modlin's engines
>were the dominant view rather than minority voices, as though solipsystem
>computer logic was established doctrine.
Every attempt at AI based on a fixed rule base is solipsistic. This
includes most everything called "expert systems" in the 1980s.
>To my limited knowledge, only during the last year has this newsgroup
>seriously entertained the idea that robust AI might emerge without *any*
>human-dependent constructs about "the world."
It's just behaviorism/reductionism writ large. That goes back a
hundred years. It can also be looked at as eliminativism, that goes
back at least to ancient Greece, or as essentialism, that goes back
even further. If this newsgroup had ten thousand years of retention,
you'd see that.
>Why do you speak of kids not sitting quietly for "solipsistic crap?"
Because "expert systems" are still (sometimes) preached in the late
1990s, and I was perhaps a little too terse in my pejorative
reference.
>Within twenty years, in my opinion, self-programmed AI is going to be so
>clearly intelligent that none of our opinions on the matter will make much
>difference.
Pinker goes on about how modern education encourages kids to
self-program in math and reading. He doesn't like the idea. Neither
do I. Maybe someday we'll have AI that self-programs like humans, but
one can debate to what degree humans self-program. Why hang about in
school for sixteen years or more, if we self-program? Why not sit
quietly and wait for the emergence?
> Perhaps, as several, including Seth, have proposed, the
>Internet itself will become the parent/global "body" of a 21st Century AI.
Gibson, "Neuromancer" and "Count Zero". And that's about what it's
good for.
Joshua Stern
JRS...@gte.net
JRStern
On 30 Dec 1997 14:40:20 GMT, ri...@pangea.Stanford.EDU (Rick Ottolini)
wrote:
>Twenty years from now today's computers will seem very slow
>and unbelievably cumbersome.
They already seem unbelievably cumbersome.
Josh
JRStern
On 30 Dec 1997 12:29:51 -0600, ric...@cs.niu.edu (Neil Rickert)
wrote:
>To be fair to Josh,
I'm all for that.
> I introduced the term 'solipsism' into the
>discussion. Josh just continued the point. I'll grant that
>solipsism is rejected by most people. In introducing the term, I was
>being critical of philosophy.
I thought you were being critical of solipsistic systems.
On 12/29 @ 12:35 you said:
>>>One of the problems is that the availability of computers
>>>tends to encourage the basically solipsistic approach of emphasizing
>>>internal structure (algorithms, etc), and ignoring the external
>>>relation between the individual and the world.
OK, the approach is not the system. I focused on the system. I agree
solipsism is ineffective, probably inherently so, maybe tautologically
so, yet much of AI in the 1980s, and before, and today, was based on
the solipsism of fixed rule sets.
> In essence I was saying that
>philosophy is implicitly solipsistic, even though it explicitly
>rejects solipsism. That is to say, philosophy is based on internal
>standards and is largely immune to external evidence.
Huh? Philosophy has multiple domains that are both about interaction
and created by interaction, from causality to morality.
> Similarly, at
>other times I have described the Turing machine as solipsistic, since
>it starts out with complete internal information (on its tape), and
>does not have any interaction with the world.
Right, that's what I thought you were talking about, and is what I
replied about.
Joshua Stern
JRS...@gte.net
Neil Rickert
In <68e10n$kuv$2...@gte2.gte.net> JRS...@gte.net (JRStern) writes:
>On 30 Dec 1997 06:51:38 -0600, ric...@cs.niu.edu (Neil Rickert)
>wrote:
>>In another post you pointed out that a priori knowledge plays a large
>>role for Pinker. All the more reason for me to think he is on the
>>wrong track.
>Pinker quotes (well, refers to very tersely) a great deal of cognitive
>psychology experiments done since I was in school. Apparently, there
>is strong evidence that people perceive and reason differently,
>depending on whether they believe they are dealing with agents, tools,
>or things. These may be a priori categories in the human mind. This
>does not say they are ontologically necessary or real, just what
>happens to be present in one kind of advanced ape.
Yet evolutionary psychology sees no reason to give a plausible
scenario that would allow tools to evolve as an a priori category in
a species most of whose history has been as a simple
hunter-gatherer.
>The structure of a particular Turing machine is an a priori limitation
>on the details of programs it runs. But, it's not much of a
>limitation, because other Turing machines could run equivalent
>programs, if only by simulating the first machine.
My objection to a priori knowledge is not that it is a limitation.
Rather, my objection is that it is an ad hoc assumption used to
repair a failed theory of cognition. Worse still, little is done in
the way of explaining how evolutionary processes could solve a
problem that postulated cognitive processes are unable to solve. If
evolutionary processes are so powerful, why not postulate that
cognitive processes are similar and can solve similar problems?
Hans Moravec
JRStern:
> And when I look out on the lawn, I weigh a few million animal
> images, before deciding I'm looking at a blue jay? Well,
> that's one explanation. Others spring quickly to mind. How
> do your autonomous vehicles find their way, by matching to a
> few million possible scenes?
> I think not.
Actually, yes. The RALPH road following system described
in my post 1 of 3 explicitly remembers thousands of road
cross-sections (each just a 32-element vector) and compares
every one all to the current scene, moment to moment
choosing the one that best matches. If it doesn't get a
good enough match, it adds the current input to its
table, allowing it to adapt to changing road types. If it
had 50,000 MIPS instead of 50 MIPS, it could handle
millions, and would work even better.
Early vision programs were brittle primarily because they
handled too small a hypothesis space. When aspects
of the scene fit their list of possibilities they worked
well. When not, they behaved nonsensically. Today's
programs have bigger hypothesis spaces, and so fail
less often. It takes memory and MIPS to represent the
space and weigh evidence for increasingly many alternative
hypotheses.
RALPH works best at road following of anything tried so far.
The previous champion was ALVINN, which was a big neural
net that mapped image inputs to steering outputs. With
its thousands of weights, a trained ALVINN net can be viewed
as a sloppier and more restrictive encoding of a large
library of road images than RALPH. Computationally, it was
actually more expensive.
To get around the restrictiveness, i.e. inability of an
ALVINN net to simultaneously remember a large diversity of
road types, in its final years, before being supplanted by
RALPH, the ALVINN-based vans maintained several dozen
different neural nets, each trained for a different category
of road. Each net was fed the same image input, and the
suitability of each rated by the shape of its steering
output vector, which had been trained to show a nice bell
curve centered on the correct steering position. The net
showing most nearly a correct bell curve at each moment
was assumed to be the one best trained for this kind of
road, and its output was selected to control the van right
then.
...
I don't know exactly how it's neurally implemented
(but surely mostly in parallel given the number of
connections), but I'm pretty sure most of my own mental
discriminations are also made by subconsciously weighing
available sense (and other) evidence against a very
large list of possible hypotheses. That's the basic
Bayesian framework, and it is starting to work pretty well
in robotics.
Hans Moravec CMU Robotics http://www.frc.ri.cmu.edu/~hpm
Neil Rickert
In <68e10p$kuv$3...@gte2.gte.net> JRS...@gte.net (JRStern) writes:
>On Tue, 30 Dec 1997 13:57:35 GMT, cho...@idnsi.net (Chris Hooley)
>wrote:
>Because "expert systems" are still (sometimes) preached in the late
>1990s, and I was perhaps a little too terse in my pejorative
>reference.
There is nothing wrong with expert systems, except that they are
mis-advertized and their name is a misnomer. I recall an advocate of
expert systems at a colloquium about 10 years ago preaching that
computers had automated clerical work, and now they were automating
intellectual work. With my typical cynicism, I suggested that they
were merely automating the clerical aspects of intellectual work.
>>Within twenty years, in my opinion, self-programmed AI is going to be so
>>clearly intelligent that none of our opinions on the matter will make much
>>difference.
>Pinker goes on about how modern education encourages kids to
>self-program in math and reading. He doesn't like the idea. Neither
>do I.
I agree. But you are conflating two different meanings of
self-programming. Even in the strictest of back-to-basic curricula,
nobody is getting in and directly modifying neural connections. So
it is all self programming in the sense in which Hooley is using the
term.
Neil Rickert
In <68e10t$kuv$5...@gte2.gte.net> JRS...@gte.net (JRStern) writes:
>On 30 Dec 1997 12:29:51 -0600, ric...@cs.niu.edu (Neil Rickert)
>> In essence I was saying that
>>philosophy is implicitly solipsistic, even though it explicitly
>>rejects solipsism. That is to say, philosophy is based on internal
>>standards and is largely immune to external evidence.
>Huh? Philosophy has multiple domains that are both about interaction
>and created by interaction, from causality to morality.
I can't say that I see much evidence of this. Kuhn, in his 1962
book, brought evidence that refuted much of what had been assumed by
scientific epistemology. But recent books on philosophy of science
seem to be moving in the direction of explaining away the problems
that Kuhn raised, and inching ever closer to the position that was
shown by Kuhn not to correspond to how science is actually
practiced.
So it seems to me that philosophy is mostly a closed system of "Just
So" stories, largely immune to external evidence.
Rick Ottolini
>>Within twenty years, in my opinion, self-programmed AI is going to be so
>>clearly intelligent that none of our opinions on the matter will make much
>>difference.
People were making similar predictions in the 1950s when computer A.I.
first started.
For an interesting look at computer predictions look at the
Association for Computing Machinery's 50th Anniversary retrospective
"Computing: the Next 50 Years".
Hans Moravec
JRStern:
> Pinker quotes (well, refers to very tersely) a great deal
> of cognitive psychology experiments done since I was in
> school. Apparently, there is strong evidence that people
> perceive and reason differently, depending on whether they
> believe they are dealing with agents, tools, or things.
> These may be a priori categories in the human mind.
I think this is exactly why the idea of AI (or that
existing programs are exhibiting intelligence) gets such
intense and and emotional opposition, even in this
newsgroup, which is surely self-selected to be more
tolerant than average to the idea. AI trashes a natural
classification boundary in the human mind. Is a robot a
thing, a tool or an agent? The instincts say that if it is
one of those it can't be another. Engineers shape things
into tools. Things have weight, texture, taste. Tools
have purpose, functions, value. Agents have intentions,
feelings, debts. Treating a tool as an agent strikes many
as a mental violation. bsharvey's responses to my posts
on this thread are clear expressions of this feeling.
(Kasparov and the Deep Blue team can't BOTH be right!
It either has a mind or it doesn't! If it's a thing
it can't be an agent! Robots are tools, they can't
participate in societies!).
---
It's mostly forgotten now, but those of us interested
in the emergence of space travel early in this century
recognize similar elements in the debate about that
idea. Serious proposals for space travel were first
presented by Tsiolkovsky in Russia, Goddard in the US and
Oberth in Germany in the first decades of this century.
They were attacked by the academic community and the
popular press with the same knd of vehemence that often
greets AI. I think there is also a dichotomy in the human
mind between mechanisms for dealing with terrestrial events
and celestial ones. The suggestion of crossing the divide
triggers a sense of ridiculousness and even repugnance
or blasphemy, often inspiring unfounded objections that
seem to be mainly rationalizations of the aversion to the
idea. The same dichotomy probably prevented astronomers
from accepting the idea of meteorites until the early
1800s, despite a large amount of evidence for them.
(Rocks don't fall from the sky, you silly person.) It
took a giant incontrovertible shower of meteors falling
into a French field arond 1830 to finally turn the tide
for rocks from the sky. Accepting people into space took
longer.
Willy Ley, early German rocket enthusiasts and well known
popularizer of space travel, collected some of the copious
criticisms that followed a totally accurate 1923 popular
book "The Rocket into Interplanetary Space" by Oberth that,
in a very real sense, defined (via Werner von Braun) the
Apollo moon flights fifty years later.
Chapter 5 of Ley's "Rockets, Missiles and Men in Space"
summarizes early responses to Oberth's book:
... well reputed astronomers ... simply killed the idea ...
by stating that these things are very nice and interesting
but lacking in foundation since everybody knows there can be
no recoil in interplanetary space ... Another critic ...
added the idea of manned rockets was preposterous for all
time to come because people, as soon as they left the
atmosphere of earth (impossible anyway) would be subjected to
the gravity of the sun which is powerful enough to squash
their bodies. ... an aviation expert ... could not understand
why the exhaust gases should follow the rocket if the latter,
after some time, surpassed its own exhaust velocity. ... a
physicist ... said the rocket, of necessity, could not surpass
the velocity of its exhaust gases because its efficiency would
surpass 100% ... obviously impossible. ... a mathematician and
physicist published that even the most powerful explosive known
could not lift its own weight to a greater height than 400 km
...
Other reactions to space rocketry:
New York Times editorial January 13, 1920, commenting on
Robert Goddard's enthusiasm to reach Mars:
That Professor Goddard and his 'chair' at Clark College and
the countenancing of the Smithsonian Institution does not
know the relation of action to reaction and of the need to
have something better than vacuum against which to react--to
say that would be absurd. Of course, he seems only to lack
the knowledge ladled out daily in high schools ...
Dr. Vannevar Bush, Senate testimony, 1945:
In my opinion such a thing is impossible ... People have been
talking about a 3,000 mile high-angle rocket shot from one
continent to another, carrying an atomic bomb and so directed
as to be a precise weapon ... I think we can leave that out
of our thinking.
Stanley Friesen
On 30 Dec 1997 12:19:47 -0600, in comp.ai.philosophy
ric...@cs.niu.edu (Neil Rickert) wrote:
>In <68ah9k$7...@examiner.concentric.net> mod...@concentric.net writes:
>>I don't know what you mean by "pre"-represented raw input. All input
>>at all levels is just signals, caused by whatever causes it, and
>>meaningless until we find out what it means by discovering how it
>>behaves, what else it tells us to expect.
>
>It is signals until we generate symbols (or digits, or other discrete
>units). Thereafter it is symbolic representations.
>
It seems to me the two of you are using different definitions of
"symbol".
Would you say that a digitized image is a symbol, or composed of
symbols? I think Modlin would say it is not, that it is a digital
encoding of the raw signal.
Certainly, by your definition, it seems you would class the output of
the rods and cones in the eye as symbols not signals, as they are most
certainly encodings of the light levels using discrete units (the
neural impulses, which are either there or not - sort of binary code).
If your definition of symbol includes the output of rods and cones,
then it is irrelevent, as all known natural cybernetic systems perform
that translation as the first step towards processing it.
Louis Savain
In article <68e10p$kuv$3...@gte2.gte.net>, JRS...@gte.net (JRStern)
wrote:
>On Tue, 30 Dec 1997 13:57:35 GMT, cho...@idnsi.net (Chris Hooley)
>wrote:
>>I don't understand this. You speak as though Savain's and Modlin's engines
>>were the dominant view rather than minority voices, as though solipsystem
>>computer logic was established doctrine.
>
>Every attempt at AI based on a fixed rule base is solipsistic. This
>includes most everything called "expert systems" in the 1980s.
I don't get it. How can expert systems, and systems based on
production rules (fixed or not) be put into the solipsistic category?
It's all knowledge engineering, is it not? Expert system rules are
not normally created or generated by the system in true solipsistic
fashion, but are engineered from the outside by the programmer or
knowledge engineer. I don't understand how anyone can see anything
solipsistic in that.
>[...]
>>Within twenty years, in my opinion, self-programmed AI is going to be so
>>clearly intelligent that none of our opinions on the matter will make much
>>difference.
>
>Pinker goes on about how modern education encourages kids to
>self-program in math and reading. He doesn't like the idea. Neither
>do I. Maybe someday we'll have AI that self-programs like humans, but
>one can debate to what degree humans self-program. Why hang about in
>school for sixteen years or more, if we self-program? Why not sit
>quietly and wait for the emergence?
Are you saying that learning from information which originates from
a teacher is not self-programming? I would like to think that there
is a clear difference between teaching and programming. Come to think
about it, is not everything in the environment a teacher in the sense
that we get all of our sensory information from the environment? I
think that Chris Hooley had something else in mind when he wrote
"self-programming." I don't think he was referring to
"self-teaching."
I think that the environment has no power to program our neurons. I
think our neurons use information in the signals they receive from our
sensory apparatus to program themselves for various temporal
recognition tasks. That, to me, is self-programming. It does not
matter where the signals come from, as long as they have informational
content. By this, I mean that they must have correlatable properties.
Ben Sharvy
> bsharvy:
> > People need to stop conflating computation with intelligence.
> > It may turn out that they are identical (I doubt it), but
> > even so, that needs to be the *conclusion* of an argument,
> > not a premise.
> Here's my argument. Playing chess at the level of
> Kasparov (or proving hard theorems, or driving safely
> down the road) requires, and is a demonstration of
> intelligence. Computers have achieved these things.
> Thus computers (and computation) has exhibited intelligence.
Why do they require intelligence? Note that it is circular to reply
"Because they require computation." Your answer should not imply that
multiplying two numbers together requires intelligence, unless you want to
claim calculators are intelligent.
This topic requires finer semantic distinctions than this newsgroup
typically exhibits. When talking about people, we call exceptional
computational ability a sign of intelligence. By that, we do not mean that
it is a sign of having a mind, but merely that the person has good
math/logic skills. It is a mistake to then conclude *in the context of AI
philosophy* that computation is equivalent to intelligence. The
"intelligence" in "artificial intelligence" refers to having a mind (at
least philosphically; "AI" is also used to mean any decision-making aid,
such as stock investment programs).
There are of course many other kinds of skills people may have, such as
poetry or counseling, most of which strike me as being more indicative of
the existence of mind than good computation skills.
Your choice of Kasparov is open to the charge of arbitrariness. Why the
greatest player at this date, and not some other? And why chess, rather
than blackjack which would set a much earlier date for the arrival of AI,
or the game of Go, which would set a much later date? And why the emphasis
on game playing rather than poetry writing?
> > Speed would determine whether your AI's mind worked quickly
> > or slowly: it would not be relevant to whether your wanna-be
> > AI had a mind. Deep Blue can be no more "intelligent" than a
> > computer with a MIPS rating from the 70's with an equivalent
> > algorithm.
>
[answer I didn't fully undersatnd snipped]
>
> More abstractly, speed is a measure of the state space a machine can
> explore in practice. AI programs rarely run more
> than a few hours, so something they can't do in that time
> they can't do at all.
It doesn't follow that programs can't do at all what they can't do in the
amount of time humans usually allow them. We're talking about capability.
> Speed translates directly into runtime complexity, and thus
> to intelligence.
Runtime complexity per unit of time, you mean.
Neil Rickert
In <34aaac23....@pop.americasttv.com> sfri...@americasttv.com (Stanley Friesen) writes:
>On 30 Dec 1997 12:19:47 -0600, in comp.ai.philosophy
>ric...@cs.niu.edu (Neil Rickert) wrote:
>>It is signals until we generate symbols (or digits, or other discrete
>>units). Thereafter it is symbolic representations.
>It seems to me the two of you are using different definitions of
>"symbol".
Probably so.
>Would you say that a digitized image is a symbol, or composed of
>symbols? I think Modlin would say it is not, that it is a digital
>encoding of the raw signal.
As I was using the term, I would say that it is composed of symbols.
>Certainly, by your definition, it seems you would class the output of
>the rods and cones in the eye as symbols not signals, as they are most
>certainly encodings of the light levels using discrete units (the
>neural impulses, which are either there or not - sort of binary code).
I don't know enough about rods and cones to be sure this is correct.
I'm not sure that enough is known.
>If your definition of symbol includes the output of rods and cones,
>then it is irrelevent, as all known natural cybernetic systems perform
>that translation as the first step towards processing it.
Surely that is not the first step. Before the cones and rods do
anything, the eye moves in saccades. These eye movements have to be
considered part of the sensory detection. After that, what shows up
in the cone and rod output still depends on the sensitivity of the
cones and rods, and that is modified by chemical processes. And
don't forget the role of the iris in this. So even if the output of
rods and cones is properly considered to be symbolic (as I use the
term), there is a whole lot of processing going on which controls how
this symbolization is to be done. Modlin wants to ignore all of that
pre-symbolization processing as something fixed, while I consider it
potentially quite important.
Hans Moravec
hpm:
> > Here's my argument. Playing chess at the level of
> > Kasparov (or proving hard theorems, or driving safely
> > down the road) requires, and is a demonstration of
> > intelligence. Computers have achieved these things.
> > Thus computers (and computation) has exhibited intelligence
bsharvey:
> Why do they require intelligence?
Intelligence is an attribution I make on behaviors.
It is by behavior that I distinguish smart humans from
dumb potted plants. Quality chess, proving and driving are
manifestations of human intelligence. Animals and morons
can't do them. When a human exhibits such skills, I say
they're pretty intelligent in that area. I offer the same
courtesy to machines.
You, on the other hand, try to gerrymander the meaning
of intelligence so as to exclude machines. It's a
moving boundary, as machines do one thing after another that
once only intelligent humans could do.
> Note that it is circular to reply
> "Because they require computation." Your answer should
> not imply that multiplying two numbers together requires
> intelligence, unless you want to
> claim calculators are intelligent.
I think calculators have an intelligence for arithmetic.
In overall power, it is less than one trillionth of full
human brainpower, as suggested in my chart:
http://www.frc.ri.cmu.edu/~hpm/book97/ch3/All.things.075.jpg
However, humans happen to be very inefficient (stupid)
about arithmetic, so it doesn't take much directed
intelligence to beat them at it.
> Your choice of Kasparov is open to the charge of
> arbitrariness. Why the greatest player at this date, and
> not some other? And why chess, rather than blackjack
> which would set a much earlier date for the arrival of AI,
> or the game of Go, which would set a much later date? And
> why the emphasis on game playing rather than poetry writing?
I gave extensive material on three examples: robot navigation,
because it's my area, and Deep Blue and the Robbins proof
because they were recent newsworthy milestones.
I see the development of AI as a rough recapitulation of the
evolution of natural intelligence. I think flatworms have
some intelligence about food, reproduction and threats, and
calculators have some about arithmetic. 100 MIPS, properly
used, allows a machine to exhibit the overall intelligence
of a bug. 3 million MIPS, in the Deep Blue configuration,
gives 1 Kasparov worth of chess intelligence (chess also
a task humans don't do that efficiently), but nothing
else. As computer power grows, it will be able to match
the overall intelligence of increasingly larger nervous
systems. At 100 million MIPS, by my reckoning, it should
be able to match general human intelligence even in those
survival areas where human brains are very efficient, such
as percieving, moving and socially interacting.
Human-level AI did not arrive with Deep Blue. It's been on
its way since at least the beginning of the century with
the invention of telegraphy and electronics, and still has
several decades to go.
What did happened in the last few years, though, is that,
increasing computer power lifted it from the wormlike
overall intelligence rating it had languished in for decades
to the high end of the insect range. And it is now
approaching the scale of very small vertebrates.
> It doesn't follow that programs can't do at all what they
> can't do in the amount of time humans usually allow them.
> We're talking about capability.
The program that drove my Cart across a room in 1979 took
five hours of running time. I guarantee you it was not
possible to let it run for a month. Our computer crashed
more often than that, even if the other users had tolerated
it. Also, the machine had less than one megabyte of memory,
so the 100 Mbyte data structures I use today could not
possibly fit. (even on our disks!)
Not only did the old computers have vastly insufficient
memories to host today's programs, they were vastly too
slow and expensive to provide the trial and error necessary
to find the effective techniques that make today's programs
work.
Processing speed and size is the engine that powers
intelligent programs, just as nervous system size is the
engine that powers biological intelligence. You can't
expect a bug to do a man's thinking just by giving it more
time. A rubber band engine won't fly a 747.
Hans Moravec
> Just brute-force matching, no classification, indexing?
> The brain *might* get away with that, because of its
> architecture, but it might also do some hierarchical or
> weighted classifications first, like to get down to a
> 32 element vector. Is that still "comparing images"?
> And even if so, do we assume all cognitive functions
> like chess, are likely to work the same way?
> Josh
Surely there's more than one trick in there. I never
said it was images that were being compared (though
there is evidence that bees do just that, to locate
themselves in their home territory). I would
guess that various abstractions are used, themselves
produced by weighing many hypotheses at a simpler
level - like the motion, edge and orientation detections
one finds in the early visual system. (An orientation
detector weighs the multiple hypotheses that there might
be an edge pointed THIS way, or THAT way, or THIS OTHER
way ... against the sensory input). But I think it is
probable that Kasparov has millions of chess positions
stored away in some encoding. He must be laying down
memory when he thinks about chess day after day, year
after year. Expanding the hypothesis space, adjusting
the significance and move probabilities ... ?
The Bayesian framework, of multiple hypotheses whose
relative probability is adjusted by testing each for
compatibility with incoming data, is a very powerful and
general one, especially when the number of hypotheses
can be huge. And I can imagine how it could be set up
in a large regular array, perhaps looking something like
the cortical sheet ...
RALPH works with a road appearance hypothesis space,
my 3D mapping code has spatial occupancy hypotheses,
other vision programs weigh alternative edge object
configurations. The Xavier and Amelia robots trundling
around another building on on our campus (look for their
interactive web page!) weigh evidence for alternate
hypotheses of their location (Am I here, or there, or
there, ...).
In a few especially regular cases, the large hypothesis
spaces may be compressible by some mathematical shortcut.
Orientation detection benefits a little that way, being
able to use shared calculations for some alternative
hypotheses. But the more interesting higher level
spaces seem to be too irregular to admit much compression,
except by reducing resolution or fidelity, and with it
performance.
Hans Moravec
> And, in advance of seeing such, I will propose and advocate
> that it may turn out that intelligence does not scale to
> super levels. What if we could clone every person's mind
> into a computer, and build a box with five billion
> processors? You go ask it a question. One billion
> say yes, one billion say no, three billion abstain.
> How smart is it?
> Joshua Stern
Ah, but what if they also ran a billion times as fast, and
interacted in a similacrum of our civilization? And we
provided them great virtual rewards for doing our bidding?
Then they might come up in ten seconds answers or artworks
that would take our civilization 300 years. But surely
they could also figure out how to organize themselves
so they work more efficiently than that, given the
possibilities of their virtual existence.
Chris Hooley
On Wed, 31 Dec 1997 18:00:53 GMT, JRS...@gte.net (JRStern) wrote:
>Every attempt at AI based on a fixed rule base is solipsistic. This
>includes most everything called "expert systems" in the 1980s.
OK, then it's a difference of definitions. I've taken a narrow view of
solipsism which definately would *not* include fixed rule base systems. Now
that I understand how you are using the word, I essentially agree with you.
This still leaves an open area beneath our different use of the terms that
I'd like to explore.
Because humans share a biological and species specific platform, we don't
frequently look beneath our common observational framework, and it's
reasonable that we don't. This is naturally selected. Nature deselects
inefficient attention to unnecessary detail.
We jump over non-contested areas and spend our processing resources on
areas where there are differences like, well, speaking to each other on
newsgroups about definitions.
However, AI does not share this broad biological base. That fact is
critical. So, where do human dependent meanings about "the world" fit in
AI?
For the purpose of clear AI development, I say that they don't fit at all.
For the purpose of clear AI development, I say that there is no "world"
independent of the system that experiences it. That holds true for AI and
for NI.
For the purpose of clear AI development, I say that our agreements about
"the world" reflect common roots in our human observing systems rather than
*objective* truths about reality.
Of course, we could simply declare a floor and view AI as a transparent box
with programmable moving parts functioning through our a priori rules
(which are unavoidably human-dependent).
However, that's *not* AI, and its not solipsystem programming. That's
running models on computers. Nor is operationally transparent behavior,
i.e., behavior reducible to internal signals that have meaning to an
observing human, what we can call "intelligent behavior."
This is a somewhat subtle point, but it's important. Wouldn't we require of
"intelligent" behavior exactly this---that we recognize in that behavior
*another* self-maintaining, self-centered island, one that is "independent"
of our own? Isn't it precisely this behavior which seems to us that it
must have arisen *within another system*, uncontrollable by us, which
signifies intelligence?
I can't imagine calling any system that we could fully understand
"intelligent."
In this sense, intelligence is solipsism's logical companion.
Intelligence, as the behavioral evidence for the existence of "other"
solipsystems, peeks through the encompassing veil of experience, and
relieves our natural solitude.
Oops. Straying again. Worse than that, I am doing it poetically!
How long before superintelligence?
I see three phases to AI. 1.)Appliance AI, 2.)Management AI, and 3.)Living
AI. In the first phase, AI is clearly man's artifact and servant. In the
second, complex tasks like managing cities are given over to AI, and we
agree to abide by its decisions. In the third phase, AI reinvents itself,
remanufactures itself to its own design, uproots even the core human
dependent constructs imported by its intitiators.
Phase 3: AD 2090.
Regards
Chris
JRStern
On 31 Dec 1997 12:45:54 -0600, ric...@cs.niu.edu (Neil Rickert)
wrote:
>Yet evolutionary psychology sees no reason to give a plausible
>scenario that would allow tools to evolve as an a priori category in
>a species most of whose history has been as a simple
>hunter-gatherer.
Pinker's new book gives several hundred pages worth.
>My objection to a priori knowledge is not that it is a limitation.
>Rather, my objection is that it is an ad hoc assumption used to
>repair a failed theory of cognition. Worse still, little is done in
>the way of explaining how evolutionary processes could solve a
>problem that postulated cognitive processes are unable to solve. If
>evolutionary processes are so powerful, why not postulate that
>cognitive processes are similar and can solve similar problems?
I'm not repairing anything, though I may be building a similar
structure next door. And nothing is said about evolutionary processes
solving what cognition cannot, what is said is that an understanding
of evolutionary processes helps predict just what the structure of
effective cognition might be.
OTOH, on your final point, I rather agree, though in light of the
above, that leaves us proposing cognitive processes that evolution saw
no need for, which is uncomfortable. Let's say rather that, in so far
as evolution helps cognition develop and focus, it can do so because
it was fundamentally possible for cognition to work in the first
place.
I just did a search for commentary on How the Mind Works. Apparently,
most people hate it with a passion. See the review on amazon.com, or
Pinker's own home page for the book at http://www-bcs.mit.edu/~steve/
(half the links are dead, or go through logon gateways).
Joshua Stern
JRS...@gte.net
JRStern
Well said!
It's amazing, the lengths to which people will go to misread what
Pinker has written, when his tone and claims are really quite
reasonable, for the most part. Later in the book he omits some of the
qualifications and subjunctives, but it would double the length of the
text if he included them. Several reviewers have suggested it would
have been more convenient all around to shorten the book instead. I'm
glad he wrote what he did -- I don't agree with it all, of course, but
mostly because in some areas, he didn't go far enough. When it comes
right down to it, he's a psychologist, not a computer scientist, and
sometimes it shows. <g>
Joshua Stern
JRS...@gte.net
--