Newsgroups: comp.ai.philosophy
Path: cantaloupe.srv.cs.cmu.edu!nntp.club.cc.cmu.edu!godot.cc.duq.edu!news.duke.edu!news-feed-1.peachnet.edu!gatech!howland.reston.ans.net!ix.netcom.com!netcom.com!kovsky
From: kovsky@netcom.com (Bob Kovsky)
Subject: Re: Is the mind/brain deterministic?
Message-ID: <kovskyCz7q87.I4q@netcom.com>
Organization: NETCOM On-line Communication Services (408 261-4700 guest)
References: <kovskyCz0B4G.Aqr@netcom.com> <39r7u3$sk4@cantaloupe.srv.cs.cmu.edu> <kovskyCz4wy5.BKE@netcom.com> <HPM.94Nov12041452@cart.frc.ri.cmu.edu>
Date: Sun, 13 Nov 1994 15:40:55 GMT
Lines: 145

In earlier posts, Prof. Moravec and I have been exchanging badinage in a
debate where I contended that one reason AI has not achieved its promised
breakthroughs has been its turning away from study of biological function
and its insistence on the universal applicability of computational models.
I included a remark about

>>... the world-shaking 
>>breakthroughs that have emerged from Prof. Moravec's laboraroty.


Prof. Moravec responded:

>	I guess this is supposed to be sarcasm to punish me for my
>opinion that artificial neural nets or other systems are machines even
>when operating in high gain states "at the edge of chaos".  Well, I'm
>not ashamed of either my terminology, my opinions or my research life,
>so I'll call this bluff.  Then I'd like to hear what things Dr. Kovsky
>has accomplished to credential his opinions.

Prof. Moravec then posts an autobiographical note detailing his long and 
successful career in computers and artifical intelligence.  He continues 
with a report of the program in laboratory, which includes:

>	 Two decades of work have given us the means and confidence to
>attempt a mobile robot control program with enough spatial competence
>to reliably execute tasks like delivery and floor care in normal
>areas, with no special navigational cues and trained only by an
>initial run through.  We plan to develop and embed such a program in a
>robot that, given luck with certain research questions, could be a
>laboratory prototype of the first mobile robots with mass market
>potential. 
>	Our approach accumulates sensed evidence about surrounding
>geometry in a spatial grid. We plan to integrate about ten thousand 3D
>location estimates from multi-baseline stereo vision several times per
>meter traveled.  Preliminary, but very efficient, implementations of
>stereo and 3D grid algorithms suggest that 100 to 1000 MIPS of
>computer power will suffice to safely guide a walking-speed robot
>through normal indoors.
>	From 1973 to 1983 we developed stereo-vision-guided robots
>that drove through clutter by tracking a few dozen 3D points in camera
>images. Through several versions, the control program gained speed and
>accuracy, but a brittle failure mode persisted, misdirecting the robot
>in one journey in four, when chance clusters of tracking errors fooled
>geometric consistency checks.
>	We invented the so-called evidence grid approach in 1983, to
>handle data from inexpensive Polaroid sonar devices, whose wide beams
>leave angular position ambiguous. Instead of registering objects, the
>grid method accumulated occupancy evidence for an array of spatial
>locations, slowly resolving ambiguities as the robot moved. The first
>implementation worked surprisingly well, repeatedly mapping and
>guiding a robot across a cluttered test lab.  It worked on a tree
>lined path, in a coal mine, with stereo vision range data, combined
>stereo and sonar, and with probability theory replacing an ad-hoc
>formulation.  Its first failure, in uncluttered surroundings with
>smooth walls, led us to a major extension, the learning of sensor
>evidence patterns.  The evidence contributed by sonar readings had
>been hand-derived from the sensor's signal pattern, a poor model for
>interaction with mirrorlike walls. We are now able to train the
>program to work nicely in mirror surroundings, and superbly elsewhere.
>	Past work was with 2D grids of a few thousand cells, all that
>1980s computers could handle in near real time.  In 1992 we wrote a
>very efficient 3D implementation, that inserts thousands of readings
>per second into 3D grids with several million cells, on 1990s
>computers approaching 100 MIPS.  We have begun to use this new program
>with a multi-baseline stereo vision program developed by another
>research group, which provides range determinations at similar
>speed. Together they will be the core of a complete navigation
>demonstration, running on any of several robots available to us.
>
>			-- Hans Moravec   CMU Robotics

	Response:  a fly's navigational and motion-control systems 
enable it to execute complex activity both on flat surfaces and in the 
air.  What a major struggle it is to kill one!  It's brain, operating at 
a speed equivalent to a thousand instructions per second or so, and of a 
minute size, also controls bodily functions and exercises control over 
digestive and reproductive functions.  That you and your associates need 
to invest such enormous resources in order to accomplish something that 
does not even perform as well suggests that there is something missing in 
your approach.

	My sarcastic remark about "world-shaking breakthroughs" stands.  
After 25 years and billions of dollars, AI has failed to produce.  Other 
than some useful but minor "expert systems" operating in extremely 
limited domains, there is very little evidence that the model of 
universal mechanical computation has any validity.  And very little 
return on the investment.

	As for my "credentials," there are not as externally impressive as
Prof. Moravec's.  In high school, I built a computer out of pinball
machine relays (a step up from "Simon," the subject of an early book by a
fellow named Berkley, I believe) and went on to get a good introduction to
science and technology at high-class institutions of higher education.  I
concluded that the world of computers was superficial and over-hyped, found
real-world phenomena deeper and more attractive, and got degrees in
electrical engineering and physics.  While in graduate school, my interest
in directly encountering the real world led me to law school (and it was
1970 at Berkeley). 

	The intellectual structure of law is as complex as that of 
science and far different.  I was immediately struck by the way in some 
features of legal information were organized in a way that made them a 
appropriate for what were then called "associative memories" and are today 
part of the basis for artificial neural networks.  As I tried to 
construct conceptual bridges between the world-view of science and that 
of law, I gradually came to uncover and explore the deepest and (to my 
mind) most important problem of all:  freedom.

	I have spent 25 years developing a natural science of freedom.  
The work begins with the hypothesis that "we aren't smart enough to 
figure things out" and that our processes of experience incorporate 
systematic errors and limitations.  Under some circumstances, such as 
those established in a laboratory, our intellectual models can be made to 
approach reality, the closeness of the approximation governed by such 
factors as the impoverishment of the environment and the completeness of 
the constraints imposed on the system under consideration.

	The same systematic errors and limitations prevent us from
achieving a comprehensive resolution of the problems of freedom and
consciousness.  There are methods of approaching the problems, however,
that allow partial resolution.  My work on these matters is available at
the ftp site indicated below. 

	Biological systems like the fly's can accomplish feats of
navigation and motion-control beyond our understanidng, and the biological
system that operates in our heads, even further beyond, makes
consciousness and freedom available to us.  There is good research that
suggests "neuronal" networks operating on the "edge of chaos" make a good
model.  I believe that trial-and-error laboratory work (rather than the
theories of mechanical computation) would be fruitful in developing
engineering analogs. 

	As I noted in an earlier post, the potter's kiln, the cook's 
oven, and the metallurgist's smelter are all devices, but none are 
machines.  Each of these transforms "inputs" into "outputs" through 
phase transformations that have correspondences to complex systems of 
information "on the edge of chaos."  A more productive future, perhaps, 
lies in this direction.
	

-- 

*   *    *    *    *    *    *    *    *    *    *    *    *    *    *    *   * 
    Bob Kovsky          |  A Natural Science of Freedom 
    kovsky@netcom.com   |  Materials available by anonymous ftp
                        |  At ftp.netcom.com/pub/freeedom
*   *    *    *    *    *    *    *    *    *    *    *    *    *    *    *   * 
