AIList-REQUEST@SRI-AI.ARPA (AIList Moderator Kenneth Laws) (10/19/85)
AIList Digest Friday, 18 Oct 1985 Volume 3 : Issue 149
Today's Topics:
Projects - University of Aberdeen & CSLI,
Literature - New Complexity Journal,
AI Tools - Lisp vs. Prolog,
Opinion - AI Hype & Scaling Up,
Cognition & Logic - Modus Ponens,
Humor - Dognition
----------------------------------------------------------------------
Date: Thu 17 Oct 85 12:44:41-PDT
From: Derek Sleeman <SLEEMAN@SUMEX-AIM.ARPA>
Subject: University of Aberdeen Program
UNIVERSITY of ABERDEEN
Department of Computing Science
The University of Aberdeen is now making a sizeable committment to
build a research group in Intelligent Systems/Cognitive Science.
Following the early work of Ted Elcock and his co-workers, the
research work of the Department has been effectively restricted to
databases. However, with the recent appointment of Derek Sleeman
to the faculty from summer 1986, it is anticipated that a sizeable
activity will be (re)established in AI.
In particular we are anxious to have a number of visitors at
any time - and funds have been set aside for this. So we would be
particularly interested to hear from people wishing to spend Sabbaticals,
short-term Research fellowships etc.
Please contact Derek Sleeman at 415 497 3257 or SLEEMAN@SUMEX
for further details.
------------------------------
Date: Wed 16 Oct 85 17:12:46-PDT
From: Emma Pease <Emma@SU-CSLI.ARPA>
Subject: CSLI Projects
[Excerpted from the CSLI Newsletter by Laws@SRI-AI.]
CSLI PROJECTS
The following is a list of CSLI projects and their coordinators.
AFT Lexical Representation Theory. Julius Moravcsik
(AFT stands for Aitiuational Frame Theory)
Computational Models of Spoken Language. Meg Withgott
Discourse, Intention, and Action. Phil Cohen.
Embedded Computation Group. Brian Smith (3 sub groups)
sub 1: Research on Situated Automata. Stan Rosenschein
sub 2: Semantically Rational
Computer Languages. Curtis Abbott
sub 3: Representation and Reasoning. Brian Smith
Finite State Morphology. Lauri Karttunen
Foundations of Document Preparation. David Levy.
Foundations of Grammar. Lauri Karttunen
Grammatical Theory and Discourse
Structures. Joan Bresnan
Head-driven Phrase Structure Grammar. Ivan Sag and Thomas Wasow
Lexical Project. Annie Zaenen
Linguistic Approaches to Computer
Languages. Hans Uszkoreit
Phonology and Phonetics. Paul Kiparsky
Rational Agency. Michael Bratman
Semantics of Computer Language. Terry Winograd
Situation Theory and Situation
Semantics (STASS). Jon Barwise
Visual Communication. Sandy Pentland
In addition, there are some interproject working groups. These
include:
Situated Engine Company. Jon Barwise and Brian Smith
Representation and Modelling. Brian Smith and Terry Winograd
------------------------------
Date: Wed 16 Oct 85 09:56:32-EDT
From: Susan A. Maser <MASER@COLUMBIA-20.ARPA>
Subject: NEW JOURNAL
JOURNAL OF COMPLEXITY
Academic Press
Editor: J.F. Traub, Columbia University
FOUNDING EDITORIAL BOARD
K. Arrow, Stanford University
G. Debreu, University of California, Berkeley
Z. Galil, Columbia University
L. Hurwicz, University of Minnesota
J. Kadane, Carnegie-Mellon University
R. Karp, University of California, Berkeley
S. Kirkpatrick, I.B.M.
H.T. Kung, Carnegie-Mellon University
M. Rabin, Harvard University and Hebrew University
S. Smale, University of California, Berkeley
S. Winograd, I.B.M.
S. Wolfram, Institute for Advanced Study
H. Wozniakowski, Columbia University and University of Warsaw
YOU ARE INVITED TO SUBMIT YOUR MAJOR RESEARCH PAPERS TO THE JOURNAL.
See below for further information.
Publication Information and Rates:
Volume 1 (1985), 2 issues, annual institutional subscription rates:
In the US and Canada: $60
All other countries: $68
Volume 2 (1986), 4 issues, annual institutional subscription rates:
In the US and Canada: $80
All other countries: $93
Send your subscription orders to: Academic Press, Inc.
1250 Sixth Avenue
San Diego, CA 92101
(619) 230-1840
Contents of Volume 1, Issue 1:
"A 71/60 Theorem for Bin Packing" by Michael R. Garey & David S. Johnson
"Monte-Carlo Algorithms for the Planar Multiterminal Network
Reliability Problem" by Richard M. Karp & Michael Luby
"Memory Requirements for Balanced Computer Architectures" by H.T. Kung
"Optimal Algorithms for Image Understanding: Current Status and
Future Plans" by D. Lee
"Approximation in a Continuous Model of Computing" by K. Mount & S. Reiter
"Quasi-GCD Computations" by Arnold Schonhage
"Complexity of Approximately Solved Problems" by J.F. Traub
"Average Case Optimality" by G.W. Wasilkowski
"A Survey of Information-Based Complexity" by H. Wozniakowski
SUBMISSION OF PAPERS
The JOURNAL OF COMPLEXITY is a multidisciplinary journal which
covers complexity as broadly conceived and which publishes research
papers containing substantial mathematical results.
In the area of computational complexity the focus is on
problems which are approximately solved and for which optimal
algorithms or lower bound results are available. Papers which provide
major new algorithms or make important progress on upper bounds are
also welcome. Papers which present average case or probabilistic
analyses are especially solicited. Of particular interest are papers
involving distributed systems or parallel computers for which only
approximate solutions are available.
The following is a partial list of topics for which
computational complexity results are of interest: applied mathematics,
approximate solution of hard problems, approximation theory, control
theory, decision theory, design of experiments, distributed computation,
image understanding, information theory, mathematical economics,
numerical analysis, parallel computation, prediction and estimation,
remote sensing, seismology, statistics, stochastic scheduling.
In addition to computational complexity the following are
among the other complexity topics of interest: physical limits of
computation; chaotic behavior and strange attractors; complexity in
biological, physical, or artificial systems.
Although the emphasis is on research papers, surveys or
bibliographies of special merit may also be published.
To receive a more complete set of authors' instructions (with format
specifications), or to submit a manuscript (four copies please),
write to:
J.F. Traub, Editor
JOURNAL OF COMPLEXITY
Department of Computer Science
450 Computer Science Building
Columbia University
New York, New York 10027
------------------------------
Date: Tue, 15 Oct 85 22:15 EDT
From: Hewitt@MIT-MC.ARPA
Subject: Lisp vs. Prolog (reply to Pereira)
I would like to reply to Fernando Pereira's message in which he wrote:
It is a FACT that no practical Prolog system is written entirely
in Lisp: Common, Inter or any other. Fast Prolog systems have
been written for Lisp machines (Symbolics, Xerox, LMI) but their
performance depends crucially on major microcode support (so
much so that the Symbolics implementation, for example, requires
additional microstore hardware to run Prolog). The reason for
this is simple: No Lisp (nor C, for that matter...) provides the
low-level tagged-pointer and stack operations that are critical
to Prolog performance.
It seems to me that the above argument about Prolog not REALLY being
implemented in Lisp is just a quibble. Lisp implementations from the
beginning have provided primitive procedures to manipulate the likes
of pointers, parts of pointers, invisible pointers, structures, and
stack frames. Such primitve procedures are entirely within the spirit
and practice of Lisp. Thus it is not surprising to see primitive
procedures in the Lisp implementations of interpreters and compilers
for Lisp, Micro-Planner, Pascal, Fortran, and Prolog. Before now no
one wanted to claim that the interpreters and compilers for these
other languages were not written in "Lisp". What changed?
On the other hand primitive procedures to manipulate pointers, parts
of pointers, invisible pointers, structures, and stack frames are
certainly NOT part of Prolog! In FACT no one in the Prolog community
even professes to believe that they could EVER construct a
commercially viable (i.e. useful for applications) Common Lisp in
Prolog.
I certainly realize that interesting research has been done using
Planner-like and Prolog-like languages. For example Terry Winograd
implemented a robot world simulation with limited natural language
interaction using Micro-Planner (the implementation by Sussman,
Winograd, and Charniak of the design that I published in IJCAI-69).
Subsequently Fernando did some interesting natural language research
using Prolog.
My chief chief concern is that some AILIST readers might be misled by
the recent spate of publicity about the "triumph" of Prolog over Lisp.
I simply want to point out that the emperor has no clothes.
------------------------------
Date: Thu, 10 Oct 85 11:03:00 GMT
From: gcj%qmc-ori.uucp@ucl-cs.arpa
Subject: AI hype
A comment from Vol 3 # 128:-
``Since AI, by definition, seeks to replicate areas of human cognitive
competence...''
This should perhaps be read in the context of the general discussion which
has been taking place about `hype'. But it is still slightly off the mark
in my opinion.
I suppose this all rests on what one means by human cognitive competence.
The thought processes which make us human are far removed from the cold
logic of algorithms which are the basis for *all* computer software, AI or
otherwise. There is an element in all human cognitive processes which
derives from the emotional part of our psyche. We reach decisions not only
because we `know' that they are right, but also because we `feel' them to
be correct. I think really that AI must be seen as an important extension
to the thinking process, as a way of augmenting an expert's scope.
Gordon Joly (now gcj%qmc-ori@ucl-cs.arpa
(formerly gcj%edxa@ucl-cs.arpa
------------------------------
Date: Fri 18 Oct 85 10:13:10-PDT
From: WYLAND@SRI-KL.ARPA
Subject: Scaling up AI solutions
>From: Gary Martins <GARY@SRI-CSL.ARPA>
>Subject: Scaling Up
>Mr. Wyland seems to think that finding problem solutions which "scale up"
>is a matter of manufacturing convenience, or something like that. What
>he seems to overlook is that the property of scaling up (to realistic
>performance and behavior) is normally OUR ONLY GUARANTEE THAT THE
>"SOLUTION" DOES IN FACT EMBODY A CORRECT SET OF PRINCIPLES. [...]
The problem of "scaling up" is not that our solutions do not work
in the real world, but that we do not have general, universal
solutions applicable to all AI problems. This is because we only
understand *parts* of the problem at present. We can design
solutions for the parts we understand, but cannot design the
universal solution until we understand *all* of the problem.
Binary vision modules provide sufficient power to be useful in
many robot assembly applications, and simple word recognizers
provide enough power to be useful in many speech control
applications. These are useful, real-world solutions but are not
*universal* solutions: they do not "scale up" as universal
solutions to all problems of robot assembly or understanding
speech, respectively.
I agree with you that scientific theories are proven in the lab
(or on-the-job) with real world data. The proof of the
engineering is in the working. It is just that we have not
reached the same level of understanding of intelligence that
Newton's Laws provided for mechanics.
Dave Wyland
------------------------------
Date: Tue 15 Oct 85 13:48:28-PDT
From: Mike Dante <DANTE@EDWARDS-2060.ARPA>
Subject: modus ponens
Seems to me that McGee is the one guilty of faulty logic. Consider the
following example:
Suppose a class consists of three people, a 6 ft boy (Tom), a 5 ft girl
(Jane), and a 4 ft boy (John). Do you believe the following statements?
(1) If the tallest person in the class is a boy, then if the tallest is
not Tom, then the tallest will be John.
(2) A boy is the tallest person in the class.
(3) If the tallest person in the class is not Tom then the tallest
person in the class will be John.
How many readers believe (1) and (2) imply the truth of (3)?
- Mike
------------------------------
Date: Thu, 17 Oct 85 21:22:26 pdt
From: cottrell@nprdc.arpa (Gary Cottrell)
Subject: Seminar - Parallel Dog Processing
SEMINAR
Parallel Dog Processing:
Explorations in the Nanostructure of Dognition
Garrison W. Cottrell
Department of Dog Science
Condominium Community College of Southern California
Recent advances in neural network modelling have led to its
application to increasingly more trivial domains. A prominent
example of this line of research has been the creation of an
entirely new discipline, Dognitive Science[1], bringing together
the insights of the previously disparate fields of obedience
training, letter carrying, and vivisection on such questions as,
"Why are dogs so dense?" or, "How many dogs does it take to
change a lightbulb?"[2]
This talk will focus on the first question. Early results
suggest that the answer lies in the fact that most dog
information processing occurs in their brains. Converging data
from various fields (see, for example, "A vivisectionist approach
to dog sense manipulation", Seligman, 1985) have shown that this
"wetware" is composed of a massive number of slow, noisy
switching elements, that are too highly connected to form a
proper circuit. Further, they appear to be all trying to go off
at the same time like popcorn, rather than proceeding in an
orderly fashion. Thus it is no surprise to science that they are
dumb beasts.
Further impedance to intelligent behavior has been
discovered by learning researchers. They have found that the
connections between the elements have little weights on them,
slowing them down even more and interfering with normal
processing. Indeed, as the dog grows, so do these weights, until
the processing elements are overloaded. Thus it is now clear why
you can't teach an old dog new tricks, and also explains why
elderly dogs tend to hang their heads. Experience with young
dogs appears to bear this out. They seem to have very little
weight in their brains, and their behavior is thus much more
laissez faire than older dogs.
We have applied these constraints to a neural network
learning model of the dog brain. To model the noisy signal of
the actual dog neurons, the units of the model are restricted to
communicate by barking to one another. As these barks are passed
from one unit to another, the weights on the units are increased
by an amount proportional to the loudness of the bark. Hence we
____________________
[1]A flood of researchers finding Cognitive Science too hard
are switching to this exciting new area. It appears that trivial
results in this unknown field will beget journal papers and TR's
for several years before funding agencies and reviewers catch on.
[2]Questions from the Philosophy of dognitive science (dogmat-
ics), such as "If a dog barks in the condo complex and I'm not
there to hear it, why do the neighbors claim it makes a sound?"
are beyond the scope of this talk.
term this learning mechanism bark propagation. Since the weights
only increase, just as in the normal dog, at asymptote the
network has only one stable state, which we term the dead dog
state. Our model is validated by the fact that many dogs appear
to achieve this state while still breathing. We will demonstrate
a live simulation of our model at the talk.
------------------------------
End of AIList Digest
********************