%\documentstyle[11pt,twoside,informat]{article}
%\setlength{\oddsidemargin}{0pt}
%\setlength{\evensidemargin}{0pt}
%\setlength{\textheight}{8.25in}
%\setlength{\textwidth}{6.5in}

%\input{psfig}

\newcommand{\papername} {Matjaz Book}



\newcommand{\lnstretch}[1]
{\renewcommand{\baselinestretch}{#1}
 \large
 \normalsize}

\newcommand{\QED}{
\rule{4mm}{4mm}}

%\begin{document}
\bibliographystyle{/usr/local/lib/tex82/inputs/alpha}


\title{AI Progress, Massive Parallelism and Humility}

\author{J. Geller \\
{\it     Department of Computer and Information Sciences \\
    New Jersey Institute of Technology, Newark, NJ 07102\\
    geller@homer.njit.edu} }

  \titleodd{AI Progress, Massive Parallelism\ldots}
  \authoreven{J. Geller}

\keywords{massive parallelism, knowledge representation, Connection Machine}
\abstract{
In this paper we outline a view of Artificial Intelligence
that is between the extremes that ``everything is just
fine in AI'' and ``AI is hopeless, we might as well
give up.''  We present our own approach to Artificial
Intelligence, which is based on the combination of
Knowledge Representation with Massively Parallel hardware.
We conclude that Massive Parallelism might be helpful for
the development of Artificial Intelligence.  However, the
large investments necessary for the development of 
Massive Parallelism itself will require a determined
involvement of one or even several cooperating governments.
}

\date{}
\maketitle


%\noindent {\bf KEYWORDS:}

 
\large
\normalsize

%\begin{figure}
%\centerline{\psfig{figure=figure1.eps}}
%\caption
%\end{figure}



\section{Introduction}

In contrast to most other fields of science, Artificial Intelligence
is still questioned as a discipline in its entirety, even after a
41 year history.  The opinions of scholars about AI cover the widest
possible range.  On one side, many practitioners and researchers
feel that AI is developing as well as can be expected.  At the other
extreme the opinion is held that AI cannot succeed.  The conclusion
to be drawn from this is apparently that we might as well stop working
on it.  The contributors to this book certainly agree that AI is
{\it not} developing as well as it should.  However, they disagree
with each other on what the problem is, and how to deal with it.

Let us briefly address the advocates of the extreme attitude that 
``The problems of AI can never be solved, so we might as well
give up and work on something different.''  We feel that three questions
have to be raised, if we want to make a decision about terminating 
AI as a research program.
(1) Is the problem of AI important to science and humanity as a whole?
(2) Is it still true that scientific investigation in every known field
to date has resulted in some advancement?
(3) Is there positive proof that the problems of AI {\it cannot} be solved?
As long as the answers to these three questions are yes, yes, and no,
respectively, we feel justified in pursuing the goals of AI.

Where do we fit in the wide spectrum of opinions about AI?
Let us draw a comparison to cancer research.  There is wide agreement
that from the point of view of curing cancer, there has been (too) little
progress in research.  Does anybody therefore advocate that we stop working
on cancer research?  Of course not.  The problem of cancer is too important
to humanity.  We feel the same way about AI.  We feel that AI has not made 
as much progress as expected.  AI has also not made as much progress as 
its proponents have promised at different stages of AI's history.  
But the overall problem of AI is much too important for our understanding
of ourselves to give up on it.  Thus, the solution that we suggest is to 
expend {\it more} effort on AI while promising {\it less} about it.  

In Section 2 we present an argument why we feel that AI has shown
a certain lack of progress.  In Section 3 we touch very briefly on the
problem of overcommitment in AI.  In Section 4 we finally introduce
our own approach to Artificial Intelligence research.  This approach
is based on the combination of Knowledge Representation and 
Massively Parallel hardware.  Our conclusion in Section 5 is that we
need a return to ``big ticket science'' in Artificial Intelligence.  
Governments need to get involved in building the next generation of 
massively parallel computers for AI applications.


\section{Is there a Lack of Progress in AI?}

In 1995 the first AI book ever, {\it Computers and Thought} was republished
after 32 years \cite{Feigenbaum95}.
Looking at this book one gets the feeling that, let me say this carefully,
AI has not advanced as well as other fields of computer science.  Below
are two quotes, one from IJCAI 95\footnote{International
Joint Conference on Artificial Intelligence -- 1995, held in Montreal,
Canada.}, the other one from ``Computers and
Thought.''  The reader is invited to guess which one is from where.

``Both the FRS and the DBMS server were running on the same workstation,
a SPARCstation 10 model 41 with 64 MB of physical memory.'' (Quote 1).

``The IPL-IV system uses about 1200 of the JOHNNIAC's 4096 words of 
high-speed core memory and about 650 words of the 12,288-word auxiliary
drum memory.'' (Quote 2)  

The reader will probably think that I am trying
to pull his leg.  Obviously, Quote 1 is from 1995, and Quote 2 from 1963.
However, there is a point that I am trying to make.  Computer architecture,
processor design, memory technology and system software have improved 
between the time of Quote 2 and the time of Quote 1.  They have improved so 
strikingly, so ``screamingly,'' that the lives of all of us have been affected 
in almost every imaginable aspect.

Now let us try two more quotes:  ``However, even if we neglect the problem
of selecting which legitimate parsing is correct in a given instance, the
problem of discovering any legitimate parsing is itself formidable when
we deal with the tens of thousands of rules needed to describe a natural
language.  A complete set is, in fact, so large that none has yet been
devised for any natural language, although some have been under study
for thousands of years.''  (Quote 3)

``There are 19\% no-solution cases in the test.  This means that 593 
background verb-object pairs in the text are not sufficient to support
word sense disambiguation of these cases.''  (Quote 4)

The expert in Natural Language will probably recognize with ease which
one of these two quotes must be from 1995, and which one is from 1963.
For the general population, however, both quotations give a feeling
that things have not ``screamingly'' improved over the last 32 years, in
areas such as natural language understanding.  Let's try one last quote.
``Thus the statement that `George voted for ..... and is opposed to
medical care for the aged' makes it more likely that George is opposed to
the United Nations, though only slightly so.'' (Quote 5)  

My point is that I am not denying the considerable advancement in all fields
of Artificial Intelligence as seen from within the fields themselves.
However, looking at them from outside, from the point of view of their
impact on our every day lives, and doing this in comparison to the 
phenomenal development of available computing power, we must say that
core areas such as natural language understanding and scene/face recognition
have not advanced as expected.  The same thing is true for the most important
subarea of Artificial Intelligence, namely, Knowledge Representation.

If Wittgenstein was right, and ``meaning is use,'' then the meaning of
Knowledge Representation is right now ``mostly a subfield of applied logic.''
Anybody who has doubts about that may consult the proceedings of the
Knowledge Representation conferences, e.g., the recent one in Boston 
\cite{Aiello96}.
This development is understandable.  Many years of ``cute programs'' made
it harder and harder to judge correctness and originality of research.
Research in applied logics is easier to judge for its originality.  There
are original theorems with proofs, so the research results in correct
publications.  

What fields such as knowledge representation need is a return to 
programming.  Libraries of test cases and benchmarks need to be maintained, 
and competitions organized, such as in other AI areas, e.g., machine learning 
and theorem proving.  Any and all published implementation papers must
make their source code and their test data available over the Internet.
A paper that describes a system that objectively improves an existing 
system (or its own earlier version) by 10\% must be publishable in a
high quality conference, and if the improvement is based on a conceptual
advancement, even in a high quality journal.  Metrics must be developed
to compare implementations.  Again, this is in no way an original 
request.  Software engineers do it all the time, but knowledge representation
researchers don't.  Below a separate section will be devoted to our own
approach for making progress in AI, namely, the combination of Knowledge
Representation with Massive Parallelism.

Our comparisons with 1963 should not be misunderstood.
Having a notion of Knowledge Representation is a considerable
improvement over not having it.  Having a track on knowledge representation
in all major AI conferences, and a dedicated KR conference, is certainly
an improvement over what was done in 1963.  Yes, AI has influenced every
other field in computer science, and many other areas outside of computer
science, including philosophy, psychology, and linguistics.  Yes, my new (cheap)
Canon camera's directions for use mention that the camera uses AI (= Artificial
Intelligence).  There is no doubt about all these great advancements
since {\it Computers and Thought.}

But, why is there no Microsoft Semantic Network bundled together with
Windows NT?  Why is there no SUN natural language help program delivered
with SOLARIS?  Why do I still have to swipe a card to get into my
office on Sunday, instead of being able to smile into a robot camera?  In fact, 
why do people in AI still need to go to the office on Sunday?  Why do I still
have to vacuum the house if I don't go to the office on Sunday?  
Why isn't a household robot doing it for me?  All these
are questions that indicate that we have not come far enough in AI yet,
and {\it Computers and Thought} is a painful reminder of that fact.  Maybe
readers will not agree with my prescription(s) for progress, but let us
take the (re)publication of this book as a signal that AI researchers 
need to re-evaluate their own methods and their progress.

Let me conclude this section by solving the above puzzles.  Quote 3
and Quote 5 are from Computers and Thought, Quote 4 is from IJCAI 1995.
The most interesting point about Quote 5 is that even the ``popular examples''
then and now are from healthcare.




\section{Humility -- a Subjective Interlude}

Learning, the ability to improve one's behavior, is considered to be
central to any form of intelligence.  So what can we learn from 
41 years of AI history?  Maybe what we should learn is humility.
The overclaims of AI history are well known to everybody, and well
documented in other papers in this volume.  Did this failure shape
the behavior of the AI community?  In my opinion, 
the answer is ``no.''  How often do we hear conference speakers
say ``I don't know''?  How often do we find conference papers where
the authors stress the shortcomings of their own approaches?  
Not often enough, in my opinion.  I think that the field over all would 
gain credibility if results were presented with a little more humility.
As for predictions, maybe AI researchers should stop making predictions
altogether.

One event that can and should teach AI researchers humility is having
children.  It is very humiliating to see a child under the age of three process
language, solve problems, and negotiate obstacle courses in a way that
puts our natural language processing, planning, problem solving, vision
and robotics efforts to shame.  Maybe one semester in a child-care
facility should become a prerequisite for getting a PhD in AI.
(This prerequisite will be waived for parents.)


\section{Massively Parallel Knowledge Representation}

\subsection{What is Massively Parallel Knowledge Representation?}

We have spent the past seven years investigating an approach to AI
that combines ideas from Knowledge Representation with the power of
Massive Parallelism.  We call this Massively Parallel Knowledge
Representation (MaPKR)\footnote{This term was initially coined
for \cite{GellerDu91}.  It is pronounced ``mapcar''  which is intended to hint
at its LISP heritage.  Unfortunately, the abbreviation was mangled
by a copy editor and came out in the final version of the paper as
M\&PKR.} 
\cite{GellerDu91,Geller93a,Geller94a,Geller94b,Geller97}
\cite{Lee93,Lee96a,Lee96b,Lee97,Lee97b}.  
A small number of other investigators have also 
produced research which can be described by this label, e.g.,
\cite{Evett93a,Hendler95,Kitano93a,Waltz90,Shastri89,Stoffel97,Shastri97}.  
Having said 
about humility what we just said about humility, we need to stress that
we don't know whether MaPKR will contribute to the solution of AI's big
problems.  As opposed to some of the papers in this volume, we cannot
claim that this is a new approach.  The foundations of what has been done
in MaPKR go back to Fahlman \cite{Fahlman79}.  It is simply our belief that
this is an approach that still deserves some more attention.

We will now contrast Massively Parallel Knowledge Representation  
in broad strokes with other approaches to symbolic knowledge 
representation (SKR).  In SKR, a difficult problem is introduced with
the help of a (few) small example(s).  Then a solution is developed
that is found to work on problems of the same complexity as the initial
example.  When the solution is tried out on problems of considerably
higher complexity, it usually does not scale up.

In Massively Parallel Knowledge Representation, on the other hand, 
a comparatively simple problem is introduced, also with a few simple
examples.  However, the solution is developed from the outset keeping
in mind that it will have to work for problems of considerably
higher complexity than these examples.  To achieve these goals, algorithms
are developed for massively parallel hardware.  If the 
implementations of these parallel algorithms show satisfactory runtimes,
then the simple problems are replaced by more complicated problems.

Now we will contrast Massively Parallel Knowledge Representation
with neural network approaches.  Neural networks often have the term
``massively parallel'' in common with MaPKR.  However, most
neural network approaches use numeric weights that are hard to interpret. 
This is not the case for MaPKR.  Even though our own work uses a numeric
representation, as explained below, the numbers are (large) integers
which can be easily interpreted at all times \cite{GellerDu91}.  
Other approaches
to MaPKR don't use any auxiliary numeric representations.  Some research
straddles the borderline between MaPKR and neural networks, especially
\cite{Shastri88,Shastri89,Shastri97}.  
We feel that only research that has been implemented on 
massively parallel hardware really deserves the label MaPKR.  Simulators
are not acceptable.  This also seems to be in contrast with neural
network research, where simulators are often used.

\subsection{A Summary of our Research Results}

Some AI researchers have been motivated by trying to understand
how people can solve certain problems so quickly, so effortlessly,
so reflexively \cite{Shastri90,Shastri93,Shastri97}.  
AI's considerable interest in inheritance and
IS-A hierarchies is rooted in this motivation.  One problem that is even
simpler than inheritance is transitive closure reasoning.  Like inheritance,
this is a ubiquitous form of human reasoning which people perform on a
daily basis, seemingly without any effort.  For example, humans can
answer the question ``Are lions larger than pencils?'' without any hesitation.
We know that lions are larger than pencils, even though we have 
never seen a lion together with a pencil.  There are certainly many 
explanations possible for how humans are performing this reasoning step,
and we are not trying to find out which one is the right one.

What we are trying to do is to suggest a mechanism that gives similar
results as the ones obtained by humans and that works very fast.  Transitive
closure reasoning seems to qualify for this purpose.  We can build a
reasoner that ``knows'' that lions are larger than cats, and that also
knows that cats are larger than pencils.  If we apply transitivity to
these two assertions, the reasoner can derive the fact that lions are 
indeed larger than pencils.


The root of our research on transitivity reasoning was the combination of two 
surprising (for us) realizations.   One realization was that this kind
of transitive closure reasoning can actually be performed in constant
time.  We were made aware of this fact by a paper of Schubert {\it et al.}
\cite{Schubert87}.
Their solution had two problems: (1) It works only for trees;  
(2) For large knowledge bases, there is no good update algorithm for
Schubert's representation.  The second realization was that on a 
Connection Machine \cite{TMC88} it takes the same time to add the number 1
to one variable, or to add the number 1 to 4096 variables, as long as
every one of these variables sits on its own little processor.
(This assumes that the Connection Machine in question has 4k processors.
As the largest possible Connection Machines had 64k processors, this is a
very reasonable assumption.  We did most of our research on a 32k processor
Connection Machine.)

To make this more concrete, we will briefly show an example of how
constant-time transitive closure reasoning is possible.
In Figure 1 a typical IS-A hierarchy is represented.  Every
node in the graph represents a concept of a semantic network.
The node is annotated with a text string describing which concept
is represented.  Every arrow represents an IS-A (subclass) link.
An arrow from Mammal to Animal indicates that every mammal is also
an animal.  In addition, IS-A links are used for inheritance, however,
this is not relevant here.

Each node is annotated with a number pair.  The number pair encoding 
next to each node is due to Schubert \cite{Schubert87}.  The way number 
pairs are 
generated is simple, but it is not relevant to this paper, and we will 
omit a description \cite{Schubert87}.  The way the number pairs are used is
described now.  If the question is raised whether a Cheetah is
an Animal, then the question can be answered in two different ways:
(1) By following the chain of pointers from Cheetah up until either
Animal or the root (Thing) is found;  (2) By comparing the number pair
of Cheetah with the number pair of Animal.  If and only if the number
pair of Cheetah is contained in the number pair of Animal, we can
say that a Cheetah is indeed an animal.  Because [7 7] is contained
in [3 9] this is the case, and this result was obtained without any
``pointer chasing.''



\begin{figure*}
\centerline{\psfig{figure=geller.eps}}
% Page 14 Yugi Dissertation
\caption{}
\end{figure*}

Combining the two realizations -- the possible speed of transitive
closure reasoning and the power of the Connection Machine -- we found
that problem (2) from above, finding an update algorithm,
can be easily solved \cite{GellerDu91}.  In a relatively
short time all necessary algorithms were developed, implemented, tested,
and published.  Then we went on to solve problem (1).  Even though we
could draw on a very good idea that was published shortly before that
time \cite{Agrawal89}, we still had to find a way for parallelizing this idea.
We figured that all problems would be solved in a short time, about as
much time as it had taken us to implement the tree update and transitivity
algorithms \cite{GellerDu91}.

{\bf Five years and one dissertation \cite{Lee97b} later...}
In the next several years it turned out
that by advancing from trees to directed acyclic graphs (DAGs) of IS-A
links, completely unexpected phenomena occurred.  Just to give a flavor
of what happened, we found that during update, number pairs suddenly disappeared
and other number pairs appeared out of nowhere.  The PhD student working
on these problems was very happy.  She had found enough hard problems for
a thesis \cite{Lee97b} and there would even be some problems left.  All in all,
this was our lesson in humility.


\subsection{The Demise of Massive Parallelism -- A Chance Missed?}

During our research on MaPKR, the initial CM-1 Connection Machine was
replaced by the (better) CM-2 model.  After that, the CM-2 was replaced
by the CM-5, and this author attended one of the ``unveiling presentations''
of the new machine.  To our considerable dismay, the CM-5 was not
really massively parallel anymore.  It was more like a tightly networked
cluster of workstation processors that achieved much of its parallelism
by simulating it.  The model that we had access to had only 32 real
processors.  One has to admit that these processors were much more powerful
than the tiny CM-2 processors.  Still, the central idea 
of massive parallelism implemented in hardware had been abandoned
by Thinking Machines, the makers of the Connection Machine.  The reason for
this aboutface was market pressure.  It was much cheaper to develop the
next generation of Connection Machines with off-the-shelf chips.  
Thinking Machines was under considerable pressure to cut costs and turn
profits.  An insider's look at these difficulties appeared in \cite{Waltz97}.


With the end of the cold war and the continued budget crisis, the United
States government had lost its willingness and ability to fund big ticket
item research such as the Connection Machine.  Thinking Machines went 
``Chapter 11'' (the US bankruptcy statute) and was later re-created
as a software company, losing many of its best and brightest scientists
in the process.  While other researchers found it easy to convert
their efforts to the coarse grained parallelism of other computers 
\cite{Stoffel97},
we feel that this is not the right way to go for our research.


We believe that if human-like implemented Artificial Intelligence ever
becomes a reality, it will be found that one of the components in the 
Solution (with a capital S) will be a form of Massively Parallel Knowledge 
Representation.  Undoubtedly, there will be many other ingredients.  
Maybe all subfields of current day AI will have contributed to the Solution.
But AI researchers should not make any predictions.  We just hope that
massively parallel hardware, at least one order of magnitude larger than what 
we have seen so far, will eventually be developed and made available for
Artificial Intelligence research.


\section{Conclusions}

We have presented an example of why we feel that Artificial Intelligence
has not advanced as well as other areas of computer science in the
last 41 years.  While we share this opinion with others, we believe
that the problem is not that AI is fundamentally ``unsolvable''
but that a more concentrated effort is needed for solving it.  Specifically, 
it is our bias that Massively Parallel Knowledge Representation 
is a useful area for further investigation.  We believe that
massively parallel computers with at least a million processors 
will be needed to model ``effortless'' human reasoning.

Obviously, such a large project is beyond the reach of normal
channels of financing.  Venture capital, bank loans, and development
as part of a normal profit-oriented business cycle have failed before
\cite{Waltz97} and will certainly fail at a project of this size.  Governments
will have to step in, and multi-national cooperations might be necessary
for the development of this kind of computational capability.
Further development in other areas of Artificial Intelligence will
also be necessary, and the integration of these different technologies
will pose a nontrivial problem.

With all that, we are not comfortable with ``promising'' success.
At a time when government funding for research is often only
available if the proposer can ``virtually guarantee'' success,
we are making it a point that no such guarantee is possible.  After all,
maybe AI researchers should stop making predictions altogether.\\

{\small
\subsection*{Acknowledgments}

Bill Rapaport has pointed out to me initially that the re-publication of 
{\it Computers and Thought} 
engenders the feeling of a certain lack of progress in AI.  Mike Halper
has commented on a draft of this paper.


%\bibliography{/export/ai/faculty/geller/DB/db}
%\bibliography{/export/ai/faculty/geller/PAR/par}

\begin{thebibliography}{{Thi}88}

\bibitem[ABJ89]{Agrawal89}
R.~Agrawal, A.~Borgida, and H.~V. Jagadish.
\newblock Efficient management of transitive relationships in large data and
  knowledge bases.
\newblock In {\em Proceedings of the 1989 ACM SIGMOD International Conference
  on the Management of Data}, pages 253--262, Portland, OR, 1989.

\bibitem[ADS96]{Aiello96}
L.~C. Aiello, J.~Doyle, and S.~Shapiro, editors.
\newblock {\em Proceedings of the Fifth International Conference on Principles
  of Knowledge Representation and Reasoning (KR' 96)}, San Francisco, CA, 1996.
  Morgan Kaufmann.

\bibitem[EHA93]{Evett93a}
M.~P. Evett, J.~A. Hendler, and W.~A. Andersen.
\newblock Massively parallel support for computationally effective recognition
  queries.
\newblock In {\em Proceedings of the Eleventh National Conference on Artificial
  Intelligence}, pages 297--302. MIT Press, Cambridge, MA, 1993.

\bibitem[Fah79]{Fahlman79}
S.~E. Fahlman.
\newblock {\em NETL: A System for Representing and Using Real-World Knowledge}.
\newblock MIT Press, Cambridge, MA, 1979.

\bibitem[FF63]{Feigenbaum95}
E.~A. Feigenbaum and J.~Feldmann, editors.
\newblock {\em Computers and Thought}.
\newblock McGraw-Hill, New York, 1963.
\newblock Republished 1995 by AAAI with MIT Press.

\bibitem[GD91]{GellerDu91}
J.~Geller and C.~Y. Du.
\newblock Parallel implementation of a class reasoner.
\newblock {\em Journal of Experimental and Theoretical Artificial
  Intelligence}, 3:109--127, 1991.

\bibitem[Gel93]{Geller93a}
J.~Geller.
\newblock Innovative applications of massive parallelism.
\newblock {\em AAAI 1993 Spring Symposium Series Reports, AI Magazine},
  14(3):36, 1993.

\bibitem[Gel94a]{Geller94a}
J.~Geller.
\newblock Advanced update operations in massively parallel knowledge
  representation.
\newblock In H.~Kitano and J.~Hendler, editors, {\em Massively Parallel
  Artificial Intelligence}, pages 74--100. AAAI/MIT Press, 1994.

\bibitem[Gel94b]{Geller94b}
J.~Geller.
\newblock Inheritance operations in massively parallel knowledge
  representation.
\newblock In L.~Kanal, V.~Kumar, H.~Kitano, and C.~Suttner, editors, {\em
  Parallel Processing for Artificial Intelligence}, pages 95--113. Elsevier
  Science Publishers, Amsterdam, 1994.

\bibitem[GKS97]{Geller97}
J.~Geller, H.~Kitano, and C.B. Suttner, editors.
\newblock {\em Parallel Processing for Artificial Intelligence 3}.
\newblock North-Holland Elsevier, Amsterdam, 1997.

\bibitem[HCL95]{Hendler95}
J.~A. Hendler, Jaime Carbonell, and Douglas Lenat.
\newblock Very large knowledge bases - architecture vs engineering.
\newblock In {\em Proc. of the 14th Int. Joint Conference on Artificial
  Intelligence}, pages 2033--2036. Morgan Kaufmann, Montreal, Quebec, 1995.

\bibitem[Kit93]{Kitano93a}
H.~Kitano.
\newblock Challenges of massive parallelism.
\newblock In {\em Proc. of the 13th Int. Joint Conference on Artificial
  Intelligence}, pages 813--834. Morgan Kaufmann, San Mateo, CA, 1993.

\bibitem[Lee97]{Lee97b}
Y.~Lee.
\newblock {\em Massively Parallel Reasoning in Transitive Relationship
  Hierarchies}.
\newblock PhD thesis, CIS Department, New Jersey Institute of Technology, 1997.

\bibitem[LG93]{Lee93}
E.~Y. Lee and J.~Geller.
\newblock Representing transitive relationships with parallel node sets.
\newblock In Bharat Bhargava, editor, {\em Proceedings of the IEEE Workshop on
  Advances in Parallel and Distributed Systems}, pages 140--145. IEEE Computer
  Society Press, Los Alamitos, CA, 1993.

\bibitem[LG96a]{Lee96a}
E.~Y. Lee and J.~Geller.
\newblock Constant time inheritance with parallel tree covers,.
\newblock In {\em Proceedings of the FLAIRS (Florida AI Research Symposium)},
  pages 243--250, Key West, Florida, 1996.

\bibitem[LG96b]{Lee96b}
E.~Y. Lee and J.~Geller.
\newblock Parallel transitive reasoning on mixed relational hierarchy,.
\newblock In {\em Proceedings of the Knowledge Representation and Reasoning},
  Cambridge, MA, 1996.

\bibitem[LG97]{Lee97}
E.~Y. Lee and J.~Geller.
\newblock Parallel operations on class hierarchies with double strand
  representation.
\newblock In J.~Geller, H.~Kitano, and C.~Suttner, editors, {\em Parallel
  Processing for Artificial Intelligence 3}, pages 69--94, Amsterdam, 1997.
  North-Holland Elsevier.

\bibitem[SA90]{Shastri90}
L.~Shastri and V.~Ajjanagadde.
\newblock An optimally efficient limited inference system.
\newblock In {\em Proceedings of IJCAI-90}, pages 563--570, Boston, MA, 1990.

\bibitem[Sha88]{Shastri88}
L.~Shastri.
\newblock {\em Semantic Networks: an Evidential Formalization and its
  Connectionist Realization}.
\newblock Morgan Kaufmann Publishers, San Mateo, CA, 1988.

\bibitem[Sha89]{Shastri89}
L.~Shastri.
\newblock Default reasoning in semantic networks: a formalization of
  recognition and inheritance.
\newblock {\em Artificial Intelligence}, 39(3):283--356, 1989.

\bibitem[Sha93]{Shastri93}
L.~Shastri.
\newblock A computational model of tractable reasoning -- taking inspiration
  from cognition.
\newblock In {\em Proc. of the 13th Int. Joint Conference on Artificial
  Intelligence}, pages 202--207. Morgan Kaufmann, San Mateo, CA, 1993.

\bibitem[SHS97]{Stoffel97}
K.~Stoffel, J.~Hendler, and J.~Saltz.
\newblock Parka on mimd-supercomputers.
\newblock In J.~Geller, H.~Kitano, and C.~B. Suttner, editors, {\em Parallel
  Processing for Artificial Intelligence 3}, pages 95--118. North-Holland
  Elsevier, New York, NY, 1997.

\bibitem[SM97]{Shastri97}
L.~Shastri and D.~R. Mani.
\newblock Massively parallel knowledge representation and reasoning: Taking a
  cue from the brain.
\newblock In J.~Geller, H.~Kitano, and C.~B. Suttner, editors, {\em Parallel
  Processing for Artificial Intelligence 3}, pages 3--40. North-Holland
  Elsevier, New York, NY, 1997.

\bibitem[SPT87]{Schubert87}
L.~K. Schubert, M.~A. Papalaskaris, and J.~Taugher.
\newblock Accelerating deductive inference: special methods for taxonomies
  colors and times.
\newblock In N.~Cercone and G.~McCalla, editors, {\em The Knowledge Frontier},
  pages 187--220. Springer Verlag, New York, NY, 1987.

\bibitem[{Thi}88]{TMC88}
{Thinking Machines Corporation}.
\newblock {\em {*LISP} Reference Manual Version 5.0 edition}.
\newblock Thinking Machines Corporation, Cambridge, MA, 1988.

\bibitem[Wal90]{Waltz90}
D.~L. Waltz.
\newblock Massively parallel {AI}.
\newblock In {\em Proceedings of the Eighth National Conference on Artificial
  Intelligence}, pages 1117--1122. Morgan Kaufmann, San Mateo, CA, 1990.

\bibitem[Wal97]{Waltz97}
D.~L. Waltz.
\newblock Ai applications of massive parallelism: An experience report.
\newblock In J.~Geller, H.~Kitano, and C.~B. Suttner, editors, {\em Parallel
  Processing for Artificial Intelligence 3}, pages 327--339. North-Holland
  Elsevier, New York, NY, 1997.

\end{thebibliography}
}
%\end{document}