http:^^www.cs.utexas.edu^users^ai-lab^dream.html

This data set contains WWW-pages collected from computer science departments of various universities

MIME-Version: 1.0
Server: CERN/3.0
Date: Tuesday, 07-Jan-97 15:42:48 GMT
Content-Type: text/html
Content-Length: 24513
Last-Modified: Tuesday, 25-Jul-95 16:33:36 GMT

<TITLE>I had a Dream</TITLE>
<H1>I had a Dream</H1>

<hr>

<BLOCKQUOTE>
                    AAAI Presidential Address  <br>
                         19 August 1985  <p>

                          Woody Bledsoe  <br>
</BLOCKQUOTE>

<hr>

                          



Twenty five years ago I had a dream, a day dream if you will.  A
dream shared with many of you.  I dreamed of a special kind of 
computer, which had eyes and ears, and arms and legs, in addition
to its "brain".   <P>

I did not dream that this new computer friend would be a means of
making money for me or my employer, or a help for my country -- though
I loved my country then and still do, and I have no objection to
making money.  I did not even dream of such a worthy cause as helping
the poor and handicapped of the world using this marvelous new
machine.   <P>

No, my dream was filled with the wild excitement of seeing
a machine act like a human being, at least in many 
ways.   <P>

I wanted it to read printed characters on a page, and handwritten
script as well.  I could see it, or a part of it, in a small 
camera which would fit on my glasses, with an attached earplug
which would whisper into my ear the names of my friends and
acquaintances as I met them on the street. Or in a 
telephone which allowed me to converse with a friend in 
Germany, he in German and me in English.  For you see my
computer friend had the ability to recognize faces, synthesize
voice, understand spoken sentences, and translate languages,
and things like that.   <P>

I'll admit that in 1960 my computer person had a much larger 
head than I envision for it now.  Because I then didn't know about
microcomputers.   <P>

My dream computer person liked to walk and play ping pong, 
especially with me.  And I liked to teach it things - because
it could learn dexterity skills as well as mental concepts.  And
much more.   <P>

When I awoke from this day dream I found that we didn't have these
things, but we did have some remarkable computers, even then, so
I decided then and there to quit my job and set about spending 
the rest of my life helping bring this dream to reality.   <P>

Sometimes I forget this dream, and for these periods my life is
more drab.  Recently a reporter asked me  "why do you scientists
do AI research".  My answer,  "well certainly not for money,
though I wouldn't mind being rich, it goes deeper, to a 
yearning we have to make machines act in some fundamental ways 
like people."  Period.  This reminded my of my dream - and 
brightened my life again for awhile.   <P>

But you know this is not a fairy tale.  I do want to see 
my computer prove hard theorems (difficult new theorems)
in mathematics, recognize faces, diagnose diseases, reason by 
analogy, learn in many ways, engage in intelligent discourse, use
common sense, play good bridge and tennis, etc etc etc.   <P>

It is amazing how this seemed so possible in 1958, 1960, 1962. 
You know this still seems possible to me now;  that's what 
really drives me.  (Of course, some of it has 
happened).   <P>

There is a buck to be made in AI now.  And I don't mind that, in
fact the economic interest might just push forward part of the
needed research.  But I believe that it is not that buck 
(dollar, yen, Frank, Pound, Mark, Lira,Peso, ...) that drives the
few of us who can and will really make it happen.  We
march to a different drummer -  and I am proud to be part of 
that battalion that responds to that music.   <P>

I don't have to explain why I had (and have) this dream, I just do.
It does not require me to be a mathematician or an Engineer or
Computer Scientist.  I just want the results - well almost -
what I want too is to be surprised.  It does not even require that 
we do "good science" though that would probably help if it is not
overdone.   <P>

My dream is Buck Rogers, HAL, 2010, Starwars, R2D2, C3PO, the 
Turing test, Objects of the third kind, Jules Verne, all piled into
one, but with all the ridiculous things divided out, 
like breathing in a vacuum and going faster than the speed of 
light, leaving what is somehow possible by our 
present knowledge of science.  That is what is exciting, doing
the possible.   <P>

And I am not so generous;  I want to actually see these things myself 
before I die.  "Pressing science along a little bit" won't cut it with me
unless I am part of the excitement and see some of the major 
milestones.   <P>

The physical parts of the proposed computer person seemed important
in 1960.  And this still excites me now - to have a robot run,
fall and get up, act autonomously, to be a moving companion. 
Oh, I'm well aware that the real problems are those of the mind,
getting computer programs to act as if they reason, act as if they 
understand,
think, learn, plan, enjoy, hate, etc.  That is the challenge of the
age.  Some of this has happened already,  and I believe that
we have in this room the talent to bring about much of the
rest of it.  As it unfolds during the next years and decades,
let us not fail to stop occasionally to enjoy it, to "smell the
roses", to thrill as each new milestone is reached.   <P>

When I began to prepare for this talk I wondered what I would have to
say that was worth hearing.  And it was this contemplating that
brought to mind the above thoughts, this "dream".  These are things
that are important to me in my  "calling"  as an AI researcher.   <P>

These 25 years have not been totally kind to my dream:  Shaky
liked shaking more than running and thinking, and was laid aside for a season;
language translation sputtered, died, and was then resurrected;  
facial recognition was pushed back on the researcher's stack;
automatic provers showed signs of growing pains, 
which disheartened the fainthearted;  no machine stepped forward to 
try the turing test;  robot arms were duplicating block castles
instead of playing squash;  etc etc etc;  many AI researchers lost 
faith and dropped out.   <P>

But curiously I remained in the AI fold, and why?  Because these 25
years have also been fruitful and exciting.  We have much to proud of,
with much left to be done.   <P>

First and foremost we have learned what we are.  
Just as a small child
remains ineffective until it is taught (given knowledge), so it is
also with our machines.  Reasoning alone could not have enabled a  prehistoric
man to even invent the wheel, no matter how nimble his brain;  but
the space-age woman with her knowledge  of wheels, gears,
engines, computers, aerodynamics, and the like, with the same 
reasoning power can discover much more.  Because knowledge is king,
knowledge - the key to who we are.  Even reasoning itself is enhanced by
knowledge about reasoning, and knowledge about what we are reasoning
about.  Ours is in essence the knowledge business.  (Ed Feigenbaum
says that we are working on "Knowledge Application Machines".)   <P>

But you say, "Every 'smart' program, used to solve a problem, regulate
a chemical process, design a bridge, etc, has key pieces of knowledge
built into it.  So what is new about AI?"  The answer is that the
AI scientist or engineer recognizes this knowledge for what it is, 
and has, in the case of expert systems, plucked it out of the program
and placed it in a separate  "knowledge base".  Not only does the knowledge 
give the power, but it provides the functionality.  The knowledge
base acts as a new and powerful computer language,
which is used by the programmer to carry out his will.  He defines 
functionality and causes actions merely by changing this knowledge base.   <P>

So foremost we have learned that we must use knowledge, the knowledge
accumulated by mankind over these last few thousand years, if we are
to achieve these AI dreams.  And we have accomplished a great deal 
during these last 30 years;  let me mention some of it.   <P>

But first let me express my annoyance with some of our detractors
who criticise AI researchers for not  "jumping to infinity"  in one
leap.  Somehow to them it is OK to work step-by-step on the dream
of obtaining controlled thermonuclear energy or a cure for cancer
or a cure for the common cold, but no such step-by-step process is
allowed for those trying to (partially) duplicate the intelligent
behavior of human beings.  To these cynics, a Natural Language system 
which converses with us in a restricted form of English is somehow not 
a legitimate step toward passing the Turing test.  I know of no case in the
history of science where such "naysayers" actually helped with a 
new discovery.   <P>

Indeed, almost all of our AI accomplishments have been of the 
partial kind:  natural language processors which handle
a subset of English (or French, etc);  systems that recognize and
synthesize limited forms of speech;  character recognition machines
that read only typewritten characters;  expert systems which perform
a variety of tasks (but not all that a human can);   theorem provers
that can prove difficult theorems in a particular area of 
mathematics,or which can handle the inferencing needed for elementary expert 
systems, including non-monotonic reasoning;  programs that play
expert level chess;  programs that exhibit an elementary level of 
learning and reasoning by analogy.  And the list goes on.   <P>

Another key thing that we have learned and are still learning, is
the list of crucial technologies needed to continue the pursuit 
of our AI objectives. These partial results, mentioned above, 
have helped to unearth the roadblocks that stand between where we are
now and where we are trying to go with AI.  We are beginning to 
enumerate and classify these enabling technologies.   <P>

Foremost in the list is the representation and storage of knowledge,
with the added requirement that the particular design will allow:
<UL>

<LI>  Learning:  ease of acquiring and storing the knowledge <P>

<LI>  Performance:  effectiveness in using the stored knowledge
                to perform tasks, solve problems, and answer
                questions. <P>

</UL>


I believe that it is time to build large, very large, knowledge
bases.  Such a knowledge base should contain  "common sense"  
knowledge as well as encyclopedic and expert knowledge, and
be structured to handle the learning and performance 
requirements mentioned above.
(An effort of this sort headed by Doug Lenat at MCC, uses
common sense knowledge in a fundamental way and uses analogy 
to help with knowledge acquisition and problem solving.)   <P>

It is believed that such a
large structured knowledge base would not only allow the
sharing of knowledge by numerous systems, but if structured
correctly, could provide much more robustness and functionality
than is possible from a number of distinct smaller KB's.   <P>

It has been said that we cannot have true machine intelligence
until we have effective machine learning.  In that case we have a way
to go.  But a number of good researchers are beginning to make
progress in this area.  Earlier work on machine learning tended to be
too ambitious, to general, whereas the recent efforts have had more
success where the things being learned are controlled by knowledge
structures, where the machine finds values of facets within a
human-supplied framework.  But even so, until we see some real 
gain, a reasonable amplification of capacity, then some of us need to
be rethinking the learning and analogy process from scratch.
(This rethinking might also apply for some other areas of AI research.)   <P>

Twenty years ago one might have been tempted to say that it requires
only two things to build a machine which appears to think like a human
being: machine learning and natural language
understanding.  Because
such a machine can be taught, by feeding it more and more knowledge
from existing books, letting it bootstrap itself to higher and higher
levels of mental functionality.  But those two requirements are
formidable indeed.  In fact, I'm afraid that this characterization is
misleading.  It lets us believe that the major present needs of AI can
be had through machine learning.  While that might be a correct
principle for the long run, it won't do for the near term.  So we must
press on in other areas of AI as well as machine learning.  For
example, the important work on "speech acts" should be pushed now,