CS378 Assignment #2
			      26 Jan 94
			  Christian Mogensen
		       mogensen@cs.stanford.edu


ANALYSIS OF SUBMITTED PAPERS.

Raymond Tan

The author jumps right in, defining intelligence in terms of the Turing
test.  Despite the misgivings he cites, he clings to Feigenbaum's
notion that a specialized system, given enough additional knowledge,
will exhibit general intelligence.  The subsequent argument rests on
this assertion being valid.

I cannot agree with this: a chess playing program does not contain the
structures necessary to develop general intelligence - it has no way of
representing "it rained today" in its theory of game playing.  Consider
a toaster.  You can add all sorts of geegaws to a toaster, but you cannot
turn it into a house or even a kitchen, or even a stove.  Maybe a small
ineffectual stovetop or a toasteroven, but nothing good would come of it.

The other core problem here is language:  how do you express "it is raining"
in terms of chess moves?  The machine must become proficient at language
if it is to qualify for the Turing test, but how this happens is vague.

The author writes "I would first give the machine a knowledge of the rules
of the game..."  How should this be done?  How does on encode knowledge,
and how is this knowledge different from information?  On another level,
this sentence reveals the conduit metaphor hard at work, channeling 
packets of knowledge to the machine.

Later we get "after the machine has been trained to a reasonable level
of proficiency, a large body of knowledge is then supplied to the machine's
program so as to ... increase its 'intelligence'."  Now does this strike
anyone as a huge handwave?  Stuffing the encyclopedia down the toaster
does not make the toaster smart.  

We then have Schank's ideas on language.  The example used is "Burt gave
Joe a black eye for calling him a name."  This is supposed to be encoded
as BURT PROPEL FIST -> EYE OF JOE.  Ignoring the fact that fists were never
mentioned, I would like to point out that this encoding is bogus.
Burt actually had an eyeball he had painted black, which he gave to Joe.

The point is that language is ambiguous, and explicit encodings in semantic
nets quickly collapse as one hangs more and more information on them.
Similarly, the 16 classes of inference about sentences don't work in all
cases.  How the machine is supposed to repair misunderstandings is not
discussed.

We end up constructing overburdened chess machines.  A chess machine with
some encoding of "rain" is still a chess machine. 

This paper confirms my prejudices about AI rather than call them into 
question.  To put it another way, the arguments are not strong enough
to convince me I am wrong.

		       =======================

Designing the Future
Jennifer Hughes

The author starts well by nailing the western focus on duality and
binary belief-systems. We must change our beliefs if we want to create 
an intelligent machine, because the current focus on duality is hindering
our ability to analyze the problem.

Best quote:
"The computer cannot learn Hindu simply through contact with it."

We are given the tools to think anew and build an extension of 
ourselves.  The author talk of "something organic, an interceder
between psyche and world".  

The paper has made me go back and reanalyze my own beliefs in the light
of focus on duality as 'Zen and the Art of Motorcycle Maintenance' did
when I read that.


		       =======================
Nat Johnson

The paper makes the point that intelligence is a sensory experience as
well as a mental one.  A intelligent machine requires rich sensory
data to avoid insanity.  Intelligence require good filtering of the
stream of data.  Just as the human mind throws away most perceived 
data within a second, so the computer must find out what to focus on in
a data-torrent.

the dichotomy between the world and logic is pointed out.  The author's
suspicion that common sense is logical is (ironically) incorrect.  
According to some experiments, common sense misleads us in logical 
problems, such as conditional probability or lotteries.

The paper gives us an idea of the problem areas involved in building an
intelligent machine.  We learn some of the requirements for intelligence
and discuss a possible method for dealing with such a huge project.
The basic impression is that this is a really hard problem.

There is little in this paper that questions my assumptions or ideas.
I agree with most of it.

		       =======================
Trace Wax
Merging Man and Machine

The discussion is nicely framed as a dialog.  Unfortunately the paper does
not take its ideas far enough (IMHO).  As Jennifer Hughes points out, 
man and machine complement each other and form an intelligent system.
However, where Hughes moves the boundary between us and the machine to
encompass the machine, the author retains the self as a distinct entity.
The machine is a separate thing.

The paper implies a Turing definition of intelligence: if it does what
you expect, then it is intelligent.  This is a slippery definition and
one I disagree with.  The fact that the author then goes on to say
that we must adapt our language to the machine's expectations speaks
against its intelligence.  If the machine isn't smart enough to understand
us, how can it be intelligent?

The phrase 'humanlike character' implies a lot of work on the machine's
part, a lot of understanding of the human's context.  In addition the
machine is supposed to entertain us.  Hopefully it will be smart enough
to understand when don't want to be entertained.  This vagueness is
not encouraging.

I wonder if computers that understand humans more and more will come to
have communication problems amongst themselves.  As there is more and more
context to be shared, there is less and less time for actual 
communication.

I am not sure what one could build with this set of instructions.  I suspect
something like a more complex version of Microsoft BOB would be the result.
The one thing that stands out is the emphasis on a user model.  The fact
that the user model must adapt to changing circumstances is vital.  

I think it's hard to question my essay because it is mostly questions and
there are few answers.  I liked the style of this paper, and wonder if I
could redo the questions in my essay as a dialog.  I do not place
enough emphasis on the shared understanding and mental models in my paper.

		       =======================