COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	\input jmclet
C00008 ENDMK
C⊗;
\input jmclet
\jmclet
\address
Professor Lloyd D. Fosdick
Department of Computer Science
ECOT7-7 Engineering Center
Campus Box 430
Boulder, Colorado 80309-0430
\body

Dear Professor Fosdick:

	When Paul Smolensky requested me to be a referee,
my sense of my own objectivity
 felt flattered, because I have pursued and
continue to pursue a quite different paradigm in artificial
intelligence than his own.  I agreed to do it, but said
that couldn't make the effort to become familiar with his
whole list of publications and would confine myself to those
which intersect somewhat with my own interests.

	As it has turned out, I have had even less time than
I expected and have found some of the material more difficult
than I expected.  Therefore, my comments are more limited
than I expected.

	It seems to me that I should divide my comments in
two parts: on the prospects for connectionism and on his
work in particular.

	I read carefully his paper ``The Proper Treatment of
Connectionism'' and commented on it for Behavioral and Brain
Sciences under the title ``Epistemological Problems of Connectionism''.
In it I said that I wouldn't comment on learning algorithms
but only on the adequacy of what could be represented after
the learning was complete.

	Smolensky explained that there might be a ``subsymbolic
level'' in the brain between the neural level and the symbolic
level.  Also in AI it might be worthwhile to build hardware
and programs with subsymbolic levels.  These ideas seem worth
exploring.  Although I don't intend to pursue this approach,
I feel no need to be competitive about it.  AI is hard enough
so that all plausible approaches should be explored, and it
is likely that several approaches will be successful.  I'm
still betting on the direct symbolic approach through mathematical
logic.

	I thought Smolensky's analysis of the potential relation
between the symbolic level and a subsymbolic level was mainly
correct.  His analysis of what would have to be done in order to
make a subsymbolic system do symbolic calculations also seemed
correct to me.  However, I don't share his belief that a precise
theory will only be found at the subsymbolic level.  He made
sufficiently careful distinctions between what he hoped for and
what had already been accomplished.

	The other paper I tried to read was the one on using
tensor product spaces to represent the relation between variables
and their values.  I still hope to finish it, but I'm making
a trip to Japan now and I can't give it more time.  I chose it,
because I think it defines a major requirement for subsymbolic
systems to do symbolic reasoning.

	In general I consider Smolensky an active and successful
scientist.  I have no hesitation in recommending him for
promotion.  He is sufficiently versatile and objective so that
if connectionism gets stuck he will successfully switch his
attention to something else.

	I don't know about tenure yet.  Perhaps it should depend
on the competitive situation.  If the alternative were losing him,
I'd recommend considering it seriously.

	Incidentally, asking for a review like a tenure review
for a mere continuation inflates the currency.  You are not
entitled to the same amount of work that a referee would feel
obliged to put into a tenure review.  Had I known that it was
just a continuation, I might have declined.

\closing
Sincerely,
John McCarthy
Professor
\endletter
\end