COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.require "memo.pub[let,jmc]" source
C00010 ENDMK
C⊗;
.require "memo.pub[let,jmc]" source
Comments on John Haugeland's "The Nature and Plausibility of Cognitivism" 

	Haugeland's description of the attractions of cognitivism convinced
me that I have long been a cognitivist, although
I would not try to give cognitive explanations to
all mental phenomena.  His challenges to cognitivism are interesting,
and here are some tentative ideas for meeting them.

	We first distinguish %2competence cognitivism%1 from %2performance
cognitivism%1.  Competence cognitivism is tries to relate the
output of a mental process to its inputs as the result of a deduction
to its premises.  Performance cognitivism regards the actual mechanism
as a process of deduction.  I would support performance cognitivism for
some mental phenomena and competence cognitivism to many more.
Arguments that the human brain obtains some
of its results faster than can be accounted for by deductive processes are
proper challenges to %2performance cognitivism%1 but not to %2competence
cognitivism%1.
Thus we may relate the output of a skilled action
cognitivistically to its input without postulating a deductive internal
mechanism.

	Performance cognitivism needs more precision, because
it isn't clear when a mechanism is to be considered
cognitive even in machines.  In particular, pattern
matching mechanisms are quite varied, and some of them are closer than
others to deductive reasoning.

	Here is a conjecture about "understanding".  A person can deal
with a symbolic input in either of two ways.  First, he can manipulate it
according to rules he has learned, and second, he can translate it into
his "internal language".  Only in the latter case can he combine it freely
with other information to draw conclusions.  We say that someone has
information but doesn't understand it when he can manipulate the symbols
as objects but hasn't translated enough of it into his "internal
language".  The main symptom of lack of understanding, is a failure to
draw certain conclusions from symbolic information that seem "obvious" to
those who do understand.  This phenomenon offers no problems to
cognitivism of either kind provided we distinguish symbolic expressions
from the information they express.

	The most difficult of Haugeland's challenges to cognitivism is the
problem of how moods affect reasoning.  It is most acute if we imagine
moods to be chemical; e.g. a melancholy mood is just a high concentration
of %2melancholine%1 in the blood - together with its effects.  How can
this affect what long term goals seem worth pursuing - as moods often do?
My proposals here are tentative and certainly require revision.

	The presence of %2melancholine%1 might cause certain sources of
sentences to come into consciousness with higher probability, or to
present themselves with greater urgency, or to come with greater
frequency.  This assumes that the information retrieval system that
determines what comes into consciousness is partly non-logical.  Likewise
it inhibits other sources of sentences.

	The presence of %2melancholine%1 might give rise to a sentence %2I
am melancholy%1.  Perhaps this sentence has no special effect by itself,
but can trigger learned effects.

	The sentence %2"I am melancholy"%1 might be used logically by the
information retrieval mechanism.  Thus there might be rules saying
something like %2"If I am melancholy, I should list things that might go
wrong"%1.  This might work but seems subjectively implausible.

	All these hypotheses require an explanation of a
phenomenon which occurs in many other contexts.  A certain set ⊗A of
sentences can lead to a conclusion ⊗p.  Another set ⊗B consistent
with ⊗A and maybe even containing ⊗A can lead to the conclusion ⊗¬p. 
(McCarthy 1978) ascribes this to a method of conjectural reaoning
we call %2circumscription induction%1.  Briefly, applying circumscription
induction involves jumping to the conclusion that the objects whose
existence follows from a set ⊗A of sentences are all the objects
there are in a certain class.  This conclusion may lead in turn
to the conclusion ⊗p.  Enlarging ⊗A to ⊗B may bring in new objects,
and this may lead to ⊗¬p.  Without some such mechanism we wouldn't
be able to draw contradictory conclusions from consistent sets of
information merely by neglecting part of it.

Reference:

%3McCarthy, John%1 (1978), %2Circumscription Induction - A Way
 of Jumping to Conclusions%1,
Stanford Artificial Intelligence Laboratory, (to be published).
.begin verbatim

John McCarthy
Artificial Intelligence Laboratory
Stanford University
Stanford, California 94305
.end