COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	\input buslet[1,ra]
C00006 ENDMK
C⊗;
\input buslet[1,ra]
\jmclet
%\vskip 5pt
\address 

Prof. Terrence J. Sejnowski
Department of Biophysics
Johns Hopkins University
Baltimore, MD 21218

\body
Dear Terry:

	I found your ``Learning and Respresentation in Connectionist Models''
very thought-provoking, and here are some of the thoughts it provoked.

Both when the model of reading aloud is 
considered as AI, the following human ability should be considered.
One can be told that in the new Chinese
latin spelling the letter Q is
pronounced $//C//$, and that in French the letter G before A is
pronounced $//g//$ and before E is $//J//$ and use these facts 
immediately in reading aloud immediately.  Therefore, it seems unlikely that
the information about reading aloud is stored solely in connections.
One can imagine that one interprets the information the 
the fly, if there isn't too much of it, and uses this practice to train some
network.  One can also imagine a more elaborate network that has rules
as additional inputs.  In any case, the ability to be told as well as to be
trained requires discussion. 

	At the phonetic end it is more difficult to be told 
about new sounds or subtle conventions. Maybe this is only because we learn
phonetic skills at the age of two. Certainly one can learn the ability to be told
phonetic distinctions and then use them. I remember being surprised having 
difficulty teaching my daughters at ages approximately 7 and 12 the 
voiced-unvoiced correspondences.
However, once it was brought to their attention, the
7 year old learned it almost as fast as the 12 year old.

	Enclosed is some information about AAAI support of workshops.

	I enjoyed our discussions in Paris.

\closing
Sincerely,       

John McCarthy    
\annotations
%\vskip 1in
%Enclosure
%\vskip 1in
JMC/ra 
\endletter
\end