Reply to Pellionisz' "Open Letter"

[Michael Arbib]

[[ Editor's Note: Here is another response.  I appreciate the candor and
encourage others to express their opinion in a reasoned and supported
fashion. While I cannot comment on the correctness of the points either
in the original "open letter" or the various responses due to lack of
appropriate background, I am encouraged by the debate itself as process
of public education. -PM ]]

One of my colleagues sent me a copy of the letter by A. Pellionisz
complaining that Amari had not cited his earlier papers on applying
tensor analysis.  Since my reply may be of general interest, I reproduce
it here:

"Amari had sent me the original letter, and we had agreed it was a
KINDNESS to Pellionisz not to refer to his earlier work, since in doing
so Amari would have had to summarize our 1985 argument showing that
Pellionisz had misunderstood the mathematics of tensor analysis.  Anyway,
the work of Amari is NOT a generalization (why generalize a flawed
theory?!) but is a totally different application. For Pellionisz, the
tensors are the inputs and outputs to a single NN.  For Amari (applying
his work of many years on information geometry) the whole NN is an
element of the Riemannian space on which tensors are defined, and the
metric on that space is used as an information measure to guide inference
of network parameters to find a NN meeting specified criteria.  My only
criticism of Amari's paper is that it relies too much on his previous
publications directed to statisticians, and so will be very hard for NN
workers to read.  Finally, note that tensor analysis is a powerful branch
of mathematics with many applications.  The idea of applying it to NNs
does not need citation any more than does say the use of linear algebra.
But if one makes use of a specific technique in a way close to the work
of others, then full citation is appropriate.  The latter case does NOT
apply here."

Let me simply add that the Japan-bashing in Pellionisz's letter is both
distasteful and (as I need hardly add) totally without foundation.  It is
Professor Amari, not Dr. Pellionisz, who deserves a public apology.

Michael Arbib
Center for Neural Engineering
University of Southern California
Los Angeles, CA 90089-2520
USA


------------------------------
Neuron Digest   Tuesday, 10 Mar 1992
                Volume 9 : Issue 10

Today's Topics:
                   Reply to Pellionisz' "Open Letter"
             Re: Arbib's response to "open letter to Amari"
                      Re: Pellionisz' "Open Letter"
                    reply to the open letter to Amari