Systems, Man & Cybernetics - January 2015 - 23

for pattern recognition problems. The
Mark I Perceptron had a 20 # 20 pixel
image sensor, could recognize simple figures, and had 512 motor-driven
potentiometers, which were responsible for each of the variable connection weights.]
The mathematical analysis to
which they had subjected Rosenblatt's perceptron [10] was devastating: artificial neuronal networks like
those in Rosenblatt's perceptron are
not able to overcome many different
problems! For example, it could not
discern whether the pattern presented to it represented a single object
or a number of intertwined but unrelated objects. The perceptron could
not even determine whether the number of pattern components was odd
or even. Yet this should have been a
simple classification task that was
known as a "parity problem." Wasn't
this an impasse? What did you expect
to achieve in that area?
JB: To begin, I note that Dave
Block proved the first perceptron convergence theorem, I think with Nilsson at Stanford, and maybe Novikoff,
in about 1962.
[Nils John Nilsson (born 1933) is
an American pioneer in AI research.
He is the first Kumagai Professor
of Engineering (Emeritus) in Computer Science at Stanford University.
Albert Boris J. Novikoff (born 1928)
is an American mathematician. He
earned his doctorate degree at Stanford University. He is a professor at
the Courant Institute of Mathematical
Sciences at New York University.]
You can look this up; I used
to have the paper, but no more.
[Novikoff presented a convergence
proof for perceptrons at the Symposium on Mathematical Theory of
Automata, which took place 24-26
April 1962 at the Polytechnic Institute of Brooklyn; see [11].] Therefore, when I got to Cornell in 1969,
the same year that the Minsky and
Papert book came out, Dave and
others (including his best friend,
Bernie Widrow, I might add), were
in a funk about the apparent death
of NNs. [Bernard Widrow (born

1929) is an American professor of
electrical engineering at Stanford
University. Together with his Ph.D.
student Ted Hoff, he invented the
so-called Widrow-Hoff least mean
squares filter adaptive algorithm that
led to the ADALINE and MADALINE
artificial NNs and to the backpropagation technique.] Dave wanted to
continue in this field, but funding
agencies were reluctant to forge
ahead with NNs in the face of the
damning indictment (which, in hindsight, was pretty ridiculous) by Minsky and Papert.
About 1970, Richard Duda sent
Dave a draft of his book with Peter
Hart, the now and forever famous
"Duda and Hart" book, Pattern Classification and Scene Analysis, published in 1973 [12]. [Richard O. Duda
is a professor emeritus of electrical
engineering at San Jose State University renowned for his work on sound
localization and pattern recognition.
Peter E. Hart is an American computer scientist.] Duda asked Dave to
review it. Dave threw it in Joe Dunn's
in-box, and from there it made its way
to mine. So I read it from cover to
cover, trying to find corrections while
simultaneously learning the material,
and that's how I entered the field of
pattern recognition.
My best Dave Block story: In about
1971, Dave and I went over to the
Cornell Neurobiology Laboratory in
Triphammer Woods, where we met
a young enterprising neuroscientist
named Howard Moraff, who later
moved to the NSF, where I think he
still is today. [Howard Moraff is the
program director of robotics and machine intelligence at the NSF.]
Howard was hooking up various
people to electroencephalograph
(EEG) sensor nodes on their scalps-
16 sites at that time-and trying to
see if there was any information to
be gleaned from the signals. We spent
the day watching him, talking to him,
and so on. Dave was noncommittal to
Howard about the promise of this enterprise, but as we left the building,
Dave turned to me and said, "Maybe
there is some information in the sigJa nu a r y 2015

nals, Jim, but we are about 50 years
too early."
I have told this story many times
since then (43 years ago now), and I always end it by saying this: "And if Dave
could see the signals today, given our
current technology, what do you think
he would say now? He would say, 'Jim,
we are about 50 years too soon'."
Therefore, the bottom line for me
in 1971 was: don't do NNs, but clustering and classifier design with other
paradigms is okay. As it turned out,
however, I was out of the frying pan
of NNs and into the fire of fuzzy sets,
which was, in effect, a (very) rapid descent into the maelstrom of probabilistic discontent.
RS: Enrique and Jim, you both mentioned the difference between the
classical (or probabilistic) and the
fuzzy approach to clustering. Jim, at
the end of your last answer, you hinted at the keen competition between
fuzzy sets and probability theory,
and your Ph.D. dissertation is titled
"Fuzzy Mathematics in Pattern Classification" (Figure 3) [2].
Enrique, in your answer, you mentioned the discussions still ongoing
today, and in your 1969 paper "A New
Approach to Clustering," you wrote
that your "new method of representation of the reduced data, based on
the idea of 'fuzzy sets,' is proposed
to avoid some of the problems of current clustering procedures and to
provide better insight into the structure of the original data" [8, p. 22]. To
clarify this point: can you both please
try to explain why you favored the
new mathematical theory of fuzzy
sets to approach data reduction and
cluster analysis instead of the classical approaches of that time?
JB: As a graduate student, your
job is to find something new and sufficiently interesting and that you can
do. I did not consciously choose fuzzy
models over probabilistic ones. The
first 600-page draft of my Ph.D. dissertation had an entire chapter about
the connection between the probabilistic and fuzzy approaches that was
removed to shorten those original

IEEE SyStEmS, man, & CybErnEtICS magazInE

Table of Contents for the Digital Edition of Systems, Man & Cybernetics - January 2015