A Sociological Study of the Official History of the Perceptrons Controversy

Author(s): Mikel Olazaran

Source: Social Studies of Science, Vol. 26, No. 3 (Aug., 1996), pp. 611-659
Published by: Sage Publications, Ltd.
Stable URL: https://www.jstor.org/stable/285702
Accessed: 21-05-2019 15:48 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Sage Publications, Ltd. is collaborating with JSTOR to digitize, preserve and extend access to
Social Studies of Science

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 * ABSTRACT

In this paper, I analyze the controversy within Artificial Intellig

which surrounded the 'perceptron' project (and neural nets in gen
the late 1950s and early 1960s. I devote particular attention to the
and arguments of Minsky and Papert, which were interpreted as s
that further progress in neural nets was not possible, and th
approach to Al had to be abandoned. I maintain that this official
interpretation of the debate was a result of the emergence,
institutionalization and (importantly) legitimation of the symbolic Al
approach (with its resource allocation system and authority structure). At
the 'research-area' level, there was considerable interpretative flexibility.
This interpretative flexibility was further demonstrated by the revival of
neural nets in the late 1980s, and subsequent rewriting of the official
history of the debate.

A Sociological Study of the Official

History of the Perceptrons Controversy

Mikel Olazaran

The recent sociology of scientific knowledge has sh

processes of controversy often play a central role in the pro
and validation of scientific knowledge.1 Harry Collin
mended the study of 'interpretative flexibility' (that is, var
scientists' perceptions of the same results or experimen
methodological starting point in controversy studies
consensus is reached, groups of scientists from diverse t
(with their own cultures, interests and connections with
wider scientific community and the wider society) may
the same experiment (phenomenon, result, method or t
differently. Showing the interpretative flexibility of scientif
amounts to the realization that no knowledge possesses
warrant, whether from logic, experiment or practice;
always be grounds for challenging any knowledge claim.
In scientific practice, interpretative flexibility is red
processes of accumulation of cognitive and social resourc

Social Studies of Science Copyright ? SAGE Publications (London

Oaks, CA and New Delhi), Vol. 26 (1996), 611-59
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 612 Social Studies of Science

factors like the following play an important role: communication

and interaction between research specialties and disciplines, cross-
fertilization; interaction between scientific and technological
contexts (in this case, information technology and program-
ming); accumulation of organizational, institutional and rhetorical
resources; the science-policy context; and the wider scientific and
technical culture.4 By studying the effect of social factors - both
internal and external to the scientific community - on the pro-
cesses of closure of controversies, sociologists have exposed the
contingent elements in the production and validation of scientific
knowledge.5
Specific cognitive objects (such as certain 'crucial' experiments,
results or 'proofs') often play an important role in the evolution of
scientific controversies. The key move in the controversy analyzed
in this paper was the decision by Marvin Minsky and Seymour
Papert to replicate the 'Perceptron machine' built by a team led by
Frank Rosenblatt, with a view to showing its limitations. As
Collins pointed out, replication of this kind is quite unusual in
science, and it occurs only when the claim under discussion is
particularly important.6 The 'interpretative flexibility' of Minsky
and Papert's results was considerable. Standards of proper experi-
mentation and criteria of competence had not by then been
agreed, and experimental work relating to the controversial issues
was not equally compelling to all those involved in the debate.
As Trevor Pinch has pointed out, the construction of a disputed
cognitive object can be used as the defining point around which
differing groups of scientists taking part in a controversy can be
identified.7 These groups emphasize different dimensions of what
we can safely assume to be the 'same' object. Disputed cognitive
objects can be articulated at different levels. Following Pinch, I
will consider two modes of articulation, the 'research-area' mode
and the 'official-history' mode.8 The research-area mode of
articulation is used when the disputed object is part of the
immediate area of concern and practice of the scientists involved
in a controversy, whereas the official-history mode is used in
historical accounts of how a particular field evolved. The multi-
dimensional character of the object in the research-area mode (the
possibility of working on different aspects of it) is lost at the
official-history level, where results and proofs are regarded as
either valid or invalid. The official-history mode occurs mainly
at the informal level of communication in science. Using some con-
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 613

cepts and elements from the work of Nicholas Georgescu-Roegen,

Richard Whitley, Pierre Bourdieu and others, Pinch showed that the
official-history mode of articulation, with its legitimating functions,
often plays a very important role in scientific controversies and in the
underlying 'battles' for authority in science.
In this paper, I reconstruct the controversy which surrounded a
specific 'cognitive object' - namely, certain proofs and arguments
which apparently showed that progress in perceptron research was
not possible.9 The structure of the paper and its main arguments
are as follows. After a brief presentation of the neural-net
approach and its antecedents, I will look at Frank Rosenblatt's
Perceptron machine and the controversy which surrounded it.
Then I will analyze some of the main technical problems and
limitations of early neural nets. I then discuss 'impossibility' proofs
and arguments and the process of closure of the controversy,
which I reconstruct by linking the developments analyzed pre-
viously with the disciplinary, technological and funding contexts.
The emergence, institutionalization and legitimation of symbolic
AI as a research specialty was the most important factor in this
process. Finally, I will examine the process of accumulation and
cross-fertilization which has recently brought about the revival of
neural nets.

According to the official history of the controversy, in the mi

1960s Minsky and Papert showed that progress in neural nets
not possible, and that this approach had to be abandoned. In
paper I try to show that this official view emerged as a result of
closure of the perceptrons controversy. Before that, things w
not so clear at the research-area level. And neural nets were not
completely abandoned: a few researchers continued working in
this area, but they were displaced from artificial intelligence (AI
to other disciplines.
Collins has used the 'things could have been otherwise' argu-
ment in order to show the interpretative flexibility of scientifi
results. The curious thing about the perceptrons controversy is
that things were really otherwise in the recent revival of neural
nets, which I will examine at the end of the paper. As neural net
emerged as an accepted specialty, the official history of th
controversy was rewritten in order to legitimate the new socia
and cognitive structures of the AI discipline. The interpretative
flexibility of Minsky and Papert's impossibility proofs was reopened.
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 614 Social Studies of Science

Neural Nets

Neural networks are information-processing systems com

many interconnected processing units (simplified 'neuron
interact in a parallel fashion to produce a result or outpu
are called 'neural' because, in designing them, researc
'inspired' by some (sometimes only a few) simplified fea
information processing in the brain. The massively parall
tecture of these systems is remarkably different from
conventional (also called 'von Neumann') digital com
Neural nets are not programmed, but 'trained'. Training
net in some classification task involves selecting a sta
representative sample of input/output pairs, and algorit
adjusting the strengths (or 'weights') of the connections
processing units when the system does not produce the
outputs. Neural-net training is usually a long (and compu
expensive) process of cycles of input feeding, output ob
and weight adjustment.
The neural-net approach differs from the tradition w
dominated AI in the last decades, namely the symbol-pr
perspective. Within symbolic AI, intelligence and cognit
seen as processes of symbol manipulation and transform
symbolic system relies on its representational structure
the possibility of applying structure-sensitive operations
Representational structures are manipulated and tran
according to certain rules and strategies (embodied in co
programs), and the resulting expression is the solution to
problem.
Researchers expect neural nets to have considerable success in
tasks not easily programmable so far within the rule-based symbol-
processing approach, such as pattern and speech recognition. The
learning capabilities of neural nets may be especially important for
this type of task. Each unit in a neural-net system performs a
simple processing operation which can be divided into three parts:
input addition, comparison with a threshold value and, if that
value is equalled or surpassed, 'firing' or output activation (see
Figure 3, overleaf). Figure 1 shows one of the most popular neural-
net architectures - namely, the one formed by strata of units and
connections (also called 'multilayer feedforward' net, because
activation always spreads in the direction from input to output).
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 615

FIGURE 1
Multilayer Network

hidden

input ip units
units output units

The architecture of neural

von Neumann architecture,
computers. One of the main
architecture is the separatio
Von Neumann computing co
after another (that is, seque
binary expressions which ar
These transformations are m
or rules (the program) whic
basic operation of a von Neu
steps: localizing an expression
central processing unit; tran
different location of the me
Von Neumann memory is
locations, resembling a list
would stand for symbolic ex
certain expression in this li
Neural nets work rather dif
discrete locations, but distr
meters. The 'knowledge' that
of its evolution (that is, the
activation state of the proces
interconnecting them. In a
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 616 Social Studies of Science

expressions do not exist as such, their 'equivalent' (this word has

to be used cautiously) being overall patterns of activation emer-
ging from the parallel interaction between many units at the
subsymbolic level.
One of the most interesting properties of distributed, associative
memories is graceful degradation: in certain circumstances, the net
can perform at an acceptable level even if some of its processing
units do not work properly (in this, neural computing contrasts
sharply with conventional computing and programming). Another
important property of neural nets is their ability to recognize
whole patterns (or objects), even though only a part of them is
presented as an input (or when the image is distorted or in-
complete).
Neural-net research is an approach to AI and cognitive
science.l1 These related disciplines both aim at building intelligent
machines (that is, computer programs and simulations which carry
out intelligent or cognitive tasks) and at studying (that is, con-
structing and testing theories of) perception and cognition using
computational methods and tools. AI's emphasis is on building
intelligent machines, whereas cognitive science - a highly
interdisciplinary field of research - concentrates on understand-
ing cognition.
The origins of AI go back to the cybernetics movement of the
1940s and 1950s. This movement started around the idea that the
functioning of many systems, both live and artificial, can be better
understood with models based on information processing an
transfer, rather than on energy transfer. Researchers aimed at
studying the elements that automatic machines and the human
nervous system have in common - what they called 'control and
communication processes both in the animal and in the machine
To address this question, an important interdisciplinary effort wa
made, with contributions from areas like mathematics, formal
logic, computer science, psychology, electrical engineering, physi
ology and neuroscience. The foundations of cybernetics were buil
and explored by leading scientists, including Alan Turing, Warren
McCulloch, Claude Shannon, Norbert Wiener, John von Neu-
mann and Kenneth Craik.
An important aspect of the cybernetics movement was
existence of different approaches to the issue of the relations
between brain (or mental processes) and machine. During
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 617

second half of the 1950s, symbol-processing and neural nets were

emerging as the two main approaches to both studying cognition
computationally (today's cognitive science) and building intelli-
gent machines (today's AI).
The Dartmouth Conference (Hanover, New Hampshire, held as
a Summer School in 1956) is usually taken as the starting point of
symbolic AI. The emergence phase of this approach ended
towards the mid-1960s, when it entered a period of institutiona-
lization and development.12 Computers had first been used for
numerical calculation purposes, but symbolic AI exploited their
capability for implementing symbol manipulation. In these
systems, symbolic expressions stand for words, propositions and
other conceptual entities. The symbol-processing approach is
based upon the possibilities that computers offer for storing and
processing symbolic expressions. Computers are much better than
human beings at storing large quantities of symbolic expressions
and processing, manipulating and transforming them in ways
sensitive to their logico-syntactical structure. The representational
structures contained in a symbolic AI system are manipulated
according to certain rules and strategies (programs, algorithms,
heuristic rules), and the resulting expression is the solution to a
given problem or task. Here information processing occurs at the
representational level (its human equivalent would be mental
processes), and not at the neurobiological (or brain) level. Sym-
bolic AI systems simulate human mental and cognitive processes
by computational (digital, von Neumann) means. Among the most
important researchers of early (and contemporary!) symbolic AI
were John McCarthy, Allen Newell, Herbert Simon and Marvin
Minsky.
But since the early 1950s, in a process which accelerated towards
the late 1950s, some researchers had been exploring and develop-
ing a different, non-symbolic approach to AI: the so-called neural-
net perspective. These scientists and engineers did not seek to
model real neural networks as studied by neurophysiology or
neurobiology; rather, they were trying to build computational
architectures bearing some resemblance to the brain's nets of
neurons. These systems were being built employing McCulloch-
Pitts artificial or formal 'neurons', connected to each other by links
with modifiable links or 'weights' (Donald Hebb's notion of
learning by modifying the connections between neurons was
foundational in this respect).
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 618 Social Studies of Science

The Perceptrons Controversy

Single-layer Machines

In the late 1950s and early 1960s groups from several universities
and laboratories carried out research and implementation projects
in neural nets. Among the most important projects were those
headed by Frank Rosenblatt (Cornell University and Cornell
Auronautical Laboratory, CAL), Bernard Widrow (Department
of Electrical Engineering, Stanford University) and Charles Rosen
(Stanford Research Institute, SRI).
The number of neural-net projects or groups is difficult to
quantify. In their critical study of neural nets (analyzed later in this
paper), Minsky and Papert alleged that, after Rosenblatt's work,
there were perhaps as many as a hundred groups (in an interview
conversation this number went up to 'thousands').

Rosenblatt's (1958) [perceptron] schemes quickly took root, and soon there
were perhaps as many as a hundred groups, large and small, experimenting with
the model either as a 'learning machine' or in the guise of 'adaptive' or 'self-
organizing' networks or 'automatic control' systems.13

This issue is an important one in the official-history mode of

articulation of the controversy, where it is alleged that Minsky and
Papert had to react to stop such a great wave of 'misled' projects.
Later I will analyze this issue as a part of the official history. For
now it is important to point out that, even though there were
not so many projects, neural-net research was one of the main
cybernetic approaches to the brain-machine issue, and was taken
up very seriously by a significant number of groups and individuals.
This can be shown by looking at the scientific meetings of the
time, like the 'Mechanisation of Thought Processes' symposium,
organized by the British National Physical Laboratory in November
1958, and the 'Self-Organization' conferences held in 1959, 1960
and 1962.14
Early researchers made important scientific contributions,
especially regarding single-layer neural nets (these were systems
with one layer of modifiable connections, although they could
have more layers of fixed connections). The most famous machine
of this period was Rosenblatt's Perceptron, which is represented
in Figure 2. This machine had two layers of connections, but
only those from association units to output units had adjustable
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 619

FIGURE 2
Perceptron

input retina association units

output
--~ units

order = 6
modifiable
connections

weights. The machine built by Rosenblatt's group at C

eight response units, but only three of them are represe
Figure 2. The maximum number of incoming links receive
association unit (or 'order') of this system is six. Later I w
the importance of this issue.
As Figure 3 shows, an output unit fires if the sum
activation it receives from other units equals or exc
threshold value. Note that input activation (v) is multiplied
values or weights (w) of the connections.
The question of learning was very important in early neu
(these systems are not 'programmed' in the sense of conve
computers). In order for a perceptron-like net to impr
performance in some classification task, the modifiab
nections have to be adjusted according to a rule (or le
algorithm). In 1960, teams led by Frank Rosenblatt,
Bernard Widrow and Marcian Hoff, developed two very i
ant learning algorithms for single-layer neural nets.15 Ro
showed that, if a perceptron was physically capable of perf
a classification task (that is, if its parameters were cap
embodying that task), then it could be 'taught' that task in
number of training cycles.16 A training cycle involves present
of a pattern, observation of the output given by the mach
adjustment of the connections according to an algorithm.
The perceptron convergence theorem was proved for the
fied perceptron of Figure 4 (representing the adjustable part o
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 620 Social Studies of Science

FIGURE 3
Processing Unit

inputs

vi

\W1 threshold
V2 summing element
device

V3 :) output

* h
n

original perceptron after removing the fixed sensory-to-association

connections). This algorithm says that, for learning to occur, it is
necessary that the perceptron architecture be capable of embody-
ing the desired input/output classification. But proving whether a
classification can be carried out by the simplified perceptron of
Figure 4 (let alone Rosenblatt's Mark 1 Perceptron, which had a
first layer of randomly wired connections) is an NP-complete
problem - that is to say, it is exponentially intractable (the time it
takes to solve it grows exponentially with the size of the problem).
Thus, although the perceptron rule is a powerful learning algorithm,
training a single-layer neural net in a classification task is very
much an empirical, experimentation-based matter (where factors
like the input/output training sample used and the generalization
abilities required after training are very important).

The Rhetoric of the Debate

Controversy increased as Rosenblatt's work began to gain notori-

ety in the late 1950s. Frank Rosenblatt, a psychologist at Cornell
University, was the central figure of the early neural-net move-
ment, both from a scientific and from an organizational point of
view. He designed and studied the Perceptron, which was imple-
mented at CAL (Buffalo, New York; now the Arvin Calspan
Advanced Technology Center), but he was also the charismatic

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 621

FIGURE 4
Simplified Perceptron

input units

v1 _ output units

03

V -*
n

leader and most enthusiast advocate of neural nets, both within

the scientific community and in the wider society.
The Perceptron Project was funded by the US Office of Naval
Research (ONR). Rosenblatt and ONR presented it at a press
conference held in Washington on 7 July 1958. The statements
made by Rosenblatt there, which were widely reported in the mass
media, heated the controversy. The following report from The
New York Times is an example:

The Navy revealed the embryo of an electronic computer today that it expects
will be able to walk, talk, see, write, reproduce itself and be conscious of its
existence. Later perceptrons will be able to recognize people and call out their
names and instantly translate speech in one language to speech and writing in
another language, it was predicted.17

According to the official history of the controversy, Rosenblatt's

'overclaims' irritated many people in the AI community, including
some of its leaders.

Present day researchers remember that Rosenblatt was given to stead

extravagant statements about the performance of his machine. 'He was a
agent's dream', one scientist says, 'a real medicine man. To hear him tell i
Perceptron was capable of fantastic things . .,.18

Critics accused Rosenblatt of not having respected scien

standards and of having used the media in a partisan way. T
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 622 Social Studies of Science

following interview quote by Marshall Yovits, who was responsible

for the funding of the Perceptron Project at ONR, is interesting in
this respect:

Many of the people at MIT [referring to the symbolic AI leaders] felt that
Rosenblatt primarily wanted to get press coverage, but that wasn't true at all.
As a consequence many of them disparaged everything he did, and much of
what the Office of Naval Research did in supporting him. They felt that we were
not sufficiently scientific, and that we didn't use the right criteria. That was just
not true. Rosenblatt did get a lot of publicity, and we welcomed it for many
reasons. At that time, he was with Cornell Aeronautical Laboratory, and they
also welcomed it. But at ONR - as with any government organization - in
order to continue to get public support, they have to have press releases, so that
people know what you are doing. It is their right. If you do something good, you
should publicize it, leading then to more support.'9

Controversy was rather bitter at times, as scientific arguments,

rhetoric and organizational pressure were combined in the process
of the debate.

The campaign was waged by means of personal persuasion by Minsky

Papert and their allies, as well as by limited circulation of an unpublis
technical manuscript (which was later de-venomized and, after further ref
ment and expansion, published in 1969 as the book Perceptrons).20

The interview quote below, by Charles Rosen (from the

group, one of the most important neural-net centres of the tim
is another indicator of the tenseness of the debate:

Minsky and his crew thought that Frank Rosenblatt's work was a waste of t
and they certainly thought that our work at SRI was a waste of time. Min
really didn't believe in perceptrons, he didn't think it was the way to go. I k
he knocked the hell out of our perceptron business.21

Sociologists have shown that rhetoric is an inherent element

discourse and practice in scientific controversies.22 And, of cou
scientists use rhetoric when they present and justify their projec
outside the scientific community, as was the case with Rosenb
Because of the nature of AI, rhetoric has always been particul
controversial in this discipline.23 The so-called 'Dreyfus affair
one of the most interesting examples, although there are m
others. Hubert Dreyfus, a professor of philosophy at the Univ
ity of California, Berkeley, carefully studied the predictions m
by symbolic AI researchers in the 1950s and 1960s, and compa
them with the results which were really obtained. The rheto
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 623

studied by Dreyfus included symbolic AI leaders Allen Newell and

Herbert Simon's famous 1957 claims that, within ten years,
computers would win the world chess championship, compose
aesthetically valuable music, discover and prove an important
unknown mathematical theorem, and that most psychological
theories would take the form of computer programs.

There are now in the world machines that think, that learn and that create.
Moreover, their ability to do these things is going to increase rapidly until - in
a visible future - the range of problems they can handle will be coextensive
with the range to which the human mind has been applied.24

In 1965, Dreyfus wrote a much circulated mimeograph paper

which, in 1972, became the basis of his famous What Computers
Can't Do book.25 Dreyfus criticized some of symbolic AI's claims
from a philosophical point of view. Basically, he argued that the
digital, formalized and rule-governed nature of AI was inadequate
to model truly human intelligence (with its fuzzy, intuitive,
phenomenological and gestaltic aspects). Dreyfus' work provoked
a strong reaction from the symbolic AI community, and some
interesting and heated debates followed.26
In the perceptrons controversy, the contending views were often
represented by Rosenblatt and Minsky. They were not only the
leaders or spokesmen of the contending positions, but also two of
the most important members of the 'core set' of the controversy.27
Their famous confrontations have been reported in historical
accounts of AI.

Another who was irritated by Rosenblatt was Marvin Minsky, perhap

Rosenblatt's Perceptron was not unlike the neural-net approach Mins
alternately intrigued and frustrated by. Many in computing remember
spectator sport the quarrels Minsky and Rosenblatt had on the platf
scientific conferences during the late 1950s and early 1960s.28

Problems and Limitations of Early Neural Nets

Rosenblatt was aware of the problems and limitations

Perceptron machine, and acknowledged them in his paper
machine could not adequately detect similarities between f
because it classified objects according to the amount of ove
intersection in the input retina.29 Preprocessing (distinguishin
components of an image and the relationships between the
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 624 Social Studies of Science

another related problem. Lacking an adequate preprocessing

system, a set of association units had to be dedicated to the
recognition of each possible object, and so an excessively large
layer of association units was needed.30 Other limitations were
excessive learning time and lack of ability to separate parts in a
complex environment (Rosenblatt included here the figure-ground
or 'connectedness' problem, later analyzed by Minsky and
Papert).31
Rosenblatt studied more complex architectures: nets with two
layers of association units,32 'cross-coupled' nets (which had
connections among the units of the same layer), and multilayer
nets. He claimed that perceptrons' generalization capabilities
improved considerably with these changes,33 but he admitted that
very important problems concerning multilayer and 'cross-
coupled' nets remained to be solved. Rosenblatt summarized the
limitations of perceptrons in a list of fifteen problems, some of
which are reproduced below:
A number of perceptrons analyzed in the preceding chapters have been
analyzed in a purely formal way, yielding equations which are not readily
translated into numbers. This is particularly true in the case of the four-layer
and cross-coupled systems, where the generality of the equations is reflected in
the obscurity of their implications .... Those problems which appear to be
foremost at this time include the following: (1) Theoretical learning curves for
the error correction procedure .... (2) Determination of the probability that a
solution exists for a given problem .... (3) The development of optimum codes
for the representation of complex environments in perceptrons with multiple
response units. (4) Development of an efficient reinforcement scheme for
preterminal connections. ... (7) Theoretical analysis of convergence-time and
curves for adaptive four-layer and cross-coupled perceptrons .... (12) Effect of
spatial constraints in cross-coupled systems (e.g., limiting interconnections to
pairs of association units with adjacent retinal fields). Studies of possible figure-
segregation (figure-ground) mechanisms. (14) Studies of abstract concept
formation, and the recognition of topological or metrical relations ....34

Rosenblatt's most pessimistic comments were for problems 13

(connectedness) and 14 (recognition of topological relationships
and abstract concepts).
These two problems [13 and 14]... represent the most baffling impediments to
the advance of perceptron theory in the direction of abstract thinking and
concept formation. The previous questions [from the 1st to the 12th] are all in
the nature of 'mopping-up' operations in areas where some degree of perform-
ance is known to be possible.... [However] the problems of figure-ground
separation (or recognition of unity) and topological relation recognition
represent new territory, against which few inroads have been made.35
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 625

FIGURE 5

5.1 'And' Function 5.2 'Exclusive Or' Function

input

input

1
output output
1 .5> )

In points
multilaye
simply b
more advanced machines:

In the case of problem 4 ... simulation studies seem to be indicate

preliminary exploration, although it is hoped that some theoretical form
may ultimately be achieved .... The seventh question again is a th
one, although preliminary results obtained from simulation program
prove enlightening.36

The limitations of single-layer neural nets can be illustrated

a simple example (the simplest one possible). Figure 5.1 sh
simple net composed of two input units and one output un
easy to see that this net can compute the conjunction (or
logical function. The output unit fires only when it r
activation from both input units (only in this case is the
input activation bigger than the threshold value, 1.5). B
parameters of the system of Figure 5.1 (the values of the
tions and the threshold value) cannot support functions wh
not linearly separable, such as exclusive disjunction or 'ex
or'.37 The system should fire when presented input pairs (1
(0, 1), and should not fire when presented inputs (1, 1) and
But if inputs (1, 0) and (0, 1) exceed the threshold value, then i
(1, 1) will exceed it too, and the system will fire. As Figu
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 626 Social Studies of Science

shows, an intermediate or 'hidden' unit is necessary in order

to realize exclusive-or. This hidden unit would produce strong
inhibitory activation (-2) when the (1, 1) input pair is presented to
the net.
Early researchers were aware that multilayer systems had much
more classification capacity than single-layer ones, but they could
not find powerful weight adjustment rules for them.38

For example, the 'and' [function] . . . can be realized with the [single-layer]
linear-logic circuit . . . while the exclusive-or [functions] . . . require a cascade
linear logic arrangement [hidden units] .... [The limitations of single-layer
networks] are extremely severe . . . since the percentage of realizable logical
functions becomes vanishingly small as the number of input variables increases.
The chances of obtaining an arbitrary specified response are correspondingly
reduced. More sophisticated approaches must therefore be undertaken. A
number of alternatives are possible .... The most attractive appears to be
multiple-layer logical circuit arrangements, since it is known that any function
can thereby be realized. . . . However, no general criteria on the basis of which
intermediate logical layers can be taught functions required for over-all network
realization of the desired input/output relationship have been discovered.39

Classifications realized by neural nets can be represented as

decision regions in pattern space. Multilayer nets with two layers
of hidden units and three layers of modifiable connections (that
is, with one more layer of intermediate units than the system of
Figure 1) can form any decision region in pattern space - that is,
they can realize decision regions (classifications) of arbitrary
complexity (this complexity being limited by the number of units
in the system).40 In other words, a multilayer network with two
layers of hidden units can realize any input/output classification.
Early researchers were aware of the limitations of single-layer
systems (some of these will be illustrated in the following section),
and there is no doubt that they saw multilayer nets as the way to
go. Training multilayer nets was one of the main problems of the
early neural-net field.
Early neural nets also had important technological limitations,
one of the most important of them being the size of the com-
ponents. The Perceptron built by Rosenblatt, Charles Wightman
and their colleagues at CAL, had only 512 modifiable connections,
but it filled a whole laboratory room. Adjustable connections
were implemented using motor-driven potentiometers of consider-
able size, and so implementing a perceptron with thousands of
connections using this technology was not practical. Alternative
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 627

implementations (like the SRI group's magnetic cores, or Wid-

row's 'memistors') were developed, but this technology was rather
limited compared to the emerging von Neumann computer. The
advent of the digital computer affected other computer archi-
tectures too, like the analog architecture.41 In fact, certain ele-
ments of neural nets (in particular the continuously adjustable
weights) and other early cybernetic systems and 'brain models' can
be seen as 'analog'.
Digital computers could be used - and indeed started to be
used - to simulate neural nets, but the overall philosophy of
the neural-net approach, as formulated mainly by Rosenblatt,
favoured a brain-style, anti-von Neumann implementational
position.

Theorists are divided on the question of how closely the brain's methods of
storage, recall, and data processing resemble those practised in engineering
today. On the one hand, there is the view that the brain operates by built-in
algorithmic methods analogous to those employed in digital computers, while
on the other hand, there is the view [Rosenblatt's view] that the brain operates
by non-algorithmic methods, bearing little resemblance to the familiar rules of
logic and mathematics which are built into digital devices.42

The models which conceive of the brain as a strictly digital, Boolean algebra
device, always involve either an impossibly large number of discrete elements,
or else a precision of the 'wiring diagram' and synchronization of the system
which is quite unlike the conditions observed in a biological nervous system.43

But even though simulation of neural nets was possible in

principle, the association between the digital computer and sym-
bolic AI was much stronger, as I will show later. Before that I will
turn to Minsky and Papert's study of the problems of early neural
nets. As I pointed out earlier, according to the official-history
mode of articulation of the debate these researchers showed that
further progress in neural nets was not possible, and after tha
neural nets were largely abandoned.

The 'Proofs' of the Impossibility of Perceptrons

A Social Service for the AI Community

In the 1950s, several research or problem areas evolved from t

cybernetic movement, but none of them had, at that time, ye
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 628 Social Studies of Science

emerged as a research specialty. Competition became stronger in

the late 1950s, as symbolic AI started to emerge as a specialty and
neural nets were still attracting a significant amount of human and
economic resources. The importance of the problems of early
neural nets was not clear. Neural-net researchers maintained that
single-layer nets were only the beginning, and that their lim
tions, important as they were, would be overcome with m
complex systems. In the early 1960s, when controversy h
reached its highest levels, Marvin Minsky and Seymour Papert
two leading symbolic AI researchers from the prestigious MIT
group, decided to intervene in the controversy.

In the middle 1960s Papert and Minsky set out to kill the perceptron, or,
least, to establish its limitations - a task that Minsky felt was a sort of so
service they could perform for the artificial intelligence community.44

According to the official history, Minsky and Papert were worr

by the fact that many researchers were being attracted by neu
nets. Their motivating force was (according to this version) to
to stop what for them was an unjustified diversion of resources
an area of dubious scientific and practical value, and to push t
balance of AI funding and research towards the symbol-process
side.

In the late 1950s and early 1960s, after Rosenblatt's work, there was a gre
wave of neural network research activity. There were maybe thousands
projects. For example Stanford Research Institute had a good project. B
nothing happened. The machines were very limited. So I would say by 1
people were getting worried. They were trying to get money to build big
machines, but they didn't seem to be going anywhere. That's when Papert a
tried to work out the theory of what was possible for the machines without lo
[feedforward perceptrons].45

There was some hostility in the energy behind the research reporte
Perceptrons..... Part of our drive came, as we quite plainly acknowledge
our book, from the fact that funding and research energy were being dissipa
on . . . misleading attempts to use connectionist methods in practical ap
cations.46

The exaggerated statement about the number of neural-net

projects can be understood as part of the official history. Alleging
that there were thousands of projects going along such a 'deviant'
path justified symbolic AI leaders' strong reaction against neural
nets. The social functions of the official history will be analyzed
later. Here I will examine Minsky and Papert's technical argu-
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 629

ments in some detail. Minsky and Papert's work circulated in the

form of drafts and was well known by the mid-1960s, although it
was not published as a book until 1969.47 It is important to note
that Minsky and Papert's work had its effect upon the controversy
well before the book was published.
In the official-history mode, Minsky and Papert's work is
supposed to have shown that further progress in neural nets was
not possible, and that therefore this approach lacked scientific or
practical value. This is why I will use the term 'impossibility
proofs'.48 However, strictly speaking, Minsky and Papert showed
that single-layer nets, defined in a certain way, had some import-
ant limitations. On the other hand, they conjectured that progress
in multilayer nets would not be possible because of the problem of
learning. The key issue (which 1 try to elucidate below) is that their
study was widely seen as a 'knock down' proof of the impossibility
of perceptrons (and of neural nets in general).
Minsky had worked in neural nets, but in the early 1950s he
abandoned this field to embrace the symbolic approach. It is
interesting to note that in the early 1960s he (along with Papert)
went back to the neural-net field in order to 'replicate' (so to
speak) Rosenblatt's Perceptron, and thus show its limitations. As I
have already mentioned, Collins argues that this is rather unusual
in science.49 Normally, one accepts the results coming from an
area one is not directly involved with, and the farther away that
scientific area is from one's own, the bigger one's certainty about
it. Collins pointed out that the crucial and interesting cases are the
replication of controversial and important observations, and the
core-sets of scientists who are involved in the work. The Percep-
tron case satisfies this criterion.
Minsky and Papert's work was highly elaborated from a mathe-
matical point of view, and it stands as a very important contribu-
tion to neural-net theory. They studied a perceptron similar to the
one in Figure 2 (with one output unit instead of three), but they
introduced an important restriction regarding the number of
connections from input units to association units (the layer of fixed
connections in Figure 2). They maintained that the interest of
neural computing came from the fact that it was a parallel
combination of local information, and they suggested that, for this
computation to be effective, it had to be 'simple' in some mean-
ingful sense.50
The computation performed by the output unit of their percep-
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 630 Social Studies of Science

tron (a sum of incoming weighted activation in parallel plus a

comparison with a threshold) satisfied the proposed criterion. In
the case of the association units, Minsky and Papert interpreted
their 'simple combination of local information' restriction as
implying that each of these units could not receive connections
from many input units - that is to say, each association unit could
receive connections only from a small part of the input retina.
They defined the 'order' of a perceptron as the maximum number
of incoming connections received by any association unit (there-
fore, as I have already mentioned, the order of the perceptron of
Figure 2 is 6).
The implications of this 'conjunctive localness' criterion are
better understood by looking at the main examples analyzed by
Minsky and Papert: 'parity' (saying whether the number of
activated inputs in a perceptron retina like the one in Figure 6 is
odd or even), and 'connectedness' (the figure-ground problem,
consisting of saying whether a set of activated retina points belong
to the same object - that is, whether or not they are connected to
each other). The problem of parity is related to the exclusive-or
function mentioned earlier (in a network with two input units and
one output unit, computing parity is equivalent to computing
exclusive-or). Minsky and Papert proved that the order required
for their single-layer perceptron to compute parity was the whole
retina - that is, at least one association unit had to receive
connections from all the input units.51 But if one association unit
had to 'look at' all the input units in the retina, then the
computation realized by the perceptron was not based on a
combination of local information, and therefore the 'conjunctive
localness' criterion could not be satisfied.
The second main problem studied by Minsky and Papert was
'connectedness' issue. The input pattern appearing in the retina
Figure 6 (the blackened units) is connected. Minsky and Pap
proved that the order required for a perceptron to compute t
connectedness property also exceeded practical and accepta
limits. This order grew arbitrarily large as the input retina grew in
size.52
In sum, Minsky and Papert proved that the order required for a
perceptron to compute parity and connectedness was not finite; it
increased with the size of its input retina. This problem could be
seen as equivalent to a conventional computer program having to
be rewritten when changing the size of the task.53
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 631

FIGURE 6

EE
Earlier I showed that early neural-net researchers
aware of problems like connectedness (especially worry
object and letter recognition). Nevertheless, in Min
Papert's study those problems acquired an 'anomalous'
Larry Laudan has defined an 'anomalous problem' as
that both (a) resists solution within a scientific approac
has an acceptable solution within a competing research t
but in controversies, notions like 'resistance to solution' and
'acceptable solution within a competing tradition of research' are
evaluated differently by the contending groups. The anomalous
character of a problem increases if researchers agree, to compar
the solution (or the lack of solution) given by a tradition o
research with the solution given by a competing one. One
important move in Minsky and Papert's rhetoric was to claim tha
problems such as parity or connectedness could easily be solved
using conventional algorithms in serial computers.55

The predicate 'connected' seemed so important in this study that we felt

appropriate to try to relate the perceptron's performance to that of some other
fundamentally different, computation schemes. . . . We were surprised to fin
that, for serial computers, only a very small amount of memory was required.5

Many of the theorems show that perceptrons cannot recognize certain kinds of
patterns. Does this mean that it will be hard to build machines to recognize
those patterns? No. All the patterns we have discussed can be handled by quit
simple algorithms for general-purpose computers.57

By emphasizing that parity and connectedness could easily b

realized by conventional algorithms in von Neumann computers
Minsky and Papert were linking their critical position about neur
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 632 Social Studies of Science

FIGURE 7

nets with two very important factors that would later

closure factors in the controversy: symbolic AI and the
computer.

Interpretative Flexibility

But the importance of problems like parity and connecte

not so clear for neural-net researchers. They compared n
not with conventional computers, but with humans.
Figure 7.58 It is not immediately obvious whether the bla
is connected or not. Now look at the white background as
White is connected, and black is not. But this is not ob
first time one looks at the objects. A conscious, sequenti
is necessary in order to determine the connectednes
figures.
In the research-area mode of articulation, the importance of
these problems - and their alleged anomalous character - was
open to interpretative flexibility. Neural-net researchers claimed
that, if one is trying to explain and model human cognitive
capabilities, then problems like parity and connectedness are not
so worrying (let alone anomalous) after all, because human beings
are not good at recognizing them either.59 The following quote by
David Block, a mathematician from Cornell University who was a
colleague of Rosenblatt in the Perceptron project, is an example of
this.

Another indication of this difference of perspective [between Rosenblatt and

Minsky-Papert] is Minsky and Papert's concern with such predicates as parity
and connectedness. Human beings cannot perceive the parity of large sets (is the

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 633

FIGURE 8

C /T
T/E
Source: H.M. Collins, Artificial Experts: Social Knowledge an
(Cambridge, MA: MIT Press, 1990), 32.

number of dots in a newspaper photograph even or odd?

(on the cover of Minsky and Papert's book there are
connected, one is not). It is virtually impossible to
examination which is which. Rosenblatt would be conten
capabilities, and in fact would tend to regard unfavorably a
beyond them, since it is human perception he is trying

The relative importance of computing connect

circumstances can be shown by a letter reco
Sometimes connectedness is not a dominant feature of humans'
visual environment. The second letter of the two words app
in Figure 8 is the same (something between 'A' and 'H').61
the appealing properties of neural nets was that, due to pre
learnt associations, they would be capable of recognizing w
patterns (in this case the 'A' of 'cat') even though only a p
them (the unconnected second symbol of the first word o
figure) was presented as the input. One of the strong poin
neural nets is that, in certain circumstances, they can conti
see the same pattern even when bits are removed that chan
figure from connected to disconnected (just like humans!).
Neural-net researchers concentrated on the positive prop
of the single-layer perceptron (for example, its learning algorit
its brain-like character, its distributed memory, its resista
damage, its parallelism), and claimed that further resea
more complex models (systems with more than one lay
adjustable connections, with connections among the units
same layer, with backward connections, and so on) was nee
order to overcome its limitations. They were asking for tim
funding to carry out that research. The issue was, of c
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 634 Social Studies of Science

whether their arguments, claims and rhetoric were strong enough

to contest Minsky and Papert's criticism.

The simple perceptron (which consists of a set of inputs, one layer of neurons,
and a single output, with no feedback or cross coupling) is not at all what a
perceptron enthusiast would consider a typical perceptron. He would be more
interested in perceptrons with several layers, feedback and cross coupling...
The simple perceptron was studied first, and for it the 'perceptron convergence
theorem' was proved. This was encouraging, not because the simple perceptron
is itself a reasonable brain model (which it certainly is not; no existing
perceptron can even begin to compete with a mouse!), but because it showed
that adaptive neural nets, in their simplest forms, could, in principle, improve.
This suggested that more complicated networks might exhibit some interesting
behavior. Minsky and Papert view the role of the simple perceptron differently.
Thus, what the perceptronists took to be a temporary handhold, Minsky and
Papert interpret as the final structure.62

The opinions of other neural-net researchers of the time were

similar. For example, Widrow complained that Minsky and Papert
had defined the perceptron so narrowly that they could prove that
neural nets could do nothing, and he emphasized that his group
was working on networks much more complex than the single-
layer one.

When I first saw the book, years and years ago, I came to the conclusion that
they had defined the idea of a perceptron sufficiently narrowly so that they
could prove that it couldn't do anything. I thought that the book was relevant,
in the sense that it was good mathematics. It was good that somebody did that,
but we had already gone so far beyond that. Not beyond the specific
mathematics that they had done. But the structures of the networks, and the
kinds of models that we were working on were so much more complicated and
sophisticated than what they had discussed in the book. All the difficulties, all
the things that they could prove that the perceptron couldn't do were pretty
much of noninterest, because we were working with things so much more
sophisticated than the models that they were studying. The things they could
prove you couldn't do were pretty much irrelevant.63

For those actually involved in neural-network research, Minsky

and Papert's proofs were (in Widrow's words) 'pretty much
irrelevant'. In the research-area mode of articulation, the disputed
cognitive objects (Minsky and Papert's 'proofs' and arguments)
did not have the static (all-or-none, either valid or invalid)
character that is attributed to them in the official-history mode.
And, as in Pinch's case study of von Neumann's proof against
Bohm in quantum physics, after the perceptrons controversy was
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 635

closed most people used Minsky and Papert's proofs against neural
nets without ever going into them.64
As in Pinch's case, the authority of Minsky and Papert's proofs
can be linked to the importance of the axiomatic or 'arithmetic
ideal' in science,65 although in this case this ideal should be applied
not only to those specific disputed objects but also to the more
general differences between the symbolic and neural-net
approaches. Symbolic AI is based on the capabilities of the
computer for manipulating symbolic expressions in ways sensitive
to their logico-syntactical - and therefore discrete - structure.
Although the question of proving what a computer program can
do is by no means trivial,66 symbolic AI was much closer to the
arithmetic (and rationalist) ideal than the subsymbolic, environment-
driven, trained (not programmed) neural-net approach (which was
closer to self-organizing, cybernetic systems).67
So far I have analyzed Minsky and Papert's proofs about single-
layer perceptrons. But what about multilayer nets? The question
of learning in multilayer nets had been on neural-net researchers'
agenda since the late 1950s, and was widely seen by them as a
critical issue. According to the official history of the debate,
Minsky and Papert showed that progress in neural nets as a whole
(not just in single-layer systems) was not possible. But what
Minsky and Papert actually said (in the formal literature) was
much less than that.

The perceptron has shown itself worthy of study despite (and even because of!
its severe limitations. It has many features to attract attention: its linearity; it
intriguing learning theorem; its clear paradigmatic simplicity as a kind
parallel computation. There is no reason to suppose that any of these virtu
carry over to the many-layered version. Nevertheless, we consider it to be
important research problem to elucidate (or reject) our intuitive judgement tha
the extension is sterile. Perhaps some powerful convergence theorem will b
discovered, or some profound reason for the failure to produce an interesti
'learning theorem' for the multilayered machine will be found.68

By what process was this conjecture interpreted as showing th

further progress in multilayer neural nets was not possible?
Whereas neural-net researchers were asking for time and mone
for studying more complex systems and trying to solve th
problems they had, critics favouring the symbolic perspective
claimed that, because of the limits of single-layer systems and t
lack of successful learning rules for multilayer systems, progress in
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 636 Social Studies of Science

neural nets was not possible. We must now analyze the process of
closure of this debate - the process through which interpretative
flexibility was reduced, and controversy closed. In other words,
the question now is to explain the emergence of the official-history
view and its social functions.

Closure of the Controversy

Paul Edwards has recently pointed out two important aspect

the emergence of symbolic AI: on the one hand, sym
researchers' involvement in the early 'mundane practice' (
puts it) of (von Neumann) computer programming and softw
development and, on the other, ARPA's institutional suppor
symbolic AI, mainly through the Information Processing Te
niques Office (IPTO, directed by Joseph C.R. Licklider).69
The development of high-level computer languages and tim
sharing systems was especially important for symbolic AI. As
first computers became commercially available in the 19
programmers started to develop compiler and high-level lang
in order to simplify the program coding tasks that until the
been done in binary machine language (which was extrem
difficult to use and debug). Exploiting the capabilities of the
digital computers required the development of languages wh
would translate English-like commands and instructions
machine language. As the first high-level programming langu
became available in the late 1950s, researchers started to thin
computers as manipulators not just of numbers but als
symbolic expressions. Symbolic AI researchers developed
gramming languages especially suitable for symbol manipula
(such as Newell and Simon's IPL, and McCarthy's LISP).
Symbolic AI programs consumed vast quantities of memory
machine time and, due to the scarcity of the computing reso
then available, there was strong competition for computing t
In the late 1950s, computers were 'batch processors' (wh
program was running, the machine could do nothing else). In
output devices were much slower than the central processing
(CPU), so the CPU was idle most of the time. Edwards descr
this situation as follows:

Programs usually had to be run many times before all errors were foun
fixed. Since the debugging process was slower than CPU or input/output

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 637

by yet further orders of magnitude, after receiving their output and fixing their
programs, programmers would have to wait, frustrated, in a queue until the
machine was again free .... The ... small number of available computers
(especially in universities) meant intense competition for computer time.70

Around 1958, John McCarthy developed the idea of CPU time-

sharing; he wanted to provide symbolic AI researchers with the
possibility of working with LISP interactively from their terminals,
without having to deal with the 'priesthood' of computer opera-
tors. Working with the computer interactively created the possibil-
ity of on-line debugging and fixing of programs while these were
running. Thus the effects of each change became instantly visible
to the terminal user (the AI researcher).
As Edwards points out, this connection between AI and time-
sharing led to the second of the mentioned issues: ARPA's strong
support of symbolic AI, mainly through Licklider's IPTO.
ARPA's backing of interactive computing, time-sharing systems
connected symbolic AI with military projects for human-machine
interaction in electronically mediated systems of 'command and
control' and 'decision support'. Along with time-sharing, symbolic
researchers received strong funding for their scientific objectives
of high-level programming, cognitive simulation, heuristics, and
the like.

Supported by ARPA funding, the initial leading core of sym

bolic AI - a reduced group of researchers and their studen
working at a few prestigious centres such as MIT (Minsky
group), Carnegie-Mellon University, Stanford Universi
(McCarthy's group), and SRI - had a privileged access to
economic and (the then so scarce) computing resources, a
consolidated their professional and organizational network
ARPA's policy favoured resource concentration at a few centr
of excellence, and selection of projects was based neither on p
review, nor on equalizing principles for research money distrib
tion, but on the agency's own judgement about the best
researchers working on the best projects from the point of view of
the agency's military goals.72
At the same time that it was backing symbolic AI explicitly,
ARPA decided - also in an explicit manner - not to fund neural-
net research. Both neural-net and symbolic AI researchers were
well aware of this, and there is no doubt that this had an impact
upon the perceptrons debate. Controversy went beyond the limits
of the scientific community, and reached the US Government
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 638 Social Studies of Science

agencies that were funding AI - mainly ONR and, above all,

ARPA. Marvin Denicoff, who worked at ONR in the early 1960s
and was also well informed about ARPA's involvement in AI
(both agencies collaborated in some respects) told me ab

At that time [in the 1960s], the Office of Naval Research had fund
of $40K or $50K. ARPA was able to fund hundreds of thousan
millions. Rosenblatt never attracted that kind of money, becau
offering a large pay-off. By pay-off I mean not in the scientific sens
application sense, world problem solving. Again, his work was mu
would say, traditional science. The Office of Naval Research never g
kind of money that he really required, and he was not successful in
money from the Science Foundation or from ARPA. One can
conclusion that if he had had the money he would have made e
progress. That's too easy an answer, because it doesn't always follo
amounts of money make the difference. Well before the Minsky
book came, Rosenblatt was not successful in attracting more m
know for a fact.73

Jon Guice has studied the role of ARPA and the MIT-area
defence and research community in the process of clo
perceptrons controversy. He has documented in det
decision to concentrate its IPTO funding resources on
bolic AI centres from the early 1960s (Minsky's MIT g
mid-1960s (Stanford, CMU and other smaller institutio
same time as it explicitly rejected applications to fund
research.74 This decision by ARPA was a very importan
the legitimation of symbolic AI and in the closure of th
trons controversy. Guice has also pointed out the impor
unconventional, satirical paper entitled Artificial Intel
written by consultant Louis Fein in 1963.75
Fein asks the reader to imagine that a Federal agency
request to bid on research and development work in A
companies. The author then includes the request to bid
companies' replies, and an evaluation of the propo
external technical expert who advises the agency. Ps
are used for the agency (Bright Field), bidding
(Optimystica; Dandylines Enterprises; Search Limited,
Search Unlimited; and Calculated Risks, Inc.) and ev
(J.R. 'Bubbles' Piercer, from Pessimyths, Inc., a consult
The bidding companies represent different AI perspec
research groups: self-organization (Optimystica), n
(Dandylines, which could refer to the SRI neural-n
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 639

perhaps associated with other groups; it is interesting to note that

this company sees its work as a continuation of that of Rosenblatt's
group), symbolic AI (Search Limited, formerly Search Unlimited,
which could well refer to Minsky's MIT group) and probabilistic
and statistical pattern recognition, which can be seen as related
to neural nets (Calculated Risks, Inc.). 'Bright Field' could be
ARPA, and J.R. 'Bubbles' Piercer could be Licklider (who was
actually ARPA's IPTO Director from 1962 to 1964).
'Bubbles' Piercer's report contains some interesting points.
First, he criticizes Bright Field for failing to ask certain companies
to bid (apparently referring to certain neural-net and cybernetics
companies). Second, he criticizes the agency's overambitious AI
goals, and maintains that AI is in a research phase, far from
development and production (he is in favour of AI as an aid to
human intelligence, rather than a replacement of it). He points out
that AI (including symbolic AI) has made many promises but so
far it has failed to deliver. Finally, although he vaguely recom-
mends some support for Calculated Risks and Dandylines (in
particular for studying learning and storage capacity in multilayer
nets), he ends by making a strong recommendation to support
Search Limited. As Guice points out, of these three perspectives
of research only the third one (symbolic AI, starting with Minsky's
group) was actually funded by ARPA's IPTO.
The (unusual) satirical character of this paper makes it difficult
to evaluate, but it can be taken as a (humorous) account of the
competition for ARPA funding in the early 1960s between
symbolic and other AI approaches (neural nets, and related
approaches like probabilistic pattern recognition and cybernetics).
ARPA's decision to back symbol-processing and to reject neural
nets was a very important closure factor in the perceptrons
controversy. It is important to note that, for ARPA, symbolic and
heuristic systems were the way to go not only for 'data interpreta-
tion and decision-making in command and control' in general, but
also for the central areas of interest of neural-network researchers
(that is, visual pattern recognition, as applied for example to the
interpretation of satellite photographs).
The process of emergence and institutionalization of symbolic
AI as a scientific specialty was almost completed by the mid-1960s.
By then this approach had accumulated an important stock of
scientific contributions.76 At that time the perceptrons controversy
was approaching closure. From the three main early neural-net
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 640 Social Studies of Science

projects, only Rosenblatt continued his work in perceptrons.

Widrow's Stanford University group went into telecommunica-
tions engineering applications (where they employed successfully
some of their neural-net techniques), and the SRI group started an
important mobile robot project within symbolic AI. Later,
Rosenblatt's early death in 1971 in a sailing boat accident would
leave the neural-net field without its most charismatic leader and
advocate.
According to the official history of the controversy, after M
and Papert's study, the neural-net approach was rejecte
abandoned. Papert himself recognized the existence of 'uni
istic' (all-or-none) attitudes.
Its universalism made it almost inevitable for AI to appropriate our wo
proof that neural nets were universally bad. ... In fact, more than half
book is devoted to 'properceptron' findings about some very surpris
hitherto unknown things that perceptrons can do. But in a [scientific] cul
up for global judgement of mechanisms, being understood can be a fate
as death.77

Papert recognized the existence of a 'global judgement' (against

neural nets) in the closure of the perceptrons controversy, and
complained that his book with Minsky was interpreted in that
sense.

According to the official history, Minsky an

Rosenblatt's overclaiming and showed that pro
was not possible - and after that this field wa
But if, as I have shown here, Minsky and P
show that, and if (as I will point out soon) ne
completely abandoned, what was the role of th
is my view that its role can only have been th
emergence and institutionalization of the
which came to be seen as the 'right' appr
occupying the whole AI discipline. In the 1
leading researchers used the 'we are the o
argument in their rhetoric, as can be seen in
seminal paper by Newell and Simon:
The principal body of evidence for the symbolic hypo
considered [so far in this paper] is negative evidence:
competing hypotheses as to how intelligent activity m
whether by man or by machine.78

Therefore the closure of the perceptrons cont

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 641

1960s could well be the 'marker event' that Newell was looking for
in his account of the emergence of symbolic AI.

Through the early 1960s, all the researchers concerned with mechanistic
approaches to mental functions knew about each other's work and attended the
same conferences. It was one big, somewhat chaotic, scientific happening. The
four issues I have identified - continuous versus symbolic systems, problem
solving versus recognition, psychology versus neurophysiology, and perform-
ance versus learning - provided a large space within which the total field sorted
itself out. Workers of a wide combination of persuasions on these issues could
be identified. Until the mid-1950s, the central focus had been dominated by
cybernetics, which had a position on two of the issues - using continuous
systems and orientation towards neurophysiology - but no strong position on
the other two. The emergence of programs as a medium of exploration
activated all four of these issues, which then gradually led to the emergence of a
single composite issue defined by a combination of all four dimensions
[symbolic, problem solving, psychology, performance]. This process was essen-
tially complete by 1965, although I do not have any marker event. [Later Newell
points to one more 'issue'.] Most pattern recognition and self-organizing
systems were highly-parallel network structures. Many were modelled after
neurophysiological structures. Most symbolic-performance systems were serial
programs. Thus, the contrast between serial and parallel (especially highly-
parallel) systems was explicit during the first decade of AI. The contrast was
coordinated with the other four issues I have just discussed.79

The official history of the debate legitimated the authority struc-

ture which was emerging in AI, and was used by the elite of the
symbolic approach as a defence strategy against heterodox and
'deviant' interpretations and approaches.
The official history conveniently exaggerates the phenomenon
of the abandonment of neural nets. Although neural nets were
largely rejected as an approach to AI, throughout the 1970s, all
over the world, some (not many) researchers - most of them
belonging to a younger generation - continued working on neural
nets and related topics outside the AI field, in neuroscience and
psychology-oriented areas. As the Lighthill report for the UK
Science Research Council on the state of AI in the early 1970s
shows, neural-network-like research remained somewhat stronger
in Europe than in the United States.80 Researchers who worked in
neural nets (in topics such as unsupervised learning and associative
memory) in the 1970s include Christoph von der Malsburg, David
Willshaw, Teuvo Kohonen, Geoffrey Hinton and Igor Aleksander
in Europe; Michael Arbib, Stephen Grossberg, James Anderson,
Jack Cowan and Leon Cooper in the United States; and Kunihiko
Fukushima and Shun-ichi Amari in Japan.81
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 642 Social Studies of Science

Therefore the (inaccurate) view of the 'abandonment of neural

nets' can be seen as legitimating the emergence of symbolic AI,
rather than as an exact description of the result of the perceptrons
controversy. After the closure of the controversy, neural-net
activity decreased significantly and was displaced to areas outside
AI (it was considered 'deviant' within AI) but, contrary to the
official view, it did not completely disappear.

The Revival of Neural Nets

Studies of scientific controversies have shown that, once an

interpretation has emerged as dominant after the closure of a
controversy, time runs against the 'losers' as the organizational
and cognitive structures supporting the winning side develop and
institutionalize.82 As the institutionalization of a new social order
(with its resource allocation system and authority structure)
advances, it increasingly comes to be seen as the only possibility
(as a 'natural' order). This is why, in periods of stability, soci-
ologists employ methodological directives such as Everett
Hughes's 'remember that it could have been otherwise', in order
to remind themselves of the constructed character of social reality.
Collins has employed this idea in the sociology of science.83
In this paper, I have tried to show the interpretative flexibility of
Minsky and Papert's proofs and arguments about the impossibility
of perceptrons. The rejection of neural nets as an approach to AI
was a contingent social process, and therefore, in principle, 'things
could have been otherwise'. The interesting and curious thing
about neural nets is that things were actually otherwise in the
middle and late 1980s, two decades after the closure of the
perceptrons controversy. Here I can only review briefly some of
the main developments which brought about this change.84
In the early 1980s, symbolic AI went from institutionalization to
a stage of growth, applications and (the beginning of) commercial-
ization.85 International competition and interest in this specialty
increased as the US and UK governments reacted to the
announcement by the Japanese Government of its Fifth Genera-
tion Project (especially directed to areas like natural language
processing and 'knowledge engineering', or knowledge-based
systems). The rise of the expert systems application area in the
mid-1980s was one of the main developments of this period.86
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 643

Developments in information technology were a major change

affecting AI. In the early 1980s, dramatic decreases in computing
costs brought about a 'democratization' in the access to computing
resources. As a result, as James Fleck points out, the scope for the
strong symbolic AI elite to control the development of the field
was weakened, allowing outsiders to move in and pursue their own
variants of AI research.87 Eventually this permitted the use of
powerful computing resources to simulate until-then 'deviant'
approaches, such as neural nets. On the other hand, since the late
1970s, researchers from a variety of fields in the human sciences
(and later in the neurosciences) had started to use the computer as
a research tool in an emerging interdisciplinary discipline called
'cognitive science'. Although cognitive science was then based in
the symbolic approach, its interdisciplinary character helped bring
new perspectives into the computer-mind problem.
In this context, some symbolic AI researchers started to confront
the limitations of their models. Expert systems were being applied to
a great variety of problems, but symbolic AI was not so successful in
areas such as speech recognition, pattern recognition, and common-
sense and heterogeneous reasoning. Some researchers started to
look at new approaches for studying and modelling these tasks.
The conference organized in June 1979 at La Jolla (California)
by neural-net 'veterans' Geoffrey Hinton and James Anderson can
be seen as the first contact between researchers who had been
working in neural nets throughout the 1970s and research
coming from the symbolic approach, but looking for ways
solving some of its limitations. The papers presented there we
developed and published in 1981 in a book entitled Parallel Mod
of Associative Memory.88 The topics of the book are a good sam
of the perspectives which were being considered: information p
cessing in the brain, connectionist local nets, semantic nets, a
associative memory. Other topics which these researchers wer
looking at include parallelism in vision research (for examp
interaction between many local features in the interpretation of an
image) and multiple constraint systems.89 After this, the Paral
Distributed Processing (PDP) group was formed in the Universit
of California-San Diego, headed by psychologists David Rumelha
and James McClelland.
Although neural nets were not directly linked to the neu
sciences, increases of activity and interest in the latter in the 19
contributed to a more favourable context for the former. PD
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 644 Social Studies of Science

other researchers adopted a 'brain-style' style of information

processing. They argued that the information processing power of
the brain comes from its parallelism. Given the facts that neurons
are not too fast (firing frequencies range from a few to a few
hundred impulses per second) and that some complex mental
behaviour (like recognizing a face) takes 1/10 second, researchers
concluded that the brain's information processing power must
come from its parallelism.90
The advent of parallel computers and supercomputers in the
1980s as an attempt to overcome the speed limitations of sequen-
tial computers (separation between memory and central process-
ing unit in a von Neumann computer imposes a sequential, 'one
operation at a time' style of computation) added plausibility to
'brain-style' computation. As with the neurosciences, the con-
nection between parallel computers and neural nets was not
straightforward in the beginning; many of the most successful
neural-net experiments of the mid-1980s were done as simulations
in sequential computers. On the other hand, there are many
parallel computer architectures, and neural nets are one extreme
type (massively parallel).91 Nevertheless, increases in computer
power and speed due to parallelism will undoubtedly favour
neural-net research.92
The work done by the PDP group (with people like Rumelhart,
McClelland, Hinton and Terrence Sejnowski) and by 'veterans'
such as Anderson, Grossberg, Kohonen, Willshaw and von der
Malsburg started to a attract researchers from other disciplines to
the neural-net field.93 Migration is a common phenomenon when a
new area of research is emerging.94 Researchers coming from
overpopulated areas or specialties, or having widely applicable
backgrounds such as physics or mathematics, may perceive interest-
ing or non-exploited problems and career opportunities in different,
emerging areas.
The case of John Hopfield, a physicist from the California
Institute of Technology, was particularly important.95 Hopfield
used a method of the physics of collective phenomena (the Ising
model of magnetic material, or 'spin-glass') in order to develop a
new neural-net architecture with symmetric connections that could
be used as an associative content-addressable memory.96

In physical systems made from a large number of simple elements, interactions

among large numbers of elementary components yield collective phenomena
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 645

such as the stable magnetic orientations and domains in a magnetic system. Any
physical system whose dynamics in phase space is dominated by a substantial
number of locally stable states to which it is attracted can therefore be regarded
as a content-addressable memory. The physical system will be a potentially
useful memory if, in addition, any prescribed set of states can readily be made
the stable states of the system.97

Hopfield's model was later developed by Hinton and Sejnowski,

two of the most important researchers of the PDP group, into the
'Boltzmann machine' stochastic multilayer net.98 Hinton and
Sejnowski developed a learning algorithm which usually got the
best global minima and, although in the beginning it was quite
slow, they presented it as a first solution to the problem of learning
in multilayer nets.

In the Boltzmann machine, Hinton and I found a learning algorithm which

overcame the conjecture by Minsky and Papert that you couldn't generalize the
perceptron learning algorithm to a multilayered architecture.99

A learning algorithm was discovered for the Boltzmann machine that provided
the first counterexample to the conjecture by Minsky and Papert that extensions
of the perceptron learning rule to multilayered networks were not possible.'0

Both the Hopfield and the Hinton-Sejnowski cases show that

cross-fertilization and communication between neural nets and
other scientific fields (physics of collective phenomena; stoch
techniques from statistical mechanics) was very important in
neural-net revival.101 Different techniques were applied to t
study of representation and learning in nonlinear dynamical ne
net systems.
After the Boltzmann net, PDP researchers Rumelhart, Hin
and Ronald Williams developed a learning algorithm for mul
layer feedforward (that is, perceptron-like) nets, the so-called bac
propagation algorithm.102 This contribution - the most popula
the neural-net revival - triggered a new wave of neural-
research. Figure 1 earlier represents the type of architecture
which Rumelhart and his colleagues developed their techniqu
The main problem for weight adjustment in multilayer nets i
know the error made by the hidden units, in order to be ab
adjust the connections between input units and hidden units (
error made by the output units is the difference between the
output pattern and the desired one). The intuitive idea of bac
propagation is that the error made by a hidden unit should dep
on the errors made by the output units to which it is connect
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 646 Social Studies of Science

These errors are back-propagated, so that the weights between

input units and hidden units can then be adjusted. In a back-
propagation net, each output unit demands from the hidden units
exactly what it needs, and the hidden units try to accommodate the
conflicting demands.103
A very important difference between the back-propagation net
and the perceptron was the introduction of smooth or sigmoid
activation functions (in the processing units) instead of the classic
discontinuous step functions, so that it became possible to com-
pute error gradients in multilayer feedforward nets (the deriva-
tives of the error with respect to the hidden units' output could be
calculated). This small change in the assumptions defining a neural
net made possible the study of complex systems with flexible
activation surfaces.

Small reformulations of a problem can greatly change the possibilities of ma

progress. The change from threshold logic units to sigmoids might not seem
a major reformulation, but by using continuous rather than discontin
functions, it became possible to generalize the Widrow-Hoff and percept
learning algorithms to multilayered networks.04

It is interesting to note that Paul Werbos developed a techniq

equivalent to back-propagation in the 1970s, but found resist
to his idea of applying it to neural nets.'05
In 1986, PDP researchers Rumelhart and McClelland se
report to DARPA and the National Science Foundation (N
asking for funding for neural nets and warning against furt
neglect of this approach.106 DARPA's Neural Network Study,
its subsequent decision to start support for this approach, w
especially significant because of the strong role played by th
agency in the development (and legitimation) of symbolic AI
the end of the 1980s, most US European and Japanese fun
agencies had launched programmes in neural nets.
In the process of legitimation of the new neural-net movem
of the late 1980s, the PDP researchers confronted the view wh
had helped legitimate the symbolic approach (and delegiti
neural nets) in the 1960s - namely, the official history of
controversy. Rumelhart and his colleagues claimed that,
though their back-propagation net sometimes got trapped in l
(or false) minima, in practice the system led to acceptab
solutions in 'virtually every case'. They claimed that they
overcome Minsky and Papert's impossibility proofs and argum
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 647

The problem, as noted by Minsky and Papert, is that whereas there is a very
simple guaranteed learning rule for all the problems that can be solved without
hidden units, namely the perceptron convergence procedure (or the variation
originally due to Widrow and Hoff, which we call the delta rule), there is no
equally powerful rule for learning in networks with hidden units. The standard
delta rule [Widrow's LMS or delta rule algorithm] essentially implements
gradient descent in sum-squared error for linear activation functions. In this
case, without hidden units, the error surface is shaped like a bowl with only one
minimum, so gradient descent is guaranteed to find the best set of weights. With
hidden units, however, it is not so obvious how to compute the derivatives, and
the error surface is not concave upwards, so there is the danger of getting stuck
in local minima. The main theoretical contribution of this [paper] is to show that
there is an efficient way of computing the derivatives. The main empirical
contribution is to show that the apparently fatal problem of local minima is
irrelevant in a wide variety of learning tasks. Although our learning results do
not guarantee that we can find a solution for all solvable problems, our analysis
and results have shown that as a practical matter, the error propagation scheme
leads to solutions in virtually every case. In short, we believe that we have
answered Minsky and Papert's challenge and have found a learning result
sufficiently powerful to demonstrate that their pessimism about learning in
multilayer machines was misplaced.107

In a sense, PDP researchers made use of the official history for

their own benefit. They were saying something like 'after all,
Minsky and Papert did not really show that neural nets were
impossible'. They were exploiting the interpretative flexibility of
the debate to their own benefit, but it is important to note that
they were able to do this within the process of accumulation and
cross-fertilization of the middle and late 1980s. Other people who
tried to do the same before then (like Werbos with back-
propagation, or some neural-net 'veterans' with other systems)
failed.'08 The official history of the debate was rewritten in order
to legitimate the 'new order' (resource allocation system and
authority structure) resulting from the revival of neural nets, and
its emergence as an AI research specialty.
Rumelhart and his colleagues' claims reopened the controversy,
and Minsky and Papert reacted quickly.

We have the impression that many people in the connectionist community

do not understand that this [back-propagation] is merely a particular way to
compute a gradient and have assumed instead that back-propagation is a new
learning scheme that somehow gets around the basic limitations of hill-climbing.
. . Virtually nothing has been proved about the range of problems upon which
GD [the generalized delta rule, or back-propagation] works both efficiently and
dependably. ... In the early years of cybernetics, everybody understood that
hill-climbing was always available for working easy problems, but that it almost

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 648 Social Studies of Science

always became impractical for problems of larger sizes and complexities...

The situation seems not to have changed much - we have seen no contempor-
ary connectionist publication that casts much new theoretical light on the
situation .... We fear that its [back-propagation's] reputation also stems from
unfamiliarity with the manner in which hill-climbing methods deteriorate when
confronted with larger-scale problems. In any case, little good can come from
statements like 'as a practical matter, GD leads to solutions in virtually every
case' or 'GD can, in principle, learn arbitrary functions'. Such pronouncements
are not merely technically wrong; more significantly, the pretense that problems
do not exist can deflect us from valuable insights that could come from
examining things more carefully. As the field of connectionism becomes more
mature, the quest for a general solution to all learning problems will evolve into
an understanding of which types of learning processes are likely to work on
which classes of problems.09

As the neural-net revival advanced in the late 1980s, the

controversy about the validity and feasibility of neural n
old perceptrons controversy) reopened, and there were ne
sodes of interpretative flexibility.110 But this time the em
of neural nets as an AI specialty was unstoppable.111 Tech
like back-propagation were developed and applied to a
variety of practical problems in areas such as object and s
recognition.112
Debate about the relationships between the symboli
neural-net approaches continued, but the most negativ
about the neural-net field were quickly overcome.113 A
approaches were compared and developed, the strong an
points of each of them was being tested in each particular p
After a first period of quite strong competition betwe
two approaches, the situation will probably evolve into
normalized combination of competition and - increas
cooperation.

Concluding Summary

In this paper, I have analyzed the controversy which surrounded

Rosenblatt's Perceptron Project (and neural nets in general) in the
late 1950s and early 1960s. Attention has been focused on a
particular cognitive object: Minsky and Papert's proofs and argu-
ments, which were interpreted as showing that further progress in
neural nets was not possible and that therefore this approach had
to be abandoned. I have distinguished two modes of articulation of
this disputed cognitive object: the research-area mode and the
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 649

official-history mode.1l4 I have shown that the official-history

mode of articulation played a crucial role in the controversy.
At the research-area level, there was considerable interpretative
flexibility about Minsky and Papert's proofs and arguments.
Scientists using different research techniques and having different
approaches and interests interpreted those results differently.
However, as the symbolic AI approach emerged and institutional-
ized, an official interpretation emerged according to which Minsky
and Papert had shown that progress in perceptrons - and in
neural nets in general - was not possible. According to this
official-history view, neural nets were abandoned in the late 1960s.
The official-history mode of articulation of the debate can be
seen as part of the discourse of legitimation of the new AI 'order'
(with its resource allocation system and authority structure) which
emerged from the institutionalization of the symbolic approach as
a research specialty. The symbolic approach was presented as
occupying the whole AI field, and the official history of the
perceptron debate was used as a defence strategy against 'deviant'
claims and approaches (such as neural nets). Some researchers
continued working in neural-net-related topics throughout the 1970s,
but they were displaced from the AI field.
The interpretative flexibility of the debate is further shown by
the revival of neural nets (in different circumstances) in the mid-
1980s. In the recent process of emergence and legitimation of
neural nets as an AI research specialty, the official history was
revised (PDP researchers claimed that 'after all, Minsky and
Papert did not really show that progress in neural nets was
impossible') as the AI field was being socially and cognitively
redefined, and a new resource allocation system and authority
structure was developing.

* NOTES

The work on which this paper is based was supported by a Basque Gov
scholarship for doctoral research at the Department of Sociology of the U
of Edinburgh (1988-91). I would like to thank the following people fo
encouraging and helping me throughout my research: Donald MacK
supervisor), James Fleck, Alfonso Molina and David Willshaw (Uni
Edinburgh); Peter Dayan (University of Toronto); and Jesus Maria Lar

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 650 Social Studies of Science

Jesus Ezquerro (University of the Basque C

Kallich and Sue Jennings, who helped me in
Finally, 1 would like to thank both the peopl
information by letter.

1. H.M. Collins (ed.), 'Knowledge and Con

Science', Social Studies of Science, Vol. 11,
'The Place of the "Core Set" in Modern
Methodological Propriety in Science', His
Collins, 'An Empirical Relativist Progra
Knowledge', in Karin D. Knorr-Cetina
Observed: Perspectives on the Social Stud
113; Collins, Changing Order: Replication
(London: Sage, 1985); Bruno Latour, Scienc
and Engineers throughout Society (Milton K
1987); S. Leigh Star, Regions of the Mind
Scientific Certainty (Stanford, CA: Stanfor
2. Collins (1983), op. cit. note 1.
3. This is Donald MacKenzie's formulation: see D. MacKenzie, Inventing
Accuracy: A Historical Sociology of Nuclear Missile Guidance (Cambridge, MA:
MIT Press, 1990), 10.
4. My idea of the processes of accumulation of resources and cross-fertilization is
compatible with some parts of Bruno Latour's Science in Action, op. cit. note 1. I
take this book as a continuation of Collins' and others' controversy studies. Other
parts of actor-network theory - which are supposed to be 'philosophically more
radical' - are not useful for my case study.
5. As sociologists have argued, since the pioneer work of Barry Barnes and
David Bloor, this is not a characteristic of 'bad' or 'ideological' science, but of all
science: see, for example, B. Barnes, Scientific Knowledge and Sociological Theory
(London: Routledge & Kegan Paul, 1974); D. Bloor, Knowledge and Social
Imagery (London: Routledge & Kegan Paul, 1976).
6. Collins, 'Core Set', op. cit. note 1. I will come to this issue later.
7. T.J. Pinch, 'What Does a Proof Do if it Does Not Prove?: A Study of the
Social Conditions and Metaphysical Divisions Leading to David Bohm and John
von Neumann Failing to Communicate in Quantum Physics', in Everett Mendel-
sohn, Peter Weingart and Richard Whitley (eds), The Social Production of
Scientific Knowledge, Sociology of the Sciences Yearbook, Vol. 1 (Dordrecht: D.
Reidel, 1977), 171-215, at 174. In this paper, I use Pinch's case study as a parallel,
and I apply many of the categories he uses to my own case study.
8. Pinch, op. cit. note 7, 174-76.
9. For a detailed account of the evolution of neural-network research, see Mikel
Olazaran, A Historical Sociology of Neural Network Research (unpublished PhD
dissertation, Department of Sociology, University of Edinburgh, 1991); Olazaran,
'A Sociological History of the Neural Network Controversy', Advances in
Computers, Vol. 37 (1993), 335425. In the present paper, I examine just one key
issue of the evolution of neural nets: the emergence and functions of the official
history of the perceptrons controversy.
10. Neural networks are also called 'artificial neural networks', 'connectionist
networks', 'parallel distributed systems' and 'neural computing systems'.

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 651

11. Neural nets also have a very important engineering aspect, directed toward
special purpose hardware implementation. Nevertheless, the main developments of
the evolution of neural nets have occurred around AI. AI can be seen as the core of
a wider, interdisciplinary field called 'cognitive science'. Although cognitive science
did not emerge as a differentiated discipline until the late 1970s, many of its main
problems were studied under different headings earlier.
12. For the history of symbolic AI, see: James Fleck, The Structure and
Development of Artificial Intelligence: A Case Study in the Sociology of Science
(unpublished MSc dissertation, University of Manchester, 1978); Pamela McCor-
duck, Machines Who Think: A Personal Inquiry into the History and Prospects of
Artificial Intelligence (New York: W.H. Freeman, 1979); Fleck, 'Development and
Establishment in Artificial Intelligence', in Norbert Elias, Herminio Martins and
Richard Whitley (eds), Scientific Establishments and Hierarchies, Sociology of the
Sciences Yearbook, Vol. 6 (Dordrecht: D. Reidel, 1982), 169-217; Fleck, 'Post-
script: The Commercialisation of Artificial Intelligence', in Brian P. Bloomfield
(ed.), The Question of AI (London: Croom-Helm, 1987), 149-64; Paul N.
Edwards, The Closed World: Computers and the Politics of Discourse in Cold War
America (Cambridge, MA: MIT Press, Inside Technology series, 1995, forthcom-
ing).
13. M.L. Minsky and S.A. Papert, Perceptrons: An Introduction to Computa-
tional Geometry (Cambridge, MA: MIT Press, 1969), 19. Minsky and Papert refer
to F. Rosenblatt, 'The Perceptron: a Probabilistic Model for Information Storage
and Organization in the Brain', Psychological Review, Vol. 65 (1958), 386-408.
14. National Physical Laboratory (NPL), Mechanisation of Thought Processes,
Vols I and II (London: Her Majesty's Stationery Office, 1959); Marshall C. Yovits
and S. Cameron (eds), Self-organizing Systems: Proceedings of an Interdisciplinary
Conference, Chicago, IL, 5-6 May 1959 (New York: Pergamon Press, 1960); H.
von Foerster and G.W. Zopf (eds), Illinois Symposium on Principles of Self-
organization, University of Illinois, Urbana, IL, 1960 (New York: Pergamon Press,
1962); Yovits, G.T. Jacobi and G.D. Goldstein (eds), Self-organizing Systems 1962
(Washington, DC: Spartan, 1962). In the 'Mechanisation of Thought Processes'
conference there were contributions from approaches including symbolic AI (M.
Minsky and J. McCarthy), 'cybernetics' (Donald M. MacKay, W. Ross Ashby),
pattern recognition (Oliver G. Selfridge, A.M. Uttley, Warren S. McCulloch and
Wilf K. Taylor) and neural networks (F. Rosenblatt). In the 1962 conference on
self-organization, there were contributions from perspectives including neural
modelling (Leon D. Harmon), brain theory/neural networks (W.S. McCulloch,
Michael A. Arbib, Jack D. Cowan), neural networks (F. Rosenblatt, B. Widrow),
neural networks/electrophysiological experiments (B.G. Farley), symbolic AI (A.
Newell) and 'cybernetics' (D.M. MacKay). Within the cybernetics movement,
work was done which was related to neural networks in diverse ways and degrees.
Oliver Selfridge's (NPL, op. cit., 511-26) hybrid Pandemonium system is an
example; another one is work on pattern recognition in Britain by A.M. Uttley
(ibid., 119-47) For early neural-network papers, see also J.A. Anderson and
Eduard Rosenfeld (eds), Neurocomputing: Foundations of Research (Cambridge,
MA: MIT Press, 1988).
15. F. Rosenblatt, On the Convergence of Reinforcement Procedures in Simple
Perceptrons (Buffalo, NY: Cornell Aeronautical Laboratory Report VG-1196-G-4,
1960); Rosenblatt, Principles of Neurodynamics (New York: Spartan, 1962); B.
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 652 Social Studies of Science

Widrow and M.E. Hoff, 'Adaptive Switching Circuits', 1960 IRE WESCON
Convention Record (New York: IRE, 1960), 96-104.
16. Rosenblatt, Principles of Neurodynamics, op. cit. note 15, 111.
17. 'New Navy Device Learns by Doing', The New York Times (8 July 1958),
25:2. For similar statements, see: 'Electronic "Brain" Teaches Itself, The New
York Times (13 July 1958), at iv 9:6; 'Rival', The New Yorker (6 December 1958),
44-45.
18. McCorduck, op. cit. note 12, 87. For a list of people irritated by Rosen
see ibid., 88.
19. Yovits, interview, 28 November 1989.
20. Robert Hecht-Nielsen, Neurocomputing (Reading, MA: Addison-We
1990), 16-17. Hecht-Nielsen refers to Minsky & Papert, op. cit. note 13.
21. Rosen, interview, 10 November 1989.
22. See Collins (1983), op.cit. note 1, 49; Latour, op. cit. note 1; Star, op
note 1.

23. This is due in part to the fact that AI affects social discourses about
similarities and differences between human beings and machines: see J. F
'Artificial Intelligence and Industrial Robots: An Automatic End for Uto
Thought', in E. Mendelsohn and Helga Nowotny (eds), Nineteen Eighty-F
Science between Utopia and Dystopia, Sociology of the Sciences Yearbook, Vo
(Dordrecht: D. Reidel, 1984), 189-231.
24. Cited in Hubert L. Dreyfus, What Computers Can't Do: The Limi
Artificial Intelligence (New York: Harper Colophon, 2nd edn, 1979), 81-82
also Edwards, op. cit. note 12, 315 (draft publication).
25. Dreyfus, op. cit. note 24.
26. Some parts of Dreyfus's work are not far from the kind of contributions
the sociology of knowledge could make to AI and cognitive science. It is interest
to note that, although Dreyfus's work provoked a strong critical reaction from
AI leaders in the 1960s (see McCorduck, op. cit. note 12, Chapter 9), rece
someone as qualified as Minsky has implicitly recognized that such philosop
research can make positive contributions to AI. See the following quote f
Minsky's opening talk at the 1988 IEEE International Conference on N
Networks, in the heat of the neural-net revival: 'Minsky, who has been criticize
many for the conclusions he and Papert make in Perceptrons, opened his def
with the line "Everybody seems to think I'm the devil". Then he mad
statement, "I was wrong about Dreyfus too, but I haven't admitted it yet", w
brought another round of applause', from Randolph K. Zeitvogel, 'IC
Reviewed', Synapse Connection (now Neural Technology Update), Vol. 2-8 (19
10-11. For a recent discussion about AI from the perspective of the sociolog
knowledge, see H.M. Collins, Artificial Experts: Social Knowledge and Intelli
Machines (Cambridge, MA: MIT Press, 1990).
27. The use of the 'core set' concept in controversy studies is due to Co
'Core Set' & Changing Order, op. cit. note 1.
28. McCorduck, op. cit. note 12, 88.
29. Rosenblatt, op. cit. note 13, reprinted in Anderson & Rosenfeld (eds
cit. note 14, 92-114, at 96; Rosenblatt, Principles of Neurodynamics, op. cit
15, 67-70; F. Rosenblatt, 'Strategic Approaches to the Study of Brain Model
von Foerster & Zopf (eds), op. cit. note 13, 385-96, at 390-91.
30. Rosenblatt, Principles of Neurodynamics, op. cit. note 15, 306.

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 653

31. Ibid., 309-10.

32. In a perceptron with two layers of association units, the units of the first
association layer which responded to similar features in different positions would
all activate the same unit in the second association layer, and in this way a feature in
different positions could be recognized as the same.
33. Rosenblatt, Principles of Neurodynamics, op.cit. note 15, 576.
34. Ibid., 577-79. Rosenblatt uses the term 'terminal' to refer to the connections
between the second association layer and the response units, and 'preterminal' to
refer to the previous layers of connections.
35. Rosenblatt, Principles of Neurodynamics, op. cit. note 15, 580-81.
36. Ibid., 579-80.
37. For two propositions, the truth value of the exclusive disjunction function is
as follows:
p q ex-or
1 1 0
1 0 1
0 1 1
0 0 0

38. There are other impor

number and type of proce
number and type of layers
required, but the lack of
problem.
39. J.K. Hawkins, 'Self-Organizing Systems: A Review and Commentary',
Proceedings of the Institute of Radio Engineers (IRE), Vol. 49 (1961), 31-48, quote
at 45-47.

40. Richard P. Lippmann, 'An Introduction to Computing with Neural N

IEEE ASSP Magazine, Vol. 4 (1987), 4-22, at 15-18; DARPA, Darpa N
Network Study (Fairfax, VA: Armed Forces Communications and Electr
Association [AFCEA] International Press, 1988), 78-80.
41. The history of the analog computer (which was used for solving differ
equations) goes back to the 1930s, but its golden age was the 1950s and early 19
Voltage precision problems in analog computers contrasted with the advanc
accuracy, speed, memory capacity, miniaturization and programming of di
computers. By the mid-1960s, digital technology was the choice for most comp
users: see Time-Life Books, Alternative Computers (Alexandria, VA: Time
Books, 1989), 26, 27 & 39.
42. Rosenblatt, Principles of Neurodynamics, op. cit. note 15, 10.
43. F. Rosenblatt, 'Two Theorems of Statistical Separability in the Percept
NPL, op. cit. note 14, Vol. I, 421-56, at 422.
44. Jeremy Bernstein, 'Profiles: AI, Marvin Minsky', The New Yorker
December 1981), 50-126, quote at 100.
45. Minsky, interview, 25 October 1989.
46. S.A. Papert, 'One AI or Many?', in Stephen R. Graubard (ed.),
Artificial Intelligence Debate: False Starts, Real Foundations (Cambridge, MA
Press, 1988), 1-14, quote at 4-5, emphasis in original.
47. Minsky & Papert, op. cit. note 13.
48. Of course, I am not using the term 'proof in an absolute sense. He
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 654 Social Studies of Science

relativize this term, as has been done in o

7.
49. Collins, 'Core Set', op. cit. note 1, 7, 8 & 14.
50. Minsky & Papert, op. cit. note 13, 9.
51. Ibid., Chapter 3.
52. Ibid., 8, 17 & Chapter 5.
53. Igor Aleksander and Helen Morton made this comparison: 'Minsky and
Papert's central argument is that perceptrons are only good if their order remains
constant for a particular problem irrespective of the size of the input "retina". This
is similar to the requirement that a program in conventional computing, such as a
routine for sorting a list of numbers, should be largely invariant to the size of the
task. It is accepted that such a program might need to be given the length of the list
as input data, but it would be of little use if it had to be rewritten for lists of
different lengths': I. Aleksander and H. Morton, An Introduction to Neural
Computing (London: Chapman & Hall, 1990), 41.
54. Larry Laudan used the term 'anomalous problem': see L. Laudan, Progress
and its Problems: Towards a Theory of Scientific Growth (Berkeley, CA: University
of California Press, 1977), 29. I am, therefore, not using this term in its Kuhnian
sense (experimental results which do not fit within the accepted categories of a
scientific paradigm).
55. Aleksander and Morton described some simple algorithms for computing
parity and connectedness in input retinas like the one in Figure 6: '(i) Scan the
picture points line by line, left to right, starting at the top left-hand corner of the
image until the first black square is reached. [The blobs are assumed to be black on
a white background.] (ii) Mark this square and find all its black nearest neighbours.
Then mark these neighbours and all their nearest black neighbours and so on until
no new black elements can be found. [This marks all the elements of a blob.] (iii)
Remove all the marked elements [by turning them from black to white: this
removes the blob]. (iv) Scan the image again and if any black element is found, the
image is not connected. The parity task is executed just as easily: the scan-and-
remove procedure can be used as before, it then becomes merely a question of
counting the number of times the blobs have to be cleared. If this number is even,
the image possesses parity' (Aleksander & Morton, op. cit. note 53, 39-40).
56. Minsky & Papert, op. cit. note 13, 72.
57. Ibid., 227.
58. This figure is inspired by the drawing appearing on the front page of Minsky
& Papert, op. cit. note 13.
59. The relationship between human information processing (whether at the
mind or brain level) and machine information processing has been a constant and
especially important rhetorical resource throughout the history of AI, because of
the prominent role played in this discipline by the 'computer metaphor'.
60. H. David Block, 'A Review of Perceptrons', Information and Control, Vol.
17 (1970), 510-22, quote at 517, emphasis in original. Block refers to Minsky &
Papert, op. cit. note 13.
61. This example is taken from Collins' discussion about the distinction between
behaviour and action. In the case of this pseudoletter, two different intentions (the
'A' of 'cat' and the 'H' of 'the') were executed by the same behaviour (the same
movements of the hand): see Collins, op. cit. note 26, 32.
62. Block, op. cit. note 60, 513-14, emphasis in original.

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 655

63. Widrow, interview, 13 November 1989.

64. Pinch, op. cit. note 7, 189 & 205-06. The opinions of some physicists
interviewed by Pinch have strong similarities with Widrow's view; for example:
'Well, I suppose that [most physicists] regard von Neumann's book as a perfectly
adequate formal treatment for pedants, people who like that sort of thing. They
wouldn't read it themselves but they're glad someone has done all that hard work!'
(ibid., 205).
65. The concept of an 'arithmetic ideal' goes back to Georgescu-Roegen's
concept of 'arithmomorphism', and has been developed in the sociology of science
mainly by Richard Whitley: see R. Whitley, 'Changes in the Social and Intellectual
Organization of the Sciences: Professionalization and the Arithmetic Ideal', in
Mendelsohn, Weingart & Whitley (eds), op. cit. note 7, 143-69.
66. For a recent contribution to the issue of proofs of computer-system
correctness, see D. MacKenzie, 'Negotiating Arithmetic, Constructing Proof: The
Sociology of Mathematics and Information Technology', Social Studies of Science,
Vol. 23 (1993), 37-65.
67. Minsky and Papert exploited this point in their own rhetoric, and spoke
about 'mystique' and 'a certain flourish of romanticism' surrounding 'loosely
organized and distributed neural network machines': see Minsky & Papert, op. cit.
note 13, 4 & 18-19. Proving what a symbolic AI system can do is a much more open
and 'softer' question than in conventional computer science (typically, symbolic AI
systems are based on heuristic, rather than algorithmic, searches).
68. Minsky & Papert, op. cit. note 13, 231-32.
69. See Edwards, op. cit. note 12, Chapter 8. 'ARPA' (the Advanced Research
Projects Agency of the US Department of Defense, nowadays known as
'DARPA') was created by the Eisenhower administration in 1958, as a reaction to
the Sputnik launch, and in a period of unprecedented growth in funding for basic
research in the United States (late 1950s and early 1960s). As Jon Guice points out
(see below: op. cit. note 74), ARPA's early programmes were linked to Cold War
policy concerns, including nuclear test detection, and space and missile techno-
logies. From the late 1960s onwards, ARPA's programmes diversified. My interest
is in ARPA's involvement in the development of AI, as described in the text.
70. Edwards, op. cit. note 12, Chapter 8, 317 (draft pagination).
71. Fleck (1982), op. cit. note 12. Fleck points out that up until the mid-1970s,
symbolic AI research was backed almost exclusively by ARPA funding (ibid., 181).
72. McCorduck, op. cit. note 12, 110.
73. Denicoff, interview, 29 November 1989. Denicoff refers to Minsky &
Papert, op. cit. note 13.
74. Jon Guice, 'Lord ARPA and the Battle of Perceptrons: Controversy and the
State in Intelligent Computing, 1958-69', draft paper under consideration by Social
Studies of Science; Guice, Designing the Future: The US Advanced Research
Projects Agency (La Jolla, CA: Department of Sociology, University of California
at San Diego, PhD in progress, working title).
75. Louis Fein, 'The Artificial Intelligentsia', Wescon Technical Papers, Vol. 7
(11.1, Part 7) (1963), 1-7. This paper was later published with some minor changes,
as: L. Fein, 'The Artificial Intelligentsia', IEEE Spectrum (February 1964), 74-87.
76. Including, among others, the following: McCarthy's LISP and Newell's ILP
programming languages; programs which combined algorithmic and heuristic
methods, such as Newell, Simon and Shaw's General Problem Solvers; the Logic
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 656 Social Studies of Science

Theorist theorem-proving program; chess

heuristic problem solving; the Stanford DEN
molecular structure of unknown composite
mathematical problem solving.
77. Papert, op. cit. note 46, 7-8.
78. A. Newell and H.A. Simon, 'Compu
Symbols and Search', Communications of the
Vol. 19 (1976), 113-26, reprinted in John Ha
MA: MIT Press, 1981), 35-66, quotation from
79. A. Newell, 'Intellectual Issues in the H
Fritz Machlup and Una Mansfield (eds), The
ary Messages (New York: John Wiley & Son
added.

80. James Lighthill, Artificial Intelligence (London: Science Research Cou

1973). The three main areas of AI research studied by Lighthill were (neuro
and psychology-oriented) computer-based central nervous system research
and animals, symbolic AI/advanced automation, and robotics. Lighthill con
that success of work under the category 'computer-based central nervous s
research' (neural network-like research) would depend on its close relati
with psychology and neurobiology, in the same way as work on ad
automation/symbolic AI would depend on its close association with its appl
area (engineering): see ibid., 19-21. He also maintained that robotics re
should integrate in those areas.
81. Some representative neural-net contributions of the 1970s were repri
Anderson & Rosenfeld, op. cit. note 14. Vision researcher David Marr work
neural net-like topics until the early 1970s, when he switched to the sy
approach.
82. See, for example, Bill Harvey, 'Plausibility and the Evaluation of Know-
ledge: A Case-Study of Experimental Quantum Mechanics', Social Studies of
Science, Vol. 11 (1981), 95-130, at 126.
83. Collins used this argument in his study of the gravitational radiation
controversy. He pointed out that, in accepting the electrostatic calibration
measuring technique, Joseph Weber restricted the interpretative flexibility of
gravitational radiation results, and chose not to argue on certain fronts which, in
principle, were not entirely implausible: Collins, Changing Order, op. cit. note 1,
104-06. For Everett Hughes's methodological principle, see S. Leigh Star,
'Introduction: The Sociology of Science and Technology', Social Problems, Vol. 35
(1988), 197-205, at 198.
84. For a longer account, see Olazaran (1991), op. cit. note 9, Chapters 4-5, and
Olazaran (1993), op. cit. note 9, 386-417.
85. Fleck (1987), op. cit. note 12.
86. Basically, expert systems are composed of a knowledge-base (where know-
ledge relevant for a certain domain is represented) and techniques for making
inferences from that base in a particular situation or problem. In these knowledge-
based information processing systems, emphasis is laid on symbolic representation
and on the ability of the computer to carry out structure-sensitive transformations
of those representations.
87. Fleck (1987), op. cit. note 12, 153.
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 657

88. Geoffrey E. Hinton and J.A. Anderson, Parallel Models of Associative

Memory (Hillsdale, NJ: Laurence Erlbaum, 1981).
89. Dana H. Ballard, G.E. Hinton and Terrence Sejnowski, 'Parallel Visual
Computation', Nature, Vol. 306 (3 November 1983), 21-26.
90. Jerome A. Feldman and D.H. Ballard, 'Connectionist Models and their
Properties', Cognitive Science, Vol. 6 (1982), 205-54.
91. Parallel computing entails the use of more than one processor wor
concurrently on a problem. According to their 'granularity', parallel architec
can be 'coarse grain' (small number of sophisticated processors) or 'fine g
(large number of simpler processors). According to the instructions receive
each processor, they can be single instruction/multiple data (SIMD) or multi
instruction/multiple data (MIMD, with each processor receiving its own inst
tions).
92. Researchers have also started to design and use special-purpose hardware
for implementing neural nets. For Carver Mead's pioneer work on VLSI analog
circuits, see C. Mead, Analog VLSI and Neural Systems (Reading, MA: Addison-
Wesley, 1989).
93. The two PDP volumes were the manifesto of the new neural-net movement:
see David E. Rumelhart, James L. McClelland and The PDP Research Grou
Parallel Distributed Processing: Explorations in the Microstructure of Cognition
Vol. 1, Foundations (Cambridge, MA: MIT Press, 1986), and McClelland,
Rumelhart and PDP RG, ibid., Vol 2, Psychological and Biological Models
(Cambridge, MA: MIT Press, 1986). The PDP Group made an important
'marketing' effort aimed at bringing neural nets back to the AI and cognitiv
science fields.

94. See, for example: Michael J. Mulkay, 'Three Models of Scientific Develo
ment', Sociological Review, Vol. 23 (1975), 509-26; Mulkay, G. Nigel Gilbert an
S. Woolgar, 'Problem Areas and Research Networks in Science', Sociology, Vol
(1985), 187-203.
95. It has been pointed out that Hopfield was a well-recognized physicist w
could 'afford' to attempt to make a contribution in an area that had not yet b
recognized as a valuable or respectable one: see Anderson & Rosenfeld, op.
note 14, 457.
96. J.J. Hopfield, 'Neural Networks and Physical Systems with Emerge
Collective Computational Abilities', Proceedings of the National Academy
Sciences, Vol. 79 (1982), 2554-58, reprinted in Anderson & Rosenfeld, op.
note 14, 460-64. The crucial aspect of Hopfield's contribution, - a conseque
of his use of the spin-glass metaphor - was the notion of the 'energy' of
(symmetrically-connected) neural net. The energy of a Hopfield system (a glo
measure of its performance) decreases every time a unit updates its state (a l
operation), until a local minimum (a stable state of the system) is reached. Thus t
local activity of each unit contributes to the minimization of a global property of
whole system. Patterns are stored at local minima of the energy function. One
the most important properties of this type of net is that it can work as a conten
addressable memory so that, under the right circumstances, it will retrieve corre
whole patterns when presented with degraded versions of input patterns.
97. Hopfield, op. cit. note 96 (reprinted version), 460.
98. David H. Ackley, G.E. Hinton and T.J. Sejnowski, 'A Learning Algorith
This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 658 Social Studies of Science

for Boltzmann Machines', Cognitive Science, Vol. 9 (1985), 147-69, reprinted in

Anderson & Rosenfeld, op. cit. note 14, 638-49.
99. Sejnowski, interview, 8 November 1989. See also Ackley, Hinton &
Sejnowski, op. cit. note 98 (reprinted version), 641.
100. T.J. Sejnowski and Halbert White, Introduction to reprinted version of Nils
J. Nilsson, The Mathematical Foundations of Learning Machines (San Mateo, CA:
Morgan Kaufmann, 1991; original version 1965), vii-xxi, quote at xii.
101. 'The Boltzmann machine is a generalization of the Perceptron to more than
one layer. It's interesting, it turned out that the key assumptions you had to change
were two things. First that there are feedback connections a la Hopfield, so you
have symmetric connections, so it's no longer feedforward net but symmetric net
with feedback connections. And second of all, the Perceptron was a deterministic
machine, whereas the Boltzmann machine was probabilistic. So you make those
two changes, and then suddenly it's a completely different architecture, suddenly
you can prove theorems, you can discover learning algorithms, you can solve
problems that the Perceptron couldn't', Sejnowski, interview, 8 November 1989.
102. D.E. Rumelhart, G.E. Hinton and Ronald J. Williams, 'Learning Internal
Representations by Error Propagation', in Rumelhart, McClelland & PDP RG, op.
cit. note 93, 318-62.
103. A learning cycle in a back-propagation net can be summarized as follows. A
pattern p is presented, activity propagates forward throughout the units, and the
network produces an output. This output is compared with the desired output, and
the error made by the output units is calculated. Then, before any weight
adjustment is made, the backward stage starts. The errors made by the output units
are back-propagated to the hidden units, so that the error made by each hidden unit
can be calculated. Now all the connections in the system can be changed. If there
were more layers of connections, those layers would be adjusted in the same way.
By adjusting the connections of the system according to this technique, the total
error measure for a set of input/output patterns is minimized in a gradient descent
way. See D.J. Rumelhart, G.E. Hinton and R.J. Williams, 'Learning Representa-
tions by Back-propagating Errors', Nature, Vol. 323 (9 October 1986), 533-36,
reprinted in Anderson & Rosenfeld (eds), op. cit. note 14, 696-99, at 697.
104. Sejnowski & White, op. cit. note 100, xv.
105. Outside the context of the neural-net revival, Paul Werbos's algorithm for
multilayer nets was not considered practical by Minsky in the 1970s: see Olazaran
(1993), op. cit. note 9, 396-406. For Werbos's contributions see: P.J. Werbos,
Beyond Regression: New Tools for Prediction and Analysis in the Behavioral
Sciences (Cambridge, MA: unpublished PhD dissertation, Harvard University,
1974); Werbos, 'Applications of Advances in Nonlinear Sensitivity Analysis', in
R.F. Drenick and F. Kozin (eds), Systems Modelling and Optimization: Proceed-
ings of the 10th IFIP Conference, New York City, 31 August-4 September 1981
(New York: Springer-Verlag, 1982), 762-70.
106. See Richard Forsyth, 'The Brain Mimics Are Back in Business', The
Guardian (London, 12 January 1989), 25.
107. Rumelhart, Hinton & Williams, op. cit. note 93, 321, 324 & 361 (emphasis
in original).
108. The role of the PDP Group in rewriting the official history of the
controversy helps explain Grossberg's priority complaints. Grossberg claimed tha
some of his models were rediscovered in the neural-net revival, and that he did not

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms
 Olazaran: Official History of the Perceptrons Controversy 659

receive enough recognition for his previous work: see S. Grossberg, 'Competitive
Learning: From Interactive Activation to Adaptive Resonance', Cognitive Science,
Vol. 11 (1987), 23-63.
109. M.L. Minsky and S.A. Papert, Perceptrons: An Introduction to Computa-
tional Geometry (Cambridge, MA: MIT Press, 1988), 260-61; this is a second,
enlarged edition of Minsky & Papert, op. cit. note 13.
110. See, for example, Olazaran (1991), op. cit. note 9, 274-75.
111. Ibid., 282-92.
112. Ibid., 276-81.
113. Ibid., 293-307.
114. I have adopted these categories from Pinch's case study: see Pinch, op. cit.
note 7.

Mikel Olazaran carried out his doctoral research in the

Department of Sociology at the University of Edin
(UK), receiving his PhD in 1991. He is now a lectu
sociology at the University of the Basque Country,
member of the ILCLI research institute at that universit
is conducting research on the sociology and history
(he is preparing a book based on his PhD thesis), o
social dimensions of information technology, and o
science and technology system of the Basque-Na
region.
Author's address: Departamento de Sociologia, Facultad de
Ciencias Sociales y de la Comunicaci6n, Universidad del
Pais Vasco, Apartado 644, 48080 Bilbao, Spain.
Fax: +34 4 464 8299; e-mail: cipolrom@lg.ehu.es

This content downloaded from 146.169.183.232 on Tue, 21 May 2019 15:48:44 UTC
All use subject to https://about.jstor.org/terms