CONSTRUCTIVIST

FOUNDATIONS
Volume 1, Number 2
Date of publication: March 2006

Ernst
Ernst von
von Glasersfeld
Glasersfeld
A
A Constructivist
Constructivist Approach
Approach to
to
Experiential
Experiential Foundations
Foundations of
of
Mathematical
Mathematical Concepts
Concepts Revisited
Revisited
Kevin
Kevin McGee
McGee
Enactive
Enactive Cognitive
Cognitive Science
Science
Part
Part 2:
2: Methods,
Methods, Insights,
Insights, and
and Potential
Potential
Bernd
Bernd Porr,
Porr,Alice
Alice Egertony
Egertony &
&
Florentin
Florentin Wörgötter
Wörgötter
Towards
Towards Closed
Closed Loop
Loop Information:
Information:
Predictive
Predictive Information
Information

An interdisciplinary journal
http://www.univie.ac.at/constructivism/journal
 Advisory Board DESCRIPTION
William Clancey Constructivist Foundations (CF) is an independent academic peer-reviewed e-journal
NASA Ames Research Center, USA without commercial interests. Its aim is to promote scientific foundations and
Ranulph Glanville applications of constructivist sciences, to weed out pseudoscientific claims and to base
CybernEthics Research, UK constructivist sciences on sound scientific foundations, which do not equal the
Ernst von Glasersfeld scientific method with objectivist claims. The journal is concerned with the
University of Massachusetts, USA interdisciplinary study of all forms of constructivist sciences, especially radical
Vincent Kenny constructivism, cybersemiotics, enactive cognitive science, epistemic structuring of
Inst. of Constructivist Psychology, Italy experience, second order cybernetics, the theory of autopoietic systems, etc.
Klaus Krippendorff
University of Pennsylvania, USA AIM AND SCOPE
Humberto Maturana
Institute Matríztica, Chile
The basic motivation behind the journal is to make peer-reviewed constructivist
papers available to the academic audience free of charge. The “constructive”” character
Josef Mitterer
University of Klagenfurt, Austria of the journal refers to the fact that the journal publishes actual work in constructivist
sciences rather than work that argues for the importance or need for constructivism.
Karl Müller
Wisdom, Austria The journal is open to (provocative) new ideas that fall within the scope of
Bernhard Pörksen
constructivist approaches and encourages critical academic submissions to help
University of Hamburg, Germany sharpen the position of constructivist sciences.
Gebhard Rusch The common denominator of constructivist approaches can be summarized as
University of Siegen, Germany follows.
Siegfried J. Schmidt • Constructivist approaches question the Cartesian separation between objective
University of Münster, Germany world and subjective experience;
Bernard Scott • Consequently, they demand the inclusion of the observer in scientific
Cranfield University, UK explanations;
Sverre Sjölander • Representationalism is rejected; knowledge is a system-related cognitive process
Linköping University, Sweden rather than a mapping of an objective world onto subjective cognitive structures;
Stuart Umpleby • According to constructivist approaches, it is futile to claim that knowledge
George Washington University, USA approaches reality; reality is brought forth by the subject rather than passively
Terry Winograd received;
Stanford University, USA • Constructivist approaches entertain an agnostic relationship with reality, which is
considered beyond our cognitive horizon; any reference to it should be refrained
Editor-In-Chief from;
Alexander Riegler • Therefore, the focus of research moves from the world that consists of matter to
Vrije Universiteit Brussel, Belgium the world that consists of what matters;
• Constructivist approaches focus on self-referential and organizationally closed
Editorial Board systems; such systems strive for control over their inputs rather than their
Pille Bunnell outputs;
Royal Roads University, Canada • With regard to scientific explanations, constructivist approaches favor a process-
Olaf Diettrich oriented approach rather than a substance-based perspective, e.g. living systems
Center Leo Apostel, Belgium are defined by processes whereby they constitute and maintain their own
Dewey Dykstra organization;
Boise State University, USA • Constructivist approaches emphasize the “individual as personal scientist”
Stefano Franchi approach; sociality is defined as accommodating within the framework of social
University of Auckland, New Zealand interaction;
Timo Honkela • Finally, constructivist approaches ask for an open and less dogmatic approach to
Helsinki Univ. of Technology, Finland science in order to generate the flexibility that is needed to cope with today’s
Theo Hug scientific frontier.
University of Innsbruck, Austria
Urban Kordes
Institut Jozef Stefan, Slovenia
SUBMISSIONS
Albert Müller Language: Papers must be written in English. If English is a foreign language for you
University of Vienna, Austria please let the text be proofread by an English native speaker.
Herbert F. J. Müller Copyright: With the exception of reprints of “classical” articles, all papers are
McGill University, Montreal, Canada “original work,” i.e., they must not have been published elsewhere before nor must
Markus Peschl they be the revised version (changes amount to less than 25% of the original) of a
University of Vienna, Austria published work. However, the copyright remains with the author and is licensed
Bernd Porr under a Creative Commons License.
University of Glasgow, UK Author’s guidelines at http://www.univie.ac.at/constructivism/guideline.pdf
John Stewart
Univ. de Technologie de Compiègne, France Send all material to Alexander Riegler ariegler@vub.ac.be.
Wolfgang Winter
Important: Use “Constructivist Foundations” in the “Subject:” field.
Univ. of Cooperative Education, Germany
Tom Ziemke For more information please visit the journal’s website at
University of Skövde, Sweden http://www.univie.ac.at/constructivism/journal

Cover page: Untitled (Large Pacific Landscape) by Frank Galuszka. Oil, acrylic and mica on canvas (78" × 108"), 2005.
With kind permission of the artist, http://arts.ucsc.edu/faculty/fgaluszka
 cognitive–psychological radical constructivism
CONCEPTS

A Constructivist Approach to
Experiential Foundations of
Mathematical Concepts Revisited
Ernst von Glasersfeld A University of Massachusetts, evonglas@localnet.com

Purpose: The paper contributes to the naturalization of epistemology. It suggests ten-

Introduction
tative itineraries for the progression from elementary experiential situations to the In a popular lecture, given at Heidelberg in
abstraction of the concepts of unit, plurality, number, point, line, and plane. It also pro- 1870, Hermann von Helmholtz said that it was
vides a discussion of the question of certainty in logical deduction and arithmetic. the relation of geometry to the theory of cog-
Approach: Whitehead's description of three processes involved in criticizing mathe- nition that emboldened him to speak of geo-
matical thinking (1956) is used to show discrepancies between a traditional epistemo- metrical subjects.1 It is in precisely that spirit
logical stance and the constructivist approach to knowing and communication. Practical that I venture into the domain of mathemati-
implications: Reducing basic abstract terms to experiential situations should make cal thinking – not as a practitioner of that spe-
them easier to conceive for students. Key words: Foundations of mathematics, concept cific art but as a student of conceptual con-
formation, conceptual semantics. struction who also has an interest in
education. My purpose is not to discuss math-
ematics as it may appear to mathematicians
Mathematics is the science of acts number, point, line, and plane could be exercising their craft, but rather to suggest a
without things – and through this, derived from ordinary experience. I have way to think of the conceptual origin of some
of things one can define by acts added a postscript with references to and basic building blocks without which mathe-
Paul Valéry (1935, p. 811) brief comments on publications by Brian matics as we know it could not have devel-
Rotman, Penelope Maddy, and George oped. The constructivist approach puts in
Lakoff as samples of recent voices. question the notion of universal conceptual
Preface During the 1980s radical constructivism “objects” and their ontological derivation and,
gained a certain currency in the fields of sci- consequently, argues against Platonic or
This is a revised version of a paper which, as ence and mathematics education. Although Chomskyan innatism or any other metaphys-
an anonymous reviewer guessed, was writ- cognitive constructivists have occasionally ical foundationalism.
ten some time ago; in fact I wrote it in 1990/ referred to the intuitionist approach to the I realize that some may consider any inves-
91 for the 2nd International Conference on foundational problems in mathematics, no tigation of conceptual development an intru-
the History and Philosophy of Science in Sci- effort has so far been made to outline what a sion of psychologism, but even philosophers
ence Teaching. Queen’s University, King- constructivist’s own approach might be. This of mathematics speak of the derivation of
ston, Ontario, 1992. Much has happened paper attempts a start in that direction. notions: “The ideas, now in the minds of con-
since then in the philosophy of mathemat- Whitehead’s (1956, p. 393) description of temporary mathematicians, lie very remote
ics, especially regarding the “naturalization” three processes involved in criticizing mathe- from any notions which can be immediately
of the highly abstract concepts that were at matical thinking is used to show discrepancies derived by perception through the senses;
the core of the debate on the reality of math- between a traditional epistemological stance unless indeed it be perception stimulated and
ematical objects. As I understand it, natural- and the constructivist approach to knowing guided by antecedent mathematical knowl-
ization in that discipline is the attempt to and communication. The bulk of the paper edge.” (Whitehead 1956, p. 393)
illuminate the foundations of mathematics then suggests tentative itineraries for the pro- How ideas are derived is, after all, a legiti-
by mathematical rather than philosophical gression from elementary experiential situa- mate question for cognitive psychology, and to
thinking and it leads to the dismantling of tions to the abstraction of the concepts of conjecture paths that might lead from the
the Platonist notion that mathematical unit, plurality, number, point, line, and senses to mathematical abstractions seemed a
objects “exist” in an absolute sense. I fer- plane, whose relation to sensory–motor expe- tempting enterprise, because awareness of
vently agree with this dismantling, but my rience is usually ignored or distorted in math- some experiential building blocks could help
approach is on a much lower level of ematics instruction. There follows a discus- to humanize a subject that all too often seems
abstraction and focuses on how the most sion of the question of certainty in logical forbidding to students. I was encouraged to
elementary concepts, such as unit, plurality, deduction and arithmetic. pursue that question by the published evi-

Constructivist Foundations 2006, vol. 1, no. 2 61

 cognitive–psychological
CONCEPTS
dence of wide-spread dissatisfaction with the
traditional dogma among philosophers and
mathematicians themselves (Lorenzen 1974;
Wittenberg 1968; Lakatos 1976; Davis &
Hersh 1981; Quine 1969, Tymoczko 1986a,b;
Mittelstrass 1987). Since the ontological foun-
dations of mathematics have again been put
into question during recent decades, and since
more and more often it is acknowledged that
mathematics is the product of the human
mind2, the approach from the point of view of
mental operations abstracted from experience
should no longer be considered inadmissible.
Indeed, once one relinquishes Plato’s notion
that all ideas are prefigured in every newborn’s
head, one cannot avoid asking how they could
possibly be built up.
One might object that such an approach
would be an incestuous undertaking because
it obviously starts with some, albeit rudimen-
tary, ideas of what the building blocks might
be. To this I would answer that the very same
pertains to all epistemological investigations,
because questions about human knowledge
are inevitably asked and tentatively answered
by a human knower. This was inherent in
Vico’s (1710/1858) slogan “verum ipsum fac-
tum” (the true is the same as the made) and it
was independently and more explicitly formu-
lated by Kant, when he wrote “reason can
grasp only what she herself has produced
according to her design” (Kant 1787/1902, p.
XIII). My purpose, therefore, is to isolate pos-
sible preliminary steps of the construction. But
first I want to show my route of approach.
In his “proposals for reviving the philoso-
phy of mathematics,” Reuben Hersh (1986, p.
22) writes: “What has to be done in the philos-
ophy of mathematics is to explicate (from the
outside, as part of general human culture,
rather than from the inside, within mathemat-
ical terms) what mathematicians are doing.”
I am not a mathematician, and my remarks
are therefore not in mathematical terms. But it
should also be quite clear that I am not offer-
ing them as part of general human culture,
because the radical constructivist orientation
from which they spring is certainly not gen-
eral. In my view, the part of human culture
that concerns questions of knowledge and
knowing, not specifically in mathematics, but
in the entire experiential field, suffers from
precisely the same ambivalence and hypocrisy
that Hersh imputes to the philosophy of math-
ematics.
62
When Hersh writes, a few sentences after
the quoted passage, that such an explication
would present “the kind of truth that is obvi-
ous once it is said, but up to then was perhaps
too obvious for anyone to bother saying,” he
manifests faith in philosophical perspicacity
far greater than the perspicacity shown by our
general culture in the course of the two thou-
sand five hundred years since epistemology
began. My constructivist orientation is radical
because it proposes to cut the cognizing activ-
ity and its results loose from the traditional
dependence on an assumed ontology. It is an
attempt to do without the notion of truth as a
representation of an experiencer-indepen-
dent reality, material or metaphysical (cf. Gla-
sersfeld 1989). Hence, the approach I am
expounding here may upset not only “Pla-
tonist” mathematicians but all who are philo-
sophically or emotionally tied to some form
of realism. My intention, however, is simply to
contribute to a discussion that is still wide
open.
Mathematics and
communication
Concerning criticism in mathematical mat-
ters, Whitehead (1956, p. 395) explained that
“…there are always three processes to be kept
perfectly distinct in our minds.” The way
Whitehead has formulated and explicated
these three processes provides a good basis for
laying out some features of the constructivist
approach.3 We must first scan the purely
mathematical reasoning to make sure that
there are no mere slips in it – no casual illogi-
calities due to mental failure. Any mathema-
tician knows from bitter experience that in
first elaborating a train of reasoning, it is very
easy to commit a slight error which yet makes
all the difference. But when a piece of mathe-
matics has been revised, and has been before
the expert world for some time, the chance of
casual error is almost negligible. It seems dif-
ficult not to agree with this statement, since it
describes a checking procedure with which we
are quite familiar (e.g. when someone is hang-
ing a picture, drawing a map, writing a com-
puter program, and so forth). On second
thought, however, we may notice that White-
head refers to an expert’s check of another
person’s “purely mathematical reasoning,”
which he then calls “a piece of mathematics.”
radical constructivism
This may prompt us to ask where the expert
finds the entity that is to be checked. If it was
a piece of mathematical reasoning, it must
have been generated by someone’s thinking.
Since we cannot read minds, access to
another’s thinking or reasoning requires an
act of communication. In other words, before
a piece of reasoning can be checked, it must be
formulated in some language or symbols that
are known to both the author and the expert
who is to do the checking. (From the con-
structivist point of view, communication is
itself problematic, and I shall deal with it in
the context of Whitehead’s second process,
where it is even more relevant.)
For formalists, there should not be much
of a communication problem. They take the
presented symbols as they find them, and
check whether or not they have been com-
bined according to generally accepted rules.
Formalists are concerned with syntax, not
with conceptual semantics. The author’s acts
of reasoning would not be questioned,
because – although formalists do not usually
say this explicitly – from their perspective,
symbols have to be perceived but need no
interpretation, and the correctness or error of
a mathematical expression depends exclu-
sively on its formal compliance with the rules
of the chosen symbol system.
That this is not a very satisfactory
approach to questions of the mathematical
underground, has been remarked by many
critics during recent decades. Hersh (1986,
p. 19) has expressed the objection in a very
general way: “Symbols are used as aids to
thinking just as musical scores are used as aids
to music. The music comes first, the score
comes later.” That the score ought to conform
to the rules of the scoring system, therefore, is
simply a precondition to any judgment con-
cerning what the score might represent on the
conceptual level. Mathematical symbols, of
course, are far more complex and layered than
musical notation and some of the conceptual
“music” they are intended to signify presum-
ably arises on the higher levels of symboliza-
tion. In what follows, however, I want to focus
on the very lowest level, the level on which
non-mathematical experiences provide
material for the abstraction of the most ele-
mentary mathematical building blocks.
Having discussed the possibility of “mere
slips,” Whitehead turns to the starting-points
of mathematical reasoning.
Constructivist Foundations
 cognitive–psychological
CONCEPTS
He writes: “The next process is to make
quite certain of all the abstract conditions
which have been presupposed to hold. This is
the determination of the abstract premises
from which the mathematical reasoning pro-
ceeds. This is a matter of considerable diffi-
culty. In the past quite remarkable oversights
have been made, and have been accepted by
generations of the greatest mathematicians.
The chief danger is that of oversight, namely,
tacitly to introduce some condition, which it
is natural for us to presuppose, but which in
fact need not always be holding.” (Whitehead
1956, pp. 395–396).
The presupposition of unwarranted condi-
tions is something radical constructivism
claims to have unearthed in several areas that
have no obvious connection with mathemat-
ics. One area, however, that is relevant for the
discussion of thinking and, as philosophers of
mathematics have found, is indispensable for
any critique of mathematical reasoning, is the
kind of social interaction we call linguistic or
symbolic communication (cf. Davis & Hersh
1981; Tymoczko 1986a).
Where communication is concerned, we
habitually – and hence mostly tacitly – oper-
ate on the presupposition that others who use
language or symbols that we readily recognize
as such, are using them with the same mean-
ings that we have come to attribute to them.
This assumption is made habitually, because
without it, most if not all of our everyday lin-
guistic and symbolic interactions would be
futile. Indeed, we are constantly reinforced to
assume that our meanings are shared by oth-
ers, because by and large our ordinary com-
municatory interactions work remarkably
well. But the bulk of our ordinary communi-
cating is about two experiential areas. Either it
concerns sensory–motor objects, where per-
ceptual feedback helps us to avoid gross mis-
interpretation; or it concerns emotions, and in
the emotional sphere, where meanings are
notoriously vague, the margin for interpreta-
tion is so wide that we are rarely compelled to
consider feedback that upsets the pleasant
generic feeling of understanding or being
understood.
In contrast, when we are communicating
within the relatively systematized domain of a
science, we are dealing to a large extent with
abstract concepts and relations. In this area,
such feedback as we do receive about the
other’s reasoning usually springs from infer-
2006, vol. 1, no. 2
ences we draw from the technical context, as
we see it, or from the other’s actions, as we
observe them. Whether the respective concepts
are actually the same, cannot be ascertained,
but insofar as scientific training imposes con-
ventions of thinking, a certain degree of com-
patibility can be assumed.
Social adaptation and
compatibility
For Whitehead, almost twenty years after
Principia Mathematica, there was no doubt
that there was a fixed set of logical rules that
formed the solid basis of mathematics. And
the rules of logic were taken to be a priori and
therefore not only unquestionable but also
inevitably inherent in every thinker’s rational
procedures.
Like the methodology of Ceccato’s Italian
operationalist school4, radical constructivism
is an attempt to do without the assumption of
a priori categories or rules. Categories are seen
as the results of mental construction (as in
Piaget’s theory, in the case of space, time, and
causality5); the sort of rules that make possible
the repetition and checking of logical proce-
dures are based on the coordination of specific
mental operations with specific symbols. This
coordination has to be accomplished by every
single thinking subject, and the “intersubjec-
tivity” that makes possible the communica-
tion of logical procedures can be achieved only
through the individuals’ social interaction and
mutual adaptation of their subjective coordi-
nations (of mental operations and symbols).
This is what Maturana (1980) has called a
“consensual domain” generated by the “coor-
dination of the coordinations of actions.” In
this view, then, the meaning of both natural
language and mathematical symbols is not a
matter of “reference” in terms of indepen-
dently existing entities, but rather of subjective
mental operations which, in the course of
social interaction in experiential situations,
achieve a modicum of intersubjective compat-
ibility.
After his exposition of the three critical
processes, Whitehead made the very impor-
tant general remark: “…the trouble is not with
what the author does say, but with what he
does not say. Also it is not with what he knows
he has assumed, but with what he has uncon-
sciously assumed” (p. 396). This becomes par-
radical constructivism
ticularly relevant when he discusses the third
process.
“This third process of criticism is that of
verifying that our abstract postulates hold for
the particular case in question. It is in respect
to this process of verification for the particular
case that all the trouble arises. In some simple
instances, such as the counting of forty apples,
we can with a little care arrive at practical cer-
tainty. But in general, with more complex
instances, complete certainty is unattainable.
Volumes, libraries of volumes, have been writ-
ten on the subject. It is the battleground of
rival philosophers. There are two distinct
questions involved. There are particular defi-
nite things observed, and we have to make sure
that the relations between these things really
do obey certain definite exact abstract condi-
tions. There is great room for error here. The
exact observational methods of science are all
contrivances for limiting these erroneous con-
clusions as to direct matters of fact.” (p. 396)
Whitehead then discusses the problem of
ascertaining that an experiential object is actu-
ally of “the same sort,” so that one can ascribe
to it a condition that was abstracted from a
prior sample. “The theory of Induction,” he
says, “is the despair of philosophy – and yet all
our activities are based upon it.”
The problem of induction which, as
Whitehead defines it here, is the question
whether one could justify the generalization of
an idea that was abstracted from a particular
sample of experiences, concerns the entire
domain of science, not specifically mathemat-
ics. Here, I therefore merely emphasize that,
from the constructivist point of view, “partic-
ular definite things observed” and “direct mat-
ters of fact” are generated in ways that differ
from the conventional realist account implied
by this passage from Whitehead. The “practi-
cal certainty” in the case of the forty apples,
however, falls within the realm of experiences
that are crucial to the thesis I want to develop,
because according to it, the certainty does not
spring from the counted “objective” things,
but from the mental operations of the counter
(cf. for instance Piaget 1977, p. 71).
I have mentioned that, from the construc-
tivist point of view, there are problems in dis-
cussing an individual’s mathematical reason-
ing, because, in order to be discussed,
criticized, or assessed in any way, this reason-
ing must first be uttered or written and then
interpreted by the critics or evaluators. To this
63
 cognitive–psychological
CONCEPTS
I now want to add that whatever Whitehead
intended by the expression “direct matters of
fact” is not something that is an unquestion-
able given, but rather the result of an individ-
ual’s interpretation of experience which, to
that individual, appears to be compatible with
the interpretations of other individuals.
Compatibility – and this cannot be stressed
too much – is not the same as identity. The dis-
tinction is important, because it alters the
notion of “understanding,” of “shared” ideas
and conceptual structures, and thus also the
notion of “accepting a proof.” When I agree
with what another says or does, it means no
less, but also no more, than that I interpret the
statement or the action in a way that manifests
no discrepancies from what I might say or do
under the given circumstances as I see them.
This leaves room for uncounted discrepancies
that did not happen to surface; and we all
know how often, after a feeling of agreement,
a feeling of harmony and complete under-
standing, some further interaction brings out
a discrepancy that had remained hidden in all
that went before.
Tymoczko (1986b, p. 48) maintains that
the type of criticism that is necessary for the
development of proofs is essentially social:
“Without criticism, proof-ideas cannot
develop into proofs.” In other words, when
some reasoning, presented in speech or writ-
ing, is claimed to constitute a proof, this com-
munication has to be interpreted and criti-
cized by others. Then the critique has to be
interpreted and answered by the author, and
only when the critics interpretation of the
answer is judged adequate by them, will the
presented reasoning be promoted to the status
of proof. No matter how many iterations this
procedure might go to, it is clear that it can
never guarantee the identity of concepts and
conceptual relations used by the different
thinkers involved. It can at best lead to appar-
ent compatibility – and this, of course, is quite
sufficient for the practice of mathematics (or
any other domain of human cooperation).
But if compatibility is all that can be
achieved, we cannot found mathematics in an
ontological realm of ideas presumed to exist
independently of any thinking subject. Yet
Lakatos (1976/1986, p. 44) was probably right
when he remarked: “It will take more than the
paradoxes and Gödel’s results to prompt phi-
losophers to take the empirical aspects of
mathematics seriously…” From the outside,
64
however, witnessing the crumbling of what
were considered the foundations of mathe-
matics, it seems reasonable to ask what the
most elementary building blocks could be
that might serve as a basis for the constitution
of the mysterious structures or “objects” that
mathematics develops. To pursue that quest
requires an empirical investigation, where
“empirical” has its original meaning and
refers to experience. But clearly it will not be
sensory experience that matters, but the expe-
rience of mental operations. As Hersh (1986,
p. 22) put it: “mathematics deals with ideas.
Not pencil marks or chalk marks, not physical
triangles or physical sets, but ideas (which
may be represented or suggested by physical
objects).”
If we do not want to believe that ideas are
innate or God-given, but the result of subjec-
tive thinkers’ conceptual activity, we have to
devise a model of how elementary mathemat-
ical ideas could be constructed – and such a
model will be plausible only if the raw mate-
rial it uses is itself not mathematical.
To begin at the
beginning
To have the idea of counting, one needs the expe-
rience of handling coins or blocks or pebbles. To
have the idea of angle, one needs the experience
of drawing straight lines that cross, on paper or in
a sand box. (Hersh 1986, p. 24)
Unless we are able to conceive of something
that is unitary, in the sense that we distinguish
it as a discrete item in our experiential field
and are able to “recognize” other items like it,
we cannot have a plurality. If we have no plu-
rality, we have no occasion to count – and if we
did not count, it is unlikely that we should ever
have arithmetic and mathematics, because
without counting there would be no numbers.
Hence, if we want to get some inkling as to how
arithmetic arises, we may have to begin with
the concepts of unit and plurality, as well as the
activity of counting which generates the con-
cept of number. The ensuing development was
described by Paul Lorenzen (1974, p. 199):
“The foundation of arithmetic is the pre-
arithmetical praxis: the use of counting-signs
(e.g. |, ||, |||, ||||, …) in counting collections
(heaps, herds, groups, complexes, …); using
counting-signs rather than the collections
radical constructivism
themselves, to make comparisons of amounts;
adding and subtracting counting-signs
(instead of certain operations with the collec-
tions).”
Rotman (1987, p. 8) uses the same nota-
tion of counting signs, but both he and Loren-
zen seemed to take the concepts of unit and
collection for granted. Yet, the “prearithmeti-
cal praxis” obviously does not begin with
counting-signs. Before any collections can be
counted and coordinated with signs that can
be used in their stead, experiential items must
be gathered in collections such as heaps or
herds. To do this, we must distinguish more
than one experiential item, i.e., a plurality,
such that each of the individual items satisfies
whatever conditions govern membership in
the particular heap, herd, or collection we
want to form.
The first task, then, is the distinction of
individually discrete “things” in our experien-
tial field. To normal adult humans, who are
experienced managers of a more or less famil-
iar environment, it may seem absurd to sug-
gest that the segmentation of their experien-
tial world into discrete things should not be
an ontological given. But even the most
orthodox epistemologists, at least since the
days of John Locke, have discarded this com-
monsense notion. Many modern scientists,
e.g., Mach (1910/1970, p. 42) and Bridgman
(1961, p. 46), explicitly stated this, and Albert
Einstein (1954, p. 291) formulated one of the
simplest, uncompromising descriptions of
how we come to furnish our world with dis-
crete things:
“I believe that the first step in the setting of
a ‘real external world’ is the formation of the
concept of bodily objects and of bodily
objects of various kinds. Out of the multitude
of our sense experiences we take, mentally
and arbitrarily, certain repeatedly occurring
complexes of sense impressions (partly in
conjunction with sense impressions which
are interpreted as signs for sense experiences
of others), and we correlate to them a
concept – the concept of the bodily object.
Considered logically this concept is not iden-
tical with the totality of sense impressions
referred to; but it is a free creation of the
human (or animal) mind.”
Piaget (1937) provided a minute analysis
of how object concepts might be constructed
by the very young child, and from Edmund
Husserl we have a suggestion that is particu-
Constructivist Foundations
 cognitive–psychological
CONCEPTS
larly relevant to the present context. In his
Philosophie der Arithmetik (1887/1970), Hus-
serl proposed that the mental operation that
unites different sense impressions into the
concept of a “thing” is similar to the operation
that unites abstract units into the concept of a
number (p.157–168). I accept this hypothesis,
but want to point out that, in order to have
several units that can be united to form a
number, one must have a concept of plurality.
I therefore want to unravel the steps involved
in the initial development and fill in some of
the details that seem necessary.
Compounds of sensory impressions pre-
sumably acquire their first stability, as Ein-
stein suggested, through repetition. Brouwer
(1949) proposed that the perceiving subject’s
self-directed attention “performs identifica-
tions of different sensations and of different
complexes of sensations, and in this way, in a
dawning atmosphere of forethought, creates
iterative complexes of sensations” (p.1235;
emphasis in the original). The stability of
such a sensory compound manifests itself in
the subject’s ability to “recognize” it when it is
produced again. Children clearly show this in
the early stages of language acquisition. Once
they have isolated a group of sense impres-
sions and have associated them with, say, the
word “cup,” they may toddle to the kitchen
table and, pointing with their finger, say
“cup,” then point to another and say “cup,”
and repeat this procedure for every cup they
happen to perceive on the table. Psycholin-
guists call this phenomenon “labeling” and it
provides good evidence that some kind of
structure has been formed which allows chil-
dren to utter the associated word whenever
their perceptual mechanisms produce sen-
sory signals that can be fitted into that partic-
ular structure.
The abstraction of
plurality
However, to say that a child that does this has
a concept of plurality, would be reading too
much into the episode. In fact, it usually
takes at least a month or two before the child
will use the plural “cups” instead of acknowl-
edging individual cups singly. This delay
cannot be explained by the simple fact that
the child has to learn a different word (i.e.,
the plural form), because in order to use the
2006, vol. 1, no. 2
new word, the child also has to learn to iso-
late a different experiential situation. The
plural “cups” must be associated with a per-
ceptual situation that contains more than one
cup – and this “more than one” is not a per-
ceptual fact. The inference that more than
one cup is on a table is not based on the sen-
sory impressions isolated as cups, but on the
awareness that one has repeated the same
operations of isolating and recognizing
within certain boundaries of space and time.
The conception of a plurality, therefore, is
the result of a repetition of mental opera-
tions that accompany the sensory impres-
sions but are themselves not sensory. And the
basis on which the conceptual structure
called “set” can be created is laid only when,
in a further step of abstraction, the opera-
tions that generated a plurality are seen as an
operational pattern without considering the
sensory items that constituted the
collection6.
Consequently, the notion of a plurality of
things, for which language supplies the plu-
ral form of the things’ name, should not be
considered a mathematical concept. Though
it results from mental operations, it is tied to
sensory experience and does not require an
abstract concept of unit. Indeed, Husserl
makes clear that the ordinary meaning of the
word “one” (used in opposition to a plural-
ity, as in “more than one”) must be distin-
guished from the numerical concept of “one,”
which designates an abstract unit (Husserl
1887/1970, pp. 128ff).
That the abstract units required in arith-
metic are not quite the same as the units con-
stituted by the discrete objects in our experi-
ential world, was indicated also by Frege
(1884/1974, p. 58), when he said that the
things we number must be distinguishable,
whereas the units of arithmetic are not,
because they are conceptual and have to be
identical in every instance.
Brouwer suggested that the fact that
attention can be directed, enables the mind
to produce the complexes of sensory ele-
ments that we perceive as “things,” and that
this ability derives from the inherent charac-
ter of consciousness which, in his view, was
not a steady state but an oscillating function
(Brouwer 1949, p. 1235). A similar idea was
proposed quite independently, by Silvio
Ceccato (1966), who posited a pulsating
attentional mechanism that generates pat-
radical constructivism
terns of focused and unfocused pulses which
can then be used to segment sensory mate-
rial into iterable conceptual structures. I
have further developed this approach in a
hypothetical model of the construction of
the concepts of unit and number (Glasers-
feld 1981). The model leans on Piaget’s
notion of reflective abstraction7 and offers at
least an hypothetical skeleton of mental
operations that lead from the inception of
discrete sensory things to pluralities, collec-
tions, arithmetic units, set, and number.
Here, however, I want to concentrate, not on
the mechanisms of abstraction, but on expe-
riential elements that may serve as its raw
material. In what follows, I shall make the
case that counting provides the most plausi-
ble basis for the abstraction of the concept of
“numerosity” (the cardinal aspect of num-
ber).
Counting and number
If we want to agree with what Hersh says in
the quotation I have placed at the beginning
of this section, it will be necessary to make
quite clear what we mean by “counting.” The
word has been used for diverse manifesta-
tions that range from a toddler’s meaningless
recitation of a few number words to the
function of a gadget that indicates radioac-
tivity. From the constructivist point of view,
counting is a very specific, complex activity
(cf. Steffe et al. 1983). According to our def-
inition, it has three components: A conven-
tional number word sequence, a plurality of
unitary sensory items (perceived or visual-
ized), and the one-to-one coordination of
successive number words and the items in
the collection.
Imagine an ordinary, nonphilosophical
observer watching a mason who utters the
standard sequence of number words as he
points to the bricks lying beside him. If
“eighteen” turns out to be the last number
word to which a brick could be coordinated,
it should be quite clear to the observer why
the mason might now announce that there
are eighteen bricks in the counted collection.
If the observer has been attentive and neither
missed, nor found fault with, any of the steps
in the mason’s procedure, she herself has
indeed come to the same result. Neverthe-
less, if one asked either of them how they
65
 cognitive–psychological
CONCEPTS
know that there are eighteen bricks, the
chances are that they would both answer:
“Well, I just counted them!” It is unlikely
that either of them would explain in suffi-
cient detail the procedure that was carried
out and why the last number word used
could be taken to indicate the numerosity of
the collection.
We have all learned to count in early
childhood, and we take for granted that the
last number word of a count tells us how
many items there were involved. We are,
indeed, so accustomed to this, that we do not
consciously think, whenever we hear or read
a number word, that it entails a count of as
many discrete units as there are number
words in the conventional sequence leading
up to it from “one.” Yet, if we did not know
this in some way, the number word could
have no meaning for us. The fact is that num-
ber words have become symbols for us, and
as such they symbolize the counting proce-
dure that leads up to them, without our hav-
ing to carry out that procedure or even having
to think of it.8 Even in the case of numbers
that are higher than we could ever actually
reach by counting, the tacit knowledge that
there is a procedure by means of which, the-
oretically, we could reach them, constitutes
the first (but by no means the only) charac-
teristic of number as an abstract concept.
In short, I submit that the three elemen-
tary concepts of arithmetic – unit, set, and
number – are abstractions, not from physi-
cal objects or other sensory material, but
from mental operations that thinking sub-
jects must carry out themselves. At the
beginning, ontogenetically speaking, these
operations develop as corollaries of actions
which, in order to be performed, require
sensory–motor material. This material need
not be the same for all thinking subjects, it
merely provides the occasion. However, once
patterns of mental operations have been
abstracted, they become mathematical con-
cepts through association with symbols that
can “point”9 to them without invoking their
actual execution.
The concept of unit is abstracted from the
perceptual operation of combining various
sense impressions to form a “thing”; the con-
cept of set is derived by abstracting the plural-
ity of abstract units from a collection of things
(i.e. considering an experientially bounded
plurality but not the sensory items that were
66
used to generate it); the concept of number
arises when number words or numerals have
become symbols that tacitly point to a possi-
ble count that leads up to them.
I want to emphasize that this analysis
concerns the basic inception of the concepts,
and that in the vast domain of mathematics
each of them can be indefinitely enriched by
the addition of further abstractions. Besides,
there is the whole area of geometry to which
I now want to turn.
From action to
abstraction
The understanding is a wholly active power of the
human being; all its ideas and concepts are but
its creation,… External things are only occasions
that cause the working of the understanding …
the product of its action are ideas and concepts.
Kant (1787/1902, Vol.VII, p. 71)
In the beginning, geometry is usually pre-
sented as a matter of points, lines, and
planes. In ordinary language, we have no dif-
ficulty in finding experiential objects to
which we can apply these words. Most of
these objects, however, do not satisfy the
mathematician’s requirements. Hence there
are mathematical definitions, or rather,
expressions that purport to serve that pur-
pose. But even mathematicians themselves
are not always pleased with them and there-
fore add more technical formulations, based
not on experience but on a specific mathe-
matical frame of reference, such as a system
of orthogonal coordinates. Thus my Mathe-
matical Dictionary (James & James 1959,
p. 274) has the following entry for “point”:
(1) An element of geometry which has
position but no extension.
(2) An element of geometry defined by its
coordinates, such as the point (1,3).
Although it is a long time ago, I can still
remember our teacher, before we had heard
anything about coordinates, making a small
mark on the blackboard and saying, as he
turned to us: “This is a point.” Then he hesi-
tated for a moment, looked back at the mark,
and added: “Of course, a geometrical point
has no extension.” This left us wondering
about grains of sand, specks of dust, and other
smallest items, but we remained perplexed
because all of them still had some extension.
radical constructivism
After a few days the perplexity was forgotten.
We had learned to make points with our pen-
cils, and all that mattered was that they were
not noticeably wider than the lines we drew.10
The obstacle here is that, logically, it is impos-
sible to move smoothly from small and
smaller to “no extension,” and yet hold on to
a something that could be discriminated; and
it is equally impossible to move from few and
fewer to “none,” and yet hold on to the notion
of plurality. Hence, what is needed is another
approach. When we first meet “zero,” we can
see it as the number word to use when all
countable items have been taken away. And
there is an analogous approach to “point” in a
visual experience that most will have had in
one way or another. Imagine, for instance,
sailing away, on a perfectly smooth lake, from
a small floating object, say, a bottle. It gets
smaller and smaller, and suddenly you cannot
see it any longer, though you are still looking
at the point where it was. The bottle is gone,
and it would seem more adequate to identify
the point with the focus of your attention.
This gets rid of the problem of size, because it
is never the focus of attention that has a size,
but only the things one is focusing on.11
Hence I propose to think of “point” as the
very center of the area in the focus of atten-
tion. In the visual field, then, it would be the
center of an item we are focusing on. If that
item is so small that we cannot distinguish a
center from the circumference, we have an
item that can represent our concept of point,
but it is not itself a point, because we can still
imagine that it has a center, even if we cannot
see it. But then this center turns into the van-
ishing point, a conceptual construct that
derives from movement and attention.
Another way to approach the concept of
point was suggested by Ceccato in conversa-
tions we had around 1950: Think of a form of
cheese, he said, and the way one cuts it by pull-
ing a wire through it. (This, of course, was in
Italy, where large forms of cheese are always
cut in this way.). The first cut, say a vertical
one, gives you a plane. If you cut again verti-
cally, intersecting the first plane, the two cuts
give you a vertical line. And if now you cut
horizontally, the intersection of the three cuts
gives you a point. What you have to focus on,
of course, is not the wire, nor the space it
leaves, but the movements, because in move-
ments we feel direction but no lateral exten-
sion.12
Constructivist Foundations
 cognitive–psychological
CONCEPTS
To my mind, both these approaches are
more adequate than merely saying that a
point has no extension. They come closer to
describing what one can do to arrive at the
concept that has no sensory instantiation.
In the case of “line,” there is a reputable
precedent. In his Critique of pure reason, Kant
(1787/1902, p. 138) says, in order to experi-
ence a line, one must draw it. But he said it in
German, and translation, as so often,
obscures the original meaning. The English
verb “to draw” is used indiscriminately for
producing images with a pencil and for what
horses do to a cart. In German the two activi-
ties require different verbs. Kant used ziehen,
which means “to drag” or “to pull” and does
not refer to a graphic activity except in the
case of lines. Since he spaced the word (to
emphasize it), he had in mind the physical
motion.13
In the same vein, Brunschvicg (1912/1981,
p. 503) says of the straight line: “The elemen-
tary operation that is to furnish the simplest
image is the stroke (trait). The hand places
itself somewhere, it stops somewhere; from
the point of departure to the point of arrival,
the mind has not become aware of a division
or a change in the movement accomplished
by the hand. Hence there is no reason to sus-
pect that the path linking the two points
might not be a uniform and unique line, a
straight segment that could serve to measure
the distance. In fact, we know by what round-
about way geometry was led to question the
evidence that seems to support the unique-
ness of the straight line between two points;
and we understand that it was the ease and
certainty (of the act) that enabled the mind to
cling to what is given by intuition” (my trans-
lation).
Intuition, here, I suggest is precisely what
Kant meant when he said Anschauung, i.e.,
the view of an experience upon which we are
reflecting. The line, then, is a reflective
abstraction from a uniform movement we
make. To this I add, that this movement
need not be that of a hand or other visible
object, but it can be the movement of our
attention in whatever field we happen to be
considering. And the straightness of an
object can, in practice, be checked by shift-
ing the focus of attention in uniform
motion, as cabinet makers do when they
look along the edge of a board to check that
it is not warped.
2006, vol. 1, no. 2
A second dimension
Having found an experiential basis for points
and lines, we may follow the definition of
“plane” in James & James (1959, p. 273): “A
surface such that a straight line joining any
two of its points lies entirely in the surface.”
Implicit in this criterion is the requirement to
look in at least two directions, which is equiv-
alent to drawing more than one line. Hence
the axiom that a plane is defined by three
points. But each of these defining elements
itself involves only one direction and there-
fore provides no immediate experience of the
two-dimensionality of surfaces or planes.
It has also been suggested that a plane can
be constructed by moving a line sideways.
This is interesting, because in order to follow
this movement, the focus of attention has to
be widened to cover at least a certain stretch
of the line. A practical equivalent would be to
move one’s hand on a surface, and feeling no
change either in the tactile pressure of fingers
and palm or in the direction of movement. In
both cases there is an expansion of the focus
of attention in order to monitor more than
one point. And this expansion is the opposite
of the shrinking in the example of the bottle
on the lake.
Confrey (1990) proposed that the way an
object increases or decreases in size, as we
move towards or away from it, provides an
experiential basis for both exponential
change and geometric similarity. I fully accept
this idea and want to stress that, in the present
context, its most relevant aspect is the expan-
sion, respectively contraction, of the area cov-
ered by our attention. This movement pro-
vides experiential situations from which, in
the expanding direction, the two-dimension-
ality of the plane can be abstracted, whereas
the shrinking direction may lead to the
abstraction of a concept of point.
These brief suggestions are the merest
beginning of an analysis of the conceptual
foundations of geometry. Other experiential
situations can be found that may serve as raw
material for the abstraction of the traditional
“basic elements,” and none is unique in the
sense that it could not be replaced by another.
But as with the concepts of unit, plurality, and
number, I believe that the minute analysis of
elementary actions and operations that might
occasion their abstraction is a direction of
research that is well worth pursuing. If stu-
radical constructivism
dents, at the time when geometry is intro-
duced, were offered experiences of this kind,
they might come to understand that the lines
drawn on paper and the physical models of
bodies they are shown are merely occasions
for mental operations that have to be actively
carried out in order to abstract the basic con-
cepts of geometry.
A source of certainty
Mathematics, like theology and all free creations
of the Mind, obeys the inexorable laws of the
imaginary. – Gian-Carlo Rota (1980, p. XVIII)
In the preceding sections I have argued that
common non-mathematical activities, such
as isolating objects in the visual or tactual
field, coordinating operations while they are
being carried out, and generating a line by a
continuous uniform movement, are the expe-
riential raw material that provides the think-
ing subject with opportunities to abstract ele-
mentary mathematical concepts. If one
accepts this view, one is faced with the puz-
zling question how such obviously fallible
actions can lead to the certainty that mathe-
matical reasoning seems to afford.
The puzzle is not unlike the one that arises
if we write the traditional textbook syllogism
with a first premise that we assume to be
false – for instance, “All men are immortal.” If
we proceed with “Socrates is a man,” the con-
clusion that Socrates is immortal will be just
as certain and logically “true” as the opposite
conclusion, which we get when we start with
the more plausible first premise that asserts
the mortality of all humans.
This puzzle disappears if it is made clear
that the premises of a syllogism must be con-
sidered as hypotheses and should be preceded
by “if.” Their factual relation to the experien-
tial world is irrelevant for the formal func-
tioning of logic. Considering them to be “as
though” propositions, makes sure that, for the
time being and during the subsequent steps of
the procedure, one is not going to question
them. The steps of that procedure are, on the
one hand, the specific mental operations des-
ignated by terms such as “all,” “some,” “no,” “is
a,” “then,” etc., and, on the other hand, the
operation of combining the two premises.
Assuming that these operations are carried
out in the customary way, the certainty of the
67
 cognitive–psychological
CONCEPTS
conclusion springs from the fact that the situ-
ations specified by the premises are posited
and, therefore, not to be questioned during
the course of the procedure.14
Viewed in this way, the syllogism also
becomes immune to a criticism which, I
believe, was first brought up by John Stuart
Mill (1843, A system of logic, III.2). Because
he was mainly concerned with the logical
uncertainty of inductive generalization, he
remarked that, in order to be justified in say-
ing that all men are mortal, one should have
examined all members of the class called
“man.” If Socrates is rightly considered a
member of this class, one must have come
across him during the examination and one
does not need the conclusion, to know that he
is mortal. If, however, one has not examined
Socrates, then either the second premise or
the conclusion is false. – When an “if ” is
placed before the premises, this voids the
above argument because it obviates the exam-
ination of all members of the class involved;
and what the conclusion affirms is made
unquestionable, because this now derives
from what was hypothetically posited and is
therefore not dependent on an examination
of actual experiences.
The hypothetical trick
Yet there remains a question. How do we
come to feel certain that the conclusion tells
us something that was contained in the pre-
mises? This goes to the core of the deductive
procedure and is crucial for my thesis. If the
major premise is of the form “all X are B,” it
implies that there is a bounded collection of Xs,
either experiential or hypothesized. But col-
lections are the result of mental operations,
and to form a collection, we clearly must first
have formed a plurality. In turn, to form a
plurality we need to conceive of discrete uni-
tary items that have some attribute, let us say
A, in common. If we now consider a discrete
unitary item and find that it has the attribute
A, we may say: “Ah, here is another X” (and we
may proceed to examine whether it, too, is B,
and thus fits our hypothesis). But note that, to
say “another,” we must remember that at some
earlier moment we attributed A to some
item(s). Similarly, to conclude in the syllo-
gism that Socrates is mortal, we must remem-
ber that we previously formed a collection
68
called “man” and attributed mortality to it,
because this is the basis on which we now feel
certain that, if Socrates can be considered a
member of that collection, he must be mortal.
I believe it was Euler, who first used circles
to indicate the bounded collections involved
in propositions such as “Some A are B” and
“all A are B.” (Euler 1770, letters 102–108, 14
February to 7 March, 1761). And he went on
to show how the syllogistic procedure could
be visualized in this manner. The mystery of
logic, purported to be so difficult to approach,
he says, immediately strikes the eye if one uses
these figures (Euler 1770, letter 103). And he
is right. Two concentric circles do convey the
notion of containment with great force – but
at the moment of perceiving this symbolic
containment, one still has to remember that
the outer circle is intended to stand for the
major premise and the inner circle for the
minor. The circles help to make palpable the
relation but not what is being related. Hence,
no matter how we look at it, the judgment of
certainty involves faith in the flawless func-
tioning of the particular kind of memory we
use in carrying out the syllogistic procedure.
The “certainty” in arithmetic is analogous
in that it, too, depends on mental operations
and not on a fit with the experiential world. If
someone tells me that there are seven oranges
in the kitchen, three on the table and four on
the windowsill, I do not have to accept this as
an unquestionable “truth” – even if I believe
that he or she is using the number words the
way I myself would use them. There might
have been an error, either in recognizing
oranges or in counting. But I cannot question
the simple statement that 3 + 4 = 7.
Unpacked, this statement means: You count a
collection and come to “3,” then you count
another collection and come to “4”; now you
can consider the two collections as one collec-
tion, and if you count it, you will come to “7.”
Provided one’s procedure follows the stan-
dard number word sequence and coordinates
its terms to countable items in the standard
fashion, one is going to arrive at the standard
result every time. Because the operations
involved have become habitual and we are not
aware of carrying them out, we get the
impression that there is something preor-
dained about their results, something that
could not be otherwise. As Spencer Brown
said about the existence of the universe: “It
comes through a very clever trick. It depends
radical constructivism
on an elaborate procedure for forgetting just
what it was we did to make it how we find it”
(Keys 1972, p. 31).
Hence, there is the involvement of mem-
ory. In order to know (when you have finished
counting the composite collection and came
to the result “seven”) that 7 is the sum of a col-
lection of 3 and a collection of 4, you must
remember that you counted these two collec-
tions before you combined them. As in the
case of the syllogism, a graphic representation
of the procedure, say “prearithmetical signs”
like those Lorenzen suggested for counting,
may help to visualize the procedure – but in
order to tie the graphic signs to the process of
addition, you have to remember the meaning
that was attributed to them at the outset.
This leads to the conclusion that one can do
neither logic nor mathematics without doing
things which, themselves are not specifically
mathematical. Depending on the kind of
mathematical result aimed at, there will be
activities such as isolating discrete perceptual
or visualized items, moving a limb or the focus
of attention, attributing meanings to signs or
symbols, considering explicitly or implicitly
limited contexts, and remembering the con-
ceptual commitments that have been made
during these mental operations. The certainty
of the results, then, springs on the one hand
from the fact that one operates in a hypothet-
ical mode and therefore obliges oneself not to
question what one has hypothesized; and on
the other hand, on implicit faith in one’s mem-
ory of meanings attributed, of operations car-
ried out, and of the results they produced.15
Conclusion
If what I have outlined is a viable approach,
there are several things that will have to be
considered when the foundations of mathe-
matics are discussed. First, they cannot be dis-
cussed without the use of language. From the
radical constructivist point of view, there is no
neat distinction between private and public
language, because all meaning of signs and
symbols, including the linguistic kind, is built
up by individuals on the basis of their own
subjective experiences of isolating objects,
events, and the relations among them. No
doubt everyone’s meanings are modified and
adapted in the course of interaction with
other speakers, but the result of such adapta-
Constructivist Foundations
 cognitive–psychological
CONCEPTS
tion is compatibility, not identity. And the
compatibility achieved is relative to the par-
ticular interactions an individual has partici-
pated in. The expression “shared meaning” is
a deceptive fiction because sameness can
never be ascertained.
Second, some (and perhaps all) of the
indispensable elements in mathematical
thinking are conceptual constructs that were
abstracted from operations carried out with
sensory material, operations that are involved
in segmenting and ordering experience long
before we enter into the realm of mathemat-
ics.16
It seems to me that these are two good rea-
sons for considering mathematics to be an
“empirical” enterprise. Lakatos and others
(e.g., Davis & Hersh 1981, Tymoczko 1986a)
have called it “quasi-empirical” because it
does not directly deal with physical objects. To
a radical constructivist, however, physical
objects, too, are conceptual constructs
abstracted from a way of experiencing that
imposes structure on an essentially amor-
phous sensory manifold. Mathematics deals
with constructs that no longer contain sen-
sory or motor material because they are
abstracted from mental operations carried
out with that material; but this does not make
them any less experiential – and “empirical,”
after all, is but another word for “experien-
tial.” The poet/mathematician Paul Valéry has
said this with uncommon elegance in the epi-
graph I placed at the beginning.
Postscript
Penelope Maddy (1997, p. 233) proposed an
approach “that turns away from metaphysics
and towards mathematics.” What concerns
her is not the justification or ontological real-
ity of the relevant concepts but how mathe-
maticians have formed them. The formation
she describes takes place on a level of abstrac-
tion beyond the very first one that uses sen-
sory experiences as raw material (which is the
topic of my paper). Set theory has been so suc-
cessful, she claims, because it has created its
own conceptual ontology that does not
require the Platonic ontology philosophers
are mainly unwilling to relinquish. But as
Quine (1969, p. 83) put it: “The old episte-
2006, vol. 1, no. 2
mology aspired to contain, in a sense, natural
science; it would construct it somehow from
sense data. Epistemology on its new setting,
conversely, is contained in natural science, as
a chapter of psychology.” As I understand it,
Maddy’s suggestion is that foundations of
mathematics have to be found within set the-
ory.
Encouraged by the popular success of Met-
aphors we live by (Lakoff & Johnson 1980),
George Lakoff (Lakoff & Núñez 2000)
embarked on the exploration of the meta-
phors that, he believes, form the source of
mathematics. The meaning of mathematical
terms, he claims, consists in conceptual met-
aphors. Each conceptual metaphor is “a uni-
directional mapping from entities in one con-
ceptual domain to corresponding entities in
another conceptual domain” (Lakoff &
Núñez 2000, p. 42). From my point of view,
this is a highly misleading statement. In my
world, metaphors can be defined as the
attempt to transfer a property assumed to be
characteristic of one type of experiential item
to an item that is usually considered not to
have that property. (E.g. if I say: “My brother
in law is a gorilla,” it is up to you to discover
from the context which of a gorilla’s proper-
ties I am attributing to my brother in law. But
this is not the only kind of metaphor Lakoff
has in mind. He also wants to call metaphor
any relational pattern that has been
abstracted from an experiential situation.
This happens to be exactly what Peirce called
an “abstractive observation”; and Piaget some
thirty years later called it “reflective abstrac-
tion,” showing how patterns of physical
action could lead to patterns of mental oper-
ating. Piaget is mentioned in a marginal con-
text only, Peirce not at all, and, what is even
more astonishing, there is no reference to
Brouwer, who explicitly linked the origin of
number to an experiential situation (cf. Sec-
tion “To begin at the beginning” above).
Lakoff & Núñez’s Where mathematics comes
from (2000) suggests a great number of appli-
cations of his theory of metaphor to mathe-
matical concepts. The usefulness of this move
will have to be judged by mathematicians.
However, Lakoff says practically nothing
about the elementary concepts which I deal
with in my article. He takes discrete countable
things and pluralities for granted and his
radical constructivism
account of counting makes no mention of a
conventional number-word sequence, with-
out which, I would claim, no concept of num-
ber can be formed.
Brian Rotman (2000, p. 38), judging by
what he wrote in Mathematics as sign, would
agree with what I wrote in Section 5, above.
“Numbers,” he says, “appear as soon as there
is a subject who counts.” He justifies this in the
preceding paragraph: “However possible it is
for them (i.e. whole numbers) to be individu-
ally instantiated, exemplified, ostensibly indi-
cated in particular, physically present, plural-
ities such as piles of stones, collection of
marks, fingers, and so on, numbers do not
arise, nor can they be characterized, as single
entities in isolation from one another: they
form an ordered sequence, a progression. It
seems impossible to imagine what it means
for “things” to be the elements of this progres-
sion except in terms of their production
through the process of counting.” As a general
maxim he states that “A mathematical asser-
tion is a production, a foretelling of the result
of performing certain actions upon signs …
Thus, for example, the assertion ‘2 + 3 = 3 +
2’ predicts that if the Subject concatenates 1 1
with 1 1 1, the result will be identical to his
concatenating 1 1 1 with 1 1” (Rotman 2000,
p. 16). Following Deleuze and Guattari
(1994), Rotman distinguishes ordinals and
cardinals in a way that fits my thinking and
seems enlightening to me: ordinals are rhyth-
mic, directional in terms of serial continua-
tion, and hence akin to melody; cardinals, in
contrast, seem a parallel presentation, a har-
mony. I see this as matching the notion that
ordinals are formed by focusing on the repe-
tition of counting acts, whereas cardinals arise
from reflecting a counted plurality as a unit
(cf. Rotman 2000, p,146).
Much more, I am sure, has been written
during the last fifteen years that would merit
a comment; but my article was never intended
as a review of the field.
Acknowledgment
I am indebted to David Isles, Cliff Konold,
Jack Lochhead, John Richards, Les Steffe, and
an anonymous reviewer for helpful com-
ments on a draft of this paper.
69
 cognitive–psychological
CONCEPTS
Notes
1. In Newman (1956), p. 638.
2. E.g., “Mathematical objects are invented
or created by humans” (Hersh 1986, pp.
22f) and particularly his reference to Piag-
et: “…one cannot overestimate the impor-
tance of his central in-sight: that
mathematical intuitions are not absorbed
from nature by passive observation, but
rather are created by the experience of ac-
tive manipulation of objects and sym-
bols.” (Hersh 1986, p. 26)
3. I have separated and numbered the three
parts, although they form a consecutive
passage in Whitehead’s essay.
4. An interdisciplinary research group
founded by Silvio Ceccato in the 1940s,
which published the journal Methodos and
was later incorporated as the Center for
Cybernetics of Milan University.
5. Cf. Piaget’s 1937 book “La construction
du réel chez l’enfant” and half a dozen oth-
er volumes whose titles refer to these con-
cepts.
6. The explanations of the term “set” one
finds in textbooks (e.g. that sets are some-
what like a jury or the signs of the zodiac)
do not diminish the obscurity, because
they omit the crucial fact that such collec-
tions are formed on the basis of an extrin-
sic consideration that is quite
unmathematical and that the concept of
“set” requires one to take the units of the
collection and deprive them of whatever
attributes they might have beyond being
considered units.
7. Cf. Piaget’s two 1977 volumes Recherches
sur l’abstraction réfléchissante; also Beth &
Piaget (1961).
8. A full description of this pointing function
can be found in Glasersfeld (1991).
9. I am avoiding the word “refer” because of
its usual connotation of reference to real-
world objects.
10.I might add that perplexities of this kind
arose also at other times during our math-
ematics instruction, especially when we
came to differential calculus and integra-
tion. As the perplexities mounted, more
members of the class concluded that
mathematics is incomprehensible and not
much fun to pursue. This was a pity and
70
could have been avoided, if only it had
been made clear from the beginning that
geometry is conceptual and that the world
we consider external and physical does not
provide geometrical entities but perceptu-
al situations from which we may abstract
them.
11.The same, incidentally, goes for “loca-
tion.” The visual focus of attention has no
location, except relative to things one has
isolated in the visual field. But this would
lead to a consideration of the concept of
“space,” which lies beyond the scope of
this discussion.
12.Ceccato later discarded this explanation
because it did not reduce the concept’s
structure to moments of attention (Cecca-
to 1966, p. 500). However, as an experien-
tial scenario that might lead to the
abstraction of the concept it is still a good
example.
13.To an English reader, “to draw a line” is
most likely to suggest the appearance of a
mark on paper or on some other surface.
But Kant’s emphasis shows that he want-
ed to draw attention to the action of the
hand.
14.I was delighted to discover, on rereading
Beth’s introductory essay in Beth & Piag-
et (1961), that this idea was contained in
a passage he quoted from a French phi-
losopher of the 1920s, which I do not re-
member having read before:
“To deduce is to construct. One demon-
strates only hypothetical judgments; one
demonstrates that one thing is the conse-
quence of another. For that purpose one
uses the hypothesis to construct the con-
sequence. The conclusion is necessary,
… They (the premises) are propositions
that have been admitted beforehand, ei-
ther by virtue of preceding demonstra-
tions or as definitions or postulates.”
(Goblot 1922, as quoted in Beth & Piaget
1961, p. 21)
If not explicitly hypothetical, the propo-
sitions used as premises may be inductive
inferences or results of prior deductions
from inductive inferences; and since in-
ductive inferences cannot be considered
logically certain, they are in this context
equivalent to hypotheses. Also, it should
be stressed that certainty, in this context,
radical constructivism
concerns the derivation of the conclu-
sion from the premises and not the tradi-
tional notion of truth relative to a “real”
world.
15.It seems to me that this faith in human
memory is not essentially different from
the confidence we are expected to have in
the functioning of a computer that has
carried out a “proof” that would take a hu-
man longer than a life time (cf. Tymoczko
1979/1986c).
16.About the nonverbal mental images that
mathematicians make use of, Einstein
said: in his case, they were “visual and
some of muscular type.” (Cited in Had-
amard 1954, p. 143).
ABOUT THE AUTHOR
Ernst von Glasersfeld was born in Munich,
1917, of Austrian parents, and grew up in
Northern Italy and Switzerland. Briefly
studied mathematics in Zürich and Vienna
and survived the 2nd World War as farmer
in Ireland. Returned to Italy in 1946, worked
as journalist, and collaborated until 1961 in
Ceccato’s Scuola Operativa Italiana (lan-
guage analysis and machine translation).
From 1962 director of US-sponsored
research project in computational linguis-
tics. From 1970, he taught cognitive psy-
chology at the University of Georgia, USA.
Professor Emeritus, 1987. At present
Research Associate at Scientific Reasoning
Research Institute, University of Massachu-
setts. Dr.phil.h.c., University of Klagenfurt,
1997. – Books (among others): Wissen,
Sprache und Wirklichkeit, Vieweg Verlag,
Wiesbaden, 1987. Linguaggio e comunicazi-
one nel costruttivismo radicale, CLUP, Mil-
ano, 1989. Radical Constructivism: A way of
knowing and learning, Falmer Press, Lon-
don, 1995 (German translation: Suhrkamp
Verlag, Frankfurt am Main, 1996; also Portu-
guese, Korean, Italian translations.) Grenzen
des Begreifens, Benteli Verlag, Bern, 1996.
Wege des Wissens, Carl Auer Verlag,
Heidelberg, 1997. Wie wir uns erfinden
(with H. von Foerster), Heidelberg: Carl
Auer, 1999. (Italian translation: Rome,
Odradek, 2001). More than 260 paper
publications since 1960.
Constructivist Foundations
 cognitive–psychological
CONCEPTS
References
Beth, E. W. & Piaget, J. (1961) Épistémologie
mathématique et psychologie. Presses
Universitaires de France: Paris.
Bridgman, P. B. (1961) The way things are.
Compass: New York.
Brouwer, L. E. J. (1949) Consciousness, phi-
losophy, and mathematics. Proceedings
of the 10th International Congress of Phi-
losophy, Volume 1, Part 2. North Hol-
land: Amsterdam, pp. 1235–1249.
Brunschvicg, L. (1981) Les étapes de la phi-
losophie mathématique, 3rd edition.
Blanchard: Paris. Originally published in
1912.
Ceccato, S. (1966) Un tecnico fra i filosofi,
Volume 2. Marsilio: Padua.
Confrey, J. (1990) Splitting, similarity, and
rate of change: A new approach to multi-
plication and exponential functions.
Paper presented at the Annual Meeting of
the American Educational Research
Association. Boston MA.
Davis, P. J. & Hersh, R. (1980) The mathe-
matical experience. Birkhäuser: Boston.
Deleuze, G. & Guattari, F. (1994) What is
philosophy. Columbia Univ. Press: New
York.
Einstein, A. (1954) Physics and reality. In:
Einstein, A., Ideas and opinions. Bonanza
Books: New York, pp. 290–323. Origi-
nally published in 1936 in: Journal of the
Franklin Institute 221(3): 348–382.
Euler, L. (1770) Lettres a une princesse
d’Allemagne. Steidel: Mietau/Leipzig.
Frege, G. (1884) Grundlagen der Arithme-
tik. English translation: Frege, G. (1974)
Foundations of arithmetic (translated by
J. L. Austin). Blackwell: Oxford.
Glasersfeld, E. von (1981) An Attentional
model for the conceptual construction of
units and number. Journal for Research
in Mathematics Education 12: 83–94.
Glasersfeld, E. von (1989) Cognition, con-
struction of knowledge, and teaching.
Synthese 80: 121–140.
Glasersfeld, E. von (1991) Abstraction, re–
presentation, and reflection: An interpre-
tation of experience and Piaget’s
approach. In: Steffe, L. P. (ed.) Epistemo-
logical foundations of mathematical
experience. Springer: New York, pp. 45–
67.
2006, vol. 1, no. 2
Goblot, E. (1922) Le système des sciences: Le
vrai, l'intelligible et le réel. A. Colin: Paris.
Hadamard, J. (1954) The psychology of
invention in the mathematical field.
Dover: New York.
Hersh, R. (1986) Some Proposals for Reviv-
ing the Philosophy of Mathematics. In:
Tymoczko, T. (ed.) New directions in the
philosophy of mathematics. Birkhäuser:
Boston, Basel, pp. 9–28. Originally pub-
lished in 1979 in: Advances in Mathemat-
ics 31: 31–50.
Husserl, E. (1970) Philosophie der Arithme-
tik. Martin Nijhoff: The Hague. Origi-
nally published in 1887.
James, G. & James, R. (1959) Mathematics
dictionary. Van Nostrand: Princeton NJ.
Kant, I. (1902) Kritik der reinen Vernunft.
Akademieausgabe, Volume 3. Preussische
Akademie: Berlin. Original publication:
1787.
Keys, J. (alias G. Spencer Brown) (1972)
Only two can play this game. Julian Press:
New York.
Lakatos, I. (1976) Proofs and refutations:
The logic of mathematical discovery.
Cambridge University Press: Cambridge.
Lakatos, I. (1986) A renaissance of empiri-
cism in the recent philosophy of mathe-
matics? In: Tymoczko, T. (ed.) New
directions in the philosophy of mathe-
matics. Birkhäuser: Boston, Basel, pp.
29–49. Originally published in 1976 in:
British Journal for the Philosophy of Sci-
ence 27(3): 201–223.
Lakoff, G. & Johnson, M. (1980) Metaphors
we live by. Chicago University Press: Chi-
cago.
Lakoff. G. & Núñez, R. E. (2000) Where
mathematics comes from: How the
embodied mind brings mathematics into
being. Basic Books: New York.
Lorenzen, P. (1974) Konstruktive Wissen-
schaftstheorie. Suhrkamp: Frankfurt am
Main.
Mach, E. (1970) The guiding principles of
my scientific theory of knowledge and its
reception by my contemporaries. In: S.
Toulmin (ed.) Physical reality. Harper:
New York, pp. 44–53. German original
published in 1910.
Maddy, P. (1997) Naturalism in mathemat-
ics. Clarendon Press: Oxford.
Maturana, H. R. (1980) Biology of cogni-
radical constructivism
tion. In: Maturana, H. R. & Varela, F. J.,
Autopoiesis and cognition: The realiza-
tion of the living. Boston Studies in the
Philosophy of Science, Volume 42. Reidel:
Dordrecht, pp. 5–58. Originally pub-
lished in 1970 as BCL Report 9.0. Biolog-
ical Computer Laboratory. Department
of Electrical Engineering, University of
Illinois.
Mill, J. S. (1843) A system of logic, ratiocina-
tive and inductive, being a connected
view of the principles of evidence and the
method of scientific investigation. Lon-
don.
Mittelstrass, J. (1987) Philosophische
Grundlagen der Wissenschaft. In: Paster-
nack, G. (ed.) Philosophie und Wissen-
schaften: Das Problem des Apriorismus.
Peter Lang: Frankfurt/Bern.
Newman, J. R. (1956) The world of mathe-
matics (4 volumes). Tempus Books: New
York.
Piaget, J. (1937) La Construction du réel
chez l’enfant. Delachaux et Niestlé: Neu-
châtel.
Piaget, J. (1972) Psychology and epistemol-
ogy, Viking: New York. French original
published in 1970 by Édition Denoël:
Paris.
Piaget, J. (1977) Recherches sur l’abstraction
réfléchissante, Volume I & II. Presses Uni-
versitaires de France: Paris.
Quine, W. V. (1969) Epistemology natural-
ized. In: Quine, W. V., Ontological rela-
tivity and other essays. Columbia
University Press: New York, pp. 69–90.
Rota, G.-C. (1980) Introduction. In: Davis,
P. J. & Hersh, R. (1980) The mathematical
experience. Birkhäuser: Boston.
Rotman, B. (1987) Signifying nothing: The
semiotics of zero. Stanford University
Press: Stanford.
Rotman, B. (2000) Mathematics as sign:
Writing, imagining, counting. Stanford
University Press, Stanford.
Steffe, L. P., Glasersfeld, E. von, Richards, J.
& Cobb, P. (1983) Children’s counting
types: Philosophy, theory, and applica-
tion. Praeger: New York.
Tymoczko, T. (ed.) (1986a) New directions
in the philosophy of mathematics.
Birkhäuser: Boston, Basel.
Tymoczko, T. (1986b) Making room for
mathematicians in the philosophy of
71
 cognitive–psychological
CONCEPTS
mathematics. The Mathematical Intelli-
gencer 8(3): 44–50.
Tymoczko, T. (1986c) The four-color prob-
lem and its philosophical significance. In:
Tymoczko, T. (ed.) New directions in the
philosophy of mathematics. Birkhäuser:
Boston, Basel, pp. 243–266. Originally
published in 1979 in: The Journal of Phi-
losophy 76(2): 57–83.
72
Valéry, P. (1974) Cahiers, Volume II. Galli-
mard: Paris. Originally published in
1935.
Vico, G.-B. (1858) De Antiquissima Ital-
orum Sapientia, Stamperia de Classici
Latini. Naples. Originally published in
1710.
Whitehead, A. N. (1956) Mathematics as an
element in the history of thought. In:
radical constructivism
Newman, J. R. (ed.) The world of mathe-
matics, Volume 1. Simon & Schuster:
New York. Reprinted in 1988 by Tempus,
Microsoft Press.
Wittenberg, A. I. (1968) Vom Denken in
Begriffen. Birkhäuser: Basel/Stuttgart.
Received: 13 July 2005
Accepted: 3 October 2005
Constructivist Foundations
 engineering-computer scientific
SURVEY
Enactive Cognitive Science.
Part 2: Methods, Insights, and Potential
Kevin McGee A Linköping University, kevmc@ida.liu.se
Introduction
The first part of this paper (McGee 2005) con-
textualized and presented some major con-
cerns of enactive cognitive science; this part
discusses methods, insights, and potential.
The proposal for enactive cognitive sci-
ence as a research discipline occurred during
a transition period to what has since come to
be called second generation cognitive science
(Lakoff & Johnson 1999). In addition to ques-
tioning correspondence models of represen-
tation, this approach to cognitive science is
characterized by the use of certain methods,
techniques, mechanisms, and models. There-
fore, towards the end of this paper there will
also be a brief section summarizing second
generation cognitive science – and situating
the research within enactive cognitive science
relative to it.
Methodological issues
Enactive cognitive science is concerned with
the full range of phenomena typically treated
by conventional cognitive science. As noted
earlier, for the purposes of this paper, “cogni-
tive” includes phenomena ranging in scale
from an individual, biological cell to large pop-
ulations of multi-cellular organisms. Other
than this more expanded notion of cognition,
research within enactive cognitive science is
conducted on all the usual topics. Of course, its
perspective does treat such classical faculties as,
for example, memory, in particular ways. In the
case of memory, the concern is much less with
“mechanisms of storage” and rather with con-
stant viable action (Riegler 2005), which may
be one reason that different connectionist pro-
posals are appealing. Similarly, as already sug-
gested, the treatment of perception is radically
different (for different enactive perspectives on
perception, see for example Maturana 1980;
Skarda 1999; Noë 2004).
Constructivist Foundations 2006, vol. 1, no. 2
enactive cognitive science
Purpose: This, the second part of a two-part paper, describes how the concerns of enac-
tive cognitive science have been realized in actual research: methodological issues, pro-
posed explanatory mechanisms and models, some of the potential as both a theoretical
and applied science, and several of the major open research questions. Findings: Despite
some skepticism about “mechanisms” in constructivist literature, enactive cognitive sci-
ence attempts to develop cognitive formalisms and models. Such techniques as feedback
loops, self-organization, autocatalytic networks, and dynamical systems modeling are
used to develop alternatives to cognitivist models. A number of technical similarities are
starting to emerge in the different models being proposed. Research Implications:The
need to resolve the interplay between autonomy and coupling with the environment sug-
gests the need for further technical research.And the reintroduction of first-person con-
cerns into cognitive science raises some questions of method, particularly with regard to
the relationship between first-person experience, neuroscience, and methods of descrip-
tion, analsysis, and explanation. Results to date suggest that insights from enactive cogni-
tive science could lead to innovations in the design of artifacts. Keywords: embodiment,
mechanisms, dynamical systems modeling, puposive systems, autonomy, coupling, inter-
subjectivity, morphogenesis, neurophenomenology.
In some ways, the majority of research in laboratory to study “cognition in the wild”
enactive cognitive science is broadly conser- (Hutchins, 1995). This is motivated by a
vative in its methods. Researchers do observa- number of claims. For example, the claim
tional studies, make hypotheses and predic- that one sees cognitive phenomena outside
tions, construct experiments, and evaluate the laboratory that are not seen within labo-
the results. On the other hand, the enactive ratory settings, and that even those cognitive
orientation does in some cases motivate an phenomena one is able to study in laboratory
enlargement of methodological scope; the settings may be entirely misleading about
development of specific new methods, as in how cognition typically operates. Addition-
the case of Piaget and the development of the ally, the claim that certain kinds of individual
clinical interview method (see Piaget 1963, phenomena make no sense unless they are
ch. 1; or Ginsburg 1997); and finally, certain contextualized within the larger group activ-
changes with regard to the proposal and test- ity. Finally, a category of claims involves a
ing of possible cognitive models and mecha- rejection of the assumption that studying
nisms (some of which we will explore in more individuals can then be unproblematically
detail below). And, of course, the enactive ori- generalized to groups. An alternative, ethno-
entation does challenge a number of conven- graphically-inspired line of social research is
tional methodological assumptions, criteria, often associated with the claim that group
and the like. characteristics and dynamics are not simply
As part of the overall shift of second-gen- derivative of individual characteristics; or,
eration cognitive science, there have been more radically, that individual cognition is
some methodological changes from those derived (in some sense) from group-cogni-
that typified earlier cognitivist methods. As tion.
an example, recent work in cognitive science Another example of enlarging the meth-
sees a renewed interest in moving beyond the odological scope of the discipline is the recent
73
 engineering-computer scientific
SURVEY
work on methods for studying first-person
experience or consciousness (Varela & Shear
1999). In some ways this is a return to intro-
spection as a valid method for cognitive sci-
ence; in other ways, the recent work is propos-
ing methodological innovations and
constraints in an attempt to address some of
the criticisms addressed at the earlier use of
introspection in psychology.
Whereas studying “cognition in the wild”
can be seen as the re-introduction of a meth-
odology that was important in an earlier era,
some other methodological proposals are
more innovative. One such proposal is that
different classes of phenomena are not mutu-
ally reductive and therefore require different,
complementary, and incommensurate
descriptions and explanations (see, for exam-
ple, Varela 1980; Thompson & Varela 2001;
Lutz & Thompson 2003). Alicia Juarrero, for
example, notes that a number of recent cogni-
tive models are based on a claim that there is
mutual cycling and determination between
the local and the global – something also dis-
cussed in contemporary literary theory; thus,
she argues for including narrative techniques
for the analysis and description of cognitive
systems (see Juarrero 1999; and for more
work generally at the intersection of narrative
theory, hermeneutics, and psychology, see
Bruner & Kalmer 1998 and Messer et. al.
1994).
Models and mechanisms
There is often an uneasy tension between con-
structivist thought and proposals for cogni-
tive mechanisms. For example, constructivist-
oriented literature often contains statements
condemning the “machine” or “computer”
model of cognition. Although this is not
always made clear, these anti-machine and
anti-computational criticisms are often
(though, not always) directed at particular
classes of machines and particular models of
computation. For example, one conventional
assumption about scientific mechanisms
(inherited from classical physics) is that enti-
ties are made up of parts – and that the con-
stituted entity has no causal influence on the
constituting components. As noted earlier,
there is a great deal of work to develop models
that challenge this. Similarly, one standard
model of computation is that it involves an
unambiguous mapping from inputs to out-
puts. Such a model can be used to argue for
74
either environmental determinism (“the cog-
nitive machine’s behavior is determined by
the inputs it receives”) or genetic determin-
ism (“the structure of the cognitive machine
determines how it will behave in response to
inputs”).
But there are alternative, distinctive mod-
els or mechanisms (or classes of mechanisms)
– and each one is either based on observed
cognitive phenomena, has been used to guide
further study of such phenomena, or both.
The particular mechanisms may or may not
support varying research orientations,
hypotheses, observations, and the like; for our
purposes here, an important point is simply
that there is no in-principle rejection of
mechanisms as such by enactive cognitive sci-
ence. An early example of this is, of course,
Jean Piaget’s systematic efforts to to develop
constructivist formalisms (for detailed over-
view of Piaget’s formalisms, see Flavell 1963).
More recently, Maturana and Varela (1980)
were quite explicit that their autopoietic
model of life and cognition involves a partic-
ular kind of mechanism – one that can be real-
ized through chemical experiment (Luisi &
Varela 1989) or modeled as a computer simu-
lation (McMullin 2004). Nor is there even any
default rejection of particular mechanisms. In
their early work, Maturana and Varela argued
quite strongly against the idea that cognition
involved ideas such as “transmission” from
standard information theory. Later, Varela
proposed to extend the original autopoietic
formulation to include a “non-naive and use-
ful role of informational notions in the
description of living phenomena” (Varela
1980, p. 39) – and more recently, Alicia Juar-
rero (1999) has proposed a cognitive model
that combines concepts from dynamical sys-
tems theory, information theory, and self-
organization theory.
Note also that although there have been
strong tendencies to use certain techniques,
the use of particular techniques in and of
themselves do not distinguish enactive cogni-
tive science (nor first- or second-generation
cognitive science, for that matter). Further-
more, techniques and theoretical commit-
ments interact in subtle and non-trivial ways.
For example, identifying work in terms of its
commitment to recent concepts of embodi-
ment is far from straight-forward; there is
robotics research that is clearly committed to
“embodiment” in the sense that physical
enactive cognitive science
robots are involved, but which few contempo-
rary researchers would include in the class of
research aligned with second-generation
ideas about embodiment. The work of Lakoff
and colleagues on their NTL (Neural Theory
of Language) paradigm, on the other hand,
would presumably be an example of strong-
embodiment commitments, even though
most of that work to date is software-oriented
(for a detailed discussion of how fragmented
is the current discussion around “embodi-
ment”, see Ziemke 2004). Similarly, although
it is now common to use connectionist and
dynamical techniques, Gary Drescher’s
schema mechanism (Drescher 1991) was
implemented in Lisp and has aspects that
would qualify it in some ways as a typical
“symbolic” system (cf. Box 1).
Nonetheless, second generation cognitive
science is partly characterized by a renewed
interest in such earlier mechanisms as feed-
back loops, neural networks, self-organizing
systems, and autocatalytic networks. More
recent techniques and mechanisms include
those grouped under the term “natural com-
puting” (such as genetic algorithms, fuzzy
logic, and the like, see, e.g., Ballard 1999),
connectionist models (O’Reilly & Munakata
2000; Ellis & Humphreys 1999; Medler 1998);
cellular automata (Wolfram 1994), subsump-
tion architectures (Brooks 1999), deictic rep-
resentations (Agre & Chapman 1990), and
dynamical systems theory (Port & van Gelder
1995; Kauffman 1993). Taking inspiration
from chemistry and biology, there are also a
number of recent proposals based on varia-
tions of autocatalytic systems, including such
models as dissipative structures (Brooks &
Wiley 1988), chemical hypercyles (Ulanowicz
1997), and synergetics (Kelso 1995). More
narrowly, much of the recent research in enac-
tive cognitive science has converged on
dynamical systems modeling as the primary
explanatory tool. To the extent that connec-
tionist techniques are used, these tend to rely
on re-entrance (feedback) – and a conse-
quence of this is that dynamical systems mod-
eling is appropriate for such systems as well.
However, as we will see, various proposed
models make use of any number of different
techniques.
Before describing some common features
of enactive models, it is important to say a
word of caution about descriptions that read-
ers are likely to find as they approach the
Constructivist Foundations
 engineering-computer scientific enactive cognitive science
SURVEY

sure, the systems react to stimulation, but the

BOX 1: GARY DRESCHER’S SCHEMA MECHANISM
reactions are predictable only in broad out-
“Conventional learning systems define new concepts as boolean combinations, generalizations line. In the main, these systems enclose some
or specializations, or analogs or clusters of existing concepts. Any such variant of existing con- network of activity that largely autonomous
cepts resembles one or more prior concepts, differing only incrementally. Piagetian develop- from any exchanges of matter or energy.
ment, in contrast, requires the invention of concepts that differ fundamentally from all prior
concepts.” (Drescher 1991, p. 6) Coupling. Although the proposed models are
Gary Drescher’s schema mechanism is a remarkable system that deserves to be more widely autonomous, they are also tightly coupled
known and studied – and an interesting area of future research could involve synthesizing with their environments. How this is recon-
aspects of this system with others. His main contribution is a formal proposal for novel concept ciled with autonomy is one difference
formation, one in which an agent’s activity and embodiment play a crucial role. between the different proposals. For some,
The system is a software implementation consisting of a two-dimensional world occupied external phenomena only contribute a small
by an extremely limited agent with a “body” – and an object that can be moved (and that can percentage of overall influence to the ongoing
also move autonomously).The agent starts with a very limited and uncoordinated set of sensory (statistically-based) activation. But in every
and motor abilities, hand and eye movements are initially random, and the agent has no “mem- model, there are two distinctive characteris-
ory” of the object when it isn’t perceived – indeed, the agent has no concept of “object” (as tics of coupling. First, the exchanges and
something that is, for example, “the same thing” being seen and touched at the same time). interactions with an environment are
Nonetheless, as a result of the agent’s actions and the resulting constructive activities of the “minor” relative to the ongoing internal activ-
various mechanisms, the agent eventually attains some limited version of the concept of object ity. Second, the system and its environment
permanence – namely, that the object continues to exist (and can be recovered) even when it are co-specified; that is, what counts as an
is not perceived. More remarkable still, the system’s representation of the new concept is not environment only makes sense by reference to
simply a “boolean combination, generalization or specialization, or analog or cluster of existing the system, and vice versa.
concepts.”
It is extremely difficult to give an adequate treatment of this complex system and its other Feedback. All the models assume that the
significant achievements here; for a detailed description, see Drescher (1991). result of some system activity contributes to
the context within which subsequent activity
takes place. Unlike earlier cybernetics with its
related literature. It is common to find refer- ally be externally controlled. But although emphasis on negative feedback and stability,
ences to self-organization which may involve more recent models of cognitive self-control many of the current systems make use of pos-
mechanisms for self-assembly, self-mainte- are not entirely isolated from their environ- itive feedback. In Freeman and Skarda’s pro-
nance, or self-determination. These distinc- ments, environmental influence at most only posal, for example, the system is constantly
tions are particularly important because there modulates the overall dynamics of the system stabilizing and destabilizing itself. In their
is a tendency in some of the literature to place – and even then, there tends not to be a strict words, “brains makes chaos to make sense of
undue (or unclear) emphasis on the presence correlation between environmental phenom- the world” (Skarda & Freeman 1987). Also,
(or not) of a “Designer” in the creation of a ena and the system’s behavior. Such systems unlike earlier feedback systems, the function
system – as if a key issue is whether the systems are, in various ways, both autonomous of feedback now is somewhat different. It can
are artificial (“designed”) or natural (“spon- (within limits) and coupled with their envi- be that the feedback systems are more “inter-
taneously occurring”). This leaves the unfor- ronments. nal” (re-entrant modules in Edelman’s
tunate impression that certain systems cannot model). Or it could be that the feedback is
be the result of human engineering. However, Enactive modelling part of a chaotic model, where the particular
for other researchers, the key distinction is Although there are a wide variety of enactive inputs to the feedback mechanisms are not as
not along those lines, but rather whether or proposals for modeling a wide variety of cog- relevant as the overall dynamical pattern in
not forces external to the system directly nitive phenomena (see, e.g., Maturana & phase-space.
determine, modify, or control the behavior or Varela 1987; Freeman 1999; Juarrero 1999;
development of such systems. By this view, a Edelman 1992), it is possible to identify a Temporality. Related to the previous point,
designer can create an autonomous system. number of characteristics they have in com- with the introduction of dynamical systems
One reason for this confusion is that cyber- mon. theory there has been a shift in perspective.
netics proposed technical models of “purpo- Although such systems make use of feedback,
sive” and “self-steering” systems (Rosen- Autonomy. Each proposal describes how the the center of focus is not on how some feed-
blueth, Wiener & Bigelow 1943). However, system is distinct from its environment, how back mechanism can keep the system regu-
such systems were only purposive or self- it maintains its identity, and how it deter- lated (relative to some target-value). Rather, it
maintaining to the extent that they were able mines which external phenomena “count” is the pattern that the system traces through a
to use one or more external values to “guide (and how they count). In particular, it is not “space” of parameters. In particular, the most
themselves” in the face of other external dis- possible to directly control the behavior of interesting patterns are chaotic; broadly regu-
turbances. Such cybernetic systems can actu- these systems by external intervention. To be lar and predictable, but irregular and uncer-

2006, vol. 1, no. 2 75

 engineering-computer scientific
SURVEY
tain in detail. This actually introduces inter-
esting notions of temporality into such
systems; researchers often speak of such sys-
tems as “maintaining their histories.”
Downward Causality. Again, there are a num-
ber of variations on this theme, but it basically
involves ways in which the macro-level con-
strains or influences the micro-level.
The details of the different models are
more complex, of course, and also differ from
each other in various ways. But, even in the
case of more theoretical research, where no
particular models are being proposed, many
of these same concepts are invoked. Thus, for
example, Richard Lewontin has long argued
for a model of “The organism as the subject
and object of evolution” (1983). More
recently, similar views have motivated the
development of a research agenda around
developmental systems theory (Oyama, Grif-
fiths & Gray 2001).
Research insights
We now turn to a summary of some of the
research in cognitive science that has contrib-
uted to further insights about the major
themes identified in the first part of this paper.
In what follows there will be a particular
emphasis on recent work – and touching only
on a few highlights. In particular, the remain-
der of the section will look at some interesting
results from research on embodiment (and
perception), intersubjectivity, and conscious-
ness. We will not highlight the theme of
autonomy, for example, because while there
have several recent proposals attempting to
sketch a contemporary, rigorous notion of
autonomous agency (Juarrero 1999; Freeman
1999; Skarda 1999; Lakoff & Johnson 1999;
Libet, Freeman & Sutherland 1999), the most
extensive enactive treatment of autonomy is
probably still Varela’s early work (1980).
Embodiment. In many ways, “embodiment” is
becoming the distinguishing characteristic of
much research and thought in recent cogni-
tive science. Although some (e.g., Anderson
2003) argue that this need not (and does not)
call for anything radical, such claims do not
usually pass without comment (in this case,
see for example, Chrisley 2003). More typical,
for our purposes, is work that, for example,
76
argues for the bodily-basis of mathematics
(Lakoff & Núñez 2001). In a recent major
work, Shaun Gallagher (2005) has argued for
a model of embodied cognition that inte-
grates findings in perception, neuroscience,
consciousness studies, phenomenology, and
developmental psychology.
Within robotics, Tom Ziemke (2003) is
particularly noteworthy for trying to inte-
grate broadly enactive concepts and the phys-
ical requirements of robots. Beyond the cur-
rent work, it is not obvious how (or whether)
certain enactive concepts can or will be real-
ized in physical systems (e.g., “downward cau-
sation”). However, there are already systems
with the (limited) ability to physically grow
and change as a result of interaction with the
phenomenal world they inhabit. It thus seems
conceivable that at some point in the future it
will be possible to conduct physical experi-
ments that bring work in robotics closer to
“organic” insights from biology. Although
such work is in its technical infancy, there are
nonetheless relevant recent developments
around the theme of “evolutionary robotics”
(e.g., Lipson & Pollock, 2000) - and even the
initial work of Brooks on “mushy robotics”
(Brockman, 2002).
An additional dimension to the issue of
embodied cognition is something that has
largely been the concern of biology, namely,
how do we explain the fact that biological
organisms have the form they have? This was
the focus of pioneering work in the early
1900s by D’Arcy Thompson (1992), and later
by Alan Turing (1952). The rise of Artificial
Life in the late 1980s has catalyzed a renewed
interest in this, with a particular focus on
dynamical systems models of self-organizing
morphogenesis (e.g., Kauffman 1993; Good-
win 1994). Of course, most biologists accept
that, in broad terms, form is partly the result
of selection and genetic inheritance and vari-
ation. However, there are a number of (non-
religious) claims that additional mechanisms
are required. Thus, in recent times, there has
been a renewed interest in morphogenesis,
with attempts to introduce such concepts as
self-organization as an addition factor in
development. Another variation on this
theme is a proposal by Varela and Frenk
(1987) that the evolution of physical form is
the result of an the ongoing dynamic between
cellular and extra-cellular – between local and
global.
enactive cognitive science
Finally, one aspect of recent work in
embodied cognition that is particularly
interesting centers on the issue of perception.
In the tradition of Dewey, Merleau-Ponty,
Gibson, Bruner, and others, this work argues
for a constructive, active notion of percep-
tion. Representative of this perspective is the
work of Alva Noë (2004). It is actually diffi-
cult to summarize this model of perception
as it is so unlike the usual notion of “passive
sensors triggered by external signals.” Briefly,
Noë asks us to accept that vision (and percep-
tion in general) is like touch – it is the result
of actively reaching. Noë references a wide
variety of fascinating research in support of
his claim. There are numerous experiments
that collectively suggest neither “seeing” nor
“not seeing” is correlated in any simple way
with visual sensors. These results have been
demonstrated by various kinds of “blind-
ness” in otherwise healthy visual receptors –
and the ability to “see” for people who are
congenitally blind (fitted with other kinds of
sensors).
Consciousness and first-person cognition. In
recent years there has been an increase in
studies of consciousness and first-person cog-
nition (see Petitot, Pachoud & Roy 1999; Hur-
ley 1998; Gallagher & Shear 1999). One char-
acteristic of this work is an attempt to find
ways in which cognitive science and Conti-
nental philosophy can inform each other
(Thompson 2001). Although not universally
true, recent efforts have seen an attempt to
continue or extend the phenomenologically-
oriented work of such pioneers as Husserl and
Merleau-Ponty – rather than engage in tex-
tual analysis and debate about how to inter-
pret that earlier work. To be sure, much of the
recent work shows heightened awareness and
familiarity with philosophy, and certain his-
torical developments are mentioned in pass-
ing. But, to take one example, Juarrero (1999)
is indicative of this trend in that she summa-
rizes the history of Aristotle’s notion of cause
in order to highlight how certain contempo-
rary technical and conceptual advances can
overcome historical (and often hidden)
assumptions that may have prevented earlier
progress.
One of the most interesting recent devel-
opments in studies of consciousness is neu-
rophenomenology, an “experiential neuro-
science” (Varela 1996, 1999b; Varela &
Constructivist Foundations
 engineering-computer scientific
SURVEY
Thompson 2003). Although there are efforts
in consciousness studies to correlate the neu-
ral with the experiential (Metzinger 2000;
Edelman 2004), neurophenomenology pro-
poses a different method. Namely, that the
phenomenal is not reducible to the neurolog-
ical, and that each requires its own method of
investigation. And, further, that the formal
description can act as a bridge – but also par-
ticipates in a process by which all three mutu-
ally inform each other. For additional work
on this, see also (Thompson, Lutz & Cosmelli
2005; Gallagher 2002; Hanna & Thompson
2003; Rudrauf et al. 2003; Lutz & Thompson
2003).
Intersubjectivity. There has been quite a bit of
research in recent years on the issue of inter-
subjectivity (how we understand the think-
ing/mind of others). One way to understand
this work is in terms of three broad explana-
tory proposals. These suggest we “know what
other people are thinking” because we have a
particular theoretical model, we perform a
mental simulation (“how we would think if we
were in that situation and doing those
things”), or as embodied action (for a more
detailed treatment of this classification and
the various proposals, see Gallagher2005; for
additional discussion of intersubjectivity that
is resonant with the enactive perspective, see
also Thompson 1999, Zahavi 2001, and
Arisaki 2001).
One particular result that has generated
enormous interest is the recent discovery of
“mirror neurons” (Gallese & Goldman,
1998). Evidence suggests that very specific
neurons fire when a monkey performs some
action, such as biting into a fruit that tastes
bitter; those same neurons fire when the mon-
key observes another monkey performing that
action (i.e., in this case, biting and reacting to
something bitter). Mirror neurons have been
used to support each of the three explanatory
proposals for intersubjectivity (for review, see
Gallagher 2005; for further examples, see
Metzinger 2000 and Thompson 2003), but
from the enactive perspective, these are inter-
esting because they suggest that much of how
we “know others” is related to how we “know
ourselves.” That is, directly, as opposed to the-
oretically. Note that it is early days yet with
regard to the research on mirror neurons, so
it remains to be seen what will happen after
further study.
2006, vol. 1, no. 2
Why it matters
Enactive cognitive science is not proposing “a
world in the head”, rather it is proposing a rig-
orous and scientific approach to cognition
that differs from cognitivism by making some
different assumptions, proposing different
hypotheses and mechanisms, and arriving at
some different interpretations of the results.
But one may still wonder: what actual differ-
ence does it make whether one does enactive
or objectivist/subjectivist cognitive science?
There are several evident differences. First,
enactive researchers tend to emphasize the
provisional nature of scientific understand-
ing; of course, this is often true for objectivist
researchers, but there seems to be more of a
tendency to slide from “strong evidence” to
“fact” to “objective truth.” Second, as we have
seen in this paper, there is often a great differ-
ence in the kinds of explanations – and, in the
case of synthetic hypotheses, a difference in
the kinds of mechanisms (or their relation-
ship to the theoretical model as a whole).
Third, the development of new methodolo-
gies, as in the example of the Piagetian inter-
view or new techniques for studying first-per-
son cognition.
But there may even be much larger conse-
quences which we could characterize in terms
of “revisions to cognitive science” as a whole.
So far we have considered phenomena that
motivate alternatives to cognitivism, and
looked at some of the enactivist research from
the rather conventional model of scientific
method: observations, hypotheses, methods,
experiments, results, and interpretations of
those results. But work in enactive cognitive
science (and other domains) raise some ques-
tions about the scope of scientific inquiry and
the limits of current models of scientific
method.
In some cases, this is because of the phe-
nomena, as when a science expands (or
diminishes) its scope. This may require the
development of new formalisms, architec-
tures, mechanisms, or research methods to
understand or explain certain phenomena of
interest, such as autonomy, choice, inten-
tions, invention, creativity, play, metaphor,
intuition, empathy, emotion, and meaning.
To be sure, many of these are traditionally part
of psychology. However, cognitivism, much
like behaviorism before it, has largely had an
uneasy relationship with many of these
enactive cognitive science
topics – with reactions ranging from outright
embarrassment to attempts to dismiss them
as non-phenomena. (For an overview of ways
different traditions have treated the topic of
metaphor, see Lakoff & Johnson 1999.) In
other cases, the necessary shift may be more
profound, as may occur in the way various
sciences reconcile the role of the observer with
traditional notions of impartiality and objec-
tivity. In the case of enactive cognitive science,
one of the claims is that there is a fundamental
dependency between knower and the known.
This (and similar implications from contem-
porary physics) raise a number of challenges
for the classical model of scientific method
with its assumptions about objective perspec-
tives (for a discussion of this topic from an
enactive perspective, see, e.g., Maturana 1988;
Maturana & Poerksen 2004; Bunnell 2004). In
the case of cognitive science, where one may
interact with cognitive beings who change as
a result of the interactions, the methodologi-
cal challenges are perhaps even more com-
plex. (Piaget’s use of clinical method, based as
it is on dynamic and case-by-case interaction
with study-subjects, has been criticized in just
such terms.)
Another kind of consequence is related to
the possibility of applying results and insights
from enactive cognitive science to the design
of technologies. Conventionally, a distinction
is made between knowledge (science) and
application (design or engineering). Since the
enactive tradition emphasizes knowing as
doing, it may seem odd to speak of “applica-
tion” as a separate category. However, the
point here is simply that although we have
seen a number of examples of how enactive
concerns have been transformed into
research about cognitive functioning, we have
not yet looked at different ways that those
insights can form the basis of new approaches
to the development of products or services for
people to use.
To be sure, many have used such concepts
as embodiment, situatedness, and the like in
the development of working robots. But there
are very few attempts to leverage this new par-
adigm of cognition to explicitly guide the
design of artifacts. One (perhaps the only)
notable example is the work of Winograd and
Flores (1987) to design a group-coordination
system, The Coordinator, derived explicitly
from broadly enactive insights and principles.
One could argue that the design of the pro-
77
 engineering-computer scientific enactive cognitive science
SURVEY

gramming language Logo (Papert, 1980) is Coordinator, in many ways Logo would be would it mean to develop enactive cognitive
another example. However, in the case of considered the more successful of the two.) interfaces – if the notion of “interface” is even
Logo, it seems as if the design took place more To give an idea of how enactive cognitive still appropriate? Even more intriguing is to
in a spirit of respecting the autonomy of the science may open up new insights in the consider ways in which an enactive interpre-
child-as-programmer – in much the same development of technology, consider the tation of mirror neurons could inform the
spirit that has informed other constructivist work on intelligent user interfaces (Maybury & design of future embodied intelligent tech-
proposals for educational praxis (e.g., Duck- Wahlster, 1998). By and large, work in this nologies.
worth, 1987) – rather than as a result of trying area is conducted with a conventional model
to translate specific cognitive insights into of input/output and back-end intelligent pro-
design choices and guidelines. (Of course, the cessing. That is, the standard model assumes Coda
manner and degree to which assumptions that the “interface” is (and can be) largely
directly influence design choices is complex. decoupled from the “intelligence inside.” As The Embodied Mind was published during a
And, even granting that the design of Logo we have seen, enactive cognitive sciences period that ushered in what is now sometimes
involved less explicit articulation and applica- questions a number of the cognitivist called “second generation cognitive science”
tion of constructivist principles than the assumptions behind such models. Thus, what (Lakoff & Johnson 1999) and “New AI” (see
e.g., Dorffner 1997; Pfeifer & Scheier 1999;
Franklin 1995). Although there was a great
FUTURE WORK variety of work, there were certain concerns
common to much of it: looking for alterna-
There remain many open questions about “the common principles or lawful linkages between tives to classic representation, studying cogni-
sensory and motor systems that explain how action can be perceptually guided in a perceiver- tion as part of larger and more naturalistic
dependent world” (Varela, Thompson, & Rosch 1991, p. 173 ); these issues touch on every- contexts, a re-emphasis on the importance of
thing from the replication of research to the elaboration and testing of different formal mod- biological and neuro-physiological plausibil-
els for phenomena of interest (e.g., downward causation). ity, and the like.
On the one hand, it will be interesting to see whether it is possible to synthesize models The spirit of the times included revived
that appear to have potential to complement each other, such as Maturana and Varela’s auto- interest in connectionism (Rumelhart &
poietic model of the cell and Freeman and Skarda’s model of mass action in the nervous McClelland 1986; Gardner 1993) and systems
system. On the other hand, there are some vast challenges to develop suitable models of theory (Jantsch 1979); a new wave of work on
group-dynamics that are compatible with the claims of Vygotsky and radical interpretations mobile robotics (Brady & Paul 1984; Brooks
of social constructivism. Even if dynamical systems theory can be successful here, it is not 1999); artificial life as a new field (Langton
clear how such success could then be transformed into possible mechanisms. 1989); interesting suggestions about how
This last point is an example of certain methodological tensions. For example, there are “body knowledge” could inform serious sci-
currently mathematical models that are not easy to translate into mechanisms. A number of entific effort (Fox-Keller 1983) – and, coming
contemporary researchers have commented on this tension, especially with regard to pos- from the other direction, interesting chal-
sible relationships between quite abstract system dynamics models and computational mech- lenges to conventional ideas about the role of
anisms (see, e.g., Clark 1997, p. 170). This may not be an issue when the vocabulary and computers and computational concepts as
mathematics of dynamical systems models is used to explain connectionist implementations they informed notions of self (Papert 1980;
or neural functioning; however, it does raise questions if one can only explain certain kinds Turkle 1984); more expansive notions of cog-
of cognitive phenomena in terms of, say, dynamical systems theory and would like to then nition (Gilligan 1982; Lave 1988) – and appli-
use such descriptions as the basis for creating and testing hypothetical mechanisms. There is cations of social studies (Suchman 1987) and
some work to do this (see, e.g., Beer 2000; Port & van Gelder 1995), but it is thus far largely of phenomenological method (Sudnow
limited to accounts of individual-oriented accounts. In other words, dynamical systems mod- 1978); greater access to work from the Soviet
els have the potential to formally model social theories that are not bottom-up; how this can Union in such areas as psychology (Vygotsky
be translated into mechanisms is not obvious. 1978), activity theory (Leont’ev 1978), and
Finally, external criticisms of enactive cognitive science vary greatly. Many are often based literary theory (Bakhtin 1993); and, overall,
on differences of belief (“we believe there is an objective reality because the alternatives an opening-up of cognitive science to new
seem appalling”) or the inability to see an alternative to objectivism as anything other than sources of technical and philosophical insight
subjectivism (“enactive cognitive science is simply solipsism operationalized”). In other cases, (Graubard 1988; Winograd & Flores 1987).
the criticisms seem largely based on complaints that certain aspects of assumptions, method, How do things look now? There seems to
vocabulary, or interpretation are “unconventional” (e.g., Boden 2000). One of the more be a significantly larger group within cogni-
sharply-formulated criticisms is that constructivist-oriented approaches to scientific research tive science committed to exploring various
cannot be disconfirmed, and thus fail a basic criterion for being considered scientific (Popper non-objectivist and non-subjectivist alterna-
1963). It is not clear how to address such issues, where so much of the conceptual framework tives, although, to be sure, the key issue of rep-
is so different; it may be that some new synthesis is required to reconcile the differences. resentation remains as controversial as ever.
And this second generation of cognitive sci-

78 Constructivist Foundations
 engineering-computer scientific
SURVEY
ence seems to be converging on (various
interpretations of) embodiment as its distinc-
tive characteristic. Cognitivism is still alive
and well (for a recent summary, see, e.g.,
Thagard 2005), but it is no longer the only
game in town. In cognitive science in general,
there is now much more emphasis on emo-
tions, intersubjectivity, and consciousness (see,
e.g., Damasio 1999; Picard 1997; Trevarthen
1993; Zahavi 2005; Shear 1997; Chalmers
1996), and within the enactive tradition there
is growing emphasis on how these may inform
each other (see, e.g., Thompson 2001; Cleer-
emans 2003 – and for a particularly lucid and
succinct account, see Thompson 1999).
Having said that, enactive cognitive sci-
ence might not endure much longer in name.
In a paper preceding the publication of The
Embodied Mind by five years, Varela indicates
that the name enactive is provisional, “until a
better one is proposed” (Varela 1992,
endnote 35). Varela died in 2001, but towards
the end of his life, he made it clear that his
work was in the area of neurophenomenology
– and he even indicated that he felt enactive
cognitive science had been a transitional dis-
cipline (Varela 1999a). Evan Thompson occa-
sionally uses the term “enactive” to describe
his work, but he too seems to mainly situate
References
Agre, P. E. & Chapman, D. (1990) What are
plans for? Robotics and Autonomous Sys-
tems 6: 17–34.
Anderson, M. L. (2003) Embodied cognition:
A field guide. Artificial Intelligence 149(1):
91–130.
Arisaka, Y. (2001) The ontological co-emer-
gence of “self and other” in Japanese Phi-
losophy. In: Thompson, E. (ed.) Between
ourselves: Second-person issues in the
study of consciousness. Imprint Aca-
demic: Thorverton, UK.
Bakhtin, M. M. (1993) Towards a philosophy
of the act. University of Texas Press: Aus-
tin.
Ballard, D. H. (1999) An introduction to nat-
ural computation. MIT Press: Cambridge,
MA.
Beer, R. D. (2000) Dynamical approaches to
cognitive science. Trends in Cognitive Sci-
ences 4(3): 91–99.
2006, vol. 1, no. 2
his recent work in terms of neurophenome-
nology. Since the publication of the book,
Eleanor Rosch seems to have largely avoided
the term and has devoted more and more of
her efforts to first-person work that she her-
self admits is difficult to imagine fitting into
standard Western scientific or academic
research on cognition (Scharmer 1999). And
as for other researchers mentioned in this
paper, most of them present their work under
some other name. The exception, as noted, is
researchers who wish to emphasize a particu-
larly active conception of, say, perception
(e.g., Noë 2004). It may be that the term ulti-
mately comes to refer to such a research focus.
The current work in embodied cognition
seems to mark the maturation of research
ABOUT THE AUTHOR
Kevin McGee has a PhD in Media Arts and
Sciences from MIT. His main research inter-
est is the development of partner technol-
ogies based on enactive cognitive models.
He is currently affiliated with the Depart-
ment of Computer Science, Linkoping Uni-
versity and the Santa Anna Research
Institute, both in Linköping, Sweden.
Boden, M. A. (2000) Autopoiesis and life.
Cognitive Science Quarterly 1(1): 117–
145.
Brady, M. & Paul, R. (eds.) (1984) Robotics
research: The first international sympo-
sium. MIT Press: Cambridge, MA.
Brockman, J. (2002) Computation: A talk
with Rodney Brooks. Edge: The Third
Culture.
Brooks, D. & Wiley, E. (1988) Evolution as
entropy. University of Chicago Press: Chi-
cago, second edition.
Brooks, R. A. (1999) Cambrian intelligence:
The early history of the new AI. MIT Press:
Cambridge, MA.
Bruner, J. & Kalmer, J. D. (1998) Narrative
and metanarrative in the construction of
self. In: Ferrari, M. & Sternberg, R. J. (eds.)
Self-awareness: Its nature and develop-
ment. Guilford: New York, pp. 308–331.
Bunnell, P. (2004) Reflections on the ontology
of observing. Cybernetics and Human
Knowing 11(4): 72–84.
enactive cognitive science
that takes seriously many of the concerns
originally expressed in The Embodied Mind
and works by others during that period.
Hopefully this trend will continue and we
will see more expansive notions of cognition
– and further effort at a rapprochement
between the many important insights of
Western analytic sciences of the mind and
other relevant insights and practices, both
within the Western tradition (life sciences,
extended notion of relevant philosophical
traditions, and the arts) and from beyond it.
Perhaps it is not unrealistic to hope that
there will be mutual exchange on a global
scale, as the study of cognition is able practi-
cally to draw on the various many insights,
models, and traditions from the entire world
– and to hope that cognitive science can be
expansive enough to encourage and support
improved understanding and praxis, both
individual and social.
Acknowledgements
For invaluable feedback and suggestions, I
would like to thank Alexander Riegler and the
two reviewers of this paper. For the creation of
the custom-bib utility, thanks to Patrick Daly.
Chalmers, D. (1996) The conscious mind: In
search of a fundamental theory. Oxford
University Press: Oxford.
Chrisley, R. (2003) Embodied artificial intel-
ligence. Artificial Intelligence 149(1): 131–
150.
Clark, A. (1997) Being there: Putting brain,
body, and world together again. MIT
Press: Cambridge, MA.
Cleeremans, A. (ed.) (2003) The unity of con-
sciousness: Binding, integration, and dis-
sociation. Oxford University Press: New
York.
Damasio, A. (1999) The feeling of what hap-
pens: Body, emotion and the making of
consciousness. Harcourt Brace: New York.
Dorffner, G. (ed.) (1997) Neural networks
and a new AI. Chapman and Hall: London.
Drescher, G. L. (1991) Made-up minds: A
constructivist approach to artificial intelli-
gence. MIT Press: Cambridge, MA.
Duckworth, E. (1987) “The having of won-
derful ideas” and other essays on teaching
79
 engineering-computer scientific
SURVEY
and learning. Teachers College Press: New
York.
Edelman, G. M. (1992) Bright air, brilliant
fire: On the matter of the mind. Basic
Books: New York.
Edelman, G. M. (2004) Wider than the sky:
The phenomenal gift of consciousness.
Yale University Press: New Haven, CT.
Ellis, R. & Humphreys, G. (1999) Connec-
tionist psychology: A text with readings.
Psychology Press: East Sussex, UK.
Flavell, J. H. (1963) The developemental psy-
chology of Jean Piaget. D. Van Nostrand
Company: Princeton, NJ.
Fox-Keller, E. (1983) A feeling for the organ-
ism: The life and work of Barbara McClin-
tock. W.H. Freeman: San Francisco.
Franklin, S. (1995) Artificial minds. MIT
Press: Cambridge, MA.
Freeman, W. J. (1999) Consciousness, inten-
tionaly, and causality. In: Núñez, R. &
Freeman, W. J. (eds.) Reclaiming cogni-
tion: The primacy of action, intention, and
emotion. Imprint Academic: Thorverton,
UK, pp. 143–172.
Gallagher, S. (ed.) (2002) Special issue: Fran-
cisco Varela’s neurophenomenology of
radical embodiment. Phenomenology
and the Cognitive Sciences 1(2).
Gallagher, S. (2005) How the body shapes the
mind. Oxford University Press: Oxford.
Gallagher, S. & Shear, J. (eds.) (1999) Models
of the self. Imprint Academic: Thorverton,
UK.
Gallese, V. & Goldman, A. (1998) Mirror neu-
rons and the simulation theory of mind-
reading. Trends in Cognitive Sciences
2(12): 493–501.
Gardner, D. (ed.) (1993) Neurobiology of
neural networks. MIT Press: Cambridge,
MA.
Gilligan, C. (1982) In a different voice: Psy-
chological theory and women’s develop-
ment. Harvard University Press:
Cambridge, MA.
Ginsburg, H. P. (1997) Entering the child’s
mind: The clinical interview in psycholog-
ical research and practice. Cambridge
University Press: New York.
Goodwin, B. (1994) How the leopard
changed its spots: The evolution of com-
plexity. Touchstone: New York.
Graubard, S. R. (ed.) (1988) The artificial
intelligence debate: False starts, real foun-
dations. MIT Press: Cambridge, MA.
80
Hanna, R. & Thompson, E. (2003) Neurophe-
nomenology and the spontaneity of con-
sciousness. In: Thompson, E. (ed.) The
problem of consciousness: New essays in
phenomenological philosophy of mind.
University of Calgary Press: Calgary, Can-
ada, pp. 133–162.
Hurley, S. L. (1998) Consciousness in action.
Harvard University Press: Cambridge,
MA.
Hutchins, E. (1995) Cognition in the wild.
MIT Press: Cambridge, MA.
Jantsch, E. (1979) The self-organizing uni-
verse: Scientific and human implications
of the emerging paradigm of evolution.
Pergamon Press: Oxford.
Juarrero, A. (1999) Dynamics in action:
Intentional behavior as a complex system.
MIT Press: Cambridge, MA.
Kauffman, S. A. (1993) Origins of order: Self-
organization and selection in evolution.
Oxford University Press: Oxford.
Kelso, J. A. S. (1995) Dynamic patterns: The
self-organization of brain and behavior.
MIT Press: Cambridge, MA.
Lakoff, G. & Johnson, M. (1999) Philosophy
in the flesh: The embodied mind and its
challenge to western thought. Basic Books:
New York.
Lakoff, G. & Núñez, R. (2001) Where mathe-
matics comes from: How the embodied
mind brings mathematics into being.
Basic Books: New York.
Langton, C. (ed.) (1989) Artificial life: The
proceedings of an interdisciplinary work-
shop on the synthesis and simulation of
living systems, held september, 1987, in
Los Alamos, New Mexico. Addison-Wes-
ley: Redwood City, CA.
Lave, J. (1988) Cognition in practice. Cam-
bridge University Press: Cambridge.
Leont’ev, A. N. (1978) Activity, conscious-
ness, and personality. Prentice-Hall:
Englewood Cliffs, NJ.
Lewontin, R. (1983) The organism as the sub-
ject and object of evolution. Scientia 118:
65–82.
Libet, B., Freeman, A. & Sutherland, K. (eds.)
(1999) The volitional brain: Towards a
neuroscience of free will. Imprint Aca-
demic: Thorverton, UK.
Lipson, H. & Pollock, J. (2000) Automatic
design and manufacture of robotic life-
forms. Nature 406(31): 974–978.
Luisi, P. L. & Varela, F. J. (1989) Self-replicat-
enactive cognitive science
ing micelles: A chemical version of a min-
imal autopoietic system. Origins of Life
and Evolution of the Biosphere (19): 633–
43.
Lutz, A. & Thompson, E. (2003) Neurophe-
nomenology: Integrating subjective expe-
rience and brain dynamics in the
neuroscience of consciousness. Journal of
Consciousness Studies 10(9–10): 31–52.
Maturana, H. R. (1980) Introduction. In:
Autopoiesis and cognition: The realization
of the living. D. Reidel: Boston, pp. xi–xxx.
Maturana, H. R. (1988) Reality: The search
for objectivity or the quest for a compel-
ling argument. The Irish Journal of Psy-
chology 9(1): 25–82.
Maturana, H. R. & Poerksen, B. (2004) Vari-
eties of objectivity. Cybernetics and
Human Knowing 11(4): 63–71.
Maturana, H. R. & Varela, F. J. (1980) Auto-
poiesis and cognition: The realization of
the living. D. Reidel Publishing Company:
Boston.
Maturana, H. R. & Varela, F. J. (1987) The tree
of knowledge: The biological roots of
human understanding. New Science
Library: Boston.
Maybury, M. & Wahlster, W. (eds.) (1998)
Readings in intelligent user interfaces.
Morgan Kaufmann: San Francisco, CA.
McGee, K. (2005) Enactive cognitive science.
part 1: Background and research themes.
Constructivist Foundations 1(1): 19–34.
McMullin, B. (2004) 30 years of computa-
tional autopoiesis: A review. Artificial Life
10(3): 71–94.
Medler, D. (1998) A brief history of connec-
tionism. Neural Computing Surveys 1(1):
61–101.
Messer, S. B., Sass, L. A. & Woolfolk, R. L.
(eds.) (1994) Hermeneutics and psycho-
logical theory: Interpretative perspectives
on personality, psychotherapy, and psy-
chopathology. Rutgers University Press.
Metzinger, T. (ed.) (2000) Neural correlates of
consciousness: Empirical and conceptual
questions. MIT Press: Cambridge, MA.
Noë, A. (2004) Action in perception. MIT
Press: Cambridge, MA.
O’Reilly, R. C. & Munakata, Y. (2000) Com-
putational explorations in cognitive neu-
roscience: Understanding the mind by
simulating the brain. MIT Press: Cam-
bridge, MA.
Oyama, S., Griffiths, P. E. & Gray, R. D. (eds.)
Constructivist Foundations
 engineering-computer scientific
SURVEY
(2001) Cycles of contingency: Develop-
mental systems and evolution. MIT Press:
Cambridge, MA.
Papert, S. (1980) Mindstorms: Children,
computers, and powerful ideas. Basic
Books: New York.
Petitot, J., Varela, F. J., Pachoud, B. & Roy, J.-
M. (eds.) (1999) Naturalizing phenome-
nology: Issues in contemporary phenom-
enology and cognitive science. Stanford
University Press: Stanford, CA.
Pfeifer, R. & Scheier, C. (1999) Understanding
intelligence. MIT Press: Cambridge, MA.
Piaget, J. (1963) The child’s conception of the
world. Littlefield, Adams & Co: Paterson,
NJ.
Picard, R. W. (1997) Affective computing.
MIT Press: Cambridge, MA.
Popper, K. (1963) Conjectures and refuta-
tions: The growth of scientific knowledge.
Routledge: London.
Port, R. F. & van Gelder, T. (eds.) (1995) Mind
as motion: Explorations in the dynamics
of cognition. MIT Press: Cambridge, MA.
Riegler, A. (2005) Constructive memory.
Kybernetes 34(1/2): 89–104.
Rosenblueth, A., Wiener, N. & Bigelow, J.
(1943) Behavior, purpose and teleology.
Philosophy of Science 10(1): 18–24.
Rudrauf, D., Lutz, A., Cosmelli, D., Lachaux,
J.-P. & Quyen, M. L. V. (2003) From auto-
poiesis to neurophenomenology: Fran-
cisco Varela’s exploration of the biophysics
of being. Biological Research 36(1): 27–65.
Rumelhart, D. E. & McClelland, J. L. (eds.)
(1986) Parallel distributed processing:
Explorations in the microstructure of cog-
nition, volumes 1–2. MIT Press: Cam-
bridge, MA.
Scharmer, C. O. (1999) Primary knowing:
When perception happens from the whole
field: Conversation with professor Eleanor
Rosch.
Shear, J. (ed.) (1997) Explaining conscious-
ness: The hard problem. MIT Press: Cam-
bridge, MA.
Skarda, C. A. (1999) The perceptual form of
life. In: Núñez, R. & Freeman, W. J. (eds.)
Reclaiming cognition: The primacy of
action, intention, and emotion. Imprint
Academic: Thorverton, UK, pp. 79–93.
Skarda, C. A. & Freeman, W. J. (1987) How
brains make chaos in order to make sense
of the world. Behavioral and Brain Sci-
ences 10(2): 161–95.
2006, vol. 1, no. 2
Suchman, L. A. (1987) Plans and situated
actions: The problem of human–machine
communication. Cambridge University
Press: Cambridge.
Sudnow, D. A. (1978) Ways of the hand: The
organization of improvised conduct. Har-
vard University Press: Cambridge, MA.
Thagard, P. (2005) Mind: Introduction to
cognitive science. Bradford Books: Cam-
bridge, MA, second edition.
Thompson, D. W. (1992) On growth and
form: The complete revised edition. Dover
Publications: New York.
Thompson, E. (1999) Keynote address: The
intersubjectivity of human consciousness:
Phenomenology and cognitive science. In:
Conference on human consciousness:
Integrating phenomenology and cognitive
science. The Fetzer Institute: Kalamazoo,
MI.
Thompson, E. (ed.) (2001) Between our-
selves: Second-person issues in the study
of consciousness. Imprint Academic:
Thorverton, UK.
Thompson, E. (ed.) (2003) The problem of
consciousness: New essays in phenomeno-
logical philosophy of mind. University of
Calgary Press: Calgary, Canada.
Thompson, E., Lutz, A. & Cosmelli, D. (2005)
Neurophenomenology: An introduction
for neurophilosophers. In: Brook, A. &
Akins, K. (eds.) Cognition and the brain:
The philosophy and neuroscience move-
ment. Cambridge University Press: New
York.
Thompson, E. & Varela, F. J. (2001) Radical
embodiment: Neural dynamics and con-
scious experience. Trends in Cognitive Sci-
ence (5): 418–25.
Trevarthen, C. (1993) The self born in inter-
subjectivity: The psychology of an infant
communicating. In: Neisser, U. (ed.) The
perceived self: Ecological and interper-
sonal sources of self-knowledge. Cam-
bridge University Press: Cambridge.
Turing, A. M. (1952) The chemical basis of
morphogenesis. Phil. Trans. Roy. Soc. of
London Series B: Biological Sciences(237):
37–72.
Turkle, S. (1984) The second self: Computers
and the human spirit. Simon and Schuster:
New York.
Ulanowicz, R. E. (1997) Ecology: The ascen-
dent perspective. Columbia University
Press: New York.
enactive cognitive science
Varela, F. J. (1980) Describing the logic of the
living. In: Zeleny, M. (ed.) Autopoeisis: A
theory of the living organization. North-
Holland: New York, pp. 36–48.
Varela, F. J. (1992) Whence perceptual mean-
ing? A cartography of current ideas. In:
Varela, F. J. & Dupuy, J.-P. (eds.) Under-
standing origins: Contemporary views on
the origin of life, mind, and society. Klu-
wer Academic: Boston, pp. 235–263.
Varela, F. J. (1996) Neurophenomenology: A
methodological remedy for the hard prob-
lem. Journal of Consciousness Studies
3(4): 330–349.
Varela, F. J. (1999a) Ethical know-how:
Action, wisdom, and cognition. Stanford
University Press: Stanford, CA.
Varela, F. J. (1999b) Steps to a science of inter-
being. In: Watson, G., Batchelor, S. & Clax-
ton, G. (eds.) The psychology of awaken-
ing: Buddhism, science and our day-to-
day lives. Rider: London.
Varela, F. J. & Frenk, S. (1987) The organ of
form: Towards a theory of biological
shape. Journal of Social and Biological
Systems 10(1): 73–83.
Varela, F. J. & Shear, J. (eds.) (1999) The view
from within: First-person approaches to
the study of consciousness. Imprint Aca-
demic: Thorverton, UK.
Varela, F. J. & Thompson, E. (2003) Neural
synchrony and consciousness: A neu-
rophenomenological perspective. In:
Cleeremans, A. (ed.) The Unity of con-
sciousness: Binding, integration, and dis-
sociation. Oxford University Press: New
York.
Varela, F. J., Thompson, E. & Rosch, E. (1991)
The embodied mind: Cognitive science
and human experience. MIT Press: Cam-
bridge, MA.
Vygotsky, L. S. (1978) Mind in society: The
development of higher mental processes.
Harvard University Press: Cambridge,
MA.
Winograd, T. & Flores, F. (1987) Understand-
ing computers and cognition. Ablex Pub-
lishing: Norwood, NJ.
Wolfram, S. (1994) Cellular automata and
complexity. Addison-Wesley: Reading,
MA.
Zahavi, D. (2001) Beyond empathy: Phenom-
enological approaches to intersubjectivity.
Journal of Consciousness Studies 8(5–7):
151–167.
81
 engineering-computer scientific
SURVEY
Zahavi, D. (2005) Subjectivity and selfhood:
Investigating the first-person perspective.
MIT Press: Cambridge, MA.
Ziemke, T. (2003) Robosemiotics and
embodied enactive cognition. SEED Semi-
otics, Evolution, Energy, and Develop-
82
ment 3(3): 112–124.
Ziemke, T. (2004) Embodied AI as science:
Models of embodied cognition, embodied
models of cognition, or both? In: Iida, F.,
Pfeifer, R., Steels, L. & Kuniyoshi, Y. (eds.)
Embodied artificial intelligence: Interna-
enactive cognitive science
tional seminar, Dagstuhl Castle, Germany,
July 7–11, 2003. Revised Papers. Springer-
Verlag: Heidelberg, pp. 27–36.
Received: 15 September 2005
Accepted: 13 February 2006
Constructivist Foundations
 engineering-computer scientific radical constructivism
SYNTHETICAL

Towards Closed Loop Information:

Predictive Information
Bernd Porr A University of Glasgow, b.porr@elec.gla.ac.uk
Alice Egerton A Imperial College London, aegerton@imperial.ac.uk
Florentin Wörgötter A University of Göttingen and University of Stirling, worgott@chaos.gwdg.de

Motivation: Classical definitions of information, such as the Shannon information, are proper closed loop system is designed so that
designed for open loop systems because they define information on a channel which has it can cope with these disturbances and
an input and an output.The main motivation of this paper is to present a closed loop thereby restore the desired state of the organ-
information measure which is compatible with constructivist thinking. Design: Our infor- ism.
mation measure for a closed loop system reflects how additional sensor inputs are utilised We will now define the point of view of the
to establish additional sensor-motor loops during learning. Our information measure is closed loop system. In contrast to approaches
based on the assumption that it is not optimal to stay reactive and that it is beneficial to used in engineering, we are going to describe
become proactive through increased learning about the environment. Consequently our a closed loop system from its own perspective
information measure gauges the utilisation of new sensor inputs to generate anticipatory (Atmanspacher & Dalenoort 1994; Porr &
actions.We call this information measure “predictive information” (PI). Findings: Our PI Wörgötter 2005b). This internal perspective
is zero if the organism uses only its reflex reactions. It grows when the organism is able to of the organism was first radically employed
use other sensor inputs to preempt reflex reactions and is able to replace reflexes by by (Foerster 1960). The crucial difference
anticipatory reactions.This has been demonstrated with a real robot that had to learn to between the internal perspective and the
avoid obstacles. Conclusion: PI is a new measure which is able to quantify anticipatory outer perspective is how closed loop systems
learning and, in contrast to the Shannon information, is calculated only at the inputs of an “observe” the environment and themselves.
agent.This information measure has been successfully applied to a simple robot task but Forester claims that closed loop systems can
its application is neither limited to a certain task nor to a certain learning rule. only observe by using their own closed loops.
Keywords: Closed loop system, information measure, differential Hebbian learning, reac- Therefore, the only aspect which can be
tive vs proactive systems. observed is the closed loop which establishes
the feedback from the motor output to the
sensor input (see Fig. 1). To make it clearer:
fuzzy (Atmanspacher & Dalenoort 1994; Porr an organism can only use its own senses to
Introduction & Wörgötter 2005b) so that standard infor- judge if an action has been successful or not.
mation measures which assess information It cannot perceive the action itself. It can only
Since its introduction by Shannon and transmission can no longer be used. In this perceive the consequences of its actions. This
Weaver (1949), a quantitative information article we will present an information mea- leads directly to the question of how an
measure has been used in many disciplines. sure for a closed loop system which measures organism evaluates its actions. The answer is
For example, in telecommunications infor- how additional sensor inputs and therefore simple: only by its own sensor inputs. This
mation is used to quantify the quality of a additional sensor–motor loops are acquired leads directly to the statement that organisms
transmission. This definition of information during learning. can only control their inputs and not their
has also been applied to neurons where the Before we can define closed loop informa- outputs (Glasersfeld 1995). The contradic-
signal transmission from one neuron to the tion it is necessary to introduce closed loop tion between input- and output-control can
next one can be expressed in terms of infor- systems. Why do we need closed loop control? be made clearer by an example, which we call
mation (Rieke, Warland, de Ruyter van In the real world our knowledge about our the second chicken/egg problem (Porr &
Stevenick & Bialek 1997). Such an informa- environment is incomplete. We will always Wörgötter 2005b): Let us interpret the
tion measure treats the neuron as an input/ encounter situations that we cannot predict chicken as a closed loop system. The chicken
output system which is appropriate in this or in other words: the environment provides wants to keep the egg and acts in a way
case. However, when an organism acts in its “surprises” which are called disturbances designed to increase the sit-on-the-egg-time
environment its actions feed back to its sen- (Palm 2000). Such disturbances will cause a to improve the probability of successful
sors so that a closed loop perception–action deviation from the desired state of the organ- hatching. The farmer, however, wants to have
system is established. Consequently the dis- ism. For example, we might become hungry the egg. The farmer perceives the hen as an
tinction between input and output becomes or an enemy may attack us unexpectedly. A input–output system: food in and egg out.

Constructivist Foundations 2006, vol. 1, no. 2 83

 engineering-computer scientific radical constructivism
SYNTHETICAL

organism
sensor input which predicts the Formalising
sensor motor b imminent trigger of a late reac-
input h output tion. Consequently our infor- closed loop control
mation measure is the utilisation
environment of new sensor inputs to generate Reactive control
a an anticipatory reaction. We call Every closed loop control situation with neg-
P this information measure “pre- ative feedback has a so called desired state and
© Constructivist Foundations dictive information.” If the the goal of the control mechanism is to main-
Figure 1:The organism as a closed loop system. organism only uses its late reflex tain (or reach) this state as quickly as possible
h transforms sensor events into motor outputs. reaction, the predictive informa- (D’Azzo 1988). In our model we assume that
p transforms motor outputs into sensor events.The only tion is equal to zero. When the the desired state of the reflex feedback loop is
thing that can be perceived by the organism are actions organism is able to use other unchanging and defined by the properties of
fed back via (a). Anything that does not feed back cannot sensor inputs to preempt the late the reflex loop. We define it as x0 = 0. First we
be perceived (b) and cannot be used to establish goals. reflex reaction, thus enabling will discuss this system in the absence of
replacement of the reflex by an learning. Fig. 2a shows the situation of a non
anticipatory reaction, this pre- learning organism embedded into a very sim-
The hen, however, operates as a closed loop dictive information increases. ple but generic (i.e. unspecified) formal envi-
system to which the farmer is just a distur- Furthermore, the anticipatory reaction is a ronment which has a transfer function p0.
bance. As soon as the farmer removes the egg learned behaviour, rather than being innately This organism is able to react to an input only
the chicken will produce a new egg thereby available to the organism. Our information by means of a reflex.
restoring its desired state. This example shows measure therefore evaluates the learning pro- Fig. 2b illustrates a possible set of signals
that external and internal perspectives are cess, for which we use the ISO-learning model which can occur in such a system. When a dis-
fundamentally different and illustrates Foer- (Porr & Wörgötter 2003) which is able to turbance occurs, first the disturbance signal d
ster’s theory of organisms as closed loop sys- replace a reflex reaction by an anticipatory deviates from zero, and then the input x0
tems, acting only according to their internal reaction. senses this change x0 ≠ 0 and only finally the
perspective.
In the preceding paragraph we have
argued that an organism evaluates its own a b
actions via its sensor inputs. These inputs can d(t)
be used to improve future behaviour.
Although many sensor events are associated organism
with a real life situation, only a few inputs will
probably be able to improve the behaviour of t
the organism. For example, with a hot sur- x0 v(t)
face, it is beneficial to react to the early heat h0
radiation signal rather than a later pain signal
elicited by touching the heat source. Many
similar sequences of sensor events are V
encountered during the lifetime of an organ-
t
x0(t)
ism as the consequence of existing far-senses, P0 +
e.g., vision, hearing, smell, and near-senses
such as touch, taste, etc. Generally one
observes that the trigger of a near-sense is pre- d x0=0
ceded by that of a far sense, i.e., smell precedes environment t
taste, vision precedes touch, etc. Therefore
far-senses are often predictive of correspond-
reaction
© Constructivist Foundations delay
ing near-senses (Verschure & Coolen 1991).
Here, we will focus on the view that it is
Figure 2: a Fixed reflex loop: the organism transfers a sensor event x0 into a motor response
advantageous to react to the earliest of sensor
v with the help of the transfer function h0.The environment turns the motor response v again
events rather than to wait for later sensations.
into a sensor event x0 with the help of the transfer function p0. In the environment there exists
We may now introduce our information
the disturbance d which adds its signal at ⊕ to the reflex loop. b Possible temporal signal shapes
measure in this closed loop system. In the
occurring in the reflex loop when a disturbance d ≠ 0 happens. The desired state is x0 := 0.
paragraph above, we have argued that it is
The disturbance d is filtered by p0 and appears at x0 and is then transferred into a
beneficial for organisms to generate anticipa-
compensation signal at v which eliminates the disturbance at ⊕.
tory actions. These actions are generated by a

84 Constructivist Foundations
 engineering-computer scientific
SYNTHETICAL
motor output v can generate a reaction in
order to restore the desired state x0 = 0. No
matter how fast the controller is, there will
always be a reaction delay in such a system
which means that there will always be an
unwanted transient at x0. A system which
operates purely reactive will always experi-
ence this transient. However, we will show in
the next section that this transient can be used
to learn an anticipatory reaction which then
finally eliminates the transient at the input x0.
Predictive control
The reflex defined above may be prevented by
an earlier sensor signal that informs the
organism of an imminent disturbance. Fig. 3
shows how the disturbance d elicits a
sequence of sensor events: first it enters the
outer loop arriving at xp filtered by the envi-
ronment (p1), while it arrives at x0 only after
a delay T. The trigger of the reflex can be
avoided if the transfer function hv generates a
signal which compensates for the disturbance
d. In an ideal case the inner loop (h0 and p0)
will be completely eliminated so that the sen-
sor input x0 is always zero and the organism
uses the input xp instead. This means that the
motor reaction v is now solely generated by xp
and no longer by x0. However, as the weight of
input x0 is fixed, input x0 is still available: if
anything in the environment changes and
input xp is no longer triggered input x0 will
always serve as a “back-up.”
Learning
The filter hv in Fig. 3 can be implemented by
ISO learning (Porr & Wörgötter 2003) which
is able to learn temporal relations between
input signals and uses these relations to turn
reactive behaviour into proactive behaviour.
We consider a system of N + 1 linear filters
h receiving inputs x and producing outputs u.
The filters connect with corresponding
weights ρ to one output unit v (Fig. 4a).
We will use x0 to denote the reflex pathway
with its corresponding weight ρ0 = const. The
output v is then given as:
N Derivatives of low pass filtered signals have a
v = ∑ ρk uk uk = xk * hk (1)
k=0
where weight change is performed by the
following learning rule:
d u0. This derivative u'0 is then correlated with
ρ = μu j v' (2)
dt j
2006, vol. 1, no. 2
radical constructivism
a b
xP organism d(t)
hv
x0
h0(s) t
v(t)
V(s)
P0 +
t
T x0(t)
P1 + P01
x0=0
d t
environment
© Constructivist Foundations
Figure 3: Schematic diagram of the augmented closed loop feedback mechanism which now
contains a secondary loop. a h0 and p0 form the reflex loop already shown in Fig. 2.The new
aspect is the input-line xp which gets its signal via transfer function p1 from the disturbance d.
The reflex loop receives a delayed version (T) of the disturbance d.The adaptive controller hv
has the task to use the signal xp, which is earlier than and, thus, “predicts” the disturbance d at
x0, to generate an appropriate reaction at v to reduce the degree of change at x0. Consequently
the pathway via xp can be called the predictive loop. b Shows a schematic timing diagram for
the situation after successful learning when a disturbance has occurred. The output v sharply
coincides with the disturbance d and prevents a major change at the input x0.
The constant μ is adjusted such that all phase lead of the derivative u'0 , weights will
weight changes occur on a long time scale (i.e. grow when the signal at x1 precedes the signal
very slowly) compared to the decay of the at x0. However, if the timing between x1 and
responses u which ensures that the system x0 is reversed the weights will shrink. The cor-
operates in the steady state condition. The responding weight change curve is plotted in
transfer function h is that of a lowpass or Fig. 4c.
bandpass which transforms a δ-pulse input
into a damped oscillation. The bandpass has
two parameters: The frequency f and the qual- Closed loop
ity Q which determines the damping where a
low Q represents a high damping. In general
information:
we use as an initial condition for the weights Predictive information
ρ0 = 1 and ρj = 0, j > 0.
ISO learning is illustrated in Fig. 4b: At this point we can introduce an information
measure which reflects the performance of
phase lead so that they precede the low pass fil- predictive learning as outlined above. As the
tered signal. In our case (namely ISO-learn- measure should reflect the success of predic-
ing) the derivative is taken from the output v tive learning, it is reasonable to start with zero
of the neuron which is, before learning, iden- information before learning. In our model
tical to the low-pass filtered signal of the reflex this means that we start with a late reflex via
x0 which is associated with zero information.
the filtered predictive input u1. Because of the Predictive actions via xk, k > 0 are elicited
85
 engineering-computer scientific radical constructivism
SYNTHETICAL

a b c
HV x 1 x0
uN(t) Δρ
xN(t) hN(t) t
T
u1(t)
u1(t) ρN
xp(t) x1(t) h1(t)
v(t) = u0(t) T
ρ1
∑ v(t)
u0(t) ρ0
x0(t) h0(t) t
v'(t)
© Constructivist Foundations

Figure 4: a The neuronal circuit of ISO-learning. The shaded area marks the connections of the weights ρk, k ≥ 1 onto the neuron which
represent hv with xk = xp, k ≥ 1. b shows the inputs x0, x1, the impulse responses u for a choice of two different resonators h and the derivative
of the output v'. c shows the initial weight change ρ1(T)t = 0 for h1 = h0, Q = 1, f = 0.01 after having stimulated the two filters h0 and h1 with
delta pulses x1(t) = δ(t) and x0(t) = δ(t – T).This pulse-sequence was repeated every 2000 time steps. After 400,000 time steps the weight ρ
was measured and plotted against the temporal difference T. The learning rate was set to μ = 0.001.

when other sensor inputs are used in addition
to the reflex input. Consequently, our infor-
mation measurement should grow if these
sensor inputs are used by ISO-learning. If
there is more than one predictive sensor input
all additional sensor inputs should contribute
to the information measure.
We will now define our information value
by the weights of ISO learning:
N
ρj
PI = – ∑ -----------------------ln -----------------------
∑ k ∑ k
j=0
N ρ
k=0
N ρ
where N is the number of inputs to ISO
learning and ρj are the corresponding
weights. We call this information “Predictive
Information” (PI). Now, we have to test if our
information measure behaves according to
our theory. In the reflex-only situation, the
weight ρ0 is the only weight which is non zero.
With ρ0 ≠ 0, ρk = 0, k > 0 Eq. 3 becomes zero
which is the desired outcome. Note that the
predictive information is always zero as long
only one weight is non-zero.
To gain a better understanding of Eq. 3 we
have to recall when the weights ρk, k > 0 start
to grow. ISO learning implements differential
Hebbian learning (Kosco 1986) which means
that the weights grow when the correspond-
ing inputs xj, j > 0 precede the signal v and,
because of the fixed weight ρ0, precede the sig-
nal x0. Only those weights grow which are able
to generate an earlier reaction in relation to
the signal at x0 (or v). Consequently the
weights directly reflect the predictive power of
their corresponding inputs. The more
weights grow the more inputs are used for the
anticipatory actions. Our measure PI reflects
this: The larger the weights, the higher the
ρj value of PI.
(3)
k=0
For example, a naive person might only eat
when she/he gets hungry. This represents
basic reactive behaviour which utilises in our
case just the pathway through x0. Because
only the weight ρ0 is non-zero the predictive
information is zero. Eventually, the person
learns to anticipate when he/she will get hun-
gry and will eat earlier. In our formal model
this predicative information enters the learn-
ing circuit through x1. Learning will increase
the weight ρ1 which is associated with the pre-
dictive input x1 which leads to a non-zero pre-
dictive information because the two weights
ρ0 and ρ1 are now non-zero. The more predic-
tive signals x2,… are available and used via
nonzero weights ρ2,… the more the predic-
tive information will grow.
Weights stop growing when the reflex has
been successfully avoided. In this case the
value of PI should also be at its maximum
value. Let us assume for now that all inputs xj
are normalised to one. Having normalised
inputs makes it easier to compare the loop
gains in our feedback system. In case of pure
reactive control only the weight ρ0 is non-zero
and therefore defines the gain of the feedback
loop. During learning the weights ρj, j > 0
start to grow until the learning goal x0 = 0 has
been reached. It makes sense to assume that
the feedback gain established by the weights
ρj, j > 0 is similar to the gain of the feedback
loop (ρ0). Thus, it can be assumed that similar
values for ρj ≈ ρ0, j > 0 for the predictive path-
way will evolve as long as the different inputs
are independent or form a sequence of events.
Then every input xj, j > 0 establishes a feed-
back loop on its own. In such a case the value
of PI is at its maximum value because of an
equal distribution of the weights. In our
example (see Fig. 3) the input xp feeds into a
filterbank so that the corresponding weights
are not independent. In such a scenario the
sum of the weights ∑ j =0 ρ j will probably
N
generate a similar closed loop gain than the
reflex gain ρ0. Consequently PI will be lower
than for independent inputs xj, j > 0. If one
wants to compare the predictive information
from two different sensors then Eq. 3 can be
simply split up into the contributions from

86 Constructivist Foundations
 engineering-computer scientific
SYNTHETICAL
one sensor and from another sensor. Summa-
rising, the measure PI reflects the utilisation
of additional inputs xj, j > 0 to preempt the
reflex at input x0. The more inputs are used
the higher the PI but only under conditions
where the closed loop gain of the predictive
pathway(s) is comparable to the gain of the
reflex pathway.
Robot experiment
The task of the robot experiment is collision
avoidance. For a more detailed description of
the experiment we refer the reader to Porr &
Wörgötter (2003) and Porr, Ferber &
Wörgötter (2003). The built-in reflex behav-
iour of the robot is a retraction reaction after
the robot has collided with an obstacle. This
represents a typical feedback mechanism; the
desired state is that the signal at the collision
sensor should remain zero. In order to avoid
deviation from the desired state (i.e. collision)
the robot is fitted with range finders which
predict imminent collisions. ISO-learning is
employed to learn the existing temporal cor-
relation between the range-finder and the col-
lision sensor signals. After learning the robot
can generate a motor reaction in response to
the range finder signals and thereby avoid col-
lision and the retraction reflex.
The detailed circuit diagram of the robot is
given in Fig. 5. The robot has two ISO learn-
ers, one for the steering angle and one for the
speed of the robot. Here we focus on the ISO
learner and its corresponding weights which
control the steering angle. The predictive
information can be calculated using Eq. 3. To
be able to compare the predictive information
associated with the two range-finders we cal-
culated the predictive information for the left
and for the right sensor separately.
We conducted two experiments. In the
first experiment both range-finders were fully
operational so that both of them were able to
predict collisions. Comparing the traces of
the range finders (RF, lower trace) with the
signals from the collision sensors (CS, upper
trace) it is apparent that there is a strong tem-
poral correlation between both signals. Spe-
cifically, a peak at the collision sensor (CS) is
usually preceded by a peak coming from the
range finder (RF). In the presence of a strong
temporal correlation the weights from the
corresponding range-finders grow (see
2006, vol. 1, no. 2
radical constructivism
xpr hv
CS xpl hv
robot
RF –
ds dφ
CS x0 h0(s)
RF +
dφ
CS
xpr hv
xpl hv
+
speed: ds
constant + x0 h0(s)
bias
–
© Constructivist Foundations
Figure 5:The robot has three collision sensors (CS), two range finders (RF) and two output
neurons (speed ds and steering angle dφ). The reflex-behaviour is triggered by the collision
sensors (dotted lines).The corresponding weights of the reflex are adjusted in such a way that
the robot performs an appropriate retraction reaction (ρds 0 = 0.15 and ρ0 = –0.5).The two
dφ
signals from the left and the right range finder are fed into two filter-banks with N = 10
resonators with frequencies of fk = 1Hz/k; k ≥ 1 and Q = 1 throughout. The 20 signals from
the two filter banks converge on both the speed neuron (ds) and on the neuron responsible
for the steering angle (dφ).
Fig. 4c). Therefore, the robot successfully correlated to the reflex behaviour it does not
learns to avoid obstacles by using both range generate behaviour and therefore the predic-
finder signals. The corresponding predictive tive information stays low. If a sensor signal is
information grows for both ranger finders. able to predict the reflex it is used by ISO-
In the second experiment the right range- learning to generate an appropriate anticipa-
finder was “blindfolded” by a small cap over tory reaction which results in higher predic-
its sensor. Analysis of the traces in Fig. 4d tive information.
revealed that there was no longer a temporal
correlation between the signal from the right
range-finder (RF) and the signal from the col- Discussion
lision sensor (CS) and that only the left range-
finder was able to predict a collision (Fig. 4b). We have shown the successful application of
As expected the predictive information is still predictive information to a real world robotic
high for the left sensor but stays low for the paradigm where the predictive information
right sensor. reflects the utilisation of different sensor sig-
In summary, the predictive information nals to generate an anticipatory action.
reflects the utilisation of the different inputs Many other information measures have
to develop anticipatory responses in relation been defined (for a review in the context of
to the reflex response. If a sensor signal is not constructivism see Porr 2002) but the infor-
87
 engineering-computer scientific radical constructivism
SYNTHETICAL

We will now compare our

both sensors intact only left sensor operational predictive information with
the Shannon information.
a CS b CS Our information measure
left

uses certain properties of ISO

RF RF learning which can be also
found in other learning rules
like Hebbian learning (Hebb
c CS d CS 1949), which has been exten-
sively investigated regarding
information processing.
right

RF RF According to Linsker (1988)

Hebbian learning implements
t t information maximisation
10,000 11,000 10,000 11,000
which means that a neuron
e f transmits as much informa-
right left tion as possible from its inputs
predictive predictive
to its output. This so-called
information information
left “infomax principle” is equiva-
lent to the detection of the first
principle component in the
input space (Oja 1982). In
terms of weights this means
right that those weights grow whose
inputs are highly correlated.
As a result we preserve maxi-
mum variance at the output of
© Constructivist Foundations
the neuron which is equiva-
lent to maximum information
Figure 6: Robot experiment.The first 4 panels (a–d) show the signals of the collision sensors (CS, upper transmission through a chan-
traces) and of the range finders (RF, bottom traces) where in the left column both range finders were intact nel. Linsker’s model is there-
and in the right columns only the left range finder was intact. All signals are normalised to one. Panels e and fore directly concerned with
f depict the predictive information calculated separately for the left and the right range finder. The x axis information transmission as
shows time steps where on time step was 5 ms. described by the work of
Shannon and Weaver. Infor-
mation in this context is only
mation measure of Shannon & Weaver with the amount of surprise the events have limited by the signal to noise ratio and the cross
(1949) is most pertinent to the present dis- caused which is reflected by the logarithm in correlations between different channels. In our
cussion. The total average information Eq. 4. On the other hand the probabilities model, however information is defined as a rel-
which is caused by M discrete events is given could be also interpreted as the amount of bits evance measure. Information is measured
as: which are needed to describe the probabilities. against the reflex reaction and not against the
Even more interpretations arise when not only whole input space. In our case the input space
M
the information content but also the transmis- might be large (i.e. many sensors). However, if
I = – ∑ p k log ( p k ) (4) sion of information is considered. Such an these signals cannot be used to preempt the
k=1 information measure based on the transmis- feedback reaction the predictive information
sion of events operates on the conditional remains at zero. Therefore, in contrast to our
This is the average information generated probabilities p(yk|xk) of the channel. Informa- measure the information defined by Shannon
by all M events which have the probabilities pk. tion is zero in such a case when an event yk has and Weaver might be very high. For example,
It can be proven that if all events have the prob- been triggered by any event xk at the input of the Shannon-information of an eye or a video-
ability pk = 1/M the information is at its max- the channel. The information is highest if one camera is probably very high. Also the optical
imum which means that all events have maxi- even yk has been caused exactly by one event at nerve might transmit a large amount of infor-
mum uncertainty. The probabilities pk can be the input xk. In such a scenario it is also possi- mation in form of Shannon information.
interpreted in different ways, leading to differ- ble to introduce redundancy by increasing the However, the predictive information might
ent interpretations of the information mea- channel capacity or by adding an auxiliary still be zero if this information does not lead to
sure: The probabilities pk could be associated channel. learning and ultimately to behaviour.

88 Constructivist Foundations
 engineering-computer scientific
SYNTHETICAL
There is a second important difference
between Hebbian learning and ISO learning.
While the weights in Hebbian learning
undergo unlimited growth ISO learning sta-
bilises the weights at the moment when the
reflex has been successfully avoided. Our pre-
dictive information takes this into account.
The predictive information is only high when
the gain of the learned predictive loop is com-
parable to the gain of the reflex loop. For
example, if one weight grows in an unlimited
manner the predictive information goes to
zero because the system has basically only one
input left.
The third difference between Linsker’s
infomax principle and our predictive infor-
mation is that our measure is designed for
closed loop control whilst Linsker’s measure
relates to open loop scenarios. Our predictive
information demands feedback in order to
evaluate predictive actions. To our knowledge
only Touchette & Lloyd (2004) and Klyubin,
Polani & Nehaniv (2004) have presented
recently closed loop information measures.
However, in contrast to our approach they
define closed loop information by informa-
tion transmission from the sensors of the
agent to its motor output. This definition
makes is possible to compare open with
closed loop control but makes the assumption
that an organism can observe its output which
contradicts the constructivist’s view.
So far our predictive information uses
properties of ISO learning, in particular it
performs differential Hebbian learning and it
stabilises as soon as the reflex has been
avoided. Our ICO learning model may also be
used in situations where nested predictive
loops are not needed (Porr & Wörgötter
2005a, 2006). The application to reinforce-
ment learning (Sutton 1988) should be also
possible but a major difficulty arises from the
fact that the feedback from the environment
is more indirect because the critic first modi-
fies the actor which then generates motor
actions.
More generally it would be interesting to
develop a measure which is completely inde-
pendent of the learning rule. Such measure
must be calculated by just looking at the dif-
ferent input signals. This would make our
predictive information applicable to other
closed loop systems such as fuzzy control sys-
tems or systems which employ symbolic con-
trol such as classical AI systems. The work by
2006, vol. 1, no. 2
Klyubin, Polani & Nehaniv (2004) promises
interesting opportunities in this sense, as their
information measure does not depend on a
specific learning rule.
ABOUT THE AUTHORS
Bernd Porr has a degree in Physics and Com-
munication Science / Journalism (both from
the University of Bochum). In 2000 he
moved to Scotland and did his PhD in
sequence learning and predictive control at
the University of Stirling. In 2004 he took up
a post as lecturer at the University of Glas-
gow at the department of electronics and
electrical engineering. His main research
interests are in the fields of biologically
inspired adaptive control, synaptic plasticity,
image processing and radical constructivism.
Alice Egerton has a degree in Pharmacology
(University of Glasgow) and a PhD in Psy-
chopharmacology (University of Strath-
clyde). Her doctoral work has focussed on
preclinical models of schizophrenia, with par-
ticular reference to behavioural pharmacol-
ogy and underlying neurobiology at the
Yoshitomi Research Institute of Neuro-
science in Glasgow (YRING). She now works
in Psychiatry Research at Imperial College
London, where her primary research interest
is the investigation of brain dopamine sys-
tems using positron emission tomography
(PET) in relation to schizophrenia and drug
dependence.
Florentin Wörgötter has studied Biology and
Mathematics in Düsseldorf. He received his
PhD in 1988 in Essen working experimentally
on the visual cortex before he turned to
computational issues at the Caltech, USA
(1988–1990). After 1990 he was researcher
at the University of Bochum concerned with
experimental and computational neuro-
science of the visual system. Between 2000
and 2005 he had been Professor for Compu-
tational Neuroscience at the Psychology
Department of the University of Stirling,
Scotland where his interests strongly turned
towards “Learning in Neurons.” Since July
2005 he leads the Department for Compu-
tational Neuroscience at the Bernstein Cen-
ter at the University of Göttingen. His main
research interest is information processing in
closed-loop perception-action systems,
which includes aspects of sensory processing,
motor control and learning/plasticity.
radical constructivism
Acknowledgements
We thank Vi Romanes and the referees for
their helpful feedback.
References
Atmanspacher, H. & Dalenoort, G. (1994)
Introduction. In: Atmanspacher, H. &
Dalenoort, G. (eds.) Inside versus outside.
Springer: Berlin, pp. 1–12.
D’Azzo, J. J. (1988) Linear control system
analysis and design. McGraw: New York.
Foerster, H. von (1960) On self-organizing
systems and their environments. In:
Yovits, M. & Cameron, S. (eds.) Self-orga-
nizing systems. Pergamon Press: London,
pp. 31–50.
Glasersfeld, E. von (1995) Radical construc-
tivism: A way of knowing and learning.
Falmer Press: London.
Hebb, D. O. (1949) The organization of
behavior: A neurophychological study.
Wiley-Interscience: New York.
Klyubin, A. S., Polani, D. & Nehaniv, C. L.
(2004) Organization of the information
flow in the perception–action loop of
evolved agents. In: Proceedings of 2004
NASA/DoD Conference on Evolvable
Hardware. IEEE Computer Society: Seat-
tle, Washington, pp. 177–180.
Kosco, B. (1986) Differential hebbian learn-
ing. In: Denker, J. S. (ed.) Neural Networks
for computing. Volume 151 of AIP confer-
ence proceedings. American Institute of
Physics: New York, pp. 277–282.
Linsker, R. (1988) Self-organisation in a per-
ceptual network. Computer 21(3): 105–
117.
Oja, E. (1982) A simplified neuron model as a
principal component analyzer. Journal of
Mathematical Biology 15(3): 267–273.
Palm, W. J. (2000) Modeling, analysis and
control of dynamic systems. Wiley: New
York.
Porr, B. (2002) Systemtheorie und Naturwis-
senschaft Eine interdisziplinäre Analyse
von Niklas Luhmanns Werk. Deutscher
Universitäts-Verlag: Wiesbaden.
Porr, B., von Ferber, C. & Wörgötter, F. (2003)
Iso-learning approximates a solution to
the inverse-controller problem in an unsu-
pervised behavioural paradigm. Compu-
tation 15: 865–884.
89
 engineering-computer scientific
SYNTHETICAL
Porr, B. & Wörgötter, F. (2003) Isotropic
sequence order learning. Neural Compu-
tation 15: 831–864.
Porr, B. & Wörgötter, F. (2005a) Non-hebbian
synaptic plasticity allows for the stable
implementation of one-shot predicitive
learning. In: Zimmermann, H. & Kriegl-
stein, K. (eds.) Proceedings of the 6th
Meeting of the German Neuroscience
Society, 30th Göttingen Neurobiology
Conference 2005, p. 1223.
Porr, B. & Wörgötter, F. (2005b) What means
embodiment for radical constructivists?
90
Kybernetes 34(1/2): 105–117.
Porr, B. & Wörgötter, F. (2006) Strongly
improved stability and faster convergence
of temporal sequence learning by utilising
input correlations only. Neural Computa-
tion 18 (in press).
Rieke, F., Warland, D., de Ruyter van
Stevenick, R. & Bialek, W. (1997) Spikes:
Exploring the neural code. MIT Press:
Cambridge MA.
Shannon, C. E. & Weaver, W. (1949) The
mathematical theory of communication.
University of Illinois Press: Urbana.
radical constructivism
Sutton, R. (1988) Learning to predict by
method of temporal differences. Machine
Learning 3(1): 9–44.
Touchette, H. & Lloyd, S. (2004) Informa-
tion-theoretic approach to the study of
control systems. Physica A 331: 140–172.
Verschure, P. & Coolen, A. (1991) Adaptive
fields: Distributed representations of clas-
sically conditioned associations. Network
2: 189–206.
Received: 17 November 2005
Accepted: 07 February 2006
Constructivist Foundations