A Stroll Through the Worlds of Animats and Humans:
Review of Being There: Putting Brain, Body and World Together Again by Andy Clark
Anthony Chemero
Department of Scientific & Philosophical Studies of Mind
Department of Psychology
Whitely Psychology Lab
Franklin and Marshall College
Lancaster, PA 17604-3003
U.S.A.
a_chemero@acad.fandm.edu
Copyright (c) Anthony Chemero 1998
PSYCHE, 4(14), October, 1998
previously Held: http://psyche.cs.monash.edu.au/v4/psyche-4-14-chemero.html
KEYWORDS: artificial intelligence, dynamical systems theory, embodiment, autonomous agents, artificial life, mental repesentation, metaphysics.
REVIEW OF: Andy Clark (1997) Being There: Putting Brain, Body and World Together Again. MIT Press: Cambridge, Massachusetts. xiii+267pp. ISBN
0-262-03240-6. Price: $US12 pbk.
1. Introduction
Every few years Andy Clark writes a book designed to help philosophers
of mind get up to speed with the most recent developments in cognitive
science. In his first two such books, Microcognition (1989) and
Associative Engines (1993), Clark introduced the then-cutting-edge
field of connectionist networks. In his newest one, Being There:
Putting Brain, Body and World Together Again (1997), he once again
provides a concise, readable introduction to the state of the art. This
time, though, Clark has moved beyond (but not abandoned) connectionism
for what he calls 'embodied, active cognition,' in which the primary
focus of study is not the inner workings of the rational thinker, but
rather the way the autonomous, embodied agent interacts with its
environment.
This conception of the mind is radically different from those we've seen
in business-as-usual cognitive science. And if it's anything like on the
right track, it will lead to a methodological and conceptual
realignment, shaking up much of what we thought we knew about the mind.
In Being There, Clark provides a remarkably clear introduction to this
new conception of the mind and then re-maps the altered conceptual
landscape. In the remainder of this review, I will outline Clark's
introduction to embodied, active cognition and his view of its
consequences for cognitive science. Along the way, I will discuss one
place where I think Clark gets things very wrong. This criticism will
concern the particulars of Clark's take on the philosophical
consequences of embodied, active cognition; it has nothing to do with
the Clark's exposition of this new type of cognitive science, which I
think is just about perfect. Anyone interested in the foundations of
cognitive science or the philosophy of mind should read this book, as
should anyone who wants to keep up with the newest and best in cognitive
science.
2. Overview of the text
Being There is divided into three major parts: "Outing the Mind,"
"Explaining the Extended Mind," and "Further". In the first part, Clark
synthesizes several seemingly disparate, but related fields, of study
into a coherent picture of the embodied mind. Starting with autonomous
agent research (e.g., Brooks, 1991) as the centerpiece, Clark then adds
dynamical systems theory models of development (e.g., Thelen & Smith,
1994), connectionist networks (discussed in Clark, 1989, 1993),
epistemic action (Kirsh & Maglio, 1994), Vygotskyan (1986) and Gibsonian
(1979) psychology, and even phenomenology (Heidegger, 1927;
Merleau-Ponty, 1945). Each of these has been taken by its proponents to
be a "paradigm shift" in its own right. Clark, however, shows that to
rival classical cognitive science as a "paradigm" for studying the mind,
one needs them all. This first section is the best part of the book. In
it, Clark introduces and patiently explains each of these fields and
their place in the overall account of the mind, an account that
integrates the body, the brain and the environment in which they are
embedded.
The first important feature of Clark's synthesis is the focus on
embodiment, the inclusion of the body in cognition. To illustrate this,
Clark begins with a discussion of the work of Rodney Brooks and his
colleagues at MIT. In their autonomous agent research, Brooks et al.
build simple creatures, called 'mobots' or 'animats,' that are capable
of robust action in a dynamic environment. The focus of their work is on
building simple, but complete agents, rather than trying to attain
human-level competence in a very limited domain (e.g., chess-playing).
For example, Brooks's mobot "Attila" can do nothing to a human-level
competence. But it can walk around on its own and avoid objects; it
thus displays a roughly complete insect-level intelligence, rather than
a small sliver of our own.
Mobots like Attila are built by melding together a group of
quasi-independent devices, each of which has a self-contained link from
perception to action. These devices excite and inhibit one another, but
pass no complex messages, allowing simple, adaptive intelligence to
emerge with no central executive controlling the action. Much of the
behavior of humans and other animals, Clark claims, is like that of
mobots in this respect. To illustrate this, Clark cites dynamical
systems studies of development (Thelen & Smith, 1991), according to
which abilities to act skillfully (e.g., walking) are "soft-assembled"
by interactions among such factors as bodily growth, environmental
factors, learning and brain maturation. So, like the mobot behavior,
human development occurs without the control of a central executive
deciding what to do when; in both cases, the behavior is self-organized, emerging over time from the interaction of several
components.
Take these mobot bodies and add connectionist-network-style brains.
There are two features of connectionist networks that are important for
Clark's purposes here. The first of them is that because connectionist
networks are parallel processors, they support a view of cognition as
highly decentralized. Just as with the mobot bodies in which they are
housed, there is no one in charge, no central executive, so the brain's
activity, like the body's, is best viewed as self-organizing. Secondly,
cognition in connectionist networks is construed as pattern
completion, instead of classical reasoning. Connectionist networks,
like most humans, are good at recognition and action, but not
particularly good at math or logic. But if, as Clark claims, we humans
just are mobots with connectionist brains, how do we do math and
logic?
The answer to this question leads to consideration of the third part of
Clark's formula: the world. The connection between a thinker and its
world, in Clark's picture, is so intimate that it is difficult to decide
where one ends and the other begins. Indeed, given the nature of the
brain and body as described, the environment must be utilized actively
in cognition. Since there is no central executive in mobots with
connectionist brains, there will be no detailed, action-neutral
representation of the world. In most cases, agents will use the world as
its own model. We can see this in animate vision research (Ballard,
1991), for example, in which an agent uses rapid, repeated saccades to
extract information from a visual scene as needs arise, rather than
building an intricate three-dimensional model of its surroundings in its
head.
To accomplish the things we humans can accomplish without building
detailed, internal models, we must rely on epistemic action and
external scaffolding. Epistemic actions (Kirsh & Maglio, 1994)
are actions taken to alter the nature of cognitive tasks. Think, for
example, of moving Scrabble tiles around on their tray to see what
words you can spell. Doing so changes the nature of the task from one
of trying to come up with words inside one's head to one of completing
patterns, a task our connectionist brains are good at. Imagine that
you have the following letters on your tray:
O G T S O S E
The easiest way to make a word is to re-arrange the actual, concrete
tiles on the tray (an epistemic action), and try to use the re-arranged
pieces as a basis for your brain's pattern-completing abilities, like
this:
ST OO S G E
which we can quickly complete as "STOOGES". In this case, the tiles act
as external scaffolding, parts of the world we rely on to aid and abet
our thought. Both these categories, epistemic action and external
scaffolding, Clark points out, are extremely large: maps, models, tools,
language and culture can all act as external scaffolding; using any of
these pieces of scaffolding, for example, writing one large number above
another to multiply them with pen on paper, is epistemic action. In all
these cases, we act so as to simplify cognitive tasks by "leaning on"
the structures in our environment. (The third part of the book "Further"
explores the ways that such structures in our environment, most notably
public language, can be utilized to allow creatures like us to
participate in more advanced cognition.)
Clark's view of the embodied, active mind--that is, of human thinkers as
connectionist brains in mobot bodies, living in a highly scaffolded
environment--is a compelling alternative to traditional
computationalism.
What makes it so compelling is the way it incorporates the best recent
empirical work in the cognitive sciences in a far-reaching view of the
mind, one that shows what is special about humans (our abilities to
construct and utilize external scaffolding to complement our "on-board"
resources), yet respects the fact that we must be evolutionarily
continuous with other animals (there is no difference in kind between
our bodies and brains and those of other animals). And it really is
different, too: problem solving is pattern completion rather than
logical inference; there is no central system or language of thought;
the roles of the body and environment have been upgraded in status from
input device and problem space, respectively, to crucial parts of the
cognitive system. Clearly, with all these changes to our image of the
mind, there must also be changes in our science of the mind. In the
second section of the book, Clark reconstructs the conceptual
foundations of cognitive science to fit his new picture of the mind.
Clark's main concern in this reconstruction is to head off the more
radical claims made by the scientists whose theories are the building
blocks of embodied, active cognition. In particular, he is opposed to
attempts to spin situated robotics (e.g., by Brooks, 1991, or Beer,
1995) or dynamical systems theory (e.g., by Thelen & Smith, 1994, or
Port & van Gelder, 1995) into anti-representationalism, the claim that views of cognition as involving mental representation (and hence computation <1>) are just mistaken. Clark argues that anti-representationalism is a case of throwing out the baby with the bath water. What we need, he says, is not a non-representational story, but a better representational story, one that leaves room for other explanatory strategies. This better story requires three interlocking types of explanation:
A dynamical systems account of the gross behavior of the agent-environment system;
An implementation account, describing how the components of the
agent-environment system interact to produce the collective properties
described in (1);
A representational and computational account of the components
identified in (2).
The third, representational, type of explanation here will be somewhat different from the usual in that representations will be geared to particular actions; they will be what Clark calls 'action-oriented'. <2> Action-oriented representations are both local and
personal: they are local in that they relate to the circumstances currently surrounding the agent; they are personal in that they are related to the agent's needs and the skills that it has. Since action-oriented representations are attached to a particular agent and its situation, they are non-objective, they are always attached to a particular point-of-view. Put most simply, they stand for what's happening to me, right here, right now.
Clark calls this overall three-tiered explanatory strategy 'minimal
representationalism', carving a central ground between traditional
computationalism, where fully-objective representations are the main
explanatory tool, and anti-representationalism, where representations
are never invoked. <3>
Clark's picture of the embodied, active mind has metaphysical
consequences that go along with the methodological ones just sketched.
The first of these, which Clark embraces, is that the mind is not
confined to the brain, or even the body. Since the agent's body and
normal environment (including, especially, external scaffolding) are
crucial to cognition, these should be considered part of the mind. As
Clark puts it, "... it may for some purposes be wise to consider the
intelligent system as a spatio-temporally extended process not limited
by the tenuous envelope of skin and skull." (p. 221) The mind, this
means, is ontologically complex, spreading out over space and time and
including bits of the world, language and social structures, along with
the brain and body.
There is another, more serious, metaphysical consequence that Clark does
not embrace. In his discussion of the relationship between his project and
a similar one described in Varela, Thompson and Rosch (1991), Clark says
the following:
Varela et al. use their reflections as evidence against realist
and objectivist views of the world. I deliberately avoid this
extension, which runs the risk of obscuring the scientific value of
an embodied, embedded approach by linking it to the problematic
idea that objects are not independent of the mind. My claim, in
contrast, is simply that the aspects of real-world structure which
biological brains represent will often be tightly geared to
specific needs and sensorimotor capacities. (p. 173)
This casual sweeping under the rug of such an important issue is the one
place in Being There where Clark lets the reader down. The non-realist
conclusions that Varela, Thompson and Rosch reach seem genuinely to
follow from Clark's picture of the mind, a picture whose acknowledged
historical precedents (Heidegger, Merleau-Ponty and von Uexkull) were
opposed to realism. To see that the embodied, active mind leads to
non-realist conclusions, consider Clark's discussion of von Uexkull's
essay "A Stroll through the Worlds of Animals and Men" (1934). <4>
Von Uexkull and Clark both suggest that we should expect creatures
(including humans) to be sensitive only to those aspects of their environments that matter to the actions they regularly undertake; their representations, to use Clark's phrasing, will be action-oriented. The world represented by animals with much different needs than humans will be much different than the world humans represent. This is the case because throughout their evolutionary histories, animals develop perceptual systems responsive to opportunities to fulfill their needs. Because the needs of one type of animal can be are so different from those of another, the perceptual systems that result will constitute the world in very different ways, as full of barbecues and highways and myriad other things for humans, but, for example, as containing only three things--what we see as butyric acid, pressure and temperature changes--for ticks (see von Uexkull, 1934, p. 10). And given the way evolution works, we should not think of the perceptual systems (or any parts of animals) as ideal solutions to problems posed by the environment. <5> Instead, animals that survive and reproduce are those that do well enough to find food and so on. So, there is no reason to assume that any particular animal's perceptual system gets the world, as it is independently of thought, just exactly right; they all do only well enough.
Since, as Clark suggests (p. 25), there is no reason to except human
perceptual systems, there is also no reason to think the world that
humans live in has any special claim on being the true reality or
world-in-itself from which ticks, poor things, know only three
aspects.
The everyday human world, far from being an objective, action-neutral
environment, is just as "tightly geared to specific [human] needs and
sensorimotor capacities" (p. 173) as the tick's world is to specific
tick needs and sensorimotor capacities. Thus, as Varela, Thompson and
Rosch suggest, the world as registered by the perceptual systems of
humans (or any other type of animal) cannot be fully independent of
those perceptual systems. If, as Clark claims, representations are
action-oriented, it follows that cognitive agents, with their
particular phylogenies and ontogenies, their physiology and education,
play an important role in constituting their worlds.
So far Clark would probably agree. But in so doing, he must also admit
the falsehood of so-called "common-sense realism," in which the
world-in-itself is thought to correspond to everyday human categories.
For if humans and ticks and rats and so on are all able to live
successfully in the very different worlds constituted by their
perceptual systems, there can be no principled reason to privilege the
human world (or any of the others) as the world-in-itself. Indeed,
cockroaches, for example, are arguably more successful than humans. Yet
no one ever claims that we should adjust our ontology so that it meshes
with the categories that cockroaches use. Thus, the ontology implied by
our usual human categories has no claim on being the
world-in-itself.
Furthermore, it is a small step to see that "scientific realism" or
"objectivism" (see Husserl, 1970), in which the world as it is described
by the mathematical idealizations of physics is taken to be the world-in-itself, is also false. Consider that Clark argues that "higher
thought," the kind exhibited in mathematical and scientific theorizing,
depends on the scaffolding provided by public language. He also suggests
(pp. 211-13) that language is adapted to the way our brains worked pre-linguistically; human language, that is, is adapted to and built upon
action-oriented representations. But, as we have seen, these
representations are biased by pressures to fulfill human needs
throughout evolutionary history. And if the foundation on which language
is built is biased, it is overwhelmingly likely that language itself is
similarly biased. So if physics and other sciences depend upon our
language-using abilities (and Clark argues that they do), they have no
claim on being reflections of the world-in-itself.
Indeed, the commonly held belief that the physical is somehow equivalent
to the world-in-itself is especially baseless: nothing but the tools
that physicists build are able to perceive things like reflectances and
protons.
Clark resists these conclusions, and with good reason. They require
changing the way most of us, as cognitive scientists, view the world.
Cognitive scientists persuaded to accept Clark's picture of the
embodied, active mind cannot do as most scientists do; that is, they
cannot just assume common-sense or scientific realism. This is
especially true when trying to assign content to the mental states of
non-human animals (not to mention states of artificially evolved neural
networks and controllers for animats). These states may frequently stand
for objects and situations that have no clear analogue among our
common-sense and scientific categories, as these categories do not
necessarily reflect anything other than human needs and sensorimotor
abilities. This is a high cost indeed. It makes the task of cognitive
science considerably more difficult, and that alone might dissuade many
members of the cognitive science community from following Clark down the
road he so skillfully builds in the first part of Being There. What
Clark needs to do is explain why his embodied, active cognition does not
have these consequences, or explain why accepting them is not such a bad
thing. Alas, he does neither. <6>
3. Conclusion
Serious though it may be, this problem seems less significant when
compared with the obvious virtues of Clark's book. Clark's patient
exposition of the newest work in cognitive science, unfamiliar to many,
and his synthesis, from that work, of a new plan for studying the mind
is truly exciting. His rebuilding of the conceptual territory is
thorough and always reasonable.
Clark's picture of the embodied, active mind is a conservative
revolution: he raises, then rejects, the most radical claims made in its
vicinity (anti-representationalism and problems for metaphysical
realism), while foundational concepts like computation, and
representation are maintained, but rethought. The third part of the
book, in which Clark speculates widely about the roles of marketing(!),
language and the social world in advanced, human-level cognition is
provocative and always entertaining.
I recommend this book to everyone interested in the sciences of the
mind. Those who are unfamiliar with this territory will be excited by
what they've been missing; those who are specialists in one of the
fields Clark lassoes into embodied, active cognition will learn how
their work fits in to a complete picture of the mind. Being There will
also work well as a textbook for courses in the foundations of cognitive
science or the philosophy of mind, as long as it is balanced by works on
more traditional cognitive science. Most importantly, though, I
recommend Clark's book to any philosopher of mind who wants to do a
better job keeping up with developments in cognitive science.
Acknowledgements
Thanks to Andy Clark and David Chalmers for helpful comments on an
earlier draft of this review.
References
Ballard, J. (1991). "Animate vision." Artificial Intelligence,
48, 57-86.
Bechtel, W. (1998). "Representations and cognitive explanations:
Assessing the dynamicist's challenge in cognitive science."
Cognitive Science, 22, 295-318.
Beer, R. (1995). "Computational and dynamical languages for autonomous
agents." In Port and van Gelder (1995).
Brooks, R. (1991). "Intelligence without reason." Artificial
Intelligence, 47, 139-159.
Chemero, A. (1998) How to be an
anti-representationalist.
Doctoral thesis in Philosophy and Cognitive Science, Indiana
University, Bloomington, IN.
Clark, A. (1989). Microcognition. Cambridge: MIT Press.
Clark, A. (1993). Associative engines. Cambridge: MIT Press.
Dennett, D. (1995). Darwin's dangerous idea. Boston: Little, Brown.
Cambridge: MIT Press.
Gibson, J. (1979). The ecological approach to visual
perception. New Jersey: Houghton Mifflin.
Heidegger, M. (1927). Sein und Zeit. (translated by Macquarrie and
Robinson, 1962) New York: Harper and Row.
Husserl, E. (1970). The crisis of the European sciences and
transcendental phenomenology. Evanston: Northwestern University
Press.
Kirsh, D. and Maglio, P. (1994). "On distinguishing epistemic from
pragmatic action." Cognitive Science, 18, 513-549.
Port, R. and van Gelder, T. (1995). Mind as motion. Cambridge: MIT
Press.
Merleau-Ponty, M. (1945). Phenomenologie de la Perception (translated
1962). London: Routledge and Kegan Paul.
Millikan, R. (1984). Language, thought and other biological
categories. Cambridge: MIT Press.
Thelen, E. and Smith, L. (1994). A dynamic systems approach to the
development of cognition and action. Cambridge: MIT Press.
van Gelder, T. and Port, R. (1995). "It's about time." In Port and van
Gelder (1995).
Varela, F., Thompson, E. and Rosch, E. (1991). The embodied
mind. Cambridge: MIT Press.
von Uexkull, J. (1934). "A stroll through the worlds of animals and
men." In Instinctive behavior, ed. K. Lashley. New York:
International Press.
Vygotsky, L. (1986). Thought and language (translation of 1962
edition). Cambridge: MIT Press.
