Full text of "Brain-wise : studies in neurophilosophy"

See other formats

Brain-Wise

Studies in Neurophilosophy

Patricia Smith Churchland

“Sound philosophy requires a solid understanding of the nature
and origin of mind, which in turn depends on the best neuro-
science available. Patricia Churchland, with verve and exactitude,
has taken a large step toward establishing that link.'

Edward 0. Wilson, University Professor Emeritus, Harvard
University

This page intentionally left blank

Patricia Smith Churchland

Brain-Wise

Studies in Neurophilosophy

A Bradford Book
The MIT Press
Cambridge, Massachusetts
London, England

All rights reserved. No part of this book may be reproduced in any form by any
electronic or mechanical means (including photocopying, recording, and information
storage and retrieval) without permission in writing from the publisher.

This book was set in Times New Roman on 3B2 by Asco Typesetters, Hong Kong, and
was printed and bound in the United States of America.

Library of Congress Cataloging-in-Publication Data

Churchland, Patricia Smith.

Brain-Wise : studies in neurophilosophy / Patricia Smith Churchland.
p. cm.

Includes bibliographical references and index.

ISBN 0-262-03301-1 (he : alk. paper) — ISBN 0-262-53200-X (pbk : alk. paper)

1. Neurosciences — Philosophy. 2. Cognitive science — Philosophy. I. Title: Studies in
neurophilosophy. II. Title.

[DNLM: 1. Neuropsychology. 2. Knowledge. Metaphysics. 4. Neurology.

5. Philosophy. 6. Religion and Psychology. WL 103.5 C563b 2002]

RC343 .C486 2002

153'.01— dc21 2002066024

10 987654321

Contents

Preface vii

Introduction 1

I Metaphysics

An Introduction to Metaphysics 37

Self and Self-Knowledge 59

Consciousness 127

Free Will 201

II Epistemology

239

An Introduction to Epistemology 241

How Do Brains Represent? 273

How Do Brains Learn? 321

III Religion 371

9 Religion and the Brain 373

Contents

Notes 403
References 421

Index 451

Preface

A lot of water has passed over the dam since I published Neurophilosophy in
1986. Groundbreaking advances have been made in computational methods,
in neuroscientific techniques, and in cross-field connections. Fruitful inter-
actions have developed, for example, between molecular biology and neuro-
science, and between experimental psychology and neuroscience. Philosophers,
initially wary (to put it politely) of the idea that neuroscience might have some
relevance to the problems they call their own, have slowly warmed to the idea
of neurophilosophy. Two decades ago, proposing an undergraduate course in
neurophilosophy was more or less a bad joke. Now such courses are beginning
to spring up even in departments that had been proudly “antibrain.” Students
not only in philosophy but also in the sciences are signing up and eagerly
attacking philosophy’s Big Problems — such as the nature of consciousness, free
will, and the self — in full recognition that neuroscientific data are indispensable
to making progress. Alert to the change in philosophical winds, various people
began to needle me concerning the absence of an introductory, single-authored
neurophilosophy text. This book is the response to that needling.

I have assumed that an introductory text should provide a basic frame-
work for how the brain sciences — the neurosciences and cognitive science —
can interface with traditional topics in philosophy. Insofar as it is elementary,
such a text should be as compact and uncluttered as is consistent with being
pedagogically serviceable. Of necessity, this means keeping in-text references to
an almost indecent minimum; it means slimming the number of suggested
readings. It means making incendiary choices about which research best illus-
trates a point and which debates are worth recounting. Although selectivity
serves the goal of presenting a fairly clean picture of how I see things, it carries
a price, not least of which is the undying wrath of colleagues who feel stiffed by
the trade-off between spare functionality and congested citation. The chips will

Preface

have to as fall they may, however, since my primary goal is that the book be
useful to those who want a panoramic view of philosophical problems as they
appear from the vantage point of the brain sciences. I could not reasonably aim
to make this book encyclopedic; I could aim to make it coherent and compact.

To be useful as an introduction, a book ought not to presuppose very much
background knowledge of the subject. I have tried to abide by that rule. What I
have presupposed is that readers totally unacquainted with neuroscience or
with cognitive science will choose a good text to have handy in case of need. To
assist in that choice, I make some general suggestions in the reading list of the
introductory chapter, and topic-specihc suggestions in subsequent chapters.
There are excellent journals, websites, and encyclopedias to augment a begin-
ner’s background, and I have also listed a subset of those journals that contain
good review papers or that are widely considered indispensable to keeping
abreast of the developments in the brain sciences.

The book contains more neurobiological detail than one would typically hnd
in a philosophy text. The rationale derives from the need to illustrate — and not
merely preach — that understanding the neuroscientific detail is no mere frill if
you intend to do more than play at philosophical problems, such as the nature
of consciousness or learning. A continuing difficulty for philosophers is to be
sufficiently versed in basic neuroscience to be able to tell whether the results of
a reported experiment mean anything, and if so, what. Though I cannot solve
this difficulty, I might reduce its size by conveying the need to understand the
experimental design, the nature of the controls, possible flaws of interpretation,
and so forth by discussing detail from selected experiments.

Though experimental detail is crucial, it is also important not to smother
one’s cognitive operations. They need time and space to mull. As philosopher
and computer scientist Brian Smith once mused, some things that brains do
very well, they do very slowly, over long stretches of time, and in a chewing-on-
the-cud sort of way. These are typically the problem-solving and creative things
that existing computers cannot do at all. In the same vein, Francis Crick
observes that if you are too busy, you are probably wasting your time. With
this thought in mind, I have reigned in my impulse to recommend readings ad
infinitum. Since those readings I do recommend reflect my particular preju-
dices, curious readers will want to go afield for other points of view.

Again and again I have found the history of science invaluable in getting my
bearings. The fact is, neuroscience is still an immature science, in the sense that
it is still groping for the fundamental explanatory principles governing brain
function. In this respect, it contrasts with molecular biology, for example,

Preface

where the basic principles of the chemical structure of genes, how genes get
turned on and off, and how proteins get made are essentially in place. Because
neuroscience is still wet behind the ears, we probably have only the vaguest
glimmerings of what remains to be discovered and no idea how the discoveries
will change our heartfelt convictions about the nature of the mind. Heartfelt
convictions, unavoidable though they may be, can be an intellectual nuisance.
They have a way of posing as nonnegotiable certainties, as verities, and as
metaphysical truths. Despite their convincing pose, they in fact are just bits of
conventional wisdom. The history of science provides bracing tales of conven-
tional wisdom as obstructing progress, as failure of imagination, and as dogma.
History also shows both that sometimes the crackpots turn out to be right, but
that being a crackpot is no guarantee of being right.

In hopes that the history might be likewise useful for others, I found myself
telling science stories where they provide a helpful slant on current problems.
These are tales about scientihc error and scientihc discovery, about scientihc
tenacity and humility, as well as about scientific arrogance and scientific obliv-
ion. Many are stories of conventional wisdom turned arse over teakettle. Their
particular relevance pertains to the search for knowledge in the broadest sense,
irrespective of the topic. By putting some distance between us and our heartfelt
convictions, these stories give us room to think. Oddly, the history of science is
seldom taught to science students, yet it is this history that helps generate a
sense of how to ask the right questions and how progress on the tough prob-
lems can be made.

Not surprisingly, I have also found the history of philosophy invaluable in
putting current philosophical orthodoxy at arm’s length. This is not because I
subscribe to the goofy theory according to which the historical giants knew
more because they knew less. I emphatically do not. Rather, it is because some
of the greats were just a whole lot broader in their interests and a whole lot
more curious about nature in general than are many of today’s mainstream
philosophers. This is manifestly true of those oldies for whom I have enduring
fondness: Aristotle, Hume, and Peirce. My fondness is also explained by the
get-on-with-it reason that they are clear and sensible, logical and bold. While
these are not virtues for cult figures, they are virtues if one is trying to under-
stand the nature of things.

In my opinion, much of what is considered not quite mainstream philosophy
is where the exciting action is now to be found in academic philosophy. This
work is enthusiastically cross-disciplinary. It leaves the borders between aca-
demic disciplines looking like the mere administrative conveniences they should

Preface

be. Philosophy students are plugging into congenial labs, while students in
neuroscience, cognitive science, and computer science are coming to realize that
philosophical questions about the mind are at bottom just broad questions
about the mind, and they can be addressed through experimental techniques.
They are also learning that philosophy is often useful in showing where the
logical minefields lie. This trend is putting blood back into philosophy, making
it much more akin to the vigorous and expansive discipline it has been through
most of its very long history. This trend is also heartening to those students
who were lured into neuroscience by the big questions but found themselves
endlessly tagging proteins.

Over the years so many people have taught me about the brain and about
how to do science that I cannot begin properly to thank them all. Let me
start, however, with Francis Crick, who has been a constant fount of ideas,
not only of predictably ingenious ones but also, occasionally at least, of reas-
suringly flawed ones. His relentlessness in addressing a problem, accompanied
by warnings to avoid falling in love with one’s own theory, gave me the pluck
to try things I might otherwise have shied away from. Additionally, Francis has
been a consistently fair-minded critic of both my enduring enthusiasms and my
ranch-hand skepticism. His knowledge of the history of science, and especially
his personal and detailed knowledge of the history of molecular biology, has
given me a perspective on neuroscience as a science that I could not have had in
any other way.

Antonio and Hannah Damasio have patiently taught me how to think about
systems-level neuroscience, and have generously shared their insights gleaned
from clinical studies. They also firmly but kindly hoisted me out of a rut into
which I had comfortably settled. In particular, they caused me to begin looking
at consciousness from the perspective of the brain’s fundamental “coherencing”
functions, as well as from its perceptual functions. In turn, this led me to follow
them to consider subcortical brain structures, especially brainstem structures,
as the anchor for coherent behavior, and hence for self-representational
capacities.

Brain-Wise also turned out to be every inch a family endeavor. Paul
Churchland, as always, shared all his hunches and insights with me, laughed at
my mistakes, and gave me broad shoulders to stand on. He also drafted many
of the illustrations. Mark Churchland and Anne Churchland, steeped in phi-
losophy as a matter of household routine and in neuroscience as a matter of
professional training, took earlier versions of the manuscript to the woodshed.
Free of any need to be polite, they repeatedly sent me back for wholesale, and

Preface

badly needed, rethinking and rewriting. Marian Churchland did me the honor
of letting loose her cartoonist’s whimsy to compose the cover, and Carolyn
Churchland gave me sensible advice for the chapter on religion. I am pro-
foundly grateful.

I the world at large, Roderick Corriveau taught me about neural develop-
ment, and added depth to chapters on representations and knowledge. My
UCSD colleague Rick Crush has been a collaborator on several projects, and
his ideas about emulators have been a central element in my thinking about
how nervous systems self-represent. I must especially thank my friend and col-
league Clark Glymour, who, with his mixture of intellectual rigor and take-no-
prisoners honesty, taught me a lot about causation and gave me the spine to
say what I really think. David Molfese went over the manuscript page by page
and consistently suggested very smart improvements, both substantive and ed-
itorial. Save for the methodical determination of David, this book would have
been forever in progress. Ilya Farber was also wonderfully helpful, both criti-
cally and in his perspective on the integration of scientihc domains. Steve
Quartz gave me ideas about brain evolution that helped reorient my thinking
about modules and brain organization generally. Michael Stack, my long-time
philosophical chum, helped me tighten up many arguments and spotted sec-
tions that sounded pompous. Terry Sejnowski kindly read some of the manu-
script and gave me advice and ideas, especially about learning and memory,
and spatial representation. My editor at the MIT Press, Alan Thwaits, gave me
the kind of invaluable advice one gets from a top-notch editor. I owe him a
large debt of thanks.

Others who read the manuscript and commented, browbeat, or encouraged
me into improvements are Bill Casebeer, Carmen Carrillo, Lou Goble, Mitch
Gunzler, Andrew Hamilton, John Jacobson, Don Krueger, Ed McAmis, and
Clarissa Waites. The Sejnowski lab at the Salk Institute is my second home,
where I can learn about the latest developments and try out ideas. I am grateful
to all those in the lab who have taken the time to bring me up to speed on their
experiments and share their speculations, doubts, and wild ideas. I have taken
the liberty of testing the manuscript on two undergraduate classes at UCSD,
and their feedback has provoked many revisions, especially in the choice of
topics to emphasize. Too numerous to mention, these students have my grati-
tude for their comments and complaints. Pippin “Bubbles” Schupbach gave me
cheerful assistance in a vast range of chores.

UCSD has been the most exciting place in the world for me during the eigh-
teen years it has been my home, and I am deeply grateful to many colleagues

Preface

for having the kindness to teach me what they know. This is especially true of
Liz Bates, Gilles Fauconnier, “Rama” Ramachandran, Marty Sereno, and
Larry Squire. Finally, I am particularly pleased to note that when he was
chancellor at UCSD, Dick Atkinson was uncommonly encouraging, even in
the early days when my work was dismissed by mainstream philosophers as not
real philosophy. As president of the University of California, he continues to
keep abreast of what “his” faculty are thinking and doing, and gives us feed-
back. He is a visionary, and I have much to thank him for.

La Jolla, California, 2002

This page intentionally left blank

Introduction

The goal of science is not to open the door to everlasting wisdom, but to set a limit on
everlasting error.

Galileo, in Galileo, by Bertolt Brecht

1 Core Questions

Bit by experimental bit, neuroscience is morphing our conception of what we
are. The weight of evidence now implies that it is the brain, rather than some
nonphysical stuff, that feels, thinks, and decides. That means there is no soul to
fall in love. We do still fall in love, certainly, and passion is as real as it ever
was. The difference is that now we understand those important feelings to be
events happening in the physical brain. It means that there is no soul to spend
its postmortem eternity blissful in Heaven or miserable in Hell. Stranger yet, it
means that the introspective inside — one’s own subjectivity — is itself a brain-
dependent way of making sense of neural events. In addition, it means that the
brain’s knowledge that this is so is likewise brain-based business.

Given what is known about the brain, it also appears highly doubtful that
there is a special nonphysical module, the will, operating in a causal vacuum to
create voluntary choices — choices to be courageous in the face of danger, or to
run away and hght another day. In all probability, one’s decisions and plans,
one’s self-restraint and self-indulgences, as well as one’s unique individual char-
acter traits, moods, and temperaments, are all features of the brain’s general
causal organization. The self-control one thinks one has is anchored by neural
pathways and neurochemicals. The mind that we are assured can dominate
over matter is in fact certain brain patterns interacting with and interpreted by
other brain patterns. Moreover, one’s self, as apprehended intro spectively and

Introduction

represented incessantly, is a brain-dependent construct, susceptible to change as
the brain changes, and is gone when the brain is gone.

Consciousness, almost certainly, is not a semimagical glow emanating from
the soul or permeating spooky stulf. It is, very probably, a coordinated pattern
of neuronal activity serving various biological functions. This does not mean
that consciousness is not real. Rather, it means that its reality is rooted in its
neurobiology. That a brain can come to know such things as these, and in par-
ticular, that it can do the science of itself, is one of the truly stunning capacities
of the human brain.

This list catalogues but a few of the scientific developments that are revolu-
tionizing our understanding of ourselves, and one would have to be naive to
suppose that things have “gone about as far as they can go.” In general terms,
the mind-body problem has ceased to be the reliably tangled conundrum it
once was. During the last three decades, the pace of discovery in neuroscience
has been breathtaking. At every level, from neurochemicals to cells, and on-
wards to the circuit and systems levels, brain research has produced results
bearing on the nature of the mind (figures 1.1 and 1.2). Coevolving with neu-
roscience, cognitive science has probed the scope of large-scale functions such
as attention, memory, perception, and reasoning both in the adult and in the
developing infant. Additionally, computational ideas for linking large-scale
cognitive phenomena with small-scale neural phenomena have opened the door
to an integration of neuroscience, cognitive science, and philosophy in a com-
prehensive theoretical framework.

There remain problems galore, and the solution to some of these problems
will surely require conceptual and theoretical innovation of a magnitude that
will surprise the pants off us. Most assuredly, having achieved significant pro-
gress does not imply that only mopping-up operations remain. But it does
mean that the heyday of unfettered and heavy-handed philosophical specula-
tion on the mind has gone the way of the divine right of kings, a passing that
has stirred some grumbling among those wearing the mantle of philosopher-
king. It does mean that know-nothing philosophy is losing ground to empiri-
cally constrained theorizing and inventive experimentation.

If the aforementioned changes have emerged from discoveries in the various
neurosciences — including neuroanatomy, neurophysiology, neuropharmacology,
and cognitive science — wherefore philosophy! What is neurophilosophy, and
what is its role? Part of the answer is that the nature of the mind (including the
nature of memory and learning, consciousness, and free will) have traditionally
been subjects within the purview of philosophy. Philosophers, by tradition.

Introduction

1 m

10 cm

1 cm

10 cm

100 ijm

1 |jm

1 nm

Figure 1.1 Organized structures are found at many spatial scales in nervous systems.
Functional levels may be even more fine-grained. Thus dendrites are a smaller compu-
tational unit than neurons, and networks may come in many sizes, including local net-
works and long-range networks. Networks may also be classed according to distinct
dynamical properties. Icons on the right depict distinct areas in the visual system (top), a
network (middle), and a synapse (bottom). (Based on Churchland and Sejnowski 1988.)

have wrestled with these topics, and the work continues. Neurophilosophy
arises out of the recognition that at long last, the brain sciences and their
adjunct technology are sufficiently advanced that real progress can be made
in understanding the mind-brain. More brashly, it predicts that philosophy of
mind conducted with no understanding of neurons and the brain is likely to be
sterile. Neurophilosophy, as a result, focuses on problems at the intersection of
a greening neuroscience and a graying philosophy.

Another part, perhaps the better part, of the answer is that philosophy, tra-
ditionally and currently, is quintessentially the place for synthesizing results
and integrating theories across disciplinary domains. It is panoramic in its
scope and all-encompassing in its embrace. It unabashedly bites off much more
than it can chew. Any hypothesis, be it ever so revered or ever so scorned, is
considered fair game for criticism. Philosophy deems it acceptable to kick the

Introduction

DISTANCE

I 1

j I I I I I i i I I I i t t

10®

10-6

10-3

10°

Nanometer

Micrometer

Millimeter

Meter

(Micron)

TIME

1 1 j .

i 1 i

1 1

1 V 1

1 <7 1

10 ®

10-3

10 °

103 |

psec

msec

sec

Microsecond

Millisecond

Second

Minute

Hour

Day

Figure 1.2 Logarithmic scales for spatial and temporal magnitudes. Brackets indicate
the scales especially relevant to synaptic processing. (Based on Shepherd 1979.)

tires of every governing paradigm, examine every sacred cow, and peer behind
the curtains of every magic show.

Under this description, we are all philosophers from time to time. Certainly,
scientists have their philosophical hours, when they push back from the bench
and stew on the broad questions, or when they beat on the conventional wis-
dom and strike a blow for originality. Such philosophical hours prepare the
ground for the germination of new ideas and new experimental techniques.

Politely, we can consider philosophy the theoretical companion to experi-
mental science; less politely, we can consider it merely woolgathering and free-
lancing. Certainly, some philosophy w just horsing around. Yet that is no bad
thing, especially when a science is in its nascent stages. Neuroscience is a
nascent science, and theoretical innovation is needed in every subfield of that
broad fiber-field. Most theoretical ideas are bound to be losers, of course, but
unless we are courageous enough to nurture lots and lots of new ideas, the
rightful winners will never see the light of day.

This description highlights the positive side of philosophizing, but as with
anything else, there is a seamier side. This is the side revealed when one is lulled
into taking one’s untested theoretical fancies as fact, or equating theory beau-
tiful with theory true, or rejecting unorthodox ideas as heresy because they are
unorthodox, or supposing that some chummy circle has the corner on clever
ideas. If this applies to philosophy, it applies just as well to science, govern-
ment, finance, and war.

Introduction

This book is about neurophilosophy. It aims to take stock of various philo-
sophical problems concerning the nature of the mind, given the recent bonanza
of developments in neuroscience and cognitive science. In finding a path
through the thicket of relevant neuroscientific studies and discoveries, I found
material assembling itself into two classical categories: metaphysics and epis-
temology. Ethics gets a brief look in my discussion of free will and responsi-
bility, but is mainly undiscussed on this occasion. Religion is the subject of the
closing chapter, and has both a metaphysical and an epistemological dimension.

Before plunging on, we shall limber up with a few brief historical points and
a short discussion on reductionism, a pivotal concept whose clarity is no luxury
as we begin to assay the integration of hitherto separated domains. ^

2 Natural Philosophy

Greek thought in the period 600 b.c. to 200 a.d. was the fountainhead for
Western philosophy generally, as well as for modern science. In those days,
philosophy literally meant “love of wisdom,” and for the ancient Greeks, phi-
losophy targeted a vast range of questions, such as. What is the nature of
change such that water can freeze or wood burn? What is the nature of the
moon and stars, and where did Earth come from? Are there fundamental par-
ticles of which all objects are composed? How do living things reproduce? In
addition, of course, they raised questions about themselves — about what it is to
be human, to think and perceive, to reason and feel, to plan and decide, to live
a good life, to organize a harmonious and productive political state.

Theories about the natural world were considered part of natural philosophy.
By contrast, theories of ethics and politics and practical life were part of moral
philosophy. To a first approximation, this classification separates questions
about how things are from questions about what we should do. Though distinct,
these two domains share concepts and theories. In particular, sometimes ques-
tions about the mind will have one foot in each of these areas.

When did philosophy come to be considered a separate discipline? By the
end of the nineteenth century, advances in some domains of natural philosophy
had developed so extensively that separate subfields — physics, chemistry, as-
tronomy and biology — branched off as distinct sciences. With progress and
specialization, the expression “natural science” gained currency, while the more
old-fashioned term, “natural philosophy” faded from use, now being essentially
archaic. Nonetheless, this broad title can still be found on science buildings and

Introduction

doorways in older universities such as Cambridge in England and St. Andrews
in Scotland. Until the middle of this century, St. Andrews’s degrees in physics
were olRcially degrees in Natural Philosophy. The title Ph.D. {Philosophae
Doctor, or “teacher of philosophy”) is awarded not only to philosophers, but to
scientists of all sorts. It is a vestige of the older classification, which embraces
all of science as a part of natural philosophy.

If the stars, the heart, and the basic constituents of matter became under-
stood well enough to justify a separate science, what about the mind? Ancient
thinkers, such as the physician Hippocrates (460-377 b.c.), were convinced that
thoughts, feelings, and perceptions were activities of the brain. He believed that
events such as sudden paralysis or creeping dementia had their originating
causes in brain damage. And this implied, in his view, that normal movement
and normal speech had their originating causes in the well-tempered brain. On
the other hand, philosophers favoring a nonnatural framework — Plato (427-
347 B.C.), and especially later Christian thinkers such as St. Thomas Aquinas
(1225-1274) and St. Augustine (354-430) — believed the soul to be distinct
from the body and divine in origin. Plato, in perhaps the first systematic theo-
rizing on the soul, hypothesized it to have a sensible part (which determines
perceptions), an emotional part (by virtue of which we feel honor, fear, and
courage), and a rational part. This last was considered unique to humans and
allowed us to reason, think, and figure things out. Theologically minded phi-
losophers concluded that the mind (or, one might say, the soul) was a subject
for study by means other than those available to natural science. If super-
naturalism was true of the soul, then the nature of the soul could not be
revealed by natural science, though perhaps other methods — such as medita-
tion, introspection, and reason — might be useful.

Descartes (1595-1650) articulated the modern version and systematic de-
fense of the idea that the mind is a nonphysical thing. This dual-substance view
is known as dualism. Reason and judgment, in Descartes’s view, are functions
inhering in the mental, immaterial mind. He surmised that the mind and the
body connect at only two points; sensory input and output to the muscles.
Apart from these two functions, Cartesian dualism assumes that the mind’s
operations in thought, language, memory retrieval, reflection and conscious
awareness proceed independently of the brain. When clinical studies on brain-
damaged patients showed clear dependencies between brains and all these os-
tensibly brain-independent functions, classical dualism had to be reconfigured
to allow that brain-soul interactions were not limited to sensory and motor

Introduction

functions. Achieving this correction without rendering the soul explanatorily
redundant has been the bane of post-Cartesian dualism.

What about dualism appealed to Descartes? First, he was particularly im-
pressed by the human capacity for reasoning and language, and the degree to
which language use seems to be governed by reasons rather than causes. More
exactly, he confessed that he was completely unable to imagine how a me-
chanical device could be designed so as to reason and use language appropri-
ately and creatively.

What sort of mechanical devices were available to propel Descartes’s imag-
ination? Only clockwork machines, pumps, and fountains. Though some of
these were remarkably clever, even the most elaborate clockwork devices of the
seventeenth century were just mechanical. Well beyond the seventeenth-century
imagination are modern computers that can guide the path of a cruise missile
or regulate the activities of a spacecraft on Mars. In an obvious way, Des-
cartes’s imagination was limited by the science and technology he knew about.
Had he been able to contemplate the achievements of computers, had he had
even an inkling of electronics, his imagination might have taken wing. On the
other hand, the core of Descartes’s argument was revived in the 1970s by
Chomsky^ and Fodor^ to defend their conviction that nothing we will ever
understand about the brain will help us very much to understand the nature of
language production and use.

The second reason dualism appealed is closely connected to the first. Des-
cartes was convinced that exercise of free will was inconsistent with causality.
He was also sure that humans did indeed have free will, and that physical
events were all caused. So even if the body was a just a mechanical device, the
mind could not be. Minds, he believed, must enjoy uncaused choice. We can
undertake an action for a reason, but the relations between reasons and choices
are not causal. Animals, by contrast, he believed to be mere automata, without
the capacity for reason or for free choice. In its core, if not in its details, this
argument too is alive and well even now, and it will be readdressed in greater
detail in chapter 5 in the context of the general topic of free will.

Third, Descartes was impressed by the fact that one seems to know one’s
own conscious experiences simply by having them and attending to them. By
contrast, to know about your experiences, I must draw inferences from your
behavior. Whereas I know I have a pain simply by having it, I must draw an
inference to know that my body has a wound. I cannot be wrong that / am
conscious, but I can be wrong that you are conscious. I can even be wrong that

Introduction

you exist, since “you” might be nothing but my hallucination. According to
Descartes’s argument, differences in how we know imply that the thing that has
knowledge — the mind — is fundamentally different from the body. The mind,
he concluded, is essentially immaterial and can exist after the disintegration of
the body. Like the other two arguments for dualism, this argument has
remained powerful over the centuries. It has been touched up, put in modern
dress, and in general reworked to look as good as new, but Descartes’s insights
regarding knowledge of mental states constitute the core of virtually all recent
work on the nonreducibility of consciousness.'^ Because it continues to be per-
suasive, this argument will be readdressed and analyzed in detail when we dis-
cuss self-knowledge and consciousness. (See especially chapter 3, but also
chapters 4 and 6.)

How, in Descartes’s view, is the body able causally to affect the mind so that
I feel pain when touching a hot stove? How can the mind affect the body so
that when I decide to scratch my head, my body does what I intend it should
do? Although Descartes envisioned interaction as limited to sensory input and
motor output, notice that the business of interaction — any interaction — turns
out to be a vexing problem for dualism, no matter how restricted or rich the
interactions are believed to be. The interaction problem was, moreover, recog-
nized as trouble right from the beginning. How could there be any causal
interaction at all, was the question posed by other philosophers, including his
contemporary. Princess Elizabeth of Holland, who put her objection bluntly in
a letter of 10/20 June 1643: “And I admit that it would be easier for me to
concede matter and extension to the soul than to concede the capacity to move
a body and be moved by it to an immaterial thing” (Oeuvres de Descartes, ed.
C. Adam and P. Tannery, vol. Ill, p. 685). As Princess Elizabeth realized, the
mind, as a mental substance, allegedly has no physical properties; the brain, as
a physical substance, allegedly has no mental properties. Slightly updated, her
question for Descartes is this: how can the two radically different substances
interact? The mind allegedly has no extension, no mass, no force fields — no
physical properties at all. It does not even have spatial boundaries or locations.
How could a nonphysical thing cause a change in a physical thing, and vice
versa? What could be the causal basis for an interaction? Somewhat later,
Leibniz (1646-1716) described the problem as intractable:^ “When I began to
meditate about the union of soul and body, I felt as if I were thrown again into
the open sea. For I could not find any way of explaining how the body makes
anything happen in the soul, or vice versa, or how one substance can commu-
nicate with another created substance. Descartes had given up the game at this

Introduction

point, as far as we can determine from his writings” (from A New System of
Nature, translated by R. Ariew and Daniel Garber, p. 142).

Descartes almost certainly did recognize that mind-body interaction was a
devastating difficulty, and indeed it has remained a stone in the shoe of dualism
ever since. (For additional discussion, see chapter 2.)

The difficulty of giving a positive account provoked some philosophers,
Leibniz being the first, to assert that events in a nonphysical mind are simply
separate phenomena running in parallel to events in the brain. The mind causes
nothing in the brain, and the brain causes nothing in the mind. Known as psy-
chophysical parallelism, the idea was that the parallel occurrence of mental and
brain events gives the illusion of causal interaction, though in fact no such
causation ever actually occurs. What keeps the two streams in register? Some
parallelists, such as Malebranche, thought this was a job God regularly and
tirelessly performs for every conscious subject every waking hour. Leibniz, who
preferred the idea that God kicked off the two streams and then let them alone,
disparaged “occasionalists” such as Malebranche: “[Descartes’s] disciples . . .
judged that we sense the qualities of bodies because God causes thoughts to
arise in the soul on the occasion of motions of matter, and that when our soul,
in turn, wishes to move the body, it is God who moves the body for it” (p. 143).

Descartes’s best attempt to explain the interaction between mind and body
was the suggestion that some unobserved but very, very fine material — material
— in the pineal gland of the brain brokered the interaction between nonphysical
mind and physical brain. His critics, such as Leibniz, were not fooled.

Perhaps Descartes was not fooled either. Some historians argue that Des-
cartes’s defense of a fundamental difference between mind and body was actu-
ally motivated by political rather than intellectual considerations.’ Descartes
was unquestionably a brilliant scientist and mathematician. This is, after all,
the Descartes of the Cartesian coordinate system, a stunning mathematical
innovation for which he is rightly given credit. He also understood well the
bitter opposition of the Church to developments in science, and had left France
to live in Holland to avoid political trouble. It is possible that he feared that
developments in astronomy, physics, and biology would be cut off at the knees
unless the Church was reassured that the “soul” was its unassailable propri-
etary domain. Such a division of subject matter might permit science at least to
have the body as its domain. Whether this interpretation does justice to the
truth remains controversial.

Certainly some of Descartes’s arguments, both for the existence of God as
well as for the mind/body split, are sufficiently flawed to suggest that they are

Introduction

ostentatiously flawed. On this hypothesis, the genius Descartes knew the logic
full well and planted the flaws as clues for the discerning reader. And certainly
Descartes had good reason to fear the Church’s power to thwart scientiflc in-
quiry and to punish the scientist. Burning, torturing, and exiling those who
inquired beyond official Church doctrine was not uncommon. Galileo, for ex-
ample, was “shown the instruments of torture” to force him to retract his claim
that Earth revolved around the Sun, a claim based on observation and reason-
ing. Recant he did, rather than submitting to the rack and iron maiden, but
even so, he spent the rest of his life under house arrest by Church authorities.
By vigorously postulating the mind/body division, perhaps contrary to his own
best scientific judgment, Descartes may have done us all a huge, if temporary,
favor in permitting the rest of science to go forward.

And go forward it did. By the end of the nineteenth century, physics, chem-
istry, astronomy, geology, and physiology were established, advanced scientific
disciplines. The science of nervous systems, however, was a much slower affair.
Though some brilliant anatomical work had been done on nervous systems,
particularly by Camillo Golgi (1843-1926) and Santiago Ramon y Cajal
(1852-1934), even at the end of the nineteenth century, little was known about
the brain’s functional organization, and almost nothing was understood con-
cerning how neurons worked. That neurons signaled one another was a likely
hypothesis, but how and to what purpose was a riddle.

Why did progress in neuroscience lag so far behind progress in astronomy
or physics or chemistry? Why is the blossoming of neuroscience really a late-
twentieth-century phenomenon? This question is especially poignant since, as
noted, Hippocrates some four hundred years b.c. had realized that the brain
was the organ of thought, emotion, perception, and choice.

The crux of the problem is that brains are exceedingly difScult to study.
Imagine Hippocrates observing a dying gladiator with a sword wound to the
head. The warrior had lost fluent speech following his injury, but remained
conscious up to the end. At autopsy, what theoretical resources did Hippo-
crates possess to make sense of something so complex as the relation between
the loss of fluent speech and a wound in the pinkish tissue found under the
skull? Remember, in 400 b.c. nothing was understood about the nature of
the cells that make up the body, let alone of the special nature of cells that
make up the brain. That cells are the basic building blocks of the body was not
really appreciated until the seventeenth century, and neurons were not seen
until 1837, when Purkyne, using a microscope, first saw cell bodies in a section
of brain tissue (figure 1.3).® Techniques for isolating neurons — brain cells — to

Introduction

Figure 1.3 A cross-section through the mink visual cortex, with cresyl violet used
to stain all cell bodies. Cortical layers are numbered at the right. (Courtesy of S.
McConnell and S. LeVay.)

Introduction

Figure 1.4 A drawing of Golgi-stained neurons in the rat cortex. About a dozen pyra-
midal neurons are stained, a tiny fraction of the neurons packed into the section. The
height of the section depicted is about 1 mm. (Based on Eccles 1953.)

reveal their long tails and bushy arbors were not available until the second
half of the nineteenth century, when stains that filled the cell were invented by
Deiters (carmine stain) and then Golgi (silver nitrate stain) (figure 1.4). Neu-
rons are very small, and unlike a muscle cell, each neuron has long branches —
its axon and dendrites. There are about a 10^ neurons per cubic millimeter of
cortical tissue, for example, and about 10^ synapses. (A handy rule of thumb is
about 1 synapse/pm^.) Techniques for isolating living neurons to explore their
function did not appear until well into the twentieth century.®

By contrast, Copernicus (1473-1543), Galileo (1564-1642), and Newton
(1643-1727) were able to make profound discoveries in astronomy without
highly sophisticated technology. Through a clever reinterpretation of tradi-
tional astronomical measurements, Copernicus was able to figure out that
Earth was not the center of the universe, thus challenging geocentrism. With
a low-tech telescope, Galileo was able to see for the first time the moons of
Jupiter and the craters of our own moon, thus undermining the conventional
wisdom concerning the absolute perfection of the Heavens and the uniqueness
of Earth.

Introduction

Figuring out how neurons do what they do requires very high-level tech-
nology. And that, needless to say, depends on an immense scientific infra-
structure; cell biology, advanced physics, twentieth-century chemistry, and
post- 1953 molecular biology. It requires sophisticated modern notions like
molecule and protein, and modern tools like the light microscope and the elec-
tron microscope, and the latter was not invented until the 1950s. Many of the
basic ideas can be grasped quite easily now, but discovering those ideas
required reaching up from the platform of highly developed science.

To have a prayer of understanding nervous system, it is essential to under-
stand how neurons work, and that was a great challenge technically. The most
important conceptual tool for making early progress on nervous systems was
the theory of electricity. What makes brain cells special is their capacity to sig-
nal one another by causing fast microchanges in each others’ electrical states.
Movement of ions, such as Na+, across the cell membrane is the key factor in
neuronal signaling, and hence in neuronal function. Living as we do in an
electrical world, it is sobering to recall that as late as 1 800, electricity was typi-
cally considered deeply mysterious and quite possibly occult. Only after dis-
coveries by Ampere (1775-1836) and Faraday (1791-1867) at the dawn of the
nineteenth century was electricity clearly understood to be a physical phenom-
enon, behaving according to well-defined laws and capable of being harnessed
for practical purposes. As for neuronal membranes and ions and their role in
signaling, understanding these took much longer (figures 1.5 and 1.6).

Once basic progress was made on how neurons signal, it could be asked what
they signal; that is, what do the signals mean. This question too has been
extremely hard to address, though the progress in the 1960s correlating the re-
sponse of a visual-system neuron to a specific stimulus type, such as a moving
spot of light, opened the door to the neurophysiological investigation of sen-
sory and motor systems, and to the discovery of specialized, mapped areas.

Beginning in the 1950s, progress had been made in addressing learning and
memory at the systems level, and by the late 1970s, intriguing data on neuronal
changes mediating system plasticity permitted the physiology of learning and
memory to really take olf. Meanwhile the role of specific neurochemicals in
signaling and modulating neuronal function was beginning to be unraveled,
and associated with large-scale effects such as changes from being awake to
being asleep, to memory performance, to pain regulation, and to pathological
conditions such as Parkinson’s disease and obsessive-compulsive disorder. By
the 1980s, attention functions came within the ambit of neuroscience, and
changes at the neuronal level could be correlated with shifts in attention.

Introduction

Figure 1.5 Neurons have four main structural regions and five main electrophysio-
logical functions. The dendrites (2) have little spines (1) projecting from them, which are
the major sites of in-coming signals from other neurons. The soma (3) contains the cell
nucleus and other organelles involved in cell respiration and polypeptide production.
Integration of signals takes place along the dendrites and soma. If signal integration
results in a sufficiently strong depolarization across the cell membrane, a spike will be
generated on the membrane where the axon emerges and will be propagated down the
axon (4). Spikes may also be propagated back along dendritic membrane. When a spike
reaches the axon terminal, neurotransmitter may be released into the synaptic cleft (5).
The transmitter molecules diffuse across the cleft and some bind to receptor sites on the
receiving neuron. (Adapted from Zigmond et al. 1999.)

Introduction

Figure 1.6 In the neuron’s resting state (1), both the sodium (Na+) and potassium
(K+) channels are closed, and the outside of the cell membrane is positively charged
with respect to the inside. Hence there is a voltage drop across the membrane. If the
membrane is depolarized (2), sodium ions enter the cell until the cell’s polarity is
reversed; that is, the inside of the cell is positively charged with respect to the outside. In
the repolarization phase (3), the potassium channel then opens to allow eflux of potas-
sium ions, the sodium gate closes, and sodium ions are actively pumped out of the cell.
All of these activities help bring the membrane back to its resting potential. Because the
potassium gate does not close as soon as the resting potential is reached (4), the voltage
drop across the membrane briefly drops a little below the resting voltage. Equilibrium is
reached once the resting potential is restored. (Based on Campbell 1996.)

Introduction

c D

Figure 1.7 Examples of figures with subjective eontours. Each of (A) through (C)
seems to have a border (luminance eontrast) where none exists. The borders are induced
by line terminations that are consistent with the existence of an occluding figure. Thus
the tapered ends in (D) do not give rise to a subjective contour. (From Palmer 1999.)

Progress on all these cognitive functions required adapting human psycho-
physical experiments, such as detection of illusory contours, to animals such as
monkeys and cats (figure 1.7). In the animal studies, the responses of individual
neurons under highly constrained conditions could be determined in order to
test for sensitivity to a stimulus or a task (figure 1.8). And while cognitive
functions at the network and neuronal level were being explored, details con-
tinued pour in to update the story of the ultrastructure of neurons — their syn-
apses, dendrites, and gene expression within the nucleus — and how cognitive
function was related to various ultrastructural operations.

Nevertheless, many fundamental questions about how the nervous system
works remain wide open. In particular, bridging the gap between activity in
individual neurons and activity in networks of neurons has been difficult.
Macrolevel operations depend on the orchestrated activity of many neurons in
a network, and presumably individual neurons make somewhat different con-
tributions in order for the network to achieve a specific output, such as recog-

Introduction

Time (ms)

Figure 1.8 Neurons in owl visual forebrain areas (visual Wulst) respond to subjective
contours about as well as to a real contour. The four eontours (a) to (d) were randomly
presented to the owl until each was viewed 15 times. The left column illustrates the
stimuli; the right column shows the corresponding dot-raster displays for several pre-
sentations. Black dots represent the oecurrences of spikes. Arrows indicate the direction
of motion of the contours (motion onset at 0 ms). Notice that the neuron responds
poorly in (d), where there is no subjeetive contour, but responds as well to (b) and (c) as
to (a), the real eontour. (Reprinted with permission from Nieder and Wagner 1999.
Copyright by the American Association for the Advancement of Scienee.)

nition of visual motion or a command to move the eyes to a specific location.
Moreover, understanding the dynamics of patterns of activity in neural net-
works and across many networks is undoubtedly essential to understanding
how integration and coherence are achieved in brains. For example, there ap-
pear to be “competitions” between networks as the brain settles on a decision
whether to fight or flee, and if to flee, whether to run in this direction or that,
and so on. We are just beginning to feel our way toward concepts that might be
helpful in thinking about the problems of coherencing. ^ ^

Until very recently, neuronal responses could be probed only one neuron at a
time, but if we cannot access many neurons in a network, we have trouble fig-
uring out how any given neuron contributes to various network functions, and
hence we have trouble understanding exactly how networks operate. Significant
technical progress has been made in recording simultaneously from more than
one neuron, and the advent of powerful computers has made the problems

Introduction

of data analysis somewhat more tractable. Nevertheless, the search is on for
technical breakthroughs that will really mesh microlevel experimentation with
systems-level data. We are also uncertain how to identify what, among the bil-
lions of neurons, constitutes one particular network, especially since any given
neuron undoubtedly has connections to many networks, and networks are
likely to be distributed in space. To make matters yet more interesting, what
constitutes a network may change over time, through development, and even
on very short time scales, such as seconds, as a function of task demands.
Obviously, these problems are partly technical, but they are also partly con-
ceptual, in the sense that they require innovative concepts to edge them closer
to something that can motivate the right technological invention for neuro-
biological experiments.

The advent of new safe techniques for measuring brain activity in humans
has resulted in increasing numbers of fruitful collaborations between cognitive
scientists and neuroscientists. When the results of techniques such as functional
magnetic resonance imaging (fMRI)^^ and positron emission tomography
(PET)^^ converge with results from basic neurobiology, we move closer to an
integrated mind-brain science (figure 1.9). These techniques can show some-
thing about the changes in regional levels of activity over time, and if set up
carefully, the changes can track changes in cognitive functions. It is important
to understand that none of the imaging techniques measure neuronal activity
directly. They track changes in blood flow (hemodynamics). Because the evi-
dence suggests that localized increases in blood flow are a measure of local
increases in neuronal activity (more active neurons need more oxygen and more
glucose), they are believed to be an indirect indication of changes in levels of
activity in the local neuronal population. Note also that the recorded changes
are insensitive to what individual neurons in a region are doing. The best
spatial resolution of PET is about 5 mm, and in fMRI it is about 2 mm,
though these resolutions may improve. Since one mm^ of cortex contains about
100,000 neurons, the spatial resolution of these techniques does not get us very
close to single-neuron activity.

If the images from scanning techniques reflect changes across time, one con-
ceptual problem concerns how to interpret the changes, and that means figur-
ing out what should count as the baseline activity in any given test. Suppose
that a subject is awake and alert, and is given a task, for example, visually
imaging moving his hand. How do we characterize the state before he is to
begin the task? We ask the subject to just rest. But his brain does not rest. His
brain will be doing lots of things, including making eye movements, monitoring

Introduction

Size

(log mm) ^ _

3 -

Brain

2 -

Areas 1 -

0 -

Column

-1 -

Neuron

-2 -

Dendrite ^
Synapse ^

-5 -

Lesions

Microlesions

Optical imaging

Single unit

Patch clamp
Light microscopy

Molecule

Pharmacology

-3 -2 -1 0 1 2 3 4 5 6

Millisec. Sec. Min. Hr. Day

Time (log seconds)

Figure 1.9 Comparison of the temporal and spatial resolutions of various brain-
mapping techniques. MEG indicates magnetoencephalography; ERP, evoked response
potential; EROS, event-related optical signal; MRI, magnetic resonance imaging; fMRI,
functional MRI; PET, positron emission tomography; and 2-DG, 2-deoxyglucose.
(Adapted from Churchland and Sejnowski 1988.)

glucose levels, perhaps thinking about missing breakfast, feeling an itch in his
scalp, maintaining posture, and so forth. The subject cannot command the
cessation of all cognitive functions, and certainly not all brain functions.

The problem of the baseline was recognized right from the beginning, and
various strategies for reducing confounds have been developed, especially by
Michael Posner and his colleagues. These involve subtracting the level of
activity in the “rest” condition from the level in the task condition, to reveal
the difference made, presumably, by the task. There are other problems in get-
ting meaningful interpretations of image data. For example, if a region shows
increased activity during a cognitive task, does that mean it is specialized for
that task? At most, it probably shows that the region has some role in executing
the task, but this is a much weaker conclusion. Performance of the task may

Introduction

involve a fairly widely distributed network, and the noticed change may reflect
a local blip in which one segment of the network happens to have a high den-
sity of neurons that contribute, though collectively other low-density regions
may be more important to the execution of the function. Until we know more
about brain organization at the neuronal and network levels, some of these
problems in interpretation will persist.

These cautionary remarks regarding interpretation of image data should not
be taken to imply that the new imaging techniques are too problematic to be
useful. They are in fact very useful, but experiments do have to be carefully con-
trolled so as to reduce confounds, and conclusions have to be carefully stated
to avoid exaggerated claims. It is relatively easy to get image data, but very
difficult to know whether the data reveal anything about brain function and
organization. The main point is that the imaging techniques are indeed mar-
velous and are indeed useful, but not all imaging studies yield meaningful
results. What we want to avoid is drawing strong conclusions about localiza-
tion of function when only weak conclusions or no conclusions are warranted.

3 Reductions and Coevolution in Scientific Domains

The possibility that mental phenomena might be understood in a neuro-
scientiflc framework is associated with reductive explanation in science gen-
erally. An example where one phenomenon is successfully reduced to another is
the reduction of heat to molecular kinetic energy. In this case, the prereductive
science was dealing with two sets of phenomena (i.e., heat and energy of mo-
tion), and had a good deal of observational knowledge about each. It was not
initially obvious that heat had anything at all to do with motion, which seemed
a wholly separate and unrelated phenomenon. As it turned out, however, they
have quite a lot to do with each other, initial appearances notwithstanding.

An understanding of mental phenomena — such as memory, pains, dreaming,
and reasoning — in terms of neurobiological phenomena is a candidate case of
reduction, inasmuch as it looks reasonable to expect that they are brain func-
tions. Because the word “reduction” can be used in wildly different ways,
ranging from an honorific to a term of abuse, I now outline what I do and do
not mean by “reduction.”^®

The baseline characterization of scientific reduction is tied to real examples
in the history of science. Most simply, a reduction has been achieved when the
causal powers of the macrophenomenon are explained as a function of the phys-

Introduction

ical structure and causal powers of the microphenomenon. That is, the macro-
properties are discovered to be the entirely natural outcome of the nature of the
elements at the microlevel, together with their dynamics and interactions. For
example, temperature in a gas was reduced to mean molecular kinetic energy}''

Does a reduction of a macrotheory to a microtheory require that the key
words of the macrotheory mean the same as the words referring to the micro-
properties? Not at all. A common misunderstanding, especially among philos-
ophers, is that if macrotheory about a is reduced to microtheory features ji, y, 8,
then a must mean the same as and y and 8. Emphatically, this is not a re-
quirement, and has never been a requirement, in science. In fact, meaning
identity is rarely, if ever, preserved in scientific identifications. Temperature of
a gas is in fact mean molecular kinetic energy, but the phrase “temperature of a
gas” is not synonymous with “mean molecular kinetic energy.” Most cooks are
perfectly able to talk about the temperature of their ovens without knowing
about anything about the movement of molecules. Second, it often happens
that as the macrotheory and the microtheory coevolve, the meanings of the
terms change to better mesh with the discovered facts. The word “atom” used
to mean “indivisible fundamental particle.” Now we know atoms are divisible,
and “atom” means “the smallest existing part of an element consisting of a
dense nucleus of protons and neutrons surrounded by moving electrons.”^®
Usually, the meaning change is first adopted within the relevant scientific com-
munity and propagates more widely thereafter.

What does the history of science reveal about reductive explanations that
might be helpful in understanding what a reduction of psychology to neuro-
science will entail? A nagging question about the connection between cognition
and the brain is this: can we ever get beyond mere correlations to actual iden-
tification and hence reduction? If so, how? Let us try to address this question by
briefly discussing three cases. The first concerns the discovery that the identifi-
cation of temperature of a gas with the mean kinetic energy of its constituent
molecules permits thermal phenomena such as conduction, the relation of
temperature and pressure, and the expansion of heated things to get a coherent,
unified explanation. Correlations give you reasons for testing to evaluate the
explanatory payoff from identification, but without explanatory dividends,
correlations remain mere correlations. In the case of thermal phenomena, the
first explanatory success with gases allowed the extension of the same explana-
tory framework to embrace liquids and solids, and eventually plasmas and even
empty space. As a theory, statistical mechanics was far more successful than
the caloric theory, the accepted theory of heat in the nineteenth century. Let us

Introduction

look at little more closely at how people came to realize that temperature was
actually molecular motion.

It is very natural to think of heat as a kind of stuff that moves from
hot things to cold things. As natural philosophers investigated the nature of
changes in temperature, they gave the name “caloric” to the stuff that presum-
ably made hot things hot. Caloric was thought to be a genuine fluid — a funda-
mental stuff of the universe, along with atoms, and existing in the spaces
between atoms. When Dalton (1766-1844) proposed his atomic theory, his
sketches of tiny atoms showed them as surrounded by tiny atmospheres of
caloric fluid. Within this framework, a hot cannonball was understood to have
more caloric than a cold cannonball; snow has less caloric than steam.

Given that caloric is a kind of fluid, this entails that a thing should weigh
more when hot than when cold. Weighing a cannon ball before and after heat-
ing tested this theory. The results showed that no matter how hot the cannon
ball became, its weight remained the same. Faced with a possible refutation
of a very plausible theory (what else could heat be?), some scientists were
tempted by the hypothesis that caloric fluid was very special in that it had no
mass.

Heat created through friction was also a puzzle, because there was no evident
fluid source of caloric. The conventional wisdom settled on the idea that rub-
bing released the caloric fluid that was normally sequestered in the spaces be-
tween atoms. Rubbing jostled the atoms, and the jostling allowed the caloric to
escape. To test the solution to the friction puzzle. Count Rumford Benjamin
Thompson (1753-1814) traveled from England to a factory in Bavaria that
bored holes in iron cannons. The boring, of course, continuously produced a
huge amount of heat through friction, and the cannons under construction
were constantly cooled by water. Rumford reasoned that if caloric fluid was
released by friction during boring, then the caloric should eventually run out.
No additional heat should be produced by further boring or rubbing. Needless
to say, he observed that heat never ceased to be produced as the holes down the
cannon shaft were continuously bored. At no point did the caloric fluid in the
iron show the slightest sign of depletion.

Either there was an infinite amount of this allegedly massless fluid in the iron,
or something was fundamentally wrong with the whole idea of caloric. Rum-
ford realized that the flrst option was not seriously believable. Were it true,
even one’s hands would have to contain an infinite amount of caloric, since you
can keep rubbing them without decline in heat production. Rumford concluded
that not only was caloric fluid not a fundamental kind of stuff, it was not a stuff

Introduction

of any kind. Heat required a different sort of explanation altogether. Heat, he
proposed, just is micromechanical motion.^®

Notice that a really determined calorist could persist in the face of Rum-
ford’s experiments, preferring to try to develop the option that every object
really does contain an infinite amount of (massless) caloric fluid. And un-
doubtedly some believers did persist well after Rumford’s presentation. The
possibility of such persistence shows only that refutations of empirical theories
are not as straightforward as refutations of mathematical conjectures. The
caloric-fluid theory of heat was eventually rejected because its fit with other
parts of science slowly became worse rather than better, and because, in the
explanatory realm, it was vastly outclassed in explanatory and predictive power
by the theory that heat is a matter of molecular motion. The fit of the newer
theory with other parts of science, moreover, became better rather than worse.
These developments also led to the distinction between heat (energy transfer
as a result of difference in temperature) and temperature (movement of
molecules).

The explanation of the nature of light can be seen as another successful ex-
ample of scientific reduction. In this instance, visible light turned out to be
electromagnetic radiation (EMR), as did radiant heat, x-rays, ultraviolet rays,
radio waves, and so forth (see plate 1). Note also that in these examples, as in
most others, further questions always remain to be answered, even after the
reductive writing is on the wall. Hence, there is a sense in which the reduction
is always incomplete. If the eore mysteries are solved, however, that is usually
sufficient for scientists to consider an explanation — and hence a reduction — to
be well established and worthy of acceptance as the basis for further work.

Reductions can be very messy, in the sense that the mapping of properties
from micro to macro can be one-many or even many-many, rather than the
ideal one-one. While the case of light reducing to EMR is relatively clean, the
case of phenotypic traits and genes is far less clean. Genes, as we now know,
may not be single stretches of DNA, but may involve many distinct segments
of DNA. The regulatory superstructure of noncoding DNA means that identi-
fication of a stretch of coding DNA as a “gene for . . .” is a walloping sim-
plification. Additionally, a given DNA segment may participate in different
macroproperties as a function of such things as stage of development and
extracellular milieu. Despite this complexity, molecular biologists typically
see their explanatory framework as essentially reductive in character. This is
mainly because a causal route from base-pair sequences in DNA to macrotraits,
such as head/body segmentation, can be traced. The details, albeit messy.

Introduction

Macro-

theory

Micro-

theory

Conceptual

Time >

Figure 1.10 Macrolevel theories and microlevel theories coevolve through time. Ini-
tially, the connection between the macro- and microlevels may be tenuous and only
suggestive, but their interactions may increase as experiments reveal correlations be-
tween macro- and microphenomena. As the experimental and theoretical interactions
increase, the theories become increasingly interdigitated. The central concepts classifying
macrophenomena and microphenomena are inevitably revised, and when the conceptual
revision is very dramatic, this may be described in terms of a scientific revolution. Such
revolutions are crudely indicated by a tunnel in the darkening pattern.

can be expected to fill out, at least in general terms, as experimental results
come in.

This brings us to a second major point. Reductive explanations typically
emerge in the later stages of a long and complicated courtship between higher-
level and lower-level scientific domains. Earlier phases involve the coevolution
of the scientific subfields, where each provides inspiration and experimental
provocation for the cohort subfield, and where the results of each suggest
modifications, revisions, and constraints for the other (figure 1.10). As theories
coevolve, they gradually knit themselves into one another, as points of re-
ductive contact are established and elaborated. Initially, contact between a
high-level science and a lower-level science may be based merely on suggestive
correlations in the occurrences of phenomena. Some such suggestive connec-
tions may prove to be genuine; some may turn out to be coincidental.

Reductive links begin to be forged when mechanisms at one level begin to
explain and predict phenomena at another level. Not until there exist reason-

Introduction

ably well-developed theories on both levels do the reductive explanations
emerge. If you don’t know beans about the macrolevel phenomenon of heat,
you will not get very far trying to explain it in terms of some deeper and invis-
ible property of matter. Sometimes the coevolution involves major revisions to
the basic ideas dehning the sciences, and the history of science reveals a wide
spectrum of revisionary modifications. Caloric fluid, as we saw, got the boot
as thermodynamics and statistical mechanics knit themselves together. Galileo
and Newton rewrote the book on momentum and threw out the medieval con-
ception of “impetus.” Michael Faraday demonstrated, contrary to received
opinion, that electricity is fundamentally the same phenomenon, whether it is
produced by a battery, an electromagnetic generator, an electric eel, two hot
metals brought into contact, or a hand rubbing against cat fur. In reality, the
varieties of electrical phenomena are at bottom just one thing: electricity.

Reductive achievements sometimes fall short of the complete reduction of
one theory to another because the available mathematics are insufficient to
the task. Thus quantum mechanics has succeeded in explaining the macro-
properties of the elements, such as the conductivity of copper or the melting
point of lead, but not why a specific protein folds up precisely as it does.
Whether more is forthcoming depends on developments in mathematics. In the
case of quantum mechanisms, the mathematical limitations entail not that the
macroproperties of complex molecules (e.g., serotonin) are emergent in some
spooky sense, but only that we cannot now fully explain them.

It may come as a surprise that the great majority of philosophers working
now are not reductionists, and are not remotely tempted by the hypothesis that
understanding the brain is essential to understanding the mind. Such philoso-
phers typically also see the details of neuroscience as irrelevant to understand-
ing the nature of the mind.^° The reason for their skepticism about the role of
neuroscience is not rooted in substance dualism. Rather, the key idea is that the
mind is analogous to software running on a computer. Like Adobe Photoshop,
the cognitive program can be run on computers with very different hardware
configurations. Consequently, although mind software can be run on the brain,
it can also run on a device made of silicon chips or Jupiter goo. Hence, the
argument goes, there is nothing much we can learn about cognition per se from
looking at the brain.

Known as functionalism, this view asserts that the nature of a given type of
cognitive operation is wholly a matter of the role it plays in the cognitive
economy of the person. Thus the draw operation of Adobe Photoshop is
what it is solely and completely in virtue of its role in Adobe Photoshop. Its

Introduction

nature, so to speak, is exhausted by the description of its interactions when
Photoshop is running. Obviously, therefore, understanding the draw operation
in Photoshop will not be helped by understanding the capacitors and transistors
and circuits of one’s computer. Likewise, understanding what it is for a person
to want a banana or believe that cows can fly will not be helped by under-
standing neurons, circuits, or anything else about how the brain works.

Considerations of this sort motivated Jerry Fodor to emphasize the impor-
tance of experimental psychology, but also to firmly reject the relevance of
neuroscience. He defends a thesis he calls the autonomy of psychology. This is
a methodological claim. Its label embodies his conviction that psychology, as a
science, is independent in its concepts and generalizations, of the concepts and
generalizations of neuroscience. Briefly, the crux of the claim is that cognition
cannot be explained in neurobiological terms and will not be usefully explored
by neuroscientific techniques. The claim supports investigating cognition using
behavioral measures, such as reaction times, and developing theories by con-
structing models that reflect the cognitive organization supposedly revealed by
behavioral and introspective experiments. Neuroscientific data allegedly have a
bearing only on how the cognitive program can be implemented in a particular
physical arrangement, but have very little bearing on the actual nature of the
cognitive functions. Neuroscience, from this perspective, may be of clinical
interest, but it has no major significance for cognitive science.

There are many well-known criticisms of the autonomy-of-psychology
thesis.^"^ One powerful objection, repeatedly raised but never answered by those
who live by the software analogy, is that the conceptual distinction between
hardware and software does not correspond to any real distinction in nervous
systems. There are many levels of brain organization, ranging from protein
channels in membranes, to neurons, microcircuits, macrocircuits, subsystems,
and systems (see again figures 1.1 and 1.2). At many brain levels there are
operations fairly describable as computations, and none of these levels can be
singled out as the hardware level. For example, computations are performed by
parts of dendrites, as well as by whole neurons, as well as by networks of neu-
rons. Learning and memory, for example, involve computational operations at
many levels of structural organization.^® (This will be discussed in more detail
in chapter 8.) The fact is, in nervous systems there are no levels of brain orga-
nization identifiable as the software level or the hardware level. Consequently,
the linchpin analogy (mind/brain = software/hardware) is about as accurate as
saying that the mind is like a fire or the mind is like a rich tapestry. In a poetic
context, the metaphors are perhaps charming enough, but they are far too

Introduction

unconnected to the real phenomena do very much to advance the scientific
project of understanding cognition.

Another major concern is as practical as wearing boots in the snow. There is
no point in turning your back on a vast range of data that might very well
narrow your search space. To do so is perversely counterproductive. Keeping
psychology pure from the taint of neuroscience seems strangely puritanical.
Why not take advantage of every strategy, every technique, every well-
controlled and well-run experiment? Why turn up your nose at some data when
it might be useful?

Fodor, however, takes the software/hardware analogy to license assurance
that neuroscientific data will not be useful. As noted, the analogy stipulates that
neuroscientific data pertain to implementation rather than software. Unfortu-
nately, and rather obviously, this response is untenable, because the analogy is
untenable. By insisting that experimental psychology cut itself off from poten-
tially useful neurobiological data, theory dualism is steering resolutely into the
past instead of into the future. In a curious way, brain-averse functionalism
is methodologically close to Cartesianism. In place of Descartes’s nonphysical
mental substance, functionalism substituted “software.” Otherwise, things are
much the same; no interest in or search for mechanisms of cognitive functions,
no credence given to the possibility that we might learn fundamental facts
about the mind by understanding how the brain works.

Notwithstanding the strictures of functionalism, the fact is that neuroscience
and cognitive science are coevolving, like it or not. This coevolution is moti-
vated not by ideology, but by the scientific and explanatory rewards derived
from the interactions. Increasingly, this trend means that data from neuro-
science are having an impact on how we frame questions about the mind and
how we rethink how best to characterize psychological phenomena themselves.
Examples of these developments will be seen in later chapters, and they will
make us wonder whether some folk-psychological “verities” are as much in
need of revision as were the “verities” of geocentrism. Exactly how the cog-
nitive sciences and the neurosciences will knit into one another and how
coevolution will change both is not easily predicted.

Though we can expect in a general way that mental phenomena will reduce
to neurobiological phenomena, in the qualified sense of “reduction” used here,
that achievement is certainly not yet in hand and could well be thwarted by the
reality of the brain. For all we can be sure of now, a loose, if revealing, inte-
gration of domains may be the best we can achieve. Detailed explanatory
mechanisms may elude us, and we might have to settle for general explanatory

Introduction

principles that give us a story about mechanisms. Then again, maybe not.
Science often surprises us with progress we thought impossible.^’

There are still some very general worries about reduction to be addressed
and allayed in advance of further progress, and I shall turn to three of those
now.^®

3.1 If We Get an Explanatory Reduction of Mental Life in Terms of Brain
Activity, Should We Expect Our Mental Life to Go Away?

This worry is based on misinformation concerning what reductions in science
do and do not entail. The short answer to the question, therefore, is “No.”
Pains will not cease to be real just because we understand the neurobiology
of pain. That is, a reductive explanation of a macrophenomenon in terms of
the dynamics of its microstructural features does not mean that the macro-
phenomenon is not real or is scientifically disreputable or is somehow explana-
torily unworthy or redundant. Even after we achieved an explanation of light
in terms of EMR, the classical theory of optics continues to be useful, even in
discovering new things. Nobody thinks that light is not real, as result of Max-
well’s explanatory equations. Rather, we think that we understand more about
the real nature of light than we did before 1873. Light is real, no doubt about
it. But we now see visible light as but one segment of a wider spectrum that
includes x-rays, ultraviolet light, and radio waves (plate 1). We can now ex-
plain a whole lot at the macrolevel that we were unable to explain before, such
as why light can be polarized and why light is refracted by a lens.

Sometimes, however, hitherto respectable properties and substances do turn
out to be unreal. The caloric theory of heat, as we mentioned, did not survive
the rigors of science, and caloric fluid thus turned out not to be real. As neu-
roscience proceeds, the fate of our current conception of consciousness, for
example, will depend on the facts of the matter and the long-term integrity
of current macrolevel concepts.^®

3.2 Should We Expect a One-Step Integration of the Behavioral Domain with
the Neuronal Domain?

Introduction

unlikely that explanations of macroelTects such as perceiving motion will be
explained directly in terms of the lowest microlevel. More likely, high-level
network effects will be the outcome of interacting subnetworks; subnetwork
effects the outcome of participating neurons and their interconnections; neuron
effects the outcome of protein channels, neuromodulators, and neurotrans-
mitters; and so forth. One misconception about the integrationist strategy sees
it as seeking a direct explanatory bridge between the highest level and lowest
levels. This idea of “explanation in a single bound” does stretch credulity, and
neuroscientists are not remotely tempted by it. My approach predicts that
integrative explanations will proceed stepwise from highest to lowest, and that
the research should proceed at all levels simultaneously.^®

3.3 How Can You Have Any Self-Esteem If You Think You Are Just a Piece
of Meat?

The first part of the answer is that brains are not just pieces of meat. The
human brain is what makes humans capable of painting the Sistine Chapel,
designing airplanes and transistors, skating, reading, and playing Chopin. It is
a truly astonishing and magnificent kind of “wonder- tissue,” as the philosopher
Dennett jokingly puts it.^^ Whatever self-esteem justly derives from our ac-
complishments does so because of the brain, not in spite of it.

Second, if we thought of ourselves as glorious creatures before we knew that
the brain is responsible, why not continue to feel so after the discovery? Why
does the knowledge not make us more interesting and remarkable, rather than
less so? We can be thrilled by the spectacle of a volcano erupting or a calf being
born or a bone healing before we understand what volcanoes are and how re-
production and healing work. Being the creatures we are, however, commonly
we are even more thrilled in the embrace of the knowledge about volcanoes and
birth and bones. Understanding why we sleep and dream or how we distinguish
so many smells makes us so much more glorious, rather than less so. At the
same time, understanding why someone is demented or gripped by a hand-
washing compulsion or tormented by a phantom arm after amputation helps
replace superstition with sympathy and panic with calm reason.

Third, self-esteem, as we all know, depends on many complex factors,
including things that happened or didn’t happen during childhood and social
recognition of a certain kind. None of this is altered one iota by realizing that
one’s feelings are caused by brain activity. When I step on a thorn, it still hurts
in the same way, whether I know that the pain is really an activity in neurons

Introduction

or not. When a teacher sincerely compliments a student’s essay as insightful,
well-researched, and clearly written, he esteems the student’s accomplishment.
In consequence, she is entitled to self-esteem, and it would be utterly irrelevant
to add, “Too bad, though, this paper is just a product of your brain” as a de-
flationary remark.

4 Concluding Remarks

Three hypotheses underpin this book:

Hypothesis 1 Mental activity is brain activity. It is susceptible to scientific
methods of investigation.

Hypothesis 2 Neuroscience needs cognitive science to know what phenomena
need to be explained. To understand the scope of the capacity you want to
explain — such as sleep, temperature discrimination, or skill learning — it is
insufficient to simply rely on folk wisdom and introspection. Psychophysics,
and experimental psychology generally, are necessary accurately to characterize
the organism’s behavioral repertoire and to discover the composition, scope,
and limits of the various mental capacities.

Hypothesis 3 It is necessary to understand the brain, and to understand it at
many levels of organization, in order to understand the nature of the mind.

Hypothesis 1 is a front-and-center topic of the entire book. It will be contin-
ually dissected, tested, and defended when we address the nature of the self,
consciousness, free will, and knowledge. Ultimately, its soundness will be
settled by what actually happens as the mind/brain sciences continue to make
progress. Conceivably, it will turn out that thinking, feeling, and so on, are in
fact carried out by nonphysical soul stuff. At this stage of science, however, the
Cartesian outcome looks improbable. As noted earlier, hypothesis 3 is hotly
contested by those psychologists and philosophers who favor the “mind as
software” approach. Hypothesis 2, on the other hand, though it may be
embraced in principle by neuroscientists, is sometimes ignored in practice. For
example, molecular-level neuroscientists may be apt to scoff at systems-level
neuroscientists who are groping for ways to test psychophysical hypotheses in
monkeys.

The more serious problem, however, is that brain-averse philosophers and
psychologists tend to assume that those who believe hypothesis 3 (e.g., neuro-

Introduction

scientists) are bound and determined to dwbelieve hypothesis 2.^^ No such
conclusion follows, of course. The important point is that psychology and neu-
roscience are coevolving and will continue to do so. The fields are not mutually
incompatible, but mutually dependent. Temporarily focusing on one level of
organization is often a practical experimental expedient, but that is very dilfer-
ent from making it a principle of research strategy.

One further observation concerns our ideas about ourselves, including our
philosophical ideas. The main business of our brains is to help us adapt to
changing circumstances, to predict food sources and dangers, to recognize
mates and shelter, in general, to allow us to survive and reproduce. The human
brain, as a rather fancy defense against variability and disaster, also generates
stories — call them theories — to explain why things happen and thus help pre-
dict what will happen.

Some theories are better than others. The theory that bubonic plague is
God’s punishment is not as successful as the theory that it is a rat-borne bac-
terial infection. The first suggests prayer as a preventative, the second predicts
that hand washing, rat killing, and water boiling will be more effective. As
indeed they are. The theory that Zeus makes thunder by hurling luminous bolts
is not as successful as the theory that lightning causes a sudden heating of
adjacent air and therewith a sudden expansion. And so forth.

What about theories concerning ourselves — our natures? Our ideas about
why people do certain things, and indeed why one does something oneself, are
part of a wider network of story structures, with some cultural variability and
some commonality. We explain and predict one another’s behavior by relying
on stories about attitudes, will power, beliefs, desires, superegos, egos, and
selves. For example, we explain a certain basketball player’s demands for
attention in terms of his big ego; we may describe a backsliding smoker as
lacking will power, an actor as moody or as obsessed with popularity or as
having a narcissistic personality disorder, and so on. Freud (1856-1939) urged
us to explain compulsive behavior in terms of superego dysfunction. But what,
in neurobiological terms, are these states — will power, moods, personality, ego,
and superego? Are some of these categories like the categories of now-defunct
but hitherto “obvious” Aristotelian physics, categories such as “impetus” and
“natural place”?

Given scientific progress in general, along with specific evidence about the
brain and how it works, our shared conventional story structures may come
to be modified where they prove less successful than experimentally tested
theories. The details of theory modifications are essentially impossible to pre-
dict in advance. Already, however, we can see some story modification.

Introduction

In the last fifty years, we have come to realize that epilepsy is best under-
stood in neurobiological terms, not in terms of the divine touch. Hysterical pa-
ralysis is not a dysfunction of the uterus, but of the brain. In subjects’ who are
compulsive handwashers, possession by spirits or superego dysfunction explains
and predicts far less than neuromodulator levels. The discovery that highly
addictable subjects have a gene implicated in the quirks of their dopamine re-
ward system begins to hint that we will want to reconsider what exactly having
or lacking will power comes to. None of this is surprising, for what the history
of science reveals is that some theory revision is typical and pretty much inevi-
table, no matter what the domain of inquiry — astronomy, physics, biology, or
the nature of our minds. That the story structure giving shape to traditional
philosophical inquiry may itself evolve, perhaps quite profoundly, accordingly
presents an even deeper challenge to those who wish to isolate philosophy from
science.

The overarching theme of this book is that if we allow discoveries in neuro-
science and cognitive science to butt up against old philosophical problems,
something very remarkable happens. We will see genuine progress where prog-
ress was deemed impossible; we will see intuitions surprised and dogmas
routed. We will find ourselves making sense of mental phenomena in neuro-
biological terms, while unmasking some classical puzzles as preneuroscientific
misconceptions. Neuroscience has only just begun to have an impact on philo-
sophical problems. In the next decades, as neurobiological techniques are
invented and theories of brain function elaborated, the paradigmatic forms for
understanding mind-brain phenomena will shift, and shift again. These are still
early days for neuroscience. Unlike physics or molecular biology, neuroscience
does not yet have a firm grasp of the basic principles explaining its target phe-
nomena. The real conceptual revolution will be upon us once those principles
come into focus. How things will look then is anybody’s guess.

Selected Readings

Basic Introductions

Allman, J. M. 1999. Evolving Brains. New York: Scientific American Library.

Bechtel, W., and G. Graham, eds. 1998. A Companion to Cognitive Science. Oxford:
Blackwells.

Osherson, D., ed. 1990. Invitation to Cognitive Science. Vols. 1-3. Cambridge: MIT
Press.

Palmer, S. E. 1999. Vision Science: Photons to Phenomenology. Cambridge: MIT Press.

Introduction

Sekuler, R., and R. Blake. 1994. Perception. 3rd ed. New York: McGraw Hill.

Wilson, R. A., and F. Keil, eds. 1999. The MIT Encyclopedia of the Cognitive Sciences.
Cambridge: MIT Press.

Zigmond, M. J., F. E. Bloom, S. C. Landis, J. L. Roberts, L. R. Squire. 1999. Funda-
mental Neuroscience. San Diego: Academic Press.

Additional Selected Readings

Bechtel, W., P. Mandik, J. Mundale, and R. S. Stufflebeam, eds. 2001. Philosophy and
the Neurosciences: A Reader. Oxford: Oxford University Press.

Bechtel, W., and R. C. Richardson. 1993. Discovering Complexity. Princeton: Princeton
University Press.

Churchland, P. M. 1988. Matter and Comciousness. 2nd ed. Cambridge: MIT Press.

Churchland, P. S. 1986. Neurophilosophy: Towards a Unified Understanding of the Mind-
Brain. Cambridge: MIT Press.

Crick, F. 1994. The Astonishing Hypothesis. New York: Scribners.

Damasio, A. R. 1994. Descartes’ Error. New York: Grossett/Putnam.

Kandel, E. R., J. H. Schwartz, T. M. Jessell, eds. 2000. Principles of Neural Science. 4th
ed. New York: McGraw-Hill.

Moser, P. K., and J. D. Trout, eds. 1995. Contemporary Materialism: A Reader.
London: Routledge.

History

Brazier, M. A. B. 1984. A History of Neurophysiology in the 17th and 18th Centuries:
From Concept to Experiment. New York: Raven Press.

Finger, S. 1994. Origins of Neuroscience: A History of Explorations into Brain Function.
New York: Oxford University Press.

Gross, C. G. 1999. Brain, Vision, Memory: Tales in the History of Neuroscience. Cam-
bridge: MIT Press.

Young, R. M. 1970. Mind, Brain, and Adaptation in the Nineteenth Century. New York:
Oxford University Press.

Journals with Review Articles

Annals of Neurology

Cognition

Current Issues in Biology

Introduction

Nature Reviews: Neuroscience
Psychological Bulletin
Trends in Cognitive Sciences
Trends in Neurosciences

Websites

BioMedNet Magazine: http://news.bmn.com/magazine
Encyclopedia of Life Sciences: http://www.els.net

The MIT Encyclopedia of the Cognitive Seiences: http://cognet.mit.edu/MITECS
Neuroanatomy: http://thalamus.wustl.edu/course
Science: http:// scienceonline.org

The Whole Brain Atlas: http://www.med.harvard.edu/AANLIB/home.html

I Metaphysics

This page intentionally left blank

An Introduction to Metaphysics

1 Introduction

As a label for a subject, the term “metaphysics” has an odd origin. The story
is worth telling because it helps explain the miscellany of topics herded to-
gether as metaphysical. The term was first used about 100 b.c. by an editor,
probably Andronicus of Rhodes, of Aristotle’s works. Aristotle’s corpus
covered a vast range of topics, for he had something to say on virtually every-
thing, including logic, physics, the weather, the heavens, ethics, and reproduc-
tive processes in animals. A problem for the editor was that Aristotle had failed
to give a title to the material the editor regarded as following sequentially after
Physica, Aristotle’s book on physics. To solve the problem the editor entitled
this material, unpretentiously, “The Book after the Physics,” that is, Meta-
physica. Thus was born metaphysics.^ How, then, did metaphysics acquire the
status of a subdiscipline in its own right? One clue comes from the topics Aris-
totle had discussed in Metaphysica.

“Physica” means nature, and in the Physica, Aristotle addressed questions
about the nature of things. He asked why some things fall, for instance, rocks,
but other things, such as smoke, do not. He asked why a rolling ball eventually
comes to a stop, why the planets move as they do, and why fire is hot. In
Metaphysica, by contrast, he addressed questions of somewhat greater gener-
ality, such as what basic things exist, whether are there ultimately different
kinds of basic things, and if so, what accounts for the differences. He discussed
the view that earth, air, fire, and water are basic, Democritus’s contrasting view
that atoms are basic, and Pythagoras’s rather strange idea that everything is
ultimately made of numbers. Because Plato had argued for the otherworldly
existence of mathematical objects and logical truths, Aristotle also subjected

Metaphysics

those views to close criticism. Additionally, he theorized about causality: the
nature of different types of causes, whether there is a causal origin of the uni-
verse, and the fact that different sciences may give different causal explana-
tions of related phenomena. In discussing what might be the fundamental
stuff of reality, he also made some suggestions about what it means to say of
something — anything — that it actually does exist.

The collection of subjects in Aristotle’s Metaphysica is a bit of a hodge-
podge, but he regarded them as relevant to all sciences, and hence alike in that
respect. Because of their general relevance, Aristotle used the expression “first
philosophy” to signify the generality of the topics discussed in Metaphysica.
The shared generality did not, of course, entail that the topics delimit a unified
natural phenomenon of any kind, and it is fairly evident that Aristotle was
under no illusion about that.

Moreover, Aristotle did not suppose that the topics in Metaphysica were
beyond the methods of science or different in kind from the questions of the
particular sciences. Later philosophers commonly did, however. That meta-
physics not only has its proprietary subject matter — the fundamental nature of
reality — but also has its own distinct methods for getting true answers became
philosophical orthodoxy, but not because of any endorsement from Aristotle.

For convenience, I use the expression “pure” metaphysics to refer to the
school of thought that assumes that metaphysical answers are beyond the reach
of scientific methods and scientific discoveries and that the job of metaphysics
is to lay the absolute foundations for all the sciences. Pure reason and reflec-
tion, perhaps with the addition of introspection and meditation, are, according
to this perspective, the proper methods for making progress on metaphysical
questions. This view considers the very status of science itself — ultimately
sound or wrong — to depend on how the metaphysical answers turn out, and
hence ultimately to depend on the beyond-science methods of metaphysics.
Moreover, Aristotle’s rather innocent expression “first philosophy” came to
acquire a more self-important significance associated with suprascientific
methods and principles.

What is the status of metaphysics now? Briefly, its domain has shrunk in
tandem with the maturing of the various scientific disciplines. With the devel-
opment of modern physics and chemistry, theories about atoms, subatomic
particles, and force fields came to dominate serious discussion of the funda-
mental nature of reality. Following Newton’s discovery in the seventeenth cen-
tury of the laws of motion and the explanation of planetary movement, the
nature of space and time have been most productively pursued by physicists.

An Introduction to Metaphysics

theoretical as well as experimental.^ Especially in the last century, physicists
pursuing cosmology made stunning scientific progress on such issues as the
nature and origin of stars, planets, and galaxies; the age of the universe; and its
changes through time. Geologists have learned a great deal about the origin
and history of the Earth, and biologists about the origin and history of species.
In general, the various scientific disciplines have been spectacularly successful
in making progress on Aristotle’s metaphysical questions.

In view of this sort of scientific progress on various classical metaphysical
questions, some philosophers recognized that metaphysics, as construed by the
purists, is probably misguided. American pragmatists, beginning with Charles
Sanders Peirce (1839-1914), cautioned against the idea that there is a rock-
bottom foundation to all of science, where metaphysical reflection is the single
tool for laying that foundation. According to Peirce, there is, for better or
worse, nothing more adequate or basic than the method of science itself: ob-
servation, experiment, hypothesis formation, and critical analysis. That, as one
might say, is just a fact of the human condition. We use reason and science to
reexamine our earlier assumptions and make revisions where necessary, in
effect bootstrapping our way to a better and better understanding of our world.
As Clark Glymour put it, we have to start with whatever we think we know,
and work backwards and sideways, as well as upwards, to improve upon it.^

In the later part of the twentieth century, the central figure to attack the
pure, a priori conception of metaphysics was W. V. O. Quine (1908-2000). In
the spirit of C. S. Peirce, he defended the idea, scandalous to philosophers even
in the 1960s, that there is no first philosophy. There is nothing firmer and more
fundamental than science itself"^ What he meant was that we use science to
bootstrap our way to a better and better understanding of the world. Beyond
the scientific method, broadly conceived, there is no independent method for
discovering the nature of reality. Quine was not denying a role to common
sense, for he took science and common sense to be elements of the same enter-
prise: making sense of the world through experimenting, theorizing, and think-
ing things through. Science, in Quine’s view, is actually rigorous and systematic
common sense in the context of cultural evolution.

If we are persuaded by the pragmatists, we have a choice: either we abandon
metaphysics as misguided, or we break with the purists and update our char-
acterization of the subject matter. On the latter option, metaphysical questions
are best recharacterized as those questions where scientific and experimental
progress is not yet sufficient to found a flourishing explanatory paradigm. This
implies that “metaphysical” is a label we apply to a stage — an immature stage.

Metaphysics

in fact — in a theory’s scientific development, rather a distinct subject matter
with distinct methods. Until rather recently, theories about the self, conscious-
ness, and free will, for example, were at a very immature stage, since neuro-
science and cognitive science were not sufficiently advanced to get very far in
addressing these matters experimentally. Because of this relative immaturity,
these topics may still be regarded as metaphysical, but when scientific success
comes, that status will eventually be cast off" as uninformative and burdensome.

Redescribing metaphysical questions as questions in their prescientific phase
puts them on a very different footing from that favored by a priori philosophy.
It implies, for example, that whether substance dualism is probably true is
fundamentally an empirical issue, not an issue than can be resolved by pure
reason and reflection independently of scientific exploration. It implies that
whether conscious decisions lack all causal antecedents in the brain is funda-
mentally a question of empirical fact that no amount of beyond-science hand
wringing can alter.

The more we understand about brains, their evolutionary development, and
how they learn about their world, the more plausible that the pragmatists are
on the right track concerning the scope and limits of metaphysics. The expla-
nation is quite simple: We reason and think with our brains, but our brains are
as they are — hence our cognitive faculties are as they are — because our brains
are the products of biological evolution. Our cognitive capacities have been
shaped by evolutionary pressures and bear the stamp of our long evolutionary
history. If humans, and only humans, have a special, suprascientific, “meta-
physical” faculty, its origin and existence should be consistent with the facts of
evolutionary biology and neural development. Yet such a faculty looks incon-
sistent with the facts of evolutionary biology and neural development. Let us
take a closer look at the matter.

If, as seems evident, the main business of nervous systems is to allow the
organism to move so as to facilitate feeding, avoid predators, and in general
survive long enough to reproduce, then an important job of cognition is to
make predictions that guide decisions. The better the predictive capacities, the
better, other things being equal, the organism’s chance for survival. In a popu-
lation of organisms, those who are predictively adroit do better than those who
are predictively clumsy, other things being equal.

When an organism survives long enough to reproduce, its offspring inherit its
genes, and thus inherit the capacities whose structures are organizationally de-
pendent on those genes. Occasionally an offspring has tiny changes in its genes,
a mutation, that results in the organism being structurally somewhat different

An Introduction to Metaphysics

from its parents. Usually such mutations are disadvantageous. On rare occa-
sions, however, a mutation will give the offspring a change in brain or body
structure that, relative to the organism’s environment, ends up conferring a bit
of an edge in the struggle to survive. If an organism with the advantageous
mutation does survive and reproduce, its offspring will inherit the modified
capacity. This is descent with modification.

For the pure metaphysical approach to the mind, the implications of descent
with modification are troubling. For example, it implies that a fancy visual
system will not emerge just for the sheer excellence of having fancy percep-
tion. Unless improvements in visual capacity make a net contribution to the
organism’s overall capacity to survive, they will tend to vanish along with the
organism. When an offspring happens to have a mutation in its genes that dic-
tates a structural change in the nervous system that gives the organism a per-
ceptual capacity that allows it to make better predictions than its competitors
can make, then that organism is more likely to survive and pass its genes on
to its offspring. Importantly, however, if the mutation comes at a cost to the
organism — if, for example, there is a trade-off between speed of processing and
sophistication of perceptual images — then a given mutation may carry a net
loss, even though the higher degree of accuracy of perception is predictively
useful when considered alone.

From the perspective of Darwinian evolution, therefore, any beyond-science
metaphysics has to face a tough question: Would there have been evolutionary
pressure for the emergence of a special faculty with a unique route to Absolute
Metaphysical Truth? What could have been the nature of such pressure? Is
there a plausible account consistent with natural selection that can explain how
humans could come to have such a capacity? Relative to what is now known, it
is doubtful that any such account is forthcoming, even if one can envisage what
such a capacity would be like. Consequently, we do best to resign ourselves
to the probability that there is no special faculty whose exercise yields the
Absolute, Error-Free, Beyond-Science Truths of the Universe. All we can do,
though it is certainly no small thing, is to learn what the best available science
says, mindful that it may embody errors, both large and small, and then subject
it to criticism, refinement, and extension via more of the same — experimenting,
theorizing, and thinking things through (see also chap. 6, pp. 245-254).^

Still, it may be urged that one’s feeling of having made progress in supra-
scientific metaphysics should count for something, and the conviction that
such progress has been made does indeed exist. More exactly, feelings of cer-
tainty may be cited as the benchmark for having discovered a Beyond-Science

Metaphysics

Metaphysical Truth. For example, feelings of unshakeable conviction or abso-
lute certainty may accompany consideration of the hypothesis that the mind
is a nonphysical substance. Descartes, as we saw in chapter 1, seems to have
enjoyed such certainty and to have believed it warranted a specific conclusion.

Feelings of certainty, however, are no guarantee of truth. They can, of
course, motivate testing a hypothesis for truth. They can motivate continuing a
research project even in the face of scoffers. But feeling certain that a hypothe-
sis is true is, sadly, all too consistent with falsity of the hypothesis. Everyone
knows of occasions in his own life when certainty and falsity were happy bed-
fellows. Moreover, the historical record is painfully clear on this matter. At
various times, people have been completely certain that the Earth did not
move, that space is Euclidean, that atoms are indivisible, that insanity is caused
by possession of demons, that they can see into the future, and that they can
communicate with the dead. Yet all of these propositions are probably false.
That falsity and conviction coexist should not surprise us. Certainty, after all, is
but a cognitive-emotive state of the brain, one such state among many other
cognitive-emotive states of the brain.

The pragmatic conception of metaphysics may seem a bit of a disappoint-
ment, for much the same reason that it may seem disappointing that our
universe has is no such thing as Absolute Space or Lady Luck or Guardian
Angels. Having to muck on as best one can, Glymour-like, seems a lot less ro-
mantic, perhaps, than being on a quest for suprascientific Metaphysical Truth.
Nevertheless, in making progress, we abandon romantic notions when their
wheels fall off.

What metaphysical questions still remain to be resolved? On the topic of
causality, impressive mathematical and scientific progress has indeed been
made. Notably, there has been relatively little progress in the pure metaphysics
of causality. We shall look a bit more closely at this in the next section. One
subfield in physics where fundamental issues about the nature of reality remain
very much alive is quantum mechanics. On the significance and interpretation
of quantum mechanics, there is fruitful interaction between physics and the
philosophy of physics.®

Beyond these matters, the remaining metaphysical questions, traditionally
classified, are about the mind: What is the nature of consciousness, the self, free
will? Is the nonphysical mind perhaps the fundamental reality? The only fun-
damental reality? How can we come to understand the mind if we have to use it
to understand it? The three topics — consciousness, the self, and freedom of the
will — constitute the three chapters in the metaphysics part of this book. From

An Introduction to Metaphysics

the pragmatist’s perspective, we shall explore questions about consciousness,
free will, and the self as questions about the mind/brain, and we will see that
a young science is discovering things about the nature of the mind/brain that
we could never have discovered through reflection and introspection alone.
Whether any uniquely metaphysical work on these topics is left for the pure
metaphysician and whether they will go the way of questions about the origin
of the Earth and the nature of life are issues we can reconsider toward the end
of the book.

As part of the groundwork for those discussions, we need to revisit the mind-
body problem raised in chapter 1 . Note that there is a mind-body problem only
if the mind is nonphysical and the body is physical. The nub of the problem is
how the two substances can interact and have effects on one another if they
share no properties whatever. How, for example, can mental decisions have an
effect on neurons, or how can directly stimulating the cortex with an electrode
result in feeling one’s leg being touched? On the other hand, if the mind is
activity in the brain, then that particular problem, at least, does not exist. Other
problems of exist, to be sure, but not the problem of the interaction between
soul stuff and brain stuff.

2 Metaphysics and the Mind

How can we come to an informed opinion on the question of the existence of
soul stuff? The pragmatist suggests that we use the science we have to compare
the strengths and weaknesses of competing hypotheses, design good experi-
ments, and test the hypotheses. Were a soul-brain interaction to exist, there
should be some evidence of the interaction. This has not been forthcoming. The
laws of physics, as we currently understand them, include the law of conserva-
tion of mass-energy. An interaction whereby a nonphysical mental event causes
some physical effect, such as a change in the behavior of neurons, would violate
the law of conservation of mass-energy. So far, no such violations are seen in
nervous systems. This is not to say we are certain none exist, but only that there
is no convincing reason to believe that they do exist. Deflecting criticism by
postulating a shyness ejfect, according to which it is just the nature of souls to
shield themselves from experimental detection, is, needless to say, not going
to fool anybody. Since there is no independent evidence for a shyness effect,
postulating shyness is a blatant cheat to avoid facing the implications of the
absence of positive evidence.

Metaphysics

Powerful reasons for doubting substance dualism accumulate with the in-
creasingly detailed observations of the dependencies between brain structure
and mental phenomena. The degeneration of cognitive function in various
dementias such as Alzheimer’s disease is closely tied to the degeneration of
neurons. The loss of specific functions such as the capacity to feel fear or see
visual motion are closely tied to defects in highly specific brain structures in
both animals and humans. The shift from being awake to being asleep is char-
acterized by highly specific changes in patterns of neuronal activity in inter-
connected regions. The adaptation of eye movements when reversing spectacles
are worn is explained by highly predictable modifications in very specific and
coordinated regions of the cerebellum and brainstem. And example can be
piled upon example.

One of the most metaphysically profound discoveries in this century showed
that a human’s mental life is disconnected if the two hemispheres of his brain
are disconnected. In the 1960s, a group of patients with epilepsy so intractable
that it resisted control with drugs underwent a surgical procedure that sepa-
rated the two cerebral hemispheres by cutting the nerve bundle joining them.
The purpose of the surgery was to prevent the seizures from propagating from
one hemisphere to the other, and it was successful in achieving this aim (figure
2.1). In careful postoperative studies of the capacities of “the split-brain”
subjects, Roger Sperry, Joseph Bogen, and their colleagues found that each
hemisphere could have perceptual experiences or make movement decisions
independently of the other.’

To illustrate the disconnection elfect, consider the following experiment: A
picture of a snowy scene is flashed to the right hemisphere, and a picture of a
chicken’s claw is flashed to the left (figure 2.2).® An array of pictures is placed
before the subject, who is to select, with each hand, the picture that best
matches the flashed picture. In this setup, the split-brain subject does this: his
left hand (controlled by the right hemisphere) points to a shovel to go with the
snowy scene, and the right hand (controlled by the left hemisphere) selects
a chicken’s head to go with the chicken’s claw. This is described as a discon-
nection elfect, since each of the disconnected hemispheres seems to be able to
function in perception and choice much as a single person does. In another ex-
ample, Joseph Bogen reports observing a split-brain subject seated in an easy
chair. His right hand picked up a newspaper, and he began to read. The left
hand took the newspaper and tossed it to the floor. The right hand picked it up
again, and he resumed reading, only to have the left hand pull the paper away

An Introduction to Metaphysics

Figure 2.1 (A) The positions of major interhemispheric connections in the human brain

as seen in sagittal section. (B) Interhemispheric fibers largely connect homologous areas
in the two half-brains. (C) In addition, they terminate mostly in the cortical laminae
from which they arose in the opposite hemisphere. (From Gazzaniga and LeDoux, The
Integrated Mind. New York: Plenum, 1978.)

Metaphysics

Figure 2.2 The method used in presenting two different cognitive tasks simultaneously,
one to each hemisphere. The left hemisphere was required to select the match for what
it saw (the chicken claw), while the right hemisphere was to select the match for what it
saw (the snowy scene). After each hemisphere responded, the subject was asked to ex-
plain the behavior. (From Gazzaniga and LeDoux, The Integrated Mind. New York:
Plenum, 1978.)

and drop it to the floor. In light of this behavior, we can conjecture that each
hemisphere has its own integrated, coherent awareness.®

These remarkable results demonstrate that the unity of mental life is depen-
dent on the anatomical connections in the brain itself This seems reasonable
enough on the hypothesis that mental life is activity in the brain. If the hemi-
spheres are disconnected, the activity subserving mental function in the two
hemispheres is disconnected. On the other hand, if mental life is activity in a
nonphysical substance with no physical properties whatever, then why should

An Introduction to Metaphysics

splitting the brain split the mind? One could cobble together some story,
perhaps, but a story that is consilient with what else is known about mental
phenomena and neural phenomena — a story that has at least a modicum of
plausibility — is very hard to come by. So far, none has been able to get off the
ground.^®

In general, as a hypothesis about the nature of the mind, how does substance
dualism stack up against physicalism? The short answer is that substance dual-
ism chronically suffers from the lack of any positive description of the nature of
the mental substance and any positive description of the interaction between the
physical and the nonphysical. The content of the hypothesis is specified mainly
by saying what the soul is not. that is, it is not physical, not electromagnetic, not
causal, and so forth. Negative characterizations can be useful, and they may be
hne as a place to start. Evaluation of the hypothesis cannot proceed, however,
without some positive elaboration: we need to hear something about what the
proposed interaction is, where the interactions occur, and under what general
conditions. Were someone to proclaim a new theory of light that says only that
light is not electromagnetic radiation, it would be difficult to know how to test
it. Because the soul-brain hypothesis lacks a substantive, positive characteriza-
tion, it too is hard to take seriously, especially at this stage of science.

To compete with a brain-based explanation of, say, face recognition or
obsessive-compulsive disorder or dementia, some positive claims — even the
bare-bones of some positive claims — should be on the table. The slow degen-
eration of memory and cognition generally seen in Alzheimer’s patients, for
example, is currently explained within neuroscience as the progressive loss of
neurons in the cortex. How, according to its adherents, might a soul-based
story go? We are not told.

More generally, there appears to be no progress — not even significant ex-
perimental effort to make progress — in revealing the existence or properties of
soul-brain interactions. One neuroscientist, John Eccles, did briefly entertain
the conjecture that a soul-brain interaction is mediated by a special entity that
he called a “psychon.” Psychons, he believed, are the mediators of brain-soul
interactions. “ He predicted that they might work at certain synapses. But what
are the properties of psychons? On what, in the synapse, do they exert an effect?
Is the alleged effect mediated chemically? Electrically? Are psychons themselves
material entities? No answers, let alone answers consilient with neuroscience,
have been forthcoming. The psychon research program looks like a nonstarter.

The competing hypothesis — that mental phenomena are brain phenomena —
stands in a completely different evidential condition. By contrast with dualism.

Metaphysics

it does have a rich and growing positive account, an account that draws on the
entire range of neurosciences, as well as on cognitive science and molecular
biology. Selected features of this positive account will emerge in the individual
chapters on consciousness, self, and free will, and also in the later chapters on
representation and learning.

As a preliminary to later discussion, notice that Descartes’s particular ver-
sion of dualism identihes the mind with the conscious mind. If there are in fact
nonconscious mental states, this identification is on the skids. Notice that if
some mental events, such as visual-pattern recognition, can be nonconscious,
we do not have the alleged “direct access” to such mental states; i.e. we do not
know about them just by having them.

There is overwhelming evidence that nonconscious cognition plays a critical
role in memory retrieval, belief consolidation, judgment, reasoning, perception,
and language use. To evaluate the evidence, we need to touch on a range of
examples of nonconscious cognition in normal subjects, leaving aside for now
examples of nonconscious cognition in clinical subjects.

The first example is well known to all of us. When we talk, we are aware of
what we are saying, but we are not normally aware in detail of exactly what we
are going to say — exactly what words, phrases, grammatical structure — until
we say it. This is often true of writing as well, and many authors say that they
sometimes find themselves surprised by what they write. The brain’s deci-
sions governing specific choices in words and sometimes even in content are
typically nonconscious.

The second example of nonconscious cognition is also well known. In mak-
ing a judgment about someone’s approachability or attractiveness, based sim-
ply on a view of his or her face, one makes use of the degree of dilation of
the person’s pupils. If the pupil diameter is tiny, the face is judged to be less
approachable and attractive than if the pupils are dilated. Surprisingly, most
of us are aware neither of the role that this factor plays in our judgment nor of
having detected pupil size at all. This example is one of many in which eval-
uative judgments are made without the subject being aware of the basis for the
judgment.

In a diflcrent example, subjects are given a task of saying which of two lines,
presented in the peripheral visual field, is longer. Occasionally, a word will
be flashed in the very center of the visual held during the task (flgure 2.3).
As many as 90 percent of subjects report seeing nothing but the lines, even
though the presentation of a word (for example, flake) was above the subjects’
sensory/perceptual threshold. Despite not being consciously seen, the word can

An Introduction to Metaphysics

Task: Which line is longer — the horizontal or the vertical?

B Lines + word trials

C Stem completion task

FLAKE

A Lines-only trials

Figure 2.3 Subliminal effects in the inattention paradigm. The task given to the subject
is to report on the relative length of the intersecting lines (A). Occasionally, in addition
to lines, a word will appear in the center of the screen (B). Typically subjects are
attending to the task, and report no awareness of a word presented briefly under con-
ditions of inattention. When tested on a stem completion task (C), subjects for whom
the word was presented but not consciously seen are much more likely to respond with
the presented word (e.g., flake) than are control subjects who were not shown the word
on the inattention trial. (From Palmer 1999.)

be shown to have a cognitive effect. Here is how. Subjects from the line-judging
experiment as well as subjects who did not participate are given a stem com-
pletion task, in which they have to make a five-letter word given the first two

letters, for example, fl . While only about 4 percent of naive subjects make

the word flake, about 40 percent of test subjects make flake if flake was
flashed. This shows that there was significant cognitive processing of the word,
even though it was not consciously seen. This effect is referred to as inatten-
tional blindsight.^^

Consider now an example involving subthreshold stimuli. Normally, if peo-
ple are asked to indicate a preference for one of two arbitrary visual patterns,
such as a Chinese ideogram, they tend to choose the pattern to which they were
previously exposed. The psychologist Robert Zajonc (pronounced Zy-unse)
asked whether subjects would show this exposure effect if exposure were limited
to a mere 1 millisecond. Not surprisingly, if a picture of an object is flashed on a

Metaphysics

computer monitor for only 1 millisecond, you will not consciously see anything
but a blank screen. Nevertheless, the brain does detect something and does
perform some basic pattern-recognition operations. This is known because
when presented with two arbitrary visual patterns, people do indeed show the
exposure effect and indicate a preference for the subliminally exposed image. In
short, mere exposure biases choice. {Why this should be is another matter.)
This experiment has been replicated many times, and it is an important dem-
onstration that evaluative responses such as preferences can be generated by
nonconscious exposure.

Finally, one of the most intelligent ongoing behaviors is eye movements. The
vast majority of one’s eye movements (including tracking moving objects and
saccades, occurring about three times per second) are made without anything
like conscious decision or choice, and mostly in ignorance that one’s eyes are
moving at all (figure 2.4). Yet when tracked over time, saccadic eye movements
reveal themselves to be organized around discernible goals, directed to solve
visual disambiguation problems, and sensitive to attentional demands and task
complexity. For example, regions with maximum relevance are fixated early in
scanning and are frequently revisited. In walking, the gaze shift is typically a
few steps ahead; in steering a vehicle on a variably curved path, gaze shifts
forward at appropriate times to that distant point on the curve where the line of
sight is tangent to the inside curve, with the result that each segment maintains
constant curvature (figure 2.5).^^ These examples are but a few of many show-
ing that if we think of the mind as intelligent, as perceiving, recognizing, and
problem solving, then the mind cannot be equated with conscious experience,
though conscious experience is part of the mind.

In addition, there are many examples from clinical research that lead to a
similar conclusion. Before moving on to these topics, a further preparatory but
brief comment on causation rounds out this introduction to metaphysics.

3 Causation

The traditional list of metaphysical topics typically gives a prominent place to
causation. No obvious connection links causation as a metaphysical topic to
causation as a topic of interest to neuroscience. There are, however, two areas
of common concern: (1) How can neuroscience get beyond mere correlations
of events in order to confirm causal hypotheses? That is, what conditions have

An Introduction to Metaphysics

Figure 2.4 The effects of task on eye movement. A subject views a picture (A) while the
experimenter monitors the subject’s eye movements and direction of gaze. In each of the
five cases, the instructions were different: freely view (B), estimate the economic level of
the people (C), judge their ages (D), guess what they had been doing before the visitor’s
arrival (E), and remember the clothes worn by the people (F). A saccadic movement is
represented by a line, and a momentary resting position is represented by a small dot.
(From Yarbus 1967.)

Metaphysics

Figure 2.5 Steering a path of variable curvature. A driver (open circle) can merge tra-
jectory segments by shifting point of gaze (black circle) at appropriate times (T1-T4) so
that the line of sight is tangent to the inside curve. Each segment can meet the require-
ment of constant curvature, but the resultant overall trajectory can have a series of cur-
vature variations. (From Wann and Land 2000.)

to be satisfied to establish real causal dependencies between functions such as
consciousness and decision making? (2) What are the neurobiological mecha-
nisms whereby any organism, including us humans, acquires a systematic
causal map of the environs it inhabits? This is a puzzling matter, because
background knowledge is essential to distinguish mere correlations from causal
connections. On the assumption that nervous systems represent certain events
as causally connected, and not as merely coincidentally correlated, the second
question can be rephrased thus: how in fact does the nervous system deploy
relevant background knowledge — together with current observations, manipu-
lations, and interventions — to achieve a predictively powerful causal mapping
of its world?

The first question is basically a methodological question — a question about
reliable analytical or statistical tools, usable in any science, for assessing the
significance of data for a hypothesis and hence in formulating adequate expla-
nations of phenomena. These tools help us understand the importance of such
matters as controls, confounds, standard deviations, measurement error, depen-
dence and independence of variables, and sample-selection bias. They allow us
to sharpen experimental design in order to extract more meaningful results.
Work by philosophers, statisticians, and others has yielded an exuberant body
of important results available for any science, including neuroscience.^®

An Introduction to Metaphysics

The second question is very different. It concerns the cognitive neuroscience
of causal understanding; that is, it pertains to the nature of the actual processes
and operations that underlie a brain’s causal mapping of its world. Because this
is fundamentally a question about learning and knowledge, the topic best fits
into chapter 8, where it can be more productively discussed within the wider
context of learning in general.

As formulated, these two sets of problems seem straightforward enough.
Nevertheless, there is a version of these problems that is metaphysical — in the
beyond-science sense of “metaphysical” — and this version addresses the osten-
sibly deeper matter of the fundamental reality of causes. Although these two
sets of problems can be readily distinguished from their metaphysical com-
panions, discussions of causation have a tendency to drift into the metaphysical
realm. To forearm against unnecessary confusion, therefore, I shall outline the
basic problem in the metaphysics of causation and suggest how its dangers can
be avoided when neuroscientists address questions concerning either method-
ological tools or the neural mechanisms of causal inference.

Causal explanations have to do with how something came to happen or
came to be as it is. We want to understand why some people get gastric ulcers,
whether acidity in lakes reduces fish populations, or why the car has a flat tire.
We want to understand what causes a goose to begin its southern migration, or
to molt, or to imprint on the first large moving thing it sees after hatching.
Metaphysics, now and for the past two thousand years, includes causation as a
topic largely because it is not obvious what makes something a cause as
opposed to a coincidental ride-along, or what exactly in a connection makes for
a causal connection.

Causes are part of the universe, but not in the same way that rabbits and
waves and electrons are part of the universe. A rabbit can be the cause of
something, such as the footprints in the mud, and rabbits can be caused by
things, such as other rabbits. Yet if all the entities in the universe were listed,
“causes” would not be an item in the list. The reason is not that causes are
spooky. On the contrary, causality is about as real as anything gets. If the
dentist needs to do a root-canal procedure, we want him first to cause some-
thing, namely that the nerves innervating the tooth cease to respond to stimuli.

Being a cause means acting in a certain role, and we usually distinguish three
types of causal roles: (1) a precipitating cause (lightning caused the forest fire),
(2) a predisposing cause (having high blood pressure predisposes one for a
stroke) and (3) sustaining cause (proximity to the ocean and a southerly loca-
tion cause the moderate climate of San Diego). Any of these three types of

Metaphysics

casual roles may be productive (lightning caused the forest hre) or preventative
(the rain prevented the forest hre). Depending on interest, on what we already
do and do not know, as well as on the available possibilities for manipulation
or intervention, one or another of these roles can assume prominence in a dis-
cussion, and thus one event or one variable may be selected as “the cause.”

Are causes just self-evidently distinguishable from correlated occurrences of
independent factors? Not at all. It can be very difficult to nail down which
conditions are causal and which are independent. The history of science has
many stories that illustrate the point, but one will suffice. For many decades it
was hrmly believed that stomach ulcers were primarily caused by stress, along
with ingestion of irritating foods such as coffee and beer. This was not unrea-
sonable, since most patients who had pyloric ulcers were under stress and their
symptoms were made worse by coffee and beer. In the 1980s a pair of Austra-
lian physicians, Robin Warren and Barry Marshall, discovered a new strain of
bacteria, later named Helicobacter plyori, in the stomach tissue of patients with
pyloric ulcers (the pylorus connects the stomach to the duodenum, the hrst
segment of the small intestine). Because the stomach is highly acidic, physicians
naturally assumed it to be a hostile environment for bacteria, so this discovery
was astonishing. Even after the discovery, most physicians assumed that the
presence of H. pylori in patients with gastritis was entirely coincidental; that is,
it was noncausally correlated with ulcers and hence independent of the disease.

In proving the causal role of H. pylori in the formation of pyloric ulcers,
Marshall showed that if patients with pyloric ulcers were given antibiotics, their
ulcers soon cleared up.^^ He then used himself as a guinea pig. Having first
established that his own body tested negative for H. pylori, he then infected
himself with it. Shortly thereafter ulcers appeared in his pylorus. He then cured
himself with antibiotics. This intervention showed that H. pylori is probably the
cause of pyloric ulcers. Anxiety and stress probably do not play a causal role,
though the discomfort of ulcers may cause anxiety, which then occurs con-
currently. People with frequent ear infections rarely have ulcers, since they fre-
quently take antibiotics for their ear infections, which kill any H. pylori that
happen to be present. Notice, however, that although frequent ear infections
do not cause ulcer protection, they are correlated with ulcer protection via
the common cause — antibiotics. This sort of example, simplified though it
is, reminds us that distinguishing causes from coincidental correlations is not
always obvious.

Given these and related examples, one may wonder what is it about the
relations between events or conditions that makes for a causal relation? What

An Introduction to Metaphysics

in general distinguishes an event or process that is a cause from an event that is
only coincidentally connected to it? As a first pass, one might assume that
causes and their effects are necessarily connected, whereas coincidental events
are only contingently (accidentally) connected. That is, we think of the cause as
making the effect happen, or as producing the effect, or as being a special kind
of force — a causal force.

This general answer seemed more or less adequate until David Hume in the
eighteenth century pointed out that the necessity or productive force that sup-
posedly explains causality is itself as mysterious as the causal connection it is
meant to explain. What sort of property is this necessity? If necessity is just the
property of causally determining, the answer fails to move us forward but
rather merely moves us round in a circle. Moreover, Hume pressed on, neces-
sity does not appear to be an observable property of events in the world, in the
way that “being heavier than” or “being next to” are relations in the world. So
far as the observable evidence is concerned, all we can determine is that first
one thing happens then the other — first H. pylori gets into the system, then
ulcers begin to form in the pylorus. Unless we can explain what it is for neces-
sity or causal force to be in the world, opined Hume, we may have to conclude
that appeals to necessity and causal force as features of the world are just
metaphysical tomfoolery.

Since Hume’s first unflinching formulation of the problem, legions of philos-
ophers have struggled to answer him. Hume’s challenge could not be dismissed
as a joke or an idle problem, since causal explanations are at the heart of
science. Not understanding what it is in the world that underwrites the differ-
ences between causal connections and coincidental connections is troublesome.
If causality is not an objective feature of the world, how can a causal explana-
tion be a real explanation? How can it be exploited to predict and manipulate
other events?

Roughly speaking, two Hume-answering strategies have been deployed, with
many clever, but ultimately unsatisfactory, twists on each. The first aims to find
some plausible way to show how necessity really is in the world. And one way
to do that is to consider event A as causing event B as an instance of a natural
law. If the H. pylori in John’s pylorus is in fact the cause of the ulcer, then there
must be a natural regularity connecting H. pylori and ulcers such that, in gen-
eral, if H. pylori is in someone’s stomach, ulcers will form. Moreover, this
lawlike regularity implies a counterfactual statement: had John been free of
H. pylori in his stomach, he would not have gotten ulcers. If we take stress as
an independent variable, notice that this counterfactual does not follow: had he

Metaphysics

been stress-free, he would have been ulcer-free. The necessity we conventionally
attribute to causal connections therefore reduces to the fact that there are
objective regularities in nature. Genuinely causal connections are capturable
by natural laws, such as “Bacteria cause ulcers,” “Nails puncture tires,” “Iron
readily combines with oxygen to produce iron oxide,” and “Copper expands
when heated.” The idea is that whereas causes are governed by natural laws,
coincidental connections are not.

To critics of this view, it seems that the regularities exist because of the causal
powers of objects, such as the puncturing potential of nails. The natural-law
strategy seems to imply that things go in the opposite direction; the causality
derives from the lawlike regularity. Moreover, trying to give a noncircular and
scientifically coherent account of a natural law ended up having most of the
same problems as explaining what makes an event a cause. What is it that
makes “Copper expands when heated” a natural law but “All cubes of gold in
the universe are less than 1,000 miles on each side” not a natural law? Both are
true (presumably), both are generalizations, both are testable, both support
counterfactuals.

Another effort to identify the difference could take this form: “because heat-
ing really causes the expansion of copper, whereas nothing about gold makes it
necessary to have the sides of any gold cube be less than 1000 miles.” Alas, that
is the one answer you cannot give, since going around in circles is not progress.
And what of the necessity alluded to? Can we compose a sentence expressing a
truth that is not a natural law but is a necessary truth? Easily. The sentence
“All cubes of pure uranium 238 in the universe are less than 10 feet on a side”
is necessarily true, because that much U238, we know, is a critical mass, and
would explode before we could ever get our 10' x 10' x 10' cube made. So does
that sentence express a natural law? Surely not.

Few seriously doubt that there are natural laws, but hammering out a
satisfactory — and noncircular — account has so far been exceedingly difficult.
Citing examples of agreed-upon natural laws is typically sufficient to teach
students the difference between causes and coincidences, since they can then
generalize on the basis of learned prototypes. But what metaphysicians want to
know is what the prototypes have in common that enable students to general-
ize. Put another way, they would like to understand what lawful connections in
nature amount to.

More or less out of desperation, some philosophers decided it might be
worthwhile to explore the second option, namely, that necessity is not really in
the world, but in the mind. Needless to say, getting an even remotely acceptable

An Introduction to Metaphysics

Story on this option is difficult, since causal explanations for why the dinosaurs
disappeared or why a star exploded or why the average temperature of the
planet is increasing are surely about the world, not the mind. Moreover, they
seem to be about a world before there were humans to think about causality.
Most simply, the objection to this approach is that causal statements are state-
ments not about us, but about the way the world is.

Kant (1724-1804) is the watershed for a hybrid in-the-mind and in-the-
world approach. He well recognized that he needed to avoid the obvious ob-
jection to making causality an entirely subjective matter, yet he also saw the
force of Hume’s arguments. To a first approximation, Kant aimed to figure out
how necessity could be a real feature of events, yet be in the subject — as part of
the “lens” through which we see the world. Not surprisingly, he had a tricky
problem of explaining how causal necessity was not merely subjective. If this
sounds like an impossible goal, that is because at bottom it likely is. Although
Kant wrestled long and brilliantly with the problem, his project his was proba-
bly doomed from the start.

A semi-Kantian strategy rooted in evolutionary biology might hypothesize
that brains have evolved the capacity to infer causality from certain patterns of
regularity observed in experience. Because of the need to make good predic-
tions about food sources, predators, and so on, this is a reasonable hypothesis,
and it can be empirically explored. Partly because it can be empirically ex-
plored, this hypothesis is regarded by some philosophers as fundamentally
irrelevant to the genuinely metaphysical problem as outlined above. They still
want to know what the property is in the world that brains can detect.

In sum, neither strategy for answering Hume has produced universally ac-
cepted results. Although some progress has been made, causation as a meta-
physical issue remains an unsolved problem. As indicated earlier, certain
«o«metaphysical issues regarding causation have, in contrast, permitted con-
siderably more progress. This work clarified causal reasoning by demonstrating
that a given effect can have multiple causes, that events may be independent
but have a common cause, that statistical analyses are essential in cases where
we are trying to identify causally relevant factors, that certain sampling tech-
niques help eliminate confounds, and so on.

In view of the value of this kind of progress, a pragmatic approach counsels
putting Hume’s problem aside, at least for now. The pragmatist will suggest
that we adopt the working hypothesis that there are genuinely causal laws and
that we predict the course of nature better when we have discovered what they
are. Having that hypothesis in place, we can redirect our energy into improving

Metaphysics

techniques for identifying causal factors, judging the objective probability of
the occurrence of a given event relative to specific conditions, and figuring out
how brains actually make reasonable causal inferences. With luck, some of
these results may end up bringing the metaphysical issues to heel. At the very
least, they may help us understand why the metaphysical questions seem com-
pelling and how they might be reinterpreted.

Selected Readings

Bechtel, W., and G. Graham, eds. 1998. “Methodologies of cognitive science.” Part III
of A Companion to Cognitive Science, pp. 339-462. Malden, Mass.: Blackwells.

Churchland, P. M. 1988. Matter and Consciousness. 2nd ed. Cambridge: MIT Press.

Churchland, P. S. 1986. Neurophilosophy: Towards a Unified Understanding of the Mind-
Brain. Cambridge: MIT Press.

Hacking, I. 2001. An Introduction to Probability and Inductive Logic. Cambridge:
Cambridge University Press.

Rennie, J. 1999. Revolutions in Science. New York: Scientific American.

Skyrms, B. 1966. Choice and Chance: An introduction to Inductive Logic. Belmont,
Calif: Dickenson.

Williams, G. C. 1996. Plan and Purpose in Nature. London: Weidenfeld and Nicolson.
Wilson, E. O. 1998. Consilience. New York: Knopf

Websites

BioMedNet Magazine: http://news.bmn.com/magazine
Encyclopedia of Life Sciences: http://www.els.net

The MIT Encyclopedia of the Cognitive Sciences: http://cognet.mit.edu/MITECS

Self and Self-Knowledge

1 The Problem and the Internal-Model Solution
1.1 What Is the Problem?

Sliding out of the MRI tube, I noticed Dr. Hanna Damasio studying the lab’s
display screen as my brain’s images appeared. I stood by her and watched the
image on the screen. “Is that we?” (figure 3.1). Well, yes, in a certain sense, up
to a point. What I was looking at was an image of the thing that makes me me.
Somehow, starting in infancy, my brain built stories about its body, its history,
its present, and its world. From the inside, I know those stories — or perhaps
I should say, “I am one of those stories.” Surely, though, I am not just a bit
of fiction. I am about as real as things get in my world. So how do I make
sense of all this? What exactly is it that the brain constructs that enables me to
think of myself?

Descartes proposed that the self is not identical with one’s body, or indeed,
with any physical thing. Instead, he concluded that the essential self — the self
one means when one thinks “I exist” — is obviously a nonphysical, conscious
thing. To the eighteenth-century Scottish philosopher David Hume, however,
Descortes’s answer was anything but obvious. Hume proceeded to examine
whether there is evidence for something that is the self apart from the body. He
came to realize that if you monitor experience, there does not seem to be any
self thing there to perceive. What one can introspect is a continuously changing
flux of visual perceptions, sounds, smells, emotions, memories, thoughts, and
so forth.

Among all those experiences, however, there does not exist a single, contin-
uous felt experience that one can attend to and say, “That’s the self,” as one

Metaphysics

parietal lobe

frontal lobe

temporal lobe

brain stem brain stem

parietal lobe parietal lobe

cerebellum
frontal lobe

cingulate

corpus

callosum.

fdienocpw

temporal lobe

temporal lobe
brain stem f

midbrain

t)ccipital lobe
-cerebellum-

-spinal cord-

midbrain
pons —
tectum ,
medulla

pons

tectum

medulla

occipital lobe

Figure 3.1 The main divisions of the central nervous system and their crucial compo-
nents, shown in 3-D reconstructions of a human brain. The reconstructions are based on
magnetic-resonance data and on the BRAINVOX technique. Note the relative positions
of the four principal lobes, of the diencephalon (which encompasses the thalamus and
hypothalamus), and of the brain stem. Note also the position of the corpus callosum
(which joins both hemispheres across the midline) and of the cingulate cortex of each
hemisphere. The pattern of gyri (ridges) and sulci (gullies) is very similar but not identi-
cal in the left and right cerebral hemispheres. The pattern is also very similar but not
identical across normal individual humans. (Courtesy of H. Damasio.)

Self and Self-Knowledge

can attend to a felt experience and say, “That’s a headache.” One can remem-
ber events in one’s past, but active recollections are yet more current experi-
ences, albeit experiences where the pronoun “I” figures prominently. Nor is
there a single continuous spatial object that one can attend to and say, “That is
the self,” as one can attend to one’s body and say, “That is my neck.” Because
there seems to be more to me than my body, observations of my body are not
simply equivalent to observations of my self. Thus, Hume concluded, there
does not seem to be a thing that is the self, at least not in the way we unre-
flectively suppose there is.

Of course, Hume realized very well that nothing seems more evident than the
statement “I exist.” In reflection, we take for granted that a single thread of
“me-ness” runs through the entire fabric of one’s experience. If a brick falls on
my foot, I know the pain is mine. If I scold myself about jaywalking, I know
that it is me scolding myself. We generally awake from a deep sleep knowing
who we are, even if we are confused about when and where we are. We know,
without pausing to flgure it out, “This body is my own,” and “This hand and
this foot are both parts of my body.” We know very well that if we fail to plan
for future contingencies, our future selves may sulfer, and we care now about
that future self. Hume too knew all that. But he also thought that these rea-
sonable beliefs did not amount to an answer to his question “What is the self?”

So here is Hume’s conundrum: I think I am some?/;i«g, yet my self is not
anything that I can actually observe — at least not in the way that I can observe
pains or fatigue or my hands or my heart. So if my self is not an identifiable
experience, if it is not something I can observe, what is it? If the “self” is a
mental construction — a mode of thinking about my experiences — what are the
properties of this construction, and where does this construction come from?

In this century, we have the advantage of addressing Hume’s questions
within the framework of neuroscience. Thinking, we are reasonably sure, is
something the brain does. Therefore, thinking of oneself as a thing enduring
through time is also something the brain does.

At least in very general terms, therefore, we have an answer to Hume’s
question concerning where the constructed “self” comes from: the brain. Such
unity and coherence as there is in my conception of myself as a self depends on,
among other things, these neurobiological facts: (1) my body is equipped with
one brain, (2) body and brain are in close communication, and (3) activity in
diverse parts of the brain is coordinated at a range of time scales, from milli-
seconds to hours.

Metaphysics

Evolutionary biology, moreover, suggests a very general answer to the ques-
tion of why brains might construct a self-concept: it plays a role in the neuronal
organization used to coordinate movement with needs, perceptions, and mem-
ories. Such coordination is essential to an animal’s survival and well-being.
Coordination of functions ensures that inconsistent behaviors — fleeing and
feeding, for example — are not attempted at the same time. It ensures that a
hungry animal does not eat itself. For organisms with high-level cognition, self-
representational capacities constructed on the more fundamental platform help
us to think about the future, make useful plans, and organize knowledge (see
also section 1.3).

Still, these terms are very general, and much more detail is required to ex-
plain how brains work so that I can reflect on my motives, imagine myself
swimming, remember myself riding a bicycle, fall into deep sleep during which
my conscious sense of self vanishes, dream about flying, and wake up knowing
who I am. Much of the detail remains to be discovered, though what is known
so far permits us to sketch a basic framework for entertaining some reasonable
answers. It is also enough to allow us to design experiments that may unearth
more detailed answers concerning how the brain generates the / that I am.

In specifying the range of self phenomena to be addressed by neurobiology,
it may be useful first to consider how we routinely conceptualize the self In the
ground-clearing stages of our inquiry, determining what we believe about the
self by examining what we say about the self is a probe into the role of self
concepts in “coherencing” our inner life. How much of what we believe is true
is, of course, a distinct, empirical matter.

Frequently we use “self” to mean body, as in “I cut myself” and “I weighed
myself”; on other occasions, we mean to distinguish self from body, as when
you are exhorted, for example, to “talk to yourself” This ambiguity in the
word “self” rarely causes misunderstandings, since we share rich background
knowledge concerning when the word “self” does and does not refer to the
body.

In conversation about the self metaphors are the standby. Sometimes we use
object metaphors, as when we say we pushed ourselves to finish, pulled our-
selves together, fell apart, or tied ourselves in a knot. On the other hand, when
we say “I annoyed myself” or “I deceived myself” or “I talk to myself” the
person metaphor is invoked. ^

Using the self-as-person metaphor, people commonly describe themselves in
terms of a cluster of selves, such as one’s good self and bad self one’s shy self

Self and Self-Knowledge

and extrovert self, or one’s social and private selves, all of which are selves
belonging to “me.” My good self and my bad self are sometimes conceived of
as parts of the one thing, myself; sometimes as two persons in a group of many.
One may bemoan losing control of oneself or of being controlled by one’s
superego.

In describing character traits, one may refer to one’s real self, which one can
consider as hidden or revealed or transformed or inaccessible. What one con-
siders one’s real self is partly a cultural and a conventional matter, though only
partly. We do not usually talk about our “unreal self,” though we do admit to
masking or covering our true selves. Sometimes the self is conceived of as a
project, for example, when we undertake self-improvement or self-discipline.
Sometimes one’s self is analogized as a process, such as becoming mature or
wise.

Thus juxtaposed, these commonplace metaphors are strikingly diverse. A
space alien would be unable to extract from them an underlying coherent and
consistent prototheory about the self What the nonsystematic character of
metaphorical language suggests is that the self is not a thoroughly coherent,
single, unified representational scheme about which we have thoroughly coher-
ent, unified beliefs. Rather, the self is something like a squadron of capacities
flying in loose formation.^ Depending on context, it is one or another of these
capacities, or their exercise, to which we refer when we speak of the self. Some
of these capacities involve explicit memory, some involve detection of changes
in glucose or CO2 levels, others involve imagery in diverse modalities or emo-
tions of diverse valence. The fundamental capacity, however, probably consists
in coordinating needs, goals, perception, and memory with motor control.

1.2 Self-Representational Capacities

These considerations motivate recasting Hume’s problem in terms of self-
representational capacities. This removes the temptation to lapse into supposing
that the self is a thing, or if it is a representation, that it is a single representa-
tion. Self-representations may be widely distributed across brain tissue, coordi-
nated only on a “as needed” basis, and arranged in a loose hierarchy. We do
not understand yet exactly how all this works. But despite the large gaps in our
knowledge, adopting the terminology of representational capacities facilitates
the formulation of specific questions about the neural components that play a
role in some particular self-representational capacity or other.

Metaphysics

What are representations?

On the hypothesis that “the self” is actually a loosely connected set of repre-
sentational capacities, the chief workhorse in the account is representation.
Hence, we need to know what, in terms of neuroanatomy and neurophysiology,
representations are. How, in general, do brains represent? How can neural
states be representations of anything, in the sense that they point beyond
themselves to the thing they represent? These issues are addressed more fully in
chapter 7. In the meanwhile, we shall make do with a simple sketch to tide us
over.

To a first approximation, representations are states of the brain, such as
patterns of activity across groups of neurons, which carry information. For ex-
ample, a pattern of neuronal activity can embody information that something
hot touched the left hand, or that the head is moving to the right, or that food
is needed. We may consider a representational model to be a coordinated
organization of representations embodying information about a connected set
of objects and what happens to them across time. Thus a brain might have a
representational model of the body or of one’s hunting territory or of one’s clan
and the pattern of social relationships within it. A brain can also have models
of its own processes. If some neuronal activity represents a motor command to
reach for an apple, other neuronal activity represents the fact that a specific
command has been issued. If some neuronal activity represents a light touch
on the left ear, higher-order neuronal activity may represent the integration
of many lower-order representations (light touch on the left ear and buzzing
sound to left, which means that there is a mosquito, etc.).

The brain not only represents the sensations of one’s limbs; it specifically
represents the sight and feel of the limb as belonging to oneself (there is a mos-
quito on my left ear). Yet further neuronal activity may represent that repre-
sentation as a mental state (I know I feel a mosquito on my left ear). One’s
brain also has a model of one’s preferences (I know I prefer beets to cabbage),
one’s skills (I know how to tie a trucker’s knot but not how to play squash),
one’s memory (I do not know the names of my second cousins), even when one
is not now exercising those preferences and skills.

Self-representation, evidently, is not an all-or-nothing affair; it is not the
sort of thing that you either possess in its full glory or in no amount. Rather,
self-representation comes in grades, degrees, shades, and layers. The various
self-representational capacities are undoubtedly dependent on the task and the
context. For example, people normally express themselves very differently

Self and Self-Knowledge

when they are with family than when they are negotiating with business asso-
ciates. Certain self-representational capacities may wax or wane, depending on
neurochemical and endocrine conditions (e.g., depressed during abnormally
low serotonin conditions), behavioral state (being awake, being in a deep sleep,
or dreaming), task demands (during battle versus during rest), and immediate
history.

The multidimensional character of self-representation glimpsed in the array
of commonplace metaphors is well supported by neuropsychological and cog-
nitive data. Before tackling the neuroexplanatory task, therefore, it will be
useful to have a closer look at some of this data to understand a bit more about
how the multiple capacities can dissociate, malfunction, or deteriorate.

Autobiography and self

Memories of what I saw, felt, and did are part of the story of my life, part of
my autobiography. For each of us, one’s life story is an important part of who
one is now. As we shall see below, having an autobiographical memory is not
necessary for having a body representation. But is autobiographical memory
necessary for the conscious representation of oneself as an agent, as having
desires and goals, as being a person? The situation is complicated, as always in
biology, but the first-pass answer is “No.” This conclusion is based on studies
of patients who have lost essentially all autobiographical memory.

R.B., a patient studied in the Damasio lab in Iowa City, has profound
amnesia.^ Owing to an attack of herpes simplex encephalitis, R.B. suffered
massive destruction of both temporal lobes, including the overlying cortical
areas, as well as the deep structures including the amygdala and the hippo-
campus. He has no recollection of his past, save for one or two facts such as
that he has lived in Iowa. This aspect of his condition is known as “retrograde
amnesia.” Everything else — whether he was married and had children (he did),
whether he was in the army or went to college or owned a house — is beyond his
ken.

He has also lost the capacity to learn new things (and thus also has antero-
grade amnesia). He has a short-term memory of only about 40 seconds, less
if distracted. Undoubtedly, his sense of himself is different before and after
the onset of the disease, at least in the respect that he cannot recall anything
about himself or his life. Nevertheless, it is striking that R.B. does have im-
portant features of self-representation as a self among selves. That some self-
representational capacity endures is made evident by his being able, with no

Metaphysics

apparent effort, to refer to himself with the pronoun “I.” For example, he might
say, “I would like some coffee now,” or in response to a question about the
weather, “I think it is still snowing.” That some capacities are lost while others
remain reinforces the point that self-representation is a many-dimensioned phe-
nomenon, not an all-or-nothing phenomenon.

R.B. can attribute intentions and feelings to others, though his attributions
tend to be somewhat routinized. They are also slanted toward the positive
emotions, perhaps because he has lost tissue in a region of prefrontal cortex
known to have a role in negative feelings such as depression. Shown a picture
of a happy family at a birthday party, he will correctly describe it. Shown a
picture of a man striking a woman, who is shrinking from the blows, he
described the man as loving the woman very much and trying to help her stand
up. Although mistaken about their specific feelings, he does ascribe feelings to
them. Nor does R.B. ascribe negative feelings to himself. He always says he
feels just fine — not sad, lonely, disappointed or angry. Just fine\ This Pollyanna
bias may indicate some decrement in his self-representation insofar as these and
related data suggest that his ability to have those feelings is impaired. Insofar as
he is thus impaired, it is reasonable to say that his sense of himself as a person
is also impaired.

R.B. is a remarkable patient, and he does demonstrate, however impossible it
might have seemed antecedently, that one can have a basic sense of self despite
an essentially complete loss of autobiographical memory. Obviously, however,
his loss is truly devastating. He can neither reflect on his past nor wonder
whether some of his choices were regrettable, whether he has been self-deceived
about some things, or whether he squandered opportunities for self-realization.
He cannot reminisce about his children or even his own childhood; he cannot
strive to make up for lost time or make good past wrongs. He is, moreover,
unaware of his deficits. These capacities — to reminisce, self-refiect, know that
you have undergone a change — are very important aspects of self, and R.B.
has lost them all. He does not have a normal human sense of himself, but what
is remarkable is that the fundamentals of self-representation do persist, despite
the loss of life memories.

Depersonalization phenomena

A contrasting profile can be seen in certain schizophrenic patients. During
fiorid episodes, a patient may have good autobiographical memory, but may
suffer what are called depersonalization effects, where he is confused about self/

Self and Self-Knowledge

nonself boundaries. In recalling a florid episode during which she kept notes,
one computer scientist described her confusion as “not knowing where I
stopped and the world began; not knowing what, physically or mentally, was
me.” Afflicted with depersonalization, a schizophrenic may respond to a tactile
stimulus by insisting that the sensation belongs to someone else or that it exists
somewhere outside of him.

Auditory hallucinations, often considered diagnostic of schizophrenia,
may be a particularly striking example of an integrative failure in a self-
representational capacity. The hypothesis is that the “voices” heard by some
schizophrenics are in fact the subjects’ own inner speech or even their own
whispered speech not recognized as such.'^ A subclass of schizophrenics may be
quite deluded about their personal identity — about who in fact they are. For
example, a patient might be utterly convinced that he is Jesus, and he habitu-
ally conforms to his conception of Jesus’ demeanor, dress, and behavior. Aris-
totle, in pondering this phenomenon, accurately described such as person as
not knowing who he is.

Certain drugs can also trigger depersonalization phenomena.^ Surgical
patients given the anesthetic ketamine, for example, are susceptible to deper-
sonalization effects. A patient awaking from ketamine anesthesia may be con-
vinced she is dead or possessed or separated from her body or separated from
her own feelings. Similar depersonalization elfects may also occur in subjects
who have taken phencyclidine (PCP) or LSD. These elfects have prompted
some researchers to wonder whether some of the so-called “out of body” and
“near death” experiences of patients who are otherwise asymptomatic might
have fundamentally the same neurophysiological origin as disturbances caused
by ketamine or LSD (see chapter 9).

Parietal cortex lesions

A very different kind of abnormality in self-representational capacities may be
observed in some patients who suffer loss of sensation and movement in the
left-side of the body following a stroke in the region of the parietal cortex of
the right hemisphere. Such a patient may be adamant in denying that her left
hand or left leg is in fact hers, insisting that the limb really belongs to someone
else. This is termed “limb denial.” As one otherwise normal stroke patient
remarked about her left arm, “I do not know whose it is. Perhaps it belongs to
my brother, since it is hairy.” On occasion, a patient with limb denial has been
known to try to throw the arm or leg out of the bed, believing it to be alien.

Metaphysics

In a somewhat different profile, the patient with right parietal cortical dam-
age may recognize her left arm as her own, but deny that it is paralyzed. This
condition is called “anosognosia” (unawareness of illness). Ramachandran
studied one patient who, though very normal in other respects, believed that
her motor functions were entirely normal, and specifically that she could move
her paralyzed left arm and leg.® She nonchalantly explained her presence in the
hospital as owing to some minor problem. When asked to move her left hand,
she would cheerfully agree to comply, and a moment later when queried about
not complying, she would reply that indeed she just did move it. If asked to
point at Ramachandran’s nose, she would agree to do so, and later reply that
yes, she could see her hand pointing directly at his nose.

This was not a psychiatric effect in the traditional sense, but a compromise of
right parietal function. Remarkably, in some patients the denial can be made
temporarily to disappear using a minor intervention that increases the activity
in the right vestibular system. Using a technique developed by Eduardo
Bisiach, Ramachandran put cold water into the left inner ear of the patient,
thereby stimulating the right vestibular nucleus. For reasons we do not under-
stand, in these conditions, the patient’s anosognosia disappeared. During this
brief time, she acknowledged that she was in the hospital for a stroke, that her
left side was paralyzed, and that she could not move her left arm. As the effect
of the cold-water irrigation wore off, the anosognosia returned. Nor did she
remember her lucid description during the intervention.

Significantly, limb denial and anosognosia are rarely seen following a lesion
in the left parietal region, even though such a stroke results in comparable loss
of movement and feeling of the right side of the body. Nor is it seen with spinal
cord injuries, despite whole-body loss of sensation and movement. Christopher
Reeve, for example, is a quadriplegic and knows full well that his arms are his,
and knows he can neither move nor feel them.

Why is integrity of the parietal cortex of the right hemisphere necessary for
these aspects of self-representation? Although neuroscience cannot yet answer
this precisely, convergent evidence favors the hypothesis that an integrated
body representation may be fundamentally connected with integrated spatial
representations more generally. Because of a loss of movement and feeling, and
because of a breakdown of spatial integration, a patient’s brain may have no
basis for identifying the arm in her bed as hers. Evidence supporting this con-
nection to spatial representation derives from an array of sources, including
independent studies on split-brain subjects. These tests have shown that the
right parietal cortex in particular has a crucial role in spatial capacities,

Self and Self-Knowledge

including spatial memory, recognition of spatial configurations, and spatial
problem solving. The nature of this connection between normal body/self rep-
resentation and spatial representation is not yet understood, though as we shall
see in section 1.3, it would not be surprising if it were rooted in the brain’s
solution to the problem of generating appropriate movements to intercept objects
in space. (For further discussion of parietal-lobe symptoms, see chapter 7.)

The dementias

Dementia (Latin de mens, meaning mind) is an acquired loss of intellect. The
dementias are diffuse in their effects, meaning that widespread brain areas are
involved. There are many forms of dementia, and they differ in their causes
(e.g., infective agents, blows to the head, intoxicants), initial region of damage,
temporal course, and treatability.

It is the progressive forms of dementia that are particularly relevant to this
discussion. These diseases include Alzheimer’s, Pick’s, Creuztfeldt-Jakob, kuru,
HIV dementia, and Korsakoff’s (alcoholic dementia). In these dementias,
all capacities, including self-representational capacities, degenerate. Memory
losses are typical, and more recent autobiographical memories are lost before
very old memories. Patients tend to lose physical vigor, show gait disorders,
and have diminished language abilities. As the disease progresses, patients’
autobiographical memory fades, along with knowledge of their own lifelong
preferences. Personality changes are common, and are unpredictable in their
direction. There is relentless decline of specialized skills (e.g., carpentry, cook-
ing), social skills, and ultimately ordinary, everyday skills (e.g., buttoning a
shirt, tying shoelaces). Toward the end, patients become deeply confused
about who they are, as well as when and where they are. From observing an
Alzheimer’s patient over many years, it is reasonable to conclude that the
“self” progressively vanishes as the various self-representational capacities
disintegrate.

Anorexia nervosa

One very puzzling body misrepresentation is typical in subjects with severe
anorexia. An emaciated female may insist that she looks chubby and that she
needs to lose a few more pounds to look acceptable. No amount of reasoning
and explaining convinces her otherwise. What happens if you ask her to look at
her body in a mirror and make a judgment about whether she is fat? Even then

Metaphysics

patients will honestly and forthrightly size themselves up as plump, chubby,
tubby, and so forth. This seems impossible, given the unambiguous observa-
tional cues. Collectively, the data suggest that the disease involves a significant
impact on body representation. The etiology of anorexia, bulimia, and other
body dysmorphias (pathological distortions of body image) has not been
established but is under active investigation.’ There does appear to be a genetic
predisposition, but other factors are involved in precipitating the onset of ano-
rexia. One possibility under discussion is that in anorexic subjects, the nervous
system has an impaired ability to monitor internal milieu (see below, pp. 71-
76), resulting in a heightened anxiety that leads to loss of self-integration when
the brain’s standard strategies for reducing anxiety are ineffective. Whether this
is a cause, an effect, or even a consistent feature of anorexia nervosa is still
unsettled.

This discussion has touched on a fragment of the data relevant to the prop-
osition that self-representation is multidimensional. There are many other kinds
of cases I have not discussed, including brain-lesion subjects who have lost the
capacity to recognize faces and cannot recognize their own face in a mirror or
photograph, patients with focal brain damage who can no longer tell you which
digit is their index finger and which is the pinkie (finger agnosia), patients with
akinetic mutism, who have lost all inclination to say or do anything. My aim,
however, has been to illustrate the multidimensionality of self-representation
through a sample of cases of fragmentation and dissociation of self-representa-
tional capacities. The next matter to address is the neurobiological mechanisms
supporting self-representational capacities.

1.3 Self as Agent

The key to figuring out how a brain builds representations of “me” lies in the
fact that first and foremost, animals are in the moving business; they feed, fiee,
fight, and reproduce by moving their body parts in accord with bodily needs.
This modus vivendi is strikingly different from that of plants, which take life as
it comes. If an animal’s behavior is haphazard or incoherent, the animal tends
not to live long enough to reproduce. Consequently, an overarching demand on
any nervous system is that it appropriately coordinate the body: its movable
parts, its needs, its stored information, and its incoming signals.® This demand
is a powerful constraint on the evolution of neural organization.

Swimming or running, swallowing food or building a nest, stalking prey or
hiding from predators — smooth performance of any of these activities requires

Self and Self-Knowledge

tightly timed coordination of spatially dispersed muscle cells. Sometimes the
coordination must extend over long time periods, as in stalking and hiding.
Sometimes it can be handled by a reflex, such as ducking when a projectile
approaches your head. Because individual muscles move at the behest of motor
neurons, motor-neuron activity must be appropriately orchestrated. But to
serve survival, behavior must also be coherent relative to the animal’s needs
(don’t feed if you should flee), suitable to current given sensory signals (the
berries are too green to eat), and appropriate in view of relevant past experi-
ence (porcupines should be avoided; cover bee stings with mud) (figure 3.2).

Coordination can only be performed by neurons, since there is no intelligent,
extraneuronal “mini-me” inside who puts it all together. The intelligence of the
system has to emerge out of the patterns of neuronal connectivity, the response
properties of particular types of neurons, the activity-dependent modifiability of
neurons (learning), and a neuronal reward system for strengthening neuronal
connectivity when things go well and weakening connectivity when they go
awry. Given this very broad construal of coordination, this problem, one might
say, just is the problem of how the brain works. Perhaps that is true. Never-
theless, in addressing the nature of self-representational capacities, we shall
aim more modestly to characterize only gross aspects of the coordination
problem, leaving the vast wealth of related detail to be mined from the sug-
gested readings.

To depict the problem in its most fundamental aspects, we shall consider first
how the brain sets the animal’s basic goals and then how the brain solves the
basic problems of sensorimotor coordination so that it can achieve those goals.

Internal milieu, needs, and goals

In the nineteenth century the French physiologist Claude Bernard was in-
trigued by the seemingly trivial distinction between an animals’ external envi-
ronment and its internal environment. He was struck by the entirely nontrivial
fact that while the external environment can fluctuate a great deal, its inter-
nal condition is kept relatively constant. For example, human body tempera-
ture stays about 37° C, and a mere five degrees either way brings us death.
Homeostasis — the maintenance of a largely constant internal milieu — is a
buffer against environmental fluctuations. The brain keeps track of levels of
blood sugar, oxygen, and carbon dioxide, as well as blood pressure, heart rate,
and body temperature, in order to detect perturbations to the internal milieu
that are detrimental to the animal’s health. Deviation from the normal set

Metaphysics

Figure 3.2 Schematic representation of the perception-action cycle: the basic organiza-
tion of pathways at the cortical level. Notice that there are both forward projections and
back projections in both the sensory and motor cortical areas, and that beyond the pri-
mary sensory regions, there are connections between the sensory and motor regions.
Right: The basic organization of cortical motor connectivity and its subcortical connec-
tive loops. The complex, looping connectivity, including both cortico-cortical pathways
and cortico-subcortical pathways, undermines simple ideas about modularity and rigidly
staged feedforward processing, and suggests instead that the functional architecture of
nervous system “hierarchies” is actually very heterarchical. (From Fuster 1995.)

Self and Self-Knowledge

points cause an orchestrated set of neuronal responses that ultimately cause the
animal to seek either food, water, warmth, a hiding place, or the like, thereby
restoring deviant values to their normal values.

Homeostatic functions — and, in particular, the ability to switch between
the different internal configurations for fight and flight from that needed for
rest and digest — require coordinated control of heart, lungs, viscera, liver,
and adrenal medulla in a set of interconnected structures. In all vertebrates,
the brainstem is the site of convergence of afferents (input neurons) from the
viscera and the somatic sensory system, and it contains also nuclei for the
regulation of vital functions. In addition, the brainstem houses neuronal struc-
tures that regulate sleep, wakefulness, and dreaming, as well as structures
mediating attentional and arousal functions (figure 3.3). Antonio Damasio
(1999) has emphasized that this anatomical proximity of structures is an im-
portant clue to the coordinating role of the brainstem and its pivotal role in
self-representational capacities (table 3.1).

Maintaining a constant internal milieu means that the nervous system has
to “know,” in some sense, what the internal set points should be. This scare-
quoted sense of “know” can be cashed out in terms of patterns of neuronal
activity to which the system is rigged to return when perturbed. (More about
such patterns of activity will be discussed in chapter 7.) Additionally, the sys-
tem has to “know” what motor behavior would be effective in returning the
body to normal after changes in critical values. That is, the system has to be
rigged so that if low blood sugar is detected, for example, the animal’s nervous
system should prompt it to begin looking for food, not fleeing, unless fleeing is
currently necessary for survival. Pain signals should be coordinated with with-
drawal, not with approach. Cold-temperature signals should be coordinated
with shelter seeking, not with sleeping. Body-state signals have to be integrated,
options evaluated, and choices made, since the organism needs to act as a
coherent whole, not as a group of independent systems with competing interests.

By making some effects pleasant and some not, the nervous system directs
the animal’s choices. Emotions are the brain’s way of making us do and pay
attention to certain things. That is, they are assignments of value that direct us
one way rather than another, and they seem to have a role in every aspect of
self-representation, and certainly in body representation. Brief periods of oxy-
gen deprivation give rise to overwhelming feelings of needing air; extreme
hunger and thirst can make us feel so desperate as to banish all thought of
anything but water and food. Satisfaction is felt after feeding, sex, and

Metaphysics

Figure 3.3 A schematic representation of the main structures involved in coordinating
motivation, arousal, orientation, innerbody signals, musculoskeletal signals, and evalu-
ations of perceptual signals. The axis along the brainstem, hypothalamus, and cingulate
cortex constitutes the basic coordinating platform. The hypothalamus figures in basic
drives, such as sex, hunger, and thirst; the amygdala figures in processing fear evalutions
and responses; the nucleus accumbens is critical in feeling pleasure. The broken arrows
from brainstem structures to the sensory and motor corteces represent diffuse but very
broad modulating projections; the solid arrows represent more specific “information-
bearing” projections. For simplicity, only some of the pathways are indicated.

successful predator avoidance. More generally, self-preservation is underpinned
by powerful feelings.®

As neuroscientists have emphasized, this part of the system probably plays a
role not only in emergency situations, but also in providing assignments of
hedonic value in more humdrum categorizations, as when objects and events
are classed as desirable, nasty, familiar, novel, safe, dangerous, and what have
you.^° If you come into a shed and encounter a nasty smell, it will be recog-
nized first and fundamentally as dangerous, as the basic fear circuits respond
well before the cognitive niceties get deployed and long before “cool” reason
kicks in for impulse control.

Self and Self-Knowledge

Table 3.1 Main cell masses of the mesencephalon
Connections

Possible functional

Cell masses Afferent Efferent associations

Tectum

Tectum opticum (called
“superior colliculus” in
mammals)

Tegmentum
Torus semicircularis
(called “inferior colli-
culus” in mammals)

Nuclei III, IV (including
general somatic and
general visceral efferent)

Periaqueductal gray
matter, tegmental nuclei,
interpeduncular nuclei

Isthmo-optic nucleus
Nucleus isthmi (in
nonmammalian forms)

Reticular formation,
including tegmental
reticular nuclei

II, cord, bulb, sensory
nucleus of V, isthmus,
torus semicircularis or
inferior colliculus,
pretectum, thalamus,
telencephalon

Lateral line nuclei
(fish), cochlear nuclei
(tetrapods), vestibular
nuclei (fewer in higher
groups), cord, V sensory
nucleus

Vestibular nuclei,
cerebellum, tectum
(indirectly), reticular
formation
Complex, including
tectum, hypothalamus,
habenula, cord, telence-
phalon
Tectum

Tectum, probably torus
semicircularis

Cortex, pallidum,
reticular formation of
other levels, cerebel-
lum, vestibular nuclei,
cochlear nuelei, tectum,
cord

Cord, bulb, periaque-
ductal gray, reticular
formation, nucleus
isthmi, thalamus
(especially birds and
mammals), retina
(teleosts, amphibians)

Tectum, thalamus,
reticular formation

Extraocular muscles,
iris, and ciliary muscle
( parasympathetic)

Complex, including
nuclei of III, IV, VI,
pons, thalamus,
hypothalamus
Retina (in birds only)
Tectum, torus semi-
circularis tegmentum,
thalamus

Reticular formation of
other levels, thalamus,
cord

Correlation of visual,
auditory and somes-
thetic; feature extraction;
localizing stimuli
formulation of higher
reflex commands; eye
and head movements,
especially in orientation

Correlation of informa-
tion on equilibrium and
near-field aquatic dis-
placements (and electric
fields); sound sources;
localization
Movements of eyes;
accommodation;
pupillary constriction

Limbic system; affect,
visceral control

Horizontal cell response
Correlation of optic,
equilibrium, acoustic
influences

Motor control; pupil;
many other functions;
reticular activating
system

Metaphysics

Table 3.1 (continued)

Connections

Possible functional
associations

Cell masses

Afferent

Efferent

Red nucleus
Intermediate zone

Dentate, interposed
nuclei, precentral cortex
(somatotopically
organized)

Cord, bulbar reticular
formation, inferior olive,
cerebellum, thalamus
(especially from small-
celled newer part of red
nucleus)

Motor coordination,
especially righting; flexor
activity; well developed
in carnivores, poor in
primates

Substantia nigra (large
in man, small in other
mammals; only a
forerunner in reptiles)

Caudate, putamen,
subthalamus, pretectum

Striate, pallidum,
thalamus

Extrapyramidal motor;
inhibition of forced
movements; pathologic
in Parkinsonism

Source: Bullock 1977, table 10.5.

Moving, causing, and surviving

A brain that efficiently and accurately detects needs, but cannot orchestrate the
body to move in order to satisfy those needs, is a brain destined to be some
creature’s next meal. So how can a brain make successful limb and body
movements? How can sensorimotor coordination, both in short-term and in
long-term actions, be achieved by neurons? Bits of the story are known, many
of the details are not, and in this context, simplification of what is known is
unavoidable.

If you are a simple tube, such as a worm, with only circular and longitudinal
muscles and a few sensory neurons, the solution to the problem of how to move
is rather straightforward: you move down the gradient of good odors to get
food and away from noxious stimuli to avoid injury. Your behavioral reper-
toire and equipment is limited. For animals such as mammals and birds, how-
ever, the story must be gloriously complicated, given the range of components
that are coordinated by the nervous system and the complexity of the coordi-
nation needed to make a many-limbed body move in just the right way, at just
the right time, and over just the right time.

For engineering reasons, it behooves a nervous system to have an internal
representation of the body that is a kind of simulation of the relevant aspects of
a body’s movable parts, the relations between them, the relations to its sensory
input, and its goals. In this context, we shall talk in terms of an inner model

Self and Self-Knowledge

rather than an inner representation, since model talk allows us to interface with
research in control theory. This engineering subfield concerns how to organize
a complex system, such as automated take-off and landing controls in the Air-
bus, so that given the variable parameters of the system, the goals of the system
are reliably achieved.

Very roughly, the idea introduced by Daniel Wolpert is that one strategy
brains use to solve the coordination and control problem involves neuronal
simulations — inner models of the body.^^ Rick Crush calls these models emu-
lators}^ The engineering value of neural emulators springs from three main
sources. The first and most basic source of value derives from the fact that in-
put neurons, such as visual-system neurons, map the world in terms of sensory
structures, such as the retina, whereas the motor system maps the world in
terms of the body’s movable equipment — ^joint angles, muscle configuration,
and so on. Emulators can help make the sensory-to-motor transition. Emu-
lators also help you intercept a target you can no longer see and allow you to
imagine possible solutions to a problem. Finally, feedback from an emulator
can be many milliseconds faster than feedback through sensory systems, and
when time is of the essence, that can be a boon. We shall consider these more
closely.

Consider a sensorimotor problem. You see a plum on a tree, you are hungry,
and you want to grasp the plum with your hand. Simplified, the problem for a
nervous system is this: the visual system has a retina-based story about where
the plum is, but the motor system has to have a joint-angle story about where
the plum is, since it is the arm that must reach and the fingers that must
grasp the plum. So the motor system needs to know what joint-angle combi-
nation will serve to achieve the goal. Consciously, one never worries about this,
since the brain has the wherewithal for solving the problem without our con-
scious attention to it.

The easiest way to think of this problem is that the visual system represents
the spatial position of the plum in retinal coordinates, while the motor system
represents the plum’s position in joint-angle coordinates (figures 3.4 and 3.5).
On this construal of the problem, the brain’s task is to transform the position of
the plum in retinal coordinates into the position of the plum in joint-angle
coordinates so that the motor system can issue a command that will result in
the hand grasping what the eye sees.^^

More succinctly and more generally, the brain needs to be able to make co-
ordinate transformations so that the body is in the right configuration to get the
desire satisfied. The coordinate-transformation problem is referred to as the

Metaphysics

Figure 3.4 A cartoon characterization of an ultrasimple problem of sensory-motor co-
ordination. Consider a device with two rotatable “eyes” and an extendable “arm” with
one joint (A). In panel ( B), as the eyes triangulate a target (dotted lines) by assuming
angles (a,y9), the arm joints must assume angles (9,(p) so that the tip of the forearm
makes contact with the target.

kinematic problem, as it concerns only the path the arm and hand should take
and does not take into account such matters as the changes in momentum of
the arm, friction on the joints, or whether there are additional loads on the
arm, such as a bag hanging from the wrist. The part of the problem that does
take forces into account is called the dynamics, and will be set aside here as we
focus solely on the kinematic question.

The general form of the kinematic problem is well known to engineers. The
solution is to construct an inverse model, that is, a model internal to the system
that says (for goal g), “If I managed to get g, what command y would I have
used to get it?” A good inverse model will specify y, which then becomes the
motor command of the moment, and if all goes as planned, the plum is suc-
cessfully grasped. So now the question is this: how can a brain construct a

Self and Self-Knowledge

A The target in visual coordinates

B The target in motor coordinates

Upper arm angle 0

Figure 3.5 The respective configurations of the sensory and motor systems can be rep-
resented by an appropriate point in a corresponding coordinate space, as shown in (A)
for the eyes and (B) for the hand. The position of the target in visual space is not the
same as its position in motor space. Consequently, to make the arm go to the right place,
a coordinate transformation from visual space to motor space is needed.

successful inverse model to generate motor commands for a range of different
actions, including grasping the plum?

A computationally direct solution would be to completely specify all trans-
formations by a neuronal look-up table. For animals like us, alas, this solution
would involve excessive amounts of wiring and hence an excessively large head.
This is because the arm can have many different starting points, it can take any
one of many different adequate paths to the desired object from any given
starting point, and reaching the target may require moving the body as a
whole, and thus require appropriate leg movements and postural adjustments.

In addition to all that, life involves more than grasping plums. We sometimes
want to kick a ball, catch a fish, or climb the plum tree so we are within
grasping distance of the plum. And not only do we sometimes want to grasp

Metaphysics

the object we see, but we may need to do it while running, perhaps over rough
terrain. While running, we may need to throw a heavy object, such as a spear,
at another moving object, such as a deer. Moreover, the body changes its size
and shape during development or after accidents or with use of tools such as
knives and skis. So the look-up table would have to be bigger than gigantic.
ElRciency of wiring and flexibility of performance, therefore, demand a com-
pact, modifiable, accurate inverse model. How can a brain come by such a
desirable device?

One elegant engineering solution is to hook up a somewhat sloppy inverse
model with an error-predicting forward model and let the two converge on a
good answer.^"'' If, for example, the goal is to reach a plum, the inverse model
gives a first pass answer to the question. What motor command should be
issued to get my arm to contact the plum? Taking the command proposal, the
forward model calculates the error by running the command on a neuronal
emulator, and the inverse model responds to the error signal with an upgraded
command. The command from the inverse model need only be good enough to
get the hand very close, since on-line feedback can take over to make minor
corrections for the final few inches. If the forward model is also capable of
learning, this organization can be very efficient in acquiring a wide range of
sensorimotor skills (figure 3.6).

Brain circuits with forward models organized to work with inverse models
are emulators in Crush’s sense. With sufficient access to background knowl-
edge, goal priorities, and current sensory information, emulators can make a
wide range of relevant predictions. They can not only predict that on command
y your hand will miss the plum, but also that you will fall over, or that your
hand will contact nettles, or that if you grasp the plum you will pitch forward,
and so on. Probably, predictions as fancy as this will not be featured in a
leech’s nervous system, for example, but the bet is that they are found in the
brains of mammals and birds, among other animals.

The second value of the emulator is that it allows the brain to make an ap-
propriate movement even after the target has become invisible. This could
happen because, for example, the lights have suddenly gone out, or because the
early stages of executing a plan require a whole-body movement, during which
the target becomes occluded. It might routinely occur when an animal looks
into a cavity for birds’ eggs or into a hole for gophers and then uses its fore-
limbs to go after what it can no longer see. More generally, it can occur when-
ever you have a plan for the future where the target is not currently visible.

The third important point about emulators is that they can be engaged “off-
line” in evaluating very abstract behavioral options, such as shooting the rapids

Self and Self-Knowledge

Figure 3.6 An inverse model is connected to a forward model (the emulator). The in-
verse model gives a first pass answer to the question, What motor command will get my
arm to the plum? The inverse model proposes an answer and sends out a command
proposal to the forward model, which then calculates the error by running the command
on the neural emulator. The inverse model then responds to the error signal with an
upgraded command.

or portaging around them. Off-line imagining yields an advanced peek at the
likely consequences of pondered actions, which permits undesirable con-
sequences of contemplated actions to be foreseen and avoided. So, for ex-
ample, one may envision the risk of capsizing the canoe in rough water, as
against the many hours portaging up a steep grade and through dense bush.
Eventually, one’s brain settles on a decision, whereupon more detailed actions
can be planned, such as pulling out the canoe, hoisting it up on one’s shoulders,
and so forth. In the planning phase, the motor signals generated by the inverse
model are merely “what if” motor commands, not full-blown commands to
move.^® Off-line planning also permits an animal to prepare to intersect targets
that do not currently exist but are expected at some point in the future. Thus a
bird can build a nest, or a wolf pack can plan to intercept the annual migration
of caribou across a certain point of the Alsek River.

It is evident that humans regularly use sensory and motor imagery to work
out in their brains solutions to problems, both highly abstract and somewhat
concrete, before implementing the solutions in the world. Contemplating a

Metaphysics

Steep, icy ski slope, one’s emulator will make motor predictions about the like-
lihood of losing control. In building a shelter, one envisions before construction
what would be a suitable location, what materials are available, how it should
be structured to withstand wind and rain, and so on. As Crush has stressed,
this form of problem solving is essentially the brain manipulation of a body
image. Envisioning how to answer questions in an interview may be little
different, but probably draws on many of the same operations.

The final point concerns speed and the fact that in a competitive world,
speed matters. It is not the only thing, but it is an important factor. Feedback
concerning the consequences of the execution of a plan will come faster from
the emulator than from the body itself. It takes time for a motor-command
signal to reach the various muscles, for the muscles to change, and for feedback
signals from the muscles, tendons, and joints to return to the brain. If visual
feedback is used, it takes time for signals to be processed in the visual system,
which is relatively slow, since the retina takes about 25 milliseconds of pro-
cessing time. Especially when the animal is large and the distance between limb
and brain is on the order of meters, (as in humans, whales, and elephants),
feedback that comes faster than what is available via the perceptual route is
desirable. This shortcut may give the brain an additional 200-300 milliseconds,
and when getting the timing exactly right is important, those milliseconds can
make the difference between success and failure.

So far emulators have been discussed in terms of their engineering virtues.^®
What is the evidence that brains do in fact uses emulators? Neurobiological
studies at the level of the single neuron and the network strongly suggest that
posterior parietal cortex and area VIP do execute transformations from visual
to motor coordinates.^® More correctly, it seems likely that this region takes
information from a range of sensory systems — visual, auditory, vestibular,
somatic sensory — and converts it into eye-centered coordinates, head-centered
coordinates, body-centered coordinates, and world-centered coordinates, de-
pending on the body’s starting configuration and the brain’s goals (see figures
3.7 and 3.8). Pouget and Sejnowski have constructed a convincing artificial
neural network showing exactly how this could be done.^° In view of the sup-
porting physiological evidence, they propose that real neural nets in this region
are disposed to represent “where perceived objects are in my-body [egocentric]
space,” as well as where objects are in allocentric space.^^ (See pp. 309-312.)

Additionally, in this region and also in the dorsolateral region of the frontal
cortex, to which the posterior parietal region projects and from which it gets
signals, there are neurons that hold a target location on-line even when the

Self and Self-Knowledge

Movement

Figure 3.7 Coordinate transformations required to specify an arm movement toward a
visual target. The position of the target on the retina is specified in retinotopic coor-
dinates. This position needs to be remapped in joint coordinates to move the arm to the
corresponding spatial location. This transformation can be decomposed in a series of
subtransformations in which the target position is recoded in various intermediate
frames of reference. (From Pouget and Sejnowski 1997.)

Metaphysics

Head-Centered \

Map |00000»0000|

A X

Figure 3.8 A neural network for transforming a retinotopic map to a head-centered
map. The input contains a retinotopic map of the visual input, and the output consists of
a head-centered map. The eye-position units have a sigmoidal tuning to eye position and
a range of thresholds. The function represented by the network is nonlinear, as illus-
trated by the fact that the response to and of the units in the output layer is clearly
not a plane. This mapping could be implemented by middle-level units that compute the
product of a Gaussian of with a sigmoid of Cx- Such units would provide basis func-
tions of the input variables and would respond like gain-modulated neurons found in the
parietal cortex. (From Pouget and Sejnowski 1997.)

Self and Self-Knowledge

target itself has disappeared, as the emulator hypothesis predicts.^^ Moreover,
lesions to parietal cortex cause misreaching and other disturbances of visually
guided grasping, such as misshaping the hand to a target shape. Other struc-
tures directly involved in emulator function include the cerebellum and the
basal ganglia.

Additional evidence for the existence of brain emulators derives from psy-
chological experiments concerning eye movements. Here is how the story goes.
Without conscious commands, our eyes constantly scan the environment,
moving about three times per second along paths that maximize task-relevant
visual information. These eye movements are known as saccades. Other eye
movements involve tracking an object, when the object is moving or the subject
is moving, or both. Such tracking is known as smooth pursuit. Even though the
retina registers huge shifts in light patterns owing to all this eye movement,
stable objects in the visual scene appear to remain stable. That is, the brain
interprets retinally detected motion as motion of our eyes, not motion in the
world. Moreover, if things in the world are moving, the brain can distinguish
between motion due to object movement and motion due to eyeball movement.
So as I move my eyes around a scene, I can tell the difference between move-
ment of the dog and movement of my eyes. Although the computations re-
sponsible for this result must be exceedingly complex, at a conscious level
the achievement is effortless. How does the brain make these very important
distinctions?

The brain undoubtedly uses emulators. It knows from a copy of the eye-
movement command {efference copy), sent to the forward model among other
places, whether or not the eyeballs were commanded to move, and in what
direction. Here is one small piece of evidence for this hypothesis. Suppose that
the brain used only feedback from the eyeball muscles to know whether the
eyeball moved. On this supposition, as Helmholtz (1867-1925) rightly rea-
soned, if you close one eye and passively move the open eye by gently pressing
on it, you should still see stationary objects as stationary and see the movement
as due to eye movement. This is not what happens. As you can test for yourself,
stationary objects actually appear to move when you gently press the eyeball.
This simple experiment provides evidence in favor of efference copy: in the
passive-movement condition, no eye-movement command from the brain exists
to “explain” the stimulus movement with respect to the retina, so world move-
ment is perceived.

In a more decisive, but also more invasive, experiment, John Stevens and
colleagues reported in 1976 that they had used a pharmacological agent to

Metaphysics

paralyze the eye muscles. This experiment is a control condition for the Helm-
holtz passive-movement condition, for here the intention to move the eyes
exists, but the eye muscles cannot move. Three subjects (Stevens and two col-
leagues) sit and look at an object, say a colfee cup. At some random time, the
subject looks to the right, or rather, he intends to look to the right. Because of
the paralysis, the eye muscles cannot respond to the command, and hence the
eyeballs do not move. Thus he continues to see the coffee cup. As each subject
reports, however, something interesting does happen; the subject visually expe-
riences the whole scene jumping to the right. Why?

This stunning perceptual effect is at least partially explained on the efference-
copy hypothesis, and hence on the emulator hypothesis. Crudely speaking, the
brain thinks this: “I issued a command to move the eyes to the right, yet the
colfee cup is still in full view. That can only be because the whole scene — colfee
cup included — moved when the eyeballs moved.” In short, the brain makes a
prediction about a change of scene based on the eye-movement command,
which hitherto has always been followed by real eye movement. When the pre-
diction fails, the brain grabs the “best explanation.” Stevens’s experiment is
thus not only evidence for the role of elference copy; it is also an important
illustration of how the brain’s eye-motion emulator can have a powerful effect
on sensory experience itself. Incidentally, if one supposes that sensory systems
essentially mirror reality, with no top-down coloration, Stevens’s result is a
brilliant falsification of that supposition.

Is there evidence of emulators used for off-line problem solving? A rather
striking example of problem solving that probably involves off-line manipula-
tion of the body image is seen in ravens. The ethologist Bernd Heinrich tightly
tied a piece of meat to one end of a length of twine (about three feet) and tied
the other end to a trapeze. One at a time, hungry ravens were released into
the room. A bird’s only successful strategy for getting the meat is to sit on the
perch, draw up a length of twine with its beak, step on the twine, and repeat the
procedure about seven times until the meat is level with the perch. In other
words, this problem cannot be solved in a single step, and no reward is
obtained until all seven steps are complete. So a simple response-reward learn-
ing device will not find the solution. This is not a problem the birds encounter
in the wild, and hence the problem is novel.

When Heinrich did this experiment with a crow, invariably the crow would
fly at the meat and try to snatch it. This strategy is hopeless, and the crow suf-
fers the discomfort of a jerked neck. Heinrich observed a dozen crows, one by
one, stuck on this hopeless strategy, never managing to figure out how to get

Self and Self-Knowledge

the meat. So the solution is not obvious (whatever that means), even to a bird
as bright as a crow.

Ravens are legendary for their cleverness, and they did indeed respond in
a very dilferent fashion. Of six ravens, one was string-shy and would not
approach string in any condition. (Like many intelligent animals, ravens ap-
parently have irrational fears and phobias.) Five ravens solved the problem
within live minutes of being allowed into the area, and their strategies for
solving the problem followed much the same order. First, they spent a little
time just looking at the setup. Second, they pecked at the string where it was
attached to the trapeze as if trying to sever the string. No raven performed this
act on string to which no food was attached. Third, they grabbed the top of the
string and twisted it violently from side to side as if trying to break it off.
Fourth, they reached down to pull up a length of string, stepped on it, and re-
peated the procedure until the meat was at their feet. This fourth procedure
took 10-20 seconds. If Heinrich shooed them off the trapeze after they had
performed the pull-up, they dropped the meat and flew nearby. As soon as it
was safe, they returned and straightaway repeated the pull-up procedure. If,
out of sight, Heinrich rearranged the string with a pulley so that the raven had
to pull down on the string to get the meat to come up, they easily switched
modes. If the string and meat were merely laid on the trapeze but not attached,
the birds directly snatched the meat and flew off.

That none of the ravens got its neck jerked by going for the attached meat
while flying suggests that their brains expected what would happen were they to
do that and decided against it. That the ravens turned to the pull-up strategy
within minutes and were successful in pulling up the meat in one trial strongly
suggests that the ravens used body-image manipulation in causal problem
solving. As Heinrich argues, “The simplest . . . hypothesis is that the birds an-
ticipated at least some consequences of the behaviors before overtly execut-
ing them.”^"*^ The emulator hypothesis gives a very plausible explanation of this
remarkable problem-solving behavior, especially in the context of independent
evidence for neural emulators of the body.

Off-line emulation also appears to have a significant effect in skill acquisi-
tion. For example, covertly practicing a golf swing — going through the move-
ments in imagination — does improve the swing at a rate greater than doing
nothing, and almost as well as actually practicing.

Yet another bit of psychological evidence for the existence of emulators
comes from the difficulty of tickling yourself. Being tickled by someone else
feels very different from tickling oneself. As in the paralyzed-eye-muscle

Metaphysics

experiment, here too something about the brain’s internal representation of the
motor-intention signal affects the feeling itself that results from the touch.
Moreover, this is true even when the touching device is not your own hand,
which would provide sensory feedback and thus clue the brain in, but a lever
with a feather that you can move. When the subject moves the lever, the touch
still feels different from when someone else moves the lever, even though you
get the very same stimulus.

There is, however, a way to fool the brain. Sarah Blakemore and her collab-
orators rigged a self-touching device so that the experimenter can put a delay in
between when the subject pushes the lever and when the subject is touched.
The experimenter can also perturb the trajectory of the lever. The protocol is to
interleave self-touching trials with other-touching trials. When the experimenter
inserts a delay and/or the trajectory of the lever is perturbed, it feels to the
subject as though someone else were touching him, even when it is his move-
ment that causes the tickle.

Why should we be able to feel the touch as an other-touching stimulus under
these conditions? The most likely hypothesis is that in normal conditions the
brain’s action emulator says, to put it crudely, “Got a copy of the intention-to-
tickle-left-foot, so the left-foot-touch is my own.” When there is a delay or a
perturbation, the brain thinks, “Well, that can’t be me, because my command
would have been executed earlier.” This representation of the intention, along
with representation of the normal time for execution of the intention, affects the
actual feel of the stimulus. This is extremely important, since it shows that the
brain’s body model includes temporal parameters. Notice that this effect also
demonstrates again that experience itself can be altered by the cognitive repre-
sentation of an intention.^®

Other experiments tapping into the brain more directly indicate that in gen-
eral, intentions to move are integrated into the brain’s ongoing self/body
model. Suppose you raise your left hand. A number of brain areas contribute
to achieving the effect. These include the supplementary motor area (SMA), the
premotor cortex (PMC), the primary motor area, parts of the cerebellum, and
the somatosensory cortex (see figure 3.9). Studies using activity-sensitive mag-
netic resonance imaging technology (functional MRI) show that when you
merely imagine that same action, the same motor areas (the SMA and PMC)
are active.

Not surprisingly, activity in the somatosensory cortex is then much dimin-
ished, since there are no afferent signals from the extremities to indicate
changes of muscles, joints, and tendons. Incidentally, other areas that show

Self and Self-Knowledge

Supplementary

motor area Primary

\ motor cortex

Supplementary
motor area Primary

Premotor motor cortex

Monkey brain

Figure 3.9 The location in the monkey and human brains of primary and supplemen-
tary motor areas of the cortex and posterior parietal cortex.

Metaphysics

heightened activity during visual imagery, such as parts of the visual cortex,
were not above baseline in the motor imagery task. To control for nonmotor
effects, subjects were trained to avoid all visual images when imagining their
movements, attending only to kinesthetic images (arm-moving feelings). To
control for actual motor signals activating muscles, subjects were also trained
to keep their muscles relaxed when imagining a hand movement.

This range of data fits well with the hypothesis that the brain has a model
integrating body configuration, movement decisions, and expected results of
intended movements, all of which are time-sensitive. In other words, the data
support the hypothesis that neural emulators in the sense discussed exist in the
brain and are a watershed for self-representational capacities in general. This is
not to suggest that the emulators are the self, that they are the little person in
the head we jokingly envisage. Rather, emulators are one component in the
story of our self-representational capacities.

In animals with large brains, such as humans, there will be coordination on a
grander scale than in rats. Highly sophisticated coordinative functions ulti-
mately yield fancy results, such as impulse control, long-term planning, richly
detailed autobiographies, and imaginative explorations that stir emotions. At
this level, where there are representations of representations of representations,
and so on, we come upon those human self-representational capacities about
which we typically converse. These high-level networks embody one’s long-
terms plans, as well as one’s preferences, skills, attitudes, and temperament. At
bottom, however, what anchors self-representational capacities is the neuronal
organization serving coordination and coherence in “making a living,” so to
speak.

In this section I have stressed the role of intentions-to-move, while helping
myself to the assumption that the emulator has available to it a rich supply of
signals regarding the soma (Latin: body) — its postural configuration, its loca-
tion relative to other objects, its sensations and perceptions. Obviously, these
signals are extremely important. If I need to fiee, my motor system needs to
know what my current body configuration is, since the motor commands will
be different depending on whether I am starting from a sitting, standing, or
crouching position. If I am to learn skills such as being able to climb a tree or
throw a rock, my brain needs feedback from the joints, tendons, and muscles. If
I am to avoid bodily injury, I need to know when and where it hurts, whether
and where it is hot or cold. All this requires a sensory system that informs the
brain about what is going on body wise. Now we need to look in more detail at
the nature of the sensory information the brain gets about the body.

Self and Self-Knowledge

2 Inner Models of Body, Self, and Others

2.1 Sensory Systems Representing the Body

The nervous system is generally considered to have two main systems whose
function is to represent the body: the somatic sensory system, which has recep-
tors in the muscles, joints, tendons, and skin, and the autonomic system, which
innervates the cardiovascular structures; the bronchi and lungs; the esophagus,
stomach, and intestines; the kidney, adrenal medulla, liver, and pancreas; the
urinary structures; and the sweat glands in the skin. The genitalia are inner-
vated by both, but the labor is divided. For example, the autonomic system
stimulates erection and ejaculation, the somatic system carries signals of touch,
pressure, and so on.

The two systems follow different pathways from the extremities to the brain,
and they appear also to be different in their effects on conscious awareness. For
example, while both the tongue and the stomach have movement receptors, one
can be aware of the movement of one’s tongue, but one is not aware of the
stomach’s peristaltic movements. The visceral system has a motor subsystem
that functions, for example, in sweating, secreting tears, and changing the heart
rate. By contrast, motor control for the skeletal muscles is a system distinct
from the somatic sensory system, though of course the two are integrated at
various levels from the spinal cord to the cortex.

In the next two sections, we shall consider the somatic sensory and visceral
systems to see in a bit more detail what is known about how they contribute to
one’s sense of oneself.

The somatic sensory system

Patients with damage to their right parietal cortex teach us that although nor-
mally nothing could be more obvious than that this arm is mine, nevertheless
this is a judgment that the brain has to construct. How does my brain know my
body’s position? How does my brain know whether something touched my
body or whether it touched itself? The rough answer is that nervous systems
have highly specialized structures for detecting and transmitting signals; they
have highly organized wiring for connecting body to brain and brain to body.
The body-to-brain wiring keeps the brain informed about what is happening to
the body, while the brain-to-body wiring allows the brain to control the body.

Metaphysics

The patterns of connectivity yield representational models of the body that
keep track of the schedule of motor commands along with changes in body
configuration, body contact, and body needs.

The somatic sensory system is the nervous system’s primary device for telling
the brain how the body is configured, whether it has been harmed, whether it
is in contact with other objects, and what the features are of any contacted
objects. It is actually a rather diverse system comprising four submodalities,
each specialized for detecting a distinct signal type. The basic subdivisions are
light touch and pressure, temperature, proprioception (from joints, muscles,
and tendons), and pain (nociception), which has various subsystems of its own.

Each submodality can signal the intensity of the stimulus, its duration, and
its location on the body surface. Complex stimulus properties, such as textures
(e.g., rough or smooth), spatial configurations (e.g., curved or straight), and
tactile recognition (e.g., feels like a paper clip) depend on combinations of
neuronal responses. One particular combination of signals will represent that
there is something fuzzy crawling up my left ankle; a different combination will
represent that something cold and hard is pushing on my left ankle.

Each submodality has its proprietary pathway leading from the location of
the receptor in the body to the spinal cord. Maintaining their submodality
specificity en route, spinal neurons project to regions in the brainstem, the tha-
lamus, and the cortex (figure 3.10). Within the thalamus, each submodality has
a proprietary region where it makes synaptic contact. The next set of projec-
tions goes from the thalamus to the cortex, with axons from each submodality
clustering together in a typical path.

Distinct pathways of fibers carrying signals about body parameters are
mapped in an orderly way in the spinal cord, brain stem, hypothalamus, tha-
lamus, and several cortical regions (the insula, S2, SI, and the cingulate).
Adjacent neighborhoods at the periphery are mapped to adjacent neuronal
neighborhoods; e.g., the arm representation is next to the hand representation,
the middle-finger representation is between the index-finger and ring-finger
representations, and so on. In this sense there are maps — literal maps — of the
body in the brain stem, with successive remappings at a series of higher-level
structures (figure 3.11).

Specialized receptors in the skin perform different jobs and give rise to dis-
tinct sensations. Hairy skin, such as that on the back of the hand, contains
receptors that wrap around the hair follicle and respond to movement of
the hair. This is the principal device for signaling very light touch on hairy
skin. In some animals, muzzle whiskers are highly sensitive and provide

Self and Self-Knowledge

Figure 3.10 The whiskers of the rats are represented at many stages in the nervous
system. (A) The afferent pathways from whiskers to the brainstem, to the thalamus, and
then to the neocortex. (The cerebellum is removed to allow the other structures to be
visible.) (B) At each stage the order and arrangement of the whiskers on the face is pre-
served in the order and arrangement of the neurons. In the brainstem, the neuronal
groups representing individual whiskers are called barrelettes, in the thalamus they are
called barreloids, and in the cortex they are called barrels. Abbreviation: TG, trigeminal
nerve. (From Gerhardt and Kirschner 1997.)

Metaphysics

Area 6 ^rea 4

Figure 3.11 A schematic representation of the path taken by somatosensory signals to
the motor system. Input signals make synaptic connections in the ventrobasal complex
of the thalamus, which projects to the topographically mapped areas of somatosensory
area SI. From there, signals are mapped in somatosensory area SII and the posterior
parietal areas. The next stages are (1) the limbic structures (the entorhinal cortex and
hippocampus), where signals engage memory functions; (2) the limbic structures (the
amygdala, cingulate cortex, hypothalamus), which play an evaluative/cognitive role; (3)
the polysensory cortex in the superior temporal gyrus; and (4) the motor system (the
primary and supplementary motor cortex), where continuous sensory feedback to the
motor system occurs.

Self and Self-Knowledge

Merkel disks

Meissner's
corpuscles
Free nerve endings

Duct of sweat gland

Pacinian corpuscles

Figure 3.12 The location and morphology of mechanoreceptors in hairless (glabrous)
skin of the human hand. Receptors are located in the superficial skin, at the junction of
the dermis and epidermis, and more deeply in the dermis and subcutaneous tissue. The
receptors are Meissner’s corpuscles, Merkel disk receptors, and bare nerve endings.
Subcutaneous receptors include Pacinian corpuscles and Ruffini endings. Nerve fibers
that terminate in the superficial layers of the skin are branched at their distal terminals,
innervating several nearby receptor organs; nerve fibers in the subcutaneous layer
innervate only a single receptor organ. The structure of the receptor organ determines its
physiological function. (Based on Goldstein 1999.)

important data on such things as burrow diameter. Some animals, such as
rodents can also move their whiskers to actively reach out for additional
information.

Glabrous (nonhairy) skin, such as that found on the palms of the hands,
contains two types of receptors specialized for responding to touch in the ab-
sence of hairs; Meissner’s corpuscles and Merkel disks (figure 3.12). These differ
in their response style: A Meissner’s corpuscle is fast adapting, which means
that it responds abruptly to stimulus onset and then stops responding, even if
the stimulus continues. By contrast, a Merkel disk is slow adapting, which
means that, to a prolonged stimulus, it responds at stimulus onset and con-
tinues to respond to continued stimulation, though with a diminished frequency
of firing.

Metaphysics

® Curvature
(1/in)

HI n Mil I

1 1

Firing rate
(spikes/s)
2.0

• 2 .

nil II 1 1

1 1 1

1 1

4.3

# 2.7 .

mill MM

1 1 1 1 1 1 1

1 1 1 1 1 1 II

1 II 1

8.3

• 4 .

1 mil 1 II 1 II

1 1 1 II 1 II

IIIIIIIII

II II

11.3

• 8 .

IIIIIIIIIIIM 1

1 1 1 1 1 1 1 1 1 1 1

INI 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1

17.0

• 16 .

iiiiiiiiiiini II

iiiiiiiii Hill

Mil III lllllllllll

1 II 11 1 1 1 1 1

23.7

? 20r

o*-

Time (s)

Figure 3.13 The sensory systems encode four elementary attributes of stimuli —
modality, location, intensity, and timing — manifested in sensation. The four attributes
of sensation are illustrated in this figure for the somatosensory modality of touch. (A)
In the human hand the submodalities of touch are sensed by four types of mechano-
receptors. Specific tactile sensations occur when distinct types of receptors are activated.
The firing of all four receptor types produces the sensation of contact with an object.
Selective activation of Merkel disks and Ruffini endings produces sensations of steady
pressure on the skin above the receptor. When the same patterns of firing occur only in
Meissner’s and Pacinian corpuscles, the tingling sensation of vibration is perceived. (B)
Location and other spatial properties of a stimulus are encoded by the spatial distribu-
tion of the population of activated receptors. Each receptor fires an action potential only

Self and Self-Knowledge

Beneath the skin are found two additional types of receptors that respond to
mechanical deformation and contribute to the sensation of touch as well as
to the sensation of pressure. These are the Pascinian corpuscles (fast adapting)
and Rulhni endings (slow adapting). The packing density of these various
receptors varies over different regions. The greater sensitivity of the finger tips
relative to forearm, for example, is due to the higher density of Meissner’s and
Merkel receptors in the finger tips. All of the receptors also vary their response
as a function of stimulus intensity (figure 3.13).

Feeling temperature depends on two distinct classes of receptors: those for
warm stimuli and those for cold stimuli. A priori, one might have imagined
that one receptor type would suffice, with the whole range of hot to cold sig-
naled by variations in code. That is not, however, the solution evolution lit
upon.

Cold receptors in the skin are very sensitive to small decreases in tempera-
ture, and they discharge in proportion to any drop in temperature from the
normal baseline (34° C). Their standard detection range is between 1° and 20°
below the normal baseline. Below that, their response to cold temperature falls
off. Warm (higher than baseline) receptors discharge when the stimulus is be-
tween about 32° C and 45° C, falling off thereafter. How is it that we can feel
very hot things (above 45° C) as hot? Mostly because there are pain receptors
that activate at higher temperatures, but also because warm receptors located
at some remove from the hot stimulus source will respond to the diffuse
warmth. Though one cannot tell by introspection, high temperature and ex-
treme heat are mediated by different submodalities. A chemical found in chili
peppers, capsaicin, will selectively depolarize the warm receptors, and hence we
feel a hot sensation, even though nothing is actually hot. Conversely, menthol
selectively acts on the cold receptors, generating a feeling of coolness.

when the skin close to its sensory terminals is touched, i.e., when a stimulus impinges on
the receptor’s receptive field. The receptive fields of the different mechanoreceptors —
shown as shaded areas on the fingertip — differ in size and response to touch. Merkel
disks and Meissner’s corpuscles provide the most precise localization of touch, as they
have the smallest receptive fields and are also more sensitive to pressure applied by a
small probe. (C) The intensity of stimulation is signaled by the firing rates of individual
receptors, and the duration of stimulation is signaled by the time course of firing. The
spike trains below each finger indicate the action potentials evoked by pressure from a
small probe at the center of the receptive field. Two of these receptors (Meissner’s and
Pacinian corpuscles) adapt rapidly to constant stimulation, while the other two adapt
slowly. (From Kandel, Schwartz, and Jessell, Principles of Neural Science [2000].)

Metaphysics

There is also a phenomenon called paradoxical cold, which can be experi-
mentally induced thus: apply a very hot stimulus (above 45° C) to a punctate
region of skin containing only receptors for cold. Though the stimulus is in fact
hot, it will be felt as cold. This is because cold receptors will respond to a very
hot stimulus. Since they are “wired to report coldness,” to put it crudely, the
very hot stimulus is felt as cold.

When you put the flat of your hand against a granite cliff-face, the sensation
itself may seem to be “holistic” or seamless, not a vector with many com-
ponents. In fact, however, the responses of a variety of skin receptors are
involved. Cool receptors will indicate one property, Pascinian corpuscles will
discharge briefly in response to the pressure of the rock against your hand,
Ruflini receptors will also respond, but only for the duration of the pressure.
If you press hard, pain receptors will also respond. If you lay your hand on
gently, Merkel’s receptors will respond to touch, as will the fast adapting
Meissner’s receptors. This chorus of responses at the periphery will be con-
veyed to the brain and will allow you to have the feeling of roughness, coolness,
and solidity that you can identify as a rock surface.

Receptors in the muscles and joints serve to update the brain on the posi-
tion and movement of limbs (proprioception). These fiber pathways can be
destroyed by diseases (peripheral neuropathies), and a patient whose proprio-
ceptive system is damaged is seriously debilitated. They often have great diffi-
culty with even simple motor tasks, such as walking, because they do not know
where their legs and feet are. To walk, these patients have to use their eyes to
determine their body configuration. If a subject with peripheral neuropathy is
standing in a room when the lights go off, he will be unable to maintain posture
and, unless aided, will tend to fall in a heap. In their daily business, normal
subjects take all the incoming proprioceptive signals for granted. We scarcely
know we have this precious modality, and some philosophers have denied that
they are aware of proprioceptive signals, suggesting that they know without
awareness the position of their limbs. Be that as it may, disruption to proprio-
ception reminds us of how crucial these signals are to our body sense as well as
to our ability to move our limbs and whole body as we intend.

Representations of head movements are rather special, because they include
signals from a very specialized structure in the inner ear, the semicircular
canals. The three semicircular canals, roughly at right angles to each other,
detect head movement and are crucial to maintaining balance and posture. The
canals are filled with fluid, and fine hairs project into the fluid. When the head

Self and Self-Knowledge

Pitch: rotation
around y axis

Roll: rotation
around x axis

Figure 3.14 A schematic representation of the vestibular organs of the inner ears. Each
organ consists of three semicircular canals oriented approximately 90° with respect to
each other. The eanals are rigid tubes filled with fluid, and movement of the fluid is
detected by hair cells in a membrane stretched across the tube.

moves, the canal moves relative to the fluid, which tends to stay put due to
inertia. Hence, the hairs move in the fluid, and their resulting deflection depo-
larizes the receptor in which each hair is embedded (figure 3.14). The integrated
signals from the three canals tell us whether the head is moving, and in what
direction, relative to absolute space. Receptors in the neck muscles also con-
tribute to head-position representation, this time, relative to the trunk.

Adaptive effects are commonplace. As is well known, if you put one hand in
ice water and one hand in hot water for a few minutes and then plunge them
both into a pail of tepid water, to each hand the very same water feels a differ-
ent temperature. It feels quite warm to the hand originally immersed in ice
water and quite cool to the hot-water hand. Adaptation effects also take place
over a long time period. Normally, one always depresses the clutch on a car

100

Metaphysics

with the left foot, and the feel of the resistance of the clutch becomes deeply
familiar. If you now try to depress the clutch with your right (naive) foot, the
feeling is completely dilferent. In a middle-range time period, skates initially
feel heavy when laced on, whereas removing them after a few hours skating
makes one’s feet feel unusually light. Similarly, one gets used to a heavy back-
pack and feels a little like a moonwalker when it comes off at the end of the
day. The various adaptive effects reinforce the point that what we experience
is always mediated by nervous-system structures, with their own peculiar re-
sponse patterns and organization. These species-specific features are shaped by
evolutionary pressures.

How much body representation do newborn humans have? Observations of
newborns have shown a consistent order of the emergence of hand-to-face
movements in the first hours after birth. The median values are as follows:
movements to the mouth, 167 minutes after birth; then the face, 192 minutes;
the head, 380 minutes; the ears, 469 minutes; the nose, 598 minutes; and the
eyes, 1,491 minutes.^® In the first weeks of life, infants use proprioceptive sig-
nals to control posture, and they explore their bodies, especially their mouths,
toes, and fingers. Infants show some hand-eye coordination in reaching, which
steadily improves. The mouth anticipates the arrival of the hand, and the hand
can take any one of many paths to the mouth, from many different starting
points, and does not need visual guidance.

Remember, however, that the fetus does not just sit idly in the womb, but is
busily moving about, from about the tenth week of gestation. As well as kick-
ing and waving movements, it puts its hands to its mouth and makes whole
body movements, such as turning. These movements, along with the sensory
feedback, are part of what is needed to get the motor system and the somatic-
sensing system properly wired up. Many of the movements the fetus has been
practicing in utero provide the basics for bootstrapping to more sophisticated
skills in the postnatal world. (Development will be considered in chapter 8.)

One revealing dimension of infant body representation was discovered by
developmental psychologist Andrew Meltzoff. Even very young infants will
stick out their tongues in response to seeing another human stick out his
tongue.^® Meltzoff found that as early as he could test — forty-two minutes after
birth — newborns will gaze fixedly at his face as he sticks out his tongue, and
then, slowly, haltingly, out would come the infant’s tongue. Infants will also
mimic a gaping mouth and a scrunched up face (figure 3.15). Interestingly, the
infants do not respond well, if at all, unless they see the whole movement. A
static protruding tongue does not evoke the neonate’s imitation.

101

Self and Self-Knowledge

Figure 3.15 Photographs of two- to three-week-old infants imitating facial acts dem-
onstrated by an adult. (From Meltzoff and Moore 1977. Reprinted in The MIT Ency-
clopedia of the Cognitive Sciences, s.v. “Imitation.”)

This behavior implies that the brain, even at this very early stage, is able to
map what it sees of another’s facial movements onto its own sensorimotor
representations. In a loose sense, the infant brain “knows” that what it sees
(“your tongue moving out”) corresponds to what is in its mouth (“my
tongue”), and that by moving a set of mouth and tongue muscles, “I can look
like you.” Here again, the terms “know” and “I” are emphatically in scare
quotes, since the infant does not know these things in the way a three-year-old
child does. The infant’s self-representation is more fragmentary and less con-
nected than that of the three-year-old, but even so, the capacity to mimic sim-
ple facial expressions betokens a rudimentary coherence and the wiring that
supports it.

This capacity may be mediated by a special class of neurons in the prefrontal
cortex referred to as “mirror neurons.” First discovered in the 1990s by Rizzo-
latti and his colleagues in the monkey, they are neurons that respond either
when the monkey himself makes a particular hand movement, such as pick-
ing up a raisin, or when he sees another make exactly that hand movement.^®
Although the function of these mirror neurons in self-representation has not yet
been established, their unique response pattern does suggest that they might

Figure 3.16 Overview of the sympathetic division (left side of the figure) and para-
sympathetic division (right side of the figure) of the visceral motor system. The para-
sympathetic system can constrict the pupil, stimulate salivation and tears, constrict

103

Self and Self-Knowledge

well play a role in imitation, me/not-me distinctions, and social cognition more
generally. (These neurons and their possible role in imitation is discussed
below, pp. 108-1 10).

The visceral system

The other part of the story of body representation concerns the autonomic
nervous system, which regulates what we loosely call our “innards.” The di-
mension of self-representation anchored by neurons regulating visceral func-
tions is probably shared by all animals of varying complexity. Without our
conscious control, the autonomic nervous system tends to our vital functions.
We breathe, our hearts beat, our stomachs digest, our bladder muscles con-
tract. Among other things, we secrete saliva, insulin, and digestive enzymes; we
vasodilate our skeletal muscles when fleeing and vasoconstrict them when
digesting. These are all motor functions, as surely as walking or whistling are
motor functions, but they operate mainly on structures hidden from view. The
autonomic system acts on smooth muscles (e.g., in the blood vessels and intes-
tines), cardiac muscles, and glands (e.g., the adrenal gland, salivary glands,
lacrimal gland).

The autonomic system also has afferent pathways, carrying signals from our
innards to the brain and spinal cord. Among other things, these feedback sig-
nals appear to provide the input for a range of generalized feelings, such as
feeling well or ill, feeling energetic or fatigued, feeling relaxed or on the alert.

The autonomic nervous system has two major divisions, the sympathetic
system, which mobilizes the body to act in challenging conditions, and the
parasympathetic system, which allows the body to recover from strenuous
activity. The two systems tend to counterbalance each other (figure 3.16). For
example, if a predator is attacking, its pupils dilate, and its heart rate increases;
there is vasodilation of the bronchi and coronary artery, sweat secretion.

airways, slow the heart beat, stimulate digestion, dilate blood vessels in the gut, stimu-
late the bladder to contract, and stimulate sexual arousal. The neurotransmitter for this
division is acetylcholine (Ach). The sympathetic system has the opposite profile. It can
dilate the pupil, inhibit salivation and tearing, relax airways, stimulate glucose produc-
tion in the liver, stimulate secretion of epinephrine and norepinephrine from the adrenal
medulla, relax the bladder, and stimulate orgasm. It uses norepinephrine (NE) for some
of its functions (indicated by broken lines), and Ach for others (indicated by solid lines).
Abbreviations: III, oculomotor nerve; VII, facial nerve; IX, glossopharyngeal nerve; X,
vagus nerve. (Based on Heimer 1983.)

104

Metaphysics

secretion of catecholamines from the adrenal gland, and inhibition of smooth-
muscle activity in the digestive tract. If the predator’s attack is complete and
the prey caught, it settles down to its meal: its pupils constrict, secretion of
saliva returns, motor activity in the digestive tract resumes, digestive enzymes
are secreted, and the heart rate is reduced.

Neural networks in the brainstem, medulla, and hypothalamus play the cen-
tral role in such coordination. Signals from the brainstem also travel to the
amygdala via the thalamus into such cortical regions as the somatosensory
cortex, parts of the cingulate cortex, and the orbital frontal cortex. Integration
of inner body signals with signals from the somatic sensory system occurs at
many levels, from the (phylogenetically more ancient) brainstem to the (phy-
logenetically more recent) frontal regions of the cortex (figure 3.17).

Cortical representation, coordinated with cohort thalamic signals, may be a
necessary condition for awareness of these visceral feelings. Some afferent
signals, such as those representing duodenal distension and blood pressure,
seem to be inaccessible to conscious awareness. How much of the activity in the
autonomic pathways informs behavior without reaching the level of awareness
is not known, nor is it really understood what makes some signals, such as
bowel and bladder load, seem introspectively vivid and salient, while others,
such as feeling tired or full, seem more subtle and backgroundish. What

105

Self and Self-Knowledge

explains these differences in the conscious status of autonomic signals is an
unanswered, but researchable, question.

The autonomic nervous system is pretty much left to itself in the everyday
business of life, so long as it runs smoothly. Indeed, at a first pass, this part
of the nervous system might seem to have little to do with how one self-
represents. Why, one might ask, should regulation of such functions as peri-
stalsis, heart rate, glucose levels, and so forth, have anything to do with
i'e^-representation? It seems obvious that autobiographical memory has a
role in what makes me me, but visceral perception seems to be a less likely
player.

Nevertheless, the autonomic system — because of the centrality of its role
in coordinating vital functions, biasing behavior choice, and giving emo-
tional color to ongoing experience — constitutes the core of what makes an
animal a coherent biological entity. It is by no means the whole story of
self-representation, but it is a crucial component, both in the individual and
across species. The autonomic system and the somatic sensory system — along
with their connections to the brainstem, cingulate cortex, hypothalamus, and
amygdala — embody a model of an animal: its drives, its current parameter
settings, and its state of arousal. Although the self-representational capacities
we frequently talk about, such as consciously recollecting one’s earlier life
events or consciously wondering about one’s motives or preferences, seem to
be the obvious center of the self, they are likely evolution’s extensions and
elaborations of the rudimentary self model rooted in the autonomic and so-
matic sensory systems.

The main points of sections 1.3 and 2.1 are these: our basic self-representation
capacities are tied to the centrality of agency and inner regulation in an ani-
mal’s survival. Fancier self-representational capacities likewise have their roots,
long and winding though they may be, in agency and inner regulation. Cogni-
tion and regulation of the inner milieu can thus be thought of as regions on
one and the same capacity continuum. Roughly speaking, inner regulation is
essentially low-level cognition with a narrow plasticity range; high-level cog-
nition is essentially fancy regulation, with a much broader plasticity range.
Although it may not be immediately obvious how being able to add and sub-
tract, for example, has anything, ultimately or immediately, to do with sur-
vival-promoting behavior, a moment’s reflection reverses that intuition. By and
large, fancy cognitive capacities pay their way in the nervous system of a
species by permitting the animals with those capacities to be smarter or faster
or otherwise able to outcompete rivals in the survival game.

106

Metaphysics

Some self-representational capacities involve awareness; some do not. Some
incorporate high-level cognition; some do not. The body seems to be repre-
sented many times over in the brain, at different time scales, in different degrees
of generality, with different levels of computational goals, and with different
blends of motor sequences. A long list of questions remain unanswered,
including how integration of information is handled, how evaluation of goals
and motor options is achieved, how past experience influences forward models,
how learned skills play a role, and so on. What we have on the table is merely an
outline. It is, moreover, an outline highly simplified for purposes of exposition.

2.2 Myself among Other Things

Representation of body parameters, sensory and motor, is just the beginning of
the representational adventures of the complex brain. Guided by its rich post-
natal experience, the brain constructs a systematic representation of the exter-
nal world: for humans, the world of pillows and toys and grandmothers and
cookies; for robins, the world of grubs and hawks and leafy bushes in which to
hide. Think of this as elaboration from the protoself level, where the brain’s
categorizations are more or less of the form “Ouch, hurting here” and “Ooh,
smelling good,” to the level of computational sophistication where the brain
begins to identify objects in terms of their causal properties: “That wasp can
sting me” or “Those cookies taste good” or “I can catch that bird.”^^ Much of
learning consists of constructing a causal map of one’s world.

What sort of causal knowledge do baby humans have?^^ Developmental
psychologists have learned quite a lot about this. For example, if you tie a rib-
bon around a neonate’s foot and tie the other end to a mobile hanging over the
crib, the infant soon learns to make the mobile move by kicking. It has some
understanding, therefore, that its own leg movement can make something
happen. But its understanding of the causal situation is limited. For example, if
the ribbon is detached from its foot, the infant will continue to kick expec-
tantly, apparently not realizing that the ribbon must be connected to the mobile
to get the effect.^"''

With experience and maturation, babies come to have an increasingly com-
petent causal understanding of the world. Suppose that you put a toy on a
cloth so that baby can get the toy only by grabbing the cloth and pulling on it.
One-year-olds will successfully do this when the toy really is on the cloth, and
they will not bother pulling the cloth if the toy merely sits alongside the cloth.
Younger babies will pull on the cloth even if the toy merely sits alongside the

107

Self and Self-Knowledge

cloth, and may get frustrated when the toy fails to come along with the cloth.
At eighteen months, babies can use a toy rake to pull an object towards them,
but one-year-olds do not.

The world of bottles and toys is important, but there is still more to repre-
sent. Especially in gregarious creatures like ravens, wolves, monkeys, and
humans, the brain also comes to understand and represent the complex social
world in which it finds itself This is a world not just of other objects but also of
other selves, that is, other articulated bodies with complex perceptual skills,
motor skills, and their own representational capacities and practical agendas.

In the social world, it is vital to comprehend what others intend and feel and
want. Such cognitive comprehension seems to depend on modeling, in some
manner and to some degree, the internal cognitive states of others, such as what
objects they can see from their position, what they are planning during the
hunt, or what they are feeling about a threat. The advantage of such cognitive
representation is that it permits the animal to anticipate and manipulate the
behavior of these other cognitive creatures, and to navigate the social world
that, collectively, they constitute.

This level of representational capacity is more complicated than the protoself
representation of the body’s internal parameters in an “aah-feels-nice-here”
sort of framework. To a first approximation, the brain is now representing the
representational activities of brains in general, it is now capable of representing,
at least to some degree, its own activities as a representational system. Such
representation need not be very sophisticated or penetrating, from a scientific
point of view, in order to be useful. Thinking “She likes me,” “He is afraid of
me,” and “She intends to hit me” is the stuff of successful social navigation for
raven flocks and grade-school playgrounds alike.

Notice that if my brain represents you as wanting to be groomed, fearing a
snake, seeing me, or the like, this involves my seeing your facial expressions
and bodily behavior as resulting from something that I do not directly observe,
namely your feeling fear, desiring to be groomed, or the like. I can see your face
blanch and your eyes widen, but your fear is your brain state. But I think of
you as having internal states that cause these effects, and I think of your state
as like mine when I am afraid. In this respect, representational models of other
selves and the external world are more akin to using a scientific hypothesis,
such as Newton’s law of gravity, to explain why things fall, than they are to
generalizations embodied in the wiring supporting classical conditioning.

The analogy between scientific theories and a scheme to represent con-
specifics as other minds was largely invented by the American philosopher

108

Metaphysics

Wilfrid Sellars. Sellars realized, of course, that there are important dissim-
ilarities; for example, scientific theories are initially explicitly proposed as
hypotheses; “folk theories” are not. “Theories” of other minds would not have
emerged as a Cro-Magnon clan sat around the evening fire while Krong
explained his new idea that we folks have mental states. Sellars certainly was
not imagining anything as dopey as that. His real point was this: in our repre-
sentational framework for understanding minds, several features — such as the
interdependence of categories, the model’s use in observation, prediction, inter-
vention, and explanations — are importantly analogous to the role theories play
in science. Incidentally, Sellars’s insight applies not only to folk psychology; it
applies also to other domains of commonsense understanding: folk physics,
folk biology, and folk medicine. Nevertheless, it was application of his insight
to our inner states, our mental states, that loosened the death grip of a priori
philosophy of mind on our theories of the mind.

A related way of making the same point ties the notion of understanding the
intentions of others to emulator function. As we saw, the inverse model com-
ponent of emulators generates a motor command (intentions), a copy of which
goes to the forward model, where consequences are predicted and evaluated. If
I see another person begin an action, e.g., reach towards a plum, I understand
his intention by simulating that action in my own brain. To a first approxi-
mation, the forward model would predict the consequences of the observed
motion, and the inverse model would produce a “what if” command that,
while not itself executed, gives the brain the basis for interpreting the observed
movement.^® This is essentially representing others’ intentions via simulation
(what would I be up to if I were doing that?), and simulation, like off-line
planning, is surely a function that can be performed by making minor mod-
ifications to the vanilla emulator. First proposed by Alvin Goldman, the simu-
lation hypothesis improved upon on Sellars’s hypothesis especially because it
was overtly free of any notion that the simulation must be mediated either by
language functions or by explicit reasoning.

The hypothesis also connects to the discovery by the Rizzolatti lab of mirror
neurons in the premotor cortex (see above, p. 101). Rizzolatti and colleagues
have identified grasping-with-the-hand neurons (which are selective for partic-
ular kinds of grips), grasping-with-the-mouth neurons, holding neurons, tearing
neurons (figure 3.18). As noted earlier, mirror neurons respond either when the
animal sees a particular movement made by another animal or when the ani-
mal makes that specific movement. The behavior of these neurons suggests that
in seeing the other animal make the movement, the premotor cortex generates

Figure 3.18 Visual and motor responses of a mirror neuron in area F5. (a) A piece of
food is placed on a tray and presented to a monkey. The experimenter grasps the food,
then moves the tray with the food towards the monkey. Strong activation is present in
F5 during observation of the experimenter’s grasping movements and while the same
action is performed by the monkey. Note that the neural discharge (lower panel) is ab-
sent when the food is presented and again when it is moved towards the monkey, (b) A
similar experimental condition, except that the experimenter grasps the food with pliers.
Note the absence of a neural response when the observed action is performed with a
tool. Rasters and histograms show activity before and after the point at which the
experimenter touched the food (vertical bar). (From Rizzolatti, Fogassi, and Gallese
2001 .)

Metaphysics

incipient motor commands to match the movement. These signals can be
detected as intentions, albeit inhibited or “off-line” intentions, which are then
used to interpret what is seen.^^

In infants the inhibition of the motor decision is less developed than in
adults, and hence once sees infants’ imitating such behavior as sticking out the
tongue, waving, smiling, clapping, and so on. Moreover, even fourteen-month-
old infants show sensitivity to being imitated and recognize whether movements
do or do not match their own. As Meltzoff and Gopnik showed, playing the
imitation game and experimenting within it is how infants learn about others’
intentions, desires, and perspective; that is, how they acquire their folk-
psychological theory.^®

Cortical areas other than superior frontal (FI, F2, F7) and inferior frontal
areas, such as the superior temporal sulcus (STS) in humans, also have a dem-
onstrated role in social cognition. Certain neurons in this large area are
involved in recognizing another’s gaze as making eye contact or as averting eye
contact or as looking at another object. Other neurons in STS are responsive
when the subject sees mouth movements of another. Some are preferentially
responsive to specific hand movements (figure 3.19).^®

The explanatory and predictive role of our psychological understanding is so
commonplace as to pass almost unnoticed. Consider a homely but useful ex-
ample. I can offer a likely causal explanation of why Bill is walking towards the
coffee cart at 8:30 in the morning. The explanation would advert to his desire
for coffee in the morning and his belief that the coffee cart is a good place to get
coffee. If I have seen Bill go to the cart for coffee on several mornings, I can
predict that he will do it today as well, even if I have not yet seen him approach
the cart. I can predict with reasonable assurance that if I pay Bill $100 not to
drink coffee today, he will not drink coffee today. We can make generalizations
such as this: if I insult my students, they will be angry and discouraged. If you
have eaten nothing for 24 hours, you will feel great hunger. People who are
overtired are often grouchy and have poor judgment. And so on, and so on.

Like scientific theories, a folk-psychological theory can be contemplated,
tested, revised, and augmented. Carl Jung (1875-1961), for example, hoped to
augment folk psychology with the notion of the “collective unconscious” to
explain common themes in dreams and stories. In the end, it turned out to be a
weak and unconvincing proposal. Freud thought excessive hand washing
could be explained in terms of sexual repression. This seemed at first to be more
successful than Jung’s proposal, but it too has proved to be explanatorily less
effective than neurobiological explanations. Excessive hand washing is one

Ill

Self and Self-Knowledge

Figure 3.19 A sketch of a monkey brain and some areas hypothesized to be involved in
imitation. Abbreviations: Ps, principal sulcus; ALs, inferior arcuate sulcus; ASs, supe-
rior arcuate sulcus; STs, superior temporal sulcus; Cs, central sulcus; Ls, lateral sulcus;
IPs, intraparietal sulcus; MIP, medial intraparietal area; VIP, ventral intraparietal area;
LIP, lateral intraparietal area; AIP, anterior intraparietal area; SI, primary somato-
sensory cortex; SII, secondary somatosensory cortex. F areas are related to motor func-
tion. Gray areas indicate an opened sulcus. Arrows indicate known neuronal projections
between different areas of the brain; dashed arrows indicate hypothesized connections.
(Based on Schaal 1999.)

typical manifestation of obsessive-compulsive disorder, which appears to have
a neurobiological basis for which there is a genetic predisposition.

Developments in the neurobiology of addiction have put pressure on the
folk-psychological idea that smokers sulTer from “weakness of will.” We now
appreciate more clearly how nicotine causes changes in the brain’s reward
pathways that cause people to crave nicotine. Moreover, scientific theories, as
W. V. O. Quine correctly noted, are continuous with common sense; they are
common sense subjected to critical analysis, to demands for consistency and
coherence, and to the increasingly well-honed standards of experimental test
that are themselves solidly rooted in common sense.

However it happens that the brain comes by its earliest version of mind
modeling, Sellars’s analogy between mind models and scientific models gives us
a new slant on the representational interconnectedness in our understanding of
other minds. The analogy helps us understand the logic and structure of the
framework of concepts that gets us around the social world. It helps us appre-
ciate that even these familiar models of folk psychology are revisable, just as

112

Metaphysics

“folk physics” or “folk biology” were revisable. More strongly, it underscores
the fact that representational models that wear the badge of obviousness may
nevertheless be improved, sometimes in quite radical and surprising ways.

Sellars’s idea is now captured in the “theory theory” and has been widely
adopted in experimental psychology. It is used in describing what infants rep-
resent about minds, and what mature humans understand about their own and
others’ minds. This approach gives us a tool to raise questions about infant and
adult capacities, about animal capacities,'”'^ and about our prospects for
improving our everyday theory-of-mind beliefs as psychology and neuroscience
coevolve. It opens the door to a neurobiological understanding of the neural
basis for familiar phenomena such as addiction, mood swings, eating disorders,
and dreams. Perhaps the most important consequence of Sellars’s idea is that it
liberated philosophers from the entrenched assumption that how we think
about our own and others’ minds is a strictly philosophical, a priori, Platonic,
and nonempirical matter. His proposal made it not only acceptable but neces-
sary for philosophers to look outward to psychology, neuroscience, and biology
in general to try to understand how the brain represents its own activities and
capacities."”^

Earlier we noted that at 14 months babies can tell whether their own move-
ments are imitated by another person. What else do human infants understand
about the minds of others? Between about 9 and 12 months, babies begin to
display a cluster of behaviors that imply an early and developing concept of self
and others-like-myself. If the mother gazes at some object in the corner of the
room, the child will look at the mother’s face, and its gaze will follow that of
the mother. At this stage, the child points not only to what it wants, but also to
something it wants the other person to notice. By 16 months, still before they
acquire their first spoken words, children comprehend what someone is trying
to do and can screen out what is accidental in an action. For example, suppose
that the mother demonstrates how to take a new toy apart and during the
demonstration she accidentally drops the toy and then resumes taking it apart.
When the child imitates her actions, he omits the accidental dropping. In test
after test, the child appears to distinguish between what the adult was trying to
do and what was unintended or accidental. Other tests show that the child has
an understanding of what the mother knows or sees or expects.

One may thus conclude that by 16 months, the child already has some
understanding of what is in another person’s mind, so to speak. This emergence
of cognitive capacities is called the development of perspective-based repre-
sentations. They permit the child a degree of understanding of how things look

113

Self and Self-Knowledge

or feel from another’s point of view. To a first approximation, these repre-
sentations cohere as a framework for predicting and manipulating the behavior
of other cognitively complex creatures. The mind-model still has much devel-
opment to undergo, however. A two-year-old infant can be surprised that he
can still be seen when he covers his eyes, but by three, he has a clearer sense of
how to hide, given another’s point of view.

Elizabeth Bates has shown that a rudimentary contrast between self and
others is marked explicitly in language between 18 and 20 months. As Bates
notes, however, this does not mean the child has worked through all the con-
sequences of this contrast. Children may make linguistic errors, saying, for
example, “Carry you,” when they mean “You carry me,” though it is clear
behaviorally what is intended. Because of the number of distinct personal pro-
nouns and the complexity of the conventions governing them, it does take the
child some time to sort it all out. This stage is preceded by using objects to
communicate (e.g., showing Daddy the truck) at about 9 months. This merges
into the period where objects are given (10 to 12 months). At around twelve
months, this is followed by communicative pointing, where the child extends
arm and index finger to objects he wants the adult to notice. Such behavior is
clearly communicative, since the child repeats or emphasizes the pointing until
the adult acknowledges the object referred to.

Three-year-olds explain and predict what other humans do mainly by refer-
ring to desires and perceptions, but they are not yet in command of the notion
of beliefs. They can use counterfactuals about desires and easily answer such
questions as “If Billy wanted a cookie and I gave him a crayon, would he be
happy?” Not until they are about four do children begin to use the notions
of beliefs and thoughts to explain and predict what others do or would do if
they saw something. In a classic experiment, pencils are put in a candy box,
and this is shown to the child. When asked, “What will Billy think is in the
box?” three-year-olds say “pencils,” and four-year-olds say “candy.” The idea
that someone else will have a false belief based on misleading evidence is very
sophisticated and represents a new stage in the developments of the child’s
psychological theory. Notice that it requires the child to use generalizations to
arrive at a likely belief, given what one would normally see and expect, and
even when this belief is different from what she herself believes to be true. In
the relevant respects, this is like using a scientific hypothesis to predict what
would happen if various conditions were satisfied.

To the extent that the organism uses the perspectival model to plan and pre-
dict the organism’s own behavior and to think about its own feelings, the model

114

Metaphysics

permits the organism to represent its self. Does the child first apply the frame-
work to itself and then say, in effect, “Wow, mummy and Joey are just like
me, so I guess they can see and feel and want things too!” Probably not. The
aforementioned data, along with other data on self-referencing, imitation, and
so forth, suggest that the child’s development of perspectival representations
proceeds in tandem with, and positively adds to, his growing understanding of
his self."^^

Whatever the precise nature of this capacity for perspectival representations,
notice that having already learned a language is not essential. Of course, the
acquisition of language can enhance and significantly change and augment
the capacity, but the essential rudiments of the capacity seem to be language-
independent. As the infant-development research convincingly shows, a rather
rich perspectival representation system probably has to be in place for language
to be acquired at all.

Do animals have a theory of mind? Does a chimpanzee, for example, know
what can be seen or not seen from another chimpanzee’s point of view? The
answer seems to be “yes”. For example, in carefully controlled experiments,
Josep Call found that chimpanzees followed head direction of humans and
conspecifics to find a target above and behind them. In conditions of social
competition between subordinate and dominant chimps, he found that when
food was placed so that from the perspective of the dominant chimp, one
morsel was occluded and the other visible, the subordinate more frequently
retrieved the morsel unseeable by the dominant (see figure 3.20). In a second set
of experiments, the subordinate chimpanzee more often retrieved the morsel
from the workspace if the workspace was baited when the dominant chimp
could not see (figure 3.21).

Ethologist Frans de Waal discusses data showing that chimps may use
pointing gestures when they want something, though they do not use the hu-
man style of pointing: outstretched arm and index finger. Here is one of his
own observations of a chimpanzee using a gesture similar to what a human
might use to indicate that a sinister character has just joined the party:

A chimpanzee named Nikkie once communicated with me through the same subtle tech-
nique. Nikkie had gotten used to my throwing wild berries to him across the moat at the
zoo where I worked. One day, while I was recording data about the apes, I totally forgot
about the berries, which hung on a row of tall bushes behind me. Nikkie hadn’t forgotten.
He sat down right in front of me, locked his red-brown eyes into mine, and — once he had
my attention — abruptly jerked his head and eyes away from mine to fixate with equal in-
tensity on a point over my left shoulder. He then looked back at me and repeated the move.
I may be dense compared with a chimpanzee, but the second time 1 turned around to see

115

Self and Self-Knowledge

(b) Transparent barrier test

Figure 3.20 Mean percentages of pieces of food retrieved by subordinate chimpanzees
as a function of whether food pieces could be seen by the dominant chimpanzee. In the
occluder test (a), one of the pieces of food was hidden from the dominant chimpanzee,
and this increased the likelihood of its being retrieved in preference to the visible piece.
When both pieces of food could be seen by the dominant animal (b), there was no dif-
ference in retrieval percentage. (From Call 2001.)

what he was looking at, and spotted the berries. Nikkie had indicated what he wanted
without a single sound or hand gesture.'^^

Others report observations in well-controlled conditions of chimpanzees and
monkeys using body language to convey the location of prized or dangerous
objects. These and a huge range of other experiments imply that the animals
are not merely responding to specific cues but are also making use of repre-
sentations of what others can see, want, intend, and feel. A chimpanzee’s
theory of mind undoubtedly does not map smoothly onto that used by humans
or that used by baboons or dogs, especially because different animals will be
different in how they make their living and what they care about. But it is
entirely plausible that they have some measure of a theory of mind that enables
them to manipulate the behavior of others by using representations of others’
inner states.'^^

116

Metaphysics

Step

Figure 3.21 Mean percentage of pieces of food retrieved by subordinate chimpanzees
as a function of whether the dominant chimpanzee who witnessed the baiting process
was the same one who competed over the food with the subordinate. When the domi-
nant animal was switched (steps 2 and 3), the subordinate was more likely to retrieve the
food. This suggests that the subordinate is able to represent what its competitors can or
cannot see. (From Call 2001.)

What would a human brain be like if it lacked a theory of mind? Autism,
according to one prominent hypothesis, is the result. Autism is a developmental
disorder whose core characteristics include impairments in socialization, com-
munication, and imagination. Primary symptoms include failure to make eye
contact, follow pointing gestures, follow gaze, and play the imagination game
where you tell what someone would be feeling if x happened to him. Uta Frith
notes, “Most individuals with autism fail to appreciate the role of mental states
in the explanation and prediction of everyday behavior, including deception,
joint attention, and the emotional states which depend on monitoring other
people’s attitudes, for example pride.”"''® On this hypothesis, autism is a kind of
“mind blindness.”

Finding consistent brain abnormalities in the brains of autistics has been a
challenge, and so far the data have been perplexing. What has emerged in the
last few years is that there are neuronal abnormalities in limbic structures

117

Self and Self-Knowledge

(amygdala, hypothalamus, hippocampus) and in the cerebellum."''® The chief
finding is reductions in the numbers and sizes of specific types of neurons;
pyramidal cells in the limbic structures and Purkinje cells in the cerebellum. It
has been intriguing to many that cerebellar abnormalities are found. This is
because the cerebellum has traditionally been thought to serve mainly in sen-
sorimotor coordination and not to play a role in cognitive functions, tradition-
ally conceived. The cerebellar evidence makes sense, however, if simulation of
the minds of others is a spin-off function of Crush emulators and Crush emu-
lators are cerebellum-intensive.®®

Mind models, applied to oneself and others, can get increasingly complex,
at least in humans. One can engage in self-reflection, for example, to examine
one’s motives, excuses, and desires. This involves a representation of a self-
representation, and hence is a recursive capacity. Memories of specific events in
one’s past experience are also instances of such self-reflection. For example,
you can remember the perception of seeing pelicans diving for fish, or you can
remember being afraid when you heard the thunder. You can also remember
today that yesterday you remembered that you were hungry on the canoe trip
last summer (a representation of a representation of a representation). Flumans,
at least, can create iterated representational structures of this kind, though the
extent to which nonhuman animals can enjoy recursive representation is not
determined. The recursion may be rate-limited, however, for it may not be
useful to stack up representations of representations beyond four or five iter-
ations. “I believe that I remembered that I thought that I experienced a pain in
my foot” may be useful only rarely.

Although the capacity for self-reflection is important, it is not, on the Grush-
emulator hypothesis, the fundamental platform of the sense of self. That plat-
form, I have suggested, is first and foremost a matter of body regulation and
body representations. Nevertheless, the capacity for self-analysis, self-reflection,
and self-awareness — the capacity to know that I know — has seemed to betoken
something suprabiological and even supraphysical about the mind. In the next
and closing section, we return to the question of dualism and knowing what is
in our minds.

2.3 Knowing Oneself: A Philosophical Problem

Descartes believed that the (conscious) mind, and only the mind, is directly
known. The conscious mind, he was convinced, is known more certainly than
anything else is — or could be — known. He used the alleged epistemic specialness

118

Metaphysics

of the mind (directness) to defend the metaphysical specialness of the mind
(the thing known). His argument, in short, says that if the mind can know itself
directly and with certainty, it must be a dijferent sort of thing from things in the
physical world. Physical objects, he argued, we know only indirectly and with
degrees of uncertainty.

What is meant by the directness of self-knowledge? Boiled down, it means
that one typically makes judgments about what one is feeling or hearing or
seeing without first going through any explicit reasoning from evidence to con-
clusions. For example, you usually know without explicit inference that you
feel cold or see a light or smell smoke. If you feel a pain, you just know that;
you do not have to reason it out from more obvious information. These are
features of the world about which you are unable to say how it is that you dis-
criminate one such feature from another: you simply can.

Noninferential knowledge of one’s own mental states, the argument goes, is
evidence for the special nature of the mind, metaphysically speaking. This
argument has remained an intuitively powerful resource for many who wish to
resist the explanatory advances of neuroscience into the domain of the con-
scious mind. This special directness in discriminating simple features, such as
colors and pains, it may be insisted, will never go away. Hence there can never
be a science that can be as firmly known and as deeply believed as one’s
knowledge of one’s own conscious mind.

The logic of the situation, however, is this: nothing follows about the meta-
physical uniqueness of the mind from the existence of discriminable simples,
i.e., judgments made without consciousness of the computational antecedents.
First, absolutely all knowledge involves some neural processing prior to con-
scious recognition that something is an a or a f. This is so whether the cogni-
tion pertains to the mind or to the body, whether one is aware of a stimulus as
hot or as lasting for seconds or as looming towards you. There is no such thing
as unprocessed perception.

Second, when one becomes aware of the result of nonconscious processing,
one has no introspective (conscious) access to the processing steps that went
into producing the result. It is therefore entirely inevitable that there will be
some discriminations — the results of nonconscious processing — that are spon-
taneously, noninferentially, and, one might say, directly experienced.^^ That
one cannot articulate how the discrimination was made is simply explained by
the fact that there is a vast amount of nonconscious neural activity to which
one does not — and perhaps cannot — have conscious access. For example, even
feedback techniques will probably not allow me to be aware of what the ama-

119

Self and Self-Knowledge

crine cells of my retinae are now doing, anymore than I can become aware
of when hormones are released from the pituitary or what my blood pressure is.
I do not have introspective access to the processing that yields a stereoptic,
three-dimensional representation of the visual scene. I just simply see things
in stereoptic depth. I cannot tell you how I identify a melody as “Three Blind
Mice”; I just do. But so what, metaphysically speaking?

Once the “nonconscious processing” point is on the table, the case for a
metaphysically special stuff to handle direct knowledge is enfeebled. Additional
arguments weaken it further. In particular, noninferential judgment is not
confined to knowledge of one’s mind. Normally, one also knows many things
about one’s body without relying on any explicit inferences. For example, I
directly know that I am sitting or standing or that my arms are folded on my
chest. I directly know (using no explicit inference) whether my head is rotating
or tilted forward or back. I directly know whether my tongue is moving and
whether my feet are cold. Normally, you do not have to make any explicit,
overt inferences to know that you are sneezing, vomiting, choking, suffocating,
or passing water.

The hedge word “normally” fronts all of these claims, because under unusual
or pathological conditions, a person may have to fall back on reasoning to fig-
ure out what is going on with his body. Someone whose arm was amputated
may continue to feel an arm, which, he must remind himself, is not really there.
Suffering a migraine attack, someone may feel her body to be the size of a tiny
doll, but she turns on the light to check and be reassured. Affected by the
anesthesia ketamine, a patient may feel he is floating above his body. In zero
gravity, one feels as though one is constantly falling. These are all very unusual
conditions. They are, however, instances of errors about body states that one
usually knows about noninferentially; they are instances where a subject can
“reason himself” to the more correct judgment.

Unconvinced, the dualist may try another tack. Even if I do have non-
inferential (and hence direct) knowledge of my body, the dualist may argue, I
can be wrong about the state of my body, whereas I cannot be wrong about
the state of my conscious mind. I have noninferential and infallible knowledge
of “discriminable simples.” Such infallibility, the argument continues, entails
something metaphysically special about the mind.

Note that for the argument make any headway, the infallibility claim has to
be exceedingly strong. “Infallible” here has to mean not just that one is usually
right, or even that in fact one is always right. It has to mean that one cannot —
in principle — ever be wrong. As we shall see, this messes up the dualist.

120

Metaphysics

First, let’s look closely at the infallibility claim. First, if indeed I correctly
describe my mental states, it has to be proved that this is not just a contingent
fact but an a priori metaphysical truth. The favored cases for the infallibility
argument are discriminable simples, for example, feeling a hot sensation. They
are, after all, discriminable simples, so they have fewer degrees of fallibility
than, for example, recognizing something as a B-17 Bomber or a chanterelle
mushroom. Reliable identihcation of these simples is what normal nervous
systems are wired to achieve, not for metaphysical reasons but for survival.

Second, if I never consider myself to be wrong, this is partly owed to our
convention of normally giving the speaker the beneht of the doubt when he
describes his inner states, as he is usually in a privileged position. This is not a
metaphysical privilege, but the privilege of being the one whose brain has pro-
duced the sensation. It is an epistemological privilege: because my sensations
happen in my brain, I am likely to know about them before, and better than,
you.

The third and perhaps most important point is that just as there are con-
ditions, usually somewhat abnormal conditions of course, where I err in my
noninferential judgment about the body, so there are abnormal conditions
where I err in my noninferential judgments about my conscious states. Let us
canvas a few of these cases. As novelists such as George Eliot have rightly
observed, especially in censorious societies, a female may misread her feelings
of sexual attraction in one way or another; as repulsion, shyness, anger, fear, or
nervousness. A particularly inhibited person may need to learn to recognize her
own sexual feelings via inference from her flustered behavior in the presence of
a particular man. More generally, all those feelings we allegedly inhibit when in
denial are instances of fallibility.

The infallibilist, however, may want to dismiss these sorts of cases on
grounds that they are not the sorts of cases, like feeling something hot, that he
has in mind. In particular, they are not discriminable simples. Why not? Sup-
pose that he replies, “Because those are mental states you can be wrong about,
just as you say. I am talking about the ones you cannot be wrong about.” This
response makes his argument circular, since he is rejecting any counterexample
on the grounds that his infallibilist conclusion must be true. So this response
logically imperils the position. We can be generous, however. For the sake of
argument, we can allow that the Freudian counterexamples do not falsify the
infallibilist claim. Let us grant that the cases involving misidentification of
moods and emotions are off limits. There are other examples where squirming
off the hook is even more diflicult.

121

Self and Self-Knowledge

Sensations can, on odd occasions, fail to be correctly apprehended. Expect-
ing a very hot stimulus, one may at first believe that one feels a burning sen-
sation, only to quickly realize that one is actually experiencing an icy cold
sensation. Expecting a pain, I have been surprised to realize that the sensation
is actually not pain at all but only pressure. The infallibilist will, of course, in-
sist that in these instances heat really was felt, pain really was felt. But we can
be sure of this only if we assume, with circularity, that the infallibilist is right,
that we can never be wrong about what are feelings are. Even if it is only an
open question whether we are right or wrong, the claim for infallibility as a
metaphysical truth has lost ground.

Sometimes when the signal is faint or the subject is anxious, he can be unsure
whether he feels something or not. Can I be wrong about whether I hear a
sound at all? Yes, for example, as I wake up, or when I have been paying close
attention to a book, or when I am in a state of great anxiety. There are other
obvious cases where we can be wrong or unsure about certain of our mental
states. Young children are sometimes unsure, even when queried, about
whether they feel the need to empty the bladder. When very tired, children, and
adults too, may not recognize the feeling of being tired.

One could, of course, adopt a convention whereby if a subject says he feels
heat, as opposed to cold, then he really does. Adopting such a convention is
fine, but it fails to yield what the infallibilist wants, namely a metaphysical
truth about the special nature of the mind. Another line of defense is to say that
the cases supporting the infallibilist claim are those where conditions are nor-
mal, the stimulus is well above threshold, the sensations are simple, the subject
is fully awake and attentive, and he is not under the influence of drugs. Fine,
but this defense also looks circular, for it looks like identifying the cases as a
function of whether or not the subject is in fact wrong. That the subject is not
wrong in these cases is a function of how we have picked them out, not of some
metaphysical truth about the ethereal etiology of these cases. Moreover, if the
infallibilist can use this strategy, so can we. I can identify cases of physical
knowledge (e.g., knowing whether I am standing up) where I am not wrong:
conditions are normal, the stimulus is well above threshold, the subject is fully
attentive and awake, and he is not under the influence of drugs. If such cases
can show that physical knowledge can be infallible, then mental knowledge is
not special in this regard.

To move on to other counterexamples to infallibilism, it is interesting that
one is routinely and regularly wrong about what one thinks one tastes. As
neuroscience and psychology have demonstrated, most of what we regard as a

122

Metaphysics

sensation of taste is actually owed to our sense of smell, however convincingly
it otherwise appears. The “taste” of barbecued pork ribs is actually mostly the
smell of the ribs. Taste space is limited to five dimensions: sweet, salty, bitter,
sour, and umami (stimulated by monosodium glutamate). Smell space, by
contrast, runs into may hundreds of dimensions. The “taste” of a Chardonnay
wine is largely the complicated smell of that wine. It probably does not matter
for survival that smell and taste are not kept strictly separate in awareness, the
way that sight and smell, for example, normally are. Hence, the brain is not
equipped with mechanisms for the effortless and noninferential detection of the
separate components of taste and smell.

Pathological conditions give quite another dimension of error in the self-
reporting of mental states. A patient with a sudden lesion to his primary visual
cortex may fail to realize that he is blind, even when this is pointed out to him,
and even when he repeatedly stumbles into the furniture. Described as blind-
ness unawareness, Anton’s syndrome is a rare, but well-documented deficit. In
patients with Anton’s syndrome, the blindness may be transient, though after
recovery of some vision, patients are likely to say that nothing has changed in
the visual capacity. There are reports of unawareness of blindness that persist
indefinitely.

Are the patients with Anton’s syndrome really just mistaking visual imagery
for actual vision? Drawing on anatomical data and behavioral tests, most clin-
ical neurologists believe not. For one thing, the cortical regions needed for
vision are also the regions believed to be needed for visual imagery, and these
are the very ones destroyed by stroke. Not implausibly, Paul Churchland has
argued that these patients have lost the very mechanisms for knowing whether
one is seeing or not. Since the brain has no information to indicate otherwise, it
goes with the standard state of affairs. Thus these patients say, “Of course I
can see” and they will smoothly confabulate a reasonable story when asked
what they see. If asked whether the doctor is wearing glasses, the Anton’s
patient will answer with confidence, but the answer is mere guesswork. It is also
significant that confabulatory responses in Anton’s patients are restricted to the
topic of visual experience. They will be entirely frank and forthright in response
to questions on nonvisual topics. By contrast, patients with Korsakoff’s syn-
drome (alcoholic dementia) freely confabulate about any subject.

The mystery of Anton’s syndrome is worth dwelling on because visual expe-
rience seems so self-evident. If anything seems dead obvious, it is that one can
or cannot see, and it is hard to imagine being wrong about which is which.
Nevertheless, the patients with Anton’s syndrome present us with a compelling

123

Self and Self-Knowledge

case where the brain is simply in error about whether or not it has visual expe-
riences. To insist that such subjects must be having visual experiences if they
think they are, because one cannot be wrong about such things, is of course, to
argue in a circle. The question precisely at issue is whether one can ever be
wrong about such matters. Prima facie, at least, Anton’s patients present evi-
dence that one can be wrong and that there is a neurobiological reason why
they are wrong. More than a mere a priori conviction of infallibility is needed
to reverse the hypothesis or reinterpret the data.

From the point of view of cognitive neuroscience, whether or not someone’s
recognitional skills deploy explicit reasoning appears less important than cer-
tain other properties, such as the neural pathways involved, the contribution of
affective components, the nature of cross-modal and top-down effects, how
much learning has gone on, and how the brain automates cognitive skills. The
predilection, most evident in British Empiricism and German Idealism, for
taking the differences between inferential and noninferential judgments to be
a momentous metaphysical division looks about as misguided as believing, as
pre-Galilean physicists did, that the difference between the sublunary realm and
the supralunary realm marks a momentous metaphysical division concerning
the structure of the cosmos.

Of course, there is a difference between superlunary space and sublunary
space, and the difference means something to humans, because of the proximity
of the Moon to Earth. But it does not mark a metaphysical difference, or even
a difference in what principles of physics apply. Similarly, there is a differ-
ence between inferential and noninferential judgments, but we should hesi-
tate to attach profound metaphysical significance to these two types of neural
processing. (See also pp. 130-133.)

Dualism is implausible at this stage of our scientific understanding. In the
business of developing an ongoing research program, dualism has fallen hope-
lessly behind cognitive neuroscience. Unlike cognitive neuroscience, dualist
theories have not even begun to forge explanations of many features of our
experiences, such as why we mistake the smell of something for its taste, why
amputees may feel a phantom limb, why split-brain subjects show disconnec-
tion effects, why focal brain damage is associated with highly specific cognitive
and affective deficits. In truth, dualism does not really even try.

To be a player, dualism has to be able to explain something. It needs to de-
velop an explanatory framework that experimentally addresses the range of
phenomena that cognitive neuroscience can experimentally address. While it is
conceivable that this can be done, the bookies will give long odds against its

124

Metaphysics

success. Until at least some distinctly dualist hypotheses are on the table, dual-
ism looks like a flimsy hunch still in search of an active research program.

3 Conclusions

The brain makes us think that we have a self. Does that mean that the self I
think I am is not real? No, it is as real as any activity of the brain. It does
mean, however, that one’s self is not an ethereal bit of “soul stufl".” But it is as
real, for example, as the coherent neuronal activity that yields your capacity to
walk or think about global warming or And your way back from a hike in the
bush. Brain activity is an entirely real thing.

But, one might say, that is not how I am used to thinking of myself. Why
would my brain lie to me? Think of it this way. Fundamentally, your brain’s
task is to allow you to make your way in the world, and that means it needs to
be able to make reasonably good predictions, and to make them in a timely
manner. One’s scheme of representational devices need not be the best possible
in order to have practical and predictive value. It just has to be good enough so
that you can make a living, in the broadest sense of the term. In particular, for
most of the business of surviving on the planet, the details of how the brain
actually works need not be explicitly known by the brain. Brains manage rea-
sonably well by using such categories as “wants,” “fears,” “sees,” “is angry,”
etc., as the representational apparatus for understanding its own activities. For
much of the business of everyday life, human brains can manage without such
categories as “neuron,” “DNA,” “electrical current,” and so on.

Nevertheless, humans, for neurobiological reasons we do not yet understand,
have the stunning capacity to play the “ratchet game.”^^ That is, children can
learn the best their culture has to offer and can improve upon it. And their
children can start where they left off. Unlike chimpanzees, where each chimp
starts at essentially the same place where all of his ancestors started, human
children can start well ahead of where their parents started, and vastly far
ahead of where our stone-age ancestors started. They can build on what their
culture already knows. Hence in the general business of trying to understand
the reality behind appearances, humans can develop science and technology,
and can pass it on to their offspring. This gives us the unique opportunity to use
technology and science to develop increasingly abstract, scientifically penetrat-
ing categories, such as “atom,” “valence,” “DNA,” and “neuro transmitter.”

125

Self and Self-Knowledge

We have discovered that brains permit us to see, plan, walk, and wonder.
And now the ratchet game opens up the possibility of going beyond the famil-
iar categories, though they work reasonably well in the everyday business of
explaining and predicting human behavior. It allows us to ask, for example,
how a brain is organized so that by means of two-dimensional light arrays from
two retinae, we see a single image in three-dimensional depth. We can ask how
a brain organizes its information so that it has self-representational capacities.
Here, as elsewhere, scientific discoveries give us surprising new ways of looking
at familiar phenomena. For brains, as well as for the stars, fire, and the heart,
there is a reality behind the appearances, and the exciting thing is to figure out
how to think about that reality in a way that improves upon the old ways.

In this century, modern neuroscience and psychology allow us to go beyond
myth and introspection to approach the “self” as a natural phenomenon whose
causes and effects can be addressed by science. Helped by new experimental
techniques and new explanatory tools, we can pry loose a real understanding
of how the brain comes to know its own body, how it builds coherent models
of its world, and how changes in brain tissue can entail changes in self-
representational capacities. Neurobiology is beginning to reveal why some
brains are more susceptible than others to alcohol or heroin addiction, and why
some brains slide into incoherent world models. We can see progress on our
understanding of the staged emergence of self in childhood, as well as of the
cruel inch-by-inch loss of self in dementia.

Though well short of full answers, neuroscience has discovered much about
the effects of localized brain lesions on higher functions, such as complex
decision-making, speech, and voluntary behavior. Perhaps some questions will
forever exceed the neurobiological reach, though it is impossible at this stage to
tell whether such problems are just as yet unsolved or whether they are truly
unsolvabfe. In any case, incomplete but powerful answers anchored in data can
often provide a foot-hold for the next step. And that, in turn, for the next step
thereafter. But this is just how science proceeds — one step at a time.

Suggested Readings

Damasio, A. R. 1999. The Feeling of What Happens. New York: Grossett/Putnam.

Dennett, D. C. 1992. The self as a center of narrative gravity. In F. Kessel, P. Cole,
and D. Johnson, eds., Self and Consciousness: Multiple Perspectives, pp. 103-115. Hills-
dale, N.J.: Lawrence Erlbaum & Associates.

126

Metaphysics

Flanagan, O. 1996. Self Expressions: Mind, Morals, and the Meaning of Life. New
York: Oxford University Press.

Gopnik, A., A. N. Meltzoff, and P. K. Kuhl. 1999. The Scientist in the Crib. New York:
Morrow.

Hobson, J. A. 2001. The Dream Drugstore: Chemically Altered States of Consciousness.
Cambridge: MIT Press.

Jeannerod, M. 1997. The Cognitive Neuroscience of Action. Oxford: Blackwells.

Kosslyn, S. M., G. Ganis, and W. L. Thompson. 2001. Neural foundations of imagery.
Nature Reviews: Neuroscience 2: 635-642.

Le Doux, J. 1996. The Emotional Brain. New York: Simon and Schuster.

Panksepp, J. 1998. Affective Neuroscience. New York: Oxford University Press.

Rizzolatti, G., L. Fogassi, and V. Gallese. 2001. Neurophysiological mechanisms
underlying the understanding and imitation of action. Nature Reviews: Neuroscience 2:
661-670.

Schacter, D. L. 1996. Searching for Memory: The Brain, the Mind, and the Past. New
York: Basic Books.

Schore, A. N. 1994. Affect Regulation and the Origin of the Self. Hillsdale, N.J.:
Lawrence Erlbaum & Associates.

Tomasello, M. 2000. The Cultures and Origins of Human Cognition. Cambridge:
Harvard University Press.

Websites

BioMedNet Magazine: http://news.bmn.com/magazine
Comparative Mammalian Brain Collections: http://brainmuseum.org
Encyclopedia of Life Sciences: http://www.els.net
Living Links: http://www.emory.edu/living_links

The MIT Encyclopedia of the Cognitive Sciences: http://cognet.mit.edu/MITECS
Neurosciences on the Internet: http://neuroguide.com

Consciousness

1 The Problem and Empirical Directions
1.1 Introduction

When you wake up, you become aware of sights and sounds, of feelings in your
body, perhaps of limb movements. You may become aware of thoughts about
the movie you saw the previous night, of emotional residue from an earlier
dream, of the smell of breakfast cooking. From a hubbub of many voices, you
may follow only that of your child. You will be unaware other many other
events, such as changes in blood pressure and the decisions guiding eye move-
ment as you watch the birds flying outside. Dreams occurring just before
waking may be remembered, if only in fragments, while dreams occurring early
in the sleep cycle will escape conscious recall. If you are attending to what the
birds are doing, other events may go unattended or unnoticed, such as the
music playing softly in the next room, the movement of your tongue in your
mouth, the pain in your knee. Even unattended or subliminal events, however,
can have an effect on your behavior, present and future. ^

From the inside, so to speak, it is the conscious plans, decisions, memories,
and so forth, that seem to make me me. This presumption is almost unavoid-
able, since that is the only me I am aware of. In fact, however, the conscious
events are only a miniscule part of the story of my inner life. So what is going
on here? What happens when I shift attention and become aware of an irritat-
ing mosquito on my leg, and what happens when I am so preoccupied with
putting up the tent that I fail to notice the mosquitoes biting my arms and
legs?

128

Metaphysics

Why is it that no amount of trying to attend to peristaltic movements in the
small intestine results in awareness of those movements, though attention to my
heartbeat yields awareness? How is it that I can be aware of understanding
what you are saying, but I have no awareness of the processes underlying that
understanding? What happens when I learn a task such as riding a bicycle well
enough that I no longer need to pay much attention to my balance? What
happens when I am in deep sleep, unaware of the somatic, auditory, and other
signals that continue to percolate through the nervous system? In sum, what
constitutes the dilference between conscious states and unconscious states?

There are basically two attitudes that one can have toward these questions.
One attitude is, roughly speaking, pragmatic. It emphasizes the search for
revealing experiments, perhaps by understanding what happens to the brain in
a coma or during anesthesia, or how awareness changes after specific kinds of
brain damage. In other words, the pragmatist adopts the position that we try to
make scientific progress on all the aforementioned questions, while subjecting
all hypotheses to criticism and comparing the merits of competing theories.

The opposing attitude, which Flanagan refers to as “mysterian,” takes the
view that these questions cannot be answered scientifically and, indeed, cannot
be answered at all? The mysterians emphasize lack of progress rather than
actual progress, the mysteriousness of the various phenomena at issue rather
than tools for reducing the mystery, and the hopelessness of opening up exper-
imental avenues rather than the opportunities presented by new advances.
Whereas pragmatists tend to emphasize that consciousness is a natural phe-
nomenon of the brain, mysterians favor the idea that it is a supernatural phe-
nomenon, or at least is beyond the physical, in some sense or other.

Pragmatism seems the better counsel to me. At least, I favor an attitude that
says, “Let’s try for a neuroscientific explanation,” over one that is not prepared
to try. This is not dogma or an article of faith. It simply draws on past scientific
successes to predict that progress can often be made even when a problem
looks unsolvable. Nevertheless, it is essential to analyze those arguments that
profess to demonstrate the impossibility of a scientific explanation of con-
sciousness, the better to assess whether those arguments have greater force than
the arguments favoring the try-and-see approach. One cannot tell a priori
where the probabilities lie, and thus a balanced examination is necessary. The
main aim of the next section is to lay the groundwork for approaching the
problem and then to delineate several promising empirical approaches. At
the end of this chapter, in section 2.2, I shall set out and analyze the main
objections to empirical approaches to the assorted problems of consciousness.

129

Consciousness

1.2 Definitions and Science^

In its everyday use, the term “consciousness” can describe a range of somewhat
different things: not being in a coma, not being in a deep sleep, not being under
anesthesia, being aware of feelings and thoughts, and so forth. If we are to
explore a phenomenon, surely we had better know what the phenomenon is, or
we shall end up in confusion and pointless debate. To address the vagueness of
the term “consciousness” it will be tempting to propose that we defer investi-
gation until we have first determined precisely the proper definition of the term.
So perhaps we should stop here and precisely define our terms before we plunge
ahead, building and testing theories.

Although well intentioned, this recommendation is decidedly ill conceived,
especially in the early stages of inquiry. Let me explain.

Prescientifically, we classify things on the basis of their gross physical and
behavioral similarity, or on the basis of their relevance to our particular needs
and interests. Plants may be classified as edible or poisonous, and some may be
weeded by farmers in one locale but cultivated as cash crops in another. Like-
wise, animals that are docile or dangerous are more likely to be grouped with
others that are docile or dangerous, respectively. Highly unusual or distinctive
features may also call for distinct classification. So all birds with interesting
vocalizations are referred to as songbirds, even though they may come from
species as diverse as mockingbirds and nightingales. Rare, polishable sub-
stances get called gems, even though diamonds, rubies, amber, and opals differ
radically from each other in their chemical composition and nature.

Developed sciences tend to bootstrap themselves up from such early classifi-
cation schemes. As we come to understand the reality behind the appearance,
classifications are drawn according to different sorts of principles. Properties
such as “edible” or “makes a nasty smell” are not necessarily abandoned, but
they no longer serve as the basis for the taxonomies that we use in articulating
the deeper explanatory principles. The taxonomies current at a given stage of
scientific development do not, of course, seem at all superficial to us. They seem
to us to be a true and faithful reflection of how things really are.

Terms may change their range of application as new discoveries are made.
These changes in turn have an effect on perceptual recognition. The history of
the term “fire” provides a striking example of how the boundaries of a familiar
category get redrawn. Not so very long ago, the category “fire” included any-
thing that emitted light or heat, and the presence or absence of this property
could be determined just by looking or feeling. “Fire” was used to classify not

130

Metaphysics

only burning carbon stuffs, such as wood, but also activity on the sun and
various stars (which we now know is not fire at all, but nuclear fusion), light-
ning (actually, incandescence following an electrical discharge), the northern
lights (actually, spectral emission), and the flash of so-called fireflies (actually,
biophosphorescence). Moreover, these phenomena were thought to share some-
thing deep in common — their essence or essential nature — which allegedly
made them all instances of Are.

As modern science slowly came to realize, burning wood involves oxidation,
and this process has nothing in common with the processes underlying the
other assorted phenomena. That fact, however, is not something you can know
just by looking. The development of our understanding of Are as oxidation also
led us to see a deeper connection between Are on the one hand and iron rusting
and biological metabolism on the other. These processes were not originally
considered to share anything with burning. Because detectable heat was taken
to be an essential feature of the class that included burning wood, the sun, and
lightning, it would never even occur to someone to suppose that rusting might
be an instance of Are. Since you cannot feel any heat from rusting iron, it took
an understanding of the hidden reality of oxidation to reveal that rusting and
burning of wood are in fact the same process. As for metabolism, our bodily
heat was thought to be just how we are. The suggestion that heat in animal
bodies involves the same process as burning of wood struck some as obviously
ridiculous (figure 4.1).

Why does science tend to reject our everyday folk criteria in favor of others
that are arcane and of little apparent relevance to everyday life? One answer
is that the scientific categories more accurately reflect the structure of reality
itself. We consider the categories more accurate because they enable more
powerful explanations, predictions, and manipulations of the world. For ex-
ample, our production and manipulation of fire was aided by understanding
the chemical process of oxidation. But oxidation is of little use in understand-
ing what makes the sun hot, since nuclear fusion involves events at the sub-
atomic level. Additionally, the development of modern scientific categories
permits scientists to connect and unify their understanding in ways that the
primitive categories do not. Moreover, science and technology develop to-
gether, which means that our everyday life changes as well, sometimes quite
radically.

Now consider a classification fundamental in science for thousands of years
and easily observable by anyone: the distinction between the rahlunary realm
(the universe below the level of the moon) and the super\wa.xy realm (every-

131

Consciousness

“^ire’oid versus “Q/lot fire’oid

slow

oxidation

burning wood

- burning coai

sun

lightning
northern lights
fireflies
comets

fire”

. burning wood ] fas^^

» burning coal loxidationX

- rusting — ' / XlK

I caixification ^

[ bodily metabolismri/ '

rusting
caixification
bodily metabolism

NOT-FIRE„e„

nuclear fusion
thermal emission
spectral emission
biophosphorescence
reflected sunlight

Figure 4.1 In the early stages of a scientific investigation, a thing’s category member-
ship is determined largely by similarities in easily observable properties. Thus the cate-
gory “fire” initially embraced a range of phenomena involving the emission of heat or
light or both. As physics and chemistry progressed, the category fragmented, and sim-
ilarities based on theoretically informed properties became a more useful basis for new
groupings. The upper panel shows items in the old classification of “fire,” and the lower
panel shows the modern classifications. Caixification is oxidation of nonferrous metals.

thing beyond that) (figure 4.2). Completely different principles were assumed
to govern each domain. The superlunary realm was thought to be immuta-
ble, perfect, and governed by divine principles, such as the constant circular
motions of the planets. Here on sublunary Earth, in contrast, things change
unpredictably, they can become rotten and worn, perfectly circular motion
is rare, and Earthly principles such as “Nothing moves unless it has a force
acting on it” and “Everything moves to its natural place” prevail. Thus spake
medieval physics.

Rending all this asunder, Newton proposed that one set of laws could explain
motion wherever in space it occurred. The planetary motions, the trajectory of
an arrow, the movement of the moon, the falling of an apple — all are embraced
by a single set of laws. In developing this new framework, Newton dumped the
sublunary/superlunary distinction entirely.

The apparently indispensable and completely intuitive notion of Natural
Place also dropped out of the picture. On the old theory, rain falls down be-
cause it had “gravity,” and things with gravity have their Natural Place in the
center of the universe, namely Earth. Smoke, on the other hand, rises, because
it has levity and things with levity have their Natural Place away from the

132

Metaphysics

Figure 4.2 A schematic characterization of the geocentric conception of the universe,
viewed looking down on Earth, the Moon, and the embedded series of crystal (trans-
parent) spheres. The realm inside the first of the crystal spheres (the sublunary realm)
was presumed to be governed by very different physical laws than those governing
bodies and events in the superlunary realm. The “fixed” stars are attached to the outer-
most crystal sphere, which does not move, whereas the planets. Moon, and Sun are
attached to the intermediary spheres, whose rotations were believed to explain the
movement of these bodies. A major problem for medieval physics was to explain what
caused the huge glass spheres to move. This problem was abandoned with the advent of
Newton’s radically different explanation of the movements of the planets. Moon, and
Sun. (Courtesy of P. M. Churchland.)

133

Consciousness

center of the universe. Newton replaced the old conception of gravity with a
completely new conception: a reciprocal force between any two masses. The
seemingly obvious idea of Natural Place, comfortably entrenched for roughly
two thousand years, thus found its natural place in the scrap heap.

The more general lesson is this: theories about certain things and definitions
as to what in the world count as those things evolve together, hand in hand.
Firm, explicit dehnitions become available only fairly late in the game, as the
science that embeds them hrms up and matures."^

What, then, about dehning “consciousness”? If we cannot begin with a solid
definition, how do we get agreement on what phenomenon we are trying to
study? Roughly, we use the same strategy here as we use in the early stages of
any science: delineate the paradigm cases, and then try to bootstrap our way up
from there. Using common sense, we begin by getting provisional agreement on
what things count as unproblematic examples of consciousness.

First in the set of prototypically conscious states are a range of sensory per-
ceptions, such as seeing a bird fiy, feeling the pain of a burn, hearing a police
siren. The somatic sensory experiences pertaining to touch, vibration, pressure,
limb position, body orientation, and body acceleration are also included in the
prototype. Smells and tastes round out the list of sensory perceptions.

Second, we can include in our list states that, as are not usually considered
sensory experiences per se because they are not so closely tied to a specific sen-
sory organ. This list includes such states as remembering what you had for
breakfast, knowing that you can ride a bicycle, imagining a six-legged dog,
attending to the feeling in your big toe, wondering whether to eat a mango,
surprise that an expected event did not happen, and so forth. Likewise, emo-
tional states, such as feeling fear, anger, sadness, and elation, as well as drive
states, such as hunger, thirst, sexual desire, and parental love belong on the list.
In this context, we also need to distinguish between capacities,, which are dis-
positions, and the current exercise of those capacities. The contrast is between
the capacity to remember what you had for breakfast (though you are not
thinking about that now) and your remembering now that you had sausages for
breakfast.

Occupying a still less central location in the conscious awareness prototype
space are a host of other cases. Probably we are at least somewhat aware dur-
ing dreaming states, even when they are not recallable. We are not sure whether
we are aware at all during deep sleep, or whether a kind of low-level awareness
persists throughout. For example. Navy SEALS are trained to react to a threat
even before they are awake. We are not sure when the fetus’s nervous system is

134

Metaphysics

sufficiently developed that it becomes aware of sensory stimuli such as sounds.
Further, it is uncertain how we should think of conscious states such as recog-
nizing that something is unfamiliar or odd, or that something is intellectually
satisfying, morally unsettling, musically harmonious, or esthetically jarring.

Fortunately, we need not worry too much at this stage about these cases. By
identifying prototypical examples of conscious states, we gain lots of scope for
designing revealing, interpretable experiments. With some progress in hand,
less central examples may come to assume greater importance, perhaps even
gain recognition as the prototypical cases.

Cognizant of the possibility that these ostensibly obvious categories may be
reconhgured later under the pressure of new discoveries, perhaps we can agree
that this rough-and-ready delineation of prototypes provides us with a reason-
able way to get the project off the ground. Because the neuroscientific approach
to consciousness is young, the reasonable hope is for discoveries that will open
more doors and suggest fruitful experimental research. In the long haul, of
course, we want to understand consciousness at least as well as we understand
reproduction or metabolism, but in the short haul, it is wise to have realistic
goals. It is probably not realistic to expect, for example, that a single experi-
mental paradigm will solve the mystery.

1.3 Experimental Strategies

Although there are many proposals for making progress experimentally, for
convenience the strategies targeting the brain can roughly be grouped as one of
two kinds: a direct approach or an indirect approach. These strategies differ
mainly in emphasis. In any case, as will be seen, they are complementary, not
mutually incompatible. To see the strengths and weaknesses of each, I shall
outline the somewhat differing motivations, scientihc styles, and experimental
approaches.

The direct approach

It is possible, for all we can tell now, that consciousness, or at least the sensory
component of consciousness, may be subserved by a physical substrate with a
distinctive signature. In the hope that there is some distinct and discernible
physical marker of the substrate, the direct strategy aims hrst to identify the
substrate as a correlate of phenomenological awareness, then eventually to get
a reductive explanation of conscious states in neurobiological terms. The phys-

135

Consciousness

ical substrate need not be confined to one location. It could, for example, con-
sist in a pattern of activity in one or two structurally unique cell types found in
a particular layer of cortex across a range of brain areas. Or it could consist in
the synchronized firing of a special cell population in the thalamus and certain
cortical areas. On these alternatives, the mechanism would be distributed, and
hence would be more like the endocrine system, for example, than the kidney.
For convenience, I shall refer to a postulated physical substrate as a mechanism
for consciousness.

Notice also that the distinctive mechanism could reside at any of a variety of
physical levels: molecular, single cell, circuit, pathway, or some higher organi-
zational level not yet explicitly catalogued. Or perhaps consciousness is the
product of interactions between these myriad physical levels. The possibility
of a distributed mechanism, together with the opened-ended possibility con-
cerning the level of organization at which the mechanism inheres, means that
hypotheses are so far quite unconstrained. The lack of constraints is not a
symptom of anything otherworldly about this problem. It is merely a symptom
that science has a lot of work to do.

Discovering some one or more of the neural correlates of consciousness
would not on its own yield an explanation of consciousness. Nevertheless, in
biology the discovery of which mechanism supports a specific function often
means that the next step — determining precisely how the function is performed
— suddenly becomes a whole lot easier. Not easy, but easier. Were we lucky
enough to identify the hypothetical mechanism, the result would be comparable
in its scientific ramifications to identifying the structure of DNA. That discov-
ery was essentially a discovery about structural embodiment of information.
Once the structure of the double helix was revealed, it became possible to see
that the order of the base pairs was a code for making proteins, and hence to
understand the structural basis for heritability of traits. In the event that there
is a mechanism with a distinct signature identifiable with conscious states, the
scientific payoff could be enormous. The direct strategy, therefore, is worth a
good shot.

The downside, of course, is that the mechanism might be experimentally very
difficult to identify until neuroscience is much further along, since the signature
may not be obvious to the naive observer. Our current misconceptions about
the phenomena to be explained, or about the brain, may lead us to misinterpret
the data even if the mechanism with its distinct signature exists to be identified.
Or there may be other unforeseeable pitfalls to bedevil the approach. In short,
all the usual problems besetting any ambitious scientific project beset us here.

136

Metaphysics

In recent years, the direct approach has become more clearly articulated
and more experimentally attractive, in part occasioned by new techniques that
made it possible to investigate closely related functions such as attention and
working memory.

Francis Crick, probably more than anyone else, has a sure-footed scientific
sense of what the direct approach would need to succeed. He has drawn atten-
tion to the value of using low-level and systems-level data to narrow the search
space of plausible hypotheses, and of constantly prowling that search space to
provoke one’s scientific imagination to come up with testable hypotheses. Crick
has consistently recognized and defended the value of getting some sort of
structural bead on the neuroanatomy subserving conscious states, not because
he thought such data would solve the problems in one grand sweep, but be-
cause he realized it would give us a thread, which, when pulled, might begin
to unravel the problem. He argued that experiments probing such a mechanism
could make a plausible assumption, which I henceforth refer to as Crick’s
assumption:

Crick’s assumption There must be brain differences in the following two con-
ditions: (1) a stimulus is presented and the subject is aware of it, and (2) a
stimulus is presented and the subject is not aware of it.®

With the right experiments, it should be possible to find what is different
about the brain in these two conditions.

Within this lean framework, the next step is to find an experimental para-
digm where psychology and neuroscience can hold hands across the divide; in
other words, to find a psychological phenomenon that fits Crick’s assumption
and probe the corresponding neurobiological system to try to identify the neu-
ral differences between being aware and not being aware of the stimulus. This
would give us a lead into the neural correlate of consciousness and hence into
the mechanism. Fortunately, a property of the visual system known as binocular
rivalry presents just the opportunity needed to proceed on Crick’s assumption.®

What is binocular rivalry?

Suppose that you are looking at a computer monitor through special box with
a division down the middle, so each eye sees only its half of the screen. If the
two eyes are presented with the same stimulus, say a face, then what you see is
one face. If, however, each eye gets different inputs — the left eye gets a face,
and the right eye gets a sunburst pattern — then something quite surprising

137

Consciousness

Stimuli

left eye

right eye

Subject's perception

0 1 2 3 4 5 6

seconds

Figure 4.3 Bistable perception resulting from binocular rivalry. If different stimuli are
presented to eaeh eye, after a few moments of confusion, the brain settles down to per-
ceiving the stimuli in an alternating sequence, where the perception of any given stimu-
lus lasts only about 1 second. (Courtesy of P. M. Churchland.)

happens. After a few seconds, you perceive alternating stimuli: first sunburst,
then face, then sunburst, then face. The perception is bistable, favoring neither
one over the other, but switching back and forth between the two stimuli
(figure 4.3). The reversal happens about once every 1-5 seconds, though the
rate can be as long as once every 10 seconds. Many different stimuli give
bistable perceptual effects, including horizontal bars shown to one eye and
vertical bars to the other. So long as the stimuli are not too big or too small,
the effect is striking, robust, and quite unambiguous.^

For the purposes of Crick’s assumption, this setup is appealing: the opposing
stimuli (e.g., the face and sunburst pattern) are always present, but the subject
is perceptually aware of each only in alternating periods. Consider, for example,
the face. It is always present, but now I am aware of the face, now I am aware
of the sunburst pattern. Consequently, we can ask. What is the difference in the
brain between those occasions when you are aware of the face and those when
you are notl

Precisely why binocular rivalry exists is a question we leave aside for now,
as there are various speculations but no definitive answer. It is fairly certain,
however, that it is not a retinal or thalamic effect, but an effect of cortical
processing. The most convincing hypothesis, favored by Leopold and Logo-
thetis, is that binocular rivalry results from a system-level randomness that

138

Metaphysics

Figure 4.4 A diagram of human brain from the medial aspect showing the projections
from the retina to the lateral geniculate nucleus of the thalamus and midbrain (superior
colliculus and pretectum), and from the thalamus to cortical area VI of the cerebral
cortex. (Based on Kandel, Schwartz, and Jessell 2000.)

typifies exploratory behavior in general and whose function is to ensure that the
brain does not get stuck in one perceptual hypothesis.®

On the neurobiological side, what is experimentally convenient about bino-
cular rivalry is that in the visual system, cortical area STS (superior temporal
sulcus) is known to contain individual neurons that respond preferentially to
faces. This “tuning” of neurons, as it is called, is something that can be
exploited by the experimentalist in the binocular rivalry setup (figures 4.4 to
4.6). This means that the cellular responses during presentation of rival stimuli
can be recorded and monitored.

Area STS was identified, and its tuned neurons characterized, using single-
neuron recording techniques in the monkey. This technique involves inserting a
microelectrode into the cortex and recording the action potentials in the axon
of a single neuron (figure 4.7).® On the basis of lesion data and fMRI studies,
we know that human brains also have areas that are especially responsive to
faces. Although such macrolevel data are extremely important, it has to be
balanced by microlevel data from the single neuron. By and large, looking for
single neurons whose activity correlates with conscious perception is something

139

Consciousness

gyrus

Figure 4.5 Schematic representations of the temporal lobe of human brain (shaded
areas). The upper panel shows a side view (lateral aspect), and the lower panel shows the
underside (ventral aspeet). There are three general regions on the lateral surface of the
temporal lobe: the superior temporal gyrus, the middle temporal gyrus, and the inferior
temporal gyrus, which extends around to the ventral aspect of the temporal lobe. The
ventral aspect includes the fusiform gyrus, also referred to as the oecipitotemporal gyrus,
and the parahippocampal gyrus, also referred to as the lingual gyrus. Abbreviations: its,
inferior temporal sulcus; ots, occipitotemporal sulcus; sf, Sylvian fissure; sts, superior
temporal sulcus. (Based on Rodman 1998.)

140

Metaphysics

Figure 4.6 Recordings of activity of a cell with a large reeeptive field in the superior
temporal gyrus as pictures of faces, degraded faces, or nonfaces are visually presented to
a monkey. The eell responds most vigorously to faces, human or monkey or baboon.
Activity is diminished if the eyes are removed or if the faee’s features are all present but
jumbled. It responds better to a cartoon face than to the jumbled features or a nonface.
When the monkey is shown a hand or a meaningless pattern, the cell response drops to
its base firing rate. (From Bruce et al. 1981.)

141

Consciousness

Intracellular recording by mircoelectrode

Figure 4.7 An idealized experiment for measuring the potential difference aeross a cell
membrane. The electrode is a fine glass capillary with a tip no more than . 1 micrometer
in diameter, filled with a saline solution.

that must be done in monkeys. Nevertheless, by using an existing medical op-
portunity, Kreiman, Fried and Koch (2002) were able to repeat the Logothetis
experiment in fourteen human surgical patients. Each had intractable epilepsy.
To localize the seizure onset focus before surgery, eight depth electrodes were
implanted in the medial temporal lobe of each patient. Recordings from these
electrodes during bistable perception showed that about two thirds of the visu-
ally selective cells tracked the percept; none tracked the perceptually suppressed
stimulus. Macaque monkeys are a good substitute for humans in the binocular-
rivalry experiment because human and monkey brains are structurally very
similar, and in particular, their visual systems are organizationally and struc-
turally very similar. There is nevertheless a residual problem in using monkeys
instead of humans, which is that humans can verbally answer “face” when they
see a face, but the monkey cannot.

The tactic for overcoming this human/monkey difference is to train the
monkey to respond by pressing a button with its left or right hand to indicate
whether it sees a face or a sunburst. Monkeys are first trained in a standard
(nonrivalrous) paradigm in which there is a correct answer and they are
rewarded accordingly. That is the only way we have, so far, to let the monkey
know what behavior we want. Once trained, monkeys are presented with

142

Metaphysics

rivalrous stimuli (face to one eye, sunburst to the other) to see how they re-
spond. It is reassuring that monkeys’ response behavior matches that of
humans: it indicates an alternation in perception of the face versus the sunburst
at about once per second.

A specific and significant doubt remains, nonetheless. Although monkeys
may indeed be visually aware, they may not be using visual awareness to solve
this problem. We know from human psychophysics that subjects can perform
well above chance on a visual identification task even though they report that
they are merely guessing their answers rather than judging on the basis of a
conscious perception.

What adds fuel to this doubt is that the learning curves of the monkeys look
like the learning curves of operant conditioned rats. In other words, we cannot
assume that the experimenter’s intent suddenly dawned on the monkey and it
thought to itself, “Oh I get it. When I see faces I press this button, and when I
see sunbursts I press that one!” and with that insight its performance jumps
to nearly perfect. In fact, the monkeys show gradual improvement over days
and even weeks rather than an abrupt improvement indicative of insight. The
learning curves mean that the behavior of the animals is consistent with the
possibility that connectivity is strengthened between visual area STS and motor
cortex without visual awareness being part of the loop after all.

It is highly desirable to find ways to determine empirically, with a decent
degree of probability, whether the animal uses conscious visual perception to
solve the problem. Flexibility in response might be such an indicator. For ex-
ample, if the monkey uses awareness to solve problems in anything like the way
humans do, then the monkey should be able quickly to learn a new motor
action to respond to the very same stimulus. If it uses both the new and the
original response, the two should agree. The monkey should also appear sur-
prised if a particular trial is easy and it gets the answer wrong. This sort of
flexibility is characteristic of human conscious perception, and it is the kind of
thing that should be demonstrable if the monkey is using visual awareness in
solving the problem. Although we must shelve this problem for now, it is
essential to acknowledge the need for developing experimental procedures on
animals that overcome these problems.^®

Inspired by the empirical problems confronting the experimentalist, the a
priori skeptic might tender a much more tenacious skepticism about animal
awareness. For example, the skeptic might complain that the monkey can only
exhibit behavior, whereas the human can actually talk. So, the objection con-
tinues, we have no reason to think that the monkey is aware at all, ever, under

143

Consciousness

any conditions. “ The objection presupposes that speech is really a direct indi-
cation of consciousness, whereas button pressing is not.

Notice, first, that speech too is just behavior — behavior that humans have
learned to perform. Even if the monkey did show verbal behavior, the deter-
mined skeptic would still complain that we could not be certain that its speech
involves awareness as human speech does. Bonobo chimpanzees such as Kanzi
and Pambanisha do display some verbal behavior, but the a priori skeptic
waves this off as “mere conditioning.”^^ We are now venturing into Skepti-
cism, with a capital “S.”

A thoroughly general Skepticism takes the form “How do I know that
any person, let alone some monkey, is ever conscious? Indeed, how do I know
that anything other than I exists? And moreover, how do I know that / was
conscious before this very moment?” Part of the trouble with this brand of
skepticism is that no empirical controls could allay the doubt one whit, in prin-
ciple. The Skeptic thus overplays his hand, with the consequence that general
Skepticism is hard to take very seriously beyond a moment or two.

A Skeptic can insist that there is no decisive proof that one is not dreaming,
or that the universe was not created five minutes ago complete with fossil rec-
ord, memories, history books, crumbling Roman ruins, and so on. Indeed,
there is no decisive proof of the impossibility of what was just sketched. Still, as
a hypothesis about reality, it is a bit silly. Specific doubts about a specific ex-
periment are a very different matter, however, and they do indeed have to be
answered, one and all. In the absence of identifiable reasons for thinking that
only humans can be visually aware, the similarities in monkey and human
brains suggest that it is reasonable for me provisionally to assume that the
monkey has visual awareness qualitatively not very dijferent from ours. This is
not a dogmatic declaration that monkeys are indeed visually aware as we are,
but it is a useful working assumption, one that can sustain some interesting
experiments. Nonetheless, it could be false, and it could be falsified empirically.

The binocular rivalry experiments

The neural correlates of visual awareness in binocular rivalry were first ex-
perimentally probed by neuroscientists Nikos Logothetis and Jeffrey Schall in
1989. Logothetis and Schall were using upward-moving and downward-moving
gratings as stimuli. Their monkeys had been trained in advanced to indicate
what they saw by pressing specific buttons, and the recording of single cells was
done in visual cortical area MT. More recently (1997), Scheinberg and Logo-

144

Metaphysics

thetis have used a face and a sunburst pattern, and recorded in STS. Hence-
forth I shall frame the discussion around the face/sunburst stimuli, and I shall
say “The monkey sees the face” as shorthand for “The monkey presses the
button indicating its learned response to face stimuli,” and so forth.

Simplified, the results are as follows. Consider a set of neurons, Ni, . . . , N 5 ,
that were previously identified as responding preferentially to faces. (Suppose,
for simplicity in this discussion, that faces are always present to the left eye, and
sunbursts always to the right eye.) What do those neurons do when the monkey
sees the sunburst! Some of them, perhaps Ni and N 2 , continue to respond,
because of course the face is still present to the left eye, even if it is not con-
sciously seen. Other face neurons, perhaps N3 and N4, do not respond. Now
for the critical result: when and only when the monkey indicates that it does see
the face, N 3 and N 4 respond (and as always, Ni, N 2 respond so long as the face
is present) (figure 4.8).

Here is why this is interesting. Some neurons seem to be driven by the ex-
ternal stimulus; that is, they respond to the stimulus regardless of whether
the monkey consciously perceives it. Others seem to respond only when the
monkey sees — consciously sees — the stimulus. More exactly, the distribution of
responses in STS was this: about 90 percent of the face neurons fire when and
only when the monkey indicates it sees a face; the remainder always fire so long
as the face is present on the monitor.

Can we say that the responsivity of the neurons in the 90 percent pool is
correlated with visual perception (visual awareness)? Yes, but we need to go
carefully here. Over a fairly generous time scale, “correlated with” could
include events that are not identical with the state of perceptual awareness
but are part of the causal sequence. More exactly, the data do not exclude
the possibility that the responses of STS neurons are actually the causal
antecedents — or possibly causal sequelae — of neural activity that is the aware-
ness. In other words, we cannot simply conclude that this subset of STS neu-
rons is the seat of visual awareness of faces. Progress has been made, but we do
not want to overstate our conclusions.^'^

Although the binocular-rivalry experiments are a little complicated, they are
important because they illustrate something that will surprise convention-
bound philosophers. With the right experiment, you can make progress, even
at the level of the single neuron, in investigating the neural causes or neural
correlates of visual awareness. It shows, contra the naysayers, that headway,
albeit only a little, is possible. Moreover, image data, using fMRI on humans,
is consistent with the single-neuron results. With further experiments, this
beginning allows us to push on into territory that will be fruitful.

Figure 4.8 The neuronal responses of a faee cell in the monkey brain during bistable
perception. In the experiment, a monkey is trained to hold down one lever, e.g., the
right-hand lever, when it sees a face, and to hold down the other lever when it sees a
sunburst pattern. (A) The four horizontal graphs represent four observation periods, and
the dashed vertical line indicates the onset of a rivalrous presentation (e.g., face and
sunburst pattern). The animal’s behavioral response is shown below the line, the shaded
area representing the period during which animal holds down the appropriate lever. The
cell response is shown above the line. The high rate of activity of the face cell begins just
before, and ends just before, the period during whieh the animal holds down the face
lever. The period of high activity (between 0 and 50 spikes/second) lasts for about
1 second. (B) The brain areas that contained the eells whose activity correlated with the
monkey’s subjective perception when responding to stimuli known to drive cells in that
area. The greater the synaptic distance of the eortical area from the retina, the greater
the percentage of cells driven by the subjective pereeption. Abbreviations: IT, inferior
temporal; MT, middle temporal; MST, medial superior temporal sulcus; STS, superior
temporal suleus; VI, striate cortex; V2, V4, extrastriate cortex. (From Leopold and
Logothetis 1999.)

146

Metaphysics

Other experiments, similarly motivated, link up with the Logothetis results.
Here is one strategy. To get a visual perception called “the waterfall illusion,”
you stare at a waterfall for several minutes. When you look away at a still
surface, such as a gray blanket, you see upward motion, a kind of reverse, and
illusory, waterfall. Roger Tootell used this phenomenon to run an experiment
that complements the Logothetis and Schall experiment. The focus here will be
on the neural correlates of conscious perception of upward motion induced in
the absence of an externally present upward stimulus. Tootell used the non-
invasive scanning technique fMRI to determine what cortical visual area
showed greater activity when a human subject consciously perceives the water-
fall illusion. He found, not unexpectedly, that motion-sensitive areas such as
MT show increased activity with the onset of perception of the waterfall illu-
sion. In this experiment too, it remains unknown whether MT neurons are
actually neural correlates of consciousness, or whether they are just an element
in the causal antecedents or consequences thereof.^®

Hallucinations in human subjects present a different possibility for exploring
what happens in the brain when a visual experience is present but the stimulus
is not. Recently this has been elegantly pursued using fMRI by a group in
London led by Ffytche.^^ Patients who suffer eye damage, for example as a
result of detachment of the retina or glaucoma, lack normal vision. In some
cases, these patients periodically experience highly vivid visual effects, though
they are perfectly normal neuropsychiatrically. The character of the hallucina-
tions varies from subject to subject, and unlike visual imagery, the visual
objects appear to be in the outside world, and neither their appearance nor the
nature of the visual image is under voluntary control.

One subject saw cartoonlike faces; another saw colored, shiny shapes rather
like “futuristic cars.” In the fMRI scanner, subjects signaled the onset of their
visual hallucinations, and the scan data were analyzed. The data showed asso-
ciation of hallucinations with activity in the ventral visual regions, but with
little activity in early visual cortex (VI). More specifically, if a subject halluci-
nated in color, an area independently identified as important in color process-
ing was more active than if the hallucination was in black and white. Face
hallucinations were associated with cortical subareas independently known to be
involved with face processing, including the inferior temporal region (figure 4.9).

What do the Ffytche data mean? On their own, they do not solve the mystery,
of course, but they are at least consistent with the data from binocular rivalry
and from the waterfall illusion. These converging data suggest that a subset of
neurons in visual cortical areas may support conscious visual perception.

147

Consciousness

Figure 4.9 Bilateral lesions in the shaded region cause propopagnosia (loss of the
capacity to identify individual faces). (Courtesy of Hanna Damasio.)

Figure 4.10 Visual masking. As the subject views the monitor, a word is presented,
followed about 10 msee later by a noisy jumble — the mask. In these conditions, the
subject sees only the mask, not the word.

Another experimental approach, also using fMRI, involves comparing brain
activity during presentation of stimuli that are not consciously perceived and
during presentation of stimuli that are consciously perceived. The experiments
exploit an earlier behavioral result by Anthony Marcel, in which he showed
that nonperceived stimuli had a quantihable effect on subject’s task perfor-
mance. More specihcally, Marcel flashed a word for about 10 msec., then im-
mediately followed the word with a masking stimulus (a noisy visual stimulus
flashed in the same location as the stimulus). The presentation of the mask
somehow interferes with normal visual processing and the flashed item is not
seen (figure 4.10). Subsequently, subjects were given a lexical-decision task, in
which a string of letters was presented and the subject’s task was to specify
whether the string was or was not a word. Marcel showed that the subject’s
performance, measured in reaction time, was better for those words that had
been presented in the masked condition than for words never presented. More-
over, processing of the flashed stimulus went beyond the mere physical shape

148

Metaphysics

of the stimulus because the effect was case-insensitive, (i.e., “BIRD” versus
“bird”). This elegant experiment demonstrated a level of semantic processing
even when subjects reported no conscious perception of the stimulus.

Dehaene and colleagues used the Marcel paradigm and recorded activity in
normal subjects using fMRI in the masked and the visible conditions.^® They
showed that even in the masked condition, there is activity in both the fusiform
gyrus and the precentral gyrus, areas that independent experiments indicate are
active during conscious reading (see again fig. 4.9). In the condition where the
stimulus was seen and not masked, the activity in the fusiform gyrus appeared
to be about twelve times as strong as in the masked condition, and there was
additional activity in the dorsolateral prefrontal cortex. The data suggest that
the difference in brain activity in the two conditions is owed to conscious
awareness of the stimuli.

Clever as the experiment is and important though the data are, several cau-
tions are in order. First, the areas showing increased activity involve hundreds
of millions of neurons, so the data are giving us a very general portrait, not
detailed information about specific neurons or neuron-types and their role in
awareness. Second, the data are consistent with the possibility that the greater
activity in the nonmasked trial is caused by activation of a large range of neural
networks whose stored information is associated with the flashed word. The
mask may have associations too, but many fewer than a word. In the masked
case, activation of networks associated with the word is probably interrupted
by the mask, whereas the mask, being junk, provokes few associations. As
the authors rightly note, the effects of the mask appear to start very early in the
visual system, and propagate to higher levels. If the greater activity seen in
the nonmasked case reflects greater numbers of activated associations, these
associations might well be entirely nonconscious. They might be caused by a
conscious representation, or by whatever it is that causes the conscious repre-
sentation. Consequently, we cannot be sure that the greater range of activation
in the unmasked case corresponds to conscious activity per se.^®

Loops and conscious experience

An idea that has long been central to the approach of neuroscientist Gerald
Edelman^° is that loops (also referred to as re-entrant pathways and as back
projections) are essential circuitry in the production of conscious awareness.
The idea is that some neurons carry signals from more peripheral to more cen-
tral regions, such as from VI to V2, while others convey more highly processed

149

Consciousness

signals in the reverse direction, for example from V2 to VI. At an anatomical
level, it is a general rule of cortical organization that forward-projecting
neurons are matched by an equal or greater number of back-projecting neu-
rons. Back-projecting neurons are a feature of brain organization generally,
and in some instances, such as the pathway from VI to the lateral geniculate
nucleus (LGN) of the thalamus, back-projecting neurons are more numerous
by a factor of ten than the forward-projecting neurons. Anatomically, then, the
equipment is known to exist.

Why do Edelman and others think back projections have some particular
role in consciousness? Part of the rationale for this point is that perception
always involves classification; conscious seeing is seeing as}^ Normally, one
sees a fearful human face as fearful, rather than simply as a face followed
by the explicit inference, “Aha, the eyes are especially wide open, etc., so this
face is showing fear.” In fact, most of us instantly recognize a fearful face but
cannot articulate precisely what configuration of facial features is required for a
face to show fear (figure 4.11). So we could not say what an explicit inference
could use for premises, anyhow. Smells are often imbued with a hedonic
dimension of meaning. The smell of rotten meat, for example, is disgusting to
humans, whereas to vultures, it is appealing. Separating in experience the pure
odor of rotten meat from the anhedonic nastiness of the smell is impossible.

Integrating hedonic components, emotional significance, associated cognitive
representations, and so forth, with features of perception detected by the sen-
sory systems almost certainly relies on loops — pathways projecting a signal
back from structures such as the amygdala and hypothalamus (which have
powerful roles in emotions and drives) to the sensory systems themselves, and
pathways from so-called higher areas of cortex (e.g., prefrontal regions) to
lower areas (e.g., VI). That we directly perceive a face with its fearful expres-
sion implies that information about the emotion must be routed back to the
visual system at some level. A purely feedforward neural network cannot
achieve this kind of integration.

Artificial neural network (ANN) research indicates that many of the con-
sciousness-related functions — STM, attention, sensory perception, meaning —
are handled most powerfully and efficiently by networks with recurrent projec-
tions. The range of functions that back projections perform has not been pre-
cisely demonstrated in real neural networks, and there are serious technical
difficulties to be overcome before back-projection physiology can get very far.
Nevertheless, the fact that back projections in ANNs render those systems
vastly more powerful, and more powerful in the ways relevant to consciousness-
related functions, is highly suggestive.

Figure 4.11 Human facial expressions of four emotions: fear, anger, sadness, and hap-
piness. (Faces eourtesy of Dailey, Cottrell, and Reilly. Copyright 2001 California Facial
Expressions Database [CAFE].)

151

Consciousness

Figure 4.12 A schematic diagram of the human brain showing the position of VI. On
the left is the medial view; on the right is the lateral view. In the visual cortex, VI is
located in the calcarine sulcus in the medial aspect, shown in dark shading. The extra
striate cortex is shown with dotted shading. (Courtesy of Hanna Damasio.)

Experimental evidence is beginning to come in to support this idea. For ex-
ample, Pascual-Leone and Walsh exploited the fact that transcranial magnetic
stimulation (TMS) of cortical visual area VI will cause the subject to experi-
ence small flashes of light, while stimulation of cortical visual area MT will
produce flashes of light that move.^'^ The anatomical fact of importance is that
there are back projections from MT to VI. (In fact the back projections typical
of cortical organization are also seen in the brainstem and spinal cord, as well
as in structures such as the hypothalamus. They are essentially everywhere.) So
here is their experiment: stimulate MT in a manner normally adequate to pro-
duce moving flashes of light, and also stimulate VI, but at an intensity so low
that it does not cause perception of lights, but high enough to interfere with the
normal effect of back-projected signals from MT. If back-projected signals
from MT are necessary to see moving flashes, then in this condition, no moving
flashes will be seen. These are indeed the results. Subjects see flashes, but not
moving flashes.

As always, optimism must be tempered with skeptical questions. One major
question concerns what exactly is the effect of TMS at the neuronal level, how
focal the stimulation really is, and how far the effect spreads, cortically and
subcortically. A further problem arises from the nature of human brain anat-
omy. In the macaque monkey, VI is on the dorsal surface of the brain. In
human brains, VI is on the medial surface of the occipital lobe (figure 4.12).

152

Metaphysics

Consequently, if you aim to stimulate VI with TMS, you will also stimulate the
dorsal regions, and activity in the pathways from the incidentally stimulated
areas can be predicted to affect both VI and V2. The worry is that the inci-
dentally stimulated areas confound the results.

In any case, even if back projections are necessary for consciousness, we
know that they are not sujficient. Back projections function in phylogenetically
older parts of the brain, such as the spinal cord; some are active when subjects
are under anesthesia, in a deep sleep, or in a coma. If a subset of cortical back
projections are indeed subserving awareness of visual stimuli, it will be impor-
tant to determine which axons they are and what precisely is the nature of their
signals.

Theorizing and narrowing the hypothesis space

In addition to designing experiments to identify the neural correlates of con-
sciousness, pulling together data bearing on the conditions for visual expe-
rience and isolating structural and functional constraints can help narrow the
hypothesis space. Especially in the early stages of the problem, this is a very
useful strategy, particularly because some of the concepts needed to articulate a
good hypothesis undoubtedly need to be invented as the search space narrows
ever more.

Loops are likely to be one structural constraint on the substrate for con-
sciousness. As Francis Crick and Christof Koch suggest, other constraints that
emerge from the experimental literature include the following:

■ The neurons whose collective activity constitutes being aware of something
are distributed spatially. Transiently, they form a “coalition” that lasts for
the duration of the awareness of a particular perception, such as visual
awareness of Lincoln’s face. Individual neurons can be elements in different
coalitions as a function of the percepts. For example, a particular neuron
might be part of a coalition that constitutes being aware of Lincoln’s face,
but it also might be part of a coalition that constitutes being aware of a
human hand, or a coalition for a dog face.

■ Neurons in the coalition whose activity constitutes a perceptual awareness
probably need to reach a threshold in order for the coalition’s activity to
constitute perceptual awareness.

■ Normally, though perhaps not necessarily, a coalition emerges as a con-
sequence of synchrony of firing in neuron populations that project to the

153

Consciousness

coalition members. This synchrony of bring is part of the causal conditions
for reaching the threshold.

■ When neurons involved in perceptual awareness do fire above that threshold,
they continue firing for a short but sustained period of time (e.g., longer than
100 milliseconds but not as long as a minute).

■ Attention probably up-regulates the activity of the relevant neurons, getting
them closer to their threshold.

■ In awareness of a certain visual phenomenon, say the face of Lincoln, some
neurons will be activated as part of the cognitive background, while some
will be activated as essential to the experience itself. These latter neurons
Crick and Koch call “essential nodes,” to distinguish them from neurons
that contribute to the cognitive background. Included in the cognitive back-
ground are the expectation that the face is the front of the head, and various
nonconscious, tacit beliefs, e.g., that if Lincoln had been born in Australia,
he would not have been president of the United States. The cognitive back-
ground includes also various associations and inferential connections, for
example, the association with the civil war, and the capacity to infer from
“Lincoln was president of the United States in 1864,” the statement that
“Lincoln is not now president of the United States.”

■ At any given moment there is probably a competition between various
essential-node neurons for which neurons will fire at the threshold and hence
which representation will be conscious. Thus, if I am paying close attention
to events on television, I may not hear the lawnmower running next door.
This implies that the essential-node neurons in the auditory system will have
lost out in the competition to those in the visual system representing the
events on the television.

Ideally, the items in this list will jell to form a kind of prototheory of neural
mechanisms supporting perceptual awareness. In the role of prototheory, the
list may provoke experiments to confirm or disconfirm any one of its items, and
thus move us closer to understanding the nature of consciousness. Having some
sort of theoretical scaffolding is a clear improvement over groping haphazardly.
Even if none of the items on the list turns out to be part of the explanation of
consciousness, the exercise is valuable, because it orients us toward thinking of
the problem of consciousness in terms of mechanisms, that is, in terms of causal
organization. Identifying neural correlates is one thing, and likely a useful thing.

154

Metaphysics

but the goal we ultimately want to reach is identifying causal mechanisms so as
to understand how consciousness occurs.

A methodological question about neural correlates

In the foregoing experiments, there was evidence of neural activity correlated
with conscious awareness. Nevertheless, I expressed caution concerning what
such correlational evidence signifies. The major reason has already been stated:
finding correlations between neural activity and a subject’s reports of percep-
tual awareness is consistent with any of the following: (1) the neural activity is a
background condition for perceptual awareness, (2) the neural activity is part
of the cause, (3) the neural activity is part of the sequelae of the awareness,

(4) the neural activity parallels, but plays no direct role in, perceptual aware-
ness, and (5) the neural activity is what perceptual awareness can be identified
with (the identficand).

Ultimately, if we want to be able to explain the nature of consciousness in
neural terms, what we seek is the identification of some class of neural activity
with perceptual awareness. That is, we want our data to justify interpretation

(5) . As is evident, however, correlational data per se do not rule out all alter-
natives except (5). That some event v is a correlate of some phenomenon y does
tell you a little, such as that you may be on the right path for finding the iden-
tificand. For similar reasons, that some event z fails to correlate with some
phenomenon y suggests that you may be on the wrong path. This is not the
whole pudding, nor is it nothing, and one has to start somewhere.

Determining that two phenomena are systematically correlated requires test-
ing under a wide range of conditions. It is not enough, for example, to get
fMRI data showing that in awake subjects, a specific cortical visual area is
highly active whenever the subject reports visual awareness of an object. We
want also to know whether there is activity in that brain region when the sub-
ject is not conscious. For example, it is essential to know whether the brain of a
subject in a coma or in a persistent vegetative state or under anesthesia shows
activity in that brain region when a visual stimulus is presented. This is not idle
skepticism. Activity in various cortical areas is known to occur in response to
an external stimulus in precisely these unusual conditions. A patient in a per-
sistent vegetative state, for example, exhibits no signs of awareness, and in
particular, no behavioral sign of awareness when shown a familiar person.
Nevertheless, when the subject was shown familiar faces, the so-called “face
area” of the cortex showed a pattern of increased activity similar to that of

155

Consciousness

the normal subject.^® As Damasio correctly notes, such data are powerful clues
that neurons in the visual cortex may not be the generators of visual conscious
experience. Rather, their activities are representations that the subject might be
aware of if he were conscious. So until the tough cases have been excluded by
experiment, no conclusion can be drawn from correlations in the relatively easy
cases.

There is, however, the deeper problem touched on earlier; it is the problem of
knowing what you are looking at. It is reasonable to hope that there is a class
of neural activity correlated always and only with perceptual awareness, and
that such activity is identifiable as conscious awareness. Nonetheless, even if
there is such a class of activity, knowing that this measured activity belongs to
that class may be discoverable only very indirectly. In other words, we might be
looking straight at an instance of the class without in the slightest recognizing
that it is an instance. This will happen if, as is very likely, the physical substrate
does not have a property that is salient to the naive observer, but is recogniz-
able only through the lens of a more comprehensive theory of brain function.

An analogy may make this point clearer. In the nineteenth century, the na-
ture of light was a profound mystery. Suppose, to be fanciful, that nineteenth-
century physicists address the mystery by seeking the microstructural correlates
of light. They hope that there is a particular class of microstructural phenom-
ena that is always and only correlated with light, and that such activity, or
something connected to it, is identifiable as light. The rough idea is to look for
the “defining property” — the identficand, as we may refer to it.

Since those of us living now have the benefit of post-Maxwellian physics, we
know that the defining property is characterized abstractly and nonobserva-
tionally by the theory of electromagnetic radiation. That is, Maxwell realized
that the equations characterizing light matched perfectly the equations charac-
terizing radio waves, x-rays, and other electromagnetic phenomena. He rightly
concluded that light just is yet another form of electromagnetic radiation. Ob-
servable properties give no hint of this, but the match of deep, unobservable
properties gave the game away.

Here is the question: could our imagined pre-Maxwellian correlation hunters
notice, even if they looked closely, that radio waves and light share that same
deep property? Probably not, since, until they understand a good deal more
about electromagnetic radiation, they lack the conceptual resources to see what
counts as the same property. This is because they do not yet have the slightest
inkling that light is electromagnetic radiation, or that x-rays, gamma rays, etc.,
even exist (see plate 1).

156

Metaphysics

Or think of the problem this way: How would you know, independently of
Lavosier’s work on oxygen, that rusting, metabolizing, and burning are the
same microphysical process, but that sunlight and lightning are notl What
property would you look at? And if you did by luck make a guess that the first
three phenomena share a microstructural property, how would you test your
idea?

This is not to say that looking for the neural correlates of consciousness is
futile. On the contrary, at this very early stage of the neurobiological investi-
gation of consciousness, it is undoubtedly wise to give it the best shot possible.
My point is that it is also wise to recognize the pitfalls and to appreciate that
they are not merely technological, but derive also from the absence of a firmly
planted theoretical framework for understanding how the brain works.

The experiments discussed in this section, and others with a similar general
conceptual slant, are important because they have opened doors. From the
vantage point of 1980, when such experiments were barely conceivable, they
look downright spectacular. At the very least, they inspire researchers to invent
better and better experimental designs. It should be noted, however, that the
examples in this section do share a certain conceptual slant that is open to
criticism. All are focused mainly on the cerebral cortex, and all are drawn from
the visual system. This narrowing of the focus can be valuable, especially when
different experimental strategies unearth complementary results, as those dis-
cussed above do to some extent. Focusing narrowly allows us to probe deeply,
if not broadly, and that can be rewarding.

Nevertheless, for all we can tell now, it could turn out that other modalities
play a role in consciousness that is more straightforward and less complicated
than the role of vision. Possibly, exploration of olfactory or somatosensory
processing will reveal principles obscured thus far. More seriously, it could
turn out that it is not cortical neurons — or not cortical neurons alone — whose
activity is identifiable with awareness, but rather, the activity of various non-
cortical neurons in the brainstem, thalamus, hypothalamus, and so forth. It is
common knowledge that subcortical activity does figure in the causal ante-
cedents. Whether some subcortical activity is more than that, however, is a
possibility we shall explore in section 1 .4.

1.4 The Indirect Approach

Attention, short-term memory, autobiographical memory, self-representation,
perception, imagery, thought, meaning, being awake, self-referencing — all

157

Consciousness

seem to be connected in some way with being conscious of something.
The indirect approach proposes that once we understand the neurobiological
mechanisms of each of these diverse functions and the relations between them,
the story of consciousness will more or less come together on its own. That is,
once we have a more substantial theory of brain function in general, we will
have the means to develop a theory of the conditions under which these func-
tions involve conscious awareness. In this respect, therefore, the approach is
indirect. The strategy favors continuing to investigate, both neurobiologically
and behaviorally, these diverse brain functions and how they connect with each
other. Because its success depends on understanding most of the functions of
the brain, this strategy may take longer to bear fruit than the more direct route.

Is consciousness identifiable with some one of these functions, with, say, being
awake? Is consciousness anything over and above being awake? Consciousness
is not the same as being awake, since you can be awake but still not be con-
scious of your saccadic eye movements or of things you are not attending to
(such as tongue movement) or of masked or subthreshold stimuli. Moreover, we
seem to be conscious of our dreams, even though we are asleep while we dream.

Is consciousness identifiable with paying attentionl^^ Probably not, though
the two may indeed be very closely linked. There is more than one attention
system, and in both, shifts in attention can precede awareness. This implies that
consciousness cannot just be attention. In “bottom-up” attention, a subject
can normally orient to a moving object nonconsciously detected in peripheral
vision. Reading text provides another well-studied example where attention
and visual awareness do not coincide.^® In reading, the eyes do not smoothly
traverse the text, but jump from chunk to chunk (a chunk is about 17 charac-
ters in length). Remarkably, the fovea typically lands on the most informative
part of the chunk, such as on the word “cantaloupe” rather than on the word
“the” (figure 4.13). This shows that eye movements are not a lock-step opera-
tion, but are sensitive to specific features of the stimulus. How is this achieved?
At each eye-movement fixation, the subject reads the text on which the eye is
focused. The foveated text is what the reader is aware of. During this fixation
period, attention shifts to the right, and via peripheral vision, the next suit-
able foveation site is selected. Then the eyes shift, and one is aware of the next
chunk of text.

There are other examples where a brain appears to devote attentional
resources to something the subject is not aware of For example, failure to
suppress noise and irrelevant information messes up performance in tennis
and speech making and problem solving. If however, you consciously and

158

Metaphysics

This sentence shows the nature of the perceptual span.

xxxxxxxxxxxx shows the nature xxxxxxxxxxxxxxxxxxxx

•

Figure 4.13 The attentional (perceptual) span is the zone from which useful informa-
tion can be extracted on a given fixation. The fixation point is indicated by a bullet. The
zone in words (“shows the nature”) displays the width of the attentional span. Notice
that the span is asymmetric. The maximum perceptual span is 2 to 3 characters to the
left (the beginning of the current word) and about 15 characters to the right (2 words
beyond the current word). Regions comprised of xs flank the subject’s attentional span.
During a gaze shift in reading, the next 17 xs are replaced by words. In this reading ex-
periment, subjects remain unaware that xs flank the attentional span and are replaced
with words during a gaze shift. (Courtesy of John Henderson.)

purposefully pay attention to irrelevant information in order to suppress it,
you really mess up. So if suppression is an aspect of an attentional mechanism,
it is presumably a nonconscious aspect. One might counter that such suppres-
sion does not involve attention, because attention, at least for the “top-down”
system, is, by definition, conscious. This move should be avoided. In the
absence of independent supporting evidence, it is circular.

The more important point, perhaps, is that we still have a lot to learn
about the phenomena that we refer to as “attention.” It does appear, for ex-
ample, that different neurotransmitter systems are associated with distinct
aspects of attention; noradrenalin with alerting, acetylcholine with orienting,
and dopamine with suppressing conflicting information. In fact, it is still
unclear what computational tasks attention is supposed to perform.

Consciousness as global workspace

An important corollary of the indirect approach is that to be productive, re-
search should target the contrast between the roles of conscious representations
and nonconscious representations in the cognitive economy as a whole.
Greater flexibility in perception, planning, imagining, reasoning, motor control,
and, in the human case at least, for reporting what you experience has been
touted as the obvious distinctive difference made by consciousness in the
organism’s behavior. If you are not conscious of a touch or a pain, for exam-
ple, then you cannot report that you have it, and if the nonconscious represen-
tation plays any role in motor behavior, it will be a more reflexive than
considered role.

159

Consciousness

How might this flexibility owed to conscious representation be explained in
terms of brain functions? The suggestion is that conscious representations are
more broadly accessible in the brain than are nonconscious representations.
Therefore, the flexibility of cognitive function could be explained in terms of
information distribution. So if we could understand how information is more
broadly accessible, we might make progress in understanding the neurobiology
of conscious representations. Dennett has cast the idea in even stronger terms,
namely, that wide accessibility per se constitutes consciousness. As he asserts,
“Global accessibility is consciousness.”^^

This general concept of accessibility sketched by Dennett was elaborated
with more empirical detail by Baars, who proposed the global-workspace model
of consciousness.^”^ Simplified, Baars’s thought that a state is a conscious state
when its neural information is globally accessible, that is, available for use in
such diverse functions as perceptual categorization, motor control, planning,
decision making, and recollection of long-term memories. His seminal meta-
phor is that of a conventional bulletin board, where information can be posted,
making it accessible to others who can, on a need-to-know basis, access and
read the bulletin board. Readers can also post their own things on the bulletin
board. Some information is not broadcast (the nonconscious signals). His sug-
gestion was that the neurons mediating consciousness are, in their informational
aspects, like a bulletin board; that is, certain neural networks are connected so
as to have what amounts to a shared workspace.

Baars understood very well that the workspace metaphor was only a meta-
phor, and that it must be cashed out eventually in neuronal terms, that is, in
terms of real circuits, real neurons, and real activity. He speculated that the
reticular formation, a finger-like structure in the brainstem known to be essen-
tial in orienting and arousal, was the crucial part of the anatomy for determin-
ing what information got onto the bulletin board at any given time. The
thalamus, with its vast cortical projections, he suggested, is the mechanism for
bulletin-board broadcasting. Although these very general speculations shift the
discussion a little closer to testability, a difiicult part of the task is to specify
what “global access” means in neuronal terms, and to do so without surrepti-
tiously defining “global access” as “conscious access,” and hence getting stuck
with circularity. Why is this difiicult?

Brains, as we have seen, are composed of elements that send and receive
signals carrying information relevant to behavior. Another way of putting this
same point, but viewed from the receiving rather than the sending perspective,
is to say that brains are in the access business. Access, in varying degrees

160

Metaphysics

of breadth, is ubiquitous in nervous systems. But what does “access” mean,
neurobiologically?

To a first approximation, we can say that a neuron b has access to the in-
formation carried by the activity of a neuron a if the activity in a causes activity
in b and this causal link constitutes a transfer of information from aio b. This
is just a dilferent way of saying that neuron a sends information to neuron b.
We could, so far as I can tell, drop talk of access and state the hypothesis solely
in terms of sending and receiving information. Granting that we do not yet
understand how precisely to characterize the nature of information in ner-
vous systems, we can provisionally use this rough-hewn notion. On this basis,
therefore, we can say that information in retinal ganglion cells, for example, is
accessible to neurons in the thalamus but not to neurons in the spinal cord.
And we can say that neurons in the motor cortex make motor commands ac-
cessible to neurons in the red nucleus, the cerebellum, the basal ganglia, and the
spinal cord.

If “access” is the name of the brain game in general, how do we specify in
neuronal terms what global access is in a way that makes it the key to con-
sciousness? “Global” in this context does not, of course, literally mean global,
because given the nature of brain connectivity, the information does not in fact
go everywhere. The key idea must be that the information is made accessible
very broadly. In the recent discussion of the workspace approach, the term is
tied to anatomical properties: there are specific long-distance neurons (long-
axon neurons) in the parietal cortex, cingulate cortex, dorsolateral prefrontal
cortex, and superior temporal cortex that allegedly are the workspace neurons
(figure 4.14). The idea is that increased activity of these long-axon neurons
makes information available to selected populations of neurons, for example,
in the visual system. Before getting on the bandwagon, let us, in normal critical
fashion, give the hypothesis a long, hard look.

First, it is unclear whether the shifting neuronal population whose activity
allegedly constitutes awareness is the sending population (workspace neurons)
or the receiving population or both the sending and receiving populations.

Second, broad accessibility, if anatomically defined in terms of long-axon
neurons, or long-axon neurons with unusually high numbers of neurons in the
projective field, is certainly not unique to the areas cited. For example, it typi-
fies many neurons in motor and premotor cortex, as well as neurons in the
thalamus, amygdala, brainstem, and, well, just about everywhere (figure 4.15).
So defining the workspace neurons by means of these specific structural criteria
is less than satisfactory.

161

Consciousness

Figure 4.14 This schematic diagram is used by Dehaene and Naccache (2001) to illus-
trate some of the connections between the parietal cortex and the prefrontal cortex in
the monkey brain that subserve “global access.” The upper panel presents the medial
view (inverted), and the lower panel the lateral view. The cross-hatched areas receive
projections from the medial pulvinar nucleus of the thalamus, which receives projections
from a variety of areas, including visual cortical area VI and the superior colliculus. Not
shown are the many subcortical connections from the dorsolateral prefrontal and pos-
terior parietal cortical areas to striatum, claustrum, thalamus, and reticular formation.
(Based on Goldman-Rakic 1988.)

Can the workspace population be defined by a combination of structural and
functional criteria? Perhaps the functional criteria can be sketched out by
working backwards from functions made more flexible or deliberate or intelli-
gent by conscious representations. For example, planning or deliberating or
trying to perform a task is greatly hampered if a person is in a coma or in a
persistent vegetative state or asleep or given subthreshold stimuli. Accordingly,
one could hypothesize that the global-access neurons are those active during
tasks that involve attention, effort, or deliberation. On this supposition, a great
deal of effort has gone into determining which brain areas are implicated in the
performance of certain tasks that demand attention, working memory, or con-
scious perception, with the aim of identifying as workspace neurons those
involved in all these functions. Lesions studies, along with fMRI, EEG, and

162

Metaphysics

Supplementary motor area

Figure 4.15 A schematic drawing of some of the pathways between cortical areas and
neuronal structures mediating motor functions. Abbreviations: Gpe, globus pallidus
pars externa; Gpi, globus pallidus pars interna; IL, intralaminar thalamic nuclei; SC,
superior colliculus; SNpc, substantia nigra pars compacta; SNpr, substantia nigra pars
reticulata; STN, subthalamic nucleus; VA/VL, ventral anterior and ventral lateral nuclei
of the thalamus. (Based on Zigmond et al. 1999.)

single-cell studies in monkeys, have been cited as implicating the particular
class of neuronal connections specified in figure 4.14.

So far as so good, but let us test the approach not by selecting favorable
evidence from the available data pool, but by deliberately searching for coun-
terexamples. That is, let us see whether there are circuits that qualify under the
structural and functional criteria, but which even the proponents of the hy-
pothesis would not want to consider as workspace neurons.

As one possible counterexample, consider again eye movements. The ques-
tion is this: do eye-movement neurons in the frontal eye fields qualify as work-
space neurons? If they do, then what doesn’t^ If they do not, why are they

163

Consciousness

excluded? First, eye-movement signals are widely distributed, which means they
are widely accessible. The neurons in the frontal eye fields project to the pons,
prefrontal cortex, parietal cortex, and basal ganglia, requiring long axons.
Hence they satisfy the structural criteria. Moreover, the pathways go to and
from some of the same areas as those cited in the global workspace hypothesis.
Second, activity in these neurons does enhance the ability to see, make deci-
sions, learn skills, etc. Eye-movement information is important in making head
movements, postural adjustments, whole-body movements, and, if you can
make them, ear and nose movements. As noted earlier, patients in a coma or
persistent vegetative state are greatly compromised in their behavioral flexibil-
ity, but they are also compromised in their visual scanning, even if the eyes
are mechanically opened. So the functional criteria are satisfied. Apparently,
given the criteria on offer, therefore, frontal eye-field neurons should qualify as
workspace neurons.

The problem is that we seem not to be conscious of eye-movement decisions.
By and large, we do not consciously deliberate about our incessant eye move-
ments, and we are generally unaware of them. With the possible exception of
rare visits to the optometrist, I have no experience of eye-movement decisions.
The logic so far invites the conclusion that normal eye movements constitute
a counterexample to the workspace criteria for “consciousness neurons.” Al-
though it may be possible to shore up the criteria in a scientifically principled
way, it is far from obvious how to achieve that.

Unfortunately, putting the global- workspace hypothesis under scrutiny makes
it less, rather than a more, comprehensible in neuronal terms. This is especially
disappointing when we recall that according to Dennett’s promising assessment
of the workspace hypothesis, global access just is consciousness. Dennett’s own
story of global access is, of course, complicated by his conviction that the con-
sciousness we humans have is not shared by animals, because they do not have
language and cannot talk to themselves.^®

Moreover, of the features that do make the hypothesis appealing, two are
worrisome. First, much of its appeal derives from our familiarity with bulletin
boards, workspaces, accessibility of websites, and so forth, in the nonneuronal
world. We understand these things in their literal contexts. It would be conve-
nient if the metaphor were true of the brain without too much patching and
doctoring. But is it? Invaluable as metaphors are, they can seduce us into
believing we understand more than we truly do. In the hypothesis at hand,
trucking in the workspace metaphor tends to obscure the fact that it is very
unclear what “global access” means in neurobiological terms.

164

Metaphysics

Second, the allure of the metaphor invites us to celebrate the data that fit and
ignore the data that are awkward. The problem with seeking confirming data is
that just about any theory, false or true, wacky or ingenious, is consistent with
lots of data.^’ This was the case with the caloric theory of heat, the Ptolemaic
theory of the heavens, the grassy-knoll theory of the assassination of Presi-
dent Kennedy, and just about all of alien abduction stories. As Karl Popper
famously insisted, it is easy to get data that fit; what counts for the believability
of a theory, however, is passing a tough test. Finding that some data do sat-
isfyingly fit the hypothesis, we are duty bound to wade into hostile territory and
see whether the hypothesis can survive the really tough tests.

With the critical flags now hoisted, I should add that the hypothesis does
look robust enough to merit testing, and that is more than can be said about
many hypotheses regarding the physical basis of consciousness. The heartening
thing about the global-workspace hypothesis is that it has its feet solidly in the
empirical mud. In the coming years it will be essential to confront the hypoth-
esis with potentially falsifying tests, as well as to assemble supporting data.
Notwithstanding Dennett’s argument that it makes no sense to talk about
animal consciousness, animals experiments will continue to be extremely im-
portant. Like all investigations concerning cognitive function, this hypothesis
needs to be given time to develop closer ties with basic neurobiology.

Self, subjectivity, and consciousness

Antonio Damasio’s attack on the problem is launched at the systems level
rather than the neuronal level. Flis motivating insight is that the capacity for
consciousness is the outcome of high-level self-representational capacities.^®
For the conceptual backdrop to this idea, recall the discussion in chapter 3
(pp. 70-90) stressing that inner regulation and sensorimotor coordination are
the basic platform for the evolutionary development of cognition. Thus ner-
vous systems have integrative organizations for ranking goals, making behav-
ioral decisions, and evaluating relevant perceptual signals in the context of
specific behavioral plans. We used the notion of an internal model — specifically
the Crush emulator — to conceptualize self-representational capacities that de-
ploy an inner representation of the body in relation to its environment.

What, according to Damasio, is the connection between consciousness and
self-representation? Roughly, the idea is this: Under evolutionary pressure, the
sophistication of the simple integrative internal models increased, consistent
with the organism’s need for staying alive and maintaining niche-suitable co-

165

Consciousness

herence in behavior."^® At some stage, new circuitry enabled a neuronal popu-
lation to represent the internal model itself. It could represent some items in the
model (themselves representations) as standing in relation to representations of
states of the body. That is, the circuitry could represent certain of the organ-
ism’s current perceptual and emotional states as states of itself it could cate-
gorize some representations as being o/ objects external to the body, and, most
important, it could represent the relation between them.“^^

Thus if I step on a cactus, a certain class of neural events is represented as
inner (e.g., my pain), while others are represented as being of the outer world
(e.g., the visual representation of the cactus). The relation between them is also
represented, inasmuch as the brain sees the external thing (the cactus) as the
source of the inner state (the pain), and represents control over its own body to
avoid contact with the cactus as a way of avoiding pain. The properties recog-
nized visually (green, spiny) are seen as properties of the cactus, whereas the
properties of the pain are categorized as belonging to me. For convenience, we
can call such representations of relationships between representations meta-
representations, since they are higher-order representations that are about lower-
order representations. This richer neural architecture enables second-order
evaluative structures and second-order planning and predictive structure.

Why would metalevel representational categorization constitute an change
favored in the competition for survival? To a hrst approximation, because it
permits richer comparison, evaluation, and learning. With the envisioned meta-
representational capacity, I can emulate myself in various conditions and eval-
uate my options. I can envision myself as feeling hungry and hunting a rabbit
on one option, as finding a mate under another option, as portaging my canoe
under another option. Moreover, I can sequence my self-representations in
my plans so as to maximize my goal achievements. Those organisms whose
brains happen to excel in the coherencing and integration business have a better
chance at reproductive success than those whose brains coherence poorly. The
metarepresentational upgrade endows the organism with a greater range of
capacities to manipulate its body image in problem solving, developing impulse
control, making long-term plans, and drawing upon relevant stored knowledge.
In short, it makes the organism smarter.

Why should this metarepresentational integration constitute the basis for the
capacity for consciousness! After all, computers can have metarepresentations
without being conscious. As I understand him, Damasio answers thus. First,
banish the intuition that when you become conscious of something, a pain for
example, a little light in effect shines on the pain, making it conscious. If that

166

Metaphysics

metaphor creeps into your understanding, you are headed for mysticism.
Needless to say, being aware of a pain is neurobiologically different from not
being aware of it, but that difference is neither literally nor metaphorically
having the nonphysical flashlight of the soul shining down upon the pain.
Whatever the difference is, it is likely a difference identifiable at the sys-
tems level and constituted by specific activities of widely distributed neuronal
populations.

Second, metarepresentations per se do not yield consciousness. The meta-
representational capacities serving consciousness must involve self-attribution
(“This pain is mine”), self-representation (having a point of view), self-control
(“I will wait to eat”), and the recognition of the relations between inner and
outer things (“I can eat that” or “That thing can hurt me”).

Third, with the metaintegrative operations referred to, conscious experiences
just turn out to be items in the integrated schema. The hypothesis is therefore
reductionistic in the sense that it identifies consciousness of a pain, for example,
with a representation in the metarepresentational schema. That is, conscious-
ness of pain is just what you get when the representation of the relevant soma-
tosensory signal is metarepresented as standing in the “belongs to” relation to
the self-representation. According to the hypothesis, the identification is just a
biological fact about the brain, just as it is a physical fact that light is electro-
magnetic radiation or a neurobiological fact that an epileptic seizure is syn-
chronous firing of large populations of excitatory neurons.

Whence the qualitative differences between experiences, such as the difference
between the pain of a burn and the sound of a mosquito or the smell of skunk?
On this approach, these differences are the wholly natural consequence of rep-
resenting signals as having different sources (e.g., retina versus olfactory epi-
thelium) or as having different significance for the organism (e.g., safe versus
dangerous) or as having different action-relevant categorizations (e.g., spider
versus grizzly) or the like. Exactly how these qualitative differences depend on
these factors should be sorted out as neuroscience proceeds.

Other questions need to be considered. We have conscious experience of only
some among a range of internal signals. For example, one can be conscious of
bladder distension, but not of one’s blood pressure. What determines which
among the variety of internal signals are signals included in the high-level inte-
gration that enables one to experience them? This has to be answered in the
context of evolutionary biology and of what would and would not be favored
by natural selection. Consequently, the answer will depend on whether the type

167

Consciousness

of State in question is one where it made sense for Mother Nature to permit the
organism behavioral control and options. What, then, about blood pressure?

Maintaining the appropriate blood pressure is a constant priority of the ner-
vous system, never to be put behind eating or sex or satisfying one’s curiosity.
Not surprisingly, therefore, it is automatically regulated by the autonomic
nervous system, not relegated to cognitive functions that may be unreliable
on occasion or unregulated during sleep. Modern medicine aside, there are no
behavioral options for getting blood pressure to its appropriate values, as there
are behavioral strategies for finding a safe place to rest or finding water. These
kinds of considerations explain why representation of blood pressure is not
included in the metalevel integration that gives rise to representations of an
experience as “mine,” and why, therefore, there was no survival advantage in
having wiring to support conscious awareness of changes in blood pressure.
Similar observations can be made about peristaltic movements of the stomach
and intestines.

Among the kinds of states of which we can be conscious, what determines
which ones we are in fact aware of at any given point in time? The answers
here will derive from the wider neurocomputational theory of how the brain
achieves its various integrative feats — how its sets and resets priorities, how
attention can be directed top-down but can be overridden by important
bottom-up signals, and so on. Given the envisioned metalevel representational
and integrative capacity, what the organism currently is aware of is a function
of how its integrative architecture has determined what it should now watch
and listen for, what it should now do, what memories are now relevant to cur-
rent goal-planning functions, what behavioral options are now viable, and so
on. That is, the coordinated operation of neurocognitive functions — attention,
short-term memory, long-term memory, perception, emotion, choice, and
imagination — will result in the organism’s being aware of some states but not
others.

Since nothing like a full-fiedged theory of the nature of integration is on the
table, it is not surprising that the details needed to explain how and when
signals are consciously represented as “my experiences” are out of reach. Is
Damasio’s hypothesis therefore too fragmentary and sketchy to evaluate, let
alone bank on? Or is there evidence suggesting it is on the right track? There is
some evidence, and while a full discussion of that evidence would be too
lengthy, perhaps a first-pass answer and a list of references will constitute a
beginning.

168

Metaphysics

p superior colliculus j

brain-stem nuclei hypothalamus and

basal forebrain

Figure 4.16 The main structures constituting the platform for self-representation,
according to Damasio’s hypothesis. (Courtesy of Hanna Damasio.)

In addressing the testability issue, Damasio highlights evidence indicating the
special importance to consciousness of the following: nuclei in the brainstem
tegmentum, the cingulate cortices and the parietal cortices directly behind
them, the hypothalamus, and the intralaminar nuclei of the thalamus. Small
lesions to the brainstem tegmentum, hypothalamus, posterior cingulate, or the
intralaminar nuclei result in coma or persistent vegetative state (figure 4.16).
Damage to cingulate cortices, especially in the posterior sector and the adjoin-
ing parietal cortices, also compromise consciousness. By contrast, surprisingly
large segments of frontal or sensory or motor cortex can be removed without
loss of consciousness, though, of course, other deficits will appear.

In a very difficult but revealing study using positron emission tomography
(PET), Fiset and colleagues gave normal volunteers the drug propofol, an
anesthetic widely used in medical practice to induce general anesthesia.
Propofol is an experimentally promising drug because there is a precise and
known relationship between concentration and level of sedation. Very tiny

169

Consciousness

changes in concentration correlate with differences between mild sedation,
deep sedation, and unconsciousness, the latter defined as unresponsiveness
to verbal commands. Fiset and colleagues found that the brainstem nuclei,
and consequently the thalamic structures to which they project, were pref-
erentially affected by propofol. Other regions cited by Damasio as part of
the self-representational substructure, including the posterior cingulate cortex,
also showed changes correlated with propofol concentrations sufficient for
unconsciousness.

Brainstem structures are also known to mediate attentional functions, along
with levels of arousal, and to control the shifts in state from being asleep, to
dreaming, to being awake. Additionally, there is a convergence of input to
the brainstem signaling the states of the vestibular system, musculoskeletal
frame, viscera, and internal milieu (see chapter 3, figure 3.3). Drawing on ana-
tomical and physiological data, Damasio notes that specific small regions
(nuclei) in the brainstem contain integrated information about current activity
and recent changes relevant to the organism’s state and its goals. Depending on
this state-of-the-critter report, modulation of cortical activity is caused by other
brainstem nuclei. At the systems level, this to-ing and fro-ing between state-of-
the-critter profiles and cortical modulation means that attention is paid to some
things and not others, that some things are learned, that some relevant things
are actively remembered, and that some choices are favored over others as the
organism moves about its environment making its living.

Obviously, hordes of details remain to be worked out by neuroscience and
psychology even if this hypothesis is more or less in the right neck of the
woods. For starters, we would like to understand exactly how specific brain-
stem structures regulate shifts in attention, exactly how internal models are
organized, updated, interconnected, and modified. As with other hypotheses
herein entertained, Damasio’s hypothesis too must confront potentially falsify-
ing tests.

Progress in neuroscience and the indirect approach

In the last several decades, progress on all aspects of brain function has been
truly impressive. No meaningful summary of this progress is possible in a few
pages. In subsequent sections, however, I shall make specific use of relevant
neuroscientific developments in the course of analyzing and responding to those
who think that neuroscience, in principle, can never, ever, lead us to a deeper

170

Metaphysics

understanding of conscious experience. In later chapters on representation and
knowledge, I shall again draw upon relevant developments.

Impressive progress notwithstanding, it is also true that neuroscience as a
field is still young and still groping for its general, “exoskeletal” explanatory
principles. This probably means we are especially prone to losing the forest for
the trees, particularly when experimental research is vigorous — more prone
than, for example, scientists in molecular biology or cell biology. To a first ap-
proximation, those fields enjoy the benefits of having established the general
principles governing their target phenomena.

So far, the same cannot truly be said for neuroscience. Although there are
indeed fruitful ideas of a highly general sort, and although the rise of compu-
tational modeling has helped enormously in conceptualizing the problem of
how macroeflfects emerge from microphenomena, the fact remains that neuro-
science has not achieved the explanatory maturity of, for example, cell biology.
The reasons for the relative immaturity are not surprising. They include the
truly staggering complexity of the system under study and the monumentally
difficult problems confronting the development of reliable, revealing experi-
mental techniques. They also include conceptual lacunae. As noted earlier, we
do not yet really understand what the notion of information should mean in a
biological or psychological context. Moreover, we do not yet fully understand
how neurons code information, whatever information is. These issues will be
discussed a little more in chapter 7. I mention them now to balance my genuine
optimism for future discoveries regarding consciousness on both the direct and
the indirect routes. Like many neuroscientists, I view the pioneering aspects of
neuroscience as part of what makes the field so very exciting. Virtually nothing
is humdrum, so much is open territory, and surprises are an almost daily affair.

1.5 Concluding Remarks

The main aim of section 1 has been to see what happens if we consider con-
sciousness as a natural phenomenon that can be investigated scientifically as
well as introspectively. We saw that there are different strategies, driven by
different hunches and directed by differences in scientific feel for the problem.
Progress is evident in many investigations, though techniques for safely inves-
tigating the brains of humans at the micronetwork level remain to be devel-
oped. In the next section, I canvas a range of reservations about any and every
neurobiological approach to understanding consciousness.

171

Consciousness

2 Dualism and the Arguments against Neuroscientific Progress
2.1 Life and Conscious Experience

At this stage of our knowledge, none of the functions — attention, short-term
memory, being awake, perceiving, imagining — can plausibly be equated with
consciousness, but we are learning more about consciousness, bit by little bit, as
scientific progress is made on each of the topics. In this respect, the virtues of
the indirect approach to consciousness may be analogous to the virtues of the
indirect approach to the problem of what it is to be alive. Just as identifying
a micro-organizational correlate to being alive was not the winning strategy
for the problem of life, so perhaps, by analogy, trying to identify a micro-
organization correlate of consciousness may not be the winning strategy for the
problem of awareness. But is the analogy between the problem of being alive
and the problem of consciousness a useful analogy? Let’s consider how it might
be useful.

What is it for something to be alive? The fundamental answer is now avail-
able in college biology courses. Modern cell biology, molecular biology, physi-
ology, and evolutionary biology have discovered so much that a comprehensive,
if not complete, story can now be told. To be alive, cells need a cytoplasm con-
taining structures such as mitochondria, to produce energy. They need the
means of replication, such as DNA, along with microtubules to orchestrate cell
division. They need protein-manufacturing apparatus, and so need ribosomes,
enzymes, mRNA, tRNA, and DNA. They need specialized membranes, such
as bilipid layers with specialized protein channels to admit certain molecules
into the cell under specific conditions and to keep others out under certain
conditions. They need endoplasmic reticulum for metabolic processes, lyso-
somes for digestion, and Golgi apparatus for sorting, finishing, and shipping
cell products. The biochemistry segment of the course would talk about water,
carbon compounds, amino acids, and proteins. The physiology segment of the
course would discuss how tissues like muscle, and organs like kidneys, function.
At the end of the course, one would have, at least in outline, the scientific
account of what it is for something to be alive.

A biology professor winding down the course at the end of the year, might
hear this complaint; “I now understand all that, but you still have not ex-
plained to us what life itself is.” The reply is, roughly, that life is all that.
You understand what is it for something to be alive when you understand the

172

Metaphysics

physical processes of metabolism, replication, protein building, and so forth.
Once you know all that, there is no other phenomenon — livingness itself — to be
explained. Certainly, there are many questions still unanswered concerning how
cells work, but these are questions such as “How does a transmembrane pro-
tein get inserted?” not questions such as “How does the life force get into the
cell?”

Unconvinced, someone might persist, noting that textbook explanations
really involve the interactions of dead stuff — ribosomes, microtubules, etc. —
but what he wants to know is what living (being alive) itself is, what the essence
of life is. Surely, it may be contested, being alive cannot emerge from mere
dead stuff, no matter how it is arranged and organized.

The assumption behind this persistent question was a seriously debated hy-
pothesis in the not very distant past. By 1920, however, the assumption was
already seriously behind the scientific times. The assumption, known as vital-
ism, is that things are alive because they are infused with the “life force” or
“vital spirit” or “urge.” Vitalists are convinced that being alive cannot be a
function of the dynamics and organization of dead molecules. Even as late as
1955, a few scientists still clung to the conviction that a nonphysical “urge”
transforms a cell from a dead organization to a living organization.

Nevertheless, what modern biology has discovered is there is no vital spirit
over and above a complex — really complex — organization of physical prop-
erties. The urge intuition takes a beating when the details of metabolism, pro-
tein production, membrane functions, and replication are understood. When
you see how it all comes together, you see that no vital spirit is needed in the
explanation. This is an example where the nonexistence of something is estab-
lished as highly probable, not through a single experiment demonstrating its non-
existence, but through aeceptance of an explanatorily powerful framework that
has no place for it. The same thing happened to “impetus” as Newton’s physics
became accepted and, as we saw in chapter 2, to “caloric fluid.” This is not to
say that the nonexistence of caloric fluid or vital spirit has been absolutely
proved, but because these concepts play no explanatory role whatever in
science, they are deemed to be outdated theoretical curiosities.

Those who pursue the scientific approach to consciousness believe that
developments analogous to those in the biology of “life” will allow us to un-
derstand consciousness. That is, we are beginning to understand the neuro-
biology of sleep, dreaming, attention, perception, emotions, drives, moods,
autobiographical memory, perceptual imagery, motor control, motor imagery,
and self-representation. We are beginning to understand the neurobiology of

173

Consciousness

what happens under various anesthetics, in a coma, in subthreshold perception,
and in hallucinatory states. With more complete explanations of all, the nature
of conscious phenomena should be understood, at least in a general way. Lots
of detailed questions will remain, of course, but science is like that. What the
research program envisions is that this understanding is an empirical possibil-
ity, not an empirical certainty.

If, having understood all those functions, someone were to persist, “But what
about consciousness itself Consciousness cannot come out of nonconscious
physical stuff, no matter what its dynamics and organization,” we shall have to
respond more or less as we do now with the vitalists. We go back through the
relevant science all over again. If the objection under consideration assumes
that consciousness cannot be a brain function because consciousness is a soulish
thing, science may be up against dogmatism, as it was with vitalism circa 1950.

Dualism, as we know from discussion in chapters 2 and 3, is not likely to be
falsified by a single experiment or two showing the nonexistence of the soul.
Rather, dualism is rendered improbable because the explanatory framework
of psychology and neuroscience, though incomplete, and embedded within
the larger framework of physics, chemistry, and evolutionary biology, is much
more powerful than any dualist competitor. This could change, but so far the
empirical evidence does not point that way.'^^ As things stands, the concept of a
nonphysical soul looks increasingly like an outdated theoretical curiosity.

Even granting that dualism is essentially moribund, a number of philoso-
phers and scientists wish to argue that consciousness cannot ever be understood
in terms of brain function. Even if dualism is false, they claim, neurobiological
research on consciousness is a waste of time, and neurophilosophy is a snare
and delusion. Although a host of such arguments exist, I shall analyze only
those generally regarded as the strongest, the most widely held, or the most
appealing.

2.2 Nine Naysaying Arguments'^®

A common argument consists in stressing what we do not know, and using this
as a premise for concluding what we cannot know. Cohn McGinn, for exam-
ple, says that the problem of how the brain could generate consciousness is
“miraculous, eerie, even faintly comic.”'^^ Finding the problem difficult, he
concludes, “This is the kind of causal nexus we are precluded from ever
understanding, given the way we have to form our concepts and develop our
theories.” He thinks that for us to understand the nature of consciousness is

174

Metaphysics

like a mouse understanding calculus. McGinn is by no means alone here.
A number of contemporary thinkers believe they can already tell that the
question is unanswerable — not just now, not just given what we know so far,
but unanswerable ever. Zeno Vendler chides the ambitions of neuroscience by
saying that it is obvious from the nature of sensation, that our sensing selves
“are in principle beyond what science can explain.”"^® That we are trying to
unravel the mystery is, in Vendler’s view, a consequence of the overweening
assumption that there are no questions science cannot answer. How can we
respond to McGinn, Vendler, and other naysayers?

In each of the following subsections, I shall briefly entertain one naysaying
objection and try to assess its cogency."^®

I cannot imagine how science eould explain awareness!

This is one of the most popular naysaying arguments, advanced frequently by
philosophers and sometimes by scientists. What can be said in response?

In general, what substantive conclusions can be drawn when science has not
advanced very far on a problem? Not much. One of the basic skills philoso-
phers teach in logic is how to recognize and diagnose the range of nonformal
fallacies that lurk under ostensibly appealing arguments: what it is to beg the
question, what a non sequitur is, and so on. A prominent item in the fallacy
roster is argumentum ad ignorantiam — argument from ignorance. The canoni-
cal version of this fallacy uses ignorance as the key premise from which a sub-
stantive conclusion is drawn. The canonical version looks like this:

We really do not understand much about a phenomenon p. (Science is largely
ignorant about the nature of p.)

Therefore, we do know that

• p can never be explained, or

• nothing science could ever discover would deepen our understanding of p, or

• p can never be explained in terms of properties of kind

In its canonical version, the argument is obviously a fallacy: none of the
proffered conclusions follow, not even a little. Surrounded with rhetorical
flourishes, brow furrowing, and hand wringing, however, versions of this argu-
ment can hornswoggle the unwary.

From the fact that we do not know something, nothing very interesting
follows — we just don’t know. Nevertheless, the temptation to suspect that our

175

Consciousness

ignorance is telling us something positive, something deep, something meta-
physical or even radical, is ever-present. Perhaps we like to put our ignorance
in a positive light, supposing that but for the awesome complexity of the phe-
nomenon, we (smart as we are) would have knowledge. But there can be many
reasons for not knowing, and the specialness of the phenomenon is, quite reg-
ularly, not the most significant reason. I am currently ignorant of what caused
an unusual rapping noise in the woods last night. Can I conclude it must be
something special, something unimaginable, something alien, other-worldlyl
Evidently not. For all I can tell now, it might merely have been a raccoon
gnawing on the compost bin. Lack of evidence for something is just that: lack
of evidence. It is not positive evidence for something else, let alone something
of a spooky sort. That conclusion is not very thrilling, perhaps, but when
ignorance is a premise, that is about all you can grind out of it.

Moreover, the mysteriousness of a problematic phenomenon is not a fact
about the phenomenon. It is merely an epistemological fact about us. It is a fact
about where we are in current science. It is a fact is about what we currently do
and do not understand, about what, using the rest of our understanding, we
can and cannot imagine. It is not a property of the problem itself.

It is sometimes assumed that there can be a valid transition from “We
cannot now explain” to “We can never explain” if we have the help of a sub-
sidiary premise, namely, “I cannot imagine how we could ever explain.” But
the subsidiary premise does not help, and this transition remains a straight-up
application of argument from ignorance. Adding, “I cannot imagine explaining
p” merely adds a psychological fact about the speaker, from which, again,
nothing significant follows about the nature of the phenomenon in question.

Vitalists, we noted earlier, argued that life could be explained only by
invoking a nonphysical kind of thing, a vital spirit; living things have it, dead
things do not. A favored argument for vitalism ran as follows: I cannot imagine
how you could get living things out of dead molecules. Out of bits of proteins,
fats, sugars how could life itself emerge? It seemed obvious from the sheer
mysteriousness of life that the problem could have no solution in biology or
chemistry. We know now, of course, that this was all a shortsighted mistake.

Neuroscience is very much in its infancy. So if someone or other cannot
imagine a certain kind of explanation of some brain phenomenon, it is not
terribly significant. Aristotle could not imagine how a complex organism could
come from a fertilized egg. Given early science (300 b.c.), it is no surprise that
he could not imagine what it took many scientists hundreds of years to discover.
I cannot imagine how ravens can solve a multistep problem in one trial, or how

176

Metaphysics

an organism integrates visual signals across time, or how the brain manages
thermoregulation. But this is a (not very interesting) psychological fact about
me. One could, of course, use various rhetorical devices to make it seem like an
interesting fact about oneself, perhaps by emphasizing that it is a really, really
hard problem, but if we are going to be sensible about this, it is clear that one’s
inability to imagine how thermoregulation works is, at bottom, pretty boring.

The “I cannot imagine” gambit suffers in another way. Being able to imag-
ine an explanation for p is a highly open-ended and under-specified business.
Given the poverty of delimiting conditions of the operation, you can pretty
much rig the conclusion to go whichever way your heart desires. Logically,
however, that flexibility is the kiss of death.

Suppose that someone claims that he ean imagine the mechanisms for sen-
sorimotor integration in the human brain but cannot imagine the mechanisms
for consciousness. What exactly does this difference in imaginability amount
to? Can he imagine the former in detail? No, because the details are not known.
What, precisely, can he imagine? Suppose he answers that in a very general way
he imagines that sensory neurons interact with interneurons that interact with
motor neurons, and via these interactions, sensorimotor integration is achieved.
Now if that is all it takes to be able to imagine, one might as well say that one
can imagine the mechanisms underlying consciousness. Thus, “the interneurons
do it.” The point is this: if you want to contrast being able to imagine brain
mechanisms for attention, short-term memory, planning, etc., with being
unable to imagine mechanisms for consciousness, you have to do more that say
that you can imagine neurons doing one but cannot imagine neurons doing the
other. Otherwise, you simply beg the question.

There could he zombies

This time the attack on neurobiological strategies derives from a so-called
“thought experiment,” which roughly goes as follows. (1) We can imagine a
person, like us in all the aforementioned capacities (attention, short-term mem-
ory, verbal capacity, etc.), but lacking the experience of pain and the experi-
ence of seeing blue. That is, he would lack qualia (pronounce kwa-lee-a), i.e.,
the qualitative aspect of conscious experience, such as feeling pain or feeling
dizzy, seeing colors, hearing a C-minor chord. This person would be exactly
like us, save that he would be a zombie. He would even say things that we do,
such as “I have a funny feeling in my tummy” as the airplane suddenly
descends and, on a line summer afternoon, “The sky is very blue today.” The

177

Consciousness

next premise of the argument says this: (2) If the scenario is conceivable, it is
logically possible. The conclusion says, (3) Since a zombie is logically possible,
then whatever consciousness is, it is explanatorily independent of brain activ-
ities. That is, even a complete explanation of every aspect of the human brain
will not explain consciousness. This is because a true explanation must fore-
close the logical possibility of there being a zombie. (Something akin to this
was argued by Saul Kripke in the 1970s, by Joseph Levine in the 1980s, and
again by David Chalmers in the 1990s.)

To most of us, this argument is puzzling, because many things are logically
possible but not empirically possible, such as a 2-ton mouse or a spider that can
play the flute. Why should we suppose that the logical possibility of a zombie
tells us anything interesting about what research could be successful? After all,
what neurophilosophy is really interested in is the actual empirical world and
how it works. The reply depends on the pivotal claim about the standards
for an explanation, namely, that a proper explanation must foreclose logical
possibilities.

Assuming that this is the pivotal claim here, we need to recognize how
absurdly strong a claim it is. Not only does it rule out explaining consciousness
in terms of brain function, but it also rules out explaining consciousness in
terms of soul function or spooky-stuff function or quantum gravity or anything
else you might think of. So strong is the demand it places on successful expla-
nation that no scientific explanation of any phenomenon has ever met it, or
ever could meet it.

As we saw from the discussion in chapter 1, section 3, explanatory reductions
require that a new theory successfully reconstruct most of the features of the
reduced phenomenon, as antecedently understood. But this falls far short of
any logical entailments from the former to the latter such that previously con-
ceived possibilities are now logically impossible. Good explanations rule out
empirical possibilities, not logical possibilities. Historically speaking, no scien-
tific reduction/integration has ever met such an absurdly strong requirement.

A further problem with all such “conceivability” arguments is that they want
to draw an interesting conclusion about the nature of how things really are.
Nothing interesting follows, however, from the fact that some particular human
is, or is not, able to imagine something. That something seems possible does
not thereby guarantee it is a genuine possibility in any interesting sense, so why
should we think that the zombie idea is genuinely possible? To insist on its
possibility on grounds that the premises are grammatical is to confuse a real
possibility with mere grammaticality.

178

Metaphysics

For the sake of argument, I have played along with the underlying assump-
tion that we understand quite well the scope and limits of the domain of the
logically possible. Nevertheless, this assumption is deeply flawed. Quine dem-
onstrated in 1960 that such an assumption is actually just a bit of philosophical
self-deception. A few hand-picked examples of what is and is not logically
possible seem straightforward enough, but outside of these, all is fantasy, or
group-think, or depends on self-serving definition. Not surprisingly, the espe-
cially controversial cases are those where philosophers want logical possibility
to give them some real metaphysical leverage. And the argument at hand is
very much a case in point. Standing back a bit, one does find something
unconvincing in the idea that the conveniently elastic and philosophically con-
cocted notion of logical possibility should dictate to neurobiology what it can
and cannot discover — ever.

To see from a different perspective why the argument gets messed up, run an
analogous zombie argument with respect to life. It says that we can imagine a
planet where “deadbies” are things composed of cells with membranes, nuclei
with DNA, the usual organelles, and so forth. Deadbies reproduce, digest, re-
spire, metabolize, manufacture proteins, grow, and so forth, just as organisms
on Earth do. Unlike us, however, deadbies are not really alive. This is a logical
possibility. So life is explanatorily independent of biology.

Here too the premises are possible in the very weak sense that they are
grammatical, but so far as we know, they do not state a real possibility. Here
is another feeble thought experiment: imagine a planet where the velocity of
molecules in a gas increases, but lo and behold, its temperature does not. Does
this tell us that temperature is explanatorily independent of mean molecular
kinetic energy? Certainly not. What does this tell us about the actual relation
between mean molecule kinetic energy and temperature in a gas? Not a single
thing.

I take the zombie argument to be a demonstration of the feebleness of the
class of thought experiments that are factually isolated from the relevant
science but nonetheless hope to draw a scientifically relevant conclusion.^®

The problem is too hard

This objection is also very common, and is often advanced along with sundry
other objections, both those discussed above and some from those given below.
How valuable is it?

179

Consciousness

Can we tell how hard a problem is when we do not have a whole lot of
science on the subject? To till out the point, consider several lessons from the
history of science. Before the turn of the twentieth century, people regarded as
trivial the problem of explaining the precession of the perihelion of Mercury,
that is, the fact that the elliptical orbit of Mercury constantly but slowly
advances in the plane of its orbit. This movement was an annoying deviation
from what Newton’s Laws predict, but the problem was expected ultimately to
sort itself out as more data came in. Essentially, it looked like an easy problem.

With the advantage of hindsight, we can see that the assessment was quite
wrong: it took the Einsteinian revolution in physics to solve the problem of
the precession of the perihelion of Mercury. By contrast, the composition of the
stars was thought to be a really hard problem. How could a sample ever be
obtained? As soon as you try to get close enough to take a sample, you burn.
But with the advent of spectral analysis, that turned out to be a readily solvable
problem. When heated to incandescence, the elements turn out to have a kind
of fingerprint, easily seen when light emitted from a source is passed through a
prism.

Consider now a biological example. Before 1953, many people believed, on
rather good grounds actually, that to address the copying problem (transmis-
sion of traits from parents to olfspring), we would first have to solve the prob-
lem of how proteins fold, i.e., how a string of amino acids bends and twists so
that it ends up having a highly specific shape unique to that protein. The
copying problem was deemed a much harder problem than the problem of how
a string of amino acids takes on the correct shape, and many scientists believed
it was foolhardy to attack the copying problem directly. This was partly be-
cause it was generally believed that it would take something as complex as a
protein to be the carrier of hereditary information. DNA, a mere acid, was
considered too simple to qualify as a candidate.

As we all know now, the key to the copying problem lay in the base-pairing
of DNA, and the copying problem was solved first. Humbling it is to realize
that the problem of protein folding (secondary and tertiary folding) is still not
solved.

What is the point of these stories? They illustrate the fallacy in arguments
from ignorance. From the vantage point of ignorance, it is often very difiicult
to tell which problem will turn out to be more tractable than some other, and
whether we have even conceptualized the problem in the best way. Conse-
quently, our judgments about relative difficulty or ultimate tractability should

180

Metaphysics

be appropriately qualified and tentative. Guesswork has a useful place, of
course, but it is best to distinguish between blind guesswork and educated
guesswork, and between guesswork and confirmed fact. The philosophical
lesson is this: when not much is known about a topic, don’t take terribly seri-
ously someone else’s heartfelt conviction about what problems are scientifically
tractable. Learn the science, do the science, and see what happens.

How can I know what you experience?^ ^

This worry takes several closely related forms, the oldest and most familiar of
which is the so-called “inverted spectrum problem.” The general worry is that
the facts of anyone’s phenomenal experience are always underdetermined by
any and all physical facts, including all neurophysiological facts, that we might
come to know about that person. (By “p is underdetermined by q,” philoso-
phers mean that p cannot be strictly deduced from q-, q may provide evidence
for p, but not absolutely conclusive evidence.) Accordingly, the argument con-
cludes, phenomenal facts must be distinct and independent facts in their own
right, a class of facts that can never be explained in purely physical terms. This
general argument finds specific expression in the following thought experiment.

Consider the possibility that you and I share the same range of visual color
experiences, but in all those cases where I have the subjective experience of red
(as when I look at a ripe tomato in broad daylight), you have the subjective
experience of green, the experience that I get when I look at the lawn. Suppose,
moreover, that these divergent color experiences are systematic: when I look at
the rainbowlike spectrum projected by a prism, I see red on the left-hand side,
fading progressively into orange, yellow, green, and blue as I look to the right,
but in that same objective situation, you see blue on the left-hand side, fading
progressively to green, yellow, orange, and red as you look to the right. In
short, your internal spectrum of color experiences is mapped onto the external
world in a fashion that is exactly the inverse of my own. But this internal dif-
ference is hidden by the fact that we apply our shared color terms to external
objects in all of the same ways.

This, let us suppose, is entirely conceivable. But, continues the antiphysicalist
argument, this possible inversion of our respective color qualia stubbornly re-
mains perfectly conceivable no matter how much we might know about each
others’ brains, and no matter how similar we might be in our physical behavior,
our physical constitution, and our internal neural activities. Our brains could

181

Consciousness

be identical, and yet our conscious experiences could still diverge. The physical
facts, apparently, do not “logically fix” the phenomenal facts, and so the phe-
nomenal facts must be some kind of facts above and beyond the merely physi-
cal facts. Therefore, concludes the argument, we must look beyond the physical
sciences for any explanation of phenomenal experiences. They evidently con-
stitute a realm of nonphysical facts.

Is this argument compelling? To answer that, we must closely examine
its logic. First, this argument too relies on what is and is not alleged to be
imaginable/conceivable in order to generate support for its conclusion. The key
premise asserts, “Our brains could be identical in every respect, but our qualia
could differ.” Not surprisingly, the “could” is the “could” of conceivability,
not the “could” of “actually could.” As noted in analysis of the zombie argu-
ment, that something is logically possible implies absolutely nothing about
empirical or real possibility.^^

If the key premise collapses, the argument collapses. Is perhaps the premise
that our brains could be identical in every respect but our qualia could differ
just obviously true, and hence not in need of any defense? Not at all. Given the
weight of available empirical evidence showing that dilferences in conscious
experience do in fact involve dilferences in brain activity, the premise cannot be
sold as obviously true. For example, we know that if you decrease the activity
in the neurons projecting from a decayed tooth to the brainstem, the pain dis-
appears. If nothing is done, the pain persists. Direct stimulation of the hand
area of the somatosensory cortex during surgery produces changes in sensations
in the hand. We entirely lack any examples where we know the brain remains
exactly the same but the conscious experience changes. If there is a causal rela-
tionship between neuronal activity and conscious experience, as there certainly
seems to be, then the falsity of the key premise is exactly what one would predict.

Can the premise be defended by claiming that in the actual world there are
known examples where brains are identical in every respect and our qualia do
differ? That strategy would indeed begin to add real substance to the argument.
Yet it is never adopted, for the simple reason that there are no examples, there
is no factual evidence to bring to bear.

A distinct line of defense of the key premise asserts that the premise is true
because qualitative experiences are nonphysical properties. Consequently, it
is alleged, our brains could be identical in every respect but our qualia could
differ. The weakness in this defense is that it invokes dualism, which, on inde-
pendent grounds, appears to be highly improbable. Nevertheless, in a spirit

182

Metaphysics

Table 4.1 Key premise: our brains could be absolutely identical but our qualia could
dilfer

Brief defense of the key premise

Brief criticism of the defense

It is conceivable.

So what?

It is empirically well supported.

Show us the data.

Dualism is true.

Dualism is improbable.

The conclusion is true, so the premise must be
true.

Circular arguments are worthless.

By definition, qualia are independent of brain
states.

Circular arguments are worthless.

of thoroughness, we shall explore this possibility in much greater detail below
(pp. 182-192).

A last ditch effort consists in defending the key premise on grounds that
conscious experiences are not identifiable with any property of the nervous
system. This move is ineffective because this very claim is what the argument is
supposed to show, not what the argument gets to assume. If you defend the key
premise by appeal to the very conclusion your argument is supposed to estab-
lish, the argument is utterly worthless — it simply runs in a circle. The illusion of
progress can be conjured, however, especially if the defense of the key premise
is left implicit and hence hidden from inspection. Incidentally, a variant on the
circular argument consists in defining qualia as psychological states that are not
identifiable with any pattern of neuronal activity. This is no better than simply
arguing by restating the conclusion as a premise.

In sum, here is the logical fix the argument finds itself in. It cannot just help
itself to the key premise “Our brains could be identical in every respect but our
qualia could differ” on grounds that is it obviously true. The defense of the key
premise can, jointly or severally, take five forms (see table 4.1). Succinct criti-
cism of these defenses are given in table 4. 1 .

The inverted-spectrum argument gets into trouble not because it envisages
perceptual differences between subjects that are difficult to detect. The argument
gets into trouble because it wants to crank out a very strong conclusion about
the nature of things from essentially no facts; i.e., it wants to establish an a pri-
ori truth. It needs to persuade us that qualitative differences in experience are
undetectable', not just undetectable given only behavioral data, but undetectable
no matter what facts — behavioral, anatomical, physiological — are available.

183

Consciousness

Given the utter poverty of its cohort of defenses, dualism actually emerges as
the strongest argument against a neurophysiological explanation of conscious-
ness. At least the dualist has the option of launching an empirical argument
for dualism. The empirically sensitive dualist will want to argue that if the
facts prevent us from discovering whether one subject’s color experiences are
inverted with respect to those of another, then these facts constitute evidence in
favor of dualism. In the next section, therefore, we shall take a closer look at
the empirical possibility that one person’s color experience might be systemati-
cally different from that of another, and whether we could indeed discover that
this was so from the psychology and neuroscience of the visual system.

What happens if we get more empirieal?

First, the argument, as stated above, betrays a much-too-simple conception of
our actual color experiences.^^ The monochromatic linear spectrum produced
by a prism presents only a small percentage of the visual color qualia enjoyed
by normal human perceivers. That spectrum is missing brown, for example,
and pink, and chartreuse, and sky blue, and jade green, and black and white
too, for that matter. Indeed, it has been known for some time that the space of
human color qualia is not one-dimensional, or even two-dimensional, but is
fully three-dimensional. The Munsell color solid (plates 3 and 4) displays the
structure of that fairly complex space. Notice that every color discriminable by
humans occupies a unique place within that space, a place fixed by the unique
family of similarity and dissimilarity relations that it bears to all of the other
colors that surround it, both near and far. Two distinct colors could not ex-
change their respective positions without thereby fouling up many of the simi-
larity relations that structure the original space.

Notice also that the overall shape of this space is nonuniform. For whatever
reasons, tying equal distances in this space to equal increments of color discrimi-
nahility yields the decidedly nonspherical phenomenal space of plates 3 and 4.
Most notably, yellow bulges out from the central axis and up towards the white
pole, and at lower brightness levels, it fades to being indiscriminable from dark
gray more swiftly than any other color. (In fact, this is roughly how Munsell
pieced together his original model of our color space in the first place — by
asking people to judge relative similarities and just barely discriminable color
differences over a large sample of colored chips.) Additional experiments have
shown that we can make finer discriminations among diverse external stimuli
within the greenish, yellowish, and orangeish regions of our color space than
we can in the blueish regions.

184

Metaphysics

If we are going to perform a color-inversion thought experiment, then we
need to imagine something a little more complex than the one-dimensional
spectrum flip usually suggested. Specifically, we need to imagine that the sub-
ject of the proposed inversion has his phenomenal color solid either rotated
(180°, say) or mirror-inverted relative to its normal family of causal connec-
tions to external stimuli — thus, the yellow part of his color space gets activated
when he looks at the sky, the blue part of his color space gets activated when he
looks at bananas, and so forth — while all of the internal similarity relations of
his three-dimensional color space remain exactly the same (just as in the original
thought experiment). Is there anything impossible or inconceivable about this
scenario?

Not a thing. Such an inversion is perfectly conceivable. But we should notice
that it would be behaviorally (that is, physically) detectable in short order. In
comparison with normal humans, the inversion subject would be able to make
more and finer discriminations than we can among external objects that we
describe as various shades of blue, and he would suffer a relative discriminatory
deficit among objects that we describe as various shades of red, yellow, orange,
and green. (There are probably good evolutionary reasons why normal subjects
make finer discriminations among the greens, for example, than among the
blues. Birds, with a fourth cone type sensitive to wavelengths in the ultraviolet
range, will make finer discriminations among the blues than we do.) Moreover,
the inversion subject would locate the familiar color boundaries in different
places. Some external objects that have different shades of the same color,
according to us, will fall into entirely different color classes, according to him.

This follows from the refined and more accurate assumptions of our updated
thought experiment, as specified at the top of this page. The inversion subject’s
objective perceptual capacities would be systematically different as well, in
ways clearly predicted by the thought experiment, once it is properly per-
formed. Evidently, it is not true, even on present scientific knowledge, let alone
on all possible future knowledge, that our color qualia could be differently con-
nected to the external world without any physical or behavioral divergences to
herald that phenomenal inversion. The idea that our color qualia could be
inverted, completely independently of any objectively detectable effects, had a
superficial plausibility only because of our ignorance of the nonuniform struc-
ture of human phenomenal color space and the diversity in our discriminatory
capacities, across the colors, that experiments have revealed. Repair that igno-
rance, as we have done, and the dualist must take his thought experiment back
to the drawing board. That, you may be sure, the creative dualist will do.

185

Consciousness

400 450 500 550 600 650

Wavelength (nanometers)

Figure 4.17 The neural-response curves for the three types of cones. Cones for short
(S), medium (M), and long (L) wavelengths provide overlapping but differential
responses to light of different wavelengths. These curves are defined by the absorption
spectra of the three pigments found in normal cones.

Connecting qualia and neuronal organization

But let us put the dualist aside for a moment, and ask the independently inter-
esting questions, Why does human phenomenal color space have three dimen-
sions? Why those three dimensions in particular? And why does it have the
nonuniform shape that it does? What gives rise to the curious phenomenal
arrangement of plate 3 in the first place?

Apparently, the shape of phenomenal color space arises quite naturally and
inevitably from the physical organization and response profiles of the various
neurons in the brain’s visual pathways. The fundamentals of the story, accord-
ing to vision researchers, are surprisingly simple and elegant. The story begins
with the three types of light-sensitive cone cells scattered across the human ret-
ina. Unlike the rod cells, with which they are mixed, each cone type is prefer-
entially sensitive to its own narrow band of wavelengths, as illustrated in figure
4.17 and plate 2. This allows the retina to do a crude spectral analysis of the
mixture of various wavelengths entering the eye.

But this is only the first stage of color vision. The crucial stage is the next
one. The cone cells in the retina make a set of excitatory and inhibitory syn-
aptic connections, via the optic nerve, to a subsequent population of neurons in
the lateral geniculate nucleus (LGN), as illustrated schematically in figure 4.18.

186

Metaphysics

Cones in the retina

Color opponent cells

Figure 4.18 A simplified diagram of neural circuits in opponent-process theory. Oppo-
nent responses are derived from the outputs of three classes of cones. Excitatory con-
nections (+) are shown by solid lines, and inhibitory connections (— ) are shown by
broken lines.

That LGN population is also divided into three distinct kinds of cells, but their
response properties are quite different from the cones that project to them. As
you can see, the middle cell — labeled “green vs. red” — is the site of a constant
tug-of-war between the excitatory signals received from the M-cones (roughly,
the green part of the spectrum) and the inhibitory signals received from the L-
cones (roughly, the red part of the spectrum). Its resulting activity is thus an
ongoing measure of the relative balance or ratio of medium wavelengths over
long wavelengths currently hitting the relevant part of the retina. (Notice, by
the way, that the M- and L-cone curves in figure 4.17 and plate 2 overlap to a
substantial degree. This means that our green-versus-red tug-of-war cell will be
hypersensitive to small shifts, up or down, in the wavelength of monochromatic
light in the spectral regions immediately to the right and left of the crossover
point of the two curves.)

Similarly, the left-most cell — labeled “blue vs. yellow” — is the site of a tug-
of-war between the excitatory signals from the S-cones (roughly, the blue part

187

Consciousness

of the spectrum) and the inhibitory signals from both of the L- and M-cones
(very roughly, the yellow part of the spectrum). Its activity reflects the balance
of wavelengths from the shorter end of the spectrum over wavelengths from the
medium and longer end. (Notice in figure 4.17 and plate 2, however, that in
this case there is almost no crossover of the relevant curves. So our system will
display no hypersensitive discriminations within this area of the spectrum.)

Finally, the right-most cell — labeled “white vs. black” — is the site of a tug-
of-war between excitatory signals from all three types of retinal cones (the
L, M, and, to a lesser degree, the S cells) versus inhibitory signals averaged
over the stimulus-levels reaching the retinal surface as a whole. The activity of
that cell is thus a measure of how much brighter or darker is its local portion of
the retina, compared to the brightness levels hitting the retina as a whole.

These three types of LGN cells — called “color opponent cells” for reasons
that are obvious from the diagram — constitute a most interesting arena for the
coding of information about the character of the light hitting any spot on the
retina. The simultaneous activity levels of all three cell types constitute a three-
dimensional comparative analysis of the peculiar wavelength structure of the
light hitting any part of the retina to which they are connected. Not to waste
words, this analysis constitutes the brain’s initial representation of external
color. In fact, we can graphically represent any particular neural representation
of this sort as a single point in a three-dimensional space, a space whose three
axes correspond to the possible activity levels of each of the three types of
color-opponent cells, as in plate 5.

When we do that, something quite arresting emerges. The range of possible
coding triplets — that is, the range of simultaneous activation-level patterns
possible across the three kinds of color-opponent cells — does not include the
entire volume of the three-dimensional cube portrayed in plate 5. The available
range is constrained to an irregular central subvolume of that cube, as illus-
trated. This is because the three cell types have activation levels that are not
entirely independent of each other, as can be seen from the wiring diagram of
figure 4.18. The several corners of the coding cube are thus “off limits” to the
trio of color-opponent cells.

More specifically, when one calculates what the actual shape of that interior
volume will have to be (from the details of figure 4.17 and plate 5, from the
relative numbers and influence of the three cone types, and from the specific
overlap of their wavelength response profiles), that interior volume turns out to
have the same shape, and to have its various parts associated with the same

188

Metaphysics

colors, as the original Munsell solid of plate 3. Most obviously, the yellow
portion bulges out and up towards the white pole, and in and down towards
the black pole. Moreover, equal increments of discriminability within the
green, yellow, and orange regions of that neuronal activation space correspond
to finer increments of external wavelength than do equal increments within the
blue region, just as we found in our own discriminatory capacities.

To a first approximation, and at a rather abstract level, the mapping means
we are looking at the neuronal basis for our phenomenal color space. The gen-
eral characteristics of the neuronal basis for the existence, dimensionality,
global shape, and chromatic orientation of our internal space for color qualia,
as experimentally mapped out by Munsell and subsequent generations of
visual psychologists, are discernible. In this very general sense, it is mildly
tempting to hypothesize that we can discern the basic principles governing
color qualia.

It is tempting to say this because, as you have just seen, the various possible
coding triplets stand to each other in all of the same similarity/proximity rela-
tions that our color qualia stand to each other, and they stand in all of the same
causal relations to stimulus objects in the external world, and they stand in all
of the same causal relations to subsequent internal cognitive activities, such
as believing or saying that the lawn is green. Now, in general, in science, if
explanatory power is greatly enhanced by making a cross-level identification,
such as between light and electromagnetic radiation, or between temperature
and mean molecular kinetic energy, then the identification looks like a reason-
able bet.

In the case at hand, if we hypothesize that phenomenal color qualia are
identical with coding triplets across our opponent cells, then the systematic
parallels in their causal and relational properties are explicable rather than co-
incidental. The point is, the causal and relational properties displayed by qualia
and by coding triplets will be systematically the same if the qualia and the
coding triplets are themselves one and the same thing, in the same way that
temperature and mean molecular kinetic energy, light and electromagnetic
waves, and water and H 2 O are one and the same thing.

Is it possible that during inattention to color or even under anesthesia, per-
haps, the coding triplets might be active, but no color qualities are experienced?
Well, we do not know, but this is something we could find out. Additionally, it
is safe to assume that there are many other events that must be taking place
elsewhere, for example in the brainstem. To be a little more accurate, therefore.

189

Consciousness

we should restate this very provisional hypothesis to say that the coding triplets
are one component of a set of components that are jointly sujficient for color
experience, and that there are a host of background conditions, many of which
remain to be discovered.

So, yes, the hypothesis is undoubtedly too simple to be correct. Nevertheless,
my point is to emphasize the significance of the jit between the antecedently
determined qualia profile and the neuronal-coding triplet profile. I should
mention too that a range of other color perceptual phenomena, such as various
color illusions, afterimages, and the various forms of color blindness, are also
plausibly explained within this framework. This expanding range of explan-
atory success lends credence to the reductive promise of the general approach;
i.e., we have here the same sorts of evidential grounds and explanatory oppor-
tunities that standardly motivate reductive claims throughout science. And to
that degree, we can get a grip on why materialism seems more plausible than
dualism. For example, the task of the dualist’s original thought experiment — to
invert the qualia without changing anything physical or behavioral — is now
one step harder yet to imagine. If the inversion is to preserve the metric of
similarity and discriminability relations that structure our phenomenal qualia,
then it will require wholesale changes in the synaptic connections that project
to our various color-opponent cells, and/or major changes in the profile or
location of the normal response profiles of our three cone-cell populations. It is
these features of our nervous system, as we saw, that give rise to that nonuni-
form metric in the first place.

An inversion is still possible, to be sure, but if it were imposed, it would show
up not just in the subject’s color-discrimination behaviors (as we saw before); it
would also show up in the form of changed behavior in his cones and/or massive
physical adjustments to the wiring that connects his retinal cone cells to his
LGN opponent-cells. Evidently, as our understanding of the brain’s coding
activities gradually expands, the claim that qualia might be inverted among us,
without any behavioral or physical differences among us, looks less and less
plausible.

The dualist tries again

“Still,” the Dualist might say, “it remains conceivable. We need only invert as
well the metric of the similarityjdiscriminability relations, in addition to invert-
ing the causal map of color qualia onto the external world, and the inversion
will then require no synaptic adjustments and it will lead to no differences in
discriminatory behavior.”

190

Metaphysics

That is strictly true, although changing the global metric of the space of
possible qualia raises the issue of whether we have thereby made a significant
change in the nature of the qualia themselves. If every color in the original
space now bears a dijferent set of similarity and dissimilarity relations to every
other color in the original space, are we still talking about the same family of
colors that we started with? It is not clear that we are. But let us not insist on
this point. Who are we to insist that any one of the features we have been dis-
cussing is essential to the nature of color qualia, and could not conceivably be
switched around without compromising their identity? In the absence of any
settled scientific understanding of what qualia really are, any such insistence
would be premature and prejudicial. It is the job of unfolding research, in the
fullness of time, to provide us with authoritative grounds for claims about the
essential versus the accidental features of our color qualia.

The dualist is hoist by his own petard

But if this is a lesson we materialists must learn to swallow, it is a lesson no less
obligatory for the dualist. And for him it has an unwelcome edge to it, for the
dualist’s thought experiment, in all of its versions, depends on a preferential
fixing of “how they present themselves to introspective judgment” as the es-
sential feature of color qualia, while downgrading all other features of color
qualia (and, as we saw, there are quite a few of them) as inessential con-
tingencies, invertible without penalty at the drop of a thought experiment. But
this very insistence is also premature and prejudicial, however much it may re-
flect the uncritical convictions of untutored common sense. The dualist has no
more right to that premise than the functionalist has to “functional role” as the
essential feature of color qualia, or the reductionist has to “family of similarity
relations” as their essential feature. To insist on any one of these, before our
science on the matter is completed, is to do science by fiat instead of by con-
ceptual exploration and empirical evaluation.

Two points will drive this lesson home. The flrst is that the dualist himself
can be victimized by an alternate instance of his own strategy, as follows.

Intrinsic-character-as-judged-by-introspection cannot be the defining feature of
phenomenal qualia, since I can quite easily imagine that half the population
suffers from “phenomenal judgment inversion syndrome,” an undetectable
malady whereby the faculty of judgment makes systematically inverted, and
systematically mistaken, judgments about the identity of the phenomenal qualia
had by the subject. Our judgmental take on our qualia, therefore, can hardly be

191

Consciousness

definitive of their true nature. Accordingly, we shall have to dig even deeper
still to find the identifying essence of color qualia. (See pp. 118-123.)

Though I am disinclined to defend this argument, its mere existence is in-
structive. Conceivability, it seems, is a two-edged sword.

To this it may be replied, perhaps in exasperation, “But qualia are by defini-
tion those things whose appearance is the reality!” As an observation about
common usage, this claim may be strictly correct. But so were the following
historical claims, famously uttered and with equal exasperation. “But atoms
are by definition those things that cannot be split! (The Greek word “a-tom”
means not cuttable.) So you can forget about subatomic particles.” Or equally
fatuously, “But the Earth {terra firma) is by definition that-which-does-not-
move! So you can forget about its revolving around the Sun.”

These remarks illustrate the second major point. As Quine first argued, and
many others have underscored since, the meaning of words is not independent of
beliefs about what those words apply to, and also, no claim is immune to revision
or rejection in the face of sufficiently compelling new science. If science discovers,
as it did, that the Earth does in fact move, there is no point in trying to coun-
teract the evidence by saying, “But by ‘Earth’ I mean, in part, the thing-that-
does-not-move.” This strategy is futile, for the plain and simple reason that
whether the Earth moves or does not move depends on the facts of the matter.
It does not depend on an existing dictionary entry plus human resolve to pro-
tect the dictionary from revision. In the present context, this means that the
ultimate nature of phenomenal color qualia is something to be determined
by empirical research, and not by preemptive linguistic analyses and thought
experiments based on them. Thought experiments can be useful exploratory
devices, but they have no authority in dictating empirical facts.

So who is right?

None of this entails that the dualist hypothesis about the ultimate nature of
color qualia is false. Despite the emptiness of the inverted-spectrum thought
experiment as an argument for the truth of dualism, color qualia might still be a
metaphysically basic feature of some nonphysical sort. What, then, will decide
this issue?

The issue will be decided by the comparative virtues of the explanatory
theories produced by both parties to the debate. You have seen, for what it
is worth, what the physical sciences currently have to offer in the way of
explaining human color experience: the opponent-cell activation-space theory

192

Metaphysics

of human color coding. You have seen some of the evidence for it and some of
its explanatory prowess. Though far from proven, it plainly has at least some
virtues. We can reopen the discussion when the dualist produces a competing
explanatory theory (a competing explanation of the shape of the Munsell color
solid, for example), a theory with comparable specificity, supporting evidence,
and explanatory power. In the end, the issue is scientific, and competing theories
must be decided on their respective scientific merits. That is the ultimate lesson
of this section.

Doesn’t neuroscience leave something out?

Before moving on, it is useful to address one residual worry about the ability of
any purely physical theory — such as the one just examined — to account for the
qualitative character of our internal phenomenal experience. After all, knowing
the theory of neural coding across various color-opponent cells doesn’t tell me
how to recognize, in introspection, a visual sensation as a sensation of red. And
so, hasn’t it thereby left something out? After all, I could be color blind, and
thus phenomenally ignorant of that domain of experience. But learning the
neuronal theory just outlined wouldn’t help me one bit to repair that phenom-
enal ignorance.^®

This last sentence is entirely true. To have the perceptual skill of discrimi-
nating and recognizing colors requires more than just knowledge of the theory
of how our color-discriminating system works; it requires that the theory also
be true of oneself It requires that one actually possess a functioning instance of
the neuronal system that the theory describes. A color-blind person doesn’t have
that system, and so he is doomed to be phenomenally ignorant where colors are
concerned. Learning the theory of that missing system will be no help on that
score.

But this doesn’t mean there is anything inadequate about the theory, espe-
cially when that theory gives a detailed explanation of what produces the vari-
ous forms of color blindness (the lack of one or more retinal-cone types, which
leads to the partial or total loss of information reaching the several types of
color-opponent cells), and especially when the theory tells you what to do to
repair that discriminatory/representational deficit (namely, artificially induce
the genetic expression of the missing cone type(s), and induce the growth of
their missing synaptic connections with the LGN color-opponent cells).

The residual worry about leaving something out thus involves the confused
expectation that having-a-certain-cognitive-skill (namely, being able to have

193

Consciousness

and discriminate color qualia) should result from knowing-a-certain-theory
(namely, the color-opponent-cell theory). But the two are quite different things,
and simply knowing the latter (the theory) won’t give you the former (the skill).
However, if the theory at issue is true of you — if you actually have the neuronal
system that the theory describes — then you will certainly have the skill at issue.
You will be able to discriminate colors by spontaneous internal reaction to
their intrinsic qualitative natures. That ability is what needed an explanation
in the first place. And an explanation of that ability is precisely what the color-
opponent theory provides.

It is ridiculous to expect a reduction from the behavioral level directly to the
neuronal level

This observation is sometimes used to support the conclusion that conscious-
ness cannot be explained neurobiologically. The conclusion, however, just does
not follow, and hence the argument is a non sequitur. Here is why.

Nervous systems appear to have many levels of organization, ranging in
spatial scale from molecules such as serotonin, to dendritic spines, neurons,
small networks, large networks, brain areas, and integrated systems (see again
figure 1.1). Although it remains to be empirically determined what exactly are
the functionally significant levels, it is unlikely that explanations of macro-
effects such as face recognition will be explained directly in terms of the most
microlevel. More likely, high-level network effects will be the outcome of
smaller networks, and those effects in turn of the participating neurons and
their interconnections, and those in turn of the properties of protein channels,
neuromodulators, and neurotransmitters, and so forth. Emerging from efforts
to understand these levels are a range of midlevel concepts applicable to mid-
level neuronal organization and computation.

One misconception about the reductionist strategy interprets it as seeking a
direct explanatory bridge between the highest level and the very lowest levels.
This idea of “explanation-in-a-single-bound” does indeed stretch credulity, but
neuroscientists are not remotely tempted by it. In contrast, the direct and indi-
rect approaches predict that reductive explanations will proceed stepwise from
highest to lowest, both agreeing, of course, that the research should proceed at
all levels simultaneously. As more of the brain’s organizational midlevels and
their functions becomes known, a vocabulary suitable to those levels and func-
tions will certainly develop.

Microtubule
Anterograde motor
Retrograde motor
Neutrophic factor
Membrane receptor
Anterograde vesicle

Retrograde vesicle
Mitochondrion

Figure 4.19 This schematic diagram of a neuron and some of its organelles shows the
position of long (100 pm) microtubules in the axon and the shorter microtubules in the
dendrites. Axonal microtubules are oriented with the same polarity; dendritic micro-
tubules have mixed polarity. The microtubules have a diameter of about 14 nm. Notice

195

Consciousness

Consciousness is not a neural effect hut a subatomic effect

Roger Penrose, a Cambridge mathematician, and Stuart Hameroff, an Arizona
researcher on anesthesics, also harbor reservations about explaining awareness
neurobiologically, but are moved by rather different reasons (Penrose and
Hameroff 1995). They believe the dynamic properties at the level of neurons
and networks to be incapable of generating consciousness, regardless of the
complexity. For Penrose and Hameroff, the key to consciousness lies in quan-
tum events in tiny protein structures — microtubules — within neurons. Micro-
tubules are in fact found in all cells. They have a number of functions,
including mediating cell division. In neurons, they are used for the transport of
proteins up and down the axon and the dendrites. So our question is this: why
do Penrose and Hameroff believe that a subatomic phenomenon holds the
secret? And second, why do they find microtubules a particularly likely struc-
ture to mediate consciousness? I shall very briefly sketch their answers to these
questions.

The answer to the first question is that Penrose believes the nature of math-
ematical understanding transcends the kind of computation that could con-
ceivably by done by neurons and networks. As a demonstration of neuronal
inadequacy, Penrose cites the Godel Incompleteness Result, which concerns
limitations to theorem-provability in axiom systems for arithmetic. What is
needed to transcend these limitations, according to Penrose, are unique oper-
ations at the quantum level. Quantum gravity, were it to exist, could, he
believes, do the trick. Granting that no adequate theory of quantum gravity
exists, Penrose and Hameroff argue that microtubules are about the right size
to support the envisioned quantum events, and they have the right sort of sen-
sitivity to anesthetics to suggest that they do sustain consciousness (figure 4.19).

The details of Penrose and Hameroff’s theory are highly technical, drawing
on mathematics, physics, biochemistry, and neuroscience. Before investing time

that the microtubules do not extend into the synaptic end bulb of the axon. Neuro-
transmitter is synthesized in the cell body by the endoplasmic reticulum (1), packaged by
the Golgi apparatus (2), and transported down the axon or the dendrites by protein
motors on the microtubule. The speed of anterograde transport is 100-400 mm per day.
Vesicle proteins not sorted for reuse at the synpase are packaged into larger vesicles for
transport back to the soma for recycling. Neurotrophic factors collected from intra-
cellular space are also transported back to the cell body. The speed of retrograde trans-
port is about 50-200 mm per day. (3) Mitochondria are the site of the cell’s energy
production. (Based on Zigmond et al. 1999.)

196

Metaphysics

in mastering the details, most people want a measure of the theory’s “figures of
merit,” as an engineer might put it.^^ Specifically, is there any hard evidence in
support of the theory, is the theory testable, and if true, would the theory give a
clear and cogent explanation of what it is supposed to explain? After all, there
is no dearth of crackpot theories on every topic, from consciousness to sun
spots. Making theories divulge their figures of merit is a minimal condition for
further investment.

First, a brief interlude to glimpse the positive views Penrose has concerning
the question of how humans understand mathematics. In 1989 he suggested
as unblushing a Platonic solution as Plato himself proposed circa 400 b.c.;
“Mathematical ideas have an existence of their own, and inhabit an ideal
Platonic world, which is accessible via the intellect only. When one “sees”
a mathematical truth, one’s consciousness breaks through into this world of

ideas, and makes direct contact with it Mathematicians communicate ... by

each one having a direct route to truth” (1989, 428; Penrose’s italics).

As a solution to questions in the epistemology of mathematics. Platonism is
not remotely satisfactory. Given what we now know in biology, psychology,
physics, and chemistry, the Platonic story of mathematical understanding is as
much a fairy tale as the claim that Eve was created from Adam’s rib. Far better
to admit that we have no satisfactory solution than to adopt a “And God said
Lo” solution.

Let us return now to evaluating the quantum-gravity-microtubule theory of
conscious experience. The figures of merit are not encouraging. First, mathe-
matical logicians generally disagree with Penrose on what the Godel result
implies for brain function. Additionally, the link between conscious experiences
such as smelling cinnamon and the Godel result is obscure, at best.^®

Now, is there any significant evidential link between microtubules and
awareness? Hamerolf believes that microtubules are affected by hydrophobic
anesthetics in a way that causes loss of consciousness. But there is no evidence
that loss of consciousness under anesthesia depends on the envisaged changes
in the microtubules, and only indirect evidence that anesthetics do in fact — as
opposed to “could conceivably” — have any effect on microtubules. On the
other hand, plenty of evidence points to proteins in the neuron membrane as
the principal locus of action of hydrophobic anesthetics.^®

Is there any hard evidence that the subatomic effect they cite, namely quan-
tum coherence, happens in microtubules? Only that it might. Would not the
presence of cytoplasmic ions in the microtubule pore disrupt this effect? They
might not. Surely the effects of quantum coherence would be swamped by the

197

Consciousness

millivolt signaling activity in the neuronal membrane? They might not be. Can
the existence of quantum coherence in microtubules be tested experimentally?
For technical reasons, experiments on microtubules are performed in a dish ( in
vitro), rather than in the animal. If tests under these conditions failed to show
quantum coherence, would that be significant? No, because microtubules might
behave differently in the animal, where we cannot test for these effects. Does any
of this, supposing it to be true, help us explain such things as recall of past
events, filling in of the blindspot, hallucinations, and attentional effects on sen-
sory awareness? Somehow, it might.

The want of directly relevant data is frustrating enough, but the explanatory
vacuum is catastrophic. Pixie dust in the synapses is about as explanatorily
powerful as quantum coherence in the microtubules. Without at least a blue-
print, outline, prospectus, or something showing how, if true, the theory could
explain the various phenomena of conscious experience, Penrose and Hameroff
have little to tempt us. None of this shows that Penrose and Hameroff are def-
initely wrong, only that their theory needs work. Whether it is worth additional
work depends on how one assesses the theory’s figures of merit.

Science cannot solve all problems

Finally, consider Zeno Vendler’s admonition: science cannot expect to solve all
problems, answer all questions.®® Let’s agree with him — some questions may
never be answered. What, we must inquire, is entailed by this problem — the
problem of the neurobiology of consciousness? Absolutely nothing. Because
significant progress has been made by neuroscience on many questions about
the mind, it does look as though further progress is possible. We may at some
point hit the wall, but so far, at least, no reason has emerged to indicate the
wall has already been hit. What Vendler offers is not an argument, but off-the-
shelf Faustian rhetoric.

2.3 Conclusions

The principal aim of this chapter has been to convey my sense of where things
stand on the question of brain-based explanations for conscious phenomena.
The chapter does not pretend to be a survey of all views, since there are an al-
most limitless number of those. Nor are all of them equally worth discussing.
Of the theories I do discuss, some get better report cards than others. These
judgments reflect my particular, and quite possibly mistaken, opinions con-

198

Metaphysics

cerning what is productive and important, what is logically compelling or a
logical shell game.

The main philosophical argument submitted for uncompromised dissection
concerns the “inverted spectrum” argument. In my experience, this particular
problem is an unparalleled quagmire. Most of us are curious enough to want to
get into it. We vaguely sense that there is something to it, though exactly what
remains a bit foggy. We are reasonably confident that we can get to the crux of
it and come away with a clearer understanding. In the end, however, we tend
to find ourselves in an unholy jam, wondering how we got in the jam and
how we can avoid embarrassing ourselves further. My goal was to lay bare
the entire structure of the argument, following the various lines to their end,
however tangled the path. If this works, readers should be able to identify
the logical strengths or fallacies, and determine precisely what significance, if
any, the argument has for neuroscientific attacks on the problems of conscious
phenomena.

In addition to the inverted-spectrum argument, I considered a range of other
philosophical arguments allegedly demonstrating the futility of looking to neu-
roscience for answers to questions about the nature of consciousness. Though
each of the skeptical arguments considered boasts a considerable following,
and for that reason alone must be dissected carefully, none is convincing once
examined. Nor do they collectively create a credible skepticism even if individ-
ually they do not. That they are flawed does not, of course, show that neuro-
science will in fact be successful in expanding our understanding of
consciousness. It shows only that the skeptics’ conclusions regarding the mere
possibility are unconvincing. The most convincing answer to skepticism is, of
course, explanatory progress in neuroscience.

Suggested Readings

Churchland, P. M. 1988. Matter and Consciousness. 2nd ed. Cambridge: MIT Press.

Churehland, P. M. 1995. The Engine of Reason, The Seat of the Soul. Cambridge: MIT
Press.

Crick, F., and C. Koch. 2000. The uneonscious homunculus. In T. Metzinger, ed..
Neural Correlates of Consciousness, pp. 103-110. Cambridge: MIT Press.

Damasio, A. R. 1999. The Feeling of What Happens. New York: Hareourt Brace.

Dennett, D. C. 1991. Consciousness Explained. Boston: Little Brown.

Hobson, J. A. 1999. Consciousness. New York: Scientifie American Library.

199

Consciousness

Llinas, R. 2001. I of the Vortex. Cambridge: MIT Press.

Metzinger, T. 2003. Being No One: The Self-Model Theory of Subjectivity. Cambridge:
MIT Press.

Palmer, S. E. 1999. Vision Science: Photons to Phenomenology. Cambridge: MIT Press.
See especially chapter 13.

Parvizi, J., and A. R. Damasio. 2001. Consciousness and the brainstem. Cognition 79:
135-159.

Walsh, V., and A. Cowey. 2000. Transcranial magnetic stimulation and cognitive neu-
roscience. Nature Reviews: Neuroscience 1: 73-80.

Websites

BioMedNet Magazine: http://news.bmn.com/magazine
Comparative Mammalian Brain Collections: http://brainmuseum.org
Encyclopedia of Life Sciences: http://www.els.net

Higher Order Visual Areas: http://www.med.uwo.ca/physiology/courses/sensesweb/
L3HigherVisual/13v23.swf

The MIT Encyclopedia of the Cognitive Sciences: http://cognet.mit.edu/MITECS

This page intentionally left blank

Free Will

1 Introduction^

Much of human social life depends on the expectation that agents have control
over their actions and are responsible for their choices. In daily life it is com-
monly assumed that it is sensible to punish and reward behavior so long as the
person was in control and chose knowingly and intentionally. Without the
assumptions of agent control and responsibility, human social commerce is
hardly conceivable. As member of a social species, we recognize cooperation,
loyalty, honesty, and helping as prominent features of the social environment.
We react with hostility when group members disappoint certain socially sig-
nificant expectations. Inflicting disutilities (e.g., shunning, pinching) on the
socially erring and rewarding civic virtue help restore the standards.

In other social species too, social unreliability, such as a failure to reciprocate
grooming or food sharing, provoke a reaction likely to cost the erring agent. In
social mammals, at least, mechanisms for keeping the social order seem to be
part of what evolution has bequeathed to our brain circuitry. The stability of
the social-expectation baseline is sufficiently important to survival that individ-
uals are prepared to incur some cost in enforcing those expectations. Anyone
with dogs can observe the complex but general phenomenon of maintaining
social stability in dog interactions. Mature dogs will teach pups what is unac-
ceptable conduct, and typically dogs test newly encountered dogs, making it
clear what territory is theirs and what person they will defend. Just as anubis
baboons learn that tasty scorpions are to be found under rocks but cannot just
be picked up, so they learn that failure to reciprocate grooming when it is duly
expected may incur a slap. As discussed in chapter 3, much of our behavior is

202

Metaphysics

Figure 5.1 The frontal cortex (shaded) of six primates. The evolutionary relationships
among the species are indicated by the connecting lines. Although the human brain is
absolutely larger than the brains of the other primates, the proportion of frontal cortex
is roughly the same across species. Frontal cortex has an important role in planning,
impulse control, socialization, and the organization of behavior. (From Semendeferi et
al. 2002.)

guided by expectations of specific consequences of events, not only in the
physical world, and but also in the social world (figure 5.1).

If the reward and punishment system is to be effective in shaping social be-
havior, the actions for which the agent is rewarded or punished must be under
the agent’s control. The important question, therefore, is this: What is it, for us
or baboons or chimpanzees, to have control over our behavior? Are we ever
really responsible for our choices and decisions? Will neuroscientific under-
standing of the neuronal mechanisms for decision making change how we think
about these fundamental features of social commerce? These are the places
where issues about free will bump up against practical reality and our devel-
oping understanding of what is fair, what is reasonable, and what is effective in
maintaining civil society.

203

Free Will

2 Are We Responsible and in Control If Our Choices and Actions Are Caused?

One tradition bases the conditions for free will and control on a contrast be-
tween being caused to do something and not being so caused. For example,
if someone falls on me and I hit you, then my hitting you was caused by the
falling body; I did not choose to hit you. I am not responsible, therefore, for
hitting you. Were you to punish me for hitting you, it would not help me avoid
such events in the future. Examples emanating from this prototype have been
extended to the broader idea that for any choice to be free, it must be abso-
lutely uncaused. That is, it is suggested that I make a free choice when, without
any prior cause and without any prior constraints, I make a decision that
results in an action. Examples allegedly illustrating freely chosen actions are
Eisenhower’s decision to send troops into Little Rock to enforce school deseg-
regation, or my decision to go to the coffee shop for a cappuccino. This con-
tracausal construal of free choice is known as libertarianism.^ Is it plausible?
That is, are the paradigm cases of free choices actually uncaused choices?

As Hume demonstrated in 1739,^ the answer is no. Hume argued that our
free choices and decisions are in fact caused by other events in the mind;
desires, beliefs, preferences, feelings, and so forth. Thus Eisenhower’s decision
was the outcome of his beliefs about the situation and his desire to ensure that
the federal school-integration law was not flouted. His decision did not suddenly
spring uncaused into existence without preceding beliefs, thoughts, hopes, and
worries. I went to get a cappuccino because I usually have one about this time
in the afternoon, I wanted to have one, and I knew I had enough money to pay
for it, and so on. Save for these causal antecedents, albeit cognitive causal
antecedents, I would not have gone for coffee. By contrast, suppose that with-
out any antecedent causes, I suddenly enter a saloon, ask for a glass of vodka,
and gulp it down. I had no antecedent desire for vodka, no habit of going to a
saloon anytime, let alone in the afternoon, and the behavior would be consid-
ered utterly at odds with my cognitive state and temperament. Is this the para-
digm of free choice? Is this prototypically responsible behavior? Surely not.

Reflecting on these sorts of possibilities, Hume made the deeper and more
penetrating observation that an agent’s choices are not considered freely made
unless they are caused by his desires, intentions, and so forth. Randomness,
pure chance, and utter unpredictability are not preconditions for attribution
of responsible choice. Hume puts the matter with memorable compactness:

204

Metaphysics

“Where [actions] proceed not from some cause in the characters and disposi-
tion of the person, who perform’d them, they infix not themselves upon him,
and can neither redound to his honor if good, nor infamy, if evil.’”^

Logic reveals, Hume argued, that responsible choice is actually inconsistent
with libertarianism (uncaused choice). Someone may choose to climb onto his
roof because he does not want the rain to come into his house, he wants to fix
the loose shingles that allowed the rain in, and he believes that he needs to get
up on the roof to do that. His desires, intentions, and beliefs are part of the
causal antecedents resulting in his choice, even though he may not be intro-
spectively aware of them as causes. If, without any determining desires and
beliefs, he simply went up onto the roof — for no reason, as it were — his sanity
and hence his self-control would be seriously in doubt.

More generally, a choice undetermined by anything the agent believes,
intends, or desires is the kind of thing we consider out of the agent’s control,
and is not the sort of thing for which we hold someone responsible. Further-
more, desires or beliefs that are uncaused (if that is physically possible), rather
than caused by other stable features of the person’s character and tempera-
ment, likewise fail to be conditions for responsible choice. If a desire suddenly
and without antecedent connection to my other desires or my general character
were to spring into my mind — say, the serious desire to become a seamstress —
I would suspect that someone must be “messing with my mind.” The brain
presumably has no mechanism for introspectively recognizing a desire to fix the
roof as a cause, just as it has no way of detecting in introspection that growth
hormone has been released or that blood pressure is at 110/85. A desire,
nevertheless, is most certainly a cause.

Neither Hume’s argument that choices are internally caused nor his argu-
ment that libertarianism is absurd have ever been convincingly refuted. Notice,
moreover, that his arguments hold regardless of whether the mind is a sepa-
rate Cartesian substance or a pattern of activity of the physical brain. And they
hold regardless of whether the etiologically relevant states are conscious or
unconscious.

In fact, moreover, the brain does indeed appear to be a causal machine. So
far, there is no evidence at all that some neuronal events happen without any
cause. True enough, neuroscience is still in its early stages, and we cannot ab-
solutely rule out the possibility that evidence will be forthcoming at some later
stage. Given the data, however, the odds are against it. Importantly, even were
uncaused neuronal events to be discovered, it is a further, substantial matter to
show that precisely those events constitute choice. They might, for all we can

205

Free Will

know, have to do with features of growth-hormone release or variations in the
sleep/wake cycle.

Though all events in the brain may be caused, this does not imply that
actions are predictable. Causality and unpredictability are entirely compatible.
Causation concerns conditions that bring about an event, whereas predictabil-
ity concerns what we know about such conditions. When an event occurs in a
complex system, we may know that the event is causally governed, even though
on any given occasion we may not know exactly what conditions actually ob-
tain, and hence are unable to predict precisely the nature of the event. Never-
theless, despite our inability to make precise predictions, we can often make
useful general predictions. Thus I might be able to predict that a dollar bill
dropped from the top of the Eiffel Tower will fall to the ground in less than two
minutes, but I will be unable to predict exactly the fluttering pattern and its
precise downward trajectory. Those subtle changes in movement will depend
on moment-by-moment changes in air currents, and these changes will occur
much faster than I can take relevant measurements and do the relevant com-
putations, even if I were lucky enough to have very powerful computational
equipment. Every movement of the dollar bill is, nonetheless, caused.

Similarly, brain events relevant to decisions and choices are probably all
caused events, but this does not imply that I can predict with any great preci-
sion what you will say if I ask you for directions from UCSD to the Salk In-
stitute. I can predict roughly what you will say, however, if I know that you are
familiar with the area, that you are alert, paying attention, are not easily dis-
oriented, and that you tend to be forthcoming when asked for directions. I can
also predict with considerable confidence that given the opportunity, a human
will go to sleep at night for at least a few hours, that he will want to eat and
drink at some time during a 24-hour period, that he will not want to sit for very
long naked on an iceberg, and so on. I can predict that a neonate will suckle, a
puppy will chew shoes, and that most undergraduates will name carrots as the
first vegetable that comes into their minds. But these are rough and general, not
precise, predictions.

The brain is a dynamical system of enormous complexity. The human brain
is calculated to have about 10*^ neurons and about 10'^ synapses. The time
scale for neuronal events is in the millisecond range. If we assume that synaptic
events and neuronal events are the only causally relevant events, then to a first
approximation, this means that the human brain has about 10'^ parameters
that can vary over roughly 1-100 milliseconds. (This is a conservative estimate,
since there are intraneuronal events, such as gene expression, that are also

206

Metaphysics

relevant.) These figures mean that it is not physically possible to take all the
relevant measurements and perform all the relevant computations to grind out
a precise prediction in real time. So predicting on a neuron-by-neuron or syn-
apse-by-synapse basis is even less realistic than predicting the precise path and
flutter of the dropped dollar bill. The logical point, therefore, is this: causality
does not entail predictability, and unpredictability does not entail noncausality.
Put another way, causality and unpredictability are entirely consistent. ^

As we reflect on what would have to be true for us to have free choice,
we tend to be impressed by the fact that absolutely precise prediction of an
agent’s behavior is really impossible given the relevant variables and time
scales. We nurture the hunch that if you cannot predict whether I will choose a
green salad or a beet salad, or whether I will choose to say “Hi” or “Good
morning,” then my choices are really uncaused and therein lies my freedom
to choose. The hunch may be the more compelling if it gets support from this
tacit assumption: since “uncaused” implies “unpredictable,” “unpredictable”
implies “uncaused.” As I have shown, however, this is quite mistaken. The
implication goes only one way. “Unpredictable” does not imply “uncaused.”
Once the logic of the relation between causality and predictability are clarified,
no logical rationale remains for deriving expectations of noncausality from
facts of unpredictability.

Nonetheless, the idea that randomness in the physical world is somehow the
key to what makes free choice free remains appealing to those inclined to be-
lieve that free choice must be uncaused choice. With the advent of quantum
mechanics and the respectability of the idea of quantum indeterminacy, the
suggestion that somehow or other quantum-level indeterminacy is the basis for
a “solution” to the problem of free will remains attractive to some liber-
tarians.® Stripped to essentials, the hypothesis claims that although an agent
may have the relevant desires, beliefs, etc., he can still make a choice that is
truly independent of all antecedent causal conditions. On this view, the agent,
not the agent’s brain or his desires or his emotions, freely chooses between
cappuccino and latte, for example. It is at the moment of deciding that the in-
determinacy or the noncausality or the break in the causal nexus — whatever
one wants to call it — occurs. The subsequent choice is therefore absolutely free.

This is meant to be an empirical hypothesis, and as such, it needs to confront
neurobiologically informed questions. For example, what exactly, in neural
terms, is the agent who choosesl How does the idea of an agent who chooses fit
with what we understand about self and self-representational capacities in the
brain? Under exactly what conditions do the supposed noncaused events occur?

207

Free Will

Does noncausal choice exist only when I am dithering or agonizing between
two equally good — or perhaps equally bad — alternatives? What about when, in
conversation, I use the word “firm” rather than the word “stubborn”? Does it
exist with respect to the generation of desires? Why not? There are also ques-
tions from quantum physics, such as these: What is the mechanism of amplifi-
cation of the nondeterministic events? Were quantum effects of the envisioned
kind to exist, how could they fail to be swamped by thermal indeterminacy?

These are just the first snowballs in an avalanche of empirically informed
questions. Part of their effect is to expose the flagrantly ad hoc character of the
hypothesis. That is, it is based more on a desire to prop up a wobbling ideology
than on factual matters. Rather than fully discussing its merits and flaws now,
however, I shall defer a closer analysis of the hypothesis of a quantum-level
origin for uncaused choice until further details of the neurobiology of decision
making are on the table. That will allow us to see what bearing the neuro-
biological data have on the question of causality and choice in the brain, and
hence will provide a richer context for evaluating the hypothesis of noncausal
choice. We return to this hypothesis, and its critics, therefore, in section 6.

Provisionally, therefore, let us adopt the competing hypothesis, namely that
Hume is essentially right and all choices and all behavior are caused, in one
way or another. The absolutely crucial point, however, is that not all kinds of
causes are consistent with free choice; not all kinds of causes are equal before
the tribunal of responsibility. Some causes excuse us from culpability; others
make us culpable because they are part of the story of voluntary action. The
important question is what are the relevant differences among causes of be-
havior such that some kinds play a role in free choice and others play a role in
forced choice. That is, are there systematic brain-based differences between
voluntary and involuntary actions that will support the notion of agent re-
sponsibility? This is the crucial question, because we do hold people responsible
for what we take to be their actions. When those actions are intentionally
harmful to others, punishment, varying from social disapproval to execution,
may be visited upon the agent. When, if ever, is it fair to hold an agent re-
sponsible? When, if ever, is punishment justified?

Many possibilities have been explored to explain how the notions of con-
trol and responsibility can make sense in the context of causation. These fall
under the general rubric of “compatibilism,” which means that our work-a-day
notion of responsibility is, at bottom, compatible with the probable truth that
the mind-brain is a causal machine. First we shall consider some obvious but
unsuccessful attempts at squaring responsibility and causation, and then we

208

Metaphysics

shall raise the possibility that increased understanding of the brain will aid in
piecing together a plausible account.

3 Caused Choice and Free Choice: Some Traditional Hypotheses

3.1 Voluntary Causes Are Internal Causes

Can we rely on the following rule? “You are responsible if the causes are
internal, otherwise not.” No, for several reasons. A patient with Huntington’s
disease makes nonpurposeful, jerky movements as a result of internal causes.
But we do not hold the Huntington’s patient responsible for his movements,
since they are the outcome of a disease that causes destruction in the striatum.
He has no control over his movements, they are not voluntary, and they are not
consistent with his actual desires and intentions, which he cannot execute. A
sleepwalker may unplug the phone or kick the dog. Here too the causes are
internal, but the sleepwalker is not straightforwardly responsible. In a rather
attenuated sense, the sleepwalker may intend his movements, though he is ap-
parently unaware of his intentions.

3.2 Voluntary Causes Are Internal, They Involve the Agent’s Intentions, and
the Agent Must Be Aware of His Intention

This revision to the above strategy also fails. A patient with obsessive-compul-
sive disorder (OCD) may have an overwhelming urge to wash his hands. He
wants and intends to wash his hands, and he is fully aware of his desire and
intention. He knows that the desire is his desire; he knows that it is he who is
washing his hands. Nevertheless, in patients with OCD, obsessive behavior
such as hand washing or footstep counting is considered to be out of the agent’s
control. OCD patients often indicate that they wish to be rid of hand-washing
or footstep-counting behavior, but cannot stop. Pharmacological interventions,
such as Prozac, may enable the subject to have what we would all regard as
normal, free choice about whether or not to wash his hands.

3.3 Voluntary Causes Feel Different from the Inside

Another strategy is to base the distinction between voluntary causes and forc-
ing causes on felt differences in inner experience between those actions we

209

Free Will

choose to do and those over which we feel we have no control. Thus it allegedly
feels different when we evince a cry as a startle response to a mouse leaping out
of the compost heap and when we cry out to get someone’s attention and help.
Is introspection a reliable guide to responsibility? Can introspection — attentive,
careful, knowledgeable introspection — distinguish those internal causes for
which we are responsible from those for which we are not? (See also Crick 1994
and Wegner 2002.)

Probably not. There are undoubtedly many cases where introspection is no
guide at all. Phobic patients, the OCD patients just mentioned, and patients
with Tourette’s syndrome are obvious examples that muddy the waters. In a
patient with claustrophobia, the desire not to go into a cave feels as much his as
his desire not to go rafting without a life jacket. He can even give reasons for
both — it could be unsafe, avoidable injuries could happen, etc. His desire not
to go into a cave may be very strong, but so may his desire to eat when hungry
or sleep with his wife. So mere strength of desire will not suffice to distinguish
actions for which the agent has diminished responsibility and those for which
he is fully responsible.

The various kinds of addictions present a further range of difficulties. A
smoker feels that the desire for a cigarette is indeed his. His reaching for a
cigarette may feel every bit as free as reaching to turn on the television or
scratching his nose. He might wish it were not his, but so far as the feeling itself
is concerned, it is as much his as his desire to quit smoking. The increase in
intensity of sexual interest and desire at puberty is surely the result of hormonal
changes on the brain, not something over which one has much control. Yet all
of that interest, inclination, and alteration of behavior feels — from the inside at
any rate — entirely free.

More problematic perhaps, are the many examples from everyday life where
one may suppose the decision was entirely one’s own, only to discover that
subtle manipulation of desires by others had in fact been the decisive factor.
According to the fashion standards of the day, one finds certain clothes beau-
tiful, others frumpy, and the choice of wardrobe seems, intro spectively, as free
as any choice. There is no escaping the fact, however, that what is in fashion
has a huge elfect on what we find beautiful, and this alfects not only our
choices of clothes, but also such things as aesthetic judgment regarding plump-
ness or slenderness of the female body. Baseball hats worn backwards have
been in fashion for about ten years and are considered to look good, but from
another perspective, most people look less attractive if wearing a baseball cap
backwards.

210

Metaphysics

Social psychologists have produced dozens of examples that further muddy
the waters, but a simple one will convey the point. On a table in a shopping
mall, experimenters place ten pairs of identical panty hose and asked shoppers
to select a pair and then briefly explain their choice. Choosers referred to color,
denier, sheerness, and so forth, as their rationales. In fact, there was a huge
position elfect: shoppers tended to pick the pantyhose in the right-most position
on the table. None of them considered this to be a factor, none of them referred
to it as a basis for choice, yet it clearly was so. The ten pairs of panty hose
were, after all, identical to one another. Other examples of priming, subliminal
perception, and emotional manipulation also suggest that we will not get very
far with appeals to introspection to solve our problem about which behavior is
in our control and which is not.

3.4 Could Have Done Otherwise

In a dilferent attack on the problem, philosophers have explored the idea that if
the choice was free, the agent could have chosen otherwise. That is, in some
sense, the agent had the power to do something else.’ Certainly, this idea does
comport with conventional expectations about voluntary behavior, and to that
extent, it is appealing. Lyndon Johnson, historians say, could have done oth-
erwise regarding Vietnam. He could have decided to stop the war in Vietnam in
1965 when he correctly judged it to be unwinnable. I could have decided not to
get coffee, and perhaps to have water instead. Nobody forced me or coerced
me; the desire for coffee was mine. So far so good. The weakness in the strategy
shows up when we ask further, “What exactly does ‘could have done otherwise’
mean?” If all behavior has antecedent causes, then “could have done other-
wise” seems to boil down to “would have done otherwise if antecedent conditions
had been different.” Accepting that equivalence means the criterion is too weak
to distinguish between the shouted insults of a Touretter, whose tics including
such unpredicted and undirected outbursts as “idiot, idiot, idiot,” and those of
a member of parliament responding with “idiot, idiot, idiot” to another hon-
orable member’s proposal. In both cases, had the antecedent conditions been
different, the results would have been different. Nevertheless, we hold the par-
liamentarian responsible, but not the Touretter. So the proposed criterion seems
not so much wrong as unhelpful in revealing the nature of the difference between
the causes of voluntary behavior and the causes of nonvoluntary behavior.

A further problem lurking here is circularity. Testing for whether an agent
could have done otherwise seems to be exactly the same as testing whether the

211

Free Will

behavior was voluntary. Hence, specifying what counts as voluntary behavior
by referring to the possibility that the agent might have done otherwise just
goes around in a small circle. It does not seem to get us anywhere.

4 Toward a Neurobiology of Decision Making and Free Choice
4.1 Prototypes and Responsibility

In our legal as well as daily practice, we accept certain prototypical conditions
as excusing a person from responsibility, but assume him responsible unless
a definite exculpatory condition obtains. In other words, responsibility is the
default condition; excuse from and mitigation of responsibility has to be posi-
tively established. The set of conditions regarded as exculpatory can be modi-
fied as we learn more about behavior and its etiology. A dilferent but related
issue concerns what to do with someone who harms others but has diminished
responsibility.

Aristotle (384-322 b.c.), in his great work The Nicomachean Ethics, was an
early exponent of the principle that one is responsible unless there are excul-
pating reasons. And wise the principle is, so wise that the core of this approach
is still reflected in much of human practice, including current legal practice. In
his systematic and profoundly sensible way, Aristotle pointed out that for an
agent to be held responsible, it is necessary that the cause of an agent’s behav-
ior be internal to the agent, i.e., there must be intent. In addition, he charac-
terized as “involuntary” actions produced by coercion and actions produced in
certain kinds of ignorance. As Aristotle well knew, however, no simple rule
demarcates cases here. Clearly, ignorance is not considered excusable when it
may be fairly judged that the agent should have known. Additionally, in some
cases of coercion, the agent is expected to resist the pressure, given the nature
of the situation. A captured soldier is supposed to resist giving information to
the enemy. As Aristotle illustrated in his own discussion of such complexities,
we seem to deal with these cases by judging their similarity to uncontroversial
and well-worn prototypes. This is perhaps why precedent law is so useful.®

Increasingly, it seems unlikely that there is a sharp distinction — brain-based
or otherwise — between the voluntary and the involuntary, between being in
control and being out of control, either in terms of behavioral conditions or in
terms of the underlying neurobiology. This implies not that there is no distinc-
tion, but only that whatever the distinction, it is not sharp. That is, it is not

212

Metaphysics

like the distinction between having a valid California divers’ license and not
having a valid California drivers’ license. It is rather more like categories with a
prototype structure, e.g., “being a good sled dog,” “being a navigable river,”
“being a fertile valley.” These sorts of categories are useful even though we
cannot specify necessary and sufficient conditions for membership in such cate-
gories, but teach them by citing prototypical instances, along with contrasting
prototypical «o«instances.

Once we consider being in control in this light, we instantly recognize the
degrees and nuances typical of freedom of choice. An agent’s decision to
change television channels may be more unconstrained than his decision to pay
for his child’s college tuition, which may be more unconstrained than his deci-
sion to marry his wife, which may be more unconstrained than his decision to
turn off the alarm clock. Some desires or fears may be very powerful, others
less so, and we may have more self-control in some circumstances than in
others. Prolonged sleep deprivation makes it extremely hard to stay awake,
even when the need to do so is great. Hormonal changes, for example in
puberty, make certain behavior patterns highly likely, and in general, the neu-
rochemical milieu has a powerful effect of the strength of desires, urges, drives,
and feelings.

These considerations motivate thinking of control as coming in degrees, and
hence as falling along a spectrum of possibilities. Toward opposite ends of the
self-control spectrum are prototypical cases that contrast sufficiently in behav-
ioral and internal features to provide a foundation for a basic, if somewhat
rough-hewn and fuzzy-bordered, distinction between being in control and not,
between freely choosing and not, between being responsible and not. More-
over, as we consider various points on the spectrum, it seems likely that there
are in fact many parameters relevant to being in control. Consequently, we
should upgrade the simple one-dimensional notion of a spectrum to a multi-
dimensional notion of a parameter space, where the dimensions of the parame-
ter space reflect the primary determinants of in-control behavior.

In our current state of knowledge, we do not know how to specify all the
parameters or how to weight their significance. And the relations among the
parameters are not likely to be linear. We can nevertheless make a start. We
do know now that activity patterns in certain brain structures — including
the anterior cingulate cortex, hypothalamus, insula, and ventromedial frontal
cortex — are important. For example, large bilateral lesions to anterior cingu-
late abolish voluntary movement, though the patient remains aware of his sur-
roundings.® One fortunate patient recovered some voluntary function after a

213

Free Will

FUNCTIONAL SUBDIVISONS OF THE CINGULATE CORTEX

Figure 5.2 Functional division of the cingulate cortex of the rhesus-monkey brain. The
executive region (A) and the evaluative region (B) are the two major divisions. Sub-
divisions in (A): visceromotor (VMA), vocalization (VOA), nociceptive (pain) (NCA),
rostral cingulate motor (CMAr), and attention to action (AAA). Subdivisions in (B):
ventral cingulate motor (CMAv) and visuospatial (VSA). (Based on Vogt, Finch, and
Olson 1992.)

period of ininition. She also had good memories of her symptomatic episode,
during which, she explained, “nothing mattered,” and she said nothing because
she “had nothing to say.”^° Smaller lesions to the anterior cingulate are asso-
ciated with severe depression and anxiety (see figure 5.2).“ If a lesion occurs in
the middle area of the cingulate, patients may show loss of voluntary control
over a hand. In the alien-hand syndrome, as this deficit is called, the hand
behaves as though it has a will of its own. To the consternation of the patient,
the hand may grab cookies or behave in socially inappropriate ways. One pa-
tient discovered he could regain some control over his misbehaving alien hand
if he yelled at it, “Stop that!”

Imaging data implicate the anterior cingulate gyrus in the exercise of self-
control over sexual arousal. In an fMRI study, male subjects were first exposed
to erotic pictures and then were asked to inhibit their feelings of sexual arousal.
Comparisons between the two conditions show that when subjects are respond-
ing normally to erotic pictures, limbic areas show increased activation. When
subjects engage in inhibition of sexual arousal, this activation disappears, and

214

Metaphysics

the right anterior cingulate gyrus and the superior frontal gyrus become more
highly activated.

The anterior cingulate again emerges as a player in autism. One undisputed
finding is that autistics have deficits in analyzing affective signals. Because lim-
bic structures play a central role in affect, a leading hypothesis claims that
autism results primarily from defective affective evaluation, owing to structural
abnormalities in limbic system. This hypothesis has been tested by comparing
the micro structure of normal and autistic brains. Using whole-brain serial sec-
tions, researchers examined the brains of nine deceased autistic subjects. The
only cortical structure to show abnormalities was the anterior cingulate gyrus,
where the cells were smaller and the packing density greater. There were similar
abnormalities in limbic subcortical structures, including the hypothalamus,
amygdala, and mammillary bodies. Abnormalities in the cerebellum were also
seen.^"^

Additionally, it is known that levels of neuromodulators, such as serotonin
and dopamine, and of neurotransmitters, such as norepinephrine and acetyl-
choline, as well as of various hormones such as estrogen and testosterone, are
highly pertinent parameters in the well-tuned decision-making neural organi-
zation. For example, obsessive-compulsive pathologies and depressive path-
ologies involving loss of motivation can be greatly modified by increasing
serotonin levels (figure 5.3). It is also known that subjects with Klinefelter’s
syndrome (that is, those with XXY chromosomes) have poor long-term judg-
ment and impulse control, even when they are cognitively capable. Yet the
judgmental capacities of Klinefelter’s subjects improve markedly when they are
given constant administration of testosterone through a skin patch. Tourette’s
syndrome is much more controlled when patients are given serotonin agonists;
the subjects simply do not feel the same desire to engage in their customary
ticcing behavior. Since the anterior cingulate has been implicated in voluntary
behavior, it is noteworthy that both the dopamine projections and the nor-
epinephrine projections can influence activity in the anterior cingulate, and thus
have an influence on executive and attentional functions (see figure 5.4).^^

Appetite is a particularly promising parameter to consider in discovering the
brain-based differences between being or not being in control. Gluttony alleg-
edly is one of the seven deadly sins; overeating, we are repeatedly reminded, can
be controlled by sheer will power. The discovery of the role of the protein leptin
in eating, and in particular in over-eating, has provoked reconsideration of just
how much freedom of choice to push back from the table the very obese actually
have, and whether leptin-related interventions will give them greater control.^®

Figure 5.3 (A): Origin and distribution of the central noradenergic pathways in the rat brain. Note noradrenergic cell groups A1-A7,
including the locus ceruleus (A6). Abbreviations: DNAB, dorsal noradrenergic ascending bundle; VNAB, ventral noradrenergic
ascending bundle; CTT, central tegmental tract. (B): Origin and distribution of the central dopamine pathways. Note dopaminergic cell
groups A8-A10. Abbreviation: OT, olfactory tubercle. (C): Origin and distribution of the central cholinergic pathways. Note rostral
cell groups: nucleus basalis magnocellularis (NBM) (Meynert in primates), medial septum (MS), vertical limb nucleus of the diagonal
band of Broca (VDBB), horizontal limb nucleus (HDBB). Abbreviations: Icj, islands of Calleja; SN, substantia nigra; IP, inter-
peduncular nucleus; dltn, dorsolateral tegmental nucleus; tpp, tegmental pedunculopontine nucleus; DR, dorsal raphe; LC, locus cer-
uleus. (D): Origin and distribution of the central serotoninergic pathways. Note cell groups in the raphe nucleus, B4-B9. Common
abbreviations: MFB, medial forebrain bundle; PFC, prefrontal cortex; VS, ventral striatum; DS, dorsal striatum; cx, cortex. (From
Robbins and Everitt 1995.)

Percent correct Percent correct

216

Metaphysics

a Ceruleo-cortical NA loss; distraction b Mesolimbic DA loss; response latency

c Cortical cholinergic loss; accuracy d 5-HT loss; impulsive responding

Figure 5.4 Summary diagrams illustrating contrasting elfects in rats of selective dam-
age to the noradrenergic (NA), dopaminergic (DA), cholinergic, and serotoninergic (5-
HT) systems. A five-choice task was used. The diagrams highlight optimal conditions
for exposing deficits in each condition, (a) The ceruleocortical NA system. No delict on
the baseline, but accuracy reduced after distraction with white noise, (b) Mesolimbic DA
depletion. Baseline speed and overall probability of responding are primarily affected,
(c) Cortical cholinergic system. Baseline accuracy is impaired, (d) Serotoninergic deple-
tion. No effects on accuracy, but impulsive responding is increased. The control groups
are shown indicated by black bars. (From Robbins and Everitt 1995.)

217

Free Will

Intracranial
self stimulation

Ketamine

Nicotine

Sugar

Figure 5.5 Reward circuitry in the rat brain. The rat can self-stimulate by pressing a
bar that activates an electrode implanted in the region of the nucleus accumbens (Acc).
Experiments show that specific drugs act on the Acc, the ventral tegmental area of the
brainstem (VTA), and the locus ceruleus (LC). Abbreviations: DA, dopaminergic neu-
rons; Enk, enkephalin- and other opiod-releasing neurons; GABA, GABA-ergic inhibi-
tory intemeurons; NE, norepinephrine-releasing neurons; TEIC, tetrahydrocannabinal.
(Based on Gardner and Lowinson 1993.)

Leptin is a hormone released by fat cells. It acts on neurons in the hypothal-
amus that regulate feelings of hunger and satisfaction. Experiments on nonual
mice show that when the mouse has had an adequate meal, the leptin levels
increase, and the mice leave the food for other pleasures. Some mice are differ-
ent. They are obese, and they continue to eat even when their leptin levels
rise. Genetic analysis shows that the receptor to which leptin binds can have a
variety of mutations, and that the specific mutation predicts how overweight
the animal is. For example, if the mouse has the tu mutation, it is somewhat
tubby, relative to normals, and has twice the leptin levels of normals. If it has
the dh mutation, it is truly obese, and has ten times the leptin levels. There is

218

Metaphysics

something very different about the appetite regulation of the mutant animals
(figure 5.5 shows the reward pathways).

If a person is born with the db mutation of the leptin-receptor gene, and if, in
consequence, he feels as ravenous at the end of dinner as at the beginning, it
seems inevitable that he will overeat. More precisely, it seems reasonable to
assume that such a person will have less control over his eating behavior than a
person with the standard version of the leptin receptor. He may have perfectly
normal self-control when it comes to other matters, such as sex, alcohol, or
gambling, but for food, his situation is markedly different because his leptin
receptors in the hypothalamus are markedly different. The suggestion, there-
fore, is that the leptin receptor and its possible variations constitutes yet
another component in the complex neurobiological profile of “in control” sub-
jects, at least where food is concerned.

Many neural details remain to be uncovered, needless to say, but identifying
the major neurochemical players is a profoundly important beginning. Begin-
nings such as these inspire the vision that neuroscience might ultimately be able
specify a range of optimal values for the relevant parameters. When values fall
within the optimal range, the agent’s behavior is in his control. When values
are suboptimal, the agent will be unable to control his behavior. In between,
there may be gray areas where the agent is neither fully in control nor fully out
of control (figures 5.6 and 5.1)}''

Research from basic neuroscience as well as from lesion studies and scan
studies will be needed to transform this speculative parameter space into a
substantial, detailed, testable account of the features typical of in-control sub-
jects. These properties may be quite abstract, for “in control” individuals may
have different temperaments and different cognitive strategies.^® As Aristotle
might have put it, there are different ways to harmonize the soul. Nevertheless,
the prediction is that some such general features probably are specifiable. It is
relatively clear that dynamic-systems properties do distinguish between brains
that perform well or poorly such tasks as walking. What I am proposing here is
that more abstract skills, behaviorally characterized, such as being a successful
shepherd dog or a competent lead sled dog, can also be specified in terms of
dynamic-systems properties, dependent as they are on neural networks and
neurochemical concentrations. My hunch is that human skills in planning,
preparing, and cooperating, can likewise be specified. Not now, not next year,
but in the fullness of time as neuroscience and experimental psychology de-
velop and flourish.

219

Free Will

Figure 5.6 Localized lesions in limbic structures lead to specific behavioral changes.
(A) Lesions resulting in increases in aggressive behavior and in placidity. (B) Lesions
resulting in a release of oral behavior and in hypersexuality. (From Poeck 1969.)

In the next sections, we shall consider in more detail some of the evidence
that speaks in favor of this general approach.

4.2 Are We More in Control and More Responsible When Emotions Play a
Lesser Role and Reason Plays a Greater Role?

A view with deep historical roots assumes that in matters of practical decision,
reason and emotion are in opposition. To be in control, on this view, is to be
maximally rational and minimally emotional. To achieve rationality and self-
control, one must maximally suppress emotions, feelings, and inclinations. In a
metaphor sympathetic to this idea, Plato characterizes reason as a charioteer

220

Metaphysics

Some neural-level dimensions
of self control

serotonin

levels

Figure 5.7 Some dimensions of control at the neural level. The top panel is a highly
simplified representation of a subset of neural-level factors affecting self-control. The
bottom panel is a parameter space in which the axes, drawn from the larger set of
parameters, include leptin levels, amygdala-frontal connectivity, and serotonin levels.
Values in the normal range would have to be determined experimentally. The parame-
ter-space representation allows us to see that there are many ways of being normally in
control, and many ways of being out of control, as a function of the values and inter-
actions of many parameters. (Courtesy of P. M. Churchland.)

221

Free Will

who is pulled along by the appetites and emotions, and who must beat them to
avoid running amok.

Immanuel Kant is the philosopher best known for emphasizing on an oppo-
sition between reasons and emotions, and favoring the supremacy of reason. In
his moral philosophy, Kant saw human agents as attaining virtue only as they
succeed in downplaying feeling and inclination. He says, “The rule and direc-
tion for knowing how you go about [making a decision], without becoming
unworthy of it, lies entirely in your reason. This amounts to saying that you
don’t learn this rule of conduct by experience or from other people’s instruc-
tion; your own reason teaches and even tells you what to do.”^® The perfect
moral agent, on Kant’s view, is perfectly rational and entirely without emotion
and feeling.^® (Ronald de Sousa calls such an agent a “Kantian monster.”^^)

The kinds of cases that inspire Kant’s veneration of reason and his suspicion
of the passions are the familiar heart-over-head blunders. In such cases, the
impassioned do-gooder makes things worse, or one neglects long-term con-
sequences while satisfying an immediate need. The fool does not look before he
leaps. Othello, so overcome by jealousy that he failed to realize that he was
being duped, kills Desdemona. In the grip of an overwhelming bitterness,
Medea kills her children and herself. The moral failings of great tragedy are
typically character flaws involving great emotions engulfing weak reason.

Understanding the consequences of a plan, both its long-term and short-term
consequences, is obviously important, but is Kant right in assuming that feeling
is the enemy of virtue, that moral education requires learning to disregard the
bidding of inclination? Would we be more virtuous, or more educable morally,
were we without passions, feelings, and inclinations?

Not according to David Hume. Hume asserted, “Reason alone can never be
a motive to any action of the will; and secondly, it can never oppose passion in
the direction of the will” (1739, 413). As he later explains, “’Tis from the
prospect of pain or pleasure that the aversion or propensity arises towards any
object: And these emotions extend themselves to the causes and effects of that
object, as they are pointed out to us by reason and experience” (1888, 414). As
Hume understands it, reason is responsible for delineating the various eon-
sequences of a plan, and thus reason and imagination work together to antici-
pate pitfalls and payoff’s. But feelings, informed by experience, are generated by
the mind-brain in response to anticipations, and incline an agent for or against
a plan.

Common culture also finds something not quite right in the image of non-
feeling, nonemotional rationality. In the highly popular television series Star

222

Metaphysics

Trek, three of the main characters are severally portrayed as typically hot-
tempered, coldly reasonable, or moderate in all things. The pointy-eared semi-
alien Mr. Spock lacks emotion. In trying circumstances, his head is cool and his
approach is calm. He faces catastrophe and narrow escape with comparable
equanimity. He is puzzled by humans’ propensity to anger, fear, love, and
sorrow, and correspondingly fails to predict the role of emotions in human
affairs. Interestingly, Mr. Spock’s cold reason sometimes results in bizarre
decisions, even if they have a curious kind of “logic” to them.

By contrast, Dr. McCoy is found closer to the other end of the spectrum.
Individual human suffering inspires him to risk much, ignore future costs, or fly
off the handle, often to Mr. Spock’s taciturn evaluation, “But that’s illogical.”
The balance between reason and emotion is more nearly epitomized by the
legendary Captain Kirk. By and large, his judgment is wise. He can make
tough decisions when necessary; he can be merciful or courageous or angry
when appropriate. He is more nearly Aristotle’s ideal of someone who is wise in
practical matters.

4.3 A Disconnection Effect: E.V.R.

Neuropsychological studies reveal a lot about the significance of feeling in wise
decision making. Research by the Damasios and their colleagues on a number
of patients with brain damage shows that when deliberation is cut off from
feelings, decisions are likely to be impractical and disadvantageous in the long
run. Thus S.M., whose amygdala has been destroyed, has no feelings of fear
(see again figure 3.17). In complex circumstances, with no access to gut feelings
of unease and fear, she is as likely as not to make a decision that normal people
could easily foresee to be contrary to her interests.

In a rather more complex way, the point is dramatically illustrated by the
remarkable patient E.V.R. , who first came to the Damasios’ lab at the Univer-
sity of Iowa College of Medicine more than a decade ago.^^ A brain tumor in
the ventromedial region of E.V.R. ’s frontal lobes had been surgically removed,
leaving him with bilateral lesions. Following his surgery, E.V.R. enjoyed good
recovery and seemed very normal, at least superficially. For example, he scored
as well on standard IQ tests as he had before the surgery (about 140). He was
knowledgeable, answered questions appropriately, and so far as mentation was
concerned, seemed unscathed by his loss of brain tissue. E.V.R himself voiced
no complaints. In his day-to-day life, however, a troubling picture began to
emerge. Once a steady, resourceful, and efficient accountant, E.V.R. now made

223

Free Will

a mess of his tasks, came in late, failed to finish easy jobs, and so forth. Once a
reliable and loving family man, he allowed his personal life to become a sham-
bles. Because he scored well on IQ tests, E.V.R.’s problems seemed to his phy-
sician more likely to be psychiatric than neurological, and hence best treated
with psychoanalysis. As we now know, the psychiatric diagnosis turned out to
be quite wrong.

After studying E.V.R. for some time, the Damasios and their colleagues
conjectured that his lapses in practical judgment had something to do with a
disconnection between emotions and judgment. They repeatedly observed that
although E.V.R. could state the correct answer to questions concerning what
would be the best action to take (e.g., defer a small gratification now for a
larger reward later), his own behavior often conflicted with his stated con-
victions (e.g., he would seize the small reward now, missing out on the large
reward later). When they tested whether E.V.R.’s emotional responses were
in the normal range, they found intriguing abnormalities. For example, when
shown horrifying or disgusting or erotic pictures, his galvanic skin response
(GSR) was fiat.^"'' (Normals, in contrast, show a huge response while viewing
such pictures.) Curiously, if asked to say what he saw in the pictures, E.V.R.’s
emotional responses became somewhat more normal.

During the following years, new and more revealing tests were devised to
probe more precisely the relation between reasoning logically and acting in
accordance with reason. Antoine Bechara, working with the Damasios, devel-
oped a particularly revealing test. In this test, generally known as the Iowa
Gambling Task, a subject is presented with four decks of cards and told only
that his goal is to make as much profit as possible from an initial loan of
money. Money can be made and lost as a function of turning over cards, one at
a time, from any of the four decks. Subjects are not told how many cards can
be played before the game ends (a series of 100) or what the payoffs are from
any deck. One has to discover the winning strategy by trial and error. After a
card is turned over, the subject is either rewarded with an amount of money or
penalized and required to pay out money. Behind the scenes, the experimenter
designates two decks, C and D, to be low-paying ($50) and to contain some
moderate penalty cards; two other decks, A and B, pay large amounts ($100)
but contain very high penalty cards. Things are rigged so that players incur a
net loss if they play mostly A and B, but make a profit if they play mostly C
and D. Subjects cannot calculate losses and gains exactly because there is too
much mentally to keep track of (figure 5.8).

After about 15-20 trials, normal controls typically come to stick mainly with
the low-paying/low-penalty decks (C and D) and duly make a tidy profit in the

224

Metaphysics

normal

frontal

benign

stinker

benign

stinker

deck

Figure 5.8 The Iowa Gambling Task. Normal subjects begin to show autonomic-
nervous-system reactions, such as perspiring, when they reach for the bad decks, but
subjects with ventromedial frontal lesions do not. (Courtesy of P. M. Churchland.)

long run. In contrast, subjects with ventromedial frontal damage tend to end
the game with a loss. They generally work the high-paying decks, despite the
profit-eating penalty cards in those decks. Subjects with brain damage to
regions other than ventromedial behave like controls. Yet the ventromedial
patients had normal IQs.

As Bechara et al. note, even after repeated testing on the gambling task, as
long as a month or as short as 24 hours later, E.V.R continued to play the los-
ing decks heavily. When queried at the trial end, inevitably he correctly reports
that A and B are losing decks and rues his strategy. To put it rather paradoxi-
cally, rationally E.V.R. does indeed know what the best long-run strategy is, but
in exercising choice in actual situations, he goes for short-run gain, incurring
long-run loss. To make matters more difficult for the Kantian ideal, his judg-
ments of recency and frequency are flawless, his knowledge base and short-term
memory are intact. Because E.V.R. can articulate well enough the future con-
sequences of alternative actions, the problem cannot be lack of understanding
of what might happen. That his “pure reasoning,” displayed verbally, is at odds
with his “practical decision making,” displayed in behavior, suggests that the
crux of the problem lies with E.V.R. ’s lack of emotional responsivity to com-
plex plans.

Additional results came from a deeper analysis of skin conductance data
taken by a galvanometer placed on the arm of each subject during the gam-
bling task.^® In the gambling task, neither controls nor frontal patients showed
a skin response to card selections in the first few plays of the game (selections
1-10). By about the tenth selection, however, controls began to exhibit a skin

225

Free Will

response when they reached for the bad decks. When queried at this stage
about how they were making their choices, controls (and frontal patients) said
they had no idea whatever; they were just exploring. By about selection 20,
controls continued to get a consistent skin response when starting to reach for
the “bad” decks. In their verbal reports, controls said that they still did not
know what was the best strategy, but that they had a feeling that maybe decks
A and B were “funny.” By selection 50, controls typically could articulate and
follow the winning strategy. Frontal patients never did show a skin response in
reaching for any deck. They remain free of any affective guidance. What is so
striking is that for control subjects, choice was biased by feelings even before
subjects were clearly aware of their feelings, and well before they could articu-
late the winning strategy. That many of our daily choices are likewise biased,
without our being clearly aware of our feelings, seems likely.

The significance of nonconscious biasing by emotion has implications for the
economists’ favored model of “rational choice.” According to this model, the
ideally rational (wise) agent begins deliberation by laying out all alternatives,
calculating the expected utility for each alternative by multiplying the proba-
bility of each outcome by the value (goodies accruing) of each outcome. He
ends by choosing the alternative with the highest expected utility score. In light
of the data just considered, this model seems highly artificial, at least for the
ongoing daily activity of actual humans. Perhaps it is roughly correct for a
small range of somewhat artificial problems where the life-oriented computa-
tions that yield up the relevant options have already been performed. The point
is that normally in the on-going business of life, the set of options we con-
sciously consider is restricted by prior nonconscious, emotive-cognitive com-
putation, i.e., restricted by the “dirty” computation that gives us a set of
(mainly) relevant, sensible, and meaningful alternatives. This is a kind of com-
putation about which we know very little. At any rate, the economists’ model is
unlikely to come even close to giving the whole story of rational choice, though
it may be helpful once the set of reasonable alternatives is laid out.

On many occasions, one’s brain seems to have things pretty well sorted out
before conscious deliberation even begins. For example, in the grocery store I
rarely bother to consider Delicious apples, since they are usually punky; I am
not fond of eggplant, so I never pause over the eggplant bin. I never consider
drinking a can of paint; I never ponder whether to make a fur bathing suit or
porridge skis. I never consider beginning logic class with a demonstration of
how to milk a cow. And so forth. All these are descriptions of options my brain
could consciously entertain, but does not.

Subject A

Subject B

227

Free Will

One lesson taught us by E.V.R. and others with similar lesions (ventromedial
frontal) is that whatever rationality in decision making actually is, indepen-
dence from emotions is not its essence. When E.V.R. is confronted with a
question (“Should I finish this job or watch the football game?” “Should I
choose from deck A or from deck C7”), he is missing important emotional
clues that something is foolish or unwise or problematic. Normally, neurons in
the ventromedial frontal cortex project to and from areas such as the anterior
cingulate cortex, amygdala, and hypothalamus, which contain neurons signal-
ing body-state values. In patients with destruction of ventromedial cortex, the
pathways are disrupted. Their frontal lobes, needed for a complex decision,
have no access to information about the emotional valence of a complex situa-
tion, plan, or idea. Consequently, some of their behavior turns out to be foolish
and unreasonable.^® The point is not that patients like E.V.R. feel nothing at
all. Rather, it is that in those situations requiring imaginative elaboration of the
consequences of an option, feelings are not generated in response to the imag-
ined scenario, because the ventromedial frontal region needed for integration of
body-state representation and fancy scenario-spinning is disconnected from the
gut feelings. In particular, the capacity to remember relevantly similar occa-
sions with a recollection imbued with evaluative significance, is impaired.

An even more worrisome behavioral profile is seen when the prefrontal
lesions occur early in development. Anderson and colleagues reported on
two adults patients whose prefrontal lesions occurred before the age of 16
months.^® Both scored normally on various intelligence tests, but both were
severely impaired in their social behavior. In addition, they also showed defec-
tive social and moral reasoning, which suggests that the capacity to acquire
moral understanding was itself diminished by the early lesions. Whereas E.V.R.
and other late-onset lesion patients might do things that are socially inappro-
priate or foolish, they do understand and abide by moral rules (figure 5.9).

Figure 5.9 Neuroanatomical analysis, (a) A 3-D reconstructed brain of subject A.
There was a cystic formation occupying the polar region of both frontal lobes. This cyst
displaced and compressed prefrontal regions, especially in the anterior orbital sector,
more so on the left than on the right. Additionally, there was structural damage in the
right mesial orbital sector and the left polar cortices, (b) A 3-D reconstructed brain of
subject B. There was extensive damage to the right frontal lobe, encompassing pre-
frontal cortices in the mesial, polar, and lateral sectors. Both the lateral half of the
orbital gyri and the anterior sector of the cingulate gyrus were damaged. The cortex
of the inferior frontal gyrus was intact, but the underlying white matter was damaged,
especially in the anterior sector. (Reprinted from Anderson et al. 1999.)

228

Metaphysics

4.4 Agents and Self-Representational Capacities

If the various emotions play an ongoing and indispensable role in formulating
practically wise plans, both long-term and short-term, how does this lit into the
framework for agency, self-representation, and consciousness developed in
chapters 3 and 4? The answer is best laid out by referring once more to the
Crush emulator. As discussed earlier, to a first approximation, the motivation
for actions is anchored in the fundamental drives for food, sex, and survival. As
plans develop, the imagination generates representations of plan sequelae. To
these internally driven scenarios, as well as to perceptually driven representa-
tions, emotional responses are generated, via mediation of the brainstem struc-
tures, amygdala, and hypothalamus.^® The central function of the emulator is
to predict and evaluate consequences of proposed actions. As we saw, the
emulator can be employed on-line in making immediate decisions and olf-line
for high-level decisions involving longer time scales. The various emotions have
a central role in evaluating options and their consequences as threatening,
rewarding, dangerous, risky, painful, satisfying, and so forth. If these affective
states also represent the difference between threatening to him and threatening
to me, then the states are, on Damasio’s hypothesis, conscious feelings. In the
context of acquired cognitive-cum-emotional understanding about the world,
neuronal activity in these pathways calls forth certain memories, directs atten-
tion to certain perceptual and imaginative functions, and imbues certain per-
ceptions with practical significance. In contrast to the computational mode of
an iMAC or a PC, this is bio-computation, i.e., dirty, me-relevant computation.

The neural evaluation and assessment of options probably resembles less
the clean, step-by-step execution of an algorithm than it does the rough-and-
tumble jostling among puppies for access to the food supply. That is, the pro-
cess whereby neural networks settle into the next decision probably involves a
kind of competition, and the winning option moves ahead for assignment of
detailed movements. To put it crudely in the familiar framework of folk psy-
chology, a desire for immediate gratification can be outweighed by the fear of
missing out on a more valuable good in the longer run; the pain of exercise can
be endured for the sake of envisioned improvement in skiing performance; long
and dreary hours in the lab are sustained by the glimmering possibility of sat-
isfying one’s curiosity. On those occasions when a weighty decision involves
conscious deliberations, we are sometimes aware of the inner struggles, describ-
ing ourselves as having conflicting or ambivalent feelings. Some processes in
decision-making take longer to resolve than others, and hence the wisdom in

229

Free Will

the advice to “sleep on” consequential decisions. Everyone knows that sleeping
on a heavy decision tends to help us settle into the “decision minimum” we
can best live with, though exactly how and why are not understood. Are these
longer processes classically rational? Are they classically emotive? Probably
they are not fittingly described by our existing vocabulary. They are the pro-
cesses of a dynamical system settling into a stable attractor.

Introspection, as we know, is a highly limited and fallible guide to the
dynamical aspects of these inner processes, and folk psychology is at best a
crude interpretive filter in any case. Though introspection gives us some sense
of the neural hurly-burly subserving choice, we have little conscious access to
its neural nature. Nevertheless, good models of the interplay and competition
among parameters, whatever exactly they are, will probably emerge in time.

According to conventional wisdom, cognitive factors are used to predict
consequences, while emotive factors are used to evaluate the consequences, and
the two are entirely separate functions. From the point of view of the brain,
however, the situation is not that simple. The very formulation of certain gen-
eral goals, such as going to college or starting a business, likely employs an
inseparable alloy of cognitive-emotive elements. This is also true of more spe-
cific goals, such as taking the dogs to the beach or finding a babysitter. Scene
segmentation and pattern recognition in perception are, save for unusual cir-
cumstances, shot through with affect and meaning. In momentous decision
making, such as the decision to find the accused guilty or the decision to opt for
doctor-assisted suicide, the competition alluded to is never a one-dimensional
struggle between reason and emotion, but rather is a complex interplay be-
tween this cognitive-emotive consortium and that cognitive-emotive consor-
tium. The decision to have a latte rather than a cappuccino is, relatively
speaking, a completely trivial decision. Our choice really does not amount to a
row of pins. Such trivial choices are not, however, the model for those life
decisions which mark us as wise or foolish, as impulsive or measured, as lazy or
ambitious. Consequently, in developing adequate models of decision making,
we would do well not to make the latte-cappuccino choice the paradigm for
choice generally.

5 Learning What’s Reasonable and What’s Not

Aristotle would have us add here the point that there is an important relation
between self-control and habit formation. A substantial part of learning to cope

230

Metaphysics

with the world, defer gratification, show anger and compassion appropriately,
and have courage when necessary involves acquiring appropriate decision-
making habits. In the metaphor of dynamical systems, this is interpreted as
contouring the terrain of the neuronal state space so that behaviorally appro-
priate trajectories are “well grooved” or strongly attractive. Clearly, we have
much to learn about what this consists in, at both the behavioral and neuronal
levels. We do know, however, that if an infant has damage in critical regions,
such as the ventromedial frontal cortex or amygdala, then typical acquisition of
the proper “Aristotelian” contours may be next to impossible, and more direct
intervention may sometimes be necessary to achieve what normal children
routinely achieve as they grow up.^^

The characterization of a choice or an action as rational carries a strongly
normative component; it is not sheerly descriptive. In contrast, consider
describing an action as performed hurriedly, or with a hammer. Claiming that
an action is rational normally carries the implication that the choice was con-
ducive in some significant way to the agent’s interests or well-being, or to those
of kith and kin; that it properly took into account the consequences of the
action, both long-term and short-term. Thus the evaluative component.
Though a brief dictionary definition can capture some salient aspects of what it
means to be rational and reasonable, it hardly does justice to the real com-
plexity of the concept.

As children, we learn to evaluate actions as more or less rational by being
exposed to prototypical examples of rational actions, as well as of foolish or
unwise or irrational actions. Insofar as we learn by example, learning about
rationality is like learning to recognize patterns in general, whether it be rec-
ognizing what is a dog, what is food, or when a person is afraid or embarrassed
or weary. As Paul Churchland has argued, we also learn ethical concepts
such as fair and unfair, kind and unkind, by being shown prototypical cases
and slowly learning to generalize to novel but relevantly similar situations.

Peer and parental feedback fine-tune the pattern-recognition networks so
that over time they come closely to resemble the standard upheld in the wider
community. Nevertheless, as Socrates was fond of showing, articulating those
standards is awesomely difiicult, even when a person successfully uses the term
“rational,” case by case. Discriminating the reasonable from the unreasonable
is a cognitive-emotive skill, like discriminating whether the river is now navi-
gable by canoe, or whether and how attacking an enemy’s position will suc-
ceed. Using prototype knowledge, we can see how Scott’s skill in conducting
his Antarctic exploration was pitiful, while Amundsen’s was superb. Making

231

Free Will

the term “rational” precise in a way that fulfills the conditions for an algorithm
is almost certainly impossible. Failures in programming computers to conform
even roughly to common sense, or to understand what is relevant, are an indi-
cation of the nonalgorithmic, skill-based nature of rationality.

This is important, because most philosophers regard the evaluative dimen-
sion of ethical concepts to imply that their epistemology must be entirely dif-
ferent from that of descriptive concepts. What appears to be special about
learning some concepts, such as rational, impractical, and fair, is that the basic
wiring for feeling the appropriate emotion must be intact. That is, the proto-
typical situation of something’s being impractical or shortsighted typically
arouses unpleasant feelings of dismay and concern; the prospect of something’s
being dangerous arouses feelings of fear, and these feelings, along with percep-
tual features, are probably an integral part of what is learned in perceptual
pattern recognition.

Frankly dangerous situations — crossing a busy street, encountering a grizzly
with cubs — can likely be learned as dangerous without the relevant feelings. At
least that is suggested by the Damasios’ evidence from their patient S.M., who,
as a result of amygdala destruction, has no feelings of fear. Although she can
identify which simple situations are dangerous, this seems for her to be a purely
cognitive, nonaffective judgment. Her recognition is poor, however, when she
needs to detect menace or hostility or pathology in complex social or marketing
situations, where no simple formula for identifying danger is available. As
argued earlier, the appropriate feelings may be necessary for skilled application
of a concept, if not for fairly routine applications. This is perhaps why the fic-
tional Mr. Spock, lacking emotions, is plausibly poor at predicting what will
provoke strong sympathy or dread or embarrassment in humans.

Stories, both time-honored ones and those passing as local gossip, provide a
basic core of scenarios where children imagine and feel, if vicariously, the
results of such choices as failing to prepare for future hard times {The Ant and
the Grasshopper), failing to heed warnings {The Boy Who Cried Wolf), being
conned by a smooth talker {Jack and the Beanstalk), vanity in appearance
{Narcissus). As children, we can vividly feel and imagine the foolishness of
trying to please everybody {The Old Man and His Donkey), of not caring to
please anybody (Scrooge in Dickens’s A Christmas Carol), and of pleasing the
“wrong” people {Pinocchio). Many of the great and lasting stories, for example
by Shakespeare, Ibsen, Tolstoy, Aristophanes, are rife with moral ambiguity,
reflecting the fact that real life is filled with conflicting feelings and emotions.
They remind us that simple foolishness is far easier to avoid than great tragedy.

232

Metaphysics

Buridan’s dithering ass was just silly. Hamlet’s ambivalence and hesitation
was deeply tragic and all too human. In the great stories also is a reminder that
our choices are always made amidst a deep and unavoidable ignorance of many
of the details of the future, where coping with that very uncertainty is some-
thing about which one can be more or less wise. For all decisions save the
trivial ones, there is no algorithm for making wise choices. Matters such as
choosing a career or a mate, having children or not, moving to a new country
or not, deciding the guilt or innocence of a person on trial, deciding whether to
surrender or press on are usually complex constraint-satisfaction problems.

As we deliberate about a choice, we are guided by our reflection on past
deeds, our recollection of pertinent stories, and our imagining the sequence of
effects that would be brought about by choosing one option or another. Anto-
nio Damasio calls the feelings generated in the imagining-deliberating context
“secondary emotions” to indicate that they are a response not to external
stimuli, but to internally generated representations and recollections.^^ As we
learn and grow up, we come to associate certain feelings with certain types of
situations, and this combination can be reactivated when a similar set of con-
ditions arises. Recognition of a present situation as relevantly like a certain past
case has, of course, a cognitive dimension, but it also evokes feelings that are
similar to those evoked by the past case, and this is important in aiding the
cortical network to relax into a solution concerning what to do next. This is the
platform for one’s neuroconscience.

6 Uncaused Choice Considered Again

Much of this chapter has focused on the emerging account of the neurobiology
of decision making. The hypothesis on offer is that there are systematic neuro-
biological differences between being in control and being out of control, and
that these differences can be characterized in terms of fuzzy-bordered sub-
volumes of the multidimensional parameter space. The in-control subvolume of
the space may be relatively large, allowing for the fact that in-control humans
have different habits, cognitive styles, emotional tone, and so forth. Similarly,
the out-of-control subvolume may be very large, reflecting the fact that dys-
function to the reward system may yield an out-of-control profile that is very
different from that of a dysfunctional anterior cingulate cortex, which in turn
is different from that of a degenerating basal ganglia.

233

Free Will

As noted in section 2, there are spirited defenses of a totally different hy-
pothesis, namely that decisions made by in-control subjects are actually
uncaused decisions, whereas decisions made by out-of-control subjects are
caused. The most modern variation defends the idea that quantum indetermi-
nacy is at the root, somehow, of uncaused choice. Though briefly introduced
earlier, it is time now to reconsider the idea that real choice requires a break in
causality milliseconds prior to the emergence of the brain state that constitutes
the choice. An empirical hypothesis, it deserves to be weighed and evaluated as
an empirical hypothesis and compared to the rather different picture of the
brain discussed above.

Hume and his arguments aside, the credibility of the noncausal-choice hy-
pothesis depends on whether it can mesh with what is known so far about
neurons and nervous systems. Defenders of the hypothesis want it to be consis-
tent with existing well-established neurobiological data, not openly to clash
with the data. The hypothesis is that among the many details neuroscience has
not yet discovered is this fact: for quantum mechanical reasons, voluntary
choice is uncaused. Our task here is to ask whether, given what is well estab-
lished neurobiologically, this appears to be a plausible hypothesis with promis-
ing research prospects. The hypothesis classifles a choice as voluntary if and
only if it is uncaused. Caused choices, therefore, are deemed not free. As usual,
we can begin by raising questions to which the hypothesis should have some
noncontrived answers.

Why and how does a break in causality occur just for those particular brain
events that supposedly are paradigm cases of choice? How does the brain work
so that a simple behavior in conformity with good habit — routinely putting on
my seat belt, for example — is caused, whereas choosing a latte rather than a
cappuccino after dithering is not caused? What prevents these special noncausal
events from occurring when a nicotine addict reaches for another cigarette or a
child sucks its thumb or a highly trained but off-duty spy surveys his fellow
passengers for assassins? If, as is entirely likely, the brain events constituting
choice are distributed across many neurons, how is noncausality (quantum in-
determinacy) orchestrated across the relevant population? If the brain events
constituting choice are uncaused, what precisely are their relations to back-
ground desires, beliefs, habits, emotions, and so forth? Philosophical fantasies
floated in abstraction from the tough and detailed constraints of the real world
have an “in a single bound Jack was free” quality. Flippant answers to empir-
ically informed questions are, of course, always possible: “It just works like
that” or “Magic!” Unless the hypothesis can interdigitate with neurobiology

234

Metaphysics

and cognitive science to come up with nonfrivolous answers to these questions,
however, it will continue to look nakedly ad hoc.

Before the hypothesis can be taken seriously, it will have to garner empirical
confirmation and survive empirical tests. If uncaused choice is a quantum-level
effect, as may be supposed, the aforementioned questions, as well as those
raised in section 2, demand empirical answers. Under exactly what conditions
do the supposed noncaused events occur? Does noncausal choice exist only
when I am dithering or agonizing between two equally good — or perhaps
equally bad — alternatives? How do quantum-level effects know (so to speak)
when to occur and when not? Beyond the business of decisions, do quantum-
level indeterminacies exist with respect to such processes as the generation of
desires! Or beliefs! Why not? How is it they come into play with only some
conscious decisions but not others? Does this break in causality occur at the
synapse? If advocates of noncaused decision-making are serious, they will have
to do more than wave the flag of quantum-level indeterminacy and claim that
in a single bound choice is free. They will have to get into the business of em-
pirical confirmation.

7 What Happens to the Concept of Responsibility?

We need now to return to the dominant background question motivating this
chapter: if choices and decisions are caused, is anyone ever really responsible
for his actions? One very general conclusion is provoked by the foregoing dis-
cussion. On the whole, social groups work best when individuals are presumed
to be responsible agents. Consequently, as a matter of practical life, it is prob-
ably wisest to hold mature agents responsible for their behavior and for their
habits. That is, it is probably in everyone’s interest if we match up assignment
of responsibility with being in control and adopt the default assumption that
agents have control over their actions. Barring clear evidence that an agent’s
behavior was in the out-of-control subvolume of the parameter space, the agent
is liable to punishment and praise for his actions. This is, of course, a highly
complex and subtle issue, but the basic idea is that feeling the social con-
sequences of one choices is a crucial part of socialization — of learning to be in
the give-and-take of the group. It is part of acquiring the appropriate Aris-
totlean habits.^® Feeling those consequences is necessary for contouring the
parameter-space landscape in the appropriate way, and that means feeling the

235

Free Will

approval and disapproval meted out. Having social institutions that reinforce
those feelings helps maintain civil life.

A child must learn about the physical world by interacting with it and bear-
ing the consequences of his actions, or by watching others engage the world, or
by hearing about how others engage the world. Similarly, learning about the
social world involves direct or indirect cognitive-affective learning about the
nature of the social consequences of a choice. This must, of course, be consis-
tent with reasonably protecting the developing child, and also consistent with
compassion, kindness, and understanding. In short, I do not want the simplicity
of the general conclusion to mask the tremendous subtleties of child rearing.
Nevertheless, if the only known way for “social decency” circuitry to develop
requires that the subject generate the relevant feelings pursuant to social pat-
tern recognition, then the responsibility assumption may be preferable to any
version of a thorough-going assumption of nonresponsibility.

This leaves it open, of course, that under special circumstances agents should
be excused from responsibility or be granted diminished responsibility. In gen-
eral, the law courts are struggling, case by case, to make reasonable judgments
about what those circumstances are, and no simple rule really works. Neuro-
psychological data are clearly relevant here, as for example in cases where the
subject’s brain shows an anatomical resemblance to the brain of E.V.R. or S.M.
Quite as obviously, however, the data do not show that no one is ever really
responsible, that no one is really deserving of punishment or praise. Nor do
they show that when life is hard, one is entitled to avoid responsibility. To most
of us, the “Twinkie defense” seems a travesty of justice, but so does ignoring
someone’s massive lesion in the ventromedial frontal cortex.

Is direct intervention in the circuitry morally acceptable? This too is a hugely
complex and infinitely ramifying issue. My personal bias is twofold. First, in
general, at any level, be it an ecosystem or immune system, intervening in
biology always requires immense caution. When the target is the nervous sys-
tem, then caution by another order of magnitude is warranted. Yet not taking
action is still doing something, and acts of omission can be every bit as conse-
quential as acts of commission.

Second, the movie Clockwork Orange, typically conjured up by the very idea
of direct intervention by the criminal justice system, probably had a greater
impact on our collective amygdaloid structures than it deserves to have. Cer-
tainly, some kinds of direct intervention are morally objectionable. So much is
easy. But all kinds? Even pharmacological? Is it possible that some forms of
nervous-system intervention might be more humane than lifelong incarceration

236

Metaphysics

or death? I do not wish to propose specific guidelines to allow or disallow any
form of direct intervention. Nevertheless, given what we now understand about
the role of emotion in reason, perhaps the time has come to give such guide-
lines a calm and thorough reconsideration. Approaching these questions with
a careful Aristotelian determination to be as wise as possible may be prefer-
able to giving free rein to unreflective self-righteousness. Ideological fervor, on
the right or on the left, can often do greater harm than unhurried common
sense.

8 Conclusions

I have considered three vintage philosophical theses in the context of new data
from neuroscience: (1) feelings are an essential component of viable practical
reasoning about what to do (David Hume), (2) moral agents come to be mor-
ally and practically wise not by dint of “pure cognition” but by developing
through life experiences the appropriate cognitive-affective habits (Aristotle),
and (3) the default presumption that agents are responsible for their actions is
empirically necessary to an agent’s learning, both emotionally and cognitively,
how to evaluate the consequences of certain events and the price of taking risks
(R. E. Hobart, Moritz Schlick). Each of the theses has been controversial and
remains so now; each has been the target of considerable philosophical criti-
cism. Now, however, as the data come in from neuropsychology, experimental
psychology, and basic neuroscience, the empirical probability of each thesis has
increased. Consequently, many important social policy questions must be con-
sidered afresh, including those concerned with the most efficacious means,
consistent with other human values, for achieving civil harmony. Much, much
more needs to be learned, for example about the reward circuits in the brain,
about pleasure and anxiety and fear. Philosophically, the emphasis with respect
to civic, personal, and intellectual virtue has been focused almost exclusively on
the purely cognitive domain, with the affective domain largely left out of the
equation, as though the Kantian conception of reasoning were in fact correct.
In matters of education and social policy, how best to factor in feeling and
affect is something requiring a great deal of informed mulling and practical
wisdom. In any case, my hope is that understanding more about the empirical
facts of decision making, at both the neuronal and behavioral levels, may be
useful as we aim for practical wisdom and ponder improvements in our social
policy.

237

Free Will

Suggested Readings

Aristotle. 1955. The Nichomachean Ethics. Trans, by J. A. K. Thompson. Harmonds-
worth: Penguin Books.

Bechara, A., A. R. Damasio, H. Damasio, and S. W. Anderson. 1994. Insensitivity to
future consequences following damage to human prefrontal cortex. Cognition 50: 7-15.

Campbell, C. A. 1957. Has the self “free will”? In his On Selfhood and Godhood, 158-
179. London: Allen and Unwin.

Churchland, P. M. 1995. The Engine of Reason, the Seat of the Soul. Cambridge: MIT
Press.

Cooper, J. R., F. E. Bloom, and R. H. Roth. 1996. The Biochemical Batsis of Neuro-
pharmacology. 7th ed. Oxford: Oxford University Press.

Damasio, A. R. 1994. Descartes’ Error. New York: Grossett/Putnam.

Damasio, A. R. 1999. The Feeling of What Elappens. New York: Harcourt Brace.

Dennett, D. C. 1984. Elbow Room: The Varieties of Free Will Worth Wanting. Cam-
bridge: MIT Press.

Le Doux, J. 1996. The Emotional Brain. New York: Simon and Schuster.

Walter, H. 2000. Neurophilosophy of Free Will: From Libertarian Illmions to a Concept
of Natural Autonomy. Cambridge: MIT Press.

Wegner, D. M. 2002. The Illu.sion of Comcious Will. Cambridge: MIT Press.

Websites

BioMedNet Magazine: http://news.bmn.com/magazine
Encyclopedia of Life Sciences: http://www.els.net

The MIT Encyclopedia of the Cognitive Sciences: http://cognet.mit.edu/MITECS

This page intentionally left blank

II Epistemology

This page intentionally left blank

An Introduction to Epistemology

1 Introduction

Epistemology is the study of the nature of knowledge. Its core questions are
these: what things do we know, and how do we know them? Two competing
traditions, both originating in Greece in the fifth century b.c., contour the
intellectual landscape. Plato (427-347 b.c.) and Aristotle (384-322 b.c.) are
the principal sources for these separate traditions, for they adopted largely dis-
tinct strategies. To grasp the reality behind how things seem to be, Plato bet
on mathematical or a priori reasoning from the armchair, while Aristotle bet on
exploration of the natural world.

Plato contended that learning is really recollection; most of what we call
learning he thought to be the uncovering of innate knowledge. How that innate
knowledge got into our heads in the first place he never satisfactorily addressed.
Plato believed that coming to understand reality consists in the intellectual
apprehension of abstract objects, objects that exist not in the physical world
but in a timeless “realm of the intellect,” later dubbed “Plato’s Heaven.” For
Plato, the ideal kind of knowledge was mathematical knowledge. Mathematical
knowledge, in contrast to observation-based opinion about the natural world,
was certain, immutable, systematic, and universal. Or so it seemed to Plato,
and to Platonists ever since.

By contrast, Aristotle, Plato’s most celebrated pupil, emphasized the im-
portance of delineating successful methods — how best to use evidence and
reasoning — in acquiring reasonable beliefs about world and how it works. In
wondering how we perceive, reason, remember, and learn, Aristotle took a
more naturalistic approach. He believed perception and memory to be natural
functions facilitating the acquisition of knowledge, and he assumed that their

242

Epistemology

operations could be best understood by observation and experimental manipu-
lation. On the question of how memories are stored, for example, Aristotle
suggested that experiences “impress” themselves on the stuff of the mind, thus
leaving a semipermanent physical trace. When investigating our capacity for
knowledge, Aristotle focused less on metaphysical questions, such as whether
the soul survives the body’s death, and more on concrete “How does it work?”
questions asked about eyes and ears.

To a first approximation, the Plato/ Aristotle division in strategy has demar-
cated epistemology ever since. Note, however, that the distinction between the
Platonic and Aristotelian methodologies is mainly a difference in emphasis,
since Platonists perfectly well realize that if you want to know whether it is
raining, you have to look, and Aristotelians perfectly well realize that imagi-
nation and reasoning are important in figuring out how the world really works.

Where the methodologies really diverge is on problems that cannot be solved
just by looking — problems concerning the nature of the reality hidden behind
observable appearances. On such matters, those of the metaphysical predilec-
tion see the naturalists as fudging the Big Questions, while those with a natu-
ralistic bent see the metaphysicians as stuck spinning their wheels.

With the rise of modern science in the Renaissance, the Platonic strategy was
forced to make concessions regarding our knowledge of the physical world.
Empirical techniques for determining causal mechanisms for physical processes
like burning and breathing were demonstrably more successful than non-
empirical methods, such as reading sacred texts or trying to deduce the nature
of reality from metaphysical principles that were often both contentious and
obscure.

One major development concerned the nature of fire. As a result of his
explorations into combustion, the French chemist Lavoisier laid the ground-
work for the division between elements and compounds between 1772 and
1785. He figured out that during combustion, an invisible gas was combined
with burning wood. This gas was later referred to as oxygen. The long-standing
theory that this conflicted with said that combustion involved the ejection of a
substance called phlogiston, which allegedly gave off heat and light. In other
words, the conventional wisdom had the dynamics exactly backwards. With the
new understanding of burning in hand, Lavoisier then went on to develop the
equally surprising idea that animal respiration too consisted of the combination
of oxygen with carbon. This suggested a provocative and rather heretical con-
tinuity between mechanisms in living and nonliving things.

Other experimental explorations were equally surprising. By passing light
through a carefully constructed prism, Newton (1642-1727) showed ordinary

243

An Introduction to Epistemology

light to be a mixture of colored light, ranging from violet to red. In biology, the
allegedly obvious spontaneous generation of flies was demonstrated to be false
by enclosing meat in a tightly lidded container and leaving it for a few days.
The germ theory of disease gradually overturned the religiously grounded pun-
ishment theory of disease after Lister and Semmelweis showed that soap scrubs
by surgeons reduced lethal infections and Pasteur showed that heat killed
microbes in food and water. And what of the mind/brain? Not until the nine-
teenth century were empirical techniques systematically brought to bear on the
nature of psychological capacities, including the capacity to know.

2 The Rise of Empirical Philosophy in the Nineteenth Century

In his 1862 treatise, Wilhelm Wundt lamented that psychology, unlike physics
and chemistry, had scarcely advanced since Aristotle’s explorations almost two
thousand years earlier. He urged empirical psychology to liberate itself, hrst,
from the metaphysical preoccupations of philosophers and, second, from the
idea that introspection and logic are jointly sufficient to understand the nature
of the mind. Wundt’s cautions against uncritical reliance on introspection has a
modern ring: “self-observation cannot go beyond the facts of consciousness, . . .
[and] the phenomena of consciousness are composite products of the uncon-
scious psyche” (Wundt, 1862/1961, 57; my emphasis).

Wundt’s idea of composite products was a direct challenge to philosophers
using naive introspection to identify sensory simples. By “simple,” they meant
something that could not be analyzed or decomposed or reduced any further.
This was important to epistemologists because they assumed such simples must
be the foundation for knowledge. From what he knew about the sensory sys-
tem, Wundt realized that lots of nonconscious processing had to go on before
one was aware of a color or shape or sound. And, of course, Wundt was per-
fectly correct: certain favorite philosophical “simples,” such as the pain from a
burn, are now known to be combinations of somatosensory and hedonic com-
ponents processed in different brain regions. (See pp. 117-118.) These compo-
nents are dissociable with drugs or by lesion.

Besides encouraging replicable empirical studies of perception and memory,
Wundt realized that the empirical study of the mind needed to develop in three
areas in particular: cognitive development in childhood, comparisons between
the cognition of humans and other animals, and the effects of social interac-
tion on individual cognition. This latter study he called “Volkerpsychologie,”

244

Epistemology

translated as “folk psychology.” His three-fold proposal, like his caution con-
cerning introspection, was a masterstroke, though mainstream epistemology,
especially in the twentieth century, largely ignored all three domains charac-
terized by Wundt.

The early techniques for the empirical investigation of the mind/brain were
somewhat crude, at least as judged against today’s toolbox. There were no
MRI machines, no computers, and no microelectrodes for recording or stim-
ulating single neurons. Indeed, not until the end of the nineteenth century were
neurons finally identified as the cellular units of nervous activity, and even then,
the nature of the intricate communication between neurons remained a mys-
tery.

Despite these handicaps, the pioneering psychologists — such as Helmholtz,
Wheatstone, and Wundt — realized that to get significant results in a science of
the mind, they needed well defined, well controlled experiments and the instru-
mental means for measuring and quantifying the results. And stunning discov-
eries were indeed made. Thomas Young, actually quite early in the nineteenth
century, had already figured out that color vision must depend on merely three
types of light-sensitive receptor (red, green, and blue). Wheatstone demon-
strated for the first time that depth perception exploits the fact that each eye
gets a slightly different image of the world. Helmholtz showed how the sound-
induced vibration of the hairs embedded in the wide-to-narrow basilar mem-
brane of the cochlea was the physical basis for the audible tones ranging from
low to high pitches. Helmholtz also realized that what we are immediately
aware of in perception involves cognitive filtering — what he called “uncon-
scious inferences.”

Some philosophers, especially Alexander Bain, William Hamilton, and others
in the so-called Scottish school, recognized the potential in a scientific study of
the mind, and they pushed hard for its development. ^ Metaphysically inclined
philosophers, on the other hand, tended not to regard any of this as progress in
epistemology. They considered the important philosophical work to lie else-
where; for example, in figuring out a priori the “necessary conditions for
the possibility of knowledge” (the Kantians) or how a normatively justified
knowledge structure could be built on a foundation of self-justifying sensory
simples (the British Empiricists). For reasons we shall discuss presently, the
nonexperimental stream was the one that survived as the officially recognized
discipline of philosophy. Alexander Bain and the Scottish school are names
unknown to most contemporary philosophers. Like Wundt and Helmholtz,
they are now considered unimportant to the central concerns of genuine phi-
losophy, however important they might be to psychology or physiology.^

245

An Introduction to Epistemology

3 Empirical Philosophy and Darwin

Darwin’s theory of natural selection has profound epistemological implica-
tions. If humans and human brains are the product of eons of Darwinian evo-
lution, and if human capacities for perceiving, learning, and knowing are
capacities of the brain, then these capacities are products not of divine creation,
but of our evolutionary history. In particular, if humans are born with not mere
capacities but also a priori knowledge, then such knowledge would have to have
its origin in our evolutionary history, not in divine engineering. So the exis-
tence, character, and accuracy of inborn representations, if such there be, has
to lit within a comprehensive biological framework that includes what we know
about neural development in individuals and comparisons between human and
nonhuman brains (figures 6.1 and 6.2). This point was made earlier in the in-
troduction to metaphysics. I make it again and develop it further in this chapter
because mainstream epistemology, arguably the backbone of the academic dis-
cipline of philosophy, continues to do business as though Darwin never hap-
pened. That is, the profound discoveries of evolutionary biology have scarcely
touched mainstream epistemology.

Darwin published The Origin of Species in 1859. Selective breeding of
domestic animals such as dogs, horses, and sheep was a practice well developed
by farmers long before the nineteenth century. Beginning with wolves, humans
conducted generations of selective breeding to produce dogs as different as
cocker spaniels, bassets, greyhounds, and Newfoundland retrievers. There are
individual differences in offspring, and animal breeders select among individ-
uals for breeding those whose traits — color, ear length, water-loving — they
want to see in later offspring.

Darwin realized that an entirely natural process of selection, without a
breeder’s controlling selection of the breeding pairs, could also result in selec-
tive breeding, albeit much more slowly than breeding by farmers. Natural
selection, Darwin argued, could occur if an animal happened to have traits that
enhanced its capacity to thrive in a particular environment and it survived to
reproduce and pass on those favorable traits to its offspring. Most important,
he realized that variation among individuals in the offspring sometimes in-
cludes mutations, with the result that natural selection can eventually yield
completely distinct, noninterbreedable species. By and large, he knew, mutations
are deleterious, but occasionally a mutation may occur that results in a trait
that happens to be beneficial to an animal in a particular environment. Over

246

Epistemology

Telencephalon Diencephalon Mesencephalon Rhombencephalon

Forebrain Midbrain Hindbrain

Figure 6.1 Evolution of the vertebrate brain. Ancestral regions of the vertebrate
brain — the forebrain, midbrain, and hindbrain — are subdivided in animals that ap-
peared later in evolutionary history. In mammals, the forebrain radically increased
in size relative to the olfactory lobe, midbrain, medulla, and thalamus.

long periods of time, highly distinctive species could emerge, while others might
disappear, elbowed aside, as it were, by their more successful competitors. And
just as natural selection can yield animals with fur or feathers, so it can yield
animals with very fancy nervous systems.

Humans, just like other organisms on the planet, must compete for resources
to survive and reproduce. Adaptation to the environment through the evolution
of brain-based capacities, especially the capacities to perceive, learn, predict,
and solve puzzles was undoubtedly significant in the survival of our species.
Obviously, some of the capacities demonstrated by modern humans, for exam-
ple the ability to read or juggle or skate, were not selected for as such. Rather,
these culturally enabled capacities arise from other more general capacities,
such as intricate pattern recognition and motor learning, that presumably did
play an important role in the life of primitive hominids.

What is selected is really the whole animal, and that means the whole
package — weaknesses and strengths, warts and all. So if a trait is to become
common in a population, the whole animal that has that trait must reproduce
to yield offspring that also have that trait. DNA is the heritable material.
Genes are segments of DNA that code for proteins. No gene can code directly

247

An Introduction to Epistemology

Figure 6.2 Lateral views of seven vertebrate brains showing relative expansion of
major brain divisions. In humans, the olfactory bulb is not visible from the lateral
aspect, since it lies on the ventral aspect of the frontal lobes, which are greatly expanded
in humans relative to reptiles and rodents. Brains are not drawn to scale. (Based on
Northcutt 1977.)

248

Epistemology

for an organ or a capacity or a trait; genes can code only for stuff: specific pro-
teins or RNA.

It is, of course, common to speak loosely of particular functions being
selected by Darwinian evolution. Unless we are careful, this shorthand can
suggest that natural selection is a bit like a fairy godmother who reaches down
into an animal and its DNA, makes base-pair changes in DNA with a magic
wand, and lo, a whole new trait appears or an old trait is miraculously opti-
mized. That, of course, is Cinderella-style magic; it does no work in the real
biological world. Consequently, it is implausible to suppose that human lan-
guage or consciousness or decision making are owed to a wholly new software
package that just happened to get plugged into existing hardware. Certainly,
human software engineers might design a wholly new software package and
install it in a modestly upgraded computer. Biological evolution, however,
works very differently from technological evolution (figures 6.3 and 6.4).

Are human brains similar to the brains of other mammals? Indeed they are,
and the closer humans are to another species genetically, the closer the simi-
larity in brain anatomy. Humans and chimpanzees share about 98 percent of
their DNA; humans and mice share about 90 percent. The human brain and
the chimpanzee brain are, so far as is known, very similar anatomically, but
human and mouse brain, apart from size, are also similar organizationally
(figure 6.5).

There are some differences between human brains and other mammalian
brains, notably in overall brain size and in the size of certain general structures
relative to others. Rats, for example, have a cortex largely devoted to olfaction,
but the macaque-monkey cortex is largely devoted to vision. We all have
essentially the same organization in the spinal cord, brain stem, thalamus, and
cerebellum. An early hint of these similarities came from Frangois Magendie
and Charles Bell, who independently discovered as early as 1807 that the sen-
sory nerves enter the dorsal segments of the spinal cord, and the motor nerves
exit the ventral segments. This pathway specialization holds for all vertebrates,
whether rat, lizard, or human (figure 6.6).

Sensory signals in the gustatory, olfactory, somatosensory, and auditory
systems follow the much same routes, even in fish and mammals. For example,
signals from taste buds on the palate of humans and catfish travel via the
seventh cranial nerve into the medulla to the nucleus of the solitary tract and
then upwards to the peribrachial nucleus. In the brain, there are differences.
In humans and monkeys, but not in catfish, some fibers go directly from the
nucleus of the solitary tract to the thalamus.

Figure 6.3 Three stages (horizontal rows) of embryological development in eight species (vertical columns). The product of the
earliest stages in vertebrate development (the pharyngula, top row) is similar across species, with similar patterns of segmentation
and early eye development. In the middle row, limb buds have formed, and although the mammalian embryos (pig, deer, rabbit,
and human) are still quite similar to one another, they are now distinct from the fish, salamander, turtle, and chick embryos. In
the bottom row, more obvious differences have emerged between mammals and nonmammals, as well as among mammals.
(Reproduced from Haeckel 1874.)

250

Epistemology

stalk

tlx m

otx-2 m

emx-2 m

Hox-a2

Hox-b1

Hox-a1/b2

Hox-a3/b3

Hox-a4/b4

Hox-b5

en-2

Wnt-1

Dlx-2 m

FGF-8

BMP-4 m

Figure 6.4 The nervous system plan in chordates. Anterior is to the left, dorsal is up.
The list to the left identifies specific genes that are expressed (e.g., Hox genes) and signal
proteins that are secreted (BMP-4, Wnt, and FGF). The gray lines to their right show
the regions in which the genes and the signal proteins organize development of specific
divisions. (From Gerhardt and Kirschner 1997.)

No unique structures — structures without any homologues in other mammals
— are in evidence, at least so far as is known. Even frontal regions, long
believed to be suggestively larger in humans, relative to other brain structures
such as the cerebellum and thalamus, appear in recent anatomical measure-
ments to stand in much the same ratio to other brain structures across all
primates and even across all mammals. That is, we do have a larger prefrontal
cortex than rats, monkeys, and chimpanzees, but we also have a larger cere-
bellum, brainstem, sensory cortex, and motor cortex than rats, monkeys, and
chimpanzees (see again figure 5.1).^

lab pb DftJ Scr Antp Ubx abd-AAbd-B

DmwphUa

DNA

mouse

DNA

mouse

embryo

Figure 6.5 Similar spatial organization of compartments by Hox gene expression in
arthropods and chordates. Top: Drosophila adult. Center: Drosophila segmented embryo
at about 10 hours (phylotypic stage). Bottom: mouse pharyngula embryo at about 12
days (phylotypic stage). Notice that the anterior to posterior compartment order is
the same in the phylotypic stages of both Drosophila and the mouse, reflecting the 3 '-5'
order of genes on the chromosome. (Based on Gerhardt and Kirschner 1997.)

252

Epistemology

Figure 6.6 Organization of the peripheral nerves in chordates viewed in a cross-section
of the spinal cord at right angles to the cord. Sensory signals enter the cord through
neurons in the dorsal roots, and motor signals leave the cord via the motor neurons in
the ventral roots.

Some behavioral similarities between humans and other mammals indicate
similarity of wiring in certain systems. Human babies, monkeys, and even rats
“screw up the face” when given something sour to taste. They one and all
smack their lips when given something sweet. And they spit out something
bitter. We all learn bait shyness in one trial; that is, if a novel food is followed
after some hours by nausea, or if a familiar food in a novel place is followed by
nausea, we avoid the novel food or the familiar food in the novel place. This is
also true of many nonmammals, including birds, as was first shown by John
Garcia and colleagues in 1974. As Darwin emphasized, emotions such as fear,
disgust, joy, and anger are expressed in very similar ways across species (figure
6.7).

Molecular analyses have revealed that the very same neurochemicals found
in the human brain are found in the nervous system of leeches and worms, as
well as reptiles, birds, and mammals."^ Moreover, the physiology of neurons is
largely unchanged throughout the animal kingdom; neurons in spiders work

253

An Introduction to Epistemology

Figure 6.7 Kevin, an adolescent male bonobo from the San Diego Zoo, striking a
philosophical pose. (Photo by Frans de Waal.)

254

Epistemology

essentially the same as neurons in humans. The degree of genetic conservation
across species is truly surprising. (See also p. 324.)

Seen naturalistically, the mind is what it is because the brain is what it is. The
human brain is an evolved device, bearing the stamp of conserved structures
and reflecting the natural necessity to eat or be eaten and to mate successfully
or not. Some human capacities might have been fancier had a clever engineer
designed our brains from scratch. For example, a clever engineer might have
given humans a fourth cone type that allowed us to see in the ultraviolet range
and also allowed us to discriminate the blues much more finely. A fifth cone
type for discerning x-rays and a sixth for microwaves might have been handy as
well. It might have been nice to have had built into the material brain the
knowledge of mathematics as well as explicit knowledge that the brain, rather
than an immaterial soul, is the thing that thinks. Built-in knowledge of what
substances make good anesthetics would also have been a great boon to man-
kind. Personally, I have always wanted to speak Mandarin without having
laboriously to learn it. These innate capacities, however, we do not have. Either
there was no evolutionary pressure for the development of these capacities, or if
there was, no preadaptive structures were in place to be exploited in new ways.

Biological evolution thus raises a very general problem for a priori epis-
temology. According to the a priori tradition, knowledge about how the mind
works is innate. It depends on deduction, introspection, and reflection to be
made explicit. The problem is this: Given biological evolution, what could be
the Darwinian explanation for the inborn existence of factually correct knowl-
edge about the way the mind/brain works? What could have been the selection
pressures? Evidently, there was insufficient selection pressure for innate, factu-
ally correct knowledge of the nature of fire, reproduction, disease, the origin of
the Earth, and the nature of matter. All these things had to be figured out em-
pirically. Mind/brain function is a natural, biological phenomenon, and here too
the likelihood that correct factual knowledge of its mode of operation is built in
seems far-fetched. Certain perceptual or organizational dispositions might be
innate, but a priorists need the stronger claim implied by “innate knowledge.”

A priori epistemology could command some plausibility when it was thought
that the human mind was created by a divine being, who allegedly could plant
in the mind whatever he deemed appropriate. At this stage of human under-
standing, however, divine creation of the mind looks far less probable than its
biological evolution. Consequently, such plausibility as a priori epistemology
might have enjoyed in the heyday of the theological view of the world evapo-
rates in a scientific view of the world.

255

An Introduction to Epistemology

4 Why Does Nonempirical Epistemology Still Exist?

Why did traditional (nonempirical) epistemology continue to thrive as a disci-
pline well into the twentieth century? Why do we still have nonempirical epis-
temology? I discern two main reasons for this, though there are undoubtedly
others. The first explanation derives from the fact that the naturalistic project
has been very difficult to bring to maturity: brains are complex, fragile, and
very difficult to figure out. The second explanation, largely but not wholly in-
dependent of the first, concerns the rise of modern logic: its enormous power,
its Platonic proclivities, its seductive beauty, and the fact that it is technologi-
cally and computationally so very compliant — so very biddable, as one might
say. Modern logic seemed, for various reasons, to be an ideal tool for a priori
epistemology.

4.1 Slow Progress in the Natural Science of the Mind/Brain

What obstacles did the naturalistic program face? These have been touched on
in chapter 1 , but they can be drawn out a little here. Imagine looking at a slice
of cortex under a light microscope. What do you see? Unless the neurons are
selectively stained to stand out, not much. The development of stains to render
individual neurons visible was crucial to the development of neuroanatomy as a
discipline, and it was not achieved until late in the nineteenth century. Suppose
that we have applied a Golgi stain to the tissue, so that about 10 percent of the
neurons are filled with the stain and hence are visible against the teeming pop-
ulation of other cells (figure 1.4). Now what do you see? Unless you are quite
lucky, probably not a great deal. Neurons are three-dimensional, and a single
slice of neural tissue under a light microscope reveals only a two-dimensional
layer of the stained neurons. You need slices fore and aft in order to see more
of a neuron’s axons and dendrites. In a cubic millimeter of cortex, there are
about 10^ neurons and about 10^ synapses. Synapses are many, and they are
small. A synapse (about 1-2 microns) cannot be seen except with an electron
microscope, a device not invented until the mid 1950s. An invention of the
early 1990s, two-photon laser microscopy allows the experimenter see neurons
below the cortical surface, as well as calcium ions moving into a neuron fol-
lowing an action potential.

These are just the merest handful of the challenges facing anatomists. What
about the physiology! There too, highly specialized skills and techniques had to
be devised from scratch for extracting meaningful responses from individual

256

Epistemology

living neurons and from groups of neurons. Merely preventing the neuron from
dying before you got a measure of its activity was a major feat.

A lesion is an area of damaged brain tissue and can occur as a result of
stroke, tumor, closed head injury, dementing diseases, and so forth. Studying
humans with brain damage was, and still is, an extremely important technique
for addressing how the nervous system is organized, what its parts do, and how
activity in one part affects activity elsewhere. Nevertheless, in lesion studies,
problems constantly arose because the effects of lesions can be markedly dif-
ferent in babies than in adults, because the deficits caused by lesions can change
dramatically over time, and because in humans, lesions from stroke, tumors,
and so forth, could not be precisely located until autopsy. This meant that
interpretation of the behavioral data from lesion studies was persistently prob-
lematic. Scanning technology, developed in the last three decades, has largely
solved the problem of lesion-localization in living humans (pp. 18-20).

Other obstacles, more subtle perhaps, were due not to technological frus-
trations but to conceptual blocks, that is, to the lack of concepts adequate to
thinking about the problem or to articulating the right questions. Concepts are
the cognitive lenses we use to see the world and to think about it, and when
they are darkened by ignorance and distorted by error, our perception of the
world is likewise distorted and confused.

Consider an example. When William Harvey (1578-1657) began research on
the heart, his guiding question was this: exactly where in the heart are the “vital
spirits” concocted? His question reflected the respectable, conventional, too-
obvious-to-be-questioned wisdom of his time. According to this conventional
wisdom, blood was continuously and copiously made in the liver. The job of
the heart was to make vital spirits (by virtue of which life existed) by mixing air
from the lungs with blood from the liver. The reason death followed the cessa-
tion of heartbeat was that the vital spirits ceased to be concocted.

In one of the great stories of science, Harvey ended up discovering something
utterly different from what he sought. He discovered that the heart was actually
a meaty pump, blood circulates around the body, and blood is continuously
made, but not by the pint per minute and not in the liver at all. Shockingly,
Harvey’s discoveries implied that almost certainly there were no vital spirits
concocted in the heart — or anywhere else either. To come to see this, he had
to doff the conceptual lenses of the framework of spirits — vital, animal, and
natural — and don a completely different set of lenses. This he did: “Medical
schools admit three kinds of spirits: the natural spirits flowing through the veins,
the vital spirits through the arteries, and the animal spirits through the nerves, . . .
but we have found none of these spirits by dissection, neither in the veins, nerves.

257

An Introduction to Epistemology

Figure 6.8 The ventricles, as depicted by Hieronymous Brunschwig in the 1 525 edition
of his book from 1497. (From Finger 1994.)

arteries nor other parts of living animals.”^ Thenceforth the conceptual frame-
work of spirits was in decline.

From Galen (a.d. 130-200) until Vesalius (1514-1564), the conventional
wisdom about brains specified that the ventricles in the brain — the cavities
filled with fluid — were the seat of perception and cognition (figure 6.8). Now,
however, we are reasonably sure it is the neural tissue that perceives and
remembers, with the fluid in the ventricles serving a basically nutritive function.
Galen had it just backwards: he thought the holes were important for cogni-
tion, and the tissue played a supporting role. The ventricular hypothesis was
supported by the wider framework of vital spirits — animal and natural. If
animal spirits were the wherewithal for cognition, then as spirits, they were
most likely to be housed in holes, rather than in the meat. Obviously.

Conceptual obstacles included the aforementioned difficulty with specifying
mind/brain functions. In probing the mind/brain, nineteenth-century scientists
commonly assumed that there were fundamentally three functions: sensory,

258

Epistemology

motor, and the association of ideas. Too simple by orders of magnitude. To
others it seemed reasonable to suppose that the complexity in behavior is the
outcome of combinations of simple reflexes, such as the eye-blink reflex, the
gag reflex, and the knee-jerk reflex. This opinion gave rise to a reflex-based
theory of physiological mechanism. Again, this is too simple by orders of
magnitude. In this century it has seemed reasonable to suppose that speciflc
functions are handled by dedicated “centers” or modules, which, like different
computer applications, operate independently of what is going on elsewhere in
the brain. Again, too simple by orders of magnitude.® Until the 1980s and
1990s, it also seemed plausible that neurons passively receive signals in the
dendrites, integrate the signals, and at a certain threshold of current, the axon
sends an active signal. This is called the “integrate-and-flre” model. Alas, den-
drites are not passive, signals can be amplified, and the spike sent down the
axon can be propagated back up the dendrites. Additionally, more distant
dendritic segments have the means of amplifying their signals so that such sig-
nals do not decay by the time they reach the soma. Integrate-and-flre is far too
simple, even if a reasonable starting point. This has suggested to some neuro-
scientists that in fact the basic processing unit is not the neuron but segments of
dendrite. ''

The daunting nature of the problems, conceptual and experimental, has left
much scope for creative theorizing, and as noted in chapter 1, that is what
philosophy was in its heyday in ancient Greece. Evidently, theorizing, on a
small scale and grand scale, is one thing neurophilosophers — along with neu-
roscientists and empirical psychologists — must still do.

From the vantage point of hindsight, a lot of theorizing seems to be an utter
waste of time. In a certain sense, it is. Nonetheless, even getting things wrong is
not a waste of time, since falsiflcation at least helps narrow the search space.
Groping for a torch is what you have to do when you are in the dark, and until
some light appears, nothing is obvious. Exploration of conventional wisdom, of
the received framework — questioning it, beating on it, and seeing what alter-
natives look like — is just part of what has to occur until the science is suffi-
ciently established that it does not need such tumult anymore. Perhaps, of
course, even seemingly established sciences continue to benefit from some
tumult, simply because “well established” is never equivalent to “certainly
true.” The smug claims in the nineteenth century that physics was complete
remind us of that particular lesson. Nevertheless, some stages are more tumul-
tuous than others.

Benefiting from a handful of sensible hunches, Democritus in the fifth Cen-
tury B.c. argued that the reality behind the apparent diversity of objects and

259

An Introduction to Epistemology

substances must be “atoms” — indivisible, invisible units that hook up together
in different ways to yield different kinds of stuff, such as gold or hair. A waste
of time, perhaps, and yet it was a speculation that nipped at the heels of natural
philosophers until, lo and behold, by the time Dalton and Lavoisier started to
ponder the nature of elements and change, Democritus’s speculation could be
seen to be on the right track. Many other speculations, including some that
rose to orthodoxy, turned out to be dead wrong, such as that fundamentally
there are four elements — earth, air, fire, and water — or that fundamentally,
diseases arise from possession by devils or as divine punishment.

Getting the science even more or less right, however, is very, very hard,
and it is inevitable that most theories will end up on the scrap heap. What we
cannot do, however, is get along without theories altogether. We need some
concepts with which to see and think about the world; we always need some
hypotheses to frame the questions that motivate research. Pure observation
does not really exist, and idle observation generally doesn’t take you very far.
Since there are no algorithms for creating correct hypotheses, there are bound
to be many false starts. But without false starts, there are likely no starts at all.

In sum, although it has been difficult to make progress in psychology and
neuroscience, the strategy of pursuing epistemology in isolation from scientific
data is, I modestly opine, unwise. Managing as best as one can with meager
data is very different from turning your back on relevant data while justifying
such back-turning on grounds that philosophy is, after all, an a priori disci-
pline. On the positive side, philosophers have thoroughly explored the potential
for a priori theorizing about the nature of the mind. Thus its shortcomings, as
well as strengths, are clearly visible.

4.2 Logic, Recursion, and Cognition

So far my explanation of why epistemology endures has focused on the tech-
nical and scientific obstacles besetting a natural (i.e., scientific) theory of the
mind/brain. Although this explains why slow progress might have discouraged
some philosophers, it does not explain why a priori epistemology eclipsed em-
pirical epistemology in the twentieth century. There are undoubtedly a number
of pertinent sociological factors, such as the personal magnetism of Oxford
philosopher G. E. Moore (1873-1958).

Moore set a trend in philosophy that exalted what he called “common
sense.” In his view, common sense was to be valued not merely over foolish
enthusiasms, but even over scientific theories or philosophical theories that

260

Epistemology

were inconsistent with commonsense ideas. By “common sense” Moore did
not mean, however, just what the common person means; he had in mind an
understanding achievable by paying close, very close, attention to precisely
what we really mean (or perhaps should mean) by particular words. Clarity is,
to be sure, a good thing. But Moore’s strategy inspired the idea that analysis of
the meaning of the word “x” led to the truth about the nature of x. Thus phi-
losophers could claim to have a “method” that was not only a priori but even
more fundamental than the methods of science. This was sometimes referred to
as the linguistic turn in philosophy, though it was also derided by a few as
philosophy’s turning into a dead end. And now enter modern logic.

The dedicated — even celebrated — isolation of philosophy from the empirical
sciences of the mind/brain was also connected, I suspect, to a great achieve-
ment: the rise of modern logic. Logic was typically understood within an es-
sentially Platonic framework; it supposedly captured universal logical laws,
true independently of any actual human reasoning, laws true in any possible
universe.

The rise of modern logic is a story about a beautiful idea that braided
together several threads: (1) An algorithm is a “mechanical” procedure that
can be applied a finite number of times to crank tremendously complex struc-
tures out of very simple elements.® (2) If you identify the right set of basic ele-
ments and algorithms and pound out the right definitions, arithmetic and
mathematics generally can be shown to be a system whose truths are reducible
to the truths of logic. The envisioned reduction would be a great achievement,
because then mathematical truth would not be so mysterious; it would be just a
part of logic. This reductionist program is known as logicism. (3) Logic itself,
and reasoning generally, is just a complex structure resting on a finite set of
basic elements and definitions, with a finite set of rules or algorithms for
cranking out complex structures out of simpler structures. If this is true, then
we can understand the fundamentals of reasoning — and perhaps knowledge as
well — by figuring out the basic elements, rules, and definitions.

Bound together, these three ideas were appealing, for they implied that logic
is a well-defined system, composable and decomposable by dead-simple me-
chanical rules. Since Aristotle, logic had been a kind of hodge-podge of useful
rules of thumb. Modern logic pulled together a coherent and powerful system
out of the unconnected bits and pieces of ostensible logical truths. Cranking out
logical complexity from logical simplicity by recursion was surprisingly fruitful
for another reason; to the imaginative, such as Charles Babbage and later John
von Neumann and Alan Turing, it suggested mechanical computation, which
suggested computers, which suggested mechanical thinking.'^

261

An Introduction to Epistemology

The logicians realized that even if mathematics were reduced to logic, what
made logical truths true still had to be explained. Here Platonism — the theory
that the truths of logic are true because they inhabit Plato’s heaven — still
appealed to some logicians, most notably Frege. Others, such as Carnap, pre-
ferred to find an account with a lighter metaphysical burden. One attractive
idea was that logical truths are true by virtue of the meanings of their terms.
Roughly, this means that the axioms were definitionally true and the theorems
were guaranteed true by the rules operating on the axioms. The beauty of this
approach was that logical truth, and hence mathematical truth, no longer
requires the semimystical objects in Plato’s heaven. They require just under-
standing the meaning of terms in the language and using rules. Much mystery
about logic and mathematics seemed on the verge of disappearing. Thus
meaning, in the story as told by Carnap, came to take center stage.

Carnap believed that the systematic power of mathematical logic had un-
heralded potential, and he pushed to extend the application of the resources
and methods of logic from the reduction of mathematics to broader philo-
sophical issues. In particular, Carnap, as well as others with less interest in
science, such as G. E. Moore, suggested that meaning and analysis of the
meanings of terms were the keys to making progress not only about logic and
reasoning, but also about terms used in reasoning, such as “belief,” “desire,”
“reality,” and so forth. This meaning-oriented, language-dominated approach
is usually referred to as logical empiricism, and in its later incarnations, as
analytic philosophy. The explanation for the pair of names reflects the epis-
temology this approach embraced. The crux of the approach can be summa-
rized in three claims:

■ Human knowledge is made up of sentences, of which there are two kinds: (1)
analytic sentences, whose truth depends solely on the meanings of the terms
they contain, and (2) synthetic sentences, whose truth, given the meanings of
the terms, depends on how the world is.

■ There are two kinds of knowledge: (1) a priori knowledge (essentially logic
plus knowledge of the meanings of words) and (2) a posteriori knowledge
(knowledge of how the world is). The foundation of all synthetic knowledge,
including therefore scientific knowledge, consists of primitive observation
sentences whose truth we know directly, such as “This is round” or “This is
yellow.” Belief in any nonprimitive empirical sentence (“This is a lemon”) is
justified by logic and definitions. The two kinds of truth (analytic and syn-
thetic) map onto the two kinds of knowledge (a priori and a posteriori).

262

Epistemology

■ Many, perhaps most, philosophical problems disappear when the logic of the
language and the meanings of terms are properly analyzed. The meanings of
terms may not always be obvious. Much subtle analysis may be required to
reveal the deep meaning below the surface. Analysis consists of reflection,
consideration of counterexamples, thought experiments, and reasoning. It
requires philosophical training.

This set of convictions launched a program concerning how to do philosophy
and how to address classical philosophical issues. For one thing, the approach
had an appealing clarity. At the very least, it seemed far less obscure than the
theories pushed by neo-Kantians and Hegelians, who were wont to debate
issues about the unreality of space and time. Notice, however, that despite
including the word “empiricism” in its name, logical empiricism was, in certain
crucial respects, more Platonic than empirical in spirit. To begin with, the new
logic was really designed to make a hand-in-glove lit with mathematics, not
with the psychology of perceptual processing or with spatial or temporal prob-
lem solving or with the use of images in reasoning. It was the reduction of
mathematics to logic that motivated the development of the new logical ma-
chinery, and the logical machinery was shaped accordingly.

For another, questions about knowledge were walled off from empirical
studies about how people and other animals actually know and learn. “Lin-
guistic” epistemology was largely restricted to a priori reflection on the mean-
ings of such words as “knowledge,” “justification,” “person,” “mind,” and
their deductive environments, as revealed through reflection. Thought experi-
ments, as opposed to results from actual experiments, were considered to have
a special role in revealing so-called conceptual necessities, such as the “neces-
sary conditions for having a mind,” or the “necessary conditions for the possi-
bility of any knowledge.” Conceptual necessities, since they were necessities,
were supposed to tell us something beyond science about how the mind and its
conceptual framework had to work. And that meant “work in reality.” So the
search for conceptual necessities was the disguise philosophers used to sneak
past the limitations of talking about what words commonly mean, to talking
about how things actually are, without worrying about what science says on the
matter.

In the period from about 1910 to 1931 the program of the logical empiricists
appeared to go moderately well. Before long, however, several disasters struck.
The first was this: it turned out that logic plus definitions are not sufficient to
reduce mathematics. The reduction also required set theory. Well, if set theory
is axiomatically as secure as logic, is logic plus set theory so bad? Alas, the

263

An Introduction to Epistemology

needed axioms from set theory were hardly as self-evident as the axioms of
logic, such as “No sentence p is both true and false.”

For starters, several nonobvious set-theoretic propositions had to be intro-
duced as axioms to avoid crippling paradoxes on the one hand and deductive
inadequacies on the other. One such assumption was that there exists an infin-
ity of objects. This assumption not only failed to satisfy the ideal of being self-
evidently true or true by virtue of meaning alone, it looked quite possibly false.
Or at least, if it were true, its truth depends on how the empirical world is, not
on the meanings of words, nor on any Platonic objects, for that matter. If you
avoided that axiom, other axioms even less self-evident had to be invoked.
Alas, these befell a similar fate.

Second and perhaps most devastating, the original goal of Frege and Russell
proved to be impossible. Not merely difficult to achieve, but impossible. In
1931, the mathematician Kurt Godel proved that no matter how you axiomat-
ized arithmetic, as long as the axiomatization was consistent, you could not
crank out all the truths of arithmetic.^® In other words, if the axiomatization
was consistent, it was not complete. Hence, Gddel’s result is known as the
Incompleteness Result. Because the tools of modern logic enabled Godel to
prove his result, the proof was both undeniably brilliant and horribly galling.
The logicist program, ironically, was unable to succeed precisely where it had
seemed most promising.

These disasters might have motivated a quiet shift to empirical epistemology,
but oddly enough, they did not, by and large. It was as though the hive decided
to try to continue even though the queen was dead. In particular, philosophers
hoped that logic could be used to reveal the structure of knowledge of the
empirical world and how it rested on a foundation of sensation sentences. If
the bagel had a hole in it, they hoped that at least the outside was still fairly
substantial.

To keep their hopes alive, the logical empiricists needed to prove their claim
that sensation sentences are the foundations of belief structures. The idea was
that we have direct knowledge of sensations, but not of objects in space and
time, let alone of things like genes and gravity. Direct knowledge was supposed
to be unmediated by processes such as inference. (See also pp. 117-118.) So
physical-object sentences (e.g., “My cow is brown,” “The sun is hot”) had to
be reduced to or justified by sensation sentences (e.g., “Brown here now” and
“Cow smell here now”), plus definitions, along with the resources of logic.

A host of problems undermined the “direct knowledge” of sensation-
sentences story. As noted earlier (p. 117), Wundt had realized that the “phe-

264

Epistemology

Figure 6.9 The effects of perceptual organization. The letters M and W are obvious in
(A), less so in (B), and are fully camouflaged in (C). The stimuli are the same in each
case; only the spatial relations change. (From Palmer 1999, with permission.)

Figure 6.10 A grouping effect studied by Gestalt psychologists, (a) the law of good
continuation predicts that subjects visually group (i) and (ii) as forming one object,
and (iii) and (iv) as forming a second object, (b) However, when the same pattern is
embedded in a larger context, subjects see a line (iii-ii) intersecting a wave. (From Rock
1975.)

nomena of consciousness are composite products of the unconscious psyche.”
Seemingly direct knowledge is always a product of prodigious nonconscious
processing. Because empirical research on perception, learning, and reasoning
was assumed to be irrelevant to the business of philosophy, little if any atten-
tion was paid to empirical research showing that our “noninferential percep-
tions,” such seeing a set of lines as forming a specific shape (figures 6.9 and
6.10) or smelling an odor as of a rotting carcass, are in fact the results of highly
complex processing even though conscious inferences are entirely absent. (For
more detail on these problems, see chapter 7.) The “foundations” part of the
program was therefore in deep trouble.

So was the project to deduce or otherwise justify truths about objects
(“The cow is in the barn”) from truths about sensations (“Brown here now”).

265

An Introduction to Epistemology

The envisioned deduction was simply impossible to achieve, even though great
ingenuity went into drumming up suitable definitions to fuel the engines for the
sought-after deductions. Even softening the criteria for justification still left
the object sentences deductively unreachable from the sensation sentences. Not
even the more modest attempt of justifying very simple laws of nature (such as
“All mammals are warm-blooded”) from observations sentences about sensa-
tions (“Warm furry thing here”) could get to first base. Carnap’s heroic but
ultimately vain attempt in 1928 to push forward the logicist project in epis-
temology convinced many philosophers that perhaps the project was funda-
mentally misconceived. Nevertheless, philosophers by and large still did not see
any wisdom in the empirical epistemology of Wundt, Helmholtz, and the
Scottish School.

Hopes drifted towards the possibility of dispelling various “philosophical
puzzles” by analyzing the meaning of problematic words and laying bare the
clean logic of sentences. This was essentially G. E. Moore’s strategy, and it got
a boost when Wittgenstein abandoned the logicist strategy and turned to the
contemplation of meaning and the delivery of obscure aphorisms. “ Analyses
of meaning and thought experiments came to be considered philosophical tools
par excellence. Using these contemplative procedures, ill-defined though they
were, was advertised as leading to the discovery of so-called “conceptual neces-
sities” or “conceptual truths,” which were supposed to reveal a priori truths
about knowledge, perception, reasoning, causality, and so forth. They were
certainly not truths that one could acquire simply by consulting a dictionary,
for the process of meaning analysis required long philosophical training.

Unfortunately, alleged “analyses of meaning” were often thinly disguised
propaganda for someone’s doctrinal hobbyhorse. Typically, “thought ex-
periments” were unconstrained, poorly defined, and impossible to evaluate.
They legitimized a lot of clever but unprofitable wrangling concerning what
could and could not be imagined, what the significance of some “thought ex-
periment” was supposed to be, how to weigh counter-thought-experiments, and
on and on. Since criteria to evaluate underdescribed thought experiments are a
bit like criteria for evaluating the virtues of fairies over gnomes, making rec-
ognizable progress was difficult. Probably the most devastating, though
largely ignored, criticism was served up by Paul Feyerabend: analysis of the
actual meaning of “x” only tells you what some people in a certain place and
at a certain time believe about xs. It does not tell you anything about what is
true about xs.

What decisively weakened the linguistic-analysis program was a set of vexing
problems and a counterproject proposed by Harvard philosopher W. V. O.

266

Epistemology

Quine. Recall that the mainstay of the logical empiricists’ approach to meaning
was the existence of a principled “analytic/synthetic” distinction. Analytic sen-
tences are variously specified as necessarily true, true by virtue of the meanings
of the concepts (which, allegedly, sometimes requires deep philosophical analy-
sis). Analytic sentences are said to be conceptual truths. The falsity of an ana-
lytic sentence is supposed to be inconceivable, unimaginable, and so on. Hence
the reliance on thought experiments. Synthetic sentences, by contrast, are true
by virtue of the facts. If this distinction turned out to be rotten, the program
would be in tatters.

In the 1950s Quine realized that the analytic/synthetic distinction was at best
a continuum, not a genuine two-bin dichotomy. Moreover, he concluded that
no distinction between analytic sentences and synthetic sentences would do the
work that the logical empiricists required. He first observed that the expressions
“analytic truth,” “necessary truth,” “true by virtue of meanings alone,” and so
forth, were all defined, if at all, in terms of one another, holding hands in one
suspiciously small circle. This convenient circularity raised the worry that phi-
losophers were deluding themselves into supposing that the claims made for
the analytic/synthetic distinction were true and that the a priori epistemology
based on those claims was making real progress on the nature of knowledge,
representation, learning, and perception.

Having raised this suspicion, Quine then went on to argue that in fact, there
was no respectable principle for sorting truths in general into distinct analytic
and synthetic bins. The root of the problem, Quine saw, was that what we be-
lieve about a phenomenon is not neatly separable from the meanings of the
words we use in describing the phenomenon. For example, there seems to be no
clean line between the generalizations we believe are factually true of penicillin
and what we mean by the word “penicillin.” Ditto for “electrons,” “DNA,”
“gravitational fields,” “emotions,” and “memory.” So-called “conceptual
necessities” are just firm — sometimes very firm — convictions, rather than fun-
damental truths about the nature of reality as revealed by pure reason. As
mere convictions, they were no more instruments of progress than any other
conviction.

Cannot one just stipulate a boundary between what we believe about a phe-
nomenon and what we believe about the meaning of the word for the phenome-
non? Then the analytic/ synthetic story could be salvaged. That strategy is
ultimately as useless in propping up the analytic/ synthetic distinction as stip-
ulating that the Earth does not move on grounds that what we mean by
“Earth” is “thing that does not move.” Such a stipulation merely registers

267

An Introduction to Epistemology

someone’s decision to hold some sentences — the so-called analytic ones — to be
true no matter what the evidence. Such stubbornness is indefensible in view of
the proclivity of science for discovering very surprising things and overturning
very deep convictions. For example, scientists discovered that Earth moves
around the Sun, even though obvious observations powerfully suggest other-
wise. The issue of whether Earth does move around the Sun cannot be settled
by the claim, even if correct, that part of the very meaning of “Earth” is “thing
that does not move.” Such facts about meaning notwithstanding. Earth does
move around the Sun.

Quine’s point was that you cannot predict with any certainty how the evi-
dence might eventually go in the future, especially as science discovers surpris-
ing new things and develops revolutionary new theories. Declaring some
sentences unfalsifiable by any evidence either thwarts progress in science or else
is futile baying at the moon. Who can predict what evidence science might
uncover, or what meaning change might seem most reasonable in the light
of revolutionary discoveries.

When people believed — sincerely, fervently, and with complete conviction —
that atoms were indivisible, it seemed to them that the sentence “Atoms are
indivisible” was unfalsihable. Part of the original meaning of “atom” is “indi-
visible basic thing,” as reflected by its Greek origins. Once the atom was split,
however, physicists said, “Wow! I guess the atom is divisible after all.” Here is
what physicists did not say: “Well, since the sentence ‘Atoms are indivisible’ is
unfalsihable, we must use a diflerent word for the thing that was split in the
cyclotron. Let’s call it a ‘ratom.’ Ratoms are divisible, but atoms, obviously,
are not.” That would have been idiotic, not to say a waste of time. Worse, it
would have squandered an opportunity to see the profound ramifications of a
brilliant factual discovery.

The trouble with the analytic/synthetic distinction, as the logical empiricists
hoped to use it in epistemology and elsewhere, is that the history of science has
many examples of sentences that people were convinced were necessarily, con-
ceptually true — true no matter what the evidence — but that eventually turned
out to be false. For example, “The interior angles of a triangle add up to ex-
actly 180°” and “Parallel lines can never meet” are two sentences considered
necessarily true by Kant and many philosophers thereafter. Some analytic phi-
losophers insisted that it was part of the very meaning of “parallel” that par-
allel lines cannot ever converge, and hence that all empirical evidence is
irrelevant. Period. This intransigence struck Quine as nonscientific and as a sign

268

Epistemology

that so-called necessary truths and eonceptual truths were phony classifications
invoked to sustain a phony program.

Kant was convinced that it was necessarily true that parallel lines never
converge in space. He could not have known that Einstein would come along
and propose the general theory of relativity, according to which huge concen-
trations of mass in space yield a non-Euclidean metric in the space. Near a
black hole, for example, parallel lines may well converge. Some philosophers
have said, “Well, okay, but then the meaning of ‘parallel’ has changed. Ac-
cording to what Kant meant by ‘parallel,’ parallel lines in space do not con-
verge.” This is convoluted, to put it politely. Why not simply say that certain
widespread, highly probable beliefs about space, large gravitational fields, and
straight lines have turned out to be false?

Basically, Quine’s challenge is this: if you want to make the analytic/syn-
thetic distinction do a priori work in epistemology, at a minimum you owe us
a theory that distinguishes changes in meaning from changes in belief The
distinction will have to draw on a theory of meaning, and that theory will have
to have empirical support from psycholinguistics and neuroscience, not just
nonempirical invocations of alleged “conceptual truths.” Additionally, it will
have to work for the tough cases, not just the easy ones, since oodles of dud
theories can explain the easy cases.

There are essentially two ways of doing this: (a) declare someone, for exam-
ple me, as Meaning Potentate, with the power to legislate when meaning
changes versus when belief changes, or (b) base the theory on actual linguistic
practice and scientific development. The first is an example of a dud theory,
and no more need be said. As for the second, real examples typically bear out
Quine’s contention that no sentences are immune to empirically motivated re-
vision, and that meaning is not cleanly separable from unwavering conviction.
All of which means that no sentence gets counted as necessarily true, concep-
tually true, or analytically true.

Responses to Quine flooded the journals. Many of the replies boiled down to
variations on the theme that philosophers’ belief in and use of the analytic/
synthetic distinction can only be explained by its truth. Not surprisingly, Quine
had foreseen this move. He suggested that a variety of language-based theories
could explain why the distinction may seem to hold, but really does not. Al-
though Quine’s arguments came as close to a rout as anything can in this field,
few philosophers took his conclusions seriously enough to change they way
they did business. Most tried to find a way to kick up a bit of sand and then
carry on as though nothing had happened. This brief account still leaves

269

An Introduction to Epistemology

much to be said on the matter of meaning, but as meaning will be a topic of
chapter 8, further issues will be raised there.

4.3 Normative Epistemology and Making Tools

Before ending this chapter, I should note that there is a second, dilferent stream
of contemporary epistemology. It survives for completely different reasons. Es-
sentially, this branch survives because it has been successful in making progress
and in serving the day-to-day needs of contemporary science. This is the tool-
making branch of epistemology, and the tools are used for the analysis of
data — for suggesting causal theories to explain a large body of observations,
for determining what causally significant factors have what degree of impor-
tance in the production of a phenomenon, for evaluating the statistical signifi-
cance of outcomes and strength of evidence supporting an hypothesis, and so
on. This is analysis in its more customary sense. This subfield of epistemology
made productive use of modern logic and modern mathematics, and developed
an array of new evaluative techniques.

Many thinkers, including Aristotle (384-322 b.c.), Bacon (1214-1293), and
Pascal (1623-1662), addressed some version of the question. What kind of
procedures can help us make good progress in understanding the actual nature
of the world? What sorts of methods are reliable, in the sense that if we use
those methods, we will end up with theories that get us closer to the truth about
how things really are? This subfield of epistemology has been remarkably pro-
ductive, particularly from the middle of the nineteenth century. It includes
efforts to establish a sound and sensible mathematics of probability and to
characterize what sort of evidence is needed to support causal judgments.

It includes the development of statistical methods for analyzing data and
clarifying how we may best interpret the results of an experiment. It includes
developments in what are now called “game theory” and “decision theory.” It
includes trying to understand what computation is. The overlap here is with
mathematics, mathematical logic, and scientific methodology.^^ Ironically, be-
cause powerful technical results have been achieved, results that are broadly
used in all sciences, this branch is sometimes considered to have branched off
from “proper” philosophy. (“That is not what is meant by ‘philosophy’ ” is
one way to make a self-sealing claim by invoking the now-abjured analytic/
synthetic distinction.)

Whether or not to use the label “philosophy” for this subfield of epistemol-
ogy is less a substantive matter and more a pointless exercise in linguistic leg-

270

Epistemology

islation. Do we stop calling a topic “philosophy” as soon as its practitioners get
good technical results? Logicians and philosophers of science will say no, while
some who favor a “no progress” model of what counts as philosophy will say
yes.

5 Toward a Naturalized Epistemology

By the end of the twentieth century, the hope that much could be learned about
the nature of knowledge from the logical empiricist or analytical approach had
begun to dim. Naturalism — taking relevant empirical data into account when
theorizing — has at last acquired respectability within philosophy, though the a
priori tradition remains powerful under the rubric of “analytic philosophy.”

The convictions of a priori philosophers notwithstanding, progress in empir-
ical psychology and neuroscience continues to narrow the gap between tradi-
tional philosophical questions about knowledge and empirical strategies for
exploring how brains learn, remember, reason, perceive, and think. The time
is right for neuroepistemology. Because cognitive neuroscientists are in fact
addressing large-scale (philosophical) as well as small-scale questions, neuro-
epistemology has its feet well planted, whether or not they be planted in phi-
losophy departments.

As a bridge discipline, neuroepistemology is the study of how brains repre-
sent the world, how a brain’s representational scheme can learn, and what rep-
resentations and information in nervous systems amount to anyhow. This
characterization must be seen as provisional, however, for it is too early in the
game to be very confident that “representing reality” is the right way to de-
scribe the central function of the mind-brain.

In the next two chapters, we shall look at two main issues in epistemology as
they can be considered from within a biological framework. The first concerns
whether the notion of representation is needed at all, and if so, how best
to understand what representations are and how they relate to whatever it is
they represent. The second concerns learning, and in general, the question of
adaptation — via evolution as well as via experience-dependent changes in the
nervous system. This is the problem of “how meat knows,” to put it in its
starkest guise. Many of the traditional epistemological foci — skepticism,
innateness of knowledge, foundations for knowledge structures — can be use-
fully reconsidered in the neurobiological context. Other epistemological
questions — when are we are justified in believing something; how do we estab-

271

An Introduction to Epistemology

lish that something is true or probably true; how do we falsify a belief — are
now more integrated with decision theory and statistics. Will there be overlap
between this technical domain of philosophy and cognitive neuroscience on
such topics as reasoning, inductive or otherwise? My guess is yes.

One important line in the history of epistemology is the struggle to under-
stand representations and how they relate to reality: the degree to which the
one resembles the other, or is caused by the other, or is independent of the
other, or yields knowledge of the other. Especially since Kant, an important
question is how much the brain itself contributes to the character of what is
represented. This question ushers in a problem: if the brain contributes to the
character of what is represented, how can we, with our brains, separate out
what in our representations corresponds to the world and what the brain con-
tributes? If brain organization dictates the general form of experience, what do
we actually know about the real world? I regard these as problems not for pre-
Darwinian, a priori epistemology, but for post-Darwinian neuroepistemology.

Suggested Readings

De Waal, Frans. 1996. Good Matured. Cambridge: Flarvard University Press.

Gibson, Roger. 1982. The Philosophy of W. V. Quine. Tampa: University Presses of
Florida.

Glymour, Clark. 1997. Thinking Things Through. Cambridge: MIT Press.

Glymour, Clark. 2001. The Mind’s Arrows: Bayes Nets and Graphical Causal Models in
Psychology. Cambridge: MIT Press.

Medawar, Peter. 1984. The Limits of Science. Oxford: Oxford University Press.

Panksepp, Jaak, and Jules B. Panksepp. 2000. The seven sins of evolutionary psychol-
ogy. Evolution and Cognition 6: 108-131.

Panksepp, Jaak, and Jules B. Panksepp. 2001. A synopsis of “The seven sins of evolu-
tionary psychology.” Evolution and Cognition 7: 2-5.

Quine, W. V. O. 1960. Word and Object. Cambridge: MIT Press.

Quine, W. V. O. 1969. Epistemology naturalized. In his Ontological Relativity and Other
Essays. New York: Columbia University Press.

272

Epistemology

Websites

BioMedNet Magazine: http://news.bmn.com/magazine
A Brief Introduetion to the Brain: http://ifesunl.ifisiol.unam.mx/brain/
Encyclopedia of Life Sciences: http://www.els.net

The MIT Encyclopedia of the Cognitive Seienees: http://eognet.mit.edu/MITECS

How Do Brains Represent?

We have to remember that what we observe is not nature herself but nature exposed to our
method of questioning.

Werner Heisenberg

1 Introduction

However brains work, much of what they do involves representing — representing
the brain’s body, features of the world, and some events in the brain itself.
Performing computational operations on those representations serves to extract
relevant information, make decisions, remember, and move appropriately.
That brains represent and compute are working assumptions in much of cog-
nitive neuroscience. I emphasize that these are indeed assumptions, however,
not hrmly established truths. As science continues to progress, the assumptions
may be amplified, revised, or even falsified in favor of better hypotheses, as yet
dimly conceived.

The problem of the nature of representations has been attacked from two
directions. One is consilient with the neurosciences generally, and is coevolving
with them. For convenience, call this the brain-friendly approach.^ It aims to
discover how brains map and model the world by looking at all levels of orga-
nization from neurons to behavior. The second approach is wedded to the
analogy between cognitive operations and software running on a computer (see
chapter 1), and hence adheres to the autonomy of psychology. Call this the
brain-averse approach.^ Because it assumes that cognitive states are a function
of their role in the cognitive economy, and hence are independent of any par-
ticular hardware “implementation,” the brain-averse approach ignores neuro-
science as largely irrelevant to the problem of how the mind represents.

274

Epistemology

Recall that the autonomy-of-psychology thesis assumes that neuroscience
can at most reveal something about the implementation of the cognitive soft-
ware but cannot aspire to revealing anything much about the nature of cog-
nitive processes per se (see pp. 25-28). More extreme versions consider
neuroscience an actual impediment to progress. The reason is that neuro-
scientists routinely ascribe representational functions to neural structures, say-
ing such things as “Neural networks in the superior colliculus represent eye
position.” Comments like this one are alleged to be worse than confused, since
hardware allegedly does not represent anything. As we shall see, brain-averse
adherents take a linguistic entity — the sentence — as the prototype for all real
representations. Consequently, those animals that lack the capacity for lan-
guage are considered to have representations only by courtesy or as a figure of
speech, but not literally.

There is also a third approach, which is really a variant of the first. It ex-
plores the possibility that in some contexts representation-based explanations
for behavior may be replaceable, or at least augmented, by the concepts in the
framework of dynamical systems.^ Like the first, this approach is coevolving
with the neurosciences. It is motivated in part by the fact that nervous systems
are indeed dynamical systems and in part by the fact that there is no well-
developed theory about what exactly information and representation in biologi-
cal systems amounts to. Given this dearth of theory, the dynamical-systems
approach undertakes to explore how much explanatory ground can be covered
without appeal to representations, and how representational accounts, when
needed, can best be integrated within a dynamical-systems framework.'^

As we shall see, there is compelling support for the hypothesis that animals
other than humans have representational capacities, and that brains are the
platform for those representational capacities. In my opinion, therefore, the
brain-friendly approaches are likely to be more productive than brain-averse
approaches. As noted in chapter 2, the brain-friendly approaches are also more
appealing on sheerly pragmatic grounds, since they consider all data, not just a
subclass of ideologically approved data. Additionally, as we shall see, there are
certain fundamental kinds of representation, such as spatial representation,
where cognitive neuroscience is making impressive progress, but where brain-
averse approaches are relatively unrewarding.

From an evolutionary perspective, brains are buffers against environmental
stress and variability.^ Early in our history, evolution must have stumbled
upon the advantages accruing to nervous systems able to make predictions
based on past conditions, while evaluating current circumstances, both internal

275

How Do Brains Represent?

and external. In short, it helps to have a brain with the capacity to prepare for
events that will probably happen, and to organize a behavioral response to
something that is not now happening but can be expected to happen. If you are
a tree, in contrast, there is no advantage to knowing anything about where
home is; you have no options about how to hide from a predator or stalk prey,
or about which mate to choose. You take what comes. Variations in the
weather may affect you, but seeking shelter is not in your repertoire. Animals,
however, can move, and hence knowing things and representing things confers
a competitive advantage. The hypothesis on offer is that in the service of pre-
diction, neuronal activity maps the various me-relevant features of the world —
its spatial relations, social relations, food sources, shelters, and so forth. This
mapping is usefully considered representational. The next question is how ex-
actly neuronal structures accomplish this.

At the neuronal level, a major question driving the field asks precisely what
features of neuronal activity subserve the coding of information — both encod-
ing and decoding. Neurons exhibit a wide range of activity, and a theory of
representing needs to address the problem of how to distinguish genuine signals
from housekeeping activities and from mere noise.

At the network level, the predominant aim has been to find plausible models
that will mesh with the facts about neurons and their connectivity patterns, and
with psychophysical data deriving from behavioral studies. The hope is that
network models will be a bridge between what we understand about bodily
behavior and what we understand about neurons.

At the systems level, a major challenge is to understand how nervous systems
integrate information, store information, retrieve task-relevant information and
use information to make behavioral decisions. That the nervous system is a
complex dynamical system is evident, but figuring out the dynamical principles
by which it functions continues to be difficult.®

2 Do Brains Represent?

What motivates saying that the brain represents at all! Why couldn’t the story
of brain function be told without reference at all to representations and com-
putations? (Incidentally, the same discussion would follow whether, instead
of the word “representation,” we used the word “idea,” as Hume preferred, or
“thought” [Descartes], or “concepts” [Kant]. “Representation” is just the term
currently in fashion.) Although answers can be constructed from different

276

Epistemology

TRAINING

TESTING

Figure 7.1 Spatial representation in the rat. In the training condition (left), the rat
always starts in the same location and learns that food is always in the end of left arm of
the maze. In the test condition (right), a block is removed and the rat is placed in the
newly opened passage. Rats trained for a normal amount will turn right, correcting for
their reversed spatial orientation with respect to the food location. Overtrained rats or
rats with hippocampal lesions will turn left. (From Farber, Peterman, and Churchland
2001; based on Packard and Feather 1998b.)

starting points, the most compelling arguments begin with examples of cog-
nitive operations that resist explanation on the stimulus-and-response para-
digm. Competence in solving spatial problems will illustrate this explanatory
resistance.

Consider a well-controlled and revealing set of experiments by Packard and
Feather illustrating that a stimulus-response explanation predicts one result, a
representational explanation predicts the opposite, and the latter prediction
wins.’ Put a rat in a T maze for twenty trials, and always bait the left-hand
arm. After a few trials, the rat knows where the cheese is (hgure 7.1). Next,
unblock a barrier at the top, and put the rat in the full maze. If the rat has
merely acquired a conditioned response to turn to the left, then when con-
fronted with the full maze, rats should turn left. If the rat has a spatial map for
this environment, it will turn right. What do the rats do? They turn right, thus
turning in the direction opposite to their earlier responses. This behavior
implies that the animal is not displaying a conditioned response (always turn
left), but rather is using a representation of the spatial organization of the maze
relative to the room the maze is in. Additionally, to relate the rat’s behavior to
a specific brain structure, Packard and Feather showed that if the brain’s hip-

277

How Do Brains Represent?

pocampus is lesioned, either before or after the training trials, the animals do
show a conditioned response in the test condition: they turn left. On the other
hand, intact but overtrained rats {hundreds of training trials to the left-baited
arm) will turn left when put in the full maze almost as though the conditioned
response overrides the spatial reasoning when the animal is overtrained (and no
longer “thinking”?). Lesions again prove revealing, for when Packard and
Teather lesioned the striatum, the conditioned response is abolished. Given
striatum lesions, even the overtrained rats now turn right, which suggests that
when the conditioning circuits are unavailable, the brain again relies on spatial
representations.

Additional physiological data about the hippocampus and spatial mapping
makes a more compelling case. O’Keefe and Dostrovsky discovered in 1971
that there are place neurons in the hippocampus of rats. That is, an individual
hippocampal neuron fires when and only when the freely moving rat is in a
specific place, such as the upper east corner of the maze (see figure 7.2). Fol-
lowing the pioneer work of O’Keefe and his students, others went on to repli-
cate and extend their results. It was discovered that an individual neuron codes
for a place relative to a particular environment — for example, the kitchen
versus the living room. That is, a given hippocampal neuron may code for a
particular place in the kitchen, and a completely unrelated place in the living
room. As expected, rats with hippocampal lesions cannot learn spatial tasks, as
indeed humans cannot. Other hippocampal neurons that code for direction of
movement in the freely moving rat have also been found.

The above argument for a representational explanation of route-finding be-
havior in rats has assumed that the only alternative to a conditioned-response
explanation is a representational explanation. Although that assumption might
be false, the fact is we really do not have any other plausible options at this
stage, though theoretical developments could change that.

Many animals, including dogs, horses, bees, and bears, exhibit behavior that
shows good spatial representation, such as finding novel routes home. Because
many animals can aim for home without benefit of conditioning, and fre-
quently without much in the way of trial-and-error searching, it seems fair to
say that they “know where home is,” or better, perhaps, they “know how to get
home” in some sense of “know.” Nevertheless, describing the matter thus may
imply more than is intended, namely, that the animal says to itself “Hey, I
know where the cheese is: just over yonder, two left turns from Maggie’s nest.”
Nothing quite so humanlike as self-talk is really implied, however, if only

278

Epistemology

50 cm

( ]

— )

¥ —

■( )

( i

( ]

]

( ]

'( ]

( I

( )

( ]

( T

( ]

( i

( )

( ]

t ]

( J

( )

1 )

1 J

r J

1 J

( )
( )

( ]

( )

]

( ]
( ]

( ]

i:(=)

•( ]

^ J

( ]

( . ]

( •)

( 1

( )

Figure 7.2 Firing-rate maps of 25 hippocampal place cells simultaneously recorded in a
rat running on the elevated track of a figure-8 maze. Restricted regions with high firing
rates are called place fields. Thus the neuron whose response profile is depicted in the
top left box is tuned to respond when the rat is located in the upper rightish region of the
maze. Maps were computed from 7 minutes of continuous data. In each plot, scaling is
linear, with a 0 firing value corresponding to 0 in the color map and a maximum positive
value corresponding to 1 . (Courtesy of K. Zhang.)

279

How Do Brains Represent?

because these animals do not appear to have a humanlike language in which to
talk to themselves. Indeed, it is doubtful that human spatial representation is
generally languagelike, even though spatial knowledge can sometimes be pre-
sented in speech.® Consequently, we seek an understanding of representations
that does not depend on their being just like words and sentences.

Representing spatial relations is one kind cognitive function, but we com-
monly consider that the brain represents objects, such as that barking dog.
When humans plan, imagine, and dream, they represent in the absence of the
thing they are representing. You can think or dream about skiing down the
black-diamond runs at Whistler. Even if you have no skis, or even no legs, you
can still produce visual/motor images of skiing down the slopes. Representa-
tions are also typically invoked to explain how the brain perceives, that is,
how the brain makes a perceptual judgment when the relevant object is in
full and unobstructed view, in full odor, within earshot, or on the tongue.
Why, when the relevant scene is present, do neuroscientists say that the brain
represents?

Part of the reason is that nervous systems do a lot of processing of signals
received at the periphery — at the retina, skin, cochlea, nose, tongue, muscles,
tendons, and joints. Information is extracted, augmented, integrated, and gen-
erally worked so that the perceptual product we are aware of, such as the smell
of a wet dog, is really a far cry from the peripheral signals. The brain is not a
passive reflector of external stimuli; it is in various ways an active constructor
that builds the animal’s perceptual-motor world.® Many of the same mecha-
nisms are probably recruited in both visual perception and visual imagery, in
both auditory perception and auditory imagery, and in both motor control
and motor imagery. A chain of visual examples can illustrate this constructive
dimension.

2.1 Contours

In the figure 7.3, devised in 1955 by psychologist Georg Kaniza, we see a white
triangle overlying three lines and three black circles. In fact, there are no
boundaries demarcating the white triangle; the black areas are actually pac-
man shaped disk segments. A sensitive photometer run over the figure will de-
tect neither the borders of the triangles nor the increased brightness of the white
triangle. Subjective contours, as they are called, are also seen in plate 6.

Plate 6 contains red intersecting lines on the left and, on the right, the very
same red intersecting lines with black extensions. Nevertheless, on the right.

280

Epistemology

Figure 7.3 Illusory contours. You see an illusory white triangle on a background of
partly occluded circles and lines. The interior of the triangle generally appears whiter
than the ground, even though it is not. (From Palmer 1999.)

but not on the left, we also see a contour demarcating a light red disk from the
background.

Plate 7 is a stereo display consisting of two images, slightly olTset, that can be
fused to make a single image of a blue semitransparent rectangle hovering over
four circles. You can fuse the two images by defocusing your eyes, which is
easiest to accomplish by looking at the display as though it were further away
than it is. Devised by Ken Nakayama and Shinsuke Shimojo, this illusion is
especially striking because the only blue detectable by a photometer consists of
little blue arcs on the concentric circles. You will also notice that the filmy blue
rectangle curves out slightly toward you.

2.2 Ambiguous Figures

Some stimuli happen to permit two equally good interpretations. The classical
case is the Necker cube, investigated by Swiss psychologist Louis Albert
Necker in 1832. It can be seen either as having its face oriented up and to the
right, or oriented down and to the left (see figure 7.4, B).

2.3 Motion

Your brain will create a perception of motion in the following condition. A light
flashes off at one location, and within a very brief time period, a light flashes on
at a new but nearby location (figure 7.5). It looks as though a light moved
across space from the old to the new location.^®

281

How Do Brains Represent?

Figure 7.4 Ambiguous figures. (A) can be seen either as a white vase against a black
background or as a pair of black faces against a white background (the disambiguated
figures can be seen to the right). (B) can be seen as a cube viewed from above or below.
(C) can be seen as a duck (facing left) or a rabbit (facing right). (From Palmer 1999.)

O X

Figure 7.5 Subjective motion. A light flashes on first at the position marked O, and
then at the position marked X. If the time interval between when the light at O goes off
and the light at X comes on is between about 5 and 500 milliseconds, then what is seen is
a light continuously moving from position O to position X.

282

Epistemology

Figure 7.6 Visual completion behind partly occluding objects. Panel A is perceived as
consisting of a square, a circle, and a rectangle, even though the only visible regions are
those shown separated in panel B. (From Palmer 1999.)

2.4 Scene Segmentation

In A of figure 7.6 we typically see the bar as behind the circular disk, which is
partially occluded by the square. Nevertheless, the only regions that reflect light
to the retina are the partial objects shown below in B.

These examples on their own do not provide anything like a theory of the
nature of representation. Their function is merely to give us a feel for the kinds
of job that representations are thought to perform in perception. They show the
place in the explanatory scheme of things where representations seem, for now
anyhow, to be needed. For this reason, these and other examples provide a
modicum of guidance for the investigation. Assuming, therefore, that nervous
systems do indeed represent, we now need to ask how they represent. Ulti-
mately, we want a theory of representation in nervous systems. As a ground-
clearing preliminary, we can first ask what facts from biology, neuroscience,
and psychology constrain a theory of nervous-system representation.

283

How Do Brains Represent?

3 Some Empirical Constraints on a Theory of Representation

In the early stages of this enterprise, articulating constraints is desirable, since
knowing what will not work narrows the search space. That the human brain is
an evolved organ is, not surprisingly, an overarching constraint on any theory
of representation. Proposals that construe representations as literal pictures in
the head, to take an unlikely example, are not very promising as a starting
place.

3.1 Brains Are Products of Evolution

Because the human brain is the product of evolution, we can expect to find
instructive similarities and continuities between representing in infants and
adults, and between representing in human and nonhuman animals. A theory
that predicts failure of any continuity between humans and other animals in,
say, spatial or motor or perceptual representation would raise a red flag. A
theory that entails a magical origin for complex human representational
capacities will raise a red flag. More generally, in science any hypothesis that
fills a gap by saying “and then a miracle happened” is not compelling.

3.2 Represeutatious aud Language

In humans, language is an important means of communication, and fully ver-
bal humans often use language in thinking. Ultimately, any theory of repre-
sentations will need also to account for human languages, including the
capacity of children to learn language and the range of distinct language defi-
cits seen in humans with brain damage. Even if human language is unique in its
complexity and representational power, general considerations from evolution-
ary biology imply that a theory postulating the absence of any continuity be-
tween linguistic representations and nonlinguistic representations would need
skeptical scrutiny. “

A characteristic of humans, one that tends to amplify the differences that do
exist between humans and other animals, is cultural evolution. In humans,
language, along with cultural institutions, allow later generations to start learn-
ing what an earlier generation invented in its maturity. Accordingly, representa-
tional skills emerge that are strikingly novel relative to the ancestral versions.
Conventions, manners, and ideas newly created by the ancestral generation can

284

Epistemology

seem second nature, logically obvious, and even biologically necessary to the
grandchildren who learned them as just part of how the world is.

Unfortunately, we know little about the conceptual resources deployed by
hominids 3 million or even 100,000 years ago. Nevertheless, it is sobering to
remind ourselves of the many cognitive artifacts that are known to be cultural
inventions: reading, writing, mathematics (including the number zero), music,
and maps; the use of fire and metals; and the domestication of dogs, sheep,
wheat, and rice. We do not know how much of the complexity seen in human
language depends on cultural evolution. Structural similarities among human
languages are consistent with, but certainly do not entail, that there exists a
genetically specified grammar module in the human brain. Such structural
commonalities as do exist could be as well explained, so far as is known, as
arising from similarities in nonlinguistic representational resources and sim-
ilarities in fundamental aspects of human experience, such as spatiality, socia-
bility, the need for sequence assembling in forming plans and in behavioral
execution, and so forth. As Elizabeth Bates wryly commented, the similarities
among humans in getting food to the mouth by using hands rather than feet
does not imply the existence of an innate “hands for feeding” module. Rather,
the existence of a shared body plan and the ease of hand feeding relative to foot
feeding suffice to explain “feeding universal.”

3.3 Contrasting Digital Computers and Brains

Although we describe them as computing, brains, in important respects, are
profoundly unlike the familiar digital computer. For example, the spatial
knowledge of rats is not stored in their “hard drives,” because a rat brain does
not have a hard drive. Computers have a memory module independent of the
structures that process information, but nervous systems do not. More gener-
ally, our brains do not have modules in the way our desk-top computers have
modules. Brains do, however, exhibit areas of functional specialization, es-
pecially at maturity, but the specialization exists with a degree of functional
modifiability that is not at all compatible with the idea of “ecapsulated, dedi-
cated modules.” A suitably neurobiological sense of module has yet to be
characterized in detail, since much about brain organization, function, and de-
velopment, from conceptus to corpse, is still not understood. Incidentally,
inventing a new expression to replace “module” might help us avoid implicitly
importing to the neural domain the standard features of modularity in the
computer domain.

285

How Do Brains Represent?

Other dissimilarities between brains and computers want mentioning:

■ Neurons, unlike computer chips, grow and develop, or prune back or die. At
least in the hippocampus and perhaps elsewhere, new neurons are generated
even into adulthood.

■ Neurons are dynamical entities, and they change structurally as they learn,
making new contacts, abandoning old contacts, strengthening or weakening
existing contacts, and so on.

■ Changes in neuronal structure often require antecedent changes in gene ex-
pression, and certain genes are turned on as a consequence of the level of
certain activities in the neuron.

■ Neuronal events happen in the millisecond range; events in present-day
computers may be four or five orders of magnitude faster.

■ Nervous systems have a parallel organization; computers are serial
machines.

■ Computers have a clock that sets now for all components; brains, so far as
we can tell, do not have a clock that serves that function.

■ Computers were designed by humans to crunch numbers; nervous systems
evolved through natural selection to move bodies adaptively. The former
is nonsemantic or clean computation; the latter is life-oriented, dirty
computation.

■ Not all the 10*^ neurons in the human nervous system are in the representing
business. Some, for example, modulate the activity of representing neurons,
because they are part of the arousal or attentional systems, or they perform
functions we do not yet understand. Others may have a causal role in regu-
lating temperature, heart rate, growth, or appetite, without actually repre-
senting any of those things. Since human engineers designed computers, it
is relatively straightforward to determine what activities in computers are
information-bearing and what not. For brains, however, all of that has to be
figured out. You cannot tell by just looking.

4 Coding in Neurons and Networks: A First Pass

Neurons transmit information by virtue of their activity and are believed to
store information by changing aspects of their connectivity to other neurons.
The prototypical transmission is point-to-point; that is, from a site on the

286

Epistemology

Dendrites

spike initated

Figure 7.7 A summary diagram showing the location on a motor neuron of various
electrical events. In many neurons, dendrites and cell bodies respond with graded exci-
tatory postsynaptic potentials (EPSPs) or inhibitory postsynaptic potentials (IPSPs).
The action potential is triggered in the axon hillock and travels undiminished down the
axon. (Based on Thompson 1967.)

sending (presynaptic) neuron to a site on the receiving {postsynaptic) neuron.
In the classical paradigm, signals are received in the dendrites and cell bodies,
and signals are sent via the axon to the axonal terminal (figure 7.7). If, as a
function of complex interactions in the dendrites and cell body, a sufficiently
strong depolarization reaches the axon hillock, then a spike is generated and
propagated down the axon to its terminus. With some degree of probability, the
presynaptic membrane may release neurotransmitter into the synaptic cleft af-
ter the arrival of the spike. The probability of transmitter release (also called
the synapse’s reliability) may change as a function of learning.

In classical point-to-point signal transmission, neurotransmitter molecules
bind to specialized sites on the postsynaptic cell. These sites are really complex
protein molecules than span the neuronal membrane and change their shape
when bound by a ligand. This change in shape can allow an influx of positive

287

How Do Brains Represent?

ions, which will depolarize the postsynaptic membrane, thus exciting the post-
synaptic cell. Or, depending on the type of protein channel, it might prevent the
eflux of negative ions, which will cause the neuron to hyperpolarize. (See hg-
ures 1.5 and 1.6.) Communication between neurons is achieved when the neu-
rotransmitter released from the presynaptic cell affects the postsynaptic cell by
either exciting or inhibiting it. Many thousands of postsynaptic sites on a neu-
ron’s dendrites may respond with a depolarization or a hyperpolarization
within a few hundred milliseconds. Typically, many stimulus inputs are needed
to generate a current strong enough to initiate spiking at the axon hillock. Ad-
ditionally, there are other styles of signaling, some of which occur at non-
conventional receptor sites on the receiving (postsynaptic) neuron. Within the
last decade, it has emerged that communication in the nervous system spans
a continuum of speeds, effect durations, and postsynaptic cascades. These are
not just curious exceptions to the old and mainly true story. Rather, they are
central elements that profoundly rewrite the classical story.

By recording from individual sensory neurons during the presentation of a
stimulus, it has been found that many neurons display a response selectivity
when the animal is presented with specific external physical parameters, such as
vertical motion of an object (neurons in the visual cortex), or light touch on the
thumb (neurons in the somatosensory cortex), or the smell of peppermint
(neurons in the olfactory bulb). A neuron’s response specificity is often referred
to as its tuning, and hence a neuron is said to be tuned to visual motion or to
peppermint. In a casual manner, we also say that the neuron prefers pepper-
mint, or is driven by peppermint.

The receptive field of a neuron is the area on the receptor sheet (retina, skin,
etc.) that, when stimulated, causes the neuron to respond. For example, the
receptive field of a neuron in the somatosensory cortex might be a tiny region
on the tip of the thumb; the receptive held of a neuron in the visual system
might be a particular spot on the fovea of the retina. A neuron may have a
small receptive held but be broadly tuned, as is typical of neurons in the pri-
mary sensory cortexes. Such a neuron may respond maximally to a bar of light
moving vertically, respond slightly less if the light is moving somewhat off the
vertical, and respond less and less as the direction of movement of the light
converges on the horizontal.

On the other hand, a neuron in a higher^"'' visual area, such as the infe-
rior temporal region, may have a large receptive field spanning much of the
entire visual held and yet be narrowly tuned to respond only to faces, or even
more narrowly, only to one individual face albeit in many orientations. The

288

Epistemology

Figure 7.8 Projective and receptive fields. (A) A single receptor projects to many gan-
glion cells (via interneurons) in a center/surround organization. The center ganglion cell
is excited; the surrounding cells are inhibited. Thus the projective field of a receptor is
characteized. (B) Each ganglion cell receiving such connections therefore has a center-
surround receptive field. The illustrated network exhibits an excitatory center and inhib-
itory surround, but the opposite organization (inhibitory center and excitatory surround)
also exists in the retina. (From Palmer 1999.)

nonclassical receptive field refers to that region around the classical receptive
field that can modify the response to stimuli within the classical receptive field.
Stimuli restricted to the nonclassical region do not, however, drive the cell on
their own. The projective field refers to the set of neurons to which a given
neuron projects (figure 7.8).

Two problems need to be distinguished: (1) What properties in the single
neuron carry information? And (2) how is an objective parameter represented
by neurons? Traditionally, the dominant hypothesis offered to the first problem
has been rate coding-, that is, the average firing rate or spiking frequency of the
neuronal axon over a certain interval is what carries information. Although
rate coding is one strategy for carrying information, nervous systems probably
employ other strategies as well. The list of other possibilities include the timing
of a spike burst relative to the timing of other neuronal events, the interval
between spikes in one neuron, the specific pattern of spikes in an interval,
and the latency for the first spike after the stimulus. There may also be other
information-bearing neuronal changes that do not involve axonal firing at all.
Dendrites, as they receive signals from the presynaptic cell, undergo membrane
changes, and these changes must carry information. If they did not, the spike
generated from integrated inputs would not carry information either. Dendritic
responses are standardly construed as decoding the incoming signals, so pre-

289

How Do Brains Represent?

sumably the states of dendrites are themselves information-bearing. How does
that work? We are now beyond the well-trodden ground of the conventional
wisdom.

For the second problem (how is an objective parameter represented by neu-
rons?) at least two hypotheses command attention, (a) A property, for example,
the face of Woody Allen, may be coded by a single neuron, and this neuron
normally fires when and only when a Woody Allen face is presented {local
coding). To avoid losing its entire Woody Allen representation when one neu-
ron dies, the system may have spares; that is, it may have a pool of neurons
that all respond when and only when the face of Woody Allen is visually pre-
sented. Such redundancy is consistent with local coding. One drawback to local
coding as a general strategy for the nervous system is that there are too few
neurons to account for the huge numbers of things, places, and events we can
recognize. Nevertheless, for certain restricted representational purposes, such as
coding nonoverlapping values in a small range, local coding could be adequate
and efficient.

The second hypothesis, (b), says that some values are coded by a population
of neurons whose members are active in different degrees across a range of
properties {vector coding). As explained below, using vector coding, the brain
could represent the face of Woody Allen with a particular pattern of responses
in the population, and the very same population of neurons, but with a dilfer-
ent pattern of responses, may represent the face of Ghandi and the face of
Castro.

Although some general points can be made about neuronal coding on the
basis of available neurobiological data, many, many questions remain unre-
solved. In particular, neuromodulation, in contrast to classical, fast, point-
to-point transmission, is a major feature in all aspects of representational
function, from sensory input to motor output. Neuromodulation refers to elfects
on a cell’s activity by neurochemicals other than classical neurotransmitters. It
can up-regulate or down-regulate sensitivity of the cell, for example. To make
matters even more interesting, there appears to be modulation of the modu-
lators. Moreover, neurons appear to have a preferred range of activity, and
self-regulating mechanisms kick in once the neuron’s activity is shifted out of
the preferred range.

Emphasizing this in-progress character of neuroscience is crucial, if a little
daunting. At this stage in neuroscience, nothing like a well-established theory
of neuronal coding exists. We will get there, but we are not there yet. In any
case, the range of neurophysiological observations about neuronal responses

290

Epistemology

A Local
coding

B Scalar
coding

C Vector
coding

</)

Figure 7.9 Three methods of encoding information. (A) Local coding: a separate unit
is dedicated to each feature the system distinguishes. (B) Scalar encoding: features are
encoded by the firing rate of a single neuron. (C) Vector coding: features are encoded in
the pattern of activity in a population of units that have broad, overlapping tuning
curves. (From Churchland and Sejnowski 1992.)

under varying conditions is fuel for future theories. Needless to say, consider-
ation of the neurophysiological data is essential, since unconstrained theories
are just guesswork that can be wasteful of time and energy. At the same time,
reaching for theoretical perspective is also essential, even when the data are still
sparse. Explanatory theories do not automatically waft up from the data; they
have to be invented. Moreover, invention of data-inspired theories typically
motivates further experiments, with theory-revisions often following in the
wake of the experiments. Thus the familiar bootstrapping of science generally.

5 Local Coding and Vector Coding: A Fast Sketch

Conceptually, the basic idea of local coding is relatively simple. A neuron (or
pool of response-similar neurons) is dedicated to representing a specific prop-
erty. If a set of neurons were placed cheek to jowl on a one-dimensional grid,
then we could identify a neuron’s unique representational job by identifying its
unique place in the grid. Taken as a whole, the grid might be a 1 : 1 mapping of
locations on a receptor sheet, such as the cochlea of the ear. If the auditory
system used this strategy, then middle C, for example, would be represented
by the activity of the “middle-C neuron,” which would be found at a precise
location on the grid. Local coding is also referred to as place coding.

Vector coding, by contrast, depends on the idea that features are represented
in specific patterns of activity in a population of units, where each neuron has a
tuning curve, perhaps quite broad, and tuning curves overlap, perhaps quite
a lot. This is illustrated in figure 7.9. Mathematically, a vector is simply an
ordered set of numbers, (n \ , « 2 , • • ■ , «m>- The elements in a particular vector are

291

How Do Brains Represent?

values standing for properties such as the activity levels of each neuron in the
relevant population. Let us tentatively make the simplifying assumption that
the contribution to the vector made by a single neuron is its average spiking
rate over a specified interval, say 100 milliseconds. Each neuron in the relevant
population thus contributes some element (— its average firing rate) to the vec-
tor. Depending on the stimulus, the neuron may fire a little, a lot, or below
baseline firing. Accordingly, a particular vector, say <16,4,22>, might repre-
sent the hue yellowy orange, while a slightly different vector, say <16,6, 14>,
might represent the hue reddish orange (see pp. 183-187, plate 5). One and the
same neuron can participate in the representation of many different items (e.g.,
hues), and no one neuron represents a property all by itself.^®

Representation by a population of broadly tuned neurons is economical. For
a given number of neurons, vector coding gives you a larger range of values
than local coding. Suppose that a system has just five neurons, with four dis-
crete activity levels ranging between 0 and 3. If the system uses local coding, the
five neurons can represent 20 different values (4 x 5) . If they use vector coding,
625 values (= 5"*) can be represented. That is, <3, 1,0, 1> specifies a particular
pattern of activity during a certain time interval and represents one value,
<4,2,0, 1> specifies another distinct pattern and a different value, and so on.
Greater precision can be achieved with overlapping tuning curves because more
fine-grained values of a single external stimulus can be reflected in the joint
behavior of the group of cells. Notice that in the limiting case, if a vector has
only one element, vector coding and local coding amount to the same thing.

Each placeholder in the vector specifies a distinct dimension of the parameter
space. When each placeholder is filled with a specific value, the resulting vec-
tor delimits a specific point in the parameter space. Thus a three-element vector
generates a three-dimensional space; a five-element vector generates a five-
dimensional space. The latter is hard to visualize, of course, but think of it
as just more of the same. If the vector codes for neural activity in a 1,000
neuron network, then the parameter space will have 1,000 dimensions. As with
a 5-dimensional space, the mind need not boggle. You can think of this as just
a lot more of the same (figure 7.10).

What is so tremendously useful about spatiality here is that the space (3-D,
10-D or «-D) has a metric, meaning that positions in the space can be specified
as near each other or far from each other or in-between. And spaces admit of
regions, volumes, paths, and mappings. All of this makes it easier to conceptu-
alize representations, relations between representations, and relations between
representations and the world.

292

Epistemology

Figure 7.10 Diagram of a face space to illustrate the idea that faces vary along a
number of dimensions, represented as axes of the state space, and that a system might
code for faces using vectors whose elements represent such features as distance between
the eyes, fullness of the mouth, and width of the nose. Obviously, faces have many
features that are coded by mammals, and even the three features included here are
undoubtedly crude. (Courtesy of P. M. Churchland.)

In the struggle to find useful and coherent ways of thinking about how brains
represent, the vector Iparameter-space tool turns out to be conceptually power-
ful, at least at this early stage. One advantage is that reasonable explanations of
a range of behavioral capacities displayed by representing animals emerge quite
naturally, without ad hoc miracles. In particular, similarity relations, the be-all
of categories and categorial structures, though difficult to address in other
theories, gracefully deliver themselves as neighborhood relations in parameter
spaces. To put it crudely, the problem of similarity relations need not be solved
with hoked-up mechanisms or structures; they are a relatively simple conse-
quence of parameter-space representation. The color space discussed in chapter
4 (see color plates 3 and 5) illustrated the similarity relations at both the per-
ceptual level and the neuronal level, and showed the fit of perceptual space with
neuronal space. A comparable story can be told for tastes (figure 7.11). These
and other examples hint that vector coding in some manner or other is really
what the basic systems use, and that parameter spaces are in fact one repre-
sentational strategy brains exploit.

Two jobs are before us: first, we need a closer look at just how the vector-
coding and parameter-space story goes, and second, we need to see whether

293

How Do Brains Represent?

Figure 7.11 Taste space: the position of some familiar tastes. Similar tasting substances
are found in similar regions of the parameter space. Thus, sugars form a cluster at the
upper middle region, and tart substances are found in the lower rear region. (Based on
Bartoshuk and Beauchamp 1994.)

new insights about concepts and meaning might be discovered with their help. I
shall tackle these tasks in order.

6 Faces: An Artificial Neural Network for Face Recognition

How exactly can neuronal activity represent something? The basic ideas of the
vector/parameter-space approach to representation can be spelled out in a sim-
plified model, much as one can use a simple model to illustrate basic principles
of motion, digestion, or mitochondrial energy production. We shall therefore
begin with an artificial neural network (ANN) that can perform recognition
tasks on photographs of actual human faces. Face net, a three-stage ANN
developed by Garrison Cottrell and his colleagues, is schematically portrayed
in figure 7.12. Although it is not precisely known how nervous systems do in
fact represent faces, the Cottrell network is very useful for demonstrating the
basic principles of how a neural network of units might represent specific
faces. So I will set aside the real details of projection patterns, cell numbers,
cell physiology, and so forth as I outline only the conceptual resources of ANNs.

294

Epistemology

Gender (female)

Gender (male)

Facehood (yes/no) ^ \ \

Person's name

Layer 3:
Detection
(8 cells)

etc.

’ 111

Layer 2:
Face space
(80 cells)

Layer 1 :

Input image

(64 X 64 = 4,096 cells)

Figure 7.12 An artificial neural network for recognizing real faces. The input layer is at
the bottom, the output layer at the top. Although this network has 4,184 processing
units, it has a very simple organization. Each unit in the input layer connects to every
unit in the middle layer, and each unit in the middle layer connects to every unit in the
upper layer. (From P. M. Churchland 1995.)

Face net’s input layer (for our purposes, a pretend retina) is a (64 x 64)-pixel
grid whose elements each admit of 256 different levels of activation or “bright-
ness” according to the light reflected from the region in the photo to which it is
sensitive. The network’s input consists of gray-scaled photographs (figure 7.13).
When initially constructed, of course, the network cannot recognize anything,
and its response to any given input is just random noise. It is then trained on 64
different photographs of 11 different faces, along with 13 photos of nonface
scenes, after which it can perform specific face-recognition tasks. How is this
training achieved?

295

How Do Brains Represent?

Figure 7.13 Selected input images for training the face-recognition network. (Courtesy
of Gary Cottrell.)

Each input unit projects a radiating set of “axonal” end branches to each
and every one of the 80 units in the second layer, and this layer maps an ab-
stract space of 80 dimensions (a dimension for each unit) in which the input
faces are explicitly coded. (A two- or three-dimensional space is readily under-
stood; now just think of adding axes.) The second layer projects to an output
layer of merely eight units. These output units have their connection strengths
carefully adjusted so that the units can make a number of discriminations: first,
discriminating between faces and nonfaces; second, discriminating between
male and female faces; and third, responding with the person’s “name” (actu-
ally an arbitrarily assigned binary code) when re-presented any face that the
network “got to know” during training.

296

Epistemology

What actually does the work in face net is the overall configuration of
“synaptic” connections — positive and negative, weak and strong. It is these,
and only these, that progressively transform the initial (64 x 64)-element pat-
tern or vector into a second and finally a third vector that explicitly represents
the input’s facehood, sex, and name. The fundamental processing format con-
sists of mere vector-to-vector transformations determined by the configuration
of connection weights.

A crucial ambiguity must now be resolved. Sometimes “representation”
refers to cognitive events happening now, such as a visual perception; other
times it refers to the capacity {not now exercised) to have appropriate cognitive
events, such as my capacity to recognize an osprey. Patterns of activity in net-
works hook up with the first sense; configurations of connection weights (which
yield the appropriate patterns of activity when given specific inputs) hook
up with the second sense. Think of the first as displaying knowledge and the
second as the enduring structure, or background conceptual framework, that
makes the current display possible. (Other terms are “occurrent representa-
tions” versus “abeyant representations.”) One caution: because activity can
change structure, as in learning, it is wiser to think of activity and structure as
different points on a continuum, rather than as utterly distinct things. Some
structural features are, in this sense, very slow activities.

As the face net has 328,320 connections, and as its face-recognizing perfor-
mance depends on how those connections are configured, the question that
presses is this: how do the connection weights come to be configured? That
question turns out to be much the same question as this: how to you get infor-
mation into the structure of a network so that a fundamentally stupid thing — the
network — can display “knowledge”? Needless to say, this is a question about
learning. Since learning is the topic of the next chapter, the question of how
networks learn is best addressed later in the more appropriate context of neural
learning (chapter 8). The task in this chapter is to understand what conceptual
tools are suited to explain how networks that have learned do represent. Given
that purpose, it may provisionally suffice to know that scientists have dis-
covered a palette of algorithms for adjusting the connection weights in an
ANN so that it will end up representing features of the training set and be able
to generalize to new stimuli. The existence of such algorithms allows us to
comprehend that there are naturalistic solutions to the problem of how neural
networks learn. Some algorithms for automated weight adjustment in ANNs
are neurobiologically more realistic than others; some scale better than others;
some are faster than others. Some involve external feedback; some do not. But

297

How Do Brains Represent?

applied to a network whose weights are initially set at random, they all yield a
network whose structure and dynamics embody information. Regardless of
whether any algorithm devised so far truly conforms to one of the brain’s
methods, we at least understand the sort of procedures that can do the job.

Accordingly, once trained, Cottrell’s face net will transform each one of a
wide range of possible input vectors (pictures of faces) into an appropriate
output vector (see again figure 7.12). The output vector is in effect face net’s
answer to whether the input is a face at all, whether it is male or female,
whether it is Billy or Bob. Cottrell’s face net achieved 100 percent accuracy
on the 11 images in the training set, identifying facehood, sex, and identity.
An interesting question is whether the network can identify the same faces if
they appear in different angles, with different expressions, accessories, lighting
conditions, etc. On this more demanding test, its accuracy was 98 percent: it
missed the sex and identity of one female subject. This is impressive, since it
means that the responses are not canned. The network has a kind of flexible
competence.

Can face net generalize to completely novel (never-before-presented) faces
to give correct answers to the questions face/nonface and male/female? Yes.
On the novel face/nonface task, it scored 100 percent. On recognizing sex of
the novel face, it scored 81 percent, showing a tendency to misclassify some
female faces as male. Can it correctly identify a “familiar” face when partially
obscured by a bar? Yes, save in one set of cases where the bar was placed so as
to obscure each subject’s forehead, thus indicating that variations in hair posi-
tion across the forehead probably played a significant, though not critical, role
in identification.

Success on these tests indicates that, indeed, the rudiments of facial repre-
sentation are embodied in the connection weights. But how does that work, and
in particular, how can face net generalize to novel cases? Analysis of the net-
work to determine how each unit responds under various conditions yields the
answer. The units whose activity is the focus of analysis are the middle layer of
units, sometimes called the hidden units. In particular, what we want to know is
the “retinal stimulus” to which a given middle unit will give its maximal re-
sponse. We want to know this because it will tell us something about what
stimulus characteristics those units represent and hence how representation is
achieved by the population of units in the network.

We might have expected each of these middle-layer cells to become selec-
tively responsive to some localized facial feature such as nose length, mouth
width, eye separation, and so forth. Reconstructing the actual “tuning” of the

298

Epistemology

Figure 7.14 Six of many holons: the preferred stimuli of some of the eells in layer two
of the face-reeognition network. Compare these holons to the imput images shown in
figure 7.13. Note that each preferred pattern spans the entire input space. (Courtesy of
Gary Cottrell.)

80 middle-layer “face cells” reveals that the network settled into a coding
strategy very different from this.

Figure 7.14 reconstructs the preferred stimuli for six typical face cells from
layer two of Cottrell’s network. Notice that each cell comprehends the entire
surface of the input layer, rather than an isolated facial feature, such as the
nose. The result is that each of the units represents an entire facelike structure,
which Janet Metcalfe, Cottrell’s coworker on this project, calls a “holon.”
None of these holons, corresponds to individual faces in the training set. Rather,
they seem to capture somewhat holistic characteristics of facehood — diffuse,
global characteristics for which we do not have applicable vocabulary. A given
face presented at the input layer will variously activate each of these 80 cells in
the middle layer, as a function of how closely it resembles or approximates each
of these 80 “preferred stimuli.”

Identifications (this is Billy) can be made by the output layer because for each
face entered as input, the resulting middle-unit activation pattern (80-element
vector) will be unique. Different photographs of the same person will produce
highly similar vectors at the middle layer; male/female discrimination reflects
the fact that the activation vectors for female stimuli are more similar to one
another than they are to activation vectors for male stimuli. Sometimes the
network gets it wrong when a female has very short hair, a rather long jaw, etc.
Sometimes we get it wrong too.

299

How Do Brains Represent?

• Individual male face
O Individual female face
iTProtofypical male face
^ Protofypical female face
(D Gender— ambiguous face

Figure 7.15 A schematic characterization of the hierarchy of learned partitions across
the neuronal activation space of layer three. (From P. M. Churchland 1995.)

Here, then, is the basic story. The values in the input vector reflect the gray-
level values in the photographs, and the configuration of connection weights in
the middle-layer vectors embodies what is task-relevant in various aggregations
of input values. Input vectors are pushed through the configuration of weights,
transformed into abstract representations in a high-dimensional “facial param-
eter space.” These vectors are in turn pushed through the last layer of weights,
with the resulting output vector representing answers to “Is it a face or not?”
“Is it a male or female?” and “Who is it?”

Figure 7.15 is a three-dimensional diagram of the eighty-dimensional acti-
vation space of the of middle-layer units, and each point in it is relevant to
representation in the “display of knowledge” sense. By contrast, the overall
partitions within that space reflect representations in the “capacity for knowl-
edge” sense. They embody the network’s background “conceptual framework,”

300

Epistemology

within which its fleeting perceptual inputs get integrated. The activation space
shows a primary partition into two regions, one for faces, one for nonfaces.
The nonface region is small because the middle-layer cells respond minimally to
a nonface. Not illustrated is the fact that the boundaries are actually fuzzy.
Notice also that there is no cutoff value, in any one dimension, below which the
coded subject must fail to be a face. This tells us that a photo may still be coded
as a face if it scores zero on one dimension. This is efficient, since allows
unusual faces, such as caricatures, to be coded as faces nonetheless.

The face region is further partitioned into male and female subregions,
roughly equal in volume. Scattered throughout are the particular faces on
which the network was trained. The female subregion’s “center of gravity” is
the general area of the prototypical female face; mutatis mutandis for the male
subregion. Further subregions might have been found had we asked the net-
work to respond to them — categories such as happy, sad, angry, and fright-
ened. (Metcalfe and Cottrell have trained a network to respond to distinct
emotional expressions. ^ ®)

The story so far has been told in terms of activation spaces specified by the
activation profiles of the hidden units. It is in the configuration of weights,
however, that the information is stored. Consequently, we can also ask what
the weight space looks like. In this hyperspace, each “synapse” (weight), and
there are roughly 300,000 of them, corresponds to a dimension of the space
(300K dimensions) (figure 7.16).

I hasten to add that the network’s categories, for example, malejfemale, are
scarcely comparable to my categories of male and female. Mine are enriched by
layers and more layers of background knowledge acquired through many years
of experience. My brain has vastly more weights than the meager face net, and
vastly more categorial understanding. Nevertheless, the conceptual point is our
focus, and the conceptual point illustrated by face net is that a network can
have categorial representations, which are collectively embodied as positions in
weight space and displayed as points in activation space.

If permitted to speculate, we may imagine that categorial knowledge in real
neural networks is likewise embodied in their synaptic-weight configurations
and in the resulting set of partitions on their neuronal activation spaces. As
weight configurations adjust to conform to the statistics of the input patterns,
geometric shapes (the system’s categories) are sculpted in the activation spaces.
If representation may be so envisioned, then, to a first approximation, my
worldly knowledge might be conceived in terms of neural networks, connec-
tivity strengths, and patterns of activation. Undoubtedly, this sketch is too

301

How Do Brains Represent?

Synaptic weight- Neuronal activation-

3 configuration space ^ vector space

Figure 7.16 (a) Synaptic weight space, whose axes are weights. This is the space of all
possible weight combinations from the synapses in the network, (b) A schematic net-
work. (c) Activation-vector space, whose axes are hidden units. This is the space of all
possible activation vectors across the population of hidden units. In this ultrasimple
case, there is one partition dividing the space, with regions on each side of the partition
where the prototypical example is located. (Courtesy of P. M. Churchland.)

302

Epistemology

crude nine ways from Sunday, but as a beginning, it looks more promising in
its consilience with some general properties of nervous systems than various
competitors, such as Fodor’s representational model, in which representations
are sentences written in the mind’s encyclopedia.^® Of the many respects in
which it is too crude, one concerns system dynamics. Another concerns the use
of time by the nervous system to represent, integrate, and regroup.^® Although
dynamical properties are known to be important, understanding the various
ways in which time matters and is managed by nervous systems is difScult, for
both technological and conceptual reasons.

7 Neurosemantics

“Semantics” in its most general sense has to do with meaning. In the last sixty
years, philosophers have taken semantics to cover three problems:

Reference How can a word, which is one thing, be about something, which is
another thing?

Meaning What things have meaning, what is it for something to have mean-
ing, what is its meaning, how are meaning and reference connected, and what is
going on when meaning is conveyed from one person to another?

Truth What sorts of things are true or false, what makes something true or
false?

Note that as formulated, the problem considers semantics to pertain primarily
to language and secondarily to the wider class of representations. This puts the
order of problems back-to-front, since nonlinguistic representation is probably
the platform for linguistic representation. How did this focus on language as
the prototype of representation come to be?

The story involves the development of that powerful tool, modern symbolic
logic. The great Polish logician Alfred Tarski (1901-1983) invented formal
semantics to complement the formal syntax of Russell’s and Frege’s symbolic
logic, and he is sometimes credited as the source for the classical approach.
Ironically, however, Tarski developed formal semantics precisely because he
recognized that in natural languages, syntax, semantics, and background
knowledge, along with present context, are inextricably intertwined. Because
formal logic was a highly artificial “language,” he realized that it needed a
complementary artificial semantics stripped of whatever is not formalizable

303

How Do Brains Represent?

(roughly, not programmable). And phenomena like polysemy (multiple mean-
ings), background knowledge, analogy, metaphor, shared assumptions, current
conditions, and the like, are not formalizable.

Despite Tarski’s caution that formal semantics was no approximation to the
real thing, and perhaps because no other approach looked viable, many clever
and determined people tried to make it work for natural language anyhow.
Perhaps, it was thought, Tarski was wrong.

From the beginning, the fit between natural language and formal logic was
problematic at best. Rather like putting an octopus to bed, problems re-
appeared almost as soon they had been “fixed,” and some problems could be
tucked away only by pretending that they were not really problems for seman-
tics anyhow, but for some entirely other pursuit, call it “pragmatics.”

One tangle of problems derived from the “nonnegotiable” assumption that
thought, and representation generally, is languagelike. Matters got distinctly
worse if the language that all representation was supposed to resemble was the
“language” of formal logic. This languagelike assumption created unbridgeable
explanatory chasms between human representation and nonhuman representa-
tion, between nonverbal children and verbal children, and between sensory
perception and imagining on the one hand and linguistic thinking, such as
talking to oneself, on the other. One tanker-sized catastrophe occurred over
language learning. Learning a language obviously requires representations, but
all representations were allegedly languagelike, so you cannot learn a language
until you have one.

To confront the learning catastrophe, Fodor postulated an innate, and hence
unlearned, complete language — a language of thought shared by all humans.
According to Fodor, when the infant acquires its cradle tongue, it is learning
only a translation between its innate Mentalese and its encountered French or
English. It is not acquiring a language for the first time. This holds even for
concepts like gravitational field, neutrino, and virus. For a while, it was hard to
tell whether the troubles with the classical approach were just the normal frus-
trations encountered in getting a theory adequately to explain the phenomena
in its domain, or whether they signaled fatal defects that called for a new
approach. By the 1980s, however, it looked like the defects were nontrivial.

Withstanding withering scorn, several linguists/psychologists (mainly Ron
Langacker, Elizabeth Bates, Gilles Eauconnier, George Eakolf, Jeffrey Ellman
and their students) suspected that the flaws were indeed fatal. To test this pos-
sibility, they began, each in his or her own way, to challenge the assumptions
of the classical approach, including its assumptions about the independence

304

Epistemology

of syntax and semantics, the context-free nature of meaning, the language of
thought supposedly used to learn one’s natural language. They also challenged
the suitability of dumping into so-called “pragmatics” all of the nonformal-
izable stuff: background knowledge, context, current conditions, analogies, and
so forth. The dumping into pragmatics looked entirely too self-serving.

Once detached from the rhetoric and trappings of conventional wisdom, the
classical framework tended to look a bit wobbly and unimposing, rather like
King George III in his nightshirt. The problems with using formal logic and
formal semantics as a model were well understood by the British philosopher
Ryle in the 1960s, but with no competing theory of semantics to tempt research
in a new direction, Ryle’s observations went unheeded.

In simple terms, the new approach, which is often referred to as cognitive
semantics, suggested that formal logic and formal semantics are atypical arti-
facts of natural language, not its heart and soul. Second, cognitive semantics
averred that language is primarily a tool for communication, and only second-
arily a tool for representation, not the other way around. Third, it said that
mental representation has fundamentally to do with categorization, prediction,
and action-in-the-real-world; with parameter spaces, and points and paths
within parameter spaces. Fourth, cognitive semantics suggests that representing
in this manner could be done by something that operated not like a serial
computer running a formal logic-like program, but by something with mas-
sively parallel networks — something like a brain.

Charting the blow-by-blow history of the debates between cognitive seman-
tics and brain-averse semantics is not germane to our purposes. From my per-
spective, the important consideration concerns each theory’s figures of merit,
that is, the comparisons between the power of each approach to explain a wide
range of data and mesh with the rest of the cognitive sciences and neuro-
sciences, as well as with evolutionary and developmental biology. Sized up in
these ways, the emerging new paradigm appears to have the greater promise as
a scientific attack on semantics. For one thing, it loses the absurd complications
entailed by the innate-language-of-thought hypothesis (see above). It also fits
better with neural-network approaches to representation, though undoubtedly
many significant insights made within the classical paradigm can be saved and
recycled. For another, it can give unforced and compact explanation-sketches
of central semantic phenomena, such as context dependence, counterfactual
statements, indexicals, analogy, and polysemy, and these explanation-
sketches can be followed with empirical testing.

305

How Do Brains Represent?

As with any conflict between paradigms, there has been much posturing,
many skirmishes, and many boundary disputes as the brain-averse approach
discerned the shape of an impending revolution and cognitive semantics battled
entrenched ideology. What will decide the various issues in the long run, how-
ever, is neither force of rhetoric nor tonnage of scorn heaped, but evidence and
explanatory power. These latter two will be the main focus of my discussion. In
the next section, I shall briefly consider an hypothesis to explain how repre-
sentations can be about things, and how meaning might be rooted in neural
network representation.

8 Being about Things

In face net, we saw partitions in the parameter space and specific volumes
within the space as fuzzily carving out domains for female faces, males faces,
and so on. The configuration of weights, we saw, is the structural matrix that
effects the transformation of one activity pattern (vector) into another. Con-
sider now a distinct network, trained on the same set of faces but in a different
order and with its various weights in different initial random settings. Even
though this net may achieve comparable performance in recognizing those
faces, the details of its weight profile may be utterly different from that of the
first net.

The important point is this: notwithstanding the differences in “synaptic”
details, comparable partitions in activation space would be made. The configu-
ration of the activation subspaces for male/female, and of the subspaces for
each individual face, would be mutually congruent; that is, they would map onto
each other (see again figure 7.15). This means that so far as representing is
concerned, the critical thing is the overall geometry of the subspaces, wherever
they happen to be located in the wider activation space of each network. For
example, the subregions for each of the learned categories will map onto
each other so as to preserve all of the similarity and distance relations between
them.

This is significant, because it implies that the two networks represent, say,
female faces in much the same way, differences in the details of their learning
notwithstanding. In turn, this suggests that having the same representation
comes to this: there is a relation-preserving mapping between configured pa-
rameter spaces. Loosely speaking, the two representations are intertranslatable.

306

Epistemology

Do the networks have to have the same number of units, connections, and
weights to achieve this categorial or conceptual similarity? No. If face net a had
two fewer middle layer units than face net /?, the categorial conhgurations
within a and can still be very similar, or even perfectly congruent. And so
also if face net a is trained on a somewhat dilferent set of faces than face net
P, or on dilferent set of nonfaces. Of course, if face net a never sees any female
faces, or if all the men it ever sees have beards, or if all the women have
topknots, it will have a somewhat dilferently configured space from the more
normally trained face net /?. The two representational schemes will be at least
roughly “translatable,” nevertheless.^®

Perhaps this general picture holds true, very roughly at least, of humans.
As toddlers, our limited experience sometimes gives us false expectations. For
example, if the only dogs we see are black Labradors, we might predict that
all dogs resemble black Labradors (as I did). Additional experience with
Pomeranians, St. Bernards, poodles, and so forth, along with contrasts with
wolves, foxes, coyotes, and raccoons tuned up my neurons so that the catego-
rial configuration of my dog subspaces became a little more closely aligned
with those of wider community.

In general, for networks to have congruent subspaces corresponding to a
category, they need to pick up on the real similarities in the shared stimuli. For
it is the similarity relations that are refiected in the relative positions of the
many subspaces: labs are closer to retrievers than either is to schnauzers, and
all of these breeds of dog are closer to each other than any dog is to a carrot.
My category bird may have more distinguishable points in it than yours, but
fewer distinguishable points than that of my sister, a devoted bird watcher.
Still, there will be sufficient similarity in their internal geometry that we can
usually understand each other.

Is it possible that in adulthood your spider subspace and my cow subspace
might happen, through pure coincidence, to have an identical geometry? It is
extremely unlikely, especially because ours are not spaces with a mere three or
four dimensions, but hyperspaces with thousands of dimensions. Sometimes,
especially with children, a misunderstanding will arise because by sheer acci-
dent the child picked up on the wrong similarities. A toddler’s concept of
newspaper may apply to anything used to start the fire, the child realizing only
later that newspapers are read.

If people are given roughly similar experiences, then superordinate cate-
gories, such as animal, vegetable, or furniture, will also be roughly similar in
shape and have roughly the same subspaces. Psychological data suggest that

307

How Do Brains Represent?

barring highly unusual conditions, we and our neighbors share much the same
prototypes (carrots and potatoes are prototypical vegetables, chairs and tables
are prototypical furniture, and so forth). Such psychological data on category
structure could, therefore, be understood as evidence for approximate congru-
ence in hyperspace geometry. It is much more difficult to give a unified account
for these and related semantic phenomena in the classical framework.

To a first approximation, a representational framework can be about things
in the world because it maps onto the similarity structure of things in world.
More accurately, a representational framework maps onto those statistics of its
environment that the organism, given its way of life, needs to attend to in order
to survive and thrive. Distinctions between individual ravens may be tremen-
dously important to other ravens, but are not something my dog cares much
about. Given their way of life, ravens will care quite a lot about the differences
between ravens and crows (especially for mating and for cooperative jobs such
as harassing a wolf off its kill), between ravens and eagles (especially because
eagles can kill ravens), ravens and sparrows (sparrows have yummy eggs but
can harass a raven). For my dog, it does not really matter whether the birds
stealing his kibble are ravens or crows or jays. Thus the mapping between the
animal’s world knowledge and the world is not independent of what the animal
cares about and pays attention to. This mapping, “me”-relevant and behavior-
guiding as it is, makes it possible for an animal’s representations to be about
things in the world. And the similarity in activation-space geometry between
two brains is what makes it possible for one brain to share an understanding
with another.

Incidentally, although the face net example used to launch this story involved
training by examples, the causal origin of the representational geometry is not
my main focus here. Presumably, in animals an important intermixing of
genetically driven preparation and experientially driven tuning results in an
individual’s knowledge (see chapter 8). For the purposes at hand, the main
focus is on the question of how representations in brains could be about things
the world.

Seen through the lens of vector coding and parameter spaces, “aboutness”
and meaning in representation are rather like the “aboutness” and meaning of
maps. As maps can be richer and more detailed, so with world representations.
As maps can have errors, distortions, and omissions, so too can world repre-
sentations. In maps the internal relationships between the points and regions on
the map make it a map of London or the Tatshenshini River or Alaska. Maps
are for navigation, for going somewhere and doing something, and thus they

308

Epistemology

can be enriched with task-relevant features. A road map of southern Alaska
showing road quality and location of filling stations is less interesting to some-
one who plans to canoe the Tatshenshini River and needs to know the location
and nature of the river’s rapids, sandbars, and tributary infiows.^®

9 Me, This, Here, and Now

The classical approach to meaning and representation was hopelessly outfoxed
by indexical expressions such as “I” and “here” and demonstratives such as
“this” and “that.” These context-dependent expressions, so natural and easily
usable in natural language, presented a ferocious riddle for the classical ap-
proach, with its insistence on contextual independence. So the classical story
had to cobble together special mechanisms for generating “the view from
here,” so to speak, from context-free semantics. And a vexing business it was.
How much of now is now, and how much of here is here! How can I crank the
equivalent of the term “me” out of a set of descriptions, even a rather long set
of descriptions?

Neural-network researchers, however, came to the problem from a dilferent
direction. They realized that because the animal’s body and its brain are the
locus of sensory input, attention, and motor decisions, “my point of view” is
the basic representational stance. Context-free representation, on the other
hand, is a far fancier contraption and a more difficult achievement. A brain can
probably take for granted its current context, with its spatial configuration of
things and events in relation to “me” and what “I” am interested in and paying
attention to. The “me-here-now” trio, therefore, does not need to be specially
generated by contrived and devious logical mechanisms out of context-free
sentences. So let’s have a closer look at how the brain might be organized to
handle these matters.

As we saw earlier (chapter 3), spatiality is deeply connected to body repre-
sentation, in both sensory and motor domains, and body representation is fun-
damental. Understanding where things are in three-dimensional space does not
just arise supernaturally, that is for sure, nor is it just given, whatever that
might mean. Spatial understanding depends crucially on the structural organi-
zation of various receptor sheets and on how sensory signals are integrated and
represented. And this organization will have been configured to serve the needs
of motor skills, and motor control generally.

309

How Do Brains Represent?

A range of results from basic neurobiological research, behavioral research,
and neural modeling come together in a rather compelling idea developed by
Alexander Pouget and Terry Sejnowski. Their hypothesis grounds a strategy
for explaining how the primate brain integrates diverse sensory signals and
generates an objective representation; that is, a representation of where things
are in the space relative to one’s independently movable parts — legs, arms,
fingers, eyes, and so forth. I devote the next section to sketching their idea, and
what is attractive about it.

10 Spatial Representation in Primates

The three brain areas of particular interest here are the hippocampus, pre-
frontal cortex, and posterior parietal cortex. These three are also highly inter-
connected, which hints that a consilient, interlocking theory may emerge in the
long run.^® To narrow the discussion, I shall focus on the posterior parietal
cortex. This area appears to provide the fundamental “objects-out-there-
external-to-my-body” organization critical for primate sensory-motor repre-
sentation and control. The hippocampus and prefrontal areas likely use these
basic parietal representations for additional purposes (e.g., remembering the
when and where of goodies, planning movements, generating images of move-
ments, etc.).

The crux of the idea developed by Pouget and Sejnowski (introduced on
pp. 77-79) is that certain neural networks in posterior parietal cortex generate
a sort of map-on-demand, i.e., a device that takes sensory information from the
various modalities and transforms it into information that guides the motor
structures. For convenience, I think of this network as an archmapper. What
does the archmapper do?

Suppose that you have a mosquito on your right elbow that you wish to
swat. Something is felt, and your somatosensory cortex registers the sensation
in your body-surface-space. Something else, namely your left arm, needs to
move from its position at your side to smack your right elbow precisely where
the mosquito is feeding. This means that your brain needs to know how to
move this object with shoulder, elbow, wrist, and finger joints so that contact is
made. That the signal is in the “elbow” position on the somatosensory map
does not, in itself, contain that information. Very roughly, your brain’s prob-
lem, in parameter-space terms, is this: what path in joint-space has an endpoint

310

Epistemology

that maps onto the location of the stimulus in skin-spacel What your brain
needs is a mapping between joint-space and skin-space. It need, in other words,
a transformation from a path specified in joint-angle coordinates to a position
specified in skin coordinates.^®

Or suppose that the familiar whine of the mosquito is detected, and you
catch a glimpse in your peripheral vision of a flitting something. In the early
stages of the visual system (e.g., VI, V2), the location of the visual signal is
specifled in retinal coordinates; that is where the signal is on the retina. To
move the eyes and head to look at a heard or felt object, or to reach with an
arm or tongue or foot for a seen object, the brain needs to know where to go in
the appropriate coordinate system. On their own, retinal coordinates will not
suffice; cochlear coordinates will not suffice. The brain needs to know, inter
alia, where the eyeball is with respect to the head, where the head is with re-
spect to the shoulder and trunk. Coordinate transformations are needed to
specify where the eyeball should go to foveate (position the eyeball so that the
signal from the stimulus falls on the foveal region of the retina). Normally,
we reach our hands and move our eyes to a target effortlessly, and the compu-
tational resources needed to pull this off are not part of what the brain has
conscious access to. The effortlessness makes the task seem easy, but computa-
tionally it is anything but simple. The central point is that sensory coordinates
have to be transformed into motor coordinates in order to connect with a sen-
sorily specified target (figure 7.17; see also figure 3.7).

Enter the archmapper. This network has the representational resources to
take information from various sensory systems and yield “go to” locations in
the corresponding motor frameworks — eyeball, neck-and-shoulder, hand, arm,
etc. It can specify what path, given in joint-angle coordinates, will get your arm
to the right position in skin space. It can specify what configuration the eye
muscles should have so that you can foveate the mosquito. It is not dedicated
to any one specific mapping, but integrates fairly abstract sensory signals, and
delivers fairly abstract go-to signals to the motor structures. And it seems
plausible that a network of this kind provides the wherewithal for the repre-
sentation of space.

The archmapper can be deployed by different motor structures for distinct
motor chores, such as moving the eyes, hands, pinnae (ear flaps), legs, or head.
In doing so, it relies on sensory information from retinas, cochleas, joints,
muscles, tendon receptors, and so forth. The archmapper is not exactly or
merely perceptual, nor exactly motor, nor exactly egocentric (self-centered) or
allocentric (object-centered). It combines information from multiple sources in

311

How Do Brains Represent?

Coordinate

transformations

Input

signals

Figure 7.17 The role of the posterior parietal cortex in the transformation of reti-
notopic visual information into higher-order reference frames. Eye position, head posi-
tion (determined from neck proprioception and vestibular sources), and gaze position
(determined from visual sources) are used to modify retinotopic signals. The posterior
parietal cortex is thus positioned to provide an intermediate stage in the conversion of
visual and auditory information into eye-, head-, body-, and world-centered coordinate
frames. (Based on Andersen 1999.)

a way suited to multiple applications, but cannot neatly be described in every-
day terms. This is one of those examples where the function of a neuronal pool
does not correspond to any familiar, everyday function. Evidently, however, it
is essentially spatial.

Integration of somatosensory “body knowledge,” including proprioceptive
and vestibular knowledge such as “where-this-body-part-is-in-relation-to-other-
body-parts,” with visual-auditory “where-things-are-in-relation-to-my-body”
knowledge allows for general representations of “me-in-extemal-space.” And
structures in the parietal cortex seem to be part of this “me-in-external-space”
representation. As noted earlier, the spatial aspects of body representation can
be only part of the self-representation story, however, because other aspects,
involving various dimensions of feeling and homeostasis, will figure in what it
is to have a “me” representation.^^

312

Epistemology

The mathematical details of exactly how the Pouget-Sejnowski hypothesis
runs take us beyond the scope of this chapter. But some neurobiological evi-
dence is essential as background. Neurobiological studies of area 7a and 7b in
the parietal cortex have provided important clues as to how coordinate trans-
formations are accomplished by networks of neurons. Monkeys with bilateral
lesions in area 7 show poor reaching to a target (ataxia), misshaping of hands
to fit the shape of the target, and slowness of movement. They also show de-
fective eye movements, principally in foveating, and they have other impaired
spatial abilities. They are poor at finding the home cage when released, poor at
route-finding to a food source, and poor in judging spatial relationships among
objects (e.g., “the food source is the box located nearer to the can”).

Another region of the parietal cortex, area 5, contains some cells that fire
maximally to a signal when the arm is reaching and others that fire selectively
to the expectation of a stimulus (figures 7.18 and 7.19). Because a great deal of
research has probed the visual properties of this area, one tends to think of
these regions as essentially visual. Recent data reveal, however, that they are
much more than that. The response patterns of neurons in these areas can be
modified by many factors, including auditory, somatosensory, and vestibular
signals, as well as attention, intention, expectation, preparation, and execution.
This clearly indicates that these neurons are more than just sensory.

Area 7 is multimodal and contains cells individually responsive to either
visual, auditory, somatosensory, chemical, vestibular, or proprioceptive signals.
Interestingly, auditory cells in this region appear to be mapped in retinotopic
coordinates. A few cells are multimodal: a given cell may respond to visual and
auditory signals, or to somatosensory and visual signals, or to chemical and
somatosensory signals.

It is often claimed that our conception of space is unified, and sometimes
even that it is necessarily unified. Yet it is unclear what introspection, innocent
of philosophical indoctrination, actually delivers on this point. Nevertheless,
if introspection does present the “oneness” of spatial perception, then that
perception is undoubtedly illusory to some degree. Various versions of “where-
perceived-objects-are-in-my-body-space” can dissociate (largely without intro-
spective notice) as a function of precisely which perceptual modalities are
involved.

The effect has been demonstrated in a variety of experiments. For example,
in ventriloquism, speech is perceived as coming from a puppet whose mouth
merely moves in synchrony with the speech sounds. In this instance, spatial
location via auditory signals is trumped by visual-motion signals associated with

313

How Do Brains Represent?

Figure 7.18 Cytoarchitectonic cortical maps of the macaque monkey by Brodman
(1909). Note the location of areas 5 and 7. (From Faster 1995.)

speech. Other examples include changes in visual perception brought about
by vibrating the neck muscles (thus stimulating the vestibulum); and Stevens’s
production of illusory visual motion by paralyzing the eye muscles (see also
pp. 85-86).^’’' Additionally, within vision there can also be dissociation in
normal subjects between spatial coding as it is visually experienced and as it is
used for grasping.

One intriguing pattern of breakdown in spatial reasoning occurs in hemi-
neglect, a condition sometimes seen in patients with unilateral lesions of the
right parietal cortex. These patients display a marked tendency to ignore the
contralesional (i.e., left) side of their body-centered world. They tend to look
only to the right, though some can move their eyes to the left if directly asked

315

How Do Brains Represent?

to do so. Asked to make a drawing or reproduce a figure, they will omit most
or all of the left half; asked to “cancel” (cross out) all the lines on a page, they
will fail to cancel all the lines left of center; asked to bisect a horizontal line,
they will draw the transecting line off-center to the right. (See also discussion of
parietal-lobe symptoms in chapter 3.)

In one ingenious experiment (Biziach and Luzzatti 1978), hemineglect
patients were asked to imagine that they were standing in a well-known plaza
in their city, and to describe what they could see from a given vantage point.
Their descriptions omitted objects on the neglected side relative to the imagined
vantage point', told to imagine they were standing at the north end of the plaza,
they would omit the buildings to the east, and when subsequently instructed to
repeat the task from the southern vantage point, they would list the eastern
buildings and omit from their description all the westerly buildings that they
had listed immediately before. This shows that hemineglect is a deficit of spatial
reasoning and/or representation at some fairly basic level, and not just a per-
ceptual failure.

There is also a motor component to hemineglect. Neglect patients show little
or no spontaneous use of limbs on the left side of the body, though some will
reluctantly move the neglected limbs upon direct request. They will also neglect
auditory stimuli from the left, sometimes failing to acknowledge others who are
speaking to them from that side. This sort of polymodal, perceptuomotor defi-
cit is what one might expect from a lesion of parietal cortex, since it receives
and integrates inputs from multiple modalities and is known to be involved in
the coordination of perception with action.

Pouget and Sejnowski used hemineglect as a test of the predictive power of
their archmapper hypothesis (described above). They created a network model
that has the oculomotor input/output structure and response properties that
they attribute to area 7a, and then “lesioned” it by removing the units that
correspond to the right side of the brain. This left the network with a dis-
proportionately high number of neurons that were most responsive to right-
ward eye positions and/or right-visual-field stimuli. They then equipped it with
a winner-take-all output-selection mechanism and tested it on stimuli similar to
those used with neglect patients.

The network output exhibited striking similarities to the human behavioral
results. In the line-cancellation task, the network failed to cancel lines on the
side opposite the lesion. More important, the line between the cancelled and
noncancelled areas was fairly sharp, even though the underlying representation
had only a smooth gradient. The network also paralleled human behavior on

316

Epistemology

the bisection task: it was successful before the lesion, but shifted the transection
point to the right after the lesion. In other experiments, it was shown to sulfer
object-centered as well as visual-field-centered neglect. In an experiment where
the network received head-position information (instead of eye-position infor-
mation), the network exhibited the same curious eflcct found in human patients
whereby performance on left-field tasks can be improved by turning the head
to the right. Although these phenomena have been problematic for existing
theories of hemineglect, Pouget and Sejnowski’s model is able to explain how
they arise naturally from an organization of response functions that can plau-
sibly be attributed to the human parietal cortex.

To the extent that the experienced “oneness of space” is not illusory, it
highly depends on the fact that all of the signals are generated in one nervous
system, inside one body, that has one spatially linked source of signals. There is
no single objective spatial representation (of the sort standardly presupposed
by symbolic models of representation), but a distributed, multimodal repre-
sentation that fundamentally integrates perception and action, self and world.
Where constancies appear across distinct modalities, it becomes possible and
even inevitable to understand them as representing an enduring world beyond
the body.^®

Certainly, many questions about the nature of our representation of space
remain. In particular, it may be wondered whether the archmapper should be
limited to spatial information, or whether the brain would more likely have a
spacetime archmapper. Probably it should, but fieshing that hypothesis out is a
later scientific development. Further research will determine whether the basic
Pouget-Sejnowski idea succeeds in pointing us in the right direction.

11 Concluding Remarks

This chapter brings us only to the doorstep of the neurophilosophy of repre-
sentations. Even then, it provides at most a squinting, keyhole view of the ter-
rain beyond. Neural nets, for example, can be far more powerful and versatile
than the simple nets illustrated here. They can have backloops; they can add
units and connections, accommodate symbols, develop specialized subregions,
and incorporate various activity-dependent and modulatory properties seen in
real neurons (figure 7.20).^® Also undiscussed are surprising discoveries about
how smart and computationally deft real neurons are. New results on real

317

How Do Brains Represent?

outputs

Figure 7.20 Three types of networks. (A) Feedforward network with one layer of
weights eonnecting the input units to the output units. (B) Feedforward network with
two layers of weights and one layer of hidden units between the input and output units.
(C) Reeurrent network with reciprocal connections between units. (From Churchland
and Sejnowski 1992.)

neurons and their role in how the brain represents abound; here I list a few; the
addition of new neurons to hippocampal and cortical structures, the activation
of sequestered synapses as a function of the level of neuronal activity, the
modification of receptive-field resolution as a result of attentional influence,
the existence of nonspiking computation, self-regulating synaptic receptors,
activity-dependent gene expression, and neuromodulation everywhere.

How background knowledge and context figure in ongoing sensory and
motor representation is under vigorous study at several levels, as are mecha-
nisms for directing attention. Representation of causality is a monumentally
important target of intense research at the cognitive/neuroscience/philosophy
interface, though it has not been discussed here. More generally, inference and
analogy, though fundamental cognitive operations, have only recently found
their place on center stage in cognitive science."^®

318

Epistemology

One of the most profound recent developments, sometimes going under the
name “situated cognition,” has been the realization that brains do not have —
and do not need to have — a complete representation of the current situation.
Instead, brains can selectively represent the world on a need-to-know footing
and can rely on the fact that the world is mostly stable and continues to be
there, available for second looks and closer looks.

Further, the idea of distributed cognition depends on the realization that
social animals can divide cognitive labor, and that if a brain can represent
which person has which competence, this general knowledge is far more eco-
nomical than representing all of the detailed knowledge in one head. For
humans, this strategy gets extended to cultural artifacts, such as tools, stories,
books, and knowledge-preserving institutions. Because so much science and
know-how is scaffolded onto our environment, I need not learn it all from
scratch, or in many cases, I need not learn it at all. I can let the surgeon remove
my appendix, the electrician wire my house, and the pilot fly the plane. And the
Internet is only the most recent example of artifacts soaked in knowledge.

In canoeing terms, this chapter gets us only to the put-in; the real excitement
begins once the canoe is in the river. The recommended readings will help with
navigation and will also suggest yet other streams to follow.

Suggested Readings

Abbott, L., and T. J. Sejnowski, eds. 1999. Neural Codes and Distributed Representa-
tions. Cambridge: MIT Press.

Bechtel, W., P. Mandik, J. Mundale, and R. S. Stufflebeam, eds. 2001. Philosophy and
the Neurosciences: A Reader. Oxford: Oxford University Press.

Bechtel, W., and A. Abrahamsen 1991. Connectionism and the Mind. Oxford: Black-
wells. See especially chapter 2.

Churchland, Paul M. 1995. The Engine of Reason, the Seat of the Soul. Cambridge: MIT
Press.

Clark, Andy. 1993. Being There. Cambridge: MIT Press.

Fauconnier, Gilles. 1997. Mappings in Thought and Language. Cambridge: Cambridge
University Press.

Gardenfors, P. 2000. Conceptual Spaces: The Geometry of Thought. Cambridge: MIT
Press.

Geutner, D., K. J. Holyoak, and B. N. Kokinov. 2001. The Analogical Mind. Cam-
bridge: MIT Press.

319

How Do Brains Represent?

Hutchins, E. 1995. Cognition in the Wild. Cambridge: MIT Press.

Katz, Paul, ed. 1999. Beyond Neurotransmission. Oxford: Oxford University Press.

Lakoff, George. 1987. Women, Fire, and Dangerous Things. Chicago: University of
Chicago Press.

Pouget, A., and T. J. Sejnowski. 1997. Spatial transformations in the parietal cortex
using basis functions. Journal of Cognitive Neuroscience 9 (2): 212-231 .

Websites

BioMedNet Magazine: http://news.bmn.com/magazine
A Brief Introduction to the Brain: http://ifcsunI.ifisiol.unam.mx/brain/

Computational Neuroscience Lab: http://www.cnl.salk.edu
Encyclopedia of Life Sciences: http://www.els.net
Living Links: http://www.emory.edu/living_links

The MIT Encyclopedia of the Cognitive Sciences: http//cognet.mit.edu/MITECS
Science : http: // ww w .scienceonline.org.

The Whole Brain Atlas: http://www.med.harvard.edu/AANLIB/home.html

This page intentionally left blank

How Do Brains Learn?

1 What Is the Prohlem?

At the heart of traditional epistemology lie two questions: (1) what is the nature
of knowledge, and (2) where does knowledge come from? In chapter 7 we con-
sidered a neurophilosophical approach to the first problem. The whence and
wherefore of knowledge are the targets of this chapter: how does the brain
come to represent aspects of the world and, eventually, aspects of itself? More
generally, how do we come to know anything?

The range of things in the “knows” category seems as diverse as the range of
stuff at a yard sale. Some is knowledge how, some is knowledge that, some is a
bit of both, and some is not exactly either. There are things we can articulate,
such as the instructions for changing a tire, and other things we cannot, such
as how we retrieve facts from memory or how we distinguish the relevant from
the irrelevant in problem solving. To learn some things, such as how to ride a
bicycle, we have to try. By contrast, avoiding eating oysters if they made you
vomit last time just happens. Whereas knowing how to install Napster depends
on cultural artifacts that are imbued with others’ knowledge, knowing how to
clap does not.

Even if we have not have considered it before, we know that dolphins do not
knit and that Mt. McKinley is not made of yogurt. Presumably, this is because
when queried, we generate these beliefs from other things we do know. Some of
what you know are logical truths (e.g., it is false that lemons are both yellow
and not yellow), some are factual truths (e.g., bears have not been domes-
ticated), and some a bit of a mix (e.g., you cannot be in two places at the same
time). Some knowledge is language-dependent; some is not. Some knowledge
evokes strong emotions; some is emotionally pretty much neutral. Some
knowledge is conscious; some is not.

322

Epistemology

Some knowledge (e.g., the nature of electricity) is highly abstract, some (e.g.,
that nettles sting) is based directly on sensory experience. Some knowledge
seems to rely on character traits rather than general intelligence (e.g., how to
handle horses, how to make people laugh, how to tell a good tale). Some
knowledge is fleeting, but some endures a lifetime.

The dimensions of variation in what we know are legion. Of the miscellany
of categories honored in our conventional wisdom, which of them captures a
real distinction from the point of view of the brain will eventually be revealed
through scientific discoveries. That is, new taxonomies will emerge from the
coevolutionary interaction of neuroscience and psychology. At this stage of
cognitive neuroscience, it is difficult to tell which of our everyday categories
have sufficient integrity to endure as they stand and which do not. As I shall
suggest below, however, some hypotheses, drawing upon anatomical, behav-
ioral, and physiological results, are beginning to re-fence the landscape of our
everyday categorial system for thinking about cognitive matters.

2 Knowledge: Learned and Innate

Historically, much discussion in epistemology concerned how much of what we
know is based on “instinct,” and how much on “experience.” At the extremes,
some took the view that essentially all knowledge is innate (the Rationalists).
Knowledge displayed at birth is obviously a good candidate for instinctual
knowledge. A normal neonate rat scrambles to the warmest place, latches its
mouth onto a nipple, and begins to suck. When you touch the cheek of a new-
born human, its head turns toward the touch, it nuzzles for a nipple, and sucks
efficiently on any warm, nipplelike thing it finds. A kitten thrown into the air
rights itself and lands on its feet. Some of these early instinctual behaviors per-
sist through postnatal development; others do not. Other knowledge is obvi-
ously learned. That fire is rapid oxidation, that oxygen is an element, how to
milk a cow, or how to grow tomatoes are all examples of knowledge acquired
by learning, not by instinct.

Such contrasts suggest that everything we know has its origin either in the
genes or in experience, where these categories are entirely separate and ex-
haustive. Historically as well as currently, debate often revolves around the
appropriate criteria for sorting knowledge into the two presumptively separate
bins. In the absence of scientific understanding of biological evolution, neuro-

323

Now Do Brains Learn?

embryology, and the neurophysiology of experience-dependent modifications
to neurons, the debate tends to be rather less productive than intense. The more
we know about genes and development and the brain, however, the less we
need to rely on speculation and intuitions about what can and cannot be
imagined.

Neurodevelopment and neurobiology have essentially laid waste to the very
simple nature or nurture dichotomy. Biology turns out to be vastly more com-
plicated than the simple dichotomy implies. This two-bins assumption is over-
turned by a number of considerations, prominent among which is the fact that
normal development, right from the earliest stages, relies on both genes and
epigenetic conditions. Moreover, paradigmatic examples of long-term learning
rely on both gene expression and epigenetic conditions. This does not entail
either that there is no such thing as in-born instinct or that there is no such
thing as learning. This also does not entail that there are no causally significant
differences between, say, the sucking reflex and knowing how to shuck an oyster.
Indeed, there are. The important point is that the differences do not neatly line
up as caused by genes versus caused by experience. Underlying the existence of
both capacities are huge numbers of interacting, causally relevant factors, and
they do not sort as the simple two-bins assumption demands.

There exist, certainly, causally relevant differences between prenatal devel-
opment and postnatal learning, and between early development and later devel-
opment, as well as between skill learning, priming, and conditioning. Genes
versus epigenetic conditions is not a Alter for any of those differences. There is,
moreover, a compelling explanation why this should be so. The idea, explored
for example by Quartz and Sejnowski (1997), is that evolution lucked onto the
fact that regularities in the environment mean that the genome does not have to
code for everything; rather, it can rely on the existence of certain external con-
ditions to play a consistent role in regulating gene expression. If biological
evolution exploits environmental information in building a creature, why not
also for a creature’s adaptation for environmental change? To put it crudely,
why, if you were Mother Nature, would you care about a principled dichotomy
between nature and nurturel

Six important and related developments have chiefly contributed to the
appreciation that things are not as simple as the catchy phrase nature versus
nurture seems to imply.

■ What genes do is code for proteins. Strictly speaking, there is no gene for a

sucking reflex, let alone a gene for female coyness or Scottish thriftiness or

324

Epistemology

the concept of a hole. A gene is simply a sequence of base pairs whose order
contains the information that allows RNA to string together a sequence of
amino acids to make a protein. (A gene is said to be expressed when it is
transcribed into RNA products, some of which are translated into proteins.)

■ Natural selection cannot directly select particular wiring to support a par-
ticular domain of knowledge. Genes are in the cells of animals, and what
dies or lives on to reproduce is the whole animal, with its own style of per-
ception and motor control. Blind luck aside, what determines whether the
animal survives is its behavior, and its equipment, neural and otherwise,
underpins its behavior. If the animal’s behavior allows it to outwit or outrun
or outmuscle the competition, it has a chance to live on and reproduce.
Representational prowess can be selected for, albeit indirectly, only if the
representational package informing behavior was what gave the animal its
competitive edge. Hence representational sophistication and its wiring infra-
structure can be selected only via the motor output it upgrades. Thus the
resources, neural and otherwise, for motor control exert a powerful constraint
on the evolution of representational capacities.

■ There is a truly stunning and quite unpredicted degree of conservation in
structures and developmental organization across all vertebrate animals, and
a very high degree of conservation in basic cellular functions across phyla,
from worms to spiders to humans. (See figures 6.4 and 6.5.) Humans have
only about 30,000 genes, and we differ from mice in only about 3,000 genes.
Humans and chimpanzees are believed to share about 98.5 percent of their
genes. In fact, we share about 1 10 genes with bacteria. Some proteins, such as
histones, actin, and tubulin, are essentially the same in all organisms. Long
before the appearance of vertebrates, all the major protein superfamilies had
formed. Variations and elaborations within superfamilies were seen there-
after, of course, but no completely original protein superfamilies are found in
humans that might account for the cognitive differences between us and our
closest relatives, chimpanzees, or even between us and simple worms.

■ Given the high degree of conservation, whence the remarkable diversity of
multicellular organisms? Molecular biologists have discovered that some
genes regulate the expression of other genes, and are themselves regulated by
yet other genes, in an intricate, interactive, and systematic organization. The
systematicity ultimately depends on a clever trick: make some gene expres-
sion contingent on the local protein environment. But genes (via RNA)
make proteins, so you can regulate the expression of one gene by another

325

How Do Brains Learn?

gene via sensitivity to protein products. Additionally, proteins, both within
cells and in extracellular space, can interact with each other to yield further
contingencies that can figure in a regulatory cascade. An example of both the
highly conserved nature of developmental organization and of the critical
role of regulatory genes is the so-called master gene for the eye in Drosophila.
This gene regulates some 200 other genes via epigenetic contingencies (con-
ditions that exist at a given time in a given place in the embryo’s develop-
ment). Moreover, this gene is highly conserved: the “master gene” for the
mouse eye is essentially the same as the “master gene” for eyes in Droso-
phila. Implanted in Drosphila, the mouse gene will produce fruit-fly eyes. The
emergence of complex, interactive cause-effect profiles for gene expression
results in very fancy regulatory cascades that can make very fancy organisms.
Us, for example. Small differences in genes can have large and far-reaching
effects, owing to the intricate hierarchy of regulatory linkages.

■ Various aspects of development of an organism from fertilized egg to up-
and-running critter depend on where and when cells are born. This includes
cell specialization and wiring patterns of the various types of neurons. Neu-
rons originate from a daughter cell of the last mitotic division of precursor
cells. Whether such a daughter cell becomes a neuron or a glial cell depends
on its epigenetic circumstances. Which type of some hundred types of neu-
rons (e.g., excitatory pyramidal, inhibitory stellate, inhibitory basket) the
neuron becomes depends on its epigenetic circumstances. Notably, the genes,
in and of themselves, do not specify cell fate. That is, there are no genes for
Purkinje cells or for spiny stellate cells, in the sense that a specific gene is
necessary and sufflcient for the production of specific cell types. Moreover,
the manner in which neurons from one area, such as the thalamus, connect to
cells in the cortex depends very much on epigenetic circumstances, e.g., on
the spontaneous activity, and later the experience-driven activity, of the tha-
lamic and cortical neurons.

■ The successful strategy typical in development — an iterated, interactive,
organizational cascade — is continuous with regulatory cascades serving the
postnatal plasticity we typically call learning. Neurotransmitters such as glu-
tamate carry a signal from one neuron to the next, and neuromodulators can
regulate the functionality of receptors. Activity cascades, gene-expression
cascades, and feedback cascades modulate the modulators and are modu-
lated by other events. Some of the same cascades figure in both learning and
development.

326

Epistemology

For example, the NMDA receptor, a complex transmembrane protein, plays
a crucial role in the cascades leading to the synaptic strengthening in certain
forms of learning (pp. 345 If.). But the very same protein, NMDA, also plays a
crucial, if quite dilferent, role in brain development. As Corriveau and col-
leagues have recently demonstrated, in early development, NMDA regulates
genes that regulate the transition from proliferation of precursor cells to the
differentiation of neurons.^

The interaction between genes and extragenetic conditions can be unex-
pected. In certain species of turtle, for example, the sex of the turtle is deter-
mined not at fertilization but by the temperature of the sand in which the eggs
incubate. In mice, the sex of siblings adjacent on the placental fetus line in the
uterus will affect such things as the male/female ratio of a given mouse’s sub-
sequent offspring and even its longevity. Postnatal learning triggers cascades
leading to gene expression. For example, as we shall see in section 8.6, this is
true of certain cells in the amygdala as the animal is conditioned to expect a
foot shock after it hears a tone. We also know that if you are exposed to a new
sensorimotor experience during the day, then during your deep-sleep cycle, the
gene zif-26% is upregulated, and this affects how well you remember what hap-
pened to you during the day.

More generally, considerable evidence runs against the idea that brain evo-
lution consists in the selection of anatomically localized functional subsystems
(modules) that are separately heritable and are gradually optimized over gen-
erations. So far as we know. Mother Nature cannot reach into the depths of a
contingency cascade to tweak the genes to optimize particular behavioral traits,
such as forming the past tense of verbs. Instead, selection is forced to make do
with quite general neuroanatomical changes, such as changing the precursor-
cell proliferation schedules. These changes are highly constrained. For exam-
ple, as Steve Quartz (2001) astutely points out, the neocortex does not vary
across all dimensions, but retains common organizational themes such as the
horizontal 6-layer cortex, the vertical column, the general connectivity pattern
of input to layer 4 and output from layers 5 and 6 to other cortical areas and
the subcortex, and so on. Strikingly, what does change are the numbers of
neurons, and hence the numbers and sizes of cortical areas. ^ As a result, func-
tional changes, benehcial or deleterious, may be displayed by large neural
regions or even by the brain as a whole. In criticizing alleged examples of
domain-specihc behavior, such as the human female preference for a mate who
is devoted and can provide, Panksepp and Panksepp (2001a) note, “Simple
emotional systems with a modicum of some general-purpose cognitive skill

327

How Do Brains Learn?

may easily yield some of the most striking folk-psychological discoveries of
evolutionary psychology.”^

On the general question of the evolution of the human brain, Quartz sums up
his approach thus;

The evidence suggests that the selective forces underlying the evolution of human cognitive
architecture were critically connected to highly unstable climes. . . . Based on these consid-
erations, I suggest that an important feature of hominid evolution was a process I have
referred to as progressive externalization . . . , whereby the brain’s development becomes
increasingly regulated by extrinsic factors, likely mediated by variation in the scheduling
of various events in neural development. I suggest this process allows for flexible pre-
frontally mediated cognitive function, particularly in the social domain, and underlies the
rapid changes in social structure that was a response to the need for buffering ecological
instability. The upshot of this process was symbolic culture, which plays a central role in
shaping the structures underlying human cognition. (Quartz 2001)

Bear in mind that many questions in neurodevelopment and brain evolution
have not yet been answered, and new discoveries may profoundly change how
we think about these matters.'^ What I argue here is that we do know enough to
know that the nature versus nurture debate has been substantially miscon-
ceived. In sum, postnatal learning and prenatal development share mecha-
nisms; prenatal development relies on epigenetic conditions for gene regulation;
and birth allows for a much expanded range of extragenetic conditions to figure
in nervous-system self-organization. We are neither blank slates nor bundles of
instincts.^

One further — mainly semantic — point. The description “hardwired” often
takes a leading role in discussions about instinct and knowledge. What does
this expression mean? As noted earlier, the software/hardware distinction,
though applicable to manufactured computers, is hopelessly out of its depth
in describing nervous systems. So if “hardwired” means “like my computer’s
motherboard,” then it is meaningless in the context of neuroscience.

If “hardwired” means “a behavior that depends on brain wiring,” then we
need to ask, “As opposed to what?” So far as is known, all behavior depends
on brain wiring. If “hardwired” means “caused by genes,” we have already
seen unqualified versions of this idea wrecked on the shoals of developmental
complexity. Sometimes “hardwired” is used to refer to a circuit that is not
modifiable postnatally. Typically, this usage too is problematic, since virtually
all of a brain’s functions are modifiable in one way or another — by expecta-
tion, conditioning, drugs, and assorted adjustments of internal or external
conditions.® This includes perception, motor control, thermoregulation, the

328

Epistemology

vestibulo-ocular reflex, and various “set points” for anxiety, appetite, and
aspects of sleep cycles. Even the basic knee-jerk reflex, essentially run on spinal-
cord circuitry, though subject to descending influence, is modifiable to a de-
gree. For example, the amplitude of the kick can be affected by something as
simple as gritting your teeth, as well as by interruption of descending signals
from cortex. Perhaps there exists in some subculture a consistent, useful, and
unambiguous role for the expression “hardwired,” but because it is so heavily
encrusted with misconception and misdirection, it is preferable to seek more
precise terminology.

Justice cannot be done in this brief section to the large body of research on
which these comments rest. But my aim is fairly minimal: to uproot reliance on
outdated opinions about so-called innate knowledge as we move on to questions
about learning. To forestall criticism, let me emphasize that I am not saying
there is nothing to the distinction between what is learned and what is innate.
Rather, I am saying that the matter is far more complicated than we thought,
because of the interdependence of genes and epigenetic factors, prenatally and
postnatally.

3 Storing Information in Nervous Systems

The crux of the issue for this chapter is how brains know things. We can begin
by addressing the neural basis for postnatal plasticity, and more narrowly, for
postnatal plasticity that is uncontroversially experience-dependent. Flence we
shall consider what is more commonly called learning, remembering, forget-
ting, and adapting.

An appealing idea is that if you learn something, such as how to tie a
trucker’s knot, then the information will be stored in one particular location,
probably along with your other knot-tying knowledge, say between reef
knots and half-hitches. That is, after all, the general plan adopted when we
store paper files in a particular drawer at a particular location or perhaps
when we store electronic files on a computer. It is not, however, the brain’s
way. This was first demonstrated by American psychologist Karl Lashley in the
1920s.

Lashley reasoned that after a rat learned something, such as a route through
a certain maze, if the information is stored in a single, punctate location, then
by lesioning the rat’s brain in the right place, you should be able to take out the

329

How Do Brains Learn?

rat’s knowledge. Where might the right place be? Lashley trained twenty rats
on his maze. Next, he removed a different area of the cortex from each animal
and allowed the rats time to recover. He then retested each animal in the maze
to see which lesion removed the maze-knowledge. Lashley discovered that the
rat’s knowledge could not be localized to any single region. Instead, it appeared
that all the rats were somewhat impaired but all were somewhat competent. To
a first approximation, the more tissue removed, the more serious the deficit.

As follow-up discoveries revealed, spatial knowledge was not a particularly
good case for Lashley’s purposes, since it turns out to involve noncortical
structures (the hippocampus) and draws on a range of sensory modalities
(smell, vision, touch). Nevertheless, as improved experimental protocols went
on to show, Lashley’s nonlocalization conclusion was essentially correct. There
is no such thing as a dedicated “memory organ” in the brain; information is
not stored on the filing-cabinet model at all. Instead, information seems to be
distributed over many neurons. Nor is the modular organization of a conven-
tional computer — processing by one component and storage by another — the
brain’s way. The very same structures that process information are also modi-
fied to store information.

If a brain has knowledge, that knowledge depends on wiring, that is, on
neurons and how they are connected to other neurons. Additionally, the right
neurons must talk to the right neurons. If knowledge is in place prenatally,
something has to cause the wiring to be right. If knowledge is acquired in
response to experience, then existing wiring has to modify itself in the right
way. That is, the informationally relevant changes at the cellular level must
be orchestrated so that an overall coherent modification in system output is
achieved.^ Fundamentally, the heart of the problem is to explain global
changes in a brain’s output (behavior) in terms of orderly local changes in in-
dividual neurons. The local-global problem is part of the more general problem
of how to get device-cleverness out of component-stupidity. That is, the device
as a whole may respond adaptively and intelligently, but its individual compo-
nents are not themselves as intelligent as the whole system.

If global learning depends on local changes in cells, how do cells know,
without the guiding hand of intelligence, when they should change, by how
much, and where? In my discussion of artificial neural networks (ANNs) in
chapter 7, we saw how simple units could change so that the network stored
information in the pattern of synaptic weights where one set of units meets
another set of units (see again figure 7.16). Although neurobiologically unreal-
istic, these simple ANNs are conceptually useful because they successfully

330

Epistemology

demonstrate that the problem has a mechanistic solution, and they offer a first-
pass explanation of how real neural networks might do it. The basic idea that
feedback, such as punishment or reward, can initiate local modifications in
connectivity with global import suggests a range of testable hypotheses regard-
ing learning in real neural networks.

There are many possible ways neurons can change. For example, new den-
drites might sprout (figures 8.1 and 8.2). There might be some extension of
existing branches. Existing receptors could modify their structure (e.g., a
change in subunits of the protein that constitutes the receptor). Or new receptor
sites might be created. In the curtailing direction, pruning could decrease the
dendrites or bits of dendrite, and therewith decrease the number of synaptic
connections between neurons. Or the synapses on remaining branches could be
shut down altogether. Additionally, there might be modulation of sodium
channels to change the spiking profile of an axon as a function of neuron-neu-
ron interactions. These are all postsynaptic changes in the dendrites.

There may also be presynaptic changes in the axons. For example, there may
be changes in the membrane (channels might emerge or be altered), or new
axonal branches may be formed or pruned. Repeated high rates of firing will
deplete the neurotransmitter vesicles available for release, and that transient
depletion constitutes a kind of memory on the order of 2-3 seconds. One im-
portant presynaptic change involves increasing or decreasing the probability
that a vesicle of neurotransmitter will be released when a spike reaches the
axonal terminal of a neuron. The probability of neurotransmitter release when
a spike arrives at the terminal is referred to as the reliability of the synapse. For
example, a synapse may, on average, release transmitter once for every ten
spikes reaching the synapse. Reliability can be modified on a time-scale of a
few hundred milliseconds. Up-regulating or down-regulating reliability is a fast
and flexible way of changing effective synaptic connectivity. The synaptic
strength can be increased tenfold in less than a second, without having to build
new structures. Other presynaptic changes include changing the number of
vesicles released per spike or the number of transmitter molecules contained in
each vesicle. Finally, the whole neuron might die, taking with it all the synapses
it formerly supported, or in certain special regions, a whole new neuron might
be born. Every one of these changes does occur, though precisely how the var-
ious changes causally connect to input signals is still under study, and how
changes across populations of neurons are orchestrated remains baffling.

This broad range of modifiability can be conveniently condensed for this
discussion by referring simply to modification of the weights, or synapses. The

331

How Do Brains Learn?

Figure 8.1 Camera lucida drawings of basal dendrites of layer V human pyramidal
neurons: (a) newborn, (b) 3 months, (c) 6 months, (d) 15 months, (e) 24 months, (f)
adult. (From Schade and van Groenigen 1961.)

332

Epistemology

Figure 8.2 Changes in the relative densities of synapses in primary visual cortex
(broken line) and prefrontal cortex (continuous line) of the human brain as a function of
days after conception (expressed on a log scale on the y-axis). Notice that synapto-
genesis in the more anterior region continues at a high level after synaptogenesis in the
striate cortex has begun to decline. (From Bourgeois 2001; based on data from Hutten-
locher and Dabholkar 1997.)

connectivity modifications listed above ultimately involve synaptic change,
either directly or indirectly, or can be reasonably so construed.

When and where is the decision to modify synapses made? Basically, the
choices are rather limited. Essentially, the decision to change can be made
either globally (broadcast widely) or locally (with specific synapses targeted). If
it is made globally, then the signal for change will be essentially permissive, in
eflfect saying, “You may change yourself now,” but not dictating exactly where
or by how much or in what direction (stronger or weaker). If global, the deci-
sion will likely be mediated by one of the subcortical nuclei that projects very
broadly across the cortex and that have a role in regulating arousal, attention,
sleep cycles, internal milieu, emotional state, and so on.

On the other hand, if weight change is to be specific, we would predict a tight
connection in space and time between the cause and the effect. That is, the
cause and the effect should be in close spatial and temporal proximity. The next
question, therefore, is this: if spatial contiguity is critical, what temporal rela-
tions might signal a local structural modification with the result that the
weights change, and change in the right direction (either stronger or weaker)?

333

How Do Brains Learn?

These were the dominant problems raised by psychologist Donald Hebb in
his influential book The Organization of Behavior (1949). The crux of Hebb’s
insight, slightly reconstructed, is this: correlated activity of pre- and post-
synaptic cells should increase the strength of the synaptic connection; anti-
correlation should decrease the strength of the connection. Loosely speaking,
the idea is that if activation in the presynaptic cells causes activation in the
postsynaptic cell, this should tend to make all of their connections stronger. By
changing the strength of the synapse, you increase the probability of the post-
synaptic cell firing following the firing of the presynaptic cell. On its own,
however, one cell’s release of neurotransmitter is unlikely to cause the post-
synaptic cell to fire, because the postsynaptic effect of one neuron’s transmitter
volley is very small. So suppose two distinct presynaptic cells — perhaps one
from the auditory system and one from the somatosensory system — connect to
same postsynaptic cell and fire at the same time. This joint input activity creates
a larger postsynaptic effect and, if Hebb is right, a strengthening of the con-
nection. This general arrangement allows for associated world events to be
mirrored by associated neuronal events.

The actual mechanisms for modifying synaptic weights were not specified in
Hebb’s proposal, however, since neuroscience had not yet begun to catalogue
the assorted structural changes that could yield a strengthening of synaptic
connections (e.g., an increase in the number of vesicles released, an increase in
the quantity of transmitter per vesicle, an increase in receptor proteins, etc.). A
theory of learning mechanisms should itemize and specify the conditions that
must be satisfied for information to get stored. For example, it will need to
specify whether the firing of the postsynaptic cell is necessary or whether mere
depolarization by some critical amount is sufficient. Only recently have discov-
eries at this level of detail been made.

Hebb’s principle for synaptic weight change says, “When an axon of a cell A
is near enough to excite cell B or repeatedly or persistently takes part in firing
it, some growth or metabolic change takes place in both cells such that A’s
efficiency, as one of the cells firing B, is increased” (1949, 62).

The simplest formal version of the Hebb rule for changing the strength of the
weight W£j between neuron A, with a firing rate of Pj, projecting onto neuron
B, with an average firing rate of Vb, is ^wba = ^VbVa. This states that the vari-
ables relevant to synaptic change are the co-occurring activity levels, and that
increases in synaptic strength are proportional to the product of the presynaptic
and postsynaptic values. The weight changes, note, are all positive, since the
firing rates are all positive.

334

Epistemology

The simple rule admits of many variations that still qualify as Hebbian. In
particular, it can be modified to get a powerful reinforcement-lg&mmg algo-
rithm, which updates the weights as a function of a Hebbian correlation be-
tween a sensory reward (such as nectar) detected now and a representation
of whether there is an error in predicting what that reward would be. As we
shall see below, this is important because specific neural networks in bees, and
probably humans, do learn to expect a specific reward of a certain magnitude
at a certain time.

Does any kind of weight change count as Hebbian? No. To qualify as Heb-
bian, the plasticity has to satisfy two criteria: (1) it is specific to the synapse
where the pre- and postsynaptic activity occurs, and (2) it depends conjointly on
both the pre- and postsynaptic cells, but not on the activity of other (connected)
cells. Non-Hebbian plasticity will include changes that fail to satisfy either of
these two criteria. For example, if the modification occurs to the whole cell,
rather than to the specific synapse where the activity occurs, then the plasticity
is non-Hebbian. If it involves general instructions for cells to up-regulate
their synaptic connections, such plasticity is non-Hebbian.

To a first approximation, early development is characterized by mainly non-
Hebbian plasticity, whereas Hebbian plasticity probably characterizes much
of postnatal plasticity, including classical examples of learning.'^ In early child
development, both kinds of plasticity probably have an important role.

Some inquiries into how postnatal brains build world models are likely to be
more fruitful than others. One may be especially fascinated, for example, by
how human adults learn to construct proofs in modern symbolic logic. This is
unlikely, however, to be the most auspicious place to try to develop neuro-
biological hypotheses regarding learning. The main problem is that animal
models for any but the simplest logical capacities are not available, but animal
models are essential to neural-level exploration, and neural-level exploration is
essential for discovering learning mechanisms. Starting with simple forms of
learning will probably pay olf faster and also yield clues to the solution of the
more difficult problems. More tractable learning problems than theorem prov-
ing are reinforcement learning (operant conditioning), fear conditioning, and
spatial learning.

The “simple first” strategy has rarely appealed to philosophers, who tend to
feel that simpler forms of learning are irrelevant to “real” epistemology. This is
probably short-sighted. From an evolutionary perspective and from what is
known about conservation of mechanisms across species, we can infer that

335

How Do Brains Learn?

fancier kinds of learning procedures are probably modifications, upgrades, and
hitch-hikers on the simpler ones. Learning culturally dependent skills, such
as reading and theorem proving, undoubtedly engage culturally independent
learning mechanisms that are fundamental to honing skills in general, such as
visual-pattern recognition, spatial navigation, and problem solving.

4 Reinforcement Learning: An Example

To the casual observer, bees seem to visit flowers for nectar on a willy-nilly
basis. It turns out, however, that they forage methodically. Not only do they
tend to remember which individual flowers they have already visited, but in a
held of mixed flowers with varying amounts of nectar, they learn to optimize
their foraging strategy, so that they get the most nectar for the least effort.
Do they run through calculations to achieve this economy? No. The neuro-
biological basis for this cleverness is now partially understood, and these results
suggest some general hypotheses concerning reinforcement learning that may
be applicable on a broader scale that includes mammals.

In an experiment designed by Leslie Real, a small held was stocked with two
sets of plastic flowers, yellow and blue, each with a well in the center in which
precise amounts of sucrose could be deposited.® The flowers were randomly
distributed around the enclosed field, and were baited with volumes of “nectar”
according to the following rule; all blue flowers had 2 pi; 1/3 of the yellow
flowers had 6 pi; 2/3 had none. This sucrose distribution ensures that the mean
value of visiting a population of blue flowers was the same as that of the yellow
flowers, though the yellow flowers are more uncertain than the blues.

After initially randomly sampling the flowers, the bees quickly fell into a
pattern of going to the blue flowers 85 percent of the time. You can change
their foraging pattern by raising the mean value of the yellow flowers, for ex-
ample, by increasing the sucrose in the 1/3 baited yellow flowers to 10 pi. The
behavior of the bees displays a kind of trade-off between reliability of the
source type and nectar volume at the source type, with the bees showing a mild
preference for reliability. But they will forage equally between the blues and
yellows when the mean of the yellows is sufficiently high. For our purposes
here, what is interesting is this; according to the reward profile acquired in a
sample of visits, the bees adapt their strategy. How do bees — mere bees — do
this?

336

Epistemology

Figure 8.3 The single, diffusely projecting modulatory neuron VUMmxl in the bee
brain. Neuromodulatory neurons in the bee brain and dopamine projections in the
human brain play homologous roles. OE = cell body of VUMmxl. Systems like the
dopamine system in humans and the octopamine system in bees are called dijfu.se
neuromodulatory systems. “Diffuse” because the axons of the neurons are diffusely pro-
jecting, making synaptic connections throughout widespread brain regions. “Neuro-
modulatory” because the neurotransmitters released from these axons are thought to
modulate global brain states. Computational models show that neural activity in some
of these neurons distributes information about expected rewards based on previous sen-
sory experience. In both species, the diffuse neurons receive precategorized information
about rewarding events and combine this with sensory information to construct a scalar
signal that represents the error between the expected amount of an reward and the
amount actually received. Using this signal to control long-term changes in synaptic
weights allows this system to learn and store predictions rather than correlations. (From
Hammer 1993.)

The research of neuroscientist Martin Hammer provides an important piece
of the puzzle. In the bee brain he found a neuron, though itself neither sensory
nor motor, that responds positively to reward. This neuron, called “VUMmxl”
(“vum” for short), projects very diffusely in the bee brain, and its activity
mediates reinforcement learning (figure 8.3). For example, a particular odor
consistently paired with sucrose would change the weights on vum so that even-
tually vum would fire when that odor occurs alone. But how does the bee’s brain
allow it to learn the mean value and reliability of source types across a sample?

337

How Do Brains Learn?

In an artificial neural network, Montague and colleagues modeled the known
and relevant bee anatomy and behavior.^® They found that the weight-change
algorithm operating at vum did not just sum the various synaptic inputs.
Instead, the activity of vum represents prediction error, that is, the difference
between the goodies expected and the goodies received this time. Here is how it
works. Cell vum has input from the bee proboscis, where it sucks up nectar. We
can loosely think of this as the reward pathway (figure 8.6). Vum also has
inputs from sensory systems, for example visual (for color) and olfactory
(for odor). Simplifying a little, the output of vum at a given time is a func-
tion of the reward at which is expressed as r(t„), plus the combined value
of the sensory inputs (V(t„)) minus the value those inputs had just previously
at t„-i. More formally, the output of vum, d(t), is expressed thus: d(t) =

Kr) + F(r)]- F(C_i)-

Roughly, a result greater than zero corresponds to “better than expected,”
and a result less than zero corresponds to “worse than expected.” The output
of vum is the release of a neuromodulator that targets a variety of cells,
including those responsible for action selection. If that neuromodulator acts
also on the synapses connecting the sensory neurons to vum, then the synapses
will get changed according to whether the vum calculates worse than expected
(less neuromodulator) or better than expeeted (more neuromodulator) (figure
8.4). Assuming the model of Montague et al. is essentially correct, it turns out
that a surprisingly simple circuit, operating according to a fairly simple learn-
ing algorithm, underlies the bee’s adaptability to foraging conditions. (My
account here leaves out various details, but it captures the main ideas. See
Montague, Dayan, and Sejnowski 1993.)

Obviously, the bee has more flexibility if it can learn the nectar values of
flowers than if the nectar values are specified independently of experience.
Some years the Honeysuckle might do poorly and have little nectar; new nec-
tar-rich plant species might begin to invade the area; a flood might mean the
bee has to find new foraging territory where Indian Paintbrush and Lupin grow
instead of Honeysuckle and Campion; and so on. By being modifiable, the
bees’ neural networks can improve foraging. New pattern recognition (e.g.,
high nectar in the crimson flowers) enabled by reinforcement learning is a use-
ful thing.

For the bees, the correlations between flower color and nectar reward are
essentially spatial, but correlations also occur in the temporal domain. A bat,
for example, can learn that a certain sequence of tones is a reliable indicator
of tasty moths, and a dog quickly learns that one sequence of events predicts

338

Epistemology

error

Internal

states

Figure 8.4 Constructing and using a prediction error. (A) Interpretation of the ana-
tomical arrangement of inputs and outputs of the ventral tegmental area (VTA), whose
neurons project very widely and release dopamine. Ml and M2 represent two different
cortieal modalities whose output is assumed to arrive at the VTA in the form of a tem-
poral derivative (surprise signal) V(t) that reflects the degree to which the current sen-
sory state differs from the previous sensory state. (The overdot indicates the rate of
change.) The high degree of convergence forces V{t) to arrive at the VTA as a scalar
signal. Information about reward r{t) also converges on the VTA. The VTA output is a
simple linear sum S{t) = r{t) + V{t). The widespread output connections of the VTA
make the prediction error S{t) simultaneously available to structures eonstructing the
predictions. (B) Temporal representation of a sensory cue. A cue like a light is repre-
sented at multiple delays x„ from its initial time of onset, and each delay is associated
with a separate adjustable weight w„. The w„ parameters are adjusted according to the
correlations of x„, activity, and <5, and, through training, come to act as predictions. This
simple system stores predictions rather than correlations. (Reprinted with permission
from Schultz, Dayan, and Montague 1997. Copyright by the American Association for
the Advancement of Science.)

339

How Do Brains Learn?

going to the beach, while another predicts hiking on the marsh. Much con-
ditioning depends on temporal associations between events.

The dependency relations between phenomena can be much more complex
than such simple correlations as red flowers, high nectar. I may initially note a
correlation between dropping eggs and eggs breaking. I think: dropping an egg
causes it to break. Then I happen to notice that a dropped egg will not break if
I drop it on a soft pillow or in newly fallen snow. So the dependency relations
are a little fancier than I first thought. With continued exploration, I come to
appreciate that even with soft snow to land in, an egg may break if dropped
from a sufficiently great height. My simple correlations get upgraded to more
complex correlations as I learn about the world. This seems to be true of ani-
mals as well. We can learn that one event will probably follow the occurrence of
another, but not always, and the reason may be completely veiled. Part of what
we call intelligence in humans and other animals is the capacity to acquire
understanding of increasingly complex dependency relations. This allows us to
distinguish fortuitous correlations, which are not genuinely predictive in the
long run, from causal correlations, which are.

Does reinforcement learning in the humble bee have anything to do with us?
Very likely, since we too have a reward system that mediates learning about
how the world works. Wolfram Schultz has found neurons in the monkey
brainstem that, like vum, respond to reward, shift their responsiveness to a
stimulus that predicts reward, and indicate error if the reward is not forthcom-
ing. These neurons release dopamine at their axon terminals (and hence are
dopaminergic), and the dopamine is believed to modulate the excitability of the
target neurons to neurotransmitters such as glutamate or glycine.

By recording from single cells, Schultz showed that if an animal gets an
unpredicted reward — such as a squirt of juice — the dopaminergic neurons in-
crease their firing when the reward is received. With repeated trials where a
juice squirt follows the sounding of a tone, the monkey learns that the tone
predicts the juice. The response of these neurons tracks the monkey’s learning
that the tone predicts the juice. That is, the dopaminergic neurons now increase
their rate of firing when the tone is heard, which thus predicts the occurrence
of the reward. Should the reward fail to appear when predicted, then activity
in these neurons drops markedly below baseline at the time when the reward
should have appeared (see figure 8.5).

These neurons are believed to be part of the reward system. They arise from
areas in the midbrain (the ventral tegmental area or VTA) and the substantia
nigra, which receive projections from a wide range of areas. The input probably

340

Epistemology

No prediction
Reward occurs

Reward predicted
Reward occurs

Reward predicted
No reward occurs

Figure 8.5 Predictor neurons in the primate dopamine system. Each panel shows elec-
trical recordings from individual dopamine neurons from an alert primate during a task
where a sensory cue is followed 1 second later by the delivery of a juice reward. Each dot
is the occurrence of an action potential, and each horizontal row of dots represents a
single presentation of the sensory cue and reward. The histogram on top of each panel is
the total number of action potentials in a particular time bin. Top: Presentation of a
sensory cue to a naive monkey causes no change in the production of action potentials.
Delivery of a juice reward, however, causes a transient increase in the rate. Middle:
Presentation of the sensory cue causes a transient increase in spike production, but de-
livery of the reward causes no change in the firing rate. Bottom: Same as the middle
panel except that if the reward is not delivered, the dopamine neurons stop firing when
the reward would have been delivered as calculated from previous trials. The interpre-
tation is that the neurons are predicting the time and magnitude of the future reward
using information provided by the earliest predictive sensory cue. Abbreviations: CS,
conditioned stimulus; R, primary reward. (Prom Schultz, Dayan, Montague, 1997.)

341

How Do Brains Learn?

frontal

cortex

Figure 8.6 A schematic representation of the major dopaminergic tracts of the human
brain. Abbreviations: VTA, ventral tegmental area; hippo, hippocampus.

represents a surprise signal, in the sense that it measures the degree of difference
between the current sensory signal and the last sensory signal. The dopa-
minergic neurons project very diffusely to many regions of the brain involved in
goal-directed behavior and motivation, including the striatum, the nucleus
acumbens, and the prefrontal cortex, which, as we saw in chapter 3, plays a role
in emotional valence and in action selection (figure 8.6). The hypothesis is that
dopamine delivery regulates the plasticity of those neurons that make action
decisions, such as those in prefrontal cortex.

Convergent research indicates that these dopaminergic neurons are indeed
involved in reinforcement learning. In the 1950s, James Olds and Peter Milner
at Cal Tech devised a set-up whereby a freely moving rat could press a lever to
receive a small pulse of current through an electrode implanted in its brain.
Depending on the location of the electrode, rats quickly learned to do stimulate
themselves by pressing the lever. The only reinforcing reward for the behavior
was the pleasure caused by of the stimulation of the neurons. For specific
locations, such as the striatum, the nucleus acumbens, and the VTA, the rats
found the self-stimulation so pleasurable they would forgo food, sex, and water
to continue pressing the lever. Most recently, functional MRI has been used to
see whether any particular areas are especially active during reinforcement
learning. The results show regionalized increases in activity in the midbrain
dopamine system. This suggests that human reinforcement learning may have
some features in common with that of other animals.

342

Epistemology

Additional evidence confirms that the dopaminergic neurons in the VTA and
substantia nigra mediate reward and pleasurable feelings, and send prediction
error signals to the areas responsible for choice: animals injected with a sub-
stance that blocks the activity of dopamine show impaired reinforcement
learning; addictive substances such as cocaine and amphetamine increase dop-
amine levels. It is also important that VTA neurons display the prediction/error
profile to positive reward, but not to aversive stimuli. As we shall see in the next
section, a distinct brain region appears to mediate negative reinforcement
learning.

5 Fear Conditioning and the Amygdala

The amygdala plays a central role in evaluating stimuli as unpleasant. Animals
with lesions to the amygdala cannot learn that a certain innocuous stimulus
predicts an aversive stimulus, and hence they cannot learn to avoid the aversive
event. For example, amygdala-lesioned animals cannot learn that one event,
such as a tone, predicts a nasty event, such as a shock to its feet. Unlike normal
animals, they do not learn to avoid the nasty event by escaping when the tone
sounds (figure 8.7).

The necessity of amygdaloid structures for fear conditioning suggests that
some specific change may occur in the amygdala during fear conditioning. Can
specific changes in amygdala cells be seen? Yes indeed. The amygdala is not
an undilferentiated region, but consists of a number of specialized subregions
(figure 8.8). More restricted lesion studies reveal that the subregion relevant to
the cellular story of fear conditioning is the lateral amygdala (LA). Within LA
is the tiny dorsal subregion, housing two distinct populations of cells with dif-
ferent roles in fear conditioning. In the rat, each population has only about
twenty thousand cells.

Cells in one population (A) show a fast and transient change in synaptic
strength during the early phase of learning. Cells in the second population
(B) change more slowly, and their modification is more permanent. Further
manipulations indicate that the learning is Hebbian in both cases, though
mediated by distinct mechanisms. Type A cells have several types of receptor
channels to which the excitatory neurotransmitter glutamate will bind. One
type (the AMPA receptor) opens whenever glutamate binds, thus transmitting
a small signal from one cell to the other and causing the receiving cell to
depolarize.

343

How Do Brains Learn?

Figure 8.7 Classical fear conditioning can be demonstrated by pairing a sound with a
mild electric shock to the foot of a rat. In one set of trials, the rat hears a sound (left
panel), which has relatively little effect on the animal’s blood pressure or patterns of
movement. Next, the same sound is coupled with a foot shock (center). After several
pairings the rat’s blood pressure rises and the animal freezes; it does not move for an
extended period when it hears the sound. The rat has been fear-conditioned. After con-
ditioning, when the sound alone is given, it evokes physiological changes in blood pres-
sure and freezing similar to those evoked by the sound and shock together (right).
(From LeDoux 1994.)

344

Epistemology

Figure 8.8 A model of the neural circuit involved in conditioned fear. A hierarchy of
incoming sensory information converges on the lateral nucleus of the amygdala.
Through intra-amygdala circuitry, the output of the lateral nucleus is transmitted to the
central nucleus, which serves to activate various effector systems involved in the expres-
sion of emotional responses. Feedforward projections are indicated by solid lines, and
feedback projections are indicated by dashed lines. Abbreviations: BNST, bed nucleus
of the stria terminalis; DMV, dorsal motor nucleus of the vagus; NA, nucleus ambiguus;
RPC, nucleus reticularis pontis caudalis; RVL medulla, rostral ventrolateral nuclei of
the medulla. (From LeDoux 1994.)

345

How Do Brains Learn?

The other protein channel, the NMDA receptor, plays the key role in plas-
ticity. (We saw on p. 326 that during early development, NMDA is also
involved in regulating precursor-cell schedules.) This receptor is voltage sensi-
tive, which means that it will not open unless two primary conditions are sat-
isfied: (1) the neurotransmitter glutamate binds to it, and (2) the membrane
must already be a bit depolarized — typically from a second source. This con-
junction of events normally occurs only when a type A cell receives two distinct
inputs', an innocuous stimulus from one source (the tone), about when it
receives a strong stimulus from another source (the shock) (figure 8.12). When
these two events happen within a brief time window, the NMDA channel at the
“innocuous connection” opens, and stays open for about 100-200 msec. The
opening is actually a change in shape of the protein that sets free a magnesium
ion and permits calcium ions to enter the cell (figure 8.9).

Once calcium enters the cell via the NMDA receptor, a cascade of events
is launched that upregulates the response of the cell to the innocuous stimu-
lus. On a subsequent occasion, therefore, when the innocuous event occurs in
the absence of the aversive event, the A cells respond roughly as though the
aversive stimulus itself had occurred. This causal chain for strengthening the
synapse is known as long-term potentiation, or LTP. The general character-
ization of LTP is that the responsivity of the postsynaptic cell is potentiated
(increased). The elfect can last for many hours. The typical conditions for pro-
ducing LTP experimentally involve either a conjunction of inputs or a blast
from a single input. The NMDA receptor mediates LTP in some, but not all,
cells (figure 8.10).

Surprisingly, a crucial component responsible for LTP is presynaptic. It con-
sists in increasing the probability that neurotransmitter is released when a spike
reaches the axon terminal (i.e., it increases reliability). In some manner, possi-
bly by releasing nitric oxide (NO), the postsynaptic cell signals the presynaptic
cell to upregulate its probability of neuro transmitter release (figure 8.11). As
noted above, changes in the probability of release is something that can be
achieved on the order of a few hundred milliseconds. When the tone is dis-
sociated from the shock, LTP does not occur. (If the postsynaptic cell responds
even though the presynaptic cell did not send a signal, thereafter postsynaptic
responsivity is lowered. This elfect is called LTD — long-term depression.) In
sum, in type A cells of the dorsal region of the LA, we see NMDA-mediated
LTP.

What about type B cells, those that exhibit a more permanent change? In-
stead of the NMDA receptor typical of type A cells, B cells use a voltage-gated

346

Epistemology

Normal low- Induction of LTP involves

frequency firing high-frequency firing

Figure 8.9 The role of the NMDA (N-methyl-D-aspartate) receptor in the induction of
a form of neuronal plasticity known as long-term potentiation (LTP). Left: During
normal synaptic transmission, when the presynaptic neuron fires at low frequency, the
NMDA channels remain blocked by Mg^+ ions. Na+ and K+ ions can still enter
through non-NMDA channels to mediate ordinary synaptic transmission. Right: LTP is
induced when the presynaptic neuron fires at a high-frequency (a tetanus) and depo-
larizes the membrane of the postsynaptic cell sufficiently to unblock the NMDA recep-
tor channel, which allows calcium to enter the cell. (Based on Squire and Kandel 1999.)

347

How Do Brains Learn?

Prej

Post

Tetanus

NoLTP

LTP

Associative
pairing of
tetani in
pre and pre

Figure 8.10 Associative long-term potentiation, (a) Diagram depicting the spatial rela-
tionships between strong and weak synaptic input, (b) Stimulation of the strong input
produces LTP, but stimulation of the weak input alone does not. (c) When the strong
and weak inputs are paired, the depolarization produced by the strong input spreads
to the site of the weak input, which then contributes to the induction of LTP. (From
Levitan and Kaczmarek 1991.)

348

Epistemology

SHORT-TERM

Retrograde

singal generators

Figure 8.11 Mechanisms underlying LTP. When there is sufficient depolarization to
open the NMDA channels in the postsynaptic cell, Ca^+ flows into the postsynaptic cell,
and protein kinases are aetivated. Ca^+ influx changes the postsynaptic cell by aeting on
non-NMDA receptors, and it also sends retrograde messages back to the presynaptic
cell, telling it to release more transmitter. One of the retrograde messengers is believed to
be nitric oxide (NO). (Based on Squire and Kandel 1999.)

349

How Do Brains Learn?

calcium channel to trigger strengthening of the synapse. Like NMDA recep-
tors, these receptors are in the plasticity business. Like the type A cells, they
require a conjunction of inputs to open, but when they do open and further
depolarize the cell, a distinct cascade ensues, ultimately triggering gene expres-
sion and therewith the synthesis of proteins. Thus their elfect is more perma-
nent and the underlying change is presumably more structural than just
regulating the probability of release.

We need to pause and savor how these discoveries add up. Step by artful
step, they track the plasticity in fear conditioning from behavioral changes in
the animal’s ability to predict shock, to a particular brain structure (the amyg-
dala), then to highly confined subregions of the amygdala (the LA) specific to
fear conditioning, then to two distinctive populations of cells (in the dorsal LA)
whose synaptic weights are modifiable on different schedules, then to two spe-
cific voltage sensitive receptor proteins distinguishing those two cell populations.
Opening one type (NMDA) mediates shorter-term memory; opening the other
type (the voltage-gated calcium channel) appears to trigger gene expression as
part of the process of consolidating memory for the long haul. Thus the story
wends its way from behavior down through to specific proteins and changes at
the molecular level.

The fear-conditioning studies have been done mainly in rats. As noted in
chapter 5, however, in humans, the rare disease Urbach-Vitae causes bilateral
atrophy of the amygdala, which thus permits research on the effect of amyg-
dala destruction on humans. Neuroscientist Joseph Le Doux has studied a
woman with this disease. Like the amygdala-lesioned rats, she fails to acquire a
new fear-conditioned response. In normal subjects, if a mild shock on the hand
occurs a few seconds after the appearance of an innocuous blue square on the
television screen, in a few trials subjects exhibit a fear response when the blue
square appears. Le Doux’s subject does not, though she tests normally in other
respects. She does respond normally to the mild shock itself, but she shows
none of the conditioned responses to the innocuous stimuli that are typical of
normal humans exposed to blue/shock pairs: no increase in heart rate, no
sweating, and no feeling of apprehension at the appearance of the innocuous
stimulus. In response to queries, she says that she does not feel that anything
unpleasant is going to happen following the appearance of the blue square.
Interestingly, however, after continued trials she does acquire a certain level of
understanding, independently of any feeling of fear, based apparently on an
intellectual inference that there is a predictive connection between the innocu-
ous event and the unpleasant event.

350

Epistemology

Cue

Delay

ij.iuiimiiiluiliiiiiu

1 II.. .m

II 11 iimiiiiiMi II u ':iiii i .i_

llUllLllll. .11 ..

1 III ■l■lllllll■■^illll^■ '

lu mil 11 111 Jil l.. : . .JLiiMi 1111111 Jim ijiiiii, III iJ!.iMi:!iiiiuuiJiiJjiiiifli iii n i
1 , . 1 : I liil.llll , r '

Figure 8.12 Discharge of a cell in the prefrontal cortex of a monkey during five trials in
the classical delayed-response task. Arrow marks the monkey’s response at the end of
the memorization period (delay). Note that the cell is inhibited during presentation
of the cue but persistently activated throughout memorization (30 seconds in the upper
three trials, 60 seconds in the lower two). (From Fuster 1973.)

The questions driving this chapter are these: Where does knowledge come
from? How do we come to represent the world? Learning to avoid damaging
stimuli by learning to recognize an event that predicts their occurrence is an
important part of learning about the world. As we have seen, neuroscience is
beginning to reveal, in research that tracks learning changes through the vari-
ous levels of brain organization, where knowledge comes from and how we
come to represent the world.

Are there other learning capacities whose mechanisms neuroscience is prob-
ing? Indeed there are. Working memory — holding information in the ready
until the time to act is right — is another capacity where progress has been made
at the neuronal level (figures 8.12 and 8.13). This is especially intriguing be-
cause of the possible connection between awareness and working memory.

Recollection of events in one’s life is another domain where cognitive neu-
roscience has succeeded in pushing our understanding ahead. As we shall see
below, the capacity for recalling life’s episodes depends on a set of neural
structures distinct from those subserving working memory or fear conditioning.

6 Declarative Memory and the Hippocampal Structures

We can acquire skills, such as how to ride a bicycle or tie a trucker’s knot.
Memory for skills is referred to as procedural memory. Knowledge that a mild
shock will follow the appearance of a blue square is distinct kind of knowl-
edge — fear conditioning. Both of these contrast with remembering specific

351

How Do Brains Learn?

Figure 8.13 The structure of a model of short-term active memory. The soma of each
idealized unit is represented by a triangle at the right, and its input dendrite is shown at
the left with a row of synaptic contacts of a given strength or weight (w). The output
unit is the blank triangle at the upper right. The underlying triangles {H\, HI, Hn)
represent hidden units, that is, units that mediate transactions within the network and
determine its output at any given time in accord with the input it receives and its pre-
established (i.e., pretrained) functional architecture. (From Zipser et al. 1993.)

events that happened in the course of one’s life, such as that on your sixth
birthday you were given a bicycle, or where you parked your car this morning.

Conscious recollection of events and episodes is referred to as declarative
memory, one can say or “declare” what one remembers about past experiences.
It is also called explicit memory or conscious memory. In ordinary parlance, this
capacity is usually what people refer to when they talk about their memory.
Recollected events are typically indexed temporally and spatially. That is, we
remember, for example, our first kiss with Jerry in the barn after the eighth-
grade skating party. Of course, not all details of an episode are bought to mind
in conscious recollection, though often the longer we dwell on the event, the
more the associated details emerge into awareness.

Recollection of individual events, along with suitable temporal and spatial
referencing, requires the hippocampal structures in the brain. These include
the hippocampus, the entorhinal cortex, the perirhinal cortex, and the para-
hippocampal gyrus (figure 8.14). How do we know that these structures are
important for declarative memory but not for fear conditioning or positive re-
inforcement learning?

In the mid-1950s, a groundbreaking discovery was made by two Canadians
at the Montreal Neurological Institute, Brenda Milner and William Scoville.

352

Epistemology

A Perforant path

Neocorlex —

Para-

hippocamal

gyrus

Dgc CAS CA1 Subiculum

Entorhinal

cortex

mam milary bodies Parahippocampal

gyrus

l i

Neocortex

Figure 8.14 (A) A schematic representation of hippocampal circuitry. Note that input
from the neocortex reaches the hippoeampus via the parahippoeampal gyrus and the
entorhinal cortex, and output from the hippocampus reaches the neocortex via the
entorhinal and the parahippocampal gyrus. Note also a second input path that projects
from the dentate gyrus (perforant path). Its axons make synaptic contact with the CAS
neurons below the level at which the entorhinal axons make contact. This arrangement
suggests a computational matrix. (From Rolls 1989.) (B) An anatomical diagram
showing the location of the hippocampus in the temporal lobe of the brain, as viewed
from a coronal section (top is rostral, bottom is caudal), where the hippocampal struc-
tures are pulled out from the other tissue to be viewed in depth (facing section is ante-
rior). (Based on Kandel, Schwartz, and Jessel 2000.)

353

How Do Brains Learn?

They observed that a 27-year-old surgical patient, H.M., had completely lost
declarative memory for all postsurgical events. Nonetheless, his IQ was normal,
he retained a normal immediate memory, and had normal memory for events
that occurred in his early life. H.M. had undergone a bilateral surgery of the
medial aspect of the temporal lobe as a treatment for medically intractable
epilepsy. A patient with similar lesions, R.B., was discovered by the Damasios
in the 1980s and has been intensively studies by their lab over the decades. (In
chapter 3, while considering the importance of autobiographical memory to
self-representation, R.B.’s symptoms were briefly introduced.

H.M. could not recall an event that happened a minute ago, even when it
was a salient and signiflcant event, such as receiving the news that his father
had died. This loss of capacity is anterograde amnesia. Although he has
repeatedly met Milner after his surgery, he cannot remember having met her,
even if she left the room only a few minutes earlier. He shows some retrograde
amnesia for events in the few years preceding surgery, but has good recollection
for events in the more distant past. Probing H.M.’s deficits and capacities more
deeply, Milner and colleagues discovered that despite his profound anterograde
amnesia, H.M. can learn a new sensorimotor skill, such as keeping a pencil on
a moving target or tracing a star while watching his hand in a mirror. His skill
improves gradually, much as it does in normal subjects. Even so, he has no
recollection of having encountered the task or having learned the skill. He
shrugs off his newly acquired competence with comments such as, “I am good
at these sorts of things.”

This constellation of data suggested several hypotheses: hippocampal struc-
tures are necessary for learning new things, such as how to find the bathroom
in a new home, but they are not necessary for retrieval of information that was
consolidated when the hippocampal structures were intact, such as how to find
the bathroom in your old home or the details of your first kiss. Nor are they
necessary for acquiring skills, such as mirror-imaged tracing. This profile of
spared and damaged capacities raised fundamental questions: What exactly do
the hippocampal structures do? If cells in the hippocampal structures mediate
remembering experiences, might they be a test bed for Hebb’s hypothesis? How
does information come to be permanently stored in the cortex, and what is the
role of the hippocampal structures in memory?

Targeting all levels of brain organization, from systems and behavior to cells
and molecules, labs began to search for answers. Developing animal models
was crucial, for otherwise the details of anatomy and physiology remain out of
reach. Developing animal models for declarative memory, however, is much

354

Epistemology

Memory

Facts & Events & Skills Habits Classical Priming Habituation

individuals episodes conditioning sensitization

Figure 8.15 A classification of memory. Declarative (explicit) memory refers to con-
scious recollections of facts and events, and depends on the integrity of the medial tem-
poral lobe cortex. Nondeclarative (implicit) memory refers to a collection of abilities
and is independent of the medial temporal lobe. (Based on Squire and Zola-Morgan
1991.)

more difficult than developing them for fear conditioning, since animals cannot
be verbally instructed, nor can they verbally declare what they remember.
Nonverbal techniques had to be devised, but they had to permit testing of de-
clarative memory, not to be confounded with procedural memory or with con-
ditioning. Solving the problems in experimental design required ingenuity in no
small degree (figure 8.15).

One widely used test is the Morris water maze, a clever arrangement devel-
oped by Richard Morris in Edinburgh. The maze is actually a round tub filled
with chalky water containing a small submerged platform. A rat put in the tub
swims until he finds a measure of safety on the platform, as rats prefer to avoid
deep water. Both normal rats and rats with hippocampal lesions can learn the
direct route to the platform, so long as the starting location remains the same.
If the starting location is varied, normal rats still swim directly to the platform.
Rats with hippocampal lesions, however, paddle around the tub searching for
the platform as though the task were entirely new (figure 8.16).

If the platform site is then shifted, normal rats learn where to go in one trial,
whereas the hippocampal rats require many trials. This one-trail learning of
location is a rough analog of declarative memory, and the deficits in hippo-
campal rats are good, if imperfect, analogs of declarative-memory deficits in
hippocampal patients. Incidentally, one advantage of the Morris water maze is
that the results are quantifiable, since you can videorecord the search path and
directly compare the capacities of the control and experimental rats.

An extremely useful nonspatial test was developed by Howard Eichenbaum
and colleagues at Harvard.^'*' They buried cheerios in cups of sand. Each cup

355

How Do Brains Learn?

Paths taken by rats with hippocampal lesions

Figure 8.16 Spatial learning in rats. (A) Rats are placed in a circular arena (about the
size and shape of a child’s wading pool) filled with cloudy water (the Morris water
maze). The arena itself is featureless, but the surrounding environment contains such
positional cues as windows, doors, light fixtures, and so on. A small platform is located
just below the surface. As rats search for this resting place, the pattern of their swim-
ming (indicated by the traces in the figure) is monitored by a video camera. After a
few trials, normal rats swim directly to the platform on each trial. (B) The swimming
patterns of rats with impaired spatial memories — induced by hippocampal lesions —
indicate a seeming inability to remember where the platform is located. (Based on
Purves et al. 2001.)

could be scented with a distinct odor, such as cocoa, coflfee, mint, apple, or
orange. For rats, odors are a powerful cue. They have excellent odor discrimi-
nation and odor memory, which they use to guide behavior. Eichenbaum
showed that rats can learn that when, say, coffee-smelling and cocoa-smelling
cups are available, the reward is always in the coffee-smelling cup, but when
coffee and mint are presented, the reward is always in the mint-smelling cup.
And this overlapping-pairing schedule can be extended; in mint/orange pairs,
the reward is in orange. Rats display their knowledge by digging only in the
cup with the reward (figure 8.f7).

This cunning arrangement permits an interesting test: can the rats use stored
factual knowledge to handle a new situation? Here is how that can be

356

Epistemology

Training sets

The logic

1 sample

If sample = coffee, then

LqJ

cheerio is in cocoa, but

LqJ

coffee

if sample = apple, then

apple

2 choose

I cheerio is in mint

□

cocoa

mint

cocoa

mint

1 sample

LqJ

If sample = cocoa, then
cheerio is in peanut, but

cocoa

if sample = mint, then

mint

2 choose

□

cheerio is in orange

0- I

orange peanut orange peanut

Novel condition

Transitivity

1 sample

coffee

2 predict

Q or g

orange peanut

If sample = coffee, then
cheerio is in peanut, but
if sample = apple, then
cheerio is in orange

apple

E3 O' B

orange peanut

Figure 8.17 The hippocampus and its role in manipulating stored representations to
infer what scent predicts the reward. In this experiment, each training phase has two
parts: first, the rat is presented with a sample scent (a small cup containing sand, the
scent, and a buried cheerio). Second, to get another cheerio, the rat has to choose be-
tween two cups with distinct scents and to learn from the sample what smell predicts the
reward. The next segment of the training phase requires learning a new prediction, but
now the sample scent is the former rewarding scent. Finally, the rat is tested with a novel
situation to see whether it can predict which scent contains the reward on the basis of
past associative knowledge. There is no direct association between a sample scent and
the scent predicting the reward in the novel condition. Normal rats can succeed in this
task, but rats with hippocampal lesions cannot. This is a matter not of spatial learning
but of drawing an inference from earlier experienee. (Based on Bunsey and Eiehenbaum
1996 .)

357

How Do Brains Learn?

addressed; if, after learning that in cocoa/coffee pairs, the reward is in coffee,
and in coffee/mint pairs, it is in mint, and in mint/orange pairs, it is in orange,
present the rat with a novel combination of familiar odors: coffee and orange.
Can the rat correctly use past knowledge and the logic of transitivity to dig in
the mint cup?^^ Normal rats do indeed. Rats with hippocampal lesions perform
at chance.

6.1 Anatomy (Very Briefly)

In pondering what neural mechanisms can explain the behavioral data, we need
to understand the basic anatomy of the brain in general and the hippocampal
structures in particular. In the nervous system, structure is the key to mecha-
nism, and without an understanding of structure, we cannot advance very far in
understanding function.

The entorhinal, perirhinal, and parahippocampal cortices are the sites of a
convergence of inputs from polysensory regions of the frontal, temporal, and
parietal cortices. The hippocampus gets its input from these areas, predom-
inantly via the entorhinal cortex (EC), which has reciprocal connections to a
range of areas involved in emotions and attention. Hippocampal output goes
back to EC, via the subiculum. This loopy circuitry, illustrated in figure 8.14A,
suggests that information repeatedly circulates through the hippocampus and
its associated structures, perhaps involving rehearsal, perhaps involving the
selecting and cleaning up of information as it passes through again, perhaps
subserving recognition memory by filling in and pattern-completing in response
to partial cues.

Neurons in the EC project onto pyramidal neurons in the CAS field of the
hippocampus in a highly regular way; EC projects via the perforant path to the
upper dendritic regions, and via the dentate gyrus (DG) to the lower dendritic
regions (see again figure 8.14A). The perforant path EC synapses have NMD A
receptors, the DG synapses do not. Both exhibit LTP. The axons of the CAS
pyramidal neurons go to two places: (1) they project back onto themselves
{recurrent collaterals) and onto synapses between the EC and the DG con-
nections, and (2) they divide into a batch that projects onto the upper regions
(apical dendrites) of the CAl neurons and a batch that projects onto the lower
regions (basal dendrites). These synapses have NMDA receptors and exhibit
LTP. How does this orderly neuroanatomy serve declarative memory? Now we
need the neurophysiology to tell us how cells respond.

358

Epistemology

6.2 Neurophysiology (Very Briefly)

As discussed in chapter 7, when a rat enters a new environment, specific cells
will attach themselves to specific places that the rat visits, and will respond
vigorously whenever the rat revisits the preferred location. A cell will respond
to its preferred location even when the rat is passively moved in the environ-
ment. A cell tuned to respond to a specific location in one environment will
respond to an unrelated location in a diflcrent environment. As the rat travels
among various environments, the cell shows its preference relative to the envi-
ronment it is in. Overall, the spatial layout of the environment is not topo-
graphically mapped in the hippocampus (see again figure 7.2).

Spatial representation and more

Though place may be a necessary condition for the response, it is not sufficient.
Suppose that the rat is trained on a T-maze to get cheese rewards by alternating
which arm to choose (if it went left last time, it should go right this time). A hip-
pocampal “place cell” will respond when the rat is in one specific place and plans
to go left, but not when it is in that very same place but plans to go right. This
implies that the cell is coding for more than just location (Eichenbaum 1998).
What precisely do the hippocampal cells code for? That is, what are the dimen-
sions of the parameter space that characterize what hippocampal cells repre-
sent? Plan and place? Plan and time and place? Do we even have the vocabulary
adequate to describe whatever these hippocampal cells are representing?

Learning components: CAl

Using genetic techniques, Tonegawa and colleagues selectively blocked the
NMDA receptor in tissue-specific regions of the hippocampus (see Tsien et al.
1996 and McHugh et al. 1996). This means that there is no LTP or LTD in the
CAl cells. Mice that are normal save that they lack functional NMDA recep-
tors only in the CAl region are unable to learn spatial tasks (like the Morris
water maze).

Learning components: CA3

Mice that are normal save for lacking functional NMDA receptors on CA3
pyramidal neurons have a quite different profile. They can learn the spatial task.

359

How Do Brains Learn?

If, however, one or more visual cues surrounding the tub are removed, their
performance falls off. The fewer the cues, the worse the performance. In addi-
tion, after a shift of position of the platform, they cannot learn the new plat-
form location in one trial. Normal rats easily do both. Therefore, the data from
Tonegawa’s lab suggest that declarative memory can be fractionated into
functional components subserved by anatomically distinct regions. From what
we know about recurrence in artificial neural nets, we may surmise that the
recurrent loops on CAS pyramidals could subserve pattern completion, and
hence are needed if the animal is to retrieve information about platform loca-
tion with reduced cues. The DG connection on CAS pyramidals may be crucial
for one-trial learning.

Memory consolidation takes time

Inspired by human studies showing that amnesic patients have better recall for
older memories than for more recent events, Larry Squire and Stuart Zola
(1996) asked the following question: if you need hippocampal structures to
learn new facts, for how long after the exposure to the memorable facts must
those structures be functional for the facts to be retrievable from memory? This
was a fundamental probe into the functional relation between hippocampal
structures and cortical structures, and it had a very revealing answer. In mon-
keys, it turned out that unless the hippocampal structures were intact for about
7-10 weeks following exposure to the to-be-remembered event, the animals’
declarative memory was severely impaired. Thereafter, it was normal. Further
human data suggested that normal hippocampal structures could be necessary
for even longer periods (figures 8.18 and 8.19).

Spatial learning and sleep

Matt Wilson and his colleagues (1994) have shown that during the non-
dreaming phases of the sleep cycle, “place cells” in the hippocampus respond as
though the rat were actually running through the maze it had explored for re-
ward during the waking period. Although questions remain, this activity looks
suggestively like rehearsal. Interference with this activity reduces learning per-
formance. Human data show that deep sleep (stage IV in the sleep cycle) in the
early part of the night and dreaming sleep in the later part of the night are
necessary for skill acquisition. Moreover, the deep sleep and dreaming sleep
must occur within 30 hours of training if learning is to occur, since beyond
those limits, catch-up sleep on the second night fails to compensate.^®

360

Epistemology

Time between learning and surgery (weeks)

Figure 8.18 Two weeks after surgery to remove the hippocampus, monkeys had diffi-
culty remembering reeently learned objects, although their memory for objects learned
many weeks ago was as accurate as that of control monkeys not operated on. Chance
performanee would equal 50 percent correct. (From Squire and Zola-Morgan 1991.)

Learning and neurogenesis

The birth of new neurons (neurogenesis) in adults appears to be restricted to
the olfactory bulb and the hippocampus, though this is not known for sure. The
level of neurogenesis in the hippocampus increases when the animal explores an
interesting environment; it decreases with stress, boring environments, and de-
pression. (New cells can be identified by labeling with an analogue of the DNA
base thymidine: bromodeoxyuridine. This becomes incorporated into new DNA
during cell replication.) What do these new neurons have to do with new
memories? Elizabeth Gould and her colleagues (1999) have recently shown that
the birth of new neurons in the hippocampus is important for new trace con-
ditioning, i.e., learning to associate events that are separated by an interval of
time. The new neurons appear to be unrelated to learning to associate events
that overlap in time (so-called delay conditioning). Learning these latter associ-

361

How Do Brains Learn?

Fast-changing

connections

Siower

changing

connections

Figure 8.19 Information is transferred from hippocampal structures to neocortical
structures, where it gradually becomes consolidated, probably involving such structural
changes in connectivity as dendritic growth. The looping pathways between the neo-
cortex and hippocampal structures, together with data on temporally graded amnesia
following hippocampal loss, suggests that the hippocampus directs memory consolida-
tion in the neocortex by providing continual input, including, and perhaps most impor-
tantly, during sleep. (Based on Squire and Alvarez 1995.)

ations probably depends on the amygdala, not the hippocampus. The Gould
lab showed these effects by well-controlled interference with neurogenesis in the
hippocampus of rats. ^ ^

Evidently, the hippocampal story of learning and memory is becoming in-
creasingly detailed, and much more is now known than when Milner began in
the mid-1950s to explore what H.M. could and could not learn postsurgically.
Nevertheless, there is no shortage of open questions: What exactly does the
hippocampus do? How is it involved in the consolidation of memory? What
mechanisms might be responsible for the consolidation of memory in neural
networks external to hippocampal structures? How well connected must new
hippocampal neurons be to begin to function in learning? How is the right
connectivity established? Why is there a high rate of neuronal turnover in the
hippocampus? These are but a handful of questions that should find answers in
the coming decades.

S62

Epistemology

7 How Do Networks Learn: A Brief Look

Representations, as we saw in chapter 7, appear to be distributed across many
neurons in a network. ANNs, such as Cottrell’s face net, have been instrumen-
tal in providing a working example of distributed representation, and hence an
example of how the brain’s representations might be distributed in populations
of real neurons. An important question, raised but not answered in chapter 7
(p. 296), asked how the adjustment of individual synaptic weights can be ap-
propriately orchestrated across a population so that the population comes to
embody knowledge, such as the knowledge of how to distinguish a male face
from a female face or how to identify a face as that of Winston Churchill. In
short, we want to know how a neural network learns.

The weights cannot, of course, be hand-set in the brain, and they cannot
be hand-set in ANNs either once the number of weights is large enough to ser-
vice an interesting representation. So we are looking for an automated, brain-
plausible weight-adjusting procedure. ANNs are a useful tool for inventing,
exploring, and testing various procedures for changing structure to get meaning
into processing units in a network. Since we want to know how in fact real
populations of neurons adjust their weights, all procedures tested on ANNs
must ultimately be tested in the actual nervous system.

A variety of algorithms have been devised for adjusting the connection
weights to configure a network to embody knowledge about the properties of
the stimulus set. Because these algorithms take a network from a know-nothing
state where the weights are randomly configured to a state where the pattern of
connection weights embodies information allowing the network to categorize
input signals, they are called learning algorithms. Learning algorithms for
automated weight-adjustment divide into two basic kinds: supervised learning
algorithms and unsupervised learning algorithms. The essential difference con-
cerns feedback. The various supervised learning algorithms use feedback about
the network’s behavioral performance in determining weight modification,
whereas unsupervised learning algorithms use no external feedback.

Supervised learning relies on three things: input signals, the net’s internal
dynamics, and an evaluation of its weight-setting performance. Unsupervised
learning depends only on two things: input signals and the net’s internal
dynamics. In either case, the point of the learning algorithm is to produce
a weight configuration that can be said to represent something in the world,
in the sense that when activated by an input vector, the correct answer, or

363

How Do Brains Learn?

Feedback

modes

Unmonitored Monitored Monitored Unmonitored

(no internal) (internai) (internai) (no internai)

Figure 8.20 Strategies for feedback. (From Churchland and Sejnowski 1992.)

approximately correct answer, is produced by the network. Although unsu-
pervised learning algorithms have no access to external feedback, they can
use internal error feedback. When the feedback is external to the organism, the
learning is called “supervised”; when there is an internal measure of error,
the learning is called “monitored” (figure 8.20).

Consider, for example, a net required to learn to predict the next input. As-
sume that it gets no external feedback, but that it does use its previous inputs to
make its prediction. When the next input enters, the net may be able to use the
discrepancy between the predicted input and the actual input to get a measure
of error, which it can then use to improve its next prediction. This is an in-
stance of a net whose learning is unsupervised but monitored. More generally,
there may be internal measures of consistency or coherence that can also be
internally monitored and used in improving the internal representation.

Nets using unsupervised learning can be configured so that the weights
embody regularities in the stimulus domain. For example, the weights of a
two-layered net can be adjusted according to a Hebb rule, so that gradually,
without external feedback and with only input data, the net structures itself to
represent the correlation of feature A and feature B. Beyond the scope of the
simple net are higher-order statistical problems, such as “What is the correla-
tion story for {A,B,C,D} or for {EF,EH,GH}1” Going beyond first-order
correlations is highly desirable, since mapping causal structure in the world, for
example, requires higher-order statistics. To target high-order problems, the
simple two-layer architecture must be expanded to include so-called hidden
units that intervene between external input and behavioral output.

The ability of layers of hidden units to extract higher-order information
is especially valuable when the number of input units is large, as it is, for

364

Epistemology

example, in sensory systems. Suppose that an input layer has n units in a two-
dimensional array, like the retina, or a one-dimensional array, like the cochlea.
If the units are binary, then the total number of possible input patterns is 2". In
fact, neurons are not binary but many-valued, so the problem is actually worse.
Suppose that all patterns (state combinations) are equally likely to occur, and
suppose that one hidden unit represents exactly one input pattern (e.g., that H
and M are highly correlated). This would make it possible to represent any
function in the output layer by suitable connections from the hidden units. The
trouble arises when n is very large, e.g., a million, in which case the number of
possible states is so large that no physical system could contain all the hidden
units. The problem is solvable in part because in this world, not all input pat-
terns are equally likely, and of those that are highly likely, not all are equally
interesting to the animal. So only a small subset of all possible input patterns
needs to be represented by the hidden units.

Accordingly, the problem for the hidden units is to discover which features
systematically occur together or are otherwise “cohorted,” and among those,
which to ignore and which to care about and represent. By means of unsu-
pervised learning, the fundamental correlations can be found, and by means of
supervised learning (punishment and reward), the net can learn what correla-
tions it should represent. As the net runs, hidden units may be assigned states
according to either a linear or a nonlinear function. If linear, there is an opti-
mal solution called principal-component analysis. This procedure is used to find
the subset of vectors that are the best linear approximation to the set of input
vectors. Although principal-component analysis and its extensions are useful
for lower-order statistics, many of the interesting structures in the world can
be identified only via high-order statistics. Consequently, we want a learning
algorithm that can find high-order features. For example, if luminance is taken
as a zeroth-order property, then boundaries will be an example of a first-order
property, and characteristics of boundaries, such as occlusion and three-
dimensional shape, will be higher-order properties. Causal relations between
three-dimensional objects will be even higher order properties. On the face of it,
finding a suitable weight-adjustment rule looks difficult because not only are
the units hidden, they are nonlinear, so mere trial and error strategies will not
get us there. Fortunately, there are solutions.

Independent-component analysis (ICA) is a technique that uses the statistics
of the input signals to identify the independent sources of those signals when
the sources of those signals are unknown. ICA has many applications in tele-
communications and the analysis of biomedical data such as EEG recordings.

5(55

How Do Brains Learn?

If a system is completely naive about the nature of its signal sources and hence
of the parameters of its mixture of input signals, ICA allows it to the find
linear, nonorthogonal axes of its parameter space. Thus it can do blind source
separation. So if you are a tank commander in midbattle speaking to snipers in
the field, for example, ICA can separate the voice out of an extremely noisy
background. ICA uses not only first-order statistics, but also higher-order sta-
tistics, to find what variables in its “world” are statistically independent;
roughly speaking, to find out what in the world is causing its information. It
seems rather plausible that nervous systems may, at various stages of develop-
ment and for various tasks, use ICA learning algorithms to make sense of the
booming, buzzing confusion of signals from the sensory periphery. A newborn
animal is probably a system that is largely naive about the sources of its sen-
sory signals; it has to find the axes of its parameter spaces. ICA can do pre-
cisely that.

Devising a biologically plausible ICA learning algorithm for weight adjust-
ment has been both a computational desideratum and a formidable challenge.
Fortunately, in 1995 Bell and Sejnowski discovered an elegant and powerful
ICA learning algorithm. For example, the Bell and Sejnowski learning algo-
rithm can configure a network to solve face-recognition and lip-reading prob-
lems. On neurobiological realism, it also scores well, since it has been tested
against a range of real physiological data, including the emergence of organized
structures such as ocular-dominance columns in the early visual cortex.

Although research on ICA learning algorithms is still in its infancy, it is a
promising attack on the problem of what principles underlie coordinated
weight adjustment (learning) in populations of neurons. ICA learning algo-
rithms are powerful, but they are not the whole story, for a variety of reasons.
In particular, they assume a stable probability distribution of signals in the
world. Because we are constantly moving our eyes, heads, and whole body, the
probability distribution is not stable for long periods. So work remains to be
done. Nevertheless, even as they stand, ICA learning algorithms take us far
beyond what the ANN pessimists predicted.

Supervised-learning algorithms come in various grades as a function of the
format of the feedback informing the network on the quality of its perfor-
mance. The evaluation may (1) merely say “Good answer” or “Bad answer,”

(2) specify a measure of the size of the error with some degree of precision, or

(3) give rich detail, saying, in effect, “You said the answer was <1,9, 0, 3>, but
the answer should be <4, 9, 3, 3>.” Given the range available in (2), this allows
for a continuum of evaluation formats. As we saw with bee-foraging behavior,

366

Epistemology

diffusely projecting systems such as the dopamine projections from the VTA
mediate reinforcement learning. Reinforcement learning — learning via feed-
back from the environment — is under intense investigation, in both artificial
and real nervous systems. As we saw in chapter 3 in the discussion of the Crush
emulator, when a brain has an internal model of itself and its environment, it
can test tentative plans and, using internal feedback from the model, upgrade
it plans. Inner models with inner feedback allow for enormous complexity in
learning, and they probably lie at the heart of much reasoning and problem
solving.^®

8 Concluding Remarks

What, the skeptic might ask, has all this to do with epistemology, traditionally
conceived? First, note that the philosophical tradition is mu/Ptracked. The
approach embraced here fits with the tradition that reaches back to Aristotle; it
is naturalistic and pragmatic, as opposed to supernaturalistic or a priori. It has
links with the philosopher John Locke (1632-1704), who attended lectures and
brain dissections by the great British anatomist Thomas Willis (1621-1675). It
warms to theories, hypotheses, and models, while at the same time it demands
evidence, data, and testing. It looks for coherence and consilience across well-
established theories, yet it is open to revision of even the most successful
theories.

Together, neuroscience, psychology, ethology, and molecular biology are
teaching us about ourselves as knowers — about what it is to know, learn,
remember, and forget, and about how brains are configured to know, learn,
remember, and forget (figure 8.21). These questions are fundamental epistemo-
logical questions — really, the grand questions — and they are questions that
motivated Aristotle, Descartes, Hume, Kant, and Quine. They are also ques-
tions that motivated Helmholtz, Darwin, Cajal, E. O. Wilson, and Crick. I see
them, one and all, as engaged in naturalized epistemology. I don’t see that it
matters much whether they work in philosophy departments or not.

Not all epistemologists have been so motivated. For some, this is because
they believe that what we call external reality is naught but Ideas created in a
nonphysical mind, a mind that can be understood only via introspection and
reflection on its Ideas. For philosophers who are idealists in this technical sense,
the new developments in cognitive neuroscience will seem irrelevant. And per-

367

How Do Brains Learn?

Social sciences

Ethology

Psychology

Neurobiology

Developmental

biology

Genetics

Figure 8.21 The causal interactions between many levels of structural organization
involved in cognition, and the particular sciences that address the levels and the con-
nections between them. (After Plotkin and Odling-Smee 1981, Huber 2000.)

haps the idealists are right. Nevertheless, idealism, with its admiration for a
priori and introspective strategies, seems to have made little progress on the
nature and basis of knowledge. Certainly there is nothing extant in idealism
that can hold a candle to the emerging explanatory framework of cognitive
neuroscience. This does not mean that idealism is certainly wrong, for it may
merely need more time. It does mean, however, that the idealist’s introspect-
and-contemplate strategy is unappealing for those who wish to make progress
in understanding how we know things.

368

Epistemology

Figure 8.22 A luminance illusion. Four horizontal black bars are separated by white
space. There are two sets of gray bars of identical luminance. The gray bars on the left
look darker than those on the right, which look semitransparent. (From Hoffman 1998.)

There is one major element of truth in the idealist’s approach. As Kant real-
ized, the mind/brain is not just a passive canvas on which reality paints. The
brain organizes, structures, extracts, and also creates (figure 8.22). Reality is
always grasped through the lens of stacks upon dynamical stacks of neural
networks. There is no apprehending the nature of reality except via brains and
the theories and artifacts that brains devise and interpret.

From this it does not follow, however, that reality is only a mind-created
Idea. Rather, it means that we have to keep plugging along, trying to get closer
and closer to the nature of reality, and trying to make fewer and fewer pre-
dictive errors. Our brains — using whatever equipment is available: conceptual,
technological, linguistic, etc. — drum up increasingly adequate models of real-
ity, where the brain, among other things, is part of the reality modeled. We
keep questioning, and we build the next generation of theories upon the scaf-
folding of the last. How do we know the models are increasingly adequate?
Only by their relative success in predicting and explaining. We cannot doff all
lenses — perceptual, conceptual, technological — and make a direct comparison
between hypothesis and reality.

Does this mean that there is a fatal circularity in neuroscience — the brain
uses itself to study itself? Not if you think about it. I use my eyes to study the
eye, but nothing very troubling results from this necessity, since I can study
the eyes of others and reliably generalize to my own case. The brain I study is

369

How Do Brains Learn?

seldom my own, but usually that of other animals, and I can reliably generalize
to my own case. The enterprise of naturalized epistemology involves many
brains — correcting each other, testing each other, and building models that can
be rated as better or worse in characterizing the world. If a hypothesis says that
no new neurons are made in the adult human brain, that hypothesis can be
tested and falsified. If a hypothesis says that memories are one and all stored in
the hippocampus, that can be tested and falsified. Figuring out what is not true
helps us get closer to what is true, whether the subject matter is brains or the
origin of the Earth.

Is there anything left for the philosopher to do? For the neurophilosopher, at
least, there is plenty to do. Questions abound: about the integration of distinct
memory systems, how nervous systems handle time, how far associationist
principles can take us, the nature of representation, the nature of reasoning
and rationality, how information is used to make decisions, how informa-
tion is retrieved, about what information is for nervous systems, why sleep
and dreaming are necessary for learning, and on and on. These are all Big
Questions — big enough for me, anyhow. They are questions where experiment
and theoretical insight must jointly conspire, where creativity in experimental
design and creativity in theoretical speculation egg each other on to unimagined
discoveries. These are questions with deep historical roots reaching back to the
ancient Greeks in 500 b.c. and with ramifying branches extending throughout
the history of Western thought. They are, moreover, questions requiring a syn-
thesis from psychology, neuroscience, and molecular biology. And all this is
what makes them philosophical. Or so it seems to me.

Suggested Readings

Arthur, W. 1997. The Origin of Animal Body Plans: A Study in Evolutionary Devel-
opmental Biology. New York: Cambridge University Press.

Clark, A. 1997. Being There: Putting Brain, Body, and World Together Again. Cam-
bridge: MIT Press.

Cowan, W. M., T. C. Siidhof, and C. F. Stevens. 2000. Synapses. Baltimore: Johns
Hopkins University Press.

Faster, J. M. 1995. Memory in the Cerebral Cortex. Cambridge: MIT Press.

Gerhadt, J., and M. Kirschner. 1997. Cells, Embryos, and Evolution. Oxford: Blackwells.

Grafman, J., and Y. Christen, eds. 1999. Neuronal Pla.s'ticity: Building a Bridge from the
Laboratory to the Clinic. Berlin: Springer.

370

Epistemology

Heyes, C., and L. Huber, eds. 2000. The Evolution of Cognition. Cambridge: MIT Press.

Jeannerod, Marc. 1997. The Cognitive Neuroscience of Action. Oxford: Blackwells.

Johnson, M. H. 1997. Developmental Cognitive Neuroscience: An Introduction. Malden:
Blackwells.

Lawrence, Peter A. 1992. The Making of a Fly: The Genetics of Animal Design. Cam-
bridge, Mass.: Blackwells Science.

Le Doux, Joseph. 2002. Synaptic Self. New York: Viking.

Nelson, C. A., and M. Luciana, eds. 2001. Handbood of Development Cognitive Neuro-
science. Cambridge: MIT Press.

Prince, D. J., and D. J. Willshaw 2000. Mechanisms of Cortical Development. Oxford:
Oxford University Press.

Quartz, S. R. 2001. Toward a developmental evolutionary psychology: genes, develop-
ment and the evolution of the human cognitive architecture. In F. Rauscher and S. J.
Scher, eds.. Evolutionary Psychology: Alternative Approaches.

Schacter, Daniel L. 1996. Searching for Memory: The Brain, the Mind, and the Past.
New York: Basic Books.

Squire, Larry R., and Eric R. Kandel. 1999. Memory: From Mind to Molecules. New
York: Scientific American Library.

Sutton, R. S., and A. G. Barto. 1998. Reinforcement Learning: An Introduction. Cam-
bridge: MIT Press.

Websites

BioMedNet Magazine: http://news.bmn.com/magazine
A Brief Introduction to the Brain: http://ifcsunI.ifisiol.unam.mx/brain/

Comparative Mammalian Brain Collections: http://brainmuseum.org
Encyclopedia of Life Sciences: http://www.els.net

The MIT Encyclopedia of the Cognitive Sciences: http://cognet.mit.edu/MITECS

Ill Religion

This page intentionally left blank

Religion and the Brain

From a scientific point of view, we can make no distinction between the man who eats little
and sees heaven and the man who drinks much and sees snakes. Each is in an abnormal
physical condition, and therefore has abnormal perceptions.

Bertrand Russell (1935)

1 Introduction

Progress in science in general, as well as in neuroscience in particular, has had
an impact on a range of traditional philosophical issues, including the nature of
the mind, the nature of the universe, and the nature of life. One metaphysical
matter that looms large for many people concerns supernatural beings, and the
existence of God in particular. Is there anything we have learned about the
brain that bears upon questions of spirituality?

At the heart of reflections on religion and the brain are three questions: (1)
Does God exist? (2) Is there life after death? (3) What happens to morality if
God does not exist? One and all, these questions are ancient; one and all, they
remain highly current topics of discussion. I raise them here partly because
some developments in cognitive neuroscience have an impact on how we for-
mulate possible answers. These questions have both a metaphysical and epis-
temological dimension, and addressing them pulls together many of the ideas
discussed in earlier chapters. Certainly, they are preeminently philosophical
questions, and the historical tradition is rich with arguments, replies, reformu-
lations, and refutations. In this respect, it resembles the historical tradition of
natural philosophy generally, as humans struggled to figure out the nature of
physical reality and whether their beliefs about fire, life, the Earth, and the
mind are likely to be true.

374

Religion

As one’s understanding of the world expands, conflicts between beliefs are a
regular feature of cognitive life. Some of these beliefs are humdrum (e.g., it
looks like the Moon is about as big as a barn; the data show it actually has a
diameter of 2,000 miles and is 240,000 miles away). Some are more momentous
(e.g., one believes that autism is caused by cold mothering and then discovers
that it has a genetic basis). Some have tumultuous personal elfects (e.g., you
believe that your enduring melancholia is a character flaw and then discover
that you have a serotonin deficiency).

Some discoveries bear upon the belief that there is a life after death. More
specifically, there is tension between (a) the idea of the self as an immaterial and
immortal soul, created by God, and (b) the idea that the mind is what the brain
does, that the human brain is a product of natural selection, and that disinte-
gration of one’s brain in disease and death entails disintegration of one’s mind.
For most of us, it matters how we resolve these tensions, and it matters that we
resolve them in a way that is intellectually satisfying rather than flippant or
ideological.

For me to live after the death of my brain, I must be independent of my
brain. Hence the question of life after death is the springboard question for this
chapter. Nevertheless, because the possibility of an afterlife is so closely asso-
ciated with belief in a Supreme Being, the two matters are, for all intents and
purposes, inseparable. In this closing chapter, therefore, we shall take a closer
look at all three major questions within the neurophilosophical framework
developed for metaphysics and epistemology. My main purpose will be to clar-
ify the issues involved, so that the reader can more productively reflect on them
and figure out how best to resolve inconsistencies and tensions.

2 Does God Exist?

To reduce ambiguity and confusion, we must, as usual, begin with some pre-
liminary semantic geography. Granted that we cannot give a precise definition
of God, what roughly is meant by “God?” This question does need to be asked,
particularly because there are many different religions with many different
characterizations of a Supreme Being. Most readers will have some acquain-
tance with Judaism, Christianity, and Islam. They will consequently have some
conception of what, within those religions, is meant by “Supreme Being.”
Nevertheless, some very large religions, for example. Buddhism, do not really
countenance a Supreme Being with metaphysical status in anything like the

375

Religion and the Brain

way in which Christianity, Judaism, and Islam do. In addition, the early
Greeks believed in an extended family of humanlike gods. Many North Amer-
ican aboriginal cultures believed in animal-like gods. Some pantheists take the
view that God is in all of Nature, including lightning, water, plants, and bac-
teria. Other pantheists consider God to be equivalent to Nature as a whole and
reject the idea of God as a Supreme Person in any sense. The differences be-
tween religious beliefs on the matter of what God is are nontrivial.

Significantly, these differences in the belief tend to be culturally dependent.
As a matter of sociological fact, if someone has a religious belief, it tends to
resemble rather closely the one in which he was raised or to which he was
exposed when young. It is less common for someone, having reached adult-
hood, to canvas the entire range of possible religions and then, on the basis of
evidence and moral suitability, to make a choice. Because of this fact and the
fact that individuals are prone to the conviction that one’s own particular reli-
gion is the only true religion, it is all the more important in this discussion to be
mindful of the great breadth of religious beliefs.

Despite the diversity of religions and the cultural sensitivity of religious
preference, we can continue to make progress on the question by characterizing
a deity in terms sufficiently general and minimalist as to be independent of any
particular religion, so long as it does espouse, at least and at most, one Deity. It
is important to have rough agreement on what we are talking about in order to
have common ground for discussion. In keeping with this compromise, suppose
that by “God” or “Deity” we mean an entity that has some features of a
human being, in the sense that it cares about our welfare, pays attention to
prayers, and has high moral standing. In the interests of nonsectarian dis-
cussion, let us assume that such a Deity is vastly more capable than a human,
perhaps being omnipotent (all powerful), omniscient (all knowing), and omnibe-
nevolent (all good), and that God is responsible for the creation of the universe,
its laws, and the things in it. There is some latitude in each of the three
descriptions, so the qualifier “more or less” should be taken as implicitly riding
along. ^

Though this characterization is not universally satisfactory, it tries to avoid
being empty (as the description “God is everything that is” tends to be), while
meeting the religious expectation that God is the supreme creator of the uni-
verse, can intervene to change the course of the universe, has a deep under-
standing of what is going on, cares about our lives and our suffering, and is the
source of moral standards.^ Some such characterization is needed to make
sense of the subsidiary belief in the efficacy of prayer, for example. That is, if

376

Religion

God lacked power, there would be little point in praying for help. This char-
acterization does not fit the conception of pantheists (God is Nature), but it is
consistent with the theist’s conviction that praying for intervention is appro-
priate. The problem with an even more abstract characterization in terms of
“the Great All” or “something greater than ourselves,” for example, is that for
many theists, it undercuts the solace derived from belief in a caring, sympa-
thetic, personlike God. It also makes many religious practices, such as worship
and seeking redemption, otiose, and it seems too abstract to help out with a
belief in an afterlife. So although the three-ownis description will be less
than satisfactory for some theists, it is minimally satisfactory to many. On that
basis, the description is sufficiently adequate to launch a discussion of God’s
existence.

What are the grounds for believing in a Deity as described? Although there
may be almost as many different reasons as there are believers, for simplicity,
we can discern three general paths:

Path 1; evidence and analysis On the assumption that God is a real existing
thing, there should be evidence of God’s existence, in some form or other. This
will yield empirical knowledge.

Path 2: revelation God reveals himself to certain humans, and these persons
have direct knowledge of God.

Path 3: faith Belief based on faith is independent of what anyone else
observes, believes, or analyzes. Faith may be depicted as “chosen knowledge,”
in the sense that one exercises a choice to believe, perhaps on trust and regard-
less of evidence and analysis, regardless of revelation or lack thereof.

I turn now to the task of discussing, very brieffy, each of the three paths.

2.1 Path 1: Evidence and Analysis

There are two main lines of argument concerning the evidence for the existence
of a Deity. The strongest is the argument from design, and the second is the
argument from first cause.

The argument from design

According to the argument from design, the organization of the cosmos, and
in particular the existence and organization of the biological world, requires

377

Religion and the Brain

intelligent design. And that, the argument continues, requires an Intelligent
Designer. Why? Because it is inconceivable that material organization, such as
the human eye, and biological processes, such as protein manufacture, could
have come into existence by mere chance and with no higher purpose.

The main problem with this argument is that as science has progressed, we
have come to understand how the organization in matter can come about by
entirely natural means. In other words, what seemed inconceivable in the mid-
dle ages is quite well conceived now. In the biological realm, Darwinian evolu-
tion explains how purely natural interactions over many millions of years will
result in diverse biological organisms with complex structures."^ And this has
been borne out by comparative physiology, which allows us to see earlier and
simpler forms of structures, such as the eye, in organisms that appeared on the
planet many millions of years before humans. Variations in a complex struc-
ture, such as the ear as it exists in the barn owl, the human, the dolphin, and
the bull frog, are best explained in terms of biological adaptations to specific
features of an environmental niche through the process of natural selection.

As recently as the later part of the nineteenth century, some eminent scien-
tists, for example Louis Agassiz, thought that the Deity had created all existing
life forms at the same time, complete with the structural features that allowed
them to function best in their particular environmental niche. The fossil record,
the evidence of extinctions, and most recently the discoveries showing physio-
logical homologues and DNA relationships between animals has made this
view highly implausible. For example, the oldest (deepest) layers in the fossil
record contain no mammalian fossils, and no bony fishes, but only simple
organisms, such as trilobites. Dinosaurs, along with many other species, be-
came extinct long before large mammals appeared. In general, the understand-
ing of molecular biology^ and evolutionary biology® has greatly reduced the
appeal of the argument from design. This weakness does not yield a proof of
the nonexistence of God, but it does mean that a traditionally powerful argu-
ment for the existence of God has lost much of its plausibility.

Nevertheless, adherents of the argument from design might wish to vary the
argument by pointing out that there are still explanatory gaps in evolutionary
biology. In particular, science has not yet established where the first replicating
structures — presumably RNA — came from. We still do not have a satisfactory
theory of how proteins fold, or how Monarch butterflies find their ancestral
home, or how the human brain, structurally so very similar to the brains of
other primates, has the capacity for language. Surely, it may be suggested, these
mysteries point towards the intervention of a Supernatural Being. Why? Because

378

Religion

it is inconceivable that such structural complexity could have come about by
sheerly natural means. Our question must be this: how tenable is the logicl

As a preliminary point, recall from earlier discussions that what is or is
not conceivable varies as a function of what the conceiver already understands
and believes. Though it can seem otherwise, conceivability also varies as a
function of what conclusion one antecedently finds attractive. The main point
is that what is or is not conceivable by me is a psychological fact about me,
not a metaphysical fact about the nature of reality. Consider that vitalists —
even twentieth-century vitalists — confidently asserted the inconceivability of
explaining what it is to be alive without appeal to the life force. Even as late as
1910, many physicians found it inconceivable that diseases could be caused by
organisms so tiny that they were invisible to the naked eye. It is not unusual to
find people who consider it inconceivable that the continents move. As dis-
cussed in chapters 4 and 6, arguments from inconceivability need to be backed
up by knowledge, not by ignorance, if they are to make any headway.

Before the development of physical chemistry, molecular biology, and evo-
lutionary biology, the original version of the design argument could lean on
inconceivability and resonate with many people. By now, however, the incon-
ceivability argument has lost its luster even as applied to those phenomena that
are currently unexplained. The problem is that those items still on the list of the
unexplained can well be considered as merely not yet explained, rather than as
items requiring a supernatural explanation. For example, since scientists such as
Leslie Orgel and Jerry Joyce are hard at work trying to understand the etiology
leading up to early forms of RNA, and hence to understand the origin of life, it
is all too conceivable that they will succeed and answers will be found.’

Suppose that scientists such as Orgel and Joyce do not discover an answer.
Would that failure constitute evidence for supernatural intervention? No. It
would show only that the problem was not solved, not that it is unsolvable.
Even if the answer were never discovered, at most that would show that there is
something of which we are ignorant. This is a very banal conclusion, but it
is all that the premise supports. From the premise that we do not know the
originating causes of RNA, can we argue to the conclusion that we do know
the originating cause was supernatural? Alas, using ignorance as a premise is a
fallacy. More precisely, we cannot conclude that we do know the cause from
a premise asserting that we do not know the cause.

There is a further softness in the design argument. As David Hume pointed
out, if it is complexity of organisms in the natural world that motivates us to
postulate an Intelligent Designer, should we not be equally unsatisfied with an

379

Religion and the Brain

unexplained Intelligent Designer?® Should we not want to explain where that
complexity came from? If the existence of naturally occurring organizational
structure is a problem, why is the existence of divinely occurring organization
structure not a problem? If one is uncomfortable stopping the chain of causes
with scientific explanations of naturally occurring complexity, but comfortable
stopping with supernaturally occurring complexity, what, Hume inquired, is
the rationale? Why, so far as empirical evidence is concerned, is the existence of
natural complexity more in need of explanation than the existence of supernatu-
ral complexity?

The argument from first cause

Hume’s objection contains the seeds of a refutation of the first-cause argument.
The first-cause argument advances the hypothesis that the chain of causes in
the cosmos cannot be infinite. So there must be a first cause and that — the
Uncaused Cause, as it were — must be a Supernatural Being. Hume asked two
questions: First, why is an infinitely long chain of causes less plausible than a
finite chain whose beginning is supernatural? Perhaps, so far as the empirical
data reveal, the chain of causes is infinite. Second, if there is a first cause of the
events in the cosmos, why is it more plausible that it is supernatural rather than
natural, such as the Big Bang? To neither of his arguments is there an empiri-
cally grounded, satisfactory response. Notice again that this critical analysis
does not constitute a proof of the nonexistence of God. It says only that there
are logical fiaws in the argument for the existence of God.

The argument from evil

Hume also challenged the hypothesis that the empirical evidence points to the
existence of an omnipotent, omniscient, and omnibenevolent Deity. His prin-
cipal argument runs as follows: if we are to infer from available evidence that
God is benevolent, the existence of natural evil presents a huge problem.
By “naturally occurring evil,” he meant such tragedies as infants born with
horrible diseases and dying slow and painful deaths. He included the routine
struggle for existence that animals endure, as well as the miseries caused by
storms, droughts, plagues, insanity, and fioods. Listing the miseries of life on
the planet, for humans and other animals, is a long, sad business.

Hume concluded that the existence of sulfering is prima facie evidence
against the Deity as described. Either (a) he does not know of the misery in the

380

Religion

world, in which case he is not all-knowing, or (b) he does know but cannot
prevent it, in which case he was not all-powerful, or (c) he does know and can
prevent it, but prefers not to, in which case he is not all-benevolent. The argu-
ment does not prove that no such God does exists, but it does show that if one
draws on available evidence, one would never rationally infer that God is om-
nipotent, omniscient, and omnibenevolent.

A classic set of responses have been made to the argument, essentially all of
which Hume anticipated and attempted to refute. First, it might be suggested
that if you quantify good and suffering, then on balance there is more happi-
ness than misery. Hume’s answer was thorough. First, the estimated ratio of
happiness to suffering is, needless to say, only a very rough estimate. Given the
evidence of misery on the planet, it is as reasonable to estimate a preponder-
ance of misery over happiness as vice versa. Aside from the difficulties in mak-
ing the estimate precise, the misery that does exist is, even if counterbalanced
by lots and lots of happiness, a terrible lot of misery. Why, Hume asked, is
there so much misery? Would not a truly benevolent, powerful God find a way
to mitigate much of the pain and suffering of innocents, the misery and horror
inflicted on the powerless and virtuous? Therefore, misery blocks any inference
to the existence of an omnipotent and omniscient and omnibenevolent God.
Hence the argument from evil cannot be set aside.

A second response suggests that suffering exists because humans have free
will and have chosen to do evil things. Misery, therefore, is God’s just punish-
ment for sin. Hume’s answer was this: even if this response explains the misery
that happens to humans who have sinned (which he doubted), it does not ex-
plain the terrible suffering that befalls innocent human infants or animals.

A third response suggests that evil must exist in the world if humans are to
know the difference between good and evil. Hume’s answer was this: well,
surely a small amount of evil would suffice for that purpose. Additionally,
omnipotence is not a trifling capacity, so if the Deity is omnipotent, he should
be able to make that knowledge available without making the innocent suffer.
Being omniscient, he should know how to achieve this.

A fourth response argues that what we consider evil is not really evil from
God’s point of view, i.e., that suffering is not really a bad thing from the Divine
perspective. Hume suggested that this argument was a shocking refusal to take
seriously the theist’s own claim that the Deity is genuinely benevolent and cares
about our welfare. If, he argued, God does not consider the terminal cancer of
an innocent child as a bad thing, then it is hard to see how he can be considered
benevolent, in terms of what we mean by “benevolent.” If such suffering is

381

Religion and the Brain

consistent with supernatural benevolence, then that is a kind of benevolence so
alien to us that we recognize it only as evil. If God is not genuinely benevolent,
in terms of what we standardly mean by “benevolent,” then, Hume suggested,
God is not to be embraced as a moral authority.

The argument from evil, as it is called, does not constitute a proof for the
nonexistence of a benevolent God. But it does show that if one aims to use
reason and evidence to draw an inference about the nature of the Deity, there is
a prima facie problem in inferring the existence of a God who is omnipotent,
omniscient, and omnibenevolent.

This is a very fast summary of the main positions in the discussion of the exis-
tence of a Supernatural Being. One line of evidence not yet considered, how-
ever, concerns the possibility that God reveals himself only through a highly
select group of humans, who then convey their revelation to others as a basis
for religious belief. We turn now to arguments based on revelation.

2.2 Path 2: Revelation

Some individuals claim to have personal contact with a Supreme Being. In the
present context, the question is whether the reports are credible, and hence
whether one can infer the existence of God on the basis of the individual
reports of revelation. Moreover, this question arises whether one has the expe-
rience oneself or one knows of the experience only by report. It is well known
that many such reports are not credible for any of a variety of reasons. For
example, the subjects may be suffering psychiatric disorders, which are identi-
fied on completely independent grounds. If so, there are more straightforward
explanations of the alleged revelation consilient with the science of the brain.
Other subjects may be on drugs, such as LSD, peyote, or other hallucinogens.
There are reports of subjects exposed to the elements, such as lost sailors, who,
suffering physical exhaustion and the extremes of cold, thirst, and hunger,
experience a recurring sense of a nearby rescue boat, looming out of the fog,
but invisible. Mountaineers, suffering anoxia (lack of oxygen) also report
experiencing the feeling of someone marching along behind, always out of
sight, but definitely close by, and occasionally propelling the mountaineer for-
ward. Some subjects have ultimately confessed to fraud or have been shown to
have lied for profit. Some subjects have had sexual orgasm in a religious con-
text and mistakenly, if reasonably, have interpreted it as direct contact with
God.®

382

Religion

Under what conditions should we accept the report of direct knowledge of
God as a basis for belief? Since the third path, namely faith, is not yet the topic
of discussion, I shall assume that the question concerns when it would be rea-
sonable to think that such a report is highly probably true. Consequently, the
standards will be comparable to the standards for reasonable belief generally.
That is, what is the evidence for and against? Are there other more plausible
explanations for the experience or the report of the experience? What other
tests could be deployed to see whether the hypothesis survives falsification?
And so on. If someone reports an observation of something remarkable, it is
always wise to approach the claim in an open-minded but careful fashion.

By the very nature of the case, these claims are hard to test. That is, the
experiences are limited to a small number of individuals, the events at issue do
not occur with any regularity, and conditions tend not to be replicable. Caution
and skepticism are therefore particularly appropriate. It has been claimed, for
example, that Mark Anthony was touched by God, though he evidently suf-
fered from epilepsy. Epilepsy has also been suggested as the actual basis for the
conversion of St. Paul.

These difficulties notwithstanding, some neurologists have recently suggested
that there is a particular class of claims that deserve to be taken seriously as
reports of genuine revelations. Because these cases involve subjects with a neu-
rological disorder, namely temporal-lobe epilepsy, I am particularly eager to
understand and evaluate the arguments for their credibility. First, what are the
phenomena?

Epilepsy is a complicated condition in which a large population of excitatory
neurons in the cortex fire in abnormal synchrony (figure 9.1). Focal epilepsy
begins in a restricted area, such as the hippocampus or frontal cortex, and may
spread to adjacent areas. During the seizure, the subject may lose consciousness
or experience odd feelings. The effect of the seizure depends on the location of
the focus. If, for example, the focus is the primary motor cortex, then the sub-
ject may display involuntary muscle contractions; if it is the primary somato-
sensory cortex, there may be tingling or other odd sensory experiences. In a
form known as complex partial seizures, the regions involved are limbic struc-
tures of the temporal lobe, along with the orbitofrontal cortex (figure 9.2).
Subjects in whom this form of seizure occurs may briefly display automatized
behavior, such as laughter, and even some routinized behavior, such as sweep-
ing the floor. How aware they are during the seizure remains debatable, though
they tend to have no memory of events that occurred during the seizure. The

383

Religion and the Brain

Figure 9.1 Examples of EEG recordings from different forms of epilepsy. Abbrevia-
tions: LT, left temporal; RT, right temporal; EE, left frontal; RE, right frontal; LO, left
occipital; RO, right occipital. The black dots on the hemispheres indicate the approxi-
mate recording sites. (A) Normal adult EEG. (B) Brief excerpts from an EEG taken
during a grand mal seizure: (1) Normal recording preceding the attack. (2) A sense of
impending seizure, followed by onset of the attack. (3) Clonic phase of the attack during
which there may be sudden movements or cries. (4) Period of coma. Shaded areas rep-
resent regions picked up by electrodes placed on the scalp. (Erom Kolb and Whishaw
1990.)

Table 9.1 Manifestations of Complex Partial Seizures

• Affeetive (fear and anxiety most common)

• Automatisms (perseverative, do novo, gelastie, dacrystie, procursive, and other
seemingly purposeful actions)

• Autoscopy

• Cognitive dissonance (e.g., deja vu, depersonalization, dreamy states)

• Feeling of a presenee

• Epigastric and abdominal sensations, indeseribable but recognized as outside normal
experience

• Hallucinations (any modality)

• Sensory illusions and distortions of ongoing perceptions (e.g., metamorphopsia,
separation of color from its boundary, spatial extension of the form constants,
paracusia, umkehrtsehen, etc.)

• Synesthesia

• Time dilatation and contraction

• Psychosis

• Forced thinking

• Memory intrusions

• Hypersexuality and hyposexuality

• Autonomic dysregulation

• Contraversive movements

• Speech arrest and ictal aphasia

Source: Cytowic 1996.

Figure 9.2 A patient with eomplex partial seizures underwent video-EEG telemetry
monitoring, during whieh several of his usual seizures were recorded. The patient is
shown above during different phases of a typical seizure, ineluding his description of the
prodromal aura (a foul “sulfurlike” smell and taste) (A), evolving later to confused be-
havior, left-leg clonic twitching, and an attempt to elimb from the bed (B), and postietal
(Todd’s) paralysis of the left arm immediately following the event (C). (Courtesy of Drs.
Erik St. Louis and Mark Granner, Department of Neurology, Roy J. and Lucille A.
Carver College of Medicine, Elniversity of Iowa).

385

Religion and the Brain

epileptic focus may be associated with scar tissue, though often the etiology of
the focus is unknown.

In generalized epilepsy, there is simultaneous widespread synchronous activ-
ity, and subjects typically lose consciousness. Grand mal seizures involve loss of
consciousness, and subjects tend to fall down, and their limbs may jerk about.
Petit mal seizures tend to be briefer, less severe, and do not involve loss of
consciousness. Patients seem briefly vacant or “not at home” during petit mal
seizures. The root cause of generalized epilepsy is not well understood.

Focal epilepsy can be experimentally produced in animals by applying
directly to the cortex drugs that block the activity of inhibitory neurons. For
example, high doses of penicillin applied to the surface of the cortex blocks
inhibitory neurons and produces seizures. Focal seizures can also be produced
by repeated electrical stimulation of the cortex. Generalized seizures are more
difficult to produce experimentally. Intravenous doses of penicillin adminis-
tered over time can result in an animal prone to generalized seizures. In certain
baboons, a generalized seizure can be induced by flickering lights. Some breeds
of dogs, namely beagles and St. Bernards, are particularly susceptible to epi-
lepsy, and thus constitute an important experimental model. Epilepsy is nor-
mally treated with drugs that increase the activity of inhibitory neurons. This
treatment is usually effective in controlling the seizures.

Clinicians have long known that a small percentage of subjects with an epi-
leptic focus in the temporal lobe are prone to be hyperreligious. These same
subjects may also show hypersexuality and hypergraphia (they tend to write
an unusual amount). Dostoyevsky is sometimes cited as one such case, and
Ramachandran and Blakeslee (1998) discuss one such subject, Paul. There are
also reports from a small percentage of temporal-lobe epileptics that just prior
to manifesting an epileptic seizure, they experience unusual feelings. They may
say, for example, that they felt a gathering awe and dread or that they felt a
huge deluge of emotions. A handful say that their rather indescribable experi-
ences made them feel that they were connected with an overwhelmingly pow-
erful being, that they felt a great presence nearby. Some say that during the
seizure, they came in intimate contact with an invisible God. Ramachandran’s
subject did claim exactly this.

Let us consider now the possibility that in this highly restricted class of epi-
leptic patients, God does in fact make himself known to the patient during
the seizure, as Paul clearly believed. We need to consider the evidence for
and against. The strongest evidence in favor of the hypothesis is, of course, the
sincere reports of honest subjects. How strongly, if at all, does that evidence

386

Religion

support the conclusion that subjects who report contacting God during an epi-
leptic seizure truly do contact God during the seizure?

One major reservation derives from investigations by neuroscientist Michael
Persinger. His strategy was to simulate, albeit weakly, some conditions of a
temporal-lobe seizure in normal volunteers by exciting temporal-lobe neurons
using an oscillating magnetic field focused on the temporal lobe. His aim was
to see whether the experiences described by the special class of temporal-lobe
epileptics could be produced in normal subjects.^®

The results were interesting. Under such activation, subjects did report
highly unusual feelings. About 80 percent of Persinger’s subjects report feeling
as though there was a presence nearby, sometimes just out of view. Others, if
they are atheists, may say they feel a “oneness with the universe.” At least one
person had a visual hallucination involving an angelic appearance — a great
deal of light, rushing sounds, sublime feelings. A New York psychiatrist
described his feelings in nonreligious terms as a “resolution of binaries.”

Persinger’s data lend support to the conclusion that these experiences are
one and all the result of a particular kind and distribution of neural activity,
just as pain, hunger, and fear are neural effects. That seizures in the tem-
poral lobe should produce extraordinary feelings is predictable from the known
connectivity of temporal-lobe structures. That is, there are connections to
structures known to play a role in experiencing emotions; the amygdala, hypo-
thalamus, brainstem, and orbitofrontal cortex. The amygdala, as discussed in
chapter 3, is known to involve feelings of fear. The hypothalamus has sub-
regions involved in sex, hunger, thirst, and other desires and these will be
subject to increased activation in a unusual fashion if there is generalized stim-
ulation to the temporal lobe. If the activation spreads, as it does during a sei-
zure, then because of their connectivity, the cingulate and orbitofrontal cortices
are likely to suffer abnormal levels of synchronized activity. Random activation
of these cortical areas will also have a powerful role in the generation of an
odd blend of emotions and feelings. Heightened activity of the hypothalamus,
amygdala, brainstem, cingulate cortex, and orbitofrontal cortex may trigger
many strong feelings all at once, in a composition highly unusual in day-to-day
life. For example, there may be feelings of dread, joy, elation, anxiety, hunger,
and sexuality all at the same time. This pathological activation of emotion cir-
cuitry may be interpreted by the subject in many ways, depending on how his
past experiences situate him.

What are we to make of this? Persinger’s data raise the possibility that be-
cause we can induce the effect in normal subjects by altering neural activity in

387

Religion and the Brain

the temporal lobe, then probably the effect in both normals and epileptics has
nothing to do with contact by a Supreme Being. Part of our obligation in
evaluating revelation hypotheses is to determine whether other more probable
explanations for experiences of God are available. For this reason, Persinger’s
experiments are very important. They do support a natural (as opposed to
supernatural), neurally based cause. They do not prove it beyond all doubt, of
course, but they are supporting evidence. In any case, proof beyond all doubt is
rare for scientific hypotheses generally.

How might the theist refute these skeptical worries? One strategy is to say
that the Persinger data do not prove that the experiences of a special class of
temporal-lobe epileptics and those of the experimental volunteers have essen-
tially the same cause. Perhaps, it might be suggested, God really does contact
the epileptics, but not the volunteers. While this possibility may be worth
entertaining, our question is, Which hypothesis is more probable? Given Per-
singer’s results, the burden of proof is now on the theist to show why a natural
explanation for both the epileptic and the normal volunteers is not sufficient.
Consider a parallel example. Suppose that you believe your wounds heal by
divine intervention, even if those of everyone else heal by natural processes.
Then the burden of proof is on you to show why your case is different, and why
one type of explanation cannot serve all relevantly similar examples.

Another strategy for dealing with Persinger’s results is to view all the expe-
riences — those of epileptics, anoxics, and normal volunteers — as confirming
contact with God. Although this is a possible avenue, it has only a quirky ap-
peal. Both skeptics and believers find it farfetched to suppose that God would
choose to manifest himself through one particular pathological condition,
namely temporal lobe seizures. And why would he manifest himself via a
simulated temporal-lobe seizure? Is it reasonable to expect that God’s presence
can be invoked electromagnetically? Logically, the Persinger results are not, of
course, a proof of the nonexistence of God, nor even of the illusory status of the
experiences at issue. They are important because they drain probability from
the hypothesis that the experiences provoking God-reports are truly experiences
of God. Our question is whether, given the data, that hypothesis is probably
true. Given the analysis and the interpretation so far, the hypothesis is not
compelling.

Consider now a completely different argument. Suppose we say that the
temporal lobe, precisely because its stimulation can, albeit rarely, give rise
to experiences described in religious terms, must be specialized for this pur-
pose. Just as stimulation of the visual cortex gives rise to visual experiences, so

388

Religion

Stimulation of the “God module” gives rise to religious experiences. Since
purely natural selection cannot account for the emergence of such a cortical
specialization in humans, it may be argued, the explanation for its existence
must appeal to a Divine Cause. That is, God must have set in place this
neurobiological arrangement so that humans could have the capacity to know
God directly.

In response, it is important to emphasize again that it is only a tiny fraction
of subjects with temporal-lobe epilepsy who report their experiences as religious
in nature. Second, patients who come to the clinic reporting seizures are nor-
mally treated straightaway with seizure-controlling drugs, so the experience
they report is typically an inaugural event, not a recurring event. Consequently,
they cannot be observed and tested to see whether rapturous experiences occur
on later occasions, or whether there is a correlation between the severity of
an episode and its capacity to produce a rapturous experience, or whether the
religious denomination of the subject predicts the religious interpretation of
the experience. These are human subjects, not experimental animals, and we
cannot delay treatment of a potentially dangerous condition to experiment on
the nature of rapturous experiences.

A further problem, touched on earlier, is epistemological. In their reports,
subjects try to make some sense of the experience. That is, they experience
various feelings, and they usually wish to interpret those feelings. We know
from Persinger’s results that the feelings induced by temporal-lobe stimulation
are very hard to describe. Moreover, as I noted, not everyone interprets the
feelings as feelings of God. When they are given strange experiences, people
tend to look for explanations that are comparably strange, even though the
cause is ultimately neurobiological. We have to remind ourselves that strange
experiences, such as hallucinations, weird dreams, or out-of-body experiences,
may have quite ordinary explanations in terms of atypical neural activity.
Strange experiences may seem to us to be full of meaning and portent, however
humble their causal origin, but the strangeness of the experience tells us noth-
ing about whether the cause of the experience is equally strange.

Quite likely, cultural factors influence whether one interprets the temporal-
lobe-excitation experience as of God — of an external Supernatural Being — or
in some other fashion. That is, you might already have to have religious belief
of a certain kind to interpret the experience as of God. At least one would want
to know whether a pantheist temporal-lobe epileptic interprets the experience
in the same way as an epileptic who is a Baptist or Muslim or Buddhist or
Satanist or atheist. Consider also that temporal-lobe structures have a role in

389

Religion and the Brain

memory retrieval, and that memory retrieval often involves representation of
events or persons not currently present. For example, one can now remember a
particularly fearsome first-grade teacher, with all the terror, anxiety, and sense
of overwhelming foreboding experienced in early childhood. Is it possible that
part of what happens is that the emotion complex generated by Persinger-style
temporal-lobe stimulation activates recollections of individual persons who
provoked such feelings in the past, such as the fabled first-grade ogre-teacher?
This is sheer conjecture, of course, but it is conjecture with an eye toward
experiment.

Finally, though the argument depends on the idea that natural selection
could not possibly explain the existence of religious feelings, in fact it is very
easy to imagine that feelings are part of the more general neurobiological ap-
paratus that serves to bind humans into social groups, where they feel loyalty
to a leader and to the group. ^ ^ Consistent with individual variation in biology
generally, it may not be surprising if some individuals are more inclined to re-
ligious affiliation, just as some humans seem more blessed with mathematical
ability or a sense of humor than others. Some individuals may feel strong urges
to humble themselves before a great leader or blindly follow his dictates. Others
may be strongly independent and find the whole idea of worship and blind
loyalty sheerly baffling.

These considerations detract from one’s confidence that the reports in ques-
tion are confirming evidence of Divine Revelation to a select few. They do not
absolutely rule out the possibility that the experiences of religious temporal-
lobe epileptics are divinely caused, but they do generate skepticism to which
there seems to be no convincing counterargument.

2.3 Path 3: Faith

In the previous two sections, I assumed that the issue of whether God exists
is best approached by evidence and argument. My assumption itself may be
challenged, however, on grounds that the method suitable for religious belief is
not evidence and argument, but faith. To a first approximation, this means
adopting or rejecting the hypothesis on the basis of private motivation, as
opposed to evidence and argument.

To abandon evidence and argument as the basis for religious belief is no
small thing, however. For one thing, this means one could as easily have faith
that no deity exists, or that the deity that does exist is essentially evil, or that
there are an infinitely many competing gods, or any number of other variations.

390

Religion

It means that sharing in rational argument to figure out the most reasonable
answer thus far is essentially at an end. Against the satanist, who simply has
faith in the Devil and his great powers, for example, one then has no argument,
since argument is beside the point.

Hume considered this option, and his worry was that it puts an end to the
back and forth of exploratory conversation. In its place arise undesirable ele-
ments, such as pathology and exploitation. So long as one’s religion is personal
and private and has no implications beyond the life of the believing individual,
this may not matter. But as soon as the believer uses his belief to give him
moral or political authority with respects to others, then, in Hume’s view, the
trouble begins.

Once you have backed into the faith corner, you have no recourse against
terror and repression in the name of religion, no recourse against bigotry,
demagoguery, misogyny, or abuse posing as religion. You have no basis for
criticism of cruel religions. This is precisely because faith is not a matter of
evidence and analysis, not a matter of argument and criticism. It is belief
independent of those things. If the faith option works for decent folks, it works
every bit as well for scoundrels; if faith is acceptable for religion, then deeming
it as unacceptable in other domains is just special pleading. Faith has been used
not only by the charitable and the kind, but also by those who insist on their
divine right to unquestioned rule or their divine right to destroy another tribe
or enslave women. How can we reason with any of these persons if they claim
faith, and faith alone, as the basis of their reasons? As is well known, those who
adopt the faith option are often in open conflict on what the right faith is, what
range of questions should be decided by faith, and what moral standards ought
to be imposed. This is not surprising, since cultures vary and private motivation
is as varied as human kind.^"^

Is religious belief, or more specifically, belief in a Deity, universal"} Is it in-
nate! The claim that religious belief is both universal and innate is often raised
in a discussion about faith, and in particular, in a discussion of why two indi-
viduals have the same faith. Even if such claims are indeed true, it is unclear
precisely what conclusion regarding faith is to be drawn. As discussed earlier,
innateness of a belief is no guarantee of the truth of the belief. Innateness of a
belief is no guarantee even of its utility in survival, since it may be, for example,
an innocuous consequence of something else that is adaptive or even a mildly
deleterious consequence of something else that is very useful. In any case, there
is no compelling reason from child-development studies to think that such a
belief is innate. Some children, for example, respond to their first introduction
to the idea of a Deity with surprise and incredulity.

391

Religion and the Brain

Moreover, many entirely normal people lack a belief in a deity. Chinese folk
religions and Buddhism joinly have roughly a billion adherents but have no
surpraphysical niche for a being like the Christian God. In addition, there are
upwards of a billion pantheists, agnostics, atheists, and assorted nonbelievers.^^
This implies that theism is not universal. A trap awaiting the unwary is to push
the idea that atheists, agnostics, and such are really believers who pretend
otherwise. Why? Because belief in a Deity is universal. Unfortunately, the
argument has now become circular, since the universality of belief is used to
defend the universality of belief.

Even if belief in a Deity is not universal, the personification of nature is very
common. Personification is a typical response when we do not understand the
cause of important events, for example, why the Sun was eclipsed, why a comet
appeared, why tornadoes are spawned seemingly from nothing, why bubonic
plague kills a quarter of the people in a town, why apple trees are fruitful one
year but not another. Bewildered and with no better theory at hand, we fall
back on the richest and most powerful explanatory resource we have, namely
our framework of mental categories normally used to explain human and ani-
mal behavior. That is, we use our theory of minds. We ask the storm to subside,
entreat the plants to flourish, invoke the good auspices of the Moon to help
with fertility, and consider ourselves to be punished by wrathful forces. We
make sacrifices in hopes of appeasing anger or currying favor.

My point is not that this is foolish; it is not. It is a worthy attempt to make
sense of the universe, using the best explanatory resources at one’s disposal.
But the progress of science consists in the slow replacement of psychological
explanations for natural phenomena by more successful natural explanations.
Science gives rise to technology and the means for predicting a tsunami or
hurricane, for preventing spread of infections, for helping bees to pollinate
apple trees in a cold spring, and so on. By and large, these manipulations are
more effective than animal sacrifices or prayer. It is useful to remember that
most of the dominant religions came into existence in quite ancient, prescien-
tific times, when animal sacrifices to the gods seemed the best way to try to
influence fertility, the weather, health, and the course of battle.

I say all this while recognizing that for many people, faith in a deity is a
highly positive part of their lives. Their faith may be what sustains them, day
after day, in dealing with their own sorrows, anguish, and tragedies. It may be
instrumental in defeating alcoholism, coping with depression, and providing
courage to do terribly difficult things. I do not doubt that faith, of one kind or
another, can be a central element in people’s lives. In addition to religious faith.

392

Religion

the conviction of an athlete that he will win, or of a dancer that he will not
stumble, or of a soldier that he will survive, is singularly elfective in aiding
performance, even if it is no guarantee of success. Engaging the enemy half-
heartedly is a recipe for defeat. Nevertheless, in this context, in this discussion,
what is at issue is not so much the psychological role of faith and conviction,
but whether what is believed by faith is probably true and, more specifically,
whether faith that a deity exists constitutes evidence that a deity exists.

One last, but not minor point. Some have claimed, for example, Paul Davies,
that science too has its articles of faith. Davies says that scientists “accept as an
act of faith that the universe is not absurd, that there is a rational basis to
physical existence manifested as a lawlike order in nature.”^® He thinks that
this is essentially the same as faith that a Deity exists.

Davies’s suggestion to reduce the intellectual distance between religious faith
and science is contrived. All scientific hypothesis are evaluated on the basis of
evidence and argument, none are considered too sacred to be criticized or
investigated or refuted. No instrument is deemed reliable by faith alone, no
hypothesis is adopted once and for all on faith alone. If there is order in the
universe that we can understand, we do not believe this on faith but because
certain laws seem to hold, no matter how stringent the test or how often re-
peated. On any given occasion, we typically make many assumptions, certainly,
but our assumptions are always defeasible, that is, we acknowledge that they
could be false and may need to be tested on another occasion. Indeed, the his-
tory of science is full of examples where it was the seemingly safe assumption
that was ultimately overturned. Earth is the center of the universe and does not
move — these two assumptions seemed irrefutable, safe, necessarily true, known
with absolute certainty, part of the holy plan. And yet Galileo and Copernicus
convinced us that they are indeed false. The whole point about faith is that you
do not criticize or test or marshal evidence and argument. The whole point
about science and progress in science is that you do.

In the end, one makes up one’s own mind about these things. My considered
opinion is that no argument for the existence of God is even a little convincing,
and to that degree, I find the hypothesis that God exists to be improbable at
this time. I do believe this not on faith, but on the basis of evidence and argu-
ment. Like Hume, I see the price of the faith option as exorbitant in its moral
and political consequences, and hence to be avoided as a moral duty. But we
learn new things all the time, and new discoveries can take us by surprise. For
all that we can be certain of now, the hypothesis or some variant might some-
day be rendered probable.

393

Religion and the Brain

3 Is There Life after Death?

As discussed in earlier chapters, the preponderance of evidence supports the
hypothesis that mental states are brain states and mental processes are brain
processes. On this hypothesis, what thing exists to survive the death of the
brain? What kind of substance would it be, and how could it have the emo-
tions, knowledge, preferences, and memories that the brain had when it was
alive? How can it be related to those activities in the brain that make me me?
Reasonable answers need to be forthcoming if the hypothesis that there is life
after death is to win credibility.

So far as I can determine, there are no answers that cohere enough to make
some sense of the life-after-death hypothesis. The preponderance of the evi-
dence indicates that when the brain degenerates, mental functions are compro-
mised, and when the brain dies, mental functions cease. The suggestion that the
whole body is resurrected after death does address the problem, but so far the
evidence for resurrection is not persuasive. Old graves contain old bones, and
decaying flesh is devoured by scavengers.

Is there any positive evidence that something, we know not what, does in fact
live on after the death of the brain? There are, certainly, many reports that
purport to provide confirming evidence. Because I cannot consider them all
here, I shall restrict myself to the following pertinent observations. So many of
the claims that rest on the intervention of a psychic medium have been shown
to be fraudulent that a general suspicion of these claims is as prudent as the
general suspicion one has toward get-rich-quick investments. Many claims to a
previous life are either openly concocted, confabulated, or a matter of unwit-
ting selectivity of evidence. By “selectivity of evidence,” I mean that one pays
attention to events that, with suitable interpretation, could be construed as
confirming one antecedently favored hypothesis, while ignoring or explaining
away in ad hoc fashion events that could be disconfirming.

For a made-up illustration of selectivity of evidence concerning an afterlife
and a previous life, consider this story. A child draws a picture of a scene with a
farmhouse, apple trees in the yard, a dog sleeping under the tree, and so forth.
It reminds his mother of Great Grandfather Smith’s house. Indeed, little Billy
has some of Great Grandfather Smith’s physical traits, including his curly red
hair and his hot temper. The mother asks the child about his picture, and the
source of his ideas. “Do you remember ever seeing a place like this?” she
queries. If she prompts him, the child will begin to agree, as psychologists have

394

Religion

repeatedly shown, that he remembers this place, remembers the dog, and so
forthd® Later he may quite innocently embellish all these “memories” with
details from parental conversation, family albums, and so forth. Billy may even
discover, perhaps without conscious knowledge, that he is encouraged to con-
fabulate his earlier life, where his confabulations get conceptualized as the re-
covery of hidden memories.

Great Grandfather Smith, in Billy’s “recovered memories,” killed a grizzly
with a mere bowie knife, built a snow house in a blizzard, and talked to quail
and coyote. Nobody else remembers these events, but that is not troubling,
since Grandfather was somewhat reserved. His mother, we may imagine, does
not work hard to test the hypothesis that Billy is the reincarnation of Great
Grandfather Smith. When she does ask a question about Great Grandfather
Smith’s life that Billy cannot answer, this is soothingly explained away by say-
ing that Billy has forgotten that particular of his previous life. She ignores
countervailing evidence, she tends to notice or remember only confirming evi-
dence. This is not because she is openly mendacious. Quite the contrary. She
is inadvertently fooling herself She wants to believe. This fable illustrates se-
lectivity in considering evidence, and it is something to which we all are prone.
Consequently, we have to work hard to be as tough-minded with respect to
hypotheses we hope are true as we are with respect to those we fear are true.

Whether all accounts of reincarnation share the weaknesses illustrated in the
fable is not known, but because so many that have been studied do, and be-
cause one does not want to be gullible, we need to exercise careful scrutiny,
case by individual case. Why do we not all enthusiastically believe Shirley
McLaine’s claims of her earlier, colorful lives? Partly, I think, because her
accounts seem to suffer from the selectivity-of-evidence problem just outlined,
partly because her claims are conveniently untestable, but also because they have
the indelible stamp of fantasy. Her “earlier lives” are enviably glamorous; they
are not the lives of a poor peasant grubbing about with running sores and bent
back. Typically, reports of previous lives are replete with storybook appeal:
handsome heroes, beautiful queens, and romantic deeds. Surely, there were
many more hungry, stooped peasants than there were pining, gothic princesses,
yet these tend not to be the “previous lives” channeling reveals. Or is it perhaps
that only glamorous persons are reincarnated and the humble ones stay dead?

Recently, evocative descriptions provided by patients who very nearly died
have become a source of interest to our question. Visual experiences involving
tunnels with shimmering lights at the far end, feelings of great peacefulness,
feelings that one is being led on a journey, and sometimes the experience of
seeming to see one’s body below on a gurney are typical of experiences called

595

Religion and the Brain

“near-death experiences.” These experiences are alleged to be evidence that the
patients have experienced the otherworld of the afterlife. As always, we must
weigh the evidence for and against, and reflect on whether there might be more
down-to-earth explanations.

Several obstacles suggest that caution is in order. First, these experiences
seem to be somewhat unusual (about 35 percent) among those patients who are
very close to death but who revive. Selectivity of evidence makes them seem to
confirm an afterlife, despite the existence of other cases where the resuscitated
patient reports no such experiences.

Second, the conditions are not those of a controlled experiment, and one
wants to know whether any of these patients are encouraged to “remember”
events that, in their current stressful circumstances, they attribute to experi-
ences “while dead.”

Third, the reports are reports from patients whose brains are under great
stress; they are anoxic (oxygen deprived) and awash in norepinephrine, pre-
cisely because they are close to death. Brains under stress may produce many
abnormal activities, including involuntary movements, strange speech, un-
usual eye movements, and unusual experiences. Severe anoxia, for example in
drowning, is known to result in feelings of peacefulness, once the panic phase
has passed. Some people have used self-strangulation as a means of inducing
anoxic ecstasy. Anoxia resulting from breathing nitrous oxide (so-called laugh-
ing gas) can produce ecstatic feelings and feelings of having glimpsed profound
truths. William James says he experienced “metaphysical illuminations” while
intoxicated on nitrous oxide, though what he wrote on these occasions was, by
his admission, sheer gibberish. Nitrous oxide stimulates neurons that release
endorphins (the brain’s endogenous opiates), which is why it can be used as an
anesthetic. Endogenous endorphin release, along with some suggestibility per-
haps, is the probable cause of ecstatic effects.^® In this respect, therefore, the
problem is similar to the problem with the reports from the cases of temporal-
lobe epileptics who experience “religious feelings” during a seizure.

Fourth, as noted in chapter 3, out-of-body experiences, as well as other dis-
orienting and depersonalizing experiences, can be produced artificially, for ex-
ample with the anesthetic ketamine or with LSD. It is not unlikely that the
neuronal explanations for the ketamine experiences and the near-death experi-
ences are very similar. Moreover, as Francis Crick has pointed out in conver-
sation, the out-of-body claims could be tested a little more directly by asking
whether the patient saw an object that could be seen only if he was where he
said he was, such as floating out the hospital window. So far as I can tell, this
sort of test has not been systematically undertaken.

396

Religion

Although the skepticism and caution with respect to claims about past and
future lives are justified, we should keep an open mind about the possibility
that a genuinely testable case will emerge. If a prima facie case does emerge, it
will indeed be of the greatest importance to examine it carefully and systemat-
ically, to avoid inadvertent contamination of memory, to do everything possi-
ble to rule out fraud, to check the claims against what is known about the facts,
to consider other possible explanations, and so forth. The record of examined
cases makes one less than optimistic that such a case will survive scrutiny, but
one must not rule out the possibility that it will.

But is the prospect of extinction not unsettling? Is it not disappointing and
frightening? It may be all these things, but it need not be. One can live a richly
purposeful life of love and work — of family, community, wilderness, music,
and so forth — cognizant that it makes sense to make the best of this life.
Arguably, it is less painful to accept that miseries are just a part of life than that
they are punishment or trials or that one’s prayers are being ignored. Arguably,
it is comforting to assume that matters of justice and desert need to be
addressed in the here and now, not deferred to an afterlife. Finding peaceful
solutions, redressing wrongs, seeking reconciliation and compromise, express-
ing love, maximizing the significance of each day that one is alive — these things
may make more sense than pinning too much hope on an iffy hereafter. When
all is said and done, the truth is still the truth, however grim it turns out to be.
If there is no life after death — if that is the truth — then wishing it were other-
wise will not make it otherwise.

4 If God Does Not Exist, What Happens to Morality?

This question is really about the foundations of moral standards. It is a ques-
tion about why certain behavior is considered wrong or unfair or punishable,
and contrariwise, why some behavior is esteemed, praised, or encouraged. It is
about what it means to say that an action is wrong. It is a profoundly impor-
tant question, and one that has been the topic of intense discussion in many
cultures since ancient times. In the Western philosophical tradition, we are
deeply indebted to the Greek philosophers in the fourth and fifth centuries b.c.,
for it was they who launched systematic discussion of the problems.

The most insightful and concise examination of the idea of religion as the
source of ethical standards is found in Plato’s early dialogue Euthyphro. Soc-

397

Religion and the Brain

rates and Euthyphro, a high-born priest-about-Athens, meet on the steps of the
law courts. Socrates is awaiting trial for encouraging the young to inquire into
everything — orthodoxy, common assumptions, and revered authority. Offi-
cially, he is charged with “corrupting the youth of Athens.” Euthyphro is at the
law courts because he means to prosecute his father for murder. Of the attend-
ing circumstances, we learn that his father had punished a servant who, while
drunk, had killed a slave. He tied up the servant and left him into a ditch while
he went in search of advice concerning what should be done with the miscreant.
The servant died before the father returned. In contrast to the ever-perplexed
Socrates, Euthyphro is smoothly confident of his moral opinions and certain of
his superiority to the common run in moral matters. Euthyphro’s legal project
intensifies the drama of this dialogue, especially when he reveals himself to be
breathtakingly insensitive to the moral ambiguity permeating his action against
his own father.

The stage set, Socrates begins his methodical inquiry by asking Euthyphro,
“So, in virtue of what is an action right?” Euthyphro has no hesitation: “What
is right is what I am doing now; namely, prosecuting the wrongdoer.” When
Socrates urges him to provide a more general answer, Euthyphro eagerly
responds, “What is right is what is dear to the gods.” Or, in more contempo-
rary language, what is right is what the gods say is right. Moreover, he candidly
confides that he is unusually fortunate in having special knowledge of what is in
fact dear to the gods. The theory that religion is the source of morality is now
on the table.

In the ensuing conversation, Socrates extracts from the sanctimonious
Euthyphro the damaging admission that the gods do not appear to give a
single unequivocal answer concerning the propriety of Euthyphro’s legal action
against his father. Described as bringing a murderer to justice, the action may
be favored by the gods, at least to judge by some available myths. Described as
high-handed action against one’s aging and well-meaning father, it is forbidden
by the gods, at least to judge by other available myths. Neither myths nor gods
converge on a single answer concerning what is right in this case.

In his questioning, Socrates rebuffs the smugly self-righteous, wherever they
may be. At the same time, he undermines the pretension to special knowledge
of what the gods want, leading us to realize that clerical claims to special
knowledge can be crassly self-serving. Additionally, Socrates uses these argu-
ments to show what we all implicitly know and live by, namely that there
are recognizably justified, if unlistable, exceptions to any set of rules, whether
they are thought to come from the gods or not. We draw upon some deeper

398

Religion

understanding of what is right than the rule itself in order to arrive at a rea-
sonable judgment about when we are morally required to deviate from the rule.
The rule is only a superficial image of that deeper understanding.

Having shown Euthyphro to be muddled in his conviction that what is right
is what the gods say is right, Socrates then lays out the catastrophic problem
with Euthyphro’s popular answer. He points out that the claim is actually
ambiguous, and then asks which one of the two possible interpretations is the
intended meaning: (1) do the gods say something is right because it is right, or
(2) is something right because the gods say so?

On the second alternative, morality is sheerly a matter of the decision or de-
cree of the gods. This means that the gods’ decree that something is right makes
it right. For example, if the gods say, “It is right to sacrifice other humans,”
then it is right, regardless of any other feelings or thoughts humans might have
on the subject. If, on the other hand, they say it is wrong, then it is wrong. On
this interpretation, morality depends solely on the choices, whimsical or other-
wise, of the gods (or God). This gives morality a decidedly arbitrary character,
as though it is only incidentally connected to humans’ needs. Moral standards
must have more to them than that.

Consider instead the first, and more appealing, interpretation, namely, that
the gods say something is right because it is right. The trouble with this,
according to Socrates, is that it implies that the gods are merely spokesmen
concerning what is and is not morally appropriate. That is, the rightness of an
act derives from something other than the gods. Consequently, on this alterna-
tive, what makes something right must be independent of the gods, in the sense
that it would be right whether or not the gods were available to broadcast the
news. If so, points out Socrates, then the question concerning the foundation of
morality has not even begun to be answered. To put his worry another way, we
want to know why the gods say of an action that it is right. Whatever the ex-
planation might be, that is what we want to understand when inquire into the
nature of moral standards. If we cannot make progress on that, then we our-
selves do not understand what properties make some actions right and some
wrong. So the disappointment with this alternative is that the gods are not the
source of moral standards. Socrates also expects us to see the general lesson
implicit in his disambiguating the seemingly clear phrase “What is right is what
the gods say is right.” For it is only by questioning and reasoned analysis that he
unmasks the flaws infecting each of the two possible interpretations.

This short dialogue is deeply disturbing. Here is Socrates, awaiting trial for
less than heinous behavior, namely, his habit of questioning practices and
principles that the authorities do not wish to have questioned. We know that he

399

Religion and the Brain

will be found guilty and will be put to death by poison. He is calmly but keenly
aware of moral hypocrisy, self-satisfied indecency, and intolerance masquerad-
ing as morality. But withal, he is aware of the abiding necessity of morality for
civilized community life. He provokes us to see that even though moral deci-
sion making is part of everyday life, the everydayness of morality, along with
our feelings of moral certainty, should not lull us into thinking we understand
the foundations of morality and the origin of moral understanding. More spe-
cifically, we should be wary lest we deceive ourselves with self-serving ration-
alizations about special knowledge derived from special relationships with the
gods.

Socrates’ argument in Euthyphro does not prove that a Supreme Being is
not the source and basis of morality. It importance is owed to its articulation of
the many problems that beset the view that morality is grounded in a Supreme
Being. That is, it reminds us that there are (1) evidential problems with the hy-
pothesis that a deity exists, (2) problems with inferring God’s benevolence,
given natural disasters, suffering, and misery, and (3) problems in knowing
precisely what it is that God commands, given the difiiculty of access and con-
flicting accounts. The real power of the argument in Euthyphro is that it points
us in new directions to understand the nature of morality. It suggests that we
entertain a more naturalistic explanation of morality than supernatural com-
mand. It also makes us curious to understand what that explanation might be
and how it might connect with our evolutionary history.

Socrates’ challenge was taken up by Plato and his students, and by many
thinkers ever since. In particular, genuine progress was made by Aristotle, who,
perhaps better than anyone then or since, grasped the point that the codifica-
tion of moral rules can at best define the central prototypes, but that the lived
moral life requires coming to understand why those rules apply when they do,
and when exceptions are justified. Among other things, he grasped that im-
precision and inexactness in moral precepts are unavoidable and require us
continually to reflect and deepen our moral perspective. Aristotle clearly real-
ized, moreover, that inflexible laws that leave no discretionary room for wise
judgment, such as zero-tolerance laws, often do serious harm.^^ He also seems
to have understood that impulses of sympathy and caring, and for making a
moral community, are just part of our human nature, however our natures
came to be as they are. Our advantage over Aristotle is that we have some
understanding from evolutionary biology of how making a moral community is
part of human nature.

The tradition of moral thinkers who have wrestled with Socrates’ question is
rich indeed. Aristotle, perhaps the greatest of all moral thinkers, I have already

400

Religion

mentioned. The tradition also includes Hume, who outlined the fundamental
role of emotions; Kant, who tried to understand the authority of reason; John
Stuart Mill, who formulated utilitarianism; the pragmatists, John Dewey, and
Oliver Wendell Holmes, who were sensitive to evolutionary biology, the pro-
vincial nature of a person’s moral perspective, and the pragmatic need for
democratic institutions; and William Hamilton and Edward Wilson, who gave
us insight into the biological basis for altruistic behavior in nonhumans. In
the past several decades, some philosophers have tried — by drawing on molec-
ular biology, evolutionary biology, anthropology, and legal history — to achieve
a more satisfactory synthesis of the foundations and nature of morality.

An undertaking of great importance, this new synthesis blends the sciences
of who we are with pragmatic common sense and the wisdom of lives lived.
Will it produce a set of absolute rules, applicable for all times in all places?
No. Will it provide an algorithm for solving specific moral questions, such
as whether stem-cell research is morally acceptable? No. Will it constitute an
unquestioned authority of what is right? Not this either.

What it can begin to do is to provide a naturalistic perspective on the foun-
dation of moral judgment, and in so doing, it can help us disentangle ourselves
from many myths about morality. In disentangling ourselves from the myths,
we may become even more keenly aware of our obligation to think a problem
through rather than just react or blindly follow a rule. Ethics, in Aristotle’s
view, is the most difficult of subjects, not least because the exigencies of life
demand that decisions be rendered now and actions taken now, but also because
there is such a thing as moral wisdom, an understanding acquired through long
experience and relentless reflection, much of which is scarcely articulable.

In sum, the way things seem to stand is this: (1) There are overwhelming
problems with the idea that morality can be grounded in a God. (2) The best
modern candidates for understanding the grounding of morality are naturalis-
tic. So (3) morality and moral understanding are unlikely to require the exis-
tence of God.

5 Concluding Remarks

There are various kinds of feelings that, for want of a better term, we may
describe as sublime. Kant used the word “sublime” to characterize those expe-
riences one has when, for example, viewing a wild storm or soaring moun-
tain peaks, though the wilderness is but one source of such feelings.^® We feel

401

Religion and the Brain

ourselves awed by the immensity or complexity or power of things in various
conditions. In its general use, “sublime” means “of outstanding spiritual, in-
tellectual, or moral worth; tending to inspire awe, usually because of elevated
quality.”^^ In keeping with current usage, sublime feelings might also be
described as broadly spiritual, without implying anything about the existence of
spirits. They can be associated with different things, including music, art, reli-
gion, science, parenting, and intellectual discovery. By temperament, some
people may be inclined to enjoy these feelings in the wilderness rather than in
church, or in the opera house rather than in the delivery room — or they may
enjoy them in all these conditions. Whether there is a supernatural reality that
corresponds to a supernatural interpretation of these feelings is, however, a dif-
ferent matter, and in general, the existence of supernatural beings seems rather
doubtful.

The point is not that these various sublime feelings are unreal. The point is
not that because the feelings are brain effects, they are unworthy or incon-
sequential. Real, the feelings certainly are. Worthy, in and of themselves, they
also are — the more so if they inspire kindness and virtue, the less so if they
inspire cruelty and terror. Do we trivialize a sublime feeling if we appreciate its
dependence on the brain? No

Internet Archive Audio

Featured

Top

Images

Featured

Top

Software

Featured

Top

Books

Featured

Top

Video

Featured

Top

Mobile Apps

Browser Extensions

Archive-It Subscription

Save Page Now

Full text of "Brain-wise : studies in neurophilosophy"

See other formats