Académique Documents
Professionnel Documents
Culture Documents
BRADFORD BOOKS
A NEUROCOMPUTATIONAL PERSPECTIVE
The Nature of Mind and the Structure of S cie n ce
Paul M. Churchland NEUROCOMPUTATIONAL
"O ccasionally s o m e o ne does something genuinely creative in philosophy. Such a c
complishments are rare enough that w hen they do o ccur they are not too hard to spot
If doubts had remained, the publication of Paul Churchland’s A Neurocomputational
Perspective: The Nature of Mind and the Structure of Science surely s e c u r e s his
P e r s p e c tiv e
position a s one of today's most creative philosophers."
— Robert N. McCauley, Philosophical Psychology The Nature of Mind and
If w e are to solve the central problems in the philosophy of science, Paul Churchland the Structure of Science
argues, w e must d raw heavily on the resou rce s of the emerging s c ie n c e s of the mind-
brain. A Neurocomputational Perspective illustrates the fertility of the con cepts and
data drawn from the study of the brain and of artificial networks that model the brain.
T hese co n cepts bring unexpected coh ere n ce to scattered issues in the philosophy
of s cie n c e, n ew solutions to old philosophical problems, and n ew possibilities for the
enterprise of s c ie n c e itself.
Paul M. Churchland is P rofessor of Philosophy and a member of the Cognitive
S c ie n c e Faculty at the University of California at San Diego.
CHUPP 978026253106190000
0-262-53106-2
PAUL M. CHURCHLAND
A Neurocomputational Perspective
A Neurocomputational Perspective
The Nature of Mind and the
Structure of Science
Paul M. Churchland
A Bradford Book
The MIT Press
C am bridge, M assachusetts
L on d on , England
First МIT Press paperback edition, 1992
© 1989 Massachusetts Institute of Technology
All rights reserved. No part of this book may be reproduced in any form by any electro
nic or mechanical means (including photocopying, recording, or information storage
and retrieval) without permission in writing from the publisher.
This book was set in Palatino by Asco Trade Typesetting Ltd., Hong Kong, and printed
and bound in the United States of America.
p. cm.
Bibliography: p.
Includes index.
ISBN 0-262-03151-5 (HB), 0-262-53106-2 (PB)
1. Intellect. 2. Consciousness. 3. Cognition. 4. Neural
computers. 5. Science— Philosophy. 6. Psychology—
Philosophy. 7. Science and psychology. I. Title.
BF431.C48 1989
128'. 1— dc20 89-9398
CIP
For M ark and A nne,
tw o lam ps against the darkness
Contents
Preface xi
PART I
The N atu re of Mind
C h ap ter 1
Elim inative M aterialism and the Prepositional A ttitudes 1
1 W hy F olk Psychology Is a Theory 2
2 W hy F olk P sychology M ight (Really) Be False 6
3 A rgum ents against Elim ination 9
4 The Conservative N ature o f Functionalism 12
5 B eyond F olk P sychology 17
C h ap ter 2
Functionalism , Qualia, and Intentionality 23
1 F ou r P roblem s concerning Q ualia 23
2 The Problem o f N onstandard R ealizations 34
3 Functionalism and M ethodology 44
C h ap ter 3
R eduction, Qualia, and the Direct Introspection of Brain States 47
1 Intertheoretic Reduction 47
2 T heoretical Change and P erceptual Change 52
3 Thom as N agel's A rgum ents 56
4 Jackson's K now ledge A rgum ent 61
C h ap ter 4
K now ing Qualia: A Reply to Jackson 67
1 The P ersisten t Equivocation 67
2 O ther Invalid Instances 71
3 A G enuinely N onequivocal Know ledge A rgum ent 72
4 C onverting a Third-Person A ccount into a First-Person A ccount 74
viii Contents
C h apter 5
Som e Reductive Strategies in C ognitive N eurobiology 77
1 Introduction 77
2 L am inar Cortex, Vertical C onnections, and T opographic M aps 79
3 Sensorim otor Coordination 82
4 C oordinate T ransform ation: Its Physical Im plem entation 88
5 Cortex w ith M ore than Two Layers 96
6 Beyond State-Space Sandwiches 97
7 The R epresen tation al P ow er o f State Spaces 102
8 C oncluding R em arks 108
C h apter 6
Folk Psychology and the Explanation of H um an Behavior 111
1 O bjections to the T heoretical V iew 112
2 A n A lternative Form o f K now ledge R epresentation 122
3 A dden du m : C om m entary on D ennett 125
C h apter 7
R eductionism , C onnectionism , and the Plasticity of H um an
C on sciou sness 129
1 The P lasticity A rgum ent 131
2 The C ulturaTE m bedding O bjection 132
3 Conclusion 134
PART II
The Structure of Science
C h ap ter 8
The O ntological Status of O bservables: In Praise of the
Superem pirical Virtues 139
1 H ow van F raassen ’s Problem C ollapses into H ume's Problem 140
2 The Prim acy o f the Superem pirical Virtues 145
3 T ow ard a M ore R ealistic Realism 149
C h apter 9
O n the N atu re of Theories: A N eurocom putational Perspective 153
1 The C lassical View o f Theories 153
2 Problem s an d A lternative A pproaches 154
3 E lem entary Brainlike N etw orks 159
4 R epresentation and Learning in B rainlike N etw orks 163
5 Som e F unctional P roperties o f Brainlike N etw orks 171
6 H ow F aithfu lly D o These N etw orks D epict the Brain? 181
7 C om putational N euroscience: The N aturalization o f Epistem ology 188
8 C oncluding R em arks 194
Contents
C h apter 10
On the N atu re of Explanation: A PDP A p p roach 197
1 Introduction 197
2 C onceptual O rganization in P D P N etw orks 200
3 R ecognition and U nderstanding 206
4 P rototype A ctivation : A U nified Theory o f Explanation 209
5 Inference to the Best Explanation 218
6 Com parison w ith Earlier M odels 223
C h ap ter 11
Learning and C onceptual C hange 231
1 Introduction 231
2 M u ltiple C onceptual Com petence 234
3 C onceptual Change versus C onceptual R edeploym ent 236
4 W hat D rives Conceptual Change? 241
5 A utom ated Science 250
C h ap ter 12
P erceptual Plasticity and Theoretical N eutrality: A Reply to Jerry
F od or 255
1 The Etiology o f P erceptual B elief 257
2 The Sem antics o f O bservation Predicates 271
3 Conclusion 278
C h ap ter 13
C on ceptu al P rogress and W ord-W orld Relations: In Search of the
E ssen ce of N atural Kinds 281
1 N atu ral Kinds and Scientific P rogress: The P utnam -K ripke View 282
2 N atu ral Kinds as Law -B ound K inds: Som e Virtues, Consequences, and
D ifficulties 287
C h ap ter 14
M oral F acts and M oral K now ledge 297
1 The E pistem ology and O ntology o f M orals 297
2 M oral Prototypes and M oral D evelopm ent 298
3 P raxis, T heoria, and Progress 300
References 305
Index 315
г
Preface
They could do, it began to seem , just about anything. They could
perform sensorim otor coordination, direct gaze accurately despite
ch an ges in head position, recognize subtle similarities am ong sonar
retu rn s, pronou n ce w ell-articulated speech from printed text, recog
nize three-dim ensional shapes independently of the angle of illu
m ination, p arse sentences into gram m atical types, recognize voiced
p h on em es, predict the folding of protein m olecules, correctly recog
nize colors across changes in illumination, and so forth. A n d they did
all of this by learning from a store of relevant exam ples. The practical
necessity of sim ulating the relevant parallel netw orks w ithin con ven
tional serial com puters im posed limits on how large the m odel n et
w orks could be, since the num ber of com putationally active synaptic
connections explodes as roughly the square of the num ber of n eu ron
al units in the netw ork. A nd a further limit w as im posed by the n e
cessity of com puting thousands of adjustm ents for each of these
m an y connections during the course of training w ith the back-
p ropagation algorithm . Even so, netw orks of a few hun dred units
w ere readily sim ulated, and they proved to be capable, after training,
of som e quite arresting displays of cognitive sophistication.
The questions then becam e, Just how are these netw orks doing all
this? H ow do they m anage to find unity and structure in the chaos
and diversity of their sensory inputs? W h at form s of rep resentation
are being generated inside them ? W hy do som e netw orks learn better
than others on the sam e training corpus? Fortunately, sim ulated n et
w orks, unlike biological brains, can easily be tw eaked into giving up
their hidden secrets. O ne need not insert m icroelectrodes one by one
and blindly into a m ass of fragile cells. O ne can w rite the sim ulation
p ro gram to read out on com m and the sim ultaneous and undegrad ed
m icrobehavior of entire populations of artificial n eurons.
H ere the results are highly intriguing, and to the eye of a philo
sop h er of science they are very revealing. A lm ost im m ediately one
could begin to see lessons that bear directly and system atically on
issues central to the philosophy of science: the nature of theories, the
theory-laden ness of perception, the nature of conceptual unification,
the virtues of theoretical simplicity, the nature of parad igm s, the
kinem atics of conceptual change, the character of abductive infer
ence, and the nature of explanatory understanding (see especially the
discussions in ch apters 9, 10, and 11).
This brings m e to the second m ajor them e of this volum e. I began
this preface by detailing how it is no longer possible to d o m ajor w ork
in the philosophy of m ind w ithout draw ing on them es from the phi
losophy of science and the several sciences of the m ind-brain. I w ish
now to su ggest that the instruction and inform ation h as begun to
xvi Preface
The quick tour through the m ajor them es of the volum e w ould focus
on ch apters 1, 5, and 6 from part 1, and chapters 9, 10 and 11 from
part 2.
The research here presen ted w as supported by a gran t from the
Institute for A d vanced Study, by a grant from the Social Sciences and
H um anities R esearch Council of C an ada, by a sabbatical leave from
the U niversity of M anitoba, and by the U niversity of California, San
D iego. Turning finally to individuals, I am pleased to be able to thank
L arry Jord an, David Zipser, Terry Sejnowski, David R um elhart, F ran
cis Crick, Rodolfo Llinas, A nd ras Pellionisz, Philip K itcher, Stephen
Stich, and Patricia C hurchland. W ithout their w ise and generou s
com p an y to draw on, I could never have w ritten the later p ap ers in
this volum e, and w ould never have been able to com plete the argu
m ent begun in the earlier papers. Thanks in large m easure to them , I
h ave a coh erent story to tell.
Elim inative m aterialism is the thesis that our co m m on sen se con cep
tion of psychological p h enom ena constitutes a radically false theory,
a theory so fundam entally defective that both the principles and the
ontology of that theory will eventually be displaced, rath er than
sm oothly red u ced , by com pleted neuroscience. O ur m utual u n d er
standing and even our introspection m ay then be reconstituted w ith
in the conceptual fram ew ork of com pleted neuroscience, a theory
we m ay exp ect to be m ore pow erful by far than the com m on sen se
p sych ology it displaces, and m ore substantially integrated w ithin
physical science generally. M y purpose in this p aper is to explore
these projections, especially as they bear on (1) the principal elem ents
of co m m onsense psychology: the propositional attitudes (beliefs,
desires, e tc.), and (2) the conception of rationality in w hich those
elem ents figure.
This focus rep resents a change in the fortunes of m aterialism .
Tw enty years ago em otions, qualia, and "ra w feels" w ere held to be
the principal stum bling blocks for the m aterialist program . W ith these
barriers dissolving (Feyerabend 1963a; Rorty 1965; C hurchland 1979),
the locus of opposition has shifted. N ow it is the realm of the inten
tional, the realm of the propositional attitude, that is m o st com m only
held up as being both irreducible to and ineliminable in favor of any
thing from w ithin a m aterialist fram ew ork. W h eth er and w hy this is
so, w e m u st exam ine.
Such an exam ination will make little sense, h ow ever, unless it is
first ap p reciated that the relevant netw ork of com m on sen se concepts
does indeed constitute an empirical theory, w ith all the functions,
virtues, and perils entailed by that status. I shall therefore begin w ith a
brief sketch of this view and a sum m ary rehearsal of its rationale. The
resistance it en cou nters still surprises m e. A fter all, com m on sense
has yielded up m any theories. Recall the view that space has a pre-
This ap p roach entails that the sem antics of the term s in our familiar
m entalistic vocabulary is to be u nderstood in the sam e m an n er as the
sem antics of theoretical term s generally: the m eaning of an y theoret
ical term is fixed or constituted by the netw ork of laws in w hich it
figures. (This position is quite distinct from logical behaviorism . I
deny that the relevant law s are analytic, and it is the lawlike con n ec
tions generally that carry the sem antic w eight, not just the con n ec
tions w ith overt behavior. But this view does accou n t for w hat little
plausibility logical behaviorism did enjoy.)
M ore im portant, the recognition that folk psychology is a theory
provides a simple and decisive solution to an old skeptical problem ,
the problem of other m inds. The problem atic conviction that another
individual is the subject of certain m ental states is not inferred de
ductively from his behavior, nor is it inferred by inductive analogy
from the perilously isolated instance of o n e's ow n case. R ather, that
conviction is a singular explanatory hypothesis of a perfectly straight
forw ard kind. Its function, in conjunction with the background
laws of folk psychology, is to provide explanations/predictions/
u n d erstan din g of the individual's continuing behavior, and it is credi
ble to the d egree that it is successful in this regard o ver com peting
h yp oth eses. In the m ain, such h ypotheses are successful, and so the
belief that others enjoy the internal states com p reh en d ed by folk
p sychology is a reasonable belief.
K now ledge of oth er m inds thus has no essential d ep en den ce on
know ledge of on e's own m ind. A pplying the principles of our folk
p sych ology to our behavior, a M artian could justly ascribe to us the
familiar run of m ental states, even though his ow n p sych ology w ere
very different from ours. H e w ould not, therefore, be "gen eralizin g
from his ow n c a s e ."
A s well, introspective judgm ents about o n e's ow n case turn out not
to h ave any special status or integrity anyw ay. O n the p resen t v iew , a
sp on tan eous introspective judgm en t is just an instance of an acquired
habit of conceptual respon se to one's internal states, and the integrity
of any particular response is alw ays contingent on the integrity of the
acquired conceptual fram ew ork (theory) in w hich the respon se is
fram ed. A ccordingly, on e's introspective certainty that o n e 's mind is
the seat of beliefs and desires m ay be as badly m isplaced as w as the
classical m an 's visual certainty that the star-flecked sphere of the
heaven s turns daily.
A n o th er con un drum is the intentionality of m ental states. The
"p rop ositional attitu d es," as Russell called them , form the system atic
core of folk psychology, and their uniqueness and anom alous logical
p roperties have inspired som e to see here a fundam ental contrast
4 The Nature of Mind
(5) (x)(p)(q)[{(x desires that p) & (x believes that (if q then pj)
& (x is able to bring it about that q))
з (barring conflicting desires or preferred m ean s,
x brings it about that q)].
tive cultures, the behavior of m ost of the elem ents of nature w ere
un derstood in intentional term s. The w ind could know anger, the
m oon jealousy, the river generosity, the sea fury, and so forth. These
w ere not m etaphors. Sacrifices w ere m ade and auguries undertaken
to placate or divine the changing passions of the gods. D espite its
sterility, this anim istic approach to nature has dom inated our history,
and it is only in the last tw o or three thou san d years that w e have
restricted F P 's literal interpretation to the dom ain of the higher
anim als.
Even in this preferred dom ain, h ow ever, both the con ten t and the
success of FP have not advanced sensibly in tw o or three thousand
years. The FP of the G reeks is essentially the FP w e use today, and w e
are negligibly better at explaining h um an behavior in its term s than
w as Sophocles. This is a very long period of stagnation and infertility
for any theory to display, especially w hen faced w ith such an en or
m ou s backlog of anom alies and m ysteries in its ow n explanatory d o
m ain. Perfect theories, perh ap s, have no need to evolve. But FP is
profoundly im perfect. Its failure to develop its resources and extend
its range of success is therefore darkly curious, and one m u st query
the integrity of its basic categories. To use Imre Lakotos's term s, FP
is a stagnan t or degenerating research p rogram and has been for
millennia.
Exp lan atory success to date is, of course, n ot the only dim ension in
w hich a theory can display virtue or p rom ise. A troubled or stagnant
theory m ay m erit patience and solicitude on oth er grounds, for exam
ple, on groun d s that it is the only theory or theoretical ap proach that
fits well w ith oth er theories about adjacent subject m atters, or the
only one that prom ises to reduce to, or to be explained by, som e
established background theory w hose dom ain encom p asses the do
m ain of the theory at issue. In sum , it m ay rate credence b ecause it
holds prom ise of theoretical integration. H ow does FP rate in this
dim ension?
It is just here, perh ap s, that FP fares p oorest of all. If w e approach
H om o sapiens from the perspective of natural history and the physical
sciences, w e can tell a coh erent story of the species' constitution, d e
velop m ent, and behavioral capacities that encom p asses particle p h y
sics, atom ic and m olecular theory, organic chem istry, evolutionary
theory, biology, physiology, and m aterialistic neuroscience. That
story, though still radically incom plete, is already extrem ely p ow er
ful, outperform ing FP at m any points even in its ow n dom ain. A nd it
is deliberately and self-consciously coh eren t w ith the rest of our d e
veloping w orld picture. In short, the greatest theoretical synthesis in
the history of the hum an race is currently in our hands, and parts of it
Eliminative Materialism and the Propositional Attitudes 9
tion of the four essen ces— as im m aterial spirits— will ap p ear both
feckless and u nn ecessary given the p ow er of the corpuscularian tax
onom y of atom ic chem istry. A nd a reduction of the old taxon om y to
the new will appear impossible, given the extent to w hich the com
paratively toothless old theory cross-classifies things relative to the
new . Elim ination w ould thus appear the only alternative— unless
som e cunning and determ ined defender of the alchem ical vision had
the w it to suggest the following defense.
Being "en sou led by m ercu ry ," or "s u lfu r," or either of the other
two so-called spirits, is actually afu n ctio n a l state. The first, for exam
ple, is defined by the disposition to reflect light, to liquefy u n d er heat,
to unite w ith oth er m atter in the sam e state, and so forth. A n d each of
these four states is related to the others, in that the syndrom e for each
varies as a function of w hich of the other three states is also instanti
ated by the sam e substrate. Thus the level of description com
p reh ended by the alchemical vocabulary is abstract: various m aterial
substances, suitably "e n s o u le d ," can display the features of a m etal,
for exam ple, or even of gold specifically. For it is the total syndrom e
of occu rren t and causal properties that m atters, not the corpuscula
rian details of the substrate. Alchemy, it is concluded, com prehends
a level of organization in reality that is distinct from , and irreducible
to, the organization found at the level of corpuscularian chem istry.
This view m ight have had considerable appeal. After all, it spares
alchem ists the burden of defending im m aterial souls th at com e and
go; it frees them from having to m eet the v ery strong d em an d s of a
naturalistic reduction; and it spares them the shock and confusion of
outright elim ination. A lchem ical theory em erges as basically all right!
N or need the alchem ists appear too obviously stubborn or dogm atic
in this. A lchem y as it stands, they con cede, m ay need substantial
tidying up , and experience m ust be our guide. But w e n eed not fear
its naturalistic displacem ent, they rem ind us, since it is the peculiar
orchestration of the syndrom es of occu rrent and causal properties
that m akes a piece of m atter gold, not the idiosyncratic details of its
corpuscularian substrate. A further circum stance w ould have m ade
this claim even m ore plausible. For the fact is, the alchem ists did
know h ow to m ake gold, in this relevantly w eakened sense of 'gold',
and th ey could do so in a variety of w ays. Their "g o ld " w as never as
perfect, alas, as the "g o ld " n u rtured in n atu re's w om b , but w hat
m ortal can exp ect to m atch the skills of nature herself?
W hat this story show s is that it is at least possible for the constella
tion of m oves, claims, and defenses characteristic of functionalism to
constitute an outrage against reason and tru th , and to do so w ith a
plausibility that is frightening. A lchem y is a terrible theory, well de
14 The Nature of Mind
serving of its com plete elimination, and the defense of it just explored
is reactionary, obfuscatory, retrograd e, and w rong. But in historical
con text, that defense m ight have seem ed wholly sensible, even to
reasonable people.
The alchem ic exam ple is a deliberately transparent case of w hat
m ight be called "th e functionalist stra ta g e m ," and other cases are
easy to im agine. A cracking good defense of the phlogiston theory of
com bustion can also be constructed along these lines. C on stru e being
highly phlogisticated and being dephlogisticated as functional states
defined by certain syndrom es of causal dispositions; point to the
great variety of natural substrates capable of com bustion and calx-
ification; claim an irreducible functional integrity for w hat has proved
to lack any natural integrity; and bury the rem aining defects u n d er a
pledge to contrive im provem ents. A similar recipe will provide new
life for the four hum ors of m edieval m edicine, for the archeus or vital
essen ce of prem od ern biology, and so forth.
If its application in these other cases is any guide, the functionalist
stratagem is a sm oke screen for the preservation of error and confu
sion. W h en ce derives our assurance that in con tem porary journals
the sam e charade is n ot being played out on behalf of FP? The parallel
w ith the case of alchem y is in all other respects distressingly com
plete, right dow n to the parallel betw een the search for artificial gold
and the search for artificial intelligence!
Let m e not be m isunderstood on this last point. Both aim s are
w orth y aim s: thanks to nuclear physics, artificial (but real) gold is
finally w ithin our m ean s, if only in subm icroscopic quantities, and
artificial (but real) intelligence eventually will be. But just as the care
ful orchestration of superficial syndrom es w as the w ron g w ay to p ro
duce genuine gold, so m ay the careful orchestration of superficial
syndrom es be the w ron g w ay to produ ce genuine intelligence. Ju st as
w ith gold, w hat m ay be required is that our science p enetrate to the
underlying natural kind that gives rise to the total syndrom e directly.
(See chapters 5 and 9 to 11.)
In sum m ary, w hen confronted w ith the explanatory im potence,
stagnant history, and system atic isolation of the intentional idiom s of
FP, it is not an adequate or responsive defense to insist that those
idioms are abstract, functional, and irreducible in character. For one
thing, this sam e defense could have b een m ou nted with com parable
plausibility no m atter w hat hayw ire netw ork of internal states our
folklore had ascribed to us. A nd for an oth er, the defense assum es
essentially w hat is at issue: it assum es that it is the intentional idioms
of FP, plus or m inus a bit, that express the im portant features shared
by all cognitive system s. But they m ay not. Certainly it is w ron g to
Eliminative Materialism and the Propositional Attitudes 15
assum e that they do, and then argue against the possibility of a
m aterialistic displacem ent on grounds that it m u st describe m atters at
a level that is distinct from the im portant level. This just begs the
question in favor of the older fram ew ork.
Finally, it is very im p ortant to point out that eliminative m aterial
ism is strictly consistent with the claim that the essence of a cognitive
system resides in the abstract functional organization of its internal
states. The eliminative m aterialist is not com m itted to the idea that
the correct accou n t of cognition m ust be a naturalistic accou n t, though
he m ay be forgiven for exploring the possibility. W hat h e does hold
is that the correct accoun t of cognition, w heth er functionalistic or
naturalistic, will bear about as m uch resem blance to FP as m odern
chem istry bears to four-spirit alchem y.
Let us now try to deal with the argu m en t, against eliminative
m aterialism , from the norm ative dim ension of FP. This can be dealt
w ith rath er swiftly, I believe.
First, the fact that the regularities ascribed by the intentional core of
FP are predicated on certain logical regularities am ong propositions is
not by itself grounds for claiming anything essentially norm ative
about F P. To draw a relevant parallel, the fact that the regularities
ascribed by the classical gas law are predicated on arithm etical rela
tions b etw een num bers does not imply anything essentially n or
m ative about the classical gas law. A nd logical relations betw een
propositions are as m uch an objective m atter of abstract fact as are
arithm etical relations betw een num bers. In this respect, the law
(4) {x){p)(q)[((x believes that p) & (x believes that (if p then q)))
=3 (barring confusion, distraction, e tc., x believes that q)]
is entirely on a p ar with the classical gas law
A norm ative dim ension enters only because w e happen to value m ost
of the p attern s ascribed by FP. But w e do not value all of them .
Consider
our norm ative concerns. It implies only that they will h ave to be re
constituted at a m ore revealing level of u n d erstanding, the level that a
m atured neuroscience will provide. [A dded in 1989: The beginnings
of such a recon struction can be found in chapter 10, pp. 2 2 0 -2 2 3 .]
W h at a theoretically inform ed future m ight hold in store for us, w e
shall n ow turn to explore. N ot because w e can foresee m atters with
any special clarity, but b ecause it is im portant to try to break the grip
on our im agination held by the propositional kinem atics of FP. A s far
as the p resen t section is co n cern ed , w e m ay sum m arize our conclu
sion as follows. FP is nothing m ore and nothing less than a culturally
entrenched theory of h ow w e and the higher anim als w ork. It has no
special features that m ake it empirically invulnerable, no unique func
tions th at m ake it irreplaceable, no special status of any kind w hat
soever. W e shall turn a skeptical ear then, to any special pleading on
its behalf.
W hat m igh t the elim ination of FP actually involve: not just the com
paratively straightforw ard idiom s for sensation, but the entire ap p ar
atus of propositional attitudes? That d ep ends heavily on w hat
neuroscience m ight d iscover and on our determ ination to capitalize
on it. H ere follow three scenarios in w hich the operative conception
of cognitive activity is progressively divorced from the form s and
categories that characterize natural language. If the read er will in
dulge the lack of actual substance, I shall try to sketch som e plausible
form.
First, suppose that research into the structure and activity of the
brain, b oth fine-grained and global, finally does yield a new kinem a
tics and correlative dynam ics for w hat is now thought of as cognitive
activity. The theory is uniform for all terrestrial brains, n ot just hu
m an brains, and it m akes suitable conceptual contact w ith both evolu
tionary biology and nonequilibrium therm odynam ics. It ascribes to us
at any given time a set or configuration of com plex states that are
specified w ithin the theory as figurative "so lid s" w ithin a four- or
five-dim ensional p hase space. [Added in 1989: This guess has proved
to be v ery timid. The relevant cognitive statespaces typically have
hundreds, thou san ds, or even millions of distinct dim ensions, and
their partitioning into hypersolids is correspondingly com plex. See
chapter 9.] The law s of the theory govern the interaction, m otion, and
transform ation of these "so lid " states w ithin that space, and also
their relations to w h atever sensory and m otor transducers the system
possesses. A s w ith celestial m echanics, the exact specification of the
18 The Nature of Mind
O nce co n stru cted , this "lan g u ag e" proves to be learnable, it has the
p ow er projected, and in tw o generations it has sw ept the planet.
E veryone u ses the new system . The syntactic form s and sem antic
categories of so-called "n a tu ra l" language disappear entirely. A nd
w ith them disappear the propositional attitudes of FP, displaced by a
m ore revealing schem e in w hich (of course) " U bersatzenal attitu d es"
play the leading role. FP again suffers elim ination.
This second story, note, illustrates a them e w ith endless variations.
There are possible as m any different "folk psychologies" as there are
possible differently structured com m unication system s to serve as
m odels for them .
A third and even stranger possibility can be outlined as follows. W e
know that there is considerable lateralization of function betw een the
tw o cerebral hem ispheres, and that the tw o hem ispheres m ake use of
the inform ation they get from each other by w ay of the great cerebral
com m issure, the corpus callosum , a giant cable of neurons co n n ect
ing them . Patients w hose com m issure has been surgically severed
display a variety of behavioral deficits that indicate a loss of access by
one hem isphere to inform ation it used to get from the other. H ow
ever, in people with callosal agenesis (a congenital defect in which the
corresponding cable is simply absent), there is little or no behavioral
deficit, w hich suggests that the tw o hem ispheres have learned to ex
ploit the inform ation carried in other, less direct pathw ays connecting
them through the subcortical regions. This suggests that even in the
norm al case a developing hem isphere learns to m ake use of the in
form ation the cerebral com m issure deposits at its doorstep. W h at w e
have, then , in the case of a norm al hu m an , is tw o physically distinct
cognitive system s (both capable of ind ep en d en t function) responding
in a system atic and learned fashion to exchanged inform ation. A nd
w hat is especially interesting about this case is the sheer am ou nt of
inform ation exch an ged . The cable of the com m issure consists of
roughly 200 million neurons (G azzaniga and LeD oux 1975), and even
if w e assum e that each of these fibers is capable of one of only tw o
possible states each second (a m ost conservative estim ate), w e are
looking at a channel w hose inform ation capacity is g reater than
2 x 108 binary bits per second. C om p are this to the less than 500
bits/second capacity of spoken English.
N ow , if tw o distinct hem ispheres can learn to com m unicate on so
im pressive a scale, w hy shouldn't tw o distinct brains learn to do it
also? This w ould require an artificial "co m m issu re" of som e kind, but
let us suppose that w e can fashion a workable transducer for im
plantation at som e site in the brain that research reveals to be suit
able, a tran sd u cer to convert a sym phony of neural activity into (say)
Eliminative Materialism and the Propositional Attitudes 21
that p resu p p oses the integrity of FP. But formally speaking, one can
as well infer, from the incoherent result, that this (Gricean) theory of
m eaning is w hat m ust be rejected. Given the independent critique of
FP leveled earlier, this w ould even seem the preferred option. But in
any case, one cannot simply assum e this particular theory of m eaning
w ithout begging the question at issue, nam ely, the integrity of FP.
The question-begging nature of this m ove is m ost graphically illus
trated by the following analog, w hich I ow e to Patricia S. C hurchland
(1981). The issue here, placed in the seven teen th century, is w heth er
there exists such a substance as vital spirit. A t the tim e, this substance
w as held, w ithout significant aw aren ess of real alternatives, to be
w hat distinguished the anim ate from the inanim ate. Given the
m onopoly enjoyed by this conception, given the degree to w hich it
w as integrated w ith m any of our other conceptions, and given the
m agnitude of the revisions any serious alternative conception w ould
require, the following refutation of any antivitalist claim w ould be
found instantly plausible.
The antivitalist says that there is no such thing as vital spirit. But
this claim is self-refuting. The speaker can expect to be taken
seriously only if his claim can n ot. For if the claim is true, then the
speaker does not have vital spirit and m u st be dead. But if he is
dead, then his statem ent is a m eaningless string of noises, de
void of reason and truth.
agree that those properties are the principal m eans of our introsp ec
tive discrim ination of one kind of sensation from another. W h at he is
com m itted to denying is that any particular quale is essential to the
identity of any particular type of m ental state. Initially they m ay seem
to be essential, but reflection will reveal that they do not h ave, and
should not be conceded, that status. In w hat follows w e address four
distinct but not unrelated problem s. Each problem is m anageable on
its ow n, but if they are perm itted to band togeth er for collective
assault, the result is rather confusing and form idable, in the fashion
of the fabled M usicians of Brem en. W ith the problem s sep arated , our
strategy will be to explain and exploit the insight that intrinsic p ro p
erties per se are no anathem a to a functionalist theory of m ental
states.
ternal w orld, and it is evident that the introspective taxon om ies into
w hich w e eventually settle are no less shaped by considerations of
explanatory and causal coherence than are the taxon om ies of external
observation.
It is therefore a great irony, it seem s to us, that an yon e should
subsequently point to w h atever qualia our introspective m echanism s
have m an aged tenuously to fix upon as m ore or less usable indicators
of nom ologically interesting states, and claim them as constituting the
essence of such states. It is, of course, distantly possible that our
m echanism s of introspective discrim ination have lucked onto the
constituting essences of our psychological states (assum ing, contrary
to our earlier discussion, that each type has a uniform natural es
sence), b ut a priori that seem s about as likely as that the visual sys
tem lucked onto the constituting essence of tigerhood w h en it m ade
black-on-yellow stripes salient for distinguishing tigers.
It seem s very doubtful, therefore, that the type identity of any
psychological state derives from its sharing in any uniform natural
essen ce. M oreover, even if it does so share, it seem s entirely unlikely
that introspection provides any special access to that essen ce. C on se
quently, this beggars the intuition that sustains the inverted-qualia
objections.
The p receding investigation into the w eight and significance of fac
tors determ ining type identity of psychological states d oes m ore than
that, how ev er. It also enriches the com peting intuition, nam ely, that
the typ e identity of psychological states is determ ined by functional
characteristics. To rep eat the point m ade earlier, since the taxon om y
of observational qualia constructed by the questing child fo llo w s the
discovered taxon om y of states as determ ined by interesting causal
roles, it is evident that sam eness of functional role dom inates over
differences in qualitative character, so far as the type identity of
psychological states is concerned. That a single category, united by
functional considerations, can em brace diverse and disparate qualita
tive ch aracters has a ready illustration, ironically en ou gh , in the case
of pain.
C onsider the wide variety of qualia willfully lum ped togeth er in
com m on practice un d er the heading of pain. C om p are the qualitative
character of a severe electric shock w ith that of a sh arp blow to the
kneecap; com pare the character of hands dully aching from m aking
too m an y snowballs w ith the piercing sensation of a jet engine heard
at very close range; com pare the ch aracter of a frontal h ead ach e w ith
the sensation of a scalding pot grasped firmly. It is evident that w hat
unites sensations of such diverse characters is the similarity in their
functional roles. The sudden onset of any of them p rom p ts an in
28 The Nature of Mind
dom ain in w hich the discrim inations are m ade. Unlike the sensation
case, no narrow range of stim ulus-response connections will begin to
characterize the m echanism s at w ork here.
Sensations and beliefs, accordingly, m u st be introspectively dis
crim inated by entirely distinct cognitive m echanism s, m echanism s
facing quite different problem s and using quite different strategies
for their solutions. Sensations are identified by w ay of their intrin
sic properties; beliefs are identified by w ay of their highly abstract
structural features. It should not be w ond ered at, then, th at there is a
subjective contrast in the nature of our aw areness of each.
[A dded in 1989: I m ust now express a loss of confidence in this
argu m en t. The problem is that sensations now ap p ear to be decidedly
m ore various than I had originally estim ated and to h ave a m uch
m ore intricate com binatorial structure than I had earlier supposed
(see ch ap ter 5, sec. 7). A ccordingly, the con trasts on w hich the p re
ceding argu m en t places so m u ch w eight now ap p ear spurious: w hat
seem ed a large difference in kind now seem s a m ere difference in
d egree.]
they will; in other cases they will not. W h at decides the m atter is
w heth er those qualitative differences m ark any causal or functional
differences relevant to the explanation of psychological activity and
overt behavior.
So long as introspectible qualia w ere thought to be ineffable, or
epiphenom enal, or dualistic, or essential for type identity, one can
un derstan d the functionalist's reluctance to have anything to do with
them . But once w e have seen how the functionalist can acknow ledge
them and their epistem ic role, w ithin a naturalistic fram ew ork, the
reluctance should disappear. For the taxon om y of states appropriate
for psychological theory rem ains dictated entirely by causal and ex
p lanatory factors. Qualia are just accidental hooks of opportunity for
the introspective discrim ination of dynam ically significant states.
Som e of the issues arising here have already been broached in the
subsection on absent qualia. H ow ever, novel problem s arise as well,
and organization is best served by a sep arate section. All of the prob
lem s here begin with the functionalist's central contention that the
functional organization n ecessary and sufficient for person h ood is an
abstract one, an organization realizable in principle in an indefinite
variety of physical system s. Such liberalism seem s innocent enough
w h en w e contem plate the p rosp ect of hum anoid aliens, biom echani
cal androids, and electrom echanical robots w hose physical con stitu
tions are at least rough parallels of our ow n. W ho could deny that
C 3PO and R2D 2— of Star W ars fam e— are p ersons? But our liberal
intuitions are quickly flum m oxed w hen w e consider bizarre physical
system s th at m ight nevertheless realize the abstract causal organiza
tion at issue, and such cases m ove one to reconsider o n e's generosity
in the m ore familiar cases as well.
The following discussion will explore but tw o of these nonstan d ard
"p e rso n s” : N ed Block's "C h in ese n atio n " (1978) and John Searle's
"C hinese-speak in g ro o m " w ith the m onolingual anglophone locked
inside it (Searle 1980). Block is concerned w ith the absence of qualia
from states posing as sensations, and Searle is concerned w ith the
absence of intentionality from states posing as propositional attitudes.
b ody of the sim ulated p erson , and it interacts w ith its collective brain
as follows. It sends a sensory input m essage L to every single citizen
and subsequently receives a m otor output m essage O, from exactly
one citizen. W hich citizen sends w hat outp u t is determ ined as
follows.
O verh ead from a satellite som e state letter Sk is displayed in lights
for all to see. For each possible state letter Sk there is p erm anently
assigned a distinct subset of the population. In the rare ev en t w hen Sk
is displayed and input J; is received, one p erson in the Sk group, a
p erson to w hom Ij has been assigned, perform s the following p re
assigned task. She sends to the robot the unique output m essage O,
antecedently assigned to her for just this occasion, radios the satellite
to display the state letter Sf, antecedently assigned to h er for just this
occasion, and then subsides, w aiting for the next opportunity to do
exactly the sam e thing in exactly the sam e circum stances.
A s organized above, each citizen realizes exactly one square of the
machine table that specifies T m. (A machine table is a matrix or checker
board w ith state letters heading the colum ns and input letters h ead
ing the row s. A n y square is the intersection of som e Sk and Ip and it
specifies an output O, and a shift to som e state S„, w here possibly
p = k. See figure 2. 1. ) Block asks us to assum e that Tm adequately
sim ulates you r ow n functional organization. O ne is likely to grant
him this, since any inp u t-outpu t function can in principle be realized
in a suitable Turing m achine. In pondering an apparently fussy
detail, Block w on d ers, "H o w m any hom unculi are req u ired?" and
an sw ers, "P erh ap s a billion are enough; after all, there are only about
a billion n euron s in the brain" (p. 278). H ence his choice of China as
the potential artificial brain. (Block underestim ates here. The num ber
of n eu ron s in the brain exceeds 100 billion.)
Finally, Block finds it starkly implausible to sup p ose that this
nationw ide realization of Tm has states w ith a qualitative character
like pains, tastes, and so on. It is difficult n ot to agree w ith him . His
hom unculi do not even interact w ith one an other, save indirectly
through the satellite state letter and even less directly throu gh the
ad ventures of the robot body itself. The shim m ering intricacies of
on e's inner life are not to be found here.
The w ay to avoid this criticism of functionalism is just to insist that
any subject of beliefs and sensations m u st not only be Turing equiva
lent to us (that is, produ ce identical outpu ts given identical inputs); it
m ust be com putationally equivalent to us as well. That is, it m u st have
a system of inner states w hose causal interconnections m irror those in
our ow n case. This is not an arbitrary restriction. Folk psychology is,
and scientific psychology should be, realistic about our m ental states,
36 The Nature of Mind
Figure 2.1
A Turing machine table. The machine is always in some one or other internal state S„
represented by one of the vertical columns. When it receives an input l m, it executes the
instructions in the specific square where the S„ column and the I„, row intersect, and
then it waits for the next input.
and m ere parity of gross behavior does not guarantee parity of causal
organization am ong the states that produ ce it. The com putational
organization displayed in the C hinese-nation Turing m achine is not
even distantly analogous to our ow n. If it w ere analogous to our ow n,
w orries about absent qualia could be handled as outlined in the sub
section about absent qualia above. That is, w e could insist that the
qualia are there all right, but are unrecognized by us un d er their
physical descriptions.
There is a further reason w hy it is not arbitrary to insist on a com
putational organization m ore along the lines of our ow n, and w e m ay
illustrate it by exam ining a further defect in Block's exam ple. It is
dem onstrable that no T,„ realized as described in the population of
China could possibly simulate you r input-output relations. There are
not nearly enough C hinese for this job, not even rem otely enou gh . In
fact, a spherical volum e of space centered on the Sun and ending at
Functionalism, Qualia, and Intentionality 37
distinct cells in its ow n right, each with about 103 connections with
oth er cells. O ur estim ate of the num ber of distinct states of the brain
m u st be substantially in excess of ю 30-000-000, our num ber for the
retina.
O ur conclusion is that no brute-force, one-device/one-square
realization of a Turing m achine constructible in this universe could
even begin to sim ulate you r inp u t-ou tp u t organization. E ven the
humblest of creatures are beyond such simulation. A n unprepossess
ing gastropod like the sea slug A plysia californica has well in excess of
332 distinct sensory cells, and thus is clearly beyond the reach of the
crude m eth ods at issue. This does not m ean that the h um an input-
o u tp u t relations cannot be rep resented by an abstract Turing m achine
Tm. W h at it does m ean is that any physical m achine adequate to such
sim ulation m ust have its com putational architecture and executive
h ard w are organized along lines vastly different from , and m uch m ore
unified and efficient than, those displayed in Block's exam ple. That
exam ple, therefore, is not even rem otely close to being a fair test of
our intuitions. Quite aside from the question of qualia, the Chinese
Turing m achine couldn't sim ulate an earth w orm .
This w eakness in the exam ple is not adequately m ade up by allow
ing, as Block does at one point (p. 284), that each hom unculus m ight
be responsible for a wide range of inputs, each w ith corresponding
o utp uts. O n this m odification, each hom unculus w ould thus realize,
dispositionally, m an y m achine-table squares sim ultaneously. Sup
pose, then, that w e m ake each C hinese citizen responsible for one
billion squares peculiar to him (this w as the size of the original Tm).
This raises the num ber of distinct inputs processible by the system
to 109 citizens X 109 squares = 1018 possible inputs, still well short
of the lO30-000'000 w e are striving for.
W ell, how m any squares m uch each citizen realize if the nation as a
w hole is to instantiate som e Turing m achine adequate to handle the
required input? The answ er is, of course, 10<30'000'000 —9) squares each.
But how will each hom unculus-citizen handle this aw esom e load? N ot
by being a simple one-device/one-square Turing m achine in turn,
as w e have already seen. N o physical sim ulation adequate to you r
input-output relations, therefore, can avoid having the m ore unified
and efficient m odes of com putational organization alluded to in the
p receding p aragrap hs, even if they show up only as m odes of orga
nization of its various subunits. That is, any successful sim ulation of
you m u st som ew here display a com putational-executive organization
that is a m u ch m ore plausible hom e for qualitative states than Block's
exam ple w ould suggest. [Added in 1989: A good exam ple of a
Functionalism, Qualia, and Intentionality 39
is that the entire system of the room plus its contents does n ot u n d er
stand C hinese either. N othing here u n derstan ds C hinese, save those
sending and receiving the m essages, and those w ho w rote the p ro
gram . N o com putational state or output of that system h as any m ean
ing or intentionality save as it is interpretively im posed from
w ithout by those w ho interact with it.
H ow ever, concludes Searle, this system already contains every
thing relevan t to be found in the physical realization of any purely
formal p rogram . If m eaning and intentionality are m issing here, they
will be m issing in any such attem pt to sim ulate h u m an m ental
activity. Instantiating a p rogram cannot be a sufficient condition of
u n d erstanding,
The set of com m entaries published in the sam e issue provid es m any
useful and interesting criticisms of Searle's argum ent and of his con
clusions as well. The critical consensus is roughly as follows. If the
system of the room plus contents w ere u p graded so that its con versa
tional skills extend beyond a handful of topics to include the entire
range of topics a norm al hum an could be expected to know ; and if the
system w ere supplied w ith the sam e inductive capacities w e enjoy;
and if the "belief sto re" w ere integrated in the norm al fashion with
som e appropriately com plex goal stru cture; and if the room w ere
causally con nected to a body in such a fashion that its inputs reflected
appropriate sensory discrim inations and its outputs p rodu ced ap p ro
priate b ehavior; then the system of the room plus con ten ts jolly well
w ould u n derstand C hinese, and its various com putational states—
beliefs that p, desires that q— w ould indeed have m eaning and in
tentionality in the sam e w ay as with a norm al C hinese speaker.
Searle is quite willing to consider upgradings of the kind
described— he attem pts to anticipate them in his p ap er— but he is
convinced they change nothing relevant to his case. A s it em erges
clearly in his A u th or's R esponse (pp. 4 5 0 -4 5 6 ), of central im portance
to his argu m en t is the distinction betw een
Exam ples of the latter w ould be the w ord s and sentences of on e's
native tongue. These have m eaning and intentionality, allows Searle,
but only insofar as they bear certain relations to our beliefs, thoughts,
and intentions— states with intrinsic intentionality. A sim ulation of
h u m an m entality grounded in a formal p rogram m ay yield states h av
ing this derivative observer-relative brand of intentionality, concedes
Searle, but they cannot have intrinsic intentionality. A nd since they
lack a feature essential to genuine m ental states, they cannot be
genuine m ental states, and to that extent the simulation m ust be a
failure.
A s w e see it, this criticism of functionalism is profoundly in error. It
is a m istake to try to m eet it, how ever, by continuing w ith the
strategy of trying to upgrade the im agined simulation in h op es of
finally w inning Searle's concession that at last its states have achieved
intrinsic intentionality. The correct strategy is to argue that our ow n
m ental states are just as innocent of "intrinsic intentionality" as are
the states of any m achine sim ulation. O n our view , all ascriptions
of m eaning or propositional content are relative (in senses to be
explained). The notion of "intrinsic intentionality" m akes no m ore
em pirical sense than does the notion of position in absolute space.
W e shall try to explain these claims.
There are basically just tw o w ays in w hich one can assign p roposi
tional content to the representational states of an oth er organism . An
exam ple of the first is the translation of a foreign language. A n exam
ple of the second is the calibration of an instrum ent of m easurem en t
or detection.
In the case of translation, one assigns specific propositional co n
tents to the alien representations because one has found a general
m apping betw een the alien rep resentation s and our ow n such that
the netw ork of formal and m aterial inferences holding am ong the
alien representations closely m irrors the sam e netw ork holding
am ong our ow n representations. Briefly, their collected rep resen ta
tions display an intensional structure that m atch es the intensional
stru cture displayed by our ow n.
The story is essentially the sam e w hen w e are assigning proposi
tional con tent to an alien's thou gh ts, beliefs, etc. It m atters n aught
w h eth er the alien's representation is overt, as with a sentence, or
covert, as w ith a belief. W e assign a specific content, p, to one of the
Functionalism, Qualia, and Intentionality 43
A . Conceptual conservatism
N o functionalist will suppose that the functional organization reco g
nized in the collected lore of folk psychology exh au sts the functional
intricacies that m ake up our internal econ om y. All will agree that folk
psychology rep resents only a partial, and in som e respects even a
superficial, grasp of the m ore com plex organization that em pirical
psychology will eventually unravel. E ven so, there is a decided
tend en cy on all sides to suppose that, so far as it goes, folk psych ol
ogy is essentially correct in the picture that it paints, at least in basic
outlines. Em pirical psychology will add to it, and explain its princi
ples, m any exp ect, but alm ost no one exp ects it to be o verth row n or
transm ogrified by such research.
Functionalism, Qualia, and Intentionality 45
C. Reductionism
Thanks to the argum ent from abstraction, functionalists tend to be
strongly antireductionist. They deny that there can be any general
characterization of w hat m akes som ething a thinker that is expressible
in the language of any of the physical sciences. Given the variety of
possible substrates (biological, chem ical, electrom echanical) that
could realize a thinking system , it is difficult n ot to agree w ith them .
But it does not follow, from multiple instantiability per se, that no
such general characterization is possible. It follows only that the re
quired characterization cannot be exp ressed in the theoretical v oca
bulary peculiar to any one of the available substrates. It rem ains en
tirely possible that there is a level of physical description sufficiently
abstract to en com pass all of them , and y et sufficiently pow erful to
fund the characterization required.
A s it h ap p ens, there is indeed a physical theory of sufficient g en er
ality to encom p ass the activity of all of these substrates, and any
o th ers one m ight think of. The theory is therm od yn am ics— the
general theory of energy and entropy. It has already supplied us with
a profoundly illuminating characterization of w h at the nineteenth
cen tu ry called "vital activity ," that is, of the ph en om en on of life. A nd
it is far from unthinkable that it m ight do the sam e for w hat this cen
tury calls "m en tal activity" (for a brief exploration of these ideas, see
P. M . C hurchland 1982). The theoretical articulation of such a ch arac
terization w ould be a very great achievem ent. It w ould be un fortu n
ate if the search for it w ere im peded by the general conviction that it is
im possible, a conviction born of the antireductionist urgings of a
false o rth od oxy am ong functionalists.
Chapter 3
Reduction, Qualia, and the Direct Introspection
of Brain States
1 Intertheoretic Reduction
logically entails
T o-
Difficulties w ith this view begin w ith the observation that m ost re
duced theories turn out to be, strictly speaking and in a variety of
resp ects, fa lse. (Real gases d on 't really obey P V = jxRT, as in classical
therm od ynam ics; the planets d on 't really m ove in ellipses, as in
Keplerian astron om y; the acceleration of falling bodies isn't really
uniform , as in Galilean dynam ics; etc.) If reduction is deduction, m od
us tollens w ould thus require that the prem ises of the new reducing
theories (statistical therm odynam ics in the first case, N ew tonian
dynam ics in the second and third) be som eh ow false as well, in co n
tradiction to their assum ed truth.
This com plaint can be tem porarily deflected by pointing out that
the prem ises of a reduction m ust often include, not just the new re
ducing theory, but also som e limiting assum ptions or counterfactual
b ou n d ary conditions (such as that the m olecules of a gas have only
m echanical energy, or that the m ass of the planets is negligible co m
p ared to the su n 's, or that the distance any body falls is negligibly
different from zero). Falsity in the reducing prem ises can thus be co n
ced ed , since it is safely confined to those limiting or counterfactual
assum ptions.
This defense will not deal w ith all cases of falsity, h ow ever, since in
som e cases the red u ced theory is so radically false that som e or all of
its ontology m u st be rejected entirely, and the "corresp on d en ce
ru les" connecting that ontology to the n ew er ontology therefore dis
play a problem atic status. N ew ly conceived features cannot be iden
tical w ith, nor even nom ically connected w ith, old features, if the old
features are illusory and uninstantiated. For exam ple, relativistic
m ass is not identical w ith N ew tonian m ass, n or even coextensive
w ith it, even at low velocities. N evertheless, the reduction of N ew to
nian by Einsteinian m echanics is a paradigm of a successful red u c
tion. For a second exam ple, neither is caloric-fluid-pressure identical
w ith, nor even coextensive w ith, m ean m olecular kinetic energy. But
an overtly flu id therm odynam ics (i.e., one com m itted to the existence
of "calo ric") still finds a m oderately im pressive reduction w ithin
statistical therm odyn am ics. In sum , even theories w ith a nonexistent
ontology can enjoy reduction, and this fact is problem atic on the
traditional account at issue.
C ases like these invite us to give up the idea that w hat gets d e
d uced in a reduction is the theory to be red u ced . A m ore accu rate,
The Direct Introspection of Brain States 49
logically entails
(1) TN reduces T0 ,
(2) 'F' and 'G' are corresp on d en ce-ru le paired in the red u c
tion, and
(3) the reduction is sufficiently sm ooth to sustain the
ontology of T0 and thus to sustain the identity claim
'F-n ess = G -ness'.
Intuitively, and in the m aterial m ode, this m ean s that F -n ess reduces
to G-ness just in case the causal pow ers of F-n ess (as outlined in the
law s of T0 ) are a subset of the causal p ow ers of G-ness (as outlined in
the law s of TN).
Finally, a p roperty F will be said to be an em ergent p roperty (relative
to TN) just in case
Intuitively, this will happen w hen TN does n ot have the resou rces
adequate to define a property w ith all of the causal pow ers possessed
by F-n ess. Claims about the em ergen ce of certain properties are th ere
fore claims about the relative p overty in the resou rces of certain aspir
ant theories. H aving outlined these notions, w e shall turn to address
substantive questions of em ergence and irreducibility in a few m o
m ents.
The Direct Introspection of Brain States 51
red u ced property, w ould be rend ered irreducible, since the term
'tem p eratu re' does not app ear in the lexicon of statistical m echanics.
There is a further reason w hy the dem an d for direct deducibility is
too strong. The fact is, it is an historical accident that w e hum ans
currently use precisely the conceptual fram ew ork w e do use. W e
m ight have used any one of an infinite num ber of oth er conceptual
fram ew orks to describe the observable w orld, each one of w hich
could have been roughly adequate to com m on experience, and m any
of w hich w ould be roughly isom orphic (each in its different w ay) with
som e p art of the correct accou n t that a utopian theory will eventually
provide. A ccordingly, w e can legitim ately ask of a putatively correct
theory of a given objective dom ain that it accou n t for the p h enom ena
in (that is, function successfully in) that dom ain. But w e cannot insist
that it also be able to predict how this, th at, or the other conceptually
idiosyncratic h um an culture is going to conceive of that dom ain. That
w ould be to insist that the new theory do predictive cultural an thropolo
gy for u s, as well as m echanics, or electrom agnetic theory, or w hat
have you . The dem and that m olecular theory directly entail our ther
m al or color con cep ts is evidently this sam e unreasonable dem and.
All w e can properly ask of a reducing theory is that it have the
resou rces to conjure up a set of properties w hose nom ological
pow ers/roles/features are system atic analogues of the pow ers/roles/
features of the set of properties postulated by the old theory. Since
both theories presu m e to describe the sam e em pirical dom ain, these
system atic nom ological parallels constitute the best grounds there can
be for concluding that both theories have m anaged to latch onto the
sam e set of objective properties. The h yp othesized identity of the
properties at issue explains w hy and T0 are taxonom ically and
nom ically parallel: they are both at least partially correct accou n ts of
the very sam e objective properties. ZN m erely fram es that account
w ithin a m u ch m ore penetrating conceptual system — that of TN.
M oreover, it is to be expected that existing conceptual fram ew orks
will eventually be reduced or displaced by new and better ones, and
those in turn by fram ew orks better still, for w ho will be so brash as to
assert that the feeble conceptual achievem ents of our adolescent sp e
cies com prise an exhaustive accou n t of anything at all? If w e p ut aside
this conceit, then the only alternatives to intertheoretic reduction are
epistem ic stagnation or the outright elim ination of old fram ew orks as
w holly false and illusory.
Esoteric properties and arcane theoretical fram ew orks are not the
only things that occasionally enjoy intertheoretic reduction. O bserv
The Direct Introspection of Brain States 53
able properties and com m onsense conceptual fram ew orks can also
enjoy sm ooth reduction. Thus, being a m iddle-A sound is identical
w ith being an oscillation in air pressu re at 440 hertz; being red is
identical w ith having a certain triplet of electrom agnetic reflectance
efficiencies; being w arm is identical with having a certain m ean level
of m icroscopically em bodied energies, and so forth.
M oreover, the relevant reducing theory is capable of replacing the
old fram ew ork n ot just in contexts of calculation and inference. It
should be appreciated that the reducing theory can displace the old fram ew ork
in all o f its observational contexts as w ell. Given the reality of the p rop
erty identities just listed, it is quite open to us to begin fram ing our
sp on tan eou s perceptual reports in the language of the m ore sophisti
cated redu cin g theory. It is even desirable that w e begin doing this,
since th e new vocabulary observes distinctions w hich are in fact w ith
in the discrim inatory reach of our native perceptual system s, though
those objective distinctions go unm arked and unnoticed from w ithin
the old fram ew ork. W e can thus m ake m ore penetrating use of our
native perceptual equipm ent. Such displacem ent is also desirable for
a secon d reason : the greater inferential or com putational p ow er of the
n ew conceptual fram ew ork. W e can thus m ake better inferential use
of our new perceptual judgm ents than w e m ade of our old ones.
It is difficult to convey in w ords the enorm ity of such perceptual
transform ations and the naturalness of the new con cep tu al regim e
once established. A nonscientific exam ple m ay help to get the initial
point across.
C on sid er the en orm ou s increase in discrim inatory skill that spans
the gap betw een an untrained child's auditory ap prehension of a
sym p hon y, and the sam e p erson's apprehension of the sam e sym
p h on y forty years later, w hen hearing it in his capacity as con d u ctor
of the o rch estra perform ing it. W hat w as before a seam less voice is
now a m osaic of distinguishable elem ents. W hat w as before a dimly
ap p reh en d ed tune is now a rationally stru ctu red sequence of distin
guishable and identifiable chords supporting an appropriately related
m elody line. The m atured m usician h ears an entire w orld of stru c
tured detail, concerning w hich the child is both dum b and deaf.
O th er m odalities provide com parable exam ples. C on sid er the p rac
ticed and chem ically sophisticated wine taster, for w hom the category
"re d w in e " u sed by m ost of us divides into a netw ork of fifteen or
tw enty distinguishable elem ents: ethanol, glycol, fru ctose, sucrose,
tannin, acid, carbon dioxide, and so forth, w hose relative con cen tra
tions h e can estim ate w ith accuracy. O r consider the astron om er, for
w h om the speckled black dom e of h er you th has b ecom e a visible
abyss, scattering nearby planets, yellow dw arf stars, blue and red
54 The Nature of Mind
giants, distant globular clusters, and even a rem ote galaxy or tw o, all
discriminable as such and locatable in three-dim ensional space w ith
h er unaided (repeat: unaided) eye.
In each of these cases, w hat is finally m astered is a conceptual
fram ew ork— w hether m usical, chem ical, or astronom ical— a fram e
w ork that em bodies far m ore w isdom about the relevant sensory
dom ain than is im m ediately app aren t to untu tored discrim ination.
Such fram ew orks are characteristically a cultural heritage, pieced
togeth er over m any generations, and their m astery supplies a rich
n ess and penetration to our sen sory lives that w ould be impossible
in their absence. (The role of theory in perception and the sys
tem atic en h an cem en t of perception through theoretical p rogress are
exam ined at length in P. M. C hurchland 1979, sections 1 throu gh 6.)
O ur introspective lives are already the extensive beneficiaries of this
p h en om en on. The introspective discrim inations w e m ake are for the
m ost part learned; they are acquired w ith practice and experience,
often quite slowly. A nd the specific discrim inations w e learn to m ake
are those it is useful for us to m ake. G enerally, those are the dis
crim inations that others are already m aking, the discrim inations em
bodied in the psychological vocabulary of the language w e learn. The
conceptual fram ew ork for psychological states that is em bedded in
ordinary language is a m odestly sophisticated theoretical achieve
m en t in its ow n right, and it shapes our m atu red introspection p ro
foundly. If it em bodied substantially less w isdom in its categories and
connecting generalizations, our introspective apprehension of our in
ternal states and activities w ould be m u ch dim inished, though our
native discrim inatory m echanism s rem ain the sam e. Correlatively, if
folk psychology em bodied substantially m ore w isdom about our inner
nature than it actually does, our introspective discrim ination and
recognition could be very m u ch greater than it is, though our native
discrim inatory m echanism s rem ain u nch an ged .
This brings m e to the central positive suggestion of this p ap er. C on
sider now the possibility of learning to describe, conceive, and intro-
spectively apprehend the teem ing intricacies of our inner lives w ithin
the conceptual fram ew ork of a m atured n euroscience, a neuroscience
th at successfully reduces, either sm oothly or roughly, our com m on
sense folk p sychology. Suppose w e trained our native m echanism s to
m ake a new and m ore detailed set of discrim inations, a set that co r
respon d ed n ot to the primitive psychological taxon om y of ordinary
language, but to som e m ore penetrating taxon om y of states draw n
from a com pleted neuroscience. A n d sup p ose w e trained ourselves to
respon d to that reconfigured discrim inative activity w ith jud gm en ts
that w ere fram ed, as a m atter of course, in the appropriate con cep ts
The Direct Introspection of Brain States 55
only a single point of view , the subjective point of view of the ex
periencing subject. The properties of physical brain states, by co n
trast, are accessible from a variety of entirely objective points of view .
W e can n ot hope adequately to account for the form er, therefore, in
term s of properties appropriate to the latter dom ain, (see N agel 1974,
p p . 4 4 2 -4 4 4 .)
This som ew h at diffuse argum ent app ears to be an instance of the
following argum ent.
(1) Fa
(2) ~Fb
(3) а Ф b.
The problem w ith all of these argum ents is that the "p ro p e rty "
ascribed in prem ise (1) and witheld in prem ise (2) consists only in the
subject item 's being recognized, perceived, or know n as som eth in g, under
som e specific description or other. Such apprehension is n ot a genuine
feature of the item itself, fit for divining identities, since one and the
sam e subject m ay be successfully recognized und er one description
(e .g ., 'qualia of m y m ental state'), and yet fail to be recognized under
an oth er, equally accurate, coreferential description (e .g ., 'prop erty
of m y brain state'). In logician's term s, the propositional function, ‘x
is know n (perceived, recognized) by m e, as an F' is one of a large
num ber of intensional contexts w hose distinguishing feature is that
they do not alw ays retain the sam e truth value through substitution
of a coreferential or coextensive term for w hatever holds the place of
'x'. A ccordingly, that such a context (i.e ., the one at issue) should
show a difference in truth value for tw o term s 'a! an d 'V (i.e.,
'qualia of m y sensations' and 'property of m y b rainstates') is there
fore hard ly grou n d s for concluding that 'a' and 'V can n ot be core
ferential or coextensive term s! (I believe it w as Richard Brandt and
Jaegw on Kim (1967) w ho first identified this fallacy specifically in con
nection w ith the identity theory.)
This objection is decisive, I think, but it does not apply to a differ
ent version of the argum en t, w hich w e m u st also consider. It m ay be
u rged that one's brain states are m ore than m erely not (yet) know n by
introsp ection : they are not know able by introspection u n d er any cir
cu m stan ces. In corresp on dence, Thom as N agel has advised m e that
w hat he w ishes to defend is the following m odalized version of the
argu m en t.
The m atter of introspecting on e's brain states will arise once m ore
in the final section of this paper. For now , let us m ove on.
(a) M ary has m astered the com plete set of true propositions
about p eople's brain states.
(b) M ary does not have a representation of red n ess in her pre-
linguistic m edium of representation for sen sory variables.
Prem ises (a) and (b ) are com possible, even on a m aterialist view . But
they do n ot entail (c).
In su m , there are p retty clearly m ore w ays of "h av in g k n ow led ge"
than h avin g m astered a set of sentences. A nd nothing in m aterialism
p recludes this. The m aterialist can freely adm it that one h as "k n o w l
ed g e" of o n e's sensations in a w ay that is ind ep en d en t of the scien
tific theories one h as learned. This does n ot m ean that sensations are
beyond the reach of physical science. It ju s t m eans that the brain uses
more m odes and m edia o f representation than the sim ple storage o f sentences.
A nd this proposition is pretty obviously true: alm ost certainly the
brain u ses a considerable variety of m odes and m edia of rep resen ta
tion, p erh ap s h u n d red s of them . Jackson 's argu m en t, and N agel's,
exploit this variety illegitimately: both argu m en ts equivocate on
'know s ab ou t'.
This criticism is supported by the observation that, if Jackson's
form of arg u m en t w ere sound, it w ould prove far too m u ch . Suppose
that Jackson w ere arguing not against m aterialism , but against dual
ism : against the view that there exists a nonm aterial su b stan ce— call
it 'ecto p lasm '— w h ose hidden constitution and nom ic intricacies
grou n d all m ental p h en om en a. Let our cloistered M ary be an "e cto -
plasm ologist" this tim e, and let her kn ow ! (by description) every
thing th ere is to know about the ectoplasm ic processes underlying
vision. There w ould still be som ething she did not k now 2 (by ac
quaintance): w h at it is like to see red. D ualism is therefore inadequate
to accou n t for all m ental p henom ena.
This arg u m en t is as plausible as Jackson 's, and for the sam e reason :
it exploits the sam e equivocation. But the truth is, such argum ents
show nothin g, one w ay or the other, about how m ental p h en om en a
m ight be accou n ted for.
64 The Nature of Mind
hyp oth esis that those three param eters are rep resented in our visual
system s, and that our sensations-of-color are in som e direct w ay d e
term ined by them . Sensations-of-color m ay turn out literally to be
three-elem en t chords in som e neural m edium ! In the face of all this, I
do n ot see w hy it is even briefly plausible to insist that it is utterly
im possible for a conceptually sophisticated M ary accurately to im
agine, and then reliably pick out, color sensations she has not
previously enjoyed. W e can already foresee how it m ight actually
be done.
The preceding argum ent does not collapse the distinction (betw een
know ledge by description and know ledge by acquaintance) urged
earlier in the discussion of equivocation. But it does show that the
"taxo n o m ie s" that reside in our prelinguistic m edia of representation
can be profoundly shaped by the taxonom ies that reside in the lin
guistic m edium , especially if one has had long practice at the obser
vational discrim ination of item s that an sw er to those linguistically
em bodied categories. This is just a further illustration of the plasticity
of h um an perception.
I do not m ean to suggest, of course, that there will be no limits to
w hat M ary can im agine. H er brain is finite, and its specific an atom y
will have specific limitations. For exam ple, if a b at's brain includes
com putational m achinery that the h um an brain sim ply lacks (w hich
seem s likely), then the subjective ch aracter of som e of the b at's inter
nal states m ay well be beyond h um an im agination. Clearly, h ow ever,
the elusiveness of the b at's inner life here stem s not from the m eta
physical "e m e rg e n ce " of its internal qualia, but only from the finite
capacities of our idiosyncratically hum an brains. W ithin those sheerly
structural lim itations, our im aginations m ay soar far beyond w hat
Jackson, N agel, and Robinson suspect, if w e p ossess a neuroscientific
conceptual fram ew ork that is at last adequate to the intricate p h e
n om en a at issue. (See especially chapter 5, section 7, and ch apter 9,
section 4.)
I suggest, then, that those of us w ho prize the flux and con ten t
of our subjective phenom enological experience need not view the
ad van ce of materialistic neuroscience w ith fear and foreboding.
Quite the contrary. The genuine arrival of a m aterialist kinem atics
and dynam ics for psychological states and cognitive p rocesses will
constitute not a gloom in w hich our inner life is supp ressed or
eclipsed, but rather a daw ning, in w hich its m arvellous intricacies
are finally revealed— m ost notably, if w e apply ourselves, in direct
self-conscious introspection.
Chapter 4
Knowing Qualia: A Reply to Jackson
A n argu m en t form w ith one invalid instance can be exp ected to have
others. This w as the point of a subsidiary objection in m y 1985b
pap er: if valid, Jackson's argum ent, or one formally parallel, w ould
also serve to refute the possibility of substance dualism . I did n ot there
exp ress m y point w ith notable clarity, how ever, and I accep t resp on
sibility for Jackson 's quite missing m y intention. Let m e try again.
The basic point is that the canonical presentation of the know ledge
argu m en t, as outlined on p. 67 above, w ould be just as valid if the
predicate term 'P' w ere everyw h ere replaced by ' £ '. A n d the result
ing prem ises w ould be just as plausibly true if
(1) ' £ ' stood for 'is about som ething ectoplasm ic in ch aracter'
(w here 'ectoplasm ' is an arbitrary nam e for the dualist's
nonphysical substance), and
(2) the story is altered so that M ary b ecom es an exhaustive
expert on a com pleted ectoplasm ic science of h u m an nature.
The plausibility w ould be com parable, I subm it, b ecause a long dis
cursive lecture on the objective, statable, law -governed p roperties of
ectoplasm , w hatever they m ight be, w ould be exactly as useful, or
u se less , in helping M ary to know -by-acquaintance "w h a t it is like to see
re d ," as w ould a long discursive lecture on the objective, statable,
law -governed properties of the physical m atter of the brain. E ven if
substance dualism w ere true, therefore, and ectoplasm w ere its
heroic principal, an exactly parallel "k n ow ledge arg u m en t" w ould
"s h o w " that there are som e aspects of consciousness that m u st for
72 The Nature of Mind
ever escape the ectoplasm ic story. Given Jackson 's antiphysicalist in
tentions, it is at least an irony that the sam e form of argum ent should
incidentally serve to blow substance dualism out of the w ater.
Though I am hardly a substance dualist (and neither is Jackson), I
do regard substance dualism as a theoretical possibility, one that
m ight conceivably succeed in explicating the psychological ontology
of com m on sense in term s of the underlying properties and law-
governed behavior of the nonm aterial substance it postulates. A n d I
m u st p rotest that the parallel know ledge argu m en t against substance
dualism w ould be wildly unfair, and for the v ery sam e reason that its
analogue against physicalism is unfair: it w ould equivocate on
'know s about'. It w ould be no m ore effective against dualism than it
is against m aterialism .
The parallel un der exam ination contains a further lesson. If it
w orks at all, Jackson's argum ent w orks against physicalism not b e
cau se of som e defect that is unique to physicalism ; it works because no
am ount o f discursive know ledge, on any topic, w ill constitute the nondis-
cursive fo rm o f know ledge that M ary lacks. Jackson 's argum ent is one
instance of an indiscrim inately antireductionist form of argum ent. If it
w orks at all, an analog will w ork against any proposed reductive,
discursive, objective account of the natu re of our subjective experi
en ce, no m atter w hat the reducing theory m ight happen to be. I see
this as a further sym ptom of the logical pathology described earlier.
Since the argum ent "w o rk s" for reasons that have nothing essential
to do w ith physicalism , it should "w o rk " against the explanatory
aspirations of other ontologies as well. A n d so it " d o e s ." The price of
em bracing Jackson's argum ent is thus dram atically higher than first
ap p ears. For it m akes any scientific accou n t of our sensory experience
entirely im possible, n o m atter w hat the ontology em ployed.
This arg u m en t is also form ally valid, and its prem ises explicitly en
com p ass w h atever variety there m ay be in form s of know ing. W hat
can w e say about its soundness?
A ssu m e that M ary has had the upbringing described in Jackson's
story, an d thus lacks any know ledge-by-acquaintance w ith "w h a t it is
like to see re d ." Prem ise (2') will then be true, as and for the reasons
that Jack son 's story requires. W h at will be the truth value of prem ise
(1') on these assum ptions?
Prem ise (1') is now a very strong claim indeed, m u ch stron ger than
the old prem ise (1), and a m aterialist will be sure to insist that it is
false. The reason offered will be that, because of her deprived u p
bringing, M ary quite clearly lacks one form of know ledge of a certain
physical asp ect of people. Specifically, she lacks a p ro p er configura
tion of synaptic connections m eeting the neurons in the appropriate
area of h er visual cortex. She thus lacks an appropriately partitioned
activation vector space across those n eurons, and therefore has no
rep resentation , at that site, of the full range of sensory coding vectors
that m ight som eday com e from the retina and the LG N . In other
w ord s, there is som ething physical about person s (their color sen sa
tions, or identically, their coding v ectors in their visual p ath w ays),
and there is som e form of know ledge (an antecendently partitioned
prelinguistic taxon om y), such that M ary lacks that form of know ledge
of that aspect of p ersons. A ccordingly, prem ise (1 ') is false and the
conclusion (3') is not sustained.
F rom a m aterialist's point of view , it is obvious that (1') will be false
on the assum ptions of Jackson's story. For that story denies her the
upbringing that norm ally provokes and shapes the d evelopm ent of
the relevan t representation across the appropriate population of cor
tical n eu ron s. A nd so, of course, there is a form of know ledge, of a
physical asp ect of p erson s, that M ary does n ot have. A s just illus
trated, the m aterialist can even specify that form of know ledge, and
its objects, in neural term s. But this m eans that prem ise (1 '), as p ro p
erly quantified at last, is false. M ary does not have know ledge of
everyth in g physical about persons, in every w ay that is possible for
her. (That is w hy prem ise (2') is true.)
T here is, of course, no guarantee that the m aterialist's accou n t of
sensations and sensory recognition is correct (although the ex
perim ental and theoretical evidence for a view of this general kind
continues to accum ulate). But neither is Jackson in a position to insist
74 The N ature of Mind
1 Introduction
• The m ystery of how the brain represents the w orld, and how it
p erform s com putations on those representations
• The m ystery of sensorim otor coordination
• The m ystery of the brain's m icrophysical organization.
M olecular
layer
E xte rnal
g ran u lar
layer
E xternal
p yram id al
layer
Internal
g ranu lar
layer
In te rn al
p yram id al
layer
M u ltifo rm
layer
Figure 5.1
(a) A cross-section of the cerebral hemispheres showing the outer gray layer, the cere
bral cortex. (b) The internal laminar structure of the cortical layer, as revealed by three
different stains, (c) Brodmann's areas
Reductive Strategies in Cognitive Neurobiology 81
Finally, these distinct layers are system atically con n ected , in the
fashion of nails struck through plyw ood, by large num bers of vertical
ly oriented cells that perm it com m unication betw een the several
layers. These vertical cells conduct neuronal activity d ow n w ard s,
from the superficial input layers above to the output layer below.
If w e now leave our m icroscopic edgew ise perspective and look at
the cortical sheet from the outside, w e find that the cortical surface
divides into a patchw ork of sm aller regions (see figure 5.1c). These
areas are distinguished to som e degree by differences in their lam inar
cytoarch itectu res. An initial taxonom y, into w hat are called "B rod -
m an n 's are a s" after their discoverer, is based on this criterion. These
areas or subareas of these areas are of further interest b ecause several
of th em plainly constitute topographic m aps of som e aspect of the sen
sory or m otor periphery, or of som e other area of the brain. For exam
ple, the neighborhood relations holding betw een the cells in a given
layer of the visual cortex at the rear of the brain corresp on d to the
neighborhood relations holding betw een the cells in the retina from
w hich they receive inputs. The bundle of axonal projections from the
retinal cells to the cortical cells preserves the topographic organiza
tion of the retinal cells. The surface of the prim ary visual cortex thus
constitutes a topographic m ap of the retinal surface.
It is term ed a 'topograp hic m ap' rath er than simply a 'm ap ' be
cause the distance relations am ong retinal cells are generally not p re
82 The Nature of Mind
served. Typically, such m aps are m etrically deform ed, as if they w ere
m ade of rubber and then stretched in som e fashion.
M any such m aps have been identified. The so-called "visu al cor
te x " (areas 17, 18) has already b een m entioned. The upper layer of the
som atosensory cortex (area 3) is a topographic m ap of the b od y's tac
tile surface. The low er layer of the m otor cortex (area 4) is a top o
graphic m ap of the bod y's m uscle system . The auditory cortex (areas
41, 42) contains a topographic m ap of frequency space. A nd there are
m any oth er cortical areas, less well u n d erstood as to exactly w hat
they m ap, but w hose topographical re-presen tation of distant stru c
ture is plain.
This general pattern of neural organization is not confined to the
surface of the great cerebral hem ispheres. A s well, various nuclei of
"g re y m atter" in the m ore central regions of the brain— for exam ple,
the superior colliculus, the hippocam pus, and the lateral geniculate
nu cleu s— display this sam e m ultilayered, topographically organized,
vertically con n ected structure. N ot everything does (the cerebellum ,
for exam ple, is rather different, of w hich m ore later), but the p attern
described is one of the m ajor organizational patterns to be found in
the brain.
W h y this p attern ? W hat is its functional or cognitive significance?
W hat do these structures do, and h ow do they do it? W e can
approach a possible answ er to these questions by addressing the
second m ystery, the problem of sensorim otor coordination.
Figure 5.2
The tip of the arm m ust m ake contact w ith this point. If w e assum e
that b oth the forearm and the upper arm have a fixed length of 7
units, the elbow will therefore have to lie at the intersection of two
84 The Nature of Mind
Figure 5.3
Reductive Strategies in Cognitive Neurobiology 85
circles of radius 7 units: one centered at (a, b), and the oth er centered
at (0, 0), w here the upper arm projects from the crab's b od y. If w e
solve for the relevant intersection, the real-space elbow coordinates
(xe, i/e) are given by
Xe = ((49 - ((fl2 + b2)2/ 4b2) ■(1 - (( a2/b2)/{(a 2/b2) + 1 )) ) ) ^
+ ({{alb) ■((a2 + b2)/2 b ))/((a2/b2) + 1 ) Щ / ( ( а 2/Ь2) + 1)1/2
y e = (49 - xe2)1/2
The three points in real space, (a, b), (xe, y e), (0, 0), determ ine the
position of the arm , w hose upper arm and forearm angles <6, <p> are
finally given by
0 = tan ~1(y e /x e)
ip = 180 — (0 — ta n _1((t> — y e) l( a ~ xe)))
These are the desired coordinates of the arm in m otor state space. The
reader will note that the assem bled functions that yield them are
rath er tangled ones.
Tangled or not, if the crab is draw n on a com p u ter screen such that
its final arm position (draw n by the com p uter as output) is the speci
fied function of its eye positions (entered by us as inp u t), then it
constitutes a very effective and w ell-behaved sensorim otor system ,
especially if w e w rite the controlling p rogram as follows.
L et the p rogram hold the crab's arm folded against its chest (at
0 = 0°, q> = 180°), until som e suitable stim ulus registers on the fovea of
both ey es. The arm is then m oved from its initial state-sp ace position
(0°, 180°) along a straight line in m otor-state space, to its com puted
target position in m otor-state space. This is the state-space position at
w hich, in real sp ace, the tip of the arm contacts the triangulation point
of the eyes. This arran gem ent produces a m odestly realistic system
that reach es unerringly for w hatever it sees anyw here w ithin reach of
its arm (figure 5 .4 a - d ) .
The algebraic rep resentation of the crab's sensorim otor tran sform a
tion, as rep resented in the six equations listed earlier, supplies no
intuitive conception of its overall nature. A geom etrical presen tation is
m u ch m o re revealing. Let us therefore consider the projection of the
active p ortion of the crab's sensory state space (figure 5.5a) onto the
o rth ogon al grid of its m otor state space (figure 5 .5 b), as im p osed by
the function un d er discussion. That is to say, for every point in the
displayed sen sory grid, w e have plotted the corresponding arm -
position w ithin the m otor grid.
H ere w e can see at a glance the distortion of the vertical and hori
zontal lines of sensory space, as projected onto m otor space. The
86
The Nature of Mind
Figure 5.4
The crab's arm in action
Figure 5.5
Reductive Strategies in Cognitive Neurobiology
A transformation of coordinates
87
88 The Nature of Mind
topological features of the sensory space are p reserved , but its m etric
al properties are not. W hat w e see is a system atic transform ation o f
coordinates. (The heavy scored triangle and rectangle are draw n in
solely to help the reader locate corresponding positions in the d e
form ed and undeform ed grids. N ote also that the left border or Д axis
of figure 5.5a shrinks to the left radial point in figure 5.5b, and that the
top b order of figure 5.5a shrinks to the right radial point in figure
5.5b.)
Figure 5.6
89
Effecting a coordinate transformation by means of contiguous, metrically deformed, vertically connected topographic maps
90 The Nature of Mind
screen, and its conduction velocities are on the order of the speed of
light. Evidently, the m assively parallel architecture of the state-space
sandw ich buys it a large advantage in speed, even w ith vastly slower
com p onen ts.
Third, the quality of the crab's coordination will not be uniform
over its field of m otor activity, since in the m axim ally deform ed areas
of the sen sory grid, small errors in sensory registration p rodu ce large
errors in the m otor response (see again figure 5.5b). A ccordingly, the
crab is least well coordinated in the area close betw een its eyes, and to
its extrem e right and left.
All three of these functional properties are biologically realistic.
A nd the sandw ich appears biologically realistic in one further respect:
it is relatively easy to imagine such a system being grow n . Distinct
layers can confine distinct chem ical gradients, and can thus guide
distinct m orp h ogenetic p rocesses. A ccordingly, distinct top ogra
phic m ap s can app ear in closely adjacent layers. But given that the
m aps are so closely contiguous, and given that they are ap p rop riate
ly deform ed, the problem of connecting them up so as to produ ce a
functional system becom es a trivial one: the solution is just to grow
conductive elem ents that are roughly orthogonal to the layers.
Different creatures will have different m ean s of locating objects,
and different m otor system s to effect contact w ith them , but all of
them will face the sam e problem of coordinating positions in their
sen sory state space with positions in their m otor state sp ace, and the
style of solution here outlined is evidently quite general in nature. In
fact, th e coordination of distinct biological subsystem s by coordinate
transform ation is a m atter that presum ably extends far beyond the
obvious case of basic sensorim otor coordination. The sam e strategy
m ay also be useful, or even essential, in the execution of higher cogni
tive activities, as w e shall see later on. The point to be em phasized
here is that a state-space sandw ich constitutes a simple and bio
logically realistic m eans for effecting any tw o-dim ensional to two-
dim ensional coordinate transform ation, w hatever its m athem atical
com plexity and w h atever features— external or internal, abstract or
co n crete— the coordinate axes m ay rep resent to the brain. If the
transform ation can be graphed at all, a sandw ich can com p u te it. The
sen sorim otor problem solved above is m erely a tran sp aren t exam ple
of the general technique at w ork.
B eyond its functional realism , the system of intercon n ected m aps
in figure 5 .6 is suggestively similar to the know n physical structure of
typical lam inar cortex, including the m any topographic m ap s distrib
u ted across the cerebral surface. In all of these areas, inputs address a
given layer of cells, w hich layer frequently em bodies a m etrically de
92 The Nature of Mind
Figure 5.7
(a) A projection drawing of a Nissl-stained cross-section of cat superior colliculus illustrating laminar organization. Dots correspond to
collicular neurons. From Kanaseki and Sprague 1974; reprinted with permission, (b) A retinotopic map, a metrically deformed topographic
93
map of the visual hemifield, in rectangular coordinates, on the superficial layer of the right colliculus of the cat. M = medial; L = lateral;
A = anterior; P = posterior. Adapted from Schiller 1984
94 The Nature of Mind
vertical elem ents to the m otor m ap, w here it produ ces a saccade of
just the size and direction appropriate for foveating the external sti
m ulus th at provoked it. The superior colliculus thus ap p ears to be an
instance of both the structural and the functional pattern displayed in
figure 5 .6 .
O ne m ight exp ect a biological sandw ich to code the position of re
tinal stim ulations w ith an area of stim ulation in the upp er m ap , rather
than a single point, so as to be functionally persistent in the face of
small lesions and scattered cell death, as explained earlier. A ctivity in
the SC does display this pattern (M cllwain 1975; 1984, p . 268). The
schem atic m odel of figure 5 .6 also predicts that the size and direction
of the m o to r response induced by m icrostim ulation at various points
within the collicular sandw ich will be a function solely of w here in
either m ap the stim ulation occurs, and not of the m agn itu de of the
stim ulation nor of its vertical position betw een the tw o m ap s. Experi
m en tation has already yielded this result (Robinson 1972; Schiller and
Stryker 1972). The superior colliculus, it appears, is a real sensorim o
tor coordinate transform er of roughly the kind at issue. It foveates on
changing or m oving visual targets by essentially the sam e m eans
w hereby the schem atic cortex of the crab reaches out for triangulated
objects.
A w ord of caution is in order here, since the accou n t just offered
does not do justice to the full com plexity of the superior colliculus. In
m am m als, especially the higher m am m als, the SC is a tightly inte
grated part of a larger m odulating system that includes inputs from
the visual cortex and the frontal eye fields, and outputs to the neck
m uscles. The functional properties of the entire system are m ore
varied and m ore subtle than the preceding accou n t su ggests, and the
job of sorting them out is still un d erw ay (M ays and Sparks 1980;
Schiller and Sandell 1983). The preceding is subm itted as an account
of the central or m ore prim itive functions of the SC, at best.
W ith the exam ples of the crab's "c o rte x " and the superior colliculus
in m ind, it is appropriate to focus on the m any other topographically
organized, m ultilayered cortical areas scattered th rou g h o u t the brain,
and ask w hat coordinate transform ations they m ight be effecting.
H ere it is v ery im portant to appreciate that the top ograp h ic m aps w e
seek to d ecode need n ot, and generally will not, be m ap s of som e
thing anatom ically obvious, such as the surface of the retin a, or the
surface of the skin. M ore often they will be m aps of som e abstract state
space, w h ose dim ensional significance is likely to be opaque to the
casual observer, though of great functional im portance to the brain.
Two nice exam ples of such abstract m aps are the m ap of echo delays in
96 The Nature of Mind
W hile w e are discussing the biological reality of the lam inar m ech a
nism p roposed , consider the objection that our m odel cortex has only
tw o layers, w hereas typical h um an cortex has six layers and, if w e
count fine subdivisions, p erhaps eight or nine in som e areas. W hat
are they for?
There is no difficulty in perceiving a function for such additional
layers. Let us retu rn again to the superior colliculus, w hich illustrates
one of m any possibilities here. Betw een the visual and m otor m ap s of
the superior colliculus there are, in som e creatu res, one or tw o in
term ediate layers (see again figure 5.7). These ap p ear to constitute an
auditory m ap an d /or a som atosen sory m ap (a facial or w hisker m ap),
Reductive Strategies in Cognitive Neurobiology 97
w h ose function is again to orient the ey e's fovea, this tim e tow ard
the sou rce of sud den auditory an d /or som atosensory stim ulation (Gold
berg an d Robinson 1978). N ot surprisingly, these intervening m aps
are each m etrically deform ed in such a fashion as to be in rou gh co
ordinate register w ith the m otor m ap and hence w ith each other.
A ltogeth er, this elegant three- or four-layer top ograp h ic sandw ich
con stitu tes a m ultim odal sensorim otor coordinate transform er.
M ultilayered structu res have further virtues. It is plain that m aps of
several distinct m odalities, suitably deform ed and placed in collective
register w ithin a "club san d w ich ," provide a m ost effective m ean s of
cross-m od al integration and com parison. In the SC, for exam ple, this
m ultim odal arrangem ent is appropriate to the p roduction of a m otor
resp o n se to the jo in t receipt of faint but spatiotem porally coincident
auditory and visual stimuli, stimuli that, in isolation, w ould have
been su bthreshold for a m otor respon se. For exam ple, a faint sound
from a certain com pass point m ay be too faint to p rom p t the eyes into
a foveating saccade, and a tiny m ovem en t from a certain com pass
point m ay be similarly im potent, but if both the sound an d the m ove
m en t com e from the sam e com pass point (and are thus cod ed in the
SC along the sam e vertical axis), then their sim ultaneous conjunction
will indeed be sufficient to m ake the m otor layer direct the eyes
appropriately. This prediction is strongly corroborated by the recent
results of M eredith and Stein (1985).
F u rth er exploration reveals that m ultilayered sandw iches can sub
serve decidedly sophisticated cognitive functions. In an earlier pub
lication on these m atters (1986d), I have show n how a three-lay er
state-sp ace sandw ich can code, and project, the path of a m oving
object in such a fashion as to position the crab's arm to catch the
m oving target on the fly. Evidently, a m ultilayered cortex can offer
considerable advantages.
The exam ples studied above are uniform in having an input state
space of only tw o dim ensions and an ou tp u t state space of only tw o
dim ensions. This allows the required coordinate transform ation to be
achieved by a contiguous pair of sheetlike m aps. But w h at of cases
w here the subsystem s involved each have m ore than tw o p aram
eters? W h at of cases w here the coordinate transform ations are from
an inp u t space of n dim ensions to an ou tp u t space of m dim ensions,
w h ere n and m are different and both g reater than tw o? C onsider, for
exam ple, the problem of coordinating the joint angles of a limb with
three or m ore joints, and the problem of coordinating several such
98 The Nature of Mind
Figure 5.8
Vector-to-vector transformation by matrix multiplication
Figure 5.9
Effecting vector-to-vector transformations with a neural net
to yield y, and again w ith the third colum n to yield z. Thus results the
ou tp u t v ector <x, y, z>.
This algebraic operation can be physically realized quite sim ply by
the neural array of figure 5 .9 . The parallel input fibers at the right
each send a train of electrochem ical "sp ik es" tow ard the w aiting
dendritic trees. The numbers a, b, c, d represent the am ount by which
the m o m en tary spiking frequency of each of the four fibers is above
(positive num ber) or below (negative num ber) a certain baseline spik
ing frequency. The top -m ost input fiber, for exam ple, syn ap ses onto
each of the three outpu t cells, m aking a stim ulatory connection in
each case, one that tends to depolarize the cell body and m ake it send
a spike dow n its vertical output axon. The output frequency of spike
em issions for each cell is determ ined by the simple frequ en cy of input
stim ulations it receives from all incom ing synaptic con n ection s, and
100 The Nature of Mind
P A R A L LE L F IB E R S MAKE
M U L TIP LE SYNAPSES
PUR KINJE A X O N O U TP U T
Figure 5.10
A schematic section of the cerebellum (cell population and fiber density reduced for
clarity)
w hen the representational and com putational task exceeds the case of
tw o dim ensions. O n the contrary, the brain boasts neural m achinery
that is ideally suited to cases of v ery high dim ensionality. W e have,
then, at least tw o know n brain m echanism s for perform ing coordi
n ate transform ations: the state-space sandw ich specifically for
tw o-dim ensional cases, and the neural m atrix for cases of any
dim ensionality w hatever.
cifiable w ays (it fails to display a con trast betw een red and green, or
betw een blue and yellow , etc.). This question of the relative variety of
qualia displayed w ithin a given m odality raises the point that across the
familiar five m odalities there is notew orth y variation. F or exam ple,
thou gh the variety of discriminable color sensations is large, the variety
of discrim inable taste sensations is even larger, an d the vari
ety of discriminable smell sensations is larger still. Such variation
rem inds us further of the presum ed variation across species, as in
stanced in the canine's extraordinary ability to discrim inate, by smell
alone, any one of the 3.5 billion people on the planet. O ne presu m es
that the canine's continuum of olfactory sensations is som eh ow m uch
"la rg e r" than a h u m an 's, in the sense of containing a g reater variety
of discrim inable types of sensation.
H ere, then , are som e hu m dru m facts about the m anifold(s) of sub
jective sen sory qualia, facts w hich a reductive account of m ind m ight
attem p t to explain. It m ust do this by reconstructing these facts, in
som e revealingly system atic w ay, in neurobiological term s. (For a
general accoun t of the nature of cross-theoretic identities and inter
theoretic red u ction s, see chapter 3 and P. M. C hurchland 1979.)
This possibility will now be explored. For several of the relevant
m odalities, physiological and cognitive psychologists have already
sketched the outlines of such an account, and state-sp ace rep resen
tations play a p rom inent role in all of them .
C onsid er first the abstract three-dim ensional "co lo r cu b e" p ro
posed by Edwin Land (1977), within which every one of the m any hun
dreds of hum anly discriminable colors occupies a unique position or
small volum e (figure 5.11). Each axis rep resents the eye/brain's recon
struction of the objective reflectance of the seen object at one of the
three w avelengths to w hich our cones are selectively responsive. Two
colors are closely similar just in case their state-space positions w ithin
this cube are close to one another. A n d tw o colors are dissim ilar just
in case their state-space positions are distant. W e can even speak of
the d egree of the similarity, and of the dim ensions along w hich it is
reckon ed . (See also Zeki 1983.)
If the h u m an brain does possess an internal im plem entation of
such a state sp ace, it has purchased a great deal of representational
p ow er at a very low price. For exam ple, if our native discrim ination
along each axis of L an d 's color state space is only 10 distinct posi
tions, then a ternary system should be able to rep resent fully 103 dis
tinct colors. If anything, this underestim ates our capacities, so the
assu m p tion of 10-unit axial discrim ination is likely too low. In any
case, there is no trouble accounting for our broad discrim inatory p ow
ers: o n e's discrim ination w ithin L an d 's state space explodes as the
104 The Nature of Mind
Figure 5.11
Color state space
dimensional space has greater volume still, and will permit even great
er feats of discrim ination. A six-dim ensional space, at 10-unit axial
discrim ination, will perm it the discrim ination of 106 odors. A nd if w e
im agine only a seven-dim ensional olfactory space, w ith only three
tim es the h um an axial discrim ination, w hich space a dog alm ost cer
tainly possesses, then w e are contem plating a state space w ith 307, or
22 billion, discriminable positions! W ith such a space, a canine's abil
ity to distinguish by smell any one of the 3 .5 billion people on the
planet no longer p resents itself as a m ystery.
I have neither the space nor the understan din g to discuss the com
plex case of auditory qualia, but here too a state-space ap proach is
claim ed to be illuminating (see Risset and W essel 1982). D epending
on the research ers and the m odality involved, the state-space
approach is variously called 'm ultivariate analysis', 'm ultidim en
sional scaling', 'across-fiber pattern coding', 'vector coding', and so
forth. But these are all alternative incarnations of the sam e thing:
state-space representations.
Evidently, this approach to understanding sensory qualia is both
theoretically and empirically m otivated, and it lends support to the
reductive position advanced in ch apter 3 on the ontological status of
sen sory qualia. In particular, it suggests an effective m eans of ex
pressing the allegedly inexpressible. The "ineffable" pink of one's
cu rrent visual sensation m ay be richly and precisely expressible as a
95H z/80H z/80H z "c h o rd " in the relevant triune cortical system . The
"u n co n v ey ab le" taste sensation produced by the fabled A ustralian
health tonic V egam ite m ight be quite poignantly conveyed as a 85/80/
90/15 "c h o rd " in one's four-channeled gustatory system (a dark co r
ner of taste-space that is best avoided). A n d the "indescribable" olfac
tory sensation produced by a new ly opened rose m ight be quite
accurately described as a 95/35/10/80/60/55 "c h o rd " in som e six
dim ensional system w ithin on e's olfactory bulb.
This m ore penetrating conceptual fram ew ork m ight even displace
the com m on sen se fram ew ork as the vehicle of intersubjective d e
scription and spontaneous introspection. Ju st as a m usician can learn
to recognize the constitution of heard m usical chords, after internaliz
ing the general theory of their internal stru ctu re, so m ay w e learn to
recognize, introspectively, the n-dim ensional constitution of our sub
jective sensory qualia, after having internalized the general theory of
their internal structure. This analogy has the further advantage of
preem p ting the predictable response that such a reconception of the
"in ternal w orld " w ould rob it of its beauty and peculiar identity. It
w ould do so no m ore than reconceiving m usical p h enom ena in term s
of harm onic theory robs m usic of its beauty and peculiar identity. On
Reductive Strategies in Cognitive Neurobiology 107
Figure 5.12
A state-space representation of complex locomotor activity
8 Concluding R em arks
form ations suggest a pow erful and highly general m eans of u n d er
standing the cognitive activities of the n ervou s system . Especially
since the physical m echanism s appropriate to im plem ent such a sys
tem are w idespread throughou t the brain.
Chapter 6
Folk Psychology and the Explanation of
Human Behavior
Folk p sych ology, insist som e, is just like folk m echanics, folk therm o
dynam ics, folk m eteorology, folk chem istry, and folk biology. It is a
fram ew ork of concepts, roughly adequate to the d em an d s of every
day life, w ith w hich the hum ble adept com p reh en d s, explains, pre
dicts, and m anipulates a certain dom ain of p h enom en a. It is, in short,
a folk theory. A s w ith any theory, it m ay be evaluated for its virtues or
vices in all of the dim ensions listed. A n d as w ith any theory, it m ay be
rejected in its entirety if it fails the m easure of such evaluation. Call
this the "th eoretical v iew " of our self understanding.
Folk p sych ology, insist oth ers, is radically unlike the exam ples
cited. It does n ot consist of laws. It does n ot support causal explana
tions. It does not evolve over tim e. Its central purpose is norm ative
rath er than descriptive. A nd thus, it is not the sort of fram ew ork that
m ight be show n to be radically defective by sheerly em pirical find
ings. Its assim ilation to theories is just a m istake. It h as nothing to
fear, therefore, from advances in cognitive theory or the neurosci
en ces. Call this the "antitheoretical v iew " of our self u n d erstanding.
Som ebody here is deeply m istaken. The first burden of this p aper is
to argue that it is the antitheoretical view that harbors m ost, though
not all, of those m istakes. In the thirty y ears since the theoretical view
w as introduced (see especially Sellars 1956; Feyerabend 1963a; Rorty
1965; P. M . C hurchland 1970, 1979; and ch apter 1), a variety of objec
tions have b een leveled against it. The m ore interesting of those will
be ad d ressed shortly. M y current view is that these objections m oti
vate no ch an ges w h atever in the theoretical view .
The secon d and m ore im portant burd en of this pap er, h ow ever, is
to outline and repair a serious failing in the traditional expressions of
the theoretical view , m y ow n expressions included. The failing, as I
see it, lies in representing on e's com m on sen se understan din g of hu-
Clearly these hum ble generalizations, and thousands m ore like them ,
are causal/explan atory in character. They will and regularly do sup
p ort sim ple explanations, sustain subjunctive and counterfactual
conditionals, and underw rite predictions in the standard fashion.
M oreover, con cep ts of this simple sort carry perhaps the m ajor p art of
the folk-psychological burden. The com paratively com plex explana
tions involving the propositional attitudes are of central im portance,
but th ey are surrounded by a quotidean whirl of simple explanations
like these, all quite evidently of a causal/explanatory cast.
It w o n 't do, then, to insist that the generalizations of folk psycholo
gy are on the w hole nonem pirical or noncausal in ch aracter. The bulk
of them , and I m ean thousands upon thou san d s of them , are tran s
parently causal or nom ological. The best one can hope to argue is that
there is a central core of folk-psychological concepts w h ose explana
tory role is som ehow discontinuous w ith that of their fellows. The
propositional attitudes, especially belief and desire, are the perennial
candidates for such a nonem pirical role, for explanations in their
114 The Nature of Mind
(x)(V )(R )[(x is subject to a voltage o f(V )) & (x offers a resistance o f(R ))
^ ( 3 I ) ( ( x has a current o f (I)) & (I = V/R))]
The first point to notice here is that the crucial p red icates— has a resist
ance o f(R ), is subject to a voltage o f(V ), and has a current o f (I)— are w hat
m ight be called "n u m erical attitu d es": they are predicate-form ing
functors that take singular term s for num bers in the variable position.
A com plete predicate is form ed only w hen a specific num eral appears
in the relevan t position. The second point to notice is that this elec
trodynam ical law exploits a relation holding on the dom ain of n u m
bers in ord er to exp ress an im portant empirical regularity. The
cu rren t I is the quotient of the voltage V and the resistance R , w hose
values will be cited in explanation of the cu rrent. A nd the third point
to notice is that this law and the explanations it sustains are typical of
law s and explanations throughou t science. M ost of our scientific p re
dicates exp ress num erical attitudes of the sort displayed, and m ost of
our law s exploit and display relations that hold prim arily on the
abstract dom ain of num bers. N or are they limited to num bers. O ther
laws exploit the abstract relations holding on the abstract dom ain of
vectors, or on the dom ain of sets, or groups, or m atrices. But none of
this m ean s they are nonem pirical, or noncausal, or nonnom ic.
A ction explanations, and intentional explanations in general, fol
low the sam e pattern. The only difference is that here the dom ain of
abstract objects being exploited is the dom ain of propositions, and the
relations displayed are logical relations. A nd like the num erical and
vectorial attitudes typical of theories, the expressions for the proposi
tional attitudes are predicate-form ing functors. Believes that P, for
exam ple, form s a com plete predicate only w hen a specific sentence
app ears in the variable position P. The principles that com p reh en d
these p red icates have the sam e abstract and highly sophisticated
stru cture displayed by our m ost typical theories. They just exploit the
relations holding on a different dom ain of abstract objects in order to
exp ress the im portant empirical regularities com p reh en d in g the
states and activities of cognitive creatures. That m akes folk psychol
ogy a v ery interesting theory, p erh ap s, but it is hard ly a sign of
its being nontheoretical. Q uite the reverse is true. (This m atter is dis
cu ssed at greater length in Churchland 1979, section 14, and 1981a,
pp. 8 2 -8 4 .)
116 The Nature of Mind
I focus here on one of the relatively few issues that still divide
Dennett and m e: the ontological status of intentional states. W e both
accept the prem ise that neuroscience is unlikely to find "se n te n ce s in
the h e a d ," or anything else that answ ers to the structure of individual
beliefs and desires. O n the strength of this shared assum ption, I am
willing to infer that folk p sychology is false, and that its ontology is
chim erical. Beliefs and desires are of a piece w ith phlogiston, caloric,
and the alchem ical essen ces. W e therefore need an entirely new
kinem atics and dynam ics w ith w hich to com p reh en d h u m an cogni
tive activity, one d raw n, perhap s, from com putational neuroscience
and connectionist AI. Folk psychology could then be p u t aside in
favor of this descriptively m ore accurate and explanatorily m ore
pow erful portrayal of the reality w ithin. C ertainly, it will be p ut aside
in the lab and in the clinic, and eventually, p erh ap s, in the m arket
place as well.
But D enn ett declines to draw this eliminativist conclusion, despite
his firm accep tan ce of the prem ise cited, and despite his willingness
to contem plate u n orthod ox form s of cognitive theory. H e prefers to
claim a special status for the various intentional states, a status that
will perm it us to be "realists" about beliefs and desires despite their
projected absence from our strict scientific ontology.
This im pulse in D ennett continues to strike m e as arbitrary p ro tec
tionism , as ill m otivated special pleading on behalf of the old and
familiar. H is initial rationale for exem pting folk psychology from the
usual scientific standards involved assigning it a purely instrum ental
status, but this swiftly brought him all kinds of grief, as he him self
explains (1987, pp. 7 1 -7 2 ). Instrum entalism is first and forem ost an
anh'realist position, hardly a w elcom e p ort given D ennett's aim s, a
fact Dennett now appreciates in m ore detail. Accordingly, his current
rationale draw s a m uch m ore narrow ly focused analogy b etw een in
tentional states and geom etrical abstracta such as the centers of grav
ity, axes of rotation, equators, etc., that are postulated to such good
effect in m echanics. A s D ennett sees it, these latter are not real in the
sam e sense that concreta like bricks and trees are real (you can 't trip
over them , for exam ple), but they can reasonably be said to be real
even so. Intentional states are real in this sam e sen se, claims D ennett.
The reality of equators, cen ters, and rotational axes I am h ap p y to
g ran t. They are all places or loci of som e sort that are decisively sp e
cifiable by reference to the shape or behavior of the relevant concrete
object. But the alleged similarity of these item s to beliefs, desires, and
oth er intentional states escapes m e entirely. In w hat respects are they
similar, and w hy should they be grouped togeth er in advance of the
issue here at stake? That is, in advance of any hopes of finding an
exculpatory status for intentional states?
D enn ett is quick to point out that folk psychology has som e n on
trivial predictive pow er, especially in its central dom ain of norm al
h um an behavior, despite the lack of any neural concreta answ ering to
the propositional attitudes. H e em phasizes, quite correctly, that it is
an objective fact about hum ans that a significant am ount of their b e
havior is accurately predictable in intentional term s.
But I think he overvalues this fact wildly. W e m ust not forget that
all sorts of false theories, with w holly chim erical ontologies, can boast
v ery im pressive predictive pow er in various proprietary dom ains. But
this buys their ontology no special status. It is an objective fact that
m u ch of the behavior of m etals and ores is predictable in term s of the
alchem ical essen ces, that m ost of the behavior of the visible heavens
is predictable in term s of nested crystal sp heres, that m uch of the
behavior of m oving bodies is predictable in term s of im petus, and so
forth, A nd yet there are no alchem ical essen ses, nor any crystal
sp heres, nor any im petuses. W e could, of course, set about insisting
that these three "th in g s” are real and genuine after all, though m ere
abstracta to be sure. But none of us is tem p ted to salvage their reality
by such a tortu red and tran sp aren t ploy. W h y should w e be tem pted
in the case of the propositional attitudes?
This disagreem ent betw een us on the status of folk p sychology
dates from several letters now a full decad e old. H ow ever, one point
on w hich w e then agreed w as that neither of us could clearly im agine
a system atic alternative to folk psychology. A t the time I ascribed this
Folk Psychology and the Explanation of Behavior 127
This style of argum ent has a considerable history, but one of its
clearest and m ost accessible con tem porary exponents is Charles
Taylor (1970, 1987). In w hat follows, I hope to reverse the antinatural
ist, antireductionist conclusion of the argum ent. Y et this undertaking
h as a certain poignancy for m e, since I accept both of the prem ises
from w hich the argum ent proceeds. Indeed, in earlier w ritings
(C hurchland 1979, 1981a, 1985b) I have vigorously defended both the
radical plasticity of the h um an m ind, and the dram atic extent to
w hich the character of hu m an consciousness is determ ined by the
cultural surround, specifically, by the ideological, linguistic, and
practical surround. In this ch apter I shall try to show how a naturalist
can em brace the insights of the tw o prem ises cited, w ithout co m
prom ising the possibility of a strongly naturalistic, and even red u c
tionistic, science of hum an consciousness.
M y strategy here is not m erely critical, b ut positive. I propose to
illustrate the w eakness of the argum ents cited above by outlining a
recently developed theoretical and experim ental approach to the p h e
n om en a of h u m an cognition that is at once (a) naturalistic, (b) red u c
tionistic, and (c) capable of explaining both the radical plasticity of
h u m an consciousness, and its intricate depen den ce on the extended
cultural surround.
This new app roach , called 'connectionism ', resides at the interface
of com putational neuroscience, cognitive psychology, and artificial
intelligence (AI). It represents a radical break w ith the approach es
that have dom inated all three of these disciplines for the last three
decad es. It even m akes firm contact with som e of the m ajor them es of
the continental tradition in philosophy, such as the inarticulate or
nonpropositional character of the bulk of h u m an know ledge, and the
prim acy of being an endlessly active agent in a w orld of practical ex
igencies. But it does not support the antinaturalist them es to be found
in recent continental and analytical philosophies, as w e are about to
see. Let m e begin w ith a brief introduction to the central con cep ts of
this novel and very fertile approach. After w e have seen the unusual
p roperties displayed by connectionist system s, w e shall return to the
philosophical issues with w hich w e began.
[I h ere ask the read er to turn to ch apter 9, sections 3 and 4 (pp.
1 5 9 -1 7 1 ), for the m aterial on learning in brainlike netw orks. A fter
reading that m aterial, the p resen t essay m ay be picked up below .]
E nou gh exam ples. You have seen som ething of w hat netw orks of
this kind can do, and of how they do it. In both respects they contrast
sharply w ith the kinds of representational and processing strategies
th at analytical philosophers, cognitive psychologists, and AI w orkers
have traditionally ascribed to us (nam ely, sentencelike rep resen ta
The Plasticity of Human Consciousness 131
tions manipulated by formal rules.) You can see also w hy this theoret
ical and experim ental approach has captured the interest of those
w ho seek to u n derstand how the m icroarchitecture of the biological
brain p rodu ces the p h en om en a of h um an consciousness. Let us now
retu rn to philosophical issues.
This final objection begins w ith the claim that the features of the
w orld that are m axim ally im portant for explaining the behavior of
h u m an s are n ot the simple ones that can be defined in the vocabulary
of a naturalistic physics, but rath er the m u ch m ore subtle and co m
plex features that constitute our social culture. O ne does indeed dis
crim inate and respon d to light, to w arm th , to sound, and so forth.
But one also discrim inates and respond s to the phonem es peculiar to
o n e's language, to the m eanings of its w ord s, to the m oral of a story,
to the significance of gestures and facial expressions, to challenges
and obligations and social opportunities, and to all of the intricacies
that m ake up a functioning culture.
All of this is no doubt true. But again, these things are w holly con
sistent w ith a reductionist p rogram for understanding the natu re of
h u m an cognition. W hat the reductionist m u st do is explain how a
The Plasticity of Human Consciousness 133
3 Conclusion
A t several points in the reading of van F raassen 's book (1980), I feared
I w ould no longer be a realist by the time I com pleted it. Fortunately,
sheer doxastic inertia has allowed m y convictions to survive its search
ing critique, at least tem porarily, and as w e address you today, van
F raassen and I still hold different view s. I am a scientific realist, of
u n orth od o x persuasion, and van Fraassen is a constructive em piri
cist, w h ose persuasions currently define the doctrine. I assert that
global excellence of theory is the ultim ate m easure of tru th and ontol
ogy at all levels of cognition, even at the observational level. Van
Fraassen asserts that descriptive excellence at the observational level
is the only genuine m easure of any th eory's truth, and that one's
accep tan ce of a theory should create no ontological com m itm ents
w h atev er beyond the observational level.
A gain st his first claim I shall m aintain that observational excellence
or "em p irical ad eq u acy " is only one epistem ic virtue am on g others,
of equal or com parable im portance. A nd against his secon d claim I
shall m aintain that the ontological com m itm ents of any theory are
w holly blind to the idiosyncratic distinction betw een w h at is and
w hat is n ot hum anly observable, and so should be our o w n ontologi
cal com m itm ents. Criticism will be directed primarily at van Fraassen's
selective skepticism in favor of observable ontologies o ver unobserv
able ontologies; and against his view that the superem pirical th eoret
ical virtu es (simplicity, coherence, explanatory p ow er) are m erely
pragm atic virtues, irrelevant to the estim ate of a th eo ry 's truth. My
aim s are n ot m erely critical, h ow ever. Scientific realism does need
rew orking, and there are good reasons for m oving it in the direction
of v an F raassen 's constructive em piricism , as will be d iscussed in the
This paper first appeared in the Pacific Philosophical Quarterly 63 (1982), no. 3.
As the prose suggests, it was first presented at a sym posium on scientific
realism (Halifax, 1981) that focused on my 1979 book and Bas van Fraassen's
1980 book.
140 The Structure of Science
closing section of this paper. But those reason s do not support the
skeptical theses at issue.
com patible theories rem ains relative to w hich background theories are
added to the evaluative context, especially background theories that
in som e w ay revise our conception of w hat hum ans can observe. I
intend to sidestep this issue for now , h ow ever, since the m atter is
com plex and there is a m uch simpler objection to be voiced.
Let m e approach m y objection by first pointing out th at the em pir
ical adequacy of any theory is itself som ething that is radically u n d er
determ ined by any evidence conceivably available to us. Recall that,
for a theory to be empirically adequate, w hat it says about observable
things m u st be true— all observable things, in the p ast, in the indefi
nite fu ture, and in the m ost distant corners of the cosm os. But since
any actual data possessed by us m ust be finite in its scope, it is plain
that w e here suffer an underdeterm ination problem no less serious
than th at claimed above. This is H u m e's problem , and the lesson is
that even observation-level theories m u st suffer radical u n d erd eter
m ination by the evidence. A ccordingly, theories about observables
and theories about unobservables appear on a p ar again, so far as
skepticism is con cerned .
V an Fraassen thinks there is an im portant difference betw een the
tw o cases, and on e's first im pulse is to agree w ith him . W e are all
willing to concede the existence of H u m e's problem , the problem of
justifying the inference to unobserved entities. But the inference to
entities that are dow nright unobservable app ears as a different and
addition al problem .
The ap p earance is an illusion, as the following considerations will
show . C onsider som e of the different reasons w hy entities or p ro
cesses m ay go unobserved by us. First, they m ay go u n observed be
cause, relative to our natural sensory app aratus, they fail to enjoy an
appropriate spatial or tem poral position. They m ay exist in the U pper
Jurassic Period, for exam ple, or they m ay reside in the A n d rom ed a
G alaxy. Second, they m ay go unobserved becau se, relative to our
natural sensory app aratu s, they fail to enjoy the appropriate spatial or
tem poral dim ensions. They m ay be too sm all, or too brief, or too large,
or too p rotracted . Third, they m ay fail to enjoy the appropriate energy,
being too feeble, or too pow erful, to perm it useful discrim ination.
Fou rth and fifth, they m ay fail to have an appropriate w avelength, or
an ap p rop riate m ass. Sixth, they m ay fail to “ feel" the relevan t fun
dam ental fo rces our sensory apparatus exploits, as w ith our inability
to observe the background neutrino flux, despite the fact th at its en er
gy density exceeds that of light itself.
This list could be lengthened, but it is long enough to su ggest that
being spatially or tem porally distant from our sen sory ap p aratu s is
4 only one am ong m any w ays in w hich an entity or p rocess can fall
144 The Structure of Science
Van Fraassen 's position w ould exclude (3) from our rational ontology.
This h as at least som e initial plausibility. But his position w ould not
be at all plausible if it w ere com m itted to excluding both (3) and (2)
from our rational ontology. N o p arty to the p resen t discussion is will
ing to restrict rational ontology to (1) alon e. Van F raassen's position
thu s requires a principled distinction b etw een (2) and (3), a distinction
adequate to the radical difference in epistem ic attitude he w ould have
us ad op t tow ard them . The burden of m y argu m en t is that the dis
tinction betw een (2) and (3), once it is u n earth ed , is only very feebly
principled, and is w holly inadequate to b ear the great w eight that van
Fraassen puts on it.
The Ontological Status of Observables 145
tradition says, or m erely pragm atic virtues, as van Fraassen u rges. His
view p rom otes empirical adequacy, or evidence of empirical adequa
cy, as the only genuine m easure of a th eory's truth, the other virtues
(insofar as they are distinct from these) being cast as purely pragm atic
virtues, to be valued only for the h u m an needs they satisfy. D espite
certain com pelling features of the account of explanation that van
Fraassen provides, I rem ain inclined tow ard the traditional view .
M y reason is simplicity itself. Since there is no w ay of conceiving or
representing "th e empirical facts" that is com pletely independent of
speculative assum ptions, and since w e shall occasionally confront
theoretical alternatives on a scale so com prehensive that w e m u st also
choose betw een com peting m od es of conceiving w hat the empirical
facts before us are, it is clear that the epistem ic choice betw een
these global alternatives cannot be m ade by com paring the extent to
w hich they are adequate to som e com m on tou ch ston e, "th e em
pirical fa c ts ." In such a case, the choice m u st be m ade on the com
parative global virtues of the tw o global alternatives, 7 ,-p lu s-
the-observational-evidence-therein-construed, versu s T2-plus-the-
observational-evidence-therein-(differently)-construed. That is, it
m u st be m ad e on superem pirical grou nd s such as relative coh eren ce,
simplicity, and explanatory unity.
V an Fraassen has said that to "s a v e the ap p earan ces" is to exhibit
them as a fragm ent of a larger unity. W ith this I w holly agree. But I
am here pointing out that it is a decision b etw een com peting larger
unities that determ ines w hat w e count as "th e true ap p earan ces" in
the first place. There is no independent w ay to settle that question.
A n d if such global decisions can only be m ad e on w hat van Fraassen
calls "p ra g m a tic" grounds, then it w ould seem to follow that any d e
cision concerning w h at the observable w orld contains m ust be essen
tially "p ra g m a tic" also. Inflationary m etaphysics and "p ra g m a tic"
decisions begin, it seem s, as soon as w e open our eyes.
Global issues such as these are rem iniscent of C arn ap 's "e x te rn a l"
questions, and I think it likely that van Fraassen , like C arnap, does
n ot regard them as decidable in any but a second-rate sense, since
th ey can only be decided by second-rate (i.e ., by pragm atic) con sid
erations. If so, how ever, it is difficult to see how van Fraassen can
justify a selectively realist attitude tow ard "o b se rv a b le s," since, as w e
h ave seen , pragm atic considerations m ust attend their selection also.
(These issues receive extended treatm ent in C hurchland 1979, sec
tions 2, 3, 7, and 10.) W hat all of this illustrates, I think, is the poverty
of van F raassen 's crucial distinction betw een factors that are " e m
pirical, and therefore tru th -relev an t," and factors that are "su p e r-
em pirical, and therefore not tru th -relev an t."
The Ontological Status of Observables 147
elem ent in a clutch of descriptive and norm ative theories (folk psych ol
ogy, folk epistem ology, folk sem antics, classical logic), and w e can
exp ect conceptual p rogress here as appropriately as anyw h ere else.
The notion of truth is susp ect on purely m etaphysical grou n d s any
w ay. It suggests straightaw ay the notion of The C om plete and Final
True Th eory: at a m inim um , the infinite set of all true sen ten ces. Such
a theory w ould be, by epistem ic criteria, the best theory possible. But
nothing w h atever guarantees the existence of such a unique theory.
Ju st as there is no largest positive integer, it m ay be that there is no
best theory. It m ay be that for any theory w hatsoever, there is always
an even b etter theory, and so ad infinitum. If w e w ere thu s unable to
speak of the set of all true sentences, w hat sense could w e m ake of
truth sen ten ce-b y-sen ten ce?
These considerations do invite a "co n stru ctiv e " conception of
cognitive activity, one in w hich the notion of truth plays at best a
highly derivative role. The form ulation of such a conception, ade
quate to all of our epistem ic criteria, is the outstanding task of epis
tem ology. I do not think w e shall find that conception in van
F raassen 's m odel-theoretic version of "positivistic in stru m en talism ,"
nor do I think w e shall find it quickly. But the empirical brain begs
unraveling, and w e have plenty of time.
Finally, there is a question pu t to m e by Stephen Stich. If ultim ately
m y view is even m ore skeptical than van F raassen 's concerning the
relevance or applicability of the notion of truth, w h y call it scientific
realism at all? For at least tw o reasons. The term 'realism ' still m arks
the principal contrast w ith its traditional adversary, positivistic in
strum entalism . W h atever the integrity of the notion of tru th , theories
about unobservables have just as much a claim to truth, epistem ologi-
cally an d m etaphysically, as theories about observables. S econd, I re
m ain com m itted to the idea that there exists a w orld, ind ep en d en t of
our cognition, with w hich w e interact, and of w hich w e con stru ct
rep resentation s: for varying p urp oses, w ith varying penetration , and
w ith varyin g success. Lastly, our best and m ost penetrating grasp of
the real is still held to reside in the rep resentations provided by our
b est theories. Global excellence of theory rem ains the fundam ental
m easure of rational ontology. A nd that h as always b een the central
claim of scientific realism .
Chapter 9
On the Nature of Theories:
A Neurocomputational Perspective
them as fatal. I profess myself am ong the m inority that does so regard
them . In urging the poverty of "senten tial epistem ologies" for over a
decad e now (1975b, 1979, chapters 1 and 5), I h ave been m otivated
prim arily by the pattern of the failures displayed by that approach.
Those failures suggest to m e that w hat is defective in the classical
approach is its fundam ental assum ption that languagelike structures
of som e kind constitute the basic or m ost im portant form of rep re
sentation in cognitive creatures, and the correlative assum ption that
cognition consists in the m anipulation of those representations by
m ean s of structure-sensitive rules.
To be sure, not everyone saw the sam e p attern of failure, nor w ere
they prepared to draw such a strong conclusion even if they did. For
any research program has difficulties, and so long as w e lack a com
parably com pelling alternative conception of rep resentation and com
putation, it m ay be best to stick w ith the familiar research p rogram of
sen tences and rules for their m anipulation.
H ow ever, it is no longer true that w e lack a com parably com pelling
alternative approach. W ithin the last five years, there have been som e
striking theoretical developm ents and experim ental results w ithin
cognitive neurobiology and connectionist AI. These have provided us
w ith a pow erful and fertile fram ew ork w ith w hich to address p rob
lem s of cognition, a fram ew ork that ow es nothing to the sentential
paradigm of the classical view . M y m ain p u rp ose in this essay is to
m ake the rudim ents of that fram ew ork available to a w ider audience,
and to explore its far-reaching consequences for traditional issues in
the philosophy of science. Before turning to this task, let m e prep are
the stage by briefly sum m arizing the principal failures of the classical
view and the m ost prom inent responses to them .
accou nt for the rationality of m ajor shifts from one probability space
to an other, w hich is w hat the m ost interesting and im p ortan t cases of
learning am ou nt to. The rationality of large-scale conceptual change,
accordingly, seem ed beyond the reach of such ap p roach es. Fu rth er
m ore, simplicity em erged as a m ajor determ inant of theoretical
credibility on m ost accounts, but none of them could provide an
adequate definition of simplicity in syntactic term s, or give a co n
vincing explanation of w hy it w as relevant to truth or credibility in
any case. O ne could begin to question w heth er the basic factors
relevant to learning w ere to be found at the linguistic level at all.
Beyond these an noyan ces, the initial resources ascribed to a learn
ing subject by the sentential approach plainly p resu p p osed the su c
cessful com pletion of a good deal of sophisticated learning on the part
of that subject already. For exam ple, reliable observation judgm ents
do n ot just app ear out of now here. Living subjects h ave to learn to
m ake the com plex perceptual discrim inations that m ake perceptual
jud gm en ts possible. A nd they also have to learn the linguistic or prop-
ositional system within w hich their beliefs are to be constituted.
Plainly, both cases of learning will have to involve som e p rocedure
quite distinct from that of the classical account. For that accou n t p re
su p p oses an teced ent possession of both a determ inate propositional
system and a capacity for determ inate perceptual jud gm en t, w hich is
precisely w h at, prior to extensive learning, the h u m an infant lacks.
A ccordingly, the classical story cannot possibly accou n t for all cases
of learning. There m u st exist a type of learning that is prior to, and
m ore basic than, the p rocess of sentence m anipulation at issue.
Thus are w e led rath er swiftly to the idea that there is a level of
rep resentation beneath the level of the sentential or propositional atti
tud es, and to the correlative idea that there is a learning dynam ic that
operates prim arily on sub linguistic factors. This idea is reinforced by
reflection on the problem of cognition and learning in n on h u m an
anim als, none of w hich appear to have the benefit of langu age, either
the external speech or the internal stru ctures, but all of w hich en
gage in sophisticated cognition. P erhaps their cognition p roceeds
entirely w ithout benefit of any system for processing sentencelike
rep resentation s.
E ven in the h u m an case, the depiction of on e's know ledge as an
im m ense set of individually stored "se n te n c e s" raises a severe prob
lem concern in g the relevant retrieval or application of those internal
rep resentation s. H ow is it one is able to retrieve, from the millions of
sen ten ces stored, exactly the handful that is relevant to o n e 's current
predictive or explanatory problem , and how is it one is generally able
4 to do this in a few tenths of a second? This is know n as the "fram e
156 The Structure of Science
p roblem " in AI, and it arises b ecause, from the point of view of fast
and relevant retrieval, a long list of sen ten ces is an appallingly ineffi
cient w ay to store inform ation. A nd the m ore inform ation a creature
h as, the w orse its application problem becom es.
A further problem w ith the classical view of learning is that it finds
no essential connection w hatever b etw een the learning of fa cts and
the learning of sk ills. This is a problem in itself, since one m ight have
h op ed for a unified accou nt of learning, b ut it is doubly a problem
w hen one realizes that so m u ch of the business of u n d erstanding a
theory and being a scientist is a m atter of the skills one has acquired.
M em orizing a set of sentences is n ot rem otely sufficient: one m ust
learn to recognize the often quite various instances of the term s they
contain; one m ust learn to m anipulate the peculiar form alism in w hich
th ey m ay be em bedded; one m u st learn to apply the form alism to
novel situations; one m ust learn to control the instrum ents that typi
cally produ ce or m onitor the p h en om en a at issue. A s T. S. K uhn first
m ad e clear (1962), these dim ensions of the scientific trade are only
artificially separable from on e's understan din g of its current theories.
It begins to ap p ear that even if w e do harbor internal sen ten ces, they
capture only a small part of h um an know ledge.
These failures of the classical view over the full range of learning,
both in hu m an s and in nonh u m an anim als, are the m ore suspicious
given the total disconnection of the classical view from any theory
concerning the structure of the biological brain, and the m an n er in
w hich it m ight im plem ent the kind of representations and com p u ta
tions p roposed . M aking acceptable con tact w ith neurophysiological
theory is a long-term constraint on any epistem ology: a schem e of
representation and com putation that can n ot be im plem ented in the
m achinery of the h um an brain cannot be an adequate accou n t of
h um an cognitive activities.
The situation on this score used to be m u ch better than it is n ow : it
w as clear that the classical accou n t of rep resentation and learning
could easily be realized in typical digital com p u ters, and it w as
thou gh t that the h u m an brain w ould turn out to be relevantly like a
digital com p u ter. But quite aside from the fact that com p u ter im ple
m entations of sentential learning chronically produced disappointing
results, it has becom e increasingly clear that the brain is organized
along com putational lines radically different from those em ployed in
conventional digital com puters. The brain, as w e shall see below , is a
m assively parallel p rocessor, and it perform s com putational tasks of
the classical kind at issue only v ery slowly and com paratively badly.
Loosely speaking, it does not ap p ear to be designed to perform the
tasks the classical view assigns to it.
On the Nature of Theories 157
sem antic ap proach (see especially Giere 1988, w hose version does
m ake som e contact with cu rrent cognitive psychology). But it is clear
that the sem antic approach is a response to only a small subset of the
extant difficulties.
A m ore celebrated response is em bodied in K uhn's Structure o f Sci
entific R evolutions (1962). Kuhn cen ters our attention not on sets of
sen ten ces, n or on sets of m odels, but on w h at he calls paradigm s or
exem p lars, w hich are specific applications of our conceptual, m ath e
m atical, and instrum ental resou rces. M astering a theory, on this
view , is m ore a m atter of being able to perform in various w ays, of
being able to solve a certain class of problem s, of being able to recog
nize diverse situations as relevantly similar to that of the original or
paradigm atic application. K uhn's view brings to the fore the histori
cal, sociological, and psychological factors that structure our theoreti
cal cognition. Of central importance is the m anner in which one comes
to perceive the w orld as one internalizes a theory. The perceptual
w orld is redivided into new categories, and while the theory m ay
be able to provide n ecessary and sufficient conditions for being an
instance of any of its categories, the p erceptual recognition of any
instance of a category does not generally p roceed by reference to
those conditions, w hich often transcend perceptual experience.
R ather, perceptual recognition proceeds by som e inarticulable p ro
cess that registers sim ilarity to one or m ore perceptual prototypes of the
category at issue. The recognition of new applications of the ap p ar
atus of the entire theory displays a similar dynam ic. In all, a su ccess
ful theory provides a prototypical beachh ead that one attem pts to ex
p and by analogical extensions to new dom ains.
R eaction to this view has been deeply divided. Som e applaud
K u h n 's m ove tow ard naturalism , tow ard a perform ance conception
of know ledge, and aw ay from the notion of truth as the guiding co m
p ass of cognitive activity (M unevar 1981; Stich 1989). O thers deplore
his neglect of norm ative issues, his instrum entalism and relativism ,
and his alleged exaggeration of certain lessons from perceptual and
developm ental p sychology (Fod or 1984). I shall address these issues
later in this chapter.
A third and less visible reaction to the classical difficulties has
sim ply rejected the sentential or propositional attitudes as the m ost
im portan t form of representation used by cognitive creatu res, and
h as insisted on the necessity of empirical and theoretical research
into brain function in order to answ er the question of w hat are the
m ost im portant form s of representation and com putation w ithin
cognitive creatu res. Early statem ents can be found in P. M. C hurch-
land 1975b and H ooker 1975; extend ed argu m en ts app ear in P. M.
On the Nature of Theories 159
The functional atom s of the brain are cells called neu ron s (figure 9.1 ).
These have a natural or default level of activity, w hich can , h ow ever,
be m odulated up or dow n by external influences. From each neuron
Figure 9.1
x A schematic neuron
160 The Structure of Science
Figure 9.2
A neuronlike processing unit
there extends a long, thin output fiber called an axon, w hich typically
b ran ches at the far end so as to m ake a large num ber of synaptic con
nections w ith either the central cell body or the bushy dendrites of o ther
n eu ron s. Each n euron thus receives inputs from a great m an y other
n eu ron s, w hich inputs tend to excite (or to inhibit, according to the
type of synaptic connection) its norm al or default level of activation.
The level of activation induced is a function of the num ber of con n ec
tions, of their size or w eight, of their polarity (stim ulatory or inhibi
tory), and of the strength of the incom ing signals. F urth erm ore, each
n eu ron is constantly em itting an output signal along its ow n axon, a
signal w hose stren gth is a direct function of the overall level of activa
tion in the originating cell body. That signal is a train of pulses or
spikes, as they are called, w hich are p ropagated swiftly along the
axon. A typical cell can emit spikes along its axon at anything b e
tw een 0 and p erh ap s 200 h ertz. N euron s, if you like, are hum m ing to
one an other, in basso notes of varying frequency.
The netw orks to be explored attem p t to sim ulate natural neurons
w ith artifical units of the kind depicted in figure 9 .2 . These units
adm it of various levels of activation, w hich I shall assum e to vary
b etw een 0 and 1. Each unit receives input signals from other units via
"s y n a p tic " connections of various w eights and polarities. These are
rep resented in the diagram as small end plates of various sizes. For
simplicity I dispense w ith dendritic trees: the "a x o n a l" end b ranches
from oth er units all m ake connections directly to the "cell b o d y " of
the receiving unit. The total m odulating effect E im pacting on that
unit is just the sum of the contributions m ade by each of the co n n ec
tions. The contribution of a single connection is just the p rodu ct of its
w eight Wj tim es the strength s, of the signal arriving at that connec-
On the Nature of Theories 161
Figure 9.3
The sigmoid axonal output function
Figure 9.4
A simple network
or m agnitudes. For exam ple, a given stim ulus m ight produce the
v ector < .5, .3, .9, .2>.
These input activation levels are then p ropagated u p w ard , via the
ou tp u t signal in each u nit's axon, to the m iddle layer of the netw ork,
to w h at are called the "h id d en u n its." A s can be seen in figure 9 .4 ,
any unit in the input layer m akes a synaptic connection of som e
w eight or other w ith every unit at this interm ediate layer. Each hid
den unit is thus the target of several inputs, one for each cell at the
input layer. The resulting activation level of a given hidden unit is
essentially just the sum of all of the influences reaching it from the
cells in the low er layer.
The result of this upw ard propagation of the input vector is a set of
activation levels across the three units in the hidden layer, called the
"h id d en unit activation v e cto r." The values of that three-elem ent
vector are strictly determ ined by
Looking n ow at the w hole netw ork, w e can see that it is just a device
for transform ing any given input-level activation vector into a unique
ly corresp on d in g output-level activation vector. A n d w h at d eter
m ines the character of the global transform ation effected is the
peculiar set of values possessed by the m an y connection w eights.
This m u ch is easy to grasp. W hat is n ot so easy to grasp , prior to
exploring exam ples, is just how very pow erful and useful those tran s
form ations can be. So let us explore som e real exam ples.
Figure 9.5
Perceptual recognition with a large network
m ine its transform ational activity are initially set at ran d om values. A t
the beginning of this experim ent, then, the outp u t v ectors are sure to
disappoint u s. But w e proceed to teach the netw ork by m ean s of the
following procedure.
W e p rocu re a large set of recorded sam ples of various (genuine)
mine ech o es, from m ines of various sizes and orientations, and a
com parable set of genuine rock ech oes, keeping careful track of w hich
is w hich. W e then feed these echoes into the netw ork, one b y one,
and observe the output vector p rodu ced in each case. W h at interests
us in each case is the am ou nt by w hich the actual outp u t v ecto r differs
from w h at w ould have been the correct vector, given the identity of
the specific echo that produced it. The details of that erro r, for each
elem ent of the output vector, are then fed into a special rule that
co m p u tes a set of small changes in the values of the various synaptic
166 The Structure of Science
Figure 9.6
Learning: gradient descent in weight/error space. Axes are shown for only 2 of 105
synaptic weights
On the Nature of Theories 167
abstract space of m any dim ensions, one for each w eight in the n et
w ork (105 in this case), plus one dim ension for rep resentin g the o ver
all erro r of the outpu t vector on any given trial. A n y point in that
space rep resen ts a unique configuration of w eights, plus the perform
ance error that that configuration produ ces. W h at the learning rule
does is steadily nudge that configuration aw ay from erron eou s posi
tions and tow ard positions that are less erroneou s. The system inches
its w ay d ow n an "e rro r g radien t" tow ard a global error m inim um .
O nce there, it respon d s reliably to the relevant kinds of echoes. It
even respo n d s well to echoes that are similar to m ine ech oes, by
giving outp ut vectors that are closer to <1, 0> than to <0, 1>.
T here w as no guarantee the netw ork w ould succeed in learning to
discrim inate the tw o kinds of echoes, because there w as no guarantee
that rock ech oes and mine echoes w ould differ in any system atic or
detectable w ay. But it turns out that m ine echoes do indeed have
som e com p lex of relational or structural features that distinguishes
them from rock ech oes, and under the pressu re of rep eated error
corrections, the netw ork m an ages to lock onto, or b ecom e "tu n e d "
to, th at subtle but distinctive w eave of features.
W e can test w heth er it has truly succeeded in this by n ow feeding
the n etw ork som e mine and rock echoes n ot included in the training
set, ech o es it has never encountered before. In fact, the n etw ork does
alm ost as well classifying the new echoes as it does w ith the sam ples
in its training set. The "k n o w led ge" it h as acquired generalizes quite
successfully to new cases. (This exam ple is a highly simplified
accou nt of som e striking results from G orm an and Sejnowski 1988a,
1988b.)
All of this is m odestly am azing, because the problem is quite a diffi
cult o ne, at least as difficult as learning to discrim inate the p honem e
/а /. H u m an sonar operators, during a long tour of subm arine duty,
eventually learn to distinguish the tw o kinds of ech oes w ith som e
u ncertain but nontrivial regularity. But they never perform at the
level of the artificial netw ork. Spurred on by this su ccess, w ork is
currently u n d erw ay to train up a netw ork to distinguish the various
p h on em es characteristic of English speech (Zipser and Elm an 1987).
The idea is to produ ce a speech-recognition system th at will not be
troubled by the acoustic idiosyncracies of diverse speakers, as
existing speech-recognition system s are.
The su ccess of the m ine/rock netw ork is further intriguing because
the "k n o w led g e" the netw ork has acquired, concerning the distinc
tive ch aracter of m ine echoes, consists of nothing m ore than a careful
ly orch estrated set of connection w eights. A n d it is finally intriguing
4 b ecau se there exists a learning algorithm — the rule for adjusting the
168 The Structure of Science
Figure 9.7
Learned partition on hidden-unit activation-vector space. Axes are shown for only
three of seven hidden-unit activation levels
crim inatory system that perform s som ething like a living creatu re,
both in its speed and in its overall character.
I have explained this system in som e detail, so that the read er will
have a clear idea of how things w ork in at least one case. But the
netw ork described is only one instance of a general technique that
w orks well in a large variety of cases. N etw orks can be co n stru cted
w ith a larger num ber of units at the outp u t layer, so as to be able to
exp ress n ot just tw o, but a large num ber of distinct discrim inations.
O ne netw ork, aptly called NETtalk by its authors (Rosenberg and
Sejnowski 1987), takes vector codings for seven-letter segm en ts of
printed w ord s as inputs and gives vector codings for p h on em es as
ou tp u ts. These outp u t vectors can be fed directly into a sound syn
thesizer as they occur, to produce audible sounds. W hat this network
learns to do is to transform printed w ord s into audible speech.
T hough it involves no understanding of the w ords that it " r e a d s ," the
n etw ork's feat is still very im pressive, b ecause it w as given no rules
w h atev er concerning the phonetic significance of stan d ard English
-sp ellin g. It began its training period by producing a stream of unintel
170 The Structure of Science
specific orientations. This finding echoes the sem inal w ork of Hubei
and W iesel (1962), in w hich cells in the visual cortex w ere discovered
to h ave receptive fields of this sam e ch aracter. Results of this kind are
very im p ortan t, for if w e are to take these artificial netw orks as m od
els for h ow the brain w orks, then they m u st display realistic behavior
n ot ju st at the m acro level; they m ust also display realistic behavior at
the m icro level.
E n ou g h exam ples. You have seen som ething of w hat netw orks of
this kind can do and of how they do it. In both respects th ey contrast
sharply w ith the kinds of representational and processing strategies
that philosophers of science, inductive logicians, cognitive p sych o
logists, and AI w orkers have traditionally ascribed to us (nam ely, sen
tencelike rep resentations m anipulated by formal rules). Y ou can see
also w h y this theoretical and experim ental approach has cap tu red the
interest of those w ho seek to understand how the m icroarchitecture
of the biological brain produ ces the p h en om en a displayed in h um an
and anim al cognition. Let us now explore the functional properties of
these netw orks in m ore detail, and see how they bear on som e of the
traditional issues in epistem ology and the philosophy of science.
The netw orks described above are d escended from a device called the
P ercep tron (Rosenblatt 1959), w hich w as essentially just a tw o-layer
netw ork as opposed to a three-layer netw ork. D evices of this con
figuration could and did learn to discrim inate a considerable variety
of input p attern s. U nfortunately, having the input layer connected
directly to the outp u t layer im poses v ery severe limitations on the
range of possible transform ations a netw ork can perform (Minsky
and P ap ert 1969), and interest in Perceptron-like devices w as soon
eclipsed b y the m uch faster-m oving developm ents in standard
"p ro g ram -w ritin g " AI, w hich exploited the high-speed general-
p u rp ose digital m achines that w ere then starting to b ecom e widely
available. Throu ghout the seventies, research in artificial "n eu ral
n e ts" w as an u nd ergrou n d p rogram by com parison.
It has em erged from the shadow s for a num ber of reason s. O ne
im p ortan t factor is just the troubled doldrum s into w hich m ainstream
or program -w riting AI h as fallen. In m an y respects, th ese doldrum s
parallel the infertility of the classical approach to theories and learn
ing w ithin the philosophy of science. This is not surprising, since
m ain stream AI w as proceeding on m any of the sam e basic assu m p
tions about cognition, and m any of its attem p ts w ere just m achine
-im p lem en tation s of learning algorithm s proposed earlier by philo
172 The Structure of Science
sop hers of science and inductive logicians (G lym our 1988). The
failures of m ainstream A I— unrealistic learning, p oor perform ance in
com plex perceptual and m otor tasks, w eak handling of analogies,
and snail-like cognitive perform ance despite the use of very large and
fast m achin es— teach us even m ore dram atically than do the failures
of m ainstream philosophy that w e need to rethink the style of rep re
sentation and com putation that w e have b een ascribing to cognitive
creatures.
O th er reasons for the resurgen ce of interest in netw orks are m ore
positive. The introduction of additional layers of intervening or "h id
d en ” units produced a dram atic increase in the range of possible
transform ations that the netw ork could effect. A s Sejnowski et al.
describe it,
already 10105! This space is far too large to explore efficiently w ithout
som ething like the generalized delta rule and the back propagation of
error to do it for us. But w ith the delta rule, adm inistered by an auxili
ary com pu ter, research ers have show n that netw orks of the simple
kind described are capable of learning som e quite extraordinary skills,
and of displaying som e highly intriguing properties. Let m e now re
turn to an exploration of these.
A n im portant exploratory technique in cognitive and behavioral
neuroscience is to record , w ith an im planted m icroelectrode, the elec
trical activity of a single neuron during cognition or behavior in the
intact anim al. This is relatively easy to do, and it does give us tantaliz
ing bits of inform ation about the cognitive significance of neural activ
ity (recall the results of Hubei and W iesel m entioned earlier). Single
cell recordings give us only isolated bits of inform ation, how ever, and
w hat w e w ould really like to m onitor are the patterns of sim ultaneous
neural activation across large num bers of cells in the sam e subsystem .
U nfortunately, effective techniques for sim ultaneously recording
from large num bers of adjacent cells are still in their infancy. The task
is extrem ely difficult.
By con trast, this task is extrem ely easy w ith the artificial netw orks
w e have been describing. If the netw ork is real hard w are, its units are
far m ore accessible than the fragile and m icroscopic units of a living
brain. A n d if the netw ork is m erely being sim ulated within a standard
com p uter (as is usually the case), one can w rite the p rogram so that
the activation levels of any unit, or set of units, can be read out on
com m an d. A ccordingly, once a netw ork has been successfully trained
up on som e skill or other, one can then exam ine the collective b e
havior of its units during the exercise of that skill.
W e have already seen the results of one such analysis in the ro ck /
m ine netw ork. O nce the w eights have reached their optim um con
figuration, the activation vectors (i.e., the p attern s of activation) at
the hidden layer fall into tw o disjoint classes: the vector space is parti
tioned in tw o, as depicted schem atically in figure 9 .7 . But a m ere bin
ary discrim ination is an atypically simple case. The read er N ETtalk,
for exam p le, partitions its hidden-unit v ector space into fully 79 sub
spaces. The reason is simple. For each of the 26 letters in the alphabet,
there is at least one phonem e assigned to it, and for m any letters
there are several ph on em es that m ight be signified, depending on the
lexical context. A s it hap p ens, there are 79 distinct letter-to-phonem e
associations to be learned if one is to m aster the pronunciation of
English spelling, and in the successfully trained netw ork a distinct
hidden-unit activation vector o ccurs w hen each of these 79 possible
transform ations is effected.
On the Nature of Theories 175
In the case of the rock/m ine netw ork, w e noted a similarity m etric
w ithin each of its tw o hidden-unit subspaces. In the case of NETtalk,
w e also find a similarity m etric, this time across the 79 functional
hidden-unit vectors (by 'functional vector' I m ean a v ector that cor
respon d s to one of the 79 desired letter-to-phonem e transform ations
in the trained netw ork). R osenberg and Sejnowski did a cluster analy
sis of these vectors in the trained netw ork. Roughly, their p rocedure
w as as follow s. They asked, for every functional vector in th at space,
W hat o th er such vector is closest to it? The answ ers yielded about 30
v ector pairs. They then constructed a secondary vector for each such
pair, by averaging the tw o original vectors, and asked, for ev ery such
secon d ary vector, W hat oth er secondary v ector (or so far unpaired
prim ary vector) is closest to it? This p rodu ced a sm aller set of
secon d ary-vector pairs, on w hich the averaging proced u re w as re
p eated to produce a set of tertiary vectors. These w ere then paired in
turn, and so forth. This p rocedure produ ces a hierarchy of groupings
am ong the original transform ations, and it com es to an en d w ith a
grand division of the 79 original vectors into tw o disjoint classes.
A s it h ap p en s, that d eep est and m ost fundam ental division w ithin
the hidden-unit v ector space corresp on d s to the division b etw een the
con son an ts and the vow els! Looking further into this h ierarch y, into
the con son an t b ranch, for exam ple, w e find that there are sub
divisions into the principal consonant typ es, and that w ithin these
b ran ch es there are further subdivisions into the m ost similar con
sonants. All of this is depicted in the tree diagram of figure 9 .8 . W hat
the netw ork has m an aged to recover, from its training set of several
thou san d English w ord s, is the highly irregular phonological signif
icance of standard English spelling, plus the hierarchical organization
of the p h on etic structure of English speech.
H ere w e have a clear illustration of tw o things at once. The first
lesson is the capacity of an activation-vector space to em body a rich
and w ell-structured hierarchy of categories, com plete w ith a similar
ity m etric em bracing everything w ithin it. A n d the secon d lesson is
the capacity of such netw orks to em body rep resentation s of factors
and p attern s that are only partially or implicitly reflected in the cor
pus of inputs. Though I did not m ention it earlier, the rock /m in e n et
w ork provides another exam ple of this, in that the final partition
m ad e on its hidden-unit vector space correspon d s in fact to the objec
tive distinction betw een sonar targets m ade of m etal and son ar targets
m ade of nonm etal. That is the true uniform ity that lies behind the
ap p arently chaotic variety displayed in the inputs.
It is briefly tem pting to suggest that N ETtalk has the con cep t of a
tiard c, for exam ple, and that the rock/m ine netw ork has the concept
176 The Structure of Science
MEASURE OF DISSIMILARITY
( DISTANCE BETWEEN VECTORS )
Figure 9.8
Hierarchy of partitions on hidden-unit vector space of NETtalk
On the Nature of Theories 177
of m etal. But this w o n 't really do, since the vector-sp ace rep resen ta
tions at issue do not play a conceptual or com putational role rem otely
rich en ou gh to m erit their assimilation to specifically h u m an con
cep ts. N evertheless, it is plain that both netw orks have contrived a
system of internal representations that truly corresp on d s to im p or
tant distinctions and structures in the outside w orld, stru ctu res that
are n o t explicitly represented in the corpus of their sen sory inputs.
The value of those representations is that they and only they allow
the netw orks to "m ak e sen se" of their variegated and often noisy
input co rp u s, in the sense that they and only they allow the netw ork
to resp o n d to those inputs in a fashion that system atically red u ces the
error m essag es to a trickle. These, I need hardly rem ind, are the func
tions typically ascribed to theories.
W h at w e are confronting here is a possible conception of know l
edge o r understan din g that ow es nothing to the sentential categories
of cu rrent com m on sense. A n individual's overall theory-of-the-
w orld, w e m ight v entu re, is not a large collection or a long list of
stored sym bolic item s. R ather, it is a specific point in that individual's
synaptic w eight space. It is a configuration of connection w eights, a
configuration that partitions the system 's activation-vector space(s)
into useful divisions and subdivisions relative to the inputs typically
fed the system . 'U seful' here m ean s 'tend s to minim ize the error m es
sages'.
A possible objection here points to the fact that differently w eight
ed sy stem s can produ ce the sam e, or at least roughly the sam e, parti
tions on their activation-vector spaces. A ccordingly, w e m ight try to
abstract from the idiosyncratic details of a system 's connection
w eigh ts, and identify its global theory directly w ith the set of parti
tions th ey produ ce w ithin its activation-vector space. This w ould
allow for differently w eighted system s to h ave the sam e theory.
There is som e virtue in this suggestion, but also som e vice. W hile
differently w eighted system s can em body the sam e partitions and
thus display the sam e outpu t perform ance on any given input, they
will still learn quite differently in the face of a p rotracted sequence of
new and problem atic inputs. This is because the learning algorithm
that drives the system to new points in w eight space d oes n ot care
about the relatively global partitions that have been m ad e in activa
tion sp ace. All it cares about are the individual w eights and how they
relate to ap p reh en ded error. The law s of cognitive evolution, there
fore, do n ot operate prim arily at the level of the partitions, at least on
the view of things here being explored. R ather, they operate at the
level of the w eights. A ccordingly, if w e w ant our "u n it of cognition"
Ло figure in the laws of cognitive developm ent, the point in w eight
178 The Structure of Science
space seem s the w iser choice of unit. W e need only concede that
different global theories can occasionally p rodu ce identical short-term
behavior.
The level of the partitions certainly corresp on d s m ore closely to the
"co n cep tu al" level, as u n derstood in com m on sense and traditional
theory, but the point is that this seem s not to be the m ost im portant
dynam ical level, even w hen explicated in neurocom putational term s.
K now ing a creatu re's vector-sp ace partitions m ay suffice for the accu
rate sh ort-term prediction of its behavior, but that know ledge is in
adequate to predict or explain the evolution of those partitions over
the course of time and cruel experience. K now ledge of the w eights,
by con trast, is sufficient for this task. This gives substance to the con
viction, voiced earlier in section 2, that to explain the phen om en on
of conceptual change, w e need to unearth a level of subconceptual
com binatorial elem ents within w hich different concepts can be
articulated, evaluated, and then modified according to their p er
form an ce. The connection w eights provide a level that m eets all of
these conditions.
This general view of how know ledge is em bodied and accessed in
the brain has som e further appealing features. If w e assum e that the
brains of the higher anim als w ork in som ething like the fashion out
lined, then w e can explain a num ber of puzzling features of h um an
and anim al cognition. For one thing, the speed-of-relevant-access
problem sim ply disappears. A netw ork the size of a hum an brain—
w ith 1011 n eurons, 103 connections on each , 1014 total connections,
and at least 10 distinct layers of hidden units-— can be exp ected , in
the course of grow ing up, to partition its internal vector spaces into
m an y billions of functionally relevant subdivisions, each responsive
to a broad but proprietary range of highly com plex stimuli. W h en the
n etw ork receives a stim ulus that falls into one of these classes, the
netw ork produ ces the appropriate activation vector in a m atter of
only tens or hun dreds of m illiseconds, b ecau se that is all the time it
takes for the parallel-coded stim ulus to m ake its w ay through only
tw o or three or ten layers of the m assively parallel netw ork to the
functionally relevant layer that drives the appropriate behavioral re
sponse. Since inform ation is stored not in a long list that m ust som e
h ow be search ed , but rather in the m yriad connection w eights that
configure the network, relevant aspects of the creature's total infor
m ation are autom atically accessed by the coded stimuli them selves.
A third advantage of this m odel is its explanation of the functional
p ersistence of brains in the face of m inor dam age, disease, and the
norm al but steady loss of its cells w ith age. H um an cognition d e
grad es fairly gracefully as the physical plant deteriorates, in sharp
On the Nature of Theories 179
con trast to the behavior of typical com pu ters, w hich h ave a very low
fault tolerance. The explanation of this persistence lies in the m assive
ly parallel ch aracter of the com putations the brain perform s, and in
the v ery tiny contribution that each synapse or each cell m akes to the
overall com putation. In a large netw ork of 100,000 units, the loss or
m isbehavior of a single cell will not even be detectable. A nd in the
m ore dram atic case of w idespread cell loss, so long as the losses are
m ore o r less random ly distributed throu ghou t the netw ork, the gross
ch aracter of the netw ork's activity will rem ain u n ch an ged . W h at hap
p ens is that the quality of its com putations will be progressively
d egraded .
Turning now tow ard m ore specifically philosophical con cern s, w e
m ay n ote an u n expected virtue of this approach concerning the m at
ter of sim plicity. This im portant notion p resen ts tw o problem s: it is
robustly resistant to attem pts to define or m easure it, and it is not
clear w h y it should be counted an epistem ic virtue in any case. There
seem s no obvious reason , either a priori or a posteriori, w hy the
w orld should be simple rath er than com plex, and epistem ic decisions
based on the con trary assum ption thus ap p ear arbitrary and unjusti
fied. Sim plicity, conclude som e (van Fraassen 1980), is a m erely
pragm atic or aesthetic virtue, as opposed to a genuinely epistem ic
virtue. But consider the following story.
The rock /m in e netw ork of figure 9 .5 displayed a strong capacity for
generalizing beyond the sam ple echoes in its training set: it can accu r
ately discrim inate entirely new sam ples of both kinds. But trained
netw orks do n ot alw ays generalize so well, and it is interesting w hat
d eterm ines their success in this regard . H ow well the training gen er
alizes is in part a function of how m any hidden units the system p os
sesses, or u ses to solve the problem . There is, it turns out, an optim al
num ber of units for any given problem . If the netw ork to be trained is
given m o re than the optim al num ber of hidden units, it will learn to
respon d appropriately to all of the various sam ples in its training set,
but it will generalize to new sam ples only v ery poorly. O n the other
h and, w ith less than the optim al num ber, it never really learns to
respon d appropriately to all of the sam ples in its training set.
The reason is as follows. D uring the training period th e netw ork
gradually g enerates a set of internal rep resentation s at the level of the
hidden units. O ne class of hidden-unit activation v ectors is ch aracter
istic of rocklike input vectors; anoth er class is characteristic of m ine
like inp u t vectors. D uring this period the system is theorizing at the
level of the hidden units, exploring the space of possible activation
vectors, in hopes of finding som e partition or set of partitions on it
-that the outp u t layer can then exploit in turn, so as to draw the
180 The Structure of Science
The functional properties so far observed in these m odel netw orks are
an encou ragin g rew ard for the structural assum ptions that w ent into
them . But just how accurate are these m odels, as depictions of the
brain's m icrostru cture? A wholly appropriate answ er h ere is u n cer
tain, for w e continue to be uncertain about w hat features of the
brain's m icrostructure are and are not functionally relevant, and w e
are therefore uncertain about w hat is and is not a "legitim ate" sim
plifying assum ption in the m odels w e m ake. Even so, it is plain that
the m odels are inaccurate in a variety of respects. The point of the
p resen t section is to sum m arize and evaluate these failings. Let m e
182 The Structure of Science
Figure 9.9
a) A schematic section of the cerebellum (cell population and fiber density reduced for
clarity). b) Neural matrix
On the Nature of Theories 183
Figure 9.10
The equivalence of nets and matrices
184 The Structure of Science
back throu gh the netw ork to the sites of the relevant synaptic con n ec
tions in such a fashion that each w eight gets nudged up or dow n
accordingly. In a com p uter simulation of the netw orks at issue (which
is currently the standard technique for exploring their properties),
both the com putation and the subsequent w eight adjustm ents are
easily d on e: the com putation is done outside the netw ork by the host
com p u ter, w hich has direct access to and control over every elem ent
of the n etw ork being sim ulated. But in the self-contained biological
brain, w e have to find som e real source of adjustm ent signals and
som e real p athw ays to convey them back to the relevant un its. U nfor
tunately, the empirical brain displays little that an sw ers to exactly
these requirem ents.
N ot that it contains nothing along these lines: the prim ary ascen d
ing p ath w ays already described are typically m atch ed by reciprocal or
"d e scen d in g " p ath w ays of com parable density. These allow higher
layers to have an influence on affairs at low er layers. Y et th e influence
ap p ears to be on the activity levels of the low er cells them selves,
rath er th an on the m yriad synaptic connections w hose w eights need
adjusting during learning. There m ay be indirect effects on the
syn apses, of course, but it is far from clear that the b rain's wiring
d iagram an sw ers to the dem ands of the back-propagation algorithm .
The case is a little m ore prom ising in the cerebellum (figure 9 .9 a),
w hich contains a second m ajor input system in the aptly nam ed
clim bing fib ers (not show n in the diagram for reasons of clarity). These
fibers envelop each of the large Purkinje cells from below in the sam e
fashion that a climbing ivy envelops a giant oak, w ith its filam entary
tendrils reaching well up into the bushy dendritic tree of the Purkinje
cell, w hich tree is the locus of all of the synaptic connections m ad e by
the incom ing parallel fibers. The climbing fibers are thu s at least
roughly positioned to do the job that the back-propagation algorithm
requires of them , and they are distributed one to each Purkinje cell, as
consistent delivery of the error m essage requires (T hom pson 1986).
Equally, they m ight serve som e other quite different learning algor
ithm , as advocated by Pellionisz and Llinas (1985). U nfortunately,
there is as y et no com pelling reason to believe that the m odification of
the w eights of the parallel-fiber-to-Purkinje-dendrite syn ap ses is even
w ithin the causal p ow er of the climbing fibers. N or is there any
clear reason to see either the climbing fibers in the cerebellum , or the
d escen ding p ath w ays elsew here in the brain, as the bearers of any
appropriately com p u ted error-correction m essages appropriate to
need ed synaptic change.
O n the h ard w are side, therefore, the situation does not su p p ort the
idea that the specific back-propagation p roced u re of R um elhart et al.
186 The Structure of Science
w ork, but rath er a large com m ittee of distinct but parallel netw orks,
interacting in sundry w ays. It plainly com m an d s m any spaces of
stunning com plexity, and m any skills in consequence. It stands as a
glow ing invitation to m ake our hum ble m odels yet m ore and m ore
realistic, in h op es of unlocking the m any secrets rem aining.
O ne test of a new fram ew ork is its ability to throw a new and unifying
light on a variety of old ph en om ena. I will close this essay w ith an
exploration of several classic issues in the philosophy of science. The
aim is to recon stru ct them within the fram ew ork of the com putational
neuroscience outlined above. In section 5 w e saw how this could be
done for the case of theoretical simplicity. W e there saw a new w ay of
conceiving of this feature, and found a new perspective on w hy it is a
genuine epistem ic virtue. The hope in w h at follows is that w e m ay do
the sam e for other problem atic notions and issues.
A good place to begin is w ith the issue of foundationalism . H ere
the central bone of contention is w heth er our observation judgm ents
m u st alw ays be theory laden. The traditional discussion endures
largely for the good reason that a great deal hangs on the outcom e,
but also for the less m om entous reason that there is am biguity in
w h at one m ight wish to count as an "ob servation ju d g m en t" (an ex
plicitly u ttered sentence? a covert assertion? a propositional attitude?
a conscious experience? a sensation?), and a slightly different issue
em erges depending on w here the debate is located.
But from the perspective of this essay, it m akes no difference at
w h at level the issue m ight be located. If our cognitive activities arise
from a w eave of netw orks of the kind d iscussed above, and if w e
con strue a global theory as a global configuration of synaptic w eights,
as outlined in section 5, then it is clear that no cognitive activity w h at
ever takes place in the absence of v ectors being processed by som e
specific configuration of w eights. That is, no cognitive activity w h at
ever takes place in the absence of som e theory or other.
This perspective bids us see even the sim plest of anim als and the
y ou n gest of infants as possessing theories, since they too process
their activation vectors w ith som e configuration of w eights or other.
The difference betw een us and them is not that they lack theories.
R ather, their theories are just a good deal sim pler than ours, in the
case of anim als. A nd their theories are m u ch less coherent, less orga
nized, and less inform ed than ours, in the case of hum an infants.
W hich is to say, they have yet to achieve points in overall w eight
space that partition their activation-vector spaces into useful and
On the Nature of Theories 189
the view that our perceptual know ledge is both theory laden and
highly plastic. A nd it will do m ore. N otice that the activation-vector
spaces that a m atured brain has generated, and the p rototyp es they
em body, can en com p ass far m ore than the sim ple sensory types such
as p h on em es, colors, sm ells, tastes, faces, and so forth. G iven high
dim ensional spaces, w hich the brain h as in abundance, those spaces
and the p rototypes they em body can en com p ass categories of great
com plexity, generality, and abstraction, including those w ith a tem
poral dim ension, such as harm onic oscillator, projectile, traveling
w ave, Sam ba, tw elve-bar blues, dem ocratic election, six-course din
ner, courtship, elephant h u n t, civil disobedience, and stellar collapse.
It m ay be th at the input dim ensions that feed into su ch abstract
spaces will them selves often have to be the expression of som e earlier
level of processing, but that is no problem . The netw orks und er dis
cussion are hierarchically arranged to do precisely this as a m atter of
course. In principle, then, it is no h ard er for such a system to rep re
sent typ es of processes, procedures, and techniques than to rep resen t the
"sim p le" sensory qualities. From the point of view of the brain, these
are just m ore high-dim ensional vectors.
This offers us a possible m eans for explicating the notion of a p ara
digm , as used by T. S. Kuhn in his arresting characterization of the
nature of scientific understanding and d evelopm ent (K uhn 1962). A
p aradigm , for Kuhn, is a prototypical application of som e set of m ath e
m atical, conceptual, or instrum ental resou rces, an application ex
pected to have distinct but similar instances w hich it is the job of
norm al science to discover or construct. Becom ing a scientist is less a
m atter of learning a set of laws than it is a m atter of m asterin g the
details of the prototypical applications of the relevant resou rces in
such a w ay that one can recognize and generate further applications
of a relevantly similar kind.
K uhn w as criticized for the vagu en ess of the notion of a paradigm ,
and for the unexplicated criterion of similarity that clustered further
applications around it. But from the perspective of the n eu rocom pu -
tational ap p roach at issue, he can be vindicated on both cou n ts. For a
brain to com m an d a paradigm is for it to have settled into a w eight
configuration that p rodu ces som e w ell-structured similarity space
w hose central hypervolum e locates the prototypical application(s).
A n d it is only to be expected that even the m ost reflective p erson will
be incom pletely articulate on w hat dim ensions constitute this highly
com plex and abstract space, and even less articulate on w h at m etric
distributes exam ples along each dim ension. A com plete an sw er to
these questions w ould require a m icroscopic exam ination of the p er
son 's brain. That is one reason w hy exp osu re to a w ealth of exam ples
192 The Structure of Science
This essay opened w ith a survey of the problem s plaguing the classi
cal or "sen ten tial" approach to epistem ology and the philosophy of
science. I have tried to sketch an alternative approach that is free of all
or m ost of those problem s, and h as som e novel virtues of its ow n.
The following points are w orth noting. Simple and relatively small
netw orks of the sort described above have already dem onstrated the
capacity to learn a wide range of quite rem arkable cognitive skills and
capacities, som e of w hich lie beyond the reach of the older ap proach
to the nature of cognition (e .g ., the instantan eou s discrim ination of
subtle p erceptual qualities, the effective recognition of similarities,
and the real-tim e adm inistration of com plex m otor activity). W hile
the specific learning algorithm currently used to achieve these results
is unlikely to be the brain's algorithm , it does provide an existence
proof: by procedu res of this general sort, netw orks can indeed learn
w ith fierce efficiency. A nd there are m any oth er p rocedures aw aiting
our exploration.
On the Nature of Theories 195
p h y is that norm ative epistem ology cann ot be "n atu ralized " or recon
stru cted w ithin the fram ew ork of any purely descriptive scientific
theory. N otions such as "justified belief" and "ratio n ality ," it is said,
can n ot be adequately defined in term s of the nonnorm ative cate
gories to w hich any natural science is restricted, since "o u g h t" cannot
be derived from " i s ." Conclusions are then draw n from this to the
principled au ton om y of epistem ology from any natural science.
W hile it m ay be true that norm ative discourse cannot be replaced
w ithout rem ainder by descriptive discourse, it w ould be a distortion
to rep resent this as the aim of those w ho w ould naturalize epistem ol
ogy. The aim is rath er to enlighten our norm ative en d eavors by
recon stru ctin g them w ithin a m ore adequate conception of w hat
cognitive activity consists in, and thus to free ourselves from the
b u rd en of factual m isconceptions and tunnel vision. It is only the
autonom y of epistem ology that m ust be denied.
A u ton om y m ust be denied because norm ative issues are never in
d ep en den t of factual m atters. This is easily seen for our judgm ents of
instrum ental value, as these alw ays d epend on factual prem ises
about causal sufficiencies and dependencies. But it is also true of our
m ost basic norm ative con cepts and our jud gm en ts of intrinsic value,
for these have factual presuppositions as well. W e speak o f ju stifica
tion, but w e think of it as a feature of belief, and w heth er or not there
are any beliefs and w hat properties th ey have is a robustly factual
m atter. W e speak of rationality, but w e think of it as a feature of think
ers, and it is a substantive factual m atter w h at thinkers are and w hat
cognitive kinem atics they harbor. N orm ative concepts and norm ative
convictions are thus alw ays hostage to som e background factual p re
suppositions, and these can alw ays prove to be superficial, confused,
or just plain w rong. If they are, then w e m ay have to rethink w h at
ever norm ative fram ew ork has b een erected u pon them . The lesson
of the preceding pages is that the time for this has already com e.
Chapter 10
On the Nature of Explanation: A PDP Approach
N eu ral n etw ork m odels o f sensory processing and associative m em ory provide
resources that allow us to state a new theory o f w hat explanatory u n d er
standing consists in. That theory fin d s the theoretically im portant fa ctors to
reside not at the level o f propositions and the relations betw een them , but at
the level o f the activation patterns across large populations o f neurons. This
theory portrays explanatory understanding and perceptual recognition
as being d ifferent instances o f the sam e m ore gen eral sort o f cognitive achieve
m ent: p ro to ty p e activation. It thus purports to effect a unification o f the
theory o f explanation and the theory o f perception. It also fin d s a system atic
and revealin g unity am ong the w ide diversity o f types o f explanation (causal,
fu n ctio n a l, m athem atical, intentional, reductive, etc.), a diversity that has
been a chronic problem f o r older theories o f explanation.
1 Introduction
The notion of explanation has figured centrally in m ost con tem porary
accoun ts of scientific know ledge and rational belief. Explanation is
usually cited, along w ith prediction, as one of the tw o principal func
tions of our factual beliefs. A nd the rationality of such beliefs is com
m on ly said to be m easured, at least in part, by the relative range or
quality of the explanations they m ake possible. If som ething like this
is co rrect, then it is im portant for us to try to u n d erstan d w hat
explanation is, and w hat distinguishes a good explanation from a
p oor one.
Several existing accounts attem pt to m eet this challenge. They will
be ad d ressed below . The present p aper proposes a new accou n t of
the m atter— the prototype activation m odel— an account distinguished,
for starters, by its being grounded in a novel and u n orth od ox con cep
tion of w hat cognition consists in. That conception derives from cur
rent research in cognitive neurobiology and from parallel distributed
processing (PD P) m odels of brain function. (See R um elhart et al.
1986a, 1986b; C hurchland 1986a, 1989.) These PDP m odels are n ote
198 The Structure of Science
w orth y for m any reason s, but first am ong them in the p resen t context
is their alm ost com plete dissociation from the sentential or propositional
conception of w hat know ledge consists in, and from the conception
of h u m an inform ation processing as rule-governed inference. Those
venerable conceptions play a central role in all of the older accounts of
explanation, and in orth od ox accounts of cognition generally. They
will play alm ost no role in the accou n t to be p roposed .
This is n ot to say that the older accounts of explanation are entirely
w ithout integrity. O n the contrary, from the PDP perspective w e can
see w h y each of them has the appeal it does, and also w hy that appeal
is to som e d egree genuine. That is to say, the prototype activation
m odel successfully reduces/explains the m ajor features of its com peti
tors. It is also free from m ost of their celebrated defects. I will try to
illustrate these claims as w e proceed.
The p rototyp e activation m odel is focused first and forem ost on
w h at it is to have explanatory understanding of a problem atic thing,
event, or state of affairs. The linguistic exp ression , exch an ge, or p ro
duction of such understanding, should there be an y, is an entirely
secon d ary m atter. I shall approach the topic w ith the aims of an
em pirical scientist rather than w ith the aim s of a logician or con cep
tual analyst. The goal is to outline a substantive em pirical theory of
w hat explanatory understanding really is, rath er than to provide an
analysis of the concept of explanation as it currently is used or ideally
should be used. N orm ative issues will certainly be add ressed , and
som e u n exp ected insights will em erge, but norm ative con cern s are
not the prim e focus of the present chapter. R ather, w hat con cern s us
is the nature of the cognitive process that takes place inside the brain
of the creature for w hom explanatory u n d erstanding suddenly
d aw n s, and in w hom it is occasionally reactivated.
O n the prototype-activation (PA) m odel, a close approxim ation to
this p rocess is the process of perceptual recognition, as w h en one
suddenly recognizes an indistinct outline as the face of a close friend,
or as w h en one finally recognizes the faint m otion un d er the hed ge as
a foraging m ouse. O n the PA m odel, essentially the sam e kind of
com putational achievem ent underlies both perceptual recognition
and explanatory understanding. The latter is distinguished prim arily
by being a response to a w ider variety of cognitive situations: it is not
limited to sen sory inputs.
A close connection betw een perception and explanation is by now a
familiar them e in both psychology and philosophy. O ne's perceptual
jud gm en ts, and p erhaps even o n e's perceptual experiences th em
selves, have often been portrayed as the p erceiver's best explanatory
accou nt of the peripheral stimuli (G regory 1966, 1970a; Rock 1983). In
Explanation: A PDP Approach 199
R ecent y ears have seen an upsw ell of research into the functional
p roperties of "n e u ra l" netw orks. These are artificial netw orks that
sim ulate or m odel certain salient features of the neuronal organiza
tion of the brain. W hat is interesting is that even simple versions of
these netw orks have show n them selves capable of som e very striking
com putational achievem ents, and they perform these com putations
in a fashion that recalls im portant features of anim al cognition.
A prim ary feature of neuronal organization is abstractly depicted in
the "n e u ra l" netw ork of figure 10.1a. The circles in the bottom row of
the n etw ork rep resent a population of sen sory n eurons. Each of these
units p rojects a proprietary axonal fiber tow ard a second population
of neuronlike units. The axon there divides into a fan of term inal
b ran ch es, so as to m ake a synaptic connection w ith every unit in the
secon d population. Real cells are not quite so exhaustive in their
Explanation: A PDP Approach 201
Figure 10.1
a) A simple network, b) Learning: gradient descent in weight/error space. Axes are
shown for only 2 of 24 synaptic weights
Figure 10.2
a) Perceptual recognition with a large network, b) Learned partition on hidden-unit
activation-vector space. Axes are shown for only three of seven hidden-unit activation
levels
204 The Structure of Science
tion levels (specifically, <1, 0> for a m ine, < 0, 1 > for a rock) w h en fed
an echo of either kind.
The n etw ork 's initial verdicts are chaotic, since its synaptic w eights
w ere set at rand om values. But und er the p ressure of the w eight-
nudging algorithm , it gradually learns to m ake the desired distinc
tion. A nd after it has m astered the echoes in the training set, it will
generalize: it will reliably identify mine and rock echoes from outside
its training set, echoes it has never h eard before. M ine echoes, it turns
out, are indeed united by som e subtle w eave of features, to w hich
w eave the netw ork has becom e tuned during the training process.
The sam e is true for rock ech oes. (See G orm an and Sejnowski 1988.)
H ere w e have a binary discrim ination betw een a pair of diffuse and
very hard-to-define acoustic properties. Indeed, zve never did define
them ! It is the netw ork that has generated an appropriate internal
characterization of each type of sound, fueled only by exam ples. If w e
now exam ine the behavior of the hidden units during discrim inatory
acts in the trained netw ork, w e discover that the training process has
partitioned the space of possible activation v ectors across the hidden
units. (See figure 10.2b. N ote that this space is not the space of figure
10.1b. Figure 10.1b depicts the space of possible synaptic w eights.
Figure 10.2b depicts the space of possible activation vectors across the
m iddle layer.) The training process has g enerated a sim ilarity gradient
th at culm inates in tw o "h o t sp o ts "— tw o rou gh regions that rep re
sent the range of hidden-unit vector codings for a prototypical m ine
and a prototypical rock. The job of the top half of the netw ork is then
just the relatively simple one of discrim inating these tw o subvolum es
of th at vector space.
Several features of such netw orks beg em phasis. First, the output
verdict for any input is produced very quickly, for the com putation
occu rs in parallel. The global com putation at each layer of units is
distributed am on g m any sim ultaneously active processing elem ents:
the w eighted synapses and the sum m ative cell bodies. H ence the ex
p ression 'parallel distributed p rocessing'. M ost striking, the speed
of processing is entirely independent of b oth the num ber of units
involved and the com plexity of the function execu ted . Speed is d eter
m ined solely by the num ber of distinct layers in the netw ork. This
m akes for very swift processing indeed. In a living brain, w here a
typical inform ation-processing p ath w ay has som ething betw een five
and fifty layers, and each pass throu gh that hierarchy takes som e
thing betw een ten and tw enty m illiseconds p er layer, w e are looking
at overall processing tim es, even for com plex recognitional problem s,
of betw een one-tw entieth of a second and one second. Em pirically,
this is the right range for living creatures.
Explanation: A PDP Approach 205
p erh ap s as m any as a hundred layers along som e pathw ays. Furth er,
real brains divide into m any distinct processing hierarchies w orking
side by side on different problem s. A brain is not a single netw ork,
but a com m ittee of m any cooperating netw orks— p erhaps over a
thou san d of them in a typical m am m alian brain. A nd m ost im portant
for the p resen t issue, the input to a given bank of hidden units com es
not ju s t from the sensory periphery, but from elsew here in the brain
itself. The brain is a recurrent netw ork. The all-up input to any layer
will alm ost alw ays include som e "cu rren t co n te x t" inform ation that
derives from earlier processing elsew here in the brain.
Unlike the m odels in the figures, alm ost any layer of units in the
brain will receive som e "d escen d in g " axonal projections returning
from the next layer of units u p w ard in its hierarchy, and it will often
receive "h o rizo n tal" projections from layers w ithin the m any adja
cen t netw orks w orking on related inform ational tasks. A ccordingly,
w hich p rototyp e vectors get activated w ithin a given layer is generally
a function of diverse inputs, som e of w hich arrive unvarnished from
the sen sory periphery, but m any of w hich reflect con cu rren t activity
elsew here in the brain.
This frees the brain from the knee-jerk style of operation that
w orried us a few m om en ts ago. Its ultim ate behavior is a function
of factors so m any and so subtle, factors that interact in such highly
volatile w ays, that the brain's behavior has becom e predictable only
in its broadest outlines and only for very sh ort periods into the future.
M oreover, the factors controlling behavior reside w ithin the brain
itself as m uch as in the external environm ent.
To retu rn to the basic issue, w e can now see how the brain can
com m an d a large and sophisticated repertoire of prototype activation
vectors, each one representing som e com plex prototypical situation
in the external w orld. W e have seen how such vectors can be
activated by the perceptual apprehension of even a small portion of
the relevant external situation, and how those vectors can activate in
turn behaviors appropriate to the entire external situation, and n ot to
just the small part that w as initially coded in perception.
I w ish to suggest that those p rototype vectors, w hen activated, con
stitute the creatu re's recognition and con cu rren t understanding of its
objective situation, an understanding that is reflected in the creatu re's
subsequent behavior. Of course, a creature m ay fa il to recognize/
u n d erstan d its current perceptual or cognitive situation. The vector
activated at the relevant layer of hidden units m ay fall well outside
any of the prototypical volum es of the relevant state space, and the
behavior subsequently p rodu ced will therefore not be draw n from its
w ell-honed repertoire. The resulting behavior m ay be just confused.
Explanation: A PDP Approach 209
Of cou rse, the vectors them selves rep resen t nothing, save in the
context of the global configuration of synaptic w eights that p rodu ced
them , and w hich dictate their effects on subsequent layers of
n eu ron s. H ere again w e have a superastronom ical range of possibili
ties. In a given subsystem of 108 n eurons, a typical neuron will have
at least 103 synaptic connections arriving from other n eurons, for a
total of 1011 distinct synapses w ithin that subsystem . If each synapse
adm its of only 10 distinct w eights, then w e have 101011 or
20ioo-ooo-ooo,ooo distinct possible configurations of w eights for that sub
system alone.
O verall, and crudely speaking, this m ean s that a typical subsystem
will have a lexicon of 10108 possible representation s, each of w hich has
108 elem ents, and each of w hich could have any one of 101011 possible
"m e a n in g s ." W e should not balk, therefore, at the prem ise of the
following discussion, which regards it as unproblem atic that the brain
should com m an d intricate prototype representations of such things
as stellar collapse, cell m eiosis, p ositron-positron collision, redox
reaction, gravitational lens, oceanic rift, harm onic oscillator, inten
tional action, and econom ic depression. Such p h en om en a, intricate
thou gh they are, are not beyond reach of the representational re
sources described.
The discussion to this point has all been pream ble to the following
suggestion: Explanatory understanding consists in the activation of a
specific prototyp e vector in a w ell-trained netw ork. It consists in the
ap prehension of the problem atic case as an instance of a general type,
a type f o r w hich the creature has a detailed and w ell-inform ed representation.
Such a rep resentation allows the creature to anticipate aspects of the
case so far unperceived, and to deploy practical techniques appropri
ate to the case at han d (see figure 10.3, a and b). Given the preceding
discussion, this idea has som e plausibility already. It is m y aim in the
rem aind er of this ch apter to illustrate how m u ch illumination and
unity this suggestion can bring to a w ide range of cognitive p h e
n om ena.
L et m e open m y exposition by responding to a possible objection,
w hich will allow m e to em phasize an im portant feature of the v ecto
rial representations here at issue. "W h a t you have o u tlin ed ," runs the
objection, "m a y be a successful accou n t of sp ontaneous classification,
but explanatory understanding surely involves a great deal m ore than
m ere classification."
The objection evokes a process of m ere "lab elin g ," a process that
pu ts the appreh en ded situation into a featureless pigeonhole, a
process in w hich m ost of the com plex inform ation contained in the
input is lost in its reduction to a canonical neural respon se. Y et this
Explanation: A PDP Approach 211
Figure 10.3
Explanatory understanding as the activation of a prototype vector, a) Ampliative activa
tion of desert rat vector, b) ampliative activation of rotating plastic body vector
212 The Structure of Science
Figure 10.4
Understanding a subordinate prototype as one instance of a superordinate prototype
Euclid's original Elem ents— acclaim ed as the first great axiom atization
in h istory— I w as therefore struck by how very ««syn tactic and how
vividly diagram m atic the w hole affair w as.
This w as standardly p ut dow n to the sloppiness of Euclid's p re
sentation. I can no longer entirely believe that. Though quite real,
syn tactic system atization now appears as only one of several m odes
of cognitive system atization, p erhaps of relatively m inor im portance.
The basic m ode of system atization is the reconfiguration of the simi
larity m etric across the relevant v ector space, so that previously di
verse and unrelated cases are now rerep resen ted as being relevantly
similar to each other, as being slightly different instances of the sam e
m o re general p rototype. This reconfiguration can happen in tw o
w ay s: w ithin the space that contains the relevant prototyp es, so that
th ey com e to occupy closely proxim ate positions; or in the next layer
u p , so that they are seen as instances of a higher prototype located in
a subsequent layer of hidden units.
This m ore encom passing view of conceptual unification is of course
entirely com patible w ith the idea of syntactic unification through ax
iom atization. It m erely subsum es that m ode of organization und er a
m o re general m ode. The presen ted "a x io m s" of formal system s, you
will recall, are alm ost alw ays axiom schem ata, having a potential infin
ity of distinct but recognizable instances. A nd the rules of inference
are inference schem ata as well, w hose diverse applications m u st be
m astered by the person using them . Thus do w e learn to discrim inate
proofs. W h at unites the class of theorem s is that each is the last line of
som e proof or other. In this w ay, evidently, can prototype-trained
PD P netw orks handle the com binatorial and generative asp ects of
cognition. Their capacity for discrim inating relevant structure and
principled iteration m ay be limited, of course. But then, so is ours (see
Bechtel and A braham son, forthcom ing).
There is a great deal m ore to be said about explanatory unification,
b u t for now let us m ove on.
in the rain ?" asks the seven-year-old boy. "B ecau se you h ave your
new clothes on, and com pany will be here for Thanksgiving dinner in
half an h o u r" is an explanatory reply. H ere w e appeal to various p ro
totypical m isbehaviors— denying a constitutional right, b etraying a
confidence, being inappropriately turned out for a family fete— of
w hich the contem plated behavior is claimed to be an instance. A c
quiring a legal, m oral, or social sensibility is a m atter of m asterin g a
large system of such social-interaction p rototyp es, and of com ing to
perceive the social w orld from w ithin that fram ew ork. (Since w riting
this p ap er, I have learned that m oral explanations have indeed been
recognized and discussed. Sturgeon [1985], for exam ple, assim ilates
them to the D -N p attern .)
C. S. Peirce, w ho called the process "ab d u ctio n ," found the form er
p roblem especially puzzling. Peirce, one of the pioneers of m od ern
form al logic, appreciated very early that for any set of observations
there is a literal infinity of possible h ypotheses that m ight be posed in
explanation. But how can w e possibly search a space of infinite size?
Indeed, h ow can w e even begin to search it effectively w h en its ele
m en ts are not well ordered? Peirce m arveled that h u m an scientists
are able so regularly to p roduce, from this infinite sea of m ostly
irrelevant and hopeless possible candidates, hyp oth eses that are both
relevant and stand som e nontrivial chance of being tru e. From the
sentential perspective, Peirce w as right to m arvel. But from the
n eurocom putational perspective, the situation is not so m ysterious.
W e do not search an infinite space of possible explanations. In
general, w e do not search at all: in familiar cases a suitable prototype
is activated directly. A nd if the novelty of the case foils our w aiting
categories and thus forces us into search m ode, then w e search only
the com paratively tiny space com prising the set of our ow n currently
available p rototyp e vectors. Even here the search is m ostly blind and
probably stops at the first su ccess. If on e's initial en cou n ter w ith the
problem atic explanandum fails to activate directly a familiar and sub
sequently successful prototype vector, then one repeatedly reenters
the problem atic input in a variety of different cognitive con texts, in
hopes of finally activating som e prototype vector or oth er, or som e
vector close enough to an existing p rototype to give one at least som e
handle on the problem (see chapter 11).
Since the range of concurrently possible understan din gs is closed
un d er the relation "is at least w ithin hailing distance of an existing
p ro to ty p e ," then o f course any elem ent from that range will appear
both relevan t and potentially true. Peirce, and w e, are the victim s of a
perspectival effect. O ur hyp oth eses will look at least passably hopeful
to us b ecau se they are draw n from a source that collectively defines
w hat will be found plausible by us. W e should thus be w ary of
assum ing, as Peirce seem s to have assu m ed , that w e have any special
nose for truth.
I h ave raised Peirce's w orry here for a further reason : the p rocess
just described illustrates the role of analogy in the d iscovery of new
explanatory h yp oth eses. So m any of our best theories h ave had their
origins in a provocative initial analogy, perhaps strained in som e
dim ensions, that subsequently p roved to be a fruitful vehicle of
u n d erstandin g (see H esse 1966; Kuhn 1962). The p rocess finds a
read y depiction w ithin the p resent fram ew ork. A p rototyp e vector
w hose activation has hitherto been confined to one em pirical dom ain
subsequently com es to be activated w ith profit in a new dom ain.
220 The Structure of Science
M ore accurately, the new dom ain activates a vector that is close to the
old p rototyp e, close enough to evoke at least som e of its cognitive and
behavioral consequences. A n analogy betw een tw o dom ains has been
discovered and exploited. (O n this, see also ch apter 11, section 3.)
U nfortunately, talk of analogy has alw ays been hobbled by our
inability to say anything very specific about w hat constitutes the
relevant kind of similarity. W e are now in a position to be entirely
specific. A nalogy, or similarity, consists in the close proxim ity of the
respective prototype representations, activated by the tw o empirical
dom ains, in the relevant hidden-unit activation-vector space. A n d the
constituting dim ensions of that very high-dim ensional similarity are
given by the response properties of each n eu ron in the relevant hid
den layer. This m eans that any analogy is a com plex and highly subtle
m atter, easily recognized, perh ap s, even though it will typically
tran scend effective verbal description. M ost im portant, it is no longer
a m ystery w hy both of these things should be so. The m any dim en
sions of the relevant vector space are used by us, but they are not
generally know n to us.
The title of this section refers to the "b e s t" explanation. But how ,
on this PD P approach to explanation, are com peting m odes of u n d er
standing to be evaluated? W hat m akes one explanation better than
an other? H ere w e m ust answ er carefully, since w e are denied the
usual sem antic vocabulary of reference, truth, consistency, entail-
m ent, and so forth. The cognitive kinem atics here being explored
does n ot have sen ten ces or propositions as its basic elem ents; the
basic elem ents are activation vectors. The various dim ensions of epis
tem ic virtue will therefore have to be reconceived in term s that are
grou n ded in this new conception of w hat cognitive activity consists
in.
That will be no small task, and I cannot pretend seriously to u n d er
take it here. W hat I shall do is illustrate how som e aspects of the
problem can be add ressed , and how the netw ork approach supplies
un exp ected insight into at least som e perennial problem s.
A t the sim plest level, it is easy enough to describe the differential
virtue that tw o distinct prototype vectors A and B can have as a re
sp onse to a problem atic situation. If w e assum e that w e are dealing
w ith a w ell-trained netw ork, and that the integrity of the tw o p ro
totypes is not in question, then it m ay be a just criticism to say that A
is sim ply the w rong p rototyp e for the problem atic situation at h and. It
m ay be w ron g because the situation confronted is not a m em ber of
the class of situations that will reliably activate A from alm ost any
perspective, even though it happ ened to activate A on this occasion.
This can occur if the agent appreh en ded only a m isleading part of
Explanation: A PDP Approach 221
the problem atic situation, a part that led to the activation of A because
that unusual part w as relevantly similar to A 's typical activators. U n
fortunately for the agent, h ow ever, the problem atic situation is such
that, if ad dressed from any one of m any slightly different perceptual
or cognitive points of view and then reen tered into the netw ork, it
w ould reliably activate B in alm ost every case. In this statistical sense,
A m isrepresents the situation, w hereas B does not.
C orrelatively, the behavioral consequences typically activated in
turn by A m ay be highly, even lethally, inappropriate to the prob
lem atic situation in question, in contrast to those activated by B. A
real exam ple of both failures w ould be a co y ote's faulty u n d erstan d
ing of the natu re of a small tapered appendage disappearing into a
tuft of long grass. The hungry coyote und erstan ds it as the tail of a
retreating desert rat and acts accordingly. In fact, the problem atic
item is the tail of a poisonous snake, w hich coyotes generally avoid
(see again figure 1 0.3a).
P ragm atic and statistical considerations can thus provide a fairly
robust distinction betw een appropriate and inappropriate p ro
totypes, at least for simple cases. W e can even recon stru ct the distinc
tion betw een the correctness of a prototype on a given occasion, and its
w arrant on that occasion. H igh w arrant is a m atter of low am biguity in
the input. W e need to ask, Is the input vector closely similar to any
other possible input vector that w ould activate a different prototype?
If so, then the ambiguity of the current input is high and the w arrant
of the p rototype vector activated is correspondingly low. If the input
is not rem otely similar to any activationally diverse input, then its
am biguity is low and the w arran t of its activated p rototyp e is high.
Similarity am ong input vectors is m easured as usual: by their geom et
rical proxim ity w ithin the activational state space for the relevan t in
put layer.
The evaluation of occasion-specific explanations is one thing, but
w hat of evaluating the propriety of an entire class of explanations?
H ow do w e evaluate, that is, the integrity of the p rototyp es, and sys
tem s of prototypes, them selves? It is difficult to provide m u ch of an
an sw er to this question, since the answ er m u st draw on som e general
story of how to evaluate the global cognitive configurations of neural
netw orks. It m ust draw , that is, on som e analog of a general account
of "th eoretical v irtu e."
W e are as yet in no position even to sketch such an accou n t. But it is
clear that there m ay be gold in these hills. The cognitive behavior of
quite sim ple netw orks displays an u n expected connection betw een
representational econom y, conceptual unification, and the capacity
for successfully generalizing past experience to novel cases. Very
222 The Structure of Science
briefly, the story is this. Suppose w e wish a netw ork to discrim inate
F s. A nd suppose w e train that netw ork to a high level of success on
som e training set (the rock/m ine netw ork, for exam ple). D espite its
success at recognizing Fs d raw n from the training set, the netw ork
will do very poorly at recognizing new exam ples of F draw n from
outside the training set if the num ber of hidden units in the netw ork
is v ery m u ch larger than a certain optim al num ber. The reason is that,
w ith so m any hidden units to exploit, the lazy netw ork will tend to
learn a distinct and unrelated p rototyp e for each of the distinct and
slightly different exam ples of F in its training set. These unrelated
p rototyp es will sustain appropriate respon ses at the output layer, and
w ith the error m essages thus reduced to nothing, the system will
cease any further learning. In particular, nothing forces it to generate
a sin gle p rototype (at the cen ter of a unitary similarity gradient) as its
respon se to the entire range of relevantly similar exam ples of F in the
training set.
The cost of such laziness com es due w h en the netw ork is p resented
w ith a new exam ple from the relevant class of stimuli, an exam ple
draw n from outside its training set. That exam ple will be similar to all
of the oth ers, but the netw ork is in a relatively poor position to see
th at similarity, since it has not learned to see w hat is com m on to all of
the relevant exam ples in the original training set. To be sure, it will do
b etter than chance at recognizing new exam ples, since it does have
som e scattered grasp of the class at issue. But still, it perform s poorly.
By contrast, a netw ork that is denied a surplus of hidden units, a
n etw ork given too few resou rces to perm it such ad hoc and un p ro-
jectible learning, is a netw ork that is forced to continue learning until
it finds a single p rototyp e region, a region at the cen ter of a unified
similarity gradient, a region near w hich to code every exam ple of F in
its training set. A unified solution to the recognition problem is the
only solution for w hich it has adequate resou rces. After being forced
b y circum stance to find such a solution, the netw ork will subsequent
ly recognize a w ide range of novel exam ples of F alm ost as reliably as
those it w as trained on, since it possesses a similarity gradient ad e
quate to catch alm ost all of them . (For a m ore detailed discussion,
see P. M. C hurchland 1989b or chap ter 9.)
C onceptual unification, evidently, is a cognitive virtue of enorm ous
im portan ce, at least as conceived on the p resen t m odel of cognition. It
is im p ortan t for the very good reason that cognitive configurations
having that virtue do m u ch b etter at generalizing their past experi
ence to new cases. It is therefore reasonable to regard explanations
that involve p rototypes em bedded in a unified cognitive configura
Explanation: A PDP Approach 223
p lan atory understan din g then appears on the thin side, but it does
not disapp ear entirely.
For better or for w orse, the p rototyp e activation m odel seem s com
m itted to the sam e view . A prototype is by definition "g e n e ra l," in
the sam e sense in w hich a property is general: it has m an y instances.
But th at is the only sense in w hich it need be general. A prototype
v ector is not just a cluster of universally quantified conditionals in
neural disguise. A prototype, you will recall, can rep resen t a wide
range of diverse exam ples. There need be no feature that is universal,
or even nearly universal, to all exam ples in the class. The statistical
distribution of relevant features, across the class com p reh en d ed by
the p rototyp e, m ay have alm ost any profile. But w h atev er that dis
tribution is, it will be rep resented in the salience accorded each rel
evant feature in the similarity space configured at the relevant layer
of hidden units. A nd it will likely be reflected in the subsequent
cognitive and m otor behavior that has been learned by the netw ork.
A ctivation of that v ector, nonuniform though its portrait of the class
m ay b e, still rep resents the netw ork's understan din g of the input.
T hough all of this m ay be true, I do not think it accou n ts for m ore
than a small p art of the explanatory appeal of the rem ark "I nursed
tw o flu-infected children for a w eek ," relative to the question "W h y
do you have the flu?" R ather, that explanans has the clout it does
b ecause it activates an etiological p rototyp e concerning the typical
transm ission of viral infections, a prototype w hose w arran t in this
case m ay be extrem ely high, despite the general low probability of
adult illness given exposure to infected children. The w arran t m ay be
high b ecause infection from those tw o children m ay be the only
serious opportunity for infection you encountered, and w e do demand
som e infective etiology for every case of flu. The im portance of causal
stru ctures is also a m ajor them e of Salm on's m ore recent w ork within
the statistical perspective (1978, 1984). From the PD P persp ective, this
rep resents a w elcom e recognition of the im portance of etiological
p rototyp es for explanatory understanding.
O verall, h ow ever, it also rep resents a vision of explanation that is
confined to only the second of the five or six im portant kinds of p ro
totyp es outlined in section 4. C ausal stru ctu res are relevant to som e
kinds of explanations, but certainly not to all: think of explaining
m athem atical facts. A nd despite its sharing som e im portant com m it
m en ts w ith the p rototype activation m odel, the SR m odel rem ains
firmly lodged w ithin the sentential fram ew ork of folk p sychology,
and thu s is heir to all of the empirical and psychological problem s
d iscussed in section 1.
A different and increasingly popular approach to explana
tion abandons hope of finding anything very interesting,
226 The Structure of Science
pragm atics of such occasional "cognitive reco v ery " m ay begin to look
plausible. Y et even in these atypical cases it is plausible only for
language-using hu m an s. Puzzlem ent and its resolution in n on h u m an
anim als (that is, in m ost of the cognitive creatures on the planet) is
n ot plausibly p ortrayed in van F raassen 's term s. A n d n ot surprising
ly. They are the sam e term s that got the D -N m odel in trouble. They
p ortray know ledge as essentially propositional in ch aracter.
A ccordingly, w hat van Fraassen has given us is, at m ost, an
accou n t of the structure and rules of a certain gam e, an essentially
linguistic gam e, a gam e that is occasionally played by the older m em
bers of a single species of anim al on those exceptional occasions w hen
exp lan atory understan din g for som e reason eludes them . It does not
p rovide, nor does it im ply, a satisfactory accou n t of explanatory
u n d erstan ding itself.
P erh ap s because of his general com m itm ent to constructive
em piricism — w hich sees bare descriptive empirical ad eq u acy as the
only genuine epistem ic virtue— van Fraassen despaired too quickly of
finding anything unitary, interesting, and epistem ologically significant
in the general topic of explanation. O n the prototype-activation
m odel, h ow ever, explanatory understanding em erges as being on a
cognitive continuum w ith perceptual recognition, a genuinely episte-
m ological m atter if ever there w as. A dditionally, as w e saw at the end
of the p receding section, explanatory unification is an im p ortan t and
understandable determ inant of how successfully one is able to g en er
alize p ast experience to new cases, w hich is also an epistem ological
rath er than a m erely pragm atic m atter. Let us look m ore deeply into
this topic.
I have m ade conceptual unification an im p ortan t elem ent of the
accou nt p roposed. But the presen t accoun t is not the only one that
accords a central im portance to explanatory unification. Friedm an
(1974) and Kitcher (1981, 1989) have both urged that the virtue of an
explanation is a function of w hether it is, or prom ises to be, an ele
m ent or instance of a unified account of a broad range of p h en om en a.
Explan atory virtue, according to them , is not just a m atter of the rela
tion b etw een the local explanans and the local explanandum . It is also
a m atter of the global virtues of the fram ew ork of w hich the local
exp lanans is but one application. While each offers a different account
of w hat the explanatory unification of our know ledge consists in,
they are agreed in m aking participation in such unity the dom inant
virtue of any explanation.
O n the view of explanation here being defended, th ey are alm ost
certainly right. Som e will resist the elevation of unity to this level of
im p ortan ce, arguing that explanatory unity is a purely aesthetic or
228 The Structure of Science
A notable view that does suffer, h ow ever, is Fod or's view (1984),
w hich attem p ts to em brace the w orst of both positions. Fod or and
Pylyshyn (1981) have argued against the G ibsonians that the p ercep
tual p ro cess is extensively inferential, and Fodor (1984) has argued
against m e (1979), and other N ew Look sym pathizers, th at n everth e
less the process is basically theory neutral. For a b etter insight on
these issues, see Hatfield 1989.
Let u s retu rn to K itcher's account of explanation. It does m ore than
com pellingly p ortray unity as the preem inent cognitive and explana
tory virtue that it is. It also discovers and profitably deploys the idea
of p ro to ty p es. Kitcher takes an im portant step aw ay from the tradi
tional D -N picture by introducing the notion of an abstract argum ent
pattern . These are highly abstract logical structures characteristic of
the particular scientific theory being exploited, stru ctu res that will
yield a great variety of relevantly specific argum ents w h en instanti
ated to suit a specific explanatory problem . O ne is rem inded here of
Kuhn 1962. But unlike K uhn, Kitcher explains the m atter in great
detail and show s how an im pressive am ount of light can be throw n
on m any chronic problem s if w e adopt his point of view .
From the PD P perspective, K itcher's achievem ent m u st be judged
g enuine. The deploym ent of prototypes is w h at is central to explana
tory un d erstan din g, and for m axim al virtue it is vital that they be part
of a unified cognitive configuration. The problem with K itcher's story
is that it is still m uch too narrow ly linguistic in its conception of know l
edge rep resentation , and it encom p asses only a small subset of the
full range of different types of explanation. These are serious defects
in the story. E ven so, I w ould m uch rath er highlight its equally p rom
inent virtues, since I w ould like to claim that the P A account
successfully red u ces it. The incorporation of the general features
of K itcher's story further illustrates the conceptual unity that the
prototype-activation m odel brings to the topic of explanation.
If the conceptual unity it brings is genuine, then it ou gh t to perm it
the recognition and understanding of hitherto u n recognized types of
explanation. There is som e sign of this. The belated recognition of
legal, m oral, and social-etiquette explanations w as a novel experience
for m e, I am som ew h at em barrassed to say, and one m ight suggest
aesth etic explanation as a further category. Perhaps oth ers will see
m ore.
Finally, let m e close this essay by adverting to a highly general vir
tue of the prototype-activation m odel. I am com pelled to cite it, since,
as I look back over this essay, I am distressed at how fragile is the
accou n t p rop osed , and how sketchy are the few details provided.
W hat m akes m e hopeful in the face of this is the fact that the
230 The Structure of Science
1 Introduction
Figure 11.1
233
a) Weight space: the space of all possible configurations of the network's synaptic weights. Axes are
shown for only 3 of 27 synaptic weights, b) A schematic network, c) Hidden-unit activation-vector space,
the space of all possible activation vectors across the population of hidden units
234 The Structure of Science
any given input. But the situation at issue seem s to require that the
netw ork som etim es deliver a different output given the sam e input.
These d em an ds are incom patible, and the solution is to recast the
problem so that the relevant inputs are not strictly identical after all.
W ith the inputs distinguished, w e can then train the n etw ork to joint
com p etence on tw o functions with nonoverlapping dom ains.
A n exam ple will illustrate. Suppose w e w ish to train a n etw ork to
give as o u tp u t the sum of any tw o single-digit num bers en tered as
input. A typical input w ould be <2, 5>, and the desired outp u t <7>.
The relevan t function is thus a smallish set of ord ered pairs, each pair
being an input and its p roper output. N ow suppose w e also w ish to
train the netw ork to give as output the product of any tw o single-digit
num bers entered as input. A typical input w ould be <2, 5 > , and the
desired outp u t <10>. W e are now asking the im possible, unless w e
m ake it possible for the netw ork som ehow to discrim inate inputs on
w hich addition is to be perform ed from inputs on w hich m ultiplica
tion is to be perform ed.
This is easily done. A dd to each input a "con text-fixin g " elem ent,
so that w e en ter <2, 5, + > w hen w e w ant addition perform ed, and w e
enter <2, 5, X > w hen w e w ant multiplication performed. W ith the sets
of input elem ents from each function now com pletely disjoint, the
union of the tw o functions desired will be a legitim ate function also.
W e then train the netw ork on that larger function (i.e ., on that larger
set of in p u t/o u tp u t exam ples), and the result is that it learns both to
add and to multiply. Intriguingly, the typical result of such training is
the partitioning of the hidden-unit activation space into tw o distinct
subspaces, one in w hich the various m ultiplicative transform ations
are all cod ed , and the oth er in w hich the various additive tran sform a
tions are all coded. W hat trips these different cognitive behaviors is
the ap p earan ce of one of the context-fixing elem ents, " + " or " X " , as
part of each input.
This sim ple exam ple illustrates a general strategy. A n etw ork can
learn to deal w ith the "s a m e " class of inputs in a variety of different
w ays according to w hich context-indicating elem ents accom pan y
those inputs. Those context fixers can and often will be supplied by
the external environm ent. But that is n ot their only possible source.
The only firm requirem ent is that the context-fixing inform ation
som ehow reaches the hidden units along w ith the rest of the input
inform ation. This requirem ent is consistent w ith the con text fixers
having their origins som ew here else w ithin the netw ork itself, esp e
cially if w e are dealing w ith a com plex netw ork like the brain, w hich
has m an y layers and m any distinct processing path w ays. The idea is
that the total input to the relevant layer of hidden units includes
236 The Structure of Science
either “ descen d in g" or "h o rizo n tal" projections from one or m ore
distinct layers elsew here in the larger netw ork. In figure 11.1b, this
could be visualized by im agining som e additional axons com ing in
from the side to form synaptic connections onto the hidden units over
and above the connections already arriving there from the sensory
periphery.
This yields a netw ork w here the m an n er in w hich the sensory input
is processed is both variable and under the control of the netw ork
itself. Such a netw ork can p rocess a given p erceptual input in any one
of several different w ays. It can " s e e " a figure as either a duck or a
rabbit, as it w ere, depending on w hich of its activational subspaces
(conceptual resources) it has kicked into gear with the relevant
context-fixing inputs. This is true not only at the level of perceptual
p rocessing, b ut at higher levels as well. A situation draw n from
m em ory, or contem plated in im agination, or app reh en ded from a
printed description can also be u nd erstood in a variety of different
w ays, depending on w hat contextual inform ation accom panies its
ap prehension at the relevant population of p rocessing units. It is not
m ysteriou s, then, on the m odel of cognition here being explored, that
one can learn to perceive/understand the w orld in an Aristotelian
fashion, and in a N ew tonian fashion, and in an Einsteinian fashion,
and then use each fram ew ork (each subvolum e of the activation
space) by turns, to suit either aesthetic w him or practical occasion.
m inutes (D esm ond and Levy 1983). A related line of evidence con
cerns th e long-term potentiation (LTP) of neuronal resp on se in som e
cells, a p ro cess also occurring in the one-m inute range (D esm ond and
Levy 1986). W hile this evidence is encouraging, it rem ains unclear
w h eth er such changes are either large enough or fast en ou gh to solve
the problem at issue.
A secon d approach points out that the functional relation betw een
the w eight configurations on the one h an d , and the resulting parti
tions on the activation space on the other, is a highly nonlinear and
occasionally volatile relation (see p. 172). Relatively small ch an ges in
the w eigh ts can occasionally produce large and su dden ch an ges in
the partitions across the conceptual space. This fact could m itigate the
problem atic lethargy show n by the point in w eight space.
A third approach m ight attem pt to play dow n the discontinuities
claim ed for our conceptual behavior, by arguing that historical and
autobiographical descriptions of the relevant shifts have b een chroni
cally exag gerated . H ere w e w ould attem pt to deny, or play d ow n , the
p h en om en a w e had originally hoped to explain. Som e com bination of
these poin ts, perh ap s, will constitute a solution to the problem .
P erh ap s they m ay. Fortunately, the discussion of the preceding
section m akes available to us im m ediately a quite different and rather
less strained solution. There is a w ay to accou n t even for large-scale
conceptual shifts, on a time scale as short as m illiseconds, that re
quires n o m otion from the w eight-space point w h atsoever. The cru
cial idea is the idea of conceptual redeploym ent, a process in w hich a
conceptual fram ew ork that is already fully developed, and in regular
use in som e oth er dom ain of experience or com prehension, com es to
be used for the first time in a new dom ain.
E xam p les are m any and familiar. C onsider H u ygen 's seven teen th -
cen tu ry realization that optical phen om en a, previously grasp ed via
the ray traces of geom etrical optics, could be m ore com prehensively
u n d erstood as instances of w ave p h enom en a. H ere the theory of
w aves in m echanical m edia— a theory already w ell-form ed in
H u y g en 's m ind in connection w ith sound w aves and w ater w aves—
w as applied in a dom ain hitherto un ad dressed by that fram ew ork,
and w ith system atic success. There w as no need for H uygens to effect
a global reconfiguration of his synaptic w eights to achieve this con
ceptual shift. H e had only to apprehend a familiar class of ph e
nom ena in a new cognitive context, one supplied largely b y him self,
in order to have the inputs activate vectors in an area of his con cep
tual space quite different from the areas they had previously acti
vated. The difference w as the context fixers b rough t to the problem .
A nd the result w as a radically new u n d erstanding of optical p h e
238 The Structure of Science
that dictate the refractive index of any substance. The refractive in
dexes for transp aren t substances w ere already well know n, and the
system atic agreem en t w ith the predictions of the new theory w as
striking. These various indexes suddenly em erged as tran sp aren t
instances of the electric and m agnetic properties of m atter.
H ere again a familiar dom ain w as am biguous and p roved to be
understandable in m ore than one w ay. W h en addressed w ith the
appropriate context-fixing inputs (perhaps no m ore than the adm oni
tion "A n y ray of light is a train of EM w av es!"), optical p h en om en a
b egan system atically to activate vectors in an un exp ected subvolum e
of conceptual sp ace, a subvolum e that w as initially partitioned by its
extensive training on entirely nonoptical p h en om en a. M oreover,
after extended practice at approaching the old p h en om en a w ith the
new subvolum e in gear, one clearly did better at understanding
things than one did w ith the old fram ew ork. A nd finally, a m ajor
virtue of this shift, a virtue displayed in both of preceding exam ples
as well, is that one now had a unified understanding of w h at initially
ap p eared as disjoint empirical dom ains.
It is clear from these three exam ples that conceptual change is
regularly a m atter of conceptual redeploym ent, as opp osed to fun
dam ental conceptual novelty. It is also clear that such shifts can
initially take place, in a given individual, on a time scale of secon d s or
less, although the full exploration of the novel use of old resources
m ay well take years. Indeed, so m any of the historical exam ples fit
this red ep loym en t m old that one m ay begin to w on d er if history con
tains any exam ples of real conceptual novelty. I believe th at it d oes—
F arad ay 's conception of a "field of fo rce" com es quickly to m in d — but
I also su sp ect that such cases are relatively rare. The bulk of the con
ceptual discontinuities displayed in the history of science are clearly
cases of con cep tu al redeploym ent.
These can often be cases of learning, h ow ever, in the d eep er sense
and b eyond the m aking of the shift itself. The redeployed resou rces
seldom survive extended contact w ith the new dom ain entirely u n
ch anged, and the process of shaping and refocusing those resources
is a p rocess in w hich the relevant subvolum e of on e's activation space
is now subject to a new regim e of training exam ples. This will often
lead to a y et m ore subtle articulation of the an tecedent partitions,
a p rocess of learning that is com paratively slow and thu s easily
explained in term s of the gradual m otion of on e's w eight-space
position.
W hat w e have to acknow ledge is that the notion of "le a rn in g " is
starting to fragm ent in interesting w ays. Beyond the basic b ut com p ar
atively slow process of synaptic adjustm ent, there is the m ore sh ort
240 The Structure of Science
term p rocess w hereby one learns how to deal w ith a puzzling new
situation by repeatedly reapprehending it in conjunction w ith various
context-fixing auxiliary inputs, in h op es of eventually activating som e
robust p rototyp e vector w ithin a subvolum e that is already well
trained. The ten-year old takes apart the old alarm clock and after a
half-hour's pondering sees how it all w orks. The m ath student p u z
zles over a hom ew ork problem and after several false starts suddenly
sees the p ath through it. The physician confronts a confusing set of
sym p tom s and, several failed tests and incoherent diagnoses later,
finally lights on a successful one. These are all paradigm cases of
learning, in the sense of "com in g to u n d e rsta n d ," but the underlying
process here is quite different from the slow process that partitions
on e's activation spaces in the first place.
The frequency and im portance of conceptual redeploym ent re
quires us to acknow ledge a further divergence, w hich I have been
suppressing to this point, b etw een the partitions across on e's activa
tion space(s) and the input-output function one instantiates. Plainly it
is possible for tw o people to have closely similar partitions, but w ide
ly divergent deploym ents: they m ay com m an d essentially the sam e
conceptual resources, but apply them to quite different dom ains. Tw o
physicists (N ew ton and H uygens, say) m ay have a com parable com
m and of both projectile m echanics and w ave m echanics, and yet one
chronically un d erstan ds light as the high-speed ballistic m otion of
tiny corpuscles, while the oth er chronically u n d erstan ds light as a
train of com pression w aves in the ether. The sam e sensory inputs
produ ce different conceptual respon ses in each, and thus different
behavior from each , since particles and w aves often call for different
techniques of m anipulation and behave differently in m any circum
stances. Because of their different applications, the tw o physicists will
h ave a different input-output function, despite com m anding identical
con cep tu al resou rces.
A m ore familiar case concerns the dom ain of application of the
resou rces of folk psychology. W e all share a m ore or less com m on
conception of intentional agents, and yet som e of us, the devoutly
religious, chronically find intentional significance in a wide range of
ph en om en a that are perfectly natural to the rest of u s. The interp re
tation of natural disasters and natural blessings, both large and small,
as the deliberate punishm ents, rew ards, and m essages of a hidden
intentional agent (God) w ho has specific expectations of us is an inter
pretation still em braced by a large portion of the con tem porary p op u
lation. N ot surprisingly, their behavior is often quite different from
those w ho deploy the relevant resou rces m ore narrow ly: they dis
play p rayer, sacrifice, p enance, resignation, and so forth. H ere again
Learning and Conceptual Change 241
the only defect that need be felt in on e's cu rrent conceptual resources
is the fact that they are still diverse rather than unitary.
Exam ples are com m on enough. Classical, or "p h en om en o log ical,"
therm od yn am ics w as enorm ously successful (it helped to produce
the industrial revolution), but this did not dissuade the tradition of
Bernoulli, Joule, Kelvin, and Boltzm ann from repeatedly trying to re-
conceive therm al ph enom ena w ithin the broader fram ew ork of kine
tic and corpuscular theories. N ew tonian m echanics had conquered
m otion at both the astronom ical and the h u m an scales. O ne had to
w on d er if it also held true at the subm icroscopic scale. The possibility
of ap prehending heat as m echanical en ergy at the m olecular level w as
therefore v ery inviting.
A n unusual sensitivity to failures of unity seem s to h ave driven the
g reater p art of Einstein's theoretical w ork. Special relativity w as an
attem pt to bring m echanics and electrodynam ics togeth er un d er a
com m on and internally coh eren t roof. G eneral relativity w as an
attem pt to unify the physics of both accelerated and unaccelerated
reference fram es. In both cases the new conceptual perspective w as
provided by four-dim ensional geom etry w ith n on stan d ard m etrics.
But in neither case w as the search for this m ore unitary perspective
driven by any prom inent experim ental failing in any of the older
view s. His later search for a unified field theory is a further instance
of the sam e general yearning.
This im pulse tow ard unity is vitally im portant in any cognitive cre
ature, an im pulse coequal with sensitivity to the d ata, for reasons w e
can now u n derstand (see chapter 9, section 5). It is curious that the
relative strengths of these tw o im pulses seem so variable across indi
viduals. A m ajor im balance in either direction yields a familiar pathol
ogy. Valuing unity at the radical expense of local empirical success
yields a castle-in-the-air fantasy w orld for its victim to live in. A nd
valuing local empirical success at the radical expense of synoptic
unity traps its victim in a disconnected set of small and w indow -
less room s: it yields a hidebound and narrow vision that will not
generalize successfully to unfamiliar cases.
Finally, conceptual redeploym ent is occasionally prom p ted by
som e fortunate novelty in o n e's experience. A m p ere's observation
that a cylindrical coil of current-carrying w ire produ ced a bipolar
m agnetic field m oved him to reconceive the long-familiar case of
m agnetized iron bars as having circular cu rrents som ehow flowing
inside of them as well, just as in the coil. F resn el's striking d em on
stration of concentric circles of light at the cen ter of a tiny circular
sh ad ow m oved m any thinkers to foresake N ew ton ian corpuscles and
to set about reconceiving light on a w ave m odel. A nd Einstein's m uch
Learning and Conceptual Change 243
Supervised Learning
This class of learning algorithm s requires a suitable n um ber of
ord ered pairs, draw n from the function to be learned, for p resen ta
tion to the netw ork. Of crucial im portance is the availability of the
correct ou tp u t for any of the given inputs. The m ost p opular algor
ithm of this kind is currently the back-propagation p rocedure described
in ch apter 9, section 5. H ere the desired outpu t vector is com p ared
elem ent-by-elem ent with the actual outp u t vector p rodu ced by the
student n etw ork in response to a training input. The difference be
tw een the tw o outputs is then used to com pute a small and propri
etary adju stm ent for each w eight in the entire netw ork. That change
is such as to nudge the netw ork's global point in w eight space a small
distance d ow n the steepest slope of the local "e rro r su rfa ce ." The
path of rep eated nudgings, produced by repeated presen tation s of
the inp u t-ou tp u t pairs in the training set, is a path that eventually
leads to a position of low error or accurate perform ance on all of the
training exam ples.
This learning procedu re is highly effective, but it does have several
severe sh ortcom in gs, both as an account of how biological brains
learn, and even as a technology for training artifical nets. First, the
requirem ent that the correct output be available to the learning n et
w ork in every case is clearly unrealistic. After failing to solve a prob
lem , real brains do not generally get to look at the correct an sw ers at
the back of the book. Second, the brain show s no plausible m ech an
244 The Structure of Science
by con trast, the presentation of each training pair p rodu ces only a
tiny ch ange in the w eight-space position and only a tiny im prove
m en t in the netw ork 's p erform ance.)
In ord er to lubricate this p rocess, and to free it from getting cau gh t
in purely local error m inim a of poor final p erform ance, each w eight is
rep eatedly subjected to ran d om increases and decreases during the
course of settling. These ran d om changes are of significant size at the
beginning of the settling p rocess, but are slowly reduced to zero as it
p roceed s. It is as if the netw ork 's critical elem ents w ere h eated up at
the beginning and slowly allowed to cool in the p ro cess of finding
their m inim ally stressful m utual accom odation of w eights. This p ro
cess is called "sim ulated an n ealin g." In all, the system is evolving
tow ard the m ore probable of its possible configurations relative to the
clam ped constraints. That is w h y the p rocess is called "B oltzm an n
learn in g ."
The p ro cess described constitutes only one training cycle. The
m odified net m ust now be clam ped with the next in p u t-ou tp u t pair,
and the p rocess rep eated. A nd so with all of the pairs in the training
set. The secon d cycle m ay well obliterate m any or m ost of the gains
m ade in the first, but the overall p rocess eventually drives the w eight-
space p oin t to an error m inim um that is com m on to all of the training
pairs. The w eight-space point thus follows a m ore lengthy and
ad ven tu rou s p ath than it does under back propagation, b ut the result
is m u ch the sam e.
There are ad van tages to this proced ure. For one, it does not require
a globally com p u ted error, nor any global system to distribute it.
Global inform ation about error filters across the w hole netw ork , but
by m ean s of m any purely local effects com pu ted locally, in d ep en d en t
ly, and sim ultaneously. For this reason, Boltzm ann learning scales up
to large netw orks rather b etter than does back propagation. The local
ch aracter of the teaching m ay also be m ore plausible from a biological
point of view .
O n th e oth er h an d , Boltzm ann learning is still slow for very large
netw ork s, since they can take a long time to settle to equilibrium, and
a distinct settling cycle m u st be done for each training pair. This diffi
culty can be circum vented w ith artifical netw orks, since the settling
p rocess can be au tom ated to exceptionally high speeds in electronic
analog system s. But it rem ains a prim a facie problem for the p ro ce
dure as an accou n t of learning in biological brains. Fu rth er, learning
can n ot take place unless the correct answ ers (outputs) are som ehow
supplied. Boltzm ann learning, as described above, still d ep en ds on
having the answ ers at the back of the book.
246 The Structure of Science
R einforced learning
W e can escape the unreality of an om niscient teach er by exploring
learning proced u res in w hich the error m essages are less well in
form ed. Indeed, w e can use either of the tw o p rocedures just d e
scribed, w ith the difference that the output v ectors to w hich they are
trained are variously degraded versions of the desired output vectors:
they are partial, or they are distorted by ran d om influences. Surpris
ingly, this can actually im prove the quality of learning in som e cases,
since adding a very small am ount of noise to the teaching signals can
sm ooth out the m isleading idiosyncracies of the exam ples in a given
training set (Zipser and Elm an 1987). O n the w hole, h ow ever, d e
grading the teaching signals produces a proportional (or w orse) d e
gradation in the learning process. For the m ost part, the w eight-space
point is still being nudged dow n an error gradient, but it is no longer
taking the steepest path available at each step. It thus takes longer to
find the relevant error m inim um . W e buy a greater realism at the
expense of slow er learning.
N either the back-propagation nor the Boltzm ann p rocedure is
n ecessary for reinforced learning, how ever. A n y p rocedure that sets
the w eight-space point in m otion, p erh ap s even in ran d om m otion,
and then modifies that m otion in response to som e m easure of su c
cess or failure, counts as an instance of reinforced learning. This will
include a w ide range of possibilities.
rath er than w ith Hebb rules them selves. For Hebb rules can perfectly
well subserve learning in supervised as well as in unsupervised n et
w orks. The Boltzm ann procedure described earlier, for exam ple, uses
principles of local synaptic adjustm ent that are already one instance
of a Hebb rule. Such rules also scale very well to large netw orks. All
told, Hebb rules appear to buy us speed, biological reality, and func
tional fecundity.
That functional fecundity rem ains largely unexp lored , h ow ever,
and few lessons can yet be draw n concerning the character of learn
ing and rational m ethodology in science. E ven so, one or tw o very
general lessons em erge from the preceding discussion. They are
w orth noting.
The first lesson concerns the character of the factors that drive
synaptic adjustm ent and conceptual change. A re they exhausted by
considerations of m ere “ professional in te re st," as som e theorists
have argued (see Pickering 1 9 8 1 ,1 9 8 4 ), or does the w orld itself exert a
robust influence on the process? In the m any netw ork m odels that
have ap p eared in the research p rogram under discussion, it is clear
that it is the w orld itself that is driving the learning process, w heth er
by m eans of back propagation of m easured error, by m eans of the
progressive reduction of cognitive dissonance with im posed output
vectors, or by m eans of a progressive accom odation to the objective
statistical distribution am ong inputs.
It m u st im m ediately be adm itted, of course, that these netw orks are
not functioning in a com plex social w orld, as is a real scientist. A nd it
m u st also be adm itted that for a netw ork as com plex and sensitive as
a h u m an brain, the pressure to instantiate socially acceptable func
tions can often be overw helm ing. But while the character of social
p ressu res will have a vital role to play in any adequate accou n t of
learning in scientific com m unities, there is no reason w h atever to
regard them as exhausting the dynam ical pressu res. W e know that in
nonsocial cases of learning (artificial netw orks, simple anim als), it is
the nonsocial w orld itself that is the instructor, a relentless and often
highly successful instructor. A n d unless institutionalized science
som eh ow rep resents a total corruption of a process that show s
system atic integrity elsew here, there is no reason to em brace the
extrem ely skeptical, antirealist social determ inism suggested above.
O n the contrary, science has outperform ed those p u rer but simpler
creatures.
O n the oth er hand, from the perspective of the p resen t chapter, the
naturalism of the “ strong p ro g ram " in the sociology of know ledge
(Bloor 1976) appears entirely justified. T h roughout this chap ter w e
have b een exploring causal accounts of the learning p ro cess—
Learning and Conceptual Change 249
skill that a standard 103-unit netw ork could h ave; it needs only to be
w eighted properly.
A s w e noted, such a netw ork com m ands, at a bare m inim um , a
thou san d dram atically different possible conceptual configurations,
and it has this capacity with only 103 units and som ething like (103)2
or 106 synaptic connections. A ssum ing, as before, 10 distinct possible
values for each synaptic w eight, such a system has 10106 distinct
possible w eight configurations. This is a very large num ber indeed,
but it is a paltry fraction of the figure for a h u m an brain. In principle,
our com binatorial options are greater than the small artificial n et
w ork's by a factor of 1010I4/1 0 106 = 10<1014- 106) = 999,000,000. T h e
num ber of "dram atically different conceptual configurations" open to
us should therefore be greater than the small netw ork 's by the sam e
factor.
O n e's excitem ent at the extent of the opportunities available here is
quickly joined by a dism ay at the problem of how to explore that
space effectively. If w e suppose that one could m ake an arbitrary
change in each one of one's synaptic w eights 10 tim es every second
and did so for every second of o n e's life (1 0 10 u pdates), one w ould
still have visited a total of only l/lO^100'000'000'000'000-10Hh portion of
the functional positions available. Figures like this, and those in the
p reced in g paragrap hs, change o n e's perspective on things, for they
begin to p ut a recognizable m etric on the space. A m axim al reckoning
of any possible hum an cognitive excursion com p reh en d s but an
infinitesim al part of a minimally reckoned cognitive space.
A s w ith astronom ical space, it is clear that the effective exploration
of cognitive space will require m ajor instrum ental help. W e cannot
run fast en ou gh , jum p high enou gh , or see far enough to explore the
h eaven s w ithout technological augm entation of our native resou rces.
W e n eed m anned spacecraft, u n m anned probes, and optical and
radio telescopes. The sam e is true for the exploration of cognitive
space. O ur native resources are inadequate to the task, by m any
ord ers of m agnitude. But w e need not be limited by our native re
sources. Let us discuss the possibilities.
The ad ven t of artificial neural nets, and of autom ated p rocedures for
teaching them , opens the possibility of autom ating aspects of the sci
entific enterprise itself. C om p uters, of course, have been helping us
to assem ble, organize, and filter data for decad es. But teachable n et
w orks prom ise returns far in excess of these hum ble duties. For they
prom ise to do som ething conceptual w ith the data, som ething similar
Learning and Conceptual Change 251
to w hat intelligent creatures do w ith it. They prom ise the possibility
of effectively autom ating, for the first tim e, the theoretical p art of the
scientific enterprise.
The p ro sp ects here cover a wide range of possible achievem ents.
Let us begin w ith som e of the simplest. "E xp ert sy stem s" are now a
part of the m arketplace. These are carefully w ritten p rogram s, typi
cally diagnostic in their practical applications, that attem p t to en
capsulate and to exploit the expert know ledge available in som e
dom ain. They are regularly good enough to be useful, b ut chronically
they are p oo r enough to be frustrating. Part of the problem is that
such p ro gram s inevitably rep resent som eon e's attem p t to articulate
the available w isdom in the relevant dom ain. Such recon stru ction s
typically fall well short of the detailed expertise of a skilled profes
sional, thou gh in p rogram m ed form they do display the virtues of
tirelessness, speed, and uniform ity of treatm ent. They fall short
because m u ch of an exp ert's w isdom is inarticulate: it consists of
know ledge that is not stored in linguistic form and is difficult both
to recov er from the exp ert and to recast in the idiom of a p ro gram
m ing language (see D reyfus 1979; D reyfus and D reyfus 1986).
From the perspective of know ledge representation in neural nets,
these difficulties are not surprising. O ne's capacities for discrim inat
ing subtle and com plex patterns typically resides in a v ery high
dim ensional representation space, a space w hose individual dim en
sions each codify som e intricate feature of the input space (rem em ber
that each hidden unit receives w eighted inputs from thou san d s of
sensory units). Recovering all of this inform ation from a living expert
is effectively im possible, and exploiting all of it effectively in a serial
m achine m ight take too m u ch time even if it could be recovered .
The solution is to forget the task of trying to articulate the desired
know ledge within a set of explicit rules. Instead, train up an artificial
neural n etw ork on the sam e data set that trained the h u m an expert.
For exam p le, if m edical diagnosis is the expertise being m odeled,
then w h at is needed is a large num ber of pairs in w hich the input is
the profile of metabolic param eters and pathological sym p tom s of a
real patient, and the output is the correct diagnosis as to his disease.
H ere no attem p t is m ade to articulate rules that will con n ect com plex
sym p tom s to specific diseases. The netw ork is left to generate its ow n
"ru le s" in response to the patterns implicit in the large d ata set. W e
do, of co u rse, exploit the h um an exp ert's know ledge to provide
the diagnoses on w hich the netw ork is trained, but it is the m any
exam ples that do the w ork: no attem pt is m ade to articulate that
know ledge. W e w ish only to re-create it— and p erhaps to exceed it in
speed, ran g e, and reliability— in the trained artificial netw ork.
252 The Structure of Science
O nce that expertise has been achieved, w e can read out the con
figuration of w eights that sustains it and then fix those values im
m ediately into any num ber of new netw orks. The expertise can thus
be m ass-p rod u ced w ithout further training. M ore im portantly, w e
can also read out the partitions effected across the various activation
sp aces of the various layers of hidden units, in order to discover w hat
taxon om ic strategies w ere found by the netw ork as its solution to the
general problem set it. P erhaps its taxonom ies will parallel our ow n,
and p erh ap s they will not. The netw ork m ay find new groupings of
old cases, and it m ay identify, as diagnostically im portant, features
that w ent unnoticed by h um an exp erts. In this w ay m ight artificial
netw orks provide us with new insights into the taxon om y and causal
structure of the w orld, even in dom ains w e already com m and.
This approach is repeatable in a wide variety of contexts: psycholog
ical, chem ical, geological, econom ic, m eteorological, and industrial.
A nd no doubt it will be, since PD P expert system s will offer real
ad van tages over the conventional program m ed exp ert system s avail
able today. Instead of struggling to equal h u m an expertise in a speci
fic dom ain, they prom ise to exceed it in alm ost every respect.
Y et these m inor prognostications are not w h at m otivate this closing
section. They serve only to introduce the shape of the larger project.
W hat w e need to address is the problem of training netw orks to a
useful understan din g of dom ains w here h u m an experts have no
u n d erstanding, or none that is satisfactory. This will require that our
artificial netw orks use learning strategies that place negligible reliance
on an teced en t know ledge and expert teach ers, beyond w hat instru c
tion the w orld itself can provide. If w e can co n stru ct genuinely paral
lel hardw are realizations of the large netw orks that will be needed,
and if w e can autom ate such learning p roced u res so that they will
take place m any orders of m agnitude faster than they do now , either
in serial m achines or in hum an brains, then w e can turn such system s
loose on existing data sets like stellar and galactic surveys, the be
havior of national econom ies, the properties of millions of chem ical
com p oun d s, and the varieties of psychological dysfunction. P resen ta
tion of the data m u st be autom ated as well, to exploit the netw ork's
great speed . W e can then exam ine w ith interest w hat order our artifi
cial netw orks m an age to find in such com plex and teem ing dom ains.
For they will be able to explore the space of cognitive possibilities—
large subspaces of it, an yw ay— far m ore swiftly and extensively than
w e can ever hope to explore it w ithout their help.
W h at sort of symbiotic relationships m ay em erge here, betw een
existing hum an brains on the one h an d , and very large and fast n eu
ral nets on the oth er, is an engaging question that invites the im agina
Learning and Conceptual Change 253
tion for a ride. M aking a netw ork equal in all respects to the hum an
brain, but just faster, seem s still m uch too hard a job to be com pleted
in the n ear future. So w e m ust n ot expect to have som ething to w hich
w e can sim ply talk. If w e are willing to settle for less familiar kinds of
interactions, h ow ever, then netw orks large and fast en ou gh to be
useful seem designable and buildable right here and now . The diffi
culty will lie in m aking accessible to us the cognitive achievem ents we
m ay exp ect them to m ake. The goods will alw ays lie in the structured
partitions that em erge, in the course of learning, w ithin the hidden-
unit activation spaces. But as the netw orks get larger and the dim en
sionality of those spaces goes up, it will be progressively h ard er to
display in accessible w ays, and to m ake sense of, the stru ctu res that
develop w ithin them . For w e can expect them , after all, to develop
conceptual resources that are alien to us. Internalizing a penetrating
new fram ew ork m ay thus take som e tim e, even if it is han d ed to us
on a platter.
Even if w e can develop such turbocharged versions of, or adjuncts
to, our native cognitive capacities, the space confronting us rem ains
abyssal. A n electronic or optical realization of the neural organization
of the h u m an brain will have transm ission velocities 107 tim es faster
than axonal velocities, and this m ay allow the artificial system to learn
107 tim es faster than a hum an brain. This w ould be an im pressive
gain. But the conceptual space it could explore in a lifetime w ould still
com prise but a miniscule portion of the space available: it w ould dis
charge less than a single zero in the exp on en t of the den om in ator of
the tiny fraction discussed earlier. There is little p rosp ect, therefore,
that a "fin al, true theory of the co sm o s" is som ething w e can ever
expect to discover. N othing guarantees that any point in hum an
cognitive space is such as to yield a netw ork w ith zero error on every
p erform an ce. A n d should such a point exist, w hich I v ery m uch
doubt, it w ould still be a needle in a m onum ental haystack.
A nd y e t, if there is no real prosp ect of an end to our cognitive jou r
n ey, there is every p rosp ect that our conceptual fram ew orks can con
tinue to get better and better, ad infinitum, especially since w e can
alw ays artificially expand the num ber of neuronal units and synaptic
connections available to a given thinker, and thus expand the con
ceptual sp ace to be searched. In the long ru n , this m ay be a m ore
effective incentive to intellectual progress than the prosp ect of a final
resting place could ever be.
Chapter 12
Perceptual Plasticity and Theoretical Neutrality:
A Reply to Jerry Fodor
The doctrine that the character o f ou r perceptual know ledge is plastic, and can
vary substantially w ith the theories em braced by the perceiver, has been criti
cized in a recent paper by Fodor. H is argum ents are based on certain ex
perim ental fa cts and theoretical approaches in cognitive psychology. M y aim
in this p aper is threefold: (1) to show that Fodor's view s on the im penetrabil
ity o f perceptu al processing do not secure a theory-neutral fo u n d ation fo r
kn ow ledge; (2) to show that his view s on im penetrability are alm ost certainly
fa ls e ; an d (3) to provide som e additional argum ents fo r , and illustrations of,
the theoretical character o f all observation ju dgm en ts.
The idea that observational know ledge alw ays and inevitably in
volves som e theoretical presuppositions or prejudicial p rocessing is
an idea that has provoked m uch discussion in recent years, for its
consequences are profound. If observation cannot provide a theory-
n eutral access to at least som e aspects of reality, then our overall epis-
tem ic adven tu re contains both greater peril, and g reater prom ise,
than w e m ight have thought. The first and p erhaps the m o st im p or
tant consequence is that w e m ust direct our attention aw ay from
foundational epistem ologies, and tow ard epistem ologies that tell a
m ore global story of the natu re of theoretical justification and rational
belief. A second consequence is that our current observational ontol
ogy is just one such ontology out of an indefinitely large num ber of
alternative observational ontologies equally com patible w ith our n a
tive sen sory ap p aratus. A nd a third consequence is that, since som e
theoretical fram ew orks are m arkedly superior to oth ers, the quality of
our observational know ledge is in principle im provable. If the con
ceptual fram ew ork in w hich our perceptual respon ses to the w orld
This paper first appeared in Philosophy o f Science 55, no. 2 (Ju n e 1988). It is
here expanded by some new material on neuroscience added to section 1.2
and som e material on m eaning added to section 2.2.
256 The Structure of Science
Figure 12.1
я) Schematic preparation for illusory color contrast, b) Illusory contours and brightness
contrast
a) b)
Figure 12.2
a) The Muller-Lyer illusion, b) The Muller-Lyer illusion in a realistic setting (bold, ver
tical lines)
em bracing a novel belief? N one of the theorists cited in Fod or's p aper
have defended such an unrealistic view . A nd all of u s, at som e point
or oth er, have em phasized the im portance of long familiarity w ith the
novel idiom , of repeated practical applications of its principles, and of
socialization w ithin a like-minded group of research ers.
K uhn is quite explicit (1962, chapters 5 and 10) that the enveloping
p aradigm that shapes the scientist's perception is n ot constituted
solely by a set of explicit law s, but by an entire disciplinary m atrix that
includes standard w ays of applying and using the resources of the
p aradigm , skills acquired during a long apprenticeship. A nd m y ow n
d iscussion of the plasticity of perception (1979, ch apter 2) has the rel
evant com m unity learning their nonstan dard observational vocabul
ary from birth, in an ongoing practical setting w here no oth er idiom is
even contem plated.
I confess to having used one exam ple w here a tem p orary shift in
p erception can be m ade fairly swiftly: the exam ple of reperceiving the
organization of the solar system in a heliocentric rath er than a geocen
tric fashion (1979, pp. 3 0 -3 4 ). This case is rath er closer to the N ecker
cube in ch aracter than to the case of the inverting lenses. But even
here it w as em phasized that simply having the relevant C opernican
beliefs is not enou gh ; one m u st learn how to see the changing
h eaven s as an unfolding instance of the C opernican organization, as
view ed from our peculiar perspective within it. H aving the relevant
beliefs is one thing: w e are all of us C opernicans, after all. Reshaping
on e's perception is quite another.
The point is a general one. A physics stud en t does not com e to see
the m otions of com m on objects in a new w ay sim ply by m em orizing
N ew ton 's three laws. M ost freshm an physics students do m em orize
those law s, but relatively few have their perceptions m u ch altered.
The few w ho do are distinguished by having practiced the skills of
applying those laws in a wide variety of circum stances. They do com e
to perceive a com m on pattern in the behavior of m oving bodies that
w as hitherto invisible to them , but m em orizing the law s w as only the
first step in a fairly lengthy p rocess. There are sudden flashes of in
sight, to be sure, as w h en one first grasp s how the p attern is instanti
ated in som e typical case. But on the w hole, the process of reshaping
o n e's perception takes time, and it requires m ore than the m ere ad op
tion of a belief or three.
To sum m arize these points, if Fodor is attacking the view that p er
ceptual processing alw ays (or even usually) respon d s directly and im
m ediately to changes in on e's theoretical com m itm ents, then he is
attacking a straw m an. This is not a view that an yon e has defended.
O n the oth er h and, if Fodor is denying that perceptual processing is
Perceptual Plasticity: A Reply to Fodor 265
optic nerve to the lateral geniculate nucleus (LG N ), and stepw ise
from there by other pathw ays to the prim ary visual cortex, to the
secon d ary visual cortex, and from there to a variety of other areas
even higher in the p rocessing hierarchy.
But these “ ascen d in g" p ath w ays are alm ost invariably m atched by
"d e scen d in g " p ath w ays that lead us stepw ise back through the in
term ediate brain areas and all the w ay out to the earliest processing
sy stem s at the retina. The descending projections from the visual cor
tex back to the LG N , for exam ple, are even greater in num ber than
those in the ascending direction. A nd though the claim is not well
established, there is som e evidence that fully 10 percen t of the axonal
fibers in the h um an optic nerve are descending projections from the
LG N back out to the retinal surface itself, the very first tran sd u cer in
the processing hierarchy (W olter 1965; W olter and Lund 1968; Sacks
and Lindenberg 1969).
T here are similar chains of d escending p ath w ays— from the v ar
ious areas top m ost in the inform ation-processing hierarchy, dow n
throu gh all of the interm ediate processing stages, and all the w ay out
to the periphery— for all of the oth er sen sory m odalities as well. This
organizational pattern is typical in m am m als and also in birds (Living
ston 1978, pp. 4 5 -4 9 ). Prim a facie, the function of these descending
p ath w ays is “ centrifugal co n tro l." They allow for the m odulation of
low er level neural activity as a function of the dem an d s sent dow n
from levels higher in the cognitive h ierarchy. Experim entation on
their functional significance is so far lim ited, but lesions confined to
the d escending optic nerve p ath w ays (from LG N to retina) are know n
to cause perceptual deficits in birds, even though the descending
fibers in their case constitute only 1 percen t of the optic nerve total.
Lesioned birds are less able than intact birds to distinguish edible
seeds from oth er m inute objects in dim light (Rogers and Miles 1972).
If such descending path w ays w ere alw ays sharply confined close to
the sen sory p eriphery, or if they w ere to be found scattered only here
and there in the inform ation-processing hierarchy, then w e m ight
have som e realistic hope of dism issing any backw ard loop as an ele
m en t of w hat is still an "en cap su lated m o d u le" from a functional
point of view . But descending p athw ays are the rule in the processing
hierarchy of the brain, not the exception. They app ear to con n ect the
u p p er levels in the hierarchy to m ost and perh ap s to all of the low er
o nes, in each and every one of the sen sory m odalities. In sum , the
w iring of the brain relative to its sen sory p eriphery certainly does not
su ggest the encapsulation and isolation of p erceptual processing. A s
w ith the psychological data discussed earlier, it strongly suggests
exactly the opposite arrangem ent.
Perceptual Plasticity: A Reply to Fodor 267
leased in great quantities by fires and by friction; it alw ays flows from
high-pressure bodies to low -pressure bodies; a sufficiently high
pressure causes the boiling of w ater; a sufficiently low p ressure
cau ses w ater to freeze; and so forth.
M ost im portant, all m em bers of this com m unity regard the p res
sure of caloric as an observable feature of the w orld: w here y ou , upon
touching a sim m ering kettle, have the sp on tan eou s perceptual belief
that this kettle is hot, they have the sp ontaneous perceptual belief that
this kettle has a high caloric flu id pressure. In sum , w e have here a com
m unity using the conceptual fram ew ork of early classical th erm o
d ynam ics as a com m onsense observation fram ew ork for that sam e
range of p hen om en a com m only ad dressed by us with "folk
th erm o d y n am ics"— the familiar fram ew ork of hot and cold. The
virtues of this caloric fram ew ork, and its persistence in the face of
criticism , will not be repeated here, but they are considerable.
I cite this exam ple because the people of this com m unity are m ak
ing sp ontaneous "ob servation al" judgm ents that are obviously laden
w ith theory. M oreover, the theory at issue is know n by us to be false,
and so the prejudicial character of all of their p erceptual beliefs in this
area is m ade even m ore dram atic. C onsider now a philosopher of this
society, Jerry Caloric, w ho argues as follows.
"C o n sid er the illusion produced w hen on e's left hand is allowed to
rest in a bucket of w ater at high caloric p ressu re, and on e's right hand
in a bucket at low caloric pressu re, and then both hands are im m ersed
in a bucket at an interm ediate caloric p ressu re. If you judge w ith the
left h an d , this w ater has a low caloric p ressu re; if you judge w ith the
right, it has a high caloric pressure.
"O f cou rse, both w e and G ranny know how this illusion is p ro
duced. The nerves of the tw o han ds becom e differently fatigued by
the extrem e caloric pressures in the first tw o buckets, and thus each
gives a different and false response to the interm ediate pressure of
the third bucket. But notice that possession of this inform ation does
absolutely nothing to dispel the illusion.
"W e m ay conclude, therefore, that our peripheral m odules are cog
nitively im penetrable. A ccordingly, our perceptual judgm ents about
the caloric-fluid pressures of com m on objects are in an im portant sense
theory neutral. The theories w e em brace have no effect on caloric
perception , and all hum ans w ith norm al perceptual system s will thus
perceive the w orld in exactly this sam e w a y ."
I have here recreated the form of F od or's argum ent in a setting
w here the conclusion is clearly false. The point is to highlight som e of
the w ays his argum ent fails in its original setting. The first lesson is
that Fod or, like the philosopher just quoted, fails to appreciate the
Perceptual Plasticity: A Reply to Fodor 271
Implicit in the preceding rem arks is the view that the m eaning of an
observation term derives not prim arily, nor even p erhaps at all, from
the typical etiology of its observational application, but rath er from
the n etw ork of general beliefs and assum ptions in w hich it is em
b ed d ed. Because the contents of such em bedding netw orks can vary
substantially, so also can the m eaning of our observation term s.
Fod or correctly identifies this approach to m eaning as a m ajor ele
m en t in m y argum ent for the theory-laden character of perceptual
jud gm en t (I call it "th e conceptual-role theory of m ean in g " or "th e
n etw ork ap p ro ach "; he calls it "m ean in g h olism "). The argu m en t is
simple and quickly stated.
d eterm ined in that fashion, such sentences will still be stuck w ith a
significant burd en of prejudicial theory. To achieve a truly theory-
neutral foundation for know ledge, Fod or needs a class of sentences,
or term s, none of w hose sem antic properties is dep en den t on theory.
Second, there are decisive reasons in support of the claim that at
least som e of any observation term 's sem antic properties m u st be
determ ined by the netw ork of beliefs that em beds it. C onsider the
following argum ent.
If a term 'F' is to be a m eaningful observation term , then its p re
dication in ‘F a’ m u st have som e m aterial consequences: it m u st im ply
som e fu rther sentences, it m u st be incom patible w ith som e others,
and so forth. The sentence 'Fa' will clearly have this p rop erty if it is
asserted in a con text w here general sen ten ces such as '(x)(Fx=>Gx)',
‘(x)((Fx & Hx)=> ~ K x )', and so forth, are already assu m ed . 'Fa' will
then im ply 'Ga', be incom patible w ith '(Ha & K a)', and so forth.
But if 'F' figures in no such background beliefs or assum ptions
w h atsoev er, then 'Fa' will be entirely w ithout consequence or sig
nificance for anything. It will have no bridges to link its assertion or
denial w ith the assertion or denial of any other sentence. It will be a
w heel th at turns nothing, a coin w eightless in every balance, an
assertion em p ty by any m easure. Less figuratively, its assertion will
be com putationally inert. It will be w ithout com putational significance
for the v ery cognitive system that asserts it.
M eaningful observation term s, therefore, will always be em bedded
w ithin som e set of assum ptions. A nd since there is no analytic/
synthetic distinction, those assum ptions will alw ays be speculative
and corrigible. M eaningful observation term s, w e seem bou n d to co n
clude, will alw ays be laden with theory.
It will be pointed out, of course, that, even in the absence of any
b ackground assum ptions, 'Fa' will have a h ost of purely fo rm a l co n
seq u en ces, such as T v ~ P ', 'Q=>Fa', and so on. But the p attern of
these trivial consequences is exactly the sam e for 'Fa' as it is for any
oth er putative observation sentence: 'Ga', 'Ha', etc. These con se
quences thu s cannot serve to bestow any distinct significance on 'Fa'.
It is the m aterial consequences of 'Fa' (i.e., the ones that flow from
substantive, or nonform al, background assum ptions) that do that.
W hich is an oth er w ay of stating the central claim of the netw ork
theory: w h at determ ines the m eaning of any term is the peculiar clus
ter of beliefs in w hich the term figures, and the peculiar p attern of
inferences they m ake possible.
G iven the clear inevitability of an ideological com p on en t in the
m ean ing of any observation term , one m ay begin to w on d er at the
relative contribution of that com ponent as against an oth er possible
274 The Structure of Science
the reason. The robust causal connection betw een a high m ean
m olecular kinetic energy and the alien's application of the term 'high
caloric-fluid p ressu re' contributes nothing that I can discern to the
sem antic conten t of the quoted expression. N either does it provide a
reference for that expression (it is quite em p ty of reference), nor does
it guarantee the truth of any of its typical applications (they are all
false). All th at causal connection does is prom pt the aliens spon
taneou sly to deploy the expression at issue. Its m eaning, its reference,
and the truth of its applications (to the extent they have any) arise
from oth er sources entirely.
A n oth er reason for looking skeptically at causal accou n ts of m ean
ing arises from their regular failure to assign sem antic contents that
are consistent w ith the ideology of the speaker. The bronze-age re
port 'Thor is hurling heavenly fire and pounding his h am m er' m ay
be a reliable indicator that (i.e ., have the calibrational content that)
there is a sudden large-scale flux of electrons w ith accom panying
atm ospheric shock w ave. But this latter description w ould radically
m isrepresent the m eaning of the b ronze-age native's rep ort. That
m eaning is fixed by the assem bled beliefs about Thor, fire, and so
forth, em braced by the native. The point here is just the obvious one
that w e can system atically misconceive w hat w e p e r c e iv e , a fact that
causal accou n ts of m eaning are constitutionally inclined to suppress.
A further reason that causal accounts of m eaning are suspect is
that, on such accounts, sem antic con tent becom es com putationally
im potent. O ne's com putational econom y has access to the form al and
structural features of one's observation judgm ents or rep o rts, but it
has no access to their distal causal an tecedents. The com putational or
inferential consequences of an observation judgm ent, therefore, will
be indepen dent of its "c o n te n t," if sem antic con ten t is assigned accord
ing to distal causal antecedents.
A final problem w ith causal accounts is the am biguity of their
assignm ents of m eaning. A n u tterance typically has a unique m ean
ing, at least given context, but an utterance is alw ays the last m em ber
of a long sequence of internal and external causal an teced en ts. W hich
of the m an y causal anteced en ts that m ake up the chain supplies the
"re a l" m eaning of the utterance, and on w hat principle is the selec
tion m ade?
Part of the initial appeal of causal accounts of m eaning derives from
the fact that, w hen w e translate or assign content to the observation
reports of an other speaker, w e typically exp ect his observation re
p orts, as translated by us, to be roughly reliable indicators of the
w orld around him as perceived by us. Thus the im pulse to assign
con ten t on a causal/calibrational basis. But while quite reasonable in
276 The Structure of Science
3 Conclusion
tio n s." The second level concerns the adequacy of the em bedding
fram ew ork overall, as a system adequate to rep resent the range of
nom ologically possible configurations that the objective w orld m ight
assum e. These correspond roughly to w hat C arnap has called "e x te r
nal qu estion s" (Do the categories of m y fram ew ork capture the
objective divisions in reality? Do the basic generalizations of m y
fram ew ork exp ress genuine laws of nature?).
External questions are rath er daunting. Local standards of evalua
tion are b oth inapplicable and question b egging, and global standards
are vagu e and elusive. G ood positivist that he w as, C arnap reacted by
denying that such questions are factual in ch aracter, claim ing that the
decision to u se a given conceptual fram ew ork is ultim ately just a p rac
tical question, to be decided on pragm atic grounds. But C arn ap w as
m istaken in seeing a fundam ental difference betw een the tw o kinds
of questions, and b etw een tw o kinds of grou n d s for accep tan ce. So-
called "extern al qu estions" are just large-scale theoretical questions,
to be decided on empirical and system atic grou nd s like anything else.
A nd so-called "p rag m atic" considerations attend epistem ic decisions
at every level of inquiry, even the m ost hum ble and m u n d an e.
In fact, external questions are confronted by hum ans and dealt with
on a daily basis, by scientists inventing and evaluating new fram e
w orks for understan din g this or that dom ain, and also by infants
and children, w ho m ust evolve a conceptual fram ew ork adequate
to cond u ct a life in the Lebensw elt of con cu rren t h um an society. This
m ean s that our conceptual fram ew orks can and regularly do und er
go change, both within the lifetime of an individual and in society
as a w hole, over historical periods. To use a H egelian figure, the
jou rn ey of the h um an spirit is essentially the story of our evolving
conception of the w orld, and of our ow n place w ithin it. O ur eyes are
little different from a baboon's or a ch im p an zee's, but our p erceptual
know ledge is profoundly superior to theirs. O ur m otor system s are
little different from those of any other prim ate, but our practical cap a
bilities and intentional actions encom pass universes quite closed to
them . The m ain difference lies in the dram atically superior con cep
tual fram ew orks w e have evolved epigenetically, and not w ithout
m isad ventu re, over the course of the last 500,00 0 years.
If w e h ave com e this far, m u st the journ ey end h ere? M anifestly
not. The long aw akening is potentially endless. The h u m an spirit will
continue its breathtaking adventure of self-reconstruction, and its
percep tu al and m otor capacities will continue to develop as an inte
gral part of its self-reconstruction. But only if w e try hard to see new
opportunities, and only if w e w ork hard at leaving old fram ew orks
behind.
Chapter 13
Conceptual Progress and Word-World Relations:
In Search of the Essence of Natural Kinds
This paper first appeared in the Canadian Journal o f Philosophy 15, no. 1 (March
1985).
282 The Structure of Science
1 N atu ral Kinds and Scientific P rogress: The P utnam -K ripke View
W h at w ants rejecting in the Putnam -K ripke story is not just the speci
fic m odels of how the extensions of natural-kind term s are secured
(by a dem onstrative ostension or dubbing cerem on y, plus som e re
cursive device), but also the background vision of the aims of science,
and of the nature of scientific progress, in w hich that claim is m ad e.
Their choice of exam ples is significant: gold, w ater, and heat dom i
nate the discussion. These are all kinds that are well entrench ed in
the com m on vocabulary, and all have enjoyed a relatively sm ooth
reduction to som e category provided by m icrotheory. W ater is H 20
m olecules, gold is A u atom s, and heat is m olecular kinetic energy.
The picture w e get, if w e generalize innocently from these exam
ples, is one of a stable anteced en t taxon om y of reality, a fram ew ork of
co m m on sen se term s each in firm referential contact with an objective
ly real class, w here the advance of science provides us w ith nothing
m ore than a new and m ore penetrating accou n t of w hat unites that
already palpated class. The aim of science em erges as the explication of
natural classes already secured, rath er than as the discovery of novel
classes that finally do display natural integrity, classes that m ay well
fail to correspond , even roughly, to the specious categories w e
antecedently em braced.
This conservative picture tends to suppress the im portant fact that
the an teced en t taxonom y provided by com m on sense is as richly
theoretical, conjectural, and provisional as are the m ore obviously
"th eo retical" taxonom ies that m ay, or m ay not, reduce it. A nd it tends
to suppress the fact that the "m an ifest" taxon om y of com m on sense
has neither a better, nor a firm er, nor even a different handle on the
hidden natural kinds than do the novel theoretical taxonom ies that
aspire to replace it.
Since all know ledge is conjectural and theoretical, any fram ew ork
of com m onsense term s at any stage of our history will have the sam e
tenuous sort of claim to have as extensions the real natural classes.
That claim will have the sam e sort of grounding at any stage, nam ely,
the explanatory and theoretical success of the current taxonom ic
In Search of the Essence of Natural Kinds 283
in our quotidian affairs, since nothing in nature answ ers to the col
lected law s of "com m on sen se th erm od y n am ics."
I am keenly aw are that this claim is initially im plausible, but I stand
by it. Rather than enter again upon the intricacies of its defense,
how ever, let m e here address w h at makes it an upsetting claim. It
u p sets because it holds up the p rosp ect that entire dom ains of our
co m m on sen se observational vocabulary m ight be utterly w ithout
reference or extension. It holds up the sp ecter of large chunks of
language, even an entire language, unconnected to the w orld by w ay of
the reference of its singular term s and the extension of its general
term s. This is not only shocking to contem plate; it leaves it a m ystery
how any language could function in such a disconnected predicam ent.
Shocking or no, let us contem plate it. (W e will address the m ystery
of how a language can function in that predicam ent in a m o m en t.)
W ithout im pugning anything that currently passes for com m on
sense, w e can easily appreciate the possibility of such m assive ref
erential disconnection if w e imagine a radically false scientific theory
taught and used as an observation fram ew ork by ordinary people. To
take an easy exam ple, suppose w e w ere to use the fram ew ork of calor
ic flu id to conceptualize and talk about therm al ph en om en a. (This
possibility is explored at length in m y 1979, p p . 1 6 -2 5 .) M ore specifi
cally, suppose that w e w ere trained to exp ress our spontaneous
observation judgm en ts and rep orts in the vocabulary of caloric theory,
e .g ., "T h is porridge has a high-caloric fluid pressu re, M o m m y ."
Since you are antecedently convinced that caloric fluid does not exist
and that the vocabulary of caloric theory is em pty of reference and
extension, you will agree that such a linguistic regim e w ould m eet the
conditions at issue: a vocabulary functioning in an "ob servation al"
m od e, but w holly w ithout reference or extension.
There is no question that the vocabulary of caloric theory could
function sm oothly in the role described. (In fact, it w ould function
even m ore sm oothly than our cu rrent co m m on sen se conceptions.)
A nd the reason it could do so is that, while it w ould be disconnected
from the w orld referentially, it w ould rem ain richly connected to the
w orld causally, by w ay of our conceptual and linguistic respon ses to
sensory input, and by w ay of our m otor respon ses to our deliber-
ational output. The prim ary connection betw een any language and
the w orld is thus a causal one, and successful functioning requires
only that the language constitute a useful subsystem of our sen sor
im otor p ath w ays. If a language also enjoys referential connections
w ith the w orld, that is a fortunate and a secondary achievem ent, one
that relatively few linguistic fram ew orks can boast. M assive referen
tial disconnection, therefore, is a real possibility, even in a sm oothly
In Search of the Essence of Natural Kinds 285
"P h log isto n is w hatever bears the 'sam e-spirit' relation to the
phlogiston right here" (speaker points to the vaporous outp ou r
ings above the fire).
W h eth er this secures anything at all in the w ay of a natural extension
will be a function of the theoretical integrity of the notions of "p h lo
286 The Structure of Science
gisto n " and "is the sam e spirit a s ." A s it h ap p ens, neither notion has
any genuine integrity, and so nothing at all is secured. The sam e is
true of oth er exam ples:
least for the m ost part, true. A n d exactly w hich extension gets secured
is a function of just w hat those collected assum ptions a re .
Since a term 's intension (the set of sem antically im p ortan t sen
tences in w hich that term figures) m ay occasionally include indexical
or dem onstrative sentences, it m ust be conceded that exactly w hat
extension gets secured will be determ ined in part by w h ere, and in
w hose brain, that intension is asked to function as p art of an ongoing
cognitive econ om y. Extension is determ ined by intension plus con
text. If this is a concession to P utnam , then I think w e m u st m ake it.
H ow ever, none of this implies that our natural-kind term s enjoy a
firm and stable referential connection to a natural class of objects in
d ep endently of the intensions that govern the use of those term s.
N one of this guarantees that our term s have a n on em p ty extension.
N one of this implies that the extensions of our term s do n ot differ as
they ap p ear in com peting theories. A nd none of this implies that the
extensions of our term s do n ot change as w e change w hich theories
w e em brace.
A ccordingly, w e m u st learn to live, after all, w ith a roughly Fregean
conception of the relation betw een intensions and extensions, even in
the case of natural kinds. There is no royal road to securing natural
classes as extensions for our term s. O ur science m ust achieve refer
ence the old-fashioned w ay: it m ust earn it, by slow plodding
throu gh a variety of very different conceptions of the w orld, in hopes
of fin d in g a theory that successfully carves nature at h er system atic
joints, a theory w hich displays that success in its unexcelled explana
tory and predictive pow ers. O ur referential access to natural kinds is
by w ay of the corrigible system s of sentences w e accept, and of w hat
an sw ers to them in the w orld. This does m ake our access to natural
kinds fluid, uncertain, and problem atic. But that is as it should be.
The extensions of our term s are stably fixed neither by analytic truths,
as in the o rth od ox em piricist tradition, nor by indexical/recursive
pointings, as in the Putnam -Kripke alternative. They are n ot stably
fixed by anything, since they are not stably fixed at all.
This concludes m y discussion of our access to natural kinds. Let m e
now ad d ress the question of the essence of natural kinds. W h at are
natural kinds? W hat distinguishes them from less substantial kinds of
kinds?
2 N atu ral K inds as Law -B ound K inds: Som e Virtues, Consequences, and
D ifficulties
is a law of n ature, and hence that m ass, fo rce, and velocity are natural
kinds or properties. But now consider the sentence
Is this a law of nature? Is chair thus a natural kind? If it is, then a lot of
things are going to count as laws, and a lot of kinds are going to be
natural: toothbrushes, spaceships, toasters, etc. Surely chair is not a
natural kind. But how do w e deny natural-law status to (2)? Sentence
(2) follows from (1) and
W e can all agree that (3), though true, is n ot a natural law, and
p erh ap s here lies the key to denying natural-law status to (2).
Suppose w e identify som e natural laws as basic law s, on grounds of
their relative inexplicability, and then stipulate that nothing else
cou nts as a genuine natural law unless it is a purely logical con se
quence of some one or more-basic laws. This will rule out (2) as a natural
law, since it is neither basic, nor a consequence of purely natural laws
In Search of the Essence of Natural Kinds 289
alone: the nonnatural "law ” (3) is also required for its deduction. This
solution has independent plausibility. To explain a specific chair's
rectilinear m otion through space, one does not point to the fact that it
is a chair, and then cite (2). R ather, one points out th at it has mass,
and then one appeals to (1). It seem s to be (1) that carries the explana
tory clout, n ot (2).
This accom m odation w ith the problem has considerable appeal,
but if w e buy it, w e m u st be prepared to live w ith its consequences.
Specifically, natural kinds turn out to constitute a very small and
exclusive com pany, m uch smaller than one m ight have thought
initially. Let m e ap proach this claim by w ay of som e exam ples.
In learning elem entary m echanics, one learns to deduce lawlike ex
pressions for the rotational inertia of various kinds of bodies. Starting
from the relatively basic law that the rotational inertia / of a point
particle of m ass m at a distance r from the axis of rotation is m r 2, one
can d ed u ce general expressions for the rotational inertia of bodies
w ith a m ore com plex configuration. If M is the m ass of the body, and
R and L the relevant radius and length, w e have, variously,
F or any hoop, I = M R 2
F o r any solid cylinder, I = M R 2/2
F o r any thin rod, I = M L 2/ 12
For any solid sphere, I = 2M R 2/5
For any dumbbell, I = 2 M R 2/5 + M L 2!A
F or any hollow spherical shell w ith a thin rod stuck through it
tran sversely, I = 2 M R 2/3 + ML2/12
Figure 13.1
Rotating bodies and axes of rotation
But (7) seem s to be of a piece w ith (4) and (5), and thus (8) em erges as
292 The Structure of Science
M oral know ledge has long suffered from w hat seem s an unflattering
contrast w ith scientific or other genuinely "factu al" form s of know l
edge. It is not hard to appreciate the appearan ce. O ne has no obvious
sense organ for m oral facts, as one does for so m any of the facts dis
played in the m aterial w orld, and so there is an im m ediate epistem o
logical problem about m oral facts. H ow does one ap p reh en d them ?
C on n ected to this epistem ological problem is an ontological problem .
For em pirical statem ents, one typically finds an objective configura
tion of objects or properties to w hich the statem ent, if tru e, corres
ponds. F or statem ents of m oral truth, such as 'O ne ough t to keep
on e's p rom ises', one seem s not to find com parable objective con
figurations, lying in obliging corresp on dence. A nd even if one rejects
the corresp on d en ce conception of truth implicit in this objection, the
w idely accep ted principle that "o u g h t" cannot be derived from " is "
w ould seem to leave the truth of m oral statem ents grou n d ed in
som ething other than the w ay the m aterial universe h ap p ens to be
configured.
This ap p earan ce has tended to provoke one of tw o possible reac
tions. O n the one h and, w e have a long history of noncognitivist and
oth er deflationary or overtly skeptical accou n ts of m oral "k n ow l
e d g e ." A n d on the other, w e have a long history of inflationary
accou n ts th at try to locate the ground of m oral truth in abstract gen er
al principles that are som ehow certified by reason as opp osed to
em pirical fact. This rough division does not capture every ap p roach :
utilitarianism , perh ap s, is both nondeflationary and nonrationalist.
But the division captures a good deal. M y ow n inclination is to resist
the ap p earan ce that tends to produce these tw o reactions, an d thus to
avoid the m otivation for both pathologies. M oral truths, I shall argue,
are rou gh ly as robust and objective as oth er instances of tru th , but
this objectivity is not secured by their being grounded in pu re reason
or in som e other nonem pirical support. It is secured in som ething
very like the w ay in w hich the objectivity of scientific facts is secured.
298 The Structure of Science
to only the sim plest of environm ental features (see ch apter 9, section
5). This m ean s that, for m an y properties in the w orld, it is the entire
n etw ork that m anages to m ake the relevant discrim inations: the job is
n ot done by the units of the input layer alone.
Let u s turn now to biological brains. It has becom e evident that
very few , if any, of the properties w e com m only regard as observa
tional are distinguished by the peripheral cells alone. M ost of them
involve the activity of several subsequent layers in the processing
h ierarchy. A nd this is certainly true of relatively subtle properties.
C onsider you r observing that the sky is threatening, that a b an an a is
ideally ripe, that the car's engine is still cold, that M ary is em bar
rassed , that the lamb chops on the grill are ready, that the class is
bored, th at an infant is overtired, and so forth. These are the sorts of
im m ediate and autom atic discrim inations that one learns to m ake,
and on w hich on e's practical life depends. To be sure, th ey are am-
pliative discrim inations relative to the often m eager peripheral sti
m ulation that triggers them , and they are highly corrigible for that
very reason . But they are not the result of applying abstract general
principles, nor the result of draw ing covert discursive inferences, at
least in a w ell-trained individual. They rep resent the norm al and
alm ost instantan eous operation of a m assively parallel netw ork that
has b een trained over time to be sensitive to a specific range of en
vironm ental features.
The discrim ination of social and m oral features is surely an instance
of the sam e process, and it is m ade possible by training of a similar
kind. Children learn to recognize certain prototypical kinds of social
situations, and they learn to produce or avoid the behaviors proto-
typically required or prohibited in each. Young children learn to recog
nize a distribution of scarce resources such as cookies or candies as a
fa ir or u nfair distribution. They learn to voice com plaint in the latter
case, an d to w ithhold com plaint in the form er. They learn to recog
nize that a found object m ay be som eone's property, and that access is
limited as a result. They learn to discrim inate unprovoked cruelty, and
to dem an d or expect punishm ent for the tran sgressor and com fort for
the victim . They learn to recognize a breach o f prom ise, and to how l in
p rotest. They learn to recognize these and a hundred oth er p roto
typical sodal/m oral situations, and the w ays in which the embedding
society generally reacts to those situations and expects them to react.
H ow the learning child subsequently reacts to a novel social situa
tion will be a function of w hich of her m an y prototyp es th at situation
activates, and this will be a m atter of the relative similarity of the new
situation to the various p rototyp es on w hich she w as trained. This
m eans that situations will occasionally be am biguous. O ne and the
300 The Structure of Science
Churchland, P. M. 1979. Scientific Realism and the Plasticity o f Mind. Cambridge: Cam
bridge University Press.
Churchland, P. M. 1980. "Plasticity: Conceptual and Neuronal," Behavior and Brain
Sciences 3, no. 1:133-134.
Churchland, P. M. 1981a. "Eliminative Materialism and the Propositional Attitudes."
Journal o f Philosophy 78, no. 2: 67-90.
Churchland, P. M. 1981b. "Functionalism, Qualia, and Intentionality." Philosophical
Topics 12, no. 1: 121-145.
Churchland, P. M. 1982. "Is Thinker a Natural Kind?" Dialogue 21, no. 2: 223-238.
Churchland, P. M. 1984. Matter and Consciousness. Cambridge: MIT Press.
Churchland, P. M. 1985a. "The Ontological Status of Observables: In Praise of the
Superempirical Virtues." In P. M. Churchland, and C. A. Hooker, eds., Images o f
Science, pp. 35-47. Chicago: University of Chicago Press.
Churchland, P. M. 1985b. "Reduction, Qualia, and the Direct Introspection of Brain
States." Journal o f Philosophy 82, no. 1: 8-28.
Churchland, P. M. 1986a. "Som e Reductive Strategies in Cognitive Neurobiology."
Mind 95, no. 379: 279-309.
Churchland, P. M. 1986b. "O n the Continuity of Science and Philosophy." Mind and
Language 1, no. 1: 5-14.
Churchland, P. M. 1986c. "Firewalking and Physics." The Skeptical Inquirer 10, no. 3:
284-285.
Churchland, P. M. 1986d. "Cognitive Neurobiology: A Computational Hypothesis for
Laminar Cortex." Biology and Philosophy 1, no. 1: 25-51.
Churchland, P. M. 1986e. "Sem antic Content: In Defense of a Network Approach." The
Behavioral and Brain Sciences 9, no. 1: 139-140.
Churchland, P. M. 1988a. "Folk Psychology and the Explanation of Human Behavior."
Proceedings o f the Aristotelian Society, suppl. 62:209-221.
Churchland, P. M. 1988b. "Perceptual Plasticity and Theoretical Neutrality: A Reply to
Jerry Fodor." Philosophy o f Science 55, no. 2: 167-187.
Churchland, P. M. 1989a. "O n the Nature of Theories: A Neurocomputational Perspec
tive." In C. W. Savage, ed., The Nature o f Theories, Minnesota Studies in the Phi
losophy of Science, vol. 14. Minneapolis: University of Minnesota Press.
Churchland, P. M. 1989b. "Simplicity: The View from the Neuronal Level." In N.
Rescher, ed., Aesthetic Values in Science. Forthcoming.
Churchland, P. M ., and Churchland, P. S. 1983. "Content: Semantic and Information-
Theoretic." The Behavioral and Brain Sciences 6, no. 1: 67-68.
Churchland, P. M ., and Hooker, C. A., eds. 1985. Images o f Science: Essays on Realism
and Empiricism. Chicago: University of Chicago Press.
Churchland, P. S. 1978. "Fodor on Language Learning." Syntheses 38, no. 1: 149-159.
Churchland, P. S. 1980a. "A Perspective on Mind-Brain Research." Journal o f Philosophy
77, no. 4: 185-207.
Churchland, P. S. 1980b. "Language, Thought, and Information Processing." No Us
14:147-169.
Churchland, P. S. 1980c. "Neuroscience and Psychology: Should the Labor Be Di
vided?" Behavioral and Brain Sciences 3, no. 1: 133.
Churchland, P. S. 1981. "Is Determinism Self-Refuting?" Mind 90:99-101.
Churchland, P. S. 1986. Neurophilosophy: Toward a Unified Understanding o f the Mind-
Brain. Cambridge: MIT Press.
Cynader, M., and Berman, N. 1972. "Receptive Field Organization of Monkey Superior
. Colliculus." Journal o f Neurophysiology 35:187-201.
Davidson, D. 1973. "O n the Very Idea of a Conceptual Schem e." Proceedings and
Addresses o f the American Philosophical Association 47.
References 307
Dennett, D. C. 1971. "Intentional System s." Journal o f Philosophy 68, no. 4: 87-106.
Reprinted in D. C. Dennett, Brainstorms, pp. 3 -2 2 (Montgomery, Vt.: Bradford
Books, 1978).
Dennett, D. C. 1975. "Brain Writing and Mind Reading." In K. Gunderson, ed.,
Language, Mind, and Knowledge, Minnesota Studies in the Philosophy o f Science, vol. 7,
pp. 403-415. Minneapolis: University of Minnesota Press.
Dennett, D. C. 1981. "Three Kinds of Intentional Psychology.” In R. Healey, ed.,
Reduction, Time, and Reality, pp. 37-61. Cambridge: Cambridge University Press.
Reprinted in D. C. Dennett 1987.
Dennett, D. C. 1987, The Intentional Stance. Cambridge: MIT Press.
Desmond, N ., and Levy, W. 1983. "Synaptic Correlates of Associative Potentiation/
Depression: An Ultrastructural Study in the Hippocampus." Brain Research 265:21-
30.
Desn.^nd, N ., and Levy, W. 1986. "Changes in the Numerical Density of Synaptic
Contacts with Long-Term Potentiation in the Hippocampal Dentate Gyrus." Jour
nal o f Comparative Neurology 253:466-475.
Dretske, F. 1981. Knowledge and the Flow o f Information. Cambridge: MIT Press.
Dreyfus, H. 1979. What Computers Can't Do. New York: Harper & Row.
Dreyfus, H ., and Dreyfus, S. 1986. Mind over Machine. New York: Macmillan.
Fales, E. 1982. "Natural Kinds and Freaks of N ature." Philosophy o f Science 49, no. 1:
67-90.
Feyerabend, P. K. 1963a. "Materialism and the Mind-Body Problem ." Review o f
Metaphysics 17:49-66. Reprinted in C. V. Borst, ed., The Mind/Brain Identity Theory,
pp. 142-156. Toronto: Macmillan, 1970.
Feyerabend, P. K. 1963b. "H ow to Be a Good Empiricist." In B. Baumrin, ed., Philoso
phy o f Science, vol. 2, pp. 3-19. New York: Interscience Publications. Reprinted in
H. Morick, ed., Challenges to Empiricism (Belmont, California: Wadsworth, 1972),
pp. 164-193.
Feyerabend, P. K. 1965. "Reply to Criticism: Comments on Smart, Sellars, and Put
n am ." In M. Wartofsky, ed., Boston Studies in the Philosophy o f Science. New York:
Reidel. Reprinted in P. K. Feyerabend, Realism, Rationalism and Scientific Method,
Philosophical Papers, vol. 1 (Cambridge: Cambridge University Press, 1981), pp.
104-131.
Feyerabend, P. K. 1969. "Science without Experience." Journal o f Philosophy 66, no. 22:
791-794.
Feyerabend, P. K. 1970. "Against Method: Outline of an Anarchist Theory of Knowl
edge." In Radner and Winokur, eds., Analyses o f Theories and Methods o f Physics and
Psychology. Minnesota Studies in the Philosophy o f Science, vol. 4, pp. 17-130. Min
neapolis: University of Minnesota Press.
Feyerabend, P. K. 1980. "Consolations for the Specialist." In I. Lakatos, and A. Mus-
grave, eds., Criticism and the Growth o f Knowledge. Cambridge: Cambridge Universi
ty Press: 197-230.
Fodor, J. A. 1968. Psychological Explanation. New York: Random House.
Fodor, J. A. 1975. The Language o f Thought. New York: Crowell.
Fodor, J. A. 1984. "Observation Reconsidered." Philosophy o f Science 51, no. 1: 23-43.
Fodor, J. A.1988a. Psychosemantics: The Problem o f Meaning in the Philosophy o f Mind.
Cambridge: MIT Press.
Fodor, J. A. 1988b. "A Reply to Churchland's 'Perceptual Plasticity and Theoretical
N eutrality.'" Philosophy o f Science 55, no. 2: 188-194.
Fodor, J. A. and Pylyshyn, Z. 1981. "How Direct Is Visual Perception? An Examination
of Gibson's 'Ecological A pproach.'" Cognition 9, no. 2: 139-196.
308 References
Pellionisz, A., and Llinas, R. 1982. "Space-Time Representation in the Brain: The Cere
bellum as a Predictive Space-Time Metric Tensor." Neuroscience 7, no. 12: 2949-
2970.
Pellionisz, A., and Llinas, R. 1985. "Tensor Network Theory of the Metaorganization of
Functional Geometries in the Central Nervous System ." Neuroscience 16, no. 2:
245-274.
Pfaff, D. W ., ed. 1985. Taste, Olfaction, and the Central Nervous System. New York: Rock
efeller University Press.
Pickering, A. 1981. "The Hunting of the Quark." Isis 72:216-236.
Pickering, A. 1984. Constructing Quarks: A Sociological History o f Particle Physics. Chicago:
University of Chicago Press.
Popper, K. 1959. The Logic o f Scientific Discovery. New York: Harper & Row.
Popper, K. 1972. Objective Knowledge. New York: Oxford University Press.
Popper, K ., and Eccles, J. 1978. The Self and Its Brain. New York: Springer Verlag.
Putnam, H. 1964. "Robots: Machines or Artificially Created Life?" Journal o f Philosophy
61, no. 11: 668-691.
Putnam, H. 1970. "Is Semantics Possible?" In H. Kiefer, and M. Munitz, eds., Lan
guages, Belief, and Metaphysics. Albany: State University of New York Press.
Reprinted in H. Putnam, Mind, Language and Reality, pp. 139-152 (Cambridge:
Cambridge University Press).
Putnam, H. 1971. "The Nature of Mental States." In D. Rosenthal, ed., Materialism and
the Mind-Body Problem. New Jersey: Prentice-Hall. Reprinted in H. Putnam, Mind,
Language, and Reality, pp. 429-440 (Cambridge: Cambridge University Press).
Putnam, H. 1975. "The Meaning of 'M eaning.'" In K. Gunderson, ed., Language, Mind
and Knowledge, Minnesota Studies in the Philosophy of Science, vol. 7. Reprinted in
H. Putnam, Mind, Language and Reality, pp. 215-271 (Cambridge: Cambridge Uni
versity Press).
Putman, H. 1981. Reason, Truth, and History. Cambridge: Cambridge University Press.
Pylyshyn, Z. 1980. "Computation and Cognition: Issues in the Foundation of Cogni
tive Science." Behavioral and Brain Sciences 3, no. 1: 111-134.
Quine, W. V. O. 1960. Word and Object. Cambridge: MIT Press.
Quine, W. V. O. 1969. "Natural Kinds." In W. V. O. Quine, Ontological Relativity and
Other Essays. New York: Columbia University Press.
Randi, J. 1982. Flim-Flam! Psychics, ESP, Unicorns, and Other Delusions. Buffalo, N .Y.:
Prometheus Books.
Risset, J. C., and Wessel, D. L. 1982. "Exporation of Timbre by Analysis and Syn
thesis." In D. Deutsch, ed., The Psychology o f Music. New York: Academic Press.
Robinson, D. A. 1972. "Eye Movements Evoked by Collicular Stimulation in the Alert
M onkey." Vision Research 12:1795-1808.
Robinson, H. 1982. Matter and Sense. New York: Cambridge University Press.
Rock, I. 1983. The Logic o f Perception. Cambridge: MIT Press.
Rogers, L. J., and Miles, F. A. 1972. "Centrifugal Control of Avian Retina: Effects of
Lesions of the Isthmo-Optic Nucleus on Visual Behaviour." Brain Research 48:147-
156.
Rorty, R. 1965. "Mind-Body Identity, Privacy, and Categories." Review o f Metaphysics
19:24-54.
Rosch, E. 1981. "Prototype Classification and Logical Classification: The Two Sys
tem s." In E. Scholnick, ed., New Trends in Cognitive Representation: Challenges to
Piaget's Theory. New Jersey: Lawrence Erlbaum.
Rosenberg, C. R., and Sejnowski, T. J. 1987. "Parallel Networks That Learn to Pro
nounce English Text." Complex Systems 1:145-168.
Rosenblatt, F. 1959. Principles o f Neurodynamics. New York: Spartan Books.
312 References
Strawson, P. F. 1958. "P ersons." In H. Feigl, M. Scriven, and G. Maxwell, eds., Con
cepts, Theories, and the Mind-Body Problem, Minnesota Studies in the Philosophy of
Science, vol. 2, pp. 330-353. Minneapolis: University of Minnesota Press.
Sturgeon, N. 1985. "Moral Explanation." In D. Copp, and D. Zimmerman, eds., Moral
ity, Reason, and Truth. Southern Journal o f Philosophy, suppl. to vol. 24.
Suppe, F. 1974. The Structure o f Scientific Theories. Chicago: University of Illinois Press.
Sur, M ., Garraghty, P., and Roe, A. 1988. "Experimentally Induced Visual Projections
into Auditory Thalamus and Cortex." Science 242, no. 4884: 1437-1441.
Taylor, C. 1970. "Mind-Body Identity: A Side Issue?" In C. V. Borst, ed., The Mind/
Brain Identity Theory, pp. 231-241. Toronto: Macmillan.
Taylor, C. 1987. "Overcoming Epistemology." In K. Baynes, J. Bohman, T. McCarthy,
eds., After Philosophy: End or Transformation? pp. 464-488. Cambridge: MIT Press.
Teller, P. 1974. "O n W hy-Questions." Nods 8:371-380.
Thompson, R. F. 1986. "The Neurobiology of Learning and M emory." Science 233:941-
947.
Van Essen, D. C., and Maunsell, J. 1983. "Hierarchical Organization and Functional
Streams in the Visual Cortex." Trends in Neuroscience 6:370-375.
Van Fraassen, B. 1977. "The Pragmatics of Explanation." American Philosophical Quarter
ly 14:143-150.
Van Fraassen, B. 1980. The Scientific Image. Oxford: Oxford University Press.
Van Fraassen, B. 1981. Critical Notice of Paul Churchland, Scientific Realism and the
Plasticity o f Mind. Canadian Journal o f Philosophy 11, no. 3: 555-567.
Wilkes, K. 1981. "Functionalism, Psychology, and the Philosophy of M ind." Philo
sophical Topics 12, no. 1.
Wilkes, K. 1984. "Pragmatics in Science and Theory in Common Sense." Inquiry 27, no.
4: 339-361.
Wilson, M. 1980. "The Observational Uniqueness of Some Theories." Journal o f Philoso
phy 77, no. 4: 208-232.
Wilson, M. 1985. "W hat Can Theory Tell Us about Observation?" In P. M. Churchland
and C. A. Hooker, eds., Images o f Science, pp. 222-242. Chicago: University of
Chicago Press.
Wolter, J. R. 1965. "The Centrifugal Nerves in Human Optic Tract, Chiasm, Optic
Nerve, and Retina." Transactions o f the American Opthalmological Society 63:678-707.
Wolter, J. R., and Lund, O. E. 1968. "Reaction of Centrifugal Nerves in the Human
Retina." American Journal o f Opthalmology 66:221-232.
Zeki, S. 1980. "The Representation of Colours in the Cerebral Cortex." Nature 284:412-
418.
Zeki, S. 1983. "Colour Coding in the Cerebral Cortex: The Reaction of Cells in Monkey
Visual Cortex to Wavelengths and Colours." Neuroscience 9, no. 4: 741-765.
Zipser, D ., and Elman, J. D. 1987. "Learning the Hidden Structure of Speech." Journal
o f the Acoustical Society o f America 83, no. 4: 1615-1626.
Index
U nderdeterm ination
of truth, 142,143
of reference, 276
Unity, 7 7 ,1 4 6 ,1 9 7 , 210, 22 3 -2 3 0 ,
241, 242, 302, 303
Utilitarianism , 297
V an Essen, D. C ., 69
V an Fraassen, В. C ., 1 39-151, 157,
179, 226, 227, 228
V ector coding, 39, 9 2 ,1 0 6 , 209. See
also State-space position
Vector space, 76, 169, 1 7 4 -1 7 8 ,1 8 0 ,
181, 1 8 9 ,1 9 1 ,1 9 3 -1 9 5 , 2 03-207,
216, 220, 233, 247
partitions on, 168-169, 174-178,
2 0 3 -2 0 4 , 232 -2 3 4
Vital spirit, 14, 22