Philosophy of Science, 77 ( July 2010) pp. 419–456. 0031-8248/2010/7703-0005$10.00
Copyright 2010 by the Philosophy of Science Association. All rights reserved.

419

Neuroscience and the Multiple
Realization of Cognitive Functions*

Carrie Figdor†‡

Many empirically minded philosophers have used neuroscientific data to argue against
the multiple realization of cognitive functions in existing biological organisms. I argue
that neuroscientists themselves have proposed a biologically based concept of multiple
realization as an alternative to interpreting empirical findings in terms of one-to-one
structure-function mappings. I introduce this concept and its associated research frame-
work and also how some of the main neuroscience-based arguments against multiple
realization go wrong.

1. Introduction. Many nonreductive physicalists have long been united
by the belief that mental properties or states are multiply realizable.1 This
consensus was established largely on the basis of the intuitions and anal-
ogies presented by Putnam (1967), Block and Fodor (1972), and Fodor

*Received October 2009; revised December 2009.

†To contact the author, please write to: Department of Philosophy, 260 English-Philos-
ophy Building, University of Iowa, Iowa City, IA 52242; e-mail: carrie-figdor@uiowa.edu.

‡I wish to thank Jennifer Mundale, John Bickle, and the audience at the 2003 Society
for Philosophy and Psychology annual meeting for comments on a distant ancestor of
this paper; at least two anonymous reviewers at Philosophy of Science; Tom Polger and
other members of the October 2008 Workshop on Multiple Realization at the University
of Cincinnati, including Ken Aizawa, John Bickle, Carl Craver, Carl Gillett, Larry Sha-
piro, and Jacqueline Sullivan; and my colleagues in the Philosophy Department and the
Program in Cognitive Neuroscience at the University of Iowa.

1. In this literature, the terms “property” and “type” are used interchangeably: e.g.,
mental properties are realized by physical properties, but the one-to-one relation con-
trasted with multiple realization is type-type identity. This mixed usage is harmless if we
assume, as is usual, that particular (or token) states have properties in virtue of which
they are classified by or under types, whatever the ontological status of the properties. I
will follow customary usage to the extent possible to avoid drawing attention to this
issue. I will use the terms “cognitive,” “psychological,” and “mental” interchangeably.
Finally, I also adopt the common (if not universal) current usage of the term “multiple
realizability” to signify possible (e.g., extraterrestrial) as well as actual (e.g., biological)
cases and “multiple realization” to signify actual cases.



420 CARRIE FIGDOR

(1974). As many critics have recently pointed out, intuition and analogy
remain a primary means to argue for multiple realization (MR) in existing
creatures.2 This is rather striking, as the scientific—particularly neurosci-
entific—advances that have occurred since 1967 have prompted many of
those interested in the mind-body problem to wonder how this research
might affect the debate.

Many philosophers who have looked at this research see bad news
for MR. Bechtel and Mundale (1999), Bickle (2003), Polger (2004), Sha-
piro (2004), and others have used data from cognitive neuroscience,
cellular and molecular neuroscience, and vision science to argue that
MR in the cognitive systems of evolved biological organisms is unob-
vious, doubtful, implausible, or false. The cumulative effect of their
arguments merits close attention, aside from its challenge to the role of
intuition and analogy in the debate. First, by focusing on MR in bio-
logical creatures, the implications of the debate for clinical and exper-
imental practices are emphasized. Second, if they are right, they will
have shown that the main argument for nonreductive physicalism is
scientifically unviable with regard to such creatures. The position would
become as practically untenable as phlogiston theory after the discovery
of oxygen. (Of course, some philosophers have long considered it un-
tenable—e.g., P. M. Churchland 1981; P. S. Churchland 1986.) Efforts
to clarify the metaphysics of realization (e.g., Gillett 2003) would also
lose a good deal of their motivation.

The main aims of this article are to demonstrate the empirical viability
of the MR hypothesis and to clarify the terms of an empirically based
debate about its possibility in the cognitive states in existing biological
organisms. To this end, I focus on research in cognitive neuroscience
for two reasons. First, since neuroscience in general has been the primary
source of evidence used in “empirically based” arguments against MR,
this focus obeys a de facto restriction to an “evolved-biological” debate.
A truly empirically based debate would also include cognitive systems
we can engineer, biologically or artificially. Second, since cognitive neu-
roscience in particular is directly concerned with linking mind and brain
in existing animals, its hypotheses, findings, and methods are prima facie
relevant to a debate about the nature of that link in the cognitive systems
of organisms in that class. My third aim is to rebut some prominent
arguments against MR in the cognitive systems of existing animals that
are also based on cognitive neuroscientific data. Since data from the
same science are being used to draw the opposing conclusion from mine,
it is incumbent on me to show how these arguments go astray.

2. See Shapiro (2000), Clapp (2001), and Gillett (2003); other intuitive examples are
found in Kim (1993), while Keeley (2000) examines MR in electric fish.



MULTIPLE REALIZATION 421

The article has three main sections. In section 2, I introduce the cog-
nitive neuroscientific research program that aims to map brain structures
to cognitive processes. In section 3, I motivate and explain cognitive
neuroscience’s version of MR. In section 4, I assess some of the cognitive
neuroscientific evidence used to argue against MR in biological organ-
isms.

2. The Cognitive Neuroscientific Background of an Empirical MR Debate.
In what follows, I adopt more or less intact the coarse-grained vocabulary
of “structure” and “function” that cognitive neuroscientists use when
discussing the entities (properties and relations) that are among what those
philosophers refer to as realizers and realizees (or realized properties).
Most broadly, they are used in ways analogous to the philosopher’s “phys-
ical-mental” distinction. The term “structure” is particularly protean, of-
ten being used to introduce any physical substrate, with precise terms
employed later in context; to put the point in terms borrowed from the
philosophical lexicon, “c-fibers” and “c-fibers firing” may both be initially
called “structures.” Following Bechtel and Mundale’s (1999) working
translation of the philosophical term “brain state” as “activity in a brain
area,” a working translation of not-otherwise-specified structure-function
talk in cognitive neuroscience is that realizing structures are the neuro-
logical or neurophysiological properties of brain areas (networks) inves-
tigated in neuroscience, and realized functions are the cognitive capacities,
functions, or dispositions (the differences among these will not matter
here) investigated in cognitive psychology. The multilevel nature of neu-
roscientific explanation (see, e.g., Craver 2007; Aizawa and Gillett 2009)
strongly implies that at least some realizer properties will be instantiated
by components of entities with realized properties (e.g., properties and
relations of components of a brain area may in combination realize a
cognitive capacity assigned to the brain area). However, nothing in my
discussion turns on whether one adopts this “dimensioned” view of re-
alization (Gillett 2003), as opposed to the “flat” view that restricts realizer
and realized properties to properties of the same individual. Nor does
anything depend on my background assumption that the realization re-
lation in cognitive neuroscience (if not universally) is properly analyzed
in terms of causal role-playing (roughly, F realizes G if and only if F
plays, or contributes to the playing of, the causal role that individuates
G).

The dominant research program in contemporary cognitive neurosci-
ence is localizationism, which hypothesizes that cognitive systems and
brains (in particular, cerebral cortex) have parts and that particular cog-
nitive parts are realized in or by particular brain (especially cortical) parts.
This program was inspired by spectacular discoveries of specific cognitive



422 CARRIE FIGDOR

losses in a few patients with localized brain damage.3 Localizationism
involves more than a commitment to functional specialization. Functional
specialization is compatible with more than one brain area being spe-
cialized to subserve the same function and one brain area being able to
subserve more than one function. However, while functional specialization
without localization has long been a recognized position (Phillips, Zeki,
and Barlow 1984), researchers have often adopted assumptions (discussed
below) that support inferences from mappings in which a particular brain
region is solely responsible for a sole cognitive task. Localizationism is
functional specializationism plus these stronger assumptions.4

However, it is also widely accepted that an adequate explanation of
how the brain subserves cognition will require understanding neural con-
nectivity, without which, some argue, localizationism is just the “new
phrenology” (Phillips et al. 1984, 339; Uttal 2001; Friston 2002). In func-
tional integration, researchers investigate the ways in which coactivated
anatomical regions influence each other’s activity.5 Integrationists do not
claim that cognitive processes are properties of the whole brain’s operation
or large portions of it, which is a holist position—and even holists (e.g.,
Uttal 2001) reject Lashley’s (1929, 1950) theory of equipotentiality, in
which any cortical anatomical region can in principle subserve any func-
tion. Integrationists hold that the brain can be divided into functionally
specialized anatomical areas but that the activity of individual areas will
not suffice to explain cognition. In effect, we can consider the dominant
research program to be composed of an effort (associated with localiza-

3. These include Phineas Gage, a Vermont railroad foreman who suffered frontal lobe
damage in an 1848 accident and retained much of his intelligence but whose personality
changed dramatically; Broca’s patient “Tan,” who after damage to an area of his left
hemisphere could understand language but could only utter the syllable “tan”; and H.
M., whose 1953 lobotomy led to his inability to form new memories. H. M. was revealed
to be Henry Gustav Molaison after his death on December 2, 2008 (New York Times,
December 4, 2008, A1).

4. That said, in the empirical literature “localization” and “functional specialization” are
often used interchangeably. Here, and in my discussion of “brain area” below, I aim to
clarify concepts, not legislate usage. For similar reasons, I use “brain” and “cortical”
interchangeably: although “brain structure” includes obviously distinct brain parts such
as the amygdala or hippocampus, much localization research into “brain structure” seeks
functionally significant divisions in cortex, which are more precisely called “cortical struc-
tures” (see also Mundale 2002; Ward 2006, 62–63).

5. Friston (1997, 21) defines functional specialization as “the expression of stereotyped
patterns of neuronal activity in response to specific attributes of a stimulus, cognitive
processing, or motor behaviour by specialized cortical areas, subareas or neuronal pop-
ulations” and functional integration as “the interactions among specialized neuronal
populations and how these interactions depend upon the sensorimotor or cognitive
context.”



MULTIPLE REALIZATION 423

tionism) to identify focal neural contributions to cognition and an effort
(associated with integrationism) to identify systems-level models of their
interaction.

Many of the empirical results of localizationism are on vivid display
in the form of images of cross-sections of brains color coded to indicate
areas of cortical activity that have been correlated with specific cognitive
functions or processes.6 In fMRI, the images are the fruit of a complex
process that measures a net decrease in deoxyhemoglobin in a cortical
area or areas. This is taken to indicate an increase in blood flow to, hence
more neural activity in, that area, which in turn is taken to indicate the
relevance of the area or areas for performing the cognitive task(s) being
explored in the imaging experiment.7 The broad contours of an empirically
based MR debate emerge from some basic features of the structure-func-
tion mappings represented in these images.8

First, the images represent two distinct mapping projects: neuroana-
tomy and cognitive (or functional) neuroanatomy. Both projects employ
the term “brain area” (or “cortical area,” “structure,” “region”) to pick
out their results but use different criteria to individuate areas. In neu-

6. Functional magnetic resonance imaging (fMRI) is popular because it is noninvasive
and has relatively high spatial resolution of the cortical surface. Positron emission to-
mography (PET) requires injecting subjects with radioactive tracers. Other noninvasive
technologies include electroencephalography, magnetoencephalography, transcranial
magnetic stimulation, transcranial electrical stimulation, and diffusion (tensor or spec-
trum) imaging. Invasive procedures used on humans include direct cortical stimulation
during neurosurgery and implantation of electrodes to locate foci of epileptic seizures;
single-cell recordings and induced lesions are limited to nonhuman animals. Since different
technologies have different degrees of spatial or temporal resolution, researchers increas-
ingly use results from more than one method when developing experiments and inter-
preting data. See Ward (2006) for accessible explanations of technologies and Savoy
(2001) for extended critical discussion of the benefits and drawbacks of each.

7. Although PET and fMRI signals are both hemodynamic, PET measures changes in
blood flow while fMRI measures blood oxygenation levels (and changes in blood flow
only indirectly). Both measure such changes at a scale of millions of neurons within
volumes of a few cubic millimeters, with signals sampled every few minutes (PET) or
seconds (fMRI) from many (e.g., 100,000) cortical positions (voxels). Note that much
cognitive processing, which occurs at speeds of milliseconds, is invisible even at sampling
rates of every few seconds. Attwell and Iadecola (2002) suggest the fMRI signal reflects
postsynaptic neurotransmitter activity, not (as usually assumed) energy use in presynaptic
terminals or glia.

8. Throughout this article, the term “structure-function mapping” is intended to be neu-
tral regarding the differences in the direction of inference (function to structure or structure
to function) that leads to a given mapping in a particular experimental paradigm. In the
empirical literature in general, reference to “one-to-many” and “one-to-one” mappings
are often and easily disambiguated in context, although only the former phrase threatens
any genuine confusion.



424 CARRIE FIGDOR

roanatomy, brain areas are individuated by distinctions in cellular mech-
anisms, cytoarchitecture, morphology, myelination, axonal projections
and connectivity, and neurophysiology, without essential reference to an
area’s possible role in supporting cognition, even if such a role motivates
the mapping effort. For example, Korbinian Brodmann’s (1909/1914) map
of 47 human cortical areas, still used as a reference point in both mapping
projects, was based on purely anatomical criteria (differences in cell types
and their distribution, or cytoarchitecture). Similarly, Hagmann et al.
(2008) propose a neuroanatomical map of major axonal connections in
human cortex on the basis of anatomical data from diffusion imaging,
which traces the diffusion of water through brain tissue to reveal the
orientation of axon fibers. Neuroanatomical maps contain what I will call
“anatomical areas,” which have names like “inferotemporal cortex” or
“Brodmann’s area 44.”

In cognitive neuroanatomy, brain areas are individuated using anatom-
ical and cognitive-functional criteria. The resulting areas are actually
structure-function mappings, and cognitive-neuroanatomical area names
pick out these mappings. For example, the functions of “early” vision—
edge, orientation, contrast and brightness detection—are not mapped to
V1; they are mapped to the medial calcarine sulcus in humans, and this
mapping is called V1. Similarly, a pioneering cognitive neuroanatomical
study by Felleman and Van Essen (1991b) distinguished 32 visual areas
in macaque cortex on the basis of connectivity, architectonics, topographic
organization of the visual field (retinotopy), distinct receptive fields of
neurons, lesion and stimulation studies, and prior studies that distin-
guished visual cortical areas using similarly varied criteria. This suite of
criteria is widely used in vision research, and similar multiple criteria are
used for individuating neurocognitive areas in general.9 Cognitive neu-
roanatomical maps contain what I will call “cognitive areas,” which have
names like “V1” or “Broca’s area.”10

Second, hypothesized structure-function mappings fix the reference, not
the meaning, of cognitive-area names. That is why it is not a conceptual

9. Major individuative criteria for visual areas include (1) cyto- and myeloarchitecture,
(2) connectivity, (3) retinotopic organization, and (4) cognitive function, as revealed by
single-cell recordings, lesion studies, and neuroimaging analyses (Orban, Van Essen, and
Vanduffel 2004; see also Felleman and Van Essen 1991a, 5). Retinotopic organization
(retinotopy) refers to the layout of cortical neurons processing visual information relative
to the layout of the input on retinal cells (e.g., points close in space on the retina are
close in space in V1). Keeley’s (2002) multicriterial analysis for individuating the senses
may be considered a special case of cognitive-area individuation methods.

10. Henson (2005) distinguishes “areas” (anatomical divisions, e.g., Brodmann’s) from
“regions” (functionally significant anatomical divisions); Phillips et al. (1984) use “cortical
area” as I use “cognitive area.”



MULTIPLE REALIZATION 425

truth that vision is the function of visual cortex or conceptually incoherent
for auditory information to be processed in visual cortex (Von Melchner,
Pallas, and Sur 2000; Burton 2003). Either or both elements in a hy-
pothesized cognitive area may be modified following further research. For
example, Broca (1861/1960/2001) hypothesized that speech production
was located at the superior temporal gyrus at approximately Brodmann’s
areas 44 and 45. This mapping is called Broca’s area. However, speech
production is now thought to be in more distributed prefrontal areas, and
Broca’s area also appears to play a role in processing natural language
syntax, musical syntax, perception of rhythmic motion, imaging move-
ment trajectories, and conducting local visuospatial searches (Marshall
and Fink 2003; Grodzinsky and Santi 2008). Such modifications of named
structure-function mappings exemplify difficult problems in the theory of
reference (Field 1973)—in particular, the issue of when a term refers to
the same entity that has changed over time or whether it has changed its
reference. However, since an empirical MR debate at this time involves
examining the results of rapidly developing disciplines, the possibility of
at least some referential indeterminacy must be expected and tolerated
by both sides.

Third, functions may be mapped to anatomical networks as well as
areas, but the term “network” is also ambiguous (Henson 2005, 215–16).
A weak cognitive network is a set of coactivated anatomical or cognitive
areas that subserve a cognitive function. A difference in component areas
suffices for a difference in a weak network.11 A strong cognitive network
is a set of coactivated anatomical or cognitive areas that exhibit effective
connectivity, or activity- and time-dependent influences on processing be-
tween areas. A difference in the type of interaction between component
areas suffices for a difference in strong networks. Networks of either type
can share components, but the contribution of an area to subserving a
function may differ when it participates in different networks.12 Also,

11. Thus, Kosslyn (1999, 1284) notes that it is often misleading when researchers talk
of brain circuits: “in most studies all that is revealed are a set of activated (and/or
deactivated) areas, with no information about the flow of information between the areas.
Thus, what we are seeing are the footprints of components of the functional architecture
that are evoked during the task, but we do not see a specific circuit.”

12. For example, suppose each of several areas makes a simple cognitive contribution
to the performance of object recognition (Kosslyn et al. 1994). It does not follow that
each area is unipotential, or specialized to perform one simple function. Activity in the
component areas may shift in response to different conditions of processing or to tem-
porary or permanent damage, and there may be considerable idiosyncrasy in individuals
or across species regarding which areas realize which simple functions. In such cases,
simple functions could be realized in more than one area within and across species, even
though the function they subserve as a whole remains the same. These cases provide
possibilities for degeneracy, discussed below.



426 CARRIE FIGDOR

component areas may not individually suffice for a cognitive function,
but if at least one does—that is, if a component is a cognitive area and
not just an anatomical area—any cognitive network in which it partici-
pates would constitute an additional cognitive-area layer. In principle
there may be many such layers.

Outside of extensively studied peripheral cognitive areas such as visual,
sensorimotor, auditory, and motor cortex, few results of localizationism
to date are entirely uncontroversial. Researchers are acutely aware that
to implicate an anatomical region in the performance of a function is not
to localize that function in that region. Cognitive neuroanatomy is thus
fertile ground as the basis of an empirical MR debate. As I will argue
below, however, such a debate is not adequately framed merely by re-
stricting the relevant evidence to that gleaned from cognitive neuroscience.
Inter alia, it requires an appropriately stated MR hypothesis. As it hap-
pens, biology provides us with one.

3. MR as Degeneracy in Cognitive Neuroanatomy. The need for an MR
hypothesis suited to an empirically based debate is motivated by the in-
adequacy in that context of the various theses regarding multiple real-
izability that can be derived from the philosophical literature. These in-
clude:

Weak MR. At least some creatures that are not exactly like us in
their physical composition can be conscious.
SETI (search for extraterrestrial intelligence) MR. Some creatures
that are significantly different from us in their physical composition
can be conscious.
Standard MR. Systems of indefinitely (perhaps infinitely) many phys-
ical compositions can be conscious.
Radical MR. Any (every) suitably organized system, regardless of its
physical composition, can be conscious (Polger 2004, 6).13

None of these hypotheses is appropriate in a debate about MR in existing
biological organisms. Implicitly, all reflect their origin in an intuition-
based debate about multiple realizability that included extraterrestrial
beings, computers, advanced robots, and other hypothetical possibilities
as well as evolved biological animals. MR, like its main rival the identity
theory, is fundamentally a claim about the relation between mental states
and their physical substrates. It is not about the scope of this relation
(even when restricted to evolved biological creatures). So pace Weak,

13. Polger focuses on consciousness, but these theses can be suitably reformulated for
any mental state. Note that Polger uses “MR” for multiple realizability, not multiple
realization (see n. 1).



MULTIPLE REALIZATION 427

SETI, and Standard MR, it is not essentially about the number of different
kinds of minded creatures or the potential number of different physical
substrates for which this relation may hold. The issue of differences be-
tween creatures is independent of the issue of differences between physical
substrates, and, empirically speaking, MR may turn out to be true within
one biological species only. Nor do qualitative differences (between crea-
tures or cognitive systems) have a place in an empirical debate: “not
exactly like” and “significantly different” are not subject to empirical tests
unless quantified, and “indefinitely (perhaps infinitely) many” is not sub-
ject to empirical tests at all. Radical MR suffers somewhat from the
qualitative nature of what counts as “suitably” organized, but its main
problem is that it is consistent with the possibility of one suitably organized
substrate for each mental state and so cannot capture what is essential
to MR. Finally, all four theses depend on an empirically irrelevant dis-
tinction between composition and structure (in the intended sense of “ar-
rangement”) in the individuation of physical realizers. As noted above,
anatomical areas are individuated on the basis of multiple criteria that
include both: cytoarchitecture is a type of composition; connectivity is a
type of structure.14

Instead, the relevant sciences can provide us with an appropriately
stated hypothesis. In biology, degeneracy is defined as the ability of ele-
ments that are structurally different to perform the same function or yield
the same output (Edelman and Gally 2001).15 Thus broadly defined, de-
generacy is found at all levels of biological organization, from the mo-
lecular, cellular, and genetic levels up to the level of organism: different
nucleotide sequences encoding the same polypeptide, different antibodies
binding the same antigen, different patterns of muscular contraction yield-
ing the same movement, and different encodings of the same message. As
a biological hypothesis, degeneracy has been posited to explain a number
of studies of biological organisms, from yeast to humans, in which striking
structural differences at various suborganism levels appear to have little
or no organism-level effects.16 Edelman and Gally argue that degeneracy

14. I return to the issue of which physical differences matter for MR below.

15. Tononi, Sporns, and Edelman (1999, 3257) note that the term “degeneracy” is taken
from immunology, where it refers to the ability of different antibodies to bind to the
same antigen.

16. See Edelman and Gally (2001) and references therein for detail on the following
studies. In “knock-out” mice, in up to 30% of cases there is little or no phenotypic
difference in mice that lack the genes to produce myoglobin, tenascin C, vimentin, and
other important proteins. In yeast, systematic screening of single-gene deletions at more
than 500 gene loci shows that fewer than half of the cultures had any quantitative growth
defects. In Drosophila, when either the gene for fasciclin (a cell-adhesion protein on the
surface of Drosophila neurons) or the gene for the cytoplasmic Abelson tyrosine kinase



428 CARRIE FIGDOR

is a prerequisite of natural selection, which requires genetic dissimilarity
in a population to operate but must avoid the likely lethality of most
mutations if individual genes are wholly and uniquely responsible for
phenotypic traits. At the genetic level, a natural solution is overlapping
networks of unrelated genes that can, given appropriate conditions for
gene expression, produce the same outcome.

The biological concept of degeneracy has been appropriated and de-
veloped by cognitive neuroanatomists as an alternative hypothesis to one-
to-one mappings (Price and Friston 2002; Friston and Price 2003; Nop-
peney, Friston, and Price 2004). In cognitive neuroanatomy, degeneracy
is the claim that, for a given cognitive function F, there is more than one
nonisomorphic (nonidentical) structural element that can subserve F, ei-
ther within an individual at a time, across individuals, or within an in-
dividual across times.17 In an empirically based MR debate centered on
neuroscience, MR is just degeneracy in cognitive neuroanatomy, or, more
precisely, since there are cases of degeneracy that do not or may not count
as MRs, MRs are special, perhaps paradigmatic, cases of degeneracy. In
the rest of this section, I will introduce this concept as it is used in cognitive
neuroscience, before turning to issues in the metaphysics of realization
and MR that bear on this conceptual assimilation.

Direct empirical motivation for hypothesizing degeneracy in cognitive
neuroanatomy stems in part from long-standing anomalies for localiza-
tionism based on lesion studies. These problems include differences in
deficits with similarly located lesions (and vice versa) and restitution of
function after damage (a form of plasticity). Such anomalies are what led
Lashley to propose his theories of equipotentiality and mass action

is completely deleted, there are no gross abnormalities in nervous system development,
even though these proteins have no obvious structural or functional similarity, but major
defects result when both are deleted. In humans, subjects who had exhibited no psycho-
logical abnormalities were found in fMRI scans to lack the corpus callosum that normally
connects the two cerebral hemispheres (although subsequent detailed psychological testing
did uncover subtle abnormalities in their functioning).

17. I take this to include degeneracy in comparative cognitive neuroanatomy (i.e., across
species): degenerate functions may be unique to humans (e.g., reading), human but not
uniquely so (e.g., motion detection), or (in principle) nonhuman (e.g., echolocation). The
researchers cited in the text do not mention cross-species possibilities, presumably since
the goal of their research, like that of localizationism in general, is to explain human
cognition. (I discuss animal models in sec. 4.) Also, if degeneracy’s requirement of more
than one structure seems weak—e.g., compared to the demand for indefinitely (perhaps
infinitely) many realizers in Standard MR—recall that Kim’s (1993) influential discussion
of jadeite/nephrite has long been accepted as sufficient both to characterize multiple
realizability and to raise the theoretical issues that have dominated the philosophical
debate and threatened the nonreductive physicalism that multiple realizability (or reali-
zation) is used to support (e.g., Fodor 2000).



MULTIPLE REALIZATION 429

(whereby lesion size, not location, determined the deficit) in the first
place.18

A major motivation, however, has been anomalous results from the
imaging studies that have come to dominate cognitive neuroscientific re-
search in recent decades.19 These studies frequently show areas of acti-
vation that differ across subjects (or within a subject at different times)
within the same experimental paradigm (see fig. 1). These differences,
which are usually ignored as “noise” or “random error,” are not simply
a reflection of the inevitable differences between individual brains in the
precise location of gyri and sulci. Such anatomical idiosyncrasies are usu-
ally standardized by fitting individual imaging results to a common brain
template (e.g., the Talairach and Tournoux [1988] atlas).20 In some cases,
the imaging results replicate the earlier anomalies based on lesion data.

18. Phillips et al. (1984, 328–33) also trace Goltz’s resistance to localization to the res-
titution of function and generalized (rather than specific) impairments after lesions. (Goltz
used dogs in his lesion studies.)

19. Lesion and imaging studies are broadly complementary. In imaging, we manipulate
function in order to infer (ideally) that certain cortical structures are sufficient for a
function. In lesion studies, we manipulate cortical or subcortical structure (or it gets
“manipulated” by stroke, etc.) in order to infer (ideally) that certain structures are nec-
essary for a function. However, even in the best of cases (e.g., no intersubject or intertrial
variation in imaging studies, no difference in functional deficit in lesion studies), many
alternative interpretations of the results of manipulation are available. For example, in
fMRI, the blood-oxygen-level-dependent signal does not distinguish excitation from in-
hibition (Ward 2006) or neural codes that involve timing and synchronization. Thus,
images can indicate areas that are active not because of what they are doing but because
they are being prevented from doing something or are trying to do something but not
succeeding or simply because a task requires more of a general resource (e.g., attention).
Inference to the role of a structure from lesion studies is also difficult; to borrow Gregory’s
(1961) famous analogy: if removing a resistor from a radio causes it to emit strange
howls, it would be incorrect to ascribe to the resistor the function of howl suppression.
There may be no deficit if the lesioned area’s function is subserved by undamaged systems
that modify their operations or by newly created components (these are cases of degen-
eracy, explained in the text). Also, a lesioned area may facilitate performance without
being necessary for it. Other concerns include whether data from single cases or groups
should be used in lesion studies (Caramazza 1986) and whether it is legitimate to “stan-
dardize” brains (e.g., using the Talairach and Tournoux [1988] brain template) or average
imaging results across individuals (Savoy 2001). Of course, not all localization claims are
unreliable; e.g., we know the hippocampus plays an essential role in memory formation,
and the occipital cortex processes visual information. The point is that many inferences
to structure-function mappings (in either direction) from lesion or imaging data are not
highly confirmed at this time.

20. The idiosyncrasies can be significant. On the basis of a survey of abstracts for the
2000 International Conference on the Functional Mapping of the Human Brain, Savoy
(2001, 26) notes that irregularities of cortical size, shape, and foldings across subjects are
a serious concern for cross-subject image comparisons.



430 CARRIE FIGDOR

Figure 1. Normals based on previously published studies involving a semantic task
(e.g., matching words and pictures). a, Results averaged over 12 subjects, without
distinguishing between areas found in one subject or more than one; b, results common
to all 12 subjects only; c, data from subject 1 only; d, data from subject 2 only. Arrows
indicate areas of activation in individuals that do not appear in averaged images (a
and b). Source: Price and Friston (2002). Color version available as an online en-
hancement.

For example, double dissociations have been found between Broca’s
aphasics and neural activity in Broca’s area and between Wernicke’s
aphasics and neural activity in Wernicke’s area—that is, there are patients
with lesions in Broca’s area who do not have Broca’s aphasia and patients
with Broca’s aphasia who do not have lesions in Broca’s area (ditto,



MULTIPLE REALIZATION 431

mutatis mutandis, for Wernicke’s aphasia and area; Dronkers, Redfern,
and Knight 2000).21

In other cases, new anomalies arise from combining imaging and lesion
data. Price and Friston (2002) examined results from fMRI studies of
normal subjects performing semantic-processing tasks (e.g., picture nam-
ing) and fMRI studies of lesion patients capable of performing the same
tasks at normal levels. The patients’ lesions were plotted to the same
standard brain plan used for normal subjects. Although the lesions were
located in the areas activated in the normal subjects’ performance of the
tasks, the fMRI data from the patients showed entirely distinct areas
activated during their performance (see fig. 2). They conclude that none
of the cortical areas activated in the normals are in fact necessary for the
tasks.

The degeneracy hypothesis can explain these puzzling processing dif-
ferences and data showing undeniable cross-subject specialization of func-
tion in cortex that have made Lashley’s hypotheses untenable. Comple-
mentary to degeneracy is the concept of pluripotentiality: when a single
structure subserves more than one function. Pluripotential structures fill
a spectrum of degrees of functional specialization between unipotentiality
and equipotentiality. There can be degeneracy without pluripotentiality
(if two unipotential structures subserve the same function) and pluripo-
tentiality without degeneracy. But degeneracy and pluripotentiality are
considered closely associated for evolutionary reasons. Degenerate struc-
tures will typically exhibit the variability that selection processes require
and so will not be duplicates. As a result, many degenerate structures will
be such that they can produce the same output in one processing context
but different outputs in another. For example, a given structure may
subserve function F when activated within one network and function G
when activated within another, or if there are two structures that normally
subserve F and G, respectively, and the G-structure is damaged or inhib-
ited, the F-structure must be pluripotential if it can step in to subserve
G. Thus, for any cognitive function F, degeneracy can occur if there is
(i) more than one unipotential nonduplicate area (or network), each suf-

21. Classic double dissociations are cases in neuropsychology when one (lesioned) subject
or group (A) demonstrates one type of cognitive loss or impairment (F1) while another
capacity (F2) is left intact or relatively so, and a second subject or group (B) demonstrates
the loss or impairment of F2 while F1 is intact or relatively so. A classic case is phonological
dyslexia and surface dyslexia: subjects with surface dyslexia have trouble reading irregular
words (“chef”) but no (or less) difficulty reading nonwords (“mave”), while subjects with
phonological dyslexia exhibit the opposite difficulty. The concept has been extended to
include cognitive areas, even though there is significant controversy as to what legitimately
can be inferred about cognitive organization from double dissociations (Shallice 1988;
Farah 1994; Plaut 1995; Coltheart and Davies 2003; Dunn and Kirsner 2003).



F
ig

u
re

2.
T

op
ro

w
,

si
te

,
ty

p
e,

an
d

ex
te

n
t

o
f

le
si

o
n

s
(m

ea
su

re
d

in
vo

xe
l-

b
as

ed
m

o
rp

h
o

lo
gy

)
in

ar
ea

s
ac

ti
va

te
d

in
n

o
rm

al
s

d
u

ri
n

g
th

e
se

m
an

ti
c

p
ar

ad
ig

m
u

se
d

fo
r

th
e

im
ag

es
in

fi
gu

re
1.

P
er

ce
n

ta
ge

s
sh

o
w

p
er

fo
rm

an
ce

b
y

th
e

p
at

ie
n

ts
o

n
th

es
e

se
m

an
ti

c
ta

sk
s.

N
o

n
e

o
f

th
e

ar
ea

s
ac

ti
va

te
d

in
n

o
rm

al
s

ap
p

ea
r

n
ec

es
sa

ry
.

S
o

u
rc

e:
P

ri
ce

an
d

F
ri

st
o

n
(2

00
2,

41
9)

.
C

o
lo

r
ve

rs
io

n
av

ai
la

b
le

as
an

o
n

li
n

e
en

h
an

ce
m

en
t.



MULTIPLE REALIZATION 433

ficient for F; (ii) more than one pluripotential nonduplicate area (or net-
work), each of which suffices for F in certain contexts; or (iii) more than
one combination of unipotential or pluripotential nonduplicate areas (or
networks) that in combination suffice for F. This last possibility includes
degenerate networks that partially overlap in their component areas. Over-
lapping networks may be considered cases of only partial MR if each
common area makes the same contribution to each network but not if
their contributions differ between networks.

A measure of the degree of degeneracy of a particular structure-function
relationship is given by the number of sufficient wholly disjoint elements
(areas or networks) that can produce the same output. This number—
the order of degeneracy—can be determined behaviorally by the minimum
number of anatomical areas that must be lesioned before a behavioral
deficit can be observed.22 The order number is not the same as the number
of sufficient systems that may subserve the same function since lesioning
one area common to more than one partially overlapping degenerate
network can result in a cognitive deficit and would still be a case of first-
order degeneracy. In second-order degeneracy, there are at least two wholly
disjoint structures subserving the same function. In short, the same order
number (the same degree of degeneracy) may reflect different numbers of
realizations.

Within this general framework, not all cases of degeneracy do or may
count as MRs. For example, when degenerate systems that perform the
same function are coactivated, they are said to be functioning redundantly
(i.e., inefficiently), even if they perform the function in distinct ways. Since
degenerate systems can be latent, such that only the prepotent system
operates unless deactivated, redundant functioning entails degeneracy but
not vice versa. However, duplicate anatomical areas subserving the same
function, whether or not they function redundantly, would no more count
as MRs than the kidneys, which are both anatomically and functionally
redundant.

However, further differences may stem from the selection of the relevant
functions and structural elements for mapping. Degeneracy, like MR, is
relative to the levels of psychological function and biological organization
in a given mapping. It is an open question whether there is a right or
optimal level (or range) for mapping functions and structures, even within

22. Note that this measure can only determine the order of degeneracy of a mapping
within an individual.



434 CARRIE FIGDOR

the limits of cognitive neuroscience.23 In cognitive neuroanatomy, the levels
of function are implicitly determined by the task analysis and the psy-
chological and neurophysiological measurements being employed in a
given study. Currently, behavioral or behavior-based measures—for ex-
ample, direct responses to perceptual stimuli, performance on standard
neuropsychological test batteries, double dissociations—are used to de-
compose and characterize functions.24 Such measures have yielded a level
of cognitive “strategies” (“routes,” “pathways”) in addition to functions
or outputs because of the discovery or hypothesizing of more than one
way of generating the same output. Outstanding among dual-route models
is Ungerleider and Mishkin’s (1982) discovery of two cortical routes for
visual processing after initial processing in visual cortex. Similarly, reading
may be performed via a lexico-semantic route or an orthographic-pho-
nological route. However, while dual-route models are considered cases
of degeneracy, they may not be MRs if cognitive functions are individ-
uated by routes (e.g., reading by the semantic route vs. reading by the
phonological route). A conservative policy would hold that there is MR
only if, for any strategy that yields F, the same strategy (or the same step
in at least one strategy) is subserved by different structures.25

23. Price and Friston (2002, 418) and Noppeney et al. (2004, 434–35) recognize that
degeneracy is sensitive to the levels of function and structure. The fundamental problem
is that MR (or the identity theory, for that matter) becomes trivial if psychological
functions at any level of abstraction can be mapped to physical substrates at any level
of abstraction in realization-relevant mappings. Henson (2005, 217–19), also noting the
need to identify “the appropriate level of functional/structural abstraction” for mappings,
suggests defining the appropriate level as that at which one assumes a priori that there
is a one-to-one mapping; experiments would then test whether this hypothesis is correct
(see also n. 32 below). This proposal has the merit of voiding Bechtel and Mundale’s
(1999) charge that philosophers made MR seem plausible by mapping coarse-grained
psychological types to fine-grained neural types based on their intuitions.

24. This dependence on behavioral studies for labeling cognitive processes has been
challenged by Price and Friston (2005). However, they suggest augmenting such data,
not eliminating them, and explicitly leave open the possibility of degeneracy, even in an
improved naming system. My point here is to illustrate one of the ways in which degen-
eracy and MR can come apart.

25. This may well be too conservative. Shapiro (2000, 644; 2004) has argued that what
matters for realizer individuation are differences in causally relevant (or R) properties—
“in properties that make a difference to how they contribute to the capacity under
investigation.” Whether or not this criterion rules out many intuition-based examples of
MR, it implies that reading is multiply realized if the dual-route model is correct since
the causally relevant properties used in cognitive neuroscientific individuation include
connectivity, and connectivity between two routes obviously differs. Dual-strategy models
have also been proposed for (e.g.) object constancy and face constancy (Ward 2006, 110),
mental rotation (Kosslyn et al. 1998), and verbal response selection (Raichle et al. 1994).
The terms “process,” “strategy,” and “route” are often used interchangeably (and may
be referred to as “functions”).



MULTIPLE REALIZATION 435

On the neuroanatomical side, single cells, neuronal populations, ana-
tomical areas, or anatomical networks are among the “structural ele-
ments” that may appear in degenerate mappings. This list also may be
too inclusive for MR. For example, Edelman and Gally (2001, 13765)
consider cognitive neuroanatomy very likely to be highly degenerate, but
they appear to individuate neural structures such that a single difference
in connectivity suffices for a distinct structure.26 But presumably MR could
occur within an individual area if it contains multiple functionally spe-
cialized neuronal populations that can perform the same function or if
cells within the same population switch between different encodings (“re-
map”) at short timescales (Johnson et al. 2009). A conservative policy
would count only properties of anatomical areas and networks as (possibly
degenerate) realizers.

However, even a focus on anatomical areas or networks, or more gen-
erally to structural units well above molecular mechanisms, is neither
arbitrary nor restrictive. Obviously, if specific cognitive phenomena are
produced neither by the whole brain’s operation nor by individual neu-
rons, there must be intermediate units to which functions can be mapped.
Anatomical areas fit this basic requirement. More important, the multiple
criteria used to individuate anatomical areas are drawn from different
levels of neuroscience, not by cortical analogues of latitude and longitude
(see n. 9 and associated text). Even if cellular and molecular neuroscience
are the central disciplines within or of neuroscience (Bickle 2003), cognitive
neuroscience incorporates the lower-level findings: mechanisms are al-
ready among the criteria for individuating anatomical areas, and single
areas can contain multiple functionally specialized units. This is why it
is at best an unfounded assumption to think of anatomical areas as a
mere structural stopgap until we find out more about (e.g.) mechanisms,
such that these lower-level discoveries will settle the debate.27 There is no

26. In this light of Edelman and Gally’s (2001) very fine-grained individuation, it is ironic
that Bechtel and Mundale (1999, 178) diagnose the apparent plausibility of MR as the
result of a “methodological error” by philosophers through “mismatching” coarsely in-
dividuated functions with finely individuated realizers based on their intuitions; they
conclude that “when a common grain size is insisted on, as it is in scientific practice, the
plausibility of multiple realizability evaporates.”

27. Bickle (2003) employs detailed evidence from cellular and molecular neuroscience
mainly to make a case for reduction. His direct attack against MR (131–61) rests on
empirical findings that the consolidation of memory-like capacities (including sensitiza-
tion—the heightened responsiveness to noxious stimuli by an organism’s defenses—and
classical conditioning) in the sea slug (Aplysia californica) and fruit flies (Drosophila
melanogaster) is controlled by the same cellular and molecular mechanisms, plus the claim
that evolution conserves molecular mechanisms. This may all be true, but it does not
show that MR is false unless (inter alia) these mechanisms are exclusive and play the
same role in higher animals. Even if evolution conserves molecular mechanisms, if Ed-



436 CARRIE FIGDOR

empirical reason to think the list of individuative criteria of anatomical
areas is closed or that it will be pruned to a single (lower-level) criterion,
effectively eliminating the relevance of anatomical areas in structure-func-
tion mappings. Neuroscientific practice suggests that future discoveries
will lead to a progressively more articulated taxonomy of anatomical
areas, not abandonment of these structural units. The increasingly fine-
grained individuation of cognitive areas within visual cortex illustrates
and foreshadows this methodology.28 There is also no empirical reason
why distinctions at even lower levels, for example, in basic metabolic
processes, might not play a role ( pace Bechtel 2006, 498). If metabolic
differences (or subatomic particle differences, for that matter) can be
manipulated to make a measurable cognitive-functional difference in the
relevant behavioral tests, the most likely result is that such processes will
be added to the list of individuative criteria and weighed against the other
items when conflicts arise.

Thus, while it is likely that degeneracy and MR in the cognitive neu-
roscientific context do not exactly coincide, I have largely left open the
extent to which they may come apart. But trying to pry them apart now
would essentially require using individuation criteria that either are not
employed in the relevant sciences or, if they are, are among multiple
individuation criteria that span biological levels. Multiple criteria imply
that it is an open empirical possibility that when two neural structures
count as the same by one criterion, they may count as different based on
other criteria. Which of these physical differences will triumph in indi-
viduation in particular cases of criterial conflict cannot be determined a
priori. (In sec. 4, I discuss a case of conflict of this sort.) In short, at least

elman and Gally (2001) are right, it will conserve degenerate molecular mechanisms.
Moreover, Bickle assumes the behavior of these relatively simple creatures essentially
requires psychological explanation—a claim that is controversial, even for chimpanzees
and other higher primates (Andrews 2008). So even if sea slugs and fruit flies have a
mechanism that explains simple behaviors that are broadly analogous to what creatures
with memory may do in broadly analogous situations, and even if we also have this
mechanism, nothing follows immediately about MR.

28. Savoy (2001, 10–12), displaying a 1957 map of cognitive areas in human cortex based
on data from lesion studies and cortical stimulation during neurosurgery, notes that the
map is “remarkably accurate” and that new technologies have enabled us to “refine” the
map and add information about subcortical structures. Chemistry provides an indepen-
dent scientific precedent (Le Poidevin 2000). As new behavioral differences between mol-
ecules were discovered, chemists individuated chemical kinds progressively more finely
by adding criteria—moving from individuation by proportions of each kind of atom
(expressed in the molecular formula) to individuation by proportions and arrangement
and kinds of bonds (expressed in the structural formula) to individuation by proportions,
arrangement, bonds, and orientation in space (expressed in conventional notations rep-
resenting three-dimensional arrangements of atomic groups).



MULTIPLE REALIZATION 437

some degenerate systems are very likely to count as cases of MR, however
the issue of realizer individuation is settled.

4. Cognitive Neuroscience and Arguments against MR. The preceding
sketch of degeneracy in cognitive neuroanatomy suffices to show that MR
is a viable empirical hypothesis within cognitive neuroscience. In this
section, I respond to arguments based on data from cognitive neuroscience
that conclude that MR is empirically implausible. First, I show how a
popular strategy for using empirical data in the MR debate goes awry.
This strategy involves drawing an implication from MR regarding the
autonomy (in some sense) of psychology from neuroscience and then using
neuroscientific data to show that this implication is false. Both Bechtel
and Mundale (1999) and Shapiro (2004) develop ANA (Arguments from
Nonautonomy, as I call them), although for space reasons I can only
discuss Bechtel and Mundale’s version here.29 I then turn to a second
argument by Bechtel and Mundale that cognitive neuroscientific meth-
odologies show that MR is implausible. I show that they do not.

Bechtel and Mundale (1999) defend the following two claims, each of
which plays a central role in the empirically based arguments against MR
just sketched:

Claim 1. Neuroscientific information has been useful in psychology
(e.g., in guiding the decomposition and understanding of cognitive
systems).30

Claim 2. Cognitive researchers assume (implicitly) that MR is actually
false.31

I will be arguing that the truth of claim 1 does no harm to MR and that
claim 2 is false.

Claim 1 reflects Bechtel and Mundale’s (1999) main concern with the

29. Given his interpretations of MR and autonomy, Shapiro’s version (ANA Shap) may
be stated as follows: (1) MR claims that humanlike minds can be realized in very many
humanlike or nonhuman-like brains that might have evolved on earth consistently with
physical law. (2) If MR is true, we should be able to infer little or nothing about human
brains from facts about human psychology. (3) But we can infer facts about human brains
from facts about human psychology. (4) So MR is “much less obvious than philosophers
suppose” (2004, xiii).

30. “We have tried (i) to demonstrate that the claim that psychological states are multiple
[sic] realized has not been demonstrated, at least within animal life forms, (ii) to show
how denying MR allows fruitful use of neuroscience in guiding the decomposition and
understanding of cognitive systems, and (iii) to diagnose why multiple realizability has
been so widely accepted” (Bechtel and Mundale 1999, 204).

31. I provide textual evidence of this claim below.



438 CARRIE FIGDOR

“common corollary” of MR regarding autonomy, which appears as prem-
ise 2 in their Argument from Nonautonomy (ANAB&M):

1. MR claims that the same psychological state (process) can be re-
alized by different brain states.32

2. If MR is true, information about the brain should be of little or
no relevance to understanding psychological processes (cognitive
systems).33

3. But neuroscientific data have been useful in guiding the decom-
position and understanding of cognitive systems (psychological
processes; claim 1).

4. So MR is empirically implausible (or false).

This argument is either unsound or invalid depending on how premise 1
is interpreted. If premise 1 is understood as an expression of the hypothesis
of degeneracy in cognitive neuroanatomy, it does not imply the consequent
in premise 2. Degeneracy is cognitive neuroscience’s yes answer to the
question of whether more than one anatomical structure may subserve
the same cognitive function. It cannot rule out the use of whatever in-
formation cognitive neuroscientists want to use to determine whether that
answer is correct or imply that a hybrid subdiscipline like cognitive neu-
roanatomy should not be possible. So on this interpretation, premise 3 is
perfectly compatible with premise 1. (The implicit assumption in the con-
sequent of premise 2 that psychologists should not find neuroscientific
data useful is discussed below.)

If premise 1 is interpreted as any one of the intuition-based MR theses
stated in section 3 (which seems to be Bechtel and Mundale’s intent), the
argument involves an equivocation.34 Intuition-based MR (in any version)

32. “The claim of multiple realizability is the claim that the same psychological state
can be realized by different brain states. Thus, it is claimed that there is a many-to-
one mapping from brain states to psychological states” (Bechtel and Mundale 1999,
176).

33. “One common corollary of this rejection of the identity thesis [i.e., MR] is the
contention that information about the brain is of little or no relevance to understanding
psychological processes” (Bechtel and Mundale 1999, 176).

34. “The guiding assumption [in artificial intelligence] was that if mental activities
could be characterized in terms of operations in a system, then they should be able to
be implemented in different hardware or wetware, thereby providing alternative real-
izations. Taking this a step further, many philosophers became convinced that the same
mental activities could be realized in brains of aliens with radically different compo-
sition from ours. The upshot of these speculations about artificial and alien minds is
a metaphysical claim that mental processes are the operations themselves, and are not
identified with whatever biological or other substances realize them. For the most part,
we will have nothing to say about these speculative arguments, nor are we primarily
concerned with the metaphysical claim. Our primary concern, rather, is with the im-



MULTIPLE REALIZATION 439

and its “common corollary” (assuming that premise 2 correctly states that
corollary) were formulated in a context that included alien and artificial
cognitive systems—that is, “silicon-based extraterrestrials, computers, an-
droids, robots, and other brainless science fictional beings,” as Bickle
(1998, 114) puts it.35 So if premises 1 and 2 are interpreted in their original
forms, they are true (if true at all) of the cognitive systems of these beings
as well as those of existing biological organisms. But premise 3 is only
about the cognitive systems of the latter. In effect, the consequents of
premise 2 and premise 3 are related as genus to species; put otherwise,
the term “brain” in premises 1 and 2 means roughly “physical substrate,”
while in premise 3 “cognitive systems” means “evolved-biological brain.”
Neuroscientific information may be helpful in understanding biological-
brain-based cognitive systems, but this is compatible with the claim that
this information will not be helpful in understanding cognitive systems
in general. So on this interpretation, all the premises can be true, but the
conclusion does not follow.

It is worth noting that the empirical studies Bechtel and Mundale (1999)
use to support premise 3 do not actually do so. These include Ungerleider
and Mishkin’s (1982) influential research showing two distinct neural
pathways for post-V1 visual information processing in the macaque. This
research led to the discovery of similar neural routes in post-V1 human
vision. The error is to interpret this research as showing that the discovery
of two cortical pathways (neuroscience) suggested that vision splits into
two functional streams (psychology). Ungerleider and Mishkin used psy-
chological (visual) as well as neural information to obtain their cognitive
neuroanatomical results. In other words, their research (and the other
studies Bechtel and Mundale cite, involving cell-level research in early
visual areas) is of a hybrid nature from the start. Otherwise they would
have had no reason to distinguish (or even ability to identify) these two

plication drawn from the multiple realizability argument that information about the
brain is of little or no relevance to understanding psychological processes” (Bechtel
and Mundale 1999, 176).

35. In the philosophical debate, MR was often taken to imply autonomy in the sense
of the denial of intertheoretic reduction (e.g., Fodor 1975, 2000). This sense of auton-
omy is compatible with as much theorizing using neuroscientific data as a psychologist
might want. The sense of autonomy in premise 2, in which psychological theorizing
is not constrained at all by developments in neuroscience, is far stronger. Kim (2006,
117) expresses this sense as follows: “Perhaps there might be non-carbon-based or non-
protein-based biological organisms with mentality, and we cannot a priori preclude the
possibility that nonbiological electromechanical systems, like the ‘intelligent’ robots
and androids in science fiction, might be capable of having beliefs, desires, and even
sensations. All this suggests an interesting feature of mental concepts: They seem to
carry no constraint on the actual physical-biological mechanisms that, in a given system,
realize or implement them.”



440 CARRIE FIGDOR

neural pathways from all the others. So if premise 3 is intended to make
a claim that goes beyond the fact that cognitive neuroscience exists, em-
pirical support for it would have to come from studies showing the es-
sential use of data from a nonhybrid subdiscipline of neuroscience in
psychological theorizing.36 As a matter of fact, however, cognitive psy-
chologists differ sharply regarding the utility, if any, of neuroscientific (in
particular, neuroimaging) data in psychological theorizing, quite inde-
pendently of their positions vis-à-vis MR. For example, Henson (2005;
discussed below) defends the use of fMRI data in experimental psychol-
ogy; Coltheart (2004) defends the ‘ultracognitivist’ view in cognitive neu-
ropsychology, which sees no use for data regarding where or how functions
are realized; and Shallice (1988, 213–14) grants only limited use of ana-
tomical data in cognitive neuropsychology.37

If ANAB&M shows anything, it is that we need a new concept of au-
tonomy appropriate to an empirical MR debate. Since degeneracy is a
claim about the relation between psychological and neural hierarchies,
this new concept might highlight the difference it makes to our ability to
make inferences from knowledge of one hierarchy to hypotheses about
the other if the relation between their parts is largely degenerate rather
than largely one to one. Obviously, if degeneracy is generally true of a
kind of cognitive system, its cognitive and neural hierarchies will not
mirror each other. That is, degeneracy is compatible with incommensurate
(nonisomorphic) decompositions of functions and structures, while one-
to-one mappings imply commensurate (isomorphic) decompositions. So
if degeneracy is largely true of a system, we cannot reliably infer from a
task analysis to a decomposition of the neural structure that realizes it
or vice versa. Conversely, if one-to-one mappings are largely true, these
inferences should be reliable. This is a major difference: one-to-one map-
pings imply that evolved cognitive systems possess a degree of efficiency
in realization that would be surprising in evolved structures if not engi-
neered ones. This implication of one-to-one mappings may be the em-

36. By “essential use,” I mean that the neuroscientific data must be useful because
they are neuroscientific; obviously, scientists can use information from anywhere in
their theorizing without implying anything about the relation between the source of
that information and their theory: Kekule’s dream of a snake eating its tail inspired
his model of benzene’s ring structure, but that does not make psychological or her-
petological information (qua psychological or herpetological) useful in chemistry.

37. See also Ward (2006, 79) on the historical “schism” within cognitive neuropsy-
chology. In computational neuroscience, which develops highly abstract implementa-
tions of functional models, Schwartz (1990, x–xii) summarizes a history of disagreement
between theoreticians and experimentalists over the relevance of anatomical or other
experimental (e.g., behavioral) data to neural net models.



MULTIPLE REALIZATION 441

pirical equivalent of Putnam’s (1975, 436) oft-cited claim that the identity
theory “is certainly an ambitious hypothesis.”

I turn now to claim 2 and Bechtel and Mundale’s (1999) methodology-
based case against MR. This claim, which is independent of claim 1 and
ANAB&M, is inconsistent with the fact that cognitive neuroscientists have
proposed the degeneracy hypothesis. It is not entirely clear how Bechtel
and Mundale intend claim 2 to be used to argue against MR, but a
reasonable reading would embed it in the following inductive “success”
argument, which might be called an Argument from Empirical Fruitful-
ness (AEF):

1. MR claims that the same psychological state can be realized by
different brain states.

2. Cognitive researchers assume (implicitly) that MR is actually false
(claim 2).

3. Their assuming its falsity has allowed the fruitful use of neuroscien-
tific data in guiding the decomposition and understanding of cog-
nitive systems (Bechtel and Mundale 1999, 204).

4. So (probably) MR is false (or implausible).

Whether Bechtel and Mundale endorse AEF or not, I will only discuss
their empirical support for premise 2, without which premise 3 does not
get off the ground. A related success argument by Henson, discussed
below, will also help show how AEF goes wrong.

Bechtel and Mundale claim (1999, 177) that the denial of MR is implicit
in some of the assumptions built into localizationists’ research method-
ologies. These methodologies are such that “(1) the appeal to function,
especially psychological function, is an essential part of both the [mapping]
project and its tools, and (2) the cartographic project itself is frequently
carried out comparatively—across species” (177–78).38 According to Bech-
tel and Mundale, the implicitly MR-denying assumptions behind 1 include

38. Bechtel and Mundale continue (1999, 177–78): “For multiple realization to be a
serious option, brain taxonomy would have to be carried out both independently of
psychological function, and without comparative evaluation across species.” The first
conjunct is ambiguous. If “brain taxonomy” refers to neuroanatomy, then it is carried
out independently of cognitive function (even if it is motivated by the desire to explain
cognition). If it refers to cognitive neuroanatomy, then it is not carried out indepen-
dently of cognitive function, and degeneracy cannot imply that it should not be. The
second conjunct is discussed in the text. I should note that while Bechtel and Mundale
clearly distinguish anatomical and cognitive areas (e.g., “areas delineated by gyri and
sulci” and “functionally significant areas”), as well as neurobiology and cognitive
neuroscience, they do not disambiguate their term “brain area.” Thus, their claim that
“brain areas” are individuated partly by the cognitive function they subserve may be
expressed unambiguously by saying that the individuation of cognitive areas involves
functional and anatomical criteria.



442 CARRIE FIGDOR

pure insertion in the experimental paradigm in imaging studies called cog-
nitive subtraction, while the implicitly MR-denying assumption behind 2
is that of homology. I will argue that these assumptions do not implicitly
deny MR and that their roles in specific experimental paradigms and in
cognitive neuroscience generally are compatible with the degeneracy hy-
pothesis. I will start with homology since that concept is more familiar.

According to Bechtel and Mundale (1999), the use of comparative data
in cognitive neuroscience relies on assuming cross-species anatomical com-
monalities: “One might think, at first glance, that the ability to make
comparisons across species actually depends upon multiple realizability.
In fact, it is the very similarity (or more precisely, homology) of brain
structures which permits us to generalize across certain species. So in this
latter respect, in the context of neuroscientific research, they are not mul-
tiply realized” (177–78). “Most neuroimaging to date is performed on
humans while the most detailed neuroanatomical and neurophysiological
work (using, e.g., single-cell recording) has been done on other species.
As a result, researchers often have to try to coordinate the imaging work
on humans with neuroanatomy from other primate species (especially the
macaque), and thus are assuming that cognitive functions are not differ-
ently realized in the two species” (190). The assumption of anatomical
commonalities across species, they argue, corresponds to the way psy-
chologists and philosophers ignore vast cognitive and behavioral differ-
ences when they classify (e.g.) fear, pain, or hunger as a single type of
psychological state across species: neuroscientists ignore vast physical dif-
ferences between species in order to classify anatomical areas as being of
the same cross-species type.

Bechtel and Mundale are undoubtedly correct about the importance of
comparative data in cognitive neuroscience. Brodmann’s maps of human
and other species’ brains were developed from preparations from dozens
of mammalian species. Contemporary researchers use cross-species ana-
tomical data as converging evidence for structure-function mappings (Fel-
leman and Van Essen 1987; Kanwisher, McDermott, and Chun 1997;
LeDoux 2000) and for human structural imaging (Hagmann et al. 2008).
Cognitive-functional imaging data from awake behaving animals (pri-
marily macaques) are increasingly being used as further converging evi-
dence for structure-function mappings (Vanduffel et al. 2001; Orban et
al. 2004).

But using comparative data on the basis of presumed homology is not
equivalent to typing anatomical areas as the same across species. First,
homologies—in this case, brain structures in distinct species that derive
from a structure possessed by a common ancestor—are hypothesized, not
assumed. What is assumed is the theory of evolution. Evolution justifies
and motivates the search for structural, neurophysiological, and cognitive-



MULTIPLE REALIZATION 443

functional homologies between different species.39 But, second, evolu-
tionary theory does not ipso facto commit researchers to the existence,
extent, or nature of specific homologies or to the role that this historical
relation may play in the individuation of structural, neurophysiological,
or cognitive types or (consequently) structure-function mappings.40 In
particular, it does not commit them to ignoring vast physical differences
between possible or actual homologues to classify them as a single cross-
species anatomical-area type.41 Tootell, Tsao, and Vanduffel (2003), who
study the macaque visual system, express a concern that justifies this lack
of commitment: “The level of accepted evolutionary similarity between
possibly homologous cortical regions is likely to decrease (not increase)
as we learn more. It is easy to assume that macaque area X is equivalent
to human region Y, if almost nothing is known about region Y. However,
further study will (by definition) reveal more detailed features, any of
which may differ across species” (3986). It is an open question whether
in any particular case the similarities will outweigh the known or expected
differences when individuating even homologous anatomical areas.42

39. Kim (2002) argues that only homoplasies—similar structures that developed in-
dependently in species with distinct evolutionary lineages, such as bird and bat wings—
count as relevant evidence for or against multiple realizability; Couch (2004) counters
that both homologies and homoplasies are relevant. Bechtel and Mundale (1999) refer
specifically to homologies, so I set homoplasies aside for the sake of argument.

40. Orban et al. (2004, 317) note: “Cortical areas in humans and macaques are con-
sidered homologous if they derive from areas present in a common primate ancestor.
For areas that existed in this common ancestor, the challenge is to identify the ho-
mologous areas in monkeys and humans despite whatever divergences have occurred
in structure, function and geographic location.” Functional homologies also are not
assumed. For example, it is a matter of debate whether human language derived from
a homologous capacity in a primate ancestor or arose de novo via the redeployment
of other capacities (Deacon 2004).

41. Trivially, any anatomical area of (e.g.) macaque and human brains can already be
classified under the cross-species types “is possessed by a primate,” “is derived from
a common ancestor,” or even “is smaller than a bread box,” but presumably these are
not the cross-species types localizationists qua localizationists are interested in. Their
interest is in properties that ground cognition (or, as Polger [2004, 249 n. 13] puts it,
those which “ground the regularities of psychological explanation”). It follows that
not every cross-species property, physical or otherwise, counts as a cross-species realizer,
on pain of trivializing the debate. This suggests that Bechtel and Mundale’s diagnosis
of the intuitive appeal of MR is inadequate: even if philosophers did mismatch tax-
onomic levels, thereby making MR appear more plausible than it is, Bechtel and
Mundale err in thinking that any one-to-one mapping affects the MR debate equally.

42. Thus, many if not all references to (e.g.) “primate visual cortex” are most plausibly
construed as references to species-specific cognitive areas that (i) are possessed by
primates and (ii) as a working hypothesis involve functions that are similar across
species.



444 CARRIE FIGDOR

Moreover, even granting for the sake of argument that a well-confirmed
homology between anatomical areas commits neuroscientists to classifying
these areas under the same cross-species type, it is a separate step to the
conclusion that the cognitive functions assigned to these anatomical areas
will be the same (let alone homologous). Vision research, which provides
some of the best empirical support we have for cross-species anatomical-
area types so far, shows that some structure-function mappings are iden-
tical across species and others are not. In the macaque, visual area V3 is
moderately motion and direction sensitive, while V3A is not; in the human,
the functions performed by these areas are reversed (Felleman and Van
Essen 1987; Tootell et al. 1997, 2003; Vanduffel et al. 2001; Orban et al.
2004; Vanduffel et al. [2002] discuss other macaque-human disanalogies).
So even anatomical areas that are homologous and (by assumption) cross-
species-type identical can be mapped to distinct functions.43

These prima facie cases of MR are based on the same methodologies
that result in prima facie one-to-one mappings across species (e.g., V1 in
macaques and humans may be an example). Nor should the V3/V3A case
be considered exceptional. Tootell et al. (2003) also raise a general concern
for future mappings in visual areas: “Thus, the retinotopy defining a region
is not absolutely linked to the functional properties of the same region.
When such properties differ, do we assume homology based on the re-
tinotopic criteria or the functional criteria? Here the retinotopy appears
more fundamental (conserved)” (3982). That is, in the V3/V3A case, re-
tinotopy determines which cognitive areas (e.g., V3) are typed as the same
across species, and the function of motion sensitivity is mapped to distinct
cognitive areas in each species.44 Thus, shared anatomical (or cognitive)
areas and functions do not entail shared structure-function mappings. In
general, where multiple criteria are used in individuation, in cases of
conflict the taxonomic result cannot be determined a priori. And there

43. Tootell et al. (1997, 7075) remark: “This distinction [between V3 and V3A in
humans and macaques regarding motion sensitivity] is so salient that one anonymous
reviewer suggested switching the (retinotopically based) names V3 and V3A in humans
to match the ‘reversal’ of motion selectivity in humans. Although this suggestion has
merit, we have chosen to leave the human names V3 and V3A consistent with their
retinotopic counterparts in the macaque, implicitly assuming that the motion selectivity
has become wired differently in V3 and V3A in monkeys, as compared to humans.”

44. This introduces additional complexity to the notion of a structure-function map-
ping: cognitive areas can be individuated (in part) by reference to one cognitive criterion
(organization of visual field) and have other (nonindividuating) cognitive functions
assigned to them.



MULTIPLE REALIZATION 445

are plenty of cases with potential for conflict.45 After all, simpler functions
are generally easier to realize in more than one kind of structure (Kary
and Mahner 2002). Cognitive functions simple enough to be typed across
species (e.g., motion detection) may be the psychological equivalent of
McJobs, in the same way that many different kinds of things can uncork
a wine bottle or be an “and” gate (or man a fry station).

But Bechtel and Mundale (1999) also claim to find support for claim
2 from the methodologies used in lesion (or deficit) studies and imaging
studies (typically involving human subjects): “In interpreting [lesion-based
cognitive] deficits, researchers implicitly reject multiple realization among
human brains and assume that damage to a brain area in anyone will
result in a deficit to a particular cognitive function that is performed by
that area in undamaged brains” (184). “Many [imaging] researchers em-
ploy the subtractive method to focus on brain activities associated with
component psychological processes.” After describing this method, which
I explain below, they continue: “Identifying brain areas through neuroim-
aging depends critically on the cognitive tasks subjects are asked to per-
form; thus, the possibility of multiple realizability is restricted at the out-
set” (190). The latter quotation refers to the pure insertion assumption
used in cognitive subtraction. The former quotation is consistent with
three different assumptions—transparency (in lesion studies), universality,
and unipotentiality—that Bechtel and Mundale do not explicitly intro-
duce. I will describe these but focus on pure insertion.

Bechtel and Mundale also note that it is standard in imaging studies
to transform individual-subject results, voxel by voxel, into the coordi-
nates of a standard brain map and to average results across subjects. That
these procedures leave common areas of statistically significant activation
across subjects, Bechtel and Mundale (1999, 190) argue, “suggests much
less variability than the multiple realizability arguments would allow.”
However, these procedures are as worrisome to many researchers as meth-
odological assumptions because of significant individual differences that
are lost on averaging (see fig. 3).

Before turning directly to Bechtel and Mundale’s claims, it is instructive

45. For example, although humans and macaques both have ventral (object recognition,
“what”) and dorsal (spatial location, “where”) pathways after V1, the object-recognition
stream in humans, unlike in the macaque, runs almost entirely on the ventral surface of
the temporal lobe and does not extend as far forward, while the spatial-location stream
runs along a more superior route in parietal cortex in humans than in the macaque
(Haxby et al. 1991; Ungerleider, Courtney, and Haxby 1998). Ungerleider et al. (1998)
also propose a dual-pathway model for working memory, with object working memory
in ventrolateral prefrontal cortex and spatial working memory in dorsolateral prefrontal
cortex. But their postulated human area for spatial working memory is located both
more superior and more posterior to that of the macaque.



446 CARRIE FIGDOR

Figure 3. fMRI study of normals engaged in a semantic-processing task. Enlarged
images are of a Talaraich z-level slice in one subject (left) and averaged across 12
subjects (right). Source: Savoy (2001, 29). Color version available as an online en-
hancement.

to compare AEF with a somewhat similar “success” argument from Hen-
son (2005). Henson argues that a “working hypothesis” of “strong sys-
tematicity” (“a one-to-one mapping between functional and structural
units”; 193) is necessary in experimental psychology in order to justify
the use of imaging data to inform psychological models—specifically, to
justify inductions from the activation of a structure associated with a
function in one experimental paradigm to the performance of the same
function in other experimental contexts in which that structure is acti-
vated. But to assume strong systematicity is not to assume MR is false;
Henson explicitly acknowledges the possibility of degeneracy within and
across subjects. The assumption, rather, is that an anatomical region (or
network) is not pluripotential: “For inferences of the ‘structure-function
induction’ type . . . one must assume a one-to-one mapping between
function and structure, in order to discount the possibility that the same
structure implements different functions across experimental contexts in
which the [structure-to-function] inference is made” (220).46 As noted,
unipotentiality is compatible with degeneracy, as it does not rule out latent

46. Kosslyn (1999, 1290) raises the same concern about the validity of structure-to-
function inferences given the possibility of pluripotentiality: even for sensory and motor
cortices, “we simply do not know enough to exclude the possibility of multiple roles
for any piece of cortical real estate.”



MULTIPLE REALIZATION 447

degenerate regions (networks) or overlapping degenerate networks within
subjects or degeneracy across subjects. Moreover, Henson adds that the
assumption of strong systematicity “cannot be proved on independent
grounds and is probably best evaluated by the success of the [localization]
enterprise as a whole.” This is because if we do not already know (or
hypothesize) the functional organization of the mind when doing imaging
studies, we cannot know (hypothesize) that an activated area has been
associated with just the right function, at the appropriate level of gen-
eralization. We must assume we have got it right and “see how much
progress is made” (219). In short, Henson argues, if we assume unipo-
tentiality (not the falsity of MR), imaging data can be of use in developing
cognitive models (not that this use is already fruitful), with the hope that
partly as a result of using these data a consistent functional hierarchy will
emerge. If so, then the assumption would be justified.

Bechtel and Mundale’s argument for claim 2, however, rests largely on
the methodology of cognitive subtraction used in imaging studies. Since
the whole brain is always active, localization of function using imaging
data requires measuring relative, not absolute, differences in neural ac-
tivity. The method of cognitive subtraction involves comparing activity
measured during a baseline (or comparison) task and activity in a closely
matched experimental (or activation) task that is thought to differ from
the baseline task by one component.47 For example, the baseline task may
be passively seeing a written word, and the experimental task is seeing a
written word and saying it aloud. Activity imaged during the baseline
task is “subtracted” from that imaged during the experimental task. (Often
only the peaks or centers of mass of statistically significant areas of ac-
tivation left after subtraction—and after averaging and smoothing pro-
cedures—are shown in published imaging results.) The function has been
localized when we infer that the new cognitive component in the exper-
imental task is subserved by the region(s) of additional activation re-
maining after subtraction.

The legitimacy of this inference depends on the assumption of pure

47. Baseline activity is neural activity above a chosen threshold of statistical signifi-
cance detected during performance of the baseline task; activity detected in the ex-
perimental condition is relative to the same threshold. Developing a baseline that
involves all and only the same processes as the experimental task (except the process
of interest) is problematic (Price and Friston 1997), as is determining an appropriate
threshold for statistical significance, since depending on the choice of threshold, dif-
fering areas of activity may be implicated in subserving a cognitive function. To borrow
Savoy’s (2001, 28–30) analogy: if your task is to count islands, how many you find
and how large they are will depend on the sea level.



448 CARRIE FIGDOR

insertion.48 Whenever a task is added, there is processing involved due to
the interaction between the new and the old tasks. The pure insertion
assumption is that the amount of additional activity attributable to the
interaction is zero. At the psychological level, the assumption is that
cognition proceeds via isolable steps (in the simplest cases, in a serial,
feed-forward manner), such that adding a task has no effect on the way
previous tasks are performed. At the neural level, the assumption is that
the difference in neural activity imaged during the baseline and experi-
mental tasks is due entirely to the new task and does not represent any
influences on or interaction with the baseline activity. In short, the pure
insertion assumption is that both functional and structural systems are
modular.49 Pure insertion rules out the possibility that the additional neu-
ral activity may be neither sufficient nor necessary for the presumed-to-
be purely additional task, thereby making an inference from a one-to-one
(new task : new activity) mapping possible.

An analogous assumption in cognitive neuropsychology is called the
transparency assumption (Caramazza 1986). Cognitive neuropsycholo-
gists study patterns of cognitive deficits to develop models of normal
cognitive function that can be mapped to the lesioned areas. (Double
dissociations are patterns of deficit considered especially helpful; see n.
21.) However, the legitimacy of using lesioned subjects (e.g., stroke vic-
tims) to understand normal cognition must be justified. The transparency
assumption claims that the same cognitive system exists through change;
there is not a prelesion system and a distinct reorganized postlesion system.
Transparency rules out the possibility that the lesion site subserves part
of a new reorganized system rather than the specific deficit(s). This as-
sumption has also been questioned (see Ward 2006, 85) but is best ex-
amined in the context of a discussion of plasticity, which I set aside here.50

When researchers average the imaging results from multiple subjects
and plot them to a common brain atlas to identify areas of statistically

48. The assumption of pure insertion originates in experimental psychology (Sternberg
1969), where it is part of the “additive factors” method of testing hypotheses about
processing stages on the basis of differences in reaction times when comparing baseline
and experimental tasks. This assumption is still widely used in experimental psychology
(Henson 2005, 212) but not in imaging studies (Ward 2006, 56–62).

49. This is modularity without the usual bells and whistles (Fodor 1983; Carruthers
2006). I cannot go into the modularity debate here, but a skeptic about modularity
should be skeptical about pure insertion.

50. “Plasticity” refers to the brain’s ability to change in response to experience. This
includes both its development over a lifetime (but especially in childhood) and the
restitution of function after damage. Plasticity has been used to argue for MR in existing
biological organisms (see Shapiro [2004] for a detailed response), but for space reasons
this debate cannot be adequately assessed here.



MULTIPLE REALIZATION 449

significant activity across subjects, they also assume universality: that all
(human) cognitive systems are basically the same (Caramazza 1986). This
assumption legitimizes generalizing from a small experimental sample to
a larger group. Since it equally legitimizes inferring from degeneracy in
a sample to degeneracy in a larger group, it cannot tell against degeneracy.

Cognitive subtraction, and the assumption of pure insertion on which
it rests, has enabled researchers to use imaging data to begin to unravel
the mystery of brain-based cognition. But pure insertion is undermined
by research indicating nonnegligible interactions between areas, such as
in cases of effective connectivity and the influence of directed attention
(Friston 1997), the effect of practice (Raichle et al. 1994), and the relation
between the current task(s) and what the subject has just done before
testing (Kosslyn 1999, 1291–92; see also Pachella 1974 and Uttal 1998).
In these cases, the regions associated with the tasks shared in the baseline
and experimental conditions differ when a new processing component is
added. For example, in a direct test of pure insertion, Jennings et al.
(1997) found that the areas associated with the baseline task differed
depending on how subjects were asked to respond. Subjects were given a
semantic task (determining whether a word represents something living)
and a letter-judgment task (determining whether a word contained the
letter “a”) and, in both cases, were asked to give a yes or no response in
three ways: silent thought, clicking a mouse, or answering aloud. The
pair-wise subtractions were between (e.g.) reading, making a semantic
judgment, and responding yes or no by clicking a mouse versus reading,
making a semantic judgment, and responding yes or no by saying yes or
no aloud. If pure insertion were true, the areas associated with the se-
mantic task should remain the same across response modes. Instead, the
active areas differed significantly. Although the left inferior frontal cortex
was activated in all semantic-processing conditions (albeit to different
degrees, depending on the response condition), other areas were unique
to semantic processing given a particular response mode or to only two
of them. And even the existence of a common area of activation is con-
sistent with the hypothesis that this region is a necessary part of multiple
degenerate networks for semantic processing.

In imaging studies, experimental paradigms that avoid pure insertion
are now favored. For example, in the cognitive conjunction paradigm
(Price and Friston 1997), which is a modification of cognitive subtraction,
several task pairs are developed that share only the cognitive component
of interest.51 Since activity due to interaction will be specific to each pair,

51. Other methods include using factorial and parametric designs (Friston 1997). In
a factorial design, the aim is to identify areas of interaction explicitly by comparing
activation during performance of tasks combined as factors or variables (e.g., if P p



450 CARRIE FIGDOR

any activity that remains across subtractions can be associated with the
component of interest. In their PET study, they imaged subjects perform-
ing subtraction tasks that shared only the process of phonological retrieval
from visually presented stimuli (defined as activating the name attached
to a visual stimulus or concept). For example, one pair-wise subtraction
was reading a visually presented word and saying a prespecified word to
visually presented strings of false font; another was naming a visually
presented Arabic letter and saying the same prespecified word to single
false-font characters. Significant activation surviving subtraction across
all task pairs was found in the left posterior basal temporal lobe, left
frontal operculum, and midline cerebellum. It does not follow that pho-
nological retrieval can be localized to these regions. Because of the pos-
sibility of degeneracy, Price and Friston (2002) themselves deny the le-
gitimacy of inferring from the necessity of an activated area (or areas)
for the component of interest. The common area(s) may not be necessary
(e.g., if there is latency) or sufficient (e.g., if it participates in multiple
degenerate networks). Nevertheless the data are valuable for helping to
determine the role(s) of the activated cortical areas without presuming
which structure-function theory may eventually prove correct.

In addition, Noppeney et al. (2004) have proposed a new methodology
designed specifically to find evidence of degeneracy within subjects (see
fig. 4). In their iterated model, normal subjects are imaged to identify
candidate regions of interest and generate hypotheses about potential
degenerate neural systems. Lesions to these regions (in patients or tem-
porarily induced in normals) then test whether a specific behavioral deficit
occurs after damage to one region (prima facie evidence of its necessity)
or more than one (prima facie evidence of degeneracy). If lesioning does
not yield a deficit, a latent degenerate system may be operating. Imaging
patients with lesions in the original regions of interest but who can perform
the task at normal levels can reveal these latent systems.

To summarize, cognitive neuroscience methodologies do not rely even
implicitly on the assumption that MR is false. Nor do the revised and
new methods presuppose that degeneracy is true. They aim to provide
better data from which to infer more reliably whichever of the competing
theories is more explanatory.

It is also worth noting that the validity of inferences made in imaging
studies in particular are likely to be affected by technological develop-
ments. One, already noted, is the increasing use of diffusion imaging to

phonological retrieval and O p object recognition, a factorial design would evaluate
activity during conditions P&O, ∼P&O, P&∼O, and ∼P&∼O). Parametric designs treat
variables as dimensions (i.e., with degrees of activation) rather than categories (i.e.,
active or not). See also Ward (2006, 56–62).



MULTIPLE REALIZATION 451

Figure 4. Proposed method for identifying degenerate neuronal systems by combining
functional imaging and neuropsychological lesion data from normal subjects and lesion
patients in an iterative procedure. Bottom left arrow (after “Deficit?”) is no, and the
bottom right arrow is yes. Source: Noppeney, Friston, and Price (2004, 438).

reveal axonal connectivity, not just activated regions, in normal human
subjects. Another is the recently authorized use on human subjects of
fMRI machines that can generate magnetic fields of up to 8 tesla. Most
fMRI machines in current experimental use operate at 1 or 1.5 tesla, a
level at which just one in 100,000 hydrogen atoms in the subject’s brain
aligns with the generated magnetic field.52 The stronger machines raise an
interesting problem (Savoy 2001, 30–31). Even when keeping the choice
of threshold, experimental paradigm, and method of data analysis fixed,
additional statistically significant activation is likely to be detected with
stronger fMRI machines. It follows that at least some regions of activation
(and current cortical cognitive areas) may be artifacts of the power of
most fMRI machines used experimentally up to now. Not only are iden-
tified areas likely to morph and grow; some are likely to be statistically
significantly active (if not maximally so) during other tasks at 8 tesla,
even if this activity is below threshold at 1.5 tesla. In other words, an

52. In fMRI, the aligned hydrogen nuclei are temporarily knocked 90 degrees off
alignment by an electrical pulse; their movement back to alignment generates the signal
from which we infer statistically significant differences in blood oxygenation levels
between brain regions and from there to significant differences in neural activity, which
may in turn be associated with a task that the imaged subject performs.



452 CARRIE FIGDOR

area that “looks” unipotential at 1.5 tesla may turn out to be clearly
pluripotential at 8 tesla.53

5. Conclusion. I have argued that (i) MR is a live empirical hypothesis
within cognitive neuroscience and (ii) some of the main cognitive neu-
roscientifically based arguments against the plausibility of MR fail. There
is no basis for claiming (as Bechtel and Mundale [1999] do) that the
empirical plausibility of MR “evaporates” in the light of neuroscientific
research. To the contrary, it is a genuine mystery why so much of the
philosophical literature on the implications of cognitive neuroscience for
MR has been so negative in its conclusions. As far as I can tell, Putnam’s
intuitions may ultimately be vindicated for evolved biological organisms,
without even considering what may be the case for Martians and robots.
At the very least, supporters of MR, and of nonreductive physicalism,
have no reason to shy away from defending their positions on cognitive
neuroscientific grounds, leaving intuition and analogy far behind.

REFERENCES

Aizawa, Kenneth, and Carl Gillett. 2009. “Levels, Individual Variation, and Massive Mul-
tiple Realization in Neurobiology.” In The Oxford Handbook of Philosophy and Neu-
roscience, ed. John Bickle, 539–81. New York: Oxford University Press.

Andrews, Kristin. 2008. Stanford Encyclopedia of Philosophy, s.v. “Animal Cognition,” http://
plato.stanford.edu/entries/cognition-animal/.

Attwell, David, and Costantino Iadecola. 2002. “The Neural Basis of Functional Brain
Imaging Signals.” Trends in Neurosciences 25 (12): 621–25.

Bechtel, William. 2006. “Critical Notice: The Mind Incarnate.” Philosophy and Phenome-
nological Research 73 (2): 497–500.

Bechtel, William, and Jennifer Mundale. 1999. “Multiple Realizability Revisited: Linking
Cognitive and Neural States.” Philosophy of Science 66:175–207.

Bickle, John. 1998. Psychoneural Reduction: The New Wave. Cambridge, MA: MIT Press.
———. 2003.Philosophy and Neuroscience: A Ruthlessly Reductive Account. Dordrecht: Klu-

wer.
Block, Ned J., and Jerry A. Fodor. 1972. “What Psychological States Are Not.” Philosophical

Review 8 (2): 159–81.
Broca, Paul. 1861/1960/2001. “Remarques Sur le Siege de la Faculte du Langage Articule;

Survies d’une Observation d’Aphemie.” Bulletin de la Societe Anatomique 6:330–57.
Trans. G. von Bonin. 1960. “Remarks on the Seat of the Faculty of Articulate Language,
Followed by an Observation of Aphemia.” In Some Papers on the Cerebral Cortex, ed.
G. von Bonin, 49–72. Springfield, IL: Thomas. Repr. in Philosophy and the Neurosci-
ences: A Reader, ed. William Bechtel, Pete Mandik, Jennifer Mundale, and Robert
Stufflebeam, 87–99. Malden: Blackwell.

Brodmann, Korbinian. 1909/1994. Vergleichende Lokalisationslehre der Grosshirnrinde. Leip-
zig: Barth. Trans. L. J. Garey. 1994. Brodmann’s Localisation in the Cerebral Cortex.
London: Smith-Gordon.

53. Also, and in contrast with Bechtel and Mundale’s claim that averaged imaging
data provide evidence of commonality, Savoy (2001, 31) considers averaged data a
harbinger of this problem. In images showing averaged data from various subjects, the
areas of maximal activity tend to be both more numerous and larger.



MULTIPLE REALIZATION 453

Burton, H. 2003. “Visual Cortex Activity in Early and Late Blind People.” Journal of
Neuroscience 23 (10): 4005–11.

Caramazza, Alfonso. 1986. “On Drawing Inferences about the Structure of Normal Cog-
nitive Systems from the Analysis of Patterns of Impaired Performance: The Case for
Single-Patient Studies.” Brain and Cognition 5:41–66.

Carruthers, Peter. 2006. The Architecture of the Mind. New York: Oxford University Press.
Churchland, Patricia. 1986. Neurophilosophy: Toward a Unified Science of the Mind/Brain.

Cambridge, MA: MIT Press.
Churchland, Paul. 1981. “Eliminative Materialism and the Propositional Attitudes.” Journal

of Philosophy 78 (2): 67–90.
Clapp, Leonard. 2001. “Disjunctive Properties: Multiple Realizations.” Journal of Philosophy

98 (3): 111–36.
Coltheart, Max. 2004. “Brain Imaging, Connectionism, and Cognitive Neuropsychology.”

Cognitive Neuropsychology 21 (1): 21–25.
Coltheart, Max, and Martin Davies. 2003. “Inference and Explanation in Cognitive Neu-

ropsychology.” Cortex 39:188–91.
Couch, Mark B. 2004. “Discussion: A Defense of Bechtel and Mundale.” Philosophy of

Science 71:198–204.
Craver, Carl F. 2007. Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience.

New York: Oxford University Press.
Deacon, Terrence. 2004. “Monkey Homologues of Human Language Areas: Computing

the Ambiguities.” Trends in Cognitive Sciences 8 (7): 288–90.
Dronkers, Nina F., Brenda B. Redfern, and Robert T. Knight. 2000. “The Neural Archi-

tecture of Language Disorders.” In The New Cognitive Neurosciences, ed. Michael
Gazzaniga, 949–58. Cambridge, MA: MIT Press.

Dunn, John C., and Kim Kirsner. 2003. “What Can We Infer from Double Dissociations?”
Cortex 39:1–7.

Edelman, Gerald M., and Joseph A. Gally. 2001. “Degeneracy and Complexity in Biological
Systems.” Proceedings of the National Academy of Sciences of the USA 98 (24): 13763–
68.

Farah, Martha J. 1994. “Neuropsychological Inference with an Interactive Brain: A Critique
of the ‘Locality’ Assumption.” Behavioral and Brain Sciences 17 (1): 43–104.

Felleman, Daniel J., and David C. Van Essen. 1987. “Distributed Hierarchical Processing
in the Primate Cerebral Cortex.” Cerebral Cortex 1:1–47.

———. 1991a. “Distributed Hierarchical Processing in the Primate Cerebral Cortex.” Ce-
rebral Cortex 1:1–47.

———. 1991b. “Receptive Field Properties of Neurons in Area V3 of Macaque Monkey
Extrastriate Cortex.” Journal of Neurophysiology 57:889–920.

Field, Hartry. 1973. “Theory Change and the Indeterminacy of Reference.” Journal of
Philosophy 70:462–81.

Fodor, Jerry. 1974/1975. “Special Sciences; or, The Disunity of Science as a Working Hy-
pothesis.” Synthese 28:97–115. Rev. and repr. in The Language of Thought, 9–26. Cam-
bridge, MA: Harvard University Press.

———. 1983. The Modularity of Mind. Cambridge, MA: MIT Press.
———. 2000. “Special Sciences: Still Autonomous after All These Years; A Reply to Jaegwon

Kim’s ‘Multiple Realization and the Metaphysics of Reduction.’” In Critical Condition,
9–24. Cambridge, MA: MIT Press. Repr. from Philosophical Perspectives 11: Mind,
Causation, and World, ed. J. Tomberlin. Atascadero, CA: Ridgeview.

Friston, Karl. 1997. “Imaging Cognitive Anatomy.” Trends in Cognitive Sciences 1 (1): 21–
27.

———. 2002. “Beyond Phrenology: What Can Neuroimaging Tell Us about Distributed
Circuitry?” Annual Review of Neuroscience 25:221–50.

Friston, Karl, and Cathy Price. 2003. “Degeneracy and Redundancy in Cognitive Anatomy.”
Trends in Cognitive Sciences 7 (4): 151–52.

Gillett, Carl. 2003. “The Metaphysics of Realization, Multiple Realizability, and the Special
Sciences.” Journal of Philosophy 100 (11): 591–603.

Gregory, R. L. 1961. “The Brain as an Engineering Problem.” In Current Problems in Animal



454 CARRIE FIGDOR

Behaviour, ed. W. H. Thorpe and O. L. Zangwill, 307–30. London: Cambridge Uni-
versity Press.

Grodzinsky, Yosef, and Andrea Santi. 2008. “The Battle for Broca’s Region.” Trends in
Cognitive Sciences 12 (12): 474–80.

Hagmann, Patric, Leila Cammoun, Xavier Gigandet, Reto Meuli, Christopher J. Honey,
Van J. Wedeen, and Olaf Sporns. 2008. “Mapping the Structural Core of Human
Cerebral Cortex.” Public Library of Science (PLoS) Biology 6 (7): 1–15.

Haxby, James V., Cheryl I. Grady, Barry Horwitz, Leslie G. Underleider, Mortimer Mishkin,
Richard E. Carson, Peter Herscovitch, Mark B. Schapiro, and Stanley I. Rapoport.
1991. “Dissociation of Object and Spatial Visual Processing Pathways in Human Ex-
trastriate Cortex.” Proceedings of the National Academy of Sciences of the USA 88 (5):
1621–25.

Henson, Richard. 2005. “What Can Functional Neuroimaging Tell the Experimental Psy-
chologist?” Quarterly Journal of Experimental Psychology 58A (2): 193–233.

Jennings, Janine M., Anthony R. McIntosh, Shitij Kapur, Endel Tulving, and Sylvain Houle.
1997. “Cognitive Subtractions May Not Add Up: The Interaction between Semantic
Processing and Response Mode.” Neuroimage 5:229–39.

Johnson, Adam, Andre A. Fenton, Cliff Kentros, and A. David Redish. 2009. “Looking
for Cognition in the Structure within the Noise.” Trends in Cognitive Sciences 13 (2):
55–64.

Kanwisher, Nancy, Josh McDermott, and Marvin Chun. 1997. “The Fusiform Face Area:
A Module in Human Extrastriate Cortex Specialized for Face Perception.” Journal of
Neuroscience 17 (1): 4302–11.

Kary, Michael, and Martin Mahner. 2002. “How Would You Know if You Synthesized a
Thinking Thing?” Minds and Machines 12:61–86.

Keeley, Brian L. 2000. “Shocking Lessons from Electric Fish: The Theory and Practice of
Multiple Realization.” Philosophy of Science 67:444–65.

———. 2002. “Making Sense of the Senses: Individuating Modalities in Humans and Other
Animals.” Journal of Philosophy 99 (1): 5–28.

Kim, Jaegwon. 1993. “Multiple Realization and the Metaphysics of Reduction.” In Super-
venience and Mind: Selected Philosophical Essays, 309–35. New York: Cambridge Uni-
versity Press.

———. 2006. Philosophy of Mind. 2nd ed. Cambridge, MA: Westview.
Kim, Sungsu. 2002. “Testing Multiple Realizability: A Discussion of Bechtel and Mundale.”

Philosophy of Science 69:606–10.
Kosslyn, Stephen M. 1999. “If Neuroimaging Is the Answer, What Is the Question?” Phil-

osophical Transactions of the Royal Society B 354 (1387): 1283–94.
Kosslyn, Stephen M., Nathaniel M. Alpert, William L. Thompson, Christopher F. Chabris,

Scott L. Rauch, and Adam K. Anderson. 1994. “Identifying Objects Seen from Different
Viewpoints: A PET Investigation.” Brain 117:1055–71.

Kosslyn, Stephen M., Gregory J. DiGirolamo, William L. Thompson, and Nathaniel M.
Alpert. 1998. “Mental Rotation of Objects versus Hands: Neural Mechanisms Revealed
by Positron Emission Tomography.” Psychophysiology 35:151–61.

Lashley, Karl S. 1929. Brain Mechanisms and Intelligence. Chicago: University of Chicago
Press.

———. 1950. “In Search of the Engram.” Symposia of the Society for Experimental Biology
4:454–82.

LeDoux, Joseph E. 2000. “Emotion Circuits in the Brain.” Annual Review of Neuroscience
23:155–84.

Le Poidevin, Robin. 2000. “Space and the Chiral Molecule.” In Of Minds and Molecules:
New Philosophical Perspectives on Chemistry, ed. Nalini Bhushan and Stuart Rosenfeld,
129–41. Oxford: Oxford University Press.

Marshall, John C., and Gereon R. Fink. 2003. “Cerebral Localization, Then and Now.”
NeuroImage 20:S2–S7.

Mundale, Jennifer. 2002. “Concepts of Localization: Balkanization in the Brain.” Brain and
Mind 3:1–18.



MULTIPLE REALIZATION 455

Noppeney, Uta, Karl Friston, and Cathy Price. 2004. “Degenerate Neuronal Systems Sus-
taining Cognitive Functions.” Journal of Anatomy 205:433–42.

Orban, Guy A., David Van Essen, and Wim Vanduffel. 2004. “Comparative Mapping of
Higher Visual Areas in Monkeys and Humans.” Trends in Cognitive Sciences 8 (7):
315–24.

Pachella, Robert G. 1974. “The Interpretation of Reaction Time in Information-Processing
Research.” In Human Information Processing: Tutorials in Performance and Cognition,
ed. Barry H. Kantowitz, 41–82. Hillsdale, NJ: Erlbaum.

Phillips, C. G., S. Zeki, and H. B. Barlow. 1984. “Localization of Function in the Cerebral
Cortex: Past, Present and Future.” Brain 104 (1): 328–61.

Plaut, David C. 1995. “Double Dissociation without Modularity: Evidence from Connec-
tionist Neuropsychology.” Journal of Clinical and Experimental Neuropsychology 17 (2):
291–321.

Polger, Thomas. 2004. Natural Minds. Cambridge, MA: MIT Press.
Price, Cathy J., and Karl J. Friston. 1997. “Cognitive Conjunction: A New Approach to

Brain Activation Experiments.” NeuroImage 5:261–70.
———. 2002. “Degeneracy and Cognitive Anatomy.” Trends in Cognitive Sciences 6 (10):

416–21.
———. 2005. “Functional Ontologies for Cognition: The Systematic Definition of Structure

and Function.” Cognitive Neuropsychology 22 (3): 262–75.
Putnam, Hilary. 1967. “The Nature of Mental States.” In Mind, Language and Reality:

Philosophical Papers, vol. 2, 429–40. New York: Cambridge University Press. Repr.
“Psychological Predicates.” In Art, Mind and Religion, ed. W. H. Capitan and Daniel
D. Merrill, 37–48. Pittsburgh: University of Pittsburgh Press.

———. 1975. “Philosophy and Our Mental Life.” In Mind, Language and Reality: Philo-
sophical Papers, vol. 2, 291–303. New York: Cambridge University Press.

Raichle, Marcus E., Julie E. Fiez, Tom O. Videen, Ann-Mary K. MacLeod, Jose V. Pardo,
Peter T. Fox, and Steven E. Petersen. 1994. “Practice-Related Changes in Human Brain
Functional Anatomy during Nonmotor Learning.” Cerebral Cortex 4 (1): 8–26.

Savoy, Robert L. 2001. “History and Future Directions of Human Brain Mapping and
Functional Neuroimaging.” Acta Psychologica 107:9–42.

Schwartz, Eric L., ed. 1990. Computational Neuroscience. Cambridge, MA: MIT Press.
Shallice, Tim. 1988. From Neuropsychology to Mental Structure. Cambridge: Cambridge

University Press.
Shapiro, Lawrence A. 2000. “Multiple Realizations.” Journal of Philosophy 97 (12): 635–

54.
———. 2004. The Mind Incarnate. Cambridge, MA: MIT Press.
Sternberg, Saul. 1969. “The Discovery of Processing Stages: Extensions of Donders’

Method.” Acta Psychologica 30:276–315.
Talairach, J., and P. Tournoux. 1988. Co-planar Stereotaxic Atlas of the Human Brain. New

York: Thieme Medical.
Tononi, Giulio, Olaf Sporns, and Gerald M. Edelman. 1999. “Measures of Degeneracy and

Redundancy in Biological Networks.”Proceedings of the National Academy of Sciences
of the USA 96:3257–62.

Tootell, Roger B. H., Janine D. Mendola, Nouchine K. Hadjikhani, Partick J. Ledden,
Arthur K. Liu, John B. Reppas, Martin I. Sereno, and Anders M. Dale. 1997. “Func-
tional Analysis of V3A and Related Areas in Human Visual Cortex.” Journal of Neu-
roscience 17 (18): 7060–78.

Tootell, Roger B. H., Doris Tsao, and Wim Vanduffel. 2003. “Neuroimaging Weighs In:
Humans Meet Macaques in ‘Primate’ Visual Cortex.” Journal of Neuroscience 23 (10):
3981–89.

Ungerleider, Leslie, Susan Courtney, and James V. Haxby. 1998. “A Neural System for
Human Visual Working Memory.” Proceedings of the National Academy of Sciences of
the USA 95:883–90.

Ungerleider, Leslie, and Mortimer Mishkin. 1982. “Two Cortical Visual Systems.” In Anal-
ysis of Visual Behavior, ed. David J. Ingle, Melvyn A. Goodale, and Richard J. W.
Mansfield, 549–86. Cambridge, MA: MIT Press.



456 CARRIE FIGDOR

Uttal, William R. 1998. Toward a New Behaviorism: The Case against Perceptual Reduc-
tionism. Mahwah, NJ: Erlbaum.

———. 2001. The New Phrenology: The Limits of Localizing Cognitive Processes in the
Brain. Cambridge, MA: MIT Press.

Vanduffel, W., D. Fize, J. B. Mandeville, K. Nelissen, P. Van Hecke, B. R. Rosen, R. B.
Tootell, and G. A. Orban. 2001. “Visual Motion Processing Investigated Using Contrast
Agent-Enhanced fMRI in Awake Behaving Monkeys.” Neuron 32:565–77.

Vanduffel, W., D. Fize, H. Peuskens, K. Denys, S. Sunaert, J. T. Todd, and G. A. Orban.
2002. “Extracting 3D from Motion: Differences in Human and Monkey Intraparietal
Cortex.” Science, n.s., 298 (5592): 413–15.

Von Melchner, Laurie, Sarah Pallas, and Mrisanka Sur. 2000. “Visual Behaviour Mediated
by Retinal Projections Directed to the Auditory Pathway.” Nature 404:871–76.

Ward, Jamie. 2006. The Student’s Guide to Cognitive Neuroscience. Hove: Psychology.