Neuron

Article

A Magnitude Code Common

to Numerosities and Number Symbols
in Human Intraparietal Cortex
Manuela Piazza,1,2,4,5,* Philippe Pinel,1,2,4 Denis Le Bihan,2,4 and Stanislas Dehaene1,2,3,4
1
Institut National de la Santé et de la Recherche Médicale (INSERM), Cognitive Neuroimaging Unit, Service Hospitalier Frédéric
Joliot, F-91401 Orsay, France
2
Commissariat à l’Energie Atomique (CEA), Département de Recherche Médicale (DRM), Direction des Sciences du Vivant (DSV),
Service Hospitalier Frédéric Joliot, 91406 Orsay, France
3
Collège de France, 75005 Paris, France
4
IFR 49 (Institut Fédératif de Recherche en Imagerie Neurofonctionnelle)
5
Present address: Functional NeuroImaging Laboratory, Center for Mind Brain Sciences, University of Trento, Via Tartarotti 7,
38068 Rovereto (TN), Italy.
*Correspondence: [email protected]
DOI 10.1016/j.neuron.2006.11.022

SUMMARY set (numerosity) also involved in representing the meaning

of a given numerical symbol? It has been proposed that,
Activation of the horizontal segment of the in- in the course of learning to speak, read, or write, children
traparietal sulcus (hIPS) has been observed learn to map spoken and written numerals onto a core
in various number-processing tasks, whether representation of numerosity in the intraparietal sulcus
numbers were conveyed by symbolic numerals (Butterworth, 1999; Dehaene, 1997). It is likely that sym-
(digits, number words) or by nonsymbolic dis- bolic and concrete depictions of number are linked
together in the adult human brain, perhaps in the form of
plays (dot patterns). This suggests an abstract
notation-independent assemblies of neurons coding for
coding of numerical magnitude. Here, we criti-
number at a purely conceptual level (cardinality). The
cally tested this hypothesis using fMRI adapta- goal of the present study is to put this hypothesis to a crit-
tion to demonstrate notation-independent cod- ical test using fMRI adaptation.
ing of numerical quantity in the hIPS. Once Recent research has suggested that concepts of num-
subjects were adapted either to dot patterns ber form a semantic category which dissociates from other
or to Arabic digits, activation in the hIPS and categories of concepts (for example colors, living things, or
in frontal regions recovered in a distance- tools). In neuropsychology, a double dissociation exists
dependent fashion whenever a new number between numbers and other semantic categories. On the
was presented, irrespective of notation changes. one hand, there are patients (with cortical atrophy mainly
This remained unchanged when analyzing the of temporal lobes) with heavily deteriorated semantic pro-
cessing but spared calculation and number comprehen-
hIPS peaks from an independent localizer
sion (Butterworth et al., 2001; Halpern et al., 2004; Thioux
scan of mental calculation. These results sug-
et al., 1998). On the other hand, there are patients (mainly
gest an abstract coding of approximate number with lesions or atrophy in the parietal cortex) that show
common to dots, digits, and number words. impaired understanding of numbers but otherwise pre-
They support the idea that symbols acquire served knowledge of other semantic domains (Cipolotti
meaning by linking neural populations coding et al., 1991; Dehaene and Cohen, 1997; Halpern et al.,
symbol shapes to those holding nonsymbolic 2004; Zamarian et al., 2006).
representations of quantities. Functional imaging techniques have also revealed some
degree of segregation between numbers and other cate-
INTRODUCTION gories such as animals and colors (Eger et al., 2003; Thi-
oux et al., 2002). Furthermore, they have helped clarify
Numerical quantities, like many other categories of con- the organization of number-related processes in the pari-
cepts, can be expressed symbolically or analogically. etal lobe. fMRI studies have suggested a crucial role of
The numerical quantity 3, for example, can be expressed regions situated along the horizontal segment of the intra-
symbolically by the digit 3 or by the spoken or written parietal sulcus (hIPS) of both hemispheres for the repre-
word ‘‘three.’’ It can also be conveyed in concrete form sentation of numerical quantities (Dehaene et al., 2003).
by a set of three visual objects, a sequence of three tones, Independent lines of research have pointed to parietal cor-
or three self-generated movements. Are the neural popu- tex as crucial for coding numerical quantity both when it is
lations that extract and represent the cardinality of a given conveyed by number symbols (digits, number words; Eger

Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc. 293
 Neuron
A Magnitude Code in the Human Parietal Cortex

et al., 2003; Pinel et al., 2001) or by nonsymbolic displays

of dots patterns (Ansari et al., 2006; Cantlon et al., 2006;
Piazza et al., 2004). Those results have been taken to im-
ply that the hIPS contains a modality-independent quan-
tity representation. However, symbolic and nonsymbolic
quantities were often tested in different experimental set-
tings and subjects; only one recent fMRI study showed
common activations in the anterior IPS bilaterally, left pos-
terior IPS, medial frontal gyrus, and left precentral gyrus
for symbolic and nonsymbolic addition (Venkatraman
et al., 2005). Most crucially, coactivation of the same vox-
els need not imply a common neural code but might sim-
ply be due to the activation of distinct neural populations Figure 1. 2 3 2 Design for Adaptation to Approximate Number
intermixed at the same cortical location (Cohen Kadosh In distinct sequences, subjects adapted to numerical values close to
et al., 2005; Pinel et al., 2004). 20 or close to 50 and were then tested with those values presented
Another line of research pointing to a possible conver- as deviants. Brain regions coding for approximate number should
gence of symbolic and nonsymbolic representations of react more when the deviant is far from the adaptation value than
when it is close to it. In the complete design, two additional factors
numbers arises from chronometric investigations. Evi-
were manipulated, as both the adaptation and the deviant numbers
dence from the number comparison task has suggested could be presented either in symbolic notation (Arabic digits) or non-
that the internal metrics governing the mental representa- symbolically (as sets of dots).
tion of nonsymbolic quantities and of numerical symbols
are similar. In both cases, response times and error rates onstrated subliminal fMRI repetition priming across Arabic
show numerical distance and numerical magnitude ef- digits and written words. The key addition of the present
fects: numerical judgments become more difficult when study is to test whether those symbolic and nonsymbolic
the numerical distance between two values decreases, adaptation effects occur at a shared level of parietal rep-
and this effect is exacerbated as their absolute magnitude resentation. For each subject, we also acquired a short
increases (Buckley & Gillman, 1974; Dehaene et al., 1990; localizer scan to identify regions active during symbolic
Koechlin et al., 1999; Shepard et al., 1975). Such effects mental calculation (subtractions of numbers presented
are captured by Weber’s law, which states that the thresh- visually or auditory). This localizer scan provided an inde-
old of discrimination between two stimuli scales with their pendent definition of intraparietal regions of interest,
magnitude. They point to an internal coding of number where we investigated the presence of fMRI adaptation
by a distribution of activation on an internal compressed effects.
number ‘‘line’’ or ‘‘scale.’’ These observations led some
authors to the conclusion that symbolic and nonsymbolic RESULTS
notations converge onto a common format of representa-
tion (Buckley and Gillman, 1974; Dehaene et al., 1998; During short (2 min) sequences, participants were pas-
Piazza and Dehaene, 2004; Shepard et al., 1975). Once sively exposed to a fixed set of numerical quantities, all
again, however, such behavioral evidence is indirect in very close to a given value (e.g., 17, 18, or 19; see Figure 1).
nature and leaves open the possibility that there are two In a given sequence, these adaptation numerosities were
separate systems for representing symbolic and nonsym- presented in a fixed notation, either as Arabic digits or as
bolic numerical magnitudes that are simply governed by sets of dots. Recovery from adaptation was investigated
a similar metric. Indeed, magnitude effects are shared at two different temporal scales (Figure 2). First, after two
properties of a great variety of mental representations minutes of adaptation with a fixed approximate quantity
(e.g., size, sound intensity, etc.) that probably have little (hereafter called period A1), the adaptation numerosities
in common (Cohen Kadosh et al., 2005; Pinel et al., 2004). changed abruptly (e.g., to 47, 48, or 49), with or without a
To directly evaluate the presence of an abstract, nota- concomittent change in notation, and remained approxi-
tion-independent code for numerical magnitude in the mately fixed for another 2 min period (hereafter called
hIPS, we measured brain activity with fMRI while 14 period A2). We examined whether this unforeseen change
healthy volunteers passively observed both nonsymbolic led to a durable rebound in the fMRI signal. Second, within
(dot patterns) and symbolic (Arabic digits) numbers. Using each block, following a fixed adaptation part, sparse devi-
an adaptation paradigm, we investigated whether a region ant stimuli were occasionally introduced. We examined
in the bilateral intraparietal sulcui adapts to approximate whether each deviant event led to a local response depen-
quantity and shows crossnotation recovery whenever dent on the distance between the adaptation and deviant
the number changes. Previous research has shown num- values. Crucially, all combinations of adaptation and devi-
ber-related adaptation in the hIPS, both in adults and in 4- ant notations were used, resulting in two within-notation
year-old subjects for numbers presented as sets of dots conditions (dots-to-dots and Arabic-to-Arabic) as well as
(Ansari et al., 2006; Cantlon et al., 2006; Piazza et al., two crossnotation conditions (dots-to-Arabic and Arabic-
2004). Furthermore, Naccache and Dehaene (2001) dem- to-dots).

294 Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc.
Neuron
A Magnitude Code in the Human Parietal Cortex

Figure 2. General Structure of Each Experimental Block and Example of Specific Stimuli
Each block comprised two successive adaptation sequences (A1 and A2). Within each sequence, the first 25 stimuli always presented the designated
adaptation numbers. Deviants then occurred at pseudorandom moments (12 deviants amongst 60 adaptation stimuli). Each sequence ended with six
trials with adaptation numbers only. At the transition between the sequences A1 and A2, the adaptation numbers abruptly changed without any break
or warning. On half such transitions, notation also changed. The example (bottom) depicts a case in which adaptation stimuli in both A1 and A2 are
Arabic numerals in the range 17–19 in A1 and in the range 47–49 in A2.

Initial Adaptation related decreasing activity were mostly observed in bilat-

To investigate the presence of adaptation, we searched for eral parietal and frontal cortices, with additional effects in
regions where activity decreased linearly during the first occipito-temporal, cerebrellar, and subcortical thalamic
30 s of periods A1 and A2, both after rest and after a change regions (Table 1). Given our a priori focus on parietal cortex
in number (see Figures 2 and 3). Regions showing number- (Piazza and Dehaene, 2004; Piazza et al., 2004), we then

Figure 3. Time Course of Bilateral Parietal Activation during an Experimental Block (A1 Followed by A2), Averaged across Subjects
and Conditions
The activation shown is the average across subjects and hemispheres of the voxels where, for each subject, the largest overall adaptation effect was
observed. Error bars indicate ±1 standard error of the mean (SEM). Shaded areas represent the period in which the adaptation numbers were repeat-
edly presented. The signal begins to rise as soon as numbers are presented following the rest period. Following a period of adaptation, a global
rebound is clearly seen, both at the transition between A1 and A2, as well as during the time periods where deviants begin to occur. *p < 0.05;
**p < 0.01; ***p < 0.001.

Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc. 295
 Neuron
A Magnitude Code in the Human Parietal Cortex

Table 1. Regions Showing an Overall fMRI Adaptation jects, the mean activation in three time windows of 12 s:
Effect (Decreasing Activation with Repetition of the during baseline, around the peak, and at the end of the
Same Approximate Quantity during the First 30 s of adaptation sequence. This analysis showed a significant
Periods A1 and A2) increase after rest (paired t test T(13) = 5.11, p < 0.001)
x y z Z Score Cortical Region and a significant decrease of the activation during the
last adaptation period compared to activity around the
34 54 44 4.87 Left parietal
peak (T(13) = 3.02, p < 0.01).
20 68 42
32 64 40 Rebound Effect
34 64 54 4.19 Right parietal The activation profile of parietal peaks in Figure 3 sug-
34 58 50 gests a rebound effect, with a sudden recovery of the
fMRI signal after a change in the adaptation numbers
46 46 40 (from A1 to A2), followed by a new period of adaptation
28 64 32 4.46 Left cerebellum during the subsequent repetitions of those new numbers.
36 72 26 4.64 Right cerebellum We compared, across subjects, the mean signal averaged
over the 12 s before the change in number with the mean
34 80 2 3.46 Left occipital signal average over the first 12 s after the change in num-
24 80 4 4.20 Right occipital ber. This analysis showed a statistically significant in-
36 46 20 3.60 Right inferior temporal crease in activation (T(13) = 2.14, p < 0.05). The adapta-
tion effect, a subsequent decrease in activation, was also
36 60 14 3.44 Left inferior temporal
highly significant (t test comparing the activity during the
54 24 30 3.51 Left DLPF first and the last 12 s of the initial deviant-free period of
38 26 20 4.50 Right DLPFC A2, T(13) = 4.93, p < 0.001).
To further explore this rebound effect across our exper-
14 2 2 3.43 Putamen
imental conditions, we then isolated, for each subject,
14 2 0 3.35 within the two IP regions identified by the group analysis,
the voxel where the overall largest fMRI rebound signal
was observed. The amount of rebound was quantified
isolated, for each subject, within the two intraparietal re- as the slope of a linear regression between the bold signal
gions identified by the group analysis, the voxel where and a linearly decreasing vector going from 1 to 1 over
the largest overall adaptation effect was observed (mean the 30 s deviant-free section of period A2. An ANOVA
and standard deviation [SD] of coordinates = 31(6), with A1 notation and A2 notation as dependent variables,
62(5), 48(5) in the left hemisphere and 32(4), 64(6), showed that rebound was invariant to notation (effect of
47(7) in the right hemisphere). Figure 3 shows the time A2 notation, left, F(1,13) = 0.4, p = 0.54; right, F(1,13) =
course of the BOLD signal in those voxels, averaged 0.09, p = 0.77; effect of A1 notation, left, F(1,13) = 0.05,
across hemispheres. The plot revealed that number- p = 0.83; right, F(1,13) = 0.13, p = 0.73) and also, crucially,
related adaptation in the parietal cortex is a slow, continuous invariant to changes in notation (A1*A2 notation interac-
process that takes several tens of seconds: activation con- tion, left, F(1,13) = 1.16, p = 0.3; right, F(1,13) = 0.54, p =
tinued to drop during the entire period where a given ap- 0.47). In other words, the increased activity due to a
proximate numerical magnitude was repeated, both after change in number was identical with or without a concom-
rest and after a recovery from adaptation due to a change itant change in notation (e.g., from Arabic digits to sets of
in number. The activation attained its peak around 10 s af- dots, from sets of dots to Arabic digits, or within each no-
ter the stimuli onset and then slowly decreased down to tation; see Figure 4). This observation provides a first
a minimum about 26 s after the peak. The adaptation effect piece of evidence for an abstract coding of approximate
was then tested statistically by comparing, across sub- number in the parietal cortex. At this level of analysis,

Figure 4. Left and Right Parietal Regions

Showing a Significant Effect of Adapta-
tion in the First 30 s of Periods A1 and A2
Plots show the amount of rebound activation in
those regions at the beginning of period A2, as
a function of the notation used for adaptation
stimuli during periods A1 and A2 (A = Arabic
digits, D = sets of dots). Error bars represent 1
SEM.

296 Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc.
Neuron
A Magnitude Code in the Human Parietal Cortex

our design, in order to spare experimental time, did not Table 2. Regions Showing a Distance-Dependent
incorporate a control condition where the transition be- Recovery from Adaptation during Presentation of the
tween A1 and A2 was not accompanied by a change in Numerical Deviants (Far > Close)
number. Nevertheless, on same-notation trials, the num- x y z Z Score Cortical Region
ber changed without any warning or any break in the stim-
34 62 60 3.75 Left parietal
ulus sequence. It thus seems unlikely that the rebound
would have occurred spontaneously at this precise mo- 26 50 40
ment in time, if it was not related specifically to the change 24 58 58
in number.
48 44 38 3.63 Right parietal

Response to Sparse Deviants 54 40 52

More unambiguous evidence for an abstract numerical 42 44 52
effect came from an analysis of the local event-related
36 22 4 3.79 Left insula
recovery from adaptation during the presentation of rare
(deviant) stimuli in the second part of periods A1 and A2. 34 20 6
Although the deviants always differed in number from 36 22 6 5.01 Right insula
the adaptation stimuli, the distance between the deviant
44 24 4
and the adaptation numbers was either very small (a dis-
tance of an average of 2 units ±1) or large (32 ± 1 units; 8 12 46 4.01 Medial precentral/cingulate
see Figure 1). We reasoned that presenting an occasional 4 12 30
deviant number should lead to a local recovery of fMRI re-
44 28 28 3.70 Right DLPF/precentral
sponses only if the deviant differed from the adaptation
value by a sufficiently large amount. Based on our prior re- 40 4 42
sults with nonsymbolic stimuli, which indicate a Gaussian 32 2 60
profile of recovery from adaptation in the hIPS, we ex-
30 48 18 3.44 Left anterior frontal
pected little or no recovery for close deviants but a large
recovery effect for far deviants. Accordingly, we looked 44 52 38 3.66 Left cerebellum
for regions that responded more to large distances than 28 60 28
to small ones. In a whole-brain search, this analysis re-
38 66 38 3.52 Right cerebellum
vealed mostly a bilateral fronto-parietal-cerebellar net-
work (Table 2 and Figure 5). No brain region was more 26 68 28
active for number change than for notation change. In 4 90 12 3.14 Primary visual cortex
the converse direction, notation change yielded a supple-
mental increased activation in bilateral inferior and middle
occipito-temporal cortex (Table 3), probably reflecting ad- had an effect on the activity of IP cortex. However, impor-
aptation to the very different overall shapes of digits and tantly, in the right hemisphere the effect of numerical dis-
dot patterns (Grill-Spector et al., 1999). tance did not interact either with deviant notation or with
Given that the goals of the present study were to char- adaptation notation (p > 0.1), indicating that the effect of
acterize the parietal number system, we focus here solely number change and of notation change are additive.
on the bilateral parietal clusters responding to number In the left hemisphere, a large change in number gener-
change. To study whether and how numerical coding in ally yielded stronger activation than a small change (see
the IPS was modulated by the format of presentation of Figure 5), but there was an exception when deviant dots
the numbers we isolated, for each subject, within the were presented among digits. This was reflected in a triple
two IPS regions identified by the group analysis, the voxel interaction of adaptation notation, deviant notation, and
where the largest fMRI response to numerical distance numerical distance (F(1,13) = 8.00, p < 0.05). When deviant
was found (mean and SD of the coordinates across sub- dots were presented among digits, there was a large
jects = 32(5), 57(6), 46(10) in the left hemisphere and recovery from adaptation even for small numerical dis-
46(4), 42(5), 47(6) in the right hemisphere). We then plot- tances, contrary to the symmetrical situation where devi-
ted activation in these voxels as a function of our experi- ant digits were presented among dots for which a small
mental conditions (see Figure 5). In both hemispheres, numerical distance did not yield recovery from adaptation.
overall activation was higher when the numerical distance In order to visualize this effect more easily, we calculated
between deviant and adaptation values was large than and plotted for each subject and hemisphere the size of
when it was small (main effect of numerical distance the distance effect (the difference between the activation
F(1,13) = 73.48, p < 0.000 and F(1,13) = 74.74, p < 0.000 of far versus close deviants; see Figure 5C). A paired t
for the left and right hemispheres, respectively). Moreover, test confirmed that in the left hemisphere the distance ef-
adaptation and deviant notation interacted (F(1,13) = 37.6, fect for dots amongst digits was smaller than in the right
p < 0.000 for the left and F(1,13) = 11.45, p < 0.005 for the hemisphere (T(13) = 1.87, p < 0.05), where it did not differ
right hemisphere), indicating that changing notation also in size from the other conditions. This effect was small but

Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc. 297
 Neuron
A Magnitude Code in the Human Parietal Cortex

Figure 5. Distance Effect in the Recovery from Adaptation to the Sparse Deviants
(A) Regions showing a distance-dependent response to the deviant stimuli.
(B) Amount of recovery to deviant stimuli at the peak voxels in left and right parietal cortex, as a function of adaptation notation, deviant notation, and
the numerical distance between the deviant and adaptation stimuli.
(C) Distance effect (activation to far close deviants) as a function of adaptation and deviant notation of the same peaks as in (B). (error bars represent
±1 SEM).

present in 10 out of 14 subjects (70%). It was only found by to simply attend to visually or auditory presented short
the region-of-interest method, as no other region showed sentences (‘‘The sailors threw the anchor into the bay’’).
a significant interaction of distance and notation in the The contrast between mental calculation and sentence
whole-brain analysis. In the discussion, we consider a ten- comprehension is a reproducible localizer of parietal acti-
tative interpretation of this unexpected asymmetric adap- vations (P. Pinel et al., 2006, poster presented to NUMBRA/
tation effect in terms of the degree of precision of the ESCOP Summer School ‘‘Neuroscience of number
internal representations of symbolic and nonsymbolic processing’’). We used this localizer to isolate, for each
quantities (Verguts and Fias, 2004). Most importantly, subject, within the parietal cluster from the random-effect
however, for the issue of notation-independent coding is group analysis of the localizer contrast, the peak voxel that
the fact that distance-dependent crossnotation fMRI re- best responded to mental calculation across modalities
covery was significant in both left and right parietal corti- (see Table 4 for the parietal coordinates from the random
ces when deviant digits were presented amongst dots effect analysis and Figure 6 for a picture of the whole
(t(13) = 3.99, p < 0.01 and t(13) = 3.64, p < 0.01, for the
left and right parietal cortices, respectively). These results
Table 3. Notation Change > Number Change
show that the magnitude code of the parietal cortices is
common to numerosities and numerical symbols. x y z Z Score Region
While the above analysis focused on the peak of overall 28 42 8 4.18 Right fusiform gyrus
recovery from adaptation, very similar results were ob- 32 56 10
served when we selected the parietal regions of interest
32 48 8 4.02 Left fusiform gyrus
on the basis of a completely independent data set. All
but one subject performed an additional final 5 min long 26 36 16
event-related fMRI scan during which they were asked 14 60 10 3.52 Left lingual gyrus
to perform mental calculations with visually or auditory
16 52 2
presented number words (‘‘Subtract five from eleven’’) or

298 Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc.
Neuron
A Magnitude Code in the Human Parietal Cortex

Table 4. Mental Arithmetic > Sentence Comprehension not interact with deviant notation or with adaptation nota-
(Parietal Clusters) tion, indicating that the code for number is notation invari-
ant. The triple interaction, however, approached signifi-
x y z Z Score Region
cance, again only in the left hemisphere (F(1,13) = 3.76,
45 48 48 4.88 Left parietal p = 0.07), in line with the previous analysis.
30 72 39
21 60 54
DISCUSSION
45 45 45 3.95 Right parietal
33 54 45 Our results indicate an important role for parietal cortex in
the coding of symbolic and nonsymbolic quantities. We in-
30 72 42
vestigated adaptation to number as well as recovery from
adaptation in the parietal cortex at two different temporal
scales. We now successively discuss those phenomena
circuit for the contrast calculation > sentence comprehen- and their implications for issues of semantic-level repre-
sion across modalities). The mean and standard deviation sentation and domain specificity in the number domain.
of the coordinates across subjects were of 37(8),
54(10), 46(7) in the left hemisphere, and 40(8), 49(14),
45(5) in the right hemisphere (for the one subject for whom Adaptation
we did not have the localizer scan we used the maxima We showed that adaptation of IPS activity to a fixed ap-
from the group analysis). Those coordinates fell quite proximate quantity is a slow, continuous process that
close to those isolated by the effect of numerical distance takes several tens of seconds: activation continued to
in the main adaptation experiment (indeed, 72% of the drop during the entire 30 s period where a given approxi-
voxels responding to a distance effect in left hemisphere mate numerical magnitude was repeated. Although our
and 53% in the right hemisphere were also active in the experiment was not designed to separate adaptation ef-
calculation-sentence contrast), and the profile of activa- fects specifically related to number from those associated
tion across conditions was very similar (compare Figures with low-level visual repetition effects, the fact that activity
5 and 6). An ANOVA confirmed that in both hemispheres eventually decreased following repetition of the same ap-
the activation increased with the numerical distance be- proximate numbers is not trivial, given the large variability
tween deviant and adaptation values (F(1,13) = 5.6, p < in the visual properties of the stimuli during the repetition
0.05 and F(1,13) = 10.37, p < 0.01 for the left and right phase. For digits, we randomized across trials not only
hemispheres, respectively). Adaptation and deviant nota- the identity of the stimuli (17-18-19 or 47-48-49), but
tion interacted (F(1,13) = 13.9, p < 0.05 and F(1.13) = 14.4, also their size and position. For nonsymbolic stimuli, the
p < 0.05 for the left and right hemispheres, respectively), number, size, and position of the dots also varied from trial
indicating that changing notation has also an effect on to trial. Given such variability, it might not be surprising
the activity of IP cortex. Finally, numerical distance did that activation takes a long time to adapt and stabilize.

Figure 6. Localization and Response Profile of Parietal Regions Involved in Calculation

Left, brain regions with greater activity during mental calculation than during sentence comprehension. Insets show the activation of the peak voxels
in parietal cortex for calculation (Calc.) and sentence comprehension (S.), showing identical responsivity whether the stimuli were presented in the
visual or auditory modalities. Right, amount of recovery to deviant stimuli in the same parietal voxels as a function of adaptation notation, deviant
notation, and numerical distance between the deviant and adaptation stimuli (error bars represent ±1 SEM).

Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc. 299
 Neuron
A Magnitude Code in the Human Parietal Cortex

The slow adaptation dynamics may shed some light on cortex, the present study also revealed extended dis-
a recent controversy surrounding numerical adaptation. tance-related activation of bilateral prefrontal and inferior
Our group previously used fMRI adaptation to demon- frontal cortices, whereas only two small parietal clusters
strate approximate coding of numerosity in the parietal were observed in our previous study (Piazza et al.,
cortex (Piazza et al., 2004), and this was later replicated 2004). A possible explanation for this difference across
by another group (Cantlon et al., 2006), but others re- studies might be that number changes were much more
ported the absence of such an effect (Shuman and Kanw- evident in the present study, due both to the use of a large
isher, 2004). Crucially, the paradigms differed. In their ad- distance and to their presentation in digital format. The
aptation experiment (experiment 2), Shuman and detection of a large (semantic) difference might have
Kanwisher (2004) used a block design with blocks of 16 amplified the activation of a prefrontal cingulate atten-
s during which they either repeatedly presented the tional-arousal system. In our previous study, by contrast,
same number or very different numbers of objects. Piazza number changes were not mentioned in the instructions
et al. (2004), however, used an event-related design and and remained undetected by most subjects. Differences
analyzed activity triggered by rare deviant numbers within in conscious awareness of changes may thus explain
long blocks of several minutes with a fixed numerosity the difference in the extent of the distance related activa-
(Cantlon et al. [2006] then adopted this design). According tion across studies. It is possible, however, that beyond
to present results, the activation after 16 s of repetition of parietal cortex, other regions contain populations of
the same number (the length of a block in Shuman and neurons that also code for number, as observed in the
Kanwisher’s study) is still about 85% of its peak height. prefrontal cortex of macaque monkeys (Nieder et al.,
The maximal reduction of the signal, down to 40% of the 2002).
initial peak height, is seen only 36 s after the onset of The results in the Arabic-to-Arabic condition also ex-
stimulation. Thus, the short adaptation period used by tended earlier work from our laboratory on subliminal rep-
Shuman and Kanwisher (2004) might explain why they etition priming for Arabic digits and number words (Nacc-
failed to observe an adaptation effect. Furthermore, ache and Dehaene, 2001). In a number comparison task
Shuman and Kanwisher (2004) examined only the mean where each target was preceded by a subliminal prime,
activity averaged over a whole block of 16 s, which might Naccache and Dehaene (2001) showed reduced activa-
have further reduced the chances to detect a small adap- tion (repetition suppression) confined to bilateral parietal
tation effect. According to the present results, this effect regions during repetition priming (e.g., prime 1 followed
presumably might have become barely visible at the end by target 1) compared to nonrepetition trials (e.g., prime
of the 16 s block. 4, target 1). Response time measures indicate that such
priming varies continuously with the distance between
Rebound Effect the prime and target (Koechlin et al., 1999; Reynvoet
When the adaptation numerosities changed abruptly (e.g., et al., 2002). The present results are the first to show
from 17, 18, 19 to 47, 48, 49), we observed a durable re- such numerical distance-based priming in fMRI. Note
bound in the fMRI signal. Since this rebound was identical that we carefully selected the adaptation and deviant
whether there was a concomitant change in notation or sets so that, on both close and far deviant trials, there
not, it is already suggestive of a notation-invariant code was an equal amount of change in the physical properties
in the IPS. of the stimuli for the two distance conditions (e.g., adapta-
tion to 17, 18, 19, deviants 20 versus 50). Thus, the ob-
Response to Close and Far Deviants served recovery of adaptation in parietal cortex can only
In the critical trials, we measured the presence of a local be attributed to semantic proximity, not to visual resem-
recovery from adaptation when rare deviant stimuli were blance or to generic attentional mechanisms.
presented among adaptation stimuli. Furthermore, we Third, and crucially, we observed crossnotation adapta-
compared close deviants, which fell when within the tion and recovery, particularly in the right parietal cortex,
known coarseness of numerosity coding in the IPS and supporting the idea that shared neural populations
where we therefore expected continuing adaptation, encode nonsymbolic quantities and symbolic stimuli
with far deviants, for which we expected recovery from (Dehaene et al., 2003; Verguts and Fias, 2004). Converg-
adaptation. The results confirmed to this prediction. In ing evidence for a notation-independent code for number
the dots-to-dots condition (deviant dots presented among in the parietal cortex comes from a study showing overlap-
dots), they reproduced earlier findings of numerical adap- ping activation of the IPS when participants performed
tation and distance-related recovery for nonsymbolic sets mental arithmetic on both digits and dice dot patterns
of dots (Piazza et al., 2004; Cantlon et al., 2006). The pres- (Venkatraman et al., 2005). In principle, however, overlap
ent activation extended further laterally and anteriorily in of activation need not necessarily imply shared neural
parietal cortex, a difference which may merely be due to substrates. fMRI adaptation by contrast, implies that the
interindividual difference between subjects participating neural populations that were adapted to one notation gen-
in the two experiments (P. Pinel et al., 2006, poster pre- eralized their responses to the other notation, thus provid-
sented to NUMBRA/ESCOP Summer School ‘‘Neurosci- ing a more valid inference for a shared notation-invariant
ence of number processing’’). However, beyond parietal mechanism.

300 Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc.
Neuron
A Magnitude Code in the Human Parietal Cortex

Note that due to the limitation of the present fMRI reso- population of neurons would be adapted. Its breadth
lution we cannot exclude that within the region individu- might be sufficient for a transfer of adaptation to nearby
ated in the present study there might be separate subas- Arabic numerals, especially given that a range of adapta-
semblies of neurons that each code for a given input tion values were used (e.g., 17, 18, 19, followed by deviant
format but are highly interconnected. According to this 20). However, most of the broad population code for the
scenario, under the present experimental circumstances, corresponding dot patterns would not have been adapted,
where notations were mixed in the same runs, activation of thus resulting in a large recovery in the digits-to-dots
one given population (say, for example, coding for dots) condition on both close and far trials. Hence, our unex-
would quickly spread to the other population (say, for ex- pected finding can in fact be seen as a natural prediction
ample, coding for digits), thus leading to crossnotation of the independently motivated Verguts and Fias (2004)
adaptation. With the present experiment we cannot disen- model.
tangle between these two possibilities, which, de facto, do An alternative scenario which could also lead to the
not differ substantially at the population level. This issue present observations supposes that what differs between
might be ultimately be resolved only by means of higher- symbolic and nonsymbolic representations of numbers in
resolution fMRI (Grill-Spector et al., 2006) or by single- the left hemisphere is the relative number of neurons that
unit recordings. Interestingly, preliminary results on sin- code for each notation. In order to account for our obser-
gle-unit recordings in macaque monkeys show that, after vations, one would need to assume that, for any given
the animal has undergone extensive training associating number, there are more neurons coding for dots arrays
sets of dots with Arabic digits, there are neurons in the than for Arabic digits. The present experiment does not
IPS that code for the preferred numerical value irrespective allow distinguishing between the two alternatives. For
of whether it was presented by dot displays or numerals the time being, given that fMRI can only observe activity
(I. Diester and A. Nieder, 2006, FENS Abstr., abstract). pooled across large number of neurons, we should refrain
These results suggest that even the macaque brain can from further speculation on this point and note simply that
integrate numerical information across symbolic and non- our results point to a population code invariant to the
symbolic notations at the level of the single neuron. notation used for number presentation.

Asymmetry in Recovery from Adaptation Hemispheric Asymmetry

We unexpectedly observed an asymmetry in fMRI re- Existing models of number processing do not explicitly
sponses to deviant stimuli in the left parietal lobe; there address the question of the hemispheric asymmetry in
was a normal recovery when a distant digit was presented the precision of the neural code for number. Interestingly,
amongst dots, but there was an abnormal recovery inde- the asymmetry in crossnotation priming was only found in
pendent of numerical distance whenever deviant dots, the left hemisphere, which may suggest, according to the
whether numerically close or far, were presented among Verguts and Fias (2004) model, that only the left parietal
digits. Although this effect was small and is therefore in representation has been affected by the acquisition of
need of replication, an interesting tentative interpretation number symbols and has acquired a refined precision,
is in terms of the precision of numerical coding. Dehaene while the right parietal representation has kept a coarse
(1997) suggested and Verguts and Fias (2004) demon- representation for both symbolic and nonsymbolic nota-
strated in a neural network simulation that the neural tions. Several previous results bring some support to
code for symbolic stimuli might be more precise than this hypothesis. In a developmental fMRI study, Rivera
the neural code for nonsymbolic stimuli. Crucially, in Ver- et al. (2005) found that activation during a calculation
guts and Fias’s network, each number neuron has a pre- task with symbolic digits increased with age in the left pa-
ferred numerosity which is identical for symbolic and non- rietal and left occipito-temporal cortices, but not in the
symbolic numerical displays. However, the neuron has a right parietal lobe. Along this same line, Cantlon et al.
broad tuning curve over numerosities when those are pre- (2006) reported that the only region showing overlapping
sented as dot patterns (reproducing the electrophysiolog- activation in preschool children and adults during pro-
ical findings of Nieder et al. [2002]), and a very sharp, cessing of numerosity was the right IPS. Both results
though still distance-dependent tuning curve when num- together suggest that the right IPS is predominant for non-
bers are presented in symbolic format. How would such symbolic numerosity coding in childhood and that the left
a population code explain the observed asymmetrical IPS develops as a function of experience with numerical
priming? During adaptation to dots, given the neuron’s symbols. In adults, Piazza et al. (2004) also observed
broad tuning curves, a large population of number neu- that although numerical adaptation was present in both
rons would adapt. The adapted population, being broadly left and right IPS, the precision of the representation, as
tuned, would clearly include the narrow population of measured by the Gaussian tuning curve for recovery of
neurons responsive to the close symbolic numerals pre- adaptation, tended to be higher in the left than in the right
sented as deviants, hence the transfer of adaptation to IPS. Finally, the two hemispheres seems to be differen-
close symbolic numerals. In the converse direction, how- tially involved in approximate and exact numerical judg-
ever, adaptation to digits would not lead to the adaptation ments: approximate judgments (in both the visual and
of the population code for dots. Only a relatively narrow auditory domain) correlate with stronger activation in the

Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc. 301
 Neuron
A Magnitude Code in the Human Parietal Cortex

right than in the left IPS, while exact judgments correlate (in the cases of brain lesions studies) has been found con-
with more activation in the left versus right IPS (Piazza sistent across modalities, notations (pictures and words),
et al., 2006). Along the same line, neuropsychologically, and tasks (naming, matching, reading; Caramazza and
a superiority for the left hemisphere in exact calculation Shelton, 1998; Gorno-Tempini et al., 1998; Perani et al.,
and for the right hemisphere in approximate calculation 1999).
has been reported (Cohen and Dehaene, 1996; Dehaene
and Cohen, 1991). Moreover, transcranial magnetic stim- The Issue of Domain Specificity
ulation suggests that it is sufficient to stimulate the left pa- Shuman and Kanwisher (2004) analyzed several ROIs
rietal cortex to produce deficits for precisely coding num- from a mental calculation localizer similar to the one
bers, while it is necessary to stimulate bilaterally to disrupt used in the present study and failed to observe a stronger
approximate numerical judgments (Andres et al., 2005). response for number tasks than for closely matched color
Imaging studies also tend to show right-lateralized parietal tasks. However, this logic supposes that an entire patch of
activation in tasks that involve comparisons and left later- intraparietal cortex is specialized for numerical process-
alized activation when retrieving of exact arithmetical facts ing, a hypothesis that we find unnecessary. The present
(Chochon et al., 1999; Dehaene, 1996; Pinel et al., 2001; results indicate a response to number change, with appro-
Rickard et al., 2000). Outside the number domain, a similar priate controls to suggest that this response can only
hemispheric asymmetry has been proposed, whereby the come from neural populations coding for an abstract rep-
left hemisphere would be superior for exact or categorical resentation of numbers. They are, however, completely
judgments, and the right for approximate, continuous or neutral relative to the issue of whether, within the same
coordinate-based judgments (Kimura, 1996; Kosslyn voxels, there might be other neural populations coding
et al., 1989; McGlone and Davidson, 1973; Pasini and e.g., for color, size, space, time, or other such parameters.
Tessari, 2001; Piazza et al., 2004; Warrington and James, Indeed, previous work has reported an important overlap
1967; Young and Bion, 1979). in the neural coding of number and object size (Cohen
While there is thus tentative support for a difference in Kadosh et al., 2005; Pinel et al., 2004). In theory, deciding
the precision of left and right hemisphere number codes, whether a given region is ‘‘specific’’ for any given category
further work will be needed to directly establish the tuning would require a systematic comparison of the target cat-
curves for Arabic digits and for dot patterns. The fMRI egory (e.g., number) against a potentially infinite list of
adaptation method could again be used for that purpose. alternatives. Moreover, it is not clear which level of spatial
While we used here only two levels of distance (close and precision is needed in order to test claims for specificity
far deviants), a more continuous variation of deviancy, as (voxels, columns, or single neurons; see Grill-Spector
in our previous work (Piazza et al., 2006), would allow one et al. [2006]). Even at the level of individual neurons, recent
to trace the precision of the metric of proximity between electrophysiological data shows that a subset of number-
numbers, separately for symbolic and nonsymbolic stim- coding neurons in the macaque parietal cortex also re-
uli, and for the left and right hemispheres, thus directly sponded to flow field stimuli in a direction selective fashion
testing the tuning curves predicted by Verguts and Fias (I. Diester and A. Nieder, 2006, FENS Abstr., abstract).
(2004). These results suggest that the question of domain speci-
ficity might be an ill-posed question or, at the very least,
Criteria for Semantic-Level Representation one very difficult to answer with fMRI alone.
The present experiment used two independent criteria,
both of which associate intraparietal cortex with a seman- The Symbol-Grounding Problem
tic level of representation: semantic metric (activation A classical problem in semantics is the ‘‘symbol-ground-
varies with proximity of meaning, here defined by numer- ing’’ problem of attaching meaning to the arbitrary shapes
ical distance) and notation independence (activation is and sounds selected, in a given culture, to serve as written
identical across major changes in input notation). We pro- and spoken words (Harnad, 1999). Harnad proposed that
pose that those two criteria could be used to define se- symbolic representations are grounded bottom-up in
mantic-level representation in domains other that numer- nonsymbolic representations of two kinds: (1) ‘‘iconic rep-
ical cognition (Maess et al., 2002). The IPS voxels isolated resentations,’’ which are analogs of the proximal sensory
in the independent mental calculation localizer were also projections of distal objects and events, and (2) ‘‘categor-
activated by both visual and auditory presented stimuli ical representations,’’ which are learned or innate feature-
(see insets in Figure 6). This is further evidence for con- detectors that pick out the invariant features of object and
vergence toward a representation of numerical quantity event categories from their sensory projections. Elemen-
that is independent from the task (mental calculation or tary symbols (such for example ‘‘2’’ or ‘‘nine’’) are the
passive viewing), from the modality of stimuli presentation names of these object and event categories, assigned on
(auditory or visual), and from the notation (symbolic or the basis of their (nonsymbolic) categorical representa-
nonsymbolic). In this respect, our findings parallel those tions. Harnad’s second case may provide a solution to
on the representation of object categories such as animals the symbol-grounding problem for numbers. Our results
or tools in the occipito-temporal cortex, where activation show that, at least in the adult brain, numerical symbols
(in the cases of brain imaging studies) and impairment and nonnumerical numerosities converge onto shared

302 Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc.
Neuron
A Magnitude Code in the Human Parietal Cortex

neural representations. Perhaps we attach meaning to range as the adaptation, for example 20 among 17, 18, and 19), or a
symbols by physically linking populations of neurons sen- large deviation (ratio of 2.4, for example 20 among 47, 48, and 49). Fur-
thermore, items in the deviant sets could be of the same notation or
sitive to symbol shapes to preexisting neural populations
of a different notation of the adaptation sets, thus defining fully orthog-
holding a nonsymbolic representation of the correspond- onal notation-change and number-change factors (see Figure 1 for
ing preverbal domain (e.g., numerosity). Behavioral find- examples of stimuli).
ings indeed suggest that the nonsymbolic numerosity rep- The experiment was divided into four runs. Each run consisted in 206
resentation is present in infants and adults prior to the stimuli and started with a 12 s resting period, during which a small cen-
acquisition of number words and symbols (Brannon, tered fixation cross, which remained visible throughout all the experi-
ment, was presented on the screen. Each run consisted in two blocks
2006; Feigenson et al., 2004; Pica et al., 2004) and is
separated by a 12 s rest period. Finally, each block comprised two dif-
thus available to serve as the foundation for symbol
ferent adaptation sequences (A1 and A2; see Figure 2). The first 25 and
grounding in the number domain. An unresolved difficulty, the last 6 stimuli of each sequence were adaptation stimuli only, while
however, is that symbolic numerals do not merely refer to in the central part of each sequence, deviants (12 stimuli overall)
approximate numerosities but ultimately come to acquire appeared among the adaptation stimuli (60 overall). A2 followed A1
exact meanings (e.g., exactly seventeen), which do not without a break.
seem available in the absence of language and education. To avoid decision and response confounds, participants were sim-
ply instructed to fixate and to pay attention to the quantity conveyed
The nature of this ‘‘crystallization’’ of exact number con-
by the stimuli. They were informed that they would be shown quantities
cepts remains an unsolved issue, for which the present in different formats and that their approximate values would be 20
methods might ultimately turn out to be useful. and 50. Moreover, immediately prior to the scanning session, sub-
jects were shown approximately four exemplars of each numerosities
EXPERIMENTAL PROCEDURES (17:20 and 47:50 dots) and informed about their approximate range
(20 and 50, respectively) in order to calibrate them.
Participants Thirteen subjects (out of fourteen) performed an additional 5 min
Fourteen healthy human adults participated in the study after giving long event-related fMRI scan for isolate individual neural correlates
written informed consent. All were right handed (Edimburgh Inventory) of mental calculation. This short functional localizer sequence was rou-
and had normal or corrected-to-normal vision. The study was ap- tinely used to map various individual cortical networks involved in mo-
proved by the regional ethical committee (Hopital de Bicêtre, France). tor action, reading, language comprehension, and mental calculation.
Subjects were engaged in various tasks such as left or right clicking
Stimuli and Procedure after audio or video instruction, mental calculation (subtraction) after
Stimuli were dot patterns and Arabic digits. Both were presented for video or audio instruction (‘‘Subtract five from eleven’’), sentence com-
150 ms at a constant rate of one every 1200 ms, white on a black back- prehension from audio or visual modality (‘‘The sailors threw the anchor
ground, varying in size and position within an invisible circle of 5 ra- into the bay’’), and passive viewing of horizontal or a vertical checker-
dius around fixation. Dot patterns were designed so that, aside from boards. For this paper, we only considered the calculation task and
the number change, all deviant stimuli were equally novel with respect used sentence comprehension as a control. For the mental calculation
to all physical parameters. In half of the blocks, total luminance and task, subjects were asked to perform the operation silently (‘‘in their
total occupied area (extensive parameters) were equated across the head’’) and not to utter the result, while for the sentence comprehen-
deviant stimuli. This means that dots in the deviant number 50 had sion they were asked to simply listen (or read) attentively.
on average smaller individual item sizes and smaller inter-item spac-
ing. However, the latter parameters (intensive parameters) were varied fMRI Parameters
randomly and equated on average across the adaptation stimuli: ad- The experiments were performed on a 3T fMRI system (Bruker, Ger-
aptation stimuli were generated with item size and inter-item spacing many). Functional images sensitive to blood oxygen level-dependent
values drawn randomly from fixed distributions that spanned all the contrast were obtained with a T2*-weighted gradient echo-planar im-
range of values used for the deviant stimuli. As a result, all of the pa- aging sequence (TR [repetition time] = 2.4 s, TE [echo time] = 40 ms,
rameter values that occurred in the deviants had already been pre- angle = 90 , FOV [field of view] 192 3 256 mm, matrix = 64 3 64).
sented equally often during adaptation and were therefore equally non- The whole brain was acquired in 40 slices with a slice thickness of
novel. Therefore, the only novel aspect of the deviant stimuli was 3 mm. High-resolution images (3D gradient echo inversion-recovery
number. In the other half of the blocks, the parameters were controlled sequence, TI [inversion time] = 700 mm, TR = 2400 ms, FOV = 192 3
in a symmetric fashion (e.g., the extensive parameters were equated 256 3 256 mm, matrix = 256 3 128 3 256, slice thickness = 1 mm)
across the adaptation and the intensive parameters across the deviant were also acquired.
stimuli). An automated program (freely available on our website [http://
www.unicog.org/main/pages.php?page=Documentation]) generated Image Processing and Statistical Analysis
random configurations within those constraints, so that stimuli were Data were analyzed with SPM2 (http://www.fil.ion.ucl.ac.uk/spm). The
never repeated identically during the experiment. (see also Piazza first four volumes were discarded for each experiment. All other vol-
et al. [2004]). Arabic digits were presented in Arial font and varied, umes were realigned using the first volume as reference, normalized
from trial to trial, in size (from 16 to 33 point size) and position (within to the standard template of the Montreal Neurological Institute using
a circle of 5 degrees radius around fixation). an affine transformation, resampled (2 3 2 3 2 mm), spatially
There were two types of periods, those where the majority of the smoothed (6 mm), and low-pass (4 s) filtered. Activations for the
stimuli were sets of dots and those were they were Arabic numbers, main experiment were modeled by a linear combination of (1) eight
both with a fixed (though approximate) quantity. Occasionally, a devi- functions derived by convolution of the standard hemodynamic func-
ant stimulus occurred randomly, with the constraint that two succes- tion with the known onsets of the different types of deviants (2 3 2 3
sive deviants were separated by at least three and at most seven ad- 2 design with factors of distance [close, far], deviant notation [dots,
aptation stimuli. Adaptation numbers varied randomly between 17, 18, digits], and adaptation notation [dots, digits]) and (2) sixteen linearly
and 19 in half of the experiment, and between 47, 48, and 49 in the decreasing regressors, modeling the adaptation in the first part of
other half. Deviant stimuli always differed from the adaptation number, each adaptation period (A1 and A2). Parameters of no interest were
but there was either a small deviation (the deviant was in the same also entered, coding for the horizontal and vertical location of the dot

Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc. 303
 Neuron
A Magnitude Code in the Human Parietal Cortex

and digit stimuli on screen, and for the size of the digits. Random effect Dehaene, S., and Cohen, L. (1991). Two mental calculation systems: A
analyses were then applied to several contrasts: the main effects of case study of severe acalculia with preserved approximation. Neuro-
number and notation change in deviants, and the main effect of the psychologia 29, 1045–1074.
regressors modeling the beginning of each adaptation period. Dehaene, S., and Cohen, L. (1997). Cerebral pathways for calculation:
For the additional functional localizer, activations were modeled by double dissociation between rote verbal and quantitative knowledge
eight functions derived by convolution of the standard hemodynamic of arithmetic. Cortex 33, 219–250.
function with the known onsets of the different types of task and mo-
dality trials. In the present analysis, we looked for regions showing in- Dehaene, S., Dupoux, E., and Mehler, J. (1990). Is numerical compar-
creased activation for subtraction relative to sentence comprehension, ison digital: analogical and symbolic effects in two-digit number com-
in both the visual and auditory modality (random effect analyses of the parison. J. Exp. Psychol. Hum. Percept. Perform. 16, 626–641.
contrast looking for the main effect of calculation across modalities). Dehaene, S., Dehaene-Lambertz, G., and Cohen, L. (1998). Abstract
Data for both experiments are reported at p < 0.05 corrected for mul- representations of numbers in the animal and human brain. Trends
tiple comparisons at the cluster level, p < 0.01 at the voxel level. Neurosci. 21, 355–361.
Dehaene, S., Piazza, M., Pinel, P., and Cohen, L. (2003). Three parietal
ACKNOWLEDGMENTS circuits for number processing. Cogn. Neuropsychol. 20, 487–506.
Eger, E., Sterzer, P., Russ, M.O., Giraud, A.L., and Kleinschmidt, A.
This work was supported by INSERM, CEA, a Marie Curie fellowship (2003). A supramodal number representation in human intraparietal
of the European Community QLK6-CT-2002-51635 (M.P.), and a cortex. Neuron 37, 719–725.
McDonnell Foundation centennial fellowship (S.D.).
Feigenson, L., Dehaene, S., and Spelke, E. (2004). Core systems of
number. Trends Cogn. Sci. 8, 307–314.
Received: June 30, 2006
Revised: October 19, 2006 Gorno-Tempini, M.L., Price, C.J., Josephs, O., Vandenberghe, R.,
Accepted: November 21, 2006 Cappa, S.F., Kapur, N., and Frackowiak, R.S. (1998). The neural sys-
Published: January 17, 2007 tems sustaining face and proper-name processing. Brain 121, 2103–
2118.
REFERENCES Grill-Spector, K., Kushnir, T., Edelman, S., Avidan, G., Itzchak, Y., and
Malach, R. (1999). Differential processing of objects under various
Andres, M., Seron, X., and Olivier, E. (2005). Hemispheric lateralization viewing conditions in the human lateral occipital complex. Neuron
of number comparison. Brain Res. Cogn. Brain Res. 25, 283–290. 24, 187–203.
Ansari, D., Dhital, B., and Siong, S.C. (2006). Parametric effects of nu- Grill-Spector, K., Sayres, R., and Ress, D. (2006). High-resolution im-
merical distance on the intraparietal sulcus during passive viewing of aging reveals highly selective nonface clusters in the fusiform face
rapid numerosity changes. Brain Res. 1067, 181–188. area. Nat. Neurosci. 9, 1177–1185.
Brannon, E.M. (2006). The representation of numerical magnitude. Halpern, C.H., Glosser, G., Clark, R., Gee, J., Moore, P., Dennis, K.,
Curr. Opin. Neurobiol. 16, 222–229. McMillan, C., Colcher, A., and Grossman, M. (2004). Dissociation of
Buckley, P.B., and Gillman, C.B. (1974). Comparison of digits and dot numbers and objects in corticobasal degeneration and semantic
patterns. J. Exp. Psychol. 103, 1131–1136. dementia. Neurology 62, 1163–1169.

Butterworth, B. (1999). The Mathematical Brain (London: Macmillan). Harnad, S. (1999). The symbol grounding problem. Physica D. 42,
335–346.
Butterworth, B., Cappelletti, M., and Kopelman, M. (2001). Category
specificity in reading and writing: the case of number words. Nat. Kimura, D. (1996). Dual functional asymmetry of the brain in visual
Neurosci. 4, 784–786. perception. Neuropscyhologia 4, 275–285.
Cantlon, J.F., Brannon, E.M., Carter, E.J., and Pelphrey, K.A. (2006). Koechlin, E., Naccache, L., Block, E., and Dehaene, S. (1999). Primed
Functional imaging of numerical processing in adults and 4-y-old chil- numbers: Exploring the modularity of numerical representations with
dren. PLoS Biol. 4, e125. 10.1371/journal.pbio.0040125. masked and unmasked semantic priming. J. Exp. Psychol. Hum. Per-
Caramazza, A., and Shelton, J.R. (1998). Domain-specific knowledge cept. Perform. 25, 1882–1905.
systems in the brain: the animate-inanimate distinction. J. Cogn. Neu- Kosslyn, S.M., Koenig, O., Barrett, A., Cave, C.B., Tang, J., and Gabri-
rosci. 10, 1–34. eli, J.D. (1989). Evidence for two types of spatial representations:
Chochon, F., Cohen, L., van de Moortele, P.F., and Dehaene, S. (1999). hemispheric specialization for categorical and coordinate relations.
Differential contributions of the left and right inferior parietal lobules to J. Exp. Psychol. Hum. Percept. Perform. 15, 723–735.
number processing. J. Cogn. Neurosci. 11, 617–630. Maess, B., Friederici, A.D., Damian, M., Meyer, A.S., and Levelt, W.J.
Cipolotti, L., Butterworth, B., and Denes, G. (1991). A specific deficit (2002). Semantic category interference in overt picture naming: sharp-
for numbers in a case of dense acalculia. Brain 114, 2619–2637. ening current density localization by PCA. J. Cogn. Neurosci. 14,
455–462.
Cohen, L., and Dehaene, S. (1996). Cerebral networks for number pro-
cessing: evidence from a case of posterior callosal lesion. Neurocase McGlone, J., and Davidson, W. (1973). The relation between cerebral
2, 155–174. speech laterality and spatial ability with special reference to sex and
hand preference. Neuropsychologia 11, 105–113.
Cohen Kadosh, R., Henik, A., Rubinsten, O., Mohr, H., Dori, H., van de
Ven, V., Zorzi, M., Hendler, T., Goebel, R., and Linden, D.E. (2005). Are Naccache, L., and Dehaene, S. (2001). The priming method: imaging
numbers special? The comparison systems of the human brain inves- unconscious repetition priming reveals an abstract representation of
tigated by fMRI. Neuropsychologia 43, 1238–1248. number in the parietal lobes. Cereb. Cortex 11, 966–974.
Dehaene, S. (1996). The organization of brain activations in number Nieder, A., Freedman, D.J., and Miller, E.K. (2002). Representation of
comparison: event-related potentials and the additive-factors the quantity of visual items in the primate prefrontal cortex. Science
methods. J. Cogn. Neurosci. 8, 47–68. 297, 1708–1711.
Dehaene, S. (1997). The Number Sense (New York: Oxford University Pasini, M., and Tessari, A. (2001). Hemispheric specialization in quan-
Press). tification processes. Psychol. Res. 65, 57–63.

304 Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc.
Neuron
A Magnitude Code in the Human Parietal Cortex

Perani, D., Schnur, T., Tettamanti, M., Gorno-Tempini, M., Cappa, specialization in the left inferior parietal cortex. Cereb. Cortex 15,
S.F., and Fazio, F. (1999). Word and picture matching: a PET study 1779–1790.
of semantic category effects. Neuropsychologia 37, 293–306. Shepard, R.N., Kilpatrick, D.W., and Cunningham, J.P. (1975). The in-
Piazza, M., and Dehaene, S. (2004). From number neurons to mental ternal representation of numbers. Cognit. Psychol. 7, 82–138.
arithmetic: the cognitive neuroscience of number sense. In The Cogni- Shuman, M., and Kanwisher, N. (2004). Numerical magnitude in the
tive Neurosciences, 3rd Edition, M. Gazzaniga, ed. (New York: Nor- human parietal lobe; tests of representational generality and domain
ton), pp. 865–875. specificity. Neuron 44, 557–569.
Piazza, M., Izard, V., Pinel, P., Le Bihan, D., and Dehaene, S. (2004). Thioux, M., Pillon, A., Samson, D., de Partz, M.-P., Noel, M.-P., and
Tuning curves for approximate numerosity in the human intraparietal Seron, X. (1998). The isolation of numerals at the semantic level.
sulcus. Neuron 44, 547–555. Neurocase 4, 371–389.
Piazza, M., Mechelli, A., Price, C., and Butterworth, B. (2006). Exact Thioux, M., Pesenti, M., De Volder, A., and Seron, X. (2002). Cate-
and approximate judgements of visual and auditory numerosity: an gory-specific representation and processing of numbers and
fMRI study. Brain Res. 1106, 177–188. animal names across semantic tasks: a PET study. Neuroimage
Pica, P., Lemer, C., Izard, W., and Dehaene, S. (2004). Exact and ap- 13, S617.
proximate arithmetic in an Amazonian indigene group. Science 306, Venkatraman, V., Ansari, D., and Chee, M.W. (2005). Neural correlates
499–503. of symbolic and nonsymbolic arithmetic. Neuropsychologia 43,
Pinel, P., Dehaene, S., Riviere, D., and LeBihan, D. (2001). Modulation 744–753.
of parietal activation by semantic distance in a number comparison Verguts, T., and Fias, W. (2004). Representation of number in animals
task. Neuroimage 14, 1013–1026. and humans: a neural model. J. Cogn. Neurosci. 16, 1493–1504.
Pinel, P., Piazza, M., Le Bihan, D., and Dehaene, S. (2004). Distributed Warrington, E.K., and James, M. (1967). Tachistoscopic number
and overlapping cerebral representations of number, size, and lumi- estimation in patients with unilateral cerebral lesions. J. Neurol. 30,
nance during comparative judgments. Neuron 41, 983–993. 468–474.
Reynvoet, B., Brysbaert, M., and Fias, W. (2002). Semantic priming in Young, A.W., and Bion, P.J. (1979). Hemispheric laterality effects in the
number naming. Q. J. Exp. Psychol. A 55, 1127–1139. enumeration of visually presented collections of dots by children.
Rickard, T.C., Romero, S.G., Basso, G., Wharton, C., Flitman, S., and Neuropsychologia 17, 99–102.
Grafman, J. (2000). The calculating brain: an fMRI study. Neuropsy- Zamarian, L., Karner, E., Benke, T., Donnemiller, E., and Delazer, M.
chologia 38, 325–335. (2006). Knowing 7 3 8, but not the meaning of ‘elephant’: Evidence
Rivera, S.M., Reiss, A.L., Eckert, M.A., and Menon, V. (2005). Develop- for the dissociation between numerical and nonnumerical semantic
mental changes in mental arithmetic: evidence for increased functional knowledge. Neuropsychologia 44, 1708–1723.

Neuron 53, 293–305, January 18, 2007 ª2007 Elsevier Inc. 305