www.elsevier.com/locate/ynimg

NeuroImage 24 (2005) 369–383

The neural basis for novel semantic categorization$

Phyllis Koenig,a,* Edward E. Smith,b Guila Glosser,a Chris DeVita,a Peachie Moore,a

Corey McMillan,a Jim Gee,c and Murray Grossmana

aDepartment of Neurology, University of Pennsylvania, USAbDepartment of Psychology, University of Michigan, USAcDepartment of Radiology, University of Pennsylvania, 19104-4283, USA

Received 5 April 2004; revised 23 July 2004; accepted 30 August 2004

Available online 10 December 2004

We monitored regional cerebral activity with BOLD fMRI during

acquisition of a novel semantic category and subsequent categorization

of test stimuli by a rule-based strategy or a similarity-based strategy.

We observed different patterns of activation in direct comparisons of

rule- and similarity-based categorization. During rule-based category

acquisition, subjects recruited anterior cingulate, thalamic, and

parietal regions to support selective attention to perceptual features,

and left inferior frontal cortex to helps maintain rules in working

memory. Subsequent rule-based categorization revealed anterior

cingulate and parietal activation while judging stimuli whose con-

formity with the rules was readily apparent, and left inferior frontal

recruitment during judgments of stimuli whose conformity was less

apparent. By comparison, similarity-based category acquisition showed

recruitment of anterior prefrontal and posterior cingulate regions,

presumably to support successful retrieval of previously encountered

exemplars from long-term memory, and bilateral temporal-parietal

activation for perceptual feature integration. Subsequent similarity-

based categorization revealed temporal–parietal, posterior cingulate,

and anterior prefrontal activation. These findings suggest that large-

scale networks support relatively distinct categorization processes

during the acquisition and judgment of semantic category knowledge.

D 2004 Elsevier Inc. All rights reserved.

Keywords: BOLD fMRI; Rule-based category; Similarity-based category

Introduction

Semantic memory involves semantic content and categorization

processes that act on that content (Tulving, 1972). Categorization is

1053-8119/$ - see front matter D 2004 Elsevier Inc. All rights reserved.

doi:10.1016/j.neuroimage.2004.08.045

$ Portions of this work were presented at the annual meeting of the

Cognitive Neuroscience Society in San Francisco, April, 2002. This work

was supported in part by grants from the US Public Health Service

(AG15116, AG17586, and NS35867).

* Corresponding author. Department of Neurology, University of

Pennsylvania, 3 West Gates, 3400 Spruce Street, Philadelphia, PA 19104-

4283. Fax: +1 215 349 8464.

E-mail address: plkoenig@mail.med.upenn.edu (P. Koenig).

Available online on ScienceDirect (www.sciencedirect.com).

essential to organizing semantic content in a meaningful way

(Smith, 1995). At least two broad forms of categorization processes

have been proposed. In similarity-based categorization, category

membership is determined by overall similarity to established

exemplars or to a prototype (Medin and Schaffer, 1978; Medin

et al., 1993; Smith and Medin, 1981). This process appears to

entail exemplar or prototype retrieval, as well as organization of

perceptual features contributing to an object’s overall appearance

(Ashby et al., 1991). Rule-based categorization is employed when

specific features must be considered (Bruner et al., 1956; Smith

et al., 1998a,b). This process is resource demanding, since it

presumably involves executive processes such as selective attention

(to focus on requisite features), inhibitory control (to suppress

irrelevant features), and working memory (to retain assessments of

various features). Behavioral studies show that healthy subjects are

capable of implementing either rule-based or similarity-based

semantic categorization processes, even for the same stimulus set

(Allen and Brooks, 1991; Grossman et al., 2003b; Koenig et al.,

2002; Rips, 1989; Smith and Sloman, 1994). Qualitative distinc-

tions between these two categorization processes suggest that they

depend on partially distinct neural substrates. The present study

used BOLD fMRI in an event-related design to examine cortical

recruitment patterns in healthy young subjects during the acquis-

ition and categorization of novel, meaningful stimuli by rule-based

and by similarity-based strategies.

Behavioral studies of patients with central nervous system

disease provide some evidence for neurologically separable

categorization systems. In two recent studies, patients with

Alzheimer’s disease and patients with frontotemporal dementia

showed impaired rule-based categorization of object descriptions

(Grossman et al., 2003b) and of novel visual items (Koenig et al.,

2004, submitted for publication), although they were able to

categorize by similarity-based processing in a relatively normal

manner. This is consistent with the executive resource limitations

observed in both Alzheimer’s disease (Fisher et al., 1990; LaFleche

and Albert, 1995; Perry and Hodges, 1999) and frontotemporal

dementia (Elfgren et al., 1993; Pasquier et al., 1999).

Functional neuroimaging work also implicates somewhat dis-

tinct brain regions for the processes supporting rule-based and

P. Koenig et al. / NeuroImage 24 (2005) 369–383370

similarity-based categorization. Neuroimaging studies of clinical

tasks involving rule-based processes like Wisconsin Card Sorting

and Raven’s Progressive Matrices emphasize activation of anterior

cingulate and dorsolateral prefrontal cortices (Elliott et al., 1999;

Rao et al., 1997; Savage et al., 2001). Some investigators relate

executive processes such as selective attention and inhibitory control

(i.e., processes characteristic of categorizing by rules) to activation

of frontal regions (Petrides et al., 1993; Vendrell et al., 1995;Wagner

et al., 2001). Switching processes that govern alternating between

attributes of an object, objects in a category, and processes applied to

an object or a category have been related to dorsolateral prefrontal

cortex (D’Esposito et al., 1995; Garavan et al., 2000; Monchi et al.,

2001; Savage et al., 2001; Seger et al., 2000a,b).

By comparison, similarity-based categorization, as purportedly

used in classifying dot patterns or learning artificial grammars,

appears to implicate parietal as well as frontal-striatal brain regions

(Aizenstein et al., 2000; Poldrack and Gabrieli, 2001; Reber et al.,

1998a; Seger et al., 2000a,b). The inferior parietal component may

support feature integration necessary for global comparisons

(Wilkinson et al., 2002). Neuroimaging work also suggests the

contribution of posterior cingulate cortex, which may support

retrieval of previously encountered exemplars (Rugg et al., 1998;

Tulving et al., 1996) and anterior prefrontal cortex (Buckner et al.,

1998; Lepage et al., 2002) for the successful retrieval of

information from long-term memory.

Very few imaging studies have explicitly compared rule-based

and similarity-based processing. A recent fMRI study (Grossman

et al., 2002a,b) examined cortical activation during rule-based and

similarity-based category assignments of written descriptions of

objects that bore equal resemblance to either of two choice

categories, but violated the rule-like specifications of one. For

example, a bround object 3 in. in diameterQ equally resembled a

bpizzaQ or a bquarter,Q but, by the rule governing a quarter’s

invariable size, could only unequivocally belong to the category

bpizza.Q A direct contrast of rule-based minus similarity-based

categorization decisions showed left dorsolateral prefrontal activa-

tion, consistent with support for executive resources during rule-

based categorization.

Closer in structure to the present study is a recent PET study

examining activation during categorization of novel visual stimuli

(cartoon-like animals) by rule- and similarity-based processes

(Patalano et al., 2001). Both categorization conditions showed

activation in visual cortex, while only the rule-based categorization

showed bilateral superior parietal, right dorsolateral prefrontal, and

bilateral supplementary motor cortex activation. The authors

argued that these regions support selective attention to spatial

position, shifting attentional focus, and working memory.

The present study departs from the previous ones in several

ways. We used a visual category of novel animals, although unlike

the stimuli in the Patalano et al. study, the animals were relatively

naturalistic. We imaged category acquisition as well as category

use, whereas Patalano et al. imaged only the latter. Hence, the

present study assesses cortical activation during learning by rule-

and similarity-based processes as well as decision-making by those

processes subsequent to learning. We designed the study so that we

could directly compare the two categorization processes: Unlike

Patalano et al., we employed a single category, structured such that

subjects would generally endorse the identical items as members

across both the rule-based and similarity-based categorization

conditions. Finally, we employed empirical assessments of the

relative contribution of particular stimulus features to inter-

stimulus resemblance. This allowed us to construct a non-arbitrary

category; that is, our category members were subjectively

perceived as similar, and our rules of category membership

reflected this category coherence. Our intention was to make our

category easily learnable, so that patterns of activation would

reflect the two categorization processes rather than subjects’ efforts

to master a difficult task.

With these materials, we expected to observe distinct neural

networks for each semantic categorization process: For rule-based

categorization, we anticipated a network centering around inferior

frontal and anterior cingulate regions to sustain working memory

(e.g., for retaining the rules) and executive resources such as

selective attention (e.g., for identifying relevant features). For

similarity-based categorization, we anticipated a network including

lateral temporal-parietal cortex supporting perceptual feature inte-

gration, and posterior cingulate and anterior prefrontal activation

supporting successful retrieval of previously encountered exemplars.

Methods

Subjects

Twenty-eight healthy young volunteers participated, including

11 males and 17 females with a mean (FSD) age of 22.2 (F3.3)

years and mean education of 15.2 (F1.7) years. All were right-

handed native speakers of English. We were specifically interested

in examining cortical activation supporting rule-based and sim-

ilarity-based categorization, rather than participants’ spontaneous

choice of categorization strategy. Hence, we excluded data from

participants whose behavioral results suggested that they had not

complied with our instructions. These included participants who

seemed to focus on a single feature (i.e., they apparently applied a

simple rule) in the Similarity condition. These also included

participants who apparently did not cooperate, that is, made grossly

inaccurate or random judgments. Eleven (6 males, 5 females, mean

age of 22.0 (F2.6) years, mean education of 15.1 (F1.2) years)

participated in the Rule condition: Fourteen participants were

originally run in this condition, but three were excluded because

their behavioral responses at test were less accurate than the group

mean by at least two standard deviations. Nine (2 males, 7 females,

mean age of 22.9 (F4.7) years, mean education of 15.3 (F2.5)

years) participated in the Similarity condition: Fourteen were

originally run, but data from 1 participant were excluded because

responses at test did not differ from chance. Four additional

participants (2 males, 2 females, mean age of 21.3 (F2.6) years,

mean education of 15.3 (F1.9) years) were excluded from the

Similarity condition because they appeared to spontaneously

employ a consistent rule-like strategy at test. These four exhibited

clearly systematic patterns in their judgments, that is, focusing

exclusively on a single feature to determine category membership,

or granting membership in exact proportion to how many features

each item shared with the prototype. We do not report the results of

these participants here since the group may be too small to

demonstrate a reliable imaging profile, but the results are available

from the authors.

Stimuli

Stimuli were taken from a set of 64 realistic images of

biologically plausible novel animals. Examples can be seen in

P. Koenig et al. / NeuroImage 24 (2005) 369–383 371

the training stimuli shown in Fig. 1. The stimulus set, and the

method of obtaining empirical rankings of the features’ contribu-

tion to judged resemblance (briefly described below), are described

in detail elsewhere (Koenig et al., 2002). The set represented all

possible combinations of six dichotomous features, that is, color

(red-brown or yellow-green), legs (long or short), neck (low or

raised), snout (long or short), tail (straight or curled), and teeth

(fangs or tusks). All instances of a feature were identical, for

example, all straight-tailed animals had tails of identical dimen-

sions. All other features (e.g., texture, torso shape, facial details)

were consistent throughout the set. Prior to this study, a cohort of

healthy young subjects rated all pair-wise combinations in the set

for resemblance. We then used multidimensional scaling to obtain

a ranking of the six features’ contribution to resemblance

judgments.

We identified the four highest-ranking features from the multi-

dimensional scaling procedure, which served as bcontributingQfeatures. These were snout, legs, color, and neck type. One animal

was then chosen at random to serve as a prototype of the category.

We defined category MEMBERS as those animals that matched the

Fig. 1. Illustration of screens used during the training session. Panel A: rule-

based training, showing the four criterial features in the upper half of the

screen and two crutters in the lower half (MEMBER on left). Panel B:

Similarity-based training, showing a prototype in the upper half of the

screen and two crutters in the lower half (MEMBER on left).

prototype in at least three of the four contributing features, that is, the

same type of legs, snout, color, and/or neck. LOW DISTORTION

items matched the prototype in two of the contributing features, for

example, the same type of legs and snout but a different color and

neck type. HIGH DISTORTION items matched the prototype on no

more than one contributing feature. The two lowest-ranked features

(i.e., teeth and tail) served as bdistractorQ features that were irrelevantto category membership and were equally distributed across

MEMBERS, LOWDISTORTION items, and HIGHDISTORTION

items. Thus, our category structure was in accordance with

perceived resemblance: membership reflected the presence of salient

features shared with the prototype, rather than simply the number of

features in common.

It has been suggested that the nature of features, particularly

integral versus non-integral, differentially contribute to how

conducive a category is to similarity- or rule-based learning

(Kemler and Smith, 1979). We intended our category to be

learnable by either process. Pilot studies suggest that the individual

features are readily discernible (e.g., non-integral), hence making

the category accessible by rule-based learning. In addition, pilot

studies suggest that subjects are generally not explicitly aware of

the bases for their resemblance-based judgments, even when they

appear confident of their decision. This is characteristic of

categories with integral features; hence, the category is accessible

by similarity-based learning.

Training items were 40 unique pairs consisting of a

MEMBER and a HIGH DISTORTION item, created from

recombinations of eight items of each type. The MEMBER in

each pair had three contributing features and at least one

distractor feature in common with the prototype. Each contribu-

ting feature in common with the prototype (e.g., a long snout)

was present in six of the eight of the training set MEMBERS.

The HIGH DISTORTION items had only one contributing feature

and not more than one distractor feature in common with the

prototype. Each contributing feature differing from the prototype

(e.g., a short snout) was present in six of the eight training set

HIGH DISTORTION items. The eight MEMBERS and eight

HIGH DISTORTION items were recombined to create five

different sets of eight pairs. Within each set, pairs were randomly

ordered with the stipulation that each of the two occurrences of a

particular combination of three contributing features appeared

once among the first four pairs and once among the last four.

This distribution of feature configurations was designed to make

it highly unlikely that participants could demonstrate apparent

success during the training phases (see below) by adopting

simplified response strategies such as focusing on a single

feature, as such strategies would lead to misjudgments at frequent

intervals. For instance, a subject who consistently focused on a

particular feature, such as endorsing all yellow items, would

choose correctly no more than 75% of the time for each

consecutive set of four training trials, as one of the four training

MEMBERS would be red. The subject’s accuracy could be

further decreased by the presence of the preferred feature in one

out of every four consecutive training HIGH DISTORTION

items. Similarly, subjects who focused on a different individual

feature for each trial had maximally a 75% chance of choosing

correctly each time. In addition, subjects who employed a feature-

focus strategy would show a consistent level of accuracy

throughout training, rather than exhibiting learning over time.

Across sets, MEMBERS were always paired with different HIGH

DISTORTION items. The sets were presented continuously and

P. Koenig et al. / NeuroImage 24 (2005) 369–383372

in sequence. Stimuli were video-projected onto a screen, which

subjects viewed via a system of mirrors while in the magnet bore.

Presentation was controlled by a computer using PsyScope

software (Cohen et al., 1993), which recorded responses.

We presented these MEMBER/HIGH DISTORTION pairs in

each of two training conditions, described below.

Rule training procedure

Participants were told that an animal called a bcrutterQ had to

have at least three of four particular features, and that they would

be shown pairs of animals, one of which would be a crutter and one

of which would not. Their task was to decide which animal in each

pair was a crutter, based on the bat-least-three-of-four featuresQrule. The training session in the scanner followed. The 40-item

sequence of MEMBER/HIGH DISTORTION training pairs was

presented. Each pair was arrayed horizontally in the lower half of

the projected image, and was accompanied by brief written

descriptions and outlines of the requisite features (e.g., the words

blong snoutQ with a sketch of a long snout) in the upper half. The

description of the color feature (i.e., byellow colorQ) was illustratedby a yellow rectangle rather than a sketch. The sketches were black

outlines of the contours of the features, intended to clarify the

verbal descriptions by clearly showing which feature variants they

referred to. At the same time, the sketches’ lack of visual detail

ensured that subjects could not rely on a perceptual identity match

when identifying the features in the stimulus items. The

MEMBERS and HIGH DISTORTION items were equally dis-

tributed in the left and right positions. The sentence bWhich has 3

features?Q appeared at the top of the screen. Fig. 1 panel A shows a

sample Rule training stimulus. Stimuli appeared every 12 s.

Participants’ responses were indicated by a button press on a hand-

held box, with the buttons corresponding spatially to the location

(left or right) of the chosen item. The stimulus remained until the

subject responded, after which the screen became blank; response

times never exceeded 6 s. Nine seconds after a stimulus item first

appeared, the item reappeared for 2 s with a black rectangle

surrounding the correct choice (i.e., the category member), as a

means of providing feedback. We analyzed only items to which

participants responded (five participants each missed one training

trial).

Prior to entering the scanner, participants received a practice

session at a laptop computer. Although we expected the training

tasks to be readily comprehendible, we wished to ensure that

subjects’ efforts during the first few trials would be directed

towards performing the training task rather than acclimating

themselves to the procedure. Hence, the practice stimuli were

designed to familiarize subjects with the training procedure without

exposing them to the animal stimuli. They consisted of simple

abstract geometric figures with three variable features: large or

small size, circular or triangular appendage, and striped or dotted

pattern. None of these features coincided with features of the

animal stimuli. The practice stimuli were presented in a series of

images that were configured analogously to the training stimuli;

that is, written descriptions and sketches of the three features

appeared above a pair of items, one of which contained two of the

features and one of which contained one. Participants were told

that an object called a bdortQ had to have at least two of the

features, and they were asked to indicate which item in each pair

was a dort. Feedback was analogous to the actual training trials,

that is, a black rectangle appearing around the correct choice.

Participants were exposed to three practice trials, and all of them

indicated understanding of the procedure.

Similarity training procedure

Participants were told that they would see pictures of an animal

called a bcrutterQ along with pairs of pictured animals, one of which

would also be a crutter and one of which would not. Their task was

to decide which animal in each pair was a crutter, based on which

animal was bmore similarQ to the sample crutter. Subjects were

urged to make a quick decision based on their initial impressions.

The training session in the scanner followed. A sequence of stimuli

was presented consisting of the prototype in the upper half of the

projected image and the same horizontally arrayed MEMBER/

HIGH DISTORTION pairs as in the Rule training procedure in the

lower half. The caption bcrutterQ appeared directly below the

prototype and the sentence bWhich is a crutter?Q appeared at the

top of the screen. The stimulus displays for the Similarity and Rule

training conditions thus each contained pictorial and verbal

elements (including sentences and captions), and were designed

to minimize display discrepancies between these inherently differ-

ent training processes. Fig. 1 panel B shows a sample Similarity

training stimulus. All other aspects (e.g., timing, mode of response,

feedback, and data collection) were the same as in the Rule training

session. We analyzed only trials to which participants responded

(three participants missed a total of seven trials).

Participants practiced at a laptop computer prior to this

training procedure as well. The practice items were the same as

those used prior to the Rule training session. The stimuli

consisted of a prototype geometric configuration in which the

three requisite features were represented; this appearing above the

same pairs used in the Rule practice. Participants were told that

the prototype was a bdortQ and were asked to indicate which item

in each pair was also a dort based on greater similarity to the

sample dort. All participants indicated understanding of this

training task as well.

Testing procedure

Both training conditions were followed by the identical test.

Participants saw a sequential presentation of individual animals in

a fixed semi-random order, and judged whether each was a crutter.

They were instructed to judge in accordance with how they had

been trained; that is, participants in the Rule condition were to

judge category membership by an item’s adherence to the bthree-out-of-four featuresQ rule, and participants in the Similarity

condition were to judge category membership by an item’s overall

resemblance to the training example. This sequence consisted of a

50-item subset of the complete set of 64 animals. We used a

reduced set so that the test session would not exceed an imaging

run that is a maximum of 10 min in duration. The subset contained

20 MEMBERS, 14 LOW DISTORTION items, and 16 HIGH

DISTORTION items. Items of a particular type (e.g., LOW

DISTORTION) appeared no more than three times sequentially.

Stimuli appeared every 12 s, and remained visible until the subject

responded. Participants indicated their choice by button press, with

the right and left buttons corresponding, respectively, to byesQ andbno.Q No feedback was provided. Responses and reaction times

were recorded. Analyses included only trials in which subjects

responded (only one Rule participant did not respond to one LOW

DISTORTION stimulus).

P. Koenig et al. / NeuroImage 24 (2005) 369–383 373

Imaging data acquisition and statistical analysis

The experiment was carried out at 4T on a GE prototype

Echospeed scanner capable of ultrafast imaging. Firm foam

padding was used to restrict head motion. Each imaging protocol

began with a 10–15 min acquisition of 5 mm thick adjacent slices

for determining regional anatomy, including sagittal localizer

images (TR = 500 ms, TE = 10 ms, 192 � 256 matrix), T2-

weighted axial images (FSE, TR = 2000 ms, TEeff = 85 ms), and

T1-weighted axial images of slices used for fMRI anatomic

localization (TR = 600 ms, TE = 14 ms, 192 � 256 matrix).

Gradient echo echoplanar images were acquired for detection of

alterations of blood oxygenation accompanying increased mental

activity. All images were acquired with fat saturation, a rectangular

FOV of 20 � 15 cm, flip angle of 908, 5 mm slice thickness, an

effective TE of 50 ms, and a 64 � 40 matrix, resulting in a voxel

size of 3.75 � 3.75 � 5 mm. The echoplanar acquisitions consisted

of 18 contiguous slices covering the entire brain every 2 s. To

minimize susceptibility artifact, a separate acquisition lasting 1–2

min was needed for phase maps to correct for distortion in

echoplanar images (Alsop, 1995). We also inspected raw data of

individual subjects to check for artifacts. Raw data were stored by

the MRI computer on DAT tape and then processed off-line.

Initial data processing was carried out with Interactive Data

Language (Research Systems) on a Sun (Santa Clara, CA) Ultra 60

workstation. Raw image data were reconstructed using a 2D FFT

with a procedure that minimized artifact due to magnetic field

inhomogeneities. Individual subject data were then prepared for

statistical parametric mapping (SPM 99) developed by the Well-

come Department of Cognitive Neurology (Frackowiak et al.,

1997). This system, operating on a MatLab platform, combines

raw images from subjects into a statistical t map. Briefly, the

images in each subject’s time series were registered to the initial

image in the series. The images were then aligned to a standard

coordinate system (Talairach and Tournoux, 1988). The data were

spatially smoothed with a 12-mm Gaussian kernel to account for

small variations in the location of activation and sulcal anatomy

across subjects. Low-pass temporal filtering was implemented to

control auto-correlation with a first-order auto-regressive method.

We used an event-related approach to data acquisition. Each

event was 12 s in length, thus capturing most of the hemodynamic

response function associated with each stimulus. Hence, the initial

images acquired for each 12-s trial were used in the analyses. The

exclusive use of the initial image per trial also insured that

activation results would not be affected by between-condition

differences in time until the button-press response. Event-related

data acquisition allowed us to group and compare images

corresponding to specific stimuli. For the training phase, we

assessed condition-specific cerebral activation. We also examined

changes in activation with increased exposure to stimuli during the

course of training by analyzing activation for the first eight and last

eight of the 40 training trials (i.e., the first and fifth quintiles).

Thus, we used a condition (rule-based, similarity-based) � learning

phase (first quintile, last quintile) factorial design. For the test

phase, we grouped MEMBERS and HIGH-DISTORTION stimuli.

This was because their membership status was relatively trans-

parent, according to similarity-based as well as by rule-based

processing. Hence, even though the two categorization conditions

elicited different decision-making processes, MEMBERS should

generally be endorsed to comparable extents across conditions, and

HIGH-DISTORTION should stimuli should generally be rejected

to comparable extents across conditions. To help ensure that

patterns of activation reflected subjects’ employing the catego-

rization processes of interest, we analyzed activation only for

accurate responses to these stimuli. As shown below, this included

the vast majority of MEMBER and HIGH-DISTORTION stimuli

in both conditions. We included all LOW-DISTORTION stimuli in

our analyses, regardless of subjects’ category judgments, since

membership or non-membership status was equally plausible by

similarity-based criteria. Moreover, as shown below, proportions of

member/nonmember judgments for LOW DISTORTION items

differed depending on the categorization condition, and we did not

want our imaging analyses to be biased by the inclusion of

different numbers of a particular type of stimulus across conditions.

Test data were analyzed with a condition (rule-based, similarity-

based) � stimulus (MEMBER/HIGH-DISTORTION, LOW-DIS-

TORTION) factorial design. The resulting statistical contrasts were

converted to z-scores for each compared voxel. We used a

statistical threshold of P b 0.001 uncorrected for multiple

comparisons. The extent criterion was set to a 20 voxel threshold.

Results

Training phase

Behavioral results

Participants were accurate at judgments MEMBERS under both

training conditions. Accuracy in the Rule training condition was

96% (F6%), which differed from chance [t(10) = 28.09; P b

0.001]; accuracy in the Similarity training condition was 85%

(F9%), likewise differing from chance [(8) = 11.08; P b 0.001].

The higher accuracy for identifying MEMBERS in the Rule

condition relative to the Similarity condition reached significance

[t(17) = 3.0, P = 0.009]. Moreover, participants took significantly

longer to respond in the Rule condition [3946 ms F 914] than in

the Similarity condition [2886 ms F 1003; t(18) = 2.5, P = 0.02].

These results are consistent with the qualitative differences

between the two categorization processes: Rule-based training,

which depends on working memory and executive processes such

as selective attention, appears to be more cognitively demanding,

and yields judgments with a higher rate of identifying MEMBERS.

We also compared judgment accuracy during the first and fifth

quintiles for each condition. For the Rule condition, accuracy was

89% during the first quintile and 100% during the fifth. First

quintile accuracy differed from 100% in a one-sample t test, t(10) =

2.83, P b 0.02. (We could not use paired t tests to directly compare

first versus fifth quintile accuracy in the Rule condition because the

100% accuracy in the fifth quintile had no variance.) For the

Similarity condition, accuracy was 81% during the first quintile

and 94% during the fifth. These differed significantly by paired t

test, t(8) = 3.62, P = 0.007. Thus, participants became more

successful at correctly identifying the MEMBER by the latter part

of the session in both training conditions.

Imaging results

We analyzed training data with a condition (Rule, Similarity) �learning phase (early, late) factorial analysis. Table 1 summarizes

the activation peaks using Talairach coordinates, the associated

Brodmann areas, and the z scores of the heights for each activation

peak during training. We found significant differences for the main

effect of condition, as illustrated in Fig. 2. Fig. 2 Panel A shows

Table 1

Distribution of activation during the training phase of category acquisition

Anatomic area Brodmann area Talairach coordinates # Voxels Z score

x y z

MAIN EFFECTS

RULE minus SIMILARITY

L inferior frontal 6/44 �56 0 36 17 3.47

L anterior cingulate 24/32 �8 8 40 69 4.11

L inferior parietal 40 �44 �48 40 26 3.43

L parietal/occipital 19/7 �8 �76 32 236a 4.23

R superior parietal 7 28 �64 52 236a 4.64

L thalamus �8 �20 4 39 3.52

SIMILARITY minus RULE

L temporal–parietal 39 �40 �72 12 25 3.59

R temporal–parietal 39 56 �60 12 36 3.32

R anterior prefrontal 8 20 28 44 12 3.30

L anterior prefrontal 10 �16 56 8 23 3.20

EARLY minus LATE

R superior parietal 7 4 �72 48 30 3.28

R superior temporal 22 60 �28 4 13 3.13

LATE minus EARLY

L temporal–parietal 40/22 �44 �20 20 241 4.09

L posterior cingulate 23/29 �8 �36 20 55 3.43

R superior temporal/insula 42 28 �20 16 91 3.31

L thalamus �20 �24 4 30 3.13

INTERACTION EFFECTS

RULE: EARLY minus LATE

R superior parietal 7 8 �72 48 40 3.65

L parietal–occipital 19 �32 �88 16 15 3.48

L thalamus �4 �8 8 85 3.27

RULE: LATE minus EARLY

L inferior frontal 6/44/45 �44 0 4 82 3.06

SIMILARITY: LATE minus EARLY

L temporal–parietal 40/22 �52 �20 24 279 3.99

R temporal–parietal 40/22 56 �16 20 29 3.40

L posterior cingulate 29/23 �8 �36 20 103 4.26

R striatum 20 0 �4 49 3.83

a These regions were part of the same cluster.

P. Koenig et al. / NeuroImage 24 (2005) 369–383374

that the Rule condition, relative to the Similarity condition,

activated left inferior frontal, left anterior cingulate, left inferior

parietal, bilateral parietal/occipital, and left thalamus regions. By

comparison, Fig. 2 Panel B shows that the Similarity condition,

relative to the Rule condition, activated bilateral temporal–parietal

and bilateral anterior prefrontal regions. These findings suggest

different patterns of activation depending on the training condition.

We also found that activation for each training condition varied

depending on the early versus later phase of training. Thus, the

factorial analysis showed significant main effects for the early

versus late portions of the learning phase, that is, changes over the

course of acquisition collapsed across conditions. As summarized

in Table 1, the early portion, relative to the late portion of the

learning phase, activated right superior parietal and right superior

temporal areas. The late portion of training, relative to the early

portion, activated left temporal–parietal, left posterior cingulate,

right temporal/insula, and left caudate regions. The overall

interaction effects in the factorial analysis involved bilateral

striatum including right putamen and left caudate (x = 20, y = 4,

z = �4; z score = 4.02), and left inferior frontal cortex in a manner

that approached significance (x = �48, y = 4, z = �4; z score =

2.88). More detailed evaluation of the interaction effect is

illustrated in Fig. 2. As shown in Fig. 2 Panel C, greater activation

later in the Rule training phase, relative to earlier, was seen in left

inferior frontal cortex. For the Similarity condition, Fig. 2 Panel D

shows relatively greater activation later in training compared to the

early in bilateral temporal–parietal, left posterior cingulate, and

right striatal regions. Thus, distinct activation patterns appeared to

become consolidated over time in part in different brain regions,

depending on the specific training condition. These patterns

suggest that left inferior frontal cortex supports increased use of

executive resources such as working memory as rule-based training

progressed. In contrast, later activation of temporal–parietal,

posterior cingulate, and striatal regions during similarity-based

training suggests an increasing tendency towards feature config-

uration and integration along with successful recall of previously

seen training items.

Test phase

Behavioral results

We calculated b% yesQ scores, that is, how often participants

granted membership to each item type (i.e., MEMBER, LOW

DISTORTION, and HIGH DISTORTION items) at test. Perform-

ance during the test phase is summarized in Fig. 3. As can be

seen, patterns of responses differed across conditions and types of

Fig. 2. Distribution of activation during training phase. Panel A: Main effect: Rule-based training minus Similarity-based training. Panel B: Main effect:

Similarity-based training minus Rule-based training. Panel C: Interaction effect: Rule-based training Late minus Early. Panel D: Interaction effect: Similarity-

based training Late minus Early.

P. Koenig et al. / NeuroImage 24 (2005) 369–383 375

items in a manner consistent with the two qualitatively distinct

categorization processes. The Rule condition thus showed a sharp

category boundary, reflecting the requirement that only items

meeting the three-out-four feature criteria be endorsed as

members; the Similarity condition showed graded responses

across types of items with declining membership endorsement

as the number contributing of features shared with the prototype

declined. A repeated measures ANOVA, with a condition (Rule,

Similarity) � stimulus type (MEMBER, LOW DISTORTION,

HIGH DISTORTION) design, revealed a main effect of stimulus

Fig. 3. Category membership judgments during the test phase following

Rule-based training or Similarity-based training, illustrating the percent of

time that subjects accepted MEMBERS, LOW DISTORTION stimuli, and

HIGH DISTORTION stimuli as category members.

type [F(2,36) = 371.5, P b 0.001] and a significant condition x

stimulus type interaction [F(2,36) = 22.0, P b 0.001]. There was

no main effect of condition [F(1,18) = 2.26, ns]. The t tests

revealed that judgments differ significantly across conditions for

all three stimulus types: participants in the Rule condition

responded byesQ more often to MEMBERS [t(18) = 4.0, P =

0.001] and less often to LOW DISTORTION items [t(18) = 3.1,

P = 0.006] and HIGH DISTORTION items [t(18) = 2.5, P =

0.02] than participants in the Similarity condition. Linear trend

analyses confirmed the different shapes of the performance

curves: We saw blinearityQ in both the Rule condition [F(1,30) =

777.8, P b 0.01] and the Similarity condition [F(1,24) = 78.2, P b

0.01]. However, bdeviation from linearityQ was seen in the Rule

condition [F(1,30) = 157.3, P b 0.01] but not the Similarity

condition [F(1,24) = 3.2, ns]. These findings confirmed that

participants respond differently across conditions for the different

types of items, reflecting the sharp categorical effect of rule-based

categorization and the graded responses of similarity-based

categorization.

There was a trend towards longer reaction times in the Rule

condition (2453 ms) relative to the similarity condition (2144 ms)

that did not reach significance, t(18) = 0.79, ns. Reaction times did

not differ among item types in the Similarity condition [F(2,8) =

0.86, ns]. However, reaction times did differ among different item

types in the Rule condition [F(2,10) = 10.1, P = 0.001. While

reaction times for MEMBERS and HIGH DISTORTION items

were equivalent [t(10) = 1.71, ns], participants responded more

slowly to LOW DISTORTION items than to either MEMBERS

[ t(10) = 2.31, P b 0.05,] or to HIGH DISTORTION items [ t(10) =

7.09, P b 0.001.] This finding suggests that judging MEMBERS

P. Koenig et al. / NeuroImage 24 (2005) 369–383376

and HIGH DISTORTION items may have been facilitated by prior

exposure to such items during training.

We compared judgment accuracy for old versus new items in

both test conditions; that is, MEMBERS that had been seen five

times during training with MEMBERS seen for the first time at

test, and similarly, HIGH DISTORTION items seen during training

with those newly seen at test. To ensure that the old and new

stimulus items were comparable, we included as new items only

those with exactly three contributing features appropriately

matching or not matching the prototype; that is, we did not include

test MEMBERS matching the prototype on all four features or test

HIGH DISTORTION items matching it on none. Accuracy for old

versus new items was equivalent for each item type in each

condition: Rule MEMBERS, t(10) = 1.00, P = 0.34; HIGH

DISTORTION, t(10) = 1.18, P = 0.86; Similarity MEMBERS,

t(8) = .67, P = 0.52; HIGH DISTORTION, t(8) = 0.33, P = 0.75.

This suggests that participants were using categorization processes

at test for all items, rather than relying on recognition of training

items.

Imaging results

We analyzed categorization data with a condition (rule-based,

similarity-based) � stimulus type (MEMBERS+HIGH DISTOR-

TION, LOW DISTORTION) factorial analysis. Table 2 summa-

rizes the activation peaks using Talairach coordinates, the

Table 2

Distribution of activation during the test phase of categorization

Anatomic area Brodman

MAIN EFFE

RULE minus SIMILARITY

L inferior frontal 6/44

L anterior cingulate 24/32

SIMILARITY minus RULE

L temporal–parietal 39/22

R posterior temporal 21

L anterior prefrontal 10

R anterior prefrontal 10

R posterior cingulate 23/29/30

MEMBER + HIGH DISTORTION minus LOW DISTORTION

No significant contrasts –

LOW DISTORTION minus MEMBER + HIGH DISTORTION

L inferior frontal 6

R inferior frontal 6

B anterior cingulate 24/32

L dorsolateral prefrontal 46

L striatum

R striatum

INTERACTION

RULE minus SIMILARITY: MEMBER + HIGH DISTORTION

L inferior parietal 40

L parietal–occipital 19

L anterior cingulate 24/32

SIMILARITY minus RULE: MEMBER + HIGH DISTORTION

L temporal–parietal 39/22

R temporal–parietal 39/22

RULE minus SIMILARITY: LOW DISTORTION

L inferior frontal 6/44

SIMILARITY minus RULE: LOW DISTORTION

L temporal–parietal 39/21/19

a These regions were part of the same cluster.

associated Brodmann areas, and the z scores of the heights for

each activation peak during the test phase. The activations

associated with the main effects are illustrated in Fig. 4.

Resembling the activation patterns seen during training, the two

conditions differed in activated areas during the test phase of the

study. Panel A reveals activation during the Rule condition,

relative to the Similarity condition, in left inferior frontal and left

anterior cingulate regions. Panel B shows activation for the

Similarity condition, relative to the Rule condition, in bilateral

temporal–parietal regions that is more prominent on the left, in

bilateral anterior prefrontal regions, and in right posterior cingulate

cortex.

We also found that these condition-specific activations differed

somewhat, depending on the types of items being judged. Thus,

the main effect for item type revealed activation for LOW

DISTORTION items, compared to MEMBERS+HIGH DISTOR-

TION items, in left dorsolateral prefrontal, bilateral anterior

cingulate, and bilateral striatal areas, as well as bilateral inferior

frontal areas that is marginal on the left. The overall interaction

effect for the factorial analysis showed left inferior frontal (x =

�40, y = 28, z = 0; z score = 2.82) and right dorsolateral

prefrontal (x = 32, y = 40, z = 20; z score = 2.90) recruitment. We

examined interaction effects more closely, as summarized in Table

2 and illustrated in Fig. 4. For MEMBERS+HIGH DISTORTION

items, Panel C shows that the Rule condition, relative to the

n area Talairach coordinates # Voxels Z score

x y z

CTS

�60 4 28 59 3.84

�4 8 44 61 3.50

�32 �64 8 261 5.08

40 �44 4 44 3.08

�24 56 20 20 3.39

16 44 �4 112 3.08

8 �48 8 24 3.33

– – – – –

�32 �12 48 25 2.90

40 �16 40 186a 3.34

0 4 48 186a 3.28

�52 24 24 54 3.20

�16 0 0 45 3.32

28 20 0 39 3.24

EFFECTS

�52 �36 40 49 3.14

�32 �76 40 23 3.40

�4 8 40 24 2.99

�32 �64 8 1411a 4.75

32 �60 8 1411a 4.46

�64 0 20 91 3.19

�44 �68 8 108 4.12

Fig. 4. Distribution of activation during the test phase. Panel A: Main effect: Rule-based categorization minus Similarity-based categorization. Panel B: Main

effect: Similarity-base categorization minus Rule-based categorization. Panel C: Interaction effect: Rule-based minus Similarity-based categorization for

MEMBER + HIGH DISTORTION items. Panel D: Interaction effect: Rule-based minus Similarity-based categorization for LOW DISTORTION items. Panel

E: Interaction effect: Similarity-based minus Rule-based categorization for MEMBER + HIGH DISTORTION items. Panel F: Interaction effect: Similarity-

based minus Rule-based categorization for LOW DISTORTION items.

P. Koenig et al. / NeuroImage 24 (2005) 369–383 377

Similarity condition, activated left inferior parietal, left parietal–

occipital, and left anterior cingulate regions. For LOW DISTOR-

TION items, Panel D shows activation for the Rule condition

minus the Similarity condition in left inferior frontal cortex. Panel

E shows that the Similarity condition, compared to the Rule

condition, activated bilateral temporal–parietal cortex for MEM-

BERS+HIGH DISTORTION items. Panel F shows activation in

left temporal–parietal cortex for LOW DISTORTION items in the

Similarity condition minus the Rule condition. These findings

suggest that relatively distinct patterns of activation were

associated with each categorization condition, consistent with

support for executive resources for rule-based categorization and

perceptual feature integration for similarity-based categorization.

In addition, the condition-specific patterns vary somewhat

depending on the type of stimulus item being judged, suggesting

that, for each condition, different strategic shifts are required for

P. Koenig et al. / NeuroImage 24 (2005) 369–383378

judging items whose membership status is readily apparent versus

judging equivocal items.

We expected our participants to make category judgments at

test in accordance with how they had been trained. Overlap in areas

of activation within condition from training to test would reflect

this continuity. We looked for such overlap by comparing the

Talairach coordinates of peak activation between Rule minus

Similarity, and between Similarity minus Rule, for training and for

test. In both cases, two areas were observed for which all three

respective coordinates differed by less than 12 mm, our unit of

spatial smoothing, between the corresponding training and test

phases. As can be seen in Tables 1 and 2, Rule-based processing,

relative to Similarity-based, recruited left inferior frontal (BA 6/44)

and left anterior cingulate (BA 24/32) regions, suggesting that

executive processing and monitoring of competing responses were

common to both rule-based learning and subsequent test. The high

levels of accuracy in category membership judgments in this

resource-demanding task are consistent with a constant level of

response monitoring. Similarity-based processing, relative to Rule-

based, recruited left temporal–parietal (BA 39) and left anterior

prefrontal (BA 10) regions, suggesting that overall feature

configuration and recall of previously encountered exemplars were

common to both similarity-based learning and subsequent test. Our

experimental design was not conducive to subtractions between the

training and test phases. The stimulus displays differed; for

instance, subjects saw multiple items along with written material

during training, but only single items at test. The tasks also

differed; for instance, subjects chose between two repeatedly

presented items during training but evaluated unique individual

items at test. In addition, training trials were consistent, while test

trials involved evaluating different item types in a random pattern.

These confounds would make it impossible to reliably interpret the

results of such subtractions, and hence we cannot draw conclusions

concerning the non-overlapping areas of activation between

training and test.

Discussion

We taught two groups of participants a novel semantic category

using two different categorization strategies. Participants’ beha-

vioral responses were consistent with the qualitative distinctions

between rule-based and similarity-based categorization processes

that have been demonstrated elsewhere (Allen and Brooks, 1991;

Grossman et al., 2003b; Koenig et al., 2002; Rips, 1989; Smith and

Sloman, 1994): Categorization judgments following rule-based

training thus resulted in a clearly defined all-or-none category,

while categorization judgments following similarity-based training

resulted in graded membership that reflected an item’s resemblance

to the prototype. We monitored regional cerebral activity with

fMRI while participants first learned a novel semantic category

under rule-based or similarity-based conditions, and while they

subsequently categorized these items in accordance with their

training. We found distinct condition-specific activation profiles

during training, and unique patterns of activation during the

subsequent test phase that overlapped with the corresponding

training activation profiles. These findings are consistent with the

hypothesis that semantic categorization can be performed by at

least two qualitatively different processes, and that these processes

rely on distinct large-scale neural networks. We discuss each of

these processes and the associated brain activation patterns below.

Activation patterns during rule-based semantic categorization

Rule-based acquisition

There is considerable evidence that brain regions associated

with working memory and executive processes such as selective

attention are recruited during rule-based categorization. The rule

condition thus activated dorsal portions of left inferior frontal

cortex (BA 6 and 44). This region shows activation during many

tasks involving verbal working memory (Smith and Jonides,

1999; Smith et al., 1998a,b), and appears to be associated in

particular with maintaining information in an active state in

working memory (Awh et al., 1996; Smith and Jonides, 1999).

Recruitment may serve to maintain in verbal working memory

the features needed to diagnose category membership in

accordance with rules. Involvement of the prefrontal cortex

has been found in single-cell recordings of nonhuman primates

(i.e., monkeys), apparently specific to the category membership

status of items in recently learned all-or-none categories (Miller

et al., 2002). The important contribution of working memory is

emphasized by the significantly increasing left inferior frontal

activation during the course of rule-based training. This is in

keeping with the increased accuracy of judgments observed in

the latter part of the training phase, and suggests that the need

for working memory to employ the rules continued despite

subjects’ increasing familiarity with the visual appearance of the

stimuli. Brain regions associated with working memory are often

lateralized depending on the verbal or visual perceptual-spatial

nature of the material, with right-sided dominance for spatial

working memory and left-sided activation for verbal working

memory (Smith, 1997; Smith et al., 1996). An example of such

lateralization is apparent in the left-sided activation seen during

rule-based decision-making with verbally presented stimuli

(Grossman et al., 2002a,b), and in the right-sided activation

during rule-based decision-making with images (Patalano et al.,

2001). In the present study, although the novel category was

visual, left-lateralized activation may reflect the presentation of

the features relevant to the rule as verbal descriptions.

We also observed left anterior cingulate activation during rule-

based training. Some studies associate this brain region with

components of attention important for selecting and maintaining

sensitivity to specific features (Benedict et al., 1998; Corbetta

et al., 1993; Coull et al., 1996), compatible with Posner’s

executive component of attention (Posner and Petersen, 1990).

Thus, anterior cingulate may be important for selectively attending

to specific, critical features that determine membership in the new

category. Another line of work emphasizes the importance of the

anterior cingulate for governing the response to competing

candidate choices during task performance (Braver et al., 2001;

Carter et al., 1998; MacDonald et al., 2000), although competition

between response alternatives was also present during similarity-

based training. Regardless of the specific basis for anterior

cingulate recruitment, activation of this area and inferior frontal

cortex emphasizes the important role of executive processes during

rule-based categorization.

We also found left thalamic activation during rule-based

training. This region also appears to play an important role in

selective (Frith and Friston, 1996; LaBerge, 1995; Shulman et al.,

1997), a necessary component in the feature-selection process

inherent in rule-based, but not similarity-based, categorization. The

connectivity pattern of the thalamus is consistent with this structure

serving an attentional gating role, since it appears to modulate the

P. Koenig et al. / NeuroImage 24 (2005) 369–383 379

projection of sensory-motor information to the cerebral cortex and

cortical–cortical connectivity.

Parietal activation in BA 7 was observed during rule-based

learning, particularly during the early phase of training. Recruit-

ment of this parietal area has been associated with perception of the

spatial relationships among the feature elements of the stimuli

(Grady et al., 1994; Horwitz et al., 1992; Wilkinson et al., 2002),

and with visual selective attention (Bushnell et al., 1981). This

activation may diminish over time as subjects accumulate familiar-

ity with the spatial configuration of the stimuli, similar to

observations of diminishing cortical activation during learning in

studies of repetition priming (Buckner and Koutsaal, 1998;

Buckner et al., 1998; Raichle et al., 1994). It is noteworthy that

diminution during the course of learning was also found in the

thalamus. This finding suggests that the attention-gating role

played by the thalamus may be linked to diminishing cortical

activation during learning. The link between diminished activation

in these areas and learning over the course of training is consistent

with the increase in judgment accuracy that participants showed

during the final training trials.

Rule-based categorization

The pattern of activation in the test phase for rule-based

categorization judgments appears to represent recruitment of a

large-scale neural network that partially incorporates that which

was recruited during training. We observed recruitment of dorsal

portions of left inferior frontal cortex during the rule condition.

Inferior frontal activation suggests that participants maintained the

rules guiding categorization decisions in verbal working memory

(Awh et al., 1996; Smith and Jonides, 1999). Inferior frontal

activation was more evident during categorization of LOW

DISTORTION than MEMBER + HIGH DISTORTION items.

These items, which had two of the four requisite features, may

depend more on an exhaustive analysis of the rules maintained in

working memory because endorsing them as members requires

inspection of all four relevant features, in contrast with items which

can be accepted when three defining features have been identified.

It is also possible that assessing MEMBER and HIGH DISTOR-

TION items was facilitated because of their resemblance to items

that were repeatedly seen during training, in contrast to LOW

DISTORTION items. Hence, subjects may have had to rely more

exclusively on the application of rules in classifying the unfamiliar

LOW DISTORTION items. The greater reaction times for these

items are consistent both of these interpretations.

Anterior cingulate, also recruited during rule-based acqui-

sition, was again recruited in the left hemisphere during rule-

based categorization relative to similarity-based categorization.

Previous studies involving rule-based categorization also dem-

onstrated anterior cingulate activation (Elliott et al., 1999; Rao

et al., 1997; Savage et al., 2001). Anterior cingulate activation

was more strongly associated with rule-based categorization of

MEMBER + HIGH DISTORTION than with LOW DISTOR-

TION items. Together with the left parietal activation that was

also was seen during rule-based decisions about MEMBER +

HIGH DISTORTION items, these observations begin to empha-

size the contribution of selective attention to the diagnostic value

of specific perceptual features during rule-based processing of

objects.

Our findings thus suggest a pivotal role for executive resources

during rule-based categorization. Evidence for this first emerges

during acquisition. An anterior cingulate-thalamic-parietal network

appears to support selective attention to specific perceptual

features, and inferior frontal cortex is recruited increasingly during

rule-based learning to maintain rules in an activated state in

working memory. During subsequent categorization, anterior

cingulate and parietal cortex continue to play a particularly

important role in selectively attending to perceptual features of

stimuli, and inferior frontal cortex to maintain rules in working

memory.

Activation patterns during similarity-based semantic

categorization

Similarity-based acquisition

Similarity-based training was associated with bilateral anterior

prefrontal activation. This region is recruited in retrieval from long-

term memory (Buckner and Koutsaal, 1998; Buckner et al., 1998;

Lepage et al., 2002). We found activation in a posterior cingulate

distribution during similarity-based training as well. As with

anterior prefrontal cortex, activation of this area has been

associated with retrieval from episodic memory, particularly when

retrieval has been successful (Kapur et al., 1995; Rugg et al., 1998;

Tulving et al., 1996). It is noteworthy that posterior cingulate

recruitment increased over the course of learning, as did the

accuracy of participants’ judgments. Together with anterior

prefrontal activation, posterior cingulate recruitment during si-

milarity-based training, but not rule-based training, is consistent

with subjects’ increasingly depending on the recall of specific

exemplars during similarity-based training; that is, unlike rule-

based training in which each trial imposes the same rules on

successive training items, similarity-based training involves the

gradual formation of the category with exposure to a variety of

members.

We also observed bilateral temporal–parietal activation. Some

studies have emphasized the role of this area in the overall

configuration of features that may be important for similarity-

based judgments (Wilkinson et al., 2002). Relatedly, other work

associates activation of this region with object meaning

(Grossman et al., 2003a; Martin et al., 1995; Owen et al.,

1996a,b). Temporal–parietal cortex is a heteromodal association

region with a connectivity pattern that involves reciprocal

projections with modality-specific association cortices (Mesulam

et al., 1977; Pandya and Kuypers, 1969; Seltzer and Pandya,

1978). Activation in this brain region thus appears to be well suited

to integrating multiple features of a complex stimulus such as a

novel category for the purpose of deriving object meaning. The

increasing activation of these temporal–parietal regions over the

course of similarity-based training appears to emphasize the

importance of this integrative function in similarity-based acqui-

sition strategies.

Similarity-based categorization of a semantic category

The brain regions associated with similarity-based categoriza-

tion during the test phase overlapped the areas activated by

subjects during similarity-based training. Posterior cingulate and

anterior prefrontal activation thus were also evident during

categorization. This may continue to reflect dependence on

successful retrieval of previously encountered exemplars, that is,

the training items and the presented prototype, from memory

(Braver et al., 2001; Buckner and Koutsaal, 1998; Lepage et al.,

2002; Rugg et al., 1998; Tulving et al., 1996). The similarity

condition was also associated with bilateral temporal–parietal

P. Koenig et al. / NeuroImage 24 (2005) 369–383380

recruitment. This region appears to play a continuing role in

integrating multiple features crucial for overall similarity-based

comparisons (Grossman et al., 2002a,b, 2003a).

Our observations thus suggest that similarity-based semantic

categorization activates brain regions important for feature

integration and comparisons with previously encountered exem-

plars. During both similarity-based acquisition and subsequent

categorization, this network was seen to include anterior prefrontal,

posterior cingulate, and temporal–parietal regions.

Activation patterns common to rule- and similarity-based

categorization

Several other brain regions associated with executive resources

were recruited during judgments of LOW DISTORTION items

relative to MEMBER + HIGH DISTORTION items across both the

rule-based and similarity-based conditions at test. These include

left dorsolateral prefrontal cortex and bilateral striatum. MacDo-

nald et al. (2000) have suggested that left dorsolateral prefrontal

cortex selectively attends to information relevant to task demands,

and that such activation can be increased by monitoring of

response conflict by anterior cingulate cortex. We suggested above

that, for rule-based categorization, LOW DISTORTION items may

put greater demands on maintenance of the specific rules in

working memory because all four requisite features must be

inspected. In the similarity-based condition, subjects may have

become sensitive to the conflicting status of LOW DISTORTION

test items that equally resemble MEMBER and HIGH DISTOR-

TION items, and subsequently resolved the category membership

status of the LOW DISTORTION items in a manner mediated in

part by dorsolateral prefrontal cortex. These bequivocalQ stimulus

items, then, may have placed greater demands on executive

processes in both categorization conditions, albeit for different

decision-making strategies.

Some researchers have associated striatal activation with rule-

based categorization (Ashby and Ell, 2001; Ashby et al., 1991).

For example, caudate activation was seen in a study of rule-based

card sorting (Rao et al., 1997). To the extent that judging stimuli

typical of neither members nor nonmembers is inherently resource-

demanding, LOW DISTORTION stimuli may require activation of

the striatum regardless of the rule-based or similarity-based nature

of categorization that we elicited.

Comparison with previous rule and similarity imaging studies

There are few neuroimaging studies directly comparing rule-

and similarity-based categorization. Most studies that examine one

or both of these processes differ critically from the present study,

making informative between-study comparisons difficult. For

instance, Tracy et al. (2003) compared categorization by either

family resemblance (i.e., similarity) or criterion attribute (i.e., an

exhaustive rule). They observed medial parietal, inferior frontal,

and anterior temporal activation in their rule-based task, and

extrastriate and cerebellar activation in their similarity-based task.

However, their stimuli were letter strings; the categories were

structured differently for each condition; subjects had to deduce the

category structure; and finally, the same subjects participated in

both conditions, in alternate blocks. In another example, Reber

et al. (1998b) found decreased posterior occipital activity

associated with classifying members in a resemblance-based

categorization task. However, this was in comparison with a

recognition memory task, rather than a rule-based task. In addition,

the task used semantically meaningless dot patterns which subjects

learned to categorize by an implicit, rather than an explicit,

process. Hence, studies such as these seem unlikely to draw on the

same neural mechanisms as the present study.

The fMRI study by Grossman et al. (2002a,b), described in

Introduction, compared rule-based and similarity-based categori-

zation of minimally described real world objects. Thus, that study

differs from the current study in many ways, including its use of

verbal stimuli, familiar categories (and hence no acquisition

phase), and subject-derived rather than experimenter-imposed rules

(e.g., participants must rely on their prior knowledge to realize that

bquarters must be of a certain size,Q and then reason accordingly).

These differences would seem to account for activation discre-

pancies across the two studies. For instance, left dorsolateral

prefrontal, left anterior cingulate, and left inferior parietal

activation observed in both processing conditions in the Grossman

et al. (2002a,b) study could reflect processing requirements that are

common to both conditions in that study, but not in the present one.

These could include resource-demanding reasoning, responses that

must be monitored in the similarity condition because they are

equally plausible and in the rule condition because they are

potentially error-prone, and feature configuration necessary to

make judgments about a verbally presented category. Nonetheless,

the Rule condition, relative to the Similarity condition, activated

left frontal areas, albeit more dorsal than those activated in the

present study.

The PET imaging study by Patalano et al. (2001) seems

closer in design to the present study, and also shows left inferior

frontal activation during rule-based categorization that is not

observed during similarity-based categorization. The discrepan-

cies between other observed patterns of activation across the two

studies could reflect the many methodological differences: The

current study is event-related, and therefore activation patterns

reported for one condition are relative to those in the other. The

Patalano et al. study, which was PET rather than fMRI, contrasts

activation with within-condition baseline tasks. That study found

visual association cortex activity in both conditions; in the

present study, any such activation common to both conditions

would have been subtracted out. The Patalano et al. study’s

similarity-based condition employed exemplars that were highly

similar to training items (which consisted of two categories),

with a large number of features in common, and no equivocal

items. This may account for the lack of temporal–parietal

activation in that study, which we suggest reflects the processes

of feature configuration necessary to forming the new category

in the current study. The strong item similarity, coupled with the

several day delay between training and test in the Patalano et al.

study, may account for the lack of posterior cingulate activity,

which, in the present study, we attribute to the successful recall

of previous exemplars: Memory for specific exemplars in the

Patalano et al. study may not have been preserved over the

lengthy delay, while such memories were presumably available

in the present study. The Patalano et al. study’s rule condition

consisted of items that either closely resembled or looked greatly

unlike category members; these latter items presumably required

that participants inhibit perceptual dissimilarity in order to

follow the rules. In addition, participants had 3 s, in contrast

to the current study’s 12 s, to make a decision at test. Rule-

condition participants’ relatively low accuracy rates (78%,

compared with 98% in the current study) suggest that they

P. Koenig et al. / NeuroImage 24 (2005) 369–383 381

were unsuccessful in applying rules at least 20% percent of the

time, and it presumably cannot be known whether they

attempted to apply the rules during those times, responded to

perceptual similarity, or simply became inattentive. The current

study includes only successful trials. This may account for the

anterior cingulate activation observed in the current study, which

we attribute to response monitoring.

Conclusion

The observations of the present study extend previous findings

(Grossman et al., 2002a,b; Patalano et al., 2001) of distinct cerebral

activation patterns for rule-based and similarity-based categoriza-

tion to the acquisition and utilization of a new semantic category.

Our training stimuli were reasonably matched across conditions for

verbal and visual content, and contained the identical training

items; our test stimulus displays were completely identical across

conditions. Hence, qualitatively different responses across con-

ditions, particularly in the test phase, argue in favor of multiple

categorization processes. Our overall observation of distinct

patterns of decision-making paired with distinct patterns of cortical

activation thus is difficult to reconcile with approaches to semantic

categorization advocating a single mechanism for all forms of

categorization, with differing responses being attributed to varying

difficulty of task demands (Nosofsky and Johansen, 2000;

Nosofsky and Palmeri, 1998).

We found neural networks unique to each categorization

process. It appears that the acquisition of a new category by a

rule-based strategy is associated with an anterior cingulate–

thalamic–parietal network important for selectively attending to

specific perceptual features, and inferior frontal cortex that

maintains rules in working memory during categorization. During

subsequent categorization of stimuli following rule-based learning,

selective attention to perceptual features, supported by anterior

cingulate and parietal cortex, apparently plays a particularly

important role in the categorization of items with readily apparent

category membership. By comparison, categorization of LOW

DISTORTION stimuli is associated more closely with left inferior

frontal activation, presumably to maintain the rule-based criteria

and evaluated features in an active state in working memory.

Similarity-based categorization seems to consistently activate

temporal–parietal regions, supporting the integration of features

necessary for making overall comparisons, and an anterior

prefrontal–posterior cingulate network to support successful

retrieval from memory of previously encountered exemplars which

determine the structure of the category.

References

Aizenstein, H.J., MacDonald, A.W., Stenger, V.A., et al., 2000. Comple-

mentary category learning systems identified using event-related

functional MRI. J. Cognit. Neurosci. 12, 977–987.

Allen, S.W., Brooks, L.R., 1991. Specializing the operation of an explicit

rule. J. Exp. Psychol. Gen. 12, 3–17.

lsop, D., 1995. Correction of ghost artifacts and distortion in

echoplanar MR with an iterative reconstruction technique. Radiology

194, 338.

Ashby, F.G., Ell, S.W., 2001. The neurobiology of human category

learning. Trends Cognit. Sci. 5, 204–210.

Ashby, F.G., Alfonso-Reese, L.A., Turken, U., Waldron, E.M., 1991. A

neuropsychological theory of multiple systems in category learning.

Psychol. Rev. 105, 442–481.

Awh, E., Jonides, J., Smith, E.E., Schumacher, E.H., Koeppe, R.A., Katz, S.,

1996. PET evidence for a dissociation between the storage and rehearsal

components of verbal working memory. Psychol. Sci. 7, 25–31.

Benedict, R.H., Lockwood, A.H., Shucard, J.L., Shucard, D.W., Wack,

D.S., Murphy, B.W., 1998. Functional neuroimaging of attention in the

auditory modality. NeuroReport 9, 121–126.

Braver, T.S., Barch, D.M., Gray, J.A., Molfese, D.L., Snyder, A., 2001.

Anterior cingulate cortex and response conflict: effects of frequency,

inhibition, and errors. Cereb. Cortex 1, 825–836.

Bruner, J., Goodnow, J., Austin, G., 1956. A Study of Thinking. Wiley,

New York.

Buckner, R.L., Koutsaal, W., 1998. Functional neuroimaging studies of

encoding, priming, and explicit memory retrieval. Proc. Natl. Acad. Sci.

U. S. A. 95, 891–898.

Buckner, R.L., Goodman, J., Burock, M.A., et al., 1998. Functional–

anatomic correlates of object priming in humans revealed by rapid

presentation event-related fMRI. Neuron 20, 285–296.

Bushnell, M.C., Goldberg, M.E., Robinson, D.L., 1981. Behavioral

evidence of visual response in monkey cerebtral cortex: I. Modulation

in posterior parietal cortex related to selective visual attention.

J. Neurophysiol. 46, 755–772.

Carter, C.S., Braver, T.S., Barch, D.M., Botvinick, M.M., Noll, D., Cohen,

J.D., 1998. Anterior cingulate cortex, error detection, and the online

monitoring of performance. Science 280, 747–749.

Cohen, J.D., MacWhinney, B., Flatt, M.R., Provost, J., 1993. PsyScope: a

new graphic interactive environment for designing psychology experi-

ments. Behav. Res. Meth. Instrum. Comput. 25, 101–113.

Corbetta, M., Miezin, F.M., Shulman, G.L., Petersen, S.E., 1993. A PET

study of visuospatial attention. J. Neurosci. 13, 1202–1226.

Coull, J.T., Frith, C.D., Frackowiak, R.S.J., Grasby, P.M., 1996. A

frontal–parietal network for rapid visual information processing: a

PET study of sustained attention and working memory. Neuro-

psychologia 34, 1085–1095.

D’Esposito, M., Detre, J.A., Alsop, D.C., Shin, R.K., Atlas, S.W.,

Grossman, M., 1995. The neural basis of the central executive system

of working memory. Nature 378, 279–281.

Elfgren, C., Passant, U., Riberg, J., 1993. Neuropsychological findings in

frontal lobe dementia. Dementia 4, 214–219.

Elliott, R., Rees, G., Dolan, R.J., 1999. Ventromedial prefrontal cortex

mediates guessing. Neuropsychologia 37, 403–411.

Fisher, L.M., Freed, D.M., Corkin, S., 1990. Stroop color-word test

performance in patients with Alzheimer’s disease. J. Clin. Exp.

Neuropsychol. 12, 745–758.

Frackowiak, R.S.J., Friston, K.J., Frith, C.D., Dolan, R.J., Mazziotta, J.C.,

1997. Human Brain Function. Academic Press, San Diego.

Frith, C.D., Friston, K.J., 1996. The role of the thalamus in btop-downQmodulation of attention to sound. NeuroImage 4, 210–215.

Garavan, H., Ross, T.J., Li, S.-J., Stein, E.A., 2000. A parametric

manipulation of central executive functioning. Cereb. Cortex 10,

585–592.

Grady, C.L., Maisog, J., Horwitz, B., et al., 1994. Age-related changes in

cortical blood flow activation during visual processing of faces and

location. J. Neurosci. 14, 1450–1462.

Grossman, M., Koenig, P., DeVita, C., et al., 2002a. The neural basis

for category-specific knowledge: an fMRI study. NeuroImage 15,

936–948.

Grossman, M., Smith, E.E., Koenig, P., et al., 2002b. The neural basis for

categorization in semantic memory. NeuroImage 15, 1549–1561.

Grossman, M., Koenig, P., Glosser, G., et al., 2003a. Neural basis for

semantic memory difficulty in Alzheimer’s disease: an fMRI study.

Brain 126, 292–311.

Grossman, M., Smith, E.E., Koenig, P., Glosser, G., Rhee, J., Dennis, K.,

2003b. Categorization of object descriptions in Alzheimer’s disease and

frontotemporal dementia. Cognit. Affect. Behav. Neurosci. 3, 120–132.

Horwitz, B., Grady, C.L., Haxby, J.V., Schapiro, M.B., Ungerleider, L.G.,

P. Koenig et al. / NeuroImage 24 (2005) 369–383382

Mishkin, M., 1992. Functional associations among human posterior

extrastriate brain regions during object and spatial vision. J. Cognit.

Neurosci. 4, 311–322.

Kapur, S., Craik, F.I.M., Jones, C., Brown, G.M., 1995. Functional role of

the prefrontal cortex in retrieval of memories: a PET study. Neuro-

Report 6, 1880–1884.

Kemler, D., Smith, L., 1979. Accessing similarity and dimensional

relations: effects of integrality and separability on the discovery of

complex concepts. J. Exp. Psychol. Gen. 108, 133–150.

Koenig, P., Smith, E.E., Moore, P., Dennis, K., Grossman, M., 2002.

Category acquisition in Alzheimer’s disease, frontotemporal dementia,

and corticobasal degeneration. Neurology 58, A353.

LaBerge, D., 1995. Attentional Processing. Harvard Univ. Press, Cam-

bridge, MA.

LaFleche, D., Albert, M.S., 1995. Executive function deficits in mild

Alzheimer’s disease. Neuropsychology 9, 313–320.

Lepage, M., Ghaffar, O., Nyberg, L., Tulving, E., 2002. Prefrontal cortex

and episodic retrieval mode. Proc. Natl. Acad. Sci. U. S. A. 97,

506–511.

MacDonald, A.W., Cohen, J.D., Stenger, V.A., Carter, C.S., 2000.

Dissociating the role of the dorsolateral prefrontal and anterior cingulate

cortex in cognitive control. Science 288, 1835–1838.

Martin, A., Haxby, J.V., Lalonde, F.M., Wiggs, C.L., Ungerleider, L.G.,

1995. Discrete cortical regions associated with knowledge of color and

knowledge of action. Science 270, 102–105.

Medin, D.L., Schaffer, M.M., 1978. A context theory of classification

learning. Psychol. Rev. 85, 207–238.

Medin, D.L., Goldstone, R.L., Gentner, D., 1993. Respects for similarity.

Psychol. Rev. 100, 254–278.

Mesulam, M.-M., van Hoesen, G.W., Pandya, D.N., Geschwind, N., 1977.

Limbic and sensory connections of the inferior parietal lobule (Area

PG) in the rhesus monkey: a study with a new method of horseradish

peroxidase histochemistry. Brain Res. 136, 393–414.

Miller, E.K., Freedman, D.J., Wallis, J.D., 2002. The prefrontal cortex:

categories, concepts, and cognition. Philos. Trans. R. Soc. Lond., Ser. B

357, 1123–1136.

Monchi, O., Petrides, M., Petre, V., Worsley, K.J., Dagher, A., 2001.

Wisconsin card sorting revisited: distinct neural circuits participating in

different stages of the task identified by event-related functional

magnetic resonance imaging. J. Neurosci. 21, 7733–7741.

Nosofsky, R.M., Palmeri, T.J., 1998. A rule-plus-exception model for

classifying objects in continuous-dimension spaces. Psychon. Bull. Rev.

5, 345–369.

Nosofsky, R.M., Johansen, M.K., 2000. Exemplar-based accounts of

bmultiple-systemQ phenomena in perceptual categorization. Psychon.

Bull. Rev. 7, 375–402.

Owen, A.M., Doyon, J., Petrides, M., Evans, A., 1996a. Planning and

spatial working memory: a positron emission tomography study in

humans. Eur. J. Neurosci. 8, 353–364.

Owen, A.M., Milner, B., Petrides, M., Evans, A.C., 1996b. Memory for

object features vs. memory for object location: a positron-emission

tomography study of encoding and retrieval processes. Proc. Natl.

Acad. Sci. U. S. A. 93, 9212–9217.

Pandya, D.N., Kuypers, H.G.J.M., 1969. Cortico-cortical connections in the

rhesus monkey. Brain Res. 13, 13–36.

Pasquier, F., Lebert, F., Lavenu, I., Guillaume, B., 1999. The clinical picture

of frontotemporal dementia: diagnosis and follow-up. Dement. Geriatr.

Cogn. Disord. 10, 10–14.

Patalano, A., Smith, E.E., Jonides, J., Koeppe, R., 2001. PET evidence for

multiple strategies of categorization. Cogn. Affect. Behav. Neurosci. 1,

360–370.

Perry, R.J., Hodges, J.R., 1999. Attention and executive deficits in

Alzheimer’s disease: a critical review. Brain 122, 383–404.

Petrides, M., Alivisatos, B., Evans, A.C., Meyer, E., 1993. Dissoci-

ation of human mid-dorsolateral from posterior dorsolateral frontal

cortex in memory processing. Proc. Natl. Acad. Sci. U. S. A. 90,

873–877.

Poldrack, R.A., Gabrieli, J.D.E., 2001. Characterizing the neural mecha-

nisms of skill learning and repetition priming: Evidence from mirror

reading. Brain 124, 67–82.

Posner, M., Petersen, S., 1990. The attention system of the human brain.

Annu. Rev. Neurosci. 13, 25–42.

Raichle, M.E., Fiez, J.A., Videen, T.O., et al., 1994. Practice-related

changes in human brain functional anatomy during nonmotor learning.

Cereb. Cortex 4, 8–26.

Rao, S.M., Bobholz, J.A., Hammeke, T.A., et al., 1997. Functional MRI

evidence for subcortical participation in conceptual reasoning skills.

NeuroReport 8, 1987–1993.

Reber, P.J., Stark, C.E.L., Squire, L.R., 1998a. Cortical areas supporting

category learning identified using functional MRI. Proc. Natl. Acad.

Sci. U. S. A. 95, 747–750.

Reber, P.J., Stark, C.E.L., Squire, L.R., 1998b. Contrasting cortical activity

associated with category memory and recognition memory. Learn.

Mem. 5, 420–428.

Rips, L.J., 1989. Similarity, typicality, and categorization. In: Vosniadou, S.,

Ortony, A. (Eds.), Similarity and Analogical Reasoning. Cambridge

Univ. Press, Cambridge, pp. 21–59.

Rugg, M.D., Fletcher, P.C., Allan, K., Frith, C.D., Frackowiak, R.S.J.,

Dolan, R.J., 1998. Neural correlates of memory retrieval during

recognition memory and cued recall. NeuroImage 8, 262–273.

Savage, C.R., Deckersbach, T., Heckers, S., et al., 2001. Prefrontal regions

supporting spontaneous and directed application of verbal learning

strategies: evidence from PET. Brain 124, 219–231.

Seger, C.A., Poldrack, R.A., Prabhakaran, V., Zhao, M., Glover, G.H.,

Gabrieli, J.D.E., 2000a. Hemispheric asymmetries and individual

differences in visual concept learning as measured by functional MRI.

Neuropsychologia 38, 1316–1324.

Seger, C.A., Prabhakaran, V., Poldrack, R.A., Gabrieli, J.D.E., 2000b.

Neural activity differs between explicit and implicit learning of artificial

grammar strings: an fMRI study. Psychobiology 2, 283–292.

Seltzer, B., Pandya, D.N., 1978. Afferent cortical connections and

architectonics of the superior temporal sulcus and surrounding cortex

in the rhesus monkey. Brain Res. 149, 1–24.

Shulman, G.L., Corbetta, M., Buckner, R.L., et al., 1997. Common blood

flow changes across visual tasks: I. Increases in subcortical structures

and cerebellum but not in nonvisual cortex. J. Cognit. Neurosci. 9,

624–647.

Smith, E.E., 1995. Concepts and categorization. In: Osherson, D., Smith,

E.E. (Eds.), Invitation to Cognitive Science, Thinking, vol. 2. MIT

Press, Cambridge, pp. 1–25.

Smith, E.E., 1997. Working Memory: A View from Neuroimaging. Cogn.

Psychol. 3, 5–42.

Smith, E.E., Jonides, J., 1999. Storage and executive processes in the

frontal lobes. Science 283, 1657–1661.

Smith, E.E., Medin, D.L., 1981. Categories and Concepts. Harvard Univ.

Press, Cambridge.

Smith, E.E., Sloman, S.A., 1994. Similarity-versus rule-based categoriza-

tion. Mem. Cogn. 2, 377–386.

Smith, E.E., Jonides, J., Koeppe, R., 1996. Dissociating verbal and spatial

working memory using PET. Cereb. Cortex 6, 11–20.

Smith, E.E., Jonides, J., Marshuetz, C., Koeppe, R.A., 1998a. Components

of verbal working memory: evidence from neuroimaging. Proc. Natl.

Acad. Sci. U. S. A. 95, 876–882.

Smith, E.E., Patalano, A.L., Jonides, J., 1998b. Alternative strategies for

categorization. Cognition 65, 167–196.

Talairach, J., Tournoux, P., 1988. Co-Planar Stereotaxic Atlas of the Human

Brain. Thieme Medical Publishing Company, New York.

Tracy, J.I., Mohamed, F., Faro, S., Pinus, A., Tiver, R., Harvan, J., Bloomer,

C., Pyrros, A., 2003. Differential brain responses when applying

criterion attribute versus family resemblance rule learning. Brain Cogn.

51, 276–286.

Tulving, E., 1972. Episodic and semantic memory. In: Tulving, E.,

Donaldson, W. (Eds.), Organization of Memory. Academic Press,

New York.

P. Koenig et al. / NeuroImage 24 (2005) 369–383 383

Tulving, E., Markowitsch, H.J., Craik, F.I.M, Habib, R., Houle, S., 1996.

Novelty and familiarity activations in PET studies of memory encoding

and retrieval. Cereb. Cortex 6, 71–79.

Vendrell, P., Junque, C., Pujol, J., Jurado, M.A., Molet, J., Grafman, J.,

1995. The role of the prefrontal regions in the Stroop task. Neuro-

psychologia 33, 341–352.

Wagner, A.D., Maril, A., Bjork, R.A., Schacter, D.L., 2001. Prefrontal

contributions to executive control: fMRI evidence for functional dis-

tinctions within lateral prefrontal cortex. NeuroImage 14, 1337–1347.

Wilkinson, D.T., Halligan, P.W., Henson, R.N.A, Dolan, R.J., 2002. The

effects of interdistracter similarity on search processes in the superior

parietal cortex. NeuroImage 1, 611–619.