Published online 30 July 2002

The prefrontal cortex categories conceptsand cognition

Earl K Miller David J Freedman and Jonathan D WallisCenter for Learning and Memory RIKEN-MIT Neuroscience Research Center and Department of Brain

and Cognitive Sciences Massachusetts Institute of Technology Cambridge MA 02139 USA

The ability to generalize behaviour-guiding principles and concepts from experience is key to intelligentgoal-directed behaviour It allows us to deal ef ciently with a complex world and to adapt readily tonovel situations We review evidence that the prefrontal cortexmdashthe cortical area that reaches its greatestelaboration in primatesmdashplays a central part in acquiring and representing this information The prefrontalcortex receives highly processed information from all major forebrain systems and neurophysiologicalstudies suggest that it synthesizes this into representations of learned task contingencies concepts andtask rules In short the prefrontal cortex seems to underlie our internal representations of the lsquorules ofthe gamersquo This may provide the necessary foundation for the complex behaviour of primates in whomthis structure is most elaborate

Keywords neurophysiology cognition volition attention control

1 INTRODUCTION

Although our brains have developed exquisite mechanismsfor recording speci c experiences it is not always advan-tageous for us to take the world too literally A brain lim-ited to storing an independent record of each experiencewould require a prodigious amount of storage and burdenus with unnecessary details Instead we have evolved theability to detect the commonalities among experiences andstore them as abstract concepts general principles andrules This is an ef cient way to deal with a complex worldand allows the navigation of many different situations witha minimal amount of storage It also allows us to deal withnovelty By extracting the essential elements from ourexperiences we can generalize to future situations thatshare some elements but may on the surface appearvery different

For example consider the concept lsquocamerarsquo We do nothave to learn anew about every camera that we mayencounter Just knowing that the item is a camera com-municates a great deal of knowledge about its parts func-tions and operations Or consider the set of rules invokedwhen we dine in a restaurant such as lsquowait to be seatedrsquolsquoorderrsquo and lsquopay the billrsquo These rules are long divorcedfrom the speci c circumstances in which they were learnedand thus give us an idea about what to expect (and whatis expected of us) when we try out a new restaurant Hear-ing that a lsquocoup drsquoetatrsquo has occurred communicates thelsquogistrsquo of what happened without having to hear the details

While much is known about the encoding of physicalattributes and speci c experiences relatively little isknown about how abstract information is encoded in the

Author for correspondence (ekmaimitedu)

One contribution of 14 to a Discussion Meeting Issue lsquoThe physiologyof cognitive processesrsquo

Phil Trans R Soc Lond B (2002) 357 1123ndash1136 1123 Oacute 2002 The Royal SocietyDOI 101098rstb20021099

brain This may be because this is relatively more dif cultto study than the neural correlates of physical attributessuch as shape By de nition these categories and conceptsare labels that transcend physical appearance Think of allthe wildly different-looking objects that are considered tobe chairs However it is not always easy to disentangleencoding of categories from similarity Let us say that wediscover some neurons somewhere in a monkeyrsquos brainthat become activated whenever it views a tree Are theseneurons really encoding the category lsquotreersquo They mightbe encoding the fact that trees happen to look more likeone another than many other objects Further develop-ment of abstract representations requires a considerableamount of experience learning a general principle requiresa wide range of experiences so that underlying rules canbe extracted

Our laboratory has conducted experiments to establishhow abstract information is represented in the brain Wetrained monkeys on tasks that allowed them to group differ-ent stimuli and experiences into categories or behaviour-guiding rules We summarize some of that work and dis-cuss its implications for an understanding of a neural basisof high-level cognitive function We have focused on abrain region that is central to high-level cognitive functionthe PFC

2 THE PREFRONTAL CORTEX

The PFC is an ideal place to look for neural correlatesof abstract information It occupies a far greater pro-portion of the human cerebral cortex than in other ani-mals suggesting that it might contribute to those cognitivecapacities that distinguish humans from animals (Fuster1995 gure 1) On initial examination PFC damage hasremarkably little overt effect patients can perceive andmove there is little impairment in their memory and theycan appear remarkably normal in casual conversation

1124 E K Miller and others Concepts and the prefrontal cortex

squirrelmonkey

cat

dog

rhesusmonkey

chimpanzee

man

pscs

pf

pfprf

as

cs

gprpf

Figure 1 The relative size of the PFC in different animalsAbbreviations as arcuate sulcus cs cingulate sulcusgpr gyrus proreus pf presylvian ssure ps principalsulcus prf proreal ssure From Fuster (1995)

However despite the super cial appearance of normalityPFC damage seems to devastate a personrsquos life They havedif culty in sustaining attention in keeping lsquoon taskrsquo andseem to act on whims and impulses without regard tofuture consequences This pattern of high-level de citscoupled with a sparing of lower-level basic functions hasbeen called a lsquodysexecutive syndromersquo (Baddeley amp DellaSala 1996) and lsquogoal neglectrsquo (Duncan et al 1996)

Indeed the anatomy of the PFC suggests that it is well-suited for a role as the brainrsquos lsquoexecutiversquo It can synthesizeinformation from a wide range of brain systems and exertcontrol over behaviour (Nauta 1971) The collection ofcortical areas that comprise the PFC have intercon-nections with brain areas processing external information(with all sensory systems and with cortical and subcorticalmotor system structures) as well as internal information(limbic and midbrain structures involved in affect mem-ory and reward) (see gure 2) Correspondingly its neu-rons are highly multimodal and encode many differenttypes of information from all stages of the perceptionndashaction cycle (Fuster 1995) They are activated by stimulifrom all sensory modalities before and during a variety ofactions during memory for past events in anticipation ofexpected events and behavioural consequences and aremodulated by internal factors such as motivational andattentional state (see the review in Miller amp Cohen 2001)Because of its highly multimodal nature and its apparentrole in higher mental life the PFC seemed like an idealplace to begin our search for neural correlates of theabstract information needed for intelligent behaviour

Phil Trans R Soc Lond B (2002)

3 THE PREFRONTAL CORTEX AND PERCEPTUALCATEGORIES

Because perceptual categories often group together verydifferent looking things their representation must involvesomething beyond the sort of neural tuning that underliesthe encoding of physical attributes that is gradualchanges in neural activity as certain attributes graduallychange (eg shape orientation direction) Instead categ-ories have sharp boundaries (not gradual transitions) andmembers of the same category are treated as equivalenteven though their physical appearance may vary widely

Consider a simple example of a perceptual categorycrickets sharply divide a certain range of pure tones intolsquomatersquo versus lsquobatrsquo (a predator) (Wyttenbach et al 1996)Even though the input varies along a continuum behav-iour is binary Across a wide range of lower frequenciescrickets will turn towards the sound source because it maybe a potential mate However at a certain point (16 kHz)the behaviour suddenly ips crickets begin to turn awayfrom the sound source because it could be a bat Cricketsmake virtually no distinction between frequencies over awide range on either side of the boundary they approachor avoid with equal reliability This type of representationis illustrated schematically in gure 3 Presumably theability to transform the raw sensory inputs into distinctcategories evolved because it is advantageous in this caseit optimizes reproductive behaviour while minimizing fatalmistakes Similar effects are evident in humansrsquo percep-tion of lsquobrsquo versus lsquoprsquo (Lisker amp Abramson 1970)

The elaborate behavioural repertoire of advanced ani-mals naturally depends on more elaborate categorizationabilities The mental lexicon of primates for exampleincludes abstract categories that are characterized alongmultiple dimensions that are often dif cult to de ne pre-cisely such as lsquotoolrsquo In addition advanced animals havean enormous capacity to learn and adapt Most of ourcategories are acquired through experience (we learn whata lsquochairrsquo is) and we can continually modify and update ourcategories as we learn more about them The ability ofmonkeys to learn complex perceptual categories has beencatalogued in studies that have taught them categoriessuch as animal versus non-animal (Roberts amp Mazmanian1988) food versus non-food (Fabre-Thorpe et al 1998)tree versus non-tree shes versus non- shes (Vogels1999a) and ordinal numbers (Orlov et al 2000) Pigeonsalso have a remarkable ability to learn such distinctions(Bhatt et al 1988 Young amp Wasserman 1997)

Where such categories are encoded in the brain isunclear In primates they could be represented and storedin the same areas of the visual cortex that analyse formand are critical for remembering individual objects suchas the ITC They might also be evident in the brainregions that receive the results of visual processing fromthe ITC and are critical for planning and guiding behav-iour such as the PFC Both the ITC and PFC containneurons selective for complex stimuli such as trees shesfaces brushes etc (Desimone et al 1984 Tanaka et al1991 Vogels 1999b) But whether or not this selectivityre ects category information per se has not been determ-ined With a large amorphous category (eg food humanetc) the category boundaries are unknown Thus charac-teristics diagnostic of category representations (sharp tran-

Concepts and the prefrontal cortex E K Miller and others 1125

mid-dorsal

area 9

dorsolateral

area 46

prefrontal cortex

ventolateral

area 12 45

orbital and medial

areas 10 11

13 14

area 8

(FEF)

motor

structures

basal

ganglia

thalamusmedial temporal

lobe

visual

dorsal

ventral

somatosensory

caudal parietal lobe

auditory

superior temporal

gyrus

multimodal

rostral superior

temporal sulcus

sensory cortex

Figure 2 Schematic diagram of extrinsic and intrinsic connections of the PFC Most connections are reciprocal theexceptions are noted by arrows From Miller amp Cohen (2001)

sitions and little within-category distinction) cannot betested This is not to say that ITC neural selectivity wouldnot make important contributions to category represen-tation it is just not clear whether it represents categorymembership per se

To test for neural correlates of perceptual categories wetrained monkeys to categorize computer-generated stimuliinto two categories lsquocatsrsquo and lsquodogsrsquo (Freedman et al2001 gure 4) A novel three-dimensional morphing sys-tem was used to create a large set of parametric blends ofsix prototype images (three species of cats and threebreeds of dogs) (Beymer amp Poggio 1996 Shelton 2000)By blending different amounts of cat and dog we couldsmoothly vary shape and precisely de ne the boundarybetween the categories (greater than 50 of a given type)As a result stimuli that were close to but on oppositesides of the boundary were similar whereas stimuli thatbelong to the same category could be dissimilar (eg thelsquocheetahrsquo and lsquohouse catrsquo)

Phil Trans R Soc Lond B (2002)

Two monkeys performed a DMC task ( gure 5) thatrequired judgement of whether successively presentedsample and test stimuli were from the same category ornot For training samples were chosen from throughoutthe cat and dog morph space After training classi cationperformance was high (ca 90 correct) even when thesamples were close to the category boundary The monk-eys classi ed dog-like cats (60 cat 40 dog) correctlyca 90 of the time and misclassi ed them as dogs only10 of the time and vice versa Thus the monkeysrsquobehaviour indicated the sharp boundary that is diagnosticof a category representation The dog-like cats weretreated as cats even though they were more similar inappearance to the cat-like dogs just across the categoryboundary than they were to the prototype cats

We recorded in the lateral PFC the PFC region directlyinterconnected with the ITC and found many examplesof neurons that seemed to encode category membershipTwo examples are shown in gure 6 Note that their

1126 E K Miller and others Concepts and the prefrontal cortex

category boundary

16

tone frequency (kHz)

avo

idan

ceap

pro

ach

beh

avio

ur

Figure 3 Schematic representation of categorical perceptionusing cricketsrsquo responses to a continuum of pure tones as anexample Based on Wyttenbach et al (1996)

activity was different from dog-like (60) cats and cat-like(60) dogs yet similar between these stimuli and theirrespective prototypes In other words PFC neuronsseemed to make the same sharp distinctions that were evi-dent in the monkeysrsquo behaviour and are indicative of cat-egorical representations and collected different stimulitogether irrespective of their exact physical experienceCategory information seemed to predominate in the PFCacross the neural population tuning was signi cantlyshifted towards representing category rather than individ-ual stimuli (Freedman et al 2001)

As our monkeys had no experience with cats or dogsprior to training it seemed likely that these effects resultedfrom training To test for learning effects we retrainedone monkey on the DMC task after de ning two new cat-egory boundaries that were orthogonal to the originalboundary ( gure 4) This created three new classes eachcontained morphs centred around one cat prototype andone dog prototype (eg the cheetah and the lsquodobermanrsquo)After training we found that PFC neural activity shiftedto re ect the new but not the old categories An exampleof a single neuron is shown in gure 7 It showed a sig-ni cant effect of category during the delay period whendata were sorted according to the (currently relevant)three-category scheme it distinguished one of the categor-ies from the other two ( gure 7a) By contrast when thedata were sorted using the old category scheme there wasno differentiation between the now-irrelevant cat and dogcategories ( gure 7b)

Our results illustrate that with experience categoryinformation can become incorporated at the single-neuronlevel much as physical attributes of stimuli are This didnot have to be the case in principle categories might havebeen encoded in another fashion For example categoriesmight have been encoded at the ensemble level as anemergent property of neurons that represent their de ningfeatures This ability to carve category membership intothe tuning of single neurons may allow for the quick andeffortless classi cation of familiar items

Our results might re ect a relative specialization of the

Phil Trans R Soc Lond B (2002)

PFC in encoding category membership Categories afterall are typically de ned by their behavioural relevanceand the PFC plays a central part in planning voluntarybehaviours Conversely the traditional roles of the PFCand ITC are in cognitive functions versus object visionand recognition respectively PFC damage causes de citsin attention working memory and response inhibition butusually spares object recognition long-term memory andlsquohigh-levelrsquo visual analysis (Fuster 1989 Miller amp Cohen2001) By contrast ITC damage causes de cits in visualdiscrimination and learning (Gross 1973 Mishkin 1982)and category-speci c agnosias (eg for faces) in humans(Gainotti 2000) It might be that the category informationin the PFC was retrieved from long-term storage in theITC for its immediate use in the task Interactionsbetween the PFC and ITC underlie the storage andorrecall of visual memories and associations (Rainer et al1999 Tomita et al 1999) Tomita et al (1999) demon-strated that top-down signals from the PFC were neededto activate long-term visual memories stored in the ITCA similar relationship may exist for the recall of visual cat-egories In either case it seems that category informationis strongly represented in the PFC a nding consistentwith its role in high-level cognitive functions and in guid-ing behaviour The relative contribution of the ITCremains to be determined

4 THE PREFRONTAL CORTEX AND RULES

It is not only useful to group different sensory stimuliinto meaningful categories it is also useful to group spe-ci c experiences of our interactions with the environmentalong common themes that is as behaviour-guidingprinciples or rules To this end our brains have evolvedmechanisms for detecting and storing often-complexrelationships between situations actions and conse-quences By gleaning this knowledge from past experi-ences we can develop a lsquogame planrsquo that allows us toextrapolate and infer which goals are available in similarsituations in the future and what actions are likely to bringus closer to them

A standard behavioural test of rule learning in monkeysis conditional associative learning (Passingham 1993)This refers to a class of tasks that require learning associat-ive relationships that are arbitrary and extend beyond thesimple one-to-one stimulusndashresponse mappings thatunderlie re exive reactions to the environment In con-ditional learning tasks a given input does not invariablylead to a given output Whether or not a given response issuccessful depends on additional contextual informationFor example reaching for popcorn can be rewarding butonly if one takes other information into account if thepopcorn belongs to another person the result could bedisastrous Taking into account complex relationships inorder to decide between alternative actions is presumablywhy volition evolved

To make predictions about which actions are likely toachieve a given goal in a given situation we need to forma pattern of associations between their internal represen-tations that describes their logical relationship (Dickinson1980) Decades of behavioural research have illustratedthat the brain has learning mechanisms that are exquisitelysensitive to behaviourally informative associations (and

Concepts and the prefrontal cortex E K Miller and others 1127

C2

C1 C3

2-class boundary

D1 D3

D2

3-class boundaries

100 C1 80 C1 60 C1 60 D1 80 D1 100 D1

cats

dogs

(a)

(b)

Figure 4 (a) Monkeys learned to categorize randomly generated lsquomorphsrsquo from the vast number of possible blends of sixprototypes For neurophysiological recording 54 sample stimuli were constructed along the 15 morph lines illustrated here(b) Morphs along the C1ndashD1 line From Freedman et al (2001)

fixation

500 ms sample

600 ms delay

1000 ms

test

(non-match)

600 msdelay

600 ms test

(match)

(match)

Figure 5 The delayed match-to-category task A sample was followed by a delay and a test stimulus If the sample and teststimulus were the same category (a match) monkeys were required to release a lever before the test disappeared If they werenot there was another delay followed by a match Equal numbers of match and non-match trials were randomly interleavedFrom Freedman et al (2001)

Phil Trans R Soc Lond B (2002)

1128 E K Miller and others Concepts and the prefrontal cortex

13

10

7

4

1

fixation sample delay choice

firi

ng

rat

e (H

z)

dog 100dog 80dog 60

cat 100cat 80cat 60

_500 0 500 1000 1500 2000

dog 100dog 80dog 60

cat 100cat 80cat 60

70

60

50

40

30

20

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)

(a)

(b)

Figure 6 The average activity of two single neurons tostimuli at the six morph blends The vertical linescorrespond (from left to right) to sample onset offset andtest stimulus onset Activity is pooled over match and non-match trials

35

30

25

20

15

20

15

10

5left right

20

15

10

5

45

40

35

30left right

spik

es s

_ 1

Phil Trans R Soc Lond B (2002)

fixation sample delay choice10

8

6

4

2

0

10

8

6

4

2

0plusmn500 0 500 1000 1500 2000

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)fi

rin

g r

ate

(Hz)

(a)

(b)

Figure 7 Activity of a single PFC neuron after the monkeywas trained on the three-category scheme (see gure 4)(a) The neuronrsquos activity when the data were sorted basedon the three-category scheme (category A black linecategory B grey line category C dotted line) (b) The sameneuronrsquos activity was data-sorted by the now-irrelevant two-category scheme (black line cats grey line dogs)

Figure 8 The activity of four single prefrontal neurons wheneach of two objects on different trials instructed either asaccade to the right or a saccade to the left The linesconnect the average values obtained when a given objectcued one or the other saccade The error bars show thestandard error of the mean Note that in each case theneuronrsquos activity depends on both the cue object and thesaccade direction and that the tuning is nonlinear orconjunctive That is the level of activity to a givencombination of object and saccade cannot be predicted fromthe neuronrsquos response to the other combinations Adaptedfrom Asaad et al (1998)

insensitive to or even discount associations that are notinformative) The underlying neural ensemble of a goal-directed task then might be comprised of neurons whoseactivity re ects task contingencies Many studies haveshown that prefrontal neurons do have this property Thiswork has focused on the lateral PFC because it seems tobe a site of convergence of the information needed to solveconditional sensorindashmotor tasks It is directly intercon-

Concepts and the prefrontal cortex E K Miller and others 1129

80

60

40

20

0

_2000 _1500 _1000 _500 0 500 1000 1500

40

30

20

10

0

20

15

10

5

0

10

8

6

4

2

0

spatialobjectassociative

ITI fix cue delay

trial time (ms) block number

1 2 3 4 5 6 7 8

spik

es s

_ 1(a)

(c)

(b)

(d)

Figure 9 Spike rate versus time histograms for two neurons each sorted by task The nal second of the three-second ITI isrepresented by the rst 1000 ms ( 2 2000 to 2 1000 ms) Fixation occurs soon after (ca 2 1000 to 2 800 ms) Cuepresentation occurs at the time-point marked 0 ms Task-related differences in baseline ring rate were generally observed tobegin in the xation period While the activity of some neurons diverged almost coincident with initial xation (c) the activityof others diverged progressively as the appearance of the cue became more imminent (a) The bar graphs (bd) demonstratethe reproducibility of these small task-speci c changes in activity across multiple repetitions of the same task The mean xation-period ring rate (with standard errors) for each block of trials is shown for the two neurons in (a) and (c) The barsare colour-coded to re ect the task being performed in each block and the colours match those in the histograms to the leftThe light grey line superimposed over these bars shows the activity of these neurons during the second immediately precedingthe ITI

nected with the higher-order sensory and motor cortexand indirectly connected (via the ventromedial PFC) withlimbic structures that process lsquointernalrsquo information suchas memory and reward (Goldman-Rakic 1987 Pandya ampBarnes 1987 Fuster 1989 Barbas amp Pandya 1991) Theneural activity in the lateral PFC re ects this many of itsneurons exhibit multimodal responses (Vaadia et al 1986Watanabe 1992 Rao et al 1997 Rainer et al 1998abWhite amp Wise 1999) Further the lateral PFC is criticalfor normal learning of conditional associations betweensensory cues and voluntary actions (Petrides 1985a 1990Gaffan amp Harrison 1988 Eacott amp Gaffan 1992 Parker ampGaffan 1998b) Indeed following training on conditionallearning tasks as many as 50 of the neurons in the lat-eral PFC show conjunctive tuning for learned associationsbetween cues voluntary actions and rewards

For example Watanabe (1990 1992) trained monkeysto perform tasks in which visual and auditory cues sig-nalled in different trials whether a reward would orwould not be delivered The majority of lateral PFC neu-rons were found to re ect the association between a cueand a reward A given neuron might be activated by a cuebut only when it signalled a reward By contrast another

Phil Trans R Soc Lond B (2002)

neuron might be activated only by a cue that signalled lsquonorewardrsquo In our own experiments we have trained mon-keys to associate in different blocks of trials each of twocue objects with an eye saccade to the right or to the left(Asaad et al 1998) We found that the activity of 44 ofrandomly selected lateral PFC neurons re ected associ-ations between objects and the saccades that theyinstructed ( gure 8) Other neurons had activity thatre ected the cues or the saccades alone but they werefewer in number Fuster et al (2000) have recently shownthat PFC neurons can also re ect learned associationsbetween visual and auditory stimuli

Importantly these changes do not require a prodigiousamount of experience Changes in PFC neural propertiesare evident after one dayrsquos experience and can be detectedafter just a few minutes of training For example Bichotet al (1996) studied the FEFs part of Brodmannrsquos area 8that is important for voluntary eye movements Normallyneurons in this area re selectively to saccade targetsappearing in certain visual eld locations However whenmonkeys were trained to search for a target de ned by aparticular visual attribute (eg red) the neurons in theFEFs acquire sensitivity to that attribute (Bichot et al

1130 E K Miller and others Concepts and the prefrontal cortex

1996) When monkeys were trained to search for a differ-ent target every day neurons not only discriminated thecurrent target but also distracting stimuli that had been atarget on the previous day relative to stimuli that had beentargets even earlier (Bichot amp Schall 1999) Monkeys werealso more likely to make errors in choosing that distractingstimulus It was as though the previous dayrsquos experienceleft an impression in the brain that in uenced neuralactivity and the monkeyrsquos behaviour

We also observed evidence for rapid plasticity in ourstudy of learning of conditional objectndashsaccade associ-ations in the PFC (Asaad et al 1998) Initially the monk-eys chose their responses at random but learned thecorrect cuendashresponse pairing over a few (5ndash15) trials Asthey learned the association neural activity representingthe forthcoming saccadic response appeared progressivelyearlier in successive trials In other words the initiation ofresponse-related delay activity gradually shifted withlearningmdashfrom a point in time just before the executionof the response and reward delivery to an earlier point intime nearly coincident with the presentation of the cue

Further support for a role for PFC neurons in rep-resenting task demands comes from training monkeys toalternate between different task rules This adds anotherlevel of complexity beyond the conditional tasks describedabove Now there is more than one rule assigned to eachcue and another cue tells the monkey which rule to usein a given trial For example following a given cue monk-eys can learn to direct a response to either the cuersquoslocation (spatial matching rule) or an alternative locationassociated with the cue (associative rule) depending onwhich rule is currently in effect When tested in thisfashion many PFC neurons show rule-speci c activityFor example a PFC neuron might respond to a given vis-ual cue when the monkey is using an associative rule butexhibit weak or no activity under identical sensory andattentional conditions that differed only in that the mon-key was using a spatial rule instead (White amp Wise 1999Asaad et al 2000) Also when monkeys switch betweendifferent tasks many PFC neurons show shifts in baselineactivity that communicates which task is currently beingperformed (Asaad et al 2000 gure 9) Such effects havebeen found in the PFC for associative versus spatial rulesfor object matching versus spatial matching versus associ-ative rules and for shape matching versus object matchingrules (Hoshi et al 1998 White amp Wise 1999 Asaad etal 2000)

In all of these cases however the rules are relativelyliteral or concrete A certain cue or set of cues alwayssignals a speci c response As noted knowledge from ourpast experiences can be applied to a wider range of futurecircumstances if we abstract general principles or rulesrather than speci c cuendashresponse contingencies Theability of the PFC to represent abstract rules those nottied to speci c stimuli or actions was recently addressedin our laboratory (Wallis et al 2001)

Monkeys were trained to use two abstract rules lsquomatchrsquoversus lsquonon-matchrsquo They faced a computer screen andviewed two successively presented pictures If the matchrule was in effect the monkeys released the lever if thepictures were the same and continued to hold the lever ifthe pictures were different If the non-match rule was ineffect the reverse was true monkeys released if the pic-

Phil Trans R Soc Lond B (2002)

tures were different and held if they were the same Therule was randomly instructed in each trial by the presen-tation of a cue at the same time as the rst picture waspresented To disambiguate neuronal responses to thephysical properties of the cue from responses to the rulethat the cues instructed cues signifying the same rule weretaken from different modalities while cues signifying dif-ferent rules were taken from the same modality ( gure10) The monkeys could perform this task well abovechance levels even when they were seeing a stimulus forthe very rst time This indicates that they had abstractedtwo overarching principles of the task that could then beapplied to novel stimulimdashthe minimal de nition of anabstract rule

The most prevalent activity across the PFC was theencoding of the current rule Figure 11 shows a goodexample of a rule-selective neuron that exhibited greateractivity when the match rule was in effect than when thenon-match rule had been indicated This activity cannotbe explained by the physical properties of the cue or thepicture since activity was the same regardless of whichcue was used to instruct the monkey and regardless ofwhich picture the monkey was remembering It cannot berelated to the upcoming response since the monkey didnot know whether the second-presented picture wouldrequire a response Nor could it be related to differencesin reward expectation since the expectation of reward wasthe same regardless of which rule was in effect Further-more the performance of the monkeys was virtually ident-ical for the two types of rules (error rates differed by lessthan 01 and reaction times by less than 7 ms) Thusthe most parsimonious explanation is that the differencesin activity re ected the abstract rule that the monkey wascurrently using to guide its behaviour

What function does the ability to abstract a rule serveIt is a form of generalization that permits a shortcut inlearning thereby allowing the animal to maximize theamount of reward available from a particular situation Toillustrate this consider the above task The monkey couldpotentially solve the task as a series of paired associates(in fact 16 associations consisting of four different pic-tures each paired with four different stimuli) Forexample the monkey might learn that whenever the chefis presented with a drop of juice then at the test phasethe correct response is to choose the chef But notice thatthis type of learning tells the monkey nothing about whichresponse is appropriate to a lion appearing with a drop ofjuice In other words unless the monkey abstracts the rulethat juice indicates that the monkey should match theneach time new pictures are used the monkey would haveto learn an entirely new set of 16 associations by trial anderror The problem with this trial-and-error learning isthat errors are lost opportunities for reward Given thatthe monkeys performed well above chance when theyencountered novel pictures it is clear that they are notengaging in trial-and-error learning but rather haveabstracted two rules that they can then apply as required

This shortcut that abstraction of the rule permits is alsore ected in the neuronal encoding It would be entirelypossible for the monkey to solve the task without singlecells encoding the rule For example there might be twopopulations of cells one encoding the match and non-match rule when the cues are presented in the auditory

Concepts and the prefrontal cortex E K Miller and others 1131

+ juice

+ low tone

or

+ juice

+ high tone

or

match

non-match

Figure 10 Schematic diagram of the abstract rules taskMonkeys switched between choosing a test stimulus that didor did not match the sample depending on whether thematch or non-match rule was in effect The cues thatsignalled these rules are shown on the left

modality and one encoding this information in the tastemodality But such a solution is computationally expens-ive since if a third modality was introduced a third popu-lation of cells would be required It is more ef cient toabstract a rule that cues presented in different modalitiescommonly instruct and indeed this is the solution thatthe brain uses The prevalence of neurons encoding suchrules in the PFC is consistent with the loss of exibilitythat is observed after prefrontal damage in both monkeysand humans It is not inconsistent with studies emphasiz-ing the role of the PFC in working memory or planningbut indicates that an important component of these pro-cesses might be the use of behaviour-guiding rules

5 PREFRONTAL CORTEX AND COGNITIVECONTROL

The results presented above suggest that PFC activityre ects categories and rules This seems consistent withconjectures that a cardinal PFC function may be theacquisition and representation of the formal demands oftasks the guiding concepts and principles that provide afoundation for complex intelligent behaviour In order tounderstand this we must turn to theories of cognitive con-trolmdashthe ability of the brain to coordinate processingamong its millions of neurons in order to direct themtoward future goals

(a) Controlled versus automatic behavioursIn order to understand what we mean by lsquocognitive con-

trolrsquo it is important to understand the distinction betweencontrolled and automatic behaviours Much of our behav-iour is automatic that is direct reactions to our immediateenvironment that do not tax our attention For exampleif someone suddenly throws a baseball at your face youmight re exively duck You may not have willed thisbehaviour it just seems to happen Many such re exiveautomatic processes are lsquowiredrsquo into our nervous systemsby evolution However others can be acquired through agreat deal of experience as learning mechanisms graduallyimprint highly familiar behaviour If you are walking ahighly familiar route and traf c is light you may traversea great distance (and even negotiate turns) with little

Phil Trans R Soc Lond B (2002)

awareness of having done so In these cases your behav-iour is driven in a lsquobottom-uprsquo fashion largely determinedby stimuli in the immediate environment and their abilityto trigger behaviours with which they are strongly associa-ted In neural terms they are dependent on well-established neural pathways waiting to be red off by thecorrect input

However if something unexpected happens on yourwalk you need to lsquotake chargersquo of your actions You payattention to your surroundings and try to anticipate andaccommodate the action of others you may even decideto take an alternative route In this case your behaviour isnot governed by simple inputndashoutput stimulusndashresponserelationships You use your knowledge of the world thecurrent objective (arriving at work intact) and results fromprevious experiences to weigh the alternatives and conse-quences Because these behaviours tax your attention andseem to be driven by lsquointernalrsquo information (knowledgeabout unseen goals and how to achieve them) and thusare initiated by us and not by the environment they arereferred to as lsquocontrolledrsquo The same basic sensory mem-ory and motor processes that mediated automatic behav-iour can be engaged However now they are not simplytriggered by the environment they are shaped and con-trolled in a top-down fashion by our knowledge of howthe world works The observations that PFC neuronsencode acquired information about task contingenciescategories and rules suggest it as a source of top-down sig-nals

The observation that humans with PFC damage seemimpulsive makes sense in light of this distinction betweencontrolled and automatic behaviours Without the PFC toprovide top-down signals about expectations of goals andrequired behaviours the patient simply reacts to theirenvironment with whatever behaviours are strongly asso-ciated with the cues that are immediately presentShallice amp Burgess (1991) examined this by using a lsquoshop-ping testrsquo They described frontal lobe-damaged patientswho are able to execute simple routines in which clearsensory cues could elicit a familiar action (eg buy a loafof bread) However they were unable to carry out anerrand that involved organizing a series of such routinesbecause they kept going lsquooff taskrsquo They would forexample enter shops that were irrelevant to the errandjust because they happened to be passing them Anotherexample is utilization behaviour A patient with PFC dam-age will impulsively use items placed in front of them suchas a comb or in (hopefully) one case a urinal It seemsthat the basic elements of behaviour are intact but thatthe patients override prepotent re exive responses tocoordinate behaviour in accord with an unseen goal

A classic test of the ability to learn and follow goal-orientated rules is the Wisconsin Card Sorting Test(Milner 1963) Subjects are instructed to sort cardsaccording to the shape colour or number of symbolsappearing on them They start with one rule (eg colour)and once that is acquired the rule changes until all of thecards have been sorted using all possible rules Normalhumans have little dif culty with this task By contrasthumans with prefrontal damage can learn the rst sortingcriterion (a relatively simple mapping between a stimulusattribute and a response) but then are unable to escape itthey cannot override the previous behaviour and do not

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

1124 E K Miller and others Concepts and the prefrontal cortex

squirrelmonkey

cat

dog

rhesusmonkey

chimpanzee

man

pscs

pf

pfprf

as

cs

gprpf

Figure 1 The relative size of the PFC in different animalsAbbreviations as arcuate sulcus cs cingulate sulcusgpr gyrus proreus pf presylvian ssure ps principalsulcus prf proreal ssure From Fuster (1995)

However despite the super cial appearance of normalityPFC damage seems to devastate a personrsquos life They havedif culty in sustaining attention in keeping lsquoon taskrsquo andseem to act on whims and impulses without regard tofuture consequences This pattern of high-level de citscoupled with a sparing of lower-level basic functions hasbeen called a lsquodysexecutive syndromersquo (Baddeley amp DellaSala 1996) and lsquogoal neglectrsquo (Duncan et al 1996)

Indeed the anatomy of the PFC suggests that it is well-suited for a role as the brainrsquos lsquoexecutiversquo It can synthesizeinformation from a wide range of brain systems and exertcontrol over behaviour (Nauta 1971) The collection ofcortical areas that comprise the PFC have intercon-nections with brain areas processing external information(with all sensory systems and with cortical and subcorticalmotor system structures) as well as internal information(limbic and midbrain structures involved in affect mem-ory and reward) (see gure 2) Correspondingly its neu-rons are highly multimodal and encode many differenttypes of information from all stages of the perceptionndashaction cycle (Fuster 1995) They are activated by stimulifrom all sensory modalities before and during a variety ofactions during memory for past events in anticipation ofexpected events and behavioural consequences and aremodulated by internal factors such as motivational andattentional state (see the review in Miller amp Cohen 2001)Because of its highly multimodal nature and its apparentrole in higher mental life the PFC seemed like an idealplace to begin our search for neural correlates of theabstract information needed for intelligent behaviour

Phil Trans R Soc Lond B (2002)

3 THE PREFRONTAL CORTEX AND PERCEPTUALCATEGORIES

Because perceptual categories often group together verydifferent looking things their representation must involvesomething beyond the sort of neural tuning that underliesthe encoding of physical attributes that is gradualchanges in neural activity as certain attributes graduallychange (eg shape orientation direction) Instead categ-ories have sharp boundaries (not gradual transitions) andmembers of the same category are treated as equivalenteven though their physical appearance may vary widely

Consider a simple example of a perceptual categorycrickets sharply divide a certain range of pure tones intolsquomatersquo versus lsquobatrsquo (a predator) (Wyttenbach et al 1996)Even though the input varies along a continuum behav-iour is binary Across a wide range of lower frequenciescrickets will turn towards the sound source because it maybe a potential mate However at a certain point (16 kHz)the behaviour suddenly ips crickets begin to turn awayfrom the sound source because it could be a bat Cricketsmake virtually no distinction between frequencies over awide range on either side of the boundary they approachor avoid with equal reliability This type of representationis illustrated schematically in gure 3 Presumably theability to transform the raw sensory inputs into distinctcategories evolved because it is advantageous in this caseit optimizes reproductive behaviour while minimizing fatalmistakes Similar effects are evident in humansrsquo percep-tion of lsquobrsquo versus lsquoprsquo (Lisker amp Abramson 1970)

The elaborate behavioural repertoire of advanced ani-mals naturally depends on more elaborate categorizationabilities The mental lexicon of primates for exampleincludes abstract categories that are characterized alongmultiple dimensions that are often dif cult to de ne pre-cisely such as lsquotoolrsquo In addition advanced animals havean enormous capacity to learn and adapt Most of ourcategories are acquired through experience (we learn whata lsquochairrsquo is) and we can continually modify and update ourcategories as we learn more about them The ability ofmonkeys to learn complex perceptual categories has beencatalogued in studies that have taught them categoriessuch as animal versus non-animal (Roberts amp Mazmanian1988) food versus non-food (Fabre-Thorpe et al 1998)tree versus non-tree shes versus non- shes (Vogels1999a) and ordinal numbers (Orlov et al 2000) Pigeonsalso have a remarkable ability to learn such distinctions(Bhatt et al 1988 Young amp Wasserman 1997)

Where such categories are encoded in the brain isunclear In primates they could be represented and storedin the same areas of the visual cortex that analyse formand are critical for remembering individual objects suchas the ITC They might also be evident in the brainregions that receive the results of visual processing fromthe ITC and are critical for planning and guiding behav-iour such as the PFC Both the ITC and PFC containneurons selective for complex stimuli such as trees shesfaces brushes etc (Desimone et al 1984 Tanaka et al1991 Vogels 1999b) But whether or not this selectivityre ects category information per se has not been determ-ined With a large amorphous category (eg food humanetc) the category boundaries are unknown Thus charac-teristics diagnostic of category representations (sharp tran-

Concepts and the prefrontal cortex E K Miller and others 1125

mid-dorsal

area 9

dorsolateral

area 46

prefrontal cortex

ventolateral

area 12 45

orbital and medial

areas 10 11

13 14

area 8

(FEF)

motor

structures

basal

ganglia

thalamusmedial temporal

lobe

visual

dorsal

ventral

somatosensory

caudal parietal lobe

auditory

superior temporal

gyrus

multimodal

rostral superior

temporal sulcus

sensory cortex

Figure 2 Schematic diagram of extrinsic and intrinsic connections of the PFC Most connections are reciprocal theexceptions are noted by arrows From Miller amp Cohen (2001)

sitions and little within-category distinction) cannot betested This is not to say that ITC neural selectivity wouldnot make important contributions to category represen-tation it is just not clear whether it represents categorymembership per se

To test for neural correlates of perceptual categories wetrained monkeys to categorize computer-generated stimuliinto two categories lsquocatsrsquo and lsquodogsrsquo (Freedman et al2001 gure 4) A novel three-dimensional morphing sys-tem was used to create a large set of parametric blends ofsix prototype images (three species of cats and threebreeds of dogs) (Beymer amp Poggio 1996 Shelton 2000)By blending different amounts of cat and dog we couldsmoothly vary shape and precisely de ne the boundarybetween the categories (greater than 50 of a given type)As a result stimuli that were close to but on oppositesides of the boundary were similar whereas stimuli thatbelong to the same category could be dissimilar (eg thelsquocheetahrsquo and lsquohouse catrsquo)

Phil Trans R Soc Lond B (2002)

Two monkeys performed a DMC task ( gure 5) thatrequired judgement of whether successively presentedsample and test stimuli were from the same category ornot For training samples were chosen from throughoutthe cat and dog morph space After training classi cationperformance was high (ca 90 correct) even when thesamples were close to the category boundary The monk-eys classi ed dog-like cats (60 cat 40 dog) correctlyca 90 of the time and misclassi ed them as dogs only10 of the time and vice versa Thus the monkeysrsquobehaviour indicated the sharp boundary that is diagnosticof a category representation The dog-like cats weretreated as cats even though they were more similar inappearance to the cat-like dogs just across the categoryboundary than they were to the prototype cats

We recorded in the lateral PFC the PFC region directlyinterconnected with the ITC and found many examplesof neurons that seemed to encode category membershipTwo examples are shown in gure 6 Note that their

1126 E K Miller and others Concepts and the prefrontal cortex

category boundary

16

tone frequency (kHz)

avo

idan

ceap

pro

ach

beh

avio

ur

Figure 3 Schematic representation of categorical perceptionusing cricketsrsquo responses to a continuum of pure tones as anexample Based on Wyttenbach et al (1996)

activity was different from dog-like (60) cats and cat-like(60) dogs yet similar between these stimuli and theirrespective prototypes In other words PFC neuronsseemed to make the same sharp distinctions that were evi-dent in the monkeysrsquo behaviour and are indicative of cat-egorical representations and collected different stimulitogether irrespective of their exact physical experienceCategory information seemed to predominate in the PFCacross the neural population tuning was signi cantlyshifted towards representing category rather than individ-ual stimuli (Freedman et al 2001)

As our monkeys had no experience with cats or dogsprior to training it seemed likely that these effects resultedfrom training To test for learning effects we retrainedone monkey on the DMC task after de ning two new cat-egory boundaries that were orthogonal to the originalboundary ( gure 4) This created three new classes eachcontained morphs centred around one cat prototype andone dog prototype (eg the cheetah and the lsquodobermanrsquo)After training we found that PFC neural activity shiftedto re ect the new but not the old categories An exampleof a single neuron is shown in gure 7 It showed a sig-ni cant effect of category during the delay period whendata were sorted according to the (currently relevant)three-category scheme it distinguished one of the categor-ies from the other two ( gure 7a) By contrast when thedata were sorted using the old category scheme there wasno differentiation between the now-irrelevant cat and dogcategories ( gure 7b)

Our results illustrate that with experience categoryinformation can become incorporated at the single-neuronlevel much as physical attributes of stimuli are This didnot have to be the case in principle categories might havebeen encoded in another fashion For example categoriesmight have been encoded at the ensemble level as anemergent property of neurons that represent their de ningfeatures This ability to carve category membership intothe tuning of single neurons may allow for the quick andeffortless classi cation of familiar items

Our results might re ect a relative specialization of the

Phil Trans R Soc Lond B (2002)

PFC in encoding category membership Categories afterall are typically de ned by their behavioural relevanceand the PFC plays a central part in planning voluntarybehaviours Conversely the traditional roles of the PFCand ITC are in cognitive functions versus object visionand recognition respectively PFC damage causes de citsin attention working memory and response inhibition butusually spares object recognition long-term memory andlsquohigh-levelrsquo visual analysis (Fuster 1989 Miller amp Cohen2001) By contrast ITC damage causes de cits in visualdiscrimination and learning (Gross 1973 Mishkin 1982)and category-speci c agnosias (eg for faces) in humans(Gainotti 2000) It might be that the category informationin the PFC was retrieved from long-term storage in theITC for its immediate use in the task Interactionsbetween the PFC and ITC underlie the storage andorrecall of visual memories and associations (Rainer et al1999 Tomita et al 1999) Tomita et al (1999) demon-strated that top-down signals from the PFC were neededto activate long-term visual memories stored in the ITCA similar relationship may exist for the recall of visual cat-egories In either case it seems that category informationis strongly represented in the PFC a nding consistentwith its role in high-level cognitive functions and in guid-ing behaviour The relative contribution of the ITCremains to be determined

4 THE PREFRONTAL CORTEX AND RULES

It is not only useful to group different sensory stimuliinto meaningful categories it is also useful to group spe-ci c experiences of our interactions with the environmentalong common themes that is as behaviour-guidingprinciples or rules To this end our brains have evolvedmechanisms for detecting and storing often-complexrelationships between situations actions and conse-quences By gleaning this knowledge from past experi-ences we can develop a lsquogame planrsquo that allows us toextrapolate and infer which goals are available in similarsituations in the future and what actions are likely to bringus closer to them

A standard behavioural test of rule learning in monkeysis conditional associative learning (Passingham 1993)This refers to a class of tasks that require learning associat-ive relationships that are arbitrary and extend beyond thesimple one-to-one stimulusndashresponse mappings thatunderlie re exive reactions to the environment In con-ditional learning tasks a given input does not invariablylead to a given output Whether or not a given response issuccessful depends on additional contextual informationFor example reaching for popcorn can be rewarding butonly if one takes other information into account if thepopcorn belongs to another person the result could bedisastrous Taking into account complex relationships inorder to decide between alternative actions is presumablywhy volition evolved

To make predictions about which actions are likely toachieve a given goal in a given situation we need to forma pattern of associations between their internal represen-tations that describes their logical relationship (Dickinson1980) Decades of behavioural research have illustratedthat the brain has learning mechanisms that are exquisitelysensitive to behaviourally informative associations (and

Concepts and the prefrontal cortex E K Miller and others 1127

C2

C1 C3

2-class boundary

D1 D3

D2

3-class boundaries

100 C1 80 C1 60 C1 60 D1 80 D1 100 D1

cats

dogs

(a)

(b)

Figure 4 (a) Monkeys learned to categorize randomly generated lsquomorphsrsquo from the vast number of possible blends of sixprototypes For neurophysiological recording 54 sample stimuli were constructed along the 15 morph lines illustrated here(b) Morphs along the C1ndashD1 line From Freedman et al (2001)

fixation

500 ms sample

600 ms delay

1000 ms

test

(non-match)

600 msdelay

600 ms test

(match)

(match)

Figure 5 The delayed match-to-category task A sample was followed by a delay and a test stimulus If the sample and teststimulus were the same category (a match) monkeys were required to release a lever before the test disappeared If they werenot there was another delay followed by a match Equal numbers of match and non-match trials were randomly interleavedFrom Freedman et al (2001)

Phil Trans R Soc Lond B (2002)

1128 E K Miller and others Concepts and the prefrontal cortex

13

10

7

4

1

fixation sample delay choice

firi

ng

rat

e (H

z)

dog 100dog 80dog 60

cat 100cat 80cat 60

_500 0 500 1000 1500 2000

dog 100dog 80dog 60

cat 100cat 80cat 60

70

60

50

40

30

20

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)

(a)

(b)

Figure 6 The average activity of two single neurons tostimuli at the six morph blends The vertical linescorrespond (from left to right) to sample onset offset andtest stimulus onset Activity is pooled over match and non-match trials

35

30

25

20

15

20

15

10

5left right

20

15

10

5

45

40

35

30left right

spik

es s

_ 1

Phil Trans R Soc Lond B (2002)

fixation sample delay choice10

8

6

4

2

0

10

8

6

4

2

0plusmn500 0 500 1000 1500 2000

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)fi

rin

g r

ate

(Hz)

(a)

(b)

Figure 7 Activity of a single PFC neuron after the monkeywas trained on the three-category scheme (see gure 4)(a) The neuronrsquos activity when the data were sorted basedon the three-category scheme (category A black linecategory B grey line category C dotted line) (b) The sameneuronrsquos activity was data-sorted by the now-irrelevant two-category scheme (black line cats grey line dogs)

Figure 8 The activity of four single prefrontal neurons wheneach of two objects on different trials instructed either asaccade to the right or a saccade to the left The linesconnect the average values obtained when a given objectcued one or the other saccade The error bars show thestandard error of the mean Note that in each case theneuronrsquos activity depends on both the cue object and thesaccade direction and that the tuning is nonlinear orconjunctive That is the level of activity to a givencombination of object and saccade cannot be predicted fromthe neuronrsquos response to the other combinations Adaptedfrom Asaad et al (1998)

insensitive to or even discount associations that are notinformative) The underlying neural ensemble of a goal-directed task then might be comprised of neurons whoseactivity re ects task contingencies Many studies haveshown that prefrontal neurons do have this property Thiswork has focused on the lateral PFC because it seems tobe a site of convergence of the information needed to solveconditional sensorindashmotor tasks It is directly intercon-

Concepts and the prefrontal cortex E K Miller and others 1129

80

60

40

20

0

_2000 _1500 _1000 _500 0 500 1000 1500

40

30

20

10

0

20

15

10

5

0

10

8

6

4

2

0

spatialobjectassociative

ITI fix cue delay

trial time (ms) block number

1 2 3 4 5 6 7 8

spik

es s

_ 1(a)

(c)

(b)

(d)

Figure 9 Spike rate versus time histograms for two neurons each sorted by task The nal second of the three-second ITI isrepresented by the rst 1000 ms ( 2 2000 to 2 1000 ms) Fixation occurs soon after (ca 2 1000 to 2 800 ms) Cuepresentation occurs at the time-point marked 0 ms Task-related differences in baseline ring rate were generally observed tobegin in the xation period While the activity of some neurons diverged almost coincident with initial xation (c) the activityof others diverged progressively as the appearance of the cue became more imminent (a) The bar graphs (bd) demonstratethe reproducibility of these small task-speci c changes in activity across multiple repetitions of the same task The mean xation-period ring rate (with standard errors) for each block of trials is shown for the two neurons in (a) and (c) The barsare colour-coded to re ect the task being performed in each block and the colours match those in the histograms to the leftThe light grey line superimposed over these bars shows the activity of these neurons during the second immediately precedingthe ITI

nected with the higher-order sensory and motor cortexand indirectly connected (via the ventromedial PFC) withlimbic structures that process lsquointernalrsquo information suchas memory and reward (Goldman-Rakic 1987 Pandya ampBarnes 1987 Fuster 1989 Barbas amp Pandya 1991) Theneural activity in the lateral PFC re ects this many of itsneurons exhibit multimodal responses (Vaadia et al 1986Watanabe 1992 Rao et al 1997 Rainer et al 1998abWhite amp Wise 1999) Further the lateral PFC is criticalfor normal learning of conditional associations betweensensory cues and voluntary actions (Petrides 1985a 1990Gaffan amp Harrison 1988 Eacott amp Gaffan 1992 Parker ampGaffan 1998b) Indeed following training on conditionallearning tasks as many as 50 of the neurons in the lat-eral PFC show conjunctive tuning for learned associationsbetween cues voluntary actions and rewards

For example Watanabe (1990 1992) trained monkeysto perform tasks in which visual and auditory cues sig-nalled in different trials whether a reward would orwould not be delivered The majority of lateral PFC neu-rons were found to re ect the association between a cueand a reward A given neuron might be activated by a cuebut only when it signalled a reward By contrast another

Phil Trans R Soc Lond B (2002)

neuron might be activated only by a cue that signalled lsquonorewardrsquo In our own experiments we have trained mon-keys to associate in different blocks of trials each of twocue objects with an eye saccade to the right or to the left(Asaad et al 1998) We found that the activity of 44 ofrandomly selected lateral PFC neurons re ected associ-ations between objects and the saccades that theyinstructed ( gure 8) Other neurons had activity thatre ected the cues or the saccades alone but they werefewer in number Fuster et al (2000) have recently shownthat PFC neurons can also re ect learned associationsbetween visual and auditory stimuli

Importantly these changes do not require a prodigiousamount of experience Changes in PFC neural propertiesare evident after one dayrsquos experience and can be detectedafter just a few minutes of training For example Bichotet al (1996) studied the FEFs part of Brodmannrsquos area 8that is important for voluntary eye movements Normallyneurons in this area re selectively to saccade targetsappearing in certain visual eld locations However whenmonkeys were trained to search for a target de ned by aparticular visual attribute (eg red) the neurons in theFEFs acquire sensitivity to that attribute (Bichot et al

1130 E K Miller and others Concepts and the prefrontal cortex

1996) When monkeys were trained to search for a differ-ent target every day neurons not only discriminated thecurrent target but also distracting stimuli that had been atarget on the previous day relative to stimuli that had beentargets even earlier (Bichot amp Schall 1999) Monkeys werealso more likely to make errors in choosing that distractingstimulus It was as though the previous dayrsquos experienceleft an impression in the brain that in uenced neuralactivity and the monkeyrsquos behaviour

We also observed evidence for rapid plasticity in ourstudy of learning of conditional objectndashsaccade associ-ations in the PFC (Asaad et al 1998) Initially the monk-eys chose their responses at random but learned thecorrect cuendashresponse pairing over a few (5ndash15) trials Asthey learned the association neural activity representingthe forthcoming saccadic response appeared progressivelyearlier in successive trials In other words the initiation ofresponse-related delay activity gradually shifted withlearningmdashfrom a point in time just before the executionof the response and reward delivery to an earlier point intime nearly coincident with the presentation of the cue

Further support for a role for PFC neurons in rep-resenting task demands comes from training monkeys toalternate between different task rules This adds anotherlevel of complexity beyond the conditional tasks describedabove Now there is more than one rule assigned to eachcue and another cue tells the monkey which rule to usein a given trial For example following a given cue monk-eys can learn to direct a response to either the cuersquoslocation (spatial matching rule) or an alternative locationassociated with the cue (associative rule) depending onwhich rule is currently in effect When tested in thisfashion many PFC neurons show rule-speci c activityFor example a PFC neuron might respond to a given vis-ual cue when the monkey is using an associative rule butexhibit weak or no activity under identical sensory andattentional conditions that differed only in that the mon-key was using a spatial rule instead (White amp Wise 1999Asaad et al 2000) Also when monkeys switch betweendifferent tasks many PFC neurons show shifts in baselineactivity that communicates which task is currently beingperformed (Asaad et al 2000 gure 9) Such effects havebeen found in the PFC for associative versus spatial rulesfor object matching versus spatial matching versus associ-ative rules and for shape matching versus object matchingrules (Hoshi et al 1998 White amp Wise 1999 Asaad etal 2000)

In all of these cases however the rules are relativelyliteral or concrete A certain cue or set of cues alwayssignals a speci c response As noted knowledge from ourpast experiences can be applied to a wider range of futurecircumstances if we abstract general principles or rulesrather than speci c cuendashresponse contingencies Theability of the PFC to represent abstract rules those nottied to speci c stimuli or actions was recently addressedin our laboratory (Wallis et al 2001)

Monkeys were trained to use two abstract rules lsquomatchrsquoversus lsquonon-matchrsquo They faced a computer screen andviewed two successively presented pictures If the matchrule was in effect the monkeys released the lever if thepictures were the same and continued to hold the lever ifthe pictures were different If the non-match rule was ineffect the reverse was true monkeys released if the pic-

Phil Trans R Soc Lond B (2002)

tures were different and held if they were the same Therule was randomly instructed in each trial by the presen-tation of a cue at the same time as the rst picture waspresented To disambiguate neuronal responses to thephysical properties of the cue from responses to the rulethat the cues instructed cues signifying the same rule weretaken from different modalities while cues signifying dif-ferent rules were taken from the same modality ( gure10) The monkeys could perform this task well abovechance levels even when they were seeing a stimulus forthe very rst time This indicates that they had abstractedtwo overarching principles of the task that could then beapplied to novel stimulimdashthe minimal de nition of anabstract rule

The most prevalent activity across the PFC was theencoding of the current rule Figure 11 shows a goodexample of a rule-selective neuron that exhibited greateractivity when the match rule was in effect than when thenon-match rule had been indicated This activity cannotbe explained by the physical properties of the cue or thepicture since activity was the same regardless of whichcue was used to instruct the monkey and regardless ofwhich picture the monkey was remembering It cannot berelated to the upcoming response since the monkey didnot know whether the second-presented picture wouldrequire a response Nor could it be related to differencesin reward expectation since the expectation of reward wasthe same regardless of which rule was in effect Further-more the performance of the monkeys was virtually ident-ical for the two types of rules (error rates differed by lessthan 01 and reaction times by less than 7 ms) Thusthe most parsimonious explanation is that the differencesin activity re ected the abstract rule that the monkey wascurrently using to guide its behaviour

What function does the ability to abstract a rule serveIt is a form of generalization that permits a shortcut inlearning thereby allowing the animal to maximize theamount of reward available from a particular situation Toillustrate this consider the above task The monkey couldpotentially solve the task as a series of paired associates(in fact 16 associations consisting of four different pic-tures each paired with four different stimuli) Forexample the monkey might learn that whenever the chefis presented with a drop of juice then at the test phasethe correct response is to choose the chef But notice thatthis type of learning tells the monkey nothing about whichresponse is appropriate to a lion appearing with a drop ofjuice In other words unless the monkey abstracts the rulethat juice indicates that the monkey should match theneach time new pictures are used the monkey would haveto learn an entirely new set of 16 associations by trial anderror The problem with this trial-and-error learning isthat errors are lost opportunities for reward Given thatthe monkeys performed well above chance when theyencountered novel pictures it is clear that they are notengaging in trial-and-error learning but rather haveabstracted two rules that they can then apply as required

This shortcut that abstraction of the rule permits is alsore ected in the neuronal encoding It would be entirelypossible for the monkey to solve the task without singlecells encoding the rule For example there might be twopopulations of cells one encoding the match and non-match rule when the cues are presented in the auditory

Concepts and the prefrontal cortex E K Miller and others 1131

+ juice

+ low tone

or

+ juice

+ high tone

or

match

non-match

Figure 10 Schematic diagram of the abstract rules taskMonkeys switched between choosing a test stimulus that didor did not match the sample depending on whether thematch or non-match rule was in effect The cues thatsignalled these rules are shown on the left

modality and one encoding this information in the tastemodality But such a solution is computationally expens-ive since if a third modality was introduced a third popu-lation of cells would be required It is more ef cient toabstract a rule that cues presented in different modalitiescommonly instruct and indeed this is the solution thatthe brain uses The prevalence of neurons encoding suchrules in the PFC is consistent with the loss of exibilitythat is observed after prefrontal damage in both monkeysand humans It is not inconsistent with studies emphasiz-ing the role of the PFC in working memory or planningbut indicates that an important component of these pro-cesses might be the use of behaviour-guiding rules

5 PREFRONTAL CORTEX AND COGNITIVECONTROL

The results presented above suggest that PFC activityre ects categories and rules This seems consistent withconjectures that a cardinal PFC function may be theacquisition and representation of the formal demands oftasks the guiding concepts and principles that provide afoundation for complex intelligent behaviour In order tounderstand this we must turn to theories of cognitive con-trolmdashthe ability of the brain to coordinate processingamong its millions of neurons in order to direct themtoward future goals

(a) Controlled versus automatic behavioursIn order to understand what we mean by lsquocognitive con-

trolrsquo it is important to understand the distinction betweencontrolled and automatic behaviours Much of our behav-iour is automatic that is direct reactions to our immediateenvironment that do not tax our attention For exampleif someone suddenly throws a baseball at your face youmight re exively duck You may not have willed thisbehaviour it just seems to happen Many such re exiveautomatic processes are lsquowiredrsquo into our nervous systemsby evolution However others can be acquired through agreat deal of experience as learning mechanisms graduallyimprint highly familiar behaviour If you are walking ahighly familiar route and traf c is light you may traversea great distance (and even negotiate turns) with little

Phil Trans R Soc Lond B (2002)

awareness of having done so In these cases your behav-iour is driven in a lsquobottom-uprsquo fashion largely determinedby stimuli in the immediate environment and their abilityto trigger behaviours with which they are strongly associa-ted In neural terms they are dependent on well-established neural pathways waiting to be red off by thecorrect input

However if something unexpected happens on yourwalk you need to lsquotake chargersquo of your actions You payattention to your surroundings and try to anticipate andaccommodate the action of others you may even decideto take an alternative route In this case your behaviour isnot governed by simple inputndashoutput stimulusndashresponserelationships You use your knowledge of the world thecurrent objective (arriving at work intact) and results fromprevious experiences to weigh the alternatives and conse-quences Because these behaviours tax your attention andseem to be driven by lsquointernalrsquo information (knowledgeabout unseen goals and how to achieve them) and thusare initiated by us and not by the environment they arereferred to as lsquocontrolledrsquo The same basic sensory mem-ory and motor processes that mediated automatic behav-iour can be engaged However now they are not simplytriggered by the environment they are shaped and con-trolled in a top-down fashion by our knowledge of howthe world works The observations that PFC neuronsencode acquired information about task contingenciescategories and rules suggest it as a source of top-down sig-nals

The observation that humans with PFC damage seemimpulsive makes sense in light of this distinction betweencontrolled and automatic behaviours Without the PFC toprovide top-down signals about expectations of goals andrequired behaviours the patient simply reacts to theirenvironment with whatever behaviours are strongly asso-ciated with the cues that are immediately presentShallice amp Burgess (1991) examined this by using a lsquoshop-ping testrsquo They described frontal lobe-damaged patientswho are able to execute simple routines in which clearsensory cues could elicit a familiar action (eg buy a loafof bread) However they were unable to carry out anerrand that involved organizing a series of such routinesbecause they kept going lsquooff taskrsquo They would forexample enter shops that were irrelevant to the errandjust because they happened to be passing them Anotherexample is utilization behaviour A patient with PFC dam-age will impulsively use items placed in front of them suchas a comb or in (hopefully) one case a urinal It seemsthat the basic elements of behaviour are intact but thatthe patients override prepotent re exive responses tocoordinate behaviour in accord with an unseen goal

A classic test of the ability to learn and follow goal-orientated rules is the Wisconsin Card Sorting Test(Milner 1963) Subjects are instructed to sort cardsaccording to the shape colour or number of symbolsappearing on them They start with one rule (eg colour)and once that is acquired the rule changes until all of thecards have been sorted using all possible rules Normalhumans have little dif culty with this task By contrasthumans with prefrontal damage can learn the rst sortingcriterion (a relatively simple mapping between a stimulusattribute and a response) but then are unable to escape itthey cannot override the previous behaviour and do not

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

Concepts and the prefrontal cortex E K Miller and others 1125

mid-dorsal

area 9

dorsolateral

area 46

prefrontal cortex

ventolateral

area 12 45

orbital and medial

areas 10 11

13 14

area 8

(FEF)

motor

structures

basal

ganglia

thalamusmedial temporal

lobe

visual

dorsal

ventral

somatosensory

caudal parietal lobe

auditory

superior temporal

gyrus

multimodal

rostral superior

temporal sulcus

sensory cortex

Figure 2 Schematic diagram of extrinsic and intrinsic connections of the PFC Most connections are reciprocal theexceptions are noted by arrows From Miller amp Cohen (2001)

sitions and little within-category distinction) cannot betested This is not to say that ITC neural selectivity wouldnot make important contributions to category represen-tation it is just not clear whether it represents categorymembership per se

To test for neural correlates of perceptual categories wetrained monkeys to categorize computer-generated stimuliinto two categories lsquocatsrsquo and lsquodogsrsquo (Freedman et al2001 gure 4) A novel three-dimensional morphing sys-tem was used to create a large set of parametric blends ofsix prototype images (three species of cats and threebreeds of dogs) (Beymer amp Poggio 1996 Shelton 2000)By blending different amounts of cat and dog we couldsmoothly vary shape and precisely de ne the boundarybetween the categories (greater than 50 of a given type)As a result stimuli that were close to but on oppositesides of the boundary were similar whereas stimuli thatbelong to the same category could be dissimilar (eg thelsquocheetahrsquo and lsquohouse catrsquo)

Phil Trans R Soc Lond B (2002)

Two monkeys performed a DMC task ( gure 5) thatrequired judgement of whether successively presentedsample and test stimuli were from the same category ornot For training samples were chosen from throughoutthe cat and dog morph space After training classi cationperformance was high (ca 90 correct) even when thesamples were close to the category boundary The monk-eys classi ed dog-like cats (60 cat 40 dog) correctlyca 90 of the time and misclassi ed them as dogs only10 of the time and vice versa Thus the monkeysrsquobehaviour indicated the sharp boundary that is diagnosticof a category representation The dog-like cats weretreated as cats even though they were more similar inappearance to the cat-like dogs just across the categoryboundary than they were to the prototype cats

We recorded in the lateral PFC the PFC region directlyinterconnected with the ITC and found many examplesof neurons that seemed to encode category membershipTwo examples are shown in gure 6 Note that their

1126 E K Miller and others Concepts and the prefrontal cortex

category boundary

16

tone frequency (kHz)

avo

idan

ceap

pro

ach

beh

avio

ur

Figure 3 Schematic representation of categorical perceptionusing cricketsrsquo responses to a continuum of pure tones as anexample Based on Wyttenbach et al (1996)

activity was different from dog-like (60) cats and cat-like(60) dogs yet similar between these stimuli and theirrespective prototypes In other words PFC neuronsseemed to make the same sharp distinctions that were evi-dent in the monkeysrsquo behaviour and are indicative of cat-egorical representations and collected different stimulitogether irrespective of their exact physical experienceCategory information seemed to predominate in the PFCacross the neural population tuning was signi cantlyshifted towards representing category rather than individ-ual stimuli (Freedman et al 2001)

As our monkeys had no experience with cats or dogsprior to training it seemed likely that these effects resultedfrom training To test for learning effects we retrainedone monkey on the DMC task after de ning two new cat-egory boundaries that were orthogonal to the originalboundary ( gure 4) This created three new classes eachcontained morphs centred around one cat prototype andone dog prototype (eg the cheetah and the lsquodobermanrsquo)After training we found that PFC neural activity shiftedto re ect the new but not the old categories An exampleof a single neuron is shown in gure 7 It showed a sig-ni cant effect of category during the delay period whendata were sorted according to the (currently relevant)three-category scheme it distinguished one of the categor-ies from the other two ( gure 7a) By contrast when thedata were sorted using the old category scheme there wasno differentiation between the now-irrelevant cat and dogcategories ( gure 7b)

Our results illustrate that with experience categoryinformation can become incorporated at the single-neuronlevel much as physical attributes of stimuli are This didnot have to be the case in principle categories might havebeen encoded in another fashion For example categoriesmight have been encoded at the ensemble level as anemergent property of neurons that represent their de ningfeatures This ability to carve category membership intothe tuning of single neurons may allow for the quick andeffortless classi cation of familiar items

Our results might re ect a relative specialization of the

Phil Trans R Soc Lond B (2002)

PFC in encoding category membership Categories afterall are typically de ned by their behavioural relevanceand the PFC plays a central part in planning voluntarybehaviours Conversely the traditional roles of the PFCand ITC are in cognitive functions versus object visionand recognition respectively PFC damage causes de citsin attention working memory and response inhibition butusually spares object recognition long-term memory andlsquohigh-levelrsquo visual analysis (Fuster 1989 Miller amp Cohen2001) By contrast ITC damage causes de cits in visualdiscrimination and learning (Gross 1973 Mishkin 1982)and category-speci c agnosias (eg for faces) in humans(Gainotti 2000) It might be that the category informationin the PFC was retrieved from long-term storage in theITC for its immediate use in the task Interactionsbetween the PFC and ITC underlie the storage andorrecall of visual memories and associations (Rainer et al1999 Tomita et al 1999) Tomita et al (1999) demon-strated that top-down signals from the PFC were neededto activate long-term visual memories stored in the ITCA similar relationship may exist for the recall of visual cat-egories In either case it seems that category informationis strongly represented in the PFC a nding consistentwith its role in high-level cognitive functions and in guid-ing behaviour The relative contribution of the ITCremains to be determined

4 THE PREFRONTAL CORTEX AND RULES

It is not only useful to group different sensory stimuliinto meaningful categories it is also useful to group spe-ci c experiences of our interactions with the environmentalong common themes that is as behaviour-guidingprinciples or rules To this end our brains have evolvedmechanisms for detecting and storing often-complexrelationships between situations actions and conse-quences By gleaning this knowledge from past experi-ences we can develop a lsquogame planrsquo that allows us toextrapolate and infer which goals are available in similarsituations in the future and what actions are likely to bringus closer to them

A standard behavioural test of rule learning in monkeysis conditional associative learning (Passingham 1993)This refers to a class of tasks that require learning associat-ive relationships that are arbitrary and extend beyond thesimple one-to-one stimulusndashresponse mappings thatunderlie re exive reactions to the environment In con-ditional learning tasks a given input does not invariablylead to a given output Whether or not a given response issuccessful depends on additional contextual informationFor example reaching for popcorn can be rewarding butonly if one takes other information into account if thepopcorn belongs to another person the result could bedisastrous Taking into account complex relationships inorder to decide between alternative actions is presumablywhy volition evolved

To make predictions about which actions are likely toachieve a given goal in a given situation we need to forma pattern of associations between their internal represen-tations that describes their logical relationship (Dickinson1980) Decades of behavioural research have illustratedthat the brain has learning mechanisms that are exquisitelysensitive to behaviourally informative associations (and

Concepts and the prefrontal cortex E K Miller and others 1127

C2

C1 C3

2-class boundary

D1 D3

D2

3-class boundaries

100 C1 80 C1 60 C1 60 D1 80 D1 100 D1

cats

dogs

(a)

(b)

Figure 4 (a) Monkeys learned to categorize randomly generated lsquomorphsrsquo from the vast number of possible blends of sixprototypes For neurophysiological recording 54 sample stimuli were constructed along the 15 morph lines illustrated here(b) Morphs along the C1ndashD1 line From Freedman et al (2001)

fixation

500 ms sample

600 ms delay

1000 ms

test

(non-match)

600 msdelay

600 ms test

(match)

(match)

Figure 5 The delayed match-to-category task A sample was followed by a delay and a test stimulus If the sample and teststimulus were the same category (a match) monkeys were required to release a lever before the test disappeared If they werenot there was another delay followed by a match Equal numbers of match and non-match trials were randomly interleavedFrom Freedman et al (2001)

Phil Trans R Soc Lond B (2002)

1128 E K Miller and others Concepts and the prefrontal cortex

13

10

7

4

1

fixation sample delay choice

firi

ng

rat

e (H

z)

dog 100dog 80dog 60

cat 100cat 80cat 60

_500 0 500 1000 1500 2000

dog 100dog 80dog 60

cat 100cat 80cat 60

70

60

50

40

30

20

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)

(a)

(b)

Figure 6 The average activity of two single neurons tostimuli at the six morph blends The vertical linescorrespond (from left to right) to sample onset offset andtest stimulus onset Activity is pooled over match and non-match trials

35

30

25

20

15

20

15

10

5left right

20

15

10

5

45

40

35

30left right

spik

es s

_ 1

Phil Trans R Soc Lond B (2002)

fixation sample delay choice10

8

6

4

2

0

10

8

6

4

2

0plusmn500 0 500 1000 1500 2000

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)fi

rin

g r

ate

(Hz)

(a)

(b)

Figure 7 Activity of a single PFC neuron after the monkeywas trained on the three-category scheme (see gure 4)(a) The neuronrsquos activity when the data were sorted basedon the three-category scheme (category A black linecategory B grey line category C dotted line) (b) The sameneuronrsquos activity was data-sorted by the now-irrelevant two-category scheme (black line cats grey line dogs)

Figure 8 The activity of four single prefrontal neurons wheneach of two objects on different trials instructed either asaccade to the right or a saccade to the left The linesconnect the average values obtained when a given objectcued one or the other saccade The error bars show thestandard error of the mean Note that in each case theneuronrsquos activity depends on both the cue object and thesaccade direction and that the tuning is nonlinear orconjunctive That is the level of activity to a givencombination of object and saccade cannot be predicted fromthe neuronrsquos response to the other combinations Adaptedfrom Asaad et al (1998)

insensitive to or even discount associations that are notinformative) The underlying neural ensemble of a goal-directed task then might be comprised of neurons whoseactivity re ects task contingencies Many studies haveshown that prefrontal neurons do have this property Thiswork has focused on the lateral PFC because it seems tobe a site of convergence of the information needed to solveconditional sensorindashmotor tasks It is directly intercon-

Concepts and the prefrontal cortex E K Miller and others 1129

80

60

40

20

0

_2000 _1500 _1000 _500 0 500 1000 1500

40

30

20

10

0

20

15

10

5

0

10

8

6

4

2

0

spatialobjectassociative

ITI fix cue delay

trial time (ms) block number

1 2 3 4 5 6 7 8

spik

es s

_ 1(a)

(c)

(b)

(d)

Figure 9 Spike rate versus time histograms for two neurons each sorted by task The nal second of the three-second ITI isrepresented by the rst 1000 ms ( 2 2000 to 2 1000 ms) Fixation occurs soon after (ca 2 1000 to 2 800 ms) Cuepresentation occurs at the time-point marked 0 ms Task-related differences in baseline ring rate were generally observed tobegin in the xation period While the activity of some neurons diverged almost coincident with initial xation (c) the activityof others diverged progressively as the appearance of the cue became more imminent (a) The bar graphs (bd) demonstratethe reproducibility of these small task-speci c changes in activity across multiple repetitions of the same task The mean xation-period ring rate (with standard errors) for each block of trials is shown for the two neurons in (a) and (c) The barsare colour-coded to re ect the task being performed in each block and the colours match those in the histograms to the leftThe light grey line superimposed over these bars shows the activity of these neurons during the second immediately precedingthe ITI

nected with the higher-order sensory and motor cortexand indirectly connected (via the ventromedial PFC) withlimbic structures that process lsquointernalrsquo information suchas memory and reward (Goldman-Rakic 1987 Pandya ampBarnes 1987 Fuster 1989 Barbas amp Pandya 1991) Theneural activity in the lateral PFC re ects this many of itsneurons exhibit multimodal responses (Vaadia et al 1986Watanabe 1992 Rao et al 1997 Rainer et al 1998abWhite amp Wise 1999) Further the lateral PFC is criticalfor normal learning of conditional associations betweensensory cues and voluntary actions (Petrides 1985a 1990Gaffan amp Harrison 1988 Eacott amp Gaffan 1992 Parker ampGaffan 1998b) Indeed following training on conditionallearning tasks as many as 50 of the neurons in the lat-eral PFC show conjunctive tuning for learned associationsbetween cues voluntary actions and rewards

For example Watanabe (1990 1992) trained monkeysto perform tasks in which visual and auditory cues sig-nalled in different trials whether a reward would orwould not be delivered The majority of lateral PFC neu-rons were found to re ect the association between a cueand a reward A given neuron might be activated by a cuebut only when it signalled a reward By contrast another

Phil Trans R Soc Lond B (2002)

neuron might be activated only by a cue that signalled lsquonorewardrsquo In our own experiments we have trained mon-keys to associate in different blocks of trials each of twocue objects with an eye saccade to the right or to the left(Asaad et al 1998) We found that the activity of 44 ofrandomly selected lateral PFC neurons re ected associ-ations between objects and the saccades that theyinstructed ( gure 8) Other neurons had activity thatre ected the cues or the saccades alone but they werefewer in number Fuster et al (2000) have recently shownthat PFC neurons can also re ect learned associationsbetween visual and auditory stimuli

Importantly these changes do not require a prodigiousamount of experience Changes in PFC neural propertiesare evident after one dayrsquos experience and can be detectedafter just a few minutes of training For example Bichotet al (1996) studied the FEFs part of Brodmannrsquos area 8that is important for voluntary eye movements Normallyneurons in this area re selectively to saccade targetsappearing in certain visual eld locations However whenmonkeys were trained to search for a target de ned by aparticular visual attribute (eg red) the neurons in theFEFs acquire sensitivity to that attribute (Bichot et al

1130 E K Miller and others Concepts and the prefrontal cortex

1996) When monkeys were trained to search for a differ-ent target every day neurons not only discriminated thecurrent target but also distracting stimuli that had been atarget on the previous day relative to stimuli that had beentargets even earlier (Bichot amp Schall 1999) Monkeys werealso more likely to make errors in choosing that distractingstimulus It was as though the previous dayrsquos experienceleft an impression in the brain that in uenced neuralactivity and the monkeyrsquos behaviour

We also observed evidence for rapid plasticity in ourstudy of learning of conditional objectndashsaccade associ-ations in the PFC (Asaad et al 1998) Initially the monk-eys chose their responses at random but learned thecorrect cuendashresponse pairing over a few (5ndash15) trials Asthey learned the association neural activity representingthe forthcoming saccadic response appeared progressivelyearlier in successive trials In other words the initiation ofresponse-related delay activity gradually shifted withlearningmdashfrom a point in time just before the executionof the response and reward delivery to an earlier point intime nearly coincident with the presentation of the cue

Further support for a role for PFC neurons in rep-resenting task demands comes from training monkeys toalternate between different task rules This adds anotherlevel of complexity beyond the conditional tasks describedabove Now there is more than one rule assigned to eachcue and another cue tells the monkey which rule to usein a given trial For example following a given cue monk-eys can learn to direct a response to either the cuersquoslocation (spatial matching rule) or an alternative locationassociated with the cue (associative rule) depending onwhich rule is currently in effect When tested in thisfashion many PFC neurons show rule-speci c activityFor example a PFC neuron might respond to a given vis-ual cue when the monkey is using an associative rule butexhibit weak or no activity under identical sensory andattentional conditions that differed only in that the mon-key was using a spatial rule instead (White amp Wise 1999Asaad et al 2000) Also when monkeys switch betweendifferent tasks many PFC neurons show shifts in baselineactivity that communicates which task is currently beingperformed (Asaad et al 2000 gure 9) Such effects havebeen found in the PFC for associative versus spatial rulesfor object matching versus spatial matching versus associ-ative rules and for shape matching versus object matchingrules (Hoshi et al 1998 White amp Wise 1999 Asaad etal 2000)

In all of these cases however the rules are relativelyliteral or concrete A certain cue or set of cues alwayssignals a speci c response As noted knowledge from ourpast experiences can be applied to a wider range of futurecircumstances if we abstract general principles or rulesrather than speci c cuendashresponse contingencies Theability of the PFC to represent abstract rules those nottied to speci c stimuli or actions was recently addressedin our laboratory (Wallis et al 2001)

Monkeys were trained to use two abstract rules lsquomatchrsquoversus lsquonon-matchrsquo They faced a computer screen andviewed two successively presented pictures If the matchrule was in effect the monkeys released the lever if thepictures were the same and continued to hold the lever ifthe pictures were different If the non-match rule was ineffect the reverse was true monkeys released if the pic-

Phil Trans R Soc Lond B (2002)

tures were different and held if they were the same Therule was randomly instructed in each trial by the presen-tation of a cue at the same time as the rst picture waspresented To disambiguate neuronal responses to thephysical properties of the cue from responses to the rulethat the cues instructed cues signifying the same rule weretaken from different modalities while cues signifying dif-ferent rules were taken from the same modality ( gure10) The monkeys could perform this task well abovechance levels even when they were seeing a stimulus forthe very rst time This indicates that they had abstractedtwo overarching principles of the task that could then beapplied to novel stimulimdashthe minimal de nition of anabstract rule

The most prevalent activity across the PFC was theencoding of the current rule Figure 11 shows a goodexample of a rule-selective neuron that exhibited greateractivity when the match rule was in effect than when thenon-match rule had been indicated This activity cannotbe explained by the physical properties of the cue or thepicture since activity was the same regardless of whichcue was used to instruct the monkey and regardless ofwhich picture the monkey was remembering It cannot berelated to the upcoming response since the monkey didnot know whether the second-presented picture wouldrequire a response Nor could it be related to differencesin reward expectation since the expectation of reward wasthe same regardless of which rule was in effect Further-more the performance of the monkeys was virtually ident-ical for the two types of rules (error rates differed by lessthan 01 and reaction times by less than 7 ms) Thusthe most parsimonious explanation is that the differencesin activity re ected the abstract rule that the monkey wascurrently using to guide its behaviour

What function does the ability to abstract a rule serveIt is a form of generalization that permits a shortcut inlearning thereby allowing the animal to maximize theamount of reward available from a particular situation Toillustrate this consider the above task The monkey couldpotentially solve the task as a series of paired associates(in fact 16 associations consisting of four different pic-tures each paired with four different stimuli) Forexample the monkey might learn that whenever the chefis presented with a drop of juice then at the test phasethe correct response is to choose the chef But notice thatthis type of learning tells the monkey nothing about whichresponse is appropriate to a lion appearing with a drop ofjuice In other words unless the monkey abstracts the rulethat juice indicates that the monkey should match theneach time new pictures are used the monkey would haveto learn an entirely new set of 16 associations by trial anderror The problem with this trial-and-error learning isthat errors are lost opportunities for reward Given thatthe monkeys performed well above chance when theyencountered novel pictures it is clear that they are notengaging in trial-and-error learning but rather haveabstracted two rules that they can then apply as required

This shortcut that abstraction of the rule permits is alsore ected in the neuronal encoding It would be entirelypossible for the monkey to solve the task without singlecells encoding the rule For example there might be twopopulations of cells one encoding the match and non-match rule when the cues are presented in the auditory

Concepts and the prefrontal cortex E K Miller and others 1131

+ juice

+ low tone

or

+ juice

+ high tone

or

match

non-match

Figure 10 Schematic diagram of the abstract rules taskMonkeys switched between choosing a test stimulus that didor did not match the sample depending on whether thematch or non-match rule was in effect The cues thatsignalled these rules are shown on the left

modality and one encoding this information in the tastemodality But such a solution is computationally expens-ive since if a third modality was introduced a third popu-lation of cells would be required It is more ef cient toabstract a rule that cues presented in different modalitiescommonly instruct and indeed this is the solution thatthe brain uses The prevalence of neurons encoding suchrules in the PFC is consistent with the loss of exibilitythat is observed after prefrontal damage in both monkeysand humans It is not inconsistent with studies emphasiz-ing the role of the PFC in working memory or planningbut indicates that an important component of these pro-cesses might be the use of behaviour-guiding rules

5 PREFRONTAL CORTEX AND COGNITIVECONTROL

The results presented above suggest that PFC activityre ects categories and rules This seems consistent withconjectures that a cardinal PFC function may be theacquisition and representation of the formal demands oftasks the guiding concepts and principles that provide afoundation for complex intelligent behaviour In order tounderstand this we must turn to theories of cognitive con-trolmdashthe ability of the brain to coordinate processingamong its millions of neurons in order to direct themtoward future goals

(a) Controlled versus automatic behavioursIn order to understand what we mean by lsquocognitive con-

trolrsquo it is important to understand the distinction betweencontrolled and automatic behaviours Much of our behav-iour is automatic that is direct reactions to our immediateenvironment that do not tax our attention For exampleif someone suddenly throws a baseball at your face youmight re exively duck You may not have willed thisbehaviour it just seems to happen Many such re exiveautomatic processes are lsquowiredrsquo into our nervous systemsby evolution However others can be acquired through agreat deal of experience as learning mechanisms graduallyimprint highly familiar behaviour If you are walking ahighly familiar route and traf c is light you may traversea great distance (and even negotiate turns) with little

Phil Trans R Soc Lond B (2002)

awareness of having done so In these cases your behav-iour is driven in a lsquobottom-uprsquo fashion largely determinedby stimuli in the immediate environment and their abilityto trigger behaviours with which they are strongly associa-ted In neural terms they are dependent on well-established neural pathways waiting to be red off by thecorrect input

However if something unexpected happens on yourwalk you need to lsquotake chargersquo of your actions You payattention to your surroundings and try to anticipate andaccommodate the action of others you may even decideto take an alternative route In this case your behaviour isnot governed by simple inputndashoutput stimulusndashresponserelationships You use your knowledge of the world thecurrent objective (arriving at work intact) and results fromprevious experiences to weigh the alternatives and conse-quences Because these behaviours tax your attention andseem to be driven by lsquointernalrsquo information (knowledgeabout unseen goals and how to achieve them) and thusare initiated by us and not by the environment they arereferred to as lsquocontrolledrsquo The same basic sensory mem-ory and motor processes that mediated automatic behav-iour can be engaged However now they are not simplytriggered by the environment they are shaped and con-trolled in a top-down fashion by our knowledge of howthe world works The observations that PFC neuronsencode acquired information about task contingenciescategories and rules suggest it as a source of top-down sig-nals

The observation that humans with PFC damage seemimpulsive makes sense in light of this distinction betweencontrolled and automatic behaviours Without the PFC toprovide top-down signals about expectations of goals andrequired behaviours the patient simply reacts to theirenvironment with whatever behaviours are strongly asso-ciated with the cues that are immediately presentShallice amp Burgess (1991) examined this by using a lsquoshop-ping testrsquo They described frontal lobe-damaged patientswho are able to execute simple routines in which clearsensory cues could elicit a familiar action (eg buy a loafof bread) However they were unable to carry out anerrand that involved organizing a series of such routinesbecause they kept going lsquooff taskrsquo They would forexample enter shops that were irrelevant to the errandjust because they happened to be passing them Anotherexample is utilization behaviour A patient with PFC dam-age will impulsively use items placed in front of them suchas a comb or in (hopefully) one case a urinal It seemsthat the basic elements of behaviour are intact but thatthe patients override prepotent re exive responses tocoordinate behaviour in accord with an unseen goal

A classic test of the ability to learn and follow goal-orientated rules is the Wisconsin Card Sorting Test(Milner 1963) Subjects are instructed to sort cardsaccording to the shape colour or number of symbolsappearing on them They start with one rule (eg colour)and once that is acquired the rule changes until all of thecards have been sorted using all possible rules Normalhumans have little dif culty with this task By contrasthumans with prefrontal damage can learn the rst sortingcriterion (a relatively simple mapping between a stimulusattribute and a response) but then are unable to escape itthey cannot override the previous behaviour and do not

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

1126 E K Miller and others Concepts and the prefrontal cortex

category boundary

16

tone frequency (kHz)

avo

idan

ceap

pro

ach

beh

avio

ur

Figure 3 Schematic representation of categorical perceptionusing cricketsrsquo responses to a continuum of pure tones as anexample Based on Wyttenbach et al (1996)

activity was different from dog-like (60) cats and cat-like(60) dogs yet similar between these stimuli and theirrespective prototypes In other words PFC neuronsseemed to make the same sharp distinctions that were evi-dent in the monkeysrsquo behaviour and are indicative of cat-egorical representations and collected different stimulitogether irrespective of their exact physical experienceCategory information seemed to predominate in the PFCacross the neural population tuning was signi cantlyshifted towards representing category rather than individ-ual stimuli (Freedman et al 2001)

As our monkeys had no experience with cats or dogsprior to training it seemed likely that these effects resultedfrom training To test for learning effects we retrainedone monkey on the DMC task after de ning two new cat-egory boundaries that were orthogonal to the originalboundary ( gure 4) This created three new classes eachcontained morphs centred around one cat prototype andone dog prototype (eg the cheetah and the lsquodobermanrsquo)After training we found that PFC neural activity shiftedto re ect the new but not the old categories An exampleof a single neuron is shown in gure 7 It showed a sig-ni cant effect of category during the delay period whendata were sorted according to the (currently relevant)three-category scheme it distinguished one of the categor-ies from the other two ( gure 7a) By contrast when thedata were sorted using the old category scheme there wasno differentiation between the now-irrelevant cat and dogcategories ( gure 7b)

Our results illustrate that with experience categoryinformation can become incorporated at the single-neuronlevel much as physical attributes of stimuli are This didnot have to be the case in principle categories might havebeen encoded in another fashion For example categoriesmight have been encoded at the ensemble level as anemergent property of neurons that represent their de ningfeatures This ability to carve category membership intothe tuning of single neurons may allow for the quick andeffortless classi cation of familiar items

Our results might re ect a relative specialization of the

Phil Trans R Soc Lond B (2002)

PFC in encoding category membership Categories afterall are typically de ned by their behavioural relevanceand the PFC plays a central part in planning voluntarybehaviours Conversely the traditional roles of the PFCand ITC are in cognitive functions versus object visionand recognition respectively PFC damage causes de citsin attention working memory and response inhibition butusually spares object recognition long-term memory andlsquohigh-levelrsquo visual analysis (Fuster 1989 Miller amp Cohen2001) By contrast ITC damage causes de cits in visualdiscrimination and learning (Gross 1973 Mishkin 1982)and category-speci c agnosias (eg for faces) in humans(Gainotti 2000) It might be that the category informationin the PFC was retrieved from long-term storage in theITC for its immediate use in the task Interactionsbetween the PFC and ITC underlie the storage andorrecall of visual memories and associations (Rainer et al1999 Tomita et al 1999) Tomita et al (1999) demon-strated that top-down signals from the PFC were neededto activate long-term visual memories stored in the ITCA similar relationship may exist for the recall of visual cat-egories In either case it seems that category informationis strongly represented in the PFC a nding consistentwith its role in high-level cognitive functions and in guid-ing behaviour The relative contribution of the ITCremains to be determined

4 THE PREFRONTAL CORTEX AND RULES

It is not only useful to group different sensory stimuliinto meaningful categories it is also useful to group spe-ci c experiences of our interactions with the environmentalong common themes that is as behaviour-guidingprinciples or rules To this end our brains have evolvedmechanisms for detecting and storing often-complexrelationships between situations actions and conse-quences By gleaning this knowledge from past experi-ences we can develop a lsquogame planrsquo that allows us toextrapolate and infer which goals are available in similarsituations in the future and what actions are likely to bringus closer to them

A standard behavioural test of rule learning in monkeysis conditional associative learning (Passingham 1993)This refers to a class of tasks that require learning associat-ive relationships that are arbitrary and extend beyond thesimple one-to-one stimulusndashresponse mappings thatunderlie re exive reactions to the environment In con-ditional learning tasks a given input does not invariablylead to a given output Whether or not a given response issuccessful depends on additional contextual informationFor example reaching for popcorn can be rewarding butonly if one takes other information into account if thepopcorn belongs to another person the result could bedisastrous Taking into account complex relationships inorder to decide between alternative actions is presumablywhy volition evolved

To make predictions about which actions are likely toachieve a given goal in a given situation we need to forma pattern of associations between their internal represen-tations that describes their logical relationship (Dickinson1980) Decades of behavioural research have illustratedthat the brain has learning mechanisms that are exquisitelysensitive to behaviourally informative associations (and

Concepts and the prefrontal cortex E K Miller and others 1127

C2

C1 C3

2-class boundary

D1 D3

D2

3-class boundaries

100 C1 80 C1 60 C1 60 D1 80 D1 100 D1

cats

dogs

(a)

(b)

Figure 4 (a) Monkeys learned to categorize randomly generated lsquomorphsrsquo from the vast number of possible blends of sixprototypes For neurophysiological recording 54 sample stimuli were constructed along the 15 morph lines illustrated here(b) Morphs along the C1ndashD1 line From Freedman et al (2001)

fixation

500 ms sample

600 ms delay

1000 ms

test

(non-match)

600 msdelay

600 ms test

(match)

(match)

Figure 5 The delayed match-to-category task A sample was followed by a delay and a test stimulus If the sample and teststimulus were the same category (a match) monkeys were required to release a lever before the test disappeared If they werenot there was another delay followed by a match Equal numbers of match and non-match trials were randomly interleavedFrom Freedman et al (2001)

Phil Trans R Soc Lond B (2002)

1128 E K Miller and others Concepts and the prefrontal cortex

13

10

7

4

1

fixation sample delay choice

firi

ng

rat

e (H

z)

dog 100dog 80dog 60

cat 100cat 80cat 60

_500 0 500 1000 1500 2000

dog 100dog 80dog 60

cat 100cat 80cat 60

70

60

50

40

30

20

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)

(a)

(b)

Figure 6 The average activity of two single neurons tostimuli at the six morph blends The vertical linescorrespond (from left to right) to sample onset offset andtest stimulus onset Activity is pooled over match and non-match trials

35

30

25

20

15

20

15

10

5left right

20

15

10

5

45

40

35

30left right

spik

es s

_ 1

Phil Trans R Soc Lond B (2002)

fixation sample delay choice10

8

6

4

2

0

10

8

6

4

2

0plusmn500 0 500 1000 1500 2000

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)fi

rin

g r

ate

(Hz)

(a)

(b)

Figure 7 Activity of a single PFC neuron after the monkeywas trained on the three-category scheme (see gure 4)(a) The neuronrsquos activity when the data were sorted basedon the three-category scheme (category A black linecategory B grey line category C dotted line) (b) The sameneuronrsquos activity was data-sorted by the now-irrelevant two-category scheme (black line cats grey line dogs)

Figure 8 The activity of four single prefrontal neurons wheneach of two objects on different trials instructed either asaccade to the right or a saccade to the left The linesconnect the average values obtained when a given objectcued one or the other saccade The error bars show thestandard error of the mean Note that in each case theneuronrsquos activity depends on both the cue object and thesaccade direction and that the tuning is nonlinear orconjunctive That is the level of activity to a givencombination of object and saccade cannot be predicted fromthe neuronrsquos response to the other combinations Adaptedfrom Asaad et al (1998)

insensitive to or even discount associations that are notinformative) The underlying neural ensemble of a goal-directed task then might be comprised of neurons whoseactivity re ects task contingencies Many studies haveshown that prefrontal neurons do have this property Thiswork has focused on the lateral PFC because it seems tobe a site of convergence of the information needed to solveconditional sensorindashmotor tasks It is directly intercon-

Concepts and the prefrontal cortex E K Miller and others 1129

80

60

40

20

0

_2000 _1500 _1000 _500 0 500 1000 1500

40

30

20

10

0

20

15

10

5

0

10

8

6

4

2

0

spatialobjectassociative

ITI fix cue delay

trial time (ms) block number

1 2 3 4 5 6 7 8

spik

es s

_ 1(a)

(c)

(b)

(d)

Figure 9 Spike rate versus time histograms for two neurons each sorted by task The nal second of the three-second ITI isrepresented by the rst 1000 ms ( 2 2000 to 2 1000 ms) Fixation occurs soon after (ca 2 1000 to 2 800 ms) Cuepresentation occurs at the time-point marked 0 ms Task-related differences in baseline ring rate were generally observed tobegin in the xation period While the activity of some neurons diverged almost coincident with initial xation (c) the activityof others diverged progressively as the appearance of the cue became more imminent (a) The bar graphs (bd) demonstratethe reproducibility of these small task-speci c changes in activity across multiple repetitions of the same task The mean xation-period ring rate (with standard errors) for each block of trials is shown for the two neurons in (a) and (c) The barsare colour-coded to re ect the task being performed in each block and the colours match those in the histograms to the leftThe light grey line superimposed over these bars shows the activity of these neurons during the second immediately precedingthe ITI

nected with the higher-order sensory and motor cortexand indirectly connected (via the ventromedial PFC) withlimbic structures that process lsquointernalrsquo information suchas memory and reward (Goldman-Rakic 1987 Pandya ampBarnes 1987 Fuster 1989 Barbas amp Pandya 1991) Theneural activity in the lateral PFC re ects this many of itsneurons exhibit multimodal responses (Vaadia et al 1986Watanabe 1992 Rao et al 1997 Rainer et al 1998abWhite amp Wise 1999) Further the lateral PFC is criticalfor normal learning of conditional associations betweensensory cues and voluntary actions (Petrides 1985a 1990Gaffan amp Harrison 1988 Eacott amp Gaffan 1992 Parker ampGaffan 1998b) Indeed following training on conditionallearning tasks as many as 50 of the neurons in the lat-eral PFC show conjunctive tuning for learned associationsbetween cues voluntary actions and rewards

For example Watanabe (1990 1992) trained monkeysto perform tasks in which visual and auditory cues sig-nalled in different trials whether a reward would orwould not be delivered The majority of lateral PFC neu-rons were found to re ect the association between a cueand a reward A given neuron might be activated by a cuebut only when it signalled a reward By contrast another

Phil Trans R Soc Lond B (2002)

neuron might be activated only by a cue that signalled lsquonorewardrsquo In our own experiments we have trained mon-keys to associate in different blocks of trials each of twocue objects with an eye saccade to the right or to the left(Asaad et al 1998) We found that the activity of 44 ofrandomly selected lateral PFC neurons re ected associ-ations between objects and the saccades that theyinstructed ( gure 8) Other neurons had activity thatre ected the cues or the saccades alone but they werefewer in number Fuster et al (2000) have recently shownthat PFC neurons can also re ect learned associationsbetween visual and auditory stimuli

Importantly these changes do not require a prodigiousamount of experience Changes in PFC neural propertiesare evident after one dayrsquos experience and can be detectedafter just a few minutes of training For example Bichotet al (1996) studied the FEFs part of Brodmannrsquos area 8that is important for voluntary eye movements Normallyneurons in this area re selectively to saccade targetsappearing in certain visual eld locations However whenmonkeys were trained to search for a target de ned by aparticular visual attribute (eg red) the neurons in theFEFs acquire sensitivity to that attribute (Bichot et al

1130 E K Miller and others Concepts and the prefrontal cortex

1996) When monkeys were trained to search for a differ-ent target every day neurons not only discriminated thecurrent target but also distracting stimuli that had been atarget on the previous day relative to stimuli that had beentargets even earlier (Bichot amp Schall 1999) Monkeys werealso more likely to make errors in choosing that distractingstimulus It was as though the previous dayrsquos experienceleft an impression in the brain that in uenced neuralactivity and the monkeyrsquos behaviour

We also observed evidence for rapid plasticity in ourstudy of learning of conditional objectndashsaccade associ-ations in the PFC (Asaad et al 1998) Initially the monk-eys chose their responses at random but learned thecorrect cuendashresponse pairing over a few (5ndash15) trials Asthey learned the association neural activity representingthe forthcoming saccadic response appeared progressivelyearlier in successive trials In other words the initiation ofresponse-related delay activity gradually shifted withlearningmdashfrom a point in time just before the executionof the response and reward delivery to an earlier point intime nearly coincident with the presentation of the cue

Further support for a role for PFC neurons in rep-resenting task demands comes from training monkeys toalternate between different task rules This adds anotherlevel of complexity beyond the conditional tasks describedabove Now there is more than one rule assigned to eachcue and another cue tells the monkey which rule to usein a given trial For example following a given cue monk-eys can learn to direct a response to either the cuersquoslocation (spatial matching rule) or an alternative locationassociated with the cue (associative rule) depending onwhich rule is currently in effect When tested in thisfashion many PFC neurons show rule-speci c activityFor example a PFC neuron might respond to a given vis-ual cue when the monkey is using an associative rule butexhibit weak or no activity under identical sensory andattentional conditions that differed only in that the mon-key was using a spatial rule instead (White amp Wise 1999Asaad et al 2000) Also when monkeys switch betweendifferent tasks many PFC neurons show shifts in baselineactivity that communicates which task is currently beingperformed (Asaad et al 2000 gure 9) Such effects havebeen found in the PFC for associative versus spatial rulesfor object matching versus spatial matching versus associ-ative rules and for shape matching versus object matchingrules (Hoshi et al 1998 White amp Wise 1999 Asaad etal 2000)

In all of these cases however the rules are relativelyliteral or concrete A certain cue or set of cues alwayssignals a speci c response As noted knowledge from ourpast experiences can be applied to a wider range of futurecircumstances if we abstract general principles or rulesrather than speci c cuendashresponse contingencies Theability of the PFC to represent abstract rules those nottied to speci c stimuli or actions was recently addressedin our laboratory (Wallis et al 2001)

Monkeys were trained to use two abstract rules lsquomatchrsquoversus lsquonon-matchrsquo They faced a computer screen andviewed two successively presented pictures If the matchrule was in effect the monkeys released the lever if thepictures were the same and continued to hold the lever ifthe pictures were different If the non-match rule was ineffect the reverse was true monkeys released if the pic-

Phil Trans R Soc Lond B (2002)

tures were different and held if they were the same Therule was randomly instructed in each trial by the presen-tation of a cue at the same time as the rst picture waspresented To disambiguate neuronal responses to thephysical properties of the cue from responses to the rulethat the cues instructed cues signifying the same rule weretaken from different modalities while cues signifying dif-ferent rules were taken from the same modality ( gure10) The monkeys could perform this task well abovechance levels even when they were seeing a stimulus forthe very rst time This indicates that they had abstractedtwo overarching principles of the task that could then beapplied to novel stimulimdashthe minimal de nition of anabstract rule

The most prevalent activity across the PFC was theencoding of the current rule Figure 11 shows a goodexample of a rule-selective neuron that exhibited greateractivity when the match rule was in effect than when thenon-match rule had been indicated This activity cannotbe explained by the physical properties of the cue or thepicture since activity was the same regardless of whichcue was used to instruct the monkey and regardless ofwhich picture the monkey was remembering It cannot berelated to the upcoming response since the monkey didnot know whether the second-presented picture wouldrequire a response Nor could it be related to differencesin reward expectation since the expectation of reward wasthe same regardless of which rule was in effect Further-more the performance of the monkeys was virtually ident-ical for the two types of rules (error rates differed by lessthan 01 and reaction times by less than 7 ms) Thusthe most parsimonious explanation is that the differencesin activity re ected the abstract rule that the monkey wascurrently using to guide its behaviour

What function does the ability to abstract a rule serveIt is a form of generalization that permits a shortcut inlearning thereby allowing the animal to maximize theamount of reward available from a particular situation Toillustrate this consider the above task The monkey couldpotentially solve the task as a series of paired associates(in fact 16 associations consisting of four different pic-tures each paired with four different stimuli) Forexample the monkey might learn that whenever the chefis presented with a drop of juice then at the test phasethe correct response is to choose the chef But notice thatthis type of learning tells the monkey nothing about whichresponse is appropriate to a lion appearing with a drop ofjuice In other words unless the monkey abstracts the rulethat juice indicates that the monkey should match theneach time new pictures are used the monkey would haveto learn an entirely new set of 16 associations by trial anderror The problem with this trial-and-error learning isthat errors are lost opportunities for reward Given thatthe monkeys performed well above chance when theyencountered novel pictures it is clear that they are notengaging in trial-and-error learning but rather haveabstracted two rules that they can then apply as required

This shortcut that abstraction of the rule permits is alsore ected in the neuronal encoding It would be entirelypossible for the monkey to solve the task without singlecells encoding the rule For example there might be twopopulations of cells one encoding the match and non-match rule when the cues are presented in the auditory

Concepts and the prefrontal cortex E K Miller and others 1131

+ juice

+ low tone

or

+ juice

+ high tone

or

match

non-match

Figure 10 Schematic diagram of the abstract rules taskMonkeys switched between choosing a test stimulus that didor did not match the sample depending on whether thematch or non-match rule was in effect The cues thatsignalled these rules are shown on the left

modality and one encoding this information in the tastemodality But such a solution is computationally expens-ive since if a third modality was introduced a third popu-lation of cells would be required It is more ef cient toabstract a rule that cues presented in different modalitiescommonly instruct and indeed this is the solution thatthe brain uses The prevalence of neurons encoding suchrules in the PFC is consistent with the loss of exibilitythat is observed after prefrontal damage in both monkeysand humans It is not inconsistent with studies emphasiz-ing the role of the PFC in working memory or planningbut indicates that an important component of these pro-cesses might be the use of behaviour-guiding rules

5 PREFRONTAL CORTEX AND COGNITIVECONTROL

The results presented above suggest that PFC activityre ects categories and rules This seems consistent withconjectures that a cardinal PFC function may be theacquisition and representation of the formal demands oftasks the guiding concepts and principles that provide afoundation for complex intelligent behaviour In order tounderstand this we must turn to theories of cognitive con-trolmdashthe ability of the brain to coordinate processingamong its millions of neurons in order to direct themtoward future goals

(a) Controlled versus automatic behavioursIn order to understand what we mean by lsquocognitive con-

trolrsquo it is important to understand the distinction betweencontrolled and automatic behaviours Much of our behav-iour is automatic that is direct reactions to our immediateenvironment that do not tax our attention For exampleif someone suddenly throws a baseball at your face youmight re exively duck You may not have willed thisbehaviour it just seems to happen Many such re exiveautomatic processes are lsquowiredrsquo into our nervous systemsby evolution However others can be acquired through agreat deal of experience as learning mechanisms graduallyimprint highly familiar behaviour If you are walking ahighly familiar route and traf c is light you may traversea great distance (and even negotiate turns) with little

Phil Trans R Soc Lond B (2002)

awareness of having done so In these cases your behav-iour is driven in a lsquobottom-uprsquo fashion largely determinedby stimuli in the immediate environment and their abilityto trigger behaviours with which they are strongly associa-ted In neural terms they are dependent on well-established neural pathways waiting to be red off by thecorrect input

However if something unexpected happens on yourwalk you need to lsquotake chargersquo of your actions You payattention to your surroundings and try to anticipate andaccommodate the action of others you may even decideto take an alternative route In this case your behaviour isnot governed by simple inputndashoutput stimulusndashresponserelationships You use your knowledge of the world thecurrent objective (arriving at work intact) and results fromprevious experiences to weigh the alternatives and conse-quences Because these behaviours tax your attention andseem to be driven by lsquointernalrsquo information (knowledgeabout unseen goals and how to achieve them) and thusare initiated by us and not by the environment they arereferred to as lsquocontrolledrsquo The same basic sensory mem-ory and motor processes that mediated automatic behav-iour can be engaged However now they are not simplytriggered by the environment they are shaped and con-trolled in a top-down fashion by our knowledge of howthe world works The observations that PFC neuronsencode acquired information about task contingenciescategories and rules suggest it as a source of top-down sig-nals

The observation that humans with PFC damage seemimpulsive makes sense in light of this distinction betweencontrolled and automatic behaviours Without the PFC toprovide top-down signals about expectations of goals andrequired behaviours the patient simply reacts to theirenvironment with whatever behaviours are strongly asso-ciated with the cues that are immediately presentShallice amp Burgess (1991) examined this by using a lsquoshop-ping testrsquo They described frontal lobe-damaged patientswho are able to execute simple routines in which clearsensory cues could elicit a familiar action (eg buy a loafof bread) However they were unable to carry out anerrand that involved organizing a series of such routinesbecause they kept going lsquooff taskrsquo They would forexample enter shops that were irrelevant to the errandjust because they happened to be passing them Anotherexample is utilization behaviour A patient with PFC dam-age will impulsively use items placed in front of them suchas a comb or in (hopefully) one case a urinal It seemsthat the basic elements of behaviour are intact but thatthe patients override prepotent re exive responses tocoordinate behaviour in accord with an unseen goal

A classic test of the ability to learn and follow goal-orientated rules is the Wisconsin Card Sorting Test(Milner 1963) Subjects are instructed to sort cardsaccording to the shape colour or number of symbolsappearing on them They start with one rule (eg colour)and once that is acquired the rule changes until all of thecards have been sorted using all possible rules Normalhumans have little dif culty with this task By contrasthumans with prefrontal damage can learn the rst sortingcriterion (a relatively simple mapping between a stimulusattribute and a response) but then are unable to escape itthey cannot override the previous behaviour and do not

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

Concepts and the prefrontal cortex E K Miller and others 1127

C2

C1 C3

2-class boundary

D1 D3

D2

3-class boundaries

100 C1 80 C1 60 C1 60 D1 80 D1 100 D1

cats

dogs

(a)

(b)

Figure 4 (a) Monkeys learned to categorize randomly generated lsquomorphsrsquo from the vast number of possible blends of sixprototypes For neurophysiological recording 54 sample stimuli were constructed along the 15 morph lines illustrated here(b) Morphs along the C1ndashD1 line From Freedman et al (2001)

fixation

500 ms sample

600 ms delay

1000 ms

test

(non-match)

600 msdelay

600 ms test

(match)

(match)

Figure 5 The delayed match-to-category task A sample was followed by a delay and a test stimulus If the sample and teststimulus were the same category (a match) monkeys were required to release a lever before the test disappeared If they werenot there was another delay followed by a match Equal numbers of match and non-match trials were randomly interleavedFrom Freedman et al (2001)

Phil Trans R Soc Lond B (2002)

1128 E K Miller and others Concepts and the prefrontal cortex

13

10

7

4

1

fixation sample delay choice

firi

ng

rat

e (H

z)

dog 100dog 80dog 60

cat 100cat 80cat 60

_500 0 500 1000 1500 2000

dog 100dog 80dog 60

cat 100cat 80cat 60

70

60

50

40

30

20

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)

(a)

(b)

Figure 6 The average activity of two single neurons tostimuli at the six morph blends The vertical linescorrespond (from left to right) to sample onset offset andtest stimulus onset Activity is pooled over match and non-match trials

35

30

25

20

15

20

15

10

5left right

20

15

10

5

45

40

35

30left right

spik

es s

_ 1

Phil Trans R Soc Lond B (2002)

fixation sample delay choice10

8

6

4

2

0

10

8

6

4

2

0plusmn500 0 500 1000 1500 2000

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)fi

rin

g r

ate

(Hz)

(a)

(b)

Figure 7 Activity of a single PFC neuron after the monkeywas trained on the three-category scheme (see gure 4)(a) The neuronrsquos activity when the data were sorted basedon the three-category scheme (category A black linecategory B grey line category C dotted line) (b) The sameneuronrsquos activity was data-sorted by the now-irrelevant two-category scheme (black line cats grey line dogs)

Figure 8 The activity of four single prefrontal neurons wheneach of two objects on different trials instructed either asaccade to the right or a saccade to the left The linesconnect the average values obtained when a given objectcued one or the other saccade The error bars show thestandard error of the mean Note that in each case theneuronrsquos activity depends on both the cue object and thesaccade direction and that the tuning is nonlinear orconjunctive That is the level of activity to a givencombination of object and saccade cannot be predicted fromthe neuronrsquos response to the other combinations Adaptedfrom Asaad et al (1998)

insensitive to or even discount associations that are notinformative) The underlying neural ensemble of a goal-directed task then might be comprised of neurons whoseactivity re ects task contingencies Many studies haveshown that prefrontal neurons do have this property Thiswork has focused on the lateral PFC because it seems tobe a site of convergence of the information needed to solveconditional sensorindashmotor tasks It is directly intercon-

Concepts and the prefrontal cortex E K Miller and others 1129

80

60

40

20

0

_2000 _1500 _1000 _500 0 500 1000 1500

40

30

20

10

0

20

15

10

5

0

10

8

6

4

2

0

spatialobjectassociative

ITI fix cue delay

trial time (ms) block number

1 2 3 4 5 6 7 8

spik

es s

_ 1(a)

(c)

(b)

(d)

Figure 9 Spike rate versus time histograms for two neurons each sorted by task The nal second of the three-second ITI isrepresented by the rst 1000 ms ( 2 2000 to 2 1000 ms) Fixation occurs soon after (ca 2 1000 to 2 800 ms) Cuepresentation occurs at the time-point marked 0 ms Task-related differences in baseline ring rate were generally observed tobegin in the xation period While the activity of some neurons diverged almost coincident with initial xation (c) the activityof others diverged progressively as the appearance of the cue became more imminent (a) The bar graphs (bd) demonstratethe reproducibility of these small task-speci c changes in activity across multiple repetitions of the same task The mean xation-period ring rate (with standard errors) for each block of trials is shown for the two neurons in (a) and (c) The barsare colour-coded to re ect the task being performed in each block and the colours match those in the histograms to the leftThe light grey line superimposed over these bars shows the activity of these neurons during the second immediately precedingthe ITI

nected with the higher-order sensory and motor cortexand indirectly connected (via the ventromedial PFC) withlimbic structures that process lsquointernalrsquo information suchas memory and reward (Goldman-Rakic 1987 Pandya ampBarnes 1987 Fuster 1989 Barbas amp Pandya 1991) Theneural activity in the lateral PFC re ects this many of itsneurons exhibit multimodal responses (Vaadia et al 1986Watanabe 1992 Rao et al 1997 Rainer et al 1998abWhite amp Wise 1999) Further the lateral PFC is criticalfor normal learning of conditional associations betweensensory cues and voluntary actions (Petrides 1985a 1990Gaffan amp Harrison 1988 Eacott amp Gaffan 1992 Parker ampGaffan 1998b) Indeed following training on conditionallearning tasks as many as 50 of the neurons in the lat-eral PFC show conjunctive tuning for learned associationsbetween cues voluntary actions and rewards

For example Watanabe (1990 1992) trained monkeysto perform tasks in which visual and auditory cues sig-nalled in different trials whether a reward would orwould not be delivered The majority of lateral PFC neu-rons were found to re ect the association between a cueand a reward A given neuron might be activated by a cuebut only when it signalled a reward By contrast another

Phil Trans R Soc Lond B (2002)

neuron might be activated only by a cue that signalled lsquonorewardrsquo In our own experiments we have trained mon-keys to associate in different blocks of trials each of twocue objects with an eye saccade to the right or to the left(Asaad et al 1998) We found that the activity of 44 ofrandomly selected lateral PFC neurons re ected associ-ations between objects and the saccades that theyinstructed ( gure 8) Other neurons had activity thatre ected the cues or the saccades alone but they werefewer in number Fuster et al (2000) have recently shownthat PFC neurons can also re ect learned associationsbetween visual and auditory stimuli

Importantly these changes do not require a prodigiousamount of experience Changes in PFC neural propertiesare evident after one dayrsquos experience and can be detectedafter just a few minutes of training For example Bichotet al (1996) studied the FEFs part of Brodmannrsquos area 8that is important for voluntary eye movements Normallyneurons in this area re selectively to saccade targetsappearing in certain visual eld locations However whenmonkeys were trained to search for a target de ned by aparticular visual attribute (eg red) the neurons in theFEFs acquire sensitivity to that attribute (Bichot et al

1130 E K Miller and others Concepts and the prefrontal cortex

1996) When monkeys were trained to search for a differ-ent target every day neurons not only discriminated thecurrent target but also distracting stimuli that had been atarget on the previous day relative to stimuli that had beentargets even earlier (Bichot amp Schall 1999) Monkeys werealso more likely to make errors in choosing that distractingstimulus It was as though the previous dayrsquos experienceleft an impression in the brain that in uenced neuralactivity and the monkeyrsquos behaviour

We also observed evidence for rapid plasticity in ourstudy of learning of conditional objectndashsaccade associ-ations in the PFC (Asaad et al 1998) Initially the monk-eys chose their responses at random but learned thecorrect cuendashresponse pairing over a few (5ndash15) trials Asthey learned the association neural activity representingthe forthcoming saccadic response appeared progressivelyearlier in successive trials In other words the initiation ofresponse-related delay activity gradually shifted withlearningmdashfrom a point in time just before the executionof the response and reward delivery to an earlier point intime nearly coincident with the presentation of the cue

Further support for a role for PFC neurons in rep-resenting task demands comes from training monkeys toalternate between different task rules This adds anotherlevel of complexity beyond the conditional tasks describedabove Now there is more than one rule assigned to eachcue and another cue tells the monkey which rule to usein a given trial For example following a given cue monk-eys can learn to direct a response to either the cuersquoslocation (spatial matching rule) or an alternative locationassociated with the cue (associative rule) depending onwhich rule is currently in effect When tested in thisfashion many PFC neurons show rule-speci c activityFor example a PFC neuron might respond to a given vis-ual cue when the monkey is using an associative rule butexhibit weak or no activity under identical sensory andattentional conditions that differed only in that the mon-key was using a spatial rule instead (White amp Wise 1999Asaad et al 2000) Also when monkeys switch betweendifferent tasks many PFC neurons show shifts in baselineactivity that communicates which task is currently beingperformed (Asaad et al 2000 gure 9) Such effects havebeen found in the PFC for associative versus spatial rulesfor object matching versus spatial matching versus associ-ative rules and for shape matching versus object matchingrules (Hoshi et al 1998 White amp Wise 1999 Asaad etal 2000)

In all of these cases however the rules are relativelyliteral or concrete A certain cue or set of cues alwayssignals a speci c response As noted knowledge from ourpast experiences can be applied to a wider range of futurecircumstances if we abstract general principles or rulesrather than speci c cuendashresponse contingencies Theability of the PFC to represent abstract rules those nottied to speci c stimuli or actions was recently addressedin our laboratory (Wallis et al 2001)

Monkeys were trained to use two abstract rules lsquomatchrsquoversus lsquonon-matchrsquo They faced a computer screen andviewed two successively presented pictures If the matchrule was in effect the monkeys released the lever if thepictures were the same and continued to hold the lever ifthe pictures were different If the non-match rule was ineffect the reverse was true monkeys released if the pic-

Phil Trans R Soc Lond B (2002)

tures were different and held if they were the same Therule was randomly instructed in each trial by the presen-tation of a cue at the same time as the rst picture waspresented To disambiguate neuronal responses to thephysical properties of the cue from responses to the rulethat the cues instructed cues signifying the same rule weretaken from different modalities while cues signifying dif-ferent rules were taken from the same modality ( gure10) The monkeys could perform this task well abovechance levels even when they were seeing a stimulus forthe very rst time This indicates that they had abstractedtwo overarching principles of the task that could then beapplied to novel stimulimdashthe minimal de nition of anabstract rule

The most prevalent activity across the PFC was theencoding of the current rule Figure 11 shows a goodexample of a rule-selective neuron that exhibited greateractivity when the match rule was in effect than when thenon-match rule had been indicated This activity cannotbe explained by the physical properties of the cue or thepicture since activity was the same regardless of whichcue was used to instruct the monkey and regardless ofwhich picture the monkey was remembering It cannot berelated to the upcoming response since the monkey didnot know whether the second-presented picture wouldrequire a response Nor could it be related to differencesin reward expectation since the expectation of reward wasthe same regardless of which rule was in effect Further-more the performance of the monkeys was virtually ident-ical for the two types of rules (error rates differed by lessthan 01 and reaction times by less than 7 ms) Thusthe most parsimonious explanation is that the differencesin activity re ected the abstract rule that the monkey wascurrently using to guide its behaviour

What function does the ability to abstract a rule serveIt is a form of generalization that permits a shortcut inlearning thereby allowing the animal to maximize theamount of reward available from a particular situation Toillustrate this consider the above task The monkey couldpotentially solve the task as a series of paired associates(in fact 16 associations consisting of four different pic-tures each paired with four different stimuli) Forexample the monkey might learn that whenever the chefis presented with a drop of juice then at the test phasethe correct response is to choose the chef But notice thatthis type of learning tells the monkey nothing about whichresponse is appropriate to a lion appearing with a drop ofjuice In other words unless the monkey abstracts the rulethat juice indicates that the monkey should match theneach time new pictures are used the monkey would haveto learn an entirely new set of 16 associations by trial anderror The problem with this trial-and-error learning isthat errors are lost opportunities for reward Given thatthe monkeys performed well above chance when theyencountered novel pictures it is clear that they are notengaging in trial-and-error learning but rather haveabstracted two rules that they can then apply as required

This shortcut that abstraction of the rule permits is alsore ected in the neuronal encoding It would be entirelypossible for the monkey to solve the task without singlecells encoding the rule For example there might be twopopulations of cells one encoding the match and non-match rule when the cues are presented in the auditory

Concepts and the prefrontal cortex E K Miller and others 1131

+ juice

+ low tone

or

+ juice

+ high tone

or

match

non-match

Figure 10 Schematic diagram of the abstract rules taskMonkeys switched between choosing a test stimulus that didor did not match the sample depending on whether thematch or non-match rule was in effect The cues thatsignalled these rules are shown on the left

modality and one encoding this information in the tastemodality But such a solution is computationally expens-ive since if a third modality was introduced a third popu-lation of cells would be required It is more ef cient toabstract a rule that cues presented in different modalitiescommonly instruct and indeed this is the solution thatthe brain uses The prevalence of neurons encoding suchrules in the PFC is consistent with the loss of exibilitythat is observed after prefrontal damage in both monkeysand humans It is not inconsistent with studies emphasiz-ing the role of the PFC in working memory or planningbut indicates that an important component of these pro-cesses might be the use of behaviour-guiding rules

5 PREFRONTAL CORTEX AND COGNITIVECONTROL

The results presented above suggest that PFC activityre ects categories and rules This seems consistent withconjectures that a cardinal PFC function may be theacquisition and representation of the formal demands oftasks the guiding concepts and principles that provide afoundation for complex intelligent behaviour In order tounderstand this we must turn to theories of cognitive con-trolmdashthe ability of the brain to coordinate processingamong its millions of neurons in order to direct themtoward future goals

(a) Controlled versus automatic behavioursIn order to understand what we mean by lsquocognitive con-

trolrsquo it is important to understand the distinction betweencontrolled and automatic behaviours Much of our behav-iour is automatic that is direct reactions to our immediateenvironment that do not tax our attention For exampleif someone suddenly throws a baseball at your face youmight re exively duck You may not have willed thisbehaviour it just seems to happen Many such re exiveautomatic processes are lsquowiredrsquo into our nervous systemsby evolution However others can be acquired through agreat deal of experience as learning mechanisms graduallyimprint highly familiar behaviour If you are walking ahighly familiar route and traf c is light you may traversea great distance (and even negotiate turns) with little

Phil Trans R Soc Lond B (2002)

awareness of having done so In these cases your behav-iour is driven in a lsquobottom-uprsquo fashion largely determinedby stimuli in the immediate environment and their abilityto trigger behaviours with which they are strongly associa-ted In neural terms they are dependent on well-established neural pathways waiting to be red off by thecorrect input

However if something unexpected happens on yourwalk you need to lsquotake chargersquo of your actions You payattention to your surroundings and try to anticipate andaccommodate the action of others you may even decideto take an alternative route In this case your behaviour isnot governed by simple inputndashoutput stimulusndashresponserelationships You use your knowledge of the world thecurrent objective (arriving at work intact) and results fromprevious experiences to weigh the alternatives and conse-quences Because these behaviours tax your attention andseem to be driven by lsquointernalrsquo information (knowledgeabout unseen goals and how to achieve them) and thusare initiated by us and not by the environment they arereferred to as lsquocontrolledrsquo The same basic sensory mem-ory and motor processes that mediated automatic behav-iour can be engaged However now they are not simplytriggered by the environment they are shaped and con-trolled in a top-down fashion by our knowledge of howthe world works The observations that PFC neuronsencode acquired information about task contingenciescategories and rules suggest it as a source of top-down sig-nals

The observation that humans with PFC damage seemimpulsive makes sense in light of this distinction betweencontrolled and automatic behaviours Without the PFC toprovide top-down signals about expectations of goals andrequired behaviours the patient simply reacts to theirenvironment with whatever behaviours are strongly asso-ciated with the cues that are immediately presentShallice amp Burgess (1991) examined this by using a lsquoshop-ping testrsquo They described frontal lobe-damaged patientswho are able to execute simple routines in which clearsensory cues could elicit a familiar action (eg buy a loafof bread) However they were unable to carry out anerrand that involved organizing a series of such routinesbecause they kept going lsquooff taskrsquo They would forexample enter shops that were irrelevant to the errandjust because they happened to be passing them Anotherexample is utilization behaviour A patient with PFC dam-age will impulsively use items placed in front of them suchas a comb or in (hopefully) one case a urinal It seemsthat the basic elements of behaviour are intact but thatthe patients override prepotent re exive responses tocoordinate behaviour in accord with an unseen goal

A classic test of the ability to learn and follow goal-orientated rules is the Wisconsin Card Sorting Test(Milner 1963) Subjects are instructed to sort cardsaccording to the shape colour or number of symbolsappearing on them They start with one rule (eg colour)and once that is acquired the rule changes until all of thecards have been sorted using all possible rules Normalhumans have little dif culty with this task By contrasthumans with prefrontal damage can learn the rst sortingcriterion (a relatively simple mapping between a stimulusattribute and a response) but then are unable to escape itthey cannot override the previous behaviour and do not

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

1128 E K Miller and others Concepts and the prefrontal cortex

13

10

7

4

1

fixation sample delay choice

firi

ng

rat

e (H

z)

dog 100dog 80dog 60

cat 100cat 80cat 60

_500 0 500 1000 1500 2000

dog 100dog 80dog 60

cat 100cat 80cat 60

70

60

50

40

30

20

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)

(a)

(b)

Figure 6 The average activity of two single neurons tostimuli at the six morph blends The vertical linescorrespond (from left to right) to sample onset offset andtest stimulus onset Activity is pooled over match and non-match trials

35

30

25

20

15

20

15

10

5left right

20

15

10

5

45

40

35

30left right

spik

es s

_ 1

Phil Trans R Soc Lond B (2002)

fixation sample delay choice10

8

6

4

2

0

10

8

6

4

2

0plusmn500 0 500 1000 1500 2000

time from sample stimulus onset (ms)

firi

ng

rat

e (H

z)fi

rin

g r

ate

(Hz)

(a)

(b)

Figure 7 Activity of a single PFC neuron after the monkeywas trained on the three-category scheme (see gure 4)(a) The neuronrsquos activity when the data were sorted basedon the three-category scheme (category A black linecategory B grey line category C dotted line) (b) The sameneuronrsquos activity was data-sorted by the now-irrelevant two-category scheme (black line cats grey line dogs)

Figure 8 The activity of four single prefrontal neurons wheneach of two objects on different trials instructed either asaccade to the right or a saccade to the left The linesconnect the average values obtained when a given objectcued one or the other saccade The error bars show thestandard error of the mean Note that in each case theneuronrsquos activity depends on both the cue object and thesaccade direction and that the tuning is nonlinear orconjunctive That is the level of activity to a givencombination of object and saccade cannot be predicted fromthe neuronrsquos response to the other combinations Adaptedfrom Asaad et al (1998)

insensitive to or even discount associations that are notinformative) The underlying neural ensemble of a goal-directed task then might be comprised of neurons whoseactivity re ects task contingencies Many studies haveshown that prefrontal neurons do have this property Thiswork has focused on the lateral PFC because it seems tobe a site of convergence of the information needed to solveconditional sensorindashmotor tasks It is directly intercon-

Concepts and the prefrontal cortex E K Miller and others 1129

80

60

40

20

0

_2000 _1500 _1000 _500 0 500 1000 1500

40

30

20

10

0

20

15

10

5

0

10

8

6

4

2

0

spatialobjectassociative

ITI fix cue delay

trial time (ms) block number

1 2 3 4 5 6 7 8

spik

es s

_ 1(a)

(c)

(b)

(d)

Figure 9 Spike rate versus time histograms for two neurons each sorted by task The nal second of the three-second ITI isrepresented by the rst 1000 ms ( 2 2000 to 2 1000 ms) Fixation occurs soon after (ca 2 1000 to 2 800 ms) Cuepresentation occurs at the time-point marked 0 ms Task-related differences in baseline ring rate were generally observed tobegin in the xation period While the activity of some neurons diverged almost coincident with initial xation (c) the activityof others diverged progressively as the appearance of the cue became more imminent (a) The bar graphs (bd) demonstratethe reproducibility of these small task-speci c changes in activity across multiple repetitions of the same task The mean xation-period ring rate (with standard errors) for each block of trials is shown for the two neurons in (a) and (c) The barsare colour-coded to re ect the task being performed in each block and the colours match those in the histograms to the leftThe light grey line superimposed over these bars shows the activity of these neurons during the second immediately precedingthe ITI

nected with the higher-order sensory and motor cortexand indirectly connected (via the ventromedial PFC) withlimbic structures that process lsquointernalrsquo information suchas memory and reward (Goldman-Rakic 1987 Pandya ampBarnes 1987 Fuster 1989 Barbas amp Pandya 1991) Theneural activity in the lateral PFC re ects this many of itsneurons exhibit multimodal responses (Vaadia et al 1986Watanabe 1992 Rao et al 1997 Rainer et al 1998abWhite amp Wise 1999) Further the lateral PFC is criticalfor normal learning of conditional associations betweensensory cues and voluntary actions (Petrides 1985a 1990Gaffan amp Harrison 1988 Eacott amp Gaffan 1992 Parker ampGaffan 1998b) Indeed following training on conditionallearning tasks as many as 50 of the neurons in the lat-eral PFC show conjunctive tuning for learned associationsbetween cues voluntary actions and rewards

For example Watanabe (1990 1992) trained monkeysto perform tasks in which visual and auditory cues sig-nalled in different trials whether a reward would orwould not be delivered The majority of lateral PFC neu-rons were found to re ect the association between a cueand a reward A given neuron might be activated by a cuebut only when it signalled a reward By contrast another

Phil Trans R Soc Lond B (2002)

neuron might be activated only by a cue that signalled lsquonorewardrsquo In our own experiments we have trained mon-keys to associate in different blocks of trials each of twocue objects with an eye saccade to the right or to the left(Asaad et al 1998) We found that the activity of 44 ofrandomly selected lateral PFC neurons re ected associ-ations between objects and the saccades that theyinstructed ( gure 8) Other neurons had activity thatre ected the cues or the saccades alone but they werefewer in number Fuster et al (2000) have recently shownthat PFC neurons can also re ect learned associationsbetween visual and auditory stimuli

Importantly these changes do not require a prodigiousamount of experience Changes in PFC neural propertiesare evident after one dayrsquos experience and can be detectedafter just a few minutes of training For example Bichotet al (1996) studied the FEFs part of Brodmannrsquos area 8that is important for voluntary eye movements Normallyneurons in this area re selectively to saccade targetsappearing in certain visual eld locations However whenmonkeys were trained to search for a target de ned by aparticular visual attribute (eg red) the neurons in theFEFs acquire sensitivity to that attribute (Bichot et al

1130 E K Miller and others Concepts and the prefrontal cortex

1996) When monkeys were trained to search for a differ-ent target every day neurons not only discriminated thecurrent target but also distracting stimuli that had been atarget on the previous day relative to stimuli that had beentargets even earlier (Bichot amp Schall 1999) Monkeys werealso more likely to make errors in choosing that distractingstimulus It was as though the previous dayrsquos experienceleft an impression in the brain that in uenced neuralactivity and the monkeyrsquos behaviour

We also observed evidence for rapid plasticity in ourstudy of learning of conditional objectndashsaccade associ-ations in the PFC (Asaad et al 1998) Initially the monk-eys chose their responses at random but learned thecorrect cuendashresponse pairing over a few (5ndash15) trials Asthey learned the association neural activity representingthe forthcoming saccadic response appeared progressivelyearlier in successive trials In other words the initiation ofresponse-related delay activity gradually shifted withlearningmdashfrom a point in time just before the executionof the response and reward delivery to an earlier point intime nearly coincident with the presentation of the cue

Further support for a role for PFC neurons in rep-resenting task demands comes from training monkeys toalternate between different task rules This adds anotherlevel of complexity beyond the conditional tasks describedabove Now there is more than one rule assigned to eachcue and another cue tells the monkey which rule to usein a given trial For example following a given cue monk-eys can learn to direct a response to either the cuersquoslocation (spatial matching rule) or an alternative locationassociated with the cue (associative rule) depending onwhich rule is currently in effect When tested in thisfashion many PFC neurons show rule-speci c activityFor example a PFC neuron might respond to a given vis-ual cue when the monkey is using an associative rule butexhibit weak or no activity under identical sensory andattentional conditions that differed only in that the mon-key was using a spatial rule instead (White amp Wise 1999Asaad et al 2000) Also when monkeys switch betweendifferent tasks many PFC neurons show shifts in baselineactivity that communicates which task is currently beingperformed (Asaad et al 2000 gure 9) Such effects havebeen found in the PFC for associative versus spatial rulesfor object matching versus spatial matching versus associ-ative rules and for shape matching versus object matchingrules (Hoshi et al 1998 White amp Wise 1999 Asaad etal 2000)

In all of these cases however the rules are relativelyliteral or concrete A certain cue or set of cues alwayssignals a speci c response As noted knowledge from ourpast experiences can be applied to a wider range of futurecircumstances if we abstract general principles or rulesrather than speci c cuendashresponse contingencies Theability of the PFC to represent abstract rules those nottied to speci c stimuli or actions was recently addressedin our laboratory (Wallis et al 2001)

Monkeys were trained to use two abstract rules lsquomatchrsquoversus lsquonon-matchrsquo They faced a computer screen andviewed two successively presented pictures If the matchrule was in effect the monkeys released the lever if thepictures were the same and continued to hold the lever ifthe pictures were different If the non-match rule was ineffect the reverse was true monkeys released if the pic-

Phil Trans R Soc Lond B (2002)

tures were different and held if they were the same Therule was randomly instructed in each trial by the presen-tation of a cue at the same time as the rst picture waspresented To disambiguate neuronal responses to thephysical properties of the cue from responses to the rulethat the cues instructed cues signifying the same rule weretaken from different modalities while cues signifying dif-ferent rules were taken from the same modality ( gure10) The monkeys could perform this task well abovechance levels even when they were seeing a stimulus forthe very rst time This indicates that they had abstractedtwo overarching principles of the task that could then beapplied to novel stimulimdashthe minimal de nition of anabstract rule

The most prevalent activity across the PFC was theencoding of the current rule Figure 11 shows a goodexample of a rule-selective neuron that exhibited greateractivity when the match rule was in effect than when thenon-match rule had been indicated This activity cannotbe explained by the physical properties of the cue or thepicture since activity was the same regardless of whichcue was used to instruct the monkey and regardless ofwhich picture the monkey was remembering It cannot berelated to the upcoming response since the monkey didnot know whether the second-presented picture wouldrequire a response Nor could it be related to differencesin reward expectation since the expectation of reward wasthe same regardless of which rule was in effect Further-more the performance of the monkeys was virtually ident-ical for the two types of rules (error rates differed by lessthan 01 and reaction times by less than 7 ms) Thusthe most parsimonious explanation is that the differencesin activity re ected the abstract rule that the monkey wascurrently using to guide its behaviour

What function does the ability to abstract a rule serveIt is a form of generalization that permits a shortcut inlearning thereby allowing the animal to maximize theamount of reward available from a particular situation Toillustrate this consider the above task The monkey couldpotentially solve the task as a series of paired associates(in fact 16 associations consisting of four different pic-tures each paired with four different stimuli) Forexample the monkey might learn that whenever the chefis presented with a drop of juice then at the test phasethe correct response is to choose the chef But notice thatthis type of learning tells the monkey nothing about whichresponse is appropriate to a lion appearing with a drop ofjuice In other words unless the monkey abstracts the rulethat juice indicates that the monkey should match theneach time new pictures are used the monkey would haveto learn an entirely new set of 16 associations by trial anderror The problem with this trial-and-error learning isthat errors are lost opportunities for reward Given thatthe monkeys performed well above chance when theyencountered novel pictures it is clear that they are notengaging in trial-and-error learning but rather haveabstracted two rules that they can then apply as required

This shortcut that abstraction of the rule permits is alsore ected in the neuronal encoding It would be entirelypossible for the monkey to solve the task without singlecells encoding the rule For example there might be twopopulations of cells one encoding the match and non-match rule when the cues are presented in the auditory

Concepts and the prefrontal cortex E K Miller and others 1131

+ juice

+ low tone

or

+ juice

+ high tone

or

match

non-match

Figure 10 Schematic diagram of the abstract rules taskMonkeys switched between choosing a test stimulus that didor did not match the sample depending on whether thematch or non-match rule was in effect The cues thatsignalled these rules are shown on the left

modality and one encoding this information in the tastemodality But such a solution is computationally expens-ive since if a third modality was introduced a third popu-lation of cells would be required It is more ef cient toabstract a rule that cues presented in different modalitiescommonly instruct and indeed this is the solution thatthe brain uses The prevalence of neurons encoding suchrules in the PFC is consistent with the loss of exibilitythat is observed after prefrontal damage in both monkeysand humans It is not inconsistent with studies emphasiz-ing the role of the PFC in working memory or planningbut indicates that an important component of these pro-cesses might be the use of behaviour-guiding rules

5 PREFRONTAL CORTEX AND COGNITIVECONTROL

The results presented above suggest that PFC activityre ects categories and rules This seems consistent withconjectures that a cardinal PFC function may be theacquisition and representation of the formal demands oftasks the guiding concepts and principles that provide afoundation for complex intelligent behaviour In order tounderstand this we must turn to theories of cognitive con-trolmdashthe ability of the brain to coordinate processingamong its millions of neurons in order to direct themtoward future goals

(a) Controlled versus automatic behavioursIn order to understand what we mean by lsquocognitive con-

trolrsquo it is important to understand the distinction betweencontrolled and automatic behaviours Much of our behav-iour is automatic that is direct reactions to our immediateenvironment that do not tax our attention For exampleif someone suddenly throws a baseball at your face youmight re exively duck You may not have willed thisbehaviour it just seems to happen Many such re exiveautomatic processes are lsquowiredrsquo into our nervous systemsby evolution However others can be acquired through agreat deal of experience as learning mechanisms graduallyimprint highly familiar behaviour If you are walking ahighly familiar route and traf c is light you may traversea great distance (and even negotiate turns) with little

Phil Trans R Soc Lond B (2002)

awareness of having done so In these cases your behav-iour is driven in a lsquobottom-uprsquo fashion largely determinedby stimuli in the immediate environment and their abilityto trigger behaviours with which they are strongly associa-ted In neural terms they are dependent on well-established neural pathways waiting to be red off by thecorrect input

However if something unexpected happens on yourwalk you need to lsquotake chargersquo of your actions You payattention to your surroundings and try to anticipate andaccommodate the action of others you may even decideto take an alternative route In this case your behaviour isnot governed by simple inputndashoutput stimulusndashresponserelationships You use your knowledge of the world thecurrent objective (arriving at work intact) and results fromprevious experiences to weigh the alternatives and conse-quences Because these behaviours tax your attention andseem to be driven by lsquointernalrsquo information (knowledgeabout unseen goals and how to achieve them) and thusare initiated by us and not by the environment they arereferred to as lsquocontrolledrsquo The same basic sensory mem-ory and motor processes that mediated automatic behav-iour can be engaged However now they are not simplytriggered by the environment they are shaped and con-trolled in a top-down fashion by our knowledge of howthe world works The observations that PFC neuronsencode acquired information about task contingenciescategories and rules suggest it as a source of top-down sig-nals

The observation that humans with PFC damage seemimpulsive makes sense in light of this distinction betweencontrolled and automatic behaviours Without the PFC toprovide top-down signals about expectations of goals andrequired behaviours the patient simply reacts to theirenvironment with whatever behaviours are strongly asso-ciated with the cues that are immediately presentShallice amp Burgess (1991) examined this by using a lsquoshop-ping testrsquo They described frontal lobe-damaged patientswho are able to execute simple routines in which clearsensory cues could elicit a familiar action (eg buy a loafof bread) However they were unable to carry out anerrand that involved organizing a series of such routinesbecause they kept going lsquooff taskrsquo They would forexample enter shops that were irrelevant to the errandjust because they happened to be passing them Anotherexample is utilization behaviour A patient with PFC dam-age will impulsively use items placed in front of them suchas a comb or in (hopefully) one case a urinal It seemsthat the basic elements of behaviour are intact but thatthe patients override prepotent re exive responses tocoordinate behaviour in accord with an unseen goal

A classic test of the ability to learn and follow goal-orientated rules is the Wisconsin Card Sorting Test(Milner 1963) Subjects are instructed to sort cardsaccording to the shape colour or number of symbolsappearing on them They start with one rule (eg colour)and once that is acquired the rule changes until all of thecards have been sorted using all possible rules Normalhumans have little dif culty with this task By contrasthumans with prefrontal damage can learn the rst sortingcriterion (a relatively simple mapping between a stimulusattribute and a response) but then are unable to escape itthey cannot override the previous behaviour and do not

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

Concepts and the prefrontal cortex E K Miller and others 1129

80

60

40

20

0

_2000 _1500 _1000 _500 0 500 1000 1500

40

30

20

10

0

20

15

10

5

0

10

8

6

4

2

0

spatialobjectassociative

ITI fix cue delay

trial time (ms) block number

1 2 3 4 5 6 7 8

spik

es s

_ 1(a)

(c)

(b)

(d)

Figure 9 Spike rate versus time histograms for two neurons each sorted by task The nal second of the three-second ITI isrepresented by the rst 1000 ms ( 2 2000 to 2 1000 ms) Fixation occurs soon after (ca 2 1000 to 2 800 ms) Cuepresentation occurs at the time-point marked 0 ms Task-related differences in baseline ring rate were generally observed tobegin in the xation period While the activity of some neurons diverged almost coincident with initial xation (c) the activityof others diverged progressively as the appearance of the cue became more imminent (a) The bar graphs (bd) demonstratethe reproducibility of these small task-speci c changes in activity across multiple repetitions of the same task The mean xation-period ring rate (with standard errors) for each block of trials is shown for the two neurons in (a) and (c) The barsare colour-coded to re ect the task being performed in each block and the colours match those in the histograms to the leftThe light grey line superimposed over these bars shows the activity of these neurons during the second immediately precedingthe ITI

nected with the higher-order sensory and motor cortexand indirectly connected (via the ventromedial PFC) withlimbic structures that process lsquointernalrsquo information suchas memory and reward (Goldman-Rakic 1987 Pandya ampBarnes 1987 Fuster 1989 Barbas amp Pandya 1991) Theneural activity in the lateral PFC re ects this many of itsneurons exhibit multimodal responses (Vaadia et al 1986Watanabe 1992 Rao et al 1997 Rainer et al 1998abWhite amp Wise 1999) Further the lateral PFC is criticalfor normal learning of conditional associations betweensensory cues and voluntary actions (Petrides 1985a 1990Gaffan amp Harrison 1988 Eacott amp Gaffan 1992 Parker ampGaffan 1998b) Indeed following training on conditionallearning tasks as many as 50 of the neurons in the lat-eral PFC show conjunctive tuning for learned associationsbetween cues voluntary actions and rewards

For example Watanabe (1990 1992) trained monkeysto perform tasks in which visual and auditory cues sig-nalled in different trials whether a reward would orwould not be delivered The majority of lateral PFC neu-rons were found to re ect the association between a cueand a reward A given neuron might be activated by a cuebut only when it signalled a reward By contrast another

Phil Trans R Soc Lond B (2002)

neuron might be activated only by a cue that signalled lsquonorewardrsquo In our own experiments we have trained mon-keys to associate in different blocks of trials each of twocue objects with an eye saccade to the right or to the left(Asaad et al 1998) We found that the activity of 44 ofrandomly selected lateral PFC neurons re ected associ-ations between objects and the saccades that theyinstructed ( gure 8) Other neurons had activity thatre ected the cues or the saccades alone but they werefewer in number Fuster et al (2000) have recently shownthat PFC neurons can also re ect learned associationsbetween visual and auditory stimuli

Importantly these changes do not require a prodigiousamount of experience Changes in PFC neural propertiesare evident after one dayrsquos experience and can be detectedafter just a few minutes of training For example Bichotet al (1996) studied the FEFs part of Brodmannrsquos area 8that is important for voluntary eye movements Normallyneurons in this area re selectively to saccade targetsappearing in certain visual eld locations However whenmonkeys were trained to search for a target de ned by aparticular visual attribute (eg red) the neurons in theFEFs acquire sensitivity to that attribute (Bichot et al

1130 E K Miller and others Concepts and the prefrontal cortex

1996) When monkeys were trained to search for a differ-ent target every day neurons not only discriminated thecurrent target but also distracting stimuli that had been atarget on the previous day relative to stimuli that had beentargets even earlier (Bichot amp Schall 1999) Monkeys werealso more likely to make errors in choosing that distractingstimulus It was as though the previous dayrsquos experienceleft an impression in the brain that in uenced neuralactivity and the monkeyrsquos behaviour

We also observed evidence for rapid plasticity in ourstudy of learning of conditional objectndashsaccade associ-ations in the PFC (Asaad et al 1998) Initially the monk-eys chose their responses at random but learned thecorrect cuendashresponse pairing over a few (5ndash15) trials Asthey learned the association neural activity representingthe forthcoming saccadic response appeared progressivelyearlier in successive trials In other words the initiation ofresponse-related delay activity gradually shifted withlearningmdashfrom a point in time just before the executionof the response and reward delivery to an earlier point intime nearly coincident with the presentation of the cue

Further support for a role for PFC neurons in rep-resenting task demands comes from training monkeys toalternate between different task rules This adds anotherlevel of complexity beyond the conditional tasks describedabove Now there is more than one rule assigned to eachcue and another cue tells the monkey which rule to usein a given trial For example following a given cue monk-eys can learn to direct a response to either the cuersquoslocation (spatial matching rule) or an alternative locationassociated with the cue (associative rule) depending onwhich rule is currently in effect When tested in thisfashion many PFC neurons show rule-speci c activityFor example a PFC neuron might respond to a given vis-ual cue when the monkey is using an associative rule butexhibit weak or no activity under identical sensory andattentional conditions that differed only in that the mon-key was using a spatial rule instead (White amp Wise 1999Asaad et al 2000) Also when monkeys switch betweendifferent tasks many PFC neurons show shifts in baselineactivity that communicates which task is currently beingperformed (Asaad et al 2000 gure 9) Such effects havebeen found in the PFC for associative versus spatial rulesfor object matching versus spatial matching versus associ-ative rules and for shape matching versus object matchingrules (Hoshi et al 1998 White amp Wise 1999 Asaad etal 2000)

In all of these cases however the rules are relativelyliteral or concrete A certain cue or set of cues alwayssignals a speci c response As noted knowledge from ourpast experiences can be applied to a wider range of futurecircumstances if we abstract general principles or rulesrather than speci c cuendashresponse contingencies Theability of the PFC to represent abstract rules those nottied to speci c stimuli or actions was recently addressedin our laboratory (Wallis et al 2001)

Monkeys were trained to use two abstract rules lsquomatchrsquoversus lsquonon-matchrsquo They faced a computer screen andviewed two successively presented pictures If the matchrule was in effect the monkeys released the lever if thepictures were the same and continued to hold the lever ifthe pictures were different If the non-match rule was ineffect the reverse was true monkeys released if the pic-

Phil Trans R Soc Lond B (2002)

tures were different and held if they were the same Therule was randomly instructed in each trial by the presen-tation of a cue at the same time as the rst picture waspresented To disambiguate neuronal responses to thephysical properties of the cue from responses to the rulethat the cues instructed cues signifying the same rule weretaken from different modalities while cues signifying dif-ferent rules were taken from the same modality ( gure10) The monkeys could perform this task well abovechance levels even when they were seeing a stimulus forthe very rst time This indicates that they had abstractedtwo overarching principles of the task that could then beapplied to novel stimulimdashthe minimal de nition of anabstract rule

The most prevalent activity across the PFC was theencoding of the current rule Figure 11 shows a goodexample of a rule-selective neuron that exhibited greateractivity when the match rule was in effect than when thenon-match rule had been indicated This activity cannotbe explained by the physical properties of the cue or thepicture since activity was the same regardless of whichcue was used to instruct the monkey and regardless ofwhich picture the monkey was remembering It cannot berelated to the upcoming response since the monkey didnot know whether the second-presented picture wouldrequire a response Nor could it be related to differencesin reward expectation since the expectation of reward wasthe same regardless of which rule was in effect Further-more the performance of the monkeys was virtually ident-ical for the two types of rules (error rates differed by lessthan 01 and reaction times by less than 7 ms) Thusthe most parsimonious explanation is that the differencesin activity re ected the abstract rule that the monkey wascurrently using to guide its behaviour

What function does the ability to abstract a rule serveIt is a form of generalization that permits a shortcut inlearning thereby allowing the animal to maximize theamount of reward available from a particular situation Toillustrate this consider the above task The monkey couldpotentially solve the task as a series of paired associates(in fact 16 associations consisting of four different pic-tures each paired with four different stimuli) Forexample the monkey might learn that whenever the chefis presented with a drop of juice then at the test phasethe correct response is to choose the chef But notice thatthis type of learning tells the monkey nothing about whichresponse is appropriate to a lion appearing with a drop ofjuice In other words unless the monkey abstracts the rulethat juice indicates that the monkey should match theneach time new pictures are used the monkey would haveto learn an entirely new set of 16 associations by trial anderror The problem with this trial-and-error learning isthat errors are lost opportunities for reward Given thatthe monkeys performed well above chance when theyencountered novel pictures it is clear that they are notengaging in trial-and-error learning but rather haveabstracted two rules that they can then apply as required

This shortcut that abstraction of the rule permits is alsore ected in the neuronal encoding It would be entirelypossible for the monkey to solve the task without singlecells encoding the rule For example there might be twopopulations of cells one encoding the match and non-match rule when the cues are presented in the auditory

Concepts and the prefrontal cortex E K Miller and others 1131

+ juice

+ low tone

or

+ juice

+ high tone

or

match

non-match

Figure 10 Schematic diagram of the abstract rules taskMonkeys switched between choosing a test stimulus that didor did not match the sample depending on whether thematch or non-match rule was in effect The cues thatsignalled these rules are shown on the left

modality and one encoding this information in the tastemodality But such a solution is computationally expens-ive since if a third modality was introduced a third popu-lation of cells would be required It is more ef cient toabstract a rule that cues presented in different modalitiescommonly instruct and indeed this is the solution thatthe brain uses The prevalence of neurons encoding suchrules in the PFC is consistent with the loss of exibilitythat is observed after prefrontal damage in both monkeysand humans It is not inconsistent with studies emphasiz-ing the role of the PFC in working memory or planningbut indicates that an important component of these pro-cesses might be the use of behaviour-guiding rules

5 PREFRONTAL CORTEX AND COGNITIVECONTROL

The results presented above suggest that PFC activityre ects categories and rules This seems consistent withconjectures that a cardinal PFC function may be theacquisition and representation of the formal demands oftasks the guiding concepts and principles that provide afoundation for complex intelligent behaviour In order tounderstand this we must turn to theories of cognitive con-trolmdashthe ability of the brain to coordinate processingamong its millions of neurons in order to direct themtoward future goals

(a) Controlled versus automatic behavioursIn order to understand what we mean by lsquocognitive con-

trolrsquo it is important to understand the distinction betweencontrolled and automatic behaviours Much of our behav-iour is automatic that is direct reactions to our immediateenvironment that do not tax our attention For exampleif someone suddenly throws a baseball at your face youmight re exively duck You may not have willed thisbehaviour it just seems to happen Many such re exiveautomatic processes are lsquowiredrsquo into our nervous systemsby evolution However others can be acquired through agreat deal of experience as learning mechanisms graduallyimprint highly familiar behaviour If you are walking ahighly familiar route and traf c is light you may traversea great distance (and even negotiate turns) with little

Phil Trans R Soc Lond B (2002)

awareness of having done so In these cases your behav-iour is driven in a lsquobottom-uprsquo fashion largely determinedby stimuli in the immediate environment and their abilityto trigger behaviours with which they are strongly associa-ted In neural terms they are dependent on well-established neural pathways waiting to be red off by thecorrect input

However if something unexpected happens on yourwalk you need to lsquotake chargersquo of your actions You payattention to your surroundings and try to anticipate andaccommodate the action of others you may even decideto take an alternative route In this case your behaviour isnot governed by simple inputndashoutput stimulusndashresponserelationships You use your knowledge of the world thecurrent objective (arriving at work intact) and results fromprevious experiences to weigh the alternatives and conse-quences Because these behaviours tax your attention andseem to be driven by lsquointernalrsquo information (knowledgeabout unseen goals and how to achieve them) and thusare initiated by us and not by the environment they arereferred to as lsquocontrolledrsquo The same basic sensory mem-ory and motor processes that mediated automatic behav-iour can be engaged However now they are not simplytriggered by the environment they are shaped and con-trolled in a top-down fashion by our knowledge of howthe world works The observations that PFC neuronsencode acquired information about task contingenciescategories and rules suggest it as a source of top-down sig-nals

The observation that humans with PFC damage seemimpulsive makes sense in light of this distinction betweencontrolled and automatic behaviours Without the PFC toprovide top-down signals about expectations of goals andrequired behaviours the patient simply reacts to theirenvironment with whatever behaviours are strongly asso-ciated with the cues that are immediately presentShallice amp Burgess (1991) examined this by using a lsquoshop-ping testrsquo They described frontal lobe-damaged patientswho are able to execute simple routines in which clearsensory cues could elicit a familiar action (eg buy a loafof bread) However they were unable to carry out anerrand that involved organizing a series of such routinesbecause they kept going lsquooff taskrsquo They would forexample enter shops that were irrelevant to the errandjust because they happened to be passing them Anotherexample is utilization behaviour A patient with PFC dam-age will impulsively use items placed in front of them suchas a comb or in (hopefully) one case a urinal It seemsthat the basic elements of behaviour are intact but thatthe patients override prepotent re exive responses tocoordinate behaviour in accord with an unseen goal

A classic test of the ability to learn and follow goal-orientated rules is the Wisconsin Card Sorting Test(Milner 1963) Subjects are instructed to sort cardsaccording to the shape colour or number of symbolsappearing on them They start with one rule (eg colour)and once that is acquired the rule changes until all of thecards have been sorted using all possible rules Normalhumans have little dif culty with this task By contrasthumans with prefrontal damage can learn the rst sortingcriterion (a relatively simple mapping between a stimulusattribute and a response) but then are unable to escape itthey cannot override the previous behaviour and do not

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

1130 E K Miller and others Concepts and the prefrontal cortex

1996) When monkeys were trained to search for a differ-ent target every day neurons not only discriminated thecurrent target but also distracting stimuli that had been atarget on the previous day relative to stimuli that had beentargets even earlier (Bichot amp Schall 1999) Monkeys werealso more likely to make errors in choosing that distractingstimulus It was as though the previous dayrsquos experienceleft an impression in the brain that in uenced neuralactivity and the monkeyrsquos behaviour

We also observed evidence for rapid plasticity in ourstudy of learning of conditional objectndashsaccade associ-ations in the PFC (Asaad et al 1998) Initially the monk-eys chose their responses at random but learned thecorrect cuendashresponse pairing over a few (5ndash15) trials Asthey learned the association neural activity representingthe forthcoming saccadic response appeared progressivelyearlier in successive trials In other words the initiation ofresponse-related delay activity gradually shifted withlearningmdashfrom a point in time just before the executionof the response and reward delivery to an earlier point intime nearly coincident with the presentation of the cue

Further support for a role for PFC neurons in rep-resenting task demands comes from training monkeys toalternate between different task rules This adds anotherlevel of complexity beyond the conditional tasks describedabove Now there is more than one rule assigned to eachcue and another cue tells the monkey which rule to usein a given trial For example following a given cue monk-eys can learn to direct a response to either the cuersquoslocation (spatial matching rule) or an alternative locationassociated with the cue (associative rule) depending onwhich rule is currently in effect When tested in thisfashion many PFC neurons show rule-speci c activityFor example a PFC neuron might respond to a given vis-ual cue when the monkey is using an associative rule butexhibit weak or no activity under identical sensory andattentional conditions that differed only in that the mon-key was using a spatial rule instead (White amp Wise 1999Asaad et al 2000) Also when monkeys switch betweendifferent tasks many PFC neurons show shifts in baselineactivity that communicates which task is currently beingperformed (Asaad et al 2000 gure 9) Such effects havebeen found in the PFC for associative versus spatial rulesfor object matching versus spatial matching versus associ-ative rules and for shape matching versus object matchingrules (Hoshi et al 1998 White amp Wise 1999 Asaad etal 2000)

In all of these cases however the rules are relativelyliteral or concrete A certain cue or set of cues alwayssignals a speci c response As noted knowledge from ourpast experiences can be applied to a wider range of futurecircumstances if we abstract general principles or rulesrather than speci c cuendashresponse contingencies Theability of the PFC to represent abstract rules those nottied to speci c stimuli or actions was recently addressedin our laboratory (Wallis et al 2001)

Monkeys were trained to use two abstract rules lsquomatchrsquoversus lsquonon-matchrsquo They faced a computer screen andviewed two successively presented pictures If the matchrule was in effect the monkeys released the lever if thepictures were the same and continued to hold the lever ifthe pictures were different If the non-match rule was ineffect the reverse was true monkeys released if the pic-

Phil Trans R Soc Lond B (2002)

tures were different and held if they were the same Therule was randomly instructed in each trial by the presen-tation of a cue at the same time as the rst picture waspresented To disambiguate neuronal responses to thephysical properties of the cue from responses to the rulethat the cues instructed cues signifying the same rule weretaken from different modalities while cues signifying dif-ferent rules were taken from the same modality ( gure10) The monkeys could perform this task well abovechance levels even when they were seeing a stimulus forthe very rst time This indicates that they had abstractedtwo overarching principles of the task that could then beapplied to novel stimulimdashthe minimal de nition of anabstract rule

The most prevalent activity across the PFC was theencoding of the current rule Figure 11 shows a goodexample of a rule-selective neuron that exhibited greateractivity when the match rule was in effect than when thenon-match rule had been indicated This activity cannotbe explained by the physical properties of the cue or thepicture since activity was the same regardless of whichcue was used to instruct the monkey and regardless ofwhich picture the monkey was remembering It cannot berelated to the upcoming response since the monkey didnot know whether the second-presented picture wouldrequire a response Nor could it be related to differencesin reward expectation since the expectation of reward wasthe same regardless of which rule was in effect Further-more the performance of the monkeys was virtually ident-ical for the two types of rules (error rates differed by lessthan 01 and reaction times by less than 7 ms) Thusthe most parsimonious explanation is that the differencesin activity re ected the abstract rule that the monkey wascurrently using to guide its behaviour

What function does the ability to abstract a rule serveIt is a form of generalization that permits a shortcut inlearning thereby allowing the animal to maximize theamount of reward available from a particular situation Toillustrate this consider the above task The monkey couldpotentially solve the task as a series of paired associates(in fact 16 associations consisting of four different pic-tures each paired with four different stimuli) Forexample the monkey might learn that whenever the chefis presented with a drop of juice then at the test phasethe correct response is to choose the chef But notice thatthis type of learning tells the monkey nothing about whichresponse is appropriate to a lion appearing with a drop ofjuice In other words unless the monkey abstracts the rulethat juice indicates that the monkey should match theneach time new pictures are used the monkey would haveto learn an entirely new set of 16 associations by trial anderror The problem with this trial-and-error learning isthat errors are lost opportunities for reward Given thatthe monkeys performed well above chance when theyencountered novel pictures it is clear that they are notengaging in trial-and-error learning but rather haveabstracted two rules that they can then apply as required

This shortcut that abstraction of the rule permits is alsore ected in the neuronal encoding It would be entirelypossible for the monkey to solve the task without singlecells encoding the rule For example there might be twopopulations of cells one encoding the match and non-match rule when the cues are presented in the auditory

Concepts and the prefrontal cortex E K Miller and others 1131

+ juice

+ low tone

or

+ juice

+ high tone

or

match

non-match

Figure 10 Schematic diagram of the abstract rules taskMonkeys switched between choosing a test stimulus that didor did not match the sample depending on whether thematch or non-match rule was in effect The cues thatsignalled these rules are shown on the left

modality and one encoding this information in the tastemodality But such a solution is computationally expens-ive since if a third modality was introduced a third popu-lation of cells would be required It is more ef cient toabstract a rule that cues presented in different modalitiescommonly instruct and indeed this is the solution thatthe brain uses The prevalence of neurons encoding suchrules in the PFC is consistent with the loss of exibilitythat is observed after prefrontal damage in both monkeysand humans It is not inconsistent with studies emphasiz-ing the role of the PFC in working memory or planningbut indicates that an important component of these pro-cesses might be the use of behaviour-guiding rules

5 PREFRONTAL CORTEX AND COGNITIVECONTROL

The results presented above suggest that PFC activityre ects categories and rules This seems consistent withconjectures that a cardinal PFC function may be theacquisition and representation of the formal demands oftasks the guiding concepts and principles that provide afoundation for complex intelligent behaviour In order tounderstand this we must turn to theories of cognitive con-trolmdashthe ability of the brain to coordinate processingamong its millions of neurons in order to direct themtoward future goals

(a) Controlled versus automatic behavioursIn order to understand what we mean by lsquocognitive con-

trolrsquo it is important to understand the distinction betweencontrolled and automatic behaviours Much of our behav-iour is automatic that is direct reactions to our immediateenvironment that do not tax our attention For exampleif someone suddenly throws a baseball at your face youmight re exively duck You may not have willed thisbehaviour it just seems to happen Many such re exiveautomatic processes are lsquowiredrsquo into our nervous systemsby evolution However others can be acquired through agreat deal of experience as learning mechanisms graduallyimprint highly familiar behaviour If you are walking ahighly familiar route and traf c is light you may traversea great distance (and even negotiate turns) with little

Phil Trans R Soc Lond B (2002)

awareness of having done so In these cases your behav-iour is driven in a lsquobottom-uprsquo fashion largely determinedby stimuli in the immediate environment and their abilityto trigger behaviours with which they are strongly associa-ted In neural terms they are dependent on well-established neural pathways waiting to be red off by thecorrect input

However if something unexpected happens on yourwalk you need to lsquotake chargersquo of your actions You payattention to your surroundings and try to anticipate andaccommodate the action of others you may even decideto take an alternative route In this case your behaviour isnot governed by simple inputndashoutput stimulusndashresponserelationships You use your knowledge of the world thecurrent objective (arriving at work intact) and results fromprevious experiences to weigh the alternatives and conse-quences Because these behaviours tax your attention andseem to be driven by lsquointernalrsquo information (knowledgeabout unseen goals and how to achieve them) and thusare initiated by us and not by the environment they arereferred to as lsquocontrolledrsquo The same basic sensory mem-ory and motor processes that mediated automatic behav-iour can be engaged However now they are not simplytriggered by the environment they are shaped and con-trolled in a top-down fashion by our knowledge of howthe world works The observations that PFC neuronsencode acquired information about task contingenciescategories and rules suggest it as a source of top-down sig-nals

The observation that humans with PFC damage seemimpulsive makes sense in light of this distinction betweencontrolled and automatic behaviours Without the PFC toprovide top-down signals about expectations of goals andrequired behaviours the patient simply reacts to theirenvironment with whatever behaviours are strongly asso-ciated with the cues that are immediately presentShallice amp Burgess (1991) examined this by using a lsquoshop-ping testrsquo They described frontal lobe-damaged patientswho are able to execute simple routines in which clearsensory cues could elicit a familiar action (eg buy a loafof bread) However they were unable to carry out anerrand that involved organizing a series of such routinesbecause they kept going lsquooff taskrsquo They would forexample enter shops that were irrelevant to the errandjust because they happened to be passing them Anotherexample is utilization behaviour A patient with PFC dam-age will impulsively use items placed in front of them suchas a comb or in (hopefully) one case a urinal It seemsthat the basic elements of behaviour are intact but thatthe patients override prepotent re exive responses tocoordinate behaviour in accord with an unseen goal

A classic test of the ability to learn and follow goal-orientated rules is the Wisconsin Card Sorting Test(Milner 1963) Subjects are instructed to sort cardsaccording to the shape colour or number of symbolsappearing on them They start with one rule (eg colour)and once that is acquired the rule changes until all of thecards have been sorted using all possible rules Normalhumans have little dif culty with this task By contrasthumans with prefrontal damage can learn the rst sortingcriterion (a relatively simple mapping between a stimulusattribute and a response) but then are unable to escape itthey cannot override the previous behaviour and do not

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

Concepts and the prefrontal cortex E K Miller and others 1131

+ juice

+ low tone

or

+ juice

+ high tone

or

match

non-match

Figure 10 Schematic diagram of the abstract rules taskMonkeys switched between choosing a test stimulus that didor did not match the sample depending on whether thematch or non-match rule was in effect The cues thatsignalled these rules are shown on the left

modality and one encoding this information in the tastemodality But such a solution is computationally expens-ive since if a third modality was introduced a third popu-lation of cells would be required It is more ef cient toabstract a rule that cues presented in different modalitiescommonly instruct and indeed this is the solution thatthe brain uses The prevalence of neurons encoding suchrules in the PFC is consistent with the loss of exibilitythat is observed after prefrontal damage in both monkeysand humans It is not inconsistent with studies emphasiz-ing the role of the PFC in working memory or planningbut indicates that an important component of these pro-cesses might be the use of behaviour-guiding rules

5 PREFRONTAL CORTEX AND COGNITIVECONTROL

The results presented above suggest that PFC activityre ects categories and rules This seems consistent withconjectures that a cardinal PFC function may be theacquisition and representation of the formal demands oftasks the guiding concepts and principles that provide afoundation for complex intelligent behaviour In order tounderstand this we must turn to theories of cognitive con-trolmdashthe ability of the brain to coordinate processingamong its millions of neurons in order to direct themtoward future goals

(a) Controlled versus automatic behavioursIn order to understand what we mean by lsquocognitive con-

trolrsquo it is important to understand the distinction betweencontrolled and automatic behaviours Much of our behav-iour is automatic that is direct reactions to our immediateenvironment that do not tax our attention For exampleif someone suddenly throws a baseball at your face youmight re exively duck You may not have willed thisbehaviour it just seems to happen Many such re exiveautomatic processes are lsquowiredrsquo into our nervous systemsby evolution However others can be acquired through agreat deal of experience as learning mechanisms graduallyimprint highly familiar behaviour If you are walking ahighly familiar route and traf c is light you may traversea great distance (and even negotiate turns) with little

Phil Trans R Soc Lond B (2002)

awareness of having done so In these cases your behav-iour is driven in a lsquobottom-uprsquo fashion largely determinedby stimuli in the immediate environment and their abilityto trigger behaviours with which they are strongly associa-ted In neural terms they are dependent on well-established neural pathways waiting to be red off by thecorrect input

However if something unexpected happens on yourwalk you need to lsquotake chargersquo of your actions You payattention to your surroundings and try to anticipate andaccommodate the action of others you may even decideto take an alternative route In this case your behaviour isnot governed by simple inputndashoutput stimulusndashresponserelationships You use your knowledge of the world thecurrent objective (arriving at work intact) and results fromprevious experiences to weigh the alternatives and conse-quences Because these behaviours tax your attention andseem to be driven by lsquointernalrsquo information (knowledgeabout unseen goals and how to achieve them) and thusare initiated by us and not by the environment they arereferred to as lsquocontrolledrsquo The same basic sensory mem-ory and motor processes that mediated automatic behav-iour can be engaged However now they are not simplytriggered by the environment they are shaped and con-trolled in a top-down fashion by our knowledge of howthe world works The observations that PFC neuronsencode acquired information about task contingenciescategories and rules suggest it as a source of top-down sig-nals

The observation that humans with PFC damage seemimpulsive makes sense in light of this distinction betweencontrolled and automatic behaviours Without the PFC toprovide top-down signals about expectations of goals andrequired behaviours the patient simply reacts to theirenvironment with whatever behaviours are strongly asso-ciated with the cues that are immediately presentShallice amp Burgess (1991) examined this by using a lsquoshop-ping testrsquo They described frontal lobe-damaged patientswho are able to execute simple routines in which clearsensory cues could elicit a familiar action (eg buy a loafof bread) However they were unable to carry out anerrand that involved organizing a series of such routinesbecause they kept going lsquooff taskrsquo They would forexample enter shops that were irrelevant to the errandjust because they happened to be passing them Anotherexample is utilization behaviour A patient with PFC dam-age will impulsively use items placed in front of them suchas a comb or in (hopefully) one case a urinal It seemsthat the basic elements of behaviour are intact but thatthe patients override prepotent re exive responses tocoordinate behaviour in accord with an unseen goal

A classic test of the ability to learn and follow goal-orientated rules is the Wisconsin Card Sorting Test(Milner 1963) Subjects are instructed to sort cardsaccording to the shape colour or number of symbolsappearing on them They start with one rule (eg colour)and once that is acquired the rule changes until all of thecards have been sorted using all possible rules Normalhumans have little dif culty with this task By contrasthumans with prefrontal damage can learn the rst sortingcriterion (a relatively simple mapping between a stimulusattribute and a response) but then are unable to escape itthey cannot override the previous behaviour and do not

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

1132 E K Miller and others Concepts and the prefrontal cortex

time from sample onset (ms)

firi

ng r

ate

(Hz)

(a) (b)

(c) (d)

(a) (b)

(c) (d)

sample delay

sample delay

sample delay

sample delay

50

40

30

20

10

0

50

40

30

20

10

02000150010005000_500 2000150010005000_500

2000150010005000_500

50

40

30

20

10

02000150010005000_500

50

40

30

20

10

0

time from sample onset (ms)

Figure 11 A neuron exhibiting rule selectivity The neuron shows greater activity during match trials regardless of which cuesigni ed the rule or which object was remembered (a) Sample object 1 (b) sample object 2 (c) sample object 3 (c) sampleobject 4 The vertical grey bar marks the cue epoch Match (juice) and match (low tone) are represented by thick black andgrey bars respectively Non-match (juice) and non-match (high tone) are represented by thin black and grey bars respectively

realize that when the rule changes and they must learn anew one (Milner 1963) Monkeys with PFC lesions areimpaired in similar tasks (Dias et al 1996) PFC damagein humans or monkeys or disconnecting the PFC fromits sensory inputs in monkeys also produces de cits ina standard test of rule learning the aforementionedconditional learning tasks (Petrides 1985ab Gaffan ampHarrison 1988 1991 Murray amp Wise 1997 Parker ampGaffan 1998a Murray et al 2000)

(b) A theory of the prefrontal cortex and cognitivecontrol

In summary we have seen that PFC damage seems todisrupt cognitive control the ability of animals to directaction toward unseen goals and leaves them at the mercyof the environment We discussed neurophysiologicalstudies that indicate that the PFC represents task-relevantknowledge such as categories and rules How theseproperties are acquired and how they are used for cogni-tive control have been addressed in a model of PFC byMiller and Cohen (Miller amp Cohen 2001)

In this view the ability to form representations of theformal demands of behaviour stems from the position ofthe PFC at the top of the cortical hierarchy (Fuster 1995)The PFC is a network of neural circuits that is intercon-nected with cortical regions that analyse virtually all typesof sensory inputs and with regions involved in generatingmotor outputs It is also in direct contact with a wide array

Phil Trans R Soc Lond B (2002)

of subcortical structures that process among other thingslsquointernalrsquo information such as motivational state The PFCthus provides a venue in which information from distantbrain systems can interact through relatively local cir-cuitry During learning reward-related signals could acton the PFC to strengthen pathwaysmdashthe associativelinksmdashbetween the neurons that processed the infor-mation that led to a reward As a result the PFC rapidlyconstructs a pattern of activity that represents goals andthe means to achieve them This in essence is a represen-tation of the logic of the task a task model that re ectsthe constellation of relevant information and their inter-relations In neural terms this could amount to a lsquomaprsquoof the neural pathways in the brain that is needed to solvethe task Cognitive control results from the PFC sendingexcitatory signals from this representation back to thebrain structures that provide the PFC with input Thesesignals arise from the ability of many PFC neurons to sus-tain their activity These chronic signals re ecting the taskdemands can enhance the activity of neurons that processtask-relevant information (that match the model) in otherbrain systems and thereby select the forebrain neural path-ways that are needed to solve the task at hand

To understand how this selection takes place considervisual attention In the visual system neurons processingdifferent aspects of the visual scene compete with eachother for activation by mutually inhibiting one anotherThis is thought to be important for enhancing contrast

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

Concepts and the prefrontal cortex E K Miller and others 1133

C1 R1

R2

C2

C3

PFC

Figure 12 Schematic diagram of a posited role for the PFCin cognitive control Information on sensory inputs currentmotivational state memories etc (eg lsquocuesrsquo such as C1C2 and C3) as well as information about behaviour (eglsquoresponsesrsquo such as R1 and R2) is indicated Reward signalsfoster the formation of a task model a neural representationthat re ects the learned associations between task-relevantinformation A subset of the information (eg C1 and C2)can then evoke the entire model including informationabout the appropriate response (eg R1) Excitatory signalsfrom the PFC feed back to other brain systems to enabletask-relevant neural pathways Thick lines indicate activatedpathways thin lines indicate inactive pathways

and separating a gure from its background The neuronsthat lsquowinrsquo the competition and remain active are those thatincur a higher level of activity The biased competitionmodel posits that visual attention exploits this circuitry(Desimone amp Duncan 1995) In voluntary shifts of atten-tion a competitive advantage comes from excitatory sig-nals (thought to originate from the PFC) that representthe expected stimulus These excitatory signals enhancethe activity of neurons in the visual cortex that process thatstimulus and by virtue of the mutual inhibition suppressactivity of neurons processing other stimuli This notionof excitatory bias signals that resolve local competition canbe extended from visual attention to cognitive control ingeneral (Miller 1999 2000) By enhancing the activity ofneurons representing task-relevant information those rep-resenting irrelevant information are simultaneously sup-pressed and neural activity is steered down the pathwaysneeded to solve the task at hand

For an illustrative example of how this might work con-sider the cartoon shown in gure 12 Processing units areshown that correspond to cues (C1 C2 C3) They canbe thought of as neural representations of sensory eventsinternal states stored memories etc in correspondingneural systems Also shown are processing units that cor-respond to the motor circuits mediating two responses (R1and R2) We have set up the sort of exible situation forwhich the PFC is thought to be important Namely onecue (C1) can lead to one of two responses (R1 or R2)depending on the situation (C2 or C3) Imagine that yousuddenly decide that you want a beer (and let us considerthat to be cue C1) If you are at home (C2) then you getup and get one (R1) But if you are in a pub (C3) youask for one instead (R2) These conditional associationsform the lsquoifndashthenrsquo rules that are fundamental buildingblocks of voluntary behaviour (Passingham 1993) Howdoes the PFC construct these representations

Phil Trans R Soc Lond B (2002)

In an unfamiliar situation information ows into thePFC relatively unchecked But then reward-related signalsfrom successful (rewarded) experiences foster the forma-tion of associations between the PFC neurons that hadprocessed the information immediately preceding rewardThis signal may be an in ux of dopamine from the mid-brain VTA neurons that are sensitive to reward andthrough the basal ganglia in uence the PFC (Passingham1993 Schultz amp Dickinson 2000) As this neural ensemblebecomes established it becomes self-reinforcing It sendssignals back to other brain systems biasing their pro-cessing towards matching information and thus re ningthe inputs to the PFC As learning proceeds reward-related signals from VTA neurons appear progressivelyearlier as they become evoked by the events that rstpredict reward (Schultz amp Montague 1997) Throughrepeated iterations of this process more and more task-relevant information is linked into the PFC represen-tation it lsquobootstrapsrsquo from direct associations with rewardto a multivariate network of associations that can describea complex goal-directed task

Once the PFC representation is established a subset ofthe information (such as the cues) can activate the remain-ing elements (such as the correct response) So if we wanta beer (C1) and we are at home (C2) the correspondingPFC representation containing the correct response (R1)is activated and sustained until the response is executedThe resulting excitatory bias signals from the PFC thenfeed back to other brain regions selecting the appropriatepathway needed for the task (eg C1ndashR1) A different pat-tern of cues (eg cues 1 and 3 in gure 12) evokes a differ-ent PFC model and a different pattern of bias signalsselects other neural pathways (C1ndashR2) With repeatedselection of these pathways they can become establishedindependently of the PFC As this happens the PFCbecomes less involved and the behaviour becomes habitualor automatic

This particular view of PFC function is not without peeror precedent Fuster rst proposed that PFC neuronsencode task-relevant contingencies between stimuli andorresponses particularly when they are separated by gapsin time (as so often happens with extended goal-directedbehaviours) (Fuster 1985) Neurons that explicitly encodetask contingencies and rules are used in neural networkmodels of cognitive control by Changeux and Dehaene(Changeux amp Dehaene 1993 Dehaene et al 1998) Themodels of Cohen and colleagues use a layer (thought tocorrespond to the PFC) that represents task demands orlsquocontextrsquo (Cohen amp Servan-Schreiber 1992) Wise et al(1996) proposed that a cardinal PFC function is theacquisition of behaviour-guiding rules Shimamura inde-pendently proposed a role for the PFC directly analogousto the Miller and Cohen model (Shimamura 2000)

Central to our model and indeed all physiologicallyinspired models of PFC function is the ability of PFCneurons to sustain their activity for several seconds in theabsence of further stimulation This is crucial for severalreasons As previously mentioned gaps in time are aninevitable consequence of extended goal-directed behav-iours Thus sustained activity allows PFC neurons tolearn relationships (associations) between stimuli andorresponses that are separated in time (Fuster 1985) It alsoallows task rules to be maintained until the task is com-

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

1134 E K Miller and others Concepts and the prefrontal cortex

pleted But conveying information by sustained activationaffords more than the ability to bridge temporal gapsOrsquoReilly and Munakata have pointed out that it is an idealformat for transmitting the knowledge needed for cogni-tive control to other brain systems (OrsquoReilly amp Munak-ata 2000)

A tenet of modern neuroscience is that long-term stor-age in the brain depends on strengthening some neuralconnections and weakening others that is by changingsynaptic weights Encoding information in structuralchanges has obvious advantages for information storageit allows for very long-term memories But the resultingmemories are relatively in exible once a neural circuit isestablished it will tend to re in the same way every timeit is triggered Also changing the strength of a synapseonly affects the neurons that share the synapse Thus theinformation encoded in a pattern of synaptic weights onlyaffects the ring of the neurons that form that particularcircuit and is only expressed when that circuit is redCognitive control however requires that a given patternof information (the task demands) affect many brain cir-cuits it is used to orchestrate processing in many differentbrain systems It is apparent therefore that the infor-mation needs to be encoded in a different format Sus-tained activity is such a format Because information isencoded in a pattern of sustained activity (rather than onlyin a pattern of synaptic weights) it can be propagatedacross the brain Thus the ability of sustained activity totonically in uence other brain systems is probablyimportant for coordinating diverse processing around aspeci c goal It also affords exibility if cognitive controlstems from a pattern of information maintained in thePFC changing behaviour is as easy as changing the pat-tern (OrsquoReilly amp Munakata 2000 Miller amp Cohen 2001)

Finally the central role of sustained activity mightexplain the severely limited capacity of controlled pro-cesses While we can carry out a number of automatic pro-cesses simultaneously our ability to carry out controlledprocesses is limited by the low capacity of our attentionIf the information for cognitive control is expressed in aunique pattern of ongoing activity distributed across manysimultaneously active neuronsmdasha population codemdashthenthere will be a natural capacity limitation Trying to rep-resent more than just a few items at the same time woulddegrade information because the unique patterns im-pinging on a given set of neurons would overwrite andinterfere with one another

6 SUMMARY AND CONCLUSIONS

The ability to take charge of onersquos actions and directthem towards future unseen goals is called cognitive con-trol Virtually all theories of cognition posit that cognitiondepends on functions specialized for the acquisition ofinformation about goals and the means to achieve themThese functions exert a top-down in uence on the lower-level automatic processes that mediate sensory analysismemory storage and motor outputs orchestrating anddirecting them toward a given goal

The PFC a brain structure that reaches its greatestcomplexity in the primate brain seems to have a centralrole in cognitive control It has access to and the meansto in uence processing in all major forebrain systems and

Phil Trans R Soc Lond B (2002)

can provide a means to synthesize the diverse informationrelated to a given goal As we have established PFC neu-rons seem to have a crucial ability for cognitive controlthey convey the knowledge that animals acquire about agiven goal-directed task Their ability to develop ab-stracted representations frees the organism from speci cassociations and endows it with the ability to generalizeand develop overarching concepts and principles Thisability is consistent with observations of a loss of exibilityafter PFC damage and may form a foundation for thecomplex intelligent behaviour that is often seen in pri-mates

This work has been supported by the National Institutes ofHealth the RIKEN-MIT Neuroscience Research Center thePew Foundation The McKnight Foundation The SloanFoundation and the Whitehall Foundation The authors thankMarlene Wicherski for her valuable comments on the manu-script

REFERENCES

Asaad W F Rainer G amp Miller E K 1998 Neural activityin the primate prefrontal cortex during associative learningNeuron 21 1399ndash1407

Asaad W F Rainer G amp Miller E K 2000 Task-speci cactivity in the primate prefrontal cortex J Neurophysiol 84451ndash459

Baddeley A amp Della Sala S 1996 Working memory andexecutive control Phil Trans R Soc Lond B 351 1397ndash1403

Barbas H amp Pandya D 1991 Patterns of connections of theprefrontal cortex in the rhesus monkey associated withcortical architecture In Frontal lobe function and dysfunction(ed H S Levin H M Eisenberg amp A L Benton) pp 35ndash58 New York Oxford University Press

Beymer D amp Poggio T 1996 Image representations for visuallearning Science 272 1905ndash1909

Bhatt R S Wasserman E A Reynolds W F amp KnaussK S 1988 Conceptual behavior in pigeons categorizationof both familiar and novel examples from four classes ofnatural categories J Exp Psych Anim Behav Process 14219ndash234

Bichot N P amp Schall J D 1999 Effects of similarity and his-tory on neural mechanisms of visual selection Nat Neurosci2 549ndash554

Bichot N P Schall J D amp Thompson K G 1996 Visualfeature selectivity in frontal eye elds induced by experiencein mature macaques Nature 381 697ndash699

Changeux J P amp Dehaene S 1993 Formal models for cogni-tive functions associated with the prefrontal cortex InExploring brain functions models in neuroscience (ed T APoggio amp D A Glaser) Chichester Wiley

Cohen J D amp Servan-Schreiber D 1992 Context cortexand dopamine a connectionist approach to behavior andbiology in schizophrenia Psychol Rev 99 45ndash77

Dehaene S Kerszeberg M amp Changeux J P 1998 A neu-ronal model of a global workspace in effortful cognitive tasksProc Natl Acad Sci USA 95 14 529ndash14 534

Desimone R amp Duncan J 1995 Neural mechanisms of selec-tive visual attention A Rev Neurosci 18 193ndash222

Desimone R Albright T D Gross C G amp Bruce C 1984Stimulus-selective properties of inferior temporal neurons inthe macaque J Neurosci 4 2051ndash2062

Dias R Robbins T W amp Roberts A C 1996 Primate ana-logue of the Wisconsin Card Sorting Test effects of excito-toxic lesions of the prefrontal cortex in the marmoset BehavNeurosci 110 872ndash886

Concepts and the prefrontal cortex E K Miller and others 1135

Dickinson A 1980 Contemporary animal learning theory Cam-bridge University Press

Duncan J Emslie H Williams P Johnson R amp Freer C1996 Intelligence and the frontal lobe the organization ofgoal-directed behavior Cogn Psychol 30 257ndash303

Eacott M J amp Gaffan D 1992 Inferotemporalndashfrontal dis-connectionndashthe uncinate fascicle and visual associative learn-ing in monkeys Eur J Neurosci 4 1320ndash1332

Fabre-Thorpe M Richard G amp Thorpe S J 1998 Rapidcategorization of natural images by rhesus monkeys Neuro-Report 9 303ndash308

Freedman D J Riesenhuber M Poggio T amp Miller E K2001 Categorical representation of visual stimuli in the pri-mate prefrontal cortex Science 291 312ndash316

Fuster J M 1985 The prefrontal cortex mediator of cross-temporal contingencies Hum Neurobiol 4 169ndash179

Fuster J M 1989 The prefrontal cortex New York RavenFuster J M 1995 Memory in the cerebral cortex Cambridge

MA MIT PressFuster J M Bodner M amp Kroger J K 2000 Cross-modal

and cross-temporal association in neurons of frontal cortexNature 405 347ndash351

Gaffan D amp Harrison S 1988 Inferotemporalndashfrontal discon-nection and fornix transection in visuomotor conditionallearning by monkeys Behav Brain Res 31 149ndash163

Gaffan D amp Harrison S 1991 Auditoryndashvisual associationshemispheric specialization and temporalndashfrontal interactionin the rhesus monkey Brain 114 2133ndash2144

Gainotti G 2000 What the locus of brain lesion tells us aboutthe nature of the cognitive defect underlying category-speci c disorders a review Cortex 36 539ndash559

Goldman-Rakic P S 1987 Circuitry of primate prefrontalcortex and regulation of behavior by representational mem-ory In Handbook of physiology the nervous system (ed FPlum) pp 373ndash417 Bethesda MD American PhysiologySociety

Gross C G 1973 Visual functions of inferotemporal cortexIn Handbook of sensory physiology (ed R Jung) BerlinSpringer

Hoshi E Shima K amp Tanji J 1998 Task-dependent selec-tivity of movement-related neuronal activity in the primateprefrontal cortex J Neurophysiol 80 3392ndash3397

Lisker L amp Abramson A 1970 The voicing dimension someexperiments in comparing phonetics Prague Academia

Miller E K 1999 The prefrontal cortex complex neuralproperties for complex behavior Neuron 22 15ndash17

Miller E K 2000 The neural basis of top-down control ofvisual attention in the prefrontal cortex In Attention and per-formance 18 (ed S Monsell amp J Driver) Cambridge MAMIT Press

Miller E K amp Cohen J D 2001 An integrative theory of pre-frontal function A Rev Neurosci 24 167ndash202

Milner B 1963 Effects of different brain lesions on card sort-ing Arch Neurol 9 100ndash110

Mishkin M 1982 A memory system in the monkey PhilTrans R Soc Lond B 298 83ndash95

Murray E A Bussey T J amp Wise S P 2000 Role of pre-frontal cortex in a network for arbitrary visuomotor map-ping Exp Brain Res 133 114ndash129

Murray E A amp Wise S P 1997 Role of the orbitoventralprefrontal cortex in conditional motor learning SocNeurosci Abstr 27 12

Nauta W J H 1971 The problem of the frontal lobe a re-interpretation J Psychiatr Res 8 167ndash187

OrsquoReilly R C amp Munakata Y 2000 Computational explo-rations in cognitive neuroscience understanding the mind Cam-bridge MA MIT Press

Orlov T Yakovlev V Hochstein S amp Zohary E 2000

Phil Trans R Soc Lond B (2002)

Macaque monkeys categorize images by their ordinal num-ber Nature 404 77ndash80

Pandya D N amp Barnes C L 1987 Architecture and connec-tions of the frontal lobe In The frontal lobes revisited (ed EPerecman) pp 41ndash72 New York The IRBN Press

Parker A amp Gaffan D 1998a Memory after frontaltemporaldisconnection in monkeys conditional and non-conditionaltasks unilateral and bilateral frontal lesions Neuropsychologia36 259ndash271

Parker A amp Gaffan D 1998b Memory after frontaltemporaldisconnection in monkeys conditional and non-conditionaltasksunilateral and bilateral frontal lesions Neuropsychologia36 259ndash271

Passingham R 1993 The frontal lobes and voluntary actionOxford University Press

Petrides M 1985a De cits in non-spatial conditional associat-ive learning after periarcuate lesions in the monkey BehavBrain Res 16 95ndash101

Petrides M 1985b De cits on conditional associative-learningtasks after frontal- and temporal-lobe lesions in manNeuropsychologia 23 601ndash614

Petrides M 1990 Nonspatial conditional learning impaired inpatients with unilateral frontal but not unilateral temporallobe excisions Neuropsychologia 28 137ndash149

Rainer G Asaad W F amp Miller E K 1998a Memory eldsof neurons in the primate prefrontal cortex Proc Natl AcadSci USA 95 (15) 15 008ndash15 013

Rainer G Asaad W F amp Miller E K 1998b Selective rep-resentation of relevant information by neurons in the primateprefrontal cortex Nature 393 577ndash579

Rainer G Rao S C amp Miller E K 1999 Prospective codingfor objects in the primate prefrontal cortex J Neurosci 195493ndash5505

Rao S C Rainer G amp Miller E K 1997 Integration of whatand where in the primate prefrontal cortex Science 276821ndash824

Roberts W A amp Mazmanian D S 1988 Concept learning atdifferent levels of abstraction by pigeons monkeys andpeople J Exp Psychol Anim Behav Proc 14 247ndash260

Schultz W amp Dickinson A 2000 Neuronal coding of predic-tion errors A Rev Neurosci 23 473ndash500

Schultz W amp Montague P R 1997 A neural substrate of pre-diction and reward Science 275 1593ndash1599

Shallice T amp Burgess P W 1991 De cits in strategy appli-cation following frontal lobe damage in man Brain 114727ndash741

Shelton C 2000 Morphable surface models Int J Comp Vis38 75ndash91

Shimamura A P 2000 The role of the prefrontal cortex indynamic ltering Psychobiology 28 207ndash218

Tanaka K Saito H Fukada Y amp Moriya M 1991 Codingvisual images of objects in the inferotemporal cortex of themacaque monkey J Neurophysiol 66 170ndash189

Tomita H Ohbayashi M Nakahara K Hasegawa I ampMiyashita Y 1999 Top-down signal from prefrontal cortexin executive control of memory retrieval (see comments)Nature 401 699ndash703

Vaadia E Benson D A Hienz R D amp Goldstein Jr M H1986 Unit study of monkey frontal cortex active localizationof auditory and of visual stimuli J Neurophysiol 56 934ndash952

Vogels R 1999a Categorization of complex visual images byrhesus monkeys Part 1 behavioural study Eur J Neurosci11 1223ndash1238

Vogels R 1999b Categorization of complex visual images byrhesus monkeys Part 2 single-cell study Eur J Neurosci11 1239ndash1255

Wallis J D Anderson K C amp Miller E K 2001 Single neu-

1136 E K Miller and others Concepts and the prefrontal cortex

rons in the prefrontal cortex encode abstract rules Nature411 953ndash956

Watanabe M 1990 Prefrontal unit activity during associativelearning in the monkey Exp Brain Res 80 296ndash309

Watanabe M 1992 Frontal units of the monkey coding theassociative signi cance of visual and auditory stimuli ExpBrain Res 89 233ndash247

White I M amp Wise S P 1999 Rule-dependent neuronalactivity in the prefrontal cortex Exp Brain Res 126 315ndash335

Wise S P Murray E A amp Gerfen C R 1996 The frontalndashbasal ganglia system in primates Crit Rev Neurobiol 10317ndash356

Wyttenbach R A May M L amp Hoy R R 1996 Categoricalperception of sound frequency by crickets Science 2731542ndash1544

Phil Trans R Soc Lond B (2002)

Young M E amp Wasserman E A 1997 Entropy detectionby pigeons response to mixed visual displays after same-different discrimination training J Exp Psychol AnimBehav Process 23 157ndash170

GLOSSARY

DMC delayed match-to-categoryFEF frontal eye eldITC inferior temporal cortexITI inter-trial intervalPFC prefrontal cortexVTA ventral tegmental area