REVIEWS

THE PREFRONTAL CORTEX AND

COGNITIVE CONTROL
Earl K. Miller
One of the enduring mysteries of brain function concerns the process of cognitive control. How
does complex and seemingly wilful behaviour emerge from interactions between millions of
neurons? This has long been suspected to depend on the prefrontal cortex — the neocortex at
the anterior end of the brain — but now we are beginning to uncover its neural basis. Nearly all
intended behaviour is learned and so depends on a cognitive system that can acquire and
implement the ‘rules of the game’ needed to achieve a given goal in a given situation. Studies
indicate that the prefrontal cortex is central in this process. It provides an infrastructure for
synthesizing a diverse range of information that lays the foundation for the complex forms of
behaviour observed in primates.

TOP-DOWN Humans and other animals can do more than reflex- mon internal theme. So a key function of the neural cir-
Brain signals that convey ively react to sensory information that is immediate cuitry mediating cognitive control is to extract the goal-
knowledge derived from prior and salient. We engage in complex and extended relevant features of our experiences for use in future cir-
experience rather than sensory
behaviours geared towards often far-removed goals. To cumstances. It has been proposed that the prefrontal
stimulation.
do so, we have evolved mechanisms that can override cortex — a neocortical region that finds its greatest
or augment reflexive and habitual reactions in order to elaboration in humans — is centrally involved in this
orchestrate behaviour in accord with our intentions. process4–8.
These mechanisms are commonly referred to as ‘cogni- The prefrontal cortex (PFC) (FIG. 1) is an intercon-
tive’ in nature and their function is to control lower- nected set of neocortical areas that have a unique, but
level sensory, memory and/or motor operations for a overlapping, pattern of connectivity with virtually all
common purpose. So cognitive control is essential for sensory neocortical and motor systems and a wide
what we recognize as intelligent behaviour. range of subcortical structures9–12. This provides an
Insight into the neural mechanisms for cognitive ideal infrastructure for synthesizing the diverse range
control may come from what is arguably their most of information needed for complex behaviour. The
important feature: they are sculpted by experience. PFC also has widespread projections back to these sys-
Virtually all intended behaviours are learned and so tems that may allow it to exert a ‘TOP-DOWN’ influence on
depend on a cognitive system that can acquire the rules a wide range of brain processes9–12. Indeed, the effects
of the game — what goals are available and what means of PFC damage are most apparent when cognitive con-
can be used to achieve these goals1–4. Take, for example, trol is most needed — when the knowledge about a
dining in a restaurant. We are not born knowing that given situation must be used to select the appropriate
Center for Learning and
Memory, RIKEN-MIT this can be a rewarding experience or how to act in this goal-directed actions (BOX 1).
Neuroscience Research situation. Instead, our experiences arm us with expecta- Here I review recent neurophisiological studies in
Center, Department of Brain tions about the important sensory information deserv- monkeys that have explored the neural basis of cogni-
and Cognitive Sciences, ing our attention (for example, the wine list), typical tive control. They indicate that a major function of the
Massachusetts Institute of
Technology, Cambridge,
events, appropriate actions and expected consequences PFC is to extract information about the regularities
Massachusetts 02139, USA. (for example, paying the bill). This knowledge allows across experiences and so impart rules that can be used
e-mail: [email protected] diverse brain processes to be orchestrated along a com- to guide thought and action8,13–15.

NATURE REVIEWS | NEUROSCIENCE VOLUME 1 | OCTOBER 2000 | 5 9

© 2000 Macmillan Magazines Ltd
REVIEWS

Posterior parietal cortex Motor structures Many lateral PFC neurons reflect these learned asso-
ciations18,28,29. For example, Watanabe used a set of tasks
in which visual and auditory cues signalled, on different
Auditory cortex trials, whether reward would or would not be deliv-
ered18,28. Most lateral PFC neurons were found to reflect
8 the association between a cue and reward. A given neu-
9 46 ron might be activated by a cue, but only when it sig-
45 nalled ‘reward’. In contrast, another neuron might be
46 activated only by a cue that signalled ‘no reward’.
12 Similarly, we trained monkeys to associate, in different
blocks of trials, each of two cue objects with a SACCADE to
11 13
the right or left29, and found that the activity of 44% of
lateral PFC neurons reflected associations between
objects and the saccades they instructed (FIG. 2). Other
neurons had activity that reflected the cues or the sac-
Medial temporal structures Inferior temporal cortex
cades alone, but they were fewer in number. Fuster and
Figure 1 | Integrative anatomy of the macaque monkey colleagues30 have also shown that PFC neurons can
prefrontal cortex. Numbers refer to sub-regions within reflect learned associations between visual and auditory
prefrontal cotex as defined by Brodmann. Different PFC stimuli.
subregions have unique, but overlapping, patterns of
Striking examples of experience-dependent neural
connections with other brain regions. For example, the more
posterior and dorsal portions of the lateral PFC are more heavily plasticity come from Bichot and Schall’s studies of the
interconnected with cortical areas that emphasize processing of frontal eye fields, part of Brodmann’s area 8 that is
visuospatial and motor information. Ventral and anterior lateral important for voluntary shifts of gaze. Normally, neu-
regions are more heavily interconnected with cortical areas that rons in this area fire selectively to saccade targets appear-
emphasize information about visual form and stimulus identity. ing in certain visual field locations. However, when mon-
The ventral (orbitofrontal) PFC is more associated with
keys were trained to search for a target defined by a
subcortical structures that process ‘internal’ information such
as homeostasis. Above and beyond this regional emphasis,
particular visual attribute (for example, red), the neu-
however, there is also multimodal convergence. Many PFC rons in the frontal eye fields acquired sensitivity to that
areas receive converging inputs from at least two sensory attribute31. Bichot and Schall32 trained monkeys to search
modalities94,95 and there are ample interconnections between for a different target every day and found that neurons
different PFC areas (illustrated by the purple lines) that could not only discriminated the current target, but also dis-
bring together results from a wide range of brain processes. For tracting stimuli that had been a target on the previous
simplicity, this figure only shows a subset of PFC areas and a
subset of their connections. Areas on the medial surface are not
day, relative to stimuli that had been targets even earlier.
GOAL-DIRECTED BEHAVIOUR
Behaviour directed toward shown or discussed in this review. Monkeys were also more likely to make errors in choos-
attainment of a future state (for ing that distracting stimulus. It was as if the previous
example, obtaining a graduate day’s experience left an impression in the brain that
degree). Associations, conjunctions and rules influenced neural activity and task performance.
GOAL-DIRECTED BEHAVIOUR requires predictions about events, But monkeys and humans do more than remember
INTERNAL STATES
INTERNAL STATES and actions that are likely to achieve a goal. simple contingencies. They can discern the regularities
Brain information not directly
related to a sensory input or But to make these predictions, we need to form associa- across them to extract general principles or rules. This is
motor output; for example, tions between their internal representations16. A neural reflected in PFC activity as well. White and Wise21 found
homeostatic information such as ensemble of a task, then, might be composed of neurons that the activity of up to half of PFC neurons depended
hunger, thirst or other
motivational influences.
whose activity reflects learned associative relationships on whether the monkey was guiding its behaviour by a
between these goal-relevant elements, that is, the TASK CON- spatial rule (a cue’s location indicated where the target
TASK CONTINGENCIES TINGENCIES (BOX 2). Prefrontal neurons do have this proper- would appear) or an associative rule (the identity of the
The logical structure of a given ty — they show conjunctive tuning for learned associa- cue indicated the target’s location). Hoshi et al.33 found
task (for example, if the light is
tions between cues, voluntary actions and rewards. that many PFC neurons were modulated by which rule
green, cross the street).
Prefrontal neurons even show tuning for complex, (matching shape or location) the monkey was currently
LIMBIC STRUCTURES behaviour-guiding rules. So they may help form neuron using. We have also observed lateral prefrontal neurons
A collection of subcortical ensembles that represent the regularities across experi- with rule-dependent activity (BOX 3)34,35. These neurons
structures important for ences that describe the principles needed to achieve a par- could correspond to the ‘rule-coding’ units in the mod-
processing memory and
emotional information.
ticular goal in a particular situation. els of Dehaene and Changeux13,36.
Prominent structures include For example, the lateral PFC is directly interconnect- So PFC neurons convey information about the for-
the hippocampus and amygdala. ed with higher-order sensory and motor cortex, and mal demands of tasks, a possible foundation for the
indirectly connected (through the ventromedial PFC) complex forms of behaviour of primates. The mecha-
MULTIMODAL RESPONSES
Neural activity elicited by more with LIMBIC STRUCTURES that process ‘internal’ information nisms that guide the formation of these representations
than one sensory modality. such as reward9–12. The neural activity in the lateral PFC are discussed in the next section.
reflects this — many of its neurons show MULTIMODAL
SACCADE 17–22
A rapid, ballistic eye movement
RESPONSES . Furthermore, the lateral PFC is critical for Reward signals and rule representations
from one point of gaze to normal learning of arbitrary associations between sen- If PFC neural ensembles reflect goal-relevant informa-
another. sory cues, rewards and voluntary actions23–27. tion, their construction is probably guided by reward.

60 | OCTOBER 2000 | VOLUME 1 www.nature.com/reviews/neuroscience

© 2000 Macmillan Magazines Ltd
 REVIEWS

Reward information does have a pervasive influence on

Box 1 | Behavioural effects of prefrontal cortex damage
PFC activity — activity in the lateral PFC and ventro-
Humans with prefrontal damage can seem strikingly normal upon superficial medial PFC conveys the identity and size of expected
examination. They can carry on a conversation, often have normal IQ scores and can rewards18,28,37–39. A major source of reward-related sig-
perform familiar routines without difficulty. However, despite their good performance nals may be the dopamine-mediated innervation of
on standard neuropsychological tests of perceptual, memory and motor skills, their the PFC from a group of cells situated in the ventral
ability to organize their lives is profoundly impaired. They are impulsive and irresponsible tegmental area(VTA) of the midbrain.
and consequently can have trouble holding a job, remaining married and so on. Careful VTA neurons have properties that are ideal for pro-
testing has revealed that the behaviour of humans and monkeys with prefrontal damage viding a signal that guides acquisition of goal-relevant
can be described as stimulus-bound. Their behaviour is captured by salient sensory cues information. Initially, they give a burst of activity to
that reflexively elicit strongly associated actions. They are unable to override these unpredicted rewards40,41. With experience, they become
impulses to engage in behaviours that depend on knowledge of a goal and the means to activated by cues that predict reward and not by the
achieve it, that is, behaviours that are weakly established, complex, changing, or that
rewards themselves42. These neural responses that have
must be extended over time4,7,10,11,86. For example, consider a classic test of prefrontal
been transferred to the cues also wane with further
impairment, the Wisconsin Card Sorting Task. Subjects are instructed to sort cards
training, perhaps because they transfer to environmen-
according to the shape, colour or number of symbols appearing on them. They start
tal cues that are earlier predictors of reward43. VTA neu-
with one rule (for example, colour) and, once that is acquired, the experimenter changes
rons are also inhibited when an expected reward is
the rule (for example, shape) without telling the subject. Rules are acquired and changed
until all the cards have been sorted using all possible rules. Normal people have little
withheld44. This codes the degree to which a reward, or
difficulty with this task. In contrast, people with prefrontal damage can learn the first a cue that predicts reward, is surprising. As the aim of
rule but then they are unable to escape it: they make a great deal of errors because they the organism is to predict the means to achieve reward,
lapse back to the earlier rule91. The ability of monkeys with PFC lesions to perform an this ‘prediction error’ indicates when the associative
analogue of this task is also impaired92. Shallice and Burgess described patients with learning that underlies this ability should occur45.
damage to the frontal lobes who are able to execute simple routines in which clear The resulting dopamine influx into the PFC could
sensory cues could elicit a familiar action (for example,‘buy a loaf of bread’)93. However, affect plasticity through several plausible mechanisms.
they were unable to carry out an errand that involved organizing a series of such For example, dopamine could augment NMDA (N-
routines. They would, for example, enter shops that were irrelevant to the errand. In methyl-D-aspartate) receptor-mediated glutamatergic
these cases, the basic elements of behaviour are intact but it seems that they are missing transmission, which has been directly implicated in
the flexibility to shift between different rules and so override PREPOTENT RESPONSES to plasticity46. Dopamine may also help augment and
persist toward a goal. Here, I suggest that the PFC allows for this flexibility by sustain PFC activity47,48, allowing activity-dependent
dynamically establishing task-relevant neural pathways in other brain systems (BOX 2). plasticity mechanisms to work.

Box 2 | A suggested role for the prefrontal cortex in cognitive control

The figure shows processing units representing
cues such as sensory inputs, current motivational
state, memories and so on (C1, C2 and C3); units Prefrontal
cortex
representing two voluntary actions (for example,
‘responses’ R1 and R2); and internal or ‘hidden’
units representing intervening stages of
processing. The PFC is shown as being connected
C2
to the hidden units because it is interconnected
with higher-order ‘association’ and premotor
C3
cortices, not with primary sensory or motor
cortices. A situation in which the PFC seems
C1 R1
particularly important is pictured here: when the
same cue (C1) could lead to one or another
R2
response (R1 or R2) depending on some other
item of information (C2 or C3). For example, if
the phone rings (C1) and you are at home (C2), you answer it (that is, C1…R1). But if the phone rings (C1) and you
are a guest in someone else’s home (C3), you do not (C1…R2). During learning, reward signals may strengthen the
connections between PFC neurons that process the information that leads to reward, resulting in a pattern of activity
that reflects the pattern of associations between goal-relevant information that is unique to each situation (that is, the
task contingencies). Once established, a subset of the information (for example, C1 and C2) can activate the entire
representation (for example, the constellation of PFC ‘units’ shown in red), including information about the
appropriate response (for example, R1). Bias signals from the PFC task representations may then select task-relevant
neural pathways in other brain systems (for example, C1–R1). A different set of cues (C1 and C3) would activate a
different PFC representation (shown in blue) and, consequently, a different pattern of bias signals selects a different set
of neural pathways (C1–R2). By providing a bias signal to the intermediate (hidden) units, the PFC favours the
PREPOTENT RESPONSES pathways in the posterior neocortex and other brain areas that are appropriate for the task. So task-relevant pathways
Reflexive actions, either innate can be dynamically and flexibly established because they depend on the current pattern of PFC activity. A loss of
or well established through a flexibility is a hallmark of PFC damage (BOX 1).
great deal of experience.

NATURE REVIEWS | NEUROSCIENCE VOLUME 1 | OCTOBER 2000 | 6 1

© 2000 Macmillan Magazines Ltd
REVIEWS

35 20 attention55 (BOX 1). This may reflect the loss of mecha-

nisms for maintaining goal-relevant information, a
Spikes per second

Spikes per second

30 process known as ‘WORKING MEMORY’. This has been
15
explored in a variety of neurophysiological studies in
25 monkeys19,20,49–53.
10 Many cortical areas seem to have some sort of short-
20 term buffering ability. What sets working memory apart
as being more ‘cognitive’ is that it can retain information
15 5 over potentially distracting events. PFC neurons do have
Left Right Left Right this ability. For example, when monkeys are required to
20 45 sustain the memory of a sample object across a delay
period filled with visual distractors that each require
attention and processing, sustained activity within the
Spikes per second

Spikes per second

15 40 PFC acts to maintain the sample memory52. In contrast,
sustained activity in extrastriate visual areas seems to be
more easily disrupted by the presence of distractors —
10 35
following presentation of a distractor, neural activity in
the INFERIOR TEMPORAL CORTEX and POSTERIOR PARIETAL CORTEX
5 30
no longer reflected the sample object that the monkey
Left Right Left Right was retaining in memory52,56,57.
Figure 2 | Conjunctive tuning in the prefrontal cortex. The activity of four single PFC neurons How does the PFC ‘latch’ onto goal-relevant informa-
when each of two objects instructed either a saccade to the right or a saccade to the left on tion and maintain it without disruption? Several models
different trials. The lines connect the average values obtained when a given object cued one or the have been suggested48,58–61. They typically use a form of a
other saccade. The error bars show the standard error of the mean. Note that in each case, the gating signal that instructs the network when to main-
neuron’s activity depends on both the cue object and the saccade direction and that the tuning is tain a given activity state. Dopamine influx into the PFC
nonlinear or conjunctive. That is, the level of activity to a given combination of object and saccade
may again be involved. Its neuromodulatory effects
cannot be predicted from the neuron’s response to the other combinations. (Adapted from REF. 29.)
could strengthen current representations, protecting
them against interference from disruption by irrelevant,
PFC activity elicited by a transient event can be sus- distracting information until another dopamine influx
tained for many seconds19,20,49–53. This allows PFC neu- reinforces another representation48,59,61.
rons to form associations between events separated in
time29,30. When the dopamine influx reaches the PFC, it Bias signals and top-down control
could strengthen connections — associative links — The ability to sustain task information is of little use
between neurons that were activated by the event that unless the PFC can somehow use it to control process-
elicited the midbrain dopamine burst and the event that ing in other brain systems. PFC activity could exert a
preceded it. Iteration of this process could drive the pro- top-down influence by providing an excitatory signal
gressively earlier generation of the dopamine signal that biases processing in other brain systems towards
from the VTA neurons54. During learning, as dopamine task-relevant information. To understand how this
arrives progressively earlier, more and more informa- might work, consider selective visual attention. In the
tion could be linked into an increasingly multivariate visual system, neurons processing different aspects of
PFC representation that will ultimately describe the the visual scene compete with each other for activation.
constellation of goal-relevant task features. This is thought to be important for enhancing contrast
A possible neural correlate of this phenomenon was and separating objects from the background. The neu-
observed29 in an experiment in which we recorded rons that ‘win’ the competition and remain active are
neural activity from the lateral PFC of monkeys learn- those that incur a higher level of activity. The biased
ing associations between each of two cue objects and competition model proposes that visual attention
WORKING MEMORY each of two saccadic eye movements. As the monkeys exploits this circuitry62. In voluntary shifts of attention a
The representation of items held learned, neural activity reflecting the forthcoming sac- competitive advantage is conferred by excitatory signals
in consciousness during cadic response appeared progressively earlier in the trial (thought to originate from the PFC) that represent the
experiences or after retrieval of
(FIG. 3). The initiation of saccade-related activity shifted ‘to be attended’ stimulus. These excitatory signals
memories. Short-lasting and
associated with active rehearsal with learning, from just before the execution of the sac- enhance the activity of neurons in the visual cortex that
or manipulation of information. cade (and acquisition of reward) to an earlier point in process that stimulus and, by virtue of the mutual inhi-
time, nearly coincident with the cue that instructed the bition, suppress activity of neurons processing other
INFERIOR TEMPORAL CORTEX
response that led to reward acquisition. stimuli. This idea of excitatory bias signals that resolve
A neocortical region responsible
for high-level analysis of form local competition can be extended from visual attention
information. Keeping to a task to cognitive control in general8,63.
The capacity for PFC neurons to sustain activity is Several studies have indicated that the PFC exerts a
POSTERIOR PARIETAL CORTEX important not only for learning, but also for persisting top-down influence over other neocortical regions.
A region of the visual cortex
thought to be involved in
towards goals. One of the classic signs of PFC damage is Deactivation of the lateral PFC attenuates the activity of
visuospatial, visuomotor and increased distractibility: subjects seem unable to focus extrastriate neurons to a behaviourally relevant cue64,65.
attentional processes. on a task when other, irrelevant events compete for their Tomita et al.66 showed that top-down signals originating

62 | OCTOBER 2000 | VOLUME 1 www.nature.com/reviews/neuroscience

© 2000 Macmillan Magazines Ltd
 REVIEWS

Because task representations in the PFC include dis-

Box 3 | Task-dependent activity in the lateral prefrontal cortex
parate information, the excitatory signals from this area
We trained monkeys to alternate between tasks that used the same cues and responses but could be involved in selecting particular sensory inputs
three different rules: matching (delayed matching to sample), associative (conditional (attention), memories (recall) or motor outputs
visuomotor) and spatial (spatial delayed response)34. The first two tasks shared common (response selection). By simultaneously biasing process-
cue stimuli, but differed in how these cues were used to guide behaviour, whereas the ing in different brain systems towards a common
latter two used different cues to instruct the same behaviour. All three required the same ‘theme’ (the task), the PFC can select the neural path-
motor responses. The associative task required the monkeys to associate a foveally ways needed to perform the task (BOX 2).
presented cue stimulus with a saccade either to the right or left. The cue–response pairings
were reversed within each session in order not to confound the influence of cue stimulus
Practice and automaticity
and response direction on neural activity. The object task used the same cue stimuli as the
The PFC may have a key role in task acquisition, but it is
associative task; however, in this case the monkeys needed only to remember the identity
unlikely to be the long-term repository of all task infor-
of the cue and then saccade to the test object that matched it. Conversely, the spatial task
mation. Plasticity is evident throughout the neocortex,
used small spots of light to explicitly cue a saccade to the right or left and so required the
monkeys to simply remember the response direction. We found that over half of lateral even early in sensory processing68–70. As task-relevant
PFC neurons were task-dependent. A given neuron might be activated by a cue object neural pathways in other brain systems are repeatedly
during one task (for example, the associative task), but be unresponsive when the same selected by PFC bias signals, activity-dependent plastici-
cue appeared under identical sensory conditions during another task (for example, the ty mechanisms could strengthen and establish them
object task). Also, the baseline activity of many neurons varied with the task — a given independently of the PFC. When this happens, the PFC
neuron might consistently show higher baseline activity whenever the monkey performed may become less involved and the task less taxing on
the object task, for example. These results indicate that PFC neurons do not simply code a our limited cognitive resources; that is, its performance
stimulus or forthcoming action. Rather, they also convey their behavioural context, the becomes automatic. Indeed, PFC damage often impairs
pattern of associated information that is unique to a particular task. new learning while sparing well-practised tasks71, and
neuroimaging and neurophysiological studies have
from the PFC are required to activate (recall) a long- found greater PFC activation during initial learning
term memory stored in the inferior temporal cortex. with weaker activity to familiar stimuli or during per-
Other suggestive evidence comes from Miller and formance of well-practised tasks29,72–75. An example is an
Desimone’s52,56,67 investigation into the respective roles experiment in which monkeys were required to learn to
of the PFC and inferior temporal cortex in working associate each of two novel cue objects with a saccade to
memory. Monkeys were trained to hold a sample object the right or left29. They also performed this task with
‘in mind’ while they viewed a sequence of objects. They well-practised object–saccade associations — two high-
were required to respond when the sample was repeat- ly familiar cue objects that were used throughout
ed and to ignore other irrelevant object repetitions. As months of training and whose associations with sac-
noted above, sustained activity in the prefrontal, but cades had therefore been well established. The average
not inferior temporal, cortex maintained the sample activity across the entire population of 254 lateral pre-
memory across intervening stimuli52,56. However, many frontal neurons studied in this experiment is shown in
inferior temporal neurons showed an enhancement of FIG. 4. Novel objects that required new associative learn-
their neural responses to the sample repetition but not ing elicited, on average, more activity than the familiar
to irrelevant repetitions52,67. This indicated that sus- objects for which the associations were already well
tained activity to the sample in the PFC might have learned. Weaker responses to familiar stimuli are not
enhanced responses to its repetition in the inferior tem- unique to the PFC: neurons in the inferior temporal
poral cortex52. cortex also show this property76–79.
The PFC may remain critical for implementing task
a b information, particularly in situations when familiar
behaviours need to be flexibly combined into a coher-
Latency to half-max selectivity

30 800
30
25 700 ent sequence. In addition, the PFC is required to acti-
25
Selectivity index

vate long-term visual memories stored in the temporal

Trial number

20 20 600
lobe66,80. The PFC could retain links to stored represen-
15
15 500 tations that allow it to bring visual memories and other
10
10 400 task knowledge ‘online’ when needed.
5
Indeed, the PFC does not work alone. It is inter-
5 0 300
connected with other structures that make unique
0 –5 200
250 500 750 1000 1250 1500 0 5 10 15 20 25 30
contributions to cognition. For example, the hippo-
Time Trial number campus seems to bind stimuli into long-term memo-
Figure 3 | Change in latency of response-related activity with learning. a | The average ries of specific episodes81: it has neurons that show
degree of saccade-direction selective activity for 64 lateral PFC neurons is shown in this surface conjunctive tuning for the co-occurrence of sensory
plot. Directional selectivity appeared earlier (further to the left) with increasing trial number. Each features82 (see Eichenbaum, this issue). In contrast, I
individual box represents the average selectivity index for 25 ms of one trial. The trials are aligned suggest that the PFC represents not specific episodes
on the initiation of the saccadic eye movement and include the cue and delay intervals. The black
bar in the lower right corner illustrates the average standard error of the mean for all the data
but the regularities across them that describe task
points. b | The time at which half of the maximal selectivity was reached within each trial is plotted rules. Furthermore, the PFC engenders flexibility.
along with the fitting sigmoid function. Note that the largest change in latency occurred for trials Unlike the hippocampus, which seems to consolidate
5–15, which is exactly when the monkeys were learning the associations. (Adapted from REF. 29.) ‘permanent’ connections in the neocortex, the PFC

NATURE REVIEWS | NEUROSCIENCE VOLUME 1 | OCTOBER 2000 | 6 3

© 2000 Macmillan Magazines Ltd
REVIEWS

150
Cue Delay response based on them. Here, I have extended and
modified this idea to include not just recent sensory
125 inputs, but also task contingencies and rules. Such infor-
Novel objects with newly
100 learned associations mation must be maintained until the goal at hand is
achieved. Maintenance of information is also critical

Activity index
75
Familiar objects with because learning rules typically involves forming associ-
well-learned associations
50 ations between disparate events separated in time.
25
Fuster30,86 has emphasized the importance of the PFC in
temporal integration and this idea is central to the
0 model I have proposed here. Petrides and Owen have
–25 explored the role of the PFC in the monitoring and
0 0.5 1.0 1.5 manipulation of information held ‘in mind’, an impor-
Time (s) tant cognitive faculty87,88. The ability of the PFC to flexi-
bly form associations in accord with a current goal may
Figure 4 | Stimulus familiarity and prefrontal neurons. Plot
be the neural implementation of this capacity.
of the normalized activity of 254 prefrontal neurons in trials in
which unusual objects were used (red) versus trials that used PFC organization could provide important clues to
highly familiar objects (blue). Activity was normalized by PFC function. One possibility is that different PFC
expressing it as a percentage change over average baseline regions conduct qualitatively different operations87–89.
firing rate during the inter-trial interval (not pictured). The Other possibilities include organization on the basis of
shaded area represents the time of cue presentation. The bin stimulus dimension90. These schemes are not mutually
width was 10 ms.
exclusive. The model proposed here does not address
this issue directly, but it does make related claims. I have
dynamically selects among existing pathways. The suggested that the PFC is involved in representing
BASAL GANGLIA and CEREBELLUM are important structures acquired relationships between various pieces of infor-
for automating behavioural and cognitive routines, mation, a function essential for intelligent behaviour.
particularly their timing7,83,84. Mechanisms that deter- This allows for the possibility of a relative regional
mine when to exert control are also critical and this emphasis of certain stimulus domains or processes, but
may depend on the ANTERIOR CINGULATE CORTEX85. it also indicates that disparate information cannot be
divided into separate PFC modules. Also, the functions I
Conclusions have ascribed to the PFC indicate that learning will be
One of the brain’s great mysteries is cognitive control. important in the formation of its representations, and
How does the brain produce behaviour that seems hence in its organization.
organized and wilful? Here, I have reviewed evidence In conclusion, I intended to convey here a general
BASAL GANGLIA
that cognitive control stems from patterns of activity in view of the type of mechanisms that might underlie the
A collection of interconnected
subcortical structures
the PFC that represent goals and the means to achieve role of the PFC in cognitive control. Virtually all com-
reciprocally connected to the them. Bias signals are provided to other brain structures plex behaviour involves constructing relationships
prefrontal cortex. that can flexibly guide the flow of activity along task-rel- between diverse, arbitrary pieces of information that
evant neural pathways, so establishing appropriate have no intrinsic connection. Insight into the role of the
CEREBELLUM
mappings between inputs, internal states and outputs PFC in cognition can surely be gained from a better
A structure overlying the pons
that is important for needed to perform a given task. understanding of this process.
sensorimotor coordination. This account of PFC function complements other
theories. Goldman-Rakic and colleagues have empha-
ANTERIOR CINGULATE CORTEX
sized the role of the PFC in holding sensory informa- Links
A structure lying close to, and
connected with, the prefrontal
tion online temporarily through sustained activity10. ENCYCLOPEDIA OF LIFE SCIENCES Neural activity and the
cortex, which is involved in error This is important because sensory events are often fleet- development of brain circuits | Learning and memory |
detection. ing, but we must frequently wait to make a decision or a Dopamine

1. Barsalou, L. W. & Sewell, D. R. Contrasting the 7. Wise, S. P., Murray, E. A. & Gerfen, C. R. The frontal-basal of automatic processes: A parallel distributed processing
representation of scripts and categories. J. Mem. Lang. 24, ganglia system in primates. Crit. Rev. Neurobiol. 10, model of the Stroop effect. Psychol. Rev. 97, 332–361
646–665 (1985). 317–356 (1996). (1996).
2. Abbott, V., Black, J. B. & Smith, E. E. The representation of 8. Miller, E. K. The prefrontal cortex: complex neural properties 15. Shimamura, A. P. The role of the prefrontal cortex in dynamic
scripts in memory. J. Mem. Lang. 24, 179–199 (1985). for complex behavior. Neuron 22, 15–17 (1999). filtering. Psychobiology (in the press).
3. Norman, D. A. & Shallice, T. in Consciousness and 9. Pandya, D. N. & Barnes, C. L. in The Frontal Lobes 16. Dickinson, A. Contemporary Animal Learning Theory
Self–Regulation: Advances in Research and Theory (eds Revisited (ed. Perecman, E.) 41–72 (IRBN Press, New York, (Cambridge Univ. Press, 1980).
Davidson, R. J., Schwartz, G. E. & Shapiro, D.) 1–18 1987). 17. Vaadia, E., Benson, D. A., Hienz, R. D. & Goldstein, M. H. Jr
(Plenum, New York, 1986). 10. Goldman-Rakic, P. S. in Handbook of Physiology: The Unit study of monkey frontal cortex: active localization of
4. Grafman, J. in Handbook of Neuropsychology (eds Boller, F. Nervous System (ed. Plum, F.) 373–417 (American auditory and of visual stimuli. J. Neurophysiol. 56, 934–952
& Grafman, J.) 187 (Elsevier, Amsterdam, 1994). Physiological Society, Bethesda, 1987). (1986).
5. Cohen, J. D. & Servan-Schreiber, D. Context, cortex, and 11. Fuster, J. M. The Prefrontal Cortex (Raven Press, New York, 18. Watanabe, M. Frontal units of the monkey coding the
dopamine: A connectionist approach to behavior and 1989). associative significance of visual and auditory stimuli. Exp.
biology in schizophrenia. Psychol. Rev. 99, 45–77 (1992). 12. Barbas, H. & Pandya, D. in Frontal Lobe Function and Brain Res. 89, 233–247 (1992).
An account of how task ‘context’ representations can Dysfunction (eds Levin, H. S., Eisenberg, H. M. & Benton, A. 19. Rao, S. C., Rainer, G. & Miller, E. K. Integration of what and
be implemented in a model of cognitive control that L.) 35–58 (Oxford Univ. Press, New York, 1991). where in the primate prefrontal cortex. Science 276,
further proposes that dopamine may be central in this 13. Dehaere, S. & Changeux, J. P. The Wisconsin Card Sort 821–824 (1997).
process. Test: Theoretical analysis and modelling in a neuronal 20. Rainer, G., Asaad, W. F. & Miller, E. K. Memory fields of
6. Passingham, R. The Frontal Lobes and Voluntary Action network. Cerebral Cortex 1, 62–79 (1991). neurons in the primate prefrontal cortex. Proc. Natl Acad.
(Oxford Univ. Press, Oxford, 1993). 14. Cohen, J. D., Dunbar, K. & McClelland, J. L. On the control Sci. USA 95, 15008–15013 (1998).

64 | OCTOBER 2000 | VOLUME 1 www.nature.com/reviews/neuroscience

© 2000 Macmillan Magazines Ltd
 REVIEWS

21. White, I. M. & Wise, S. P. Rule-dependent neuronal activity in 44. Hollerman, J. R. & Schultz, W. Dopamine neurons report an 69. Merzenich, M. M. & Sameshima, K. Cortical plasticity and
the prefrontal cortex. Exp. Brain Res. 126, 315–335 (1999). error in the temporal prediction of reward during learning. memory. Curr. Opin. Neurobiol. 3, 187–196 (1993).
Demonstration of ‘rule-tuned’ neurons in the primate Nature Neurosci. 1, 304–309 (1998). 70. Gilbert, C. D. Plasticity in visual perception and physiology.
prefrontal cortex. Monkeys were trained to acquire a 45. Schultz, W. & Dickinson, A. Neuronal coding of prediction Curr. Opin. Neurobiol. 6, 269–274 (1996).
target using either a ‘spatial’ or ‘associative’ rule. errors. Annu. Rev. Neurosci. 23, 473–500 (2000). 71. Rushworth, M. F., Nixon, P. D., Eacott, M. J. & Passingham,
22. Rainer, G., Asaad, W. F. & Miller, E. K. Selective A review of evidence that dopamine neurons provide R. E. Ventral prefrontal cortex is not essential for working
representation of relevant information by neurons in the a ‘prediction error’ signal that can orchestrate memory. J. Neurosci. 17, 4829–4838 (1997).
primate prefrontal cortex. Nature 393, 577–579 (1998). learning of the means to acquire rewards. 72. Knight, R. T. Decreased response to novel stimuli after
23. Petrides, M. Deficits in non-spatial conditional associative 46. Cepeda, C., Buchwald, N. A. & Levine, M. S. prefrontal lesions in man. Clin. Neurophys. 59, 9–20 (1984).
learning after periarcuate lesions in the monkey. Behav. Neuromodulatory actions of dopamine in the neostriatum 73. Yamaguchi, S. & Knight, R. T. Anterior and posterior
Brain Res. 16, 95–101 (1985). are dependent upon the excitatory amino acid receptor association cortex contributions to the somatosensory
24. Petrides, M. Nonspatial conditional learning impaired in subtypes activated. Proc. Natl Acad. Sci. USA 90, P300. J. Neurosci. 11, 2039–2054 (1991).
patients with unilateral frontal but not unilateral temporal lobe 9576–9580 (1993). 74. Knight, R. T. Distributed cortical network for visual attention.
excisions. Neuropsychologia 28, 137–149 (1990). 47. Williams, G. V. & Goldman-Rakic, P. S. Modulation of J. Cogn. Neurosci. 9, 75–91 (1997).
25. Gaffan, D. & Harrison, S. Inferotemporal-frontal memory fields by dopamine D1 receptors in prefrontal 75. Shadmehr, R. & Holcomb, H. Neural correlates of motor
disconnection and fornix transection in visuomotor cortex. Nature 376, 572–575 (1995). memory consolidation. Science 277, 821–824 (1997).
conditional learning by monkeys. Behav. Brain Res. 31, 48. Braver, T. S. & Cohen, J. D. in Attention and Performance 18 76. Riches, I. P., Wilson, F. A. & Brown, M. W. The effects of
149–163 (1988). (eds Monsell, S. & Driver, J.) (MIT Press, Cambridge, visual stimulation and memory on neurons of the
26. Eacott, M. J. & Gaffan, D. Inferotemporal-frontal Massachusetts, in the press). hippocampal formation and the neighboring
disconnection — the uncinate fascicle and visual associative 49. Fuster, J. M. Unit activity in prefrontal cortex during delayed- parahippocampal gyrus and inferior temporal cortex of the
learning in monkeys. Eur. J. Neurosci. 4, 1320–1332 (1992). response performance: neuronal correlates of transient primate. J. Neurosci. 11, 1763–1779 (1991).
27. Parker, A. & Gaffan, D. Memory after frontal/temporal memory. J. Neurophysiol. 36, 61–78 (1973). 77. Miller, E. K., Gochin, P. M. & Gross, C. G. Habituation-like
disconnection in monkeys: conditional and non-conditional 50. Niki, H. Differential activity of prefrontal units during right and decrease in the responses of neurons in inferior temporal
tasks, unilateral and bilateral frontal lesions. left delayed response trials. Brain Res. 70, 346–349 (1974). cortex of the macaque. Vis. Neurosci. 7, 357–362 (1991).
Neuropsychologia 36, 259–271 (1998). 51. Funahashi, S., Bruce, C. J. & Goldman-Rakic, P. S. 78. Li, L., Miller, E. K. & Desimone, R. The representation of
28. Watanabe, M. Prefrontal unit activity during associative Mnemonic coding of visual space in the monkey’s stimulus familiarity in anterior inferior temporal cortex. J.
learning in the monkey. Exp. Brain Res. 80, 296–309 (1990). dorsolateral prefrontal cortex. J. Neurophysiol. 61, 331–349 Neurophysiol. 69, 1918–1929 (1993).
29. Asaad, W. F., Rainer, G. & Miller, E. K. Neural activity in the (1989). 79. Miller, E. K., Li, L. & Desimone, R. A neural mechanism for
primate prefrontal cortex during associative learning. Neuron 52. Miller, E. K., Erickson, C. A. & Desimone, R. Neural working and recognition memory in inferior temporal cortex.
21, 1399–1407 (1998). mechanisms of visual working memory in prefrontal cortex Science 254, 1377–1379 (1991).
Neural information about a cue object and the of the macaque. J. Neurosci. 16, 5154–5167 (1996). 80. Rainer, G., Rao, S. C. & Miller, E. K. Prospective coding for
saccade it instructed merged together in prefrontal 53. Romo, R., Brody, C. D., Hernandez, A. & Lemus, L. objects in primate prefrontal cortex. J. Neurosci. 19,
activity in this neurophysiological study of associative Neuronal correlates of parametric working memory in the 5493–5505 (1999).
learning. prefrontal cortex. Nature 399, 470–473 (1999). 81. Squire, L. R. & Zola-Morgan, S. The medial temporal lobe
30. Fuster, J. M., Bodner, M. & Kroger, J. K. Cross-modal and 54. Schultz, W., Dayan, P. & Montague, P. R. A neural substrate memory system. Science 253, 1380–1386 (1991).
cross-temporal association in neurons of frontal cortex. of prediction and reward. Science 275, 1593–1599 (1997). 82. Eichenbaum, H., Dudchenko, P., Wood, E., Shapiro, M. &
Nature 405, 347–351 (2000). 55. Duncan, J., Emslie, H., Williams, P., Johnson, R. & Freer, C. Tanila, H. The hippocampus, memory, and place cells: Is it
Demonstration that prefrontal neurons reflect Intelligence and the frontal lobe: The organization of goal- spatial memory or a memory space? Neuron 23, 209–226
learned cross-modal associations. Many prefrontal directed behavior. Cogn. Psychol. 30, 257–303 (1996). (1999).
neurons were selectively responsive to a visual 56. Miller, E. K., Li, L. & Desimone, R. Activity of neurons in 83. Ivry, R. B. The representation of temporal information in
stimulus and the auditory stimulus with which it was anterior inferior temporal cortex during a short-term memory perception and motor control. Curr. Opin. Neurobiol. 6,
associated. task. J. Neurosci. 13, 1460–1478 (1993). 851–857 (1996).
31. Bichot, N. P., Schall, J. D. & Thompson, K. G. Visual feature 57. Constantinidis, C. & Steinmetz, M. A. Neuronal activity in 84. Graybiel, A. M. The basal ganglia and chunking of action
selectivity in frontal eye fields induced by experience in posterior parietal area 7a during the delay periods of a sequences. Neurobiol. Learn. Mem. 70, 119–136 (1998).
mature macaques. Nature 381, 697–699 (1996). spatial memory task. J. Neurophysiol. 76, 1352–1355 (1996). 85. Botvinick, M., Nystrom, L. E., Fissell, K., Carter, C. S. &
Neurophysiological study showing learning-induced 58. Zipser, D., Kehoe, B., Littlewort, G. & Fuster, J. A spiking Cohen, J. D. Conflict monitoring versus selection-for-action
response properties for neurons in the frontal eye network model of short-term active memory. J. Neurosci. in anterior cingulate cortex. Nature 402, 179–181 (1999).
fields. Monkeys trained to look for a particular colour 13, 3406–3420 (1993). 86. Fuster, J. M. Memory in the Cerebral Cortex (MIT Press,
developed neurons sensitive to that colour. 59. Durstewitz, D., Kelc, M. & Gunturkun, O. A Cambridge, Massachusetts, 1995).
32. Bichot, N. P. & Schall, J. D. Effects of similarity and history neurocomputational theory of the dopaminergic modulation 87. Petrides, M. Functional organization of the human frontal
on neural mechanisms of visual selection. Nature Neurosci. of working memory functions. J. Neurosci. 19, 2807–2822 cortex for mnemonic processing — Evidence from
2, 549–554 (1999). (1999). neuroimaging studies. Ann. NY Acad. Sci. 769, 85–96
33. Hoshi, E., Shima, K. & Tanji, J. Task-dependent selectivity of 60. Wang, X. J. Synaptic basis of cortical persistent activity: the (1995).
movement-related neuronal activity in the primate prefrontal importance of NMDA receptors to working memory. 88. Owen, A. M., Evans, A. C. & Petrides, M. Evidence for a
cortex. J. Neurophysiol. 80, 3392–3397 (1998). J. Neurosci. 19, 9587–9603 (1999). two-stage model of spatial working memory processing
34. Asaad, W. F., Rainer, G. & Miller, E. K. Task-specific neural 61. Durstewitz, D., Seamans, J. K. & Sejnowski, T. J. within the lateral frontal cortex: A positron emission
activity in the primate prefrontal cortex. J. Neurophysiol. 84, Dopamine-mediated stabilization of delay-period activity in a tomography study. Cereb. Cortex 6, 31–38 (1996).
451–459 (2000). network model of the prefrontal cortex. J. Neurophysiol. 83, Evidence from human functional imaging that
Demonstration that the responses of prefrontal 1733–1750 (2000). different prefrontal regions are involved in simple
neurons to cues and actions are highly task-specific. 62. Desimone, R. & Duncan, J. Neural mechanisms of selective maintenance versus the monitoring and manipulation
This indicates that prefrontal neurons may participate visual attention. Ann. Rev. Neurosci. 18, 193–222 (1995). of information held ‘in mind’.
in neural ensembles that represent tasks, not just A review of the neural mechanisms for focal attention. 89. Petrides, M. Specialized systems for the processing of
stimuli and forthcoming motor acts. The authors suggest that bias signals from the PFC mnemonic information within the primate frontal cortex. Phil.
35. Wallis, J. D., Anderson, K. C. & Miller, E. K. Neuronal resolve neural competition between items vying to Trans. R. Soc. Lond. B 351, 1455–1461 (1996).
representation of abstract rules in the orbital and lateral reach awareness. 90. Goldman-Rakic, P. S. in Vision and Movement Mechanisms
prefrontal cortices (PFC). Soc. Neurosci. Abstr. (in the press). 63. Miller, E. K. in Attention and Performance 18 (eds Monsell, in the Cerebral Cortex (eds Caminiti, R., Hoffman, K. P.,
36. Dehaene, S., Kerszeberg, M. & Changeux, J. P. A neuronal S. & Driver, J.) (MIT Press, Cambridge, Massachusetts, in Lacquaniti, F. & Altman, J.) 162–172 (HFSP, Strasbourg,
model of a global workspace in effortful cognitive tasks. the press). 1996).
Proc. Natl Acad. Sci. USA 95, 14529–14534 (1998). 64. Fuster, J. M., Bauer, R. H. & Jervey, J. P. Functional 91. Milner, B. Effects of different brain lesions on card sorting.
37. Watanabe, M. Reward expectancy in primate prefrontal interactions between inferotemporal and prefrontal cortex in Arch. Neurol. 9, 90 (1963).
neurons. Nature 382, 629–632 (1996). a cognitive task. Brain Res. 330, 299–307 (1985). 92. Dias, R., Robbins, T. W. & Roberts, A. C. Primate analogue
38. Tremblay, L. & Schultz, W. Relative reward preference in 65. Chafee, M. V. & Goldman-Rakic, P. S. Inactivation of parietal of the Wisconsin Card Sorting Test: effects of excitotoxic
primate orbitofrontal cortex. Nature 398, 704–708 (1999). and prefrontal cortex reveals interdependence of neural lesions of the prefrontal cortex in the marmoset. Behav.
39. Leon, M. I. & Shadlen, M. N. Effect of expected reward activity during memory-guided saccades. J. Neurophysiol. Neurosci. 110, 872–886 (1996).
magnitude on the response of neurons in the dorsolateral 83, 1550–1566 (2000). 93. Shallice, T. & Burgess, P. W. Deficits in strategy application
prefrontal cortex of the macaque. Neuron 24, 415–425 66. Tomita, H., Ohbayashi, M., Nakahara, K., Hasegawa, I. & following frontal lobe damage in man. Brain 114, 727–741
(1999). Miyashita, Y. Top-down signal from prefrontal cortex in (1991).
40. Mirenowicz, J. & Schultz, W. Importance of unpredictability executive control of memory retrieval. Nature 401, 699–703 94. Jones, E. G. & Powell, T. P. S. An anatomical study of
for reward responses in primate dopamine neurons. J. (1999). converging sensory pathways within the cerebral cortex of
Neurophysiol. 72, 1024–1027 (1994). Neurophysiological study showing that ‘top-down’ the monkey. Brain 93, 793–820 (1970).
41. Mirenowicz, J. & Schultz, W. Preferential activation of signals from the PFC are required to activate long- 95. Chavis, D. A. & Pandya, D. N. Further observations on
midbrain dopamine neurons by appetitive rather than term memories stored in the inferior temporal cortex. cortico–frontal connections in the rhesus monkey. Brain
aversive stimuli. Nature 379, 449–451 (1996). 67. Miller, E. K. & Desimone, R. Parallel neuronal mechanisms Res. 117, 369–386 (1976).
42. Schultz, W., Apicella, P. & Ljungberg, T. Responses of for short-term memory. Science 263, 520–522 (1994).
monkey dopamine neurons to reward and conditioned 68. Recanzone, G. H., Merzenich, M. M. & Jenkins, W. M. Acknowledgements
stimuli during successive steps of learning a delayed Frequency discrimination training engaging a restricted skin I thank Wael Asaad, Jonathan Cohen, Peter Dayan, John Duncan,
response task. J. Neurosci. 13, 900–913 (1993). surface results in an emergence of a cutaneous response Howard Eichenbaum, David Freedman, Tomaso Poggio,
43. Schultz, W. Predictive reward signal of dopamine neurons. zone in cortical area 3a. J. Neurophysiol. 67, 1057–1070 Maximilian Riesenhuber and Marlene Wicherski for valuable com-
J. Neurophysiol. 80, 1–27 (1998). (1992). ments and discussions.

NATURE REVIEWS | NEUROSCIENCE VOLUME 1 | OCTOBER 2000 | 6 5

© 2000 Macmillan Magazines Ltd