Opinion

The What and How of prefrontal

cortical organization
Randall C. OReilly
Department of Psychology and Neuroscience, University of Colorado Boulder, 345 UCB, Boulder, CO 80309, USA

How is the prefrontal cortex (PFC) organized such that it ations of the other two dimensions have been extensively
is capable of making people more flexible and in control discussed in the literature and are not themselves novel,
of their behavior? Is there any systematic organization but they interact in potentially interesting ways with the
across the many diverse areas that comprise the PFC, or first dimension, and are discussed in that context.
is it uniquely adaptive such that no fixed representa- The What versus How idea is developed below, and then
tional structure can develop? Going against the current related to the two other axes of PFC organization in
tide, this paper argues that there is indeed a systematic subsequent sections.
organization across PFC areas, with an important func-
tional distinction between ventral and dorsal regions What versus How
characterized as processing What versus How infor- There are two broad frameworks for understanding the
mation, respectively. This distinction has implications ventral versus dorsal organization of the posterior cortex:
for the rostro-caudal and mediallateral axes of organ- the What versus Where distinction advanced by Ungerlei-
ization as well. The resulting large-scale functional map der and Mishkin [6], and the later What versus How model
of PFC could prove useful in integrating diverse data, and of Goodale and Milner [5]. The key distinction between
in generating novel predictions. these two frameworks is in characterizing the role of the
dorsal pathway (principally the parietal cortex) is it
The What-How, Abstraction, Cold/Hot (WHACH) model primarily about spatial representations (Where) or is it
of PFC organization primarily about transforming perception into action
The prefrontal cortex (PFC) is known to be important for (How)? Some of the data motivating the How model showed
cognitive control, enabling behavior to be at once flexible that people with ventral pathway damage could not
yet task-focused [1,2]. One of the principal means of un- describe shape information (e.g. the angle of a slot that
derstanding how it achieves these remarkable feats is by was rotated in different ways), but could nevertheless
characterizing the nature of its underlying neural repres- clearly express shape knowledge when it was used to
entations. The central question addressed here is: are PFC constrain their actions (e.g. putting a card into the rotated
representations systematically organized across areas, slot) [5].
and if so, what is the nature of this organization? Although Thus, Goodale and Milners key insight was that shape
many have attempted to answer the former in the affir- information can be processed by both pathways, and the
mative, with a variety of different organizational schemes, crucial distinction is how the information is used the
the broadest consensus in the field seems to be that, if there dorsal pathway extracts visual signals relevant for driving
is any organization there, it is extremely difficult to charac- motor behavior ( perception for action), whereas the ventral
terize. Indeed, some go so far as to argue that PFC should pathway extracts information relevant for identification
not exhibit any kind of systematic organization, by virtue of and other forms of semantic knowledge. Note that spatial
its very nature [3,4] (Box 1). information is often very relevant for guiding motor beha-
However, there are also strong reasons to believe that vior, and this could explain the prevalence of Where infor-
the PFC should have some kind of stable systematic mation in the dorsal pathway, such that the How model can
organization (Box 1), and considerable data appear to be be considered a generalization of the Where model.
consistent with a specific proposal advanced here: the The notion of taking a ventral versus dorsal distinction
What-How, Abstraction, Cold/Hot (WHACH) model from posterior cortex forward into the PFC was pioneered
(Figure 1). This model is organized along 3 major axes: by Goldman-Rakic and others, in the context of the original
ventral versus dorsal (What versus How), rostral versus What versus Where model [79]. A significant motivation
caudal (Abstraction), and lateral versus medial (Cold ver- for doing so is that the appropriate ventral-to-ventral and
sus Hot). The ventral versus dorsal distinction is the dorsal-to-dorsal connectivity patterns between PFC and
primary focus of the paper. The key idea is to bring the corresponding posterior cortical areas are predominant [9
What versus How distinction between ventral and dorsal 11] (Figure 2). But the evidence in support of this What
pathways in posterior cortex, developed by Goodale and versus Where distinction in PFC, in both monkeys and
Milner [5], forward into the PFC in terms of ventrolateral humans, has not been very consistent [12,13].
(VLPFC) versus dorsolateral (DLPFC). The characteriz- The main claim of this paper is that Goldman-Rakics
overall strategy was correct, but that the characterization
Corresponding author: OReilly, R.C. ([email protected]) of posterior cortex she used was too narrow: if you instead
0166-2236/$ see front matter ! 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.tins.2010.05.002 Trends in Neurosciences 33 (2010) 355361 355
Opinion Trends in Neurosciences Vol.33 No.8

Box 1. Dedicated versus dynamic PFC representations

The adaptive coding hypothesis offered by John Duncan, building in the volume of our voices, but nothing else: this is how neurons
part on the ideas and work of Earl Miller, suggests that PFC neurons communicate. This volume-based communication is appropriate
are capable of rapidly adapting to new tasks, and thus explains how given that a given neuron is listening to roughly 10,000 other neurons
the PFC can play such an important role in supporting flexible, task- at the same time: it would be impossible to decode and integrate
relevant behavior [3,4]. This hypothesis accounts for the evident more complex signals with this level of parallelism.
adaptability of monkey PFC neurons when trained on new tasks In effect, neurons operate within a giant social network, where the
(although such training typically takes a long time), and considerable whole game is to become a reliable source of information that other
neuroimaging data showing that many diverse PFC functions activate neurons can learn to trust (Figure I). The meaning of the neural
highly-overlapping brain areas [4]. But it also appears incompatible message is all about the identity of the sender and the relationship of
with the idea that the PFC has the kind of systematic long-term its neural firing to all the others within the network. Receiving
organizational structure proposed here. neurons gradually learn to respond to reliable patterns of incoming
However, the adaptive coding hypothesis ignores a crucial aspect of activation, and pass along filtered and transformed versions thereof
neural processing: neurons do not communicate using symbols that to other neurons. If a given neuron were to suddenly change its
have intrinsic meaning. Thus, it is not possible for a neuron to rapidly representation, there would be no way for other neurons to interpret
change what it encodes, because it cannot communicate this new its new meaning (its spikes would look the same as ever). They have
representation to other neurons. Instead, meaning in neural networks to learn over time the new relationships between this neuron and the
must be learned over time and in relationship to larger patterns of others. This poses a strong constraint on the adaptive coding
overall neural activity, and this strongly favors stable long-term hypothesis.
representations and larger-scale structuring thereof. There are several alternative ways to account for data supportive of
Neurons broadcast simple spike impulses, each the same as any the adaptive coding hypothesis within this framework, including that
other, with at least the majority of the information contained in the PFC representations are (a) highly abstract, and thus encompass
firing rate over time [67]. This contrasts strongly with human spoken many specific tasks, (b) highly context-sensitive, taking on different
language, where we have elaborate phonological distinctions that activation patterns in different tasks or other contexts, and (c) strongly
multiply over time to convey words and sentences that have modulated by the basal ganglia to maintain different information in a
[(Box_1)TD$FIG]
independent meaning. Imagine instead if we could only modulate very dynamic manner [2].

Figure I. Social network analogy for neural communication and learning. In panel (a), the author is in a central node and communicates with various colleagues. Over
time, the author has come to learn which people in his network are reliable, what their expertise is, and how to evaluate the kinds of things they might say. Hence, if the
social network were to undergo dynamic reconfiguration (b), the consequences could be devastating for the authors ability to trust, filter, and otherwise build upon the
information coming from the network.

take the What versus How characterization of ventral Miller and colleagues showed that monkeys trained on tasks
versus dorsal pathways forward into the PFC, it provides that integrate both What and Where information had highly
a much more compelling fit with the available data. Specifi- overlapping distributions of What and Where cells over a
cally, the proposal is that VLPFC (Brodmann areas large area of lateral PFC [13]. Romanski [9] provides a nice
44,45,47,12) (denoted as What* where the * indicates discussion of these and other relevant data, and argues that
the cognitive control and robust active maintenance abil- a major part of the problem is that the VLPFC in the
ities associated with the PFC) provides a cognitive control macaque that actually interconnects with the ventral
system for ventral What pathway processing, and DLPFC posterior cortex is quite far ventral relative to where many
(Brodmann areas 8, 9, 46, How*) provides cognitive control relevant neural recordings took place. Furthermore, many
for the dorsal pathway. This idea has not been explored of these recordings include the region between the clearly-
very extensively a thorough literature search revealed defined VLPFC and DLPFC areas, corresponding roughly to
only one relatively brief discussion of this specific idea [14], inferior frontal junction (IFJ) in humans, and this could
and another paper discusses the Goodale and Milner ideas represent a more polymorphous bridging region between
more broadly in relation to PFC [15]. What and How pathways [16]. The What* versus How*
distinction should definitely be considered as more of a
What about What? continuum. Also, the posterior aspect of VLPFC (e.g. along
The first obstacle for the What* versus How* model is that the 44/6 border) is a grey area because adjacent motor areas
several of the arguments against the What/Where idea in (organized somatotopically) can drive very different repres-
PFC were focused on the shared What aspect. For example, entations, such as a spatial eye field map [17].

356
(Figure_1)TD$IG][ Opinion Trends in Neurosciences Vol.33 No.8

for monitoring the sensory signals that indicate when to

inhibit, rather than the inhibitory process itself [23]. This
account is highly consistent with the What* model for
VLPFC. Perhaps the lateralization differences have more
to do with the left hemisphere being dominant for the
primary task, whereas the right takes on secondary tasks
(as discussed more later).

How about How?

The advantage of the DLPFC How* side of the story
relative to the earlier Where* account is that it encom-
passes a much broader range of cognitive processing one
Figure 1. The complete set of broad functional organizations discussed, along should expect to see DLPFC activation whenever the
each of the three major axes: dorsal [How = perception for action = Stimulus (S) to parietal cortex requires extra cognitive control (such as
Response (R) mappings] versus ventral [What = Stimulus-driven semantic
representations (S)]; within PFC rostral (abstract) versus caudal (concrete); working memory and top-down biasing) to carry out the
medial (Hot value representations) versus lateral (Cold cognitive calculations). processing of sensory information to guide action outputs.
How* (DLPFC) indicates a control system for the How posterior pathway, as What*
For example, studies show that parietal cortex plays an
(VLPFC) does for the What pathway. Grey numbers indicate Brodmann areas on
the lateral surface. important role in guiding the encoding and retrieval of
memories [24], representing number and mathematical
transformations [25], and encoding various forms of
Human neuroimaging studies of the VLPFC provide relationships (spatial and more abstract) [26]. Thus, the
ample evidence to support a clear What* role. Numerous definition of perception for action must include a broad
studies have shown that the VLPFC plays a crucial role in range of more abstract cognitive actions such as memory
guiding the selection and retrieval of semantic/linguistic retrieval and mathematical transformations, in addition to
knowledge that is almost certainly encoded in the infer- more concrete motor actions. As discussed later, this
otemporal cortex (i.e. ventral What posterior cortex) [18 account of DLPFC aligns well with Petrides theory [27].
20]. From a computational perspective, the most straight- The relevant DLPFC data are generally consistent with
forward account of these data involves the active mainten- the How* view. Starting at the most posterior end of DLPFC,
ance of stimulus information in VLPFC, and this produces areas 6 (which is properly not prefrontal, just frontal) and 8
a top-down biasing effect to drive selection and retrieval have been shown to encode simple action rules [2830], and
dynamics in posterior cortex [2,21]. also spatial maps associated with saccade planning and
Most of the above-mentioned VLPFC data come from spatial working memory [17]. Moving somewhat more
studies with predominant left hemisphere activation anterior, monkey neurophysiology data from studies of
what about the right VLPFC? This area has been charac- temporal order processing show that VLPFC neurons
terized as important for response inhibition [22], and this respond selectively to object identity, whereas DLPFC
would seem at odds with the What* account. However, neurons respond selectively to sequential order [31,32].
recent data suggest that this area could be more important Sequential order is a good example of higher-level How
[(Figure_2)TD$IG]

Figure 2. Summary of anatomical connectivity (in the macaque, where most of the anatomical work has been done) suggesting that dorsal versus ventral distinctions in
posterior cortex should influence prefrontal cortex, due to dominant dorsaldorsal and ventralventral connectivity. Reproduced, with permission, from Ref. [9].

357
(Figure_3)TD$IG][ Opinion Trends in Neurosciences Vol.33 No.8

Figure 3. Results from the Wager and Smith meta-analysis [12], highlighting the selectivity of (a) VLPFC areas (Brodmann areas 44, 45; circled) for object and verbal content
information, consistent with the What pathway account; and (b) DLPFC areas (Brodmann areas 6, 8, 9, 46) for updating and order processing, consistent with the How
pathway account, in that these are broad categories of cognitive actions that also activate parietal cortex areas.

pathway information it is important for coordinating A recent study on response versus semantic selection
motor commands, and probably involves number-line-like difficulty manipulations provides additional converging
representations in parietal cortex to organize order infor- evidence for dorsal versus ventral PFC involvement [40].
mation [33]. An even more abstract example of DLPFC task- They found that DLPFC was modulated by response selec-
rule encoding comes from neurons that encoded a repeat tion difficulty, but not by semantic selection difficulty, and
stay/changeshift strategy, independent of specific stimuli vice-versa for VLPFC, exactly as would be predicted by the
or actions involved [34]. More generally, DLPFC regions are What versus How model. Intrinsic functional connectivity
widely reported for action selection and behavioral rule analyses also reveal distinct dorsal versus ventral fronto-
performance tasks [3539], consistent with a role as a parietal networks [41].
top-down bias on parietal How processing pathways (and Taken together, these data are consistent with the
all of these studies find strong co-activation between DLPFC What* versus How* distinction between VLPFC and
and parietal cortex). DLPFC, and this in turn is based on the predominant
connectivity of these regions with corresponding ventral
Direct dorsal/ventral contrasts and dorsal pathways in posterior cortex. In the next two
A major source of data on PFC functional organization is sections we explore how this What* versus How* account
the Wager and Smith [12] meta-analysis of neuroimaging might interact with other widely-discussed axes of PFC
studies. Although they conclude that the What/Where functional organization: a rostralcaudal gradient of
model is inconsistent with the data, a careful re-examin- abstraction, and a lateralmedial Cold versus Hot distinc-
ation of their findings in light of the What/How framework tion.
suggests that it could be quite consistent (Figure 3).
Specifically, areas 44 and 45 (VLPFC) showed exclusively The rostralcaudal axis
object and verbal encoding (with appropriate rightleft Two predominant ideas about the rostralcaudal organiz-
lateralization), with no spatial sensitivity. Although these ation of PFC are in terms of gradients of abstraction
differences did not achieve statistical criterion, there is [30,4245] or rule complexity [4649]. According to the
clearly a strong numerical trend across a large number of abstraction idea, more anterior PFC areas encode more
studies that is very consistent with a What* encoding in abstract information, in terms of having broader categories
VLPFC. Conversely, when studies were sorted according to (e.g. color versus red versus brick red), or otherwise
various processing-oriented categories, DLPFC areas 6, 8 being more distantly removed from concrete physical
and 9 showed significant activation differences, but objects (e.g. beauty versus sunset). By contrast, rule
VLPFC areas did not. This is consistent with a How* role complexity refers to the number of different elements that
for DLPFC. These same DLPFC areas did not exhibit must be taken into account to generate a task-appropriate
strong content selectivity according to object, verbal or response. For example, the rule hit the left button if the
spatial information. This result is damaging for a Where* previous stimulus was an A and the current one is an X
account of DLPFC function, but not for a How* account a requires two items to be integrated (A and X) to determine
crucial feature of the How* framework is that dorsal path- the response. Interestingly, the abstraction gradient
way representations are specifically not thought to be appears to align with the VLPFC What* domain, in that
organized according to stimulus input categories, and it is focused more on semantic and categorization issues.
instead are best categorized in terms of transforming a By contrast, rule complexity is more response-focused and
wide range of sensory inputs into appropriate motor thus seems to fit better within the DLPFC How* domain.
outputs. Consistent with this view, they also found parietal Given this alignment, it seems possible that there are
cortex activations for spatial and non-spatial executive actually two parallel rostralcaudal gradients, one within
tasks, whereas IT activations were more consistently for DLPFC organized according to rule-complexity, and
visual object based tasks. another within VLPFC organized according to abstraction.

358
Opinion Trends in Neurosciences Vol.33 No.8

Computational modeling and other work suggests that and corresponding lateral PFC regions, where the DMPFC
each of these organizational gradients can emerge from areas are organized along a parallel hierarchy as discussed
hierarchical connectivity patterns (where more anterior in the previous section [59].
PFC areas provide contextualization to more posterior In sum, there appears to be a homology in the overall
PFC areas) [27,5052]. But what about the extant data? functions of the dorsal and ventral MPFC areas each
Unfortunately, the relevant fMRI data come from task forms affective value associations, and these value repres-
paradigms that do not cleanly distinguish between entations then drive behavior, for example to approach
abstraction and rule complexity as characterized above. positive stimuli and avoid negative ones (VMPFC What),
For example, Badre and DEsposito [30] describe their and to perform successful, rewarding actions and avoid
hierarchy specifically in terms of abstraction, but their difficult, risky behaviors (or exert more cognitive control
task manipulation includes a strong rule complexity con- when attempting them) (DMPFC How). This account can
found (which they acknowledge). Consistent with this also help to explain the apparent asymmetry between
confounding of the two factors, their hierarchy crosses over VMPFC and DMPFC value representations, where
between DLPFC (area 6) for the most concrete response- DMPFC areas seem to have generally more negative value
oriented 1st level, to anterior VLPFC (area 47) for the 3rd coding (such as error and conflict, generally the cost of
level, with the 2nd level being intermediate on the dorsal taking an action), whereas VMPFC value representations
ventral axis. Interestingly, a similar but not identical are more a mixture of positive and negative: actions are
hierarchical pattern was found by Koechlin and colleagues often not intrinsically rewarding, and generally incur effort
[29], but their 2nd level was more clearly in VLPFC (area and other costs, whereas stimuli are often the end goals
44). Both studies also found activation in fronto-polar PFC that drive actions (towards reward and away from punish-
(area 10) for the highest level, but this area is not clearly ment).
dorsal or ventral in nature, and is thus outside the scope of
this paper. Relationship to other frameworks
In short, more careful experimental work is needed to The What/How distinction can be related to other preva-
determine if the What versus How distinction might help to lent ways of describing the division between ventral and
clarify the hierarchical structure of representations across dorsal PFC. For example, Petrides [27] has argued, based
lateral PFC. on a wide range of data, that VLPFC is important for
simple working memory of stimulus information and
The mediallateral axis other first-order executive functions such as selection
Across the cortex, medial areas tend to be more directly and comparison, whereas DLPFC is important for more
connected to subcortical systems, and play a role in limbic complex higher-order information-processing operations.
or affective/motivational systems, whereas lateral areas This maps well onto the What/How framework, in that PFC
tend to be more involved in sensory/motor processing. executive control over What pathway information is likely
Thus, it is uncontroversial to characterize the lateral to be manifest as apparently simpler functions such as
PFC areas (which have been the focus of discussion to this active maintenance of sensory stimuli, and top-down bias-
point) as Cold cognitive processing, whereas the medial ing to select information in IT cortex. By contrast, PFC
PFC contains Hot emotional and motivational areas. In executive control over dorsal How processing will appear
this context, we again ask if the What versus How ventral/ more complex, involving coordination of motor and cogni-
dorsal distinction could have some currency. The answer tive actions over time, and selection of relational, spatial,
appears to be yes. and mathematical operations encoded in parietal cortex.
Considerable animal data suggest that ventral medial A different take on DLPFC versus VLPFC comes from
PFC (VMPFC) (e.g. orbital frontal cortex, OFC) is special- Corbetta and Shulman [60], who characterize DLPFC as
ized for representing the emotional and motivational value important for endogenous attentional focusing, whereas
of stimuli (i.e. What-based processing) [53], whereas VLPFC is more exogenously driven. Interestingly, the
DMPFC areas (including the anterior cingulate cortex, endogenous cases associated with the DLPFC that they
ACC), appear to be important for encoding affective values considered involved spatial and motion information that is
of actions [54,55]. In human neuroimaging, the specific role processed in the dorsal visual pathway, whereas the
of the ACC within DMPFC has been characterized in many exogenous cases were more stimulus-oriented tasks invol-
different ways, including conflict and error monitoring ving low-frequency oddball tasks. Thus, their actual data
[56,57]. A way of reconceptualizing these ideas about are broadly consistent with the What/How distinction, even
ACC that is consistent with a putative HotHow control if their overall characterization does not obviously map
area, is to think of it as providing a motivational cost signal onto it.
associated with prospective actions that could be taken. If a The Stuss and Alexander framework [61], based on
given action is associated with high levels of conflict or lesion data, also does not appear at first to be related to
error, then more cognitive effort will be required, and it is the What/How model, but is in fact quite compatible. They
less likely that the action will be taken. The Brown and argue for a lateralization account, where left lateral PFC is
Braver model of ACC, where it learns to associate arbitrary important for task setting, whereas right lateral PFC is
task cues with conflict, error, and other difficulty signals, is important for monitoring. Alternatively, one can think of
particularly compatible with this view [58]. Also consistent these as reflecting the same basic function (active main-
are recent neuroimaging data from Koechlin and col- tenance of task context and top-down biasing of appropri-
leagues that show a relationship between DMPFC regions ate processing), but the left hemisphere is more important

359
Opinion Trends in Neurosciences Vol.33 No.8

for the task being currently performed, whereas the right Box 2. Outstanding Questions
hemisphere could be more important for maintaining infor-
mation about other possible tasks that might be relevant at ! To what extent does the What/How distinction correctly char-
some other point in time (i.e. monitoring; see also Ref. [23] acterize PFC organization?
as discussed above). They also argue that medial PFC is ! Where exactly are the boundaries between What and How, and
how do these pathways interact to produce overall cognitive
responsible for energization, and this is compatible with control?
the Hot role for example, dorsal medial PFC areas are ! Are there two parallel forms of hierarchical organization along the
important for motivational selection of task-appropriate rostro-caudal axis, one within DLPFC and another within VLPFC,
action plans. or do both dorsal and ventral areas fit within one larger hierarchy?
! Can the What* model provide a unifying account of both left and
Banich and colleagues [62] have developed a cascade of
right VLPFC, or for example does right VLPFC really have a
control model that involves DLPFC (attentional task set), specific role in response inhibition that cannot be accounted for in
VLPFC (stimulus feature representations), and ACC terms of monitoring?
(response selection and evaluation) these functions are ! Is left VLPFC engaged for the dominant, focal task, whereas right
generally compatible with the What versus How model, VLPFC takes on secondary tasks? Can this be tested using task
switching and dual task paradigms?
and we are currently designing experiments to further test
! Does the medial Hot organization of brain mirror the lateral
the relationship between these frameworks. surface (e.g. in terms of a rostralcaudal hierarchy as suggested
Lebedev and Wise [15] discuss the Goodale and Milner by Koechlin et al. [59]), or does the affective/motivational side
framework with a focus on perceptual awareness. Based in have its own separate hierarchy, for example, a goal hierarchy?
part on experimental results showing no clear anatomical ! How do dedicated neural representations support the rapid
adaptation phenomena used by Duncan to develop his adaptive
dissociation in PFC for cells responsive or not to a visual coding hypothesis [4] (Box 1)?
illusion [63], they conclude that this dorsal versus ventral
distinction does not carry forward as such into the PFC,
and is instead more intermixed. Given that they recorded associated posterior cortical (and basal ganglia) areas, [66].
in the bridging area between dorsal and ventral (similar to Many important questions remain (Box 2), but hopefully
Ref. [13], as discussed earlier), this could be taken as future empirical research, guided in part by predictions
further evidence for intermixing in this area. from computational models, can begin to answer them and
Finally, Tanji and Hoshi [32] provide a thorough test the validity of this overall framework.
review of anatomy and functional data regarding ventral
versus dorsal PFC organization, and this is highly com- Acknowledgements
patible with What/How distinction. However, they and Supported by Office of Naval Research (ONR) grant N00014-07-1-0651
others (e.g. Ref. [64]) continue to reject the notion that and National Institutes of Health (NIH) grant MH079485. Thanks for
PFC is important for working memory (WM), based on comments and discussion: Silvia Bunge, Chris Chatham, Tom Hazy, Seth
Herd, Rich Ivry, Yuko Munakata, Wolfgang Pauli, and the Determinants
studies showing intact WM following PFC lesions. But
of Executive Function and Dysfunction (DEFD) research group.
such results can instead be accounted for in terms of
lingering memory traces in posterior cortical areas, and References
these are sufficient when there are no intervening dis- 1 Miller, E.K. and Cohen, J.D. (2001) An integrative theory of prefrontal
tractors or other forms of processing, but are otherwise cortex function. Annu. Rev. Neurosci. 24, 167202
not as robust as PFC-mediated WM representations [65]. 2 OReilly, R.C. (2006) Biologically based computational models of high-
level cognition. Science 314, 9194
In any case, working memory has many different mean-
3 Miller, E.K. (2000) The prefrontal cortex and cognitive control. Nat.
ings to different researchers, so it is clearer to instead use Rev. Neurosci. 1, 5965
the more precise computationally-explicit terminology of 4 Duncan, J. (2001) An adaptive coding model of neural function in
robust active maintenance of neural firing over time, and prefrontal cortex. Nat. Rev. Neurosci. 2, 820829
this seems to be crucial for many PFC functions beyond 5 Goodale, M.A. and Milner, A.D. (1992) Separate visual pathways for
perception and action, Trends Neurosci. 15, 2025
simple short-term maintenance, including top-down
6 Ungerleider, L.G. and Mishkin, M. (1982) Two cortical visual systems.
attention, action selection, manipulation, control, and In The Analysis of Visual Behavior (Ingle, D.J. et al., eds), pp. 549586,
so forth [2,21]. MIT Press
7 Funahashi, S. et al. (1993) Prefrontal neuronal activity in rhesus
Conclusion monkeys performing a delayed anti-saccade task. Nature 365, 753756
8 Smith, E. and Jonides, J. (1999) Storage and executive processes in the
The proposal outlined above and summarized in Figure 1 frontal lobes. Science 283, 16571661
constitutes a comprehensive map of the functional proper- 9 Romanski, L.M. (2004) Domain specificity in the primate prefrontal
ties of the PFC along multiple dimensions: ventral What* cortex. Cogn. Affect. Behav. Neurosci. 4, 421429
versus dorsal How*; anterior Abstract versus posterior 10 Petrides, M. and Pandya, D.N. (2002) Comparative cytoarchitectonic
Concrete; lateral Cold versus medial Hot (i.e. the WHACH analysis of the human and the macaque ventrolateral prefrontal cortex
and corticocortical connection patterns in the monkey. Eur. J.
model). The primary What versus How distinction seems to Neurosci. 16, 291310
have relevance for understanding issues within the other 11 Cavada, C. and Goldman-Rakic, P.S. (1989) Posterior parietal cortex in
two axes of organization, and thus could represent a pro- rhesus monkey. II. Evidence for segregated corticocortical networks
ductive theoretical framework for integrating a diverse linking sensory and limbic areas with the frontal lobe. J. Comp. Neurol.
287, 422445
range of empirical data across many different content
12 Wager, T.D. and Smith, E.E. (2003) Neuroimaging studies of working
areas and task domains. We are currently formalizing this memory: A meta-analysis. Cogn. Affect. Behav. Neurosci. 3, 255274
framework by developing integrated computational 13 Rao, S.C. et al. (1997) Integration of what and where in the primate
models of dorsal and ventral PFC areas interacting with prefrontal cortex. Science 276, 821824

360
Opinion Trends in Neurosciences Vol.33 No.8

14 Sakagami, M. et al. (2006) Behavioral inhibition and prefrontal cortex 40 Nagel, I.E. et al. (2008) Functional MRI investigation of verbal
in decision-making. Neural Netw. 19, 12551265 selection mechanisms in lateral prefrontal cortex. Neuroimage 43,
15 Lebedev, M.A. and Wise, S.P. (2002) Insights into seeing and grasping: 801807
distinguishing the neural correlates of perception and action. Behav. 41 Vincent, J.L. et al. (2008) Evidence for a frontoparietal control system
Cogn. Neurosci. Rev. 1, 108129 revealed by intrinsic functional connectivity. J. Neurophysiol. 100,
16 Brass, M. et al. (2005) The role of the inferior frontal junction area in 33283342
cognitive control. Trends Cogn. Sci. 9, 314316 42 Badre, D. and DEsposito, M. (2009) Is the rostro-caudal axis of the
17 Kastner, S. et al. (2007) Topographic maps in human frontal cortex frontal lobe hierarchical? Nat. Rev. Neurosci. 10, 659669
revealed in memory-guided saccade and spatial working-memory 43 Badre, D. et al. (2009) Hierarchical cognitive control deficits following
tasks. J. Neurophysiol. 97, 34943507 damage to the human frontal lobe. Nat. Neurosci. 12, 515522
18 Thompson-Schill, S.L. et al. (1997) Role of left inferior prefrontal cortex 44 Christoff, K. et al. (2009) Prefrontal organization of cognitive control
in retrieval of semantic knowledge: A re-evaluation. Proc. Natl. Acad. according to levels of abstraction. Brain Res. 1286, 94105
Sci. U. S. A. 94, 1479214797 45 OReilly, R.C. et al. (2002) Prefrontal cortex and dynamic categorization
19 Thompson-Schill, S.L. (2003) Neuroimaging studies of semantic tasks: representational organization and neuromodulatory control.
memory: inferring how from where. Neuropsychologia 41, 280292 Cereb. Cortex 12, 246257
20 Wagner, A.D. et al. (2001) Recovering meaning: left prefrontal cortex 46 Bunge, S.A. et al. (2006) A brain-based account of the development of
guides controlled semantic retrieval. Neuron 31, 329338 rule use in childhood. Curr. Dir. Psychol. Sci. 14, 118121
21 OReilly, R.C. et al. (1999) A biologically based computational model of 47 Burgess, P.W. et al. (2007) The gateway hypothesis of rostral prefrontal
working memory. In Models of Working Memory: Mechanisms of Active cortex (area 10) function. Trends Cogn. Sci. 11, 290298
Maintenance and Executive Control (Miyake, A. and Shah, P., eds), pp. 48 Koechlin, E. and Summerfield, C. (2007) An information theoretical
375411, Cambridge University Press approach to prefrontal executive function. Trends Cogn. Sci. 11, 229
22 Aron, A.R. et al. (2007) Triangulating a cognitive control network using 235
diffusion-weighted magnetic resonance imaging (MRI) and functional 49 Christoff, K. and Gabrieli, E. (2000) The frontopolar cortex and human
MRI. J. Neurosci. 27, 37433752 cognition: Evidence for a rostrocaudal hierarchical organization within
23 Hampshire, A. et al. (2010) The role of the right inferior frontal gyrus: the human prefrontal cortex. Psychobiology 28, 168186
inhibition and attentional control. Neuroimage 50, 13131319 50 Fuster, J.M. (1997) The Prefrontal Cortex: Anatomy, Physiology and
24 Ciaramelli, E. et al. (2008) Top-down and bottom-up attention to Neuropsychology of the Frontal Lobe (3rd edn), LippincottRaven
memory: a hypothesis (atom) on the role of the posterior parietal 51 Botvinick, M.M. (2007) Multilevel structure in behaviour and in the
cortex in memory retrieval. Neuropsychologia 46, 18281851 brain: a model of fusters hierarchy. Philos. Trans. R. Soc. Lond. B Biol.
25 Dehaene, S. et al. (2004) Arithmetic and the brain. Curr. Opin. Sci. 362, 16151626
Neurobiol. 14, 218224 52 Reynolds, J.R. and OReilly, R.C. (2009) Developing PFC
26 Chafee, M.V. et al. (2007) Representing spatial relationships in representations using reinforcement learning. Cognition 113, 281292
posterior parietal cortex: single neurons code object-referenced 53 Murray, E.A. et al. (2007) What we know and do not know about the
position. Cereb. Cortex 17, 29142932 functions of the orbitofrontal cortex after 20 years of cross-species
27 Petrides, M. (2005) Lateral prefrontal cortex: architectonic and studies. J. Neurosci. 27, 81668169
functional organization. Philos. Trans. R. Soc. Lond. B Biol. Sci. 54 Rushworth, M.F.S. et al. (2007) Contrasting roles for cingulate and
360, 781795 orbitofrontal cortex in decisions and social behaviour. Trends Cogn.
28 Brasted, P.J. and Wise, S.P. (2004) Comparison of learning-related Sci. 11, 168176
neuronal activity in the dorsal premotor cortex and striatum. Eur. J. 55 Rudebeck, P.H. et al. (2008) Frontal cortex subregions play distinct
Neurosci. 19, 721740 roles in choices between actions and stimuli. J. Neurosci. 28, 13775
29 Koechlin, E. et al. (2003) The architecture of cognitive control in the 13785
human prefrontal cortex. Science 302, 11811185 56 Botvinick, M. et al. (1999) Conflict monitoring versus selection-for-
30 Badre, D. and DEsposito, M. (2007) Functional magnetic resonance action in anterior cingulate cortex. Nature 402, 179181
imaging evidence for a hierarchical organization of the prefrontal 57 Botvinick, M.M. et al. (2004) Conflict monitoring and anterior cingulate
cortex. J. Cogn. Neurosci. 19, 20822099 cortex: an update. Trends Cogn. Sci. 8, 539546
31 Ninokura, Y. et al. (2004) Integration of temporal order and object 58 Brown, J.W. and Braver, T.S. (2005) Learned predictions of error
information in the monkey lateral prefrontal cortex. J. Neurophysiol. likelihood in the anterior cingulate cortex. Science 307, 11181121
91, 555560 59 Kouneiher, F. et al. (2009) Motivation and cognitive control in the
32 Tanji, J. and Hoshi, E. (2008) Role of the lateral prefrontal cortex in human prefrontal cortex. Nat. Neurosci. 12, 939945
executive behavioral control. Physiol. Rev. 88, 3757 60 Corbetta, M. and Shulman, G.L. (2002) Control of goal-directed and
33 Walsh, V. (2003) A theory of magnitude: common cortical metrics of stimulus-driven attention in the brain. Nat. Rev. Neurosci. 3, 201215
time, space and quantity. Trends Cogn. Sci. 7, 483488 61 Stuss, D.T. and Alexander, M.P. (2007) Is there a dysexecutive
34 Genovesio, A. et al. (2005) Prefrontal cortex activity related to abstract syndrome? Philos. Trans. R. Soc. Lond. 362, 901915
response strategies. Neuron 47, 307320 62 Banich, M. (2009) Executive function: the search for an integrated
35 Rowe, J.B. et al. (2005) The prefrontal cortex shows context-specific account. Curr. Dir. Psychol. Sci. 18, 8994
changes in effective connectivity to motor or visual cortex during the 63 Lebedev, M.A. et al. (2001) Prefrontal cortex neurons reflecting reports
selection of action or colour. Cereb. Cortex 15, 8595 of a visual illusion. J. Neurophysiol. 85, 13951411
36 Bunge, S.A. et al. (2002) Dissociable contributions of prefrontal and 64 Wise, S.P. (2008) Forward frontal fields: phylogeny and fundamental
parietal cortices to response selection. Neuroimage 17, 15621571 function. Trends Neurosci. 31, 599608
37 Bunge, S.A. (2004) How we use rules to select actions: a review of 65 Miller, E.K. and Desimone, R. (1994) Parallel neuronal mechanisms for
evidence from cognitive neuroscience. Cogn. Affect. Behav. Neurosci. 4, short-term memory. Science 263, 520522
564579 66 Pauli, W.M. et al. (2010) Integrating what & how/where with
38 Moore, T.L. et al. (2009) Effects on executive function following damage instrumental and pavlovian learning: a biologically-based
to the prefrontal cortex in the rhesus monkey (macaca mulatta). Behav. computational model. In Cognition and Neuropsychology (Vol. 1)
Neurosci. 123, 231241 (Frensch, P.A. and Schwarzer, R., eds), Psychology Press, (in press)
39 Buckley, M.J. et al. (2009) Dissociable components of rule-guided 67 Rolls, E.T. et al. (2003) An information theoretic approach to the
behavior depend on distinct medial and prefrontal regions. Science contributions of the firing rates and the correlations between the
325, 5258 firing of neurons. J. Neurophysiol. 89, 28102822

361