《认知神经科学》课程教学资源（参考文献）[Miller, E. K., & Cohen, J. D.（2001）]An integrative theory of prefrontal cortex function.

AN INTEGRATIVE THEORY OF PREFRONTAL CORTEX FUNCTION Earl K.Miller Center for Learning and Memory.RIKEN-MIT Neuroscience Research Center and Department of Brain and Cognitive Sciences,Massachusetts Institute of Technology. Cambridge,Massachusetts 02139:e-mail:ekm@ai.mit.edu Jonathan D.Cohen Center for the Study of Brain,Mind,and Behavior and Department of Psychology. Princeton University.Princeton.New Jersey 08544:e-mail:jdc@princeton.edu Key Words frontal lobes,cognition,executive control,working memory,attention Abstract The prefrontal cortex has long been suspected to play an important role in cognitive control,in the ability to orchestrate thought and action in accordance with internal goals.Its neural basis,however,has remained a mystery.Here,we propose that cognitive control stems from the active maintenance of patterns of activity in the prefrontal cortex that represent goals and the means to achieve them.They provide bias signals to other brain structures whose net effect is to guide the flow of activity along neural pathways that establish the proper mappings between inputs,internal states,and outputs needed to perform a given task.We review neurophysiological neurobiological,neuroimaging,and computational studies that support this theory and discuss its implications as well as further issues to be addressed. INTRODUCTION One of the fundame tal,pur oseful behavio arises from the distrib ted activity of billions of neurons in the brair Simple behaviors can rely on relatively straightforward interactions between the brain's input and output systems.Animals with fewer than a hundred thousand neurons(in the human brain there are 100 billion or more neurons)can approach food andavoid predators.For animals with a arger brains,behavior is more flex ible.But flexibility carries a cost:Although our elab nsory and moto systems provide detailed information about the external world and make avail able a large repertoire of actions,this introduces greater potential for interference and confusion.The richer information we have about the world and the greater number of options for behavior require appropriate attentional,decision-making and coordinative functions,lest uncertainty prevail.To deal with this multitude of 0147-006X/01/0301-0167514.00 167

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 Annu. Rev. Neurosci. 2001. 24:167–202 Copyright c 2001 by Annual Reviews. All rights reserved AN INTEGRATIVE THEORY OF PREFRONTAL CORTEX FUNCTION Earl K. Miller Center for Learning and Memory, RIKEN-MIT Neuroscience Research Center and Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139; e-mail: ekm@ai.mit.edu Jonathan D. Cohen Center for the Study of Brain, Mind, and Behavior and Department of Psychology, Princeton University, Princeton, New Jersey 08544; e-mail: jdc@princeton.edu Key Words frontal lobes, cognition, executive control, working memory, attention ■ Abstract The prefrontal cortex has long been suspected to play an important role in cognitive control, in the ability to orchestrate thought and action in accordance with internal goals. Its neural basis, however, has remained a mystery. Here, we propose that cognitive control stems from the active maintenance of patterns of activity in the prefrontal cortex that represent goals and the means to achieve them. They provide bias signals to other brain structures whose net effect is to guide the flow of activity along neural pathways that establish the proper mappings between inputs, internal states, and outputs needed to perform a given task. We review neurophysiological, neurobiological, neuroimaging, and computational studies that support this theory and discuss its implications as well as further issues to be addressed. INTRODUCTION One of the fundamental mysteries of neuroscience is how coordinated, purposeful behavior arises from the distributed activity of billions of neurons in the brain. Simple behaviors can rely on relatively straightforward interactions between the brain’s input and output systems. Animals with fewer than a hundred thousand neurons (in the human brain there are 100 billion or more neurons) can approach food and avoid predators. For animals with larger brains, behavior is more flex￾ible. But flexibility carries a cost: Although our elaborate sensory and motor systems provide detailed information about the external world and make avail￾able a large repertoire of actions, this introduces greater potential for interference and confusion. The richer information we have about the world and the greater number of options for behavior require appropriate attentional, decision-making, and coordinative functions, lest uncertainty prevail. To deal with this multitude of 0147-006X/01/0301-0167$14.00 167 Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only

168 MILLER■COHEN possibilities and to curtail confusion,we have evolved mechanisms that coordinate ower-level sensory and motor processes along a common theme,an internal goal This ability for cognitive control no doubt involves neural circuitry that extends over much of the brain,but it is commonly held that the prefrontal cortex(PFC) ymportant that is most elaborated in primates,animals known for thei coordinate a wide range of neural processes:The PFC is a collection of intercon nected neocortical areas that sends and receives projections from virtually all cor tical sensory systems,motor systems,and many subcortical structures(Figure 1). Neurophysiological studies in nonhuman primates have begun to define many of the detailed properties of PFC,and hu neuropsyc ogy and neuroimaging studies have begun to provide a broad view of the task conditions under which it is engaged.However,an understanding of the mechanisms by which the PFC executes control has remained elusive.The aim ofthis article is to describe a theory of PFC function that integrates these diverse findings,and more precisely defines its role in cognitive control. The Role of the PFC in Top-Down Control of Behavior The PFC is not critical for performing simple,automatic behaviors,such as our mechanisms potentiate existing pathw ays or form ne ones "hardwired pathways are advantageous because they allow highly familiar behaviors to be executed quickly and automatically (i.e.without demanding attention).How- ever,these behaviors are inflexible,stereotyped reactions elicited by just the right stimulus.They do not generalize well to novel situations,and they take extensi time and experience to develop.These sorts ofauto omatic behaviors can be though of as relying primarily on"bottom-up"processing;that is,they are determined largely by the nature of the sensory stimuli and well-established neural pathways that connect these with corresponding responses. By contrast,the PFC is important when"top-down"processing is needed:that is when behaviormust be guided by internal statesor intentions.The PFCiscritical in situations when veen sensory inputs,thoughts and actionseithe re weakly established relative to other existing ones or are rapidly chan is when we need to use the "rules of the game,"internal representations of goals and the means to achieve them.Several investigators have argued that this is a cardinal function of the PFC(Cohen Servan-Schreiber 1992 Passingham 1993 Grafman 1994.Wise et al 1996,Miller 1999).Two classic tasks illustrate this point:the Stroop task k and the Wisconsin card sort task(WCST) In the Stroop task(Stroop 1935,MacLeod 1991),subjects either read words or name the color in which they are written.To perform this task,subjects must selectively attend to one attribute.This is especially so when naming the color

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 168 MILLER ¥ COHEN possibilities and to curtail confusion, we have evolved mechanisms that coordinate lower-level sensory and motor processes along a common theme, an internal goal. This ability for cognitive control no doubt involves neural circuitry that extends over much of the brain, but it is commonly held that the prefrontal cortex (PFC) is particularly important. The PFC is the neocortical region that is most elaborated in primates, animals known for their diverse and flexible behavioral repertoire. It is well positioned to coordinate a wide range of neural processes: The PFC is a collection of intercon￾nected neocortical areas that sends and receives projections from virtually all cor￾tical sensory systems, motor systems, and many subcortical structures (Figure 1). Neurophysiological studies in nonhuman primates have begun to define many of the detailed properties of PFC, and human neuropsychology and neuroimaging studies have begun to provide a broad view of the task conditions under which it is engaged. However, an understanding of the mechanisms by which the PFC executes control has remained elusive. The aim of this article is to describe a theory of PFC function that integrates these diverse findings, and more precisely defines its role in cognitive control. The Role of the PFC in Top-Down Control of Behavior The PFC is not critical for performing simple, automatic behaviors, such as our tendency to automatically orient to an unexpected sound or movement. These behaviors can be innate or they can develop gradually with experience as learning mechanisms potentiate existing pathways or form new ones. These “hardwired” pathways are advantageous because they allow highly familiar behaviors to be executed quickly and automatically (i.e. without demanding attention). How￾ever, these behaviors are inflexible, stereotyped reactions elicited by just the right stimulus. They do not generalize well to novel situations, and they take extensive time and experience to develop. These sorts of automatic behaviors can be thought of as relying primarily on “bottom-up” processing; that is, they are determined largely by the nature of the sensory stimuli and well-established neural pathways that connect these with corresponding responses. By contrast, the PFC is important when “top-down” processing is needed; that is, when behavior must be guided by internal states or intentions. The PFC is critical in situations when the mappings between sensory inputs, thoughts, and actions either are weakly established relative to other existing ones or are rapidly changing. This is when we need to use the “rules of the game,” internal representations of goals and the means to achieve them. Several investigators have argued that this is a cardinal function of the PFC (Cohen & Servan-Schreiber 1992, Passingham 1993, Grafman 1994, Wise et al 1996, Miller 1999). Two classic tasks illustrate this point: the Stroop task and the Wisconsin card sort task (WCST). In the Stroop task (Stroop 1935, MacLeod 1991), subjects either read words or name the color in which they are written. To perform this task, subjects must selectively attend to one attribute. This is especially so when naming the color Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only

PREFRONTAL FUNCTION 169 .29al9 Sensory cortex Mid-dorsal area9 Dorsal Motor Dorsolatera Ventral structures area 46 Area ventrolateral (FEF) areas 12,45 Auditory Superior temporal gyrus Orbital and medial aeas10.11. Multimodal Basal 13.14 Ganglia Rostral superior 0a0装Nmmam temporal sulcus Thalamus Medial tempor lobe Figure 1 Schematic diagram of some of the extrinsic and intrinsic connections of the prefrontal cortex.The partial convergence of inputs from many brain systems and internal connections ofthe prefrontal cortex(PFC)may allow it to play a central role in the synthesis of diverse information needed for complex behavior.Most connections are reciprocal,the exceptions are indicated by arrows.The frontal eye field (FEF)has variously been considered either adjacent to,or part of,the PFC.Here,we compromise by depicting it as adjacent to, yet touching,the PFC

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 PREFRONTAL FUNCTION 169 Figure 1 Schematic diagram of some of the extrinsic and intrinsic connections of the prefrontal cortex. The partial convergence of inputs from many brain systems and internal connections of the prefrontal cortex (PFC) may allow it to play a central role in the synthesis of diverse information needed for complex behavior. Most connections are reciprocal; the exceptions are indicated by arrows. The frontal eye field (FEF) has variously been considered either adjacent to, or part of, the PFC. Here, we compromise by depicting it as adjacent to, yet touching, the PFC. Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only

170 MILLER■COHEN ofa conict stimulus (the word GREEN displayed n red),because there is a strong prepotent tendency to read the word("green").which competes with the response to the color("red").This illustrates one of the most fundamental aspects of cognitive control and goal-directed behavior:the ability to select a weaker task-relevant response(or source of information)in the face of competition from ,but task-irrelevantone.Patients with frontal h this task (eg.Perrett 19,CohenServan-Schreiber Vendrell et al 1995),especially when the instructions vary frequently(Dunbar& Sussman 1995,Cohenetal 1999),which suggests that they have difficulty adhering to the goal of the task or its rules in the face of a competing stronger (ie.more salient or habitual)response. Similar findir ng s are evident in the WCST.Subjects are instructed to sort card according to the shape,color,or number of symbols appearing on them and the sorting rule varies periodically.Thus,any given card can be associated with several possible actions,no single stimulus-response mapping will work,and the correct one changes and is dictated by whichever rule is currently in effect.Humans with PEC dar nage show stere ch di culty but ar unable to adapt their beh Mner 1)Monkeys with PFClesonsrrd n analog of this task(Dias et al 1996b,1997)and in others when they must switch between different rules(Rossi et al 1999). The Stroop task and WCST are variously described as tapping the cognitive functions of either selective atte or rule-bas workin ed or goal-directed behavio we functions depend on the representation of goals and rules in the form of patterns of activity in the PFC,which configure processing in other parts of the brain in accordance with current task demands.These top-down signals favor weak(but task-relevant)stimulus-response mappings when they are in competition with more habitual,strongero he Stroop task),especially whe needed (such as in the WCST).We believe that this can account for the wide range of other tasks found to be sensitive to PFC damage,such as A-not-B(Piaget 1954. Diamond Goldman-Rakic 1989),Tower of London(Shallice 1982,1988;Owen et al 1990).and others (Duncan 1986.Duncan et al 1996).Stuss Benson 1986). We build on the fundamental principle that processing in the brain is compet itive:Different carrying different source s of inf ormation,compete for expression in behavior,and the winners are those with the strongest sources ofsup port.Desimone Duncan(1995)have proposed a model that clearly articulates such a view with regard to visual attention.These authors assume that visual corti- cal neurons processing different aspects ofa scene compete witheachother viamu- tually inhibi inter action The neur ns that"win active reach higher levels hey share inhibitory interactions.Voluntary shifts of attention result from the influence of excitatory top-down signals representing the to-be-attended features of the scene.These bias the competition among neurons representing the scene,increasing the activity of

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 170 MILLER ¥ COHEN of a conflict stimulus (e.g. the word GREEN displayed in red), because there is a strong prepotent tendency to read the word (“green”), which competes with the response to the color (“red”). This illustrates one of the most fundamental aspects of cognitive control and goal-directed behavior: the ability to select a weaker, task-relevant response (or source of information) in the face of competition from an otherwise stronger, but task-irrelevant one. Patients with frontal impairment have difficulty with this task (e.g. Perrett 1974, Cohen & Servan-Schreiber 1992, Vendrell et al 1995), especially when the instructions vary frequently (Dunbar & Sussman 1995, Cohen et al 1999), which suggests that they have difficulty adhering to the goal of the task or its rules in the face of a competing stronger (i.e. more salient or habitual) response. Similar findings are evident in the WCST. Subjects are instructed to sort cards according to the shape, color, or number of symbols appearing on them and the sorting rule varies periodically. Thus, any given card can be associated with several possible actions, no single stimulus-response mapping will work, and the correct one changes and is dictated by whichever rule is currently in effect. Humans with PFC damage show stereotyped deficits in the WCST. They are able to acquire the initial mapping without much difficulty but are unable to adapt their behavior when the rule varies (Milner 1963). Monkeys with PFC lesions are impaired in an analog of this task (Dias et al 1996b, 1997) and in others when they must switch between different rules (Rossi et al 1999). The Stroop task and WCST are variously described as tapping the cognitive functions of either selective attention, behavioral inhibition, working memory, or rule-based or goal-directed behavior. In this article, we argue that all these functions depend on the representation of goals and rules in the form of patterns of activity in the PFC, which configure processing in other parts of the brain in accordance with current task demands. These top-down signals favor weak (but task-relevant) stimulus-response mappings when they are in competition with more habitual, stronger ones (such as in the Stroop task), especially when flexibility is needed (such as in the WCST). We believe that this can account for the wide range of other tasks found to be sensitive to PFC damage, such as A-not-B (Piaget 1954, Diamond & Goldman-Rakic 1989), Tower of London (Shallice 1982, 1988; Owen et al 1990), and others (Duncan 1986, Duncan et al 1996), Stuss & Benson 1986). We build on the fundamental principle that processing in the brain is compet￾itive: Different pathways, carrying different sources of information, compete for expression in behavior, and the winners are those with the strongest sources of sup￾port. Desimone & Duncan (1995) have proposed a model that clearly articulates such a view with regard to visual attention. These authors assume that visual corti￾cal neurons processing different aspects of a scene compete with each other via mu￾tually inhibitory interactions. The neurons that “win” the competition and remain active reach higher levels of activity than those with which they share inhibitory interactions. Voluntary shifts of attention result from the influence of excitatory top-down signals representing the to-be-attended features of the scene. These bias the competition among neurons representing the scene, increasing the activity of Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only

PREFRONTAL FUNCTION 171 neurons representing the to-be-attended features and,by virtue of mutual inhi bition,suppressing activity of neurons processing other features.Desimone Duncan suggest that the PFC is an important source of such top-down biasing. However,they left unspecified the mechanisms by which this occurs.That is the focus of thisarticle. We begin by outlining a theory that extends the notion of biased competition and proposes that it provides a fundamental mechanism by which the PFC exerts control over a wide range ofprocesses in the service of goal-directed behavior.We describe the minimal set of functional properties that such a system must exhibit if it can serve as a mechanism of cognitive control.We then review the existing literature that provides suppor sion of recent computational modeling efforts that illustrate how a system with these properties can support elementary forms of control.Finally,we consider unresolved issues that provide a challenge for future empirical and theoretical research. Overview of the Theory We assume that the PFC serves a specific function in cognitive control:the active maintenance of patterns of activity that represent goals and the means to achieve them.They provide bias signals throughout much of the rest of the brain,affecting oyisual processes but also other sensory modalities.as well as systems responsible for r exe on,memory retrieva The aggregate effect of these bias signals is to guide the flow of neural activity along pathways that establish the proper mappings between inputs,internal states, and outputs needed to perform a given task.This is especially important whenever stimuli are ambiguous (ie they activate more than one input representation) or when multiple responses are possible and the task-appropriate resp compete w biases-which resolves competition,guides activity along appropriate pathways and establishes the mappings needed to perform the task-can be viewed as the neural implementation of attentional templates,rules,or goals,depending on the target of their biasing influence To heln understand how this Figure 2 They can be thought of as neural representations of sensory events.internal states (e.g.stored memories,emotions,etc),or combinations of these.Also shown are units corresponding to the motor circuits mediating two responses(RI and R2). pro cessing pathy e and response units We have set up the type of situ ation thought to be important.Namely,one cue(C1)can lead to either of two responses (RI or R2)depending on the situation(C2 or C3),and appropriate behavior de- pends on establishing the correct mapping from CI to RI or R2.For example imagine you are standing at the corner of a street(cue C1).Your natural reaction is to look left before crossing(R1),and this is the correct thing to do in most

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 PREFRONTAL FUNCTION 171 neurons representing the to-be-attended features and, by virtue of mutual inhi￾bition, suppressing activity of neurons processing other features. Desimone & Duncan suggest that the PFC is an important source of such top-down biasing. However, they left unspecified the mechanisms by which this occurs. That is the focus of this article. We begin by outlining a theory that extends the notion of biased competition and proposes that it provides a fundamental mechanism by which the PFC exerts control over a wide range of processes in the service of goal-directed behavior. We describe the minimal set of functional properties that such a system must exhibit if it can serve as a mechanism of cognitive control. We then review the existing literature that provides support for this set of properties, followed by a discussion of recent computational modeling efforts that illustrate how a system with these properties can support elementary forms of control. Finally, we consider unresolved issues that provide a challenge for future empirical and theoretical research. Overview of the Theory We assume that the PFC serves a specific function in cognitive control: the active maintenance of patterns of activity that represent goals and the means to achieve them. They provide bias signals throughout much of the rest of the brain, affecting not only visual processes but also other sensory modalities, as well as systems responsible for response execution, memory retrieval, emotional evaluation, etc. The aggregate effect of these bias signals is to guide the flow of neural activity along pathways that establish the proper mappings between inputs, internal states, and outputs needed to perform a given task. This is especially important whenever stimuli are ambiguous (i.e. they activate more than one input representation), or when multiple responses are possible and the task-appropriate response must compete with stronger alternatives. From this perspective, the constellation of PFC biases—which resolves competition, guides activity along appropriate pathways, and establishes the mappings needed to perform the task—can be viewed as the neural implementation of attentional templates, rules, or goals, depending on the target of their biasing influence. To help understand how this might work, consider the schematic shown in Figure 2. Processing units are shown that correspond to cues (C1, C2, C3). They can be thought of as neural representations of sensory events, internal states (e.g. stored memories, emotions, etc), or combinations of these. Also shown are units corresponding to the motor circuits mediating two responses (R1 and R2), as well as intervening or “hidden” units that define processing pathways between cue and response units. We have set up the type of situation for which the PFC is thought to be important. Namely, one cue (C1) can lead to either of two responses (R1 or R2) depending on the situation (C2 or C3), and appropriate behavior de￾pends on establishing the correct mapping from C1 to R1 or R2. For example, imagine you are standing at the corner of a street (cue C1). Your natural reaction is to look left before crossing (R1), and this is the correct thing to do in most Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only

172 MILLER■COHEN of the world(C2).However,if you are in England(C3),you should look righ (R2).This is a classic example of a circumstance requiring cognitive control which we assume depends on the PFC.How does the PFC mediate the correct behavior? We assume that cues in the environment activate internal representations within 0 facti habitual or more salient)but produces the incorrect behavior.Thus,standing at the corner(C1),your"automatic"response would be to look left(RI).However,other cues in the environment"remind"you that you are in England(C3).That is,the cues activate the corresponding PFC representation,which includes information about the appropriate action.This produ activity along the pathway leading you to look right (e.g.CI that activation of this PFC representation is necessary for you to perform the correct behavior.That is,you had to keep"in mind"the knowledge that you were in England.You might even be able to cross a few streets correctly while keeping presentation antained nhorekel to revert to themore if this activity s tha is,ifyou you are in England tual response and look left.Repeated selection can strengthen the pathway from CI to R2 and allow it to become independent of the PFC.As this happens,the behavior becomes more automatic,so you can look right without having to keep in mind that you are in En gland.An rtant question is how the PFC develops the needed to prod repre sentations ce the contextually appropriate response In an unfamiliar situation you may try various behaviors to achieve a desired goal,perhaps starting with some that have been useful in a similar circumstance (looking to the left for oncoming traffic)and,if these fail,trying others until you meet with success (e.g.by looking right).We assume that each of these is as- iated with son patter of activity within the PFC(as in Figure )When a pattern of activity by strengthening connections between the PFC neurons acti vated by that behavior.This process also strengthens connections between these neurons and those whose activity represents the situation in which the behavior was useful,establishing an association between these circumstances and the PFC eated iterations the PFC representation can be further slaborated as subter combi and rep nations of events and contingencies between them and the requisite actions are learned.As is discussed below.brainstem neuromodulatory systems may provide the relevant reinforcement signals,allowing the system to"bootstrap"in this way Obviously,many details ed to he added hef we fully unders stand the c plexity of cognitive control.But we believe that this general notio can explair many of the posited functions of the PFC.The biasing influence of PFC feedback signals on sensory systems may mediate its role in directing attention(Stuss Benson 1986:Knight 1984,1997;Banich et al 2000),signals to the motor system

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 172 MILLER ¥ COHEN of the world (C2). However, if you are in England (C3), you should look right (R2). This is a classic example of a circumstance requiring cognitive control, which we assume depends on the PFC. How does the PFC mediate the correct behavior? We assume that cues in the environment activate internal representations within the PFC that can select the appropriate action. This is important when the course of action is uncertain, and especially if one of the alternatives is stronger (i.e. more habitual or more salient) but produces the incorrect behavior. Thus, standing at the corner (C1), your “automatic” response would be to look left (R1). However, other cues in the environment “remind” you that you are in England (C3). That is, the cues activate the corresponding PFC representation, which includes information about the appropriate action. This produces excitatory bias signals that guide neural activity along the pathway leading you to look right (e.g. C1 →···→ R2). Note that activation of this PFC representation is necessary for you to perform the correct behavior. That is, you had to keep “in mind” the knowledge that you were in England. You might even be able to cross a few streets correctly while keeping this knowledge in mind, that is, while activity of the appropriate representation is maintained in the PFC. However, if this activity subsides—that is, if you “forget” you are in England—you are likely to revert to the more habitual response and look left. Repeated selection can strengthen the pathway from C1 to R2 and allow it to become independent of the PFC. As this happens, the behavior becomes more automatic, so you can look right without having to keep in mind that you are in England. An important question is how the PFC develops the representations needed to produce the contextually appropriate response. In an unfamiliar situation you may try various behaviors to achieve a desired goal, perhaps starting with some that have been useful in a similar circumstance (looking to the left for oncoming traffic) and, if these fail, trying others until you meet with success (e.g. by looking right). We assume that each of these is as￾sociated with some pattern of activity within the PFC (as in Figure 2). When a behavior meets with success, reinforcement signals augment the corresponding pattern of activity by strengthening connections between the PFC neurons acti￾vated by that behavior. This process also strengthens connections between these neurons and those whose activity represents the situation in which the behavior was useful, establishing an association between these circumstances and the PFC pattern that supports the correct behavior. With time (and repeated iterations of this process), the PFC representation can be further elaborated as subtler combi￾nations of events and contingencies between them and the requisite actions are learned. As is discussed below, brainstem neuromodulatory systems may provide the relevant reinforcement signals, allowing the system to “bootstrap” in this way. Obviously, many details need to be added before we fully understand the com￾plexity of cognitive control. But we believe that this general notion can explain many of the posited functions of the PFC. The biasing influence of PFC feedback signals on sensory systems may mediate its role in directing attention (Stuss & Benson 1986; Knight 1984, 1997; Banich et al 2000), signals to the motor system Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only

PREFRONTAL FUNCTION 173 may be responsible for response selection and inhibitory control (Fuster 1980. Diamond 1988),and signals to intermediate systems may support short-term (or working)memory (Goldman-Rakic 1987)and guide retrieval from long-term memory (Schachter 1997,Janowsky et al 1989,Gershberg Shimamura 1995). Wthou th PFC.the most used (and thus best established) eural path ways would predominate or,where these don'texist,behavior would be haphazard Such impulsive,inappropriate,or disorganized behavior is a hallmark of PFC dys- function in humans(e.g.Bianchi 1922,Duncan 1986,Luria 1969,Lhermitte 1983, Shallice Burgess 1996.Stuss Benson 1986). Minimal Requirements for a Mechanism of Top-Down Control There are several critical features of our theory.First,the PFC must provide a source of activity that can exert the required patter of biasing signals to other structures.We n thus think of PFC fund memory in the service of contre ol."It follows,theref e tha at the PF aintain its activity robustly against distractions until a goal is achieved,yet also be flexible enough to update its representations when needed.It must also house the appropriate representations, those that can select the neural pathways needed for the task.Insofar as primates are capable of tasks that involve diverse combinations of stimuli,internal states and respons ntations in the PFC must hav access to and be able to influence a similarly wide range of information in other brain regions.That is. PFC representations must have a high capacity for multimodality and integration Finally,as we can acquire new goals and means,the PFC must also exhibit a high degree of plasticity Of course.it must be possible to exhibit all these properties without the need to invoke some other mechanism of control to explain them,lest our theory be subject to perennial conc ns of a hidder nculus The rapidly accumulating body of findings regarding the PFC suggests that it meets these requirements.Fuster(1971,1973,1995),Goldman-Rakic(1987 1996),and others have extensively explored the ability of PFC neurons to main- tain task-relevant information.Miller et al(1996)have shown that this is robust to interferer e fron has lo eated the role of the PFCin integrating diverse ormation(Fuster 19 1995) The earliest de criptions c the effects of frontal lobe damage suggested its role in attention and the control of behavior(Ferrier 1876,Bianchi 1922),and investigators since have interpreted the pattern of deficits following pec damage as a loss of the ability to acquire and use rules (Shallice 1982.Duncan 1986.Passingham 1993,Grafmar 1994,Wise et al 1996).Recent empirical studies have begun identify correlates of plasticity in the PFC(Asaad et al 1998,Bichot et al 199,Schultz& Dickinson 2000),and recent computational studies suggest how these may operate as mechanisms for self-organization(Braver Cohen 2000,Egelman et al 1998) Our purpose in this article is to bring these various observations and arguments together,and to illustrate that a reaso coherent,and mechanistically explicit

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 PREFRONTAL FUNCTION 173 may be responsible for response selection and inhibitory control (Fuster 1980, Diamond 1988), and signals to intermediate systems may support short-term (or working) memory (Goldman-Rakic 1987) and guide retrieval from long-term memory (Schachter 1997, Janowsky et al 1989, Gershberg & Shimamura 1995). Without the PFC, the most frequently used (and thus best established) neural path￾ways would predominate or, where these don’t exist, behavior would be haphazard. Such impulsive, inappropriate, or disorganized behavior is a hallmark of PFC dys￾function in humans (e.g. Bianchi 1922, Duncan 1986, Luria 1969, Lhermitte 1983, Shallice & Burgess 1996, Stuss & Benson 1986). Minimal Requirements for a Mechanism of Top-Down Control There are several critical features of our theory. First, the PFC must provide a source of activity that can exert the required pattern of biasing signals to other structures. We can thus think of PFC function as “active memory in the service of control.” It follows, therefore, that the PFC must maintain its activity robustly against distractions until a goal is achieved, yet also be flexible enough to update its representations when needed. It must also house the appropriate representations, those that can select the neural pathways needed for the task. Insofar as primates are capable of tasks that involve diverse combinations of stimuli, internal states, and responses, representations in the PFC must have access to and be able to influence a similarly wide range of information in other brain regions. That is, PFC representations must have a high capacity for multimodality and integration. Finally, as we can acquire new goals and means, the PFC must also exhibit a high degree of plasticity. Of course, it must be possible to exhibit all these properties without the need to invoke some other mechanism of control to explain them, lest our theory be subject to perennial concerns of a hidden “homunculus.” The rapidly accumulating body of findings regarding the PFC suggests that it meets these requirements. Fuster (1971, 1973, 1995), Goldman-Rakic (1987, 1996), and others have extensively explored the ability of PFC neurons to main￾tain task-relevant information. Miller et al (1996) have shown that this is robust to interference from distraction. Fuster has long advocated the role of the PFC in integrating diverse information (Fuster 1985, 1995). The earliest descriptions of the effects of frontal lobe damage suggested its role in attention and the control of behavior (Ferrier 1876, Bianchi 1922), and investigators since have interpreted the pattern of deficits following PFC damage as a loss of the ability to acquire and use behavior-guiding rules (Shallice 1982, Duncan 1986, Passingham 1993, Grafman 1994, Wise et al 1996). Recent empirical studies have begun to identify neural correlates of plasticity in the PFC (Asaad et al 1998, Bichot et al 1996, Schultz & Dickinson 2000), and recent computational studies suggest how these may operate as mechanisms for self-organization (Braver & Cohen 2000, Egelman et al 1998). Our purpose in this article is to bring these various observations and arguments together, and to illustrate that a reasonably coherent, and mechanistically explicit, Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only

174 MILLER■COHEN theory of PFC function is beginning to emerge.The view presented here draws on previous work that has begun to outline such a theory (e.g.Coh n Servan Schreiber 1992:Cohen et al 1996;O'Reilly et al 1999;Miller 1999,2000).In the sections that follow,we review neurobiological,neuropsychological,and neu- roimaging findings that support this theory,and computational modeling studies that have begun to make explicit the processing mechanisms involved. PROPERTIES OF THE PFC Convergence of Diverse Information One of the critical features for a system of cognitive control is the requirement that it have access to diverse information about both the internal state of the and the of the ord.The PFCs natomically well sit this requirement.The cytoarchitectonic areas that comprise the monkey PFC are often grouped into regional subdivisions,the orbital and medial,the lateral,and the mid-dorsal(see Figure 1).Collectively,these areas have interconnections with virtually all sensory systems,with cortical and subcortical motor system structures, and with limbic and midbrain structures involved in affect,me mory,and reward The subdivisions have partly unique,but overlapping.patterns of connections with the rest of the brain,which suggests some regional specialization.However,as in much of the neocortex many pFC connections are local there are extensive connections between different PFC areas that are likely to support an intermixing of di parate information.Such intermixing provides a basis for synthesizing results from,andc oordinating the regulation of ide variety ofbra s would be required of a brain area responsible Sensory Inputs The lateral and mid-dorsal PFC is more closely ass ociated with sensory neocortex than is the ventromedial PFC(see Figure 1).It receives visual somatosensory,and auditory information from the occipital,temporal,and pari etal cortices (Barbas Pandya 1989,1991;Goldman-Rakic Schwartz 1982 Pandya Barnes 1987:Pandya Yeterian 1990:Petrides Pandya 1984.1999: Pandya 199).Many PFC areas receive converging inputsfm two sen (Cha vis Pa 11970).Fo lethe(DL (areas8gana4 and(12级） PFC both receive projections from visual,auditory,and somatosensory cortex Furthermore,the PFC is connected with other cortical regions that are themselves sites of multimodal convergence.Many PFC areas (9,12,46,and 45)receive inputs from the ostral s mporal sulcus,which has n with bimoda or trimodal (visual,auditory,and somatosensory)responses (Br ce et al 1981 Pandya Barnes 1987).The arcuate sulcus region(areas 8 and 45)and area 12 seem to be particularly multimodal.They contain zones that receive overlapping inputs from three sensory modalities(Pandya Barnes 1987).In all these cases

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 174 MILLER ¥ COHEN theory of PFC function is beginning to emerge. The view presented here draws on previous work that has begun to outline such a theory (e.g. Cohen & Servan￾Schreiber 1992; Cohen et al 1996; O’Reilly et al 1999; Miller 1999, 2000). In the sections that follow, we review neurobiological, neuropsychological, and neu￾roimaging findings that support this theory, and computational modeling studies that have begun to make explicit the processing mechanisms involved. PROPERTIES OF THE PFC Convergence of Diverse Information One of the critical features for a system of cognitive control is the requirement that it have access to diverse information about both the internal state of the system and the external state of the world. The PFC is anatomically well situated to meet this requirement. The cytoarchitectonic areas that comprise the monkey PFC are often grouped into regional subdivisions, the orbital and medial, the lateral, and the mid-dorsal (see Figure 1). Collectively, these areas have interconnections with virtually all sensory systems, with cortical and subcortical motor system structures, and with limbic and midbrain structures involved in affect, memory, and reward. The subdivisions have partly unique, but overlapping, patterns of connections with the rest of the brain, which suggests some regional specialization. However, as in much of the neocortex, many PFC connections are local; there are extensive connections between different PFC areas that are likely to support an intermixing of disparate information. Such intermixing provides a basis for synthesizing results from, and coordinating the regulation of, a wide variety of brain processes, as would be required of a brain area responsible for the orchestration of complex behavior. Sensory Inputs The lateral and mid-dorsal PFC is more closely associated with sensory neocortex than is the ventromedial PFC (see Figure 1). It receives visual, somatosensory, and auditory information from the occipital, temporal, and pari￾etal cortices (Barbas & Pandya 1989, 1991; Goldman-Rakic & Schwartz 1982; Pandya & Barnes 1987; Pandya & Yeterian 1990; Petrides & Pandya 1984, 1999; Seltzer & Pandya 1989). Many PFC areas receive converging inputs from at least two sensory modalities (Chavis & Pandya 1976; Jones & Powell 1970). For ex￾ample, the dorsolateral (DL) (areas 8, 9, and 46) and ventrolateral (12 and 45) PFC both receive projections from visual, auditory, and somatosensory cortex. Furthermore, the PFC is connected with other cortical regions that are themselves sites of multimodal convergence. Many PFC areas (9, 12, 46, and 45) receive inputs from the rostral superior temporal sulcus, which has neurons with bimodal or trimodal (visual, auditory, and somatosensory) responses (Bruce et al 1981, Pandya & Barnes 1987). The arcuate sulcus region (areas 8 and 45) and area 12 seem to be particularly multimodal. They contain zones that receive overlapping inputs from three sensory modalities (Pandya & Barnes 1987). In all these cases, Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only

PREFRONTAL FUNCTION 175 the PFC is directly connected with secondary or"association"but not primary sensory cortex. Motor Ouputs The dorsal PFC,particularly DL area 46,has preferential con nections with motor system structures that may be central to how the PFC exerts control over behavior.The DL area 46 is interconnected (a)with motor areas in the medial frontal lobe such as the supplementary motor area,the pre-supplementary motor area,and the rostral cingulate,(b)with the premotor cortex on the lateral and (c)with cerebellum olliculus(Bates&Goldm Rakic 193.GoldmanNauta 176.Lu.SchmahmannPandya 1997).The DL area 46 also sends projections to area 8,which contains the frontal eye fields,a region important for voluntary shifts of gaze.There are no direct connections between the PFC and primary motor cortex,but they are extensive with pre otor a eas that,in turn,send pr pjections to primary motor cortex and the spinal ord.Also mporant are the dense intercc be een the PFC and basal ganglia(Alexander et al 1986),a structure that is likely to be crucial for automating behavior.The basal ganglia receives inputs from much of the cerebral cortex,but its major output(via the thalamus)is frontal cortex(see Figure 1). Limbic Connections The orbital and medial PFC are closely associated with medial temporal limbic structures critical for long-term memory and the processing of internal states,such as affect and motivation.This includes direct and indirect (via the medial dorsal thalamus)connections with the hippocampus and associated ygdala,and theh P naral Price 1984.Barbas 990,Bar &Pandya 1989,Goldman-Rakic et al 1984,Porrino eta 1981,Van Hoesenetal 1972).Other PFCregions have access to these systems both through connections with the orbital and medial PFC and through other intervening structures. Intrinsic Connections Most PFC regions are interconnected with most other PFC regions.There are not only interconnections between all three major subdi- visions(ventromedial,lateral,and mid-dorsal)but also between their constituent areas (Barbas Pandva 1991 Pandva barnes 1987)The lateral pec is partic ularly well connected.Ventrolateral areas 12 and 45 are int nnected with DI well as with ventromedial areas and13 Intrinsic connections within the PFC allow information from regional afferents and processes to be distributed to other parts of the PFC.Thus,the PFC provides a venue by which information from wide-ranging brain systems can interact through relatively local circuitry Convergence and Plasticity Given that goal-directed behavior depends on our ability to piece together rela- tionships between a wide range of exte mal and internal information,it stands to

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 PREFRONTAL FUNCTION 175 the PFC is directly connected with secondary or “association” but not primary sensory cortex. Motor Outputs The dorsal PFC, particularly DL area 46, has preferential con￾nections with motor system structures that may be central to how the PFC exerts control over behavior. The DL area 46 is interconnected (a) with motor areas in the medial frontal lobe such as the supplementary motor area, the pre–supplementary motor area, and the rostral cingulate, (b) with the premotor cortex on the lateral frontal lobe, and (c) with cerebellum and superior colliculus (Bates & Goldman￾Rakic 1993, Goldman & Nauta 1976, Lu et al 1994, Schmahmann & Pandya 1997). The DL area 46 also sends projections to area 8, which contains the frontal eye fields, a region important for voluntary shifts of gaze. There are no direct connections between the PFC and primary motor cortex, but they are extensive with premotor areas that, in turn, send projections to primary motor cortex and the spinal cord. Also important are the dense interconnections between the PFC and basal ganglia (Alexander et al 1986), a structure that is likely to be crucial for automating behavior. The basal ganglia receives inputs from much of the cerebral cortex, but its major output (via the thalamus) is frontal cortex (see Figure 1). Limbic Connections The orbital and medial PFC are closely associated with medial temporal limbic structures critical for long-term memory and the processing of internal states, such as affect and motivation. This includes direct and indirect (via the medial dorsal thalamus) connections with the hippocampus and associated neocortex, the amygdala, and the hypothalamus (Amaral & Price 1984, Barbas & De Olmos 1990, Barbas & Pandya 1989, Goldman-Rakic et al 1984, Porrino et al 1981, Van Hoesen et al 1972). Other PFC regions have access to these systems both through connections with the orbital and medial PFC and through other intervening structures. Intrinsic Connections Most PFC regions are interconnected with most other PFC regions. There are not only interconnections between all three major subdi￾visions (ventromedial, lateral, and mid-dorsal) but also between their constituent areas (Barbas & Pandya 1991, Pandya & Barnes 1987). The lateral PFC is partic￾ularly well connected. Ventrolateral areas 12 and 45 are interconnected with DL areas 46 and 8, with dorsal area 9, as well as with ventromedial areas 11 and 13. Intrinsic connections within the PFC allow information from regional afferents and processes to be distributed to other parts of the PFC. Thus, the PFC provides a venue by which information from wide-ranging brain systems can interact through relatively local circuitry. Convergence and Plasticity Given that goal-directed behavior depends on our ability to piece together rela￾tionships between a wide range of external and internal information, it stands to Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only

176 MILLER■COHEN reason that top-down control must come from PFC representations that reflect a wide range of learned associations. There is mounting neurophysiological evi dence that this is the case.Asaad et al (1998)trained monkeys to associate,on different blocks of trials,each of two cue objects with a saccade to the right or a saccade to the left They found relatively few lateral pe neurons whose activity ween a visual cue and a directional saccade it instructed.For example,a given cell might only be strongly activated when object“A”instructed“saccade left”and not when object“B”instructed the same saccade or when object“A”instructed another sac cade(Figure 3A).Lateral PFC neurons can also convey the degree of association between a cue e and a response(Quintana&Fuster 1992) Other studies indicate that PFC neurons acquire selectivity for features to which they are initially insensitive but are behaviorally relevant.For example,Bichot etal(1996)observed that neurons in the frontal eye fields(in the bow of the arcuate sulcusordinarily not selective to the form and color ofstimuli-became so as the animal learned eye movements that were contingent on these features.Similarly Watanabe(1990 1992)ha ed monkeys to recogn that certa visual and auditory stimuli signaled whether or not,on different trials,a reward(a drop of juice)would be delivered.He found that neurons in lateral PFC(around the arcuate sulcus and posterior end of the principal sulcus)came to reflect specific cue-reward associations.For example,a given neuron could show strong activation to one of the two auditory (and the visual)cu signaled Other neuror strongly modulated by their reward status. More complicated behaviors depend not on simple contingencies between cues and responses or rewards but on general principles or rules that may involve more. complex mapping.PFC activity also seems to represent this information.Barone oseph(199)observed cells near the arcuate sulcus that were responsive to spe cific light stimuli,but only when they occurred at a particular point in a particular Figure 3 ()Shown is the activity of four single prefrontal (PF)neurons when each of two objects,on different trials,instructed either a saccade to the right or a saccade to the left The lines connect the average values obtained when a given object cued one or the other saccade.The error bars show the standard error of the mean.Note that in each case,the neuron's activity depends on both the cue object and the saccade direction and that the tuning is nonlinear or conjunctive.That is,the level of activity to a given combination of object and saccade cannot be predicted from the neuron's response to the other combinations [Adapted from Asaad et al(1998).](B)A PF neuron whose neural response to a cue object was highly dependent on task context.The bottom half shows an example of a single PF neuron's response to the same cue object during an object task(delayed matching to sample)and during an associative task(conditional visual motor).Note that the neuron is responsive to the cue during one task but not during the other,even though sensory stimulation is identical across the tasks.[Adapted from Asaad et al (2000).]

P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 176 MILLER ¥ COHEN reason that top-down control must come from PFC representations that reflect a wide range of learned associations. There is mounting neurophysiological evi￾dence that this is the case. Asaad et al (1998) trained monkeys to associate, on different blocks of trials, each of two cue objects with a saccade to the right or a saccade to the left. They found relatively few lateral PF neurons whose activity simply reflected a cue or response. Instead, the modal group of neurons (44% of the population) showed activity that reflected the current association between a visual cue and a directional saccade it instructed. For example, a given cell might only be strongly activated when object “A” instructed “saccade left” and not when object “B” instructed the same saccade or when object “A” instructed another sac￾cade (Figure 3A). Lateral PFC neurons can also convey the degree of association between a cue and a response (Quintana & Fuster 1992). Other studies indicate that PFC neurons acquire selectivity for features to which they are initially insensitive but are behaviorally relevant. For example, Bichot et al (1996) observed that neurons in the frontal eye fields (in the bow of the arcuate sulcus)—ordinarily not selective to the form and color of stimuli—became so as the animal learned eye movements that were contingent on these features. Similarly, Watanabe (1990, 1992) has trained monkeys to recognize that certain visual and auditory stimuli signaled whether or not, on different trials, a reward (a drop of juice) would be delivered. He found that neurons in lateral PFC (around the arcuate sulcus and posterior end of the principal sulcus) came to reflect specific cue-reward associations. For example, a given neuron could show strong activation to one of the two auditory (and none of the visual) cues, but only when it signaled reward. Other neurons were bimodal, activated by both visual and auditory cues but also strongly modulated by their reward status. More complicated behaviors depend not on simple contingencies between cues and responses or rewards but on general principles or rules that may involve more￾complex mapping. PFC activity also seems to represent this information. Barone & Joseph (1989) observed cells near the arcuate sulcus that were responsive to spe￾cific light stimuli, but only when they occurred at a particular point in a particular −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→ Figure 3 (A) Shown is the activity of four single prefrontal (PF) neurons when each of two objects, on different trials, instructed either a saccade to the right or a saccade to the left. The lines connect the average values obtained when a given object cued one or the other saccade. The error bars show the standard error of the mean. Note that in each case, the neuron’s activity depends on both the cue object and the saccade direction and that the tuning is nonlinear or conjunctive. That is, the level of activity to a given combination of object and saccade cannot be predicted from the neuron’s response to the other combinations. [Adapted from Asaad et al (1998).] (B) A PF neuron whose neural response to a cue object was highly dependent on task context. The bottom half shows an example of a single PF neuron’s response to the same cue object during an object task (delayed matching to sample) and during an associative task (conditional visual motor). Note that the neuron is responsive to the cue during one task but not during the other, even though sensory stimulation is identical across the tasks. [Adapted from Asaad et al (2000).] Annu. Rev. Neurosci. 2001.24:167-202. Downloaded from arjournals.annualreviews.org by University of California - Los Angeles on 03/27/06. For personal use only