PREFRONTAL FUNCTION 177 A 25 Righ B F Delay. Associative task Object task 00 0 500 1000 1500 Time(ms)
P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 PREFRONTAL FUNCTION 177 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
178 MILLER■COHEN sequence that the monkey had to imitate.White&Wise(1999)trained a monkey to orient to a visual target according to two different rules.One of four cue patterns briefly appeared at one of four locations.The cue indicated where a target would eventually appear.It did so by one of two rules,aspatial rule(the cue appeared at the same location that the target would appear)or anassociative rule(the identity ofthe ue instructed the locatio to昌Crshow rule.Another example was provided by Asaad et al(2000),who trained monkeys to alternate between tasks that employed the same cues and responses but three different rules:matching(delayed matching to sample),associative(conditional visuomotor),and spatial (spatial delayed response).Over half of lateral PFC neurons were rule dependen Neur de ofen depended on which uleurntu coming the baseline activity of many neurons(54%)varied with the rule.Hoshiet al(1998) have also observed PFC neurons that were modulated by whether the monkey was using a shape-matching or location-matching rule.Recently.Wallis et al (2000) have shown that lateral and orbitofrontal PFC neurons reflect whether the monkey Sui of onkeysand umh PPC is currently using a"matc ing”or"nonma critical for learning rules.For example.Petrides found that following PFC damage patients could no longer learn arbitrary associations between visual patterns and hand gestures (Petrides 1985,1990).In monkeys,damage to ventrolateral area 12 or to the arcuate sulcus region also imp rs the abi response associations(Halsband d&Passingham 1985,Mu des 1982,1985).Leaming of visual stimulus-response conditional associations is also impaired by damage to PFC inputs from the temporal cortex(Eacott Gaffan 1992 Gaffan Harrison 1988,Parker Gaffan 1998).Passingham(1993)argues that most,ifnot all,tasks that are disrupted following PFC damage depend onacquiring conditional associations(if-the en rules). In sum,these results indicate that PFC neural activity represents the rules or mappings required to perform a particular task,and not just single stimuli or forthcoming actions.We assume that this activity within the PFC establishes these mappings by biasing competition in other parts of the brain responsible for actually perforn the task.These sigals favor task-relevant sensory inputs(). Feedback to Other Brain Areas provide these feedback signals;it sends projections to much of the neocortex (Pandya Barnes 1987,Pandya Yeterian 1990).Physiological studies have yielded results consistent with this notion
P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 178 MILLER ¥ COHEN sequence that the monkey had to imitate. White & Wise (1999) trained a monkey to orient to a visual target according to two different rules. One of four cue patterns briefly appeared at one of four locations. The cue indicated where a target would eventually appear. It did so by one of two rules, a spatial rule (the cue appeared at the same location that the target would appear) or an associative rule (the identity of the cue instructed the location, e.g. cue A indicated the right, cue B the left, etc). They found that up to half of lateral PFC neurons showed activity that varied with the rule. Another example was provided by Asaad et al (2000), who trained monkeys to alternate between tasks that employed the same cues and responses but three different rules: matching (delayed matching to sample), associative (conditional visuomotor), and spatial (spatial delayed response). Over half of lateral PFC neurons were rule dependent. Neural responses to a given cue or forthcoming saccade often depended on which rule was current in the task (Figure 3B). Plus, the baseline activity of many neurons (54%) varied with the rule. Hoshi et al (1998) have also observed PFC neurons that were modulated by whether the monkey was using a shape-matching or location-matching rule. Recently, Wallis et al (2000) have shown that lateral and orbitofrontal PFC neurons reflect whether the monkey is currently using a “matching” or “nonmatching” rule to select a test object. Studies of monkeys and humans with PFC damage also suggest that the PFC is critical for learning rules. For example, Petrides found that following PFC damage, patients could no longer learn arbitrary associations between visual patterns and hand gestures (Petrides 1985, 1990). In monkeys, damage to ventrolateral area 12 or to the arcuate sulcus region also impairs the ability to learn arbitrary cueresponse associations (Halsband & Passingham 1985, Murray et al 2000, Petrides 1982, 1985). Learning of visual stimulus-response conditional associations is also impaired by damage to PFC inputs from the temporal cortex (Eacott & Gaffan 1992, Gaffan & Harrison 1988, Parker & Gaffan 1998). Passingham (1993) argues that most, if not all, tasks that are disrupted following PFC damage depend on acquiring conditional associations (if-then rules). In sum, these results indicate that PFC neural activity represents the rules, or mappings required to perform a particular task, and not just single stimuli or forthcoming actions. We assume that this activity within the PFC establishes these mappings by biasing competition in other parts of the brain responsible for actually performing the task. These signals favor task-relevant sensory inputs (attention), memories (recall), and motor outputs (response selection) and thus guide activity along the pathways that connect them (conditional association). Feedback to Other Brain Areas Our model of PFC function requires feedback signals from the PFC to reach widespread targets throughout the brain. The PFC has the neural machinery to provide these feedback signals; it sends projections to much of the neocortex (Pandya & Barnes 1987, Pandya & Yeterian 1990). Physiological studies have yielded results consistent with this notion. 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 179 Fuster et al (1985)and Chafee&Goldman-Rakic (2000)have found that de- activating the lateral PFC cortex attenuates the activity of visual cortical(inferior temporal and posterior parietal)neurons to a behaviorally relevant cue.Tomita et al (1999)directly explored the role of top-down PFC signals in the recall of visual memorie stored in the inferior temporal ( Ap earance of a cue object instructed monkeys to recall and then choose another object that was as sociated with the cue during training.In the intact brain,information is shared between II cortices in the two cerebral hemispheres.By severing the connecting fibers each IT cortex could only"see"(receive bottom-up inputs from)visual the pEc in each hemisphere were left intact.When Tomita et al examined activity of single ne rons in an IT cortex that could not "see"the cue,it nonetheless reflected the recalled object,albeit with a long latency.It appeared that visual information took a circuitous route,traveling from the opposite IT cortex(which could"see" the cue)to the still-connected PFC in each hemisphere and then down to the as s confirmed by sev the PFC in the here and eliminating the feedback,whicha ITactivity and disrupted task performance. Other evidence suggestive of PFC-IT interactions also comes from investiga- tions,by Miller Desimone (1994)and Miller et al (1996),of the respective roles of the PC and I cortex in working memoryin each trial.monkevs were shown first a sar mple stimu Then,one t stimuli appeared in sequence.If a test stimulus matched the sample,the monkey indicated so by releasing a lever.Sometimes,one of the intervening nonmatch stimuli could be repeated.For example,the sample stimulus“A”might be followed by“B.…B. C...A."The monkey was only rewarded for responding to the final match ("A") and thus had to maintain a specific rep presentation of the sample rather than re spond to any repetition any stimulus.Asnoted in the next sectio n,neurons found in the PFC that exhibited sustained sample-specific activity that survived the presentation of intervening distractors.This was not so for IT cortex.However, neurons in both areas showed a selective enhancement of responses to a match of the sample.The fact that IT neurons had not maintained a representation of this ssuggests that their enhanced response to the match n interactions the representation maintained in the PFC ight have resulted with the recent finding indicating that,in a target detection task,target-specific ac- tivity appears simultaneously within the PFC and the visual cortex(Anderson et al 1999).Together,these findings suggest that identification of an intended stimulus relies on nteractions between the PFC and the posterior cortex Active Maintenance If the PFC represents the rules of a task in its pattern of neural activity,it must maintain this activity as long as the rule is required.Usually this extends beyond the eliciting event and must span other intervening,irrelevant,and potentially
P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 PREFRONTAL FUNCTION 179 Fuster et al (1985) and Chafee & Goldman-Rakic (2000) have found that deactivating the lateral PFC cortex attenuates the activity of visual cortical (inferior temporal and posterior parietal) neurons to a behaviorally relevant cue. Tomita et al (1999) directly explored the role of top-down PFC signals in the recall of visual memories stored in the inferior temporal (IT) cortex. Appearance of a cue object instructed monkeys to recall and then choose another object that was associated with the cue during training. In the intact brain, information is shared between IT cortices in the two cerebral hemispheres. By severing the connecting fibers, each IT cortex could only “see” (receive bottom-up inputs from) visual stimuli in the contralateral visual field. The fibers connecting the PFC in each hemisphere were left intact. When Tomita et al examined activity of single neurons in an IT cortex that could not “see” the cue, it nonetheless reflected the recalled object, albeit with a long latency. It appeared that visual information took a circuitous route, traveling from the opposite IT cortex (which could “see” the cue) to the still-connected PFC in each hemisphere and then down to the “blind” IT cortex. This was confirmed by severing the PFC in the two hemispheres and eliminating the feedback, which abolished the IT activity and disrupted task performance. Other evidence suggestive of PFC-IT interactions also comes from investigations, by Miller & Desimone (1994) and Miller et al (1996), of the respective roles of the PFC and IT cortex in working memory. During each trial, monkeys were shown first a sample stimulus. Then, one to four test stimuli appeared in sequence. If a test stimulus matched the sample, the monkey indicated so by releasing a lever. Sometimes, one of the intervening nonmatch stimuli could be repeated. For example, the sample stimulus “A” might be followed by “B ... B ... C ... A.” The monkey was only rewarded for responding to the final match (“A”) and thus had to maintain a specific representation of the sample rather than respond to any repetition of any stimulus. As noted in the next section, neurons were found in the PFC that exhibited sustained sample-specific activity that survived the presentation of intervening distractors. This was not so for IT cortex. However, neurons in both areas showed a selective enhancement of responses to a match of the sample. The fact that IT neurons had not maintained a representation of this stimulus suggests that their enhanced response to the match might have resulted from interactions with the representation maintained in the PFC. This is consistent with the recent finding indicating that, in a target detection task, target-specific activity appears simultaneously within the PFC and the visual cortex (Anderson et al 1999). Together, these findings suggest that identification of an intended stimulus relies on interactions between the PFC and the posterior cortex. Active Maintenance If the PFC represents the rules of a task in its pattern of neural activity, it must maintain this activity as long as the rule is required. Usually this extends beyond the eliciting event and must span other intervening, irrelevant, and potentially 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
180 MILLER■COHEN interfering events.The capacity to support sustained activity in the face of interference is one of the distinguishing characteristics of the PFC Sustained neural activity within the PFC was first reported by Fuster(1971) and Kubota Niki (1971)and has subsequently been reported in a large number of studies These have demonstrated that neurons within the pec remain active ented cue and the later execution of c to aparticular type of information,such as the location and/or identity of a stimulus(di Pellegrino& Wise 1991;Funahashi et al 1989;Fuster 1973:Fuster Alexander 1971 Kubota Niki 1971;Rainer et al 1998a,b,1999:Rao et al 1997;Romo et al 1999),forthcoming actions (Asaad et al 1998,Ferrera et al 1999,Quintana Fuste r1992). expected rewards (Leon&Shadlen 1999,Tremblay et al 199 Watanabe )and more-compepropertiessuch as the sequential positionof a stimulus within an ordered series (Barone Joseph 1989)or a particular as sociation between a stimulus and its corresponding response(Asaad et al 1998). Functional neuroimaging studies have begun to yield similar results with humans (Cohen et al 1997,Courtney et al 1997,Prabhakaran et al 2000). Other areas ofthe brainexhibit a form ofsusta ned activ Forexample in many cortical visual areas,abriefvisual stimulus will evoke activity that persists from several hundred milliseconds to several seconds (Fuster Jervey 1981 Gnadt Andersen 1988,Miller et al 1993.Miyashita Chang 1988).What appears to distinguish the PFC is the ability to sustain such activity in the face intervenin no dis ons.When monk eys must sustain the ry of as object over adelay filled with visual distractors,eachofw hich mu processed,sustained activity in the PFC can maintain the sample memory across the distractors(Miller et al 1996).By contrast,sustained activity in extrastriate visual areas(such as the IT and posterior parietal cortex)is easily disrupted by distractors (Constantinidis Steinmetz 1996:Miller et al 1993 1996)Thus pos sterior cor eflect the most recent input regardless of it relevance,whereas the PFC selectively maintains task-relevant information. Learning“Across Time”Within the PFC Typically.the internal representationofgoalsandassociated rulesmust beactivated avior they govem.Furthermo as we have seen,rule often involve learningassociations between stimuliand behaviors that are separated in time.How can associations be learned between a rule or event that occurs at one point in time and contingent behaviors or rewards that occur later?The capacity of the PFC for active maintenance,coupled with its innervation by brainstem dopa tems,suggest e way in which this might occur The capacityto actively maintain representations over time is fundamenta associative learning,as it allows information about fleeting events and actions to comingle that would otherwise be separated in time(Fuster 1985).For example, consider the Asaadet al(1998)study discussed above,in which the monkey needed
P1: FXZ January 12, 2001 14:38 Annual Reviews AR121-07 180 MILLER ¥ COHEN interfering events. The capacity to support sustained activity in the face of interference is one of the distinguishing characteristics of the PFC. Sustained neural activity within the PFC was first reported by Fuster (1971) and Kubota & Niki (1971) and has subsequently been reported in a large number of studies. These have demonstrated that neurons within the PFC remain active during the delay between a transiently presented cue and the later execution of a contingent response. Such delay period activity is often specific to a particular type of information, such as the location and/or identity of a stimulus (di Pellegrino & Wise 1991; Funahashi et al 1989; Fuster 1973; Fuster & Alexander 1971; Kubota & Niki 1971; Rainer et al 1998a,b, 1999; Rao et al 1997; Romo et al 1999), forthcoming actions (Asaad et al 1998, Ferrera et al 1999, Quintana & Fuster 1992), expected rewards (Leon & Shadlen 1999, Tremblay et al 1998, Watanabe 1996), and more-complex properties such as the sequential position of a stimulus within an ordered series (Barone & Joseph 1989) or a particular association between a stimulus and its corresponding response (Asaad et al 1998). Functional neuroimaging studies have begun to yield similar results with humans (Cohen et al 1997, Courtney et al 1997, Prabhakaran et al 2000). Other areas of the brain exhibit a simple form of sustained activity. For example, in many cortical visual areas, a brief visual stimulus will evoke activity that persists from several hundred milliseconds to several seconds (Fuster & Jervey 1981, Gnadt & Andersen 1988, Miller et al 1993, Miyashita & Chang 1988). What appears to distinguish the PFC is the ability to sustain such activity in the face of intervening distractions. When monkeys must sustain the memory of a sample object over a delay filled with visual distractors, each of which must be attended and processed, sustained activity in the PFC can maintain the sample memory across the distractors (Miller et al 1996). By contrast, sustained activity in extrastriate visual areas (such as the IT and posterior parietal cortex) is easily disrupted by distractors (Constantinidis & Steinmetz 1996; Miller et al 1993, 1996). Thus, posterior cortical neurons seem to reflect the most recent input regardless of its relevance, whereas the PFC selectively maintains task-relevant information. Learning “Across Time” Within the PFC Typically, the internal representation of goals and associated rules must be activated in anticipation of the behavior they govern. Furthermore, as we have seen, rules often involve learning associations between stimuli and behaviors that are separated in time. How can associations be learned between a rule or event that occurs at one point in time and contingent behaviors or rewards that occur later? The capacity of the PFC for active maintenance, coupled with its innervation by brainstem dopaminergic systems, suggests one way in which this might occur. The capacity to actively maintain representations over time is fundamental to associative learning, as it allows information about fleeting events and actions to comingle that would otherwise be separated in time (Fuster 1985). For example, consider the Asaad et al (1998) study discussed above, in which the monkey needed 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