Cognition and control in schizophrenia: a computational model of dopamine and prefrontal function

Abstract

Behavioral deficits suffered by patients with schizophrenia in a wide array of cognitive domains can be conceptualized as failures of cognitive control, due to an impaired ability to internally represent, maintain, and update context information. A theory is described that postulates a single neurobiological mechanism for these disturbances, involving dysfunctional interactions between the dopamine neurotransmitter system and the prefrontal cortex. Specifically, it is hypothesized that in schizophrenia, there is increased noise in the activity of the dopamine system, leading to abnormal “gating” of information into prefrontal cortex. The theory is implemented as an explicit connectionist computational model that incorporates the roles of both dopamine and prefrontal cortex in cognitive control. A simulation is presented of behavioral performance in a version of the Continuous Performance Test specifically adapted to measure critical aspects of cognitive control function. Schizophrenia patients exhibit clear behavioral deficits on this task that reflect impairments in both the maintenance and updating of context information. The simulation results suggest that the model can successfully account for these impairments in terms of abnormal dopamine activity. This theory provides a potential point of contact between research on the neurobiological and psychological aspects of schizophrenia, by illustrating how a particular physiological disturbance might lead to precise and quantifiable consequences for behavior.

Introduction

Some of the most prominent clinical symptoms exhibited by patients with schizophrenia include: distractibility, loosening of associations, and disorganized or socially inappropriate behavior. A number of investigators have postulated that these symptoms might relate to core cognitive deficits in a number of domains, such as attention Cornblatt and Keilp 1994, Kornetsky and Orzack 1978, Nuechterlein 1991, working memory Park and Holzman 1992, Weinberger et al 1986, episodic memory Gold et al 1992, Goldberg et al 1993 and inhibition (Abramczyk et al 1983, Barch et al in press a; Carter et al 1993; Chapman et al 1964). However, the underlying mechanisms that contribute to both clinical symptomatology and cognitive impairments in schizophrenia are still not well understood, at either the psychological or neurobiological level.

In previous work, we have argued that many cognitive deficits in schizophrenia (and the clinical symptoms stemming from them) can be interpreted as reflecting a failure to exert control over thoughts and actions, and that a central feature of cognitive control is the ability to properly maintain and update internal representations of task-relevant context information (Cohen and Servan-Schreiber 1992). This theory has been made explicit in the form of connectionist computational models of behavioral tasks in which patients with schizophrenia exhibit specific cognitive deficits. Furthermore, these models have provided a conceptual mapping between the psychological processes thought to be impaired in schizophrenia and their neurobiological underpinnings. Specifically, they have shown how: a) cognitive control can emerge from the biasing influence of representations of goal-related, or context, information actively maintained in PFC on more posterior pathways responsible for task performance; b) dopamine (DA) can exert a neuromodulatory influence on this function of PFC; and c) disturbances in this neuromodulatory function can produce disturbances in cognitive performance that closely match those observed in patients with schizophrenia.

In our initial modeling work, the role of DA was treated in a relatively abstract form, as influencing the responsivity of processing units to their afferent input—the “gain” hypothesis Cohen and Servan-Schreiber 1993, Servan-Schreiber et al 1990. In recent work we have begun to refine our model of DA function, driven simultaneously by the pressure to conform more closely to accumulating neurobiological evidence, and to arrive at a more powerful and complete theory of the computational mechanisms underlying cognitive control. This has led us to a new hypothesis: that DA serves a “gating” function in PFC, regulating access of context representations into active memory (Braver and Cohen in press a). This gives DA an important control function, responsible for the flexible updating of active memory in PFC, while retaining protection against interference. Here, we suggest an important component of the pathophysiology of schizophrenia may be an increase in the noise in the DA system, and that this increased variability leads to disturbances in both the updating and maintenance of context information within working memory. Below, we briefly review the literature on cognitive deficits in schizophrenia, our theory of cognitive control, and its ability to account for empirical findings in the schizophrenia literature. Finally, we present the results from a new set of computer simulation modeling studies, that implement our new theory of DA function, and use it to address empirical data regarding the behavioral deficits observed in patients with schizophrenia during performance of a specific cognitive control task.

A large literature on cognitive function in schizophrenia suggests that patients with this illness display deficits in several different cognitive domains. In particular, much recent work has focused on deficits in attention, working memory, episodic memory, and executive functions. In the attentional domain, the research to date suggests that the most reliable impairments exhibited by schizophrenia occur when attention must be exerted in a selective and controlled manner to both facilitate the processing of task-relevant information or inhibit the processing of task-irrelevant information. In the cognitive psychology literature, researchers often use the Stroop (Stroop 1935) color naming task as a paradigmatic measure of selective/controlled attention (MacLeod 1991). In this task, participants are presented with words printed in colors, and are told to either: 1) read the word and ignore the print color; or 2) name the print color, and ignore the word. When asked to read the word, participants can effectively ignore the print color, as evidenced by the fact that print color has little influence on reading time. When participants are asked to name the print color, they have difficulty suppressing the effects of the word. In particular, when the word and its color conflict (such as RED displayed in blue print) participants are slower than when there is no such conflict. This effect is called Stroop interference, and is thought to result from the obligatory nature of word reading disrupting color naming performance (MacLeod 1991). A similar, but smaller effect can be observed in improved performance for congruent stimuli (e.g., RED displayed in red print), referred to as Stroop facilitation.

A large number of studies have used the Stroop task to test patients with schizophrenia, employing both the traditional card-based method Abramczyk et al 1983, Everett et al 1989, Golden 1976, Wapner and Krus 1960, Wysocki and Sweet 1985, as well as more modern methods using tachistoscopic presentation and the on-line monitoring of response times and accuracy Barch et al in press a, Barch et al in press b, Carter et al 1992, Cohen et al 1999, Schooler et al 1997, Taylor et al 1996. These have consistently produced reliable evidence of enhanced Stroop interference or facilitation among patients, indicative of an impairment in selective/controlled attention in schizophrenia. Deficits in selective/controlled attention have also been observed using other tasks, such as the anti-saccade task in which subjects must inhibit the prepotent response to make a saccade to the location of visual stimulus, and instead saccade to a location in the opposite visual field Clementz et al 1994, Katsanis et al 1997, McDowell and Clementz 1997, Radant et al 1997.

Working memory (WM) has also been shown to be impaired in schizophrenia. WM is commonly defined as the collection of processes responsible for the on-line maintenance and manipulation of information necessary to perform a cognitive task (Baddeley and Hitch 1994). A growing number of studies suggest that patients with schizophrenia show deficits on tasks designed to measure WM Cohen et al 1999, Gold et al 1997, Goldberg et al 1998, Park and Holzman 1992, Park and Holzman 1993, Stone et al 1998, Wexler et al 1998. For example, several recent studies have shown impairments in spatial WM among schizophrenia patients Keefe et al 1997, Park and Holzman 1992, Stone et al 1998. Studies have also shown deficits in verbal WM Servan-Schreiber et al 1996, Wexler et al 1998, although there is some suggestion that deficits of verbal WM may be most evident when patients are challenged with a high information load (Carter et al 1998), when they have to deal with distraction (Keefe et al 1997), or when they are required to manipulate maintained information in WM Cohen et al 1999, Gold et al 1997. Based on such findings, several researchers have suggested that a deficit in WM may be a fundamental cognitive defect present in schizophrenia Cohen and Servan-Schreiber 1992, Goldman-Rakic 1991, Weinberger and Gallhofer 1997.

A third type of deficit that has been implicated in schizophrenia is in episodic memory Duffy and O’Carrol 1994, Gold et al 1992, Goldberg et al 1993, Hutton et al 1998. Episodic memory refers to the ability to encode or retrieve newly learned information. Some research suggests that this function is more seriously disturbed in schizophrenia than general intellectual ability McKenna et al 1990, Tamlyn et al 1992 or other cognitive functions Saykin et al 1991, Saykin et al 1994. Based on such findings, some investigators have argued that episodic memory deficits are a core cognitive deficit in schizophrenia, either in addition to or instead of WM deficits Clare et al 1993, McKay et al 1996. One possible interpretation of episodic memory deficits in schizophrenia is that they reflect an inability to use contextual cues to organize information at either encoding or retrieval, rather than a fundamental inability to encode new information (O’Reilly et al 1999). For example, some studies have shown that schizophrenia patients are more impaired on recall than recognition memory tasks Calev 1984, Goldberg et al 1989, Koh 1978, Paulsen et al 1995, Rizzo et al 1996, Rushe et al 1998, consistent with the hypothesis that patients’ performance improves with the addition of cue information. In addition, among patients with schizophrenia, recall performance benefits from the addition of explicit cues to organize the information at encoding (Koh 1978), as well as from additional cues at retrieval (Sengel and Lovallo 1983).

A fourth area that schizophrenia patients show deficits is the domain of executive functioning. This includes functions such as set switching, planning, and dual-task coordination. Patients with schizophrenia show consistent deficits in each of these domains. For example, two classic tasks associated with the measurement of set switching ability are the Wisconsin Cart Sorting Task and the Trails B task. It has long been known that schizophrenia patients show deficits on both of the tasks, and many researchers consider deficits on these tasks to be a hallmark sign of schizophrenia Berman et al 1986, Gold et al 1997, Goldberg et al 1988, Weinberger et al 1986. Patients also display disturbances on tasks that measure planning ability, such as the Tower of London task (Andreasen et al 1992). In addition, recent research utilizing dual-task paradigms has also provided evidence that patients are impaired when required to perform two tasks simultaneously (Granholm et al 1996), or to alternate between two different tasks (Smith et al 1998). In all of the tasks involving executive function, a central feature is the requirement that goal-related information must be both represented and updated at appropriate junctures.

As the above review suggests, the literature on cognitive function in schizophrenia points to impairments in a set of basic cognitive functions including: 1) attentionally-mediated selection of task-relevant information, and suppression of task-irrelevant information; 2) maintenance and manipulation of information in WM; 3) context-based organization of cues for memory encoding and retrieval; and 4) updating and switching of internally represented goal-related information These findings could be interpreted as evidence that patients with schizophrenia suffer cognitive dysfunction in a variety of qualitatively distinct domains. However, we have argued that all of the cognitive functions considered above may rely on a common process: the internal representation and use of context information in the service of exerting control over behavior. Thus, it is possible that deficits in attention, memory, and executive function all reflect the disturbance of a single underlying processing mechanism that is central to cognitive control. Below, we review our arguments in support of this hypothesis.

The need for a control mechanism in cognition has been long noted within psychology. Virtually all theorists agree that some mechanism is needed to guide, coordinate, and update behavior in a flexible fashion—particularly in novel or complex tasks (e.g., Norman and Shallice 1986). More specifically, control over processing requires that information related to behavioral goals be actively represented and maintained, such that these representations can bias behavior in favor of goal-directed activities over temporally-extended periods. Moreover, goal-related information must be: 1) appropriately selected for maintenance; 2) maintained for arbitrary lengths of time; 3) protected against interference; and 4) updated at appropriate junctures. The recognition that active representation and maintenance of goal-related information are central components of cognitive control can be seen in many theories. The best known of these is Baddeley’s working memory executive model (Baddeley 1986), that includes a specific sub-component, “the central executive,” responsible for utilizing goal-related information in the service of control. The postulation of a cognitive system involved in executive control closely parallels theorizing regarding the nature of frontal lobe function Bianchi 1922, Damasio 1985, Luria 1969, based on the clinical observation that patients with frontal lesions often exhibit impairments in tasks requiring control over behavior—the so-called “dysexecutive syndrome.” Traditional theories have not specified the mechanisms that the executive operates.

Theories aimed at providing a more explicit computational account of human behavior have also included goal representations as a central component. For example, in production system models, goal states represented in declarative memory are used to coordinate the sequences of production firings involved in complex behaviors (Anderson 1983). One critical feature of goal representations in production systems is that they are actively represented and maintained throughout a sequence of behaviors. Shallice Norman and Shallice 1986, Shallice 1982, Shallice 1988 has relied upon the production system framework to put forth his Supervisory Attentional System (SAS) as a mechanism by which complex cognitive processes are coordinated and non-routine actions are selected.

In our own work, we have suggested that the active maintenance of context information is critical for cognitive control Braver et al 1999, Cohen et al 1996, Cohen and Servan-Schreiber 1992. We have defined context as any task-relevant information that is internally represented in such a form that it can bias processing in the pathways responsible for task performance. Goal representations are one form of such information, that have their influence on planning and overt behavior. We use the more general term context to include representations that may have their effect earlier in the processing stream, on interpretive or attentional processes. For example, in the sentence “To keep his chickens, the farmer needed a pen,” the words “chicken” and “farmer” may elicit a context representation that is used to constrain the interpretation of the word pen to its weaker, but relevant meaning (i.e., “fenced enclosure”). Thus, context representations may include a specific prior stimulus, or the result of processing a sequence of stimuli, as well as task instructions or a particular intended action. Because context representations are maintained on-line, in an active state, they are continually accessible and available to influence processing. Consequently, context can be thought of as one component of WM. Specifically, context can be viewed as the subset of representations within WM that govern how other representations are used. Representations of context are particularly important for situations where there is strong competition for response selection. These situations may arise when the appropriate response is one that is relatively infrequent, or when the inappropriate response is prepotent (such as in the Stroop task). In this respect, context representations are closely related to goal representations within production system architectures. Maintenance of internal goal representations, or goal-related knowledge, is critical for initiating the selection of “weak” behaviors, and for coordinating their execution over temporally extended periods, while at the same time suppressing competing, possibly more compelling behaviors. Next, we discuss evidence that context information is actively maintained within PFC.

Over a hundred years of neuropsychological studies on patients with PFC lesions have provided strong evidence of the involvement of this brain region in the regulation of behavior. In recent years, a large body of converging evidence from neurophysiology and neuroimaging studies have suggested a more specific role for PFC in the active maintenance of task-relevant information. Single-cell recording studies in nonhuman primates have typically examined the active maintenance properties of PFC through the use of delayed-response paradigms, in which the animal must maintain a representation of a cue stimulus over some delay, to respond appropriately at a later point. It is now well-established that during performance of these tasks, populations of neurons in monkey PFC exhibit sustained, stimulus-specific activity during the delay period Fuster and Alexander 1971, Kubota and Niki 1971. The mnemonic properties of these neurons have been demonstrated by showing both that: 1) local and reversible lesions to PFC impair task performance; and 2) performance errors in intact animals are correlated with reduced delay-period activity Bauer and Fuster 1976, Fuster 1973. Neuroimaging studies have confirmed and extended these findings in humans. For example, in humans, PFC activity has been shown to: 1) increase as delay interval increases (Barch et al 1997); 2) increase as memory load increases (Braver et al 1997b); and 3) be sustained over the entire delay interval Cohen et al 1997, Courtney et al 1997.

In addition to these other properties, PFC also seems to be particularly specialized to maintain information in the face of interference, whereas still allowing for flexible updating of stored information. Recently, Miller and colleagues (Miller et al 1996) have provided direct evidence for this hypothesis. They trained monkeys to respond to repeats of a prespecified cue (e.g., A) when presented with sequences such as A–B–B–A. This task clearly required the ability to identify the cue on each trial, and maintain it across intervening distractors. They observed cue-specific delay period activity for units in both inferotemporal cortex (IT) and PFC after initial presentation. Subsequent stimuli obliterated this activity in IT, whereas it was preserved in PFC until a match occurred. The crucial role of PFC in updating and interference-protection can also clearly be seen in studies of PFC pathology. Increased distractibility and perseveration are hallmarks of PFC damage Damasio 1985, Engle et al 1999, Milner 1963, Owen et al 1991, Stuss and Benson 1986, as well as a classic symptom of schizophrenia Malmo 1974, Nuechterlein and Dawson 1984. Together, these findings support the idea that there are specialized mechanisms within PFC for active memory, as well as for protecting maintained information from both perseveration and interference. More specifically, we hypothesize that representations of context are housed within PFC and actively maintained there.

From a computational viewpoint there are a number of different processing mechanisms that could support short-term maintenance of information. The most commonly employed and well-understood of these are fixed-point attractor networks Hopfield 1982, Zipser 1991. Such networks possess recurrent connections, that “recirculate” activation among units, and are thus capable of supporting sustained activity. The state of such networks typically settles into “attractors,” defined as stable states in that a particular pattern of activity is maintained. Thus, attractors can be used to actively store information. Indeed, a number of computational models of simple maintenance tasks have demonstrated that both physiological and behavioral data regarding PFC function can be captured using an attractor-based scheme Braver et al 1995, Dehaene and Changeux 1989, Moody et al 1998, Zipser et al 1993.

Simple attractor systems have a number of limitations that create problems in more complex maintenance tasks. These limitations can be traced to the fact that the state of an attractor system is determined by its inputs, so that presentation of a new input will drive the system into a new attractor state, thereby overwriting previously maintained information Bengio et al 1993, Mozer 1993. Although attractor networks can be configured to display resistance to disruption from input (i.e., hysteresis), this impairs their ability to be updated in a precise and flexible manner. One way that attractor networks can overcome these difficulties is through the addition of a gating mechanism. Such systems only respond to inputs, and change their attractor state, when the “gate” is opened. Computational analyses suggest that gating mechanisms provide the most effective way to stably maintain information in an active state, protect this information from interference, and still retain the ability of flexible updating. For example, Hochreiter and Schmidhuber (1997) compared gated recurrent neural networks with other types of attractor systems, and concluded that networks with a gating mechanism were able to learn and perform complex short-term memory tasks better than simple attractor networks, especially when the tasks involved noisy environments, frequent updating, and relatively long periods of storage. Thus, computational studies suggest that a gated attractor system is the optimal one for active memory. Moreover, the physiological evidence reviewed above is consistent with the hypothesis that PFC implements such a gated system. Indeed, in previous work, Zipser and colleagues Moody et al 1998, Zipser 1991, Zipser et al 1993 have proposed a gated attractor model and have used it to successfully simulate the pattern of delay period activity observed for PFC neurons. However, the Zipser model has not specified the source of the gating signal. In the following section, we suggest that phasic increases in DA activity serve as a gating signal within PFC.

The DA system has been implicated in a wide range of effects on behavior. The most prominent of these is the linkage of DA with motor function. It is well-established that disturbances to the subcortical DA system cause severe movement-related disorders such as Parkinson’s disease. Further, stimulants such as amphetamine and apomorphine (that are thought to act by stimulating DA release; Kelly et al 1975) have clear effects on motor behavior. For example, in animals, these drugs produce consistent changes in both locomotor activity (Segal 1975), and the repertoire of behaviors exhibited (Norton 1973), with high doses inducing species-specific stereotypies (Randrup and Munkvad 1970). There are also many studies documenting the effect of DA activity on response activity in goal-directed tasks, such as operant conditioning paradigms Heffner and Seiden 1980, Louilot et al 1987. A number of investigators have hypothesized that, together, these findings suggest a function for DA in selecting or initiating new motor response patterns Iversen 1984, Oades 1985.

Another commonly postulated function of DA is that this neurotransmitter mediates the processing of reward information. This reward-based account of DA activity is supported by findings that suggest that DA facilitates a number of primary motivation behaviors, such as feeding, drinking and sexual activity (Willner and Scheel-Kruger 1991). Conversely, spontaneous engagement in these behaviors has been shown to result in increased DA transmission (Heffner et al 1980). In addition, innumerable studies have shown that the electrical self-stimulation paradigm is primarily dependent on stimulation of DA pathways Mora and Cobo 1990, Phillips and Fibiger 1989. This finding is consistent with the pharmacological evidence that many drugs of addiction act through the DA system (Koob and Bloom 1988). Taken together, these findings have led some researchers to postulate a crucial role for DA in conveying information regarding the rewarding or reinforcing properties of specific behaviors (Wise and Rompre 1989).

The literature on the behavioral effects of DA is not limited to studies of motor and reward-related behaviors. There have also been a number of reports of DA effects on cognitive function. In humans, systemic administration of DA agonists have been associated with improvements on various cognitive tasks Callaway et al 1994, Klorman et al 1984. In particular, the most consistent effects of DA on cognition have been in tasks relying on WM. DA effects in WM have been seen systemically in humans Luciana et al 1995, Luciana et al 1992, and through local manipulations in nonhuman primates Brozoski et al 1979, Sawaguchi and Goldman-Rakic 1994. These local effects in primates have focused on DA activity selective to PFC. For example, Goldman-Rakic and colleagues have found that pharmacologically blocking DA receptors in circumscribed areas of PFC produced reversible deficits in task performance (Sawaguchi and Goldman-Rakic 1991). Moreover, microiontophoresis of DA agonists and antagonists, and even DA itself has been found to directly affect the activity patterns of PFC neurons Sawaguchi and Goldman-Rakic 1991, Sawaguchi et al 1990. Goldman-Rakic and others have concluded from these findings that DA activity serves to modulate the cognitive functions mediated by PFC Cohen and Servan-Schreiber 1992, Goldman-Rakic 1991.

The literature on DA involvement in motor, reward and cognitive functions reveals the wide-spread influence of this neural system on behavior. Further, the disparate nature of these three domains suggests that DA may perform multiple, unrelated behavioral functions. A more parsimonious explanation is also possible: DA activity plays a unitary function in the central nervous system that is expressed in different domains as a result of its interaction with the different brain systems that the DA system projects (i.e., striatal, limbic, and cortical). Specifically, we propose that the function of the DA system is to provide a means for the organism to learn about, predict, and respond appropriately to events that lead to reward. The DA system serves this function through simple neuromodulatory effects in the neural populations that it targets. One effect modulates the responsivity of the target neurons to other inputs, and the other effect modulates the synaptic strength between the target neuron and its other inputs. The DA effects on synaptic strength serve to drive the learning of predictors of reinforcement, whereas the effects on responsivity serve to transiently bias on-going processing. Most importantly, we propose that through its projection to PFC, the responsivity effect of DA serves to gate access to active memory, whereas its coincident learning effect allows the system to discover what information must be actively maintained for performance of a given task. In previous reviews of the literature Braver and Cohen in press a, Braver and Cohen in press b, Cohen et al 1996, we have discussed a number of lines of research supporting this hypothesis, including evidence that argues that: 1) DA exerts a modulatory effect on target neurons; 2) This effect is of a type that could be exploited to perform a gating function in PFC; and 3) The role of the DA system in reward-prediction learning provides it with particular activation dynamics and timing that are required of a gating signal.

Taken together, the properties of DA and PFC reviewed above suggest the outlines of a theory regarding the neural and computational mechanisms of cognitive control. In particular, we refine our previous work on active maintenance in PFC by integrating it with the work of Montague et al (1996) on reward-based learning. This integration provides a means of accounting for the relevant data regarding DA activity dynamics and reward functions as well as the modulatory role of DA in active memory. Specifically, the following refinements are made to our original theory (Cohen and Servan-Schreiber 1992):

• DA gates access to active memory in PFC to provide flexible updating while retaining interference protection.

• Phasic changes in DA activity mediate gating and learning effects in PFC.

• Both effects occur locally at the synapse, and rely on similar neuromodulatory mechanisms (possibly through different receptor subtypes).

• The gating effect occurs through the transient potentiation of both excitatory afferent and local inhibitory input.

• The learning effect occurs through Hebbian-type modulation of synaptic weights, and is driven by errors between predicted and received rewards.

• The temporal coincidence of the gating and learning signals produces cortical associations between the information being gated, and a triggering of the gating signal in the future.

The power of this new theory is that it provides a framework that may be able to account for specific patterns of normal behavioral performance across a wide-range of tasks requiring cognitive control. At the same time, by making close contact with the known physiological properties of both the DA system and PFC, it allows for more detailed and biologically realistic investigations of the neural basis of control. In the following section we discuss the relationship of DA and PFC function to the pathophysiology of schizophrenia.

The centrality of PFC function to schizophrenic cognitive deficits is a common theme in recent research. Structural abnormalities have been observed in PFC of schizophrenic patients Andreasen et al 1986, Weinberger et al 1980, and these have been linked to reductions in resting PFC metabolism Andreasen et al 1986, Buchsbaum et al 1982, Farkas et al 1984, Franzen and Ingvar 1975, Ingvar and Franzen 1974, Morihisa and McAnulty 1985, Weinberger et al 1980. These findings have been supported by two reviews Andreasen et al 1992, Buchsbaum 1990 that suggest that decreased metabolic activity or diminished cerebral blood flow in PFC is reliably present in schizophrenia. Despite this seemingly impressive literature, the presence of hypofrontality in schizophrenia remains somewhat controversial. For example, Gur and Gur (Gur and Gur 1995) cited several recent studies that did not find decreased frontal metabolism or rCBF in schizophrenia patients. These authors strongly questioned the presence of resting hypofrontality in schizophrenia, but did acknowledge that functional hypofrontality, defined as a failure to activate frontal cortex during cognitive activity, “may still merit further investigation.” Indeed, studies using functional neuroimaging techniques have greater sensitivity for detecting PFC disturbances in schizophrenic subjects, and relating these to cognitive disturbances. In particular, a number of recent functional neuroimaging studies in schizophrenia have linked disturbance in PFC function to impaired performance on tasks measuring WM Carter et al 1998, Stevens et al 1998, controlled/selective attention (Carter et al 1996), planning (Andreasen et al 1992), and set switching Berman et al 1986, Weinberger et al 1986.

Disturbances to the DA system have also long been regarded as a fundamental pathophysiological component of schizophrenia. Most of the support for this viewpoint comes from observations regarding the therapeutic efficacy of neuroleptics. The finding that the clinical potency of traditional neuroleptics is highly correlated with its affinity for DA receptors (Creese et al 1976) strongly implicates this neurotransmitter in schizophrenic symptomatology. In addition, long-term usage of drugs that stimulate DA activity in the CNS can lead to schizophreniform psychoses (Snyder 1972). Although the view that dopamine plays a role in schizophrenia is a long-standing one, it is clear that not all data on the pathophysiological disturbances present in schizophrenia are consistent with this hypothesis. The DA projection to PFC in particular has been a recent focus of attention in schizophrenia research. Specifically, a common current viewpoint is that schizophrenia is not simply caused by a hyperactive dopamine system. Rather, it has been postulated that many of the cognitive impairments seen in schizophrenia are related to reduced DA activity in PFC Davis et al 1991, Goldman-Rakic 1991.

Because our theory of DA and PFC function is conceptualized in terms of explicit computational mechanisms, it can be explored through simulation studies. In recent work, we have conducted simulations that tested the computational validity of the theory and that investigated the influence of different gating parameters on updating, maintenance, and interference protection (Braver and Cohen in press b). We also provided support for the hypothesis that DA implements both gating and learning effects, and that these can work synergistically to provide a mechanism for how cognitive control might be learned through experience (Braver and Cohen in press a). Specifically, in these simulations, the timing of the gating signal developed as a function of reward-prediction errors using in a temporal difference algorithm (Sutton 1988). This algorithm enabled the network to chain backward in time to find the earliest predictor of reward, that was a cue stimulus that also had to be maintained in active memory to receive the reward. Because this cue triggered a phasic response in the gating/reward-prediction unit, the information provided by the cue was allowed access to active memory.

In addition to providing an account of the neural mechanisms underlying normal cognitive control, our theory provides an explicit framework for testing ideas regarding the particular neurobiological disturbances that may underlie schizophrenia and their consequences for behavior. Most importantly for understanding schizophrenia, our prior work provided useful insights into the relationship between gating unit activity and active maintenance (Braver and Cohen in press b). In particular, previous simulations have demonstrated three significant effects: 1) Reduced phasic activity during the presentation of “task-relevant” stimuli leads to perseveratory behavior, by decreasing the probability that the previous context will be replaced by the current context; 2) Increased phasic activity during the presentation of “irrelevant” stimuli produces interference effects, by increasing the probability that these stimuli will disrupt the currently maintained context; and 3) Increased tonic activity during delay periods produces a delay-related decay of active memory, by increasing the probability that the current context will deactivate over time. Together, these three effects may provide a model of how DA impairment influences cognitive control. Indeed, perseverations, poor interference control, and maintenance deficits are three symptoms commonly associated with schizophrenia Malmo 1974, Nuechterlein and Dawson 1984. Our theory provides an explicit mechanism that can explain how these cognitive deficits arise in schizophrenia. In the remainder of this article, we report a study that tests this idea directly, by incorporating the gating mechanism into a model of performance on a simple cognitive control task and examining whether disturbances to this mechanism can account for the patterns of behavioral impairments observed in schizophrenia patients.

The model suggests that the pattern of deficits observed in patients is consistent with a decrease in phasic and increase in tonic DA activity. The gating hypothesis predicts that tonically increased DA activity should produce deficits in active or WM, whereas decreased phasic DA activity should produce perseveration and interference-effects. Here, we directly test these predictions by conducting simulations of behavioral performance on a simple cognitive control task that requires both active maintenance and frequent updating of context information. The task is an “AX” variant of the Continuous Performance Test (CPT, Rosvold et al 1956). We have collected extensive behavioral data regarding the performance of both healthy subjects and patients with schizophrenia on this task Barch et al 1998, Braver et al 1999, Cohen et al 1999, Servan-Schreiber et al 1996. Simulations of behavioral data were conducted by adding a gating mechanism to an existing computational model of the task Braver et al 1995, Braver et al 1999, Cohen et al 1996.