On the Control of Control: The Role of Dopamine in Regulating Prefrontal Function and Working Memory

An Important aspect of cognitive control is the ability to appropriately select, update, and maintain contextual information related to behavioral goals, and to me this information to coordinate processing over extended periods. In our novel, neurobiologicaUy based, connectionist computational model, the selection, updating, and maintenance of context occur through Interactions between the prefrontal cortex @'PC) and dofamlne (DA) neuiotiuismitter system. Phasic DAactivity serves two simultaneous and synergistic functions: (1) a gating function, which regulates the access of information to active memory mechanisms subserved by PFC; and (2) a learning function, whlcn auowa me system to discover wast uifonnation is relevant for selection as context. We present a simulation that establishes the computational viability of these postulated neurobiological mechanisms for subserving control functions. 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 fashionparticularly in novel or complex tasks (Norman and Shallice 1986). In particular, control over processing requires that information related both to current context and to behavioral goals be actively represented, such that these representations can bias behavior in favor of goal-directed activities over extended periods. Indeed, most computationally explicit theories of human behavior have included such a mechanism as a fundamental component. For example, in production system models, goal states represented in declarative memory are used to coordinate the' sequence of production firings involved in complex behaviors (e.g., Anderson 1983); One critical feature of goal representations in production systems is that they must be actively represented and maintained throughout the course of a sequence of behaviors. Such formulations of a control (or "executive") mechanism closely parallel theorizmg about the nature of frontal lobe function (Bianchi 1922; Darnasio 1985; Luria 19691, and clinical observations of patients with frontal lesions who often exhibit impairments in tasks requiring control over behavior-the so-called dysexecutive syndrome. Shallice (Nonnan and Shallice 1986; Shallice, 1982,1988) explicitly noted this relationship, using the production system framework to describe his theory of a "supervisory attentional system" (SAS) as a mechanism by wmcn the frontal lobes coordinate complex cognitive processes and select nonroutine actions. While these efforts have provided insights into the types of processes that may be engaged by cognitive control, they d6 not map transparently onto underlyingneural mechanisms. They have also not fully addressed several critical issues, such as how a control system can develop through learning. A number of recently proposed connectionist models of prefrontal function incorporate some of the central features of control processes in production system models, such as the active maintenance of goal representations (Dehaene and Changeux 1992; Guigon et al. 1991; Levine and Prueitt 1989). Connectionist models have the advantage of both being mechanistically explicit and using a computational architecture that maps more naturally onto neural mechanisms than traditional production system models. In this chapter, we report on work that uses this framework to address a critical question about cognitive control: How can a system learn to choose and appropriately update representations in active memory that can be used to control behavior? This is an extension of our ongoing effort to specify the neural underpinnings of cognitive control (Braver et al. 1995a; Cohen, Braver, and WReilly 1996; Cohen and Servan-Schreiber 1992), reviewed briefly below as background. A central hypothesis in our work is that a cardinal function of prefrontal cortex (PFC) is to actively maintain context information. We use the general term context to include not.only goal representations, which have their influence on planning and overt behavior, but also representations that may have their effect earlier in the processing stream, on interpretive or attentional processes. We assume that a primary function of PFC is tomaintain task-relevant context represmtations in an active state. These active context representations .serve to mediate control by modulating the flow of information within task-specific pathways such that processing in the task-relevant pathway is favored, over a (possibly stronger) competing pathway. This function of PFC can also be thought of as a component of working memory 0, commonly defined as the collection of mechanisms respowible for the on-line maintenance and manipulation of information necessary to perform a cognitive task (Baddeley and Hitch 1994). From this perspective, context can be viewed as the subset of representations within WM that govern how other representations are used. As noted above, there is long-standing recognition that control involves representation and maintenance of context information (e.g., goals). However, a more complete account of cognitive control has additional requirements. Here we focus on four. Context information must be (1) appropriately selected for maintenance; (2) held for, arbitrary lengths of time; (3) protected against interference; and (4) updated at appropriate junctures. Inasmuch as we assume that context infopt ion is represented in PEC, our interest is in the mechanisms that regulate the selection and updating of representations in PFC. One type of system meeting