Computational Neuroscience Models: Error Monitoring, Conflict Resolution, and Decision Making

Abstract

A critical part of the perception–reason–action cycle is performance monitoring, in which the outcomes of decisions and actions are monitored with respect to how well the actions are (or are not) achieving the desired goals. If the current behavior is not achieving the desired goals, then the problems must be detected as early as possible, and corrective action must be implemented as quickly as possible. The cognitive process of performance monitoring refers to the ability to detect such problems, and cognitive control (or executive control) refers to the processes of directing actions toward their intended goals and correcting failures to do so. In this chapter, the basic neural mechanisms of goal-directed behavior are re-viewed through the lens of computational neural models on the topic, and the neural mechanisms of decision-making are reviewed briefly. We then go on to discuss how top-down goal representations are chosen, instantiated, strengthened, and when necessary, changed. Progress on computational neuroscience models of cognitive control began with earlier, more abstract qualitative models. Norman and Shallice (1986) delineated a hierarchy of control with several components. In their framework, schema formed the basic building blocks that linked stimuli to corresponding responses. In some cases, schemamight conflictwith each other, and this required a contention schedul- ingmechanism to resolve the conflicts. Thiswas conceived of as operating by lateral inhibition and competition between schemas, so that the dominant schema gains exclusive control of the output. In some cases of especially complex or novel situ- ations, an appropriate schema might not be readily available, and so a supervisory attentional system must be invoked. This system redistributes attention to the most relevant stimuli, which then leads to the activation of an appropriate schema. This seminal paper byNorman and Shallice delineated qualitative roles for the interaction of bottom-up stimuli in driving responses (via schema and contention scheduling), vs. top-down cognitive control (via the supervisory attentional system). A line of cognitive models called production systems has continued to develop in this com- putational cognitive tradition (Newell 1991; Meyer and Kieras 1994; Anderson et al. 2004). The main underlying theme of these models is the existence of a set of behavioral goals, which in turn motivate a set of rules that define how environ- mental states are mapped to actions. A set of control processes and mechanisms serve to instantiate these goals, and these mechanisms correspond essentially to the supervisory attentional system in the Norman and Shallice framework. A cognitive system may have a variety of goals ranging from simple behaviors such as pressing a button to complex goals such as finding a mate. Goals may involve attaining a reward or avoiding an aversive event. A general requirement for adaptive behavior is that an agent should be capable of modifying its behavior moment-by-moment in order to achieve some goal while simultaneously adjusting for errors and changing environmental conditions. The ability to direct ongoing be- havior toward specific desirable goals is seemingly a simple observation dating back to Thorndike’s law of effect (Thorndike 1911). Nonetheless, the existence of more complex goal-directed behavior spawns a number of additional related questions. Once a goal is selected, how is progress toward the goal evaluated? Relatedly, what events, internal or external, underlie performance monitoring? In the course of on- going behavior, when is cognitive control exerted? How does an organism learn what indicates the need for cognitive control? These questions are related in that an answer to one implies answers to the others. In a simple instrumental conditioning task, for instance, changes in behav- ioral strategiesmay be indicated by the failure to receive a reward after generating a previously rewarded response, even without an explicit cue regarding what the cor- rect response should be. Subsequently, a single error of this sort may drive changes in the responses the organism generates (Grant and Berg 1948; Rabbitt 1967; Laming 1968). In this case, cognitive control is exerted only after the erroneous response, and the need to learn the association between environmental cues and outcomes isminimal. This kind of cognitive controlmay be termed reactive (Braver et al. 2007). On the other hand, what if we instead suppose that cognitive control is deployed as means to avoid an error in the first place rather than after the fact? This kind of control may be termed proactive (Braver et al. 2007). Often it is advantageous for the organism not only to adjust behavior in response to an error but also to learn what environmental factors are associated with the possibility of an error. This in turn generates additional questions:what factors in the environment does the organ- ism associate with potential error? How is this association learned? The remainder of the chapter describes recent progress on computational models of performance monitoring and cognitive control.