Meeting ID: 892 5037 7091
Computational modeling of Cognitive Control over Rule-guided behavior
Cognitive Control is the general capacity of an organism to use top-down control signals to inhibit the dominant behavior in favor of a contextually relevant response, in accordance with internally described goals (which could result from environmental or motivational factors).
Various experimental studies and computational models have tried to understand the neural mechanisms and structures that enable flexible and adaptive behavior by exerting cognitive control. Nevertheless, a unifying theory that explains these mechanisms remains elusive. The degree of adaptability that cognitive control provides varies from humans to nonhuman animals. We elaborate this gradation of cognitive control in a conceptual framework, and then use biologically plausible computational models to identify key computational processing requirements at each stage.
In the first model, we use a basic actor-critic model, to show how the default behavior of exploration in mice, needs to be overridden in order for a rodent (agent) to learn a simple tactile rule in a radial maze. Based on the decision making literature on rodents, we then show through a series of incrementally complex tasks, the necessity of working and episodic memory systems. This is done by extending the previous model with an elementary abstraction of these memory systems in order to make concrete the underlying mechanisms and criteria of cognitive control in rodents. As a third step, we highlight the need to form explicit mental representations of “context” from implicitly acquired rules, to enable contextually guided behavior, using a simple recurrent neural network trained on a sensorimotor task. Finally, to understand how the PFC supports contextual learning and the full capacity of cognitive control in humans, we develop a hierarchical computational model that explains the role of selective and sustained attention in learning abstract rules, and selects the appropriate concrete rules by manipulating representations, or task sets, and monitoring these representations and prediction errors.