A computational model of the interaction of neurobiological circuits for category learning

Abstract

The goal of this proposal is to design a neurobiologically-based model that describes the switching mechanism in category learning based on existing category learning theory and model. COVIS is a neurobiologically-detailed theory of multiple systems in category learning. COVIS postulates two systems that compete throughout learning—a frontal-based declarative hypothesis-testing system that uses logical reasoning and depends on working memory and executive attention, and a basal ganglia-mediated system that uses procedural learning. However, no biological mechanism has been proposed to account for the interaction between the COVIS systems. We propose a model that employs a neurobiological-based circuit that describes the interaction and switching between the hypothesis-testing system and the procedural learning systems in COVIS. With the hypothesis-testing system and procedural learning system modeled as black boxes, the model focuses on the network that facilitates switching. In COVIS, both learning systems generate a response signal in each trial based on the stimuli given. Our model incorporates the Izhikevich firing model that represents the activity of the neuronal cells from the hyperdirect pathway of the cortico-basal ganglia network. The hyperdirect pathway acts as a gate for the response signal of the procedural learning system to reach the premotor units for action selection. We propose that the procedural learning system’s response is inhibited from approaching the premotor units when the hypothesis-testing system is in control of the response. However, if rule-based strategies fail, inhibition to the procedural system’s response is reduced. The reduction in inhibition results in the acceptance of responses from both learning systems in the premotor units. To validate the proposed model, we fit the model to two groups of participants in a perceptual category-learning task. One group of participants used the optimal procedural strategy in the task and the other used a suboptimal rule-based strategy. The categorization task was an information-integration task, whereby participants had to switch away from rule-based strategies and learn to integrate the stimulus dimensions to be able to perform optimally. We were able to differentiate the switchers from the non-switchers by adjusting the parameters in the model. In addition, we fitted another task to the model in which participants from different age groups with or without Parkinson’s disease were asked to switch between rule-based and procedural strategies on a trial-by-trial basis. We were able to match the learning curve, accuracy switch cost, and proportion of switchers of the different groups of participants.