Belief as a product of cortical Hebbian processes: a theoretical and experimental approach in monkeys
Director of Research, Physiology and Pathophysiology of Executive Functions, CNRS and University of Bordeaux
While we were investigating learning behaviour of monkeys in a two-armed bandit task, we found that sometimes the animals persisted with preferring the less rewarding option. We hypothesized that this perseverance is based on wrong estimation (belief) of the outcome of the possible options. We propose here an analysis of this behaviour in light of the "two actors-one critic" computational model of the decision making network we developed recently. Briefly, it is made of 2 “actor” modules – namely, a cortical network nested in a cortico-basal ganglia (CBG) loop – and a “critic” module. The critic drives reinforcement learning in the CBG loop, which in turn drives Hebbian plasticity at the cortical level according to the input of the CBG loop. According to this model, the CBG loop is necessary during learning of new contingencies in two-armed bandit task, while, once learning is over, cortex can choose by itself without subcortical feedback. Our model is supported by robust experimental data. Based on this model, we propose that beliefs can be operationalize in animals by luring the system with ambiguous outcomes. Once the belief has been consolidated in the cortex by Hebbian associations, it becomes much less sensitive to learning from experience.