Neuro-computational account of arbitration between imitation and emulation during human observational learning

In observational learning (OL), organisms learn from observing the behavior of others. There are at least two distinct strategies for OL. Imitation involves learning to repeat the previous actions of other agents, while in emulation, learning proceeds from inferring the goals and intentions of others. While putative neural correlates for these forms of learning have been identified, a fundamental question remains unaddressed: how does the brain decides which strategy to use in a given situation? Here we developed a novel computational model in which arbitration between the strategies is determined by the predictive reliability, such that control over behavior is adaptively weighted toward the strategy with the most reliable prediction. To test the theory, we designed a novel behavioral task in which our experimental manipulations produced dissociable effects on the reliability of the two strategies. Participants performed this task while undergoing fMRI in two independent studies (the second a pre-registered replication of the first). Behavior manifested patterns consistent with both emulation and imitation and flexibly changed between the two strategies as expected from the theory. Computational modelling revealed that behavior was best described by an arbitration model, in which the reliability of the emulation strategy determined the relative weights allocated to behavior for each strategy. Emulation reliability - the models arbitration signal - was encoded in the ventrolateral prefrontal cortex, temporoparietal junction and rostral cingulate cortex. Being replicated across two fMRI studies, these findings suggest a neuro-computational mechanism for allocating control between emulation and imitation during observational learning.