Objectives: To use model-based methods from reward learning to test the above hypothesis, using behavioral and fMRI data.
Methods: We designed a task that compared monetary reward with directly social reward. The social rewards and punishments consisted of friendly and unfriendly faces together with verbal statements, whereas the monetary rewards and punishments consisted of winning or losing money. Both social and monetary rewards were used in an instrumental reward learning task that was structurally identical: participants had to learn to choose among three slot machines each associated with different probability distributions over the rewards (one resulting in punishment, one resulting in reward, and one being neutral). We measured people’s choices over time on this task while we acquired BOLD-fMRI. The data were then modeled using reinforcement-learning models to compare directly the possible stages in reward processing. We tested a group of 9 high-functioning people with autism and compared their data to those from 9 healthy matched controls.
Results: We found that both controls and people with autism quickly learned to choose in favor of the most rewarding slot machine on our task, regardless of whether the outcome was monetary or social reward. In general the autism group performed behaviorally very comparably to the controls on measures such as learning rate, switching, and asymptotic task performance. Moreover, ratings of the valence of the stimuli was very similar between groups.
fMRI results showed that in controls a common network of brain structures is activated when processing either monetary or social reward. This included parts of the medial prefrontal cortex and the ventral striatum. Initial fMRI results from the autism group using the same model-based approach suggest a very different pattern of brain activations. We are exploring the hypothesis that people with autism may be able to generate very similar behavioral performances on this task, but through a very abnormal strategy that is possible more model-based as opposed to resulting from standard reward learning of the value of the slot machines.
Conclusions: Whereas high-functioning people with autism can perform normally in learning to choose rewarding or avoid punishing outcomes, and can do so both with respect to monetary or social outcomes, the underlying neural computations appear to be very different from those seen in healthy controls