32141
Social Learning Relies on Distinct Cognitive Mechanisms in Adolescents with and without Autism
Many of our efforts in social interactions are dedicated to learning about other persons’ preferences, mental states, and behaviors—abilities referred to as Theory of Mind (ToM). ToM abilities continue to develop throughout adolescence and are associated with positive peer relationships and emotional wellbeing. Adolescents with ASD have core deficits in ToM, which have been a main focus in social skill trainings and interventions for this age group. Despite the large body of research specifying ToM deficits and intervention outcomes of adolescents with ASD, a mechanistic understanding of how individuals integrate environmental cues to learn about others is lacking.
Objectives:
Here we aim at revealing the cognitive mechanisms underlying learning about others’ preferences in adolescents with ASD using a computational modeling approach. In a next step, we will explore the neural implementation of these cognitive mechanisms.
Methods:
We devised a novel preference task, in which TD adolescents (N=23), and adolescents with ASD (N=21) rated how much three peers liked a variety of items (e.g. activities, fashion items and food) in the scanner. After each rating, they received feedback about the peer’s preference ratings. Participants could improve predictions about a new item by integrating feedback about similar past items into their judgements. After this learning task, participants rated their own preferences for the same items outside the scanner.
The preference profiles used in the learning task were of real adolescents who participated in an online preference survey. In total, we obtained 103 preference profiles. Three profiles were selected for the learning task and the remaining 100 were used to compute learning priors (e.g. an individual’s prior knowledge about their peers’ preferences).
Results:
Prediction errors, the difference between participants’ judgements and the trial-by-trial feedback, did not differ between groups. To test finer-grained group differences in learning strategies, we devised various computational models: non-learning regression models and reinforcement learning (RL) models. RL models describe participants’ judgements over time based on previous feedback alone, or based on a combination of feedback and either own preferences or priors. We also tested more sophisticated models, which assume that participants represent similarities between item preferences for their peer group (e.g., participants know that preferences often are similar for apples and pears but not necessarily so for apples and kale). Prediction errors are then scaled according to these fine-grained similarities between the items (similarity-RL model). Bayesian model comparison revealed that preference ratings of TD adolescents relied on a combination of similarity-RL and participants’ own preferences. In contrast, preferences of ASD adolescents were best described by a non-learning model that relied on simple priors (i.e. peers’ average preference ratings) for each item.
Conclusions:
We show that computational models are well suited to differentiate between social learning strategies of TD and ASD samples. In a next step, we will explore how parameters derived from the winning model for either group are implemented in brain activity. This will provide biological validity for the models and help specify differences in the underlying neural mechanisms of social learning in adolescents with ASD.