Objectives: To explain flexibility deficits by studying probabilistic reinforcement, transitive interference, and hierarchical learning, based on computational models of interactions between the basal ganglia, hippocampus and the prefrontal cortex (PFC), which is believed to represent progressively more abstract action rules within a gradient of progressively more rostral PFC regions
Methods: Age, IQ, and gender matched young adults with ASD (n=28) and typical development (TYP; n=31) participated in Studies (i) and (ii). In Study (i) they completed a probabilistic selection (PS) task with three stimulus pairs, AB, CD, and EF. They learned to choose one of two stimuli based on probabilistic feedback valid 80%, 70%, and 60% of the time. PS performance is thought to involve the inter-workings of the basal ganglia and the orbito-frontal cortex (OFC). In Study (ii) participants completed a transitive inference (TI) task. They were trained on a partially-overlapping stimulus hierarchy with four pairs: A+B-, B+C-, C+D-, and D+E-. During a subsequent test phase, novel combinations BD and AE were tested. This task relies on the hippocampus, basal ganglia, parietal cortex, OFC, and rostrolateral PFC. In Study (iii), 12 participants with ASDs and 15 TYP completed a hierarchical learning task. They learned stimulus-response mappings under two conditions: "flat" (appropriate mappings could not be described by more general rules, and thus had to be learned individually) and "hierarchical" (learning could be greatly facilitated by uncovering more abstract rules to guide responding). This task involves intact functioning of the basal ganglia, hippocampus, and pre-PMD. Data analysis for all tasks used both univariate methods and Bayesian state-space models.
Results: In Study (i), ASDs acquired the simplest AB (80%) pair at rates equivalent to TYP, but exhibited slower learning for the more difficult CD (70%) pair. They out-performed TYP on the EF stimulus pair for which valid feedback only was provided 60% of the time. This suggests ASDs exhibit OFC deficits. In Study (ii), ASDs learned simple stimulus response associations comparably to TYP, but showed reduced interference from intervening training trials. At test, they showed no significant differences on the novel BD pair, although their performance on the AE pair was worse. Again, findings suggest PFC impairments. In Study (iii), ASD performed comparably to the TYP group in the flat condition, but showed deficits on the hierarchical task, suggesting an inability to uncover the more abstract rules, which is consistent with an impairment in PFC function.
Conclusions: This pattern suggests that ASDs have cognitive control related learning deficits. They rely on rote learning-based strategies (related to intact basal ganglia and hippocampal function) rather than using more flexible PFC-based strategies reliant on abstracting subtle patterns of reward-related information; on rapidly updating reward contingencies; and on integrating task information in the service of goal directed behavior.