Statistical Learning of Dynamic Contingencies in Autism Spectrum Conditions

Background: The capacity to implicitly process statistical regularities in one’s environment is a core aspect of perception and cognition. Prior experiences of environmental regularities are used in conjunction with incoming sensory information to arrive at the perceived interpretation of the external environment.

There have been suggestions that there may be impairments in the ability to build these ‘priors’ in autism. Despite this, studies looking at implicit learning have failed to find significant differences between people with autism and the typical population. However, it is important to note that these studies predominantly focus on implicit learning under deterministic and stable conditions.

Objectives: We set out to assess i) whether differences in implicit learning occur in individuals with autism when underlying regularities are probabilistic rather than deterministic and ii) whether individuals with autism show differences in their ability to update these priors when the underlying regularities change.

Methods: We used a probabilistic serial reaction time task in which participants were asked to use key presses to respond to a visual target which appeared on the screen in one of four possible locations.

During the task, trial outcomes were determined by a probabilistic Markov chain. This was designed so that for each 2-back context, there was a probable and improbable target location for the subsequent trial. Acquiring implicit knowledge of the underlying structure of the trial sequence leads to reduced response times during probable trials when compared to improbable trials.

The participants (15 with a diagnosis of autism and 20 controls) were asked to complete a primary phase of 8 blocks (of 120 trials each) before moving on to a secondary phase of 8 blocks in which probable and improbable locations were reversed for all contexts.

Results: Individual differences in implicit acquisition of the underlying sequence in the task were assessed by comparing response times for probable and improbable trial types. To assess differences in the ability to adapt to changes in the underlying predictive structure, we calculated gain scores for performance in the second session relative to the first session.

We conducted a Bayesian Independent T-Test on the gain scores between the autism group (ASC) and the control group (CTR). The CTR group had higher gain scores on average than the ASC group (M = 0.67 and -1.21 respectively), with a Bayes factor of 19.90 in favor of the alternative hypothesis suggesting that there is strong evidence in support of a group difference.

We then used a computational model to assess how perceptual expectations are influenced by using different lengths of temporal window when calculating the level of uncertainty (Entropy) in the task environment.

Conclusions: Under stable probabilistic conditions, the ASC group showed increased rates of implicit learning relative to the CTR group. However, when the underlying probabilistic structure changed the ASC group were slower to update these expectations.

We discuss how data from computational models can be used to infer how attending to statistical regularities across different time scales might lead to the observed group differences in the behavioral data.