Examining and extending Bayesian theories of autism

Abstract

Recent approaches to understanding autism view it through the lens of the Bayesian brain framework. Within this framework, it is hypothesized that beliefs about the world are represented probabilistically in the brain and are updated using Bayesian statistics. Perception is understood a process of inferring the most likely cause of sensory input by combining sensory information with prior beliefs about the environment. Bayesian theories of autism propose that its heterogeneous symptomatology arises from an imbalance in the weighting or ‘precision’ of priors and likelihoods, favouring the latter. While this is a promising approach, the relevant research is highly diverse, occasionally resulting in contradictory findings, with some studies finding no evidence of such imbalances in autism and others showing more nuanced and complex results.

The aim of this work is twofold. First, we conducted a comprehensive examination of previous research, starting by conducting a systematic review of Bayesian studies of autism and autistic traits. The results were mixed, with a slight majority of studies finding no difference in the integration of Bayesian priors and likelihoods. Methodologically, many studies had low statistical power and inconsistent approaches. In a subsequent paper, we clarified the meaning of ‘sensory precision’, a central term in Bayesian theories of autism. Using a simple Bayesian perception model, we examined the role of two possible interpretations in the inferential process and their relevance to autism theories. We reanalysed data from an old study under this new light.

The second part of this work expands upon the main Bayesian theories of autism to better approximate the behaviours of autistic individuals. Informed by the literature review, we designed an experiment to compare implicit and explicit learning across autistic traits, manipulating the presence and content of instructions across conditions. We found that the presence of instructions exerted a significant influence on participant priors until the end of the task, while the effect of their content had only a weak, temporary influence. Our results showed no relationship between autistic traits and response biases, but offered hints of a weak correlation between autistic traits and participant uncertainty about implicit environmental regularities. In another study, we applied the circular inference model, a model of belief overconfidence, to autistic traits. Our findings did not reveal any significant differences in circularity across the autism spectrum. To further explore this model in a more autism-relevant context, we designed a follow-up task that incorporated social elements. Surprisingly, the pilot results indicated that participants did not engage in Bayesian inference, instead they simply averaged the information obtained from different sources.

These results show that, while imbalance theories of autism are prominent within the field, their simplest form is weakly supported by experimental evidence. Both the literature review and our study on implicit and explicit learning highlighted the need for more nuanced approaches that focus on prior development. We discuss two such theories of volatility processing impairments in autism, which imply higher uncertainty in strong autistic traits, as hinted by our experimental findings. We also discuss methodological issues commonly present in the field and the need for thorough computational modelling.