Disentangling “Bayesian brain” theories of autism spectrum disorder

Abstract

BackgroundBayesian theories of perception have provided a variety of alternative mechanistic explanations for autistic symptoms, including (1) overprecise sensations, (2) imprecise priors, (3) inflexible priors, and (4) altered hierarchical learning. Here, we designed a set of experiments to systematically test predictions from each of the four hypotheses in individuals with autism spectrum disorder (ASD).MethodsTwo versions of a two-alternative forced-choice (2AFC) random dot motion task were developed to disentangle the influence of sensory inputs and prior expectations and test implications of the four Bayesian hypotheses. Using a cross-sectional observational study design, behavioural data were obtained from participants with autism spectrum disorder (N=47) and a control group (N=50). Analyses used mixed effects models and a drift diffusion model.ResultsContrary to the sensory overprecision hypothesis, individuals with ASD did not differ from controls in performance or parameter estimates during the perceptual 2AFC task. When adding cues to the task, individuals with ASD profited less from this prior information than controls, as predicted by the imprecise prior hypothesis. However, individuals with ASD were still able to update their initial priors, contradicting the inflexible prior hypothesis. Finally, in line with the hierarchical learning hypothesis, volatility differentially modulated expectation effects, exerting a significantly smaller effect in the ASD group compared to the control group.ConclusionOur findings support two mutually compatible alterations of perception in ASD (imprecise priors and, to a lesser degree, altered hierarchical learning). By contrast, they are not compatible with the notions of sensory overprecision and inflexible priors.

Testing predictive coding theories of autism spectrum disorder using models of active inference

“…Recent neuro-computational theories have identified a variety of mechanisms that could explain atypical sensory processing [ 1 ] and movement control [ 8 ] in autism, but these theoretical accounts are yet to be adequately tested using the types of naturalistic behavioural tasks that autistic people find challenging (but see [ 74 ] for comparison of Bayesian theories of autism in a purely perceptual task). To probe the generative processes responsible, we applied a generative modelling approach to active inference behaviours from both an object lifting paradigm and an interceptive movement task.…”

Testing predictive coding theories of autism spectrum disorder using models of active inference

“…Indeed, although the location of predictive gaze fixations may generally be atypical in 4 But see (Schneebeli et al, 2022) for comparison of Bayesian theories of autism in a purely perceptual task.…”

A common computational and neural anomaly across mouse models of autism

“…To address the phenotypic heterogeneity, computational psychiatry [8][9][10][11][12][13][14][15][16][17] has recently cast ASD with the language of probabilistic inference 18 and suggested that individuals within the spectrum (1) have attenuated expectations 19 (2) inflexibly weight predictions relative to sensory observations 20 and/or (3) over-estimate the volatility of their sensory environment and thus are less surprised by statistically unlikely events 21 . Critically, each of these computational accounts suggest that the learning of context-dependent statistical regularities (i.e., "contextual priors 22 ") is abnormal (e.g., slow 23,24 ) in ASD.…”