Predictive coding in ASD: inflexible weighting of prediction errors when switching from stable to volatile environments

Abstract

The nature of predictive-processing differences between individuals with autism spectrum disorder (ASD) and typically developing (TD) individuals is widely debated. Some studies suggest impairments in predictive processing in ASD, while others report intact processes, albeit with atypical learning dynamics. Here, we assessed duration reproduction tasks in high- and low-volatility settings to examine the updating dynamics of prior beliefs and sensory estimates. Employing a two-state Bayesian model, we differentiated how individuals with ASD and TD controls update their priors and perceptual estimates, and how these updates affect long-term prediction and behavior. Our findings indicate that individuals with ASD use prior knowledge and sensory input similarly to TD controls in perceptual estimates. However, they place a greater weight on sensory inputs specifically for iteratively updating their priors. This distinct approach to prior updating led to slower adaptation across trials; individuals with ASD relied less on their priors in perceptual estimates during the first half of sessions but achieved comparable integration weights as TD controls by the end of the session. By differentiating these aspects, our study highlights the importance of considering inter-trial updating dynamics to reconcile diverse findings of predictive processing in ASD. In consequence to the current findings, we suggest the distinct iterative updating account of predictive processing in ASD. Significance StatementResearch on predictive processing in Autism Spectrum Disorder (ASD) remains controversial. The current study employed a two-state Bayesian model in varied volatility settings to explore inter-trial updating dynamics in ASD compared to typically developing (TD) peers. We found that individuals with ASD, while utilizing prior knowledge similarly to TD controls, place a disproportionate emphasis on sensory inputs when updating their priors. This unique pattern of slower adaptation during iterative updating leads to significant behavioral differences in the first half of trials between the two groups, but comparable levels by the end of the session. These findings not only highlight the importance of considering different timescales and dynamic updating processes in ASD, but also suggest that the predictive processing framework in ASD involves unique prior updating mechanisms that is likely associated with increased sensory reliance.