Differentiating Bayesian model updating and model revision based on their prediction error dynamics

Abstract

Within predictive processing learning is construed as Bayesian model updating with the degree of certainty for different existing hypotheses changing in light of new evidence. Bayesian model updating, however, cannot explain how new hypotheses are added to a model. Model revision, unlike model updating, makes structural changes to a generative model by altering its causal connections or adding or removing hypotheses. Whilst model updating and model revision have recently been formally differentiated, they have not been empirically distinguished. The aim of this research was to empirically differentiate between model updating and revision on the basis of how they affect prediction errors and predictions over time. To study this, participants took part in a within-subject computer-based learning experiment with two phases: updating and revision. In the updating phase, participants had to predict the relationship between cues and target stimuli and in the revision phase, they had to correctly predict a change in the said relationship. Based on previous research, phasic pupil dilation was taken as a proxy for prediction error. During model updating, we expected that the prediction errors over trials would be gradually decreasing as a reflection of the continuous integration of new evidence. During model revision, in contrast, prediction errors over trials were expected to show an abrupt decrease following the successful integration of a new hypothesis within the existing model. The opposite results were expected for predictions. Our results show that the learning dynamics as reflected in pupil and accuracy data are indeed qualitatively different between the revision and the updating phase, however in the opposite direction as expected. Participants were learning more gradually in the revision phase compared to the updating phase. This could imply that participants first built multiple models from scratch in the updating phase and updated them in the revision phase.

Beyond the Adult Mind: A Developmental Framework for Predictive Processing in Infancy

“…To account for these crucial structural operations, a new type of learning has been recently proposed as an addition to predictive processing: structure learning (Friston et al, 2017;Kwisthout et al, 2017;Rutar et al, 2022;Rutar et al, 2023;Smith et al, 2020). Structure learning, unlike model updating, results in a change in the structure of a generative model by removing, adding, merging or splitting variables, and by adding or removing a causal connection between two variables (Rutar et al, 2022).…”

“…There are many other forms of learning which cannot be described by a simple updating of an existing probability distribution. If updating were the only learning mechanism in human cognition, prediction error minimisation would always be gradual, and there would be no possibility of insight or paradigm shift, but stronger still, there could be no new categories or new hypotheses, and no new explanations for existing data (Rutar et al, 2022;Rutar et al, 2023). Such forms of learning are a crucial part of development, but cannot be encapsulated by the simple model-updating process that is currently described by Predictive Processing.…”