Learning to predict based on self- versus externally induced prediction violations: a direct comparison using a Bayesian inference modelling approach

Abstract

Predictive processing is quickly gaining ground as a theory of perception and attention. From this perspective the brain is cast as an organisms predictive model of how its world works and will continue to work in the future. However, research on the brains predictive capacities remains beholden to traditional research practices in which participants are passively shown stimuli without their active involvement. The current study is an investigation into ways in which self-generated predictions may differ from externally induced predictions. Participants completed a volatile spatial attention task under both conditions on different days. We used the Hierarchical Gaussian Filter, an approximate Bayesian inference model, to determine subject-specific parameters of belief-updating and inferred volatility. We found preliminary evidence in support of self-generated predictions incurring a larger reaction time cost when violated compared to predictions induced by sensory cue, which translated to participants increased sensitivity to changes in environmental volatility. Our results suggest that internally generated predictions may be afforded more weight, but these results are complicated by session order and duration effects, as well as a lack of statistical power. We discuss the limitations of our study preventing us from replicating previous research, and ways to remedy these shortcomings in future studies.