A recurrent neuronal model for the effect of predictions on sensory processes

Predictions, prior knowledge, and expectations strongly influence sensory processing and decision-making, yet the neural mechanisms underlying these effects remain a central question in sensory neuroscience. Although the well-established predictive coding model has gained traction as a computational framework for understanding these effects, empirical findings are not always consistent with its predictions. Here, we present an alternative recurrent neuronal model that implements predictive processing through feature-tuned neural populations integrating feedforward sensory input with feedback signals via circuit dynamics consistent with established features of cortical organization. Model parameters were first constrained by fitting the model to behavioral threshold data from a cue-based expectation paradigm using faces and houses as stimuli, allowing the model to capture the characteristic pattern of behavioral prediction effects observed in that study. Without further adjustment of the participant-specific parameters estimated from the behavioral data, the model was then used to simulate neural responses under the experimental protocols of three independent fMRI studies drawn from the literature. Despite differences in stimulus categories and task designs across these studies, the model accounted for key aspects of their main neuroimaging findings, demonstrating generalization from behavior to neural activity. Overall, these results show that a single recurrent neuronal mechanism can account for both behavioral and neural effects of prediction, offering a theoretical account of how prior knowledge, predictions, and expectations influence sensory processing.