Large-Scale Cortical Networks for Hierarchical Prediction and Prediction Error in the Primate Brain

Abstract

The predictive-coding theory proposes that cortical areas continuously generate and update predictions of sensory inputs, and emit predictionerror signals when the predicted and actual sensory inputs differ. According to the theory, the computations of predictions and prediction errors are carried out by hierarchically organized neuronal populations and transmitted between hierarchies via specific frequency channels. However, these multilevel processes are simultaneous and interdependent, making it difficult to disentangle their constituent neural network organization. Here, we test the theory by using hemisphere-wide, high-density electrocorticography (ECoG) to provide a large-scale characterization of the cortical networks for hierarchical auditory prediction and prediction-error processing in macaque monkeys. Broadband neuronal signals were collected during an auditory “local-global” paradigm in which the temporal regularities of the stimuli and their violations were controlled at two hierarchical levels. Using an automatized decomposition method for cortical activations and corticocortical interactions, we identified three distinct structures in the violation responses and further evaluated their functional interactions within and across trials. Structure 1, representing bottom-up processing of lower-level prediction errors, was reflected in γ oscillations (>40Hz) in the auditory cortex. Structure 2, representing the subsequent bottom-up processing of higher-level prediction errors, was reflected in γ oscillations in the anterior temporal cortex. Lastly, structure 3, representing a top-down updating of those predictions, was reflected in α/β-band interactions (<30Hz) from the prefrontal cortex back to the temporal cortex. Our findings provide strong support for hierarchical predictive coding and outline how it is dynamically implemented in signals using distinct areas and frequency bands.