Figure 2. Predictive processing from computation to whole-brain dynamics.

(A) Schematic depiction of the original predictive coding model (Rao & Ballard, 1999). Sensory input drives feedforward error signals which are processed by predictive estimators (PE). PE modules consist of neurons whose activity (r) is compared to top-down activity from higher levels (i.e., prediction; rtd) and the difference of which is propagated forward to the next level as error (r-rtd). (B) Flow of prediction and prediction error signals between cortical columns based on cortical lamination gradients. Using anterograde and retrograde tracers, Barbas and colleagues showed that the relative difference in laminar structure between two communicating cortical columns predicts whether the information flow is a feedback (prediction) or a feedforward (prediction error) signal. Prediction signals (in green) originate in the deep layers (layers V and VI) of less differentiated cortical areas (such as agranular cortex with undifferentiated layers II and III and without a layer IV, as depicted in the red column) and terminate in superficial layers of areas with a more developed laminar structure (such as dysgranular cortices with differentiated layers II and III and a rudimentary layer IV or granular cortices with differentiated layers II and III and a well-defined layer IV, depicted in the yellow column). Prediction error signals (in purple) flow in the other direction, originating in the superficial layers (II–III) with more laminar differentiation and terminating in middle deep layers (V–VI) of areas with less differentiated laminar architecture. This structural model successfully predicts the flow of information in frontal, temporal, and parietal cortices in experiments with monkeys and cats see Barbas, 2015, for a review). (C) Whole-brain estimate of flow of prediction and prediction error signals. Based on whole-brain cortical granularity data (Triarhou, 2008; von Economo, 2009). Predictions flow from cortical regions with less laminar differentiation to regions with increasing laminar differentiation (e.g., from limbic cortices to motor, interoceptive, and primary somatosensory, auditory and visual cortices). Many of the anatomical and computational details are still under investigation, such as whether each individual neuron is capable of coding for internal representations (i.e., predictions) and comparing those predictions to incoming inputs from lower in the hierarchy (i.e., prediction errors) or whether predictions and prediction errors are coded by different neurons in the cortex.