Fig. 5. A computational model incorporating hierarchical predictive coding and attractor networks reproduces main experimental findings.

A Model consists of three layers (sensory, concept, and prior), each containing neural populations whose activity is tuned to one or the other percept. Communication between layers is in the form of bottom-up prediction error and top-down prediction. Communication between populations tuned to different percepts only occurs within the concept layer in the form of mutual inhibition. Excitatory interaction is displayed as arrows, inhibitory interaction as circles. For model details see methods. B Example model output shows how rates of the populations tuned to the preferred (green) and non-preferred (purple) percept, respectively, change over time. The currently experienced percept is defined by the population with higher firing rate within the concept layer and is indicated by shading. Prediction-error inputs received by the prior and concept layers are also shown. C Preferred percepts have longer durations than non-preferred percepts. D The distribution of the strengths of prediction and prediction error signals between layers during the preferred (green) and non-preferred (purple) percept. E Summary of D: during preferred percept top-down prediction is stronger between layers, and during non-preferred percept bottom-up prediction error is stronger.