Figure 4.

Model simulations. The optimal representation based on hierarchical Bayesian inference reproduces V4 and PFC responses in the experiments. (A) The network model schematic as in Fig. 1A. The solid rectangle shows the initial feedforward-only signal computation. The dotted rectangle encompasses the computations for the delayed response inferences that integrate the bottom-up sensory inputs and the top-down predictions from PFC. The corresponding optimal representations are shown in solid (initial, feedforward-only) and dotted (delayed, feedforward+feedback) lines in D–E. (B) Illustration of the input stimuli– shape A with varying degrees of occlusion. The actual images were not used as the input; the κ-dependent population response distributions of V4 neurons were used to represent the shape stimuli. Note that the occluders are of a different color than the shape or the background, and activate a group of V4 cells selective for the color. (C) Inferred PFC responses increase as occlusion level increases, in accordance with experiments. A weak shape selectivity is present, as PFC unit 1 responds at higher rates than PFC unit 2 to the presented shape A across the occlusion levels. (D) Inferred responses of the shape-selective V4 units before (solid) and after (dotted) the top-down prediction. The green lines are the optimal responses of the V4 population selective for the test shape– shape A (V4 unit 1), and the blue lines are those of the non-preferred V4 population that responds preferentially to shape B (V4 unit 2). (E) Model prediction of average firing rates of the occluder-selective V4 population (V4 unit 3), as a function of occlusion level. The salient occlusion activates this class of V4 neurons. Note that the x-axis shows fraction unoccluded.