Figure 6.

Shape discriminability under occlusion increases with the top-down prediction. The optimal average firing rates across degrees of occlusion as in Fig. 4D (yellow), projected onto the state space of V4 unit 1 (preferred) and unit 2 (non-preferred) responses. For each occlusion level, 200 responses were generated with a white noise with the mean at the optimal average value (yellow) and standard deviation of 2 arbitrary chosen for illustration purpose (blue: low occlusion, green: high occlusion). When the population responses are under the unity line (dotted black), r_v4,1 > r_v4,2, and the animal concludes that the test shape presented is shape A. The opposite is true for r_v4,2 > r_v4,1. Before the top-down prediction (A), the noisy responses under high occlusion (green dots) lie close to the unity line, obscuring the shape identity. With the top-down prediction included (B), the average optimal responses to occluded stimuli are moved horizontally to larger r_v4,1 values (yellow). Thus the noisy responses are more squeezed and moved away from the unity line, clarifying the shape identity.