Figure 4.

Results of simulated lesion studies using the birdsong model of figure 3. (i) The percept in terms of the predicted sonograms and (ii) the corresponding prediction error (at both the levels); these are the differences between the incoming sensory information and the prediction and the discrepancy between the conditional expectation of the second-level cause and that predicted by the second-level hidden states. (a) The recognition dynamics in the intact bird. (b) The percept and corresponding prediction errors when the connections between the hidden states at the second level and their corresponding causes are removed. This effectively removes structural priors on the evolution of the attractor manifold prescribing the sensory dynamics at the first level. (c) The effects of retaining the structural priors but removing the dynamical priors by cutting the connections that mediate inversion in generalized coordinates. These results suggest that both structural and dynamical priors are necessary for veridical perception.