Fig. 10.

A hierarchical generative model capturing multiple timescales and reference distributions at each level. Without addressing empirical questions about neural hierarchies, here we employ a model with L = 4 levels to match Fig. 6. For l ∈ [1. . L − 1] each $s_t^{(l)}$ node denotes an unobserved latent state, and each ${\rho }_t^{(l-1)}$ represents parameters of a reference distribution for $s_t^{(l-1)}$. o_t represents observed sensory outcomes, and a_t represents the closed-loop control actions generated by motor reflexes. Arrows between random variables denote conditional dependencies. Arrows stretch further to the right when they denote change over longer time scales.