Recall that the motivation for state estimation in optimal control is to finesse problems with noisy and delayed sensory input. However, FRAX597 there are also delays in descending control signals from the motor cortex. These can be discounted if we consider classical reflex arcs to be solving the easy (intrinsic) inverse problem. In other words, if motor neurons are wired to suppress proprioceptive prediction errors in the dorsal horn of the spinal cord, they effectively implement an inverse model, mapping from desired sensory consequences to causes in intrinsic (muscle-based) coordinates. In this simplification of conventional schemes, descending motor commands become top-down predictions of proprioceptive

sensations conveyed by primary and secondary sensory afferents. Note that this is not an open-loop scheme, because top-down predictions are part of a closed loop that optimizes estimates of hidden states using bottom-up (e.g., visual) sensations. This simplification speaks to the recursive and hierarchical anatomy of the motor system (Grafton and Hamilton, 2007 and Shipp, 2005) and acknowledges the role of nested, closed-loop dynamics at both peripheral and central levels.

In this scheme, optimal control signals prescribe action indirectly Trichostatin A mw through predictions about desired proprioceptive consequences. This means that their role is to provide predictions about changes in hidden states that minimize cost. These predictions (from the forward model in Figure 1) require optimal control to solve the hard (extrinsic) inverse problem. However, this is no longer necessary because control signals are not required in intrinsic coordinates (because the intrinsic consequences of extrinsic predictions drive action). It is therefore sufficient to provide the forward model

with predictions about desired trajectories in an Phenibut extrinsic frame of reference. This means that we do not have to solve the hard problem of working out how (intrinsic) muscle contractions produce (extrinsic) movements; we only have to solve the forward problem of how (extrinsic) movements stretch (intrinsic) muscles. In other words, the inverse model (optimal control) is unnecessary. This brings us to active inference. Active inference eschews the hard inverse problem by replacing optimal control signals that specify muscle movements (in an intrinsic frame) with prior beliefs about limb trajectories (in an extrinsic frame). The resulting scheme is shown in Figure 3, where the forward model now maps from prior beliefs about desired trajectories to their sensory consequences. This model is formally identical to hierarchical models used for perceptual inference. Here, motor commands become descending predictions of proprioceptive sensations, while their exteroceptive homologs become corollary discharges (see left panel of Figure 4).