Fig. 2.

Generative process and model of oculomotor pursuit movements. This schematic illustrates the process (left panel) and generative model of that process (right panel) used to simulate Bayes optimal pursuit. The graphics on the left show a putative predictive coding scheme (with superficial pyramidal cells in red and deep pyramidal cells in black in the pontine nuclei) processing proprioceptive information during oculomotor pursuit. These cells receive proprioceptive information from an inverse model in the subcortical oculomotor system and respond reflexively to minimise proprioceptive prediction error through action. This prediction error rests on descending predictions from the generative model on the right. The actual movement of the target is determined by a hidden cause (target location), which determines the visual input for any given direction of gaze. The generative model entails beliefs about how the target and eyes move. In brief, this model includes an invisible location that attracts the target, causing it to move. Crucially, the agent believes that its centre of gaze is attracted to this location (and the target), where the forces of attraction may (or may not) depend upon occlusion of the target and its attracting location. These forces of attraction are illustrated with lilac arrows in the top right; the arrows are labelled with their respective multipliers from the equations directly below. Please see main text for a description of the variables in the equations describing the motion of hidden states and how they depend upon hidden causes. Note that real states that are hidden from observation in the real world are in bold, whereas the hidden states assumed by the generative model are in italics. (For interpretation of the references to color in the text, the reader is referred to the web version of this article.)