Fig. 2.

This figure summarizes hierarchical neuronal message passing in predictive coding using the (simplified) neuroanatomy of a songbird. Neuronal activity encodes expectations about the causes of sensory input, where these expectations minimize prediction error. Prediction error is the difference between (ascending) sensory input and (descending) predictions of that input. Here, sensory input is represented by a sonogram encoding the amplitude of different acoustic frequencies over time. Minimising prediction error rests upon recurrent neuronal interactions among different levels of the cortical hierarchy. The available evidence suggests that superficial pyramidal cells (red triangles) compare the expectations (at each level) with top-down predictions from deep pyramidal cells (black triangles) of higher levels. Left panel: these equations represent the neuronal dynamics implicit in predictive coding. Prediction errors at the i-th level of the hierarchy are simply the difference between the expectations encoded at that level and top-down predictions of those expectations. The expectations per se are driven by prediction errors so that they reduce the sum of squared (precision weighted) prediction error. See the appendix for a detailed explanation of these equations and the variables in this figure. Right panel: this provides a schematic example in the auditory system of a songbird: it shows the putative cells of origin of ascending or forward connections that convey (precision weighted) prediction errors (red arrows) and descending or backward connections (black arrows) that construct predictions. In this example, area X sends predictions to HVC (c.f., high vocal centre), which projects to the auditory thalamus. However, the HVC also sends proprioceptive predictions to the hypoglossal nucleus, which are passed to the syrinx to generate vocalisation through classical reflexes. These predictions can be regarded as motor commands, while the descending predictions of auditory input correspond to corollary discharge. Note that every top-down prediction is reciprocated with a bottom-up prediction error to ensure predictions are constrained by sensory information. The neuroanatomy implicit in this schematic should not be taken too seriously: we have simply transcribed a generic hierarchical message passing scheme onto the key connections among the cardinal regions implicated in the processing of birdsongs (Nottebohm, 2005).