FIGURE 3.
A template circuit diagram to implement predictive coding. A neural interpretation of the computational architecture shown in Figure 1. The schematic is modeled upon a non-primary area of primate visual cortex, such as area V2 or V4. It shows the correspondence of a subset of the computational interactions in Figure 1 – those updating scenic causes – to circuit elements bearing the same numbers. These neural pathways include both extrinsic and intrinsic connections, some of them polysynaptic. The extrinsic connections are hierarchical and involve reciprocal connections of one area, level (i), with the levels above (i+1) and below (i–1). The intrinsic connections form a canonical microcircuit that has been summarized as successive relays from layer 4 to 3 (and 2), and 3 to 5, with layer 6 receiving more diffuse input from the layers above. All computational units are designated to be subclasses of excitatory pyramidal neurons: expectation units (rectangular, ‘Exp’) prediction error units (circular, ‘Err’), or precision units (triangular, ‘Prc’). Two pathways (8 and 14) are explicitly indicated to require inhibitory transmission, but all pathways are likely to subsume contacts on both pyramidal neurons and interneurons. The circuitry may be summarized as follows. Forward, extrinsic connections terminate in layers 4 and 3B (pathway 6). Activity in expectation units is maintained by recurrent interactions between these units (pathway 5) and is informed by prediction error signals: excitation by the ascending prediction error (pathway 6) and inhibition by local error units in layer 3B (pathway 8), which are also the source of ascending prediction error routed to the level above (pathway 6). These superficial error units in layer 3B implement a subtraction of two signals: an excitatory signal received from their associated expectation units (pathway 13) and a descending prediction, relayed by a local interneuron (pathway 14). Interlaminar connections from layer 3 to layers 5 and 6 may possibly transmit all three classes of signal, but only the relay of expectation signals is shown here. These contact the deep pyramidal neurons that are the source of descending predictions (pathway 14). Precision signals are shown to arise from layers 2 and 3A, and form a descending chain of transmission through the superficial layers (pathway 12); these signals are capable of modulating pyramidal error units via their apical dendrite. Figure reproduced with modification from Shipp et al. (2013).