Neural Elements for Predictive Coding

Abstract

Predictive coding theories of sensory brain function interpret the hierarchical construction of the cerebral cortex as a Bayesian, generative model capable of predicting the sensory data consistent with any given percept. Predictions are fed backward in the hierarchy and reciprocated by prediction error in the forward direction, acting to modify the representation of the outside world at increasing levels of abstraction, and so to optimize the nature of perception over a series of iterations. This accounts for many ‘illusory’ instances of perception where what is seen (heard, etc.) is unduly influenced by what is expected, based on past experience. This simple conception, the hierarchical exchange of prediction and prediction error, confronts a rich cortical microcircuitry that is yet to be fully documented. This article presents the view that, in the current state of theory and practice, it is profitable to begin a two-way exchange: that predictive coding theory can support an understanding of cortical microcircuit function, and prompt particular aspects of future investigation, whilst existing knowledge of microcircuitry can, in return, influence theoretical development. As an example, a neural inference arising from the earliest formulations of predictive coding is that the source populations of forward and backward pathways should be completely separate, given their functional distinction; this aspect of circuitry – that neurons with extrinsically bifurcating axons do not project in both directions – has only recently been confirmed. Here, the computational architecture prescribed by a generalized (free-energy) formulation of predictive coding is combined with the classic ‘canonical microcircuit’ and the laminar architecture of hierarchical extrinsic connectivity to produce a template schematic, that is further examined in the light of (a) updates in the microcircuitry of primate visual cortex, and (b) rapid technical advances made possible by transgenic neural engineering in the mouse. The exercise highlights a number of recurring themes, amongst them the consideration of interneuron diversity as a spur to theoretical development and the potential for specifying a pyramidal neuron’s function by its individual ‘connectome,’ combining its extrinsic projection (forward, backward or subcortical) with evaluation of its intrinsic network (e.g., unidirectional versus bidirectional connections with other pyramidal neurons).

Auditory predictions and prediction errors in response to self-initiated vowels

“…Other current accounts of surprising (e.g., oddball or deviant) responses in a more general, modality-independent framework are usually cast in terms of predictive coding (e.g. Brown et al 2013;Friston, 2018;Shipp, 2016). In these models, the brain uses a hierarchical forward or generative model inferring the causes of sensations from its sensory consequences.…”

Top‐down versus bottom‐up attention differentially modulate frontal–parietal connectivity

“…A key example of this is negative extrinsic (between regions) cortical connectivity, which is mediated by glutamatergic neurotransmission. This means that a negative directed (effective) connection has to be mediated by inhibitory neurons, which are usually the target of backward or descending extrinsic connections (Shipp, 2016). Similar arguments pertain to changes in connectivity.…”

“…Furthermore, anatomical tracing studies suggest that FEF can occupy a relatively low level in the visual hierarchy, based upon asymmetries in the laminar specificity of extrinsic connections (Anderson, Kennedy, & Martin, 2011;Vezoli et al, 2004). Finally, detailed analysis of the microcircuitry in the motor cortex leads to the conclusion that (motor) executive components of sensorimotor hierarchies are hierarchically subordinate to regions such as the IPS (Shipp, 2016;Shipp, Adams, & Friston, 2013). On this view, the excitatory forward connections from FEF to IPS are consistent with the driving influences of ascending prediction errors (Lee & Mumford, 2003;Mumford, 1992;Rao & Ballard, 1999).…”

Laminar-specific cortico-cortical loops in mouse visual cortex

“…The lack of specificity for looped neurons in L2/3 is surprising in light of experimental evidence associating this layer with prediction errors. Neural responses consistent with predictive processes have been identified in L2/3 in several cortical areas [45][46][47][48] and the layer is considered critical for generating prediction error signals 5,6,49 . Predictive models require excitatory and inhibitory looped interactions of FB inputs with lower-level neurons to signal violations of expectations 6 .…”