Reflections on agranular architecture: predictive coding in the motor cortex
2013; Elsevier BV; Volume: 36; Issue: 12 Linguagem: Inglês
10.1016/j.tins.2013.09.004
ISSN1878-108X
AutoresStewart Shipp, Rick A. Adams, Karl Friston,
Tópico(s)Motor Control and Adaptation
Resumo•Predictive coding explains the recursive hierarchical structure of cortical processes.•Granular layer 4, which relays ascending cortical pathways, is absent from motor cortex.•Perceptual inference results if ascending sensory data modify sensory predictions action, if spinal reflexes enact descending motor and/or proprioceptive predictions.•Motor layer 4 regresses as motor predictions inherently require less modification. The agranular architecture of motor cortex lacks a functional interpretation. Here, we consider a 'predictive coding' account of this unique feature based on asymmetries in hierarchical cortical connections. In sensory cortex, layer 4 (the granular layer) is the target of ascending pathways. We theorise that the operation of predictive coding in the motor system (a process termed 'active inference') provides a principled rationale for the apparent recession of the ascending pathway in motor cortex. The extension of this theory to interlaminar circuitry also accounts for a sub-class of 'mirror neuron' in motor cortex – whose activity is suppressed when observing an action –explaining how predictive coding can gate hierarchical processing to switch between perception and action. The agranular architecture of motor cortex lacks a functional interpretation. Here, we consider a 'predictive coding' account of this unique feature based on asymmetries in hierarchical cortical connections. In sensory cortex, layer 4 (the granular layer) is the target of ascending pathways. We theorise that the operation of predictive coding in the motor system (a process termed 'active inference') provides a principled rationale for the apparent recession of the ascending pathway in motor cortex. The extension of this theory to interlaminar circuitry also accounts for a sub-class of 'mirror neuron' in motor cortex – whose activity is suppressed when observing an action –explaining how predictive coding can gate hierarchical processing to switch between perception and action. Motor cortex was localised to the precentral gyrus of apes by Sherrington in 1901 [1Sherrington C.S. Grunbaum A.S.F. An address on localisation in the 'motor' cerebral cortex as exemplified in the anthropoid apes.BMJ. 1901; ii: 1857-1859Crossref Scopus (13) Google Scholar], and was first identified histologically the following year by Campbell, using the brains of Sherrington's subjects [2Macmillan M. Alfred Walter Campbell and the visual functions of the occipital cortex.Cortex. 2012; https://doi.org/10.1016/j.cortex.2012.10.007Abstract Full Text Full Text PDF PubMed Scopus (10) Google Scholar]. Although Campbell emphasised the prominent fibre architecture of motor cortex, it is the cytoarchitectural tag 'agranular' cortex, coined by Brodmann [3Brodmann, K., translated by Garey, L.J. (1909/1994) Brodmann's Localisation in the Cerebral Cortex, Smith-GordonGoogle Scholar] to describe his areas 4 and 6, that has proved the more enduring. Both authors used variations in cortical architecture for cartographic purposes, but these variants have rarely, if ever, been interpreted functionally; despite its 'fame', the dramatic recession of granular layer 4 in motor cortex has not attracted a single functional hypothesis. Cortical layers are identified by cytoarchitecture, and further characterised by patterns of intrinsic axonal and dendritic arborisation [4Shipp S. Structure and function of the cerebral cortex.Curr. Biol. 2007; 17: 443-449Abstract Full Text Full Text PDF Scopus (144) Google Scholar]. Laminar distribution distinguishes consistent types of extrinsic corticocortical connection, classified as ascending, descending, and lateral [5Felleman D.J. Van Essen D.C. Distributed hierarchical processing in the primate cerebral cortex.Cereb. Cortex. 1991; 1: 1-47Crossref PubMed Scopus (5507) Google Scholar]. These patterns are sufficiently conserved to identify a hierarchical organisation of areas in sensory systems (Box 1). Initial descriptions of sensorimotor hierarchies placed premotor above primary motor cortex (M1), with areas 3a and 3b (components of the primary somatosensory area, S1) at the lowest levels [5Felleman D.J. Van Essen D.C. Distributed hierarchical processing in the primate cerebral cortex.Cereb. Cortex. 1991; 1: 1-47Crossref PubMed Scopus (5507) Google Scholar]. Our own survey aimed to establish the polarity of key reciprocal connections, but not to arrange areas into discrete tiers [6Adams R.A. et al.Predictions not commands: active inference in the motor system.Brain Struct. Funct. 2013; 218: 611-643Crossref PubMed Scopus (414) Google Scholar]. The absence of a distinct granular layer in primary motor cortex calls for some modification of the laminar criteria, but the presence of a cryptic layer 4 [7Sloper J.J. et al.A qualitative and quantitative electron microscopic study of the neurons in the primate motor and somatic sensory cortices.Philos. Trans. R. Soc. Lond. B: Biol. Sci. 1979; 285: 141-171Crossref PubMed Scopus (64) Google Scholar, 8Skoglund T.S. et al.The existence of a layer IV in the rat motor cortex.Cereb. Cortex. 1997; 7: 178-180Crossref PubMed Scopus (38) Google Scholar, 9Mao T. et al.Long-range neuronal circuits underlying the interaction between sensory and motor cortex.Neuron. 2011; 72: 111-123Abstract Full Text Full Text PDF PubMed Scopus (318) Google Scholar] justifies the treatment of terminal patterns that target the layer 3/5 border zone as forward connections (or backward, if the pattern avoids this zone). Similar arguments apply to premotor cortex (Brodmann's area 6), sometimes described as 'dysgranular' [10Watanabe-Sawaguchi K. et al.Cytoarchitecture and intrafrontal connections of the frontal cortex of the brain of the hamadryas baboon (Papio hamadryas).J. Comp. Neurol. 1991; 311: 108-133Crossref PubMed Scopus (49) Google Scholar], owing to a rudimentary granular layer.Box 1Laminar specific connectivity and hierarchical distanceFigure I shows the laminar sources and distributions of ascending connections (green) and descending connections (red, violet, and blue), originating from a certain level (i) in a hierarchical chain.The basic laminar patterns distinguishing ascending and descending connections were originally established by studies of primate visual cortex [5Felleman D.J. Van Essen D.C. Distributed hierarchical processing in the primate cerebral cortex.Cereb. Cortex. 1991; 1: 1-47Crossref PubMed Scopus (5507) Google Scholar, 97Rockland K.S. Pandya D.N. Laminar origins and terminations of cortical connections of the occipital lobe in the rhesus monkey.Brain Res. 1979; 179: 3-20Crossref PubMed Scopus (693) Google Scholar, 98Maunsell J.H.R. Van Essen D.C. The connections of the middle temporal area and their relationship to a cortical hierarchy in the macaque monkey.J. Neurosci. 1983; 3: 2563-2586PubMed Google Scholar]. Systematic variations with hierarchical distance were later formulated as a 'distance rule' [37Markov N.T. Kennedy H. The importance of being hierarchical.Curr. Opin. Neurobiol. 2013; 23: 187-194Crossref PubMed Scopus (105) Google Scholar, 99Barone P. et al.Laminar distribution of neurons in extrastriate areas projecting to visual areas V1 and V4 correlates with the hierarchical rank and indicates the operation of a distance rule.J. Neurosci. 2000; 20: 3263-32681PubMed Google Scholar]: with regard to origins, the proportion of superficial (layer 3A) neurons forming a backward projection decreases with greater distance spanned by the projection [99Barone P. et al.Laminar distribution of neurons in extrastriate areas projecting to visual areas V1 and V4 correlates with the hierarchical rank and indicates the operation of a distance rule.J. Neurosci. 2000; 20: 3263-32681PubMed Google Scholar, 100Sousa A.P.B. et al.Topographic organization of cortical input to striate cortex in the Cebus monkey: a fluorescent tracer study.J. Comp. Neurol. 1991; 308: 665-682Crossref PubMed Scopus (80) Google Scholar, 101Rockland K.S. Van Hoesen G.W. Direct temporal-occipital feedback connections to striate cortex (V1) in the macaque monkey.Cereb. Cortex. 1994; 4: 300-313Crossref PubMed Scopus (154) Google Scholar, 102Perkel D.J. et al.Topography of the afferent connectivity of area 17 in the macaque monkey: a double-labelling study.J. Comp. Neurol. 1986; 253: 374-402Crossref PubMed Scopus (137) Google Scholar], illustrated in Figure I by the 'red' terminals failing to contact level (i-2). The backward projection originating from deep layers reaches further, but these descending terminations (blue) show a progressive shift of focus upon layers 1 and 6 [51Rockland K.S. et al.Divergent feedback connections from areas V4 and TEO in the macaque.Vis. Neurosci. 1994; 11: 579-600Crossref PubMed Scopus (117) Google Scholar]. In the opposite direction, levels (i+1) and (i+2) show a progressive shift of focus of ascending terminations (green) upon layer 4 [97Rockland K.S. Pandya D.N. Laminar origins and terminations of cortical connections of the occipital lobe in the rhesus monkey.Brain Res. 1979; 179: 3-20Crossref PubMed Scopus (693) Google Scholar].The differential contribution of superficial and deep sources to superficial and deep terminations in descending projections is not well established, because few studies have used tracers with subtotal layer deposition to study interareal connectivity. At minimum, the rule may be that like connects with like, laminar-wise. Layer 6, for instance, receives the densest input when the source of the descending projection includes layer 6 of the higher area [103Coogan T.A. Burkhalter A. Hierarchical organization of areas in rat visual cortex.J. Neurosci. 1993; 13: 3749-3772PubMed Google Scholar, 104Henry G.H. et al.Projections from areas 18 and 19 to cat striate cortex: divergence and laminar specificity.Eur. J. Neurosci. 1991; 3: 186-200Crossref PubMed Scopus (40) Google Scholar]. However, layer 1 can receive descending input from deep layers in systems as diverse as primate visual and rodent somatomotor cortex [105Cauller L.J. et al.Backward cortical projections to primary somatosensory cortex in rats extend long horizontal axons in layer I.J. Comp. Neurol. 1998; 390: 297-310Crossref PubMed Scopus (156) Google Scholar, 106Angelucci A. et al.Circuits for local and global signal integration in primary visual cortex.J. Neurosci. 2002; 22: 8633-8646PubMed Google Scholar], and layer 5 can receive descending input from superficial sources, at least in cat and rat area V1 [103Coogan T.A. Burkhalter A. Hierarchical organization of areas in rat visual cortex.J. Neurosci. 1993; 13: 3749-3772PubMed Google Scholar, 104Henry G.H. et al.Projections from areas 18 and 19 to cat striate cortex: divergence and laminar specificity.Eur. J. Neurosci. 1991; 3: 186-200Crossref PubMed Scopus (40) Google Scholar]. These patterns are summarised in Figure I by the violet tone of descending terminations to layer 5 and superficial layers in level (i-1), indicating a mix of superficial (red) and deep (blue) sources at level (i). The blueing of terminals in deeper layers of level (i-1), and all layers in level (i-2), indicates a progressive domination of deep layer sources from level (i). Figure I shows the laminar sources and distributions of ascending connections (green) and descending connections (red, violet, and blue), originating from a certain level (i) in a hierarchical chain. The basic laminar patterns distinguishing ascending and descending connections were originally established by studies of primate visual cortex [5Felleman D.J. Van Essen D.C. Distributed hierarchical processing in the primate cerebral cortex.Cereb. Cortex. 1991; 1: 1-47Crossref PubMed Scopus (5507) Google Scholar, 97Rockland K.S. Pandya D.N. Laminar origins and terminations of cortical connections of the occipital lobe in the rhesus monkey.Brain Res. 1979; 179: 3-20Crossref PubMed Scopus (693) Google Scholar, 98Maunsell J.H.R. Van Essen D.C. The connections of the middle temporal area and their relationship to a cortical hierarchy in the macaque monkey.J. Neurosci. 1983; 3: 2563-2586PubMed Google Scholar]. Systematic variations with hierarchical distance were later formulated as a 'distance rule' [37Markov N.T. Kennedy H. The importance of being hierarchical.Curr. Opin. Neurobiol. 2013; 23: 187-194Crossref PubMed Scopus (105) Google Scholar, 99Barone P. et al.Laminar distribution of neurons in extrastriate areas projecting to visual areas V1 and V4 correlates with the hierarchical rank and indicates the operation of a distance rule.J. Neurosci. 2000; 20: 3263-32681PubMed Google Scholar]: with regard to origins, the proportion of superficial (layer 3A) neurons forming a backward projection decreases with greater distance spanned by the projection [99Barone P. et al.Laminar distribution of neurons in extrastriate areas projecting to visual areas V1 and V4 correlates with the hierarchical rank and indicates the operation of a distance rule.J. Neurosci. 2000; 20: 3263-32681PubMed Google Scholar, 100Sousa A.P.B. et al.Topographic organization of cortical input to striate cortex in the Cebus monkey: a fluorescent tracer study.J. Comp. Neurol. 1991; 308: 665-682Crossref PubMed Scopus (80) Google Scholar, 101Rockland K.S. Van Hoesen G.W. Direct temporal-occipital feedback connections to striate cortex (V1) in the macaque monkey.Cereb. Cortex. 1994; 4: 300-313Crossref PubMed Scopus (154) Google Scholar, 102Perkel D.J. et al.Topography of the afferent connectivity of area 17 in the macaque monkey: a double-labelling study.J. Comp. Neurol. 1986; 253: 374-402Crossref PubMed Scopus (137) Google Scholar], illustrated in Figure I by the 'red' terminals failing to contact level (i-2). The backward projection originating from deep layers reaches further, but these descending terminations (blue) show a progressive shift of focus upon layers 1 and 6 [51Rockland K.S. et al.Divergent feedback connections from areas V4 and TEO in the macaque.Vis. Neurosci. 1994; 11: 579-600Crossref PubMed Scopus (117) Google Scholar]. In the opposite direction, levels (i+1) and (i+2) show a progressive shift of focus of ascending terminations (green) upon layer 4 [97Rockland K.S. Pandya D.N. Laminar origins and terminations of cortical connections of the occipital lobe in the rhesus monkey.Brain Res. 1979; 179: 3-20Crossref PubMed Scopus (693) Google Scholar]. The differential contribution of superficial and deep sources to superficial and deep terminations in descending projections is not well established, because few studies have used tracers with subtotal layer deposition to study interareal connectivity. At minimum, the rule may be that like connects with like, laminar-wise. Layer 6, for instance, receives the densest input when the source of the descending projection includes layer 6 of the higher area [103Coogan T.A. Burkhalter A. Hierarchical organization of areas in rat visual cortex.J. Neurosci. 1993; 13: 3749-3772PubMed Google Scholar, 104Henry G.H. et al.Projections from areas 18 and 19 to cat striate cortex: divergence and laminar specificity.Eur. J. Neurosci. 1991; 3: 186-200Crossref PubMed Scopus (40) Google Scholar]. However, layer 1 can receive descending input from deep layers in systems as diverse as primate visual and rodent somatomotor cortex [105Cauller L.J. et al.Backward cortical projections to primary somatosensory cortex in rats extend long horizontal axons in layer I.J. Comp. Neurol. 1998; 390: 297-310Crossref PubMed Scopus (156) Google Scholar, 106Angelucci A. et al.Circuits for local and global signal integration in primary visual cortex.J. Neurosci. 2002; 22: 8633-8646PubMed Google Scholar], and layer 5 can receive descending input from superficial sources, at least in cat and rat area V1 [103Coogan T.A. Burkhalter A. Hierarchical organization of areas in rat visual cortex.J. Neurosci. 1993; 13: 3749-3772PubMed Google Scholar, 104Henry G.H. et al.Projections from areas 18 and 19 to cat striate cortex: divergence and laminar specificity.Eur. J. Neurosci. 1991; 3: 186-200Crossref PubMed Scopus (40) Google Scholar]. These patterns are summarised in Figure I by the violet tone of descending terminations to layer 5 and superficial layers in level (i-1), indicating a mix of superficial (red) and deep (blue) sources at level (i). The blueing of terminals in deeper layers of level (i-1), and all layers in level (i-2), indicates a progressive domination of deep layer sources from level (i). There is a consistent asymmetry between forward connections from sensory to motor areas and the reverse backward connections (e.g., between M1 and area 3a) [11Kunzle H. Cortico-cortical efferents of primary motor and somatosensory regions of the cerebral cortex in Macaca fascicularis.Neuroscience. 1978; 3: 25-39Crossref Scopus (90) Google Scholar, 12Stepniewska I. et al.Architectonics, somatotopic organization, and ipsilateral cortical connections of the primary motor area (M1) of owl monkeys.J. Comp. Neurol. 1993; 330: 238-271Crossref PubMed Scopus (271) Google Scholar, 13Leichnetz G.R. Afferent and efferent connections of the dorsolateral precentral gyrus (area 4, hand/arm region) in the macaque monkey, with comparisons to area 8.J. Comp. Neurol. 1986; 254: 460-492Crossref PubMed Scopus (203) Google Scholar], but the reciprocal connections among motor areas are of a distinct nature: there is a backward pattern of termination for projections from premotor areas to M1 [10Watanabe-Sawaguchi K. et al.Cytoarchitecture and intrafrontal connections of the frontal cortex of the brain of the hamadryas baboon (Papio hamadryas).J. Comp. Neurol. 1991; 311: 108-133Crossref PubMed Scopus (49) Google Scholar], yet the reverse connections (e.g., M1 to SMA, supplementary motor area) are columnar [11Kunzle H. Cortico-cortical efferents of primary motor and somatosensory regions of the cerebral cortex in Macaca fascicularis.Neuroscience. 1978; 3: 25-39Crossref Scopus (90) Google Scholar, 12Stepniewska I. et al.Architectonics, somatotopic organization, and ipsilateral cortical connections of the primary motor area (M1) of owl monkeys.J. Comp. Neurol. 1993; 330: 238-271Crossref PubMed Scopus (271) Google Scholar, 13Leichnetz G.R. Afferent and efferent connections of the dorsolateral precentral gyrus (area 4, hand/arm region) in the macaque monkey, with comparisons to area 8.J. Comp. Neurol. 1986; 254: 460-492Crossref PubMed Scopus (203) Google Scholar], of the sort normally associated with lateral connections. Hence, the premotor areas may top the hierarchy, as previously suggested [5Felleman D.J. Van Essen D.C. Distributed hierarchical processing in the primate cerebral cortex.Cereb. Cortex. 1991; 1: 1-47Crossref PubMed Scopus (5507) Google Scholar, 14Shipp S. The importance of being agranular: a comparative account of visual and motor cortex.Philos. Trans. R. Soc. Lond. B: Biol. Sci. 2005; 360: 797-814Crossref PubMed Scopus (98) Google Scholar], but there is little evidence for a classical ascending pathway through motor areas [6Adams R.A. et al.Predictions not commands: active inference in the motor system.Brain Struct. Funct. 2013; 218: 611-643Crossref PubMed Scopus (414) Google Scholar, 14Shipp S. The importance of being agranular: a comparative account of visual and motor cortex.Philos. Trans. R. Soc. Lond. B: Biol. Sci. 2005; 360: 797-814Crossref PubMed Scopus (98) Google Scholar]. In this review, we attempt to reconcile the laminar architecture and connectivity in both visual and sensorimotor hierarchies within a popular theoretical framework for describing cortical operations [15Friston K. The free-energy principle: a rough guide to the brain?.Trends Cogn. Sci. 2009; 13: 293-301Abstract Full Text Full Text PDF PubMed Scopus (1008) Google Scholar]. A percept can be regarded as a hypothesis that explains sensory input [16Gregory R.L. Perceptions as hypotheses.Philos. Trans. R. Soc. Lond. B: Biol. Sci. 1980; 290: 181-197Crossref PubMed Scopus (477) Google Scholar, 17Friston K. et al.Perceptions as hypotheses: saccades as experiments.Front. Psychol. 2011; 3: 151Google Scholar] – on occasion, an erroneous hypothesis, as demonstrated by classic illusions (Figure 1A) . The percept interprets sensory data, such that what we see is the inferred cause of the sensations, not merely an image of the data per se [18Helmholtz, H., translated by Southall, J.P.C., ed. (1860/1962) Handbuch der Physiologischen Optik (Vol. 3), DoverGoogle Scholar]. In Figure 1A, the facial features have an ambiguous depth structure that is resolved by our past experience of convex faces. The ability to infer the cause of visual sensations (e.g., a face) rests on an internal, generative model of how objects generate sensory data [19Friston K. A theory of cortical responses.Philos. Trans. R. Soc. Lond. B: Biol. Sci. 2005; 360: 815-836Crossref PubMed Scopus (2556) Google Scholar, 20Kersten D. et al.Object perception as Bayesian inference.Annu. Rev. Psychol. 2004; 55: 271-304Crossref PubMed Scopus (843) Google Scholar]. Generative models are required to finesse the problem of sensory indeterminacy (e.g., ambiguity) that illusions aptly illustrate. A generative model also has a temporal aspect: velocity is not a property of an instantaneous scene or 'snapshot', but an attribute that integrates sensory evidence over time. Biological motion detection implies recognition of complex motion patterns, such as a reach and grasp movement, or a repetitive action, such as walking [21Blake R. Shiffrar M. Perception of human motion.Annu. Rev. Psychol. 2007; 58: 47-73Crossref PubMed Scopus (681) Google Scholar]. In other words, the generative model of the brain is more like a narrative or scenario, predicting sequences of events. The scenario enables predictions about what may happen next. If a head is turning, for instance, a frontal view of a face may soon be replaced by a profile [22Perrett D.I. et al.Seeing the future: natural image sequences produce 'anticipatory' neuronal activity and bias perceptual report.Q. J. Exp. Psychol. 2009; 62: 2081-2104Crossref PubMed Scopus (55) Google Scholar]. Generative models are necessarily hierarchical (in space and time). If the visual system operates as a generative model, the percept corresponding to a particular cause is not specified at only one level, but has multiple levels of description. Take face processing, for example: a high-level face area encodes view-invariant face identity, whereas lower levels are view specific but less identity specific [23Freiwald W.A. Tsao D.Y. Functional compartmentalization and viewpoint generalization within the macaque face-processing system.Science. 2010; 330: 845-851Crossref PubMed Scopus (441) Google Scholar]. Features such as hair, eye, and skin colour are also encoded elsewhere [24Conway B.R. Tsao D.Y. Color architecture in alert macaque cortex revealed by FMRI.Cereb. Cortex. 2006; 16: 1604-1613Crossref PubMed Scopus (71) Google Scholar]. In addition, because face cells are size and position invariant [25Rolls E.T. Neurophysiological mechanisms underlying face processing within and beyond the temporal cortical visual areas.Philos. Trans. R. Soc. Lond. B: Biol. Sci. 1992; 335: 11-20Crossref PubMed Scopus (292) Google Scholar], lower areas must represent the 'filled-in surface' and 'border ownership' attributes of a percept [26Pollen D.A. Fundamental requirements for primary visual perception.Cereb. Cortex. 2008; 18: 1991-1998Crossref PubMed Scopus (24) Google Scholar, 27Poort J. et al.The role of attention in figure-ground segregation in areas V1 and V4 of the visual cortex.Neuron. 2012; 75: 143-156Abstract Full Text Full Text PDF PubMed Scopus (155) Google Scholar, 28Qiu F.T. et al.Figure-ground mechanisms provide structure for selective attention.Nat. Neurosci. 2007; 10: 1492-1499Crossref PubMed Scopus (167) Google Scholar]. In short, the gestalt of a 'face' has multiple components. In modelling terms, the high-level face area provides the highest stamp of recognition, guiding and contextualising inference about physical attributes in lower-level areas. Here, we shall use the term 'expectations' to refer to the representations of causes encoded at each level. Predictive coding schemes (e.g., [29Rao R.P. Ballard D.H. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects.Nat. Neurosci. 1999; 2: 79-87Crossref PubMed Scopus (2864) Google Scholar]) describe the inversion of a generative model, in order to recognise causes from their sensory consequences. In global terms, the model generates predictions of sensory input from high-level representations of causes; more specifically, the expectations at any given level predict the expectations at the level below. The model is inverted using a 'guess it and try it' approach (Figure 1B): each level computes 'prediction errors' by subtracting top-down predictions from its current expectations. The requisite predictions are based on expectations from the level above and conveyed by top-down or backward connections. Bottom-up prediction error signals are then passed forwards to modify expectations in the level above. This iterative, reciprocal exchange of predictions and errors minimises prediction error at every level of the hierarchy and provides a plausible explanation for visual sensations, in terms of expectations at multiple levels. In generalised formulations of hierarchical predictive coding, there are three sorts of expectation: expected 'causes', 'states', and 'precisions' [15Friston K. The free-energy principle: a rough guide to the brain?.Trends Cogn. Sci. 2009; 13: 293-301Abstract Full Text Full Text PDF PubMed Scopus (1008) Google Scholar]. Causes are invariant aspects of the world that create regularities in sensory data, such as objects in the visual scene. Their correspondence to elements of the scene is concrete at lower levels (e.g., a colour), and increasingly abstract at higher levels of the hierarchy (e.g., a smile). Whereas causes model categorical aspects of the world, states model their dynamics; that is, the fluctuations caused by the interactions among causes (e.g., motion of an object) or between cause and context (e.g., a rotating object and its illumination). Finally, precision corresponds to the reliability (inverse amplitude of random fluctuations) of causes and states. Therefore, expected precision determines the relative confidence in descending predictions and ascending prediction error. The differential equations describing predictive coding are provided elsewhere [30Adams R.A. et al.The computational anatomy of psychosis.Front Psychiatry. 2013; 4: 47Crossref PubMed Scopus (482) Google Scholar], together with the theory relating predictive coding to Bayesian inference [15Friston K. The free-energy principle: a rough guide to the brain?.Trends Cogn. Sci. 2009; 13: 293-301Abstract Full Text Full Text PDF PubMed Scopus (1008) Google Scholar, 31Friston K. The free-energy principle: a unified brain theory?.Nat. Rev. Neurosci. 2010; 11: 127-138Crossref PubMed Scopus (3544) Google Scholar]. Here, we consider the computational architecture and its implementation by neuronal circuitry. Figure 2 shows five kinds of computational unit (cf. neuronal ensembles): expectation and error units for causes and states, and units signalling expected precision. To recap, expectation units encode the expected causes and states describing events (scenarios) in the environment, whereas error units report inconsistencies between expectations at different levels or, at the sensory level, the mismatch between predictions and sensory input. Units encoding expected precision modulate the gain of error units and endow them with greater or lesser weight. This cortical gain control balances the influence of prediction errors at different levels in the hierarchy. Accordingly, precision is associated with the top-down deployment of attention [32Feldman H. Friston K.J. Attention, uncertainty, and free-energy.Front. Hum. Neurosci. 2010; 4: 215Crossref PubMed Scopus (769) Google Scholar] in the sensory domain and action selection in the context of affordance. In summary, expectation and error units interact to update beliefs about causes and states in the world, with one crucial distinction: expected causes are updated by reciprocal exchanges between hierarchical levels, whereas expected states are updated within each level. Below, we suggest a neural implementation of the predictive coding model outlined above (noting that alternative formulations could specify a different neural architecture [33Spratling M.W. Reconciling predictive coding and biased competition models of cortical function.Front. Comput. Neurosci. 2008; 2: 4Crossref PubMed Scopus (95) Google Scholar]). We prefabricate the scheme in visual cortex, as a model of hierarchical processing, before transcribing it to motor cortex and illustrating its explanatory scope through the example of mirror neurons. We now attempt to marry the computational anatomy of predictive coding with cortical microcircuitry. For simplicity, we focus on updating expected causes: our aim is not to specify exactly how such computations are performed at the synaptic level, but to indicate how they might map onto the laminar architecture of extrinsic and intrinsic cortical connections. The scheme shown in Figure 3 is inferred from anatomy alone; there is no explicit physiological categorisation of the notional expectation, error and precision units, but we make the provisional assumption that all three are represented in some form by pyramidal cells (or by excitatory, spiny stellate cells in layer 4). Extrinsic and intrinsic axonal ramifications typically contact inhibitory interneurons as well as pyramidal neurons [34Isaacson J.S. Scanziani M. How inhibition shapes cortical activity.Neuron. 2011; 72: 231-243Abstract Full Text Full Text PDF PubMed Scopus (999) Google Scholar], but the former are largely exclude
Referência(s)