the cybernetic bayesian brain - connecting … · the cybernetic bayesian brain ... ult of the...

The Cybernetic Bayesian BrainFrom Interoceptive Inference to Sensorimotor Contingencies

Anil K. Seth

Is there a single principle by which neural operations can account for perception,cognition, action, and even consciousness? A strong candidate is now takingshape in the form of “predictive processing”. On this theory, brains engage in pre-dictive inference on the causes of sensory inputs by continuous minimization ofprediction errors or informational “free energy”. Predictive processing can account,supposedly, not only for perception, but also for action and for the essential con-tribution of the body and environment in structuring sensorimotor interactions. Inthis paper I draw together some recent developments within predictive processingthat involve predictive modelling of internal physiological states (interoceptive in-ference), and integration with “enactive” and “embodied” approaches to cognitivescience (predictive perception of sensorimotor contingencies). The upshot is a de-velopment of predictive processing that originates, not in Helmholtzian percep-tion-as-inference, but rather in 20th-century cybernetic principles that emphasizedhomeostasis and predictive control. This way of thinking leads to (i) a new view ofemotion as active interoceptive inference; (ii) a common predictive framework link-ing experiences of body ownership, emotion, and exteroceptive perception; (iii)distinct interpretations of active inference as involving disruptive and disambigu-atory—not just confirmatory—actions to test perceptual hypotheses; (iv) a neuro-cognitive operationalization of the “mastery of sensorimotor contingencies” (wheresensorimotor contingencies reflect the rules governing sensory changes producedby various actions); and (v) an account of the sense of subjective reality of percep-tual contents (“perceptual presence”) in terms of the extent to which predictivemodels encode potential sensorimotor relations (this being “counterfactual rich-ness”). This is rich and varied territory, and surveying its landmarks emphasizesthe need for experimental tests of its key contributions.

KeywordsActive inference | Counterfactually-equipped predictive model | Evolutionary ro-botics | Free energy principle | Interoception | Perceptual presence | Predictiveprocessing | Sensorimotor contingencies | Somatic marker hypothesis | Synaes-thesia

Author

Anil K. [email protected] University of SussexBrighton, United Kingdom

Commentator

Wanja [email protected] Gutenberg-UniversitätMainz, Germany

Editors

Thomas [email protected] Johannes Gutenberg-UniversitätMainz, Germany

Jennifer M. [email protected] Monash UniversityMelbourne, Australia

1 Introduction

An increasingly popular theory in cognitive sci-ence claims that brains are essentially predic-tion machines (Hohwy 2013). The theory isvariously known as the Bayesian brain (Knill &Pouget 2004; Pouget et al. 2013), predictiveprocessing (Clark 2013; Clark this collection),and the predictive mind (Hohwy 2013; Hohwythis collection), among others; here we use theterm PP (predictive processing). (See Table 1for a glossary of technical terms.) At its mostfundamental, PP says that perception is the res-

ult of the brain inferring the most likely causesof its sensory inputs by minimizing the differ-ence between actual sensory signals and the sig-nals expected on the basis of continuously up-dated predictive models. Arguably, PP providesthe most complete framework to date for ex-plaining perception, cognition, and action interms of fundamental theoretical principles andneurocognitive architectures. In this paper I de-scribe a version of PP that is distinguished by(i) an emphasis on predictive modelling of in-

Seth, A. K. (2015). The Cybernetic Bayesian Brain - From Interoceptive Inference to Sensorimotor Contingencies.In T. Metzinger & J. M. Windt (Eds). Open MIND: 35(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570108 1 | 24

http://www.open-mind.net/showAuthor?author=AnilK_Seth

http://www.open-mind.net/showAuthor?author=AnilK_Seth

http://www.open-mind.net/showAuthor?author=JenniferM_Windt

http://www.open-mind.net/showAuthor?author=Thomas_Metzinger

http://www.open-mind.net/showAuthor?author=Wanja_Wiese

http://dx.doi.org/10.15502/9783958570108

http://www.open-mind.net/papers/@@chapters?nr=35

www.open-mind.net


Table 1: A glossary of technical terms.

http://dx.doi.org/10.15502/9783958570108


http://www.open-mind.net/

www.open-mind.net

ternal physiological states and (ii) engagementwith alternative frameworks under the bannerof “enactive” and “embodied” cognitive science(Varela et al. 1993).

I first identify an unusual starting pointfor PP, not in Helmholtzian perception-as-infer-ence, but in the mid 20th-century cybernetictheories associated with W. Ross Ashby (1952,1956; Conant & Ashby 1970). Linking these ori-gins to their modern expression in Karl Fris-ton’s “free energy principle” (2010), perceptionemerges as a consequence of a more funda-mental imperative towards homeostasis andcontrol, and not as a process designed to furnisha detailed inner “world model” suitable for cog-nition and action planning. The ensuing view ofPP, while still fluently accounting for (extero-ceptive) perception, turns out to be more natur-ally applicable to the predictive perception ofinternal bodily states, instantiating a process ofinteroceptive inference (Seth 2013; Seth et al.2011). This concept provides a natural way ofthinking of the neural substrates of emotionaland mood experiences, and also describes acommon mechanism by which interoceptive andexteroceptive signals can be integrated toprovide a unified experience of body ownershipand conscious selfhood (Blanke & Metzinger2009; Limanowski & Blankenburg 2013).

The focus on embodiment leads to distinctinterpretations of active inference, which in gen-eral refers to the selective sampling of sensorysignals so as to improve perceptual predictions.The simplest interpretation of active inference isthe changing of sensory data (via selectivesampling) to conform to current predictions(Friston et al. 2010). However, by analogy withhypothesis testing in science, active inferencecan also involve seeking evidence that goesagainst current predictions, or that disambigu-ates multiple competing hypotheses. A nice ex-ample of the latter comes from self-modelling inevolutionary robotics, where multiple competingself-models are used to specify actions that aremost likely to provide disambiguatory sensoryevidence (Bongard et al. 2006). I will spendmore time on this example later. Crucially,these different senses of active inference rest onthe capacity of predictive models to encode

counterfactual relations linking potential (butnot necessarily executed) actions to their expec-ted sensory consequences (Friston et al. 2012;Seth 2014b). It also implies the involvement ofmodel comparison and selection—not just theoptimization of parameters assuming a singlemodel. These points represent significant devel-opments in the basic infrastructure of PP.

The notion of counterfactual predictionsconnects PP with what at first glance seemsto be its natural opponent: “enactive” theoriesof perception and cognition that explicitly re-ject internal models or representations (Clarkthis collection; Hutto & Myin 2013; Thompson& Varela 2001). Central to the enactive ap-proach are notions of “sensorimotor contingen-cies” and their “mastery” (O’Regan & Noë2001), where a sensorimotor contingency refersto a rule governing how sensory signals changein response to action. On this approach, theperceptual experience of (for example) rednessis given by an implicit knowledge (mastery) ofthe way red things behave given certain pat-terns of sensorimotor activity. This mastery ofsensorimotor contingencies is also said to un-derpin perceptual presence: the sense of sub-jective reality of the contents of perception(Noë 2006). From the perspective of PP, mas-tery of a sensorimotor contingency corres-ponds to the learning of a counterfactually-equipped predictive model connecting poten-tial actions to expected sensory consequences.The resulting theory of PPSMC (PredictivePerception of SensoriMotor Contingencies),Seth 2014b) provides a much needed reconcili-ation of enactive and predictive theories ofperception and action. It also provides a solu-tion to the challenge of perceptual presencewithin the setting of PP: perceptual presenceobtains when the underlying predictive modelsare counterfactually rich, in the sense of en-coding a rich repertoire of potential (but notnecessarily executed) sensorimotor relations.This approach also helps explain instanceswhere perceptual presence seems to be lack-ing, such as in synaesthesia.

This is both a conceptual and theoreticalpaper. Space limitations preclude any signific-ant treatment of the relevant experimental lit-



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

erature. However, even an exhaustive treat-ment would reveal that this literature so farprovides only circumstantial support for thebasics of PP, let alone for the extensions de-scribed here. Yet an advantage of PP theoriesis that they are grounded in concrete compu-tational processes and neurocognitive architec-tures, giving us confidence that informativeexperimental tests can be devised. Implement-ing such an experimental agenda stands as acritical challenge for the future.

2 The predictive brain and its cybernetic origins

2.1 Predictive processing: The basics

PP starts with the assumption that in order tosupport adaptive responses, the brain must dis-cover information about the external “hidden”causes of sensory signals. It lacks any direct ac-cess to these causes, and can only use informa-tion found in the flux of sensory signals them-


Figure 1: A. Schemas of hierarchical predictive coding across three cortical regions; the lowest on the left (R1) andthe highest on the right (R3). Bottom-up projections (red) originate from “error units” (orange) in superficial corticallayers and terminate on “state units” (light blue) in the deep (infragranular) layers of their targets; while top-down pro -jections (dark blue) convey predictions originating in deep layers and project to the superficial layers of their targets.Prediction errors are associated with precisions, which determine the relative influence of bottom-up and top-down sig-nal flow via precision weighting (dashed lines). B. The influence of precisions on Bayesian inference and predictive cod-ing. The curves show probability distributions over the value of a sensory signal (x-axis). On the left, high precision-weighting of sensory signals (red) enhances their influence on the posterior (green) and expectation (dotted line) ascompared to the prior (blue). On the right, low sensory precision weighting has the opposite effect. Figure adaptedfrom Seth (2013).


http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

selves. According to PP, brains meet this chal-lenge by attempting to predict sensory inputson the basis of their own emerging models ofthe causes of these inputs, with prediction er-rors being used to update these models so as tominimize discrepancies. The idea is that a brainoperating this way will come to encode (in theform of predictive or generative models) a richbody of information about the sources of signalsby which it is regularly perturbed (Clark 2013).

Applied to cortical hierarchies, PP over-turns classical notions of perception that de-scribe a largely “bottom-up” process of evidenceaccumulation or feature detection. Instead, PPproposes that perceptual content is determinedby top-down predictive signals emerging frommulti-layered and hierarchically-organized gen-erative models of the causes of sensory signals(Lee & Mumford 2003). These models are con-tinually refined by mismatches (prediction er-rors) between predicted signals and actual sig-nals across hierarchical levels, which iterativelyupdate predictive models via approximations toBayesian inference (see Figure 1). This meansthat the brain can induce accurate generativemodels of environmental hidden causes by oper-ating only on signals to which it has direct ac-cess: predictions and prediction errors. It alsomeans that even low-level perceptual content isdetermined via cascades of predictions flowingfrom very general abstract expectations, whichconstrain successively more fine-grained predic-tions.

Two further aspects of PP need to be em-phasized from the outset. First, sensory predic-tion errors can be minimized either “passively”,by changing predictive models to fit incomingdata (perceptual inference), or “actively”, byperforming actions to confirm or test sensorypredictions (active inference). In most casesthese processes are assumed to unfold continu-ously and simultaneously, underlining a deepcontinuity between perception and action (Fris-ton et al. 2010; Verschure et al. 2003). This pro-cess of active inference will play a key role inmuch of what follows. Second, predictions andprediction errors in a Bayesian framework haveassociated precisions (inverse variances, Figure1). The precision of a prediction error is an in-

dicator of its reliability, and hence can be usedto determine its influence in updating top-downpredictive models. Precisions, like mean values,are not given but must be inferred on the basisof top-down models and incoming data; so PPrequires that agents have expectations aboutprecisions that are themselves updated as newdata arrive (and new precisions can be estim-ated). Precision expectations can therefore bal-ance the influence of different prediction-errorsources on the updating of predictive models.And if prediction errors have low (expected)precision, predictive models may overwhelm er-ror signals (hallucination) or elicit actions thatconfirm sensory predictions (active inference).

A picture emerges in which cortical net-works engage in recurrent interactions wherebybottom-up prediction errors are continuously re-conciled with top-down predictions at multiplehierarchical levels—a process modulated at alltimes by precision weighting. The result is abrain that not only encodes information aboutthe sources of signals that impinge upon itssensory surfaces, but that also encodes informa-tion about how its own actions interact withthese sources in specifying sensory signals. Per-ception involves updating the parameters of themodel to fit the data; action involves changingsensory data to fit (or test) the model; and at-tention corresponds to optimizing model updat-ing by giving preference to sensory data thatare expected to carry more information, whichis called precision weighting (Hohwy 2013). Thisview of the brain is shamelessly model-basedand representational (though with a finessed no-tion of representation), yet it also deeply em-beds the close coupling of perception and actionand, as we will see, the importance of the bodyin the mediation of this interaction.

2.2 Predictive processing and the free energy principle

PP can be considered a special case of the freeenergy principle, according to which perceptualinference and action emerge as a consequence ofa more fundamental imperative towards theavoidance of “surprising” events (Friston 2005,2009, 2010). On the free energy principle, or-



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

ganisms – by dint of their continued survival—must minimize the long-run average surprise ofsensory states, since surprising sensory statesare likely to reflect conditions incompatible withcontinued existence (think of a fish out of wa-ter). “Surprise” is not used here in the psycholo-gical sense, but in an information-theoreticsense—as the negative log probability of anevent’s occurrence (roughly, the unlikeliness ofthe occurrence of an event).

The connection with PP arises becauseagents cannot directly evaluate the (informa-tion-theoretic) surprise associated with anevent, since this would require—impossibly—the agent to average over all possible occur-rences of the event in all possible situations. In-stead, the agent can only maintain a lower limiton surprise by minimizing the differencebetween actual sensory signals and those signalspredicted according to a generative or predictivemodel. This difference is free energy, which, un-der fairly general assumptions, is the long-runsum of prediction error.

An attractive feature of the free energyprinciple is that it brings to the table a richmathematical framework that shows how PP canwork in practice. Formally, PP depends on estab-lished principles of Bayesian inference and modelspecification, whereby the most likely causes ofobserved data (posterior) are estimated based onoptimally combining prior expectations of thesecauses with observed data, by using a (generative,predictive) model of the data that would be ob-served given a particular set of causes (likelihood).(See Figure 1 for an example of priors and pos-teriors.) In practice, because optimal Bayesian in-ference is usually intractable, a variety of approx-imate methods can be applied (Hinton & Dayan1996; Neal & Hinton 1998). Friston’s frameworkappeals to previously worked-out “variational”methods, which take advantage of certain approx-imations (e.g., Gaussianity, independence of tem-poral scales)—thus allowing a potentially neatmapping onto neurobiological quantities (Fristonet al. 2006).1 1 Some challenging questions surface here as to whether prediction

errors are used to update priors, which corresponds to standardBayesian inference, or whether they are used to update the un-derlying generative/predictive model, which corresponds to learn-ing.

The free energy principle also emphasizesaction as a means of prediction error minimiza-tion, this being active inference. In general, act-ive inference involves the selective sampling ofsensory signals so as to minimize uncertainty inperceptual hypotheses (minimizing the entropyof the posterior). In one sense this means thatactions are selected to provide evidence compat-ible with current perceptual predictions. This isthe most standard interpretation of the concept,since it corresponds most directly to minimiza-tion of prediction error (Friston 2009). However,as we will see, actions can also be selected onthe basis of an attempt to find evidence goingagainst current hypotheses, and/or to efficientlydisambiguate between competing hypotheses.These finessed senses of active inference repres-ent developments of the free energy framework.Importantly, action itself can be thought of asbeing brought about by the minimization ofproprioceptive prediction errors via the engage-ment of classical reflex arcs (Adams et al. 2013;Friston et al. 2010). This requires transientlylow precision-weighting of these errors (or elsepredictions would simply be updated instead),which is compatible with evidence showing sens-ory attenuation during self-generated move-ments (Brown et al. 2013).

A more controversial aspect of the free en-ergy principle is its claimed generality (Hohwythis collection). At least as described by Friston,it claims to account for adaptation at almostany granularity of time and space, from macro-scopic trends in evolution, through developmentand maturation, to signalling in neuronal hier-archies (Friston 2010). However, in some ofthese interpretations reliance on predictive mod-elling is only implicit; for example the body of afish can be considered to be an implicit modelof the fluid dynamics and other affordances ofits watery environment (see section 2.3). I amnot concerned here with these broader inter-pretations, but will focus on those cases inwhich biological (neural) mechanisms plausiblyimplement explicit predictive inference via ap-proximations to Bayesian computations—namely, the Bayesian brain (Knill & Pouget2004; Pouget et al. 2013). Here, the free energyprinciple has potentially the greatest explanat-



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

ory power, especially given the convergence ofempirical evidence (see Clark 2013 and Hohwy2013 for reviews) and computational modellingshowing how cortical microcircuits might imple-ment approximate Bayesian inference (Bastos etal. 2012).

2.3 Predictive processing, free energy, and cybernetics

Typically, the origins of PP are traced to thework of the 19th Century physiologist Hermannvon Helmholtz, who first formalized the idea ofperception as inference. However, the Helmholt-

zian view is rather passive, inasmuch as there islittle discussion of active inference or behaviour.The close coupling of perception and action em-phasized in the free energy principle points in-stead to a deep connection between PP andmid-twentieth-century cybernetics. This is mostobvious in the works of W. Ross Ashby (Ashby1952; 1956; Conant & Ashby 1970) but is alsoevident more generally (Dupuy 2009; Pickering2010). Importantly, cybernetics adopted as itscentral focus the prediction and control of beha-viour in so-called teleological or purposeful ma-chines.2 More precisely, cybernetic theoristswere (are) interested in systems that appear tohave goals (i.e., teleological) and that particip-ate in circular causal chains (i.e., involving feed-back) coupling goal-directed sensation and ac-tion.

Two key insights from the first wave of cy-bernetics usefully anticipate the core develop-ments of PP within cognitive science. These areboth associated with Ashby, a key figure in themovement and often considered its leader, atleast outside the USA (Figure 2).

The first insight consists in an emphasison the homeostasis of internal essential vari-ables, which, in physiological settings, corres-pond to quantities like blood pressure, heartrate, blood sugar levels, and the like. In Ashby’sframework, when essential variables move bey-ond specific viability limits, adaptive processesare triggered that re-parameterize the systemuntil it reaches a new equilibrium in whichhomeostasis is restored (Ashby 1952). Such sys-tems are, in Ashby’s terminology, ultrastable,since they embody (at least) two levels of feed-back: a first-order feedback that homeostaticallyregulates essential variables (like a thermostat)and a second-order feedback that allostatically3

re-organises a system’s input–output relationswhen first-order feedback fails, until a newhomeostatic regime is attained. In the most ba-sic case, as implemented in Ashby’s famous“homeostat” (Figure 2), this second-order feed-back simply involves random changes to system2 This underlines the close links between cybernetics and behaviour-

ism. Perhaps this explains why cybernetics was so reluctant to bringphenomenology into its remit, an exclusion which, looking back,seems like a missed opportunity.

3 Allostasis: the process of achieving homeostasis.


Figure 2: A. W. Ross Ashby, British psychiatrist andpioneer of cybernetics (1903–1972). B. A schematic of ul-trastability, based on Ashby’s notebooks. The system Rhomeostatically maintains its essential variables (EVs)within viability limits via first-order feedback with theenvironment E. When first-order feedback fails, so thatEVs run out-of-bounds, second order “ultrastable” feed-back is triggered so that S (an internal controller, poten-tially model-based) changes the parameters of R govern-ing the first-order feedback. S continually changes R untilhomeostatic relations are regained, leaving the EVs againwithin bounds. C. Ashby’s “homeostat”, consisting offour interconnected ultrastable systems, forming a so-called “multistable” system. D. One ultrastable unit fromthe homeostat. Each unit had a trough of water with anelectric field gradient and a metal needle. Instability wasrepresented by the non-central needle positions, which onoccurring would alter the resistances connecting the unitsvia discharge through capacitors. For more details seeAshby (1952) and Pickering (2010).


http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

parameters until a new stable regime is reached.The importance of this insight for PP is that itlocates the function of biological and cognitiveprocesses in generalizing homeostasis to ensurethat internal essential variables remain withinexpected ranges.

Another way to summarize the funda-mental cybernetic principle is to say that adapt-ive systems ensure their continued existence bysuccessfully responding to environmental per-turbations so as to maintain their internal or-ganization. This leads to the second insight,evident in Ashby’s law of requisite variety. Thisstates that a successful control system must becapable of entering at least as many states asthe system being controlled: “only variety canforce down variety” (Ashby 1956). This inducesa functional boundary between controller andenvironment and implies a minimum level ofcomplexity for a successful controller, which isdetermined by the causal complexity of the en-vironmental states that constitute potential per-turbations to a system’s essential variables. Thisview was refined some years later, in a 1970 pa-per written with Roger Conant entitled “Everygood regulator of a system must be a model ofthat system” (Conant & Ashby 1970). This pa-per builds on the law of requisite variety by ar-guing (and attempting to formally show) thatthe nature of a controller capable of suppressingperturbations imposed by an external system(e.g., the world) must instantiate a model ofthat system. This provides a clear connectionwith the free energy principle, which proposesthat adaptive systems minimize a limit on freeenergy (long-run average surprise) by inducingand refining a generative model of the causes ofsensory signals. It also moves beyond Ashby’shomeostat by implying that model-based con-trollers can engage in more successful multi-level feedback than is possible by random vari-ation of higher-order parameters.

Putting these insights together providesa distinctive way of seeing the relevance of PPto cognition and biological adaptation. It canbe summarized as follows. The purpose of cog-nition (including perception and action) is tomaintain the homeostasis of essential variablesand of internal organization (ultrastability).

This implies the existence of a control mech-anism with sufficient complexity to respond to(i.e., suppress) the variety of perturbations itencounters (law of requisite variety). Further,this structure must instantiate a model of thesystem to be controlled (good regulator the-orem), where the system includes both thebody and the environment (and their interac-tions). As Ashby himself tells us “[t]he wholefunction of the brain can be summed up in:error correction” (quoted in Clark 2013, p. 1).Put this way, perception emerges as a con-sequence of a more fundamental imperative to-wards organizational homeostasis, and not asa stage in some process of internal world-model construction. This view, while high-lighting different origins, closely parallels theassumptions of the free energy principle inproposing a primary imperative towards thecontinued survival of the organism (Friston2010).

It may be surprising to consider the leg-acy of cybernetics in this light. This is be-cause many previous discussions of this legacyfocus on examples which show that complex,apparently goal-directed behaviour can emergefrom simple mechanisms interacting withstructured bodies and environments (Beer2003; Braitenberg 1984). On this more stand-ard development, cybernetics challenges ratherthan asserts the need for internal models andrepresentations: it is often taken to justify slo-gans of the sort “the world is its own bestmodel” (Brooks 1991). In fact, cybernetics isagnostic with respect to the need for deploy-ment of explicit internally-specified predictivemodels. If environmental circumstances arereasonably stable, and mappings between per-turbations and (homeostatic) responses reas-onably straightforward, then the good regu-lator theorem can be satisfied by controllersthat only implicitly model their environments.This is the case, for instance, in the Watt gov-ernor: a device that is able exquisitely to con-trol the output of (for instance) a steam en-gine, in virtue of its mechanism, and notthrough the deployment of explicit predictivemodels or representations (see Figure 3 andVan Gelder 1995; note that the governor can



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

be described as an implicit model since it hasvariables – e.g., eccentricity of the metal ballsfrom the central column – which map onto en-vironmental variables that affect the homeo-static target – engine output). However, wherethere exist many-to-many mappings betweensensory states and their probable causes, asmay be the case more often than not, it willpay to engage explicit inferential processes inorder to extract the most probable causes ofsensory states, insofar as these causes threatenthe homeostasis of essential variables.

Figure 3: The Watt governor. This system, a centralcontributor to the industrial revolution, enabled precisecontrol over the output of (for example) steam engines.As the speed of the engine increases, power is supplied tothe governor (A) by a belt or chain, causing it to rotatemore rapidly so that the metal balls have more kineticenergy. This causes the balls to rise (B), which closes thethrottle valve (C), thereby reducing the steam flow,which in turn reduces engine speed (D). The oppositehappens when the engine speed decreases, so that thegovernor maintains engine speed at a precise equilibrium.

In summary, rather than seeing PP asoriginating solely in the Helmholtzian notionof “perception as inference”, it is fruitful tosee it also as part of a process of model-basedpredictive control entailed by a fundamentalimperative towards internal homeostasis. Thisshift in perspective reveals a distinctiveagenda for PP in cognitive science, to which Ishall now turn.

3 Interoceptive inference, emotion, and predictive selfhood

3.1 Interoceptive inference and emotion

Considering the cybernetic roots of PP, togetherwith the free energy principle, leads to a poten-tially counterintuitive idea. This is that PP mayapply more naturally to interoception (the senseof the internal physiological condition of thebody) than to exteroception (the classic senses,which carry signals that originate in the ex-ternal environment). This is because for an or-ganism it is more important to avoid encounter-ing unexpected interoceptive states than toavoid encountering unexpected exteroceptivestates. A level of blood oxygenation or bloodsugar that is unexpected is likely to be badnews for an organism, whereas unexpected ex-teroceptive sensations (like novel visual inputs)are less likely to be harmful and may in somecases be desirable, as organisms navigate a del-icate balance between exploration and exploita-tion (Seth 2014a), testing current perceptualhypotheses through active inference (see section5, below), all ultimately in the service of main-taining organismic homeostasis.

Perhaps because of its roots in Helm-holtz, PP has largely been developed in thesetting of visual neuroscience (Rao & Ballard1999), with a related but somewhat independ-ent line in motor control (Wolpert &Ghahramani 2000). Recently, an explicit ap-plication of PP to interoception has been de-veloped (Seth 2013; Seth & Critchley 2013;Seth et al. 2011; see also Gu et al. 2013). Onthis theory of interoceptive inference (or equi-valently interoceptive predictive coding), emo-tional states (i.e., subjective feeling states)arise from top-down predictive inference of thecauses of interoceptive sensory signals (seeFigure 4). In direct analogy to exteroceptivePP, emotional content is constitutively spe-cified by the content of top-down interoceptivepredictions at a given time, marking a distinc-tion with the well-studied impact of expecta-tions on subsequent emotional states (see e.g.,Ploghaus et al. 1999; Ueda et al. 2003). Fur-thermore, interoceptive prediction errors can



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

be minimized by (i) updating predictive mod-els (perception, corresponding to new emo-tional contents); (ii) changing interoceptivesignals through engaging autonomic reflexes(autonomic control or active inference); or (iii)performing behaviour so as to alter externalconditions that impact on internal homeo-stasis (allostasis; Gu & Fitzgerald 2014; Sethet al. 2011).

Consider an example in which bloodsugar levels (an essential variable) fall towardsor beyond viability thresholds, reaching unex-pected and undesirable values (Gu & Fitzger-ald 2014; Seth et al. 2011). Under interocept-ive inference, the following responses ensue.First, interoceptive prediction error signalsupdate top-down expectations, leading to sub-

jective experiences of hunger or thirst (forsugary things). Because these feeling statesare themselves surprising (and non-viable) inthe long run, they signal prediction errors athierarchically-higher levels, where predictivemodels integrate multimodal interoceptive andexteroceptive signals. These models instanti-ate predictions of temporal sequences ofmatched exteroceptive and interoceptive in-puts, which flow down through the hierarchy.The resulting cascade of prediction errors canthen be resolved either through autonomiccontrol, in order to metabolize bodily fatstores (active inference), or through allostaticactions involving the external environment(i.e., finding and eating sugary things).

The sequencing and balance of theseevents is governed by relative precisions andtheir expectations. Initially, interoceptive pre-diction errors have high precision (weighting)given a higher-level expectation of stablehomeostasis. Whether the resulting high-levelprediction error engages autonomic control orallostatic behaviour (or both) depends on theprecision weighting of the corresponding predic-tion errors. If food is readily available, consum-matory actions lead to food intake (as describedearlier, these actions are generated by the resol-ution of proprioceptive prediction errors). Ifnot, autonomic reflexes initiate the metaboliza-tion of bodily fat stores, perhaps alongside ap-petitive behaviours that are predicted to lead tothe availability of food, conditioned on perform-ing these behaviours.4

3.2 Implications of interoceptive inference

Several interesting implications arise when con-sidering emotion as resulting from interoceptiveinference (Seth 2013). First, the theory general-izes previous “two factor” theories of emotionthat see emotional content as resulting from aninteraction between the perception of physiolo-4 It is interesting to consider possible dysfunctions in this process.

For example, if high-level predictions about the persistence of lowblood sugar become abnormally strong (i.e., low blood sugar be-comes chronically expected), allostatic food-seeking behavioursmay not occur. This process, akin to the transition from hallucin -ation to delusion in perceptual inference (Fletcher & Frith 2009),may help understand eating disorders in terms of dysfunctionalsignalling of satiety.


Figure 4: Inference and perception. Green arrows representexteroceptive predictions and predictions errors underpin-ning perceptual content, such as the visual experience of atomato. Orange arrows represent proprioceptive predictions(and prediction errors) underlying action and the experienceof body ownership. Blue arrows represent interoceptive pre-dictions (and prediction errors) underlying emotion, mood,and autonomic regulation. Hierarchically higher levels willdeploy multimodal and even amodal predictive models span-ning these domains, which are capable of generating mul-timodal predictions of afferent signals.


http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

gical changes (James 1894) and “higher-level”cognitive appraisal of the context within whichthese changes occur (Schachter & Singer 1962).Instead of distinguishing “physiological” and“cognitive” levels of description, interoceptiveinference sees emotional content as resultingfrom the multi-layered prediction of interocept-ive input spanning many levels of abstraction.Thus, interoceptive inference integrates cogni-tion and emotion within the powerful setting ofPP.

The theory also connects with influentialframeworks that link interoception with decisionmaking, notably the “somatic marker hypo-thesis” proposed by Antonio Damasio (1994).According to the somatic marker hypothesis, in-tuitive decisions are shaped by interoceptive re-sponses (somatic markers) to potential out-comes. This idea, when placed in the context ofinteroceptive inference, corresponds to the guid-ance of behavioural (allostatic) responses to-wards the resolution of interoceptive predictionerror (Gu & Fitzgerald 2014; Seth 2014a). Itfollows that intuitive decisions should be af-fected by the degree to which an individualmaintains accurate predictive models of his orher own interoceptive states; see Dunn et al.2010, Sokol-Hessner et al. 2014 for evidencealong these lines.

There are also important implications fordisorders of emotion, selfhood, and decision-making. For example, anxiety may result fromthe chronic persistence of interoceptive predic-tion errors that resist top-down suppression(Paulus & Stein 2006). Dissociative disorderslike alexithymia (the inability to describeone’s own emotions), and depersonalizationand derealisation (the loss of sense of realityof the self and world) may also result fromdysfunctional interoceptive inference, perhapsmanifest in abnormally low interoceptive pre-cision expectations (Seth 2013; Seth et al.2011). In terms of decision-making, it may beproductive to think of addiction as resultingfrom dysfunctional active inference, wherebystrong interoceptive priors are confirmedthrough action, overriding higher-order or hy-per-priors relating to homeostasis and organis-mic integrity. It has even been suggested that

autism spectrum disorders may originate inaberrant encoding of the salience or precisionof interoceptive prediction errors (Quattrocki& Friston 2014). The reasoning here is thataberrant salience during development coulddisrupt the assimilation of interoceptive andexteroceptive cues within generative models ofthe “self”, which would impair a child’s abilityto properly assign salience to socially relevantsignals.

3.3 The predictive embodied self

The maintenance of physiological homeostasissolely through direct autonomic regulation isobviously limited: behavioural (allostatic) inter-actions with the world are necessary if the or-ganism is to avoid surprising physiologicalstates in the long run. The ability to deploy ad-aptive behavioural responses mandates the ori-ginal Helmholtzian view of perception-as-infer-ence, which has been the primary setting for thedevelopment of PP so far. A critical but argu-ably overlooked middle ground, which mediatesbetween physiological state variables and theexternal environment, is the body. On one hand,the body is the material vehicle through whichbehaviour is expressed, permitting allostatic in-teractions to take place. On the other, the bodyis itself an essential part of the organismic sys-tem, the homeostatic integrity of which must bemaintained. In addition, the experience of own-ing and identifying with a particular body is akey component of being a conscious self (Apps& Tsakiris 2014; Blanke & Metzinger 2009;Craig 2009; Limanowski & Blankenburg 2013;Seth 2013).

It is tempting to ask whether commonpredictive mechanisms could underlie not onlyclassical exteroceptive perception (like vision)and interoception (see above), but also their in-tegration in supporting conscious and uncon-scious representations of the body and self (Seth2013). The significance of this question is un-derlined by realising that just as the brain hasno direct access to causal structures in the ex-ternal environment, it also lacks direct access toits own body. That is, given that the brain is inthe business of inferring the causal sources of



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

sensory signals, a key challenge emerges whendistinguishing those signals that pertain to thebody from those that originate from the ex-ternal environment. A clue to how this chal-lenge is met is that the physical body, unlikethe external environment, constantly generatesand receives internal input via its interoceptiveand proprioceptive systems (Limanowski &Blankenburg 2013; Metzinger 2003). This sug-gests that the experienced body (and self) de-pends on the brain’s best guess of the causes ofthose sensory signals most likely to be “me”(Apps & Tsakiris 2014), across interoceptive,proprioceptive, and exteroceptive domains (Fig-ure 4).

There is now considerable evidence thatthe experience of body ownership is highlyplastic and depends on the multisensory integ-ration of body-related signals (Apps &

Tsakiris 2014; Blanke & Metzinger 2009). Oneclassic example is the rubber hand illusion,where the stroking of an artificial hand syn-chronously with a participant’s real hand,while visual attention is focused on the artifi-cial hand, leads to the experience that the ar-tificial hand is somehow part of the body(Botvinick & Cohen 1998). According to cur-rent multisensory integration models, thischange in the experience of body ownership isdue to correlation between vision and touchoverriding conflicting proprioceptive inputs(Makin et al. 2008). Through the lens of PP,this implies that prediction errors induced bymultisensory conflicts will over time updateself-related priors (Apps & Tsakiris 2014),with different signal sources (vision, touch,proprioception) each precision-weighted ac-cording to their expected reliability, and all in


Figure 5: The interaction of interoceptive and exteroceptive signals in shaping the experience of body ownership. A. Set-upfor applying cardio-visual feedback in the rubber hand illusion. A Microsoft Kinect obtains a real-time 3D model of a sub-ject’s left hand. This is re-projected into the subject’s visual field using a head-mounted display and augmented reality (AR)software. B. The colour of the virtual hand is modulated by the subject’s heart-beat. C. A similar set-up for the full-body il-lusion whereby a visual image of a subject’s body is surrounded by a halo pulsing either in time or out of time with theheartbeat. Panels A and B are adapted from Suzuki et al. (2013); panel C is adapted from Aspell et al. (2013).


http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

the setting of strong prior expectations forcorrelated input.5

While the potential for exteroceptivemultisensory integration to modulate the exper-ience of body ownership has been extensivelyexplored both for the ownership of body partsand for the experience of ownership of the bodyas a whole (for reviews, see Apps & Tsakiris2014; Blanke & Metzinger 2009), only recentlyhas attention been paid to interactions betweeninteroceptive and exteroceptive signals. Initialevidence in this line of investigation was indir-ect, for example showing correlation betweensusceptibility to the rubber hand illusion andindividual differences in the ability to perceiveinteroceptive signals (“interoceptive sensitivity”,typically indexed by heartbeat detection tasks;Tsakiris et al. 2011). Other relevant studieshave shown that body ownership illusions leadto temperature reductions in the correspondingbody parts, perhaps reflecting altered activeautonomic inference (Moseley et al. 2008; Sa-lomon et al. 2013).

Emerging evidence now points more dir-ectly towards the predictive multisensory integ-ration of interoceptive and exteroceptive signalsin shaping the experience of body ownership.Two recent studies have taken advantage of so-called “cardio-visual synchrony” where virtual-reality representations of body parts (Suzuki etal. 2013) or the whole body (Aspell et al. 2013)are modulated by simultaneously recordedheartbeat signals, with the modulation eitherin-time or out-of-time with the actual heartbeat(Figure 5). These data suggest that statisticalcorrelations between interoceptive (e.g., cardiac)and exteroceptive (e.g., visual) signals can leadto the updating of predictive models of self-re-lated signals through (hierarchical) minimiza-tion of prediction error, just as happens forpurely exteroceptive multisensory conflicts inthe classic rubber hand illusion.

While these studies underline the plausib-ility of common predictive mechanisms underly-ing emotion, selfhood, and perception, manyopen questions nevertheless remain. A key chal-lenge is to detail the underlying neural opera-5 Interestingly the expectation of perceptual correlations seems to be

sufficient for inducing the rubber hand illusion (Ferri et al. 2013).

tions. Though a detailed analysis is beyond thescope of the present paper, it is worth notingthat attention is increasingly focused on the in-sular cortex (especially its anterior parts) as apotential source of interoceptive predictions,and also as a comparator registering interocept-ive prediction errors. The anterior insula haslong been considered a major cortical locus forthe integration of interoceptive and exterocept-ive signals (Craig 2003; Singer et al. 2009); it isstrongly implicated in interoceptive sensitivity(Critchley et al. 2004); it is sensitive to intero-ceptive prediction errors—at least in some con-texts (Paulus & Stein 2006); and it has a highdensity of so-called “von Economo” neurons,6which have been frequently though circumstan-tially associated with consciousness and self-hood (Critchley & Seth 2012; Evrard et al.2012).

3.4 Active inference, self-modeling, and evolutionary robotics

What role might active inference play in pre-dictive self-modelling? Autonomic changes dur-ing illusions of body ownership (see above) areconsistent with active inference; however theydo not speak directly to its function. In theclassic rubber hand illusion, hand or fingermovements can be considered active inferentialtests of self-related hypotheses. If these move-ments are not reflected in the “rubber hand”,the illusion is destroyed—presumably becausepredicted visual signals are not confirmed (Apps& Tsakiris 2014). However, if hand movementsare mapped to a virtual “rubber hand”—through clever use of virtual and augmentedreality—the illusion is in fact strengthened, pre-sumably because the multisensory correlation ofperi-hand visual and proprioceptive signals con-stitutes a more stringent test of the perceptualhypothesis of ownership of the virtual hand (Su-zuki et al. 2013). This introduces the idea thatactive inference is not simply about confirmingsensory predictions but also involves seeking“disruptive” actions that are most informativewith respect to testing current predictions,6 These are long-range projection neurons found selectively in hominid

primates and certain other species.



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

and/or at disambiguating competing predictions(Gregory 1980). A nice example of how thishappens in practice comes from evolutionary ro-botics7—which is obviously a very different liter-ature, though one that inherits directly fromthe cybernetic tradition.

In a seminal 2006 study, Josh Bongardand colleagues described a four-legged “starfish”robot that engaged in a process much like activeinference in order to model its own morphologyso as to be able to control its movement and at-tain simple behavioural goals (Bongard et al.2006). While there are important differencesbetween evolutionary robotics and (active)Bayesian inference, there are also broad similar-ities; importantly, both can be cast in terms ofmodel selection and optimization.

The basic cycle of events is shown in Fig-ure 6. The robot itself is shown in the centre(A). The goal is to develop a controller capableof generating forward movement. The challengeis that the robot’s morphology is unknown tothe robot itself. The system starts with a rangeof (generic prior) potential self-models (B), herespecified by various configurations of three-di-mensional physics engines. The robot performsa series of initially random actions and evalu-ates its candidate self-models on their ability topredict the resulting proprioceptive afferent sig-nals. Even though all initial models will bewrong, some may be better than others. Thekey step comes next. The robot evaluates newcandidate actions on the extent to which thecurrent best self-models make different predic-tions as to their (proprioceptive) consequences.These disambiguating actions are then per-formed, leading to a new ranking of self-modelsbased on their success at proprioceptive predic-tion. This ranking, via the evolutionary roboticsmethods of mutation and replication, gives riseto a new population of candidate self-models.The upshot is that the system swiftly developsaccurate self-models that can be used to gener-ate controllers enabling movement (D). An in-teresting feature of this process is that it is7 Evolutionary robotics involves the use of population-based

search procedures (genetic algorithms) to automaticallyspecify control architectures (and/or morphologies) of mo-bile robots. For an excellent introduction see (Bongard2013).

highly resilient to unexpected perturbations.For instance, if a leg is removed then proprio-ceptive prediction errors will immediately ensue.As a result, the system will engage in anotherround of self-model evolution (including the co-specification of competing self-models and dis-ambiguating actions) until a new, accurate, self-model is regained. This revised self-model canthen be used to develop a new gait, allowingmovement, even given the disrupted body (E,F).8

This study emphasizes that the opera-tional criterion for a successful self-model is notso much its fidelity to the physical robot, butrather its ability to predict sensory inputs undera repertoire of actions. This underlines that pre-dictive models are recruited for the control ofbehaviour (as cybernetics assumes) and not tofurnish general-purpose representations of theworld or the body.

The study also provides a concrete ex-ample of how actions can be performed, not toachieve some externally specified goal, but topermit inference about the system’s own phys-ical instantiation. Bayesian or not, this impliesactive inference. Indeed, perhaps its most im-portant contribution is that it highlights howactive inference can prescribe disruptive ordisambiguating actions that generate sensoryprediction errors under competing hypotheses,and not just actions that seek to confirm sens-ory predictions. This recalls models of atten-tion based on maximisation of Bayesian sur-prise (Itti & Baldi 2009), and is equivalent tohypothesis testing in science, where the bestexperiments are those concocted on the basisof being most likely to falsify a given hypo-thesis (disruptive) or distinguish betweencompeting hypotheses (disambiguating). Italso implies that agents encode predictionsabout the likely sensory consequences of arange of potential actions, allowing the selec-tion of those actions likely to be the most dis-ruptive or disambiguating. This concept of acounterfactually-equipped predictive modelbring us nicely to our next topic: so-called en-active cognitive science and its relation to PP.8 Videos showing the evolution of both gait and self-model are avail-

able from http://creativemachines.cornell.edu/emergent_self_models


http://creativemachines.cornell.edu/emergent_self_models


http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

4 Predictive processing and enactive cognitive science

4.1 Enactive theories, weak and strong

The idea that the brain relies on internal rep-resentations or models of extra-cranial states ofaffairs has been treated with suspicion eversince the limitations of “good old fashioned arti-

ficial intelligence” became apparent (Brooks1991). Many researchers of artificial intelligencehave indeed returned to cybernetics as an al-ternative framework in which closely coupledfeedback loops, leveraging invariants in brain-body-world interactions, obviate the need fordetailed internal representations of externalproperties (Pfeifer & Scheier 1999). The evolu-tionary robotics methodology just described is


Figure 6: An evolutionary-robotics experiment demonstrating continuous self-modelling Bongard et al. (2006). Seetext for details. Reproduced with permission.


http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

often coupled with simple dynamical neural net-works in order to realize controllers that aretightly embodied and embedded in just this way(Beer 2003). Within cognitive science, such anti-representationalism is most vociferously defen-ded by the movement variously known as “en-active” (Noë 2004), “embodied” (Gallese & Sini-gaglia 2011), or “extended” (Clark & Chalmers1998) cognitive science. Among these ap-proaches, it is enactivism that is most explicitlyanti-representationalist. While enactive theoristsmight agree that adaptive behaviour requiresorganisms and control structures that are sys-tematically sensitive to statistical structures intheir environment, most will deny that thissensitivity implies the existence and deploymentof any “inner description” or model of theseprobabilistic patterns (Chemero 2009; Hutto &Myin 2013).

This tradition has weak and strong expres-sions. At the weak extreme is the truism thatperception, cognition, and behaviour—and theirunderlying mechanisms—cannot be understoodwithout a rich appreciation of the roles of thebody, the environment, and the structured in-teractions that they support (Clark 1997;Varela et al. 1993). Weak enactivism is emin-ently compatible with PP, as seen especiallywith emerging versions of PP that stress em-bodiment through self-modelling and interocep-tion, and which emphasize the importance ofagent-environment coupling (embeddedness)through active inference. At the other extremelie claims that explanations based on internalrepresentations or models of any sort are funda-mentally misguided, and that a new explicitlynon-representational vocabulary is needed in or-der to make sense of the relations betweenbrains, bodies, and the world (O’Regan et al.2005). Strong enactivism is by definition incom-patible with PP since it rejects the core conceptof the internal model.

4.2 Sensorimotor contingency theory

A landmark in the strongly enactive approach isSMC (sensorimotor contingency) theory, whichsays that perception depends on the “practicalmastery” of sensorimotor dependencies relevant

to behaviour (O’Regan & Noë 2001). In brief,SMC theory claims that experience and percep-tion are not things that are “generated” by thebrain (or by anything else for that matter) butare, rather, “skills” consisting of fluid patternsof on-going interaction with the environment(O’Regan & Noë 2001). For instance, on SMCtheory the conscious visual experience of red-ness is given by the exercise of practical mas-tery of the laws governing how interactions withred things unfold (these laws being the“SMC”s). The theory is not, however, limited tovision: the experiential quality of the softness ofa sponge would be given by (practical masteryof) the laws governing its squishiness upon be-ing pressed.

Two aspects of SMC theory deserve em-phasis here. The first is that the concept of anSMC rightly underlines the close coupling ofperception and action and the critical import-ance of ongoing agent-environment interactionin structuring perception, action, and beha-viour. This is inherited from Gibsonian notionsof perceptual affordance (Gibson 1979) and hascertainly advanced our understanding of whydifferent kinds of perceptual experience (vision,smell, touch, etc.) have different qualitativecharacters.

The second is that mastery of an SMC re-quires an essentially counterfactual knowledge ofrelations between particular actions and the res-ulting sensations. In vision, for instance, mas-tery entails an implicit knowledge of the ways inwhich moving our eyes and bodies would revealadditional sensory information about perceptualobjects (O’Regan & Noë 2001). Here SMC the-ory has made an important contribution to ourunderstanding of perceptual presence. Percep-tual presence refers to the property whereby (innormal circumstances) perceptual contents ap-pear as subjectively real, that is, as existing. Forexample, when viewing a tomato, we see it asreal inasmuch as we seem to be perceptuallyaware of some of its parts (e.g., its back) thatare not currently causally impacting our sensorysurfaces. Looking at a picture of a tomato doesnot give rise to the same subjective impressionof realness. But how can we be aware of parts ofthe tomato that, strictly speaking, we do not



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

see? SMC theory says the answer lies in our(implicit) mastery of SMCs, which relate poten-tial actions to their likely sensory effects; and itis in this sense that we can be perceptuallyaware of parts of the tomato that we cannot ac-tually see (Noë 2006).

SMC theory has often been set againstnaïve representationalist theories in cognitivescience that propose such things as “pictures inthe head” or that (like good-old-fashioned-AI)treat accurate representations of external prop-erties as general-purpose goal states for cogni-tion. This is all to the good. Yet by dispensingwith implementation-level concepts such as pre-dictive inference, it struggles with the import-ant question of what exactly is going on in ourheads during the exercise of mastery of a sen-sorimotor contingency. 9

4.3 Predictive perception of sensorimotor contingencies

A powerful response is given by integratingSMC theory with PP, in the guise of PPSMC(Predictive Perception of SensoriMotor Contin-gencies; Seth 2014b). An extensive developmentof PPSMC is given elsewhere (see Seth 2014bplus commentaries and response). Here I sum-marize the main points. First, recall that underPP prediction errors can be minimized either byupdating perceptual predictions or by perform-ing actions, where actions are generatedthrough the resolution of proprioceptive predic-tion errors. Also recall that PP is inherentlyhierarchical, so that at some hierarchical levelpredictive models will encode multimodal andeven amodal expectations linking exteroceptive(sensory) and proprioceptive (motor) sensations.These models generate predictions about linkedsequences of sensory and proprioceptive (andpossibly interoceptive) inputs corresponding tospecific actions, with predictions becoming in-creasingly modality-specific at lower hierarchicallevels. These multi-level predictive models can

9 At a recent symposium of the AISB society that focused on SMCtheory, it was stated that “the main question is how to get the braininto view from an enactive/sensorimotor perspective. […] Addressingthis question is urgently needed, for there seem to be no accepted al-ternatives to representational interpretations of the inner processes”(O’Regan & Dagenaar 2014).

therefore be understood as instantiating the im-plicit sub-personal knowledge of sensorimotorconstructs underlying SMCs and their acquisi-tion. Put simply, hierarchical active inferenceimplies the existence of predictive models en-coding information very much like that requiredby SMC theory.

The next step is to incorporate the notionof mastery of SMCs, which, as mentioned, im-plies an essentially counterfactual kind of impli-cit knowledge. The simple solution is to aug-ment the predictive models that animate PPwith counterfactual probability densities.10 Asintroduced earlier (section 4.1), counterfactu-ally-equipped predictive models encode not onlythe likely causes of current sensory input, butalso the likely causes of fictive sensory inputsconditioned on possible but not executed ac-tions. That is, they encode how sensory inputs(and their expected precisions) would change onthe basis of a repertoire of possible actions (ex-pressed as proprioceptive predictions), even ifthose actions are not performed. The counter-factual encoding of expected precision is im-portant here, since it is on this basis that ac-tions can be selected for their likelihood of min-imizing the conditional uncertainty associatedwith a perceptual prediction. There is a math-ematical basis for manipulating counterfactualbeliefs of this kind, as shown in a recent modelwhere counterfactual PP drives oculomotor con-trol during visual search (Friston 2014; Fristonet al. 2012).11 Here the main point is that coun-terfactually-rich predictive models supply justwhat is needed by SMC theory: an answer tothe question of what is going on inside ourheads during the exercise of mastery of SMCs.

Counterfactual PP makes sense from sev-eral perspectives (Seth 2014b). As mentionedabove, it provides a neurocognitive operational-isation of the notion of mastery of SMCs that iscentral to enactive cognitive science. In doing soit dissolves apparent tensions between enactive10 See Beaton (2013) for a distinct approach to incorporating counter-

factual ideas in SMC theory. Beaton’s approach remains squarelywithin the strongly enactivist tradition.

11 There are also some challenges lying in wait here. For instance, itis not immediately clear how important assumptions like theLaplace approximation can generalize to the multimodal probab-ility distributions entailed by counterfactual PP (Otworowska etal. 2014).



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

cognitive science and approaches grounded inthe Bayesian brain, but only at the price of re-jecting the strong enactivist’s insistence that in-ternal models or representations—of any sort—are unacceptable.12 PPSMC also provides asolution to the challenge of accounting for per-ceptual presence within PP. The idea here isthat perceptual presence corresponds to thecounterfactual richness of predictive models.That is, perceptual contents enjoy presence tothe extent that the corresponding predictivemodels encode a rich repertoire of counterfac-tual relations linking potential actions to theirlikely sensory consequences.13 In other words, weexperience normal perception as world-revealingprecisely because the predictive models underly-ing perceptual content specify a rich repertoireof counterfactually explicit probability densitiesencoding the mastery of SMCs.

A good test of PPSMC is whether it can ac-count for cases where normal perceptual presenceis lacking. An important example is synaesthesia,where it is widely reported that synaesthetic “con-currents” (e.g., the inexistent colours sometimesperceived along with achromatic grapheme in-ducers) are not experienced as being part of theworld (i.e., synaesthetes generally retain intactreality testing with respect to their concurrent ex-periences). PPSMC explains this by noticing thatpredictive models related to synaesthetic concur-rents are counterfactually poor. The hidden (envir-onmental) causes giving rise to concurrent-relatedsensory signals do not embed a rich and deep stat-istical structure for the brain to learn. In particu-lar, there is very little sense in which synaestheticconcurrents depend on active sampling of theirhidden causes. According to PPSMC, it is thiscomparative counterfactual poverty that explainswhy synaesthetic concurrents lack perceptual pres-ence. SMC theory itself struggles to account forthis phenomenon—not least because it struggles toaccount for synaesthesia in the first place (Gray2003). 12 There is a more dramatic conflict with “radical” versions of enactiv-

ism, in which mental processes, and in some cases even their materialsubstrates, are allowed to extend beyond the confines of the skull(Hutto & Myin 2013).

13 Presence may also depend on the hierarchical depth of predictivemodels inasmuch as this reflects object-related invariances in percep-tion. For further discussion see commentaries and response to (Seth2014b).

There are some challenges to thinking thatperceptual presence uniquely depends on coun-terfactual richness. One might think that themore familiar one is with an object, the richerthe repertoire of counterfactual relations thatwill be encoded. If so, the more familiar one iswith an object, the more it should appear to bereal. But prima facie it is not clear that famili-arity and perceptual presence go hand-in-handlike this.14 Also, some perceptual experiences(like the experience of a blue sky) can seemhighly perceptually present despite engaging anapparently poor repertoire of counterfactual re-lations linking sensory signals to possible ac-tions. An initial response is to consider thatpresence might depend not on counterfactualrichness per se, but on a “normalized” richnessbased on higher-order expectations of counter-factual richness (which would be low for theblue sky, for instance). These considerationsalso point to potentially important distinctionsbetween perceived objecthood and perceivedpresence, a proper treatment of which movesbeyond the scope of the present paper.

5 Active inference

5.1 Counterfactual PP and active inference

Active inference has appeared repeatedly as animportant concept throughout this paper. Yet itis more difficult to grasp than the basics of PP,which involve passive predictive inference. Thisis partly because several senses of active infer-ence can be distinguished, which have not previ-ously been fully elaborated.

In general, active inference can be har-nessed to drive action, or to improve perceptualpredictions. In the former case, actions emergefrom the minimization of proprioceptive predic-tion errors through engaging classical reflex arcs(Friston et al. 2010). This implies the existenceof generative models that predict time-varyingflows of proprioceptive inputs (rather than justend-points), and also the transient reduction ofexpected precision of proprioceptive prediction

14 Thanks to my reviewers for raising this provocative point.



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

errors, corresponding to sensory attenuation(Brown et al. 2013).

In the latter case, actions are engaged inorder to generate new sensory samples, with theaim of minimizing uncertainty in perceptualpredictions. This can be achieved in several dif-ferent ways, as is apparent by analogy with ex-perimental design in scientific hypothesis test-ing. Actions can be selected that (i) are expec-ted to confirm current perceptual hypotheses(Friston et al. 2012); (ii) are expected to dis-confirm such hypotheses; or (iii) are expected todisambiguate between competing hypotheses(Bongard et al. 2006). A scientist may performdifferent experiments when attempting to findevidence against a current hypothesis thanwhen trying to decide between different hypo-theses. In just the same way, active inferencemay prescribe different sampling actions forthese different objectives.

These distinctions underline that active in-ference implies counterfactual PP. In order for abrain to select those actions most likely to con-firm, disconfirm, or decide between current pre-dictive model(s), it is necessary to encode ex-pected sensory inputs and precisions related topotential (but not executed) actions. This isevident in the example of oculomotor controldescribed earlier (Friston et al. 2012). Here, sac-cades are guided on the basis of the expectedprecision of sensory prediction errors so as tominimize the uncertainty in current perceptualpredictions. Note that this study retained thehigher-order prior that only a single perceptualprediction exists at any one time, precludingactive inference in its disambiguatory sense.

Several related ideas arise in connectionwith these new readings of active inference.Seeking disconfirmatory or disruptive evidenceis closely related to maximizing Bayesian sur-prise (Itti & Baldi 2009). This also reminds usthat the best statistical models are usuallythose that successfully account for the mostvariance with the fewest degrees of freedom(model parameters), not just those that resultin low residual error per se. In addition, disam-biguating competing hypotheses moves fromBayesian model selection and optimization tomodel comparison, where arbitration among

competing models is mediated by trade-offsbetween accuracy and model complexity (Rosaet al. 2012).

The information-seeking (or “infotropic”15)role of active inference puts a different gloss onthe free energy principle, which had been inter-preted simply as minimization of prediction er-ror. Rather, now the idea is that systems bestensure their long-run survival by inducing themost predictive model of the causes of sensorysignals, and this requires disruptive and/or dis-ambiguating active inference, in order to alwaysput the current-best model to the test. Thisview helps dissolve worries about the so-called“dark room problem” (Friston et al. 2012), inwhich prediction error is minimized by predict-ing something simple (e.g., the absence of visualinput) and then trivially confirming this predic-tion (e.g., by closing one’s eyes).16 Previous re-sponses to this challenge have appealed to theidea of higher-order priors that are incompatiblewith trivial minimization of lower-level predic-tion errors: closing one’s eyes (or staying put ina dark room) is not expected to lead to homeo-static integrity on average and over time (Fris-ton et al. 2012; Hohwy 2013). It is perhapsmore elegant to consider that disruptive anddisambiguatory active inferences imply explor-atory sampling actions, independent of anyhigher-order priors about the dynamics of sens-ory signals per se. Further work is needed to seehow cost functions reflecting infotropic activeinference can be explicitly incorporated into PPand the free energy principle.

5.2 Active interoceptive inference and counterfactual PP

What can be said about counterfactual PP andactive inference when applied to interoception?Is there a sense in which predictive models un-derlying emotion and mood encode counterfac-tual associations linking fictive interoceptivesignals (and their likely causes) to autonomic orallostatic controls? And if so, what phenomeno-15 Chris Thornton came up with this term (personal communication).16 The term “dark room problem” comes from the idea that a free-en-

ergy-minimizing (or surprise-avoiding) agent could minimize predic-tion error just by finding an environment that lacks sensory stimula-tion (a “dark room”) and staying there.



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

logical dimensions of affective experience de-pend on these associations? While these remainopen questions, we can at least sketch the ter-ritory.

We have seen that active inference in ex-teroception implies counterfactual processing, sothat actions can be chosen according to theirpredicted effects in terms of (dis)confirming ordisambiguating sensory predictions. The sameargument applies to interoception. For active in-teroceptive inference to effectively disambiguatepredictive models, or (dis)confirm interoceptivepredictions, predictive models must be equippedwith counterfactual associations relating to thelikely effects of autonomic or (at higher hier-archical levels) allostatic controls. At least inthis sense, interoceptive inference then also in-volves counterfactual expectations.

That said, there are likely to be substan-tial differences in how counterfactual active in-ference plays out in interoceptive settings. Forinstance, it may not be adaptive (in the longrun) for organisms to continually attempt todisconfirm current interoceptive predictions, as-suming these are compatible with homeostaticintegrity. To put it colloquially, we do not wantto drive our essential variables continually closeto viability limits, just to check whether theyare always capable of returning. This recalls ourearlier point (section 4.1) that predictive controlis more naturally applicable to interoceptionthan exteroception, given the imperative ofmaintaining the homeostasis of essential vari-ables. In addition, the causal structure of coun-terfactual associations encoded by interoceptivepredictive models is undoubtedly very differentthan in cases like vision. These differences mayspeak to the substantial phenomenological dif-ferences in the kind of perceptual presence asso-ciated with these distinct conscious contents(Seth et al. 2011).

6 Conclusion

This paper has surveyed predictive processing(PP) from the unusual viewpoint of cyberneticorigins in active homeostatic control (Ashby1952; Conant & Ashby 1970). This shifts theperspective from perceptual inference as fur-

nishing representations of the external world forthe consumption of general-purpose cognitivemechanisms, towards model-based predictivecontrol as a primary survival imperative fromwhich perception, action, and cognition ensue.This view is aligned with the free energy prin-ciple (Friston 2010); however it attempts to ac-count for specific cognitive and phenomenolo-gical properties, rather than for adaptive sys-tems in general. Several implications followfrom these considerations. Emotion becomes aprocess of active interoceptive inference (Seth2013)—a process that also recruits autonomicregulation and influences intuitive decision-mak-ing through behavioural allostasis. A commonpredictive principle underlying interoceptionand exteroception also provides an integrativeview of the neurocognitive mechanisms underly-ing embodied selfhood, in particular the experi-ence of body ownership (Apps & Tsakiris 2014;Limanowski & Blankenburg 2013; Suzuki et al.2013). In this view, the experience of embodiedselfhood is specified by the brain’s “best guess”of those signals most likely to be “me” acrossexteroceptive and interoceptive domains. Fromthe perspective of cybernetics the embodied selfis both that which needs to be homeostaticallymaintained and also the medium through whichallostatic interactions are expressed.

A second influential line deriving from cy-bernetics sets PP within the broader context ofmodel-based versus enactivist perspectives oncognitive science. On one hand, cybernetics hasbeen cited in support of non-representationalcognitive science in virtue of its showing howsimple mechanisms can give rise to complex andapparently goal-directed behaviour by capitaliz-ing on agent-environment interactions, mediatedby the body (Pfeifer & Scheier 1999). On theother, the cybernetic legacy shows how PP canput mechanistic flesh on the philosophical bonesof enactivism, but only by embracing a finessedform of representationalism (Seth 2014b). A keyconcept within enactive cognitive science is thatof mastery of sensorimotor contingencies(SMCs). This concept is useful for understand-ing the qualitative character of distinct percep-tual modalities, yet as expressed within enactiv-ism it lacks a firm implementation basis. “Pre-



http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

dictive Perception of SensoriMotor Contingen-cies” (PPSMC) addresses this challenge by pro-posing that SMCs are implemented by predict-ive models of sensorimotor relations, under-pinned by the continuity between perceptionand action entailed by active inference. Masteryof sensorimotor contingencies depends on pre-dictive models of counterfactual probabilitydensities that specify the likely causes of sens-ory signals that would occur were specific ac-tions taken. By relating PP to key concepts inenactivism, this theory is able to account forphenomenological features well treated by thelatter, such as the experience of perceptualpresence (and its absence in cases like synaes-thesia).

Considering these issues leads to distinctreadings of active inference, which at its mostgeneral implies the selective sampling of sensorysignals to minimize uncertainty about percep-tual predictions. At a finer grain, active infer-ence can involve performing actions to confirmcurrent predictions, to disconfirm current pre-dictions, or to disambiguate competing predic-tions. These different senses rest on the conceptof counterfactually-equipped predictive models;and they generalize the free energy principle toinclude Bayesian-model comparison as well asoptimization and inference.

In summary, the ideas outlined in this pa-per provide a distinctive integration of predict-ive processing, cybernetics, and enactivism.This rich blend dissolves apparent tensionsbetween internalist and enactivist (model-basedand model-free) views on the neural mechan-isms underlying perception, cognition, and ac-tion; it elaborates common predictive mechan-isms underlying perception and control of selfand world; it provides a new view of emotion asactive interoceptive inference, and it shows how“counterfactual” predictive processing can ac-count for the phenomenology of conscious pres-ence and its absence in specific situations. Italso finesses the concept of active inference toengage distinct forms of hypothesis testing thatprescribe different sampling actions (one bonusis that the “dark room problem” is elegantlysolved). At the same time, new and difficultchallenges arise in validating these ideas experi-

mentally and in distinguishing them from al-ternative explanations that do not rely on in-ternally-realised inferential mechanisms.

Acknowledgements

I am grateful to the Dr. Mortimer and TheresaSackler Foundation, which supports the work ofthe Sackler Centre for Consciousness Science.This work was also supported by ERC FP7grant CEEDs (FP7-ICT-2009-5, 258749). Manythanks to Thomas Metzinger and JenniferWindt for inviting me to make this contribu-tion, and for the insightful and helpful reviewercomments they solicited. I’m also grateful toKevin O’Regan and Jan Dagenaar for invitingme to speak at a symposium entitled “Con-sciousness without inner models?” (London,April 2014), which provided a feisty forum fordebating some of the ideas presented here.

References

Adams, R. A., Shipp, S. & Friston, K. J. (2013). Predic-tions not commands: Active inference in the motor sys-tem. Brain Structure and Function, 218 (3), 611-643.10.1007/s00429-012-0475-5

Apps, M. A. & Tsakiris, M. (2014). The free-energy self:A predictive coding account of self-recognition. Neuros-cience and Biobehavioral Reviews, 41, 85-97.10.1016/j.neubiorev.2013.01.029

Ashby, W. R. (1952). Design for a brain. London, UK:Chapman and Hall.

(1956). An introduction to cybernetics. London,UK: Chapman and Hall.

Aspell, J. E., Heydrich, L., Marillier, G., Lavanchy, T.,Herbelin, B. & Blanke, O. (2013). Turning the bodyand self inside out: Visualized heartbeats alter bodilyself-consciousness and tactile perception. PsychologicalScience, 24 (12), 2445-2453. 10.1177/0956797613498395

Bastos, A. M., Usrey, W. M., Adams, R. A., Mangun, G.R., Fries, P. & Friston, K. J. (2012). Canonical micro-circuits for predictive coding. Neuron, 76 (4), 695-711.10.1016/j.neuron.2012.10.038

Beaton, M. (2013). Phenomenology and embodied action.Constructivist Foundations, 8 (3), 298-313.

Beer, R. D. (2003). The dynamics of active categoricalperception in an evolved model agent. Adaptive Beha-vior, 11 (4), 209-243. 10.1177/1059712303114001


http://dx.doi.org/10.1007/s00429-012-0475-5

http://dx.doi.org/10.1016/j.neubiorev.2013.01.029

http://dx.doi.org/10.1177/0956797613498395

http://dx.doi.org/10.1016/j.neuron.2012.10.038

http://dx.doi.org/10.1177/1059712303114001


http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

Blanke, O. & Metzinger, T. (2009). Full-body illusionsand minimal phenomenal selfhood. Trends in CognitiveSciences, 13 (1), 7-13. 10.1016/j.tics.2008.10.003

Bongard, J. (2013). Evolutionary robotics. Communica-tions of the ACM, 56 (8), 74-85. 10.1145/2493883

Bongard, J., Zykov, V. & Lipson, H. (2006). Resilient ma-chines through continuous self-modeling. Science, 314(5802), 1118-1121. 10.1126/science.1133687

Botvinick, M. & Cohen, J. (1998). Rubber hands ‘feel’touch that eyes see. Nature, 391 (6669), 756-756.10.1038/35784

Braitenberg, V. (1984). Vehicles: Experiments in syn-thetic psychology. Cambridge, MA: MIT Press.

Brooks, R. A. (1991). Intelligence without reason. In J.Mylopoulos & R. Reiter (Eds.) Proceedings of the 12thinternational joint conference on artificial intelligence- volume 1 (pp. 569-595). San Francisco, CA: MorganKaufmann Publishers.

Brown, H., Adams, R. A., Parees, I., Edwards, M. & Fris-ton, K. J. (2013). Active inference, sensory attenuationand illusions. Cognitive Processing, 14 (4), 411-427.10.1007/s10339-013-0571-3

Chemero, A. (2009). Radical embodied cognitive science.Cambridge, MA: MIT Press.

Clark, A. (1997). Being there. Putting brain, body, andworld together again. Cambridge, MA: MIT Press.

(2013). Whatever next? Predictive brains, situatedagents, and the future of cognitive science. Behaviorand Brain Sciences, 36 (3), 181-204.10.1017/S0140525X12000477

(2015). Embodied prediction. In T. Metzinger &J. M. Windt (Eds.) Open MIND (pp. 1-21). Frankfurta. M., GER: MIND Group.

Clark, A. & Chalmers, D. J. (1998). The extended mind.Analysis, 58 (1), 7-19. 10.1093/analys/58.1.7

Conant, R. & Ashby, W. R. (1970). Every good regulatorof a system must be a model of that system. Interna-tional Journal of Systems Science, 1 (2), 89-97.

Craig, A. D. (2003). Interoception: The sense of thephysiological condition of the body. Current Opin-ion in Neurobiology, 13 (4), 500-505. 10.1016/S0959

(2009). How do you feel now? The anterior insulaand human awareness. Nature Reviews Neuroscience,10 (1), 59-70. 10.1038/nrn2555

Critchley, H. D., Wiens, S., Rotshtein, P., Ohman, A. &Dolan, R. J. (2004). Neural systems supporting intero-ceptive awareness. Nature Neuroscience, 7 (2), 189-195.10.1038/nn1176

Critchley, H. D. & Seth, A. K. (2012). Will studies ofmacaque insula reveal the neural mechanisms of self-awareness? Neuron, 74 (3), 423-426.10.1016/j.neuron.2012.04.012

Damasio, A. (1994). Descartes’ error. London, UK: MacMillan.

Dunn, B. D., Galton, H. C., Morgan, R., Evans, D.,Oliver, C., Meyer, M. & Dalgleish, T. (2010). Listeningto your heart. How interoception shapes emotion ex-perience and intuitive decision making. PsychologicalScience, 21 (12), 1835-1844. 10.1177/0956797610389191

Dupuy, J.-P. (2009). On the origins of cognitive science:The mechanization of mind. Cambridge, MA: MITPress.

Evrard, H. C., Forro, T. & Logothetis, N. K. (2012). Voneconomo neurons in the anterior insula of the macaquemonkey. Neuron, 74 (3), 482-489.10.1016/j.neuron.2012.03.003

Ferri, F., Chiarelli, A. M., Merla, A., Gallese, V. & Cost-antini, M. (2013). The body beyond the body: Expect-ation of a sensory event is enough to induce ownershipover a fake hand. Proceedings of the Royal Society B:Biological Sciences, 280 (1765), 20131140-20131140.10.1098/rspb.2013.1140

Fletcher, P. C. & Frith, C. D. (2009). Perceiving is believ-ing: A Bayesian approach to explaining the positivesymptoms of schizophrenia. Nature Reviews Neuros-cience, 10 (1), 48-58. 10.1038/nrn2536

Friston, K. J. (2005). A theory of cortical responses.Philosophical Transactions of the Royal Society B: Bio-logical Sciences, 360 (1456), 815-836.10.1098/rstb.2005.1622

(2009). The free-energy principle: A rough guideto the brain? Trends in Cognitive Sciences, 13 (7), 293-301. 10.1016/j.tics.2009.04.005.

(2010). The free-energy principle: A unified braintheory? Nature Reviews Neuroscience, 11 (2), 127-138.10.1038/nrn2787

(2014). Active inference and agency. CognitiveNeuroscience, 5 (2), 119-121.10.1080/17588928.2014.905517

Friston, K. J., Kilner, J. & Harrison, L. (2006). A free en-ergy principle for the brain. Journal of Physiology -Paris, 100 (1-3), 70-87.10.1016/j.jphysparis.2006.10.001

Friston, K. J., Daunizeau, J., Kilner, J. & Kiebel, S. J.(2010). Action and behavior: A free-energy formula-tion. Biological Cybernetics, 102 (3), 227-260.10.1007/s00422-010-0364-z


http://dx.doi.org/10.1016/j.tics.2008.10.003

http://dx.doi.org/10.1145/2493883

http://dx.doi.org/10.1126/science.1133687

http://dx.doi.org/10.1038/35784

http://dx.doi.org/10.1007/s10339-013-0571-3

http://dx.doi.org/10.1017/S0140525X12000477

http://dx.doi.org/10.1093/analys/58.1.7

http://dx.doi.org/10.1038/nrn2555

http://dx.doi.org/10.1038/nn1176


http://dx.doi.org/10.1177/0956797610389191


http://dx.doi.org/10.1098/rspb.2013.1140


http://dx.doi.org/10.1098/rstb.2005.1622

http://dx.doi.org/10.1016/j.tics.2009.04.005.


http://dx.doi.org/10.1080/17588928.2014.905517

http://dx.doi.org/10.1016/j.jphysparis.2006.10.001

http://dx.doi.org/10.1007/s00422-010-0364-z

http://dx.doi.org/10.1016/S0959


http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

Friston, K. J., Adams, R. A., Perrinet, L. & Breakspear,M. (2012). Perceptions as hypotheses: Saccades as ex-periments. Frontiers in Psychology, 3 (151), 1-20.10.3389/fpsyg.2012.00151

Friston, K. J., Thornton, C. & Clark, A. (2012). Free-en-ergy minimization and the dark-room problem. Fronti-ers in Psychology, 3 (130), 1-7.10.3389/fpsyg.2012.00130

Gallese, V. & Sinigaglia, C. (2011). What is so specialabout embodied simulation? Trends in Cognitive Sci-ences, 15 (11), 512-519. 10.1016/j.tics.2011.09.003

Gibson, J. J. (1979). The ecological approach to visualperception. Hillsdale, NJ: Lawrence Erlbaum.

Gray, J. A. (2003). How are qualia coupled to functions?Trends in Cognitive Sciences, 7 (5), 192-194.10.1016/S1364-6613(03)00077-9

Gregory, R. L. (1980). Perceptions as hypotheses. Philo-sophical Transactions of the Royal Society B: Biolo-gical Sciences, 290 (1038), 181-197.10.1098/rstb.1980.0090

Gu, X., Hof, P. R., Friston, K. J. & Fan, J. (2013). An-terior insular cortex and emotional awareness. Journalof Comparative Neurology, 521 (15), 3371-3388.10.1002/cne.23368

Gu, X. & Fitzgerald, T. H. (2014). Interoceptive infer-ence: Homeostasis and decision-making. Trends in Cog-nitive Sciences, 18 (6), 269-270.10.1016/j.tics.2014.02.001

Hinton, G. E. & Dayan, P. (1996). Varieties of HelmholtzMachine. Neural Networks, 9 (8), 1385-1403.10.1016/S0893

Hohwy, J. (2013). The predictive mind. Oxford, UK: Ox-ford University Press.

(2015). The neural organ explains the mind. In T.Metzinger & J. M. Windt (Eds.) Open MIND (pp. 1-22). Frankfurt a.M., GER: MIND Group.

Hutto, D. & Myin, E. (2013). Radicalizing enactivism:Basic minds without content. Cambridge, MA: MITPress.

Itti, L. & Baldi, P. (2009). Bayesian surprise attractshuman attention. Vision Research, 49 (10), 1295-1306. 10.1016/j.visres.2008.09.007

James, W. (1894). The physical basis of emotion. Psycho-logical Review, 1, 516-529.

Knill, D. C. & Pouget, A. (2004). The Bayesian brain:The role of uncertainty in neural coding and computa-tion. Trends in Neurosciences, 27 (12), 712-719.10.1016/j.tins.2004.10.007

Lee, T. S. & Mumford, D. (2003). HierarchicalBayesian inference in the visual cortex. Journal ofthe Optical Society of America A,Optics, image sci-ence and vision, 20 (7), 1434-1448.10.1364/JOSAA.20.001434

Limanowski, J. & Blankenburg, F. (2013). Minimal self-models and the free energy principle. Frontiers in Hu-man Neurosciences, 7 (547), 1-20.10.3389/fnhum.2013.00547

Makin, T. R., Holmes, N. P. & Ehrsson, H. H. (2008). Onthe other hand: Dummy hands and peripersonal space.Behavioural Brain Research, 191 (1), 1-10.10.1016/j.bbr.2008.02.041

Metzinger, T. (2003). Being no one. Cambridge, MA:MIT Press.

Moseley, G. L., Olthof, N., Venema, A., Don, S., Wijers,M., Gallace, A. & Spence, C. (2008). Psychologicallyinduced cooling of a specific body part caused by theillusory ownership of an artificial counterpart. Proceed-ings of the National Academy of Sciences of the UnitedStates of America, 105 (35), 13169-13173.10.1073/pnas.0803768105

Neal, R. M. & Hinton, G. (1998). A view of the EM al-gorithm that justifies incremental, sparse, and othervariants. In M. I. Jordan (Ed.) Learning in GraphicalModels (pp. 355-368). Dordrecht, NL: Kluwer AcademicPublishers.

Noë, A. (2004). Action in perception. Cambridge, MA:MIT Press.

(2006). Experience without the head. Clarendon,NY: Oxford University Press.

O’Regan, J. K. & Dagenaar, J. (2014). Consciousnesswithout inner models: A sensorimotor account of whatis going on in our heads. Proceedings of the AISB.http://doc.gold.ac.uk/aisb50/

O’Regan, J. K., Myin, E. & Noë, A. (2005). Skill, corpor-ality and alerting capacity in an account of sensoryconsciousness. Progress in Brain Research, 150, 55-68.10.1016/S0079-6123(05)50005-0

O’Regan, J. K. & Noë, A. (2001). A sensorimotor accountof vision and visual consciousness. Behavioral andBrain Sciences, 24 (5), 939-1031.

Otworowska, M., Kwisthout, J. & van Rooj, I. (2014).Counterfactual mathematics of counterfactual predict-ive models. Frontiers in psychology: Consciousness Re-search, 5 (801), 1-2. 10.3389/fpsyg.2014.00801

Paulus, M. P. & Stein, M. B. (2006). An insular view ofanxiety. Biological psychiatry, 60 (4), 383-387.10.1016/j.biopsych.2006.03.042


http://dx.doi.org/10.3389/fpsyg.2012.00151



http://dx.doi.org/10.1016/S1364-6613(03)00077-9

http://dx.doi.org/10.1098/rstb.1980.0090

http://dx.doi.org/10.1002/cne.23368


http://dx.doi.org/10.1016/S0893

http://dx.doi.org/10.1016/j.visres.2008.09.007

http://dx.doi.org/10.1016/j.tins.2004.10.007

http://dx.doi.org/10.3389/fnhum.2013.00547

http://dx.doi.org/10.1016/j.bbr.2008.02.041

http://dx.doi.org/10.1073/pnas.0803768105

http://doc.gold.ac.uk/aisb50/

http://dx.doi.org/10.1016/S0079-6123(05)50005-0


http://dx.doi.org/10.1016/j.biopsych.2006.03.042

http://dx.doi.org/10.1364/JOSAA.20.001434


http://dx.doi.org/10.15502/9783958570108


www.open-mind.net

Pfeifer, R. & Scheier, C. (1999). Understanding intelli-gence. Cambridge, MA: MIT Press.

Pickering, A. (2010). The cybernetic brain: Sketches ofanother future. Chicago, IL: University of ChicagoPress.

Ploghaus, A., Tracey, I., Gati, J. S., Clare, S., Menon, R.S., Matthews, P. M. & Rawlins, J. N. (1999). Dissoci-ating pain from its anticipation in the human brain.Science, 284 (5422), 1979-1981.10.1126/science.284.5422.1979

Pouget, A., Beck, J. M., Ma, W. J. & Latham, P. E.(2013). Probabilistic brains: Knowns and unknowns.Nature Neuroscience, 16 (9), 1170-1178.10.1038/nn.3495

Quattrocki, E. & Friston, K. (2014). Autism, oxytocinand interoception. Neuroscience and Biobehavioral Re-views, 47C, 410-430. 10.1016/j.neubiorev.2014.09.012

Rao, R. P. & Ballard, D. H. (1999). Predictive coding inthe visual cortex: A functional interpretation of someextra-classical receptive-field effects. Nature Neuros-cience, 2 (1), 79-87. 10.1038/4580

Rosa, M. J., Friston, K. J. & Penny, W. (2012). Post-hocselection of dynamic causal models. Journal of Neuros-cience Methods, 208 (1), 66-78.10.1016/j.jneumeth.2012.04.013

Salomon, R., Lim, M., Pfeiffer, C., Gassert, R. & Blanke,O. (2013). Full body illusion is associated with wide-spread skin temperature reduction. Frontiers in Beha-vioral Neuroscience, 7 (65), 1-11.10.3389/fnbeh.2013.00065

Schachter, S. & Singer, J. E. (1962). Cognitive, social,and physiological determinants of emotional state. Psy-chological Review, 69, 379-399. 10.1037/h0046234

Seth, A. K. (2013). Interoceptive inference, emotion, andthe embodied self. Trends in Cognitive Sciences, 17(11), 565-573. 10.1016/j.tics.2013.09.007

(2014a). Interoceptive inference: From decision-making to organism integrity. Trends in Cognitive Sci-ences, 18 (6), 270-271. 10.1016/j.tics.2014.03.006

(2014b). A predictive processing theory of sensor-imotor contingencies: Explaining the puzzle of percep-tual presence and its absence in synaesthesia. CognitiveNeuroscience, 5 (2), 97-118.10.1080/17588928.2013.877880.

Seth, A. K. & Critchley, H. D. (2013). Interoceptive pre-dictive coding: A new view of emotion? Behavioral andBrain Sciences, 36 (3), 227-228.

Seth, A. K., Suzuki, K. & Critchley, H. D. (2011). An in-teroceptive predictive coding model of conscious pres-

ence. Frontiers in Psychology, 2 (395), 1-16.10.3389/fpsyg.2011.00395

Singer, T., Critchley, H. D. & Preuschoff, K. (2009). Acommon role of insula in feelings, empathy and uncer-tainty. Trends in Cognitive Sciences, 13 (8), 334-340.10.1016/j.tics.2009.05.001

Sokol-Hessner, P., Hartley, C. A., Hamilton, J. R. &Phelps, E. A. (2014). Interoceptive ability predictsaversion to losses. Cognition and Emotion, 1-7.10.1080/02699931.2014.925426

Suzuki, K., Garfinkel, S. N., Critchley, H. D. & Seth, A.K. (2013). Multisensory integration across exterocept-ive and interoceptive domains modulates self-experi-ence in the rubber-hand illusion. Neuropsychologia, 51(13), 2909-2917.10.1016/j.neuropsychologia.2013.08.014

Thompson, E. & Varela, F. J. (2001). Radical embodi-ment: Neural dynamics and consciousness. Trends inCognitive Sciences, 5 (10), 418-425. 10.1016/S1364-6613(00)01750-2

Tsakiris, M., Tajadura-Jimenez, A. & Costantini, M.(2011). Just a heartbeat away from one’s body: Intero-ceptive sensitivity predicts malleability of body-repres-entations. Proceedings. Biological sciences / The RoyalSociety, 278 (1717), 2470-2476. 10.1098/rspb.2010.2547

Ueda, K., Okamoto, Y., Okada, G., Yamashita, H., Hori,T. & Yamawaki, S. (2003). Brain activity during ex-pectancy of emotional stimuli: An fMRI study.NeuroReport, 14 (1), 51-55.10.1097/01.wnr.0000050712.17082.1c

Van Gelder, T. (1995). What might cognition be if notcomputation? Journal of Philosophy, 92 (7), 345-381.10.2307/2941061

Varela, F., Thompson, E. & Rosch, E. (1993). The em-bodied mind: Cognitive science and human experience.Cambridge, MA: MIT Press.

Verschure, P. F., Voegtlin, T. & Douglas, R. J. (2003).Environmentally mediated synergy between perceptionand behaviour in mobile robots. Nature, 425 (6958),620-624. 10.1038/nature02024nature02024

Wolpert, D. M. & Ghahramani, Z. (2000). Computationalprinciples of movement neuroscience. Nature Neuros-cience, 3 Suppl, 1212-1217. 10.1038/81497


http://dx.doi.org/10.1126/science.284.5422.1979

http://dx.doi.org/10.1038/nn.3495

http://dx.doi.org/10.1016/j.neubiorev.2014.09.012

http://dx.doi.org/10.1038/4580

http://dx.doi.org/10.1016/j.jneumeth.2012.04.013

http://dx.doi.org/10.3389/fnbeh.2013.00065

http://dx.doi.org/10.1037/h0046234



http://dx.doi.org/10.1080/17588928.2013.877880.



http://dx.doi.org/10.1080/02699931.2014.925426

http://dx.doi.org/10.1016/j.neuropsychologia.2013.08.014

http://dx.doi.org/10.1016/S1364-6613(00)01750-2

http://dx.doi.org/10.1098/rspb.2010.2547

http://dx.doi.org/10.1097/01.wnr.0000050712.17082.1c

http://dx.doi.org/10.2307/2941061

http://dx.doi.org/10.1038/nature02024nature02024

http://dx.doi.org/10.1038/81497


http://dx.doi.org/10.15502/9783958570108


the cybernetic bayesian brain - connecting … · the cybernetic bayesian brain ... ult of the...

Documents