4.1 Representational Function
Recall the first anti-representationalist challenge introduced in Sect.
2: do the structures characterised as representations in the foregoing presentation—as “inferences,” “predictions,” and “generative models,” for example—genuinely
warrant this representational interpretation? That is, do they perform recognisably representational jobs within the cognitive architecture described by predictive processing?
Anderson and Chemero (
2013) have recently expressed scepticism on just this score. They argue that representational interpretations of predictive processing conflate “different senses of “prediction” that ought to be kept separate.” One sense of “prediction”—what they call “prediction1”—“is closely allied with the notion of correlation, as when we commonly say that the value of one variable “predicts” another,” and is “essentially model-free” (Anderson and Chemero
2013, 203). Another sense (“prediction2”), by contrast, “is allied instead with abductive inference and hypothesis testing,” and is “theory laden and model-rich.” At most, they argue, the evidence for predictive processing is evidence for the ubiquity of prediction1 in cortical activity. Conceptualising such activity in terms of prediction2 is a “theoretical choice not necessitated by the evidence” (Anderson and Chemero
2013, 204). Given that one
can describe the functional asymmetry between bottom-up and top-down signals at the core of predictive processing in a non-representational vocabulary,
18 Anderson and Chemero raise a reasonable challenge: why
bother with the representational interpretation of such cortical processes advanced above?
This challenge is easily answered, however. As Gladziejewski (
2015) has recently argued, the generative models posited by predictive processing perform robustly representational functions within the overall cognitive architecture it posits. Indeed, predictive processing “might be as representational as cognitive-scientific theories get” (Gladziejewski
2015, 561).
I won’t recapitulate every detail of Gladziejewski’s nuanced treatment here, with which I am in complete agreement. For our purposes, the core idea is relatively straightforward: generative models within predictive brains function as “
action-
guiding, detachable, structural models that afford
representational error detection” (Gladziejewski
2015, 559). Each of these characteristics should be familiar from the foregoing presentation, so I will move through them relatively quickly.
First, generative models are
structural models in exactly the sense introduced in Sect.
3.3: they are physically realised cortical networks that recapitulate the causal-probabilistic
structure of the (functionally significant) environment.
Second, this structural resemblance is actively
exploited by the brain in its proper functioning, guiding the organism’s environmental interventions. To see this, recall from Sect.
3 why brains minimize prediction error: namely, to maintain the organism within its expected states. As Gladziejewski (
2015) notes, the central thesis of predictive processing is that the brain’s ability to achieve this feat is
dependent on the resemblance between the causal-probabilistic structure of the generative model and the ambient environment. That is, effective active inference is only possible given a sufficiently accurate model of the causal-probabilistic dependence relationships among significant environmental variables (cf. Hohwy
2013, 91). As Gladziejewski and Milkowski (
2017) note in a recent paper, this makes the structural resemblance between the generative model and the environment
causally relevant to the brain’s proper functioning. Such models are thus “action-guiding” in that the organism’s ability to intervene on its environment to maintain its viability is functionally
dependent on the degree to which its cortical networks accurately recapitulate the causal-probabilistic structure of the action-relevant environment.
Third, an implication of this is that such models are “detachable.
” Specifically, it is the generative model
itself that functions as the locus of behavioural control—of the organism’s active-inference induced environmental interventions—and
not some direct coupling with the environment. As Gladziejewski (
2015) puts it, “active inferences are dictated by
endogenously-
generated hypotheses about causes in the external world.” In this way such generative models genuinely function as a
proxy or
stand-
in for the surrounding environment in much the same way that one might exploit a
map as the locus of
navigational decisions in navigating an unfamiliar terrain. Further, given the fundamentally
predictive character of generative models, this detachment is such that active inferences are guided as much by model-based
expectations (predictive simulations) of how things
will be as by estimates of how they
are.
Finally, such generative models afford
representational error detection. Specifically, they enable the brain to determine to what extent
its internal stand-in for the environment genuinely mirrors its functionally relevant causal structure. This follows from a simple fact: because the brain’s proper functioning is dependent on its ability to minimize prediction error, and this ability is in turn dependent on to what extent its internal model recapitulates the causal-probabilistic structure of the world, the brain can harness failures of prediction error to detect errors in the accuracy of its internal model. Indeed, it is this ability of predictive brains to harness their own sensory inputs as
feedback to the installation and deployment of their generative model that is one of the most attractive features of predictive processing (Hohwy
2013, 49).
As this analysis showcases, the characterisation of generative models as models within predictive processing is neither idle nor vacuous. Such structures function in a way that is robustly representational in character, enabling brains to effectively coordinate the organism’s behaviour with the surrounding environment by constructing an internal surrogate or simulation of that environment with which to predict its sensory effects and support adaptive interventions. It is thus not just that cortical networks recapitulate the causal-probabilistic structure of the environment that renders them generative models. It is the fact this structural resemblance is causally relevant to the brain’s homeostatic functioning and exploited in a way that is recognisably representational in character. Talk of “models” and “prediction” is therefore fully justified.
With this analysis in hand, consider again Anderson and Chemero’s preference for focusing exclusively on anticipatory dynamics within cortical networks in place of the representational interpretation advanced here. It should now be clear that this suggestion neglects the two most important questions in the vicinity. First, what is the function of such anticipatory dynamics? Second, how are they achieved? It is in answering these questions that the representationalist interpretation of predictive processing is required: effectively anticipating the incoming signal is necessary for the organism’s ability to intervene upon the environment to maintain homeostasis, and it is made possible by the exploitation of an internal model of the signal source. Without this representationalist interpretation, the brain’s ability to so successfully “predict1” its incoming sensory inputs is both unmotivated and unexplained. It is not enough to show that brains are “prediction machines”: predictive processing explains how and why they become this way—namely, by installing and deploying a model with which to guide the organism’s viability-preserving interventions in the world.
4.2 Content Determination
Recall now the second challenge introduced in Sect.
2: representational content cannot find a place in the natural world. After consciousness, this “problem of intentionality” constitutes the most significant challenge to a thoroughly naturalistic understanding of the mind, and it has given rise to a truly staggering amount of philosophical work. Of course, I cannot demonstrate that predictive processing solves this perennial problem here. Instead, I offer some preliminary reasons to think that it genuinely transforms the
nature of the problem in a significant way. Specifically, I argue that it situates the problem firmly in the domain of
cognitive science, not
metaphysics.
To see this, it is helpful to begin with a remark by Clark (
2015, 2) in a recent paper discussing the implications of predictive processing for the problem of content:
To naturalize intentionality… “all” we need do is display the mechanisms by which such ongoing viability-preserving engagements are enabled, and make intelligible that such mechanisms can deliver the rich and varied grip upon the world that we humans enjoy. This, of course, is exactly what PP [predictive processing] sets out to achieve.
This passage should be puzzling for two reasons. First, Clark seems to suggest that naturalizing intentionality is a matter of identifying the
neural mechanisms implicated in hierarchical prediction error minimization, which he takes to be part and parcel of the first-order research programme of predictive processing itself. This stands in stark contrast to the division of labour philosophers are accustomed to, in which cognitive scientists posit a computational architecture and philosophers explain what determines the contents of its representations (Fodor
1987; Von Eckardt
2012). Second, Clark seems to ignore all those characteristics of intentionality that have made the problem of content so difficult, reducing it instead to our ability to gain a “rich and varied grip upon the world.” What about determinacy, shared contents, and the possibility of
misrepresentation, for example (Fodor
1987; Hutto and Satne
2015)? It is common knowledge in the philosophy of mind that a mere account of internal mechanisms has little to say about such recalcitrant phenomena.
Nevertheless, I think that Clark is on to something, and it follows once more from predictive processing’s structuralist approach to internal representation.
First, recall from Sect.
2.2 that almost all work on “naturalizing content” has been concerned with linguaformal semantic properties, where the challenge has been to establish the referential properties of in-the-head symbols from which the propositional contents (truth-conditions) of intentional states are recursively constructed. At the heart of this project is a rigid distinction between the formal or “syntactic” properties of such symbol structures and their semantic properties, in which—as with all forms of digital computation—it is assumed that computational procedures are sensitive only to the former, not the latter. Those who argue that
cognition is a matter of syntax-sensitive operations on symbol structures thus need a story about how such structures acquire their contents—hence the project of “naturalistic psychosemantics” (Fodor
1987). As many have noted, however, a worry with this project is that its
starts from the view that the representational status of such structures is epiphenomenal. Worse, this worry is exacerbated by the fact that most attempts to provide a semantics for such symbol structures appeal to
extrinsic properties such as causal or informational relations that are irrelevant to the intrinsic properties by which they perform their functional roles (Bechtel
2009; O’Brien and Opie). For many, this engenders the suspicion that such forms of in-the-head digital computation are not truly representational at all (Stich
1983; Searle
1980), or that their semantic interpretation is at best part of the “informal presentation of the theory” (Chomsky
1995, 55)—what Dennett (
1987, 350) once called a “heuristic overlay” (cf. also Bechtel
2009; Egan
2013).
Structuralist accounts of internal representation of the sort implied by predictive processing fundamentally transform this situation in at least two important ways. First, the semantic properties of such models are grounded in their
intrinsic structure—in the case of predictive processing, in the intrinsic patterns of cortical activity that realise its causal-probabilistic structure (Cummins
2010). Thus the properties implicated in cognitive
processing—the intrinsic structure of the representational vehicles—are the
same properties in virtue of which they represent (through resemblance) their target (O’Brien and Opie
2010). Second, as noted in the previous section, this structural resemblance between the two systems is
causally relevant to the cognitive system’s functioning: the proper functioning of predictive brains is causally dependent on the structural resemblance between their generative model and the environment (Gladziejewski and Milkowski
2017). These two features are bound up with one another, of course: it is only
because the intrinsic structure of a predictive brain's internal model is simultaneously responsible both for its ability to represent
and for the capacities it exhibits that the former can be causally relevant to the latter.
The implication of these facts is straightforward and genuinely transformative, however: issues concerning content determination become directly relevant to the question of how such structures perform their systemic role. As O’Brien and Opie (
2010) note, representational systems that exploit a structural similarity between their internal states and their target are not merely “syntactic engines” that acquire a semantics through
interpretation or through hypothesised causal relations to environmental states; they are full-blown “
semantic engines” in which “computational processes … are driven by the very properties that determine the contents of [their internal] vehicles” (O’Brien and Opie
2010, 8).
The immediate implication of this fact is to situate questions concerning content determination firmly in the realm of cognitive neuroscience, just as Clark suggests. The question becomes
how the brain’s structural and dynamical properties can recapitulate the nested causal structure of the environment in the exploitable manner suggested above—a question upon which there has already been extensive research (Friston
2002,
2008). The problem of integrating representational properties into a scientific metaphysics thus becomes first and foremost a problem in
science, not
metaphysics. Of course, the suggestion is not that philosophers have no role to play in this project—a self-defeating suggestion in the current context, and one undermined by the recent explosion of extremely valuable work in just this area drawn upon here.
19 Rather, the claim is that this work is now firmly entangled with the explanatory concerns of first-order science in a manner largely
absent from the programme of naturalistic psychosemantics as it has been practiced in recent decades.
20
But what about those desiderata that have proven so difficult to accommodate in this project: determinacy, shared contents, the possibility of misrepresentation, and so on? How would a mere account of neural mechanisms speak to those phenomena?
This gets things backwards, however. Cognitive science—indeed, science in general—is under no obligation to accommodate folk psychological or semantic intuitions (Churchland
2012; Cummins
2010).
Contra Hutto (
2017), the mere fact (if it is a fact) that we currently have no story about how to reduce semantic properties as viewed through the lens of folk psychological intuition—namely, as fine-grained determinate truth-conditions—to purely physical properties is not
itself an objection to representationalist treatments of predictive processing. The question is whether such properties are necessary for generative models to perform their functional role. And—as a number of philosophers have noted (Churchland
2012; Cummins
2010; O’Brien and Opie
2015)—these properties in fact sit uneasily with structural representations of the sort harnessed by predictive brains. Representational media such as maps and models, for example, typically lack the fine-grained, determinate contents we pre-theoretically attribute to folk psychological states and associate with linguistic expressions, and these characteristics are likely to be carried over to representation in natural systems.
21 Further, the prospects of identical or shared contents looks hopeless in the context of predictive processing: the internal models of similar animals with similar learning histories will no doubt overlap and resemble each other to substantial degrees, but their contents will still likely be endlessly idiosyncratic (Clark
2015).
What about the notorious problem of
misrepresentation or
error? Again, I cannot hope to tackle this enormous issue here, except to note one cause for optimism: by focusing on the subservience of generative models to pragmatic success, predictive processing moves us away from a picture of internal representations as
judgements to one in which they function as representational
tools—that is, physically instantiated surrogates for the action-salient causal structure of the environment that facilitate viability-preserving environmental interventions. As many have noted, structural representations force us to shun the idea of representational evaluation as a
binary phenomenon in favour of a much looser and more graded notion of accuracy or “aptness,” where—crucially—the vehicle’s
representational value is relativized to the sort of practical application for which it exists to provide guidance (Horst
2016, 86).
22 It is a familiar theme in the philosophy of science that
models are not
true or
false; they are invariably highly idealised, selective and often purposefully distortive
stand-
ins for a domain that enable us to coordinate our practical engagements with it (Giere
2004). Representational
error must therefore be evaluated against such practical ends. As Clark (
2015, 4) puts it, “the
test of a good [generative] model is how well it enables the organism to engage the world in a rolling cycle of actions that maintain it within a window of viability” (my emphasis).
If this is right, it suggests that many of the problems associated with classical attempts to naturalize intentionality may not arise in the context of predictive processing. Clark’s suggestion is perhaps a little over-stated, but it touches on something important. The core thesis of predictive processing is that brains install and deploy a generative model of environmental causes in the service of homeostasis. If we can explain how cortical networks come to embody these pragmatic structural models, and how such models can be exploited in cognitive functioning, we will have “naturalized” intentionality in the only way that could be important to the representational status of the framework.
Before turning to the final challenge outlined in Sect.
2, it is worth introducing an objection that might naturally arise in response to the foregoing treatment. The objection is this: even if one accepts that predictive processing can avoid the first two anti-representationalist challenges in the manner I have suggested, the principal explanation of this is not anything specific to
predictive processing. Rather, it is the fact that predictive processing posits
structural representations. Such structural representations, however, are common to a much broader class of approaches in cognitive science, including both classical computational and connectionist accounts of information-processing. Thus it is not predictive processing
as such that puts an end to the representation wars, but the broader class of structural model-based approaches of which it is merely one manifestation.
23
This objection clearly gets
something right. A structural approach to internal representation has become increasingly popular in recent years—and for good reason.
24 Part of the argument I have advanced here is that predictive processing can
capitalize on the theoretical advantages it enjoys with this broader class of models.
Nevertheless, predictive processing also contributes something genuinely novel. In addition to its implication that model-based representation is the
fundamental kind of representation employed by the brain, it also
situates this compelling structural resemblance-based account of internal representation within an overarching account of neural function that can effectively answer the
third anti-representationalist challenge introduced in Sect.
2. It thus comes with a fuller package of answers to the concerns raised by those sceptical of internal representations in cognitive science. It is to this final challenge, then, that I turn next.
4.3 Cognitive Function
Superficially, at least, the third anti-representationalist challenge introduced in Sect.
2 is the most straightforward to address given the presentation of predictive processing in this paper. This challenge, recall, contends that the concept of representation implies an implausibly “reconstructive” account of perception that fails to capture the “action-oriented” character of cognition and thus the many profound ways in which contingent properties of the
organism are implicated in the contents of its experience.
First, predictive processing fully embraces the control-theoretic perspective on brain function we saw associated with the most perspicuous advocates of this “action-oriented” view in Sect.
2. Predictive brains are fundamentally
pragmatic brains, designed to maintain the organism’s viability under conditions tending towards disorder. As we saw in Sect.
3.4, any
representation that occurs in such systems is subservient to this practical end.
In addition, numerous authors have noted that predictive processing provides a literal vindication of the functional primacy many in the EEEE tradition ascribe to
action in cognition (Bruineberg et al.
2016; Clark
2016). To see this, note that
reactive or
perceptual inference—that is, the process by which brains update top-down predictions to bring them into alignment with the incoming signal—is
in itself impotent when it comes to minimizing “surprisal,” the ultimate function of prediction error minimization. As Hohwy (
2013, 85) nicely puts it, “perceptual inference can make you perceive that you are hurtling towards the bottom of the sea… but cannot do anything to change that disturbing sensory input.” It is only through
active inference that organisms can intervene upon their environments to actively minimize surprising exchanges with them. Thus “perception plays a secondary role in optimising action” (Friston and Stephan
2007, 418), just as many advocates of embodied cognition have long argued (Engel et al.
2015; Glenberg et al.
2013).
Perhaps most importantly, however, predictive processing accommodates the hostility towards “reconstructive” accounts of perception expressed by those in the EEEE tradition. As noted in Sect.
3.5, the world modelled by predictive brains is the organism’s
affective niche, the causal-probabilistic structure of the environment
as it bears upon the brain’s regulatory function and thus the organism’s physiological integrity. This concept of an “affective niche” can accommodate metaphors like “enacting a world” and “world-making” in the enactivist tradition within a thoroughly representationalist outlook on cognition. Indeed, as Barrett (
2017b, 83) puts it (characterising homeostasis as the maintenance of one’s “body budget”): “from the perspective of your brain, anything in your affective niche could potentially influence your body budget, and nothing else in the universe matters. That means, in effect, that
you construct the environment in which you live.”
Nevertheless, at this point a potential objection raises its head. If the contents of these generative models are as profoundly
organism-
relative as I have suggested, what sense can be made of the structural
resemblance that has been at the core of the view advanced here? That is, is there any prospect of independently identifying “what stands on the other side” of this alleged resemblance relation?
25 If
not, one might object that talk of
re-
presentation is not warranted: perhaps this thoroughly pragmatic perspective on brain function should force us to ditch such reconstructive talk in favour of a “performative” or “enactive” understanding of the mind. Bruineberg et al. (
2016, 15) suggest as much in their anti-representationalist treatment of predictive processing: “if my brain is a scientist,” they argue, “it is a crooked and fraudulent scientist.” Their worry is that cortical networks do not really
recapitulate the objective causal-probabilistic structure of the external environment: they are “vehicles of world-making”, not “world-mirroring devices” (Engel et al.
2015, 5).
There are two reasons this objection is misguided.
First, the fact that a model is not “neutrally specified” or “observer-independent” (Anderson
2014) does not imply it is not a model. Advocates of EEEE cognition often write as if the only kind of viable internal representations are what Clark (
2001) calls “objectivist representations,” namely perspective-independent representations of the action-neutral environment of the sort familiar from models of perception in classical cognitive science. This cannot be right, however. Most if not all models in
science are heavily idealised, partially distortive and interest-relative (Giere
2004; Horst
2016). The question is whether the relevant vehicle or vehicles are being exploited as a structural surrogate for another domain, and we have seen excellent reason to suppose they
are in the case of predictive processing: predictive brains exploit cortical activity as a stand-in for the ambient environment with which to anticipate its sensory effects and support viability-preserving interventions.
Second, the organism-relativity defended here does not imply that the elements of generative models are
imagined. It is vastly implausible that brains could generate time-pressured and adaptively valuable behaviour in hostile environments without at least partially recovering the objective structure of such environments. As Gibson (
1979) himself stressed, “affordances” are not
subjective. The point is rather that the objective structure predictive brains
do recover is interest-relative and specified relative to the organism’s practical abilities for intervention. In Anderson’s (
2014,
2016) “performative” theory of brain function, he writes that “because perception is both active and in the service of action, much of the information to which organism are attuned is not objective information of the sort one might need for model-building, but rather
relational information that is more immediately useful for guiding action in the world” (Anderson
2016, 7). The contrast here is simply confused, however:
relational information—for example, the network of complex dependence relationships between essential organismic variables, environmental states and opportunities for intervention—is perfectly
objective and
exactly the kind of information
structural models are suited to represent.
The upshot of these considerations is that predictive processing can accommodate what is
important in the third anti-representationalist challenge introduced in Sect.
2 while nevertheless preserving its robustly representational status. Predictive brains are not passive spectators: they are vehicles of pragmatic success, facilitating self-organization through the construction and exploitation of structural stand-ins for the organism’s affective niche.