On the predictive processing (PP) framework, the brain is considered to be an inference machine which uses approximations of Bayesian probability theory to maintain its own viability. PP explains action and perception to be driven by one simple tenet: minimise prediction error. Prediction error-signals signal that a given prediction is false—the sensory signal received by the brain was not the sensory signal which it predicted would be received. By minimising prediction error, the brain should be led to situations which allow it to maintain itself within prescribed homeostatic boundaries. PP understands perception to be constituted by the brain’s ‘best guess’ about the ultimate causes of its current sensory states. Importantly, this ‘best guess’ is thought to be fundamentally action-oriented because it pertains to the perception of environmental affordances (Clark 2016; Friston 2012; Wiese and Metzinger 2017; cf. Gibson 1979). Furthermore, because PP is advanced within the dynamic tradition of cognitive science, cognition is considered to come as part of the ‘package deal’ once action and perception have been explained (Clark 2016; Hurley 2001). Therefore, PP explains perception, action, and cognition to all be executed via prediction-error minimisation.
The brain’s ‘best inference’ will be the hypothesis which has the highest overall posterior probability. Posterior probability is the overall probability assigned to a hypothesis (given relevant information) and it is calculated by multiplying the hypothesis’ prior probability and its likelihood, and then dividing this amount by the prior probability that the information on which these calculations are based is correct. The prior probability is the probability given to the hypothesis independently of current events. For example, the prior probability of my being perceptually presented with a cat is quite high, because there are many cats living in my street. The prior probability of my being presented with a leopard, however, is quite low. Leopards are rarely, if ever, to be found in my street. The likelihood of a hypothesis can be calculated as follows—assume the truth of the hypothesis in question, and then determine the probability of the evidence actually acquired being collected if that hypothesis were true. For example, if I encounter a cat in my street then the hypothesis “I am perceiving a cat” will be accorded a high likelihood because my sensory stimulation is highly probable in the context of this hypothesis. Similarly, if I happen to confront an escaped leopard in my street, the hypothesis “there is a leopard in front of me” would be accorded a high likelihood. In both cases, the hypothesis in question is highly likely because the sensory stimulation received is highly probable given the hypothesis.Footnote 2
Having outlined the key posits of PP (“prediction-signal”, “error-signal”, “posterior probability”, “prior”, and “likelihood”) I am now going to explain how each can be satisfactorily explained without invoking the concept “representation”. I will do so by drawing on arguments in this vein recently proffered by Orlandi (2015; cf. 2014) and explicitly linking them in with William Ramsey’s job-description challenge. I will explain why these PP posits do not meet the job-description challenge, and so conclude that the key posits of PP can be accepted without thereby requiring adoption of representation.
Ramsey’s “job-description challenge”
When it is used in cognitive science, the concept “representation” is considered to be a psychological notion which requires the invocation of properties such as “aboutness”, “truth-conditions”, and “content”. These properties are not typically thought to exist in non-cognitive physical entities, such as stones or trees. As such, they are to be considered special, higher-level properties, which are not ubiquitous in the natural world. Accordingly, we must have good reasons to describe a given cognitive mechanism or system in terms of representation.
William Ramsey has recently argued that we can only ascribe “representation” to a cognitive mechanism or system if it passes “the job-description challenge” (2009). If a mechanism or system is to pass this challenge it must: (1) play a functional role within the cognitive system which we would pre-theoretically understand to be representational in nature; and, (2) it must be explanatorily beneficial to treat the mechanism as functioning in this manner. Ramsey argues that condition (1) must be met because, otherwise, the concept of “representation” would become empirically vacuous. If we re-define the concept “representation”, such that our use of the concept within cognitive science has nothing in common with our pre-theoretical psychological notion, then ascription of this concept would become meaningless. If condition (1) is not met, the “representational theory of mind” becomes a representational theory in name only. Condition (2) must be met, according to Ramsey, because otherwise the concept “representation” would become explanatorily vacuous. We can trivially describe any mechanism or system in terms of representation. For example, I can describe a stone rolling down a hill in terms of the stone desiring that it reaches the bottom, believing that rolling is the best way to achieve this aim, and so on. However, we do not receive any additional explanatory benefit from this representational account of stone rolling over and above that received from applying a purely non-representational, physicist’s account to the stone’s rolling behaviour. Thus, condition (2) must be met in order to avoid a complete trivialisation of the concept “representation”. Representation should only be ascribed to a cognitive mechanism if ascribing representation helps one gain a better understanding of that mechanism.
In short, “representation” (as it is used in cognitive science) is a psychological concept which is not ubiquitous throughout physical systems in nature. If we are to maintain a robust and empirically useful notion of representation then, according to Ramsey, a given cognitive mechanism should only be described in terms of representation if it deserves to be described in such terms. He suggests that we assess whether a mechanism does deserve to be described in terms of representation by submitting it to the job-description challenge. Ramsey argues that a cognitive mechanism should be described in terms of representation only if it passes this challenge.
“Prediction” and “error” signals
Let us now submit the PP concepts of “prediction-signal” and “error-signal” to the job-description challenge. Both of these concepts appear prima-facie to require representation—predictions signal that such-and-such is the case, whilst errors signal that such-and-such is not the case. Nico Orlandi argues, however, that closer inspection of the role these terms play in PP explanations reveals they are not representational posits (2015; cf.2014). They fail the job-description challenge.
Predictions are present at all levels of the perceptual hierarchy and are passed down to the level immediately below their level of origin. Error-signals, on the other hand, are passed up from their level of origin to the level immediately above it. As such, each is concerned only with proximal conditions.Footnote 3 Therefore, the signals being passed up and down the perceptual hierarchy are better understood in terms of causal covariation or correlation. In order to argue that these signals are representational in nature, one must therefore explain how or why brain-based causal covariation results in or requires representation. William Ramsey argues that accounts of brain-based representation founded upon causal covariation fail the job-description challenge, and so he concludes that mechanisms in the brain which function on the basis of causal covariation should not be considered representational (2009, ch. 4). His summary of this conclusion is worth quoting in full:
Despite its common appeal, the receptor notion of representation [Ramsey’s name for covariation based accounts] comes with a job description that, in this context, has little to do with the role of representation...When we look at the role of receptors inside of cognitive systems, as described by cognitive theories that employ them, we see that the role is better described as something like a reliable causal mediator or relay circuit which, as such, is not representational in nature. In other words, when a causal/physical system (like the brain) is described as performing various cognitive tasks by employing a structure that has the job of causing something to occur when and only when something else occurs, then the system is not, on the basis of this description alone, employing internal representations. (Ramsey 2009, p. 149, italics in original).
Ramsey’s argument, in essence, is that we gain no extra explanatory purchase by treating causal covariation within the brain in terms of representation. We do not arrive at better accounts of neural processing by giving causal covariation between neurons a representational status, because treating them as such does not provide one with any extra explanatory benefits over those one would accrue by treating them as mere non-representational causal correlations. Furthermore, treating causal covariation in terms of representation violates our pre-theoretic use of the concept. Therefore, causal correlation between neural processes fails the job-description challenge and so does not deserve a representational status.
The proponent of representation is likely to object, however, along the following lines:
The universe is stuffed with correlations and it is implausible to count them all as representations (think of accidental correlations). We agree, but note that the correlations between, for example, specific brain states and color perception look to fall onto the intuitively acceptable side of such a divide. (Clark and Toribio 1994, p. 417)
A proponent of this kind of argument will agree that causal covariation, taken alone, is not sufficient for representation. However, if this causal covariation occurs in a biological organ (like the brain) and has proven evolutionarily beneficial (has been selected for by the forces of natural selection), then it should be described in terms of representation.
One could, for example, argue for a teleosemantic account, upon which representation is thought to occur when there is causal covariation which has been selected for by evolution because it performs a fitness enhancing role (Dretske 1988; Millikan 1984). On this kind of account, we determine the representational function of a given mechanism by averting to its evolutionary history. Neural states which co-vary with a given x would be taken to represent x because they were selected for by evolution to respond to it. If the neural states happen to co-vary with y, where y is a non-natural stimulus introduced in the lab, then the neural states will misrepresent because natural selection did not select them for signalling y. Teleosemantics requires ascribing representation by determining the function of a given mechanism, with this function determined in turn by considering what the mechanism itself was selected to do by the forces of natural selection. It could therefore be argued that covariation between neural processing is representational because it has been selected for by the biological forces of natural selection.
Buttressing the concepts of “prediction-signal” and “error-signal” with teleosemantics, however, is not going to help these concepts pass the job-description challenge. Both prediction and error signals are concerned only with proximal conditions, and the functioning of both can therefore be adequately accounted for without invoking the concept “representation”. Consequently, applying teleosemantics to this particular example will not help with the job-description challenge, because teleosemantics is concerned primarily with the content of a given representation. Applying teleosemantics to a given mechanism involves discerning when representation and mis-representation occur. As such, teleosemantics is only applicable to scenarios in which the concept of “representation” has already been applied—the application of teleosemantics to a mechanism requires the assumption that a given instance of causal covariation is representational, and thereafter attempts to naturalise the particular content of that representation. The theory will not, therefore, help one in determining whether or not a given mechanism deserves to be described in terms of representation to begin with. In short, the theory of teleosemantics is only applicable once a given mechanism has already passed the job-description challenge. Consequently, teleosemantics cannot be used in an argument for the claim that instances of causal covariation pass the job-description challenge (cf. Hutto and Myin 2013).
In sum, the concepts “prediction-signal” and “error-signal” should be understood to involve mere causal correlation between neural processes. Causal correlation should not be considered sufficient for representation because it fails the job-description challenge. We do not gain any extra explanatory purchase by treating the causal covariation entailed by “prediction-signals” and “error-signals” in terms of representation. Although one can buttress causal correlation with concepts from evolutionary biology, doing so will not help with the job-description challenge. Therefore, I conclude that we can satisfactorily account for the role of prediction and error signals within the brain by conceiving of them in terms of mere causal mediation.
“Priors”, “likelihoods”, and “posterior probability” are non-representational
Having outlined Orlandi’s argument that prediction and error signals fail the job-description challenge, I am now going to explain why Orlandi thinks that the concepts “prior” and “likelihood” also fail the job-description challenge. Orlandi argues that these concepts are best understood as referring to non-representational biases present in the neuronal system:
Understanding perceptual priors, hyperpriors and likelihoods as biases means thinking that, as a result of repeated exposure to the structure of the world in the evolutionary past and in the present, the perceptual system is skewed to treat certain stimuli in a certain way. (Orlandi 2015, p. 25, italics in original)
She argues that theorists are tempted to explain biases in terms of representation largely because they are in the grips of the traditional cognitivist idea that perception is to be understood in terms of internal inferential transitions between premises and conclusions in some kind of language of thought (cf. Ramsey 2009). Orlandi argues, however, that biases are more realistically understood to fulfil “the simple function of marking a hypothesis as more or less probable. They are like valves. They skew the brain toward certain neuronal arrangements” (Orlandi 2015, p. 25).
Consider the water fountain in my back garden. This fountain is made-up of three parts: a small bowl at the top (the mouth), a large bowl at the bottom, and a pumping mechanism which connects the bottom bowl to the top. The pump plays a biasing role within the fountain by ensuring that the vast majority of water in the fountain stays pooled in the bottom bowl, with only a small amount being pumped back up to the fountain top at any given time. We would not be at all tempted to ascribe the concept of “representation” to the functioning of this pump, and this is presumably because the ascription of representation to this pump fails the job-description challenge—describing the pump’s biasing role in terms of representation does not provide one with any explanatory benefits over and above those one would accrue by simply treating it as a mere non-representational bias in a water fountain system.
Orlandi contends that a similar conclusion should be drawn in the case of the PP concepts “prior” and “likelihood”. She argues that these concepts should be taken to refer to certain biases within a neuronal system, and that one should not treat these biases in terms of representation because there is no explanatory benefit in doing so. Consequently, Orlandi concludes that the concepts “prior” and “likelihood” are better understood as referring to mechanisms which pre-dispose brains to configure themselves into specific organisational patterns in response to environmental stimulation. Once more, the argument I am presenting here is not that biases cannot be understood in terms of representation. Rather, it is that their functioning can be understood entirely without invoking representation, and that treating them in representational terms is not explanatorily beneficial. Biases do not pass the job-description challenge, and so we have no reason to treat them in terms of representation. Therefore, I conclude that the PP concepts “prior” and “likelihood” should be taken to describe non-representational biasing processes occurring within the neural system.Footnote 4
Consider, finally, the concept “posterior probability”. Although Orlandi argues for a non-representational stance toward the processes underlying PP, she concludes that the results of this processing (the resulting ‘winning hypothesis’, which is the hypothesis with the highest overall posterior probability) do deserve to be described in representational terms. Orlandi arrives at this conclusion because she thinks that the winning hypothesis fulfils the three conditions which she takes to be both necessary and sufficient for the ascription of “representation”:
[R]epresentations are only those performance-guiding structures that are de-coupled from their causes, where this fact materialises in their standing for distal or absent conditions. (Orlandi 2014, p. 133).
Orlandi claims that the winning hypothesis is concerned with distal conditions because it is formulated on the basis of sensory information received by the brain (photons, sound-waves, and so on) and yet is itself about things beyond brain-based sensory receptors (such as cats and leopards). She argues that the winning hypothesis is de-coupleable from its environmental causes because it can be deployed even in the absence of environmental causes. Finally, she argues that the winning hypothesis deserves to be treated in terms of representation because it is used by the brain to reason with and plan action. Consequently, according to Orlandi, although PP processing itself does not deserve a representational status the result of this processing does. I am inclined to reject Orlandi’s claim because I believe that it begs-the-question on two crucial points: (1) its presumes that cognition is concerned with distal, and not proximal, states-of-affairs; and, (2) it rests on an assumption of the problematic representation demarcation thesis.
Does cognition concern proximal or distal states of affairs?
Cognitivist theorists tend to assume that, in resolving a given cognitive task, the organism is restricted to the use of information and resources contained within the brain. Indeed, it is primarily for this reason that cognitivists invoke the concept of “representation” within their explanations at all: environmental input alone is generally considered too impoverished to explain successful cognition, and yet organisms do nevertheless successfully cognise. Cognitivists typically assert that the environmental poverty of stimulus for a given cognitive task is ameliorated via the presence of brain-based representations (Chomsky 1959; Fodor 1975; Marr 1982). Thus, “representation” is invoked by cognitivists in order to resolve a problem which only arises if one assumes that cognition is brain-bound. Orlandi’s claim that the PP ‘winning hypothesis’ is concerned with distal states of affairs, and her claim that it is de-coupleable from its causes, can only be made if one makes the prior cognitivist assumption that cognition is brain-bound: if cognition is brain-bound, then it follows that it will be concerned with distal states of affairs (events beyond the brain) and that internal cognitive states will be de-coupleable from their causes.
This cognitivist view of the mind can, however, be rejected. Proponents of enactive and ecological approaches toward mind, for example, deny that it is brain-bound. Rather, they claim that mind is constituted by the brain, the body, and the environment. Theorists working within these research traditions argue that an emphasis on the performative and temporally extended aspects of cognition will lead one to the realisation that there is no poverty of the stimulus in most cognitive domains (Anderson 2014; Barrett 2011; Chemero 2009; Gibson 1979; Hutto and Myin 2013; Thompson 2007). If there is no poverty of the stimulus in a given cognitive domain, then representation should be rejected in that domain because it is posited as the solution to a non-existent poverty of stimulus problem.
If one adopts such an enactive/ecological view of mind and applies it to the PP framework, the predictive brain can be considered to act as the categorical basis for cognition by enabling direct personal level cognitive contact with an organism’s environment (Anderson and Chemero 2013; Bruineberg and Rietveld 2014; Downey 2017; cf. McDowell 1994). If the environment is directly cognised, then cognition itself is concerned with proximal states of affairs and so should not be described in terms of representation. Furthermore, such instances of direct cognition cannot be de-coupled from their environmental causes (or, to be more precise, such instances cannot be de-coupled from their causes in a manner which necessitates the positing of representation).Footnote 5 Therefore, if one adopts PP within an overall enactive or ecological approach to mind, the ‘winning hypothesis’ will be considered to enable direct cognitive contact with the organism’s environment and so should not be described in terms of representation. It will be taken to concern proximal states of affairs and it will not be understood as de-coupleable from its environmental causes.
Thought, action planning, and the ‘representation demarcation thesis’
This leaves us with Orlandi’s final condition, upon which the ‘winning hypothesis’ is described in terms of representation because it is used for thought and the planning of action. This conclusion is implausible because it appears to be based upon what Ramsey (2015) has labelled the “representation demarcation thesis” (RDT). Ramsey defines RDT as “the view that cognitive processes necessarily involve inner representations and cognitive theories must thereby be about the representational states and processes” (2015, p. 4). Orlandi’s final condition for ascribing a representational status to the winning hypothesis appears to be based on acceptance of a version of the RDT because, by her reasoning, it is a conceptual truth that any states which involve the planning and execution of action are representational states.
Ramsey provides three arguments against RDT. The first reason he provides for rejecting RDT is that it requires a conceptualisation of cognition which is arrived at by largely a priori means. Obviously, whether or not a given instance of cognition is to be understood in terms of representation should primarily be an empirical matter. However, by defining cognition in terms of representation (as RDT does), one guarantees that no instance of cognition will ever be non-representational. Either we must find a representational explanation of the cognitive activity in question, or, it will not count as a cognitive activity at all. Ramsey thus rejects RDT because it requires cognitive science accept a priori constraints on its domain of study and he thinks that no serious science should accept such constraints. His second reason for rejecting RDT is that RDT undermines the empirical status of the representational theory of mind. “Representation” is proposed as a theoretical posit which is supposed to play an empirical role in providing an empirical explanation of cognition. If, however, one accepts RDT, the cognitivist research programme within cognitive science no longer looks to be empirical in nature. Representation is not being posited for empirical reasons, because it plays an important role in an empirical theory of cognition. Rather, it is proposed for conceptual reasons, because it is a priori assumed that any empirical theory of cognition must be a theory of representation. Representation thereby becomes an unfalsifiable theoretical posit, and so the cognitivist research programme loses its empirical credentials. Consequently, Ramsey’s second reason for rejecting RDT is that it requires an unscientific approach to cognitive science. Finally, Ramsey rejects RDT because it encourages a wildly deflationary understanding of representation, such that even mere causal mediation or correlation is considered to be sufficient for representation (cf. Sect. 2.1). Aside from making the concept of representation itself almost vacuous, Ramsey concludes that deflationary accounts of representation can in fact hinder our investigation and resultant understanding of cognitive systems and so should be rejected.
Thus, although Orlandi does conclude that the ‘winning hypothesis’ should be understood in terms of representation, I have argued that this conclusion is not warranted because it is reliant upon the prior assumptions of: a cognitivist view of mind (upon which cognition is taken to be brain-bound); and, the RDT. Enactive and ecological versions of PP can accept that the ‘winning hypothesis’ forms the categorical basis of personal level cognition without thereby taking it be concerned with distal events or to be de-coupleable from its causes. Similarly, although such accounts could agree that the ‘winning hypothesis’ is causally implicated in guiding and executing action, concluding that it deserves a representational status on this basis alone is misguided, because reliant on acceptance of RDT. Therefore, the concept of “posterior probability” can satisfactorily be explained within PP accounts without thereby accepting representation.
At this point, we have arrived at a non-representational version of PP. “Prediction-signal”, “error-signal”, “prior”, “likelihood”, and “posterior probability” have all been described in non-representational terms. Importantly, no explanatorily beneficial insight was lost by describing these concepts in such non-representational terms. Having explained how the key PP posits can be understood in non-representational terms, I will now turn to explaining how even PP explanations which make indispensable use of representation can be accommodated within an eliminativist framework. I will argue that such representational posits are indispensable for epistemological, and not metaphysical, reasons.