The Philosophy and Science of Predictive Processing

Dina Mendonça, Manuel Curado, Steven S. Gouveia, Dina Mendonça, Manuel Curado, Steven S. Gouveia

• Publisher: Bloomsbury Academic
• Year: 2020
• ISBN:
  9781350099777
• Edition: 1
• Topic: Filosofia
• Subtopic: Mente e corpo in filosofia

This book explores how predictive processing, which argues that our brains are constantly generating and updating hypotheses about our external conditions, sheds new light on the nature of the mind. It shows how it is similar to and expands other theoretical approaches that emphasize the active role of the mind and its dynamic function. Offering a complete guide to the philosophical and empirical implications of predictive processing, contributors bring perspectives from philosophy, neuroscience, and psychology. Together, they explore the many philosophical applications of predictive processing and its exciting potential across mental health, cognitive science, neuroscience, and robotics. Presenting an extensive and balanced overview of the subject, The Philosophy and Science of Predictive Processing is a landmark volume within philosophy of mind.

Part One

Predictive Processing: Philosophical Approaches

1

Predictive Processing and Representation: How Less Can Be More

Thomas van Es

University of Antwerp

Erik Myin

University of Antwerp

1. Introduction

Attempts to explain cognitive phenomena can be representational or non-representational. Representationalist approaches to all or some cognitive phenomena have been criticized for a few decades (see for example Di Paolo et al. 2017, Gibson 1979, Hutto and Myin 2013, 2017, Varela et al. 1991). Nonetheless, wide-ranging representationalism continues to thrive as the mainstream position in cognitive science. In this chapter, we investigate whether this is because the issues have been solved, with particular focus on predictive processing (PP). We argue that, despite significant effort, PP still can’t have the representations it lays claim on. Further, we will defend that the relevant explanatory work can be done without representations. This shows that PP gains from a non-representational interpretation.

PP, in short, is a theoretical framework that places cognition, action, and perception under a single banner of prediction error minimization (Clark 2016, Hohwy 2013). The standard interpretation is as follows. In order to maintain the system’s homeostasis, the brain is thought to actively predict the barrage of stimuli entering the system. These predictions are tested against the actual stimuli, and the brain’s primary (or only) function is to minimize the prediction errors that result from this testing. This prediction error minimization system is optimized using representational models that mirror the causal-probabilistic structure of the world. These models are continuously updated and fine-tuned as the brain detects prediction errors. Representations here are thought to perform their causal role because of their representational status (Clark 2016, Gładziejewski 2016, Hohwy 2013).

Whether it is legitimate to understand PP in a representational way is contested (Bruineberg et al. 2018, Hutto 2017, Kirchhoff and Robertson 2018). Central to the notion of representation is that it has content. On a standard understanding, contentful representation is for a representation to have a target which it represents in a way as being such that it may not be so (Travis 2004). In having a target the contentful representation is intentional; in representing this target in a way, it is intensional. Supposing contents in a naturalistic framework require being able to tell a naturalistically credible story of how both of the representational properties, intentionality and intensionality, can be natural properties. This requires being able to account for how these properties of intentionality and intensionality originate: how some naturally occurring properties genuinely are the representational properties. Moreover, it needs to show that the representations play a causal role qua representations—a causal role which derives from the fact that they are bona fide representations.

Many have argued that standard representational understandings of cognitive systems are inflated because they cannot provide an answer to these challenges about the natural origin and causal efficacy of contents (Hutto and Myin 2013, 2017, Rosenberg 2015). Nonetheless, some theorists have attempted to argue that these challenges can be met in the case of PP. According to a recent idea, representations in PP are analogous to cartographic maps, and attain their representational status by analogy to this prototype. This involves relating to the target by way of structural similarity, and further requires action guidance, decouplability, and misrepresentation (Gładziejewski 2016, Gładziejewski and Miłkowski 2017). As it seems Gładziejewski’s proposal has gained significant traction in the literature, we will take this particular proposal of representations in PP to be representative (Kiefer and Hohwy 2018, Lee 2018, Pezzulo et al. 2017, Wiese 2017, Wiese and Metzinger 2017, Williams 2018, Williams and Colling 2018, but see also Dołega 2017).

Here we will argue that this recent attempt to rescue the representational paradigm is not up to the job. Furthermore, we will argue that any explanatory advantages PP might have can be retained without representations. First, we will discuss general objections against representationalism. Second, we will discuss Gładziejewski’s (2016) account of representation and analyze whether it solves those issues. Third, we will offer a sketch of an alternative approach to PP, to show that we can move forward without mental representations. Finally, we will conclude that PP is best understood in a non-representational way.

2. General Objections against Representationalism

The account of representation endorsed by philosophers and cognitive scientists alike is one according to which representing involves describing, characterizing, picturing “things as being thus and so—where, for all that, things need not be that way” (Travis 2004: 58). In other ways, representation involves specifying in a way which can be true or false, accurate or inaccurate, or something akin. Something that is specified in some way can be compared to the world, and it can either be correct or not correct, aligned or misaligned. Moreover, representation requires some form or medium of specification, that is description, characterization, or picture—something which has “saying power.” As this characterization of representation resonates with the way representation is widely understood philosophically, we will take it as offering legitimate criteria that any candidate representation in PP should minimally meet. Further, due to the causal/explanatory context in which representations are invoked in PP, defenders of representational PP need to establish that representations play a role qua representation: contents need to be causally efficacious.

Let’s call “structural accounts of representation,” all those that construe representation as a relation between some domain of reality and some distinct structured entity, where representation is construed in terms of some mapping, causal or other, between elements and or relations in the domain of reality, and elements or relations in the structure. A minimalistic structural account would count causal covariation between some worldly entity and some other entity as representational, while a cartographical map of a terrain would be an example of a more complicated form of structural representing. Pure structural accounts of representation face some well-known difficulties for meeting the criteria of our characterization (see for example de Oliveira 2018 for more, and further references). A prominent issue is that they over-generate representations, as structural similarities are widespread. This applies not just to the most minimal, correlational, reading of similarity, but also to more complex readings. That is, there seems to be no principled limit on what can be seen as structurally similar to something else—and there seems to be no principled distinction between what can be seen to be structurally similar, and what genuinely is structurally similar. Moreover, and independent of this, structural accounts of representation should provide an explication of why the structural relations amount to being representational relations, over and above being just structural relations.

A way to accommodate these two concerns is to add a functional component to a structural basis. Such has become standard practice (see again de Oliveira 2018). That is, most accounts of representation, while including a structural component, require that the structural elements and/or relations play some role in the functioning of the system of which they are a part. They should be a “fuel for success” (Godfrey-Smith 1996, Miłkowski and Gładziejewski 2017: 339, Shea 2007, 2014). Adding function constrains the class of relations that could count as representational with respect to a pure structural account—representation is only to be found where there is function. Moreover, by introducing function, causal roles come into play, to potentially offer a solution to demand that representations be causally efficacious.

In order to assess whether structural-plus-functional theories of representation can successfully meet the challenge of showing that candidate representations meet the appropriate criteria for being a representation, the question should first be answered what is doing the representing in such accounts. Or: where are the candidate representations located in such an account?

Of course, the representation cannot just be identified with the structures, because then the allegedly different combined structural/functional view inherits all the problems of a pure structural account. That is, if the structural elements represent by themselves, all the questions raised with respect to structural representations can be raised again. So the function has to play a constitutive role for representation in a combined account. This leads to two problems, however. First, the need for adding function arises because structural similarities are cheap, and ubiquitous. Let us see whether a function plus structural similarities account sufficiently constrains the notion of representation.

Consider some examples. There are many structural similarities that will, prima facie, be excluded with this requirement of functionality. Think for example of the structural similarities between the dimensions of a countertop and that of a cardboard box. The relation is certainly not representational, and indeed, without further context, the structural similarities are not a fuel for success of an action. But what if we decide to place the box on the countertop? The structural similarities are now a fuel for success in the action of placing the box. It is because of the structural similarities between the countertop and the box that the action is successful. Had the countertop been convex, the cardboard box most likely would have slid off. In much the same way the structural similarities are a fuel for success in socks and our sliding them on our feet, door knobs and our turning of them to open doors, and our sitting on chairs. Yet none of these exploitable structural similarities are representations.

This over-generation of representation extends further and encompasses any sort of adaptive behavior. Any behavior that is adapted to the world contains structural similarities to whatever aspects of the world it is adapted to. These structural similarities, further, are, prima facie, a fuel for success. We are able to walk so easily due at least in part to the structural similarities that hold between the shape of our feet and the surface of the world. The same story can be told for any adaptive behavior.

Second, any combined structural/functional account will struggle with ascribing a causal role to representation. The move to include function, or use, in the definition of representation is prima facie convincing: if a representation is only the type of relation that enables usage, it has to have causal effect. There are now, however, two ways in which “use” is employed in representations. This invites a circle that is difficult to escape from. A representation is defined by its use. Only when a composition of driftwood is used as a map can the composition count as a representation for the structure of the terrain it is used for (Ramsey 2007). However, as representations figure in an explanatory context, the usefulness of a representation, in turn, is defined in terms of the representational relation it bears onto the target. In other words, the way the relations are used are what make them representational, and their representational status is what makes them useful. Put differently,

(A) a structural similarity is used, because of

(B) its representational relation to the world.

Further, a structural similarity is said to have

(B) a representational relation to the world, because of

(A) the way it is used.

This clearly shows the circularity of the argument. We ground (A) in (B), and, conversely, ground (B) in (A). The representational status of the structural similarities is thus grounded in our use, and our use of the structural similarities is grounded in their representational status. If representational status is to be grounded in use, then use cannot in turn be grounded in its representational status. This is viciously circular.

If the use of a relation is what grants the relation representational status, then there are two options. Either the use of the representation is based in something non-representational, or it is based in something representational. The representationalist thus faces a dilemma. She may attempt to ground the use of representations into something non-representational. One option would be to ground its usefulness in, say, covariance (or some other characteristic of the relation). Yet this means that it is not the representational relation, but the covariational relation that actually figures in the causal explanation. The representational status of the relation has become causally irrelevant. There would be no point to ascribe representational status to a relation in which only its covariation is causally relevant. The same goes for any other characteristic invoked here—structural similarity instead of covariance, for example. The representationalist may then opt for the second horn, and attempt to ground the use of representations in their representational relation to the world. This ensures a position in the explanatory causal chain for the representation, but now the account of representation can no longer be use-based. If it remains use-based, this comes at the cost of confusing explanandum and explanans: the representational status and use now serve to explain one another circularly.¹

This does not mean that we argue representational status cannot ever be grounded in use. On the contrary, public representations attain their representational status by way of social norms under which particular objects are used as stand-ins for others, and are thus grounded in our practices, our use (Hutto and Myin 2017, Van Fraassen 2008). Yet the cognitive system’s reliance on representations cannot be grounded in social norms in the same way (Tonneau 2012: 339). Mental representations in PP are thought to precede social activity, instead of being a product thereof. There is thus a clear disanalogy between a cognitive agent’s activity and our personal level use of public representations. Thus, for our current purposes, grounding representations in use is insufficient to qualify for full representational status.

3. Can PP Have Its Representations?

Standardly, PP is viewed as a representational theory of cognition (Clark 2016, Hohwy 2013). Though most theorists working on PP take this as a background assumption, few people have attempted to unearth the foundations of this supposed representational character. There has, however, been one proposal that seems to have gained traction in the field (Kiefer and Hohwy 2018). This is Gładziejewski’s (2016) proposal, in which he uses Ramsey’s (2007) compare-to-prototype strategy to attempt to answer the job description challenge.² His prototype of choice is a cartographic map, to which he claims PP’s models are analogous to. These models recapitulate the cau...