Daniel Burnston

Despite their popularity, relatively scant attention has been paid to the upshot of Bayesian and predictive processing models of cognition for views of overall cognitive architecture.  Many of these models are hierarchical ;  they posit generative models at multiple distinct "levels," whose job is to predict the consequences of sensory input at lower levels.  I articulate one possible position that could be implied by these models, namely, that there is a continuous hierarchy of perception, cognition, and action control comprising levels of generative models.  I argue that this view is not entailed by a general Bayesian/predictive processing outlook.  Bayesian approaches are compatible with distinct formats of mental representation.  Focusing on Bayesian approaches to motor control, I argue that the junctures between different types of mental representation are places where the transitivity of hierarchical prediction may be broken, and I consider the upshot of this conclusion for broader discussions of cognitive architecture.

The traditional approach to explanation in cognitive neuroscience is realist about psychological constructs, and treats them as explanatory.  On the “standard framework,” cognitive neuroscientists explain behavior as the result of the instantiation of psychological functions in brain activity.  This strategy is questioned by results suggesting the distribution of function in the brain, the multifunctionality of individual parts of the brain, and the overlap in neural realization of purportedly distinct psychological constructs.  One response to this in the field has been to employ the tools of databasing and machine learning to attempt to find and quantify specific correlations between psychological kinds such as ‘memory’ or ‘attention’ (or sub-kinds thereof) and patterns of activity in the brain.  I assess the status and prospects of these projects.  I argue that current proponents of the project are vague about their aims, vis-à-vis the standard framework, sometimes suggesting substantiation of the framework, sometimes suggesting retaining the framework but revising the ontology of mental constructs, and sometimes suggesting abandonment of the framework.  I argue that extant results from within the projects fail to substantiate the standard framework, and propose an alternative.  On my view, psychological constructs should not be viewed as explanantia, but instead as heuristic concepts that help us uncover ways that behaviors can vary and the ways that the brain implements those distinctions.  I then discuss the normative upshot of these views for databasing and brain mapping projects.

According to the Causal Theory of Action (CTA), genuine actions are individuated by their causal history. Actions are bodily movements that are causally explained by citing the agent's reasons. Reasons are then explained as some combination of propositional attitudesbeliefs, desires, and/or intentions. The CTA is thus committed to realism about the attitudes. This paper explores current models of decision-making from the mind sciences, and argues that it is far from obvious how to locate the propositional attitudes in the causal processes they describe. The outcome of the analysis is a proposal for pluralism: there are several ways one could attempt to map states like "intention" onto decision-making processes, but none will fulfill all of the roles attributed to the attitudes by the CTA.

Abstract:  Proponents of cognitive penetration often argue for the thesis on the basis of combined intuitions about categorical perception and perceptual learning.  The claim is that beliefs penetrate perceptions in the course of learning to perceive categories.  I argue that this “diachronic” penetration thesis is false.  In order to substantiate a robust notion of penetration, the beliefs that enable learning must describe the particular ability that subjects learn.  However, they cannot do so, since in order to help with learning they must instruct learners to employ previously existing abilities.  I argue that a better approach recognizes that we can have sophisticated causal precursors to perceptual learning, but that the learning process itself must operate outside of cognitive influence.

The notion of representation in neuroscience has largely been predicated on localizing the components of computational processes that explain cognitive function. On this view, which I call "algorithmic homuncularism," individual, spatially and temporally distinct parts of the brain serve as vehicles for distinct contents, and the causal relationships between them implement the transformations specified by an algorithm. This view has a widespread influence in philosophy and cognitive neuroscience, and has recently been ably articulated and defended by Shea (2018). Still, I am skeptical about algorithmic homuncularism, and I argue against it by focusing on recent methods for complex data analysis in systems neuroscience. I claim that analyses based on machine learning tools, such as principle components analysis and linear discriminant analysis, prevent individuating vehicles as algorithmic homuncularism recommends. Rather, each individual part contributes to a global state space, trajectories of which vary with important task parameters. I argue that, while homuncularism is false, this view still supports a kind of "vehicle realism," and I apply this view to debates about the explanatory role of representation.

I draw on empirical results from perceptual and motor learning to argue for an anti-intellectualist position on skill. Anti-intellectualists claim that skill or know-how is non-propositional. Recent proponents of the view have stressed the flexible but fine-grained nature of skilled control as supporting their position. However, they have left the nature of the mental representations underlying such control undertheorized. This leaves open several possible strategies for the intellectualist, particularly with regards to skill learning. Propositional knowledge may structure the inputs to sensorimotor learning, may constitute the outcomes of said learning, or may be needed for the employment of learned skill. I argue that sensorimotor learning produces multi-scale associational representations, and that these representations are of the right sort to underlie flexible and fine-grained control. I then suggest that their content is vitally indeterminate with regards to propositional content attribution, because they exhibit a kind of open-ended structure. I articulate this kind of structure, and use it to respond to the three intellectualist strategies. I then show how the perspective I advance offers insights for understanding both instruction and expert practice.

Fodor's view of the mind is thoroughly computational. This means that the basic kind of mental entity is a sententially structured or "discursive" mental representation. It also means that operations over this kind of mental representation have broad architectural scope, extending out to the edges of perception and the motor system. Still, however, in multiple epochs of his work, Fodor attempted to define a functional role for non-discursive, imagistic representation. I describe and critique his two considered proposals, both of which were initially formulated in The Language of Thought, but only the second of which is pursued thoroughly in later work. The first view says that images play a particular kind of functional role in certain types of deliberative tasks. The second says that images are solely restricted to the borders of perception, and act as a sort of medium for the fixing of conceptual reference. I argue, against the first proposal, that a broad-scope computationalism such as Fodor's renders images in principle functionally redundant, since all of their cognitively relevant properties must be represented in discursive form by both the producers and consumers of the representation. I argue, against the second proposal, that empirical evidence suggests that non-discursive representations are learned through perceptual learning, and directly inform category judgments. In each case, I point out extant debates for which the arguments are relevant. The upshot is that there is motivation for limited scope computationalism, in which some, but not all, mental processes operate on discursive mental representations.

Functional decomposition is an important goal in the life sciences, and is central to mechanistic explanation and explanatory reduction.  A growing literature in philosophy of science, however, has challenged decomposition-based notions of explanation.  ‘Holists’ posit that complex systems exhibit context-sensitivity, dynamic interaction, and network dependence, and that these properties undermine decomposition.  They then infer from the failure of decomposition to the failure of mechanistic explanation and reduction.  I argue that complexity, so construed, is only incompatible with one notion of decomposition, which I call ‘atomism’, and not with decomposition writ large.  Atomism posits that function ascriptions must be made to parts with minimal reference to the surrounding system.  Complexity does indeed falsify atomism, but I contend that there is a weaker, ‘contextualist’ notion of decomposition that is fully compatible with the properties that holists cite.  Contextualism suggests that the function of parts can shift with external context, and that interactions with other parts might help determine their context-appropriate functions.  This still admits of functional decomposition within a given context.  I will give examples based on the notion of oscillatory multiplexing in systems neuroscience.  If contextualism is feasible, then holist inferences are faulty—one cannot infer from the presence of complexity to the failure of decomposition, mechanism, and reductionism.

When doing mental ontology, we must ask how to individuate distinct categories of mental states, and then, given that individuation, ask how states from distinct categories interact.  One promising proposal for how to individuate cognitive from sensorimotor states is in terms of their representational form.  On these views, cognitive representations are propositional in structure, while sensorimotor representations have an internal structure that maps to the perceptual and kinematic dimensions involved in an action context.  This way of thinking has resulted in worries about the interface between cognition and sensorimotor systems—i.e., about how representations of these distinct types might interact in performing actions.  I claim that current solutions to the interface problem fail, because they have not sufficiently abandoned intuitions inspired by faculty psychology.  In particular, current proposals seek to show how cognitive states can enforce prior decisions on sensorimotor systems.  I argue that such “determination” views are the wrong kind of views to adopt, given the form distinction.  Instead, I offer a proposal on which propositional representations can at best bias us towards certain kinds of action.  This kind of view, I argue, appealingly distributes the explanation of action across distinctive contributions from cognitive and sensorimotor processing.

There are many theoretical choices involved in whether one argues for against cognitive penetration (CP)—the thesis that perception is sensitive, in an “intelligible way” (Stokes, 2015, p. 75) to our beliefs and other cognitive states.  These choices include how to distinguish cognition and perception, how strong a relationship between them needs to be in order to count as CP, and what kinds of cases provide evidence for one view over the other. 

Aesthetic perception is one case that has recently been argued to provide very good evidence for CP (Nanay, 2014; Stokes, 2014).  I will construe aesthetic perception very broadly—as perception of aesthetically relevant qualities such as an artworks’ being somber or dynamic, and perception of artistic categories, such as being a Monet or being Mannerist or Cubist.  On this broad view, aesthetic perception is what occurs when we use perceptual mechanisms in particularly aesthetic contexts.  This assumes that there is some connection between what goes on in “normal” (non-aesthetic) and aesthetic perception, but leaves open precisely what that connection is.

The issue of CP stands at a crossroads of philosophical and cognitive scientific analysis of aesthetic experience.  Many models within cognitive science make no place for CP, claiming that perceptual analysis of artworks occurs prior to beliefs about them (Bullot & Reber, 2013; Leder, Belke, Oeberst, & Augustin, 2004; Ramachandran & Hirstein, 1999).  Some philosophers, on the other hand, claim that CP is vital for understanding aesthetic appreciation.  According to McMahon, for instance, aesthetic perception is “concepts all the way out” (2012, p. 417).  If this view is right, then there are major revisionary consequences for empirical and philosophical theorizing about aesthetics (Nanay, 2016). 

Some factors involved in aesthetic perception make it a good test case for CP.  If aesthetic perception genuinely occurs, it is one possible case of “high-level” perceptual content, which is often taken to provide prima facie evidence for CP.  Moreover, aesthetic perception can be shaped by expertise that is mediated by social instruction and learning.  These combined features can make aesthetic perception look like a likely case for CP—but only, I will argue, given some highly questionable assumptions and pragmatic choices about how to construe the debate.  These include the tendency to systematically weaken the kinds of relationships that are taken to be evidence for CP, and the (historically well-precedented) tendency to severely underestimate the power of perceptual association and perceptual learning.  Together, these assumptions result in what I will call the scaling argument—the claim that the more high-level, categorical, socially-mediated, and learning-dependent a percept is, the more likely it is to be the result of CP.

I will argue that the scaling argument is invalid.  Given the best empirical and theoretical picture of how perception and perceptual learning work, CP is not required to explain what goes on in aesthetic perception.  Rather, we can account for even expertise-mediated percepts through purely perceptual processes and more mundane roles for cognition—such as instructing us where and how to look—that are too weak to imply revisionary theses such as CP.

In discussion of mechanisms, philosophers often debate about whether quantitative descriptions of generalizations or qualitative descriptions of operations are explanatorily fundamental. I argue that these debates have erred by conflating the explanatory roles of generalizations and patterns. Patterns are types of quantitative relationships that hold between quantities in a mechanism, over time and/or across conditions. While these patterns must often be represented in addition to descriptions of operations in order to explain a phenomenon, they are not equivalent to generalizations, because their explanatory role does not depend on any specific facts about their scope or domain of invariance.

In this paper I criticize a view of functional localization in neuroscience, which I call “computational absolutism” (CA). “Absolutism” in general is the view that each part of the brain should be given a single, univocal function ascription. Traditional varieties of absolutism posit that each part of the brain processes a particular type of information and/or performs a specific task. These function attributions are currently beset by physiological evidence which seems to suggest that brain areas are multifunctional—that they process distinct information and perform different tasks depending on context. Many theorists take this  contextual variation as inimical to successful localization, and claim that we can avoid it by changing our functional descriptions to computational descriptions. The idea is that we can have highly generalizable and predictive functional theories if we can discover a single computation performed by each area regardless of the specific context in which it
operates. I argue, drawing on computational models of perceptual area MT, that this computational version of absolutism fails to come through on its promises. In MT, the modeling field has not produced a univocal computational description, but instead a plurality of models analyzing different aspects of MT function. Moreover, CA cannot appeal to theoretical unification to solve this problem, since highly general models, on their own, neither explain nor predict what MT does in any particular context. I close by offering a perspective on neural modeling inspired by Nancy Cartwright’s and Margaret Morrison’s views of modeling in the physical sciences.

Functional localization has historically been one of the primary goals of neuroscience. There is still debate, however, about whether it is possible, and if so what kind of theories succeed at localization. I argue for a contextualist approach to
localization. Most theorists assume that widespread contextual variability in function is fundamentally incompatible with functional decomposition in the brain,
because contextualist accounts will fail to be generalizable and projectable. I argue that this assumption is misplaced. A properly articulated contextualism can ground successful theories of localization even without positing completely generalizable accounts. Via a case study from perceptual neuroscience, I suggest that there is strong evidence for contextual variation in the function of perceptual brain areas. I
then outline a version of contextualism that is empirically adequate with respect to this data, and claim that it can still distinguish brain areas from each other according to their functional properties. Finally, I claim that the view does not fail the norms for good theory in the way that anti-contextualists suppose. It is true that, on a contextualist view, we will not have theories that are completely generalizable and predictive. We can, however, have successful partial generalizations that structure ongoing investigation and lead to novel functional insight, and this success is sufficient to ground the project of functional localization.

I argue that discussions of cognitive penetration have been insufficiently clear about (i) what distinguishes perception and cognition, and (ii) what kind of relationship between the two is supposed to be at stake in the debate.  A strong reading, which is compatible with many characterizations of penetration, posits a highly specific and directed influence on perception.  According to this view, which I call the “internal effect view” (IEV) a cognitive state penetrates a perceptual process if the presence of the cognitive state causes a change to the computation performed by the process, with the result being a distinct output.  I produce a novel argument that this strong reading is false.  On one well-motivated way of drawing the distinction between perceptual states and cognitive states, cognitive representations cannot play the computational role posited for them by IEV, vis-à-vis perception.  This does not mean, however, that there are not important causal relationships between cognitive and perceptual states.  I introduce an alternative view of these relationships, the “external effect view” (EEV).  EEV posits that each cognitive state is associated with a broad range of possible perceptual outcomes, and biases perception towards any of those perceptual outcomes without determining specific perceptual contents.  I argue that EEV captures the kinds of cases philosophers have thought to be evidence for IEV, and a wide range of other cases as well.

It is a widespread assumption in philosophy of science that representations of data are not explanatory—that they are mere stepping stones towards an explanation, such as a representation of a mechanism.  I draw on instances of representational and explanatory practice from mammalian chronobiology to suggest that this assumption is unsustainable.  In many instances, biologists employ representations of data in explanatory ways that are not reducible to constraints on or evidence for representations of mechanisms.  Data graphs are used to represent relationships between quantities across conditions, and often these representations are necessary for explaining particular aspects of the phenomena under study.  The benefit of the analysis is two-fold.  First, it provides a more accurate account of explanatory practice in broadly mechanistic investigation in biology.  Second, it suggests that there is not an explanatorily “fundamental” type of representation in biology.  Rather, the practice of explanation consists in the construction of different types of representations and their employment for distinct explanatory purposes.

Intentions are commonly conceived of as discrete mental states that are the direct cause of actions. In the last several decades, neuroscientists have taken up the project of finding the neural implementation of intentions, and a number of areas have been posited as implementing these states. We argue, however, that the processes underlying action initiation and control are considerably more dynamic and context sensitive than the concept of intention can allow for. Therefore, adopting the notion of ‘intention’ neuroscientific explanations can easily lead to misinterpretation of the data, and can negatively influence investigation into the neural correlates of intentional action. We suggest reinterpreting the mechanisms underlying intentional action, and we will discuss the elements that such a reinterpretation needs to account for.

There is a long and distinguished tradition in philosophy and psychology
according to which the mind’s fundamental, foundational connection to the world is made by connecting perceptually to features in the environment. On this picture, which we’ll call feature prioritarianism, minds like ours first make contact with the colors, shapes, and sizes of distal items, and then, only on the basis of the information so obtained, build up representations of the objects that bear these features. The feature priority view maintains, then, that our perception/knowledge of features asymmetrically depends on our perception/knowledge of simple features. In this paper, we will argue that, the long and  distinguished tradition in its favor notwithstanding, feature prioritarianism amounts to an untenable picture of the way our minds are related to the world, and will propose a possible alternative. We contend that there is abundant evidence, drawn from many different aspects of the operation of the perceptual system, that collectively tells against the exclusive priority of feature representations over object representations. Moreover, we believe that this evidence leads in the direction of a more complicated conception of perception according to which our representations of features and representations of objects are highly
interactive and mutually informing. Thus, we will argue that perception is
most plausibly construed in neither feature prioritarian nor object prioritarian terms, but by an integrative view on which neither features nor objects are always representationally prior.

Many in cognitive science have noted the importance of external visualizations for reasoning and learning, and have suggested that such visualizations play a role in complex reasoning contexts such as scientific investigation. However, what cognitive role diagrams play in scientific reasoning is unclear. I suggest that mechanistic diagrams function as search organizers in active research projects. Diagrams aid in scientific reasoning by being uniquely positioned to coordinate cognitive search through multiple search spaces, both within an individual and within a field. I examine this role using a number of published diagrams from mammalian chronobiology.