Finding Useful Concepts of Representation in Cognitive Neuroscience: a New Tactic for Addressing Dynamical Critiques of Representational Models of Cognition, Action, and Perception
Total Page:16
File Type:pdf, Size:1020Kb
Finding Useful Concepts of Representation in Cognitive Neuroscience: A new tactic for addressing dynamical critiques of representational models of cognition, action and perception A dissertation submitted to the Graduate School of the University of Cincinnati in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Department of Philosophy of the McMicken College of Arts and Science by Jonathan Martin 1 Abstract Since Timothy van Gelder's 1995 paper “What Might Cognition be if not Computation?” there has been a growing distrust of representational accounts of cognition, most notably from proponents of relatively new programs like dynamical systems theory and ecological/embodied approaches to cognition. Some advocates of these programs have argued that the concept of “representation” is deeply flawed – does little explanatory work, prejudices what is researched and how results are interpreted, and loses sight of the complex interactions that hold between agents and their environments. In this dissertation, I will not try to challenge these claims. Instead, I will argue that responses from defenders of representational modeling often fail to adequately meet these challenges, underestimate their insights, and have a tendency to react by redefining “representation” in such a way that it loses its explanatory significance. Although I believe that the ultimate shape and utility of representational models of cognition will only be revealed in the course of doing science, I will advocate a new tack for defending representational modeling. I begin by acknowledging that representational terminology is indeed used inconsistently, and sometimes without warrant or explanatory power. Next, I examine cases from the cognitive neuroscience to try to show that we can distinguish between explanations whose representational features are closely tied to their explanatory ambitions and those that are not. I will argue that the representationalist should emphasize that information-processing solutions are contingent strategies available to organisms for producing adaptive behavior – as are varieties of synergistic agent-environmental coupling – and, therefore, there should be ways of marshaling evidence for or against the hypothesis that any particular cognitive/perceptual capacity is the result of a representational solution. She can 2 then point out that explanatory representational models in neuroscience provide concrete examples of the kinds of empirical considerations and observations that can be used to motivate attributions of representational significance. They do so by bringing together disparate nervous systems and their components under reoccurring principles of anatomical organization, processing strategies, encoding methods, etc., and these produce testable empirical predictions, suggest interventions, unify phenomena as instances of a more general pattern, and exemplify other commonly lauded scientific virtues. A defense of representational modeling should demonstrate that some of our best explanations in cognitive neuroscience do make central use of representational theorizing for securing their explanatory goals. This can be done while also admitting that there may be many cases where organisms adopted non- representational solutions to navigating their environment. Adjusting metaphysical definitions of “representation” to include even apparently non-representational processes is unconvincing and fails to appreciate the reasons that animate representational explanations. 3 4 Chapter 1: Introduction In attempting to clarify or discover the explanatory function of information-processing models, it will be helpful to first get a grasp of why this historically central aspect of cognitive science has recently come under fire. For purposes of brevity, I will assume a basic familiarity with the two major programs have been paramount among representational approaches to cognition – Classical Cognitive Science and Connectionism – although I will provide explanations and examples of these views. I should also note that I will use terms like “information processing,” “computation” and “representation” interchangeably, without implying any particular computational architecture –from classical Symbol Systems models, to production systems and Good Old-Fashioned Artificial Intelligence, to Connectionist Artificial Neural Networks, to neuroscientists’ claim to find “central representation of X” in some neural structure or activity. While the posited mechanisms differ, what these research programs all have in common is that they all rely heavily on the idea that at least much of cognition, action, and perception can be explained in terms of how information is taken-in, stored, and transformed mechanistically (usually in the brain). Criticism of various proposed representational posits have been a lively part of cognitive science for some time now; debates about the appropriate computational architecture for modeling cognition occupied computationalists and connectionist researchers throughout the 1980s and 90s. But, more recently, the status of representational explanations in general has been increasingly called into question. The majority of skepticism about representations has come from proponents of 5 embodied, dynamical, and ecological approaches to cognition. And, while each of these approaches is home to sometimes quite distinct sets of commitments and research interests within their own communities, the most pressing challenges to representation stem from a set of claims and observations that are common to all three approaches. Accordingly, anti- representationalist accounts often advocate a hybrid of these approaches to replace traditional representational models. In what follows, we will look at the reasons that motivate some theorists to question whether or not cognition is best explained in terms of representation, information-processing, or computation, and I will identify the main theoretical commitments and emphases that unite anti-representationalist arguments across these approaches using exemplar cases from empirical work. I will try to show that the challenge posed by this framework to traditional modeling is a serious one; not only does a survey of the cases give us good reasons to think that it may be possible to model much of cognition, action, and perception without representation or information-processing, but there seem to be good arguments for believing that representational assumptions often leads us astray. I begin with this exposition because I think that, for the most part, advocates defending representation quite often fail to appreciate the force of these arguments and, consequently, do not address them adequately. I will then try to formulate how I think the defender of representation can take their opponents arguments seriously while offering an argument in favor of some representational modeling that is grounded in neuroscientific practice. But let us start by looking first at where all this skepticism comes from. 6 Chapter 2: Tightly-Coupled Cognition and Dynamical Arguments Against Representation Arguments against representation-laden models have come from a variety of sources. Within these distinct theoretical frameworks, those who advocate abandoning representational models often rely on a common set of premises. When these premises are employed for antirepresentational purposes, I will call such views Tightly Coupled Cognition (TCC). The following claims are common to what I am labeling TCC approaches: C1: In (at least) many cases of perception and action, the environment is informationally rich enough to govern an agent's successful navigation of a cognitive task. C2: There are models whose characterization of certain organism-environment couplings provide satisfactory explanations of the accomplishment of a perceptual or cognitive task that do not implicate representation or information processing. C3: The class of dynamical systems taken as paradigmatic cases by TCC involve such intimate organism-environmental couplings that they lack the kind of isolatable, interpretable states to which representational explanations typically ascribe informational content. C1-C3 should be read as empirical claims made on the basis of existing models and explanations. each can, on its own, provide some reason for thinking that traditional representational models are not necessary for understanding at least some substantial portion of perceptual and cognitive phenomena. Taken individually, or as a group, they are insufficient to prove that representational modeling in all cases is ill-advised. Still, it is not necessary (or fair) to require that advocates of TCC demonstrate such a strong thesis conclusively in order to see the force of arguments made on the basis of C1-C3. As I will argue, the empirical results that support these claims are reason enough to shake one's confidence in the received view. There are a few important dimensions to claim C1 that ought to be pointed out before proceeding. First, C1 is partly – and most explicitly – a claim about the nature of the 7 environment. Specifically, it requires of some cognitive agent's environment – usually some causal features of objects proximal to the agent in question – that it contains information that can be exploited toward the execution of some action without calling upon some additional informational resource. It is important to note, however, that the notion of information here is often interestingly agent-centric: a room with