Handbook of Cognitive Science

An Embodied Approach

Paco Calvo, Toni Gomila

• Publisher: Elsevier Science
• Year: 2008
• eBook ISBN: 9780080914879
• Topic: Computer Science
• Subtopic: Artificial Intelligence (AI) & Semantics
• Index: Computer Science

The Handbook of Cognitive Science provides an overview of recent developments in cognition research, relying upon non-classical approaches. Cognition is explained as the continuous interplay between brain, body, and environment, without relying on classical notions of computations and representation to explain cognition. The handbook serves as a valuable companion for readers interested in foundational aspects of cognitive science, and neuroscience and the philosophy of mind. The handbook begins with an introduction to embodied cognitive science, and then breaks up the chapters into separate sections on conceptual issues, formal approaches, embodiment in perception and action, embodiment from an artificial perspective, embodied meaning, and emotion and consciousness. Contributors to the book represent research overviews from around the globe including the US, UK, Spain, Germany, Switzerland, France, Sweden, and the Netherlands.

Part I. The Interactive Architecture of Cognition: Conceptual Issues

Chapter 2. Is Embodiment Necessary?

Every known instance of a genuine cognitive agent is embodied, and it is clear that embodiment has a major influence on, and can be a major help for, artificial agents (Pfeifer & Scheier, 1999). It is also clear that the notion of embodiment has multiple interpretations (Ziemke, 2001; Wilson, 2002; Svensson & Ziemke, 2005). But these points leave open a basic question: Is there a sense of embodiment in which being embodied is necessary to cognition? Is embodiment in some way necessary to the nature of cognition, or is it (merely) an important but secondary consideration for (most) implementations of cognitive agents (whether biological or artificial)? I will argue that embodiment is in fact necessary—it is essential in the nature of representation, and, therefore, of cognition.

There are several parts to this argument. First, there are considerations of approaches to the modeling of representation that do not have as consequences any such necessity of embodiment. I will argue that these approaches do not and cannot succeed. Second, there is the development of an alternative model of the nature of representation and cognition. This will be outlined, and it is clear that it requires embodiment, because it requires genuine interaction between a cognitive agent and the world. Finally, a word or two will be in order to look at the kind of embodiment that is involved in this interactive approach to representation and cognition.

Critiques

Critiques of standard models of representation can be partially compressed because they are all heirs to an underlying error. This error has many manifestations, some of which have been known for millennia and some of which have been discovered relatively recently. Because the central error is held in common among the various models of representation on offer in the literature, only minimal particularization of the central critiques is required in order to demonstrate the applicability of those critiques to specific models.

This central error (or family of errors) has to do with the normativity of representation: the sense in which representation can be true or false. One criterion for a model of representation is that the model be able to account for the simple possibility that the representation is in error. This can be difficult if the representational relationship is purportedly constituted in some sort of factual relation between representation and represented—for example, a causal or informational or nomological relation—because the proper factual correspondence to constitute a representation cannot exist unless the environmental end of the correspondence exists, so that the representation of that existence, in such a view, must be correct. There have been multiple attempts to avoid this problem,^[1] but I will focus primarily on a strengthened variant of it that has not been addressed.

¹ Without success, I argue elsewhere (Bickhard, 1993, 2004a, in press, in preparation; Bickhard &Terveen, 1995).

This variant is the criterion of being able to account for organism (or system) detectable error—that the system can itself detect its own errors. Such possibilities of detection may be restricted to certain kinds of organism complexity, and may be quite fallible, but we know that they occur, so any model of representation that cannot account for the possibility of such detections is at best incomplete, and any model that precludes such detections is refuted. This criterion is, in fact, of central importance to any complex cognitive system because system detectable error is necessary in order for error guided behavior and learning to occur, and error guided behavior and learning underpin major portions of most species’ cognitive world.

Nevertheless, there is no attempt to address this criterion in the major approaches to representation in the literature. One reason why, so I argue, is that there is no major approach that can possibly account for system detectable error.

Some sense of the depth of the problem posed by this criterion can be found by realizing that it is equivalent to the radical skeptical argument: we cannot check our own representations for truth or falsity because, in order to do so, we would have to step outside of ourselves to obtain independent epistemic access to what we are attempting to represent and then compare what we are trying to represent with our attempted representation of it (Rescher, 1980). We cannot step outside of ourselves, so this is impossible—therefore, so this argument goes, checking our representations for error is impossible. Again, however, we know that error guided behavior and learning occur, so there must be something wrong with this argument. It is not “merely” an armchair philosophical argument: it is a long-standing manifestation of an error in fundamental assumptions about representation.

The radical skeptical argument also illustrates a second, related, criterion: we must be able to compare represented with representation, so we must have access to our own representational contents—we (or the organism, or system, or agent, or central nervous system, etc.) must have access not only to the represented (which is what the skeptical argument focuses on) but also to the content that we are applying to the represented. We must have access to this content in order to make the comparison in order to determine whether or not the content truly applies to the represented. So, in order to engage in any such comparison, we must have access to both sides that are to be compared. Most models fail this second criterion as well as the first.

Fodor

With this pair of criteria in mind, then, I will take a look at some of the central contemporary models of representation, and I begin with Jerry Fodor. Fodor's model is a version of an information semantics, with the crucial representation constituting relationship, in this case, being a nomological relationship between the represented and the representing state in the organism (Fodor, 1987, 1990, 1991, 1998, 2003). As such, the model encounters difficulties accounting for the possibility of representational error, and Fodor has proposed an ingenious attempted solution.

Fodor's model turns on the intuition that false evocations of a representation are dependent on correct evocations, but that there is no reverse dependency—the dependency is asymmetric. Thus, a cow representation may be evoked by cows, but also perhaps by a horse on a dark night. But the possibility of evocation by the horse is dependent on the possibility of evocation by cows, and this dependency is not reciprocated: evocations by cows could continue even if there were never any possibility of evocations by horses.

One problem with this proposal is that such asymmetric dependencies, even among nomological relationships, do not suffice to pick out representation at all. For example, a neurotransmitter docking with a receptor triggers ensuing activity in a nomological manner, and a mimicking poison molecule docking with the same receptor also evokes nomologically related activity—and the poison molecule's possibility of such evocation is asymmetrically dependent on the possibility of the neurotransmitter evoking such activity. Yet there is at best a biologically functional relationship here, not a representational relationship (Levine & Bickhard, 1999).

Setting this concern aside, however, we still find that the alleged representational error is characterizable as error only for an external observer who could (1) (supposedly) determine the counterfactual asymmetries involved among various families of nomological relationships in order to characterize what a representation is supposed to represent—that is, to characterize the content, (2) epistemically access the represented, and (3) compare the two in order to determine whether the content holds of the represented.

Note that this external observer is in precisely the position that the radical skeptical argument points out that no actual epistemic agent can be in for itself. First, no epistemic agent can have access to its own relevant counterfactual nomological relations in order to determine content, and, second, to access the represented in order to make the comparison with the content is precisely the representational problem all over again. This is the circularity that is at the center of the skeptical argument.

For Fodor, the consequence is that content is not accessible, the represented is not independently accessible, and system detectable error is therefore impossible. Consequently, error-guided behavior and learning are not possible. But error-guided behavior and learning occur, therefore the model is refuted.

Millikan

In Millikan's etiological model, representing X is a particular kind of function that some things or conditions might have, and having such a function is constituted in having (or being properly derived from something that has) the right kind of evolutionary selection history (Millikan, 1984, 1993). The crucial selection history is one of undergoing a sufficient number of generations of selection for having the (functional) conseq...

Table of contents

• Brief Table of Contents
• Table of Contents
• Preface
• List of Contributors
• Part I. The Interactive Architecture of Cognition: Conceptual Issues
• Part II. Robotics and Autonomous Agents
• Part III. Perceiving and Acting
• Part IV. A Dynamic Brain
• Part V. Embodied Meaning
• Part VI. Scaling-Up
• Part VII. Emotion and Social Interaction
• Appendix Plates