REPRESENTATION IN MIND NEW APPROACHES TO MENTAL REPRESENTATION
Previous books published in the Perspectives on Cognitive Science series: Creativity, Cognition and Knowledge - An Interaction
Terry Darnall Language Universals and Variation
Mengistu Amberber and Peter Collins Perspectives on Cognitive Science, Vol 2 - Theories, Experiments, and Foundations
Janet Wiles and Terry Dartnall Perspectives on Cognitive Science, Vol 1 - Theories, Experiments, and Foundations
Peter Slezak, Terry Caelli and Richard Clark
-.~mL
Perspectives on Cognitive Science ~
~
Series editor: Peter Slezak ~
~ -,,~n
mBw,.-
REPRESENTATION IN MIND NEW APPROACHES TO MENTAL REPRESENTATION
EDITED BY
H U G H CLAPIN University of Sydney, NSW, Australia
PHILLIP STAINES University of New South Wales, Sydney, Australia
PETER S L E Z A K University of New South Wales, Sydney, Australia
2004
ELSEVIER Amsterdam
- Boston
Paris - San Diego
- Heidelberg
- San Francisco
- London - Singapore
- New
York - Oxford
- Sydney
- Tokyo
ELSEVIER B.V. Sara Burgerhartstraat 25 P.O. Box 211 1000 AE Amsterdam The Netherlands
ELSEVIER Inc. 525 B Street, Suite 1900 San Diego CA 92101-4495 USA
ELSEVIER Ltd The Boulevard, Langford Lane, Kidlington Oxford OX5 1GB UK
ELSEVIER Ltd 84 Theobalds Road London WC1X 8RR UK
9 2004 Elsevier Ltd. All rights reserved. This work is protected under copyright by Elsevier Ltd, and the following terms and conditions apply to its use: Photocopying Single photocopies of single chapters may be made for personal use as allowed by national copyright laws. Permission of the Publisher and payment of a fee is required for all other photocopying, including multiple or systematic copying, copying for advertising or promotional purposes, resale, and all forms of document delivery. Special rates are available for educational institutions that wish to make photocopies for non-profit educational classroom use. Permissions may be sought directly from Elsevier's Rights Department in Oxford, UK; phone: (+44) 1865 843830, fax: (+44) 1865 853333, e-mail:
[email protected]. Requests may also be completed on-line via the Elsevier homepage (http://www.elsevier.com/locate/permissions). In the USA, users may clear permissions and make payments through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA; phone: (+1) (978) 7508400, fax: (+ 1) (978) 7504744, and in the UK through the Copyright Licensing Agency Rapid Clearance Service (CLARCS), 90 Tottenham Court Road, London W1P 0LP, UK; phone: (+44) 20 7631 5555; fax: (+44) 20 7631 5500. Other countries may have a local reprographic rights agency for payments. Derivative Works Tables of contents may be reproduced for internal circulation, but permission of the Publisher is required for external resale or distribution of such material. Permission of the Publisher is required for all other derivative works, including compilations and translations. Electronic Storage or Usage Permission of the Publisher is required to store or use electronically any material contained in this work, including any chapter or part of a chapter. Except as outlined above, no part of this work may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without prior written permission of the Publisher. Address permissions requests to: Elsevier's Rights Department, at the fax and e-mail addresses noted above. Notice No responsibility is assumed by the Publisher for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions or ideas contained in the material herein. Because of rapid advances in the medical sciences, in particular, independent verification of diagnoses and drug dosages should be made. First edition 2004 Library of Congress Cataloging in Publication Data A catalog record is available from the Library of Congress. British Library Cataloguing in Publication Data A catalogue record is available from the British Library. ISBN: 0-08-044394-X Q The paper used in this publication meets the requirements of ANSI/NISO Z39.48-1992 (Permanence of Paper). Printed in The Netherlands.
Contents
Volume Editors
vii
Contributors
ix
Editors' Note Introduction: Issues and Overview
xi
Notes Toward a Structuralist Theory of Mental Representation Gerard 0 'Brien and Jon Opie .
.
.
.
.
Representation and Indication Robert Cummins and Pierre Poirier
21
Representation and the Meaning of Life Wayne D. Christensen and Cliff A. Hooker
41
The Dynamic Emergence of Representation Mark H. Bickhard
71
New Norms for Teleosemantics Timothy Schroeder
91
Representation and Experience Frank Jackson
107
vi
.
Contents
Phenomenal Qualities and Intermodal Perception
125
Ian Gold On Folk Psychology and Mental Representation Peter Godfrey-Smith
,
10.
11.
147
The World Gone Wrong? Images, Illusions, Mistakes and Misrepresentations Peter Slezak
163
Representation, Reduction, and Interdisciplinarity in the Sciences of Memory John Sutton
187
Kant on Constructing Causal Representations
217
Patricia Kitcher
Author Index
237
Subject Index
243
Volume Editors
Hugh Clapin is an Honorary Associate in the Department of Philosophy at the University of Sydney. He holds B.Sc and Ph.D. degrees from the University of New South Wales, and has taught at the Australian National University and the University of Sydney. His publications include Connectionism Isn't Magic (Minds and Machines 1991) and Problems with Principle P(Pacific Philosophical Quarterly 1997) and he is the editor of Philosophy of Mental Representation (OUP 2002).
PhiUip Staines is Senior Lecturer in the School of Philosophy at the University of New South Wales. He teaches courses in Cognitive Science, Philosophy of Language and the foundations of Artificial Intelligence as well as an introductory course in reasoning and argument. His Ph.D. examined the conditions under which logic can be reliably used to evaluate reasoning and he has published in the areas of logic, philosophy of language and informal reasoning as well as computer and cognitive science. He co-authored Reasoning and Argument in Psychology (Routledge and Kegan Paul 1983) and Logical Psych (UNSW Press 2001). Peter Slezak is Senior Lecturer in the School of History and Philosophy of Science and Coordinator of the Graduate Program in Cognitive Science at the University of New South Wales, Sydney, Australia. His teaching and research interests include, mental representation and visual imagery, philosophy of language and mind, Descartes and early modern philosophy, Galileo, philosophy of science, sociology of scientific knowledge and science education. He has edited several volumes and published in leading journals including Philosophical
Psychology, Language & Communication, British Journal for Philosophy of Science, Behavioral & Brain Sciences, Synthese and Analysis. He was chair of the joint conference of the 4th International Conference on Cognitive Science and the 7th Australasian Society for Cognitive Science 2003 Sydney.
This Page Intentionally Left Blank
Contributors
Mark H. Bickhard
Cog,,~t~,,,~ o,~r,,t~. . . . n th,~ Ohilosophy of Knowledge, Leh aa, PA, USA
Wayne Christensen
Konrad Lorenz Institute, Altenberg, Austria
Hugh Clapin
Department of Philosophy, University of Sydney, Sydney, Australia
Robert Cummins
Philosophy Department, UC Davis, Davis, CA, USA
Peter Godfrey-Smith
Philosophy Program, Research School of Social Sciences, Australian National University, Canberra ACT, Australia
Ian Gold
School of Philosophy and Bioethics, Monash University, Victoria, Australia
Cliff Hooker
Philosophy, University of Newcastle, Callaghan, NSW, Australia
Frank Jackson
Philosophy Program, Research School of Social Sciences, Australian National University, Canberra ACT, Australia
Patricia Kitcher
Philosophy, Columbia University, New York, NY, USA
Gerard O'Brien
Department of Philosophy, University of Adelaide, SA, Australia
Jon Opie
Department of Philosophy, University of Adelaide, SA, Australia
Pierre Poirier
Philosophy Department, University of Quebec in Montreal, Montreal (Quebec), Canada
Tim Schroeder
Department of Philosophy, University of Manitoba, Winnipeg, MB, Canada
Peter Slezak
School of History and Philosophy of Science, University of New South Wales, Sydney, NSW, Australia
Phillip Staines
School of Philosophy, University of New South Wales, Sydney, NSW, Australia
John Sutton
Department of Philosophy, Macquarie University, NSW, Australia
Editors' Note This volume has its genesis in the conference "Representation in Mind: New Approaches to Mental Representation," held at the University of Sydney from 27 to 29 June 2000 and organised by Hugh Clapin. The conference formed a part of the "New Theories of Mental Representation" U2000 Research Project of the Department of Philosophy at the University of Sydney. Drafts of these papers gathered together in this collection with one exception, were presented at the conference. While Robert Cummins was not able to attend the conference as expected, the inclusion of his paper "Representation and Indication," co-authored with Pierre Poirier, completes the volume.
Introduction: Issues and Overview
... I don't believe that folk psychology says there is a language of thought. Rather, I think it is agnostic about how mental representation w o r k s - and wisely so. David Lewis (1996) This introduction proceeds by briefly considering a problem for which mental representation has been seen as an answer, sketching an influential framework, outlining some of the issues addressed in this book and then providing an overview of the papers. At first blush it seems clear that thought can cause action and (perhaps only at second blush) that thought involves representation. If so, it is natural to ask what role representations play in causing action and to further ask what kinds of things representations are. What kinds of systems could they be part of or interact with in the generation of action? How thought can cause action remains a continuing mystery m one aspect of the mind-body problem. One source of the problem is the apparent gull in this case, between cause and effect m they seem to come from completely different realms, the mental and the physical. This has been, perhaps, most vividly presented by Leibniz (in the opposite, world to mind direction) with his image of the magnified machine: Moreover, we must confess that the perception and what depends on it, is inexplicable in terms of mechanical reasons, that is through shapes and motions. If we imagine that there is a machine whose structure makes it think, sense and have perceptions, we could conceive it enlarged, keeping the same proportions so that we could enter into it, as one enters into a mill. Assuming that, when inspecting its interior, we will only find parts that push one another, and we
xii
Introduction
will never find anything to explain a perception. And so, we should seek perception in the simple substance and not in the composite or the machine. Leibniz (1714) We can see Leibniz as drawing attention to the disconnection between concepts that apply to the physical machine/brain and the kinds of concepts we use in our everyday talk about the mental. Cummins (2000) calls this "Leibniz's gap." The latter concepts, include those of belief, desire and intention m part of an informal system sometimes called "folk psychology." Concepts like these have been wielded by humans in their understanding of their fellows and the world (both animate and inanimate) for millennia. Materialists have reacted to this disconnection of mental and physical by taking thought to be on the physical side of the gulf. It can be tempting to think that mental representations can help bridge it. Being physical they are on the fight side of the gulf to makes things happen, and being representations they point to, stand for or represent, as thought does. Indeed much theorizing using mental representations can usefully be seen as located somewhere between the opposite sides of this gulf and attempting, if not to bridge them, to forge a link. Some philosophers, going back at least as far as Sellars (1953) have analysed the circumstances of a person believing a proposition p into two steps m a relation between the person and a sentence and the sentence meaning that p. The sentence represents the proposition. Where the sentence is taken to be in some internal language of thought, as in Field (1978), we have mental representation. So far, we have the statics of the idea. Its dynamics added enormously to its attraction. Put simply, Aristotle's discovery that logical deduction could be treated formally in terms of patterns of symbols and Turing's demonstration that machines could handle the symbol manipulation meant that, appropriately interpreted, machines could be seen as reasoning logically. For example, computers could be programmed to take pairs of strings of symbols (having, say, the patterns "p" and "if p then either q or r") representing two premises and produce a string of symbols (in this case having the pattern m "either q or r") that, again, appropriately interpreted, represented a conclusion which in fact followed from those premises. Using a vertical figure, some cognitive scientists have adopted a three-level framework which incorporates these ideas. The top level, called the semantic level (Pylyshyn 1989) or the knowledge level (Newell et al. 1989) corresponds to folk psychology and the bottom level is a physical or biological level. In between is the symbol level, where a symbol level system is one that "operates in terms of representations and information-processing operations on these representations"
Introduction
xiii
(Newell et al. op. cit.: 96). The term "classical cognitive science" has been used to refer to this framework. It can be useful to view approaches to cognition and mental representation in terms of their location between the physical or biological level and the semantic or knowledge level. Language of thought theories like Jerry Fodor's and indeed the symbol level of the classical approach can be seen as close to the folk psychology level. Connectionist or neural network approaches and autonomous agent robotics and dynamical systems theory can be seen as much closer to the biological or physical level. In the epigraph for this introduction, David Lewis (1996) says that folk psychology, contrary to some language of thought theorists, leaves open the way mental representation works. The ideas of the classical approach have been influential, but also subjected to intense criticism from many directions over many years. An interesting case is Daniel Dennett. He offers a three-level framework that is similar to the classical one. His hierarchy of stances one can adopt for explaining and predicting the behaviour of complex systems includes the intentional stance (roughly, folk psychology), the design stance and the physical stance. He takes there to be only a difference in emphasis (Dennett 1998: 253) between his approach and the classic approach, but for our purposes here a key difference is his reluctance to identify the design stance with the symbol level of classical cognitive science the classical level of mental representation. Recently (Dennett 2000: 372) he has expressed his position as a rejection of p r e m a t u r e representationalism: 'Early' representations - - posited, for instance, in a n i m a l cognition and in most human perception ~ is as dangerous an abuse as underage drinking. Representations are, in effect, an adult preoccupation into which one is only gradually initiated. An example of what he has in mind is mental arithmetic. "If I do long division in my head, there are representations of the numbers just as surely as there are if I do it on the blackboard" (op. cit.: 374-375). It is time now to turn to what is in this volume, for it contains papers that deal in new ways with some fundamental questions concerning mental representation: What is the relation between mental representation and the more publicly familiar non-mental representation (e.g. the tripartite scheme as studied by semioticians)? How is misrepresentation to be accommodated in a theory of mental representation? Should error be system detectable?
xiv
Introduction
What role do mental representations play in the production of (intelligent) action? If they play a role in making a system's behaviour appropriate to its environment, what would mental representations have to be like to play this role? How can problems with classical cognitive science models (as exemplified in the frame problem) be circumvented? What is the relation between representation and experience? This is of particular current interest given the resurgence of interest in the problem of consciousness. What is the relation between common-sense explanations (folk psychology) of behaviour and thought and explanations invoking mental representations? The papers in this volume abound in new ideas, fresh approaches and new criticisms of old ideas. With one or two exceptions they are arranged from the general to the particular: starting with papers that deal with more general aspects of mental representation and papers which survey the field of current and older theories when offering new approaches to mental representation and finishing with papers that focus on relatively specific topics like the role of representation in memory and the nature of causal representations. Most of the papers are organised into pairs on similar topics, for example: The papers by Gerard O'Brien and Jon Opie and by Robert Cummins and Pierre Poirier both discuss structuralist representations. The former paper argues for a framework that is used and elaborated by the latter paper. In the second paper Cummins and Poirier usefully distinguish indication from representation and contrast the two. The papers by Cliff Hooker and Wayne Christensen and by Mark B ickhard both survey existing theories and share an interest in representations explaining the interaction between systems and their environments. Frank Jackson and Ian Gold both address, in related ways, issues concerned with representation and experience. Gold brings new psychological research to bear on a long-standing philosophical problem. The volume concludes with two papers on narrower topics: John Sutton on the different treatments of memory in different disciplines and ways of unifying them and the final paper in which Patricia Kitcher brings her substantial Kantian scholarship to bear on causal representations. In between these pairs are important papers on how to think about folk psychology by Peter Godfrey-Smith, on misunderstanding semantics in mental
Introduction
xv
representation by Peter Slezak and a paper advocating cybernetic governance as the source of the natural normativity of representation by Timothy Schroeder. What follows is an introduction to some of the main ideas of this volume's papers. Gerard O'Brien and Jon Opie's paper, "Notes Toward a Structuralist Theory of Mental Representation" is a good first paper for this book. In arguing for their structuralist theory of mental representation they provide a useful survey of alternative approaches and isolate some important foundational principles. I will introduce it in more detail than those to follow. Remarkably, they argue that causal, functional and teleosemantic theories fail a natural causal constraint on mental representation, namely, that an explanation of mental representation should be consistent with its causal role in the production of "appropriate behaviour." Following Barbara Von Eckardt (1993), O'Brien and Opie explore the consequences for a theory of mental representation of Peirce's tripartite analysis of non-mental representation. Von Eckardt distinguishes the representation vehicle, the represented object and the "interpretation" as O'Brien and Opie call it. The interpretation is the cognitive effect of the vehicle on the representation user. An example of this would be the sign user thinking of the represented object. It is this last element of the analysis that is particularly problematic when one applies it to mental representation. O'Brien and Opie accept Von Eckardt's conclusion that the only adequate noncircular account of interpretation takes it to involve the modification of behavioural dispositions towards the represented object. And they follow her in taking the triadic relation of representations to be analysable for these purposes into two binary relations - - a content grounding relation between vehicle and object and a relation between the vehicle and interpretation. As O'Brien and Opie say, the important question becomes: "What grounding relations might obtain between mental representing vehicles and their represented objects such that the former are capable of disposing cognitive subjects to behave appropriately towards the latter?" Following Peirce and Von Eckardt they distinguish three kinds of grounding relations - - resemblance, causation and convention. They dismiss convention as not providing an objective grounding r e l a t i o n - content is only conferred on a vehicle by being interpreted. They argue that their causal constraint knocks this out. Briefly put, if content is constituted by its role in modifying behavioural dispositions it cannot be causally responsible for them. Causal grounding relations fall to the causal constraint in a different but related way. The argument this time is that the objective relation between vehicle and object has no effect on the intrinsic properties of the vehicle and so cannot affect the vehicle's causal powers - - in particular its capacity to effect interpretations. O'Brien and Opie don't argue in detail that teleosemantic theories are eliminated by this but suggest that because these theories seem to use a combination of
xvi
Introduction
causal and conventional grounds they too are eliminated. This leaves them with theories based on resemblance. They quickly rule out "physical or first-order" resemblance the sharing of physical properties on naturalistic grounds. " . . . (O)ur minds are capable of representing features of the world that are not replicable in neural tissue." Their favoured grounding relation is second-order resemblance. This is a relation between systems of vehicles and systems of objects in which relations between the vehicles mirror relations between their objects. The two systems resemble each other m they share the same relational organisation. O'Brien and Opie point out that causal, conceptual and functional role semantics take advantage of second-order resemblance. They argue that despite relying on a second-order resemblance, these theories nevertheless also fail the causal constraint for much the same reason as causal grounding theories do - - they reduce representation to interpretation. " . . . (M)ental representing vehicles don't possess the capacity to enter into causal relations (and thereby affect behaviour) in virtue of this second-order resemblance relation; rather the resemblance relation itself obtains in virtue of the causal roles those vehicles occupy in the cognitive economy." Their final proposal is offered conjecturally. They define structural resemblance as second-order resemblance based on the physical relations among a set of representing vehicles. The physical relations among the representing vehicles mirror the relations of whatever kind among the represented objects. Their proposal is that structural resemblance is the ground of mental representation m "it is a relation of structural resemblance between mental representing vehicles and their objects that disposes cognitive subjects to behave appropriately towards the latter." Unlike other approaches, they argue, their proposal satisfies the causal constraint. "The physical relations among representing vehicles are independent of their causal relations, and hence of the use to which they are put." Robert Cummins and Pierre Poirier's paper "Representation and Indication" is a natural companion to O'Brien and Opie's. They have similar views on the nature of representation ~ both holding that it is a kind of second-order resemblance. But where O'Brien and Opie are focused on arguing for this particular view, Cummins and Poirier's paper explores and develops the framework into which the view of representation fits ~ a framework proposed in Cummins' (1996) monograph "Representation, Targets and Attitudes." The paper contrasts two kinds of mental content which are not always distinguished m indication and representation and discusses their roles in the explanation of cognitive phenomena. Indicators are detectors, mechanisms whose function is to detect the presence of some target. Some examples are thermostats, edge detectors in the primary
Introduction
xvii
visual cortex and "idiot lights" m the indicator lights in cars for overheating, low oil pressure etc. The "detector signal" is the state or process that signals that the target has been detected, e.g. the patterns of electrical spikes emitted by edge detector cells in the primary visual cortex. They take mental representations to be elements in a "scheme of semantically individuated types whose tokens are manipulated m structurally transformed m by (perhaps computational) mental processes." As they later remark, it is the susceptibility of mental representations to disciplined structural transformation that earns their keep. As the authors note, indication is commonly taken to be a kind of representation but they think this is a mistake. They draw attention to three differences: indication is transitive, and (consequently) arbitrary and source dependent. They summarize the resulting contrasts as follows" 'Indicator signals say "My target is here," while representations say, "My target, wherever it is, is structured like so." After drawing this central distinction the paper uses it to clarify discussion of a number of topics. The first is an illuminating assessment of the rationale for and limitations of a language of thought. One consequence they draw is that language of thought systems will have no plausible neural implementation. Another is that they can represent propositions and nothing else. The next section of the paper discusses visual representations. It provides a sketch, using some recent visual research, of how visual images can be constructed from indicators like the Hubel and Wiesel edge detectors to form representations sharing structure with their targets. The last two sections discuss representations in relation to targets and accuracy. The paper finishes with a discussion of some of the differences between sentential and pictorial representation. The next two papers also have a particular focus on the role of cognition in action and their authors share some basic assumptions about the best way to think about cognition. It is easy to think of representation as modelled on reference and hence too readily think of it (almost exclusively) as a relation between internal states and external objects. These two papers argue instead that whatever representation is it should be seen as contributing to an adaptive (interactive) relation between a system and its environment. Wayne Christiansen and Cliff Hooker's paper "Representation and the Meaning of Life" begins with a valuable comparison of cognitivism (as they call the classical approach) with dynamical systems theory and autonomous agent robotics, which they think present a strong challenge to cognitivism. A recurring theme through their paper is the problem of determining the question for which representation is the answer. The brief description that follows just picks a few points from their nuanced discussion of the issues.
xviii
Introduction
Christiansen and Hooker characterise cognitivism as providing a model of the mind as essentially disembodied leaving no important role for physical processes: representation --+ semantic inference --+ action One of their conclusions is that unlike cognitivism which does provide a paradigm for research which explains epistemic capacity, rationality and adaptive capacity, dynamical systems theory is at present just a general modelling framework. One of their criticisms of the latter is that it tends to downplay the role of the internal organisation of cognitive systems. The autonomous agent robotics research of Brooks (1991) while delivering robots capable of simple physical interactions, is seen as failing to scale up to human-like interactions. Although critical of aspects of dynamical systems theory and autonomous agent robotics (which they collectively call dynamically situated models) they think that cognitivism faces fundamental problems. They mention three within its own domain: symbol grounding, the frame problem and the combinatorial explosion problem. Their paper provides a helpful analysis of the frame problem. They argue "... the frame problem cuts to the heart of the cognitivist account of high order cognition because an adequate theory of cognition ought to explain the ability of intelligent agents to determine relevancy, generalise and learn." They briefly sketch their own approach to cognitive intelligence called "selfdirectedness" reflecting a kind of "internally directed adaptive organisation" and contrasted with simple reactiveness. As does Mark Bickhard in the next chapter, Christiansen and Hooker draw attention to the importance of flexible control of behaviour in a variety of environmental conditions. Rather than recommending, as some have, a hybrid dynamical situated/cognitivist approach to the study of intelligent agents and conceding that at present the dynamical situated approach does not provide a paradigm, they do in fact favour the dynamical approach. They recommend "reconstructing" cognitive concepts like intentionality and representation in a dynamical situated context. Instead of seeing representations as "things about things" their recommendations include broadening the traditional focus to include "downstream modulation of action and interactive differentiation of adaptive relations," giving a process-oriented account of content. This account satisfies a specific requirement on content, also recommended by Bickhard: content should be accessible to the system in the sense that in principle misrepresentation by the system should be discoverable by it. Process-orientation is a theme taken up in the next related paper in this volume by Mark B i c k h a r d - "The Dynamic Emergence of Representation." As his first three section headings reveal, Bickhard's paper is wide ranging: emergence, normative emergence and representation. His main concern is to describe and
Introduction xix advocate a theory of emergent mental representation that he calls "interactive representation." The first two sections of this paper can be seen as clearing the way for the positive account of representation. They defend the possibility of the view of representation that he presents in the third section. He begins by observing that representation is an e m e r g e n t - there was a time before which there was no representation. He rejects a much discussed argument of Kim's concerning the impossibility of causal emergence. Although he considers it to be valid he nevertheless rejects it as unsound, finding fault with its assumption that causal power resides in particles. Taking the only alternative to a metaphysics of particles to be a process metaphysics he has an argument from the evident fact of emergence to this metaphysics. In the brief second section he argues that those who hold that an "ought" cannot be deduced from purely factual premises are mistaken. This opens up the possibility that norms relevant to representation can be deduced from facts. Section three begins the account of r e p r e s e n t a t i o n - an "interactivist" view. Bickhard says "Interactivism models representation as emergent in a particular kind of biological function" but although this gives a good idea of the source of the idea, the analysis is not limited to functions that are biological. It applies to any "self-maintenant" s y s t e m - systems which contribute to the maintenance of far-from-equilibrium stability by interacting with their environments. Something serves a function for such a system if it helps maintain the far-from-equilibrium stability. This is a causal notion. Bickhard proposes three requirements on content that serve to distinguish the interactivist view: representational content must be: (1) internally related to representation; (2) functionally accessible to the system; and (3) normative for the system. He argues that if the content of a representation is only externally related to its vehicle, an interpreter will be needed to connect vehicle and content with an ensuing infinite regress of interpreters. Timothy Schroeder's "New Norms for Teleosemantics" shares Bickhard's difficulty with teleosemantic approaches - - that they require whatever has meaning to have a history. He rejects the view that history is important for mindedness. Where teleosemantics takes natural selection as the source of natural norms Schroeder proposes taking cybernetic governance as the source. His account of representation is normative. Roughly, the idea is that "to be a mental representation is to be a structure which (to a first approximation) is supposed to be A when and only when the world is B." His suggestion is that the norms come from feedback-driven governance systems which he points out intuitively seem to have some normative significance. Schroeder takes the following account of cybernetic regulatory systems as both necessary and sufficient. They have three features:
xx
Introduction
(1) A capacity to carry the information that a goal state has been attained. (2) A feedback system bringing information about the systems state and output back in as input. (3) A causal dependence between the fedback information and the new performance that minimises the difference between the system's present state and its goal state. The basic proposal is that cybernetic systems create norms for the systems they govern by governing them. Towards the end of his paper Schroeder proposes an important link between cybernetic governance and intentionality. It is his view that if a particular cybernetic governance system "drives the states of one object to correspond to the states of some other object.., it points, aims or directs the first object at the second" m and hence has the essential characteristics of intentionality. The next two papers address the relation between representation and experience. Frank Jackson's paper, "Representation and Experience," might in several senses be taken to be the centre piece of this volume. In it he significantly changes his influential views on the mind. Recently, in discussing the act-object accounts of sensory experience, Bill Lycan (1996: 52) reflects (on) this influence: . . . suppose that an antimaterialist . . . . says forthrightly that there d a m n well are phenomenal individuals such as sense data m af-
terimages and so forth m that we damn well are acquainted with t h e m . . , and that not just naive intuitions but powerful arguments can be urged in their favor. (Frank Jackson [ 1977] has done just this.) To locate the subject, Jackson defines representationalism briefly as the view that "perceptual experience's representational character exhausts its experiential character." Put simply, "the redness of seeing red comes from its representing that there is something red." His paper is organised into three parts: (1) the path from sense data to representationalism and why we should take it; (2) Swampman or the externalism issue; and (3) getting "feel" from content. In the first he discusses the relation between sense datum and representational accounts of sensory experience and some arguments for the latter. His diagnosis of the tempting mistake made by sense datum theorists is illuminating: sharing a property (say redness) with what it represents does not, he now notes, give an account of "the representational nature of perceptual experience." He proposes replacing the relation of direct awareness of something
Introduction
xxi
red with a representing that something is red. He sees the main failing of the sense datum theory, not as being its reification of experience, but as being its failure "to address the representational nature of experience.., the most important part of the story." In the second part of the paper he defends, controversially, the existence of truthconditional narrow content. As he says "This means that I can admit the intuition that the sensory side of psychology supervenes on how one is from the skin in." He usefully points out that a state's content may be individuated by external things without this implying that the content does not supervene on my skin-and-inside state. Then he gives a critique of two influential lines of argument for externalism, both concerning a subject's actual or possible causal links to his or her environment. One of the nice ideas here is that in perception part of the perceptual content is that one is causally interacting with the environment. As he says "My perceptual experience is not merely that there is a biro in front of me; it is that there is a certain biro causally interacting with me in a highly distinctive way." The third section of his paper addresses the problem of getting the "feel" from the content. Here he refines the simple account of representationalism and notes that, for example, the redness of seeing red "had better be the representing that there is something red as a part of representing that things are such and such, for some very rich such and such." In addition to the causal feature mentioned above, he requires that the representation be direct and include second-order representations. The paper concludes with a rejection of the need for some special non-conceptual content as a source of the "feel." Like Frank Jackson, Ian Gold addresses the relation between perceptual experience and representation. The central concern of his paper "Phenomenal Qualities and Intermodal Perception" is to argue against a simple empiricist view of perception, a view called "phenomenism." He does this, interestingly, from empirical premises drawn from work in child psychology. Phenomenism is the view that "once the representational content of a mental state is set aside there is a remainder which is purely phenomenal." Phenomenal qualifies, while contributing to the state's representational function also "contribute a subjective feeling, a 'raw feel,' 'quale,' 'what it's like,' or what have you m to the state." For Gold's purposes two theses of the phenomenist are important: (1) Accessible representational content supervenes on phenomenal qualities. (2) Phenomenal qualities are modality specific. Now, for example, the shape of an object may be both seen and felt. But by (2) the phenomenal qualities experienced during touch and vision will be distinct.
xxii
Introduction
A question that arises for phenomenists - - Molyneaux's question w is whether someone blind from birth who later recovered his or her sight would, prior to any learning, be able to identify something as round without feeling it. Not according to phenomenists. Gold argues that while psychological evidence for the answer to this specific question is ambiguous, other research results, on neonates, do count against the phenomenist. He concludes the paper by canvassing some of the options for this approach. Peter Godfrey-Smith's paper "On Folk Psychology and Mental Representation," begins with a rather pessimistic assessment of the success of and prospects for a program of giving a theory of the physical or biological properties of the internal states of organisms that suffice to make these states representations of the world beyond them. The main concern of his paper, however, is the question of how best to think about folk psychology. The key question is: Does folk psychology make any claims about what is going on in the head? His central contribution is that we should think of folk psychology as a model. Put simply, there is a long-running and still unresolved dispute between intentional realists and interpretationists. Very roughly it's between those, like Fodor and Dretske, who think that belief/desire talk refers to internal causally effective structures and those like Dennett and Davidson who see this sort of talk as lacking this commitment. Godfrey-Smith steps back from the dispute between these people and notes that, however that goes, there are two sets of fact to be accounted for: (1) wiring (and organization) and connection (to the world) facts; (2) facts about our practices of interpretation and ascriptions of content (interpretation facts). Godfrey-Smith notes that there are different uses of folk psychology w "law court vs h i g h w a y " - - and that the former attributes more structure than the latter. He proposes to treat folk psychology as a model. In contrast with a theory, a model is a conceptual structure with a continuum of possibilities of which parts of the model are taken to have realistic counterparts in the world. A natural companion to Peter Godfrey-Smith's discussion of folk psychology is Peter Slezak's multi-faceted paper "The World Gone Wrong? Images, Illusions, Mistakes and Misrepresentations." He proposes the intriguing and challenging idea that talk of misrepresentation commits the homunculus fallacy. Misrepresentation, he holds, cannot be coherently articulated without invoking the very semantic problem to be explained m that is, philosophers commit the homunculus fallacy in formulating the problem of misrepresentation. A number of apparently
Introduction
xxiii
independent problems, including Gettier examples and Putnam's twin earth, are seen to have a common source. Here is a crucial argument: " . . . the very distinction between true and false ideas cannot be made without comparing representations and the world. Of course, this perspective is unavailable to the mind itself. Correspondingly, an explanatory theory cannot make tacit appeal to such a perspective without committing the homunculus error." Consequently Slezak rejects the conception of semantically evaluable representations intervening between mind and world. But he observes, this does not amount to abandoning an account of the content of mental representations. What he rejects is any need to "tie representations to the world" in certain ways. Slezak is centrally concerned with the distinction between understanding mental representations and explaining them. Only the latter is appropriate for cognitive science. He observes that enquiring after the truth value of a representation presupposes understanding its meaning. As he notes: "this is the crucial slip from the respectable explanatory concept of intentionality to the spurious observer-relative sense of meaningfulness. It should be clear that the role of mental symbols and our scientific concern to explain this must be independent of whether or not we are able to understand the symbols as meaningful to us." The last two papers in this volume have a narrower focus than the others, although not necessarily a narrower range. The first is on memory and the second on causation. John Sutton, in his "Representation, Reduction and Interdisciplinarity in the Science of Memory" sketches an integrated framework to house many of the different approaches to memory. Representation is after all a key concept in the study of memory. Following an introductory section, the paper is organised into three main sections. The focus of the first part of his paper is metatheoretic, centering on the possibilities of reduction. The last two sections argue against an individualist approach to memory and point to work that links some individualist ideas with the study of external memory in society. Sutton begins by noting the diversity of the disciplines that study memory disciplines ranging from neuroscience through other cognitive sciences and beyond to social sciences and elsewhere. This sets the scene for a discussion of how these differing approaches can be related and the question: what ought we to expect from interdisciplinary theory construction? Are the different disciplines even studying the same thing? Reduction is concerned with the interdependence of different theories of what can loosely be called the same phenomena and Sutton, in distinguishing different kinds of reduction, is concerned not to be seen as imposing an "artificial unity"
xxiv
Introduction
on complex and diverse phenomena. In particular he wants to reject the view that reduction implies all higher level terms and concepts will be eliminated from lower level reducing explanations and more particularly, "that ultimately only the vocabulary of fundamental physics will be legitimate." In the next section on "Constructive Remembering: source memory and development" he argues that the study of memory requires consideration of extra-individual factors, arguing for the relevance of social psychology to cognitive science. Empirical work reveals the potential for serious distortion in memory. Sutton holds, not that veridical remembering is unlikely, but that a study of the mechanisms of distortion will throw light on it. He reports work on source monitoring the process of recalling how, when and where information was acquired. This work notes that memories do not generally come with identification of their sources attached and argues that social processes play an important role in sourcing memory. In addition, there is evidence that childhood memories can be strongly influenced by shared reminiscing. He concludes "Social norms and cultural narratives are not simply downloaded into the mind, and yet each mind is more intimately tangled with such norms and narratives than traditional individualistic cognitive science has often allowed." In the final section, via a consideration of collective and external memory representations, he reaches for a general framework to connect these with more familiar information processing. This framework could be described as combining connectionism, or distributed cognition, with the "extended m i n d " - - in which "quite disparate internal and external elements are simultaneously co-opted into integrated larger cognitive systems." In the final paper, "Kant on Constructing Causal Representations," Patricia Kitcher presents and critically examines Patricia Cheng's project of showing how a priori knowledge that all events have causes can permit the deduction of causal laws from limited observation. She criticises it on three (mostly) Kantian counts: firstly on Cheng's definition of cause, secondly on her account of causal priority and thirdly on her treatment of the computational problem of causal induction. In the first two cases she proposes modifications that rescue the theory from her criticisms. Her first criticism of Cheng's account is that the notion she uses of causal power lacks clarity and would be rejected by both Hume and Kant. Kitcher suggests that Cheng should replace the notion of powers with that of rules. Rather than cognizers assuming the existence of causal powers the suggestion is that they assume as Kant put it (A 189), that "everything that happens (begins to be) presupposes something from which it follows in accordance with a rule." Secondly, Kitcher argues that Cheng is wrong to claim that because causes must exist before they can produce any consequences, causes precede effects. Kant held
Introduction
xxv
that cause and effect are simultaneous. This raises the question: how then can Kant hold that one's appreciation of cause and effect guides judgments of temporal order? She accepts the suggestion that this was because Kant was operating with a three place causal relation, relating a cause to the old and the new states of its effects. Kitcher argues that this analysis provides reason to conclude that no cause can come after its effect. Her third criticism is of Cheng's analysis of the "computational" problem of causal induction. This is the problem of specifying the pairings of input and output in causal induction. Cheng uses perceived priority as a criterion for identifying potential causes and Kitcher raises the Kantian objection that causal interpretation is already required for temporal ordering, defending it against the criticism that it is circular. Her proposed amendment for Cheng is to allow for the possibility that in some cases our views of temporal and causal relations may be determined together. Although this introduction has presented only some of the central ideas and arguments from the papers in this book, it is already clear, I think, that these eleven papers on mental representation cover a surprising amount of ground. The reader who engages with any or all of the papers in this volume will be rewarded with a view of how much else there is.*
References Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47, 139-159. Cummins, R. (1996). Representations, targets and attitudes. MIT Press. Cummins, R. (2000). How does it work? vs. What are the laws?: Two conceptions of psychological explanation. In: E C. Kiel, & A. W. Wilson (Eds), Explanation and cognition (pp. 117-144). MIT Press. Dennett, D. (1998). Brainchildren: Essays on designing minds. London: Penguin Books. Dennett, D. (2000). With a little help from my friends. In: D. Ross, A. Brook, & D. Thompson (Eds), Dennett's philosophy: A comprehensive assessment (pp. 327-388). MIT Press. Field, H. (1978). Mental representation. Erkenntnis, 13, 9-61. Leibniz, G. (1714). The monadology. In: R. Ariew, & D. Garber (Transl.), Leibniz: Basic works (1989). Indianapolis: Hackett. Lewis, D. (1996). Reduction of mind. In: S. Guttenplan (Ed.), A companion to the philosophy of mind (pp. 421-423). Oxford: Blackwell. Lycan, W. (1996). Consciousness and experience. MIT Press. *Thanks to Peter Slezak for helpful comments on this introduction.
xxvi
Introduction
Newell, A., Rosenbloom, E S., & Laird, J. E. (1989). Symbolic architectures for cognition. In: M. I. Posner (Ed.), Foundations of cognitive science (pp. 93-131). MIT Press. Pylyshyn, Z. (1989). Computing in cognitive science. In: M. I. Posner (Ed.), Foundations of cognitive science (pp. 49-91). MIT Press. Sellars, W. (1953). A semantical solution to the mind-body problem. In: J. Sicha (Ed.), Pure pragmatics and possible worlds: The early essays of wilfrid sellars (1980). Resada, CA: Ridgeview Publishing. Von Eckardt, B. (1993). What is cognitive science ? MIT Press. Phillip Staines
Chapter 1
Notes Toward a Structuralist Theory of Mental Representation Gerard O'Brien and Jon Opie
Introduction Any creature that must move around in its environment to find nutrients and mates, in order to survive and reproduce, faces the problem of sensorimotor control. A solution to this problem requires an on-board control mechanism that can shape the creature's behaviour so as to render it "appropriate" to the conditions that obtain. There are at least three ways in which such a control mechanism can work, and Nature has exploited them all. The first and most basic way is for a creature to bump into the things in its environment, and then, depending on what has been encountered, seek to modify its behaviour accordingly. Such an approach is risky, however, since some things in the environment are distinctly unfriendly. A second and better way, therefore, is for a creature to exploit ambient forms of energy that carry information about the distal structure of the environment. This is an improvement on the first method since it enables the creature to respond to the surroundings without actually bumping into anything. Nonetheless, this second method also has its limitations, one of which is that the information conveyed by such ambient energy is often impoverished, ambiguous and intermittent. Once the trick of exploiting the information carried by ambient forms of energy has been mastered, however, a third kind of control mechanism becomes available. Instead of responding directly to the information impacting on its sensory surfaces, a creature can use it to construct internal models of the environment m on-board states that "stand in for" or "represent" external objects, relations and
Representation in Mind: New Approaches to Mental Representation Edited by H. Clapin, P. Staines and P. Slezak Copyright 9 2004 by Elsevier Ltd. All rights of reproduction in any form reserved. ISBN: 0-08-044394-X
2
Gerard O'Brien and J. Opie
states of affairs. These internal representations, rather than the information-laden signals from which they were constructed, can then be pressed into service to shape behaviour. Such a decoupling of behaviour from direct environmental control confers great benefits, since internal models of the environment can have a stability and definiteness that is lacking in the signals that impact on a creature's sensory surfaces. They also enable a creature to respond to features of the world that are not immediately present, to use past experiences to shape present behaviour, to plan for the future, and, in creatures such as ourselves, to be sensitive to very abstract features of the world. The appearance of creatures with the capacity to internally represent their environment thus constitutes a very significant branch-point in evolutionary history. It amounts to nothing less than the emergence of minds on this planet. Minds are Nature's most sophisticated solution to the problem of sensorimotor control: neural mechanisms that shape behaviour by constructing and processing representations of the environment and the body. One of the first tasks confronting a science of the mind is to explain how nervous systems can be in the representing business in the first place m how brain states can be about aspects of the world. It is a commonplace in the philosophy of mind that a theory of mental representation must be naturalistic, in the sense that it must explain mental representation without appealing to properties that are either non-physical or antecedently representational. 1 What is not quite so commonplace, however, is the further injunction that such a theory must explain mental representation in a fashion consistent with its causal role in shaping appropriate behaviour. Yet this is precisely the message that issues from the simple evolutionary tale we have just told. We shall call this the causal constraint on a theory of mental representation. Remarkably, even though the well-known causal, functional, and teleosemantic theories of mental representation are all naturalistic, they all violate the causal constraint. Despite their internal differences, these theories ultimately treat mental content in terms of the appropriate behaviour that cognitive subjects are capable of exhibiting towards their environments. 2 And any theory that treats mental
i See, e.g. Cummins (1989: 127-129, 1996: 3-4), Dretske (1981: xi), Field (1978: 78), Fodor (1987: 97-98), Lloyd (1989: 19-20), Millikan (1984: 87), and von Eckardt (1993: 234-239). 2 This is a slightly different way of expressing one of the main conclusions Robert Cummins draws in his book Representations, Targets, and Attitudes (1996). Cummins convincingly argues that all of the main contenders in the contemporaryphilosophy of mind are what he calls "use" theories, in that they ultimately attempt to explain mental content in terms of the use to which representations are put by a cognitive system. Our characterisation follows from Cummins', once one observes that the use of a representationby a cognitive system is ultimately to be unpacked in terms of the role it plays in causing the system to behave appropriately.
Notes Toward a Structuralist Theory of Mental Representation
3
c o n t e n t in terms of intelligent b e h a v i o u r is in principle u n a b l e to explain h o w m e n t a l r e p r e s e n t a t i o n is causally r e s p o n s i b l e for such behaviour. 3 T h e w o r r y e x p r e s s e d in the previous p a r a g r a p h is not new, of course. It represents just one (admittedly, s o m e w h a t u n u s u a l ) w a y of f o r m u l a t i n g the m u c h d e b a t e d p r o b l e m of mental causation the p r o b l e m of e x p l a i n i n g h o w the specifically representational properties of brain states can be causally potent over and above their physical properties. 4 T h e standard r e s p o n s e to this p r o b l e m in the p h i l o s o p h y of m i n d is to accept that r e p r e s e n t a t i o n a l properties are causally inert, but to argue that there is e n o u g h r o o m b e t w e e n e x p l a n a t i o n and causation for r e p r e s e n t a t i o n a l properties to be explanatorily relevant despite their inertness. 5 In their heart of hearts, however, m o s t p h i l o s o p h e r s k n o w that this r e s p o n s e is d e e p l y unsatisfactory. O u r aim, in adopting the p r e s e n t f o r m u l a t i o n , is to avoid further fancy f o o t w o r k on the differences b e t w e e n e x p l a n a t i o n and causation. W h a t is really n e e d e d in the p h i l o s o p h y of m i n d is a c o m p l e t e l y different a p p r o a c h to m e n t a l r e p r e s e n t a t i o n m one that d o e s n ' t violate the causal constraint, and h e n c e one for w h i c h the p r o b l e m of m e n t a l c a u s a t i o n d o e s n ' t even arise. T h e a m b i t i o u s task we u n d e r t a k e in this paper is to sketch the outlines of a naturalistic t h e o r y of m e n t a l r e p r e s e n t a t i o n that is consistent with the simple e v o l u t i o n a r y story told above. We call this a structuralist theory of mental representation for reasons that will b e c o m e apparent as w e proceed.
3 Cummins' way of putting this point is to say that use theories cannot account for the explanatory appeals that cognitive science makes to mental representation (1996, especially: 47-51). The best way to see this, he thinks, is by observing that use theories cannot do justice to misrepresentation.According to Cummins, representational error occurs when a cognitive system uses a representation incorrectly. But this requires a robust distinction between how a representation is used, and what it means. Since use theories explicitly deny this distinction, they undermine the notion of representational error and with it the explanatory importance of representation (1996: 47). Again, however, once one notes that cognitive science invokes misrepresentation (representational error) in order to account for misbehaviour (inappropriate behaviour) it can be seen that, at base, it is their violation of the causal constraint that renders use theories incompatible with the explanatory appeals that cognitive science makes to mental representation. 4 In the language made familiar by this debate, the worry is that by explaining mental representation in terms of the intelligent behaviour that cognitive creatures are capable of exhibiting, all of the currently fashionable theories entail that the representational properties of brain states fail to supervene on their intrinsic physical properties. This failure entails in turn that the representational properties of brain states do not determine their causal powers. 5 See, e.g. Baker (1993), Block (1989), Dretske (1987, 1988, 1990), Fodor (1986, 1989), Heil & Mele (1991), Jackson & Pettit (1990a, b), and LePore & Loewer (1989). Even Cummins was tempted to develop this kind of response in his earlier work (see, e.g. 1989: 129-136).
4
Gerard O'Brien and J. Opie
Mental Representation" A Triadic Analysis What we are all after is a naturalistic theory of mental representation, one that explains mental representation without recourse to the non-physical or antecedently representational. To this requirement we've just added another: such a theory must explain mental representation in a way that is consistent with its causal role in shaping appropriate behaviour. Where are we to find a theory that satisfies these twin demands? Perhaps we can make some headway by examining representation as it exists in those public objects words, sentences, paintings, photographs, sculptures, maps, and so forth m with which we are all familiar. By investigating how such public objects operate as representations, we may gain some insight into the nature of mental representation. This is a strategy very effectively deployed by Barbara von Eckardt in her book What is Cognitive Science? (1993). Adapting the work of Charles Sanders Peirce, von Eckardt analyses non-mental representation as a triadic relation involving a representing vehicle, a represented object and an interpretation (1993: 145-149). 6 The representing vehicle is the physical object (e.g. spoken or written word, painting, map, sculpture, etc.) that is about something. The represented object is the object, property, relation or state of affairs the vehicle is about. And the interpretation is the cognitive effect in the subject for whom the vehicle operates as a representation, such that this subject is brought into some appropriate relationship to the vehicle's object. Usually this cognitive effect is understood in terms of the subject thinking about the object in question. What happens when we apply this triadic analysis to the special, and presumably foundational, case of mental representation? Given our commitment to naturalism, and hence to a rejection of non-physical properties, the vehicles of mental representation must be understood as brain states of some kind. As for the represented objects, the same analysis we considered above applies: these are the objects, properties, relations, and states of affairs that mental vehicles are about. But the story about the third relatum namely, interpretation m is different in the case of mental representation. If we apply the account we adopted in the non-mental case, and treat interpretation in terms of a cognitive subject thinking about a represented object, we violate the naturalism constraint by
von Eckardt actually uses the terms representation bearer, representational object and interpretant to describe the three relata implicated in representation. We prefer our terminology because it is more consistent with the literature on mental representation and with terminology we have employed elsewhere (e.g. see O'Brien & Opie 1999). 6
Notes Toward a Structuralist Theory of Mental Representation
5
interjecting a state that is already representational. 7 What we require, therefore, is another account of interpretation one that doesn't appeal to further mental representation. This is a complex matter, and one that we can't do justice to here (see von Eckardt 1 9 9 3 : 2 8 1 - 3 0 2 for a much fuller discussion). But, to cut a long story short, the only account that would seem to be available is one that treats interpretation in terms of the modification of a cognitive subject's behavioural dispositions. This acts to block the threatened regress since, presumably, it is possible to unpack such behavioural dispositions without invoking further mental representing vehicles. Not any old dispositions will do, however. The process of interpretation, remember, must bring the subject into some appropriate relationship with the represented object. Consequently, interpretation must modify a subject's behavioural dispositions towards the vehicle's represented object. The trouble with this more naturalistic account of interpretation is that it seems to make it impossible to reconcile the triadic analysis of mental representation with the causal constraint. If mental representation incorporates interpretation, and if interpretation concerns modifications to a cognitive subject's behavioural dispositions, then mental representation isn't independent of the subject's behaviour (appropriate or otherwise) and hence can't be causally responsible for it. All is not lost, however, von Eckardt observes that the triadicity of representation in general, and mental representation in particular, can be analysed into two dyadic component relations: one between representing vehicle and represented object (which she calls the content grounding relation); the other between vehicle and interpretation (1993: 149-158). 8 This suggests that any theory of mental representation must be made up of (at least) two parts: one that explains how the content of mental vehicles is grounded, and a second that explains how they are interpreted. 9 The distinction between mental content and mental interpretation is important because it shows us how the triadicity of mental representation can be rendered compatible with the causal constraint. The crucial point is this. A theory
7 This is why mental representation is foundational: any explanation of non-mental representation must ultimately appeal to mental representation to account for interpretation. A theory of representation in general thus waits upon a completed theory of mental representation. 8 von Eckardt admits that some might object to this manoeuvre, on the grounds that if representation is triadic then it can't be properly analysed into two dyadic component relations in this way. She responds by noting that if representation is genuinely triadic then it will turn out that these subrelations are not purely dyadic, and hence it will be impossible to explicate either of them without reference to the third relatum (1993: 149). 9 According to von Eckardt, most of the proposals in the literature that purport to be theories of mental representation are best understood as theories of mental content. Cummins is another theorist who thinks that mental representation decomposes into a number of different elements, each of which requires a distinct theory (1996: 20-21).
6
Gerard O'Brien and J. Opie
of mental representation satisfies the causal constraint as long as it explains how mental content is causally responsible for making a cognitive subject's behaviour appropriate to its environment. And a theory of mental representation can do this if it holds that mental representing vehicles possess the capacity to effect mental interpretations in virtue of the grounding relations they bear to their representing objects. Putting all this together, the important question would seem to be: What grounding relations might obtain between mental representing vehicles and their represented objects such that the former are capable of disposing cognitive subjects to behave appropriately towards the latter? Again taking her cue from the case of non-mental representation, von Eckardt observes that when we consider the many forms of public representation with which we are familiar, there would seem to be three types of ground: resemblance (e.g. a portrait represents a person in virtue of resembling them), causation (e.g. a knock at the door represents the presence of someone in virtue of a causal relation between them), and convention (e.g. the word "cat" represents the property of being a cat in virtue of a convention of the English language). According to von Eckardt, convention is an inappropriate ground for mental representation, since it violates the naturalism constraint (1993: 206). She never makes explicit her reason for drawing this conclusion, however. If a vehicle is related to its object by convention, the cognitive subject must deploy a rule that specifies how the vehicle is to be interpreted. In the case of non-mental representation, where for example the vehicle is a word in a natural language, the application of such a rule is a cognitive achievement that must be explained in terms of processes defined over mental representing vehicles. Perhaps this is why von Eckardt thinks that convention cannot be a naturalistic ground of mental representation. However, as Daniel Dennett has been fond of reminding us over the years, at least some of the rules that govern the behaviour of a cognitive system must be deployable without implicating further representation, on pain of an infinite regress. 10
l0 Dennett attributes this idea to Ryle: This is what Ryle was getting at when he claimed that explicitly proving things (on blackboards and so forth) depended on the agent's having a lot of knowhow, which could not itself be explained in terms of the explicit representation in the agent of any rules or recipes, because to be able to manipulate those rules and recipes there has to be an inner agent with the knowhow to handle those explicit items - - and that would lead to an infinite regress. At the bottom, Ryle saw, there has to be a system that merely has the knowhow .... The knowhow has to be built into the system in some fashion that does not require it to be represented (explicitly) in the system (Dennett 1982: 218). See also Dennett (1978:119-126).
Notes Toward a Structuralist Theory of Mental Representation
7
A n d c o m p u t e r science has even s h o w n us h o w this can be done: a c o m p u t a t i o n a l device can tacitly e m b o d y a set of primitive instructions in virtue of the w a y it is constructed, l I So it is not o b v i o u s that c o n v e n t i o n does fail the test of naturalism. 12 E v e n so, it is o b v i o u s that c o n v e n t i o n violates the causal constraint. Unlike r e s e m b l a n c e and causation, the existence of a c o n v e n t i o n a l g r o u n d d o e s n ' t entail an objective relation b e t w e e n a vehicle and its r e p r e s e n t e d object (see von E ckardt 1993: 149). This c o n n e c t i o n is forged instead by a rule that specifies h o w the vehicle is to be interpreted. In other words, c o n v e n t i o n places the w h o l e b u r d e n of r e p r e s e n t a t i o n on the shoulders of interpretation: it is the p r o c e s s of being interpreted that confers content on a m e n t a l r e p r e s e n t i n g vehicle. But recall that in the case of m e n t a l representation, interpretation is a m a t t e r of modifications to a cognitive subject's b e h a v i o u r a l dispositions. A n y theory of m e n t a l r e p r e s e n t a t i o n that invokes convention, therefore, ultimately treats a m e n t a l v e h i c l e ' s content in terms of the c o g n i s e r ' s b e h a v i o u r a l dispositions towards the r e p r e s e n t e d object. C o n s e q u e n t l y , such a theory is inconsistent with the c l a i m that m e n t a l content is causally r e s p o n s i b l e for such dispositions. 13 This w o u l d a p p e a r to leave us with r e s e m b l a n c e and causat i on as potential g r o u n d s of m e n t a l representation. O f these two, causation has b e e n m o r e p o p u l a r in the recent p h i l o s o p h y of mind, as it forms the f o u n d a t i o n of a n u m b e r of I I The Turing machine can be used to illustrate the point. The causal operation of a Turing machine is entirely determined by the tokens written on the machine's tape together with the configuration of the machine's read/write head. One of the wondrous features of a Turing machine is that computational manipulation rules can be explicitly written down on the machine's tape; this of course is the basis of stored program digital computers and the possibility of a Universal Turing machine (one which can emulate the behaviour of any other Turing machine). But not all of a system's manipulation rules can be explicitly represented in this fashion. At the very least, there must be a set of primitive processes or operations built into the system. These reside in the machine's read/write head n it is so constructed that it behaves as if it were following a set of primitive computational instructions. 12Indeed, Dennett's own theory of mental representation is based on convention n in this case conventions laid down in our brains by our evolutionary history (1987, especially: 287-321). This is why Dennett rejects the distinction between original and derived intentionality: mental phenomena have no greater representational status than the words on this page (1980). Interestingly, von Eckardt doesn't discuss Dennett's proposal. 13Dennett, whose theory of mental representation does invoke convention (see the previous footnote), is quite sanguine about this consequence. For example, he writes: There is a strong but tacit undercurrent of conviction.., to the effect that only by being rendered explicit.., can an item of information play a role. The idea, apparently, is that in order to have an effect, in order to throw its weight around, as it were, an item of information must weigh something, must have a physical embodiment .... I suspect, on the contrary, that this is almost backwards. [Representing vehicles].., are by themselves quite inert as information bearers .... They become information-bearers only when given roles in larger systems (1982: 217).
8
Gerard O'Brien and J. Opie
w e l l - k n o w n and m u c h discussed proposals. 14 Yet despite their popularity, causal theories of mental representation, like their conventional counterparts, fail to satisfy the causal constraint. The p r o b l e m this time is not that the existence of a causal ground fails to entail an objective relation between a representing vehicle and its represented object. The p r o b l e m is that the connection forged by this objective relation has no influence on the vehicle's intrinsic properties, and hence on its causal powers. Consequently, while there might be causal relations b e t w e e n represented objects and representing vehicles, the latter d o n ' t acquire the capacity to bring about interpretations in virtue of these relations. 15 In this respect, vehicles whose content is grounded by causation are in precisely the same position as those grounded by convention, in that the full burden of representation falls on their interpretation. For exactly the same reasons as before, therefore, any theory of mental representation that invokes a causal ground will inevitably treat mental contents in terms of a cognitive subject's behavioural dispositions, and in so doing transgress the causal constraint. 16
Resemblance as the Ground of Mental Representation Neither convention nor causation can satisfy the two constraints on a theory of mental representation that we have set. It remains, then, to consider resemblance. R e s e m b l a n c e would appear to be an appropriate ground of such public representing vehicles as paintings, maps, sculptures and scale models. 17 W h a t we consider in this section is whether r e s e m b l a n c e is an appropriate ground of mental representation.
14See, e.g. Fodor (1987, 1990), Dretske (1981), Stampe (1977, 1986). Teleosemantic theories such as Millikan's (1984, 1989) and the later Dretske's (1987) seem to employ a mixture of causal and conventional grounds. 15For example, the causal relation that obtains between a person and a door knock has no influence on the latter's intrinsic properties, and hence door knocks on their own are quite powerless to effect interpretations. 16For a much more detailed consideration of the failure of causal theories of mental representation in this regard, see Cummins (1996: 53-74). 17 Goodman (1969) developed a famous objection to this approach on the basis that, whereas representation is an asymmetric relation, resemblance is symmetric (a man resembles his portrait to the same extent that the portrait resembles him, for example). Note, however, that the triadic analysis of representation provides the resources to break this symmetry. In the case of a portrait, what picks it out uniquely as the representing vehicle is the interpretation placed on it by an observer. For a fuller discussion of this response to Goodman's objection see Files (1996).
Notes Toward a Structuralist Theory of Mental Representation
9
Resemblance is a fairly unconstrained relationship, because objects or systems of objects can resemble each other in a huge variety of ways, and to various different degrees. However, one might hope to make some progress by starting with simple cases of resemblance, examining their possible significance for mental representation, and then turning to more complex cases. Let us begin, then, with resemblance between concrete objects. 18 The most straightforward kind of resemblance in this case involves the sharing of one or more physical properties. Thus, two objects might be of the same colour, or mass, have the same length, the same density, the same electric charge, or they might be equal along a number of physical dimensions simultaneously. We will call this kind of relationship physical or first-order resemblance. 19 A representing vehicle and its object resemble each other at first order if they share physical properties, that is, if they are equal in some respects. For example, a colour chip a small piece of card coated with coloured ink m is useful to interior designers precisely because it has the same colour as paint that might be used to decorate a room. First-order resemblance is a promising grounding relation because it depends on a representing vehicle's intrinsic properties, unlike convention and causation. Unfortunately, first-order resemblance, while relevant to certain kinds of public representation, is clearly unsuitable as a general ground of mental representation, since it is incompatible with what we know about the brain. Nothing is more obvious than the fact that our minds are capable of representing features of the world that are not replicable in neural tissue. Moreover, even where the properties actually exemplified by neural tissue are concerned, it is most unlikely that these very often play a role in representing those self-same properties in the world. For this reason, philosophers long ago abandoned first-order resemblance as the basis of a theory of mental representation (see, e.g. Cummins 1989:31). But perhaps resemblance is not yet dead in the water. There is a more abstract kind of resemblance available to us. Consider colour chips again. Interior designers typically use sets of chips or colour charts to assist them in making design decisions. In other words, they employ a system of representations which depends on a mapping of paints onto chips according to their shared colour (their
18 We discuss systems of concrete objects below. Since mental vehicles are presumably brain states of some kind we here restrict our attention to resemblance relations between concrete objects. However, we will make some brief remarks about how concrete objects might represent conceptual objects such as numbers, theories, and formal systems. 19 We are here adapting some terminology developed by Shepard & Chipman (1970). They distinguish between first and second-order isomorphism. Isomorphism is a very restrictive way of characterising resemblance. We explain both the first-order/second-order distinction and the distinction between isomorphism and weaker forms of resemblance in what follows.
10
Gerard O'Brien and J. Opie
first-order resemblance). A useful side effect of having such a system is that when one wants to compare paints (e.g. 2-place comparisons such is "this one is bolder than that one," or 3-place comparisons such as "this one harmonises better with this one than with that one") one can do so by comparing the cards. This is because the system of chips embodies the same pattern of colour-relations as the paints. Whenever pairs or triples of paints satisfy particular colour relationships, their ink-coated proxies fall under mathematically identical relations. Similar remarks apply to surface maps. What makes a map useful is the fact that it preserves various kinds of topographic and metrical information. The way this is accomplished is by so arranging the points on the map that when location A is closer to location B than location C, then their proxies (points A, B and C on the map) also stand in these metrical relations; and when location A is between locations B and C, then points A, B and C stand in the same (3-place) topographic relation; and so on. The utility of a map thus depends on the existence of a resemblance relation that takes points on the map into locations in the world in such a way that the spatial relations among the locations is preserved in the spatial relations among the points. We will speak here of second-order resemblance. 2~ In second-order resemblance, the requirement that representing vehicles share physical properties with their represented objects can be relaxed in favour of one in which the relations among a system of representing vehicles mirror the relations among their objects. Of course, the second-order resemblance between colour charts and paints is a consequence of the first-order resemblance between individual chips and their referents. And in the case of surface maps, space is used to represent space. 21 But one can typically imagine any number of ways of preserving the pattern of relations of a given system without employing first-order resemblance. For example, the height of a column of liquid in a mercury thermometer is used to represent the temperature of any object placed in close contact with it. Here, variations in height correspond to variations in temperature.
20 Bunge (1969), in a useful early discussion of resemblance, draws a distinction between substantial and formal analogy which is close to our distinction between first and second-order resemblance. Two theorists who have kept the torch of second-order resemblance burning over the years are Palmer (1978) and Shepard (Shepard & Chipman 1970; Shepard & Metzler 1971). More recently, Blachowicz (1997), Cummins (1996), Gardenfors (1996), Johnson-Laird (1983), O'Brien (1999), and Swoyer (1991), have all sought to apply, though in different ways, the concept of second-order resemblance to mental representation. 21 Note, however, that this is already a step away from the first-order resemblance employed in the case of colour chips, since each colour chip shares a property with its represented object. Maps, on the other hand, don't preserve absolute distance, but only distance relations.
Notes Toward a Structuralist Theory of Mental Representation
II
Weather maps provide a more compelling example. On such a map, regions in the earth's atmosphere (at some specified elevation) are represented by points, and contiguous regions of equal atmospheric pressure are represented by lines known as "isobars" (see Figure 1). More significantly, the spacing of isobars corresponds to atmospheric pressure gradients, knowledge of which can be used to predict wind velocity, the movement of fronts, and so on. The representation of pressure gradients by isobar spacing is second-order, because for any two points on the map the relative spacing and orientation of isobars in the vicinity of those points corresponds to the relative size and direction of pressure gradients at the represented regions. Moreover, since geometric relations among lines and points on the map are being used to represent pressure relations this is a case of "pure" second-order resemblance m it doesn't depend on an underlying first-order resemblance. Let us be more precise. Suppose Sv -- (V, 9~v) is a system comprising a set V of objects, and a set 9~v of relations defined on the members of g. 22 The objects in V may be conceptual or concrete; the relations in 9~v may be spatial, causal, structural, inferential, and so on. For example, V might be a set of features on a map, with various geometric and part-whole relations defined on them. Or V might be set of well formed formulae in first-order logic falling under relations such as identity and consistency. We will say that there is a second-order resemblance between two systems S v - - ( V , 9~v) and So = (O, 910) if, for at least some objects in V and some relations in 9~v there is a one-to-one mapping from V to O and a one-to-one mapping from 9~v to N o such that when a relation in 9~v holds of objects in V, the corresponding relation in 9~0 holds of the corresponding objects in O. In other words, the two systems resemble each other with regard to their abstract relational organisation. As already stressed, resemblance of this kind is independent of first-order resemblance, in the sense that two systems can resemble each other at second-order without sharing properties. Second-order resemblance comes in weaker and stronger forms. As defined it is relatively weak, but if we insist on a mapping that takes every element of V onto some element of O, and, in addition, preserves all the relations defined on V, then we get a strong form of resemblance known as a homomorphism. 23 Stronger
22The relations in 9iv must have an arity greater than one. We exclude unary relations (i.e. properties) in order to maintain a clear distinction between first-order and second-order resemblance (see the definition above). 23Bunge describesthis kind of resemblanceas an "all-some... analogy" (1969:17). Swoyerrefers to it as an "isomorphic embedding" (1991: 456). A homomorphismis an injection: a one-to-one, all-some mapping, because very element of its domain maps to a unique element in its range, but not every element of the range is the image of some domain element.In other words, a homomorphismmaps the whole domain onto part of the range.
Fig,ure 1: A weather map showing a low pressure cell over South Australia. The isobars around the low are closely spaced, indicating a steep pressure gradient.
Notes Toward a Structuralist Theory of Mental Representation
13
still is an isomorphism, which is a one-to-one relation-preserving mapping such that every element of V corresponds to some element of O, and every element of O corresponds to some element of V. 24 W h e n two systems are isomorphic their relational organisation is identical. In the literature on second-order resemblance the focus is often placed on isomorphism (see, e.g. C u m m i n s 1996: 85-111), but where representation is concerned, the kind of correspondence between systems that is likely to be relevant will generally be weaker than isomorphism. In what follows, therefore, we will tend to avoid this restrictive way of characterising resemblance. The significance of second-order resemblance for mental representation is this. While it is extremely unlikely that first-order resemblance is the general ground of mental representation (given what we know about the brain) the same does not apply to second-order resemblance. Two systems can share a pattern of relations without sharing the physical properties upon which those relations depend. Second-order resemblance is actually a very abstract relationship. It is a mathematical or set-theoretic notion something which should be apparent from the way it was defined. Essentially nothing about the physical form of the relations defined over a system Sv of representing vehicles is implied by the fact that Sv resembles So at second-order; second-order resemblance is a formal relationship, not a substantial or physical one. e5 It is a little acknowledged fact that one of the more prominent approaches to mental representation in the recent literature exploits second-order resemblance. We have in mind the group of theories that go variously by the names causal, conceptual, or functional role semantics. 26 These functional role theories (as we shall call them) share a focus on the causal relations that a system of mental representing vehicles enter into; where they differ is in the class of causal relations they take to be significant for mental representation. What informs this causal focus is the idea that a system of vehicles represents a domain of objects when the former functionally resembles the latter. A functional resemblance obtains when
24 An isomorphism is, therefore, a bijection: a one-to-one, all-all (surjective) mapping. Every isomorphism is a homomorphism, and every homomorphism is a weak (some-some) second-order resemblance relation. But there are second-order resemblance relations that are not homomorphisms, and homomorphisms that are not isomorphisms. Second-order resemblance is therefore the most inclusive category, isomorphism the most restrictive. 25 A consequence of this is that a system of mentalvehicles (which by assumption is a set of brain states) is not only capable of standing in a relationship of second-order resemblance to concrete or natural systems, but also to abstract systems such as logical formalisms and theories. This is presumably a welcome outcome. 26 See, e.g., Block (1986, 1987), Cummins (1989:87-113), Field (1977, 1978), Harman (1982), Loar (1982), McGinn (1982), and Schiffer (1987).
14
Gerard O'Brien and J. Opie
the pattern of causal relations among a set of representing vehicles preserves at least some of the relations among a set of represented objects. 27 Nonetheless, while it is not always made clear (even by their proponents!) that these functional role theories of mental representation rely on second-order resemblance, 28 it is clear that they violate the causal constraint. The reason for this should now be familiar. Unlike convention, the connection between representing vehicles and represented objects forged by this grounding relation (i.e. by functional resemblance) is fully objective. But just like causation, it has no impact on the intrinsic properties of mental vehicles, and hence no influence on their causal powers. Specifically, mental representing vehicles don't possess the capacity to enter into causal relations (and thereby affect behaviour) in virtue of this second-order resemblance relation; rather, the resemblance relation itself obtains in virtue of the causal roles those vehicles occupy in a cognitive economy. This version of second-order resemblance, just like convention and causation, reduces representation to interpretation. Hence the analysis we applied to conventional and causal theories of mental representation, also applies to functional role theories. 29 With resemblance, therefore, we seem to be caught in a dilemma. On the one hand, any theory that treats first-order resemblance as the ground of mental representation, although compatible with the causal constraint, is incompatible with naturalism. On the other, while second-order resemblance is capable of forming the basis of a naturalistic theory of mental representation, it seems to give rise to theories that violate the causal constraint.
A Structuralist Theory of Mental Representation Fortunately, there is a way out of this dilemma. In all the recent discussion of functional role semantics in the philosophy of mind, another variety of second-order resemblance has been overlooked: second-order resemblance based on the physical relations among a set of representing vehicles. We will say that
27 Some theorists characterise functional role semantics in terms of afunctional isomorphism between representing vehicles and represented objects, rather than a functional resemblance (see, e.g. von Eckardt 1993:209-214). A functional isomorphism obtains when the pattern of causal relations among the set of representing vehicles is identical to the pattern of relations among the set of represented objects. We have already observed, however, that a resemblance relationship weaker than isomorphism may generally be sufficient to ground representation. 28 Cummins (1989:114-125) and von Eckardt (1993: 209-214) are important exceptions. 29 For a much more detailed consideration of the failure of functional role theories of mental representation in this regard, see Cummins (1996:29-51).
Notes Toward a Structuralist Theory of Mental Representation
15
one system structurally resembles another when the physical relations among the objects that comprise the first preserve some aspects of the relational organisation of the objects that comprise the second. Structural resemblance is quite different from functional resemblance. What determines the functional/structural distinction is the way relations in the second system are preserved by the first: by causal relations in the case of functional resemblance, by physical relations in the case of structural resemblance. In neither case is there any restriction on the kinds of relations allowed in the second s y s t e m - they can be relations among objects, properties or relations; they can be physical, causal or conceptual. 3~ The structuralist theory of mental representation is based on the conjecture that structural resemblance is the general ground of mental representation. This amounts to the claim that it is a relation of structural resemblance between mental representing vehicles and their objects that disposes cognitive subjects to behave appropriately towards the latter. Structural resemblance grounds all the various examples of representation discussed in the last section. A surface map preserves spatial relations in the world via spatial relations among map points. Since spatial relations are a species of physical relations this clearly qualifies as an instance of representation grounded in structural resemblance. Likewise, the representing power of a mercury thermometer relies on a correspondence between one physical variable (the height of the column of mercury) and another (the temperature of bodies in contact with the thermometer). And in weather maps the relative spacing of isobars is employed to represent relative pressure gradients. In each of these cases we can identify a system of vehicles whose physical relations ground a (non-mental) representing relationship. As yet we don't know enough about the brain to identify the structural properties and consequent resemblance relations that might ground mental representation. However, in our view, connectionism has taken important steps in that direction. In this context connectionist networks are to be understood as idealised models of real neural networks, which, although unrealistic in certain respects, capture what may well be the key structural features whereby the brain represents its world (see O'Brien 1998; Opie 1998). As an example consider Cottrell's face-recognition network (see Churchland 1995:38-55 for discussion). This network has a three layer feed-forward architecture: a 64 x 64 input array, fully connected to a hidden layer of 80 units, which in tum is fully connected to an output layer comprising 8 units. Each unit in the input layer can take on one of 256 distinct activation
30Note that it follows from this that both functional and structural resemblance can be asymmetric relations: one systemcan functionally/structurallyresemble a second, withoutthe converseobtaining.
16
Gerard O'Brien and J. Opie
values, so it is ideal for encoding discretised grey-scale images of faces and other objects. After squashing through the hidden layer these input patterns trigger three units in the output layer that code for face/non-face status and gender of subject, and five which encode arbitrary 5-bit names for each of 11 different individuals. Cottrell got good performance out of the network after training it on a corpus of 64 images of 11 different faces, plus 13 images of non-face scenes. He found that the network was: (i) 100% accurate on the training set with respect to faceness, gender and identity (name); (ii) 98% accurate in the identification of novel photos of people featured in the training set; and (iii) when presented with entirely novel scenes and faces, 100% correct on whether or not it was confronting a human face, and around 80% correct on gender. What is significant about the face-recognition network, for our purposes, is the way it codes faces at the hidden layer. Cluster analysis reveals that the network partitions its hidden unit activation space into face/non-face regions; within the face region into male/female regions; and then into smaller sub-regions corresponding to the cluster of patterns associated with each subject (see Figure 2). What this suggests is that the network supports a structural resemblance between activation
Figure 2: The hierarchy of learned partitions across the hidden unit activation space of Cottrell's face recognition network (after Churchland 1995: 49).
Notes Toward a Structuralist Theory of Mental Representation
17
patterns and the domain of human faces. Within the face region of activation space each point is an abstract (because compressed) representation of a face. Faces that are similar are represented by points that are close together in the space (this particularly applies to different images of a single face), whereas dissimilar faces are coded by points that are correspondingly further apart. So the relations among faces which give rise to our judgments of similarity, gender, and so on, are preserved in the distance relations in hidden unit activation space. This space, it should be remembered, is a mathematical space used by theorists to represent the set of activation patterns a network is capable of producing over its hidden layer. Activation patterns themselves are physical objects (patterns of neural firing if realised in a brain or a brain-like network), thus distance relations in activation space actually codify physical relations among activation states. Consequently, the set of activation patterns generated across any implementation of Cottrell's face-recognition network constitutes a system of representing vehicles whose physical relations capture (at least some of) the relations among human faces. All of this would be moot if this structural resemblance were causally inert. But in fact the structural resemblance embodied in Cottrell's network is arguably what powers both its computational and its behavioural capacities. Structural resemblance thus appears to ground representation in any connectionist computational system. Accordingly, if the brain is a network of connectionist networks then mental content is governed by the structural properties of neural networks, and semantic properties are finally brought home to roost in the mental representing vehicles themselves. Naturally, this is what any advocate of the structuralist theory of mental representation would hope to find. Standard approaches take the intrinsic properties of representing vehicles to be inconsequential, except in so far as these properties are consistent with the causal relations in which these vehicles are caught up. The physical relations among the vehicles are essentially arbitrary so far as their semantic properties are concerned. On the other hand, if the grounding relation for mental representation is a structural resemblance between the system of representing vehicles and the represented objects, then the intrinsic properties of the vehicles run the show. The semantic relations among the vehicles are none other than their physical relations, because it is the latter that generate the required grounding. 31 According to the structuralist theory of mental representation we can make sense of the power of our inner models to safely guide us through the world if we suppose two things:
31Paul Churchland is one neurophilosopher who has long argued for the semantic significance of neuronal activation spaces. See, for example, Churchland (1998, 2001).
18
Gerard O'Brien and J. Opie
(1) that the system(s) of mental representing vehicles in our heads stand in an objective relation of second-order resemblance to their represented objects; (2) that this resemblance is supported by the physical relations among the mental vehicles. The crucial difference between a theory of mental representation based on structural resemblance and one based on physical (first-order) resemblance, is that the former is compatible with the implementation of mental representation in the brain. The crucial difference between a theory of mental representation based on structural resemblance and one based on functional resemblance, is that the former is compatible with the causal constraint. The physical relations among a system of representing vehicles are independent of their causal relations, and hence of the use to which they are put. Indeed, it is the physical relations among representing vehicles which explain their causal powers. In this sense, structural resemblance underwrites and explains the existence of any functional resemblance between representing vehicles and their represented domain. This is just how things should be if the evolutionary story with which we started is on the right track. It's our contention that of all the possible grounds of mental representation, only structural resemblance can satisfy the twin demands of naturalism and the causal constraint. By our lights, this makes the structuralist theory of mental representation mandatory for those philosophers of mind who think this discipline is answerable to cognitive science. In this paper we have merely offered the skeletal outlines of such a theory. The task of putting some flesh on those bones will have to wait for another day (but see O'Brien & Opie in prep.).
References Baker, L. R. (1993). Metaphysics and mental causation. In: J. Heil, & A. Mele (Eds), Mental causation. Oxford University Press. Blachowicz, J. (1997). Analog representation beyond mental imagery. Journal of Philosophy, 94, 55-84. Block, N. (1986). Advertisement for a semantics for psychology. Midwest Studies in Philosophy, 10, 615-678. Block, N. (1987). Functional role and truth conditions. Proceedings of the Aristotelian Society, 61, 157-181. Block, N. (1989). Can the mind change the world? In: G. Boolos (Ed.), Meaning and method: Essays in honor of Hilary Putnam. Cambridge University Press. Bunge, M. (1969). Analogy, simulation, representation. Revue-Internationale-dePhilosophie, 23, 16-33.
Notes Toward a Structuralist Theory of Mental Representation
19
Churchland, E M. (1995). The engine of reason, the seat of the soul. MIT Press. Churchland, E M. (1998). Conceptual similarity across neural and sensory diversity: The Fodor Lepore challenge answered. The Journal of Philosophy, 95, 5-32. Churchland, E M. (2001). Neurosemantics: On the mapping of minds and the portrayal of worlds. In: S. Colonna (Ed.), The emergence of mind: Proceedings of the international symposium. Montedison. Cummins, R. (1989). Meaning and mental representation. MIT Press. Cummins, R. (1996). Representations, targets, and attitudes. MIT Press. Dennett, D. (1978). Brainstorms. MIT Press. Dennett, D. (1980). The milk of intentionality. Behavioral and Brain Sciences, 3, 428-430. Dennett, D. (1982). Styles of mental representation. Proceedings of the Aristotelian Society, New Series, 83, 213-226. Dennett, D. (1987). The intentional stance. MIT Press. Dretske, E (1981). Knowledge and the flow of information. MIT Press. Dretske, E (1987). The explanatory role of content. In: R. Grimm, & D. Merrill (Eds), Contents of thought. University of Arizona Press. Dretske, E (1988). Explaining behavior: Reasons in a worm of causes. MIT Press. Dretske, E (1990). Does meaning matter? In: E. Villanueva (Ed.), Information, semantics, and epistemology. B lackwell. Field, H. (1977). Logic, meaning, and conceptual role. Journal of Philosophy, 74, 379-409. Field, H. (1978). Mental representation. Erkenntnis, 13, 9-61. Files, C. (1996). Goodman's rejection of resemblance. British Journal of Aesthetics, 36, 398-412. Fodor, J. A. (1986). Banish DisContent. In: J. Butterfield (Ed.), Language, mind, and logic. Cambridge University Press. Fodor, J. A. (1987). Psychosemantics. MIT Press. Fodor, J. A. (1989). Making mind matter more. Philosophical Topics, 17, 59-79. Fodor, J. A. (1990). A theory of content and other essays. MIT Press. Gardenfors, E (1996). Mental representation, conceptual spaces and metaphors. Synthese, 106, 21-47. Harman, G. (1982). Conceptual role semantics. Notre Dame Journal of Formal Logic, 28, 242-256. Heil, J., & Mele, A. (1991). Mental causes. American Philosophical Quarterly, 28, 61-71. Jackson, E, & Pettit, E (1990a). Causation and the philosophy of mind. Philosophy and Phenomenological Research Supplement, 50, 195-214. Jackson, E, & Pettit, E (1990b). Program explanation: A general perspective. Analysis, 50, 107-117. Johnson-Laird, E (1983). Mental models. Harvard University Press. LePore, E., & Loewer, B. (1989). More on making mind matter. Philosophical Topics, 17, 175-191. Lloyd, D. (1989). Simple minds. MIT Press. Loar, B. (1982). Conceptual role and truth conditions. Notre Dame Journal of Formal Logic, 23, 272-283.
20
Gerard O'Brien and J. Opie
McGinn, C. (1982). The structure of content. In: A. Woodfield (Ed.), Thought and context. Oxford University Press. Millikan, R. G. (1984). Language, thought and other biological categories. MIT Press. O'Brien, G. (1998). The role of implementation in connectionist explanation. Psycoloquy, 9(06) http ://www.cogsci. soton.ac.uk/psyc-bin/newpsy ?9.06. O'Brien, G. (1999). Connectionism, analogicity and mental content. Acta Analytica, 22, 111-136. O'Brien, G., & Opie, J. (1999). A connectionist theory of phenomenal experience. Behavioral and Brain Sciences, 22, 127-148. O'Brien, G., & Opie, J. (in preparation). Structural resemblance and neural computation. Opie, J. (1998). Connectionist modelling strategies. Psycoloquy, 9(30), http://www.cogsci. soton.ac.uk/psyc-bin/newpsy?9.30. Palmer, S. (1978). Fundamental aspects of cognitive representation. In: E. Rosch, & B. Lloyd (Eds), Cognition and categorization. Lawrence Erlbaum. Schiffer, S. (1987). Remnants of meaning. MIT Press. Shepard, R., & Chipman, S. (1970). Second-order isomorphism of internal representations: Shapes of states. Cognitive Psychology, 1, 1-17. Shepard, R., & Metzler, J. (1971). Mental rotation of three-dimensional objects. Science, 171, 701-703. Stampe, D. (1977). Towards a causal theory of linguistic representation. Midwest Studies in Philosophy, 2, 42-63. Stampe, D. (1986). Verificationism and a causal account of meaning. Synthese, 69, 107-137. Swoyer, C. (1991). Structural representation and surrogative reasoning. Synthese, 87, 449-508. Von Eckardt, B. (1993). What is cognitive science ? MIT Press.
Chapter 2
Representation and Indication Robert Cummins and Pierre Poirier
Two Kinds of Mental Content This paper is about two kinds of mental content and how they are related. We are going to call them representation and indication. We will begin with a rough characterization of each. The differences, and why they matter, will, hopefully, become clearer as the paper proceeds.
Representation Some authors (e.g. Schiffer 1987) use "mental representation" to mean any mental state or process that has a semantic content. On this usage, a belief that the Normans invaded England in 1066 counts as a mental representation, as does the desire to be rich. This is not how we use the term. As we use the term, a mental representation is an element in a scheme of semantically individuated types whose tokens are manipulated-- structurally transformed p by (perhaps computational) mental processes. The scheme might be language-like, as the language of thought hypothesis asserts (Fodor 1975), or it might consist of (activation) vectors in a multidimensional vector space as connectionists suppose (e.g. Churchland 1995). Or it might be something quite different: a system of holograms, or images, for example. On one popular theory of the propositional attitudes, having the belief that the Normans invaded England in 1066 involves tokening a mental representation with the content that the Normans invaded England in 1066 writing in the Belief
Representation in Mind: New Approaches to Mental Representation Edited by H. Clapin, P. Staines and P. Slezak Copyright 9 2004 by Elsevier Ltd. All rights of reproduction in any form reserved. ISBN: 0-08-044394-X
22
Robert Cummins and Pierre Poirier
Box a representation that means that the Normans invaded England in 1066.1 That theory takes the content of the belief to be the same as the content of the implicated representation, but distinguishes representations from attitudes, taking representations, as we do, to be expressions in a scheme of representational types. We think this theory of the attitudes is seriously flawed (see C u m m i n s 1996), but one thing it surely gets right is that representations should be distinguished from attitudes. 2
Indication It is useful to begin with some influential examples. 9 Thermostats: The shape of the bimetallic element (or the length of a column of mercury) is said to indicate the ambient temperature. 9 Edge detectors: Cells in the primary visual cortex (V 1) strongly respond to edges in the visual field, that is, linear boundaries between light and dark regions. 3 9 Idiot lights: Most cars have indicator lights that come on when, e.g. the fuel level is low, or the oil pressure is low, or the engine coolant is too hot. "Indication" is just a semantic-sounding word for detection. We are going to need a way to mark the distinction between the mechanism that does the detection and the state or process that is the signal that the target has been detected. We will say that the cells studied by Hubel & Wiesel (1962) are indicators, and that the patterns of electrical spikes they emit when they fire are indicator signals. Rather less obviously, a mercury thermometer is an indicator, and the length of the mercury column is the indicator signal. Similarly, the bimetallic element found in most thermostats is an indicator, and its shape is the signal. 4
1This way of putting things m that believing that it will rain and desiring that it will rain differ only in that a mental representation meaning that it is raining is in the Belief Box in the first case and in the Desire Box in the second- is due to Schiffer (1987). 2 Conceptual role theories of mental content have difficulty distinguishing mental representations from the attitudes because they take the mental content of x to be a function of x's epistemic liaisons. Propositional attitudes have epistemic liaisons, but representations, as we use the term, do not. See Cummins (1996), Chapter 3 for a detailed discussion of this point. 3 The cells were discovered by David Hubel & Torsten Wiesel (1962). They describe the behaviors of these cells this way: "The most effective stimulus configurations, dictated by the spatial arrangements of excitatory and inhibitory regions, were long narrow rectangles of light (slits), straight-line borders between areas different brightness (edges), and dark rectangular bars against a light background." 4 Thermostats using bimetallic strips are usually designed so that when the bimetallic changes shape "enough," it closes a circuit that turns on the furnace, or air conditioner, or whatever. The electrical
Representation and Indication
23
Indication and Representation Contrasted It is commonplace to think of indication as a species of representation. Indeed, one very popular theory of representational content has it that, at bottom, representational content just is, or is inherited from, indicator content. 5 We think there are some good reasons to keep the two distinct. Indication is transitive, representation is not. If $3 indicates $2, and $2 indicates S1, then $3 indicates S1. Aim a photosensitive cell at the oil pressure indicator light in your car. Attach this to a relay that activates an audio device that plays a recording of the sentence, "Your oil pressure is low" If the light going on indicates low oil pressure, so does the recording. Indeed, there is already a chain of this sort connecting the pressure sensor and the light. Representation, on the other hand, is not transitive. A representation of the pixel structure of a digitized picture of my aunt Tilly is not a representation of my aunt Tilly's visual appearance, though, of course, it is possible to recover the later from the former. To anticipate some terminology we will use later, a representation of the pixel structure is an encoding of my aunt Tilly's visual appearance. 6 Indicator signals are arbitrary in a way that representations are not. This actually follows from the transitivity of indication. Given transitivity, anything can be made to indicate anything else (if it can be detected at all), given enough ingenuity and resources. (This is what makes it tempting to think of words as indicators: they mean something, but they are arbitrary.) It follows from the arbitrariness of indicator signals that disciplined structural transformations of them are not going to systematically alter their meanings. Susceptibility to such transformations, however, is precisely how representations earn their keep. Consider, for example, a software package that takes a digitized image of a face as input and "ages" it, i.e. retums an image of that face as it is likely to look after some specified lapse of time. Nothing like this could possibly work on an input that was required only to indicate a certain face, because there is no correlation between the physical characteristics something must have to be a signal that indicates the
pulse that closes the furnace relay is also an indicator signal, but it doesn't indicate the temperature, except relative to a fixed thermostat setting. 5 The theory is generally credited to Dennis Stampe (1977). Its most prominent advocates are Fodor (1987, 1990) and Dretske (1981, 1988). 6 It is also what Haugeland would call a recording of the picture. See Haugeland (1991). If r represents a representation r' of t, r need not be a representation of t even though, as was argued in Cummins (1996), representation is grounded in isomorphism, which is transitive. This is because the structure that r shares with r' need not, in general, be the same as the structure rs shares with t, as the pixel example illustrates.
24
Robert Cummins and Pierre Poirier
appearance of my face at age 18 and the physical characteristics of my face at age 18. You could design a device that s e e m e d to do this, of course. Given a name and a target age, it would retrieve from memory its most recent picture of the person named (if it has one), and age it by the number of years equal to the target date minus the date the picture was taken. But surely(?), respectable cognitive scientists would not be fooled by such a device. They will know that you need a representation, not just an indicator signal, to get the job done, and they will infer the internal representation. 7 Similarly, while it could conceivably be useful to detect the utterance of a particular sentence ("Look out!"), there is no possibility of using a mere indicator signal to effect an active-to-passive transformation. A representation of the sentence's phrase structure, on the other hand, actually has, by hypothesis, the relevant structure, and hence formal transformations - - transformations that systematically alter structure - - of a representation of that phrase structure can easily be constructed to give a representation of the structure of the corresponding passive. It is, therefore, a consequence of the nature of indication that the structural properties of an indicator signal (if it has any) have no significance. 8 Indicators "say" that their targets are there, but do not "say" anything about what they are like. Representations, on the other hand, mirror the structure of their targets (when they are accurate), and thus their consumers can cognitively process the structure of the target by manipulating the structure of its representation. But representations, unlike indicator signals, are typically silent concerning whether their targets are "present": they are not, except incidentally and coincidently, detector signals. I n d i c a t o r s are s o u r c e d e p e n d e n t in a w a y t h a t r e p r e s e n t a t i o n s are not. The cells studied by Hubel and Wiesel all generate the same signal when they detect a target. You cannot tell, by looking at the signal itself (the spike train), what has been detected. You have to know which cells generated the signal. This follows from the arbitrariness of indicator signals, and is therefore a general feature of indication: the meaning is all in who shouts, not in what is shouted. 9 In sum, then, indication is transitive, representation is not. It follows from the transitivity of indication that indicator signals are arbitrary and source dependent in a way in which representations are not, and this disqualifies indicator signals
7 Actually, a structural encoding (Cummins et al. 2001) of the information a representation would carry would do the trick as well. This does not affect the present point, which is that indicator signals would not do the trick. 8 The structural properties of an indicator signal might be used to type-identify it. More of this shortly. 9 We do not mean to imply here that the shape of a spike train is never significant. The point is rather that two indicators can have the same spike train, yet indicate different things.
Representation and Indication
25
as vehicles for structure dependent cognitive processing. Representation is intransitive, non-arbitrary and portable (not source dependent), and therefore suitable for structural processing. Indicator signals "say" their targets are present, but "say" nothing about them; representations provide structural information about their targets, but do not indicate their presence. Indicator signals say, "My target is here," while representations say, "My target, wherever it is, is structured like so."
Portable Indicators and the Language of Thought In principle, indicator signals can be distinguished into types so as to reduce source dependency. Rather than label the low fuel and low oil pressure lights, one could make the lights different sizes or shapes or colors. This amounts to building an arbitrary source label into the form of the signal. The result is a portable signal. It is not obvious, however, what advantage there might be to such a system beyond the fact that you could tell in the dark whether it is oil pressure or fuel that is low. To appreciate the value of portability, we have to imagine a more complex system. Consider, then, the system called LOCKE (see, e.g. Cummins 1989: 37ff) (Figure 1). A TV camera is attached to a card punch device. The pattern of holes punched in a "percept" card depends on the structure of the optical stimulus. The matcher is equipped with a set of "concept" cards. A match is achieved when every hole in a concept card matches a hole in a percept card (barring chads, of course). When a match is achieved, the word printed on the back of the concept card is displayed.
//
Master Cards Y J TV
PUNCH
MATCHER
black square DISPLAY
Percept cards passed to matcher here
Matched master passed to display here
Figure 1" The Locke machine (from Cummins 1989).
26
Robert Cummins and Pierre Poirier
It would be no mean feat to build such a device. A Frog detector, for example, would have to accommodate the fact that frogs come in various postures, colors and sizes, against various backgrounds, with various neighbors, etc. Even color and shape detection will have to accommodate variations in distance, in lighting and in angle of view. For LOCKE to work well, a huge variety of visual scenes whose only common feature is the presence of a frog or red or square will have to produce cards that have the same characteristic sub-pattern. The "analysis" box thus amounts to a sophisticated vision system capable of detecting very "high level" distal properties. As this example illustrates, the physical characteristics of the signal - - punch patterns - - can be arbitrarily different from the physical characteristics of the detector itself, and from the physical characteristics of the target. Heat detectors do not have to get hot, nor must motion detectors move. Unlike the sorts of indicators studied by Hubel and Wiesel, however, Locke's indicator signals are type identified by their structural properties (their "form"), and are therefore potentially source independent. This makes it possible to compare concept cards to each other to yield such judgments as that sisters are female, or even that frogs are green, even though none of the corresponding perceptual targets are in the offing. LOCKE, in short, has a crude language of thought, the primitive symbols of which are punch patterns ~ perhaps single holes ~ whose contents are fixed by their roles in detection. A punch pattern means square, wherever or whenever it is tokened, if it is the pattern LOCKE's detection system produces in response to squares. 10 One might then suppose that primitive patterns could be combined in ways that yield complex patterns whose meanings are functions of the meanings of their constituents and their mode of combination. The resulting complex representations will have the semantically primitive indicator signal patterns as constituents, but these will not, of course, function as indicator signals in this new role. Rather, they simply inherit their denotations from their other life as signals indicating the presence of the properties (or whatever) they are now taken to denote. As things stand, LOCKE cannot represent bindings: It cannot represent the difference between a red square plus a blue triangle on the one hand, and a blue square plus a red triangle on the other. To overcome this limitation, and others like it ~ to exploit the possibilities of combinatorics ~ LOCKE will need some syncategorematic representational machinery. The forms that distinguish bindings and the like evidently cannot themselves have their meanings (semantic
l0 Qualifications are required to turn this into an even mildly plausible definition of content. See Cummins (1989) for details.
Representation and Indication
27
functions) fixed by their roles in indication. Rather, a complex representation will count as a conjunction or predication or whatever in virtue of how it is processed (Fodor 1987). The machinery of binding is just whatever form a representation has that causes it to be processed as a binding. Portable indicator signals, then, provide one way of grounding the capacity to represent in the capacity to indicate (detect). The result is a Language of Thought (LOT) as it is defended by Fodor (1975). There is a characteristic approach to cognition entrained by LOT style representation. It is entrained by the fact that there are basically just three ways that arbitrary mental symbols of the Language of Thought variety can enter into cognitive explanations: As triggers for procedures, as cues for stored knowledge, and as constituents of complex representations. This simply follows from the fact that the structure of a LOT symbol serves only to type identify it. It carries no information about its target, not even the information that its target is present, since, freed from its source, it no longer functions as an indicator signal. Suppose you are to milk the cow. First you must find the cow. You wander around scanning until your visual system tokens a Icowl an arbitrary mental symbol that refers to cows. But to visually recognize cows, you need to know how a cow looks. A Icowl contains no information about how cows look, and so it isn't what psychologists, at any rate, would call a visual concept. But knowledge of cows is what you need, for it is knowledge of cows, including tacit knowledge about the sort of retinal projections they tend to produce, that makes it possible for your visual system to token a Icowl when you see one. So the Mentalese Icowl does no work for the object recognition system, it just signals its output, functioning as an indicator signal. Tokening a [cow I, we may suppose, triggers the next step in the plan. Needing to locate the udder, a mental word is totally useless unless it happens to function as a retrieval cue for some stored knowledge about cows. Faced with actually having to deal with a cow, the burden therefore shifts again from the symbol to your stored knowledge, because the symbol, being arbitrary, tells you nothing about cows. So it turns out that it isn't because you have a Mentalese term for cows that you find the cow and get the milking done, it is because you have some stored knowledge about cows m some in your visual analysis system, some higher up the cognitive ladder. Mentalese ]cowls could play a role in stored knowledge about cows only as constituents of complex representations m Icows have udders between their back legs], for example ~ that are, on the Mentalese story, implicated in the possession of stored knowledge about cows. LOT stories therefore lead inevitably to the idea that it is really stored knowledge, in the form of systems of LOT "sentences" that does the explanatory work. It is worth emphasizing that there is a big difference between appealing to
28
Robert Cummins and Pierre Poirier
the fact that one has a primitive mental symbol referring to cows, and appealing to the fact that one has a lot of knowledge about cows. LOT commits one to the view that representations of cows don't tell you anything about cows. On the contrary: having a mental symbol that refers to cows presupposes considerable knowledge about cows. Perhaps it isn't so bad that LOT entails that the representations that are satisfied by cows have only an indirect role in the explanation of cow cognition, for there are always mental sentences to tell us about cows. But let's just be clear about what LOT is committed to here: The view we have arrived at is that cognition is essentially the application of a linguistically expressed theory. All the serious work gets done by sets of sentences that are internal tacit theories about whatever objects of cognition there happen to be. As far as cognizing cows goes, your [cowls really don't matter; it is your tacit theory (or theories) of cows that does the work. Enough has been said to suggest how the language of thought hypothesis provides for assembling representational structures from symbols whose semantic content is grounded in their functions as (portable) indicator signals. The strategy, however, has its limitations. Two are important for the present discussion. First, portable indicator signals that are assembled into complex representations, while a reality in digital computers, are surely not in the cards in the brain. As we have seen, complex LOT representations cannot have actual source-dependent indicator signals as constituents, for this would imply that every representation indicated the presence of the targets of each of its constituents. Such "representations," indeed, would not be representations at all, but simply bundles of simultaneous feature detections. The transition from indication to representation in LOT systems is mediated by source-independent signal types that, severed from their source, denote the properties whose presence they detect when functioning as indicator signals. II But such source-independent (portable) indicator signals are neurally implausible, to say the least. What formal characteristics could distinguish the signal types? There is no evidence to support the idea that a distinctive spike train produced by a neural indicator retains its semantic significance when produced elsewhere in the brain. It is more plausible to suppose that a distinctive pattern of activation in a pool of associated neurons might retain its significance if copied, or simply repeated, elsewhere. But there is no evidence whatever for the suggestion that such patterns are not only the distinctive outputs of detection circuits, but are
11Or rather: they are, when things are working properly, processed as if they denoted the properties they detect when functioning accurately as indicator signals.
Representation and Indication
29
also assembled into complex representations. The fear, voiced early and often by connectionist critics of LOT systems, that LOT systems have no plausible neural implementation, seems well-founded. The second limitation of LOT systems has to do with their restricted representational power. Representations in the language of thought have whatever structure is implicated by their combinatorial syntax. In language like schemes, the structure in question is logical form. While propositions arguably have structures isomorphic to logical forms, it is surely the case that many, perhaps most, representational targets of significance to cognitive systems have structures of entirely different sorts. Natural images of the sort studied by Olshausen & Field (1996), for example, certainly do not have logical forms, nor do problem spaces of the sort studied by planning theory in artificial intelligence. (Newell & Simon 1972; see Cummins 1996, for more on this theme.) Representational systems whose non-syncategorematic elements get their meanings by inheriting denotations from their roles in indication are, inevitably, denotational schemes, schemes whose semantics is the familiar truth-conditional semantics. The only thing such schemes can represent are propositions or complex propositional functions. Everything else is simply denoted. You can call denotation representation, provided you keep in mind the very real difference between something that merely labels its target, and something that actually provides information about it. Grounding representation in indication by promoting portable indicator signals into the semantic constituents of complex representations inevitably leads to a scheme that represents propositions and nothing else. LOT schemes get around this limitation by encoding structure rather than representing it. To see the difference, compare LOT schemes for representing sentences in a natural or artificial language with G6del numbering. A LOT scheme represents a target sentence S by tokening a sentence in the language of thought that has the same relevant structure e.g. logical f o r m - as S. In the G6del numbering scheme, on the other hand, words are assigned natural numbers, and their positions in a sentence are encoded by prime numbers in ascending order. A word in position n assigned to m yields the number n m. This number is uniquely factorable into n, m times. The G6del number of a sentence is determined by multiplying all these uniquely factorable numbers together, yielding a uniquely factorable number from which the sentence can be recovered. For example, Assume that "John," "Mary" and "loves" are assigned 2, 3 and 4 respectively. Then "John loves Mary" is encoded as 22 x 43 x 35 - 62208. This number may be represented in any convenient number system. The result is a numeral that does not itself share structure with the sentence it encodes. Information about the constituent structure of the sentence is still there, though well disguised, and this is
30
Robert Cummins and Pierre Poirier
what makes it possible to devise arithmetical processes that manipulate the relevant structural information without actually recovering it and representing it explicitly. An active-passive transformation, for example, can be written simply as an arithmetic procedure. 12
The Assembly of Complex Visual Representations According to indicator-based theories of mental content generally, and to LOT theories in particular, mental representations either are indicator signals, or inherit their content from their roles, or the roles of their constituents, as indicator signals. We have seen however, that there are serious reasons for doubting that complex representations in the brain could be semantically composed of constituents whose meanings are inherited from their roles as indicator signals. An entirely different relation between indication and representation emerges if we examine the way in which the sort of indicators discovered by Hubel and Wiesel are implicated in the assembly of visual images. One account of this is to be found in recent research by Field and Olshausen (Field 1987, 1994; Olshausen & Field 1996, 1997, 2000). Natural images contain much statistical structure and redundancies (Field 1987) and early visual processing functions to retain the information present in the visual signal while reducing the redundancies. In the 1950s, Stephen Kuffler (1952) discovered the center-surround structure of retinal ganglion cells' response and, 40 years later, Joseph Atick (1992) showed that this arrangement serves to decorrelate these cells' responses. As Horace Barlow (1961) had suspected, sensory neurons are organized to maximize the statistical independence of their response. Bruno Olshausen and David Field recently showed that the same is true of neurons in the primary visual cortex. Hubel and Wiesel discovered that neurons in the primary visual cortex are sensitive to edges, hence their functional description as edge detectors, but could only guess at the functional relevance of this structure (Hubel 1988). According to Olshausen and Field, edge detection allows neurons in the primary visual cortex to respond to visual signals in a maximally independent fashion and thus produce sparsely coded representations of the visual field. To show this, they constructed an algorithm that could identify the minimal set of maximally independent basis functions capable of describing natural images (or small 12 by 12 pixel patches thereof) in a way that preserves all the information present in the visual signal. Because natural images tend to contain edges, and
12For an extended discussion of this theme, see Cumminset al. (2001).
Representation and Indication
31
Figure 2: Optimal basis function set to represent natural images (from Olshausen & Field 2000).
because there are reliable higher-order correlations (three-point and higher) between pixels along an edge, it turns out that natural images can be fully described with minimal resources as composites of about a hundred such basis functions (see Figure 2). Given the statistical structure of natural images in the environment, there had to be such a set of functions, but the important point is that these basis functions are similar to those Hubel and Wiesel found 40 years earlier: spatially localized and oriented edges. Recently, O'Reilly & Munakata (2000) showed how to train a neural network using a form of Hebbian learning (conditional PCA) to produce a similar set of basis functions.
32
Robert Cummins and Pierre Poirier
To understand how visual representations are constructed out of these basis functions, consider a vector of V1 cortical cells 13 connected to the same retinal area via the same small subset of L G N cells. 14 Each cell has a receptive field similar to one in the minimal set of basis functions in Figure 1 above. Visual representations (of a small retinal area) can thus be thought of as a vector of cell activations, and an observed state of affairs (the inferred distal stimuli) as a linear function of a visual representation and some noise factor. A natural, but ultimately wrong, way to understand the construction of visual representations is as the activation of a subset of such basis functions solely on the basis of the information each cell receives from the relevant retinal region. In such a case, cells in the primary visual cortex would function as indicators of activity in the L G N and, ultimately, of properties in the visual field. Indeed, it would be easy to determine what state of affairs is currently observed if the basis functions were completely independent and the noise factor was known, but neither is the case. Because of this, many distinct visual representations (sets of basis vector functions) will account for the same observed state of affairs. Lewicki & Olshausen (1999) have shown, however, that it is possible to construct (infer) a unique visual representation given prior knowledge of the observed environment, that is, prior probabilities (the probability of observed states of affairs) and likelihood functions (the probability of visual representations given observed states of affairs). Instead of constructing a visual representation from a set of indicator signals, the visual system may infer the representation from indicator signals and relevant probabilistic knowledge of the visual environment. Visual representations are thus constructed from retinal indicator signals, knowledge of the high-order correlational structure in the environment (coded in the LGN-V1 connections) and knowledge of relevant prior probabilities and likelihood functions. The picture that emerges here involves the construction of an image from a set of indicator signals that have the following characteristics:
13In reality, this vector is contained within a hypercolumn, which also contains a similar contralateral vector (for the other eye). All cells in a given hypercolumn respond to the same region of retinal space. Neighboring hypercolumns respond to neighboring regions of retinal space (Livingstone & Hubel 1988). 14 The LGN (the lateral geniculate nucleus of the thalamus) is not a mere relay station en route to the cortex since it performs important computations of its own. But these will not matter here, as they are mostly concerned with the dynamic, as opposed to structural, aspects of vision and, owing to the massive backprojections from the cortex to LGN, with visual attention processes (it is estimated that there are ten times more backprojections from the cortex to the LGN as there are projections from the LGN to the cortex, Sherman & Koch 1986).
Representation and Indication
33
9 They are surprisingly few in number. 9 They indicate multi-point correlations between adjacent pixels in the (whitened version of the) input. 9 Their prowess as detectors of their proprietary targets is due, to a large extent, to recurrent circuitry that, in effect, computes a Bayesian function in which the prior probabilities are determined by the properties of neighboring areas of the developing image. 9 Their representational content is semantically disjoint from that of the image they compose in the same way that pixel values themselves are semantically disjoint from the representational content of a computer graphic. Like maps and scale models, image representations thus constructed have their meanings in virtue of their geometry rather than their origins. This is what gives them the source-independence characteristic of representations rather than indicator signals, and what allows for the possibility of disciplined structure sensitive transformations. Because such representations literally share structure with their targets, both static and dynamic structural properties of those targets can be mirrored by learning to transform the representations in ways that mirror the ways in which nature constrains the structure of, and structural changes in, the targets. Faces can be aged, objects can be rotated or "zoomed," projections into three dimensions can be computed. None of this is thinkable in a system in which visual representations are semantically composed from constituents whose contents are determined by their roles as indicator signals.
Representations and Targets We have been emphasizing the source-independence that representations achieve in virtue of their distinctive internal structure which either mirrors or encodes the structure of what is represented, or encodes semantic combinatorics. To be useful, however, representations must typically retain a kind of source-dependence. To understand this, we need to understand the relation between a representation and its target. "x represents y" belongs to a group of relational semantic terms that are ambiguous. This can be seen by reflecting on a comment of Jerry Fodor's. "Misrepresentation," Fodor said, "is when a representation is applied to something it doesn't apply to." (Email exchange.) It is obvious that "applies to" has to be ambiguous if this remark is to make sense and be true, which it evidently does and is. Analogous observations apply to "x refers to y," "x is true of y," "x means y" and "x represents y." For example, you say, "I used that map and found my
34
Robert Cummins and Pierre Poirier
way around the city with no problem." "Which city do you mean?" ("Which city are you referring to?") I ask. Here, I am asking for your target, the city against which that map's accuracy is to be measured. "Here is the city map I was telling you about," you say. "Which city do you mean?" ("Which city are you referring to?") I ask. Here, I am asking for the content, i.e. for the city the map actually represents - - the city with which the map shares structure. Though unmarked in ordinary vocabulary, the distinction between representational targets and contents is a commonplace. We are looking through my pictures of opera singers. You ask me what Anna Moffo looked like in her prime, and I hand you a picture, which happens to be a picture of Maria Callas. (It's hard to believe, but it is just a Philosophy example.) We can distinguish here between the target of my representing M Anna Moffo, the singer the picture I handed you was supposed to represent on this occasion - - and the content of the picture I p r o d u c e d - Maria Callas. The target of a representation on a particular occasion of its use is whatever it is supposed to represent on that occasion of use. The content of a representation is whatever it actually does represent. So: a representation can be applied to something it doesn't apply to because there can be a mismatch between the content of a representation and its target on a particular occasion of its use, a mismatch between what it actually represents its representational c o n t e n t - and what it is intended to represent on that occasion. Let's call something whose function is to represent some target or class of targets an intender. Intenders come in many forms. A clear example would be the process hypothesized by standard speech-recognition models whose function is representing the phrase structure of the current linguistic input. This process has a proprietary class of targets, viz., phrase structures of the current linguistic inputs. The accuracy of the representational performances of that process are to be judged by reference to the degree of match between the current target the actual phrase structure of the current i n p u t - and the actual (as opposed to intended) content of the representation it produces on that occasion. A representation's target, as opposed to its representational content, is sourcedependent. Processes downstream of the parser lately i m a g i n e d - its clients operate on the assumption that the representations it gets from that intender are representations of the phrase structure of the current linguistic input. To know that a token representation's target is the phrase structure of the current linguistic input amounts to knowing it was produced by an intender whose business it is to represent the phrase structure of the current linguistic input. If the representation were produced by a different intender, one whose business it is to represent move-trees in checkers, for example, it would need very different treatment (though it probably wouldn't get it in this case).
Representation and Indication
35
Exactly this sort of source-dependency is exhibited by the visual representations built up in V 1. Consumers of these representations other than mere enhancers need (typically) to operate on the assumption that they are processing representations of the current visual scene, and not, say, a remembered scene. This information is not encoded in the structure of the representation itself. To be available to a consumer that has more than one source, then, the information has to be carried by the fact that a particular intender produced the representation in question. It is tempting to suppose that source information, hence target information, could be incorporated into the representation itself. To see the limitations of this idea, imagine labeling year book photographs with the names of the people represented. This gives us representations with contents of the form, "Susan Smith looks like this < p h o t o > . " There is nothing wrong with this scheme, provided consumers know that the targets of the complex representations m name plus picture are true propositions about people's appearances. 15 But imagine someone using such a representation in a reductio argument. "Suppose Mafia Callas looked like this < p h o t o > , " she starts out, exhibiting a photo of Anna Moffo. Her target is a false proposition about how Mafia Callas looked. A consumer who does not realize this will get things seriously wrong. Labeling a photo just produces a new complex representation with a new content. Its target could be anything. Though most things, e.g. the number nine, will be non-starters as the target, it doesn't follow that they couldn't be the target. It is just unlikely that any intender would be that stupid. A representation cannot say what its own target is. 16 Thus, representations have this in common with indicator signals: just as the information carried by an indicator signal depends on which indicator produces the signal, so the information carried by a token representation depends, in part, on the intender that produces it. Unlike indicator signals, however, representations carry lots of information about their targets beyond the fact that they are present. And this is precisely what makes them useful. 17
15It actually isn't clear that the targets could be propositions, since the representationshave pictures as constituents. See Cummins (1999), for an argument against the view that pictures could have propositional contents. From the fact that pictures cannot have propositional contents, it doesn'tfollow that representations with pictures as constituents cannot have propositional contents. But it makes it seem problematic. 16This is why LOTers put LOT sentences into boxes like the Belief Box. An II am richl in the Belief Box is presumed to have a true proposition as target. Not so an 1Iam rich[ in the Desire Box. 17There is much more to be said about intenders and target fixation. Some of this can be found in (Cummins 1996, 2000).
36
Robert Cummins and Pierre Poirier
Accuracy There are two different ways in which indication can be accurate or inaccurate. Simple binary indicators of the "idiot light" variety are either on or off, hence either fight or wrong. An indicator signal in such a case cannot be more or less accurate. But the indicator that produces it may be said, somewhat misleadingly, to be more or less accurate as a function of its reliability. Indicator signals often come in degrees, however. Thermometers can fail to be accurate because they have the amount of something as a target, and they can get the amounts wrong by a greater or lesser margin. The cells studied by Hubel and Wiesel present an interesting case of this kind. A cell firing at less than peak but more than is involved in normal background noise, could be interpreted as indicating: (1) a probability that its target is present; or (2) how close the current stimulus is to the target. Both interpretations allow for individual signals having a degree of accuracy that is independent of the cell's reliability, since both involve indicating the amount of something (probability, nearness). On the other hand, it may be that less than peak firing is simply ignored, in which case the cells in question are, essentially, binary indicators. Indicators say only whether (or how much of) their target is there; they say nothing about what their targets are like. Representations do carry information about what their targets are like, and this makes representational accuracy a far more complex affair than indicator accuracy. Representational accuracy is a relation between representation and a target. Once we recognize that accuracy is a relation between a representation and its target on a particular occasion of its deployment, it becomes clear that representations are not accurate or inaccurate in their own right. What we are calling accuracy thus differs sharply from truth. A sentence may express a true proposition, yet be an inaccurate representation of its target. This is precisely the situation in reductio arguments, where the proposition targeted by the sentence expressing the supposition is (if the argument is successful) a false proposition. Similarly, if you are asked to specify some false proposition about the Eiffel Tower, your performance will be accurate only if you express a false proposition. In general, targets need not be actual states of affairs, objects, events, or whatever. The target a particular token representation is aimed at is fixed by the function of the intender that tokens that representation, together with the facts that happen to obtain at the time. Thus, an intender may have the function of representing the phrase structure of the current linguistic input, and that function, together with the actual (phrase) structure of the then current input will determine what the target of a particular token happens to be, and hence the standard against which the accuracy of the representation is to be measured.
Representation and Indication
37
Propositions are peculiar targets in that they cannot be represented with greater or less accuracy: they are either hit, or missed. There is no such thing as varying degrees of accuracy when it comes to representing the proposition that the Eiffel Tower is in Paris. Someone might say it is in France, and claim that that is "closer" than saying that it is in Europe or Australia. But this is evidently not a matter of getting closer to expressing the right proposition, but a matter of specifying a location that is closer to the correct one. Proposition expressing, at least as the going theory has things, is an all-or-nothing affair (Fodor & LePore 1992). But most representation isn't like this. Pictures, maps, diagrams, graphs, scale models, and, or course, partitioned activations spaces, are more or less accurate representations of the targets they are applied to. It follows from this observation that most representations cannot have propositional contents. It might seem that pictures could express propositions, because they could be said to hold in some possible worlds and not in others. It seems that one could take a video tape of a convenience store robbery, and alter it in various ways so that the variants held in possible worlds that were "close" to the actual world, worlds in which everything went down just as it did in the actual world except that the perpetrator had a moustache, or the clock on the wall behind the clerk said ten twenty-five instead of ten twenty-two. Since a proposition can be conceived as a set of possible worlds (Stalnaker 1984), it might seem that a picture could be regarded as expressing the proposition that consists in the set of possible worlds it actually depicts accurately. But no picture depicts with perfect accuracy. This is not just a consequence of rendering three dimensions in two. One might reasonably hold that the target of a photograph, for example, is not the three dimensional spatial layout, but its two-dimensional projection at a certain plane. Even granting this, however, there is still the fact that getting the color fight often requires inaccurate illumination and a sacrifice in resolution. Depth of field issues will inevitably render some things sharper than others. A video that gets the perpetrator's face in focus but blurs the clock on the wall behind him misrepresents the scene as having a blurry clock. We are not fooled, of course: we know clocks in convenience stores are not blurry. This makes it tempting to suppose that the photo doesn't really represent the clock as blurry. But it does: a photo of things we don't antecedently know a b o u t - one taken through a microscope, for example m can leave us wondering whether we have an accurate picture of a blurry object or a depth of field problem. Similarly, there are many compromises involved in shooting moving subjects with still shots, and even with movies or video. When we take all this into account, we no longer have a set of possible worlds corresponding to a given picture, but a smear of more or less similar worlds spreading out along various dimensions. Is the photo that nails the color of Aunt Tilly's hat while, inevitably, overestimating
38
Robert Cummins and Pierre Poirier
the intensity of illumination and blurring the features of the man running past her, more or less accurate than the one that allows us to identify the man but not the hat? 18 Trade-offs of the sort illustrated by the depth of field problem in photographs are ubiquitous in non-symbolic representational schemes. Such schemes often represent many quantitatively variable properties simultaneously. Photos and models, for example, simultaneously represent color, relative distances, and sizes. It frequently happens that increasing accuracy in one dimension entails sacrificing accuracy in another. In human vision, sensitivity is in conflict with resolution and color accuracy because of the high density of the relatively insensitive cones at the fovea. (This is also why color blind individuals have better night vision.) Sentences (again, as construed by the tradition of truth-conditional semantics) can get around problems like this by abstracting away from some properties while focusing on others, and by abstracting away from troublesome differences in degree. "Roma tomatoes are red when ripe," is simply silent on size, shape, and variation in shade, intensity and saturation. A photo or model cannot do this, nor can biological vision systems. For those of us brought up to think of semantics in a linguistic setting, the striking thing about maps, diagrams, partitioned activations spaces, images, graphs and other non-linguistic representations is that they are not true or false, and that their accuracy comes in degrees. A sentence either hits its propositional target, or it fails. Non-propositional representations require a graded notion of accuracy. Moreover, such representations are typically multi-dimensional. Images, for example, represent (relative) size, shape, color and (relative) location simultaneously. The possibility thus arises that two image representations might be incomparable in over-all accuracy, since one might do better on some d i m e n s i o n s - size and shape, s a y - while the other does better on o t h e r s - color and location. 19 The fact that non-propositional representations can simultaneously represent along many dimensions probably precludes any sort of "all things considered" or "over all" accuracy scores. The concepts of truth and falsehood, and the Tarskian
18We could, perhaps, imagine a possible world whose natural laws dove-tailed with the constraints of a given representational scheme, so that, for example, things flattened into two dimensions when photographed, with objects originally at different distances losing various degrees of resolution themselves. These would be radically counter-nomic worlds. The laws of our world, and every world satisfying anything like the same laws, preclude perfect accuracy for most representational schemes. 19It seems likely that high accuracy on one dimension will often have to be paid for in lower accuracy in others, given limited resources. The eye, for example, gains considerable resolution and color information via foveation, but loses light sensitivity in the process. A map that shows all the streets of London on one page will be either too big to use in the car, or be viewable only with magnification.
Representation and Indication
39
combinatorial semantics we have come to associate with them, will be no help at all in understanding how these non-propositional representations fit or fail to fit their targets. Representational meaning for non-propositional representations will have to be understood in different terms, as will their semantic structures. It is precisely this rich area of inquiry that is opened up when we distinguish representation from indication, and contemplate mechanisms other than those presupposed by the Language of Thought for assembling representations from indicator signals.
References Atick, J. (1992). Could information theory provide an ecological theory of sensory processes? Network, 3, 213-251. Barlow, H. B. (1961). Possible principles underlying the transformation of sensory messages. In: W. A. Rosenblueth (Ed.), Sensory communication (pp. 217-234). Cambridge, MA: MIT Press. Churchland, P. M. (1995). The engine of reason, the seat of the soul: A philosophical journey into the brain. Cambridge, MA: MIT Press. Cummins, R. (1989). Meaning and mental representation. Cambridge, MA: MIT Press. Cummins, R. (1996). Representations, targets, and attitudes. Cambridge, MA: MIT Press. Cummins, R. (1999). Truth and meaning. In: J. Keim Campbell, M. O'Rourke, & D. Shier (Eds), Meaning and truth: Investigations in philosophical semantics. New York: Seven bridges Press. Cummins, R. (2000). Reply to Millikan. Philosophy and Phenomenological Research, 60, 113-128. Cummins, R., Blackmon, J., Byrd, D., Poirier, P., Roth, M., & Schwarz, G. (2001). Systematicity and the cognition of structured domains. The Journal of Philosophy, 98, 1-19. Dretske, F. I. (1981). Knowledge and the flow of information. Cambridge, MA: MIT Press. Dretske, E I. (1988). Explaining behavior: Reasons in a world of causes. Cambridge, MA: MIT Press. Field, D. J. (1987). Relations between the statistics of natural images and the response properties of cortical cells. Journal of the Optical Society of America, A, 4, 2379-2394. Field, D. J. (1994). What is the goal of sensory coding? Neural Computation, 6, 559-601. Fodor, J. A. (1975). The language of thought. New York: T. Y. Crowell. Fodor, J. A. (1987). Psychosemantics. Cambridge, MA: MIT Press. Fodor, J. A. (1990). A theory of content and other essays. Cambridge, MA: MIT Press. Fodor, J. A., & LePore, E. (1992). Holism: A shopper's guide. Oxford: Blackwell. Haugeland, J. (1991). Representational genera. In: W. Ramsey, S. Stich, & D. Rumelhart (Eds), Philosophy and connectionist theory. Hillsdale, NJ: Lawrence Erlbaum. Hubel, D. H. (1988). Eye, brain, and vision. New York: Scientific American Library.
40
Robert Cummins and Pierre Poirier
Hubel, D. H., & Wiesel, T. N. (1962). Receptive fields, binocular interaction and functional architecture in the cat's visual cortex. Journal of Physiology, 160, 106-154. Kuffler, S. (1952). Neurons in the retina: Organization, inhibition and excitatory problems. Cold Spring Harbor Symposia on Quantitative Biology, 17, 281-292. Lewicki, M. S., & Olshausen, B. A. (1999). A probabilistic framework for the adaptation and comparison of images codes. Journal ofthe Optical Society ofAmerica, A, 16, 1587-1601. Livingstone, M., & Hubel, D. (1988). Segregation of form, color, movement and depth: Anatomy, physiology, and perception. Science, 240, 740-749. Newell, A., & Simon, H. (1972). Human problem solving. Englewood Cliffs, NJ: Prentice-Hall. Olshausen, B. A., & Field, D. J. (1996). Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 381,607-609. Olshausen, B. A., & Field, D. J. (1997). Sparse coding with an overcomplete basis set: A strategy employed by V 1? Vision Research, 37, 3311-3325. Olshausen, B. A., & Field, D. J. (2000). Vision and the coding of natural images. American Scientist, 88, 238-245. O'Reilly, R. C., & Munakata, Y. (2000). Computational explorations in cognitive neuroscience. Cambridge, MA: MIT Press. Schiffer, S. (1987). Remnants of meaning. Cambridge, MA: MIT Press. Sherman, S. M., & Koch, C. (1986). The control of retino-geniculate transmission in the mammalian lateral geniculate nucleus. Experimental Brain Research, 63, 1-20. Stalnaker, R. (1984). Inquiry. Cambridge, MA: MIT Press. Stampe, D. (1977). Towards a causal theory of linguistic representation. Midwest Studies in Philosophy, 2, 42-63.
Chapter 3
Representation and the Meaning of Life Wayne D. Christensen and Cliff A. Hooker
'Forty-two!' yelled Loonquawl. 'Is that all you've got to show for seven and a half million years work?' 'I checked it very thoroughly,' said the computer, and that quite definitely is the answer. I think the problem, to be quite honest with you, is that you've never actually known what the question is.' 'But it was the Great Question! The Ultimate Question of Life, the Universe and Everything,' howled Loonquawl. 'Yes,' said Deep Thought with the air of one who suffers fools gladly, but what actually is it?' Douglas Adams, The Hitchhikers Guide to the Galaxy
1. Introduction This paper evaluates some of the implications of dynamical situated agent approaches with respect to understanding high order cognition and theories of representation. We will argue that these new approaches have fundamental and wide ranging implications for cognitive science and philosophy of mind, but that current versions are not sufficiently well-characterised or powerful enough to provide a general conceptual paradigm for cognition. Indeed, they possess limitations that become acute with respect to some of the most important issues, including that of understanding the nature of high order cognition. We explore some of these limitations and provide a brief outline of our own model of self-directedness designed to capture the genesis of high order cognition. Self-directedness is
Representation in Mind: New Approaches to Mental Representation Edited by H. Ciapin, P. Staines and P. Slezak 9 2004 Published by Elsevier Ltd. ISBN: 0-08-044394-X
42
Wayne D. Christensen and Cliff A. Hooker
concerned with anticipative action that involves a balance of reactive and internally driven processes. It is characterised firmly within a dynamically situated context, but goes beyond the conceptual boundaries that have been associated with this type of approach by explicitly modelling the organisation of internal processes of the agent and postulating asymmetries of role between system and environment in the interaction process. At present many philosophers and scientists doubt that dynamical situated approaches offer truly cognitive models: the suspicion is that they are a form of neo-behaviourism that will need to ultimately reintroduce the old cognitivist categories and principles. The limitations that we identify provide a legitimate basis for some of these suspicions. However an overly sceptical response is as misplaced as an incautiously enthusiastic one. All current approaches have serious l i m i t a t i o n s - including cognitivism m and it is very much an open question what type of concepts may play important roles in future cognitive modelling. It is a mistake on both sides, therefore, to interpret the dynamically situated research as supporting an eliminativist cognitive paradigm that would, if successful, dispatch outright all of the old cognitive concepts, like representation, intention and purpose (cf. Hendriks-Janson 1996). Instead, the implications are more subtle: cognitive concepts, old or new, must be reconstructed within a detailed system-environment characterisation in a way that clearly articulates a distinctive role for them in adaptive interaction processes. This is a severe challenge and it is unclear what older concepts will survive and how they might be transformed. In the case of representation the challenge is to pay much greater attention than is usually the case to the role that representation is supposed to play in the adaptive processes of the intelligent agent.
2. What is the Significance of the Adaptive Interaction Approach to Cognition? In the western philosophical tradition the reasoning agent is characterised in a way that is, perhaps unsurprisingly, overwhelmingly preoccupied with ideas. Rationality, and intelligent action, are characterised in terms of relations amongst ideas, and between ideas and the world. In the twentieth century this concem with ideas translated into a strong focus on representation, and a representationally based conception of the nature of cognition called cognitivism. According to cognitivism, representation is the fundamental and ubiquitous unit of all cognitive explanation. It is the key element in the answers to all questions about how we can know about the world and act intelligently in it. Perception, for instance, is
Representation and the Meaning of Life
43
modelled as the acquisition of representations and problem solving as the logical manipulation of representations. However, after the emergence of dynamical situated research in cognitive science representation has become a little like "Forty-two" in the Hitchhiker's Guide to the Galaxy m an answer in search of a question. The old assumption that it had to be the explanation for knowledge and intelligent action, because no other possibility was conceivable, has been undermined by the emergence of alternative forms of explanation and significant empirical evidence that there are other ways to achieve apparently intelligent action. This is combined with the emergence of serious difficulties in applying a representational approach to artificial intelligence. Many people continue to insist on the importance of the concept of representation for understanding cognition, but saying precisely what it is supposed to explain is no longer simple. As is sometimes gleefully pointed out, it is no longer the only game in town. The theme of our paper is that the problem of representation as it has been historically posed in philosophy is ill-formed, and discovering the nature of the question-to-which-it is-the-answer, and hence representation itself, will require embedding representation in an adaptive interaction context. This section examines the nature of the alternative explanations of intelligence that have been put forward in order to evaluate the extent to which they constitute a genuine challenge to cognitivism.
2.1. The Story So F a r . . .
Cognitivism provides a representation --+ semantic inference --+ action model of mind. Because all cognitive processes are assumed to occur after the formationof-representation stage and before the action stage, and because the nature of cognition is assumed to be essentially logical computation, cognitivism yields a self-contained, abstract model and hence a profoundly disembodied picture of cognition. The physical processes of the system play no part in cognition except as inputs, outputs or material instantiations of semantic processes. There has been a minor alternative tradition which holds that physical embodiment is in some way more important than this (e.g. Braitenberg 1984; Dreyfus 1996; Merleau-Ponty 1962), but it made little headway against the cognitivist orthodoxy until the flowering of dynamical systems and autonomous agent robotics empirical research. The impact of this research has been such that cognitivism and the conceptual panoply associated with it are seen to be under strong challenge. This new research can be grouped under the general banner of dynamical situated approaches to cognition, but because there are important differences in content
44
Wayne D. Christensen and Cliff A. Hooker
and focus between the type of research that has adopted dynamics as its general modelling framework and the research that is more specifically focussed on autonomous agent robotics we will initially consider each separately. 2.1.1. Dynamical systems theory One strand of the challenge to cognitivism has come via advocacy of dynamical systems theory (DST) as an important rival for the computer-oriented cognitivist paradigm, most prominently by Tim Van Gelder (Van Gelder 1995, 1998; Van Gelder & Port 1995). 1 Van Gelder points out that one of the main contrasts between cognitivism and the DST approach lies in assumptions about the processes by which coordinated action is achieved. In the cognitivist picture coordinated action is achieved because of commands issued from a central processing module. These commands take the form of specifications of behaviour for motor effectors the overall coordination of motor behaviour occurs because of the intelligent coordination of the commands issued by the central processing module. In contrast, dynamical systems approaches assume that behaviour is governed by high-level collective variables whose state is governed by differential equations. Coordination is not imposed from on high, it is there to begin with. 2 Van Gelder comments, The basic insight is that coordination is best thought of as explained, not as masterminded by a digital computer sending symbolic instructions at just the right time, but as an emergent property of a nonlinear dynamical system self-organizing around instabilities (1998: 616). Van Gelder identifies a number of general implications (here collapsed to three) of shifting from cognitivism to the DST approach: (1) The dynamical approach shifts the emphasis from state descriptions and discretized input-output transitions to ongoing temporally structured processes. (2) In place of (internalist) structural explanations which appeal to functional components of the system, dynamical explanations employ geometric (state space) models which appeal to positions, distances, regions and the properties offlows (ranges of dynamical pathways) of the system as a whole. (3) Interaction is treated as coupling rather than state-setting, and as occurring in parallel rather than serial sequences (1998: w I See also Beer (1995), Smith & Thelen (1993), Smithers (1995). 2 Van Gelder illustrates his discussion using Kelso's "HKB" model of finger coordination (Kelso 1995). Abstractanalysisof its conceptualstructurefails to capturethe persuasiveforce of the dynamical approach as a distinctiveand powerfulformof explanation.To appreciatethis it is necessaryto examine the range and productivenessof the specific models that have been generated. For a sample see Port & Van Gelder (1995), Smith & Thelen (1993), Beer to appear.
Representation and the Meaning of Life
45
In arguing that such a shift should occur, Van Gelder presents four reasons for believing that the DST approach is a plausible empirical hypothesis for cognitive science: (i) Dynamics has been empirically successful, both in science generally as a successful framework for modelling physical behaviour, and in cognitive science in particular in the form of a vigorous school of empirical research employing dynamical models. The fact that many researchers are pursuing the dynamical approach, as sampled in Port & Van Gelder (1995), attests to its fruitfulness. (ii) Natural cognition happens in real time, hence dynamics is better suited to model it than the a-temporal computational approach. (iii) Cognition is embedded: in a nervous system, in a body, and in an environment, whereas computationalism typically abstracts this embeddedness away, and can incorporate it in only an ad hoc manner. (iv) Dynamics can explain the emergence and stability of cognition through self-organisation, whereas cognitivism ignores the problem of cognitive emergence. However, Van Gelder largely sidesteps the issue of what it is to be cognitive. He simply "takes an intuitive grasp of the issue for granted," and rather than attempt to characterise cognition he says, "the question here is not what makes something cognitive, but how cognitive agents work" (1998: 619). Van Gelder does say that cognitive systems are dynamical systems of "special kinds" but leaves it to further empirical work to specify what those kinds are. This deflection of the issue severely compromises the touted status of the dynamical systems approach as a direct mirror of cognitivism. Indeed, it appears that in his formulation of the basic contrast between the "computational hypothesis," characterised as using the resources of computer science as its theoretical basis for modelling cognition, and the "dynamical hypothesis," characterised as using the resources of dynamical systems theory as its theoretical basis, Van Gelder misunderstands the former. Cognitivism is more than just the hypothesis that computer science supplies the appropriate modelling tools for cognitive science it involves a strong positive conception of the nature of cognitive agents. Cognitivism explains epistemic capacity (the system possesses internal symbolical states which representationally encode information about the environment), rationality (in terms of algorithms that optimise for truth preservation when computing belief formation, and for desire satisfaction when computing goal-directed output), and adaptive capacity (in terms of the rational response to the represented state of the environment). Regardless of the ultimate adequacy of these explanations, the status of cognitivism as a cognitive paradigm hinges on the fact that it provides a systematic conception of the nature of cognition. In contrast, DST has no rival cognitive paradigm to offer, only a general recommendation to pay attention to dynamical modelling when studying cognitive phenomena.
46
Wayne D. Christensen and Cliff A. Hooker
We shall return to this issue below, here we note a further technical limitation of the dynamical approach associated with the general nature of DST. DST provides no principled way of modelling system organisation, and in fact tends to suppress it. Recall that of the implications of the dynamical approach identified by Van Gelder, (2) concerned the shift from internal models of functional organisation to geometric state space models. Here the system is modelled as a point within a space whose dimensions stand for the parameters of variation which collectively describe the system's behaviour. Change in state is specified as movement within this space. It is crucial to observe that this type of model does not explicitly describe the physical organisation of the system m a chemical clock and a pendulum, for instance, may be modelled as equivalent dynamical oscillators. It is always in principle possible to capture the internal organisation of a system by modelling it as a system of dynamically coupled component systems. However there is no principled basis for: (a) specifying when such organisation should be explicitly modelled; or (b) individuating the system in a principled way. The choice of what counts as the system, and what structure should be included, is left to the discretion of the modeller. Insofar as the ability to provide generalisations about the organisation of intelligent agents is an important component of a cognitive paradigm, this is a significant defect.
2.1.2. Autonomous agent robotics Another strand of the challenge to cognitivism has come via the success of the autonomous agent robotics research (AAR) program championed by Brooks (1991). In practical application the cognitivist representation --~ semantic inference --~ action model of mind translates into what Brooks has described as a sense-think-act architecture, which he argues is unworkable as an approach to building robots capable of effective real world interaction capability. The problem is that the sense-think-act architecture, employing a representational model of the world in the central module, acts as a bottle-neck to rapid context-sensitive response to a complex environment. Firstly, it is overwhelmingly difficult to represent environments approaching realistic complexity with sufficient detail to permit effective action in the environment because of the astronomical number of things, situations and contexts which may prove relevant to action. Secondly, the sense-think-act cycle results in tortuously slow robot behaviour as the system attempts to update its representational model of the environment and compute the appropriate response at every step. This is a strategy poorly suited for a dynamic environment. Robots built according to cognitivist principles have proved slow, clumsy and brittle, and largely incapable of effective behaviour outside of highly simplified "toy worlds."
Representation and the Meaning of Life
47
The altemative that Brooks proposes is to decompose intelligence by whole system interaction activity instead of intemal functional modules, starting with simple whole systems capable of basic real-world activities and incrementally building the capabilities of the systems by adding "layers" of activity. Since at each step the systems are tested in the world there is far less opportunity for unchecked design error. The design that results from this method is called the "subsumption" architecture, and it is organised in a radically different way to the sense-think-act architecture. Its general features are as follows: (i) multiple parallel active subsystems (in contrast with passive subsystems which await extemal command); (ii) task coherency at every activity level; (iii) modulation relations between activity generating systems (rather than command relations); (iv) response of sub-systems to functional aspects of situations (rather than representations of function-independent things in the environment); and (v) a focus on behaviour production, with global coherency of the robot's state, including its "goals," being largely or entirely implicit in the operation of the robot. 3 Two powerful factors support Brooks' methodological/design approach to studying intelligence: the approach mimics evolution, increasing confidence that the design principles reflect those of natural intelligent systems, and it works m it produces capable robots (Brooks 1991: 149).
2.2. Strengths and Limitations of Dynamical Situated Research The DST and AAR approaches have much in common and their development has not been completely independent. The research of both Beer and Smithers, for instance, could fairly be described as combining the two schools (e.g. Beer 1995; Smithers 1995). However it would be a mistake to simply equate the two. DST and AAR share a common holistic orientation, but from somewhat different perspectives. Roboticists must be concemed with holistic system-environment interaction because a successful robot has to operate as a complete unit in a real environment. This echoes the holism of DST, but roboticists are also concemed with the details of the organisation of the system involved in generating macro-behaviour, since they have to build them. Roboticists have therefore been interested in the capacity to produce independent intelligent behaviour from both a dynamical and an organisational perspective, and indeed the subsumption architecture is a kind of organisational recipe for constructing adaptive robots. 3 For an extended discussion of the subsumption approach that introduces additional principles see Pfeifer & Scheier (1999), especiallyChapter 10.
48
Wayne D. Christensen and Cliff A. Hooker
As a result AAR has developed a level of theoretical generalisation that lies between the abstractness of the DST framework and the concrete specificity of particular dynamical models, such as Kelso's "HKB" model of finger coordination (Kelso 1995), or Elman's models of grammatical competency (Elman 1995). The absence of this kind of generalisation ability is an indicator of DST's weakness when considered as a cognitive paradigm. Thus, whilst DST is certainly compatible with the thesis that cognition is essentially situated, it says little about it. Likewise, it provides no information about the biological context in which intelligence emerges. In contrast AAR research has concerned itself with the evolutionary context of the acquisition of intelligent capacities, the organisation of the system required to produce them (loosely codified in the subsumption architecture principles), and the systematic exploitation of an organised environment to produce adaptive behaviour. This sharper focus makes Brooks' criticisms of cognitivism more specific and penetrating than those of Van Gelder, who concentrates primarily on the general limitations of cognitivist models for capturing the temporality of cognitive action (see also Van Gelder 1995). The silence of DST on these kinds of issues is symptomatic of its abstract g e n e r a l i t y - it is really a general modelling framework and not, per se, a paradigm about the nature of intelligence. Some important characterisations of the role of self-organisation, emergence and the interactive development of skilled activity in intelligence have been proposed by researchers using the DST approach (e.g. Smithers 1995; Thelen 1995), but they are more like partial insights than a full cognitive paradigm. AAR takes us further towards a genuine cognitive paradigm, but its characterisations are limited by the fact that it has been primarily driven by engineering considerations. For instance, although AAR has shown a broad interest in the order in which intelligent capacities are acquired (focussing on insect mobility before human language, for instance) it hasn't been particularly concerned with detailed ethological modelling of organisms or comparative phylogenetic analysis of adaptive capacities, other than as an ad hoc source of inspiration. As a result some of the confidence placed in AAR principles as a foundation for a general theory of situated cognition is misplaced m its agenda had been too narrow to serve this role. The lack of a clear positive conception of the nature of cognition is less of a barrier to dynamically oriented research than it might seem, since the context and target phenomena can be supplied by a particular discipline, such as ethology, ecological psychology, linguistics, etc., to which the modelling tools of DST are applied. Indeed, DST's lack of prescriptive content can be seen as an advantage in the sense that it doesn't impose blinkers on research. Arguably, cognitivism's strong paradigmatic content acted as a dead hand on cognitive science research by
Representation and the Meaning of Life
49
restricting ideas to a very limited set of possibilities. The dynamical approach is liberating precisely because it opens up a rich new spectrum of possibilities, and to try to close these off again might seem premature at best. Similarly, it might be argued, roboticists should be free to take, or ignore, whatever sources of inspiration they see fit, and direct their attention to whatever is necessary for building effective robots, since in the end it is the hard engineering problems that are the best guide to what is relevant. This type of reasoning provides a basis for reinterpreting Van Gelder's stance, not as a claim for DST as a cognitive paradigm in the same sense as cognitivism, but as a heuristic claim about appropriate research strategies at the current historical juncture of cognitive science. Specifically, he might be taken as arguing that: (i) there are no adequate theories of the nature of cognition to hand and it is counterproductive to try to specify one in advance of further empirical investigation; (ii) whatever cognition turns out to be, it is most plausible that it will be a kind of dynamical phenomenon; and (iii) dynamics provides tools and a "way of thinking" about cognitive phenomena which will shape the construction of a theory of cognition. This is a relatively moderate and plausible position; nonetheless, Van Gelder makes a crucial mistake in interpreting the "dynamical hypothesis" (as he formulates it) as a cognitive paradigm. At best it is a road sign pointing towards a paradigm. Moreover, notwithstanding an understandable fear of premature articulation, there are reasons why we might want to look for a positive cognitive paradigm: a paradigm could play a highly useful role in identifying key problems and focussing research effort, and assist in developing a new set of concepts shareable across the cognitive science disciplines. Sheer intellectual curiosity is also important. The absence of an adequate theory of cognition is no reason for avoiding the issue. In pursuit of such a paradigm we broadly agree with Hendriks-Jansen's (1996) strategy of embedding dynamically situated research in an evolutionary/ethological context. We differ in that we think that Hendriks-Jansen's conception of the project is too narrow. In this respect both Van Gelder and Hendriks-Jansen are mistaken in interpreting DST and AAR as conceptual paradigms in the same way that cognitivism is. This puts too much weight on their explicit conceptual structure and leads to over-interpretation of the state of research at a particular point in time. It also misses some of the more subtle methodological implications of the new kind of research. Our preference is to adopt a more open interpretation which sees the DST and AAR research as fertile orienting sources for the development of a cognitive paradigm whose detailed conceptual structure is yet to be established. Such a paradigm need not have the heavily top-down, closed nature of cognitivism, but it should provide considerably
50
Wayne D. Christensen and Cliff A. Hooker
more integrative explanatory power than is currently available to DST and AAR. Specifically, it should incorporate constraints on the nature of intelligence derived from the full set of relevant sciences, including ethology, developmental psychology, neuroscience, cognitive linguistics, and so on. It should possess an integrated set of methods and concepts that coherently address a broadly established set of problems. 4 A more open interpretation is arguably better supported by the broad range of contemporary research practice, as distinct from a few favoured examples, and is more conceptually defensible. An open interpretation doesn't support the sweeping eliminativism that has been pinned on the new research by some of its advocates. Hendriks-Jansen, for instance, argues that cognitive-science should abandon all models of behaviour selection that involve representation, intentions and purposes (Hendriks-Janson 1996: 190-195). This would seem to be a high price to pay for accepting the explanatory benefits of a dynamically situated approach to cognition, nor do the clear empirical successes of dynamically situated research of themselves warrant such a full-blooded eliminativism. A well-founded eliminativism needs to establish two further conditions: (i) Dynamically situated approaches genuinely involve no use of the older cognitive concepts. The mere fact that representation isn't explicitly mentioned, for instance, doesn't ensure that it isn't tacitly presupposed somewhere. (ii) They are complete. That is, they have the conceptual resources to explain all relevant cognitive phenomena. These are very strong conditions and there is little prospect for convincingly demonstrating them. The completeness condition is particularly questionable. The central empirical example of Hendriks-Janson (1996) is a wall-following robot designed by Mataric based on Brooks' subsumption architecture (Mataric 1992; Mataric & Brooks 1990). His argument that representations, intentions and purposes are completely eliminable from models of behaviour selection hinges on the claim that the organisational principles of the subsumption architecture are fully generalisable. In fact, relatively few roboticists other than Brooks ever believed this, and now not even Brooks does (Brooks 1997; Bryson 1999). Hendriks-Jansen's conceptual claim is that no special internal mechanisms need be or ought be postulated to explain choice b e h a v i o u r - apparently goal-directed behaviour is emergent
4 Developing genuinely integrative multidisciplinary research paradigms is a rather new exercise in science. Current evolutionary developmental biology provides a fascinating guide to the nature of such research paradigms, explicitly taking as its problem the task of integrating population genetics, phylogenetic systematics, developmental biology and "model systems" molecular-genetic research (Raft 1996, see also Christensen 2000).
Representation and the Meaning of Life
51
and not localisable to anything more specific than the full system-environment interaction complex. This picture certainly fits the highly parallel "reactive" nature of the subsumption architecture, but there is an important question as to whether this architecture can be scaled up indefinitely to produce more sophisticated intelligence, or whether new organisational principles may need to be introduced at some point. Hendriks-Jansen's claim is that it can be scaled up - - that more sophisticated intelligence is all just "more of the same." However subsumption-based robots have distinct limitations. Internal communication amongst sub-systems is deliberately kept to a minimum to facilitate rapid reactivity to the environment. This can work well if the robot has a relatively simple sensori-motor organisation, an already well-tuned design and a relatively simple task to perform. Problems arise as complexity is added. Increases in the sensori-motor complexity of the robot add degrees of freedom to its dynamics, explosively increasing the possible states of the system. This makes it much less easy to simply rely on an initially well-tuned design to produce coordinated activity, and as complexity increases this begins to impose a requirement for a degree of active self-management (Pfeifer & Scheier 1999). Similarly, increases in task complexity make it increasingly difficult to rely on the structure of the situation for action coordination. If the task involves sequenced action, achieving the right sequences based on environmental cues can require increasingly baroque robot design as the sequences become more complex, and for sufficiently complex sequenced tasks there may simply not be enough appropriate environmental cues (Bryson 1999). Complex tasks may also require persistence of activity in the face of potentially distracting stimuli. In certain situations this requires that the robot ignore stimuli it would react to in other circumstances. There may be multiple conflicting possibilities for action in a given context, requiting the robot to have preferences over action possibilities. The task might also display strong temporal sensitivity in the sense that small differences in the nature or timing of the robot's actions may make a large difference in outcome. This can require that the system be capable of learned anticipation (see Christensen & Hooker 2000a). Each of these types of task-complexity impose a requirement for increasingly active, self-directed behaviour on the part of the robot. Brooks confronted exactly these type of design issues in his Cog project (Brooks & Stein 1993). In an extraordinary shift of research gears Brooks' switched from attempting to model insect-level intelligence with subsumption-architecture robots to modelling human intelligence by building a humanoid robot capable of interacting in a human-like fashion. The goal of the project was that the robot ("Cog") would, "learn to 'think' by building on its bodily experiences to accomplish progressively more abstract tasks" (Brooks & Stein 1993: 7). In
52
WayneD. Christensen and Cliff A. Hooker
summarizing the implications of the Cog project, Brooks (1997) speaks of two shifts of view point in thinking about intelligence - - the first shift being the shift to the behaviour-based approach outlined above, while the second shift is from behaviour-based to "cognitive robotics" (or "cognobotics," as he terms it). This "second shift" signals a recognition that the subsumption architecture design strategy failed the scaling p r o b l e m - that qualitatively new design principles are required to move significantly beyond insect-level intelligence. Brooks identifies a number of the problems raised above and lists several new design principles for dealing with them, the most important here being motivation, coherence, self adaptation, and development (1997: 297). The motivation principle is the requirement that the robot have preferences that context-sensitively guide action. For example: Suppose the humanoid robot is trying to carry out some manipulation task and is foveating on its hand and the object with which it is interacting. But, then suppose some object moves in its peripheral vision. Should it saccade to the motion to determine what it is? Under some circumstances this would be the appropriate behavior, for instance when the humanoid is just fooling around and is not highly motivated by the task at hand. But when it is engaged in active play with a person, and there is a lot of background activity going on in the room this would be entirely inappropriate. If it kept saccading to everything moving in the room it would not be able to engage the person sufficiently, who no doubt would find the robot's behavior distracting and even annoying (Brooks 1997: 298).
Coherence is the associated problem of ensuring that action choices do not conflict. The problem of self-adaptation concerns the fact that the many sub-systems of the humanoid robot will be liable to change their characteristics due to changing context or temperature or drift, etc. The sheer complexity of the system prevents global pre-set calibration. Instead the system must continuously self-calibrate i.e. be self-adapting. The problem of development concems the fact that human intelligence develops through an extended series of self-scaffolding stages of coordinated learning processes. Brooks argues that any practical approach to imitating human intelligence will therefore require recapitulating this process. It is possible that Hendriks-Jansen would argue that these new design principles are really just more of the same, a matter of fine-tuning the basic subsumption architecture. Compared with classical designs Cog is still highly parallel and
Representation and the Meaning of Life
53
uses the structure of the environment to perform tasks in a richly interactive way. Therefore the fundamental principle of understanding intelligent action as emergent from interaction remains true. However this kind of defence would overlook the fact that achieving increasingly sophisticated action requires increasingly active self-management on the part of the robot. In other words more internal modulatory processes are required than recommended by the subsumption design principles. This violates the principle that there is no essential difference in role between the contribution of the system and the contribution of the environment to the overall interaction process. In fact, a special functional role is played by the self-management processes, moreover these appear to involve distinctive internal intentions and goals. Though they are not realised in the classical cognitivist way, the motivations and preferences that Brooks talks about have a clearly intentional flavour. As a consequence these processes start to appear cognitive in a somewhat old-fashioned sense. Brooks did, after all, choose to label his second shift "cognobotics." 2.2.1. Self-directedness Elsewhere we have developed a model of selfdirectedness that takes this transition from largely reactive to increasingly internally directed adaptive organisation as the genesis of cognitive intelligence (Christensen & Hooker 2000a, b). Self-directedness involves the active differentiation of adaptive relations as the basis for appropriate modification of action. The concept involves some subtle distinctions, but it essentially concerns the sophistication of the system's processes for targeting action. As systems become more self-directed they are able to interact effectively in more variable contexts, their ability to coordinate performance is more focussed, and they are able to explicitly recognise more of the adaptively significant relations involved in the interaction process. Thus cheetahs appropriately modify aspects of their hunting, such as stalking direction and distance, in response to the nature of the prey (alert or not, large or small, etc.), and terrain (e.g. wooded or open), and how hungry the cheetah, and possibly its cubs, are. The fundamental distinction involved in self-directedness is between adaptive relations that are actively differentiated by the system and those that are not. To understand this distinction it is important to realise that simply differentiating a signal is not the same as differentiating the adaptive relations that make that signal important for the system. Mosquitos differentiate CO2 concentration, but they do not seem to internally differentiate the relation between CO2 and the presence of a blood host. Instead they act like a Brooksian robot the CO2 signal governs an action module that causes the mosquito to fly up the CO2 gradient. Upon contact with the source of the signal, usually a blood host, flight behaviour is suppressed and another action module generates feeding behaviour
54
Wayne D. Christensen and Cliff A. Hooker
(Klowden 1995). Hendriks-Jansen is quite correct to argue that a great deal of adaptive action can be achieved in this way. He is wrong, however, to claim that it is all achieved this way. Bees provide a simple example of adaptive interaction processes that do involve active differentiation by the system of an adaptive relation. For bees the relationships between flower availability, flower colour, and nectar yield are both variable and adaptively significant. They discover which relations are useful interactively through foraging - - they sample the flower types within their range and preferentially visit those with an adequate nectar reward (Real 1991). In so doing they differentiate an adaptive relation - - the correlation between flower colour and nectar reward ~ and use this to appropriately modify behaviour by preferentially visiting flowers whose colour is correlated with nectar reward (see Montague et al. 1995). Self-directedness involves anticipation and evaluation: in the bee example the association of flower colour and nectar reward is the anticipation, and the gustatory signal that indicates the amount of nectar actually acquired at the flower is the evaluation. Bees are only weakly self-directed because their capacity to evaluate interaction and form anticipations is fairly limited. Stronger forms of self-directedness require more powerful learning processes for forming anticipations. This is required for performing tasks with the complex coordination requirements and sensitivities discussed above. For example, cheetahs have to coordinate many conditions if they are to hunt successfully, including stalking well enough that they do not alert the prey too early and allow it the opportunity to escape. Performing tasks of this nature imposes particularly strong demands on a system's capacity for anticipation, since the system must be responsive to extended temporal patterns in interaction. In particular, the system must continuously predictively modulate action so that interaction in the context coordinates all of the conditions required to perform the task. Thus, throughout the stalking phase of the hunt cheetahs shape their actions so as not to startle the prey, and after the chase is initiated they continuously respond to the prey's movement, e.g. seeking to separate it from the herd. In general, improvements in self-directedness allow the system to shape its actions over longer timescales and with respect to more detailed, in some cases modal, information concerning the flow of the interaction process. 2.2.2. Some preliminary conclusions The sudden appearance of new design principles in the Cog research and the self-directedness model ~ which violates prescriptions against functional analysis, postulating content, and asymmetry of role between system and environment ~ serve as cautionary points against
Representation and the Meaning of Life
55
an over-enthusiastic belief in the explanatory power of modelling concepts and techniques of the DST and AAR research programs at such a nascent state of development. Until there is much more solid comparative ethological data matched with detailed models of sophisticated intelligent capacities any claim to possess a cognitive paradigm with general explanatory power must be given a low degree of confidence. Thus, in evaluating the significance of the dynamical situated approach it is important not to be seduced into over-interpretation and over-generalisation. There is a difference between: (1) Thinking that, after the impact of dynamical situated modelling approaches has been absorbed, the conceptual landscape of cognitive science and philosophy of mind will be significantly changed in virtually all respects. (2) Thinking that the extant dynamical situated modelling approaches provide coherent, well characterised and broadly sufficient conceptual frameworks for cognitive science and philosophy of mind. We believe that the former is probably true and the latter is almost certainly not true. Specifically, as already discussed, there are serious identifiable limitations with the DST and AAR approaches that count against interpreting them as a brave new paradigm for cognitive science, or at least as a paradigm with the kind of positive conceptual content that cognitivism has. The problem with Van Gelder's "dynamical hypothesis" is that it provides no positive account of the nature of intelligence, and has trouble modelling organisation. In this respect the reluctance of dynamical modellers to countenance functional analysis can be seen to stem, at least in part, from a technical limitation of the dynamical modelling framework. Dynamical models certainly don't explicitly model functional organisation, but that doesn't mean that functional organisation is not important. Whilst it is certainly true that the complex interconnectedness of nonlinear dynamical processes and their tendency to exhibit emergent phenomena strongly mitigates against the simplistic top-down functional decomposition strategies of classical artificial intelligence, this does not mean that all analysis of functional organisation is useless. In fact, as noted above the subsumption architecture can be thought of as an organisational recipe for producing an adaptive system. Moreover there may be situations in which it is justifiable to use models of intelligent interaction that postulate somewhat localised functional specialisation, as our account of self-directedness does. The problem with using AAR research as the basis for a cognitive paradigm is that the basic subsumption architecture fails the scaling test, and fails to explain self-directed behaviour. This justifies a concern that such a paradigm leaves out phenomena of central cognitive importance. These negative conclusions should not serve to diminish the strong positive role that dynamical situated research has played in opening up and advancing
56
Wayne D. Christensen and Cliff A. Hooker
cognitive science. Insofar as Van Gelder's goal has been to provide conceptual and rhetorical "elbow room" for such research his advocacy has been valuable and successful, though the strength of the research itself has played a large role in the attention it is receiving. But the cause is not furthered by a bitter war of rhetoric between those who feel that a dynamically situated approach fundamentally changes cognitive science and those who feel that something of key importance has been left out, since both may be right.
3. Is the Glass Half Empty or Half Full? Implications for Theories of Cognition and Representation Given the limitations of current dynamical situated approaches, can cognitivists take heart and assume that the implications of the challenge are minimal? The answer, we believe, is no. One of the most straightforward minimalist positions is a hybrid view that sees dynamical situated models as suitable for capturing low order adaptive interaction processes and old-style symbolic models as accounting for high order cognition. This looks appealing because it appears to combine the strengths of dynamic situated principles in producing effective interaction with the strengths of classical approaches in modelling abstract reasoning, and some researchers are implementing the idea by developing hybrid architectures which couple parallel distributed systems with more conventional symbol processing systems (e.g. Hasemann 1995; Hexmoor et al. 1997; Malcolm 1997). However such a hybrid position fails to adequately take into account the limitations of cognitivism in relation to the strengths of dynamical situated models.
3.1. Fundamental Problems with Cognitivism
Cognitivist artificial intelligence faces fundamental problems, even in its own claimed domain, including the symbol grounding problem, the frame problem, and the combinatorial explosion problem (see Bickhard & Terveen 1995; Janlert 1987). One manifestation of these problems has been difficulty in creating agents capable of successful sensori-motor interaction in a complex, noisy, dynamic environment, 5 however the problems run deeper than an inability to produce mundane sensori-motor coordination in real time. There are also
5 See for example the Shakeyproject (Raphael 1976).
Representation and the Meaning of Life 57 difficulties modelling high order cognitive abilities like generalisation, creativity and learning. In particular, the frame problem has beset cognitivist artificial intelligence attempts to simulate realistic, flexible cognitive abilities. The central issue of the frame problem is the capacity to deal with relevancy and similarity relations, and generalise, in complex situations. Cognitive agents must be able to focus on the adaptively significant performance relations in situations where they haven't been carefully pre-specified for the agent (cf. Dennett 1984). This involves being able to differentiate relevant factors in the context from amidst a sea of details, where what is relevant can vary with changes in the environment, the agent and its goals, and the actions being performed. In addition to being sensitive to features of the environment relevant to performing a specific task, agents should be able to generalise to new contexts, employing existing skills where the relevant factors are similar, and modifying those skills, and developing new ones, where the relevant factors differ. Unfortunately, cognitivist systems have proved extremely poor at these abilities, typically not functioning effectively outside narrow, well defined problem domains. Arguably, the source of the frame problem is deep within the cognitivist framework, and lies in the limitations of formal symbolic models (Bickhard & Terveen 1995). Specifically, the completely formal conceptualisation of cognition in terms of symbolic algorithms involves the assumption that the basic substrate of cognition is a fully explicit symbol system, and that cognition takes the form of algorithms operating within this symbol system. The symbol system determines the basic representational space within which a problem domain can be defined. Problem solving algorithms operate by searching the possibilities in some way until the one constituting the solution is found. In other words all of the possible states must already be represented before problem solving can begin. Not only is the basic structure of this representational space fixed, it has no inherent topology on which similarity might be g r o u n d e d - any representation is as near or far from any other (distance in fact has no meaning). As a result, any similarity relations that are available to the system for problem solving have to be explicitly and individually encoded as representational superstructure (Bickhard & Campbell 1996). Hence the frame problem. 6 As such, the frame problem cuts to the heart of the cognitivist account of high order cognition because an adequate theory of cognition ought to explain One kind of attempt at solvingthe problemof creativityin this approachinvolvesadding meta-rules to modify lower-orderrules (Boden 1996). Howeverthe meta-rules have to be already built in and the basic rule-space definedby the symbolsystemcannotbe modified, so the fundamentalproblemis postponed rather than resolved. 6
58
WayneD. Christensen and Cliff A. Hooker
the ability of intelligent agents to determine relevancy, generalise, and learn. Indeed, it is arguably the ability to solve open or "ill-posed" problems that is the hallmark of cognitive intelligence. Complex real world cognitive problems such as driving a car, starting a business, or designing a scientific experiment, are typically multifactorial, context-sensitive, and not fully formulated for the agent in advance. To solve such problems agents must cope with idiosyncratic and partially unknown situations, and engage in interactive, self-modifying learning processes that use the environment as a resource to discover enough features of the problem that a solution can become apparent (cf. Hutchins 1995). Determining relevancy, generalising and learning are thus clearly central to sophisticated cognition, and therefore the supposed strengths of cognitivism for modelling high order cognition appear just as questionable as its ability to model basic adaptive interaction.
3.2. Strengths of Dynamical Situated Models Moreover a dynamical situated approach offers some potentially fruitful theoretical resources for understanding these phenomena. One of the most basic differences with cognitivism is that the generation of adaptive novelty is not fundamentally problematic for a dynamical situated approach, as it is for cognitivism, because there is no conceptual restriction to a fixed, intemal, fully explicit symbol system architecture. Dynamics offers a rich framework for modelling interactive, emergent properties. In particular, dynamical interaction with an organised environment provides a promising avenue for understanding constructive skill-based learning processes in which new cognitive abilities develop (see, e.g. Fischer & Biddell 1998; Smith & Thelen 1993). Likewise, a dynamical situated perspective provides (at least) three mechanisms for understanding how relevancy is achieved: (i) Because situated agents don't need to represent the environment in a generalised internal model, or function as general computers, they can be "pre-tuned" (by evolution or an engineer) to respond to the stimuli relevant to task performance. This is the strategy of the subsumption architecture. However, although performance tends to be considerably more robust than classical systems due to the advantages of continuous parallel processes over central representational control, as we saw in Section 2.2 the subsumption architecture does face its own version of the flame problem: as complexity increases the number of relevant interaction possibilities can overwhelm the system. Fortunately, (ii), more active "on board" determination of relevancy can be achieved by affective processing which evaluates interaction conditions by assigning them normative valency, in the way that bees learn to
Representation and the Meaning of Life 59 associate visiting flowers of particular colours with the amount of nectar they receive from those flowers, discussed above. The example is elementary, but in principle affective processing can be highly sophisticated. In addition, (iii), interaction guided learning can generate increasingly fine-grained differentiation of relevancy as skill development generates new information about the significant features of the environment (see the discussion of self-directed anticipative learning below). Similarity and generalisation are also more tractable for a dynamical situated approach, in part because of the continuous nature of dynamics (in contrast with the discrete nature of symbol systems). Neural nets, e.g. inherently tend to group similar stimuli and generalise prior learning to new situations. Situated agents that are (at least approximately) "pre-tuned" in the manner above can use dynamical neural net learning or equivalent to refine sensorimotor coordination so as to achieve skilful interaction ability. In the process they will tend to learn about the interactive characteristics of the environment relevant to adaptive action, with all of the dimensions of similarity and difference that this involves. In Section 2.2 we argued that for sufficiently complex agents this learning will involve anticipation, and in our view this is the locus of concept formation. Whilst the general phenomenon of emergence in dynamical systems has been widely noted as a virtue by advocates of a dynamical approach, the related but less recognised phenomenon of nonstationarity - - or change to the dynamical form of the system over time m is perhaps just as significant for understanding cognition. Quartz & Sejnowski (1997) argue that nonstationary learning systems have a number of key advantages over classical learning systems. 7 One of the fundamental problems for classical models of learning is the restriction to a fixed architecture, which means that the system must possess its full learning capacity before any learning begins, and that learning systems capable of solving complex problems must deal with the full complexity of the problem domain at the outset, resulting in an extremely large and potentially intractable search space. In contrast, nonstationary learners change their learning architecture as they learn, e.g. by adding and/or deleting neural interconnections. Rather than grappling with the full complexity of the final solution at the outset, nonstationary learners can start with a simple architecture that solves a partial, but tractable, approximation to the overall problem. As competency develops new learning architecture is progressively added, allowing the problem domain itself to guide the development of learning capacity.
7For a discussion see Christensen & Hooker(2000a).
60
Wayne D. Christensen and Cliff A. Hooker
Each of the dynamical concepts just described provide important insights for understanding cognition, but a strong part of the appeal lies in the way they can be combined to form complex integrated models. For example, Christensen & Hooker (2000a) combines the ingredients given above, including affective processing, anticipation formation, and nonstationarity, with the account of self-directedness given in Section 2.2, to develop an integrated account of self-directed anticipative learning processes. In self-directed anticipative learning processes agents learn about the nature of the problem in the process of task performance by generating and refining anticipations. In order to produce appropriate action the agent integrates affective and contextual information from multiple partially overlapping sources. The integration of these sources of information provides constraints which distinguish better from worse situations and help to focus performance. And as the interaction occurs it generates information that allows the agent to construct anticipative models of the interaction process. This allows the agent to: (a) improve its recognition of relevant information; (b) perform more focussed activity; and (c) to evaluate its performance more precisely. Over time a pushme-pullyou feedback cycle develops in which learning improves the agent's ability to anticipate, which in turn facilitates further learning. Some dynamical situated concepts for understanding high order cognition Interaction + emergence:
Novelty in learning.
Embodied biases + affect:
Relevancy, similarity.
Interaction guided learning:
Relevancy.
Interactive anticipation (dynamical neural net learning, sensorimotor coordination):
Similarity, generalisation.
Nonstationarity:
Tractable learning in complex domains, improving learning competency as part of learning process.
Affect + anticipation + nonstationarity + self-directedness:
Self-directed anticipative learning.
The discussion of the concepts which a dynamical situated approach can contribute to an account of high order cognition is summarised in the table above. Our purpose here is not to argue for their virtues in detail or to suggest that
Representation and the Meaning of Life
61
they provide a demonstrable ability to model high order cognitive phenomena. Of necessity much of the detailed modelling and empirical research has so far been oriented towards the lower end of the cognitive spectrum, and the implications for modelling high order cognition are largely suggestive. Rather, we are suggesting that a dynamical situated approach does address issues that are central to understanding high order cognition issues which have been highly problematic for cognitivism. In other words, rather than staying safely confined to low order adaptive capacities, the dynamical situated approach "leaks up" into sophisticated cognition. Thus, while current dynamical situated research does not supply a ready-made cognitive paradigm, it does offer some potentially interesting ways of tackling central cognitive phenomena that cognitivism struggles with. We started this section by raising the question of the prospects for a hybrid dynamical situated/cognitivist approach. For the reasons discussed we believe that such an approach is not the most promising research program for modelling high order cognition: the cognitivist half of the marriage is in trouble even in its own domain, and there are grounds for optimism about the potential for extending the dynamical situated half to encompass higher order cognition. Coupling parallel and traditional symbolic systems just puts off the problems, it doesn't solve them. But if not a straightforward hybrid approach, then what? A more thorough-going strategy involves taking a dynamical situated approach as the context for developing a theory of high order cognition.
4. Reconstructing Cognitive Concepts in a Dynamical Situated Context Arguably the most important general characteristic of the dynamical situated approach is that it involves modelling intelligent agents as integrated interactive systems. As such, its main contrast with classical philosophical and artificial intelligence approaches is that it does not attempt to characterise and treat separately different aspects of cognition, such as sensing, representing, planning, acting, learning, etc. The problem with such partitioning, from the dynamical situated perspective, is that it leads to models that are under-constrained and frequently taxonomically misleading because they formulate the issues at a very abstract level without paying sufficient attention to the processes by which intelligent agents actually work. (For an illustration of the difference these latter factors can make see the discussion of intentionality below.) The real force of the dynamical situated approach is not, therefore, that it motivates a sweeping
62
Wayne D. Christensen and Cliff A. Hooker
anti-representationalism, but that it presses firmly home questions about the constraints under which theories of cognition are constructed. 8 It follows, in particular, that dynamical situated research should not be thought of as automatically supporting an eliminativist prohibition on theorising cognitive concepts, instead it should be thought of as strongly challenging the cognitivist versions of those concepts. Ironically, eliminativists too easily accept cognitivism on its own terms and evaluate it as a complete package. That is, they accept that if the concepts of intentionality, representation etc., are anything they are much as the received cognitivist view says they are. Cognitivism is in trouble, the argument goes, so these concepts probably don't correspond to anything, and a radically new conceptual taxonomy is required. But by entirely excluding the conceptual resources of an agency-oriented characterisation of higher order cognition eliminativism gives up too much strategic ground. There may be other ways to understand intentionality, representation or other cognitive concepts than the way cognitivism does. For example, Christensen & Hooker (2000b) use the self-directedness model outlined above as a central component of an account of intentionality. This account conceptualises the "aboutness" of intentionality in terms of the complex relational structure of adaptive interaction processes. Adaptive systems must direct the overall interaction process flow in ways that satisfy many viability constraints (cf. energy, injury and other constraints on cheetah hunting). The relation between the processes by which action is generated and the constraints it must satisfy constitutes the basic way an organism orients itself in its world with respect to its own constitution and is thus the basic form of intentionality. In the case of simpler adaptive systems virtually all of these constraints are satisfied implicitly as outcomes of the overall interaction process. However as systems become more self-directed they increasingly begin to explicitly differentiate adaptive relations in order to guide action, and as a result their intentionality becomes, at least in part, more explicit. Further, it becomes more open to systematic elaboration as intentional content is formed through the interactive learning processes by which the system differentiates adaptively relevant relations. For example, the cheetah learns to differentiate the relation between stalking distance and catching
8 On this view it is the general modelling approach that is important, which should not be confused with the specific modelling tools. DST is one of the tools available for modelling intelligent agents as integrated systems, but not necessarily one that is fully developed or sufficient. DST may need to be extended to successfully model intelligence, and it may also need to be supplementedwith qualitatively different kinds of concepts and models. What those are remains to be seen, but organisational characterisations -- of the kind we consider below, for example -- are likely to be central to developing the generalisations about intelligence.
Representation and the Meaning of Life
63
prey. This concept of intentionality is designed to take seriously the idea that intelligent agency has a distinctive "aboutness" to it, but it is significantly different from the classical analytic conception of intentionality because it doesn't take reference as the basic intentional relation. It is in fact inspired by, and partly a development of, Merleau-Ponty's embodied conception of intentionality (Merleau-Ponty 1962). The relevant point to make here is that in developing this account of intentionality we have taken adaptive interaction processes as the primary context for framing the issues. This context serves both as the desiderata for successful e x p l a n a t i o n - intentionality must be shown to play an important explanatory role in understanding adaptive i n t e r a c t i o n - and as the resource base for constructing concepts to solve the problem. One of the benefits of this kind of approach is that by posing the problems in a more sharply defined context a dynamical situated approach can lead to the discovery of latent vagueness and ambiguity in more abstract philosophical and artificial intelligence theories. In this sense the problem with more abstract approaches is not that they don't attempt to deal with genuine and important cognitive issues, but that the issues are under-specified. 9 In the case of representation the most important poorly specified issue is the role it is supposed to play in the processes of adaptive coordination between an intelligent agent and its environment. According to cognitivism centralised representational control is supposed to explain all adaptive coordination between systems and environment, but dynamical situated research demonstrates that this is clearly not the case. At least some, and arguably a very significant amount, of the adaptive coordination involved in intelligence arises through distributed interaction processes. The success thus far of dynamical situated approaches seriously undermines the assumptions of passive-body-awaiting-command and environment-as-mere-problem which are an integral part of the cognitivist central control picture of intelligence. The new default assumption should be that cognitive processes are occurring in a dynamic and highly parallel context. Consequently a more refined specification of the problem of representation is required which points to a sub-class of adaptive interaction processes that have a distinctive
9 One kind of complaint that might be made against our call to include adaptive interaction issues as adequacy constraints on theories of representation is that the problem of understanding representation is hard enough with the constraints cognitivism treated as relevant m adding more "interaction" constraints seemingly only makes the problem worse. However such a criticism ignores both the space opened up by the rejection of several cognitivist constraints (most obviously the assumption that there is a single class of structures that underlie all cognitive capacities) and the constructive role that the additional factors can play by serving as resources for improving the characterisation of what a solution must be like. It is, after all, particularly difficult to solve a partially framed problem.
64
Wayne D. Christensen and Cliff A. Hooker
representational component. Clark's attempt to formulate the locus of representational processes in terms of "informationally hungry" interaction problems (Clark 1997: 166-170) is the fight kind of thing to do, but is itself somewhat vague and risks circularity. A more comprehensive account should involve a full system-environment specification that relates ecological possibilities and constraints with system architecture and interaction processes. More specifically, the traditional focus on covariance between system states and environment states, which lies behind a cognitivist conception of reference and intentionality, must be expanded to include downstream modulation of action and interactive differentiation of adaptive relations. I~ In the process the narrow conception of representation as a unique entity should be broadened to encompass a process-oriented account of content that recognises a spectrum of relations between system and environment that may be involved in any instance of adaptive behaviour, from relatively elementary Brooks-style environmental scaffolding of behaviour to high order goal-oriented behaviour driven by internal organisation capable of anticipatively exploring possibilities. 11 It seems very unlikely that there is a unique type of structural entity in agents that combines all of the attributes run together in standard accounts of representation, such as covariance, reference, and functional role. Nor is representation likely to be the fundamental "mark of the mental" that subserves all cognitive functions. An interactively grounded account of content should focus first on the processes by which systems govern action production, and it should relate content to interactive relations. From the perspective of the adaptive system the primary problem is to produce action appropriate to the context, not referentially individuate a signal source. Whilst teleosemantic approaches do go somewhat in this direction by adding functional role as part of the conditions for individuating content, functional role serves merely as a mechanism for picking out a correlation. It makes no substantive contribution to the nature of the putative content, which is assumed simply to be a referential relation. Moreover the account of functional role is extremely coarsely specified, in terms of a selectively favoured function. Little attention is given to the way in which the system actually shapes the interaction process (Christensen & Hooker 2000b). As such, teleosemantics fails to distinguish amongst the various ways adaptive interaction processes can be organised, or pay sufficient attention to the phenomenon of holistic interactive emergence which motivated the AAR rejection of representation. It simply makes to As is discussed in the account of self-directedness, simple signal detection does not count as differentiation of an adaptive relation. 11Cf. Grush's discussion of "off-line" modelling (Grush 1997). See Christensen & Hooker (2000b) for discussion of the spectrum of anticipative capacities.
Representation and the Meaning of Life
65
the standard cognitivist assumption that adaptively appropriate action is mediated by representation. To get beyond this and characterise a distinctive role for representation in interaction it is necessary to shift the emphasis to the detailed organisation of processes of action production. Such a shift should also broaden the account of content from a narrow focus on covariance to include action and interactive outcomes. One of the central constraints that an interactively grounded theory of content should satisfy is that the nature of the content is accessible to the system. That is, it plays a substantive role in the way the system behaves. For example, if the system misrepresents, it should in principle be able to detect the fact. 12 Including an account of the processes by which systems differentiate interactive relations provides a means for grounding content in this way. This also provides a way of relating content to learning, in terms of the processes by which systems utilise interactive success and error for modifying performance. An interactivist account of content should therefore also be a constructivist account which sees content as being formed through learning processes (see Christensen & Hooker 2000a). Again, approaching content in this way forces the account to be interactively relevant. A theory of reference could be part of this approach, but not the most basic part. An account of reference that satisfied the constraint of system-accessibility and interactive grounding would have to show how the system is able to interactively differentiate the entity referred to as a distinct entity from out of the background of a relatively continuous environment. 13 Since this will be a resource-intensive process involving considerable cross-correlational processing object permanency conditions like re-identifiability across time and from multiple perspectives must be satisfied the account should also specify the circumstances under which the nature of the interaction task makes it adaptively important for the system to have this capability. Finally, the most promising way for an interactive approach to grapple with the issue of semantic systematicity is to look for it in the development of systematically organised activity complexes, rather than internal formal rule systems. These are the kinds of constraints that interactively grounded theories of semantic content and representation should address. There is not the space to show them
12Bickhard (1993) develops an account of representationbased on this constraint with which we are in general agreement. The teleosemantic account of content doesn't satisfy the system accessibility constraint, since an adaptive system typically doesn't know its selection history and therefore doesn't have access to the content of its representations. As a result it is unable to detect whether they have succeeded or failed. See Bickhard & Terveen (1995), Christensen & Hooker (2000b). 13Bickhard & Ritchie (1983) approach this issue in terms of interactive invariances m a system differentiates an object if there is a relativelyclosed set of interactions it can have with it.
66
Wayne D. Christensen and Cliff A. Hooker
in action, as it were, and the list is not intended to be comprehensive. However, they are designed to confront the issue of finding a distinctive role for cognitive concepts in understanding dynamically situated agents. They take seriously both the implications of dynamical situated research for understanding cognition and some of the concerns of older philosophical and artificial intelligence approaches with understanding "content driven" cognitive processes. Adequately dealing with both sets of constraints requires a wider agenda than either side has so far proposed, and the process of reconstructing cognitive concepts in a dynamically situated context also requires transforming them. Some of the classical concerns must be recognised, but the eventual result may look rather non-classical. In this respect the implications go beyond cognitive science to philosophy more broadly. One of the central agendas of twentieth century analytic philosophy was the attempt to collapse meaning to reference and syntax. This is arguably one of the central flaws in cognitivism. The problem here is therefore not that cognitivism attempts to take seriously the issue of reference, but that it attempts to make reference the fundamental key from which all else follows. As we have been arguing, developing a naturalised embodied account of meaning requires broadening the agenda, and there are other philosophical traditions that can make a useful contribution. Merleau-Ponty's embodied account of intentionality provides a rich and useful alternative to the Brentano-based reference model. Moreover the overall approach to content we are advocating has a somewhat Peircian flavour because it relates meaning to expectancies about interaction, and it examines the learning processes by which concepts gain meaning. 14
5. Conclusion Cognition is ultimately an aspect of the dynamical coordination process between system and environment. It is undoubtedly embedded in, and exploits, very rich dynamical self-organising processes. Any would-be theory of cognition must give full recognition to this. It must also give full recognition to the distinctive cognitive abilities of intelligent agents. The problem for philosophy and cognitive science lies in reconciling these issues. Theories of representation aim to explain some aspect of cognition, however working out what aspect they are supposed to explain requires understanding in considerable detail the adaptive interaction
14 For a lucid discussion of Peirce's account of meaning in this vein and its relation to the analytic tradition see Legg (unpublished).
Representation and the Meaning of Life
67
processes that have been the primary focus of dynamical situated research. In effect, we must abandon our preoccupation with representation as the answer and work out what the question is.
References Beer, R. D. (1995). Computational and dynamical languages for autonomous agents. In: R. E Port, & T. Van Gelder (Eds), Mind as motion: Explorations in the dynamics of cognition. Cambridge, MA: MIT Press. Bickhard, M. H. (1993). Representational content in humans and machines. Journal of Experimental and Theoretical Artificial Intelligence, 5, 285-333. Bickhard, M. H., & Campbell, R. L. (1996). Topologies of learning and development. New Ideas in Psychology, 14(2), 111-156. Bickhard, M. H., & Ritchie, D. M. (1983). On the nature of representation: A case study of James J. Gibson's theory of perception. New York: Praeger. Bickhard, M. H., & Terveen, L. (1995). Foundational issues in artificial intelligence and cognitive science: Impasse and solution. Amsterdam: Elsevier Scientific. Boden, M. (1996). The creative mind: Myths and mechanisms. London: Weidenfeld and Nicholson. Braitenberg, V. (1984). Vehicles: Experiments in synthetic psychology. Cambridge, MA: MIT Press. Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47, 139-159. Brooks, R. A. (1997). From earwigs to humans. Robotics and Autonomous Systems, 20, 291-304. Brooks, R. A., & Stein, L. (1993). Building brains for bodies. MIT AI Lab Memo, 1439 (August). Bryson, J. (1999). The study of sequential and hierarchical organisation of behaviour via artificial mechanisms of action selection. M.Phil., University of Edinburgh 1999. Christensen, W. D. (2000). Biology as a matrix science. Presented to the Australasian Association of the History and Philosophy and Social Studies of Science, 28 June-3 July, Sydney 2000. Christensen, W. D., & Hooker, C. A. (2000a). An interactivist-constructivist approach to intelligence: Self-directed anticipative learning. Philosophical Psychology, 13, 5-45. Christensen, W. D., & Hooker, C. A. (2000b). A general interactivist-constructivist model of intentionality. In: J. Macintosh (Ed.), Contemporary naturalist theories of evolution and intentionality, Canadian Journal of Philosophy (Special Supplementary Volume). Clark, A. (1997). Being there: Putting brain, body and world together again. Cambridge, MA: MIT Press. Dennett, D. C. (1984). Cognitive wheels: The frame problem of AI. In: C. Hookway (Ed.), Minds, machines and evolution. Cambridge: Cambridge University Press.
68
Wayne D. Christensen and Cliff A. Hooker
Dreyfus, H. (1996). The current relevance of Merleau-Ponty's phenomenology of embodiment. In: H. Haber, & G. Weiss (Eds), Perspectives on embodiment. London: Routledge. Elman, J. L. (1995). Language as a dynamical system. In: R. E Port, & T. Van Gelder (Eds), Mind as motion (pp. 195-225). MIT Press. Grush, R. (1997). The architecture of representation. Philosophical Psychology, 10(1), 5-23. Hasemann, J. M. (1995). Robot control architectures: Application requirements, approaches, and technologies. Proceedings of the SPIE Intelligent Robots and Computer Vision XIV." Algorithms, Techniques, Active Vision, Material Handling (pp. 22-36). Philadelphia, PA. Hendriks-Janson, H. (1996). Catching ourselves in the act: Situated activity, interactive emergence, evolution and human thought. Cambridge, MA: MIT Press. Hexmoor, H., Horswill, I., & Kortenkamp, D. (1997). Special issue: Software architectures for hardware agents. Journal of Experimental & Theoretical Artificial Intelligence, 9(2/3). Hutchins, E. (1995). Cognition in the wild. Cambridge, MA: MIT Press. Janlert, L. E. (1987). Modeling change: The frame problem. In: Z. W. Pylyshyn (Ed.), The robots dilemma: The frame problem in artificial intelligence. Norwood, NJ: Ablex. Kelso, J. A. S. (1995). Dynamic patterns. MIT Press. Klowden, M. J. (1995). Blood, sex, and the mosquito: Control mechanisms of mosquito blood-feeding behavior. BioScience, 45, 326-331. Legg, C. (n.d.). The meaning ofmeaning-fallibilism. Unpublished. Malcolm, C. (1997). A hybrid behavioural/knowledge-based approach to robotic assembly. DAI Research Paper, 875. University of Edinburgh, Edinburgh, Scotland. Mataric, M. J. (1992). Integration of representation into goal-driven behavior-based robots. IEEE Transactions on Robotics and Automation, 8(3), 301-312. Mataric, M. J., & Brooks, R. A. (1990). Learning a distributed map representation based on navigation behaviors. Proceedings of 1990 USA Japan Symposium on Flexible Automation (pp. 499-506). Kyoto, Japan. Merleau-Ponty, M. (1962). Phenomenology of perception. C. Smith (Transl). London: Routledge & Kegan-Paul. Montague, E R., Dayan, E, Person, C., & Sejnowski, T. J. (1995). Bee foraging in uncertain environments using predictive hebbian learning. Nature, 377, 725-728. Pfeifer, R., & Scheier, C. (1999). Understanding intelligence. Cambridge, MA: MIT Press. Raft, R. (1996). The shape of life: Genes, development, and the evolution of animal form. Chicago: University of Chicago Press. Raphael, B. (1976). The thinking computer: Mind inside matter. San Francisco: W. H. Freeman. Real, L. A. (1991). Animal choice behavior and the evolution of cognitive architecture. Science, 253, 980-986. Quartz, S. R., & Sejnowski, T. J. (1997). The neural basis of cognitive development: A constructivist manifesto. Behavioural and Brain Sciences, 20(4), 537-596. Smith, L. V., & Thelen, E. (Eds) (1993). A dynamics systems approach to the development of cognition and action. Cambridge, MA: Bradford Books.
Representation and the Meaning of Life
69
Smithers, T. (1995). Are autonomous agents information processing systems? In: L. Steels & R. A. Brooks (Eds), The artificial life route to 'artificial intelligence': Building situated embodied agents. Lawrence Erlbaum. Thelen, E. (1995). Time-scale dynamics and the development of an embodied cognition. In: R. Port, & T. Van Gelder (1995). Van Gelder, T. (1995). What might cognition be, if not computation? The Journal of Philosophy, 92(7). Van Gelder, T. (1998). The dynamical hypothesis in cognitive science. Behavioral and Brain Sciences, 21(5), 615-627. Van Gelder, T., & Port, R. (1995). Its about time: An overview of the dynamical approach to cognition. In: Van Gelder, & Port (Eds), Mind as motion: Explorations in the dynamics of cognition. Bradford/MIT Press.
This Page Intentionally Left Blank
Chapter 4
The Dynamic Emergence of Representation Mark H. Bickhard
Representation did not exist moments after the Big Bang; it does now. Representation has emerged. Accounting for that emergence is among the central problems of naturalism today. The emergence of representation is a difficult problem for a number of reasons, but one of them is that representation involves issues both of fact and of normativity, and accounting for the naturalistic emergence of normativity is particularly difficult m some would say impossible. The possibility of ontological emergence of any kind is itself in question, according to some arguments, so I begin by examining one of the most powerful of those arguments and deriving a metaphysical lesson from it: accounting for non-epiphenomenal ontological emergence requires a process metaphysics. Next, a basic flaw in the argument against the possibility of deriving normativity within a naturalistic framework is pointed out, thus freeing the discussion to develop a naturalistic model of the emergence of representation.
Emergence
Jaegwon Kim has presented an argument against the possibility of genuine emergence that, if sound, would render nugatory all attempts to account for any kind of emergence (Kim 1989, 1990, 1991, 1992, 1993a, b). I contend that the argument is not correct, but that a deep lesson can be derived from it.
Representation in Mind: New Approaches to Mental Representation Edited by H. Clapin, P. Staines and P. Slezak Copyright 9 2004 by Elsevier Ltd. All rights of reproduction in any form reserved. ISBN: 0-08-044394-X
72
Mark H. Bickhard
Kim argues that genuine emergence must involve the emergence of causal power. The emergence of mentality, for example, if it were causally inert, would at best be a form of epiphenomenalism. But this creates a dilemma if we are to hold to basic naturalistic commitments: (1) Either all causal power is resident in whatever the lowest level of particles may be, in which case all higher level processes (including mental processes) are causally epiphenomenal; or (2) some new causal power does come into existence, in which case the world is not physically closed, and naturalism, at least in most conceptions of naturalism, is false. So, the dilemma is between the failure of emergence or the failure of naturalism. The first horn of the dilemma is simply that no high level entity or process, including mental processes, can be anything more than the working out of the causal interactions of the basic particles involved. Any new causal regularities that might be observable are merely regularities of that particular kind or organization of particle interaction; all causal power remains at the level of the particles per se, and there is no emergent causal power. If there is presumed to be some new causal power, then that causal power will be ad hoc relative to particle level causality, and will entail that the physical, the particle, universe is not causally closed. This introduces a kind of causal dualism (or worse) between particle causality and whatever new causality comes into existence and constitutes, among other things, a failure of naturalism. Thus the second horn of the dilemma. The argument, however, though valid, is unsound. The crucial false premise is a metaphysical premise, and Kim's argument, thus, can be taken as a reductio of this premise. The false premise is the assumption of a particle metaphysics. Put simply, there are no particles. The best contemporary physics tells us that the world is composed of quantum fields, not particles (Brown & Harr6 1988; Davies 1984; Saunders & Brown 1991; Weinberg 1977, 1995, 1996). Field models are mathematically forced by special relativity together with conservation of energy. What appear to be particles are the quantization of quantum wave phenomena; such quantized phenomena are no more particles than is the "quantization" of the number of waves in a guitar string. That is, particles are the "appearance" of quantized quantum field phenomena, and have no substantial existence beyond that. Furthermore, we know of phenomena that are simply inconsistent with a particle based physics, but are predicted by quantum field theory, such as the Casimir effect (Aitchison 1985; Sciama 1991; Weinberg 1995). Quantum field theory, therefore, provides its own empirical predictions and support. Still further, there are in-principle arguments against a strict particle physics, and, therefore, against a strict particle metaphysics: e.g. point particles would have zero probability of ever encountering each other, and, therefore, nothing would ever happen in such a universe (Bickhard 2000).
The Dynamic Emergence of Representation
73
A standard semi-sophisticated contemporary view might postulate both particles and fields, with the fields enabling the interactions among the particles. This would be incorrect, but what I argue is the crucial point is already accepted even in this move: fields are processes, and the conclusions of Kim's argument require that we move to a process metaphysics. The reason for this rests on a diagnosis of Kim's basic argument. In framing the anti-emergenfist argument within a particle metaphysics, Kim has presupposed that causal power is resident in particles. Particles do not themselves have any organization. They are points. Particles, however, can participate in organization in space and time and with respect to each other. It is precisely within such organization that particles engage in their presumed causal interactions. Such organization is the stage setting within which particle interactions proceed. The critical move here is that organization is relegated to a strictly stage setting role. Causality is possessed by entities that do not have organization, and organization itself is not a legitimate locus of causal power: it is merely stage setting. It is this delegitimation of organization as a potential locus of causal power, and, therefore, as a potential locus of emergent causal power, that drives Kim's argument. Processes, in contrast, are inherently organized. A point process is a nonsensical notion. If everything is process and if everything is quantum field theory, then everything is process then everything that has causal power is itself already organized. Organization cannot be delegitimated as a locus of causal power without emptying the universe of all causality. Furthermore, causal power varies with organization; some process organizations have different causal powers than other organizations. So, a new organization of process might well have a new causal power, by virtue of that organization. Such a new causal power will be emergent in that organization. This is a naturalistic and non-epiphenomenal emergence. In sum, then, Kim's argument shows that emergence makes sense only within a process metaphysics, and the clear fact of emergence therefore forces a process metaphysics. It is doubly fortunate, then, that contemporary physics also forces a process metaphysics.
Normative Emergence It is a standard truism that "ought" cannot be derived from "is" m norms cannot be derived from facts. But it is a false truism. Because I propose precisely to show how certain norms relevant to representation can be emergent from, and therefore derived from, facts. I will clear the way for that by showing how the argument against deriving norms from facts fails.
74
M a r k H. Bickhard
The core of the argument is the claim that normative terms cannot be validly introduced in a derivation from premises that do not already contain normative terms. Since facts, presumably, do not contain normative terms, then no normative conclusions can be derived. There is actually more than one point at which this argument can be attacked, but I will focus on one in particular. The structure of the argument assumes that the only valid form of introduction of terms is via abbreviatory definition. If every valid new term is simply an abbreviation for combinations involving prior terms, and if the base set of terms contain none that are normative, then any valid conclusion could be rendered solely in the original non-normative terms by back-substituting through the definitions. By assumption, then, the conclusion would not involve norms. But the assumption that the only valid form of definition is abbreviatory definition is false. There is also implicit definition. The model theoretic notion of implicit definition is illustrated by the sense in which a set of formal sentences implicitly defines the class of models that would satisfy those sentences. Hilbert argued, for example, that the axioms of geometry implicitly defined the terms of geometry, such as "point" and "line." Implicitly defined terms cannot be substituted away in favor of some base set of terms. This point does not demonstrate how to derive norms from facts, but it does show that a commonly presumed impossibility of doing so does not hold. There is, however, a rejoinder to this point that I would like to address. Beth's theorem (in model theory) demonstrates that implicit and explicit definition are of equal power (Chang & Keisler 1990), and this is often taken as a reason for ignoring implicit definition (Doyle 1985). This stance, however, overlooks important points. First, the equivalence between implicit and explicit definition is extensional only. Second, and more deeply, Beth's theorem holds in the case of first-order predicate logic with infinite models. This is not the only combination of logic and model theoretic framework. In every combination examined of logics, e.g. first order predicate logic, infinitary logics, fixed-point logics, and models, e.g. infinite models, finite models, implicit definition is either of equal power to explicit definition, or it is more powerful than explicit definition (Dawar et al. 1995; Hella et al. 1994; Kolaitis 1990, manuscript 1996). Implicit definition, therefore, cannot be ignored, and it suffices to render "no 'ought' from 'is' "invalid.
Representation The model of representation that will be outlined in the following pages is called interactivism. Interactivism models representation as emergent in a particular kind
The Dynamic Emergence of Representation
75
of biological function, so the first focus is to model the emergence of biological function. The normativity of representation derives from that of biological normative function (Bickhard 1993). In this structure, the interactive model resembles etiological models, such as that of Ruth Millikan (Millikan 1984, 1993), which also first model biological function and then representation as a special case. The specifics, however, differ fundamentally: (1) the interactive model of function is not etiological; and (2) the relationship between representation and function is not the same. Before proceeding to the interactive model, I will show that etiological models are not acceptable as models of the naturalistic emergence of function and representation. In particular, they are causally epiphenomenal, and, in that sense, fail naturalism (Bickhard 1993). Etiological Models of Function and Representation. In an etiological model of function, X having a (proper) function is constituted in X being the product of the requisite history of evolutionary selections. A kidney, for example, has the function of filtering blood because the evolutionary ancestors of this kidney were selected for filtering blood. Conversely, it is because kidneys have the function of filtering blood, and that the ancestors of this kidney succeeded in fulfilling that function, that this kidney exists. There are many details to attend to and sophisticated and ingenious elaborations of and derivations from these notions in the literature, but the crucial point that I wish to focus on is that function is constituted in having the fight history. To have a function is to have the fight history. This point has some counter-intuitive consequences; the consequence that I wish to focus on is one which Millikan has partially addressed with the following science fiction example. If a lion were to suddenly pop into existence in the room that was, by hypothesis, molecule for molecule identical to the lion in the zoo, nevertheless, the heart of the lion in the zoo would have a function, while none of the organs in the magical lion would have any functions at all. No part or parts of the lion that just appeared have any history at all, and certainly, therefore, no evolutionary history. Correspondingly, they do not have the fight evolutionary history to constitute them as having any functions. There would be no biological normativity involved at all in the suddenly-appearing lion, despite its being molecule for molecule identical to the zoo lion. This is strongly counter-intuitive, but quantum mechanics has taught us that counter-intuitiveness is not necessarily a good ground for rejecting a theory. It might seem, in fact, that such a consequence is a small price to pay for the power of a naturalized model of biological function, and rendering function in terms of biological evolution certainly has the prima facie look of a successful naturalization. This seems to be Millikan's stance. The example, however, illustrates an even deeper problem. Etiological function is ontologically constituted in having the fight history. Having the fight history is
76 Mark H. Bickhard not a property of the current state of a system. The two lions, for example, have identical current physical states by hypothesis, yet they do not have the same functional state because they do not have the same histories. But only current state can be causally efficacious. The two lions will have identical causal properties, but one will have organs with functions and the other will not. That is, etiological function is causally epiphenomenal; it makes no causal difference in the world. Etiological function is not a naturalization of a causally relevant function. I turn, therefore, to a model of the emergence of function that I claim is naturalistic, including accounting for emergent causal power. Autonomy, Self-Maintenance, and Function. In a substance or particle metaphysics, the stability through time of the basic substances or particles is taken for granted. In a process framework, any stability of an organization of process requires explanation. There are two general forms of such explanation: (1) Some organizations of process are in energy wells, in the sense that a change in the organization would require the introduction of energy above what is currently impinging on the process. Atoms, molecules, and much of the standard furniture of the world is temporally persisting because of such energy well stabilities. (2) The second form of such stability is that of processes that are far from thermodynamic equilibrium. Such a process will move toward equilibrium, and thus cease to exist, unless some active counterinfluence is operative. That is, they are unstable if isolated from such counter-equilibrating influences. In some cases, those influences are completely external to the system itself. A chemical bath can be maintained in a far from equilibrium condition, for example, with the pumping into the chamber of appropriate chemicals. Any stability is dependent on the continuing operation of the pumps and availability of the chemicals. Such systems can exhibit interesting properties, such as self-organization. They also illustrate a fundamental fact about stable far from equilibrium systems: they are open to their environments, not just as a matter of fact, but as a matter of ontological necessity. If cut off from their environments, they cannot remain far from equilibrium, and they cease. Self-Maintenant Systems. A more interesting case for current purposes, however, is the class of far from equilibrium systems that make their own contributions to their own stability. A canonical example is a candle flame. A candle flame maintains above combustion threshold temperatures; it vaporizes wax into flammable gases; and in standard atmospheric and gravitational conditions it induces convection, which pulls in fresh oxygen and gets rid of waste products. A candle flame is, in several ways, self-maintenant. Recursive Self-Maintenance. A self maintenant system can maintain itself over some range of conditions - - if a candle is put into a vacuum or doused with water, it ceases. Some systems, however, can, in addition, contribute to their own
The Dynamic Emergence of Representation 77 stability over a range of changes in conditions. They can change what they do to maintain stability in accordance with changes in environmental conditions. A bacterium, for example, might swim and continue swimming if it is going up a sugar gradient, but tumble if it finds itself swimming down a sugar gradient (D. T. Campbell 1990). It maintains its condition of being self-maintenant in the face of changing environmental conditions: it is recursively self-maintenant (B ickhard 1993). Recursive self-maintenance requires some way of differentiating environmental conditions, two or more differing ways of contributing to conditions for stability, and appropriate switching relationships between the differentiations and the alternative potential kinds of interaction. Such properties require that the system have infra-structure: process organizations that are on a sufficiently slower time scale that they can be taken as fixed, as structural, during the differentiations, switching, and interactions. For example, processes that are on a sufficiently slower time scale than those of the swimming and tumbling that they can control the process of switching between swimming and tumbling. The (self) maintenance of such infrastructure is what constitutes metabolism (Moreno & Ruiz-Mirazo 1999). In the case of a recursively self maintenant system, the infrastructural processes have both an informational a s p e c t - differentiation and switching and an energetic aspect m guiding energy into swimming. This suggests a more minimal form of infrastructure. Some systems may have infrastructure that accomplishes work m that guides energy - - but that has no informational aspects. An example might be a sulfur based cell that does only one thing, "burn" sulfur, but does it with complex metabolic processes. The models of self-maintenance, work infrastructure, and recursive selfmaintenance collectively constitute a graded model of autonomy (Bickhard 2000a). Autonomy is the ability of a system to maintain itself in accommodation to its environment, and self-maintenance, infrastructure, and recursive self-maintenance constitute increasing grades of such autonomy. There is now in place a sufficient model to address both function and representation. Function first: Serving a function is modeled as making a contribution to far from equilibrium stability. Serving a function, therefore, is relative to the system which is being contributed to. A heart, for example, may serve a function for a parasite, but be dysfunctional for the host. The normativity of function will be similarly contextualized. Note that serving a function contributes to the stability of a far from equilibrium process, which has distinct causal consequences in the world: this is not a model of an epiphenomenal function. It should be noted that this explication turns upside down the explicatory organization in etiological models. Etiological models take as primary the notion of having a (proper) function, and derive a notion of serving a function from it:
78
Mark H. Bickhard
X serving a function is (roughly) X accomplishing a function that X has. 1 Here, the notion of serving a function is primary. Basing a model of having a function on this model of serving a function requires one additional preliminary notion. Dynamic Presupposition. A recursively self-maintenant system may just switch from one interaction with its environment to another as differentiated conditions change, such as is the case for the swimming and tumbling of the bacterium, or it may set up indications of multiple interactions that would be appropriate in current circumstances, and engage in some more complicated process of (inter)action selection. That is, action selection can occur via simple triggering, or via more complex selection processes among indicated interaction potentialities. There is much to be addressed about such systems of action selection, but the crucial point for now is that any triggering of an interaction, or any indication of the current appropriateness of an interaction, presupposes that that interaction is in fact appropriate for the current conditions. Continuing to swim down a sugar gradient is, in general, not appropriate. Appropriateness here is a normative notion, and the normativity is a functional normativity. That is, it is derived from the norm of contributing to the maintenance of the conditions for the far from equilibrium continued existence of the system. Interaction (types) will tend to be appropriate in some conditions, and not in others. An indication of the appropriateness of an interaction, therefore, dynamically presupposes that those conditions obtain. The dynamic presuppositions of an interaction or interaction indication are those conditions that would make that interaction appropriate, that render it likely to make a functional contribution. More generally, a process dynamically presupposes whatever those conditions are, internal to the system or external to the system, that support its being functional for the system. Having a function. X has a function insofar as some function being served dynamically presupposes one or more consequences of X. X has a proper function insofar as it is dynamically presupposed by embryological or metabolic functional processes - that is, it has a proper function if it is dynamically presupposed in the biological "design" of the system. A kidney has a (proper) function of filtering blood because the organization of the other organ systems in the body dynamically presuppose that something at that relative location will filter blood. Representational content. The dynamic presuppositions of a blood circulatory system will in general be internal: hearts and kidneys, for example. The dynamic presuppositions of an interaction indication will be about the environment. If those dynamic presuppositions do not hold, then the interaction will fail. That is, if those dynamic presuppositions are false, the interaction will fail. Dynamic
1For a more detailed treatmentof these issues, see Bickhard (2000a), Christensen & Bickhard (2002).
The Dynamic Emergence of Representation
79
presuppositions, then, can be true or false, and they can be true or false about the environment. Dynamic presuppositions constitute representational content about the environment. Interaction indications, in this model, are the primitive form of representation. They predicate of the environment that the environment possesses the dynamically presupposed conditions. They predicate that content of the environment. Such an interactive representation may be false: the dynamically presupposed conditions may not be true. Furthermore, they may be (fallibly) discovered to be false: if the system engages in the indicated interaction, and it does not proceed as indicated, then the dynamic presuppositions, the content, was false. In this model, not only the possibility of error, but also of system detectable error, are trivially accounted for. This is indeed a primitive form of representation. More needs to be addressed to indicate its potential to be ground for all representation. It is also a model of representation that has several unfamiliar properties m properties not common in standard models. These too will be outlined. More Familiar Representations: Objects. I will address first how the interactive model could account for the representation of physical objects. If an organism differentiates a relevant condition in its environment, it will invoke indications of appropriate further interactive potentialities. Even when that differentiation process is inactive, however, the control infrastructure that would engage in it, and its relationships to interaction indications, are still present in the system. Such an aspect of the control structure constitutes a conditionalized indication of interaction potentialities: if XYZ differentiation is made, then QRS interactions will be indicated. Conditionalization, in turn, creates the possibility of iterating such indications: if XYZ differentiation occurs, then QRS is possible, and if QRS occurs, then ABC will be possible. 2 So, interaction indications can both be multiple they can branch and they can iterate. As such, they can form webs of interconnected conditionalized indications of interaction potentiality m perhaps vast and complex webs. Some subwebs of such a larger web may come to have special properties. In particular, they may be internally reachable, in the sense that any indicated interaction anywhere in the subweb is reachable as a direct interaction potentiality, perhaps via various intermediary conditional interactions, and that intemal reachability property may remain invariant under some relevant class of other kinds of interactions. For example, a child's toy block will afford multiple potentialities of visual scans and manipulations. Any one of these potentialities is available from
2 For a more detailed treatment of these points, see Bickhard (1993), Bickhard & Terveen (1995).
80
Mark H. Bickhard
any other - - e.g. you can always turn the block back so that an earlier visual scan is again possible so the subweb of interactive potentialities for this block is internally reachable. And that internal reachability itself remains invariant under a large class of other interactions, such as putting the toy away in the toy box, the child leaving the room, and so on. It is not invariant under all possible interactions, however, such as crushing or burning the block. This outlines the general manner in which the interactive model can scale from simple interaction possibility representations to representations of physical objects. It is a generally Piagetian, or pragmatic, 3 model of object representation, and I would suggest a generally Piagetian approach to other more complex kinds of representation, such as abstract representations - - such as of numbers. 4 What about Input Processing ? Models of representation are standardly what the pragmatists called spectator models. They are models of some homunculus staring back down the input stream, processing inputs, rather than future oriented models of interactive anticipation. But input processing clearly does occur in sensory systems, for example. If such input processing is not to be taken as somehow constituting or generating representation, what account is to be given of it? The interactive model distinguishes between two aspects of epistemic relationship to the world: contact and content. Contact with the environment is provided by the differentiations of that environment. Such differentiations are the basis for setting up indications of further interactive potentialities; they are how the system can locate itself in its web of conditional interactive indications. Without contact, no interactive content, no indications of potentiality, would have any likelihood of being appropriate for any particular environment. Such indications, in turn, constitute representational content. It is such anticipatory indications that involve dynamic presuppositions, presuppositions that can be false. It is in these presuppositions that representation is emergent. Differentiation in general is generated by the internal outcomes of previous interactions. If an interaction control system is engaged in interaction with an environment, the internal course of that interaction will be partially determined by the control system, but importantly determined by the environment. Differing
3 The interactive model is a pragmatic model in the sense of being action based rather than a spectator model (see below), but it is closer to Peirce's model of meaning as anticipatory habit than to his model of representation per se (Rosenthal 1983). Anticipations can be false, and can be (fallibly) detected to be false. 41 characterize these as "generally" Piagetian because, although one of Piaget's many massive contributions was to construct such action based representations, I don't think the details of his model are all correct (e.g. Bickhard & Campbell 1989; Campbell & Bickhard 1986).
The Dynamic Emergence of Representation
81
environments will yield differing internal courses of the interaction, and differing internal outcomes of the interaction. Any particular possible outcome of an interaction serves to differentiate those environments what would yield that outcome from those that would yield a different outcome: the outcomes differentiate types of environments. There is no other information available in such a differentiating outcome per se about what kind of environment it differentiates, but nevertheless it may be useful for setting up indications of further interactive potentialities. If so, then any such indication predicates of that environment whatever properties are dynamically presupposed by those indications. It is the future oriented indications that represent (something about) the differentiated environment, not the differentiations per se. Differentiations in general may involve full interactions, but a simple version would be a differentiation process that had no outputs, a passive differentiation. A passive differentiation is a differentiation nevertheless, and can serve as the basis for further indications of interactive potentiality. But passive differentiations are just input processing. Input processing, then, is an aspect of the interactive model just as it is for spectator models. The difference is that standard models take input processing as constituting or generating representation, while the interactive model takes it to be only a simple case of the general function of differentiation of contact. In effect, input processing models conflate contact and content; they take whatever the contact is in fact with as somehow the content of the purported representation.
Properties of Representation It is a large programmatic task to demonstrate the adequacy of the interactive model for all forms of (purported) r e p r e s e n t a t i o n - perception, memory, rational thought, language, and so on. These have been addressed elsewhere. 5 For current purposes, I will take it as demonstrated that the interactive model is a candidate for a model of the nature of representation, and proceed to examine some of the consequences of that nature. Representational Error. As pointed out earlier, the possibility of representational error is trivially accounted for in the interactive model: the dynamic presuppositions may be false. This is in contrast to correspondence models of representation, that simply do not inherently have the resources to account for
5 For example, Bickhard (1980, 1998), Bickhard & Campbell (1992), Bickhard & Richie (1983), Bickhard & Terveen (1995).
82
M a r k H. Bickhard
error, and must, at best, superimpose some additional criterion for error on the basic correspondence framework. The limitation is, simply, that if a purported representation constituting correspondence exists, then the representation exists and is correct, while if the crucial correspondence does not exist, then the representation does not exist. There are only two model possibilities m the correspondence exists or the correspondence does not exist m but there are three conditions to be m o d e l e d - the representation exists and is correct, the representation exists and is false, and the representation does not exist. 6 One attempt to introduce such an error criterion is Fodor's asymmetric dependency criterion. Consider two conditions under which a representation is invoked, one purportedly correct and the other incorrect. If the representation is constituted simply in the invocation relationship (be it causal, nomological, informational, or whatever), then the purportedly incorrect deployment of the representation is just as legitimate a participant in the representational constitution as is the "correct" object. So, if the objects are X and Y, there are no grounds for the claim that the representation is supposed to represent Xs and that its invocation for Y is in error. Instead, since both Xs and Ys activate the representation, the content should be construed as "X or Y" and the possibility of error evaporates. This so-called "disjunction problem" is just one version of the general problem of accounting for representational error. Fodor has suggested that the correct and incorrect cases can be distinguished in the following way: the incorrect invocation is dependent on the correct invocation in the sense that the incorrect deployment would never occur if the correct case didn't exist, but that dependency is not reciprocated; it is asymmetric in the sense that the correct case could occur even if the incorrect case never did. In the by now canonical example, if the COW representation is invoked by a horse on a dark night, that is in error because such "horse on dark night" invocations are asymmetrically dependent on invocations by cows (Fodor 1990, 1991). There are multiple problems with Fodor's model, but a straightforward counterexample to the asymmetric dependency criterion is the following: Consider a neurotransmitter docking on a receptor in a cell surface and evoking corresponding activities in the cell. Here we have full biological and nomological correspondences. Now consider a poison molecule that mimics the neurotransmitter and also docks on that receptor. Here is a clear case of asymmetric dependency, yet at best we have a case of functional error, not representational error. Fodor cannot account for the possibility of representational error (Bickhard 1993; Levine & Bickhard 1999).
6 See Millikan (1984) for this point.
The Dynamic Emergence of Representation
83
Millikan, because of the distinction between what something is functionally supposed to represent and what it is being used to represent currently, can model the possibility of representational error. But Millikan's model is epiphenomenal, and thus not acceptable for a different reason. Similarly, Dretske (1988) has resources for modeling representational error, but this model too is etiological (a learning etiology), thus epiphenomenal (Bickhard 1999). Cummins invokes a distinction between the target to be represented and the representation being applied to that target (Cummins 1996). In terms of this distinction, he models error as a representational content being applied to a target to which it is false. This is, I would argue, roughly the correct strategy for accounting for representational error. Cummins' distinction between target and representation is capturing the distinction in the interactive model between a differentiated condition and the interactive content attributed to it, between contact and content (Bickhard 1980, 1993). I argue, however, (see below) that Cummins' model fails to satisfy other crucial criteria. System Detectable Error. In the interactive model, if an indicated interaction is undertaken and the interaction does not proceed as indicated, then the indication is false, and is falsified for the system itself in a way that is potentially usable by that system. Representational error is system detectable. Only if error is system detectable can it be used to guide further behavior or to guide learning processes. Clearly system detectable error occurs, and, therefore, any model which cannot account for it is impeached. In general, models of representation do not even address the criterion of system detectable error. It is clear, however, that standard models cannot account for it (Bickhard 1999). No organism can take into account the evolutionary or learning history of its functional representations, or the asymmetric dependencies among potential invocations of its representations, to be able to determine what its representations are supposed to represent. Nor can they then compare that normative content with what is currently being represented to find out if the representation is being truly applied or falsely applied m to accomplish the later is the problem of representation all over again. 7 Cummins' Model. In the case of Cummins' model (Cummins 1996), the normativity of the representations is carried in the observer of the system, not in the system itself. This is illustrated by one of Cummins' examples: Consider a toy car that is to run a maze. The car's wheels are guided by a peg that slides
7Note that the claim that such systemdetectable error checking is impossible is the conclusion of the radical skeptical argument: any such check of a representation is simply re-using the representation again, and, thus, is circular.
84
Mark H. Bickhard
back and forth in a slot in a card that is inserted into the car and that is read by moving the card progressively through the location of the peg. Differing cards with differing slot patterns would, thus, guide the car through differing mazes. The slot in a card is isomorphic to the required turns in the mazes that that card correctly represents. This is an instance of Cummins' general point that representational content is constituted in structures that are isomorphic to what they are to represent. The normativity in this example, however, is carried in the assumption that the car is supposed to run through the maze, and that assumption is made by the observer, not by the car. If the goal were to hit the wall of the maze at a certain point, then the card that gets the car through the maze would no longer be correct. This example also helps to illustrate another point. Cummins assumes that the notion of structural isomorphism is relatively unproblematic, but it is in fact seriously problematic. There is no "fact of the matter" about what the relevant structure is in any material entity. Suppose that the card inserted into the car were not "read" by a peg in the slot, but, rather, by a head responding to the pattern of magnetic domains along the edge of the slot. Now the slot pattern per se is irrelevant. Structure, then, must be relative to the norm of the interactive task that the system is engaged in. That norm, in turn, must be a norm emergent in the system itself. And structure is relative to the method of "reading," which is a control system notion, not a material notion. Further, the only possible consequence that such a "structural" read out could have is to influence the course of the (interactive) processes in the system, and any structure per se is in-principle superfluous for accomplishing such process flow influence (Bickhard 1980). The crucial property is to control the interaction in such ways that it proceeds as indicated, as selected for, in the differentiated environments and that is the interactive model. Future Orientation. Correspondence models of representation are past oriented, with the input processing spectator looking backwards down the input stream. The interactive model is future oriented. Representation is, most fundamentally, of future potentialities of interaction. Future orientation is a feature of pragmatist models generally, but is rarely found in contemporary models. It is the future orientation of the interactive model that makes accounting for error and for system detectable error so immediate. Modality. Interactive representation is of future potentialities of interaction that is, representation is of possibilities. Interactive representation, then, is inherently modal. Standard models rarely address this issue, but the presumption is that representation is of actualities (whatever is actually on the other end of
The Dynamic Emergence of Representation 85 the input stream) and that modality is something to be added or dealt with later. Interestingly, young childrens' cognition is inherently modal, with actuality, possibility, and necessity being poorly differentiated, rather than being amodal with modality developing later (Bickhard 1988; Piaget 1987). Implicitness. Interactive content is the dynamic presuppositions made in indications of interactive potentiality. Those presuppositions are not explicitly represented; instead, they are implicit in the indications themselves. It can be explicit that an interaction of a particular kind, arriving at a designated outcome, indicates that one or more further interactions would be possible, but what supports those indications, what is presupposed about the environment by those indications, is not explicit. This implicitness of content is fundamentally different from standard models. Encodings cannot be encodings without explicit content. Implicitness is a source of some of the power of the interactive model m for example, I argue elsewhere that the frame problems arise largely from attempting to render implicit content in explicit form (Bickhard & Terveen 1995). The interactive model easily accounts for the possibility of representational error, as well as the possibility of an even stronger criterion: system detectable representational error. It also has the consequences that representation is future oriented, modal, and, at base, implicit. In all these respects, it differs radically from standard models.
Three Desiderata This discussion introduces three criteria that I would like to propose as requirements for an acceptable model of representation. These three are not exhaustive, nor are they strictly independent, but they do, arguably, capture necessary aspects or facets of representation, and I find them useful to illustrate problems with models in the contemporary literature. The three are: (1) Representational content must be internally related to whatever is to constitute representation. (2) Representational content must be functionally accessible to the system for which representation is to be modeled. And (3) Representational content must be normative for the system for which representation is to be modeled. For short, the relationship between representation and content must be system internal, system accessible, and system normative. Only the first of these requires
86
Mark H. Bickhard
any significant discussion, and that is because the distinction between internal and extemal relations is no longer widely understood or appreciated. 8 Representational content must be intemally related to representation in the logical sense of intemal relation. A logically external relation is one that can change without changing the relata; an intemal relation is one that the relata must be in if they are to be what they are. A glass of water on a table is in the external relation of being "on" the table; it could be taken off the table with no change in what it is. An arc of a circle is intemally related to the center of that circle: it could not be the arc that it is without having that relation to that center point. If content is externally related to representation, then, by assumption, there is nothing inherent in the representation that carries or determines that content. The content is arbitrary, whether there even is any content at all is arbitrary, to the purported representation being what it is. The representation is, at best, an arbitrary encoding. This works fine for such representations as Morse code, blueprints, and so on, but the extemality of the relation requires that an interpreter provide the connection between the representation and its content. That yields the familiar infinite regress of interpreters if it is mental content that we are trying to model. Intemally related content is the only way to avoid interpretive homunculi. The remaining two desiderata are relatively straightforward. Content must be functionally accessible to a system in order to be content for that system at all. Content must be normative for a system in order to be content for that system at all - - content is a normative phenomenon. The dynamic presuppositions of the interactive model are intemally related to the interaction indications to which they belong. Those interactions could not be what they are without having those appropriateness conditions. An input processing model, in contrast, generates an intemal state that is externally related to the history of the input flow. Assuming that the crucial correspondence relationship is causal, nomological, informational, or evolutionary does not provide an internal relationship between the representation and its content, and, therefore, does not avoid the homunculus regress (e.g. Bickhard & Richie 1983). Clearly, the interactive model of content is system a c c e s s i b l e - the content is internally related to the organization of system o r g a n i z a t i o n - and it is system
8 Russell rejected internal relations because of their role in the idealisms of Green and Bradley (Hylton 1990), and they have largely disappeared from the scene. With respect to representation, it was objected that for a representation to be internally related to what it represented was for mental representation to itself participate in determining the world that it represented: any change in that internal relation would, by assumption, involve a change in the world that was represented. Note, however, that the criterion proposed here is not that of an internal relation between a representation and that which it represents, but, rather, between a representation and its content.
The Dynamic Emergence of Representation 87 normative m it is normative relative to the processes of recursive self-maintenance of the far from thermodynamic equilibrium system. Structural isomorphism is intemally related to structure, but only if the structure is itself well defined. Wittgenstein's structural model in the Tractatus (Wittgenstein 1961) involved well defined structure, because it was logically defined structure, but it was neither system accessible nor system normative. Cummins' structural isomorphism has the advantage of being system accessible, because it is ultimately a physical or biological "structure," but it is not system normative, and that makes the identity of the relevant physical or biological structure itself indeterminate, thus not intemal, as well as making the distinction between a correct application of the representation to the target and an incorrect application impossible to model for the system itself. The point was made earlier that content is not accessible for etiological models nor for Fodor's model: relevant histories and relationships among counterfactual dependencies are not available to the typical organism. There is a claim of normativity in Millikan's model, though whether there is normativity for the organism is questionable. Certainly there is no causally efficacious normativity for the organism. The appearance of normativity in Fodor's model derives from the sense in which asymmetric dependency seems to capture something of the parasiticness of error on correctness, but none of this is determinable in the system itself, and, although one side of an asymmetric dependency is picked out just in virtue of the asymmetry, it is unclear why there is any normative force in one side of an asymmetric dependency vs. the other side, if that is all that is available for making the distinction. 9
Conclusion Interactive representation is naturalistically emergent as the solution to the problem of action selection. It is not epiphenomenal, and it emerges naturally in the evolution of biological agents. It has resources with which to model more complex forms of representation. Interactive representation is intemally related to its content, and, therefore, does not require an interpreter: it is free of homunculus problems. Interactive representation is inherently related to the functioning of the system it is certainly functionally accessible. Interactive representation 9 Needless to say, models that attempt to account for representation in terms of an external observer's gloss on, or explanation of, a system(e.g. Clark 1997;Dretske 1988)do not involve systemnormativity, and generally do not have internally related content. In being explicitly observer dependent, they do not even make a claim of being naturalistic.
88
Mark H. Bickhard
is inherently normative: it has the function of serving action selection, and that function is normative in its contributions to maintaining the organism in its far from equilibrium state. Interactive representation has truth value; it trivially accounts for the possibility of representational error; and it accounts for the possibility of system detectable error, and is thus compatible with the facts of error guided behavior and error guided learning. It also thereby avoids the radical skeptical argument. Interactive representation is a candidate for modeling the fundamental nature of representation.
References Aitchison, I. J. R. (1985). Nothing's plenty: The vacuum in modem quantum field theory. Contemporary Physics, 26(4), 333-391. Bickhard, M. H. (1980). Cognition, convention, and communication. New York: Praeger. Bickhard, M. H. (1988). The necessity of possibility and necessity. Review of Piaget's possibility and necessity. Harvard Educational Review, 58(4), 502-507. Bickhard, M. H. (1993). Representational content in humans and machines. Journal of Experimental and Theoretical Artificial Intelligence, 5, 285-333. Bickhard, M. H. (1998). Levels of representationality. Journal of Experimental and Theoretical Artificial Intelligence, 10(2), 179-215. Bickhard, M. H. (1999). Interaction and representation. Theory & Psychology, 9(4), 435--458. Bickhard, M. H. (2000). Emergence. In: P. B. Andersen, C. Emmeche, N. O. Finnemann, & P. V. Christiansen (Eds), Downward causation (pp. 322-348). Aarhus, Denmark: University of Aarhus Press. Bickhard, M. H. (2000a). Autonomy, function, and representation. Communication and Cognition ~ Artificial Intelligence. Special issue on: The contribution of artificial life and the sciences of complexity to the understanding of autonomous systems. Guest Editors: A. Exteberria, A. Moreno, & J. Umerez, 17(3-4), 111-131. Bickhard, M. H., & Campbell, R. L. (1989). Interactivism and genetic epistemology. Archives de Psychologie, 57(221), 99-121. Bickhard, M. H., & Campbell, R. L. (1992). Some foundational questions concerning language studies: With a focus on categorical grammars and model theoretic possible worlds semantics. Journal of Pragmatics, 17(5/6), 401--433. Bickhard, M. H., & Richie, D. M. (1983). On the nature of representation: A case study of James J. Gibson's theory of perception. New York: Praeger. Bickhard, M. H., & Terveen, L. (1995). Foundational issues in artificial intelligence and cognitive science ~ impasse and solution. Amsterdam: Elsevier. Brown, H. R., & Harrr, R. (1988). Philosophical foundations of quantum field theory. Oxford: Oxford University Press.
The Dynamic Emergence of Representation
89
Campbell, D. T. (1990). Levels of organization, downward causation, and the selectiontheory approach to evolutionary epistemology. In: G. Greenberg, & E. Tobach (Eds), Theories of the evolution of knowing (pp. 1-17). Hillsdale, NJ: Erlbaum. Campbell, R. L., & Bickhard, M. H. (1986). Knowing levels and developmental stages. Basel: Karger. Chang, C. C., & Keisler, H. J. (1990). Model theory. North Holland. Christensen, W. D., & Bickhard, M. H. (2002). The process dynamics of normative function. Monist, 85(1), 3-28. Clark, A. (1997). Being there. MIT Press/Bradford. Cummins, R. (1996). Representations, targets, and attitudes. MIT Press. Davies, E C. W. (1984). Particles do not exist. In: S. M. Christensen (Ed.), Quantum theory of gravity (pp. 66-77). Adam Hilger. Dawar, A., Hella, L., & Kolaitis, Ph. G. (1995). Implicit definability and infinitary logic in finite model theory. Proceedings of the 22nd International Colloquium on Automata, Languages, and Programming, ICALP 95, Szeged, Hungary, July 10-11 1995 (pp. 621-635). New York: Springer-Verlag. Dretske, E I. (1988). Explaining behavior. Cambridge, MA: MIT Press. Fodor, J. A. (1990). A theory of content. Cambridge, MA: MIT Press. Fodor, J. A. (1991). Replies. In: B. Loewer, & G. Rey (Eds), Meaning in mind: Fodor and his critics (pp. 255-319). Oxford: Blackwell. Hella, L., Kolaitis, E G., & Luosto, K. (1994). How to define a linear order on finite models. Proceedings: Symposium on logic in computer science, Paris, France, July 4-7. Los Alamitos, CA: IEEE Computer Society Press. Hylton, E (1990). Russell, idealism, and the emergence of analytic philosophy. Oxford. Kim, J. (1989). The myth of nonreductive materialism. Proceedings and Addresses of the American Philosophical Association, 63, 31-47. Kim, J. (1990). Supervenience as a philosophical concept. Metaphilosophy, 21(1-2), 1-27. Kim, J. (1991). Epiphenomenal and supervenient causation. In: D. M. Rosenthal (Ed.), The nature of mind (pp. 257-265). Oxford University Press. Kim, J. (1992). Multiple realization and the metaphysics of reduction. Philosophy and Phenomenological Research, 52, 1-26. Kim, J. (1993a). Supervenience and mind. Cambridge University Press. Kim, J. (1993b). The non-reductivist's troubles with mental causation. In: J. Heil, & A. Mele (Eds), Mental causation (pp. 189-210). Oxford University Press. Kolaitis, Ph. G. (1990). Implicit definability on finite structures and unambiguous computations. In: Proc. 5th IEEE LICS (pp. 168-180). Kolaitis, Ph. G. (manuscript 1996). Infinitary logic in finite model theory. Levine, A., & Bickhard, M. H. (1999). Concepts: Where Fodor went wrong. Philosophical Psychology, 12(1), 5-23. Millikan, R. G. (1984). Language, thought, and other biological categories. Cambridge, MA: MIT Press. Millikan, R. G. (1993). White queen psychology and other essays for Alice. Cambridge, MA: MIT Press.
90
Mark H. Bickhard
Moreno, A., & Ruiz-Mirazo, K. (1999). Metabolism and the problem of its universalization. BioSystems, 49, 45-61. Piaget, J. (1987). Possibility and necessity (Vols 1 and 2). Minneapolis: University of Minnesota Press. Rosenthal, S. B. (1983). Meaning as habit: Some systematic implications of Peirce's pragmatism. In: E. Freeman (Ed.), The relevance of Charles Peirce (pp. 312-327). La Salle, IL: Monist. Saunders, S., & Brown, H. R. (1991). The philosophy of vacuum. Oxford University Press. Sciama, D. W. (1991). The physical significance of the vacuum state of a quantum field. In: S. Saunders, & H. R. Brown (Eds), The philosophy ofvacuum (pp. 137-158). Oxford University Press. Weinberg, S. (1977). The search for unity, notes for a history of quantum field theory. Daedalus, 106(4), 17-35. Weinberg, S. (1995). The quantum theory of fields. Vol. I. Foundations. Cambridge. Weinberg, S. (1996). The quantum theory offields. Vol. II. Modern applications. Cambridge. Wittgenstein, L. (1961). Tractatus logico-philosophicus. New York: Routledge.
Chapter 5
New Norms for Teleosemantics Timothy Schroeder Abstract Teleosemantics has a problem: it holds that to have a mind one must have a history, often a long evolutionary history. The solution to the problem is for teleosemanticists to give up on natural selection as the source of natural norms (functions)for neural structures, and to find a different source of natural norms which is not essentially history-involving. Such a source in fact exists, in cybernetic governance. This paper argues for the existence of natural norms derived from cybernetic governance, parallel to natural norms derived from natural selection but distinct from them, and shows how such norms could save teleosemantics from its over-emphasis on history. Teleosemantics is in trouble. According to teleosemanticists such as Dretske, Millikan, and Papineau, 1 to have a mind it is metaphysically necessary to have an evolutionary past (or, in moderate versions of the theory, a learning history), and this leads to a number of apparently absurd conclusions. While it is certainly true that we have minds as a causal consequence of our evolutionary and learning histories, it seems bizarre to hold that a being atom-for-atom identical to a person with a mind would, if lacking a history, lack intentionality. This, however, is just what is held by teleosemanticists. 2 Teleosemantics has a number of other consequences which are similarly difficult to believe, also following from its
1 Otherprominent teleosemantic works include Cummins (1996), Lycan (1987), and Sterelny ( 1990)though Cummins only allows a teleosemantic interpretation of his theory, without committing to it. 2 And also, famously, Donald Davidson. See Davidson (1987) for a sketch of Davidson's nonteleosemantic reasons for holding history necessary to mentality. Representation in Mind: New Approaches to Mental Representation Edited by H. Clapin, P. Staines and P. Slezak Copyright 9 2004 by Elsevier Ltd. All rights of reproduction in any form reserved. ISBN: 0-08-044394-X
92
TimothySchroeder
commitment to history. One is that, on versions of teleosemantics of a strictly evolutionary bent, there must be biological functions concerning things such as monster truck pulls and math co-processors. Another is that, if mental properties are historical properties, then it seems there can be no such thing as mental causation, for the properties involved in causal relations are those of an a-historical nature: mass, charge, coming to a sharp point, and so on. While teleosemanticists have made efforts to take the sting out of these apparent absurdities, their efforts have not, by and large, convinced the doubters. This unfortunate commitment to history seems to knock teleosemantics, an otherwise promising theory of the mind, out of contention. Delivering teleosemantics from history is the aim of this paper.
Teleosemantics Teleosemantics is a branch of representationalism, distinguished from functionalist (e.g. Fodor 1975) and asymmetric-dependency (e.g. Fodor 1990) approaches to representationalism in employing normative notions: to be a mental representation is to be a structure which (to a first approximation) is supposed to be A when and only when the world is B. 3 Naturally, one wonders what makes it the case that some neural structure is supposed to do something, that it has a function, a job to do which it may fail to d o - what makes it the case that a norm applies to a neuron? In keeping with their naturalistic inclinations, teleosemanticists hold that brains get their functions through natural selection, and this is where history enters the theory. According to most contemporary philosophers of biology, biological functions arise when natural selection favours organisms having particular traits (Godfrey-Smith 1994). The structures which, in previous generations, provided adaptive advantage are, in the present generation, supposed to do whatever they once did which provided that advantage. Teleosemanticists take this idea and apply it to the case of neural structures. The neurons in MT, for instance, are (it seems) supposed to fire rapidly in response to motion in the visual field (with particular neurons being sensitive to particular directions of motion and visual-field locations) because it was so firing that made past tokens of such neurons evolutionarily advantageous. Teleosemanticists also frequently extend the
3 This approximation has the virtue of slighting all teleosemanticists equally: although it gives the flavour of the teleosemantic view, it is not an accurate characterisation of any actual teleosemanticist. Dretske (1988) holds a mental representation to be a structure with the function of carrying information; Millikan (1984) holds indicative (but not imperative) representations to be structures which must stand in a particular mathematical mapping relation to something in the world if they are to have biologically appropriate effects; etc.
New Norms for Teleosemantics
93
notion of natural selection to shorter timescales and sub-organismal populations: natural selection does not only select organisms because of their traits, but also selects particular neural connections in particular living brains because of their traits. As Dretske (1988) might put it, for a neural structure to have the function of producing some state is for it to have been recruited for the production of that state. The teleosemantic appeal to history is thus entirely an accident of current theories about biological functions and the natural extensions of such theories. If philosophers of biology had come to the conclusion that biological functions were best explained in terms of the current propensities provided by traits, then teleosemantics would be making appeal to current propensities, not histories, and it would not be facing the distinctive problems it is. Or, if teleosemanticists were convinced by Mark Bedau's arguments to the effect that biological functions cannot be fully reduced, and must be understood as the product of sui generis yet still natural norms, then they would give up the problems of history for other problems in value theory. Unfortunately for teleosemantics, these alternative scenarios seem unlikely to materialise when it comes to biological function, the historical account appears to have carried the day decisively. The alternative scenarios nonetheless provide an important moral, and one which will guide the rest of this paper: teleosemantics is not essentially a history-involving theory of mind, but merely a norm-involving theory of mind which has seen no viable theory of natural normativity outside of that provided by the theory of biological function. Fortunately for teleosemantics, there is an alternate source of natural norms just waiting to be tapped.
Social Norms and Their Natural Analogues Natural selection is a process which creates norms. 4 Why? What is special about natural selection? Processes of entropic decay, or sedimentation, or elliptical movement, or of a thousand other sorts, do not create norms. What, then, is special about natural selection? The answer seems to be that natural selection is the mechanism of natural design, and processes of design of all sorts create norms. Processes of intelligent design, carried out by people, create norms, and natural processes sufficiently analogous to be called "natural design" carry out processes
4 Somephilosophers are uncomfortablewith the idea of norms createdby non-humannaturalprocesses. However, if notions of disease and health, deformity and well-formedness,wilting and flourishing and all the rest are indispensable, then natural, biological norms are also indispensable. I will take for granted the possibility of natural norms in what follows.
94
TimothySchroeder
which, at a certain level of abstraction, are of the same type. Hence, they too create norms. Natural design, carried out through natural selection, is not exactly like intelligent human design, of course, but it is similar in a number of ways: in both, there are criteria imposed upon the object being designed, such that the object will not be produced if it does not meet the criteria; in both, the criteria imposed may be in conflict to some extent, preventing optimal satisfaction of particular constraints; in both, a well-designed product often displays a striking degree of complexity not achievable through simple physical processes; and so on for various other features which suffice, somehow, to make natural selection a function-giving, and hence norm-creating, process. What exactly the crucial shared features might be, which make both intelligent design and natural design instances of norm-creating design processes, is not important for present purposes. All that is important is that some such shared features must exist. Now think about everyday, mundane, socially created norms. Those imposed by designers' intentions are an important group, certainly, but they are not the only familiar norms. Norms imposed through intentional governance of objects or people are equally familiar, and almost as ubiquitous. If a person takes a garbage can and uses it to prop open a door, the garbage can is supposed to hold open the door, and the door is supposed to stay open, because a person has taken steps to bring these things about via taking control over the garbage can and the door. Similarly, if children are playing at being pirates, then it may be true that they are not permitted to step "in the water" (i.e. on the carpet) because they take steps to bring it about that they do not do so, for instance by reminders and complaints. Though these examples have something in common with cases of design, they are clearly distinct: design-derived norms last as long as the designed objects last, while governancederived norms last only for the duration of the governance. An oven always has its characteristic function, but a garbage can intended to hold open a door no longer has that function once no one cares any longer about the can or the door. The existence of socially-created norms derived from design and from governance, and natural norms derived from a process analogous to intentional design, hints at the possibility of there being natural norms derived from a process analogous to intentional governance. The mere fact of the parallel between intentional design and natural design does not by itself entail a corresponding parallel between intentional governance and natural governance, of course. That said, the possibility is certainly an attractive one. The intuition that feedback-driven governance systems cybernetic systems have some sort of normative significance is an intuition which has persisted, in one form or another, since cybernetic systems came to the attention of philosophers in the 1940s. If the intuition has a sound basis, then there may be natural norms of a non-historical character, derived from natural governance rather than natural design. And if this
New Norms for Teleosemantics
95
is the case, then there is reason to hope that teleosemantics can be rescued from its dependence upon history.
Cybernetic Systems as Sources of Norms A perfect analysis of what it is to be a cybernetic system has not yet been discovered, but that should not detain the philosopher of mind looking to cybernetic systems for theoretical relief from history, for several reasons. First, it is clear that cybernetic systems are real things. Engineers model them mathematically and build them, they are found in autopilots, in climate-control systems, and in robots, and they are as safe from vanishing in puffs of theoretical reasoning as any sort of complex physical system can be. Given this, the fact that there is no satisfactory analysis of what it is to be a cybernetic system does not count much against quantifying over them. Second, there is a long philosophical tradition of appealing to controversial, imperfectly understood properties (such as that of being a property, for example), in explaining other things. No one need blush about writing post-dated intellectual cheques; if we insist on waiting for perfect analyses of all our basic concepts before going on to analyse other things, we will never get started. Third, substantial progress has actually been made in solving the technical problems in early analyses of cybernetic systems. Adams (1979) proposes the following, which will almost serve for present purposes. An analysis of goal-directedness which is based on a cybernetic account maintains that a [cybernetic regulatory] system must have: (1) an internal representation of the goal-state; (2) a feedback system by which information about the system's state variables and its output values are fed back into the system as input values; (3) a causal dependence between the information which is fed back into the system and the system's performance of successive operations which minimize the difference between the present state of the system and its goal-state (Adams 1979: 506; italics in original). Although Adams' analysis is presented as only a necessary condition for something's being a cybernetic governance system, he appears to view his three conditions as both necessary and sufficient, and this is how I will treat them. The only problem with Adam's analysis for present purposes is that Adams puts a representation at the heart of all cybernetic systems, which certainly will not do
96
TimothySchroeder
for present purposes m after all, the purpose of this paper is to make it possible to analyse representation in terms of cybernetic systems. The fault is easily remedied, however. It will suffice to replace (1) with (1"): (1") a capacity to carry the information that the goal-state has been attained. This amendment is in the spirit of Adams' piece, since in a footnote (506, n. 18) he mentions that by "representation" he has something roughly like a causal co-variation account in mind. Understanding "information" in something like the sense of Dretske (1981, 1988) as something carried by a thing's state when that state guarantees (given the physical context) that the world is in some other state, the revised analysis tells us that a cybernetic system is one that has a sensitivity to whether or not the world is in a certain state the goal-state m and that it uses this sensitivity to carry out processes minimising the difference between the world as it is and the goal-state. Though one might quibble with the finer points of Adams' analysis or the proposed modification, it will do as a rough characterisation of a cybernetic system no rougher than the notions of property, causation, embodiment, or content used in the average theory of mind, and sufficient to the present task, which is finding a non-historical source of normativity for teleosemantics. Any cybernetic system S aims at some goal-state G. If the goal-state is that some object (or objects) have some feature (or features, or stand in some relation), then it follows that S is regulating the object, driving it to having that feature. Under such circumstances of regulation, a norm is created: the object regulated is supposed to have the feature mandated by the governing system. Consider a central heating system, for instance, directed at bringing the temperature in a room up to 20 ~ The room is supposed to be at 20 ~ simply because that is how the thermostat is set and the thermostat is part of a governance system driving the room toward that temperature: if the room is actually 18 ~ then it is not as warm as it is supposed to be. Or consider a homing missile, which is homing in on a certain airplane. The missile is supposed to reach the plane. If the plane eludes the missile, it defeats the missile's efforts, thwarts it; its escape is something which is not supposed to happen. These sorts of intuitions once carried the day, at the peak of cyberneticsinfluenced philosophy. 5 Nowadays, though, they are likely to be met with more
5 Though there has never, to my knowledge, been an interpretation of the basic intuitions which took exactly the form the present suggestion takes u that cybernetic systems create norms for the systems they govern in virtue of governing them.
New Norms for Teleosemantics
97
skepticism. One may agree that central heating systems and homing missiles are supposed to do certain things without agreeing that this is explained in any cybernetic fashion, for the fact that they are supposed to do certain things may be explained by appeal to their histories m specifically, by reference to what they were designed to do. The skeptic is a little too quick, however. I agree that any designed system will be supposed to do something simply because it was designed to do it, but these norms are not the ones in question. A central heating system is not its own goal-state, and so does not cybernetically determine that it is supposed to control the temperature of the room. It is the room being at a certain temperature that is the goal-state, and that is where the relevant norm is found. The room is supposed to be at 20 ~ It is supposed to be at that temperature, not because someone designed the room to be at that temperature, but because the thermostat is driving it that way. Similarly, because of the homing-missile's mechanism, the missile and airplane are supposed to collide. That is, that collision is something that is supposed to happen. But the collision is not something which was designed. It is supposed to happen simply because the missile is "seeking" it. Two lines of counter-argument might seem plausible here. First, one might argue that the room is supposed to be at 20 ~ because some person wanted it at that temperature, and took control over the room (by setting the thermostat), one might argue that some person wanted the homing missile to collide with the plane and took control over that possibility by launching the missile, and so on. I do not think that these would-be explanations succeed in every case. Of course, it is possible that a person wants the room to be 20 ~ and through the appropriate actions makes it true that the room is supposed to be 20 ~ but this should not obscure the fact that the cybernetic system driving the room to that temperature is an independent source of normative truths regarding the room. If someone jostles the thermostat setting without intending to do so, then the room is supposed to be at the newly set temperature, even though no one might want it at that temperature. (Indeed, it is likely that there are now two different temperatures at which the room is supposed to be m the one at which it is supposed to be because of the desires of a controlling person, and the one at which it is supposed to be because of the action of the thermostatic system.) If the cold air admitted by an open window prevents the central heating system from raising the temperature to the new, jostling-induced level, then it prevents the central heating system from doing what it is "trying" to do, and so puts off the warmed state of affairs which ought to obtain. Comparing intuitions about what ought to be the case after an active thermostat has been jostled, as compared to when a broken thermostat (which cannot expend energy to carry out any task) is jostled, will further strengthen the conviction that
98
Timothy Schroeder
there is some sense in which a room ought to be at a certain temperature because of the efforts of the thermostatic system alone. Imagine a long-abandoned house with a defunct central heating system: the thermostat is broken, the wires connecting it to the furnace have corroded, the furnace is no longer connected to a supply of natural gas and in any case would no longer work properly if it were connected. In such a house, the thermostat might be set to some particular temperature, but it is clear that there is no temperature a room in the abandoned house ought to have. This absence of normativity coincides with the non-existence of a cybernetic system governing room temperature, and exists even while the system designed to govern room temperature persists. It seems that the ability of the central heating system to instantiate a cybernetic system is crucial for the system to give rise to any norms. Similar things may be said for homing devices et al. The second of the two aforementioned lines of objection tries to block the above answer to the first by arguing that designed systems which are supposed to do certain things, by having these functions, make it true that other states of affairs are supposed to obtain. The fact that a thermostatic temperature-control system is designed to heat rooms makes it true in certain cases that rooms are supposed to be heated, the fact that missiles are designed to collide with airplanes makes it true that certain collisions are supposed to happen, and so on. It seems sensible to say "the job of the thermostat is to bring the room to the temperature to which it is set, and if the room does not reach that temperature, then something must be wrong, for the room ought to be at that temperature." This line of argument does not show what it pretends to. It is not the function of central heating systems to bring rooms to particular conditions; their function is simply to supply hot air until particular conditions obtain. If a room under the control of a central heating system does not come to the temperature it should be at because there is a big hole in the wall and it is cold outside, that does not mean that the central heating system is malfunctioning. The system malfunctions if it does not correctly register room temperature, if it does not turn on and turn off the furnace as needed, if it does not blow hot air when needed, and so on. Its functions do not extend all the way to the heating of particular rooms: this is shown by the fact that its malfunctions do not extend so far. However one looks at a room governed by a cybernetic system such as a central heating system, one cannot get around the fact that the room should be some temperature, and this "should" comes from the room's status as the object of a cybernetic system's attentions. A cybernetic system is a potential source of norms, just as an evolutionary history is. A cybernetic system creates a norm when it actively controls some other system, driving it into some goal-state: the driven system, because it is driven in this way, is supposed to enter the goal-state. This norm is one which is independent of human design, desire or other involvement in
New Norms for Teleosemantics
99
the system, just as the norms created by evolution are. And just as with the norms created by evolution, the norms created by cybemetic systems are analogous to certain norms created by human beings. The controlling force of a cybemetic system is akin to the control of an agent who wants the controlled system to do something, just as the selective action of evolution is akin to purposeful design. The conclusion just reached may strike the reader as surprising, unexpected even if the reader is familiar with older work on the philosophical implications of cybemetics, and quite surprising if the reader is not. What more can be said in defence of the general conclusion that cybernetic systems give rise to norms for the systems they govem? The first thing to discuss is the predictive use of normative-sounding language. Sometimes we say such things as "if I let go of this pen, it ought to fall," or "the heat should melt any plastic mixed in with the metal." These locutions are not normative but predictive: they are made when an appeal to natural necessity is apt. One might be inclined to hold that the various normative-sounding locutions surrounding cybemetic systems, even when not explainable in terms of human purposes, are explainable as predictive uses of the locutions rather than truly normative uses. This is not bome out by actual practice, however. For instance, think again of the room which is supposed to be at 20 ~ because a central-heating system is driving the room to that temperature, even though no particular person cares about the room being that temperature (perhaps the house has been abandoned, in such haste that the gas is still tumed on, and the former owner no longer cares what temperature the room is). Now imagine that a window has been left open, admitting very cold air. It might be apparent to us that the central-heating system will not drive the room temperature up to 20 ~ That is, our prediction might be that the room temperature will hover around 12 ~ Nonetheless, we can recognise what temperature the room ought to be. "It ought to be 20 ~ in here, but the open window is making it colder than it should be," we might say. "Why isn't this room the temperature it's supposed to be? Oh, the window is open" is another natural phrase under such circumstances. It is worth noting that the predictive sense of normative-seeming terms is sometimes used even when the use is contrary to actual prediction: "I know the pen should fall to the ground, but it just refuses to do so" I might say in bewilderment. However, this use of"should" is only appropriate when there is no explanation for the violation of expectations. If I discover that a fine wire is holding up the pen, then I must withdraw my claim that the pen should fall. The case of the room with the open window is like that of the pen held up by a wire, in that it may be obvious why one should predict the room not to be at 20 ~ but it is unlike the case of the pen held up by a wire in that the normative claim that the room ought to be at 20 ~ remains. This robustness of application shows that the normative-seeming talk is truly normative, and is not merely predictive.
100
Timothy Schroeder
An interesting example of normative language comes from lay ecological talk, and this provides another point of defence for the normativity of cybernetic systems. Consider the following natural-seeming claims: "there isn't supposed to be a hole in the ozone layer, .... the carbon cycle is so overloaded by the burning of fossil fuels that it is no longer operating properly," and "Southern California should not have barren tracts of alkaline land m that's the fault of over-irrigation." What is to be made of the normative language found in them? One might suggest that it is merely predictive, but all three examples seem to give the lie to that idea, for they all complain of normative violations. One might hold that normative talk in these cases is a mistake, but this seems like an explanation of last resort. Most plausible to me is the suggestion that it is the apparent cybernetic governance of the ecosystem that gives rise to this normative talk. The various nutrient cycles and the like, driven with energy from the sun and from living systems, carry out processes which tend to perpetuate the cycles, in ways that are sensitive to whether or not the cycles are being carried out. This remains true even when human intervention prevents the cybernetic goal-states from obtaining in the environment, hence the natural attribution of unmet norms in cases of environmental destruction. It has been suggested to me 6 that the scientific community has abandoned the view that the ecosystem is fruitfully thought of as a collection of cybernetically governed systems. However, the popular science and lay community has certainly not caught up with current scientific thinking in this respect. At present, many people think of ecosystems in auto-regulatory, cybernetic terms, and many people are willing to use normative language in characterising such systems. This convergence alone suffices to support the claim that governance systems are norm-making systems, whether or not it is actually warranted. If scientists are coming to have a different perspective, I would predict that their use of normative language will drop off, and that the language of the layperson will do likewise if present scientific views trickle down to the untutored. If the ecosystem is not a self-stabilising system but merely an open-ended dynamic process, then the burning of fossil fuels cannot cause the carbon cycle to break down, but merely to go through a substantial phase transition (or whatnot). Similarly, it just isn't true that there shouldn't be a hole in the ozone layer if the natural processes of the world have only maintained ozone levels coincidentally, if it is merely an accident that there have been no ozone holes for quite some time. If carbon or ozone levels are in no sense regulated, then we cannot get them "out of whack," though we may alter them in ways dangerous to ourselves, of course. This seems right, but why? The most natural explanation seems to be that, once we drop any commitment to cybernetic regulation of various
6 By Peter Godfrey-Smith,in conversation.
New Norms for Teleosemantics
101
environmental factors, we drop commitment to normative language regarding such factors. And this can only be if we recognise, perhaps implicitly, that cybernetic governance gives rise to normativity. Biology provides a further argument for the intuition that cybernetic control gives rise to norms. I have found, in an admittedly unscientific survey, that a number of people are inclined to attribute certain biological functions even if there is no evolutionary history to back the function ascription. If it turned out that people did not evolve, but were recent cosmic accidents, then presumably none of our organs would have functions as a result of natural selection. This might incline one to say that it would be impossible for our "hearts" (not of a kind with biological hearts, but resembling them; from hereon the scare-quotes will be suppressed) to become diseased, for our livers to malfunction, for there to be anything wrong with our spines, for anyone to be healthy or to pass maturational milestones on schedule. All of these things essentially involve normativity, and current thinking has it that this normativity comes exclusively from evolution. Even so, some people are inclined to think that this normative talk would remain legitimate were there to be no biological functions. I suppose that some of this normative talk could come from controlbased, intentional n o r m a t i v i t y - I use my hands to pick things up, and so they'd be broken, defective, if they were in a state that prevented them from picking things up, and so on. However, I don't use adrenaline, or my kidneys, or serotonin, or T-cells, to do anything. These things are not under my control 7 and so have nothing they are supposed to do, because I do not use them in any relevant way. Nonetheless, the intuition persists among some people that I could have an adrenaline imbalance, or malfunctioning kidneys, or normal serotonin levels, or diseased T-cells, even though I don't use these parts of my body and even if it were true that they had no evolutionary functions. Cybernetic norms can step in to accommodate this intuition. Indeed, the idea that cybernetic governance is what makes it true that, for instance, there is a salt balance that should obtain in my body (for instance), or that cybernetic systems are what determine my thermoregulatory system to be a regulatory system, is one of the early and driving intuitions behind philosophical thinking about cybernetic systems m see, e.g. Nagel (1961). If there are cybernetic norms, it is plausible that many of the body's dynamically regulated systems and structures would turn out to have things they are supposed to do, states in which they are supposed to remain, for this dynamic regulation is precisely a matter of expending energy in such a way as to bring about salt balances, thermoregulation, digestion of stomach contents and the like, and to do so via sensitivity to the
7 Typically; I may use my T-cells to do a biology experiment on occasion, bringing them under my control for that purpose, for instance.
102
Timothy Schroeder
bringing about or non-bringing about of these things. If cybernetic norms indeed exist, then some of the normative statements we make about parts of our bodies have three independent and equally good grounds: evolutionary norms, cybernetic norms, and everyday norms of use. If the evolutionary and everyday normative grounds are missing in a certain case, but our intuition that some normative talk is appropriate persists, cybernetic norms can answer the call. To those with such intuitions, the present theory of natural norms should hold extra attraction. Interestingly, not all body parts having evolutionary functions also have cybernetic functions. Any part of the body that has a job to do which it is not driven to do by other body parts in a feedback-sensitive fashion is a part which will have evolutionary but not cybernetic functions. An easy example in another species is the antler. Antlers are dead tissue, and so do not have their characteristic shapes or construction in virtue of cybernetic systems. They do, however (surely!), have evolutionary functions. This leads us to a thought experiment: what should we say about a deer-like creature with antlers, lacking an evolutionary history? Would such a creature's antlers have the job of, say, enabling the deer-like creature to fight with others for mates? It seems that, if the creature is not inclined to use its antlers in this fashion, no norm can be attributed to them by cybernetic theory, evolution, or everyday normative thinking. Here, I think common intuition would agree. Absent use and evolutionary history, this dead matter on the creature's head is merely dead matter, and cannot be deformed, the right length, too tightly spiralled, or have any other normative property. On the other hand, the creature can still run a fever (i.e. be too hot for the purpose of fighting infection) if we may attribute norms on the basis of cybernetic systems. To me, this seems at least an acceptable description of such a creature, and possibly an attractive one. Though in no way decisive, the point is one that again favours those who believe in cybernetic norms. The preceding arguments have, I hope, convinced you that there is live normative talk, that there are normative facts, which the theory presented here captures. Just as certain biological organs incline us to normative talk, so certain other features of the natural world incline us to normative talk in virtue of the ways in which they are regulated. This makes for an interesting reflection of the intentional in the non-intentional realm, but does it have anything to say to philosophers of mind with whom I began this essay? I will close with some optimistic remarks.
Cybernetics and Representation As I said at the outset, the principal problem facing teleosemantics is that it makes having a mind a matter of having a history of a particular sort. If
New Norms f o r Teleosemantics
103
cybernetic governance creates genuine natural norms, then it offers a way for a teleosemanticist to give up reference to natural selection, and so to history, while keeping norms in her theory of mind. The norms coming from cybernetic governance come into being as soon as governance is established, and go out of existence when governance ends: they are norms which are insensitive to history. Such norms are just what the teleosemanticist needs. What would a teleosemantic theory of mind look like once it abandoned evolution and adopted cybernetic governance? That, of course, would depend upon the teleosemantic theory with which one began. The only change required by adopting cybernetic norms is the abandonment of norms from natural selection; everything else remains as it was. Thus, the introduction of cybernetic norms is a conservative revision of teleosemantics: it does not impose further changes upon the theorist, but simply solves one of the problems. Admittedly, the introduction of cybernetic norms is not ontologically conservative, for it requires the existence of feedback-driven governance systems in any organism capable of thought, something not required by earlier versions of teleosemantics. Fortunately, this ontological extravagance turns out to be empirically plausible, at the very least in the case of human beings and other mammals. That neural connections in the brain are continually being maintained or modified by feedback-driven governance systems is strongly suggested by our susceptibility to operant conditioning, to take the most obvious indication. 8 Certain computer models of the mind lead to similar conclusions. Neural network models of mental processes, while differing in a number of respects, generally agree upon the need for some sort of feedbackdriven governance of connections between neurons for the creation of appropriate connections between them. Though such feedback-driven governance often takes biologically unrealistic forms in these models, other, more realistic forms of feedback are also possible. Because cybernetics-derived norms are theory-neutral, as useful to one teleosemanticist as another, I would like to explore the plausibility of cybernetic teleosemantics at a certain level of abstraction. Why should cybernetic governance have anything to do with intentionality? Or, more carefully, what is there in the nature of intentionality that has any essential connection to cybernetic governance? There is, I think, a very good answer to this question, and it is an answer which does as much to justify the cybernetic approach to teleosemantics as anything said in the previous section about the normativity of cybernetic
8This appears to be the case even for primary and secondary sensory cortex, even in adult mammals (where it was long thought that there was no cortical change after maturation). See Kaas (1991) for a review; for clear instance with detailed experiments, see, e.g. Merzenich & Jenkins (1993), Recanzone et al. (1992).
104
Timothy Schroeder
governance. The answer is that the essence of intentionality is a kind of pointing, aiming, or being directed at the world, and when a cybernetic governance system drives the states of one object to correspond to the states of some other object (to use the most genetic teleosemantic phrasing possible), it points, aims, or directs the first object at the second. To be driven to correspondence with another thing by a cybernetic system is to be directed at that other thing. For an example of what a cybernetic theory of mind might look like, consider the teleosemantic theory recently advanced by Rob Cummins (1996). According to Cummins, for there to be a neural structure whose state r represents (in the sense of also being able to misrepresent) that P is for r to be produced by a mechanism M, the (or a) function of M being to produce states isomorphic to state-of-affairs p.9 Where might M get this function? Cummins considers two options: natural selection and systemic role, though he seems completely comfortable with neither. If he opts for natural selection, his theory takes on the problem that the mind is thus rendered essentially historical. If he opts for systemic role, on the other hand, his theory is in danger of losing teleological force. The role a thing must play in order to give a containing system a particular capacity is not a fact with normative significance the fact that my computer has the capacity to annoy its user by making a squeaking sound with its fan does not entail that there is something goal-directed about the fan, does not entail that the fan has the function of squeaking. Because Cummins is not firmly committed to any particular theory of natural function, his theory presents an ideal opportunity to see how cybernetic norms might come to the rescue. The norms Cummins needs are norms for mechanisms of token-production. A mechanism which produces tokens must have the function of producing ones isomorphic to some target in order to represent (again, in the sense of "represent" which entails that misrepresentation is a possibility). On the present theory of normativity, this would be the case if there were a cybernetic system driving some neural mechanism toward producing tokens isomorphic to a particular target or targets. If some inner system monitors the production of isomorphic states by a particular mechanism, and "rewards" it for success while "punishing" it for failures (carries out processes tending to realise the goal-state), then that mechanism will have the function of producing such isomorphic tokens. If such inner governance systems were to exist, Cummins' theory would be maintainable without either abandoning normativity or making the mind essentially historical. Whether such
9 Adapted from Cummins (1996: 118). This gloss ignores a number of complications in Cummins' theory, suchas the idea that not all representationsare capable of misrepresentation,but does not change anything substantive.
New Norms for Teleosemantics
105
inner governance systems exist or not is, of course, an empirical matter, but this should not be alarming: if the essences of mental representations are fit objects for investigation at all, they should be open at least in principle to new empirical evidence. One might complain that cybernetic governance raises many of the problems raised by appeals to natural selection. Most importantly, does a version of the "disjunction problem" not arise in attempting to say what cybernetically governed neurons are supposed to do? The answer to this is yes: however problematic it is to determine just what norms natural selection imposes, it is just as problematic (though not more so) to determine the precise norms imposed by cybernetic governance. But most theories of mind are faced with some version of the disjunction problem, and one can only solve one problem at a time. If cybernetic norms can solve the biggest problem facing teleosemantics, perhaps its other problems can also be addressed. 1~ My conclusion, then, is not that theoretical incorporation of the notion of cybernetic regulation is a panacea for teleosemantics, but that it is a step forward. By moving from design-oriented thinking to control-oriented thinking, teleosemanticists can overcome one of the large obstacles on the difficult road to a final theory of the mind. The progress suggested in this paper is partial, but on such a long journey even partial progress must be some encouragement.
Acknowledgments Thanks are owed to a number of people, but most especially to Fred Dretske, John Perry, Ken Taylor, and Peter Godfrey-Smith, for numerous conversations about the ideas presented here. Comments from John Heintz, Jason Holt, and Carl Matheson strengthened earlier drafts.
References Adams, E (1979). A goal-state theory of function attributions. Canadian Journal of Philosophy, 9, 493-518. Cummins, R. (1996). Representations, targets, and attitudes. Cambridge, MA: MIT Press. Davidson, D. (1987). Knowing one's own mind. The Proceedings and Addresses of the American Philosophical Association, 60, 441-458. l0 Those who want reasons to be optimistic should see recent papers by Papineau (1998) and Rowlands (1997) on solving the disjunction problem for teleosemantics.
106
Timothy Schroeder
Dretske, E (1981). Knowledge and the flow of information. Cambridge, MA: MIT Press. Dretske, E (1988). Explaining behavior: Reasons in a world of causes. Cambridge, MA: MIT Press. Fodor, J. (1975). The language of thought. Cambridge, MA: Harvard University Press. Fodor, J. (1990). A theory of content: And other essays. Cambridge, MA: MIT Press. Godfrey-Smith, P. (1994). A modem history theory of functions. Nous, 28, 344-362. Kaas, J. (1991). Plasticity of sensory and motor maps in adult mammals. Annual Review of Neuroscience, 14, 137-167. Lycan, W. (1987). Consciousness. Cambridge, MA: MIT Press. Merzenich, M., & Jenkins, W. (1993). Reorganization of cortical representations of the hand following alterations of skin inputs induced by nerve injury, skin island transfers, and experience. Journal of Hand Therapy, 6, 89-104. Millikan, R. (1984). Language, thought, and other biological categories. Cambridge, MA: MIT Press. Nagel, E. (1961). The structure of science: Problems in the logic of scientific explanation. New York: Harcourt, Brace & World. Papineau, D. (1998). Teleosemantics and indeterminacy. Australasian Journal of Philosophy, 76, 1-14. Recanzone, G., Merzenich, M., Jenkins, W., Grajski, K., & Dinse, H. (1992). Topographic reorganization of the hand representation in cortical area 3b of owl monkeys trained in a frequency-discrimination task. Journal of Neurophysiology, 67, 1031-1056. Rowlands, M. (1997). Teleological semantics. Mind, 106, 279-303. Sterelny, K. (1990). The representational mind: An introduction. Cambridge, MA: Basil Blackwell.
Chapter 6
Representation and Experience Frank Jackson
Perceptual experience represents the world as being a certain way. The experiences of something feeling round and of something looking red and in front of one represent, respectively, that there is something round in contact with one's body, and that there is something red in front of one. But many have gone further and proposed that a perceptual experience's representational character exhausts its experiential character. The "feel" is nothing over and above how things are being represented to be. 1 Although we will see later that this requires significant elaboration, this short statement of representationalism will do for now to identify our subject. I am a Latter-Day representationalist. I am not, though, going to offer a full-scale defence of representationalism. That would take too long, and anyway I am too late on the scene. I will instead do three things. First, I outline the path from the sense datum account of perceptual experience, the account I once espoused, to representationalism. I include here an outline of why representationalism is so attractive. Secondly, I address the challenge from externalism about content to representationalism. This is, I expect, the most controversial part of the paper. In the concluding part, I defend a traditional strategy for getting "feel" or qualia from representational content.
I They include Harman (1990), Tye (1995), and Lycan (1996). (Lycan's endorsement of representationalism is a partial one. In his view, representational content delivers a good part of phenomenal content but is not the whole story.)My debt to them will be obvious, as will my debt to some opponents, notably Block (1999). My biggest debt, however, is to David Armstrong. I should have believed much more than I then did of what he was saying in the 1960s. For a recent statement of his version of representationalism, see Armstrong (1999). Representation in Mind: New Approaches to Mental Representation Edited by H. Clapin, P. Staines and P. Slezak Copyright 9 2004 by Elsevier Ltd. All rights of reproduction in any form reserved. ISBN: 0-08-044394-X
108
Frank Jackson
The Path from Sense Data to Representationalism and Why We Should Take It The sense datum theory is an act-object account of sensory experience, an account that captures the nature of experience through the properties of the objects of awareness. On the act-object view, the difference between an itch and a pain lies in the difference between what one is aware of, and not in the mode of awareness as in adverbial theories. It is, therefore, not surprising that, as a former sense datum theorist (see Jackson 1977), I find representationalism very attractive. Both theories see the nature of experience as lying in the properties of the objects of experience, with the big difference that, for representationalists, the properties of the objects of experience reside in the way that experience represents things as being. There need be nothing actually having the properties; the "objects" are intentional objects. Indeed, representationalism is what you get when you take sense datum theorists and confront them with their theory's central failing - - which, contrary to common opinion, is not that it reifies appearances. The experience of there looking to be something red in front of one represents that there is, in the world, something red in front of one. This means that the sense datum theory's analysis of the experience of there looking to be something red in front of one in terms of the direct awareness of something mental that actually is red, the sense datum, must preserve this feature. We one-time sense datum theorists thought that the requirement that there be something which is red of which the subject is directly aware, automatically captures, or part way captures, the key representational notion. This is a mistake. It is true that I can represent how I am representing something to be by using the actual way something is. For example, I might represent to you the colour I remember the murderer's coat to be by holding up an actual sample of the colour. Here I would be using the actual colour of one thing, the sample, to represent how my memory represents the colour of something else, the murderer's coat, to be; a colour which the coat may or may not have. In that sense, we have a model for understanding the sense datum theory. But, and this is the crucial point, the fact that I am using an actual sample of the colour cuts no representational ice per se. I could be using the sample to represent the one colour I do not think the murderer's coat to be. Or I could be following the convention of holding up a sample with the colour complementary to that I remember the murderer's coat to be. In the same way, standing in a certain direct-awareness relationship to a mental item with such and such properties says nothing, represents nothing per se, about how the world is. The extraordinary (as I now think) failing of the sense datum theory is that it does not start to address the representational nature of perceptual experience. It
Representation and Experience
109
somehow manages to leave out the most important part of the story. The obvious repair is to replace the sense datum's theory's positing of a direct awareness relation to something red by a representing that there is something red; this transforms the sense datum theory into representationalism. One reason that representationalism is so attractive is the diaphanous nature of sensory experience, as G. E. Moore (1922) puts it. But let me make the point via Hume's famous remarks on the Cartesian self in the Treatise. Hume found himself unable to access the self; one or another experience always got in the way. We representationalists find the same problem with phenomenal character conceived as something distinct from how things are being represented to be. Whenever we seek to access it so conceived, we find ourselves accessing the putative way experience is representing things to be. A second reason for finding representationalism attractive concerns the way difference in representational character supervenes on difference in phenomenal experience. It seems to me that whenever there is a difference in phenomenal character, there is a difference in how things are being represented to be in experience. If this is right, then, by contraposition, sameness of representational character ensures sameness in phenomenal character; equivalently, phenomenal character supervenes on representational nature. But then, somehow or other, representational character suffices for phenomenal nature. It follows that, if we are smart enough, we should be able to identify which aspects of the representational account are enough to deliver without remainder phenomenal nature. Somewhere in the representational story we must be able to find all we need to make phenomenal character. There have, of course, been attempts to describe cases where phenomenal character differs without a difference in representational content and an important exercise is the critical review of all the cases that might be thought to show the possibility of phenomenal variation without difference in representational content. I am not going to conduct this review, because I think the job has been well done by other supporters of representationalism (e.g. Tye 2000). Thirdly, there is a marked contrast between, on the one hand, the way familiar representational devices like maps and sentences represent, and, on the other, the way perceptual experience represents. There is a gap between vehicle of representation and what is represented in the first kind of case that does not exist in the second. In the case of maps and sentences, we can distinguish the features that do the representing m the gap between the isobars on a weather map, the presence of the letters "c," "a" and "t" in that order in a sentence, the green colouring on parts of a map, etc. from what they represent: a pressure gradient, a certain animal, areas of high rainfall, etc. We can, for example, describe the gap between the isobars without any reference to what it represents. But in the case of perceptual experience, we
110
Frank Jackson
cannot. W h e n I have a visual experience of a roundish object in front of me, that is what it represents. My very description of the vehicle of representation delivers how it represents things to be. I m a y or m a y not accept that things are the way they are being represented to be, but there is just the one way that things are being represented to be, and that way is an essential feature of the experience. 2 Another way of putting the point is that maps, sentences, diagrams, barometers etc. require interpretation to deliver how they represent things to be, whereas all that is required in the case of an experience is that one have it. If it were the case that a perceptual experience's nature outran how it represents things to be, there would be an "extra bit" that called for interpretation as to how, if at all, it represented things as being. And there isn't. The final argument for representationalism is an ad hominen against its opponents. Its opponents talk of there being "feels" or qualia which are essential to the perceptual and which outrun the representational. 3 But they could hardly talk the talk unless their own states represent these qualia as existing and as being too "sensory" to be captured by representational content. The path from one's own states to one's reports and beliefs about those states must go via one's own representational states. But then they must, by their own lights, admit that representation can present as being itself very "feely."
Swampman ~
O r the E x t e r n a l i s m I s s u e
If we use "content" as a term of art for how things are being represented to be, then representationalism about experience is the view that an experience's content settles its experiential nature. Content in this sense is automatically truth-conditional: it is true just if things are as they are being represented to be. M a n y hold that there are decisive arguments to show that truth-conditional content is wide. Some allow that there is narrow content, but insist that it is not truth-conditional m in which case it is not content in the sense that lies at the heart of representationalism for the reason just given. I am one of the few, it seems,
2 How things are being represented to be need not be determinate. My experience may represent that something is a roundish shape without representing that it is any particular shape - - the experience represents that there is some precise shape it has but there is no precise shape that the experience represents it to have that shape. Indeed, it is arguable that all experience has some degree or other of indeterminacy about it. The same goes for maps and most sentences, of course. 3 They say the same for the sensory in general. I am, however, in this paper setting aside the issue of whether we should extend representationalism to feelings of pain, pangs of hunger, etc., though I think we should.
Representation and Experience
111
who think that there is truth-conditional narrow content. This means that I can admit the intuition that the sensory side of psychology supervenes on how one is from the skin in. I think this is a good thing. I find it very implausible that how red something looks or how round it feels should depend, in part, on the word usages of my linguistic community, my environment over and above how it affects me, the selectional history of my internal states, and so on, for the various "outside the skin" factors that many hold in part determine the contents of my mental states. Many say that I buy this bit of comfort at a very high price; they say that there are decisive arguments for externalism. I will seek to undermine the orthodoxy by pointing out that two of the most appealing lines of argument for externalism fail. But first let me pause to note a potential semantic confusion which has, I surmise, played a subterranean role in making many regard it as all but true by definition that truth-conditional content is wide. Many of my states represent something about my surroundings; for example, that there is a tiger six metres in front of me. They are true if and only if my surroundings are as they are represented to be. It follows that their contents are individuated by the relevant ways things have to be around me iftheir contents are to be true; they are individuated by positioned scenarios in Christopher Peacocke's 1992 terminology. In this sense, they are individuated externally m the difference between the contents that there is a tiger six metres in front of me and that there is a stuffed tiger six metres in front of me concerns how things are in that spot six metres in front of me. But it does not follow from this that the contents fail to supervene on how I am from the skin in. What is being represented to obtain may concern something external to me, consistently with its so representing the nature of my surroundings supervening on, being determined without remainder by, how I am from the skin in. The first, rather abstract line of argument, starts from the position that perceptual states, and mental states in general, are located inside people's heads. The externalism in question does not have mental states smeared through the space surrounding subjects; it is an externalism about content, not about the location of the states with the content. 4 But if they are inside heads, an obvious question is how they get to represent how things are outside heads. And an obvious way into this question is in terms of head-states' actual and possible causal links to subjects' environments. But then, runs the first argument for externalism, the environment must play a role in determining content. The quickest way to see the fallacy in this line of argument is to recall that water solubility is a narrow property. If X and Y are internally alike, they are alike
4 For why smearingthe states does not help with the problem of intentionality, see Jackson (1996: 389).
112
Frank Jackson
in whether or not they are water-soluble. 5 Nevertheless, water solubility is clearly largely a matter of actual and possible causal relations to the environment (to the water in it, especially). The reason water solubility is narrow is that, although certain actual-and-possible causal relations to water are crucial for being water soluble, it does not matter whether they are actual or merely possible. What matters is that a lump of sugar would dissolve in water, not whether or not it is ever put to the test, or even whether there is any water around in which it might dissolve. In the same way, we can grant that head-states' actual-and-possible causal links to subjects' environments are major factors (though not the only factors) in determining any contents they may have, consistently with holding that these contents are narrow provided that we insist that it is the actual-and-possible causal links, without regard to whether they are in fact actual or merely possible, that are the key. Some reply that Twin Earth arguments show that, in many cases, actual links matter for content. I will discuss that popular line of thought in a moment. Others reply that any appeal to actual and possible interactions with an environment must distinguish the normal cases from the abnormal cases w it is the interactions, be they actual or merely possible, with normal environments that are crucial. But, they proceed to argue, our understanding of what is a normal environment is tied to the actual one. The suggestion is not that the actual environment is necessarily normal but that it plays a role in settling what counts as normal. But this argument only leads to externalism if we give the subjects' actual environments a privileged role in settling what is normal. Only then is part of what settles the normal something that varies with subjects' environments. But I know of no persuasive arguments that, in addressing the admittedly hard problem of specifying what a normal environment is, we should give a constitutive role to the actual environment of particular subjects. 6 And I note that tying the normal to local environment is profoundly at odds with the view we take in discussions of the problem of induction m unless we are sceptics. Hume, on the interpretation of him as a sceptic, in effect argued that any population-cum-local-environment from which a sample is drawn is as normal as any other. This is how he arrived at scepticism about induction. If we want to avoid his fate, we had better insist that we can make sense of the notion of a normal environment independently of subjects or samples' more particular environments.
5 Provided we keep the laws of nature the same. Throughout our discussion of externalism, we are thinking of the issue as one about the possible dependence of content on surroundings while keeping the laws of nature constant. My unfashionable view is that content supervenes in the intra-world sense on internal nature, not in the inter-world sense. For more on this, see Jackson & Pettit (1993: 259-282). 6 For positive reasons why we should not, see, e.g. Segal (1991).
Representation and Experience
113
The second, more concrete line of argument for externalism makes considerable appeal to intuitions about possible cases; most famously, to intuitions about Twin Earth, although I will vary the diet. I will consider this influential line of argument at a little more length. I will argue that it (a) involves a subtle misreading of certain thought experiments as showing that certain (actual) causal connections to the environment in part determine content, whereas in fact the thought experiments show that the causal connections in question are part of the content, combined (b) with a misunderstanding about the connection between content and reference. Historically, intuitions about possible cases were deployed by Hilary Putnam (1975) and Saul Kripke (1980) to unseat the description theory of reference, the view that the reference of names goes by associated descriptions. I think we should always have been highly suspicious of these arguments. Describing cases and then appealing to intuitions about what various names m "water, .... beech," "Aristotle," "Goedel," for example m refer to in the cases as described does not look like a promising way to refute the view that reference goes by associated descriptions; instead, it looks like, and in my view is, a promising way to make explicit the descriptions that determine the reference of the terms. 7 In any case, it was a further step to warp the argument across from language to make it into one for externalism about mental content. I will focus directly on externalism about perceptual content, as that is most germane to our concerns here, and take as my starting point an initially persuasive passage from a paper by Ned Block. Ifphenomenal character supervenes on the brain, there is a straightforward argument against representationalism. For arguably, there are brains that are the same as yours in internal respects.., but whose states represent nothing at all. Consider the swampman, the molecular duplicate of you who comes together by chance from particles in the swamp. He is an example of total lack of representational content. How can he refer to Newt Gingrich when he has never had any causal connection with him; he hasn't ever seen Newt or anyone else on TV, never seen a newspaper, etc. (Block forthcoming) We can, I take it, spell out the line of thought here as follows. Paul Grice (1961) "did a Gettier" i.e. produced an example as compelling as Gettier's counter-example to the true, justified belief analysis of knowledge with his example of seeing something in a mirror when there is a duplicate object behind
7 For more on this point and the description theory of reference in general, see Jackson (1998).
114
Frank Jackson
the mirror in the apparent location. Grice argued out that what you see, the object your experience refers to, is not the object in the apparent location behind the mirror, but rather the object that causes the experience, despite the fact that your experience represents the object as being behind the mirror. Everyone agrees with him. This shows that what an experience is about, what it refers to, depends crucially on causal connections. But what an experience refers to is central to its content. Ergo, runs the argument, we have a Gettier-strength reason to hold that causal connections, causal connections to a subjects' environment typically, play a major role in determining the content of subjects' perceptual states. To see where the argument goes wrong, we need first to outline the connection between the content of a perceptual state and its referential properties. The content of a perceptual state is by definition how that state represents things as being, and the state represents accurately just if things are as they are represented to be. What the state is about or refers to is then given by whatever has to exist, and how things have to be, for things to be as the state represents them in the relevant respects, together with the relevant information about which things actually fill the bill. How, on this way of thinking about the reference of contentful states, does causation enter the picture? The answer is that it is part of perceptual content that one is causally interacting with one's environment. I here and now see that there is a certain biro in front of me. My perceptual experience is not merely that there is a biro in front of me; it is that there is a certain biro causally interacting with me in a highly distinctive way. There are, and I know that there are, many biros on the desk in front of me (my desk seems to attract biros), but only one is interacting with me in this way. This is how I am able to pick it out from the many other biros that are otherwise indistinguishable from it. Opponents of the description theory of reference often give examples where we refer to something without, they argue, there being any known property that identifies the object. An example sometimes given is seeing someone coming towards you in the fog. But we know that there is a property that marks the person seen out; only he or she is the fight kind of causal origin of the experience. What is involved in being the fight kind of causal origin? This is a hard question in the sense that that it is hard to give the answer in words although our ability to correctly locate, fetch and in general identify the things we see shows that we know the answer implicitly. We can, however, say a bit about it here. Perceptual experience tracks the world, and its representational contents systematically co-vary with, and counterfactually depend on, the world it putatively represents. Consider the contrast with what happens if I shut my eyes. I will most likely be in a state that represents that the biro is in front of me because I will continue to believe that the biro is in front of me. Moreover, my belief state is caused by the biro. But it will not be causally responsive to the biro in the way distinctive of perception; it will not be actively monitoring the biro; it will not track the biro.
Representation and Experience
115
In my view, what Grice brought out so decisively is the causal element in perceptual content; our perceptual experience of the world, inter alia, represents the world as interacting with, and being tracked by, our experience. W h e n we look into a mirror, our perceptual experience represents an array of objects as standing in certain causal relations to us, as well as having various colour, location and shape properties. The causal relations are more important than the location per se, and that is why we count as seeing the object we are interacting with, rather than the one with the "fight" location. Instead of saying that causal relations to our environment in part determine content, we should say that causal relations to our environment are part of content. My experience does not simply represent that there is a biro of a certain colour, location and shape in front of me; it also represents that object as interacting with me in a way that means that the content of my perceptual experience will, all going well, track the biro's colour, location and position. Thus, when I and m y Twin Earth counterpart, tFrank, look at, respectively, water and twater, in both cases our experience represents that we are interacting with watery stuff, where "watery stuff" is convenient shorthand for stuff that typically satisfies the usual list of superficial properties associated with water, or at least those our experience represents water or twater as having on the occasion in question. The content of our experience is the same but, because the stuff we are interacting with differs, what our experience refers to differs. This is not to say that the content of the sentence "Here is some water" in m y mouth and in tFrank's mouth is the same. That is a quite different question which turns, in my view, on whether it is the A-intension or C-intension of the sentence that is in question. 8 Nor is it to say that I and tFrank have perceptual contents with the same truth-conditions. We won't: mine will be true just if there is watery stuff I am interacting with; his will be true just if there is watery stuff he is interacting with. The (egocentric) perceptual contents are the same in the sense that they agree on how our respective environments are, but the conditions under which they are true differ. Mine is true when things are the fight way around me; his content is true when they are the fight way around him. 9
8 In the terminology of Jackson (1998). The key idea goes back at least to Stalnaker's distinction between the diagonal and the horizontal proposition in his 1978. Incidentally, it is plausible that in using the word "water" to describe how experience represents things to be, both Frank and tFrank are saying that there are only a small number of watery kinds, perhaps only one, with which they are acquainted. 9 The phenomenon of egocentric content means that we believers in narrow content have to be careful when we say that ifX and Yare alike from the skin in, then their perceptual contents are the same. Either we must explain that sameness of content does not necessarily imply having the same truth-conditions, but instead implies, following Lewis's (1979) way of putting it, ascribing the same property. Or else we must express what it is to be narrow in terms of invariance of perceptual content under changes in the surroundings of any given individual who remains unchanged from the skin (or central nervous system) in.
116
Frank Jackson
Why do I say that causation is part of the content? Why not say that there is no causal element in the content, and that causation only comes into the picture as a partial determiner of the content? Part of my answer is that it is intuitively very plausible that perception represents the world as interacting with us. Think, for example, of hearing that a sound is at so and so a location. When I hear a sound as being, say, behind and to the left, my experience represents the sound as being at this location by virtue of representing it as affecting me from this location. I hear the sound as coming from this location, and coming from is a causal notion. A similar point applies to vision. Vision represents things as being located where we see them as being, as being at the location from which they are affecting us via our sense of sight. But here is an argument to add to the claim of obviousness. Suppose, against my view, that perception represents to us that things are thus and so, but there is no causal element in the thus and so, and that we get reference by adding some causal conditions. This means that we would, in effect, have two notions of reference: referencel would be fitting enough of the thus and so, and reference2 would be fitting the added-in causal story, along with fitting some of the non-causal thus and so. (Grice did not of course teach us that the causal link is all that matters.) There would then need to be some argument that the second notion of reference is "real" reference. But now there is trouble. Grice did not need to do any additional arguing to convince us that we did not see the object behind the mirror. The universal reaction to Grice's example is immediate agreement that the object we see is the one playing a certain causal role; no diversion into the theory of reference is needed. Another way to put the same point is to note that the nature of the universal reaction to Grice's example is a powerful argument in itself that the causal connection is part of the content. The upshot is that we believers in narrow perceptual content can agree that what subjects' perceptions refer to often depends greatly on their causal connections to their more or less immediate environments without conceding that the subjects' contents are in part determined by these causal links. The situation will be like that with a definite description like "the cause of this pain." Its reference depends very obviously on causal links but this does not imply that its content depends on those links. 10 I have sketched my response to two arguments for externalism. But let me conclude this section with a general consideration in favour of narrow content. It seeks to show that there is no principled bar to allowing that a head state can in itself, independently of its environment, discriminate between ways things might
10Of course, some would argue that this is a bad analogy on the ground that definite descriptions do not refer, but see Jackson (1998).
Representation and Experience
117
be. I offer it because I think the tide against narrow content is running so strongly that many have come to the view that there is something wrong with the very idea that how things are in a relatively fairly small portion of reality might, in and of itself, point to how things are elsewhere. Palaeontologists wonder whether the dinosaurs were warm-blooded or coldblooded. What makes the question hard is that there are no obvious traces that being warm-blooded or being cold-blooded would have left in the fossil record. It is unclear what might count as sign that they were, or were not, warm-blooded because it is unclear what difference it would make to how things are now. I I More generally, scientists, historians, biographers, comparative sociologists, people who want to know the score from last night's basketball game, and so on, have a potential epistemic problem when they find themselves in the following situation. (1) They are located in a region of space-time A and want to know how things are concerning P in some other region of space-time B. (2) It is known that how things are in A would be exactly the same whether or not P in B. Often they can get over the problem. To say that they cannot would be to surrender to Descartes' Evil Demon argument. It would also make knowledge of the future impossible. When we move from how things are now to how things will be in the future, we know that how things are now would be exactly the same independently of how things will be m barring backwards causation. When we argue from the nature of region A to the nature of region B, we use traces when we can, but we also project patterns, extrapolate, analogise, etc. from how things are in A to how things are in B. The nature of the relevant principles is a controversial matter, but the point that is important for our discussion is that it is how things are in region A that those in region A perforce must use in arriving at their judgements about region B. "But suppose those now in region A went to region B before making their judgements about P in B. Why should they, in this case, be restricted to how things are in region A?" But consider the two possibilities. One is that the relevant facts in region B have left no traces available to those now in region A: no diary entries were made while our travellers were in region B, or perhaps the entries have faded completely; the casts of the footprints have been lost in transit; the sea washed over the marks in the sand; the one person who witnessed the key events in B died on the way to A; or whatever. In this case, the trip to region B changes nothing. The other possibility is that the facts in region B have left traces in region A by virtue of the visit to B. In either case, the situation is as before.
II Recent research has, I understand, identified current signs that at least some dinosaurs were warmblooded. These signs are putative traces of some dinosaurs' having had hearts with four chambers. But we can pretend otherwise for the sake of the argument.
118
Frank Jackson
Judgements about how things are in B, made by those in A, depend on how things are in A. Here is a way to make the point vivid. Consider a space ship moving through space. It is collecting information about the stars, asteroids and planets around it. Suppose that at some point the ship becomes sealed in the sense that nothing outside it any longer impinges on how things are inside the capsule. What it is rational for the persons inside the ship to believe about their surroundings may change after the sealing, as a result, for example, of discussions inside the ship or detailed calculations made on computers inside the ship, but what it is rational for them to believe will not change as a result of changes outside the ship. What it is rational to believe, as opposed to which beliefs are true, about how things are outside the ship supervenes on how things are inside the ship m on the data collected before the sealing, on the course of the discussions inside the ship, and so on. Some argue that externalism about content leads to externalism about rationality. 12 They say that what it is rational to believe inside the ship can change with changes outside the ship, regardless of whether or not those external changes have any effect whatever inside the ship. I think this violates very plausible "ought implies can" principles. If the changes outside have no effect inside, it is impossible for the people inside the ship to get information about the changes outside. (That is why people get upset about breaks in transmission etc.) How then could it be the case that the outside changes mean that those inside o u g h t to believe differently? The relevance to the question of narrow content is that we are parcels of matter moving through the world. On pain of embracing scepticism, we have to allow that the narrow nature of these parcels of matter can point to how things are around us. Our bodies are "region A," and our environment is "region B." As already noted, it does not follow from this that belief or perceptual experience are in fact narrow states of persons. It does, though, follow that they might be in the sense that there is no principled bar to there being narrow states of persons which point to how things are elsewhere.
Getting "Feel" from Content Suppose I am here and now having a visual experience as of a round, red object in front of me. It will be true that it is being represented to me that there is something
12See, e.g. Millikan (1993, Chapter 14). Some opponents of internalism about justification seem to think that we internalists are committed to subjects' access to the justification and nature of what they believe being a priori. But being square is an internal property, i.e. one that supervenes on a small region of space, but our knowledge of which things are square is not a priori.
Representation and Experience
119
round and red in front of me. But this much content is not enough to get the requisite phenomenology. I can believe that there is, here and now, a round, red object in front of me without having the relevant visual experience. Perhaps my eyes are shut but I remember, or perhaps I am being told, that there is such an object here and now in front of me. Or perhaps the thought that there is such an object in front of me has simply "come into my mind," and I have boldly gone along with it. Or perhaps I am one of the blind sighted: it seems like guessing but my success rate shows that I am drawing on a subliminal representational state. Despite the identity in content, the phenomenologies in these kinds of cases are very different from seeing. One traditional approach by representationalists to "finding the feel" has been in terms of the richness and kind of content. (I take it this is Armstrong's approach.) The basic idea is that if the representational state's content is that rich and of that kind, you get the phenomenology for free; there must, that is, be the kind of experience that the blind sighted and the bold guessers lack. On this approach, it is not an interesting empirical fact that the blind sighted have rather meagre views about how things are as far as colour and position beyond the reach of their hands goes; if you suppose that their representational contents are enough like ours in richness and kind m if you put in enough detail about colour, location, and the like m you ipso facto give them vision, our vision. This is why I said at the beginning that the quick statement of representationalism that the redness of seeing red comes from its representing that there is something red needs elaboration: it had better be the representing that there is something red as a part of representing that things are such and such, for some very rich such and such. Embedding in suitably rich content arguably enables us to handle the difference between seeing something as flat and feeling it as flat. These experiences are very different but they both represent that something is flat. However, that something is flat is a component of two very different rich ways things are being represented to be in the two cases. In the first, it is packaged with colour and "at a distance" spatial properties; in the second with temperature, felt texture and impact on one's body properties. This idea is typically supplemented by adding, first, that the representation must be direct. Reading from a piece of paper that there is something of such and such a colour, location, etc. typically induces a belief that represents that there is, but does so via representing that there is a piece of paper with certain marks on it. And, second, it is supplemented by adding that consciousness of the experience m which is not to be confused with the phenomenology per se ~ is a matter of being in a second-order state that represents that one is representing that there is something of such and such a colour, location, etc. So that what the blind sighted lack is both richness of content and consciousness.
120
Frank Jackson
Second-order representations are also needed to handle a point emphasised by Peacocke. Seeing something as a square standing on one of its corners is not the same as seeing something as a regular diamond, but the representational content is the same m there is no difference in how things are being represented to be. (See e.g. Peacocke (1992: Section 3.3).) The difference lies in the difference between having a meta-representation that a first order representation represents a figure that is symmetrical about a line bisecting its sides vs. having one that represents a figure that is symmetrical about a diagonal. I am sympathetic to this way of handling the problem of finding the feel, but I have little to add to what writers like Armstrong said many years ago except that the discussion in the previous section tells us it would be good to add a causal element to the account of the fight kind of content for feel. I want to discuss an alternative approach that has become popular recently. Some representationalists tackle the problem of finding the "feel" via a distinction between conceptual and nonconceptual content. (See, e.g. and esp., Tye (1995, 2000).) The idea is that belief has conceptual content, whereas visual perception has nonconceptual content. I think that there are serious problems for this style of response. The view that beliefs have conceptual content whereas experiences have nonconceptual content can be interpreted in two very different ways. It can be thought of as the view that beliefs and experiences have content in different senses of"content." I think this is the wrong way for representationalists to go. 13 Belief is the representational state p a r excellence. If belief does not represent that things are thus and so, I do not know what does. This means that to hold that experience has content in some different sense to the sense in which belief does is to deny rather than affirm representationalism. There needs, it seems to me, to be an univocal sense of "content" at work when we discuss representationalism; a sense on which content is how things are being represented to be and on which both beliefs and experiences have content. 14 Of course, this allows that when we have other fish to fry, it might be good to introduce many kinds of content. Also, to avoid misunderstanding, I should emphasise that in saying that a belief's content is how it represents things to be I am not taking a stance against, for example, the view
131think it is the way Tye wants to go but I am unsure. But let me say that here, and in the immediately following, I draw on helpful if unresolved discussions with him. 14Some say that the belief that there is water and the belief that there is H20 agree in how things are being represented to be, but do not, on any at all plausible view, have the same content. I say that they do not agree in how things are being represented to be but that the sentences used to express their contents agree in content in the sense of agreeing in C-intension.
Representation and Experience
121
that a belief's object is a proposition, where a proposition is a set of possibilities, or a set of things and concepts, or a structured combination of universals or .... These set-theoretic entities all relate, in their different ways, to how things might be m for example, the last connects to the particular way things might be which has the universals instantiated in the relevant structured w a y - and so deliver content in the representational sense that is centre stage in discussions of representationalism. In consequence, I think representationalists should think of the distinction as one between what it takes to have a contentful belief vs. what it takes to have a contentful perceptual experience. The idea is that experience represents in a way that is independent of subjects' mastery of concepts, whereas belief does not. For example, it is observed that we can perceptually discriminate many more colours than we have names for or can remember. It is then inferred that I might have a perceptual state that represents that something is, say, red l7 without having the concept of red lT. But I could not believe that something is red17 without having the concept of redl 7. I very much doubt the claim that perceptual representation is nonconceptual in the explained sense. Perhaps - - but see below - - I can perceptually represent that something is red l7 without the concept of red l T, but I cannot represent that something is red17 without the concept of red or of colour. To perceptually represent that things are thus and so essentially involves discrimination and categorisation, and that is to place things under concepts. Again, the moral Wittgenstein drew from the duck-rabbit example is that there is no sharp distinction between how things look and the concepts your perception represents things as falling under. Against the views of my former sell I think there is no way of abstracting out a "pure" look from a perceptual experience's pointing towards its being a duck or its being a rabbit. Of course, I can see the outline purely as a geometric shape, but then the experience points towards there being something of that shape as the point is often put, all seeing is seeing as. Again, there is an intimate connection between perception and belief. If Armstrong (1968, Chapter 5), is fight, perception can be analysed in terms of belief. But, be this as it may, for each perceptual experience, there is a belief which it paradigmatically points to ~ that is part of the burden of the point that perception represents. Deny that and I am not sure what is meant by insisting that perception represents in the "no gap" way I described near the beginning. But then it must be the case that perceptional content is conceptual. Further, I doubt the claim that when I experience red17 I need not have the concept of redlT. I need not have the term "redlT" in my linguistic repertoire; I need not be representing that the colour before me is the colour correctly tagged "red lT"; and my ability to remember and identify the precise shade may be very
122
Frank Jackson
short-lived. But none of these points imply that I do not have the concept of red l7 at the time I experience it. When I learn the right term for the shade I can see, namely, the term "red 1T," it will be very different from learning about momentum, charm in physics or inertial frames, which undoubtedly do involve acquiring new concepts. It will simply be acquiring a term for something I already grasp. Any thought to the contrary would appear to conflate the concept of red l7 - - the shade with the distinct, relational concept of being indistinguishable from the sample labelled "red l7" in some colour chart. It might be objected that this latter concept is the one we have in fact been talking about all along. But if this is the case, the initial datum that we experience red l7 prior to acquaintance with colour charts is false. Prior to acquaintance with colour charts, we do not experience colours as being the same as such and such a colour on a colour chart. The same goes for shapes. It is sometimes suggested that when presented with a highly idiosyncratic shape, I will experience it but not have the concept of it. But we need to distinguish two cases. In one, I see something as having the highly idiosyncratic shape but lack a word for it. In this case, I do have the concept. What I lack is a word for it, which I can remedy by making one up for myself or by asking around to find out if there is already one in, say, English. In the second kind of case, I do not experience the shape prior to having the word and the concept. There are cases where I see that something has some complex shape or other, where that shape is in fact S, but fail to see it as having S. I simply see it as having some complex shape or other. I am then told the fight word for the shape, acquire the concept it falls under, and thereby acquire the ability to see it as having S. But then it is false that my experience represented that something is S prior to my mastery of the concept. My acquisition of the concept changes the perceptual experience. It might be suggested at this point that the claim that perceptual content is nonconceptual essentially comes to a claim about the possibility of perceptual representation without mastery of the relevant bits of a language. But then we lose the needed sharp c o n t r a s t - needed, that is, by the approach under discussion between a visual experience representing that there is something round, red, etc. and simply believing that there is, unless, of course, we take the hard road of denying that belief without language is possible. (I belong to the party that thinks it is more obvious that dogs have beliefs than is the cogency of any philosopher's argument that belief requires (public) language.) In e-mail discussion, it has been objected that there are cases where someone has one or another perceptual experience without noticing that they do. I agree; indeed, I think this is the norm. I think that most of the perceptual experiences I have when I walk across a room are one's I do not notice. What shows that I have these experiences is my ability after the event to answer various questions
Representation and Experience
123
about the location, shape and so on, of various objects along the path between me and my objective - - coffee, the tinging phone or . . . . and the way I avoid walking into them. But the same is true for belief. Corresponding to the cases we find it plausible to describe as unnoticed seeing are cases we find it plausible to describe as unnoticed believing. There is, it seems to me, no asymmetry here. 15
References Armstrong, D. M. (1968). A materialist theory of mind. London: Routledge. Armstrong, D. M. (1999). The mind-body problem. Boulder, CO: Westview Press. Block, N. (1999). Sexism, racism, ageism and the nature of consciousness. In: R. Moran, J. Whiting, & A. Sidelle (Eds), The philosophy of Sydney Shoemaker, philosophical topics (Vol. 26, Nos 1, 2). Block, N. (forthcoming). Mental paint. In: M. Hahn, & B. Ramberg (Eds), A book ofessays on Tyler Burge. MIT Press. Grice, H. P. (1961). The causal theory of perception. Proceedings of the Aristotelian Society, 35(Suppl.), 121-168. Harman, G. (1990). The intrinsic quality of experience. Philosophical Perspectives, 4. Hume, D. (????). Treatise. Treatise of human nature. Jackson, E (1977). Perception. Cambridge: Cambridge University Press. Jackson, E (1996). Mental causation: The state of the art. Mind, 105, 377-413. Jackson, E (1998). Reference and description revisited. In: J. E. Tomberlin (Ed.), Philosophical perspectives, Volume 12, Language, mind, and ontology. Cambridge, MA: Blackwell. Jackson, E, & Pettit, P. (1993). Some content is narrow. In: J. Heil, & A. Mele (Eds), Mental causation. Oxford: Clarendon Press. Kripke, S. (1980). Naming and necessity. Oxford: Basil Blackwell. Lewis, D. (1979). Attitudes De Dicto and De Se. Philosophical Review, 88, 513-543. Lycan, W. G. (1996). Consciousness and experience. Cambridge, MA: MIT Press. Millikan, R. (1993). White queen psychology; or, The last myth of the given. In: White queen psychology and other essays for Alice. Cambridge, MA: MIT Press. Moore, G. E. (1922). The refutation of idealism. In: Philosophical studies. London: Routledge & Kegan Paul. Peacocke, C. (1992). A study of concepts. Cambridge, MA: MIT Press. Putnam, H. (1975). The meaning of "meaning". In: K. Gunderson (Ed.), Language, mind, and knowledge, Minnesota Studies in the philosophy of science (No. 7). Minneapolis: University of Minnesota Press.
151 am indebted to many discussions, especially one at New York University where an earlier version of this paper was on the table.
124
Frank Jackson
Segal, G. (1991). Defence of a reasonable individualism. Mind, 100(4), 485-494. Stalnaker, R. (1978). Assertion. In: E Cole (Ed.), Syntax and semantics: Pragmatics (Vol. 9). New York: Academic Press. Tye, M. (1995). Ten problems of consciousness. Cambridge, MA: MIT Press. Tye, M. (2000). Consciousness, color, and content. Cambridge, MA: MIT Press.
Chapter 7
Phenomenal Qualities and Intermodal Perception Ian Gold
Two views about consciousness, according to Ned Block (1990, 1996), divide the philosophy of mind. On one view, conscious mental states both represent and have particular subjective qualities. On the other, mental states are exclusively representational. In this paper I raise a problem for the view that mental states have phenomenal qualities. I describe some of the commitments of the view and indicate that it is a version of a traditional empiricist picture of mental states, especially in the domain of perception. I then discuss a well-known problem in the history of perception known as Molyneux's question which embodies a thought-experiment designed to test the success of empiricism about perception. I discuss some recent advances in developmental psychology and review their consequence for traditional empiricism. I then suggest that as a form of empiricism, phenomenism is undermined by this consequence. I conclude by considering some options for phenomenism.
Phenomenal
Qualities
The question at issue in the debate about the properties of conscious mental states is: Is there anything more to a mental state than what that state represents ? Of course, mental states have lots of properties other than their representational ones. If you are a reductionist, for example, you believe that mental states are identical to states that have all sorts of neural properties. Block's question, however, is concerned Representation in Mind: New Approaches to Mental Representation Edited by H. Clapin, P. Staines and P. Slezak 9 2004 Published by Elsevier Ltd. ISBN: 0-08-044394-X
126
Ian Gold
with experiences. The question is: Is there anything more to an experience than what that experience represents ? The traditional answer to this question is that there is. While many or most mental states don't have a particular "feel" to them, some do. The paradigm case of this kind of state is perceptual. When I see the purple carpet, my perceptual state represents the colour of the carpet. But there is a particular feeling associated with this representation that distinguishes it from the feeling I get when I see a green carpet and that is entirely absent when I merely believe that the carpet is purple. This apparent fact has led many to hold that once the representational content of a mental state is set aside, there is a remainder which is purely phenomenal. Block (1996) calls this view phenomenism. According to phenomenism, perceptual states, among others, have phenomenal qualities. Further, these qualities participate in the representational function of the perceptual state but also contribute a subjective f e e l i n g - a "raw feel," "quale," "whatit'slike," or what have y o u - to the state. Block offers the striking example of orgasm. Whatever the representational content of the experience of o r g a s m - for example, that some process is occurring in one's body ~ there seems to be something in the experience that is not exhausted by that content. If the phenomenal properties are salient but the representational content is not, then it looks like the phenomenal properties "outrun," as Block (1996) puts it, the representational content: Orgasm is phenomenally impressive and there is nothing very impressive about the representational content that there is an orgasm. I just expressed it and you just understood it, and nothing phenomenally impressive happened (at least not on my end) (Block 1996: 33; emphases in original). Block (1996) characterises phenomenal qualities as "mental paint," and this analogy points to the double duty done by phenomenal qualities. Most of the phenomenal features of ordinary perceptual experiences are in the service of representation. When some purple mental paint is present in my experience, it is often there to represent the purpleness of the carpet, say, just as real purple paint can be used to represent the purpleness of a carpet. One can also attend, however, to the purpleness of the mental paint in isolation from its representational function, and sometimes purple mental paint may fail to have any representational function at all. A purple phosphene ~ the visual image you get by putting pressure on your eyeball - - can be thought of as a bit of non-representational mental paint, just as a bit of real paint can fail to represent anything in a painting. There are a number of views that are closely related to Block's phenomenism but are nonetheless different from it. One related view is that experience has both
Phenomenal Qualities and Intermodal Perception
127
phenomenal and representational properties but that there is no straightforward relation between the two. On this view, what represents the carpet as purple has nothing to do with the purple "feel" associated with seeing the carpet. A second possible view is that experience has phenomenal properties, and that phenomenal properties have a representational function, but the properties that do the representing are not those that give the experience is characteristic feel. For example, one might believe that there is mental paint involved in my experience of seeing a purple carpet, that the paint represents the purpleness of the carpet, but that the paint is not in fact purple but, say, green. This would be analogous to some actual paintings in which colored objects are represented by paint that is not everywhere the color of the object represented. It would be a mistake, however, to think that these views are Block's. Central to his view is the claim that mental paint is introspectible. Suppose you are looking at a ripe tomato in good light. There are, Block (1996) says, [m]ental properties of the experience that represent the redness of the tomato. This is mental paint. According to me, the phenomenal character of the experience is such a mental property: it represents the tomato as red. According to me, one can have introspective access to this phenomenal character (p. 29). Now if mental paint includes the property of experience that represents the tomato as red, and that property is introspectible, then it is hard to see how that property could be anything other than red in "feel." For if one introspected that property and it appeared green, then (in normal circumstances I ) it would seem impossible to explain why one nonetheless wants to say that one's experience represents the tomato as red. The phenomenist wants to say that his experience represents the tomato as red because the phenomenal property that is introspectible is itself red in feel. Many philosophers are phenomenists but not all. Harman (1990, 1996a) is a prominent defender of the opposing view according to which there are no such phenomenal qualities. On his view, when the representational content of a mental state is set aside, nothing is left over. Block (1996) calls this view representationism. Harman argues that the error of the phenomenist lies in taking the analogy between mental paint and real paint too strictly. Of course there is something in the head of the perceiver the presence of which represents the redness of a tomato. Whatever that is, however, is not accessible to the perceiver; it can only be an object of theory
1 'Normal' is meantto excludecases of spectruminversion and the like. Some such cases are discussed below.
128
Ian Gold
and experiment in psychology and neuroscience. One cannot, therefore, attend to its features the way one can attend to the features of real paint. The defender of mental paint points to the redness of a tomato and asks us to notice that this feature, though representational, can also be thought of purely qualitatively (and indeed may appear on occasion in a non-representational guise as in the case of a red phosphene). The view that denies that mental paint, construed in this way, exists holds that one can only notice the redness of the tomato and not of whatever it is in the head that represents that redness. To suggest that one can is, as Harman (1990) puts it, to confuse the properties that experience represents a scene as having with the properties of experience in virtue of which it represents that scene (see also Place 1956). The tomato is red, not the state of the visual system that represents it as red: When you see a ripe tomato your visual experience represents something as red. The redness is represented as a feature of the tomato, not a feature of your experience... Does your experience represent this redness by being itself red at a relevant place, in the way that a painting of a ripe tomato might represent the redness of the tomato with some red paint on the appropriate place on the canvas? No. That is not how visual representation works. Does your experience represent this redness by having at some place some quality.., which serves to represent the redness of the tomato in some other way, different from the way in which a painting might use red paint to represent a tomato? Well, who knows? You have no conscious access to the qualities of your experience by which it represents the redness of the tomato (Harman 1996a: 7-8). Two central commitments of the phenomenist will play a role in what follows. I consider each in turn.
Thesis 1: Accessible Representational Content Supervenes on Phenomenal Qualities As indicated above, phenomenal qualities do double duty in perception. They are the properties that give a perceptual experience its felt quality, and they are the representational building blocks of full-blooded perceptual states. Block (1996) is explicit about this: I think sensations - - almost always - - perhaps even always have representational content in addition to their phenomenal
Phenomenal Qualities and Intermodal Perception
129
character. What's more, I think that it is often the phenomenal character itself that has the representational content. What I deny is that representational content is all there is to phenomenal character (p. 20, emphasis in original). Block's view requires clarification, but he seems to be following out the analogy between mental paint and real paint in the following way. Think of a portrait of van Gough and one of Rembrandt. A variety of different configurations of paint could constitute a portrait of van Gough. Van Gough himself produced a number of them. One might think, however, that what is not possible is for a portrait of van Gough and one of Rembrandt to have exactly the same configurations of paint. 2 Usually what is represented by a painting, a drawing, a photograph, or whatever depends in part on the visual properties of the representation. One way to read Block is as advocating the same view with respect to perceptual representation. The representational content of a perceptual state is a function of the particular phenomenal qualities and their organisation in the perceptual field. Different qualities, or a different organisation of qualities, produces a different representational content. Peacocke (1983) has developed this idea in some detail. (Peacocke 1992 advances a rather different version of this idea.) He distinguishes between the properties of an object say the redness of a ripe tomato and the properties of the visual field that represent that redness, and it is the visual field that is the bearer of phenomenal qualities. The visual field has the distinct, but related, property of redtness in virtue of which it represents the redness of the tomato. One consciously experiences redtness and, as a result, one's percept has a representational content that includes the redness of the tomato. The properties of the visual field emphasised in Peacocke's view are thus systematically related to properties of the scene. A red t patch in the visual field is often, though not always, related to an area of the scene that is red. Were that area of the visual field to undergo a change in properties, the scene represented might also change in some respect. More importantly, for a change in the scene represented, there would have to be a change in the properties of the visual field. The organisation of phenomenal properties thus participates in determining the representational content of the perceptual state, and, in general, if there is to be a change in representational content, there will have to be a change in phenomenal properties.
This is not of course meantto be taken as universallytrue for works of art. Two identical black dots on a white canvas could be titled "Portrait of van Gough" and "Portrait of Rembrandt" without violating any general principle of representation. 2
130
lan Gold
It is tempting, therefore, to suggest that representational content supervenes on phenomenal qualities and their arrangement: any change in representational content requires a change in phenomenal qualities but not vice versa. If one is an internalist about mental content, then this seems right. According to internalism about mental content, content is determined entirely by states of affairs "from the skin in." The content of Mary's perceptual state, for example, is entirely determined by the non-relational properties of her mind. The external world itself plays no role in contributing to that content. Suppose, then, that you are an internalist phenomenist. You believe that the content of your perceptual state is determined entirely by the relevant states of your mind. You also believe that those states are phenomenal states because the phenomenal qualities are what bear the representational content of perceptual states. It follows that the representational content of your perceptual state supervenes on the phenomenal qualities of that state: since phenomenal qualities and nothing else determine representational content, there can be no change in representation without a change in phenomenal qualities but not vice versa. However, if externalism about mental content is true, then this supervenience thesis does not hold. Externalism holds that the content of a perceptual state, for example, depends in part on the object or property represented by that state. It is possible, therefore, to dissociate phenomenal qualities and representational content. Consider Block's (1996) "inverted earth." On inverted earth, grass is red, tomatoes are green, and so on. Suppose Mary travels to inverted earth after having had inverters implanted in her eyes. Inverters make green objects look red, and red objects look green. When Mary opens her eyes on inverted earth she sees grass looking green. But, of course, on the externalist view, the representational content of Mary's perceptual state is that grass is red. There is a change in representational content without a change in phenomenal qualities thus violating the proposed supervenience claim. Notice, however, that Mary's access to the representational content of her perceptual state is restricted by the phenomenal qualifies of her experience. If Mary is unaware that she is on inverted earth, or unaware of the properties of inverted earth and the inverters that have been implanted in her eyes, she will believe falsely that the representational content of her perceptual state is that grass is green. A different supervenience thesis is therefore suggested, namely, that the access a perceiver can have to her mental states - - and thus the beliefs that she can form about t h e m - supervenes on the phenomenal qualities of those states. Everything else being equal, without a change in phenomenal qualities, one cannot have a change in access to the content of one's perceptual states. Of course, access to one's mental states can be inadequate to form true beliefs. If externalism is true, then Mary falsely believes that her experience represents grass as green
Phenomenal Qualities and Intermodal Perception
131
when in fact it represents it as red. This is the case because perceptual experience is usually all one has to go on when forming beliefs about the world on the basis of perception. One does not, typically, have independent evidence about the properties of the objects that caused one's perceptual experience. Therefore, the beliefs that one can form about the external world on the basis of perception are typically limited by the access one has to the representational content of one's perceptual experience, and this access is typically dependent on the phenomenal qualities of one's experience. In the typical case, therefore, where there is no change in phenomenal qualities, there can be no change in one's perceptual beliefs. 3 I am not concerned to pursue the issues that arise as a result of the debate between internalism and externalism, and the weaker supervenience claim will do for what follows whether one is an internalist or an externalist. I will therefore take the weaker claim to be the one to which phenomenism is committed.
Thesis 2: Phenomenal Qualities Are Modality-Specific The phenomenist position makes a second assumption that phenomenal qualities are specific to a perceptual modality. Berkeley (1713/1975) provides a well-known expression of this view: [W]hen I hear a coach drive along the streets, immediately I perceive only the sound; but, from the experience I have had that such a sound is connected with a coach, I am said to hear the coach. It is nevertheless evident that, in truth and strictness, nothing can be heard but sound; and the coach is not then properly perceived by sense, but suggested from experience. So likewise when we are said to see a red-hot bar of iron; the solidity and heat of the iron are not the objects of sight, but suggested to the imagination by the colour and figure, which are properly perceived by that sense. In short, those things alone are actually and strictly perceived by any sense, which would have been perceived in case that same sense had then been first conferred on us (p. 161). A recent illustration of commitment to this view can be found in Clark (2000). He is concerned with a distinction made by Block (1995) between access-consciousness
3 This sort of consideration can be used to develop an argument for a rejection of externalism altogether. See, for example, Jackson (this volume).
132
lan Gold
and phenomenal-consciousness. One has access-consciousness when the content of a mental state is available for certain kinds of cognitive activity or action; phenomenal-consciousness refers to the feel of experience. Clark's purpose is to argue that there is an important relation between these two forms of consciousness. He argues as follows: I have access not just to the products of my sensory activity, but also to certain aspects of the sensory activity itself. For example, I am noninferentially aware that I am using a visual rather than a tactile modality. I am aware that I see, rather than hear or feel, the difference. But in that case, it [a perceiver] must say that there is something it is like to see the difference rather than e.g. to smell it. For in what else could direct introspective (non-inferential) access to the modality consist? To be access-aware of the act of detecting a difference requires at least saying (honestly) that the two items seem different in some modality-specific respect. So in this case, access-awareness (of the act of detecting a difference using a specific modality) seems to imply that there is (or is reported to be) something it is like to detect the difference (pp. 30-31). I am not concerned to evaluate whether Clark has in fact demonstrated a connection between access and phenomenal consciousness but only to suggest that the argument illustrates the second thesis about phenomenal qualities. Clark's argument runs from the claim that it is possible to identify the modality in which one is perceiving to a claim about one's introspective awareness of the phenomenal features of one's experience. The argument must assume, therefore, that phenomenal qualities are modality-specific. Without that assumption, the mere awareness that one is perceiving in a particular modality would have no obvious consequence for facts about phenomenal qualities. The modality-specific nature of phenomenal qualities is sometimes made in the context of asking how the different sense modalities can be distinguished. It is a commonplace that there are five distinct sense modalities, but this is usually taken to be a starting assumption rather than a matter for investigation (Stoffregen & Bardy 2001). One plausible view about how to distinguish the senses is by identifying them with the characteristic sensations they produce. This is the view, for example, of Grice (1966; see also Martin 1992). Grice suggests that there are four characteristics that might be used, individually or jointly, to distinguish one sense from the other: by the features of the world detected by a particular sense modality; by the "special introspectible character of the experiences" (p. 135) of that modality; by the nature of the stimuli to which the modality is sensitive; or by
Phenomenal Qualities and Intermodal Perception
133
the mechanistic (e.g. physiological) features of each modality. He claims that the introspectible character of experiences "cannot be eliminated" (p. 152) whatever the virtues of the other possibilities. If one accepts, therefore, that the sense modalities are distinct, it is plausible to suppose that what distinguishes them are the different phenomenal qualities they produce. Since phenomenal qualities are precisely what distinguishes a sense modality, it follows, trivially, that phenomenal qualities are modality-specific.
Phenomenal Qualities and Perceptual Epistemology Although Block characterises the phenomenism-representationism debate as one about the nature of consciousness, it can also be viewed as an instance of a traditional debate in the epistemology of perception. Philosophers of perception have historically been concerned, among other things, with the question of whether perception is "direct" or "indirect" (for a review, see Dancy 1985). Direct realists believe that the "immediate" objects of perception are the physical objects of the external world. These objects are immediate in an epistemic sense though not, of course, in a causal sense since there is a long and complex causal chain between the physical object and the perception of it by the perceiver. In contrast, the indirect realists the British empiricists most famously (see Leahey 2000; for a recent version of this view see Jackson 1977) believe that the direct objects of perception are mental items, states, or properties, and that physical objects are perceived mediately only in virtue of the immediate perception of the mental items. The perceptual intermediaries have had many names in the history of the subject including ideas, sensations, and, most infamously, sense-data. Historically, indirect realism has been closely associated with the commitment to the existence of phenomenal properties because phenomenal properties were taken to do the representational work required by a perceptual intermediary. In contrast, direct realism historically had no use for phenomenal qualities because it denied that perception involved intermediaries. Harman (1990), for example, defends an "intentional" theory of perception, and this is a special form of direct realism. According to this account, the (direct) object of one's perception is an intentional object which is usually a physical object in the environment. When you see a tomato in front of you, the intentional object of your perception is the tomato. What makes the intentional theory a special case of direct realism, however, is that it also allows for non-existent objects to be the objects of perception. When Macbeth hallucinates a dagger, the intentional object of his hallucination is a dagger. It is not a mental object or a set of mental properties that function as
134
Ian Gold
intermediaries. On the intentional view, the object of one's perception is never a mental intermediary but an intentional one even if some intentional objects, such as the dagger perceived by Macbeth, are non-existent. Given that the debate between phenomenists and representationists seems to be primarily about the existence of phenomenal intermediaries, one might wonder why the debate has focused on whether there exist nonrepresentational phenomenal qualities such as the feel of redness or orgasm. The answer is that if one can show that it is possible to be aware of a phenomenal quality that is not doing any representational work, then it will be plausible to conclude that what is doing the representational work in typical cases of perception are phenomenal qualities as well. In the absence of uncontroversial examples of non-representational phenomenal qualities, it will always be possible for the representationist to reiterate his argument that the phenomenist is confusing the properties of experience by means of which the features of the world are represented with the features of the world represented by those properties (Harman 1990). The phenomenist's strategy can thus be seen as an instance of the argument from illusion (see Dancy 1985). The argument from illusion says that in cases of illusion, or hallucination, there must be something that a perceiver is seeing. When Macbeth sees a dagger before him there must be something he is seeing. It can't be a dagger because there is no dagger before him. It must, therefore, be some mental item or items - - e.g. phenomenal qualities - - that he is perceiving. It is plausible therefore, to assume that when an agent actually sees a dagger, the agent is also experiencing a set of phenomenal qualities which in this case actually represent a physical dagger. If the phenomenist is right that there are non-representational mental qualities, then he can advance an analogy of the argument from illusion: given that there are non-representational phenomenal qualities, it is plausible that when a mental state has representational content it is the phenomenal qualities that are doing the representing. Certainly an application of Occam's razor should lead us to that view. Phenomenism is, therefore, a form of indirect realism in the empiricist tradition because it holds that perceivers experience phenomenal qualities or sensations and that these have representational content. As Block (1996) says, "it is often the phenomenal character itself that has the representational content" (p. 20). Representation of the world is thus indirect. It goes via experience of phenomenal qualities. (This does not, of course, make Block an advocate of a sense-datum theory.) Notice, however, that despite the historical association between indirect realism and a belief in the existence of phenomenal qualities, the two views can be separated. One could, for example, be a direct realist who believes that the state the perceiver is in when directly perceiving the external world has phenomenal properties. When Harman objects that positing the existence of phenomenal
Phenomenal Qualities and Intermodal Perception
135
properties represents a confusion between the properties that experience represents with the properties of experience in virtue of which it represents, he is primarily objecting to the idea of phenomenal properties as intermediaries rather than to the existence of phenomenal qualifies per se. One could continue to hold onto the existence of phenomenal properties as long as one denied that they had representational content. I briefly sketch a view of this kind in the last section.
Perceptual Learning The British empiricists thought that perception involved sensations that are specific to a single modality and that perceivers had to learn to associate sensations in one modality with sensations in another (see Leahey 2000). 4 One has to learn, for example, that a particular visual sensation of a round object picks out the same property that the tactile sensation produced by a round object does. This view is also a consequence of the two theses of phenomenism discussed above. According to Thesis 1, access to the representational content of one's perceptual state is restricted by the phenomenal properties of the state. Thesis 2 says that those properties are modality-specific. Suppose, for example, that on one occasion you look at a round object without touching it, and, on another occasion, you touch the same round object without looking at it. Your experience of the round object is different on each occasion because phenomenal qualities are modality-specific, and your experience occurs in two different perceptual modalities. Further, because your beliefs about the representational content of your perceptual states are restricted by the phenomenal qualities of those states you would, all things being equal, believe that on each occasion you were experiencing something different. It is only by having experience in which you learn that the visual sensation is associated with the tactile sensation that you come to understand that the two sensations are correlated with the same object.
4 It is not a straightforward matter to characterise empiricism because there are a number of closely related, but distinct, theses to which empiricism might be committed. For example, empiricism could be characterised as the view that the relations among perceptual modalities are contingent; or that relations must be acquired in experience; or that the relations must be learned. It is often said that empiricism is represented by the last of these views, but a careful historical analysis would be required to show this. These distinctions, however, do not affect the central argument of this paper. I am concerned to show only that if phenomenism is construed as committed to the two theses presented in first section, it is inconsistent with some of the findings of developmental psychology. I assume that these findings are also inconsistent with traditional empiricism, but should that turn out to be the case only for some versions of empiricism, it would not affect the consequence for phenomenism. (I am grateful to John Heil for helping me to clarify this issue.)
136
Ian Gold
Traditional empiricism is explicitly committed to the idea that at least some behaviours presuppose perceptual learning, and that learning is, as Hume thought, largely a matter of acquiring associations among sensations. Like traditional empiricism, phenomenism has as a consequence the view that some perceptual behaviours are impossible prior to perceptual experience. I will claim that the problem that phenomenism faces is that this consequence is false.
Intermodal Perception: Molyneux's Question Empiricism in perception, and the phenomenist version of it, says that without perceptual experience it should be impossible for a perceiver to recognise that the visual appearance of a round object and the tactile appearance of a round object have, as part of their representational content, the same object. Suppose, therefore, that a man blind from birth were to recover his sight. If he were shown a round object, would he be able to identify it as round on the basis of visual experience alone? In 1688, William Molyneux, an Irish politician and scientist, posed this question to Locke. The importance of the question in philosophy and psychology has given it a name - - Molyneux's question. Locke reported the question in a later edition of the Essay, and through this work, the question had a powerful influence on centuries of philosophical theorising about perception and on perceptual psychology (see Cassirer 1951; Degenaar 1997; Morgan 1977; see also Evans 1985). 5 Molyneux's question is a question about intermodal perception, a set of phenomena that arise as a result of the fact that some properties of objects can be perceived by more than one sensory modality (Stein & Meredith 1993). Sometimes intermodal perception involves the recognition of a property in one modality as the same as in the other modality; one can both see and feel that an object is smooth for example. It can also involve interactions between modalities. An illustration is the set of phenomena called the ventriloquism effect (Stein & Meredith 1993: 3-7). One familiar instance of the effect occurs when one is watching television. The sound of the voices of the people on the screen all emanate from the same loudspeaker but appear to the viewer to be coming from the location in space where the person's mouth is. The visual percept interacts with the auditory percept and leads to an illusion.
5 Molyneux writes to Locke (1976) as follows: "A Man, being born blind, and having a Globe and a Cube, nigh of the same bigness, committedinto his Hands, and being Told, which is Called the Globe, and which the Cube, so as easily to distinguish them by his Touch or Feeling; Then both being taken from Him, and Laid on a Table, let us suppose his Sight Restored to Him; Whether he Could, by his sight, and before he touch them, know which is the Globe and which the Cube?" (v. 3: 482-483).
Phenomenal Qualities and lntermodal Perception
137
Locke answered Molyneux's question in the negative as did the other empiricists (and Molyneux himself). Indeed, Cassirer (1951) has claimed that all of Berkeley's philosophy can be taken as an elaborate answer to Molyneux's question. In the New Theory of Vision and elsewhere Berkeley (1709/1975) famously argues that we cannot perceive distance (i.e. depth) because "distance being a line directed end-wise to the eye, it projects only one point in the fund of the eye, which point remains invariably the same, whether the distance be longer or shorter" (p. 9). How, then, does a perceiver come to see objects in depth? One of Berkeley's hypotheses is that the sensations of convergence the turning inward of the two eyes which varies with changes in object depth come to be associated with objects at different distances; the greater the convergence, the closer the object. Sensations of convergence come to be associated with sensations of touch in such a way that the convergence can be interpreted as a representation of depth. For example, a child comes to learn that greater convergence is typically associated with objects that can be touched nearer the body, and less convergence is associated with objects that can only be touched by extending one's hand or moving in a particular direction in space. The perception of depth is thus the result of visual sensations and learned associations between these sensations and sensations of touch. Philosophers and psychologists in the empiricist tradition were shaped by Berkeley's views and continued to adopt the same position with respect to the Molyneux question as Locke and Berkeley themselves (Kellman & Arterberry 1998; see also Evans 1985). For example, Piaget (1952), whose theories dominated developmental psychology until recently, would have given the same answer to Molyneux's question as the Locke and Berkeley did. He says that in perceptual development at first ... sensorial images.., only constitute spots which appear, move, and disappear without solidity or volume. They are, in short, neither objects, independent images, nor even images charged with extrinsic meaning... Still later.., the visual images acquire meanings connected with heating, grasping, touching, with all the sensorimotor and intellectual combinations (pp. 64-65). Molyneux's question was a theoretical one until early in the 18th century when cataract operations become common. These operations made it possible to restore the sight of individuals blind, or nearly blind, from birth or a young age (see Degenaar 1996; a recent case is described by Sacks 1995). The results of these operations, especially one conducted by the surgeon William Cheselden in 1728, were much discussed but inconclusive. We now know that the absence of visual
138
Ian Gold
stimulation during a so-called "critical period" of development leads to abnormal development of the visual system (Kujala et al. 2000). Cataract patients do not, therefore, have normal visual systems, and the visual abilities they have when the cataracts are removed is unlikely to be the same as that of normal perceivers. For this reason, although one can perhaps use the results of cataract surgery to answer Molyneux's question interpreted literally, the significance of the question for normal perception cannot be evaluated in this way. In the 20th century a number of empirical investigations other than cataract surgery were used to explore Molyneux's question. The two main ones are perceptual deprivation studies in animals (Kujala et al. 2000), and the use of sensory substitution systems to give a kind of sight to blind subjects either by the use of auditory or tactile stimuli (Bach-y-Rita et al. 1969; Kay 19746). Evans also (1985) envisaged the possibility of direct cortical stimulation as a way of giving blind subjects visual experiences (see Girvin 1988), but the considerations regarding abnormal brain development may be problematic here as well. Recent studies show that what is visual cortex in normal subjects is responsive to tactile stimulation in blind subjects (Hamilton & Pascual-Leone 1998), so it is an open question whether there is any properly visual cortex in the blind that can be stimulated. Because none of these investigations led to a conclusive verdict on the Molyneux question, the traditional empiricist view remained the orthodoxy in psychology until the late 1970s and early 1980s when it was discovered that newborns have perceptual abilities that developmental theory takes to be impossible. I turn to these now.
Infant Studies and Molyneux's Question A different line of research addresses Molyneux's question and does seem to provide unambiguous results. This is the study of perceptual development in human infants (Kellman & Arterberry 1998). Almost non-existent twenty years ago, the study of perceptual development has already revealed a good deal about the abilities of "newly sighted" subjects, and demonstrates that much of what has been assumed to be true of vision is controversial. Consider two well-known findings in the domain of intermodal perception and action (see Meltzoff 1993).
Visual-tactile recognition: Molyneux's task Molyneux'squestion can be asked of preverbal infants. Meltzoff & Borton (1979) gave one group of infants a normal
6 Thereis a morerecentsystemsimilarto that of Kay.See http://www.seeingwithsound.com/voice.htm.
Phenomenal Qualities and Intermodal Perception
139
Figure 1: The pacifiers used to ask the infant version of the M o l y n e u x question (from Meltzoff & Borton 1979). pacifier and another group a nubbed pacifier (Figure 1). The subjects did not see the pacifiers but could suck on them for 90 seconds. The pacifiers were then taken out of their mouths, again without their being seen, and the infants were then shown two objects, one the shape of the smooth pacifier and the other the shape of the nubbed one. The subjects stared at the object that corresponded to the pacifier they had felt with their tongues for longer than at the other object indicating that they recognised the similarity b e t w e e n the two. 7
Imitation Young children can imitate the facial expressions of adults, and it turns out that infants 12 to 21 days old can do so as well (Meltzoff & M o o r e 1977). Significantly, however, newborns are also capable of this (Meltzoff & M o o r e 1983). W h e n shown three facial expressions, lip protrusion, m o u t h opening, and tongue protrusion, neonates as y o u n g as 42 minutes old will spontaneously and successfully imitate them (see Figure 2). This is an intermodal ability in the
7 A common experimental paradigm used with infants is the habituationmethod. This method exposes the subject to a stimulus and assumes that the subject habituates to it, where "habituation" refers to the process in which repeated presentation of a stimulus leads to a diminished response in the subject relative to the first presentation. Novel stimuli hold the infant's attention for a longer time than habituated stimuli. A subsequently presented stimulus can be judged as similar or non-similar to the initial stimulus depending on whether it is treated as novel (non-similar), eliciting prolonged attention, or as habituated (similar), eliciting reduced attention. In the pacifier experiment, as in some other experiments, the results are reversed; the similar case elicits prolonged attention. It is not clear why familiarity sometimes produces the same response as novelty, but it does not matter for the present case. The differential response to the similar and non-similar conditions shows that infants detect amodal perceptual information, and that is the central point of interest (see Kellman & Arterberry 1998).
140
lan Gold
Figure 2: Imitation of facial gestures in newborns (from Meltzoff & Moore 1977).
sense that the subjects must be able to translate a perceived expression into a motor behaviour without any experience of visual feedback to confirm that they have succeeded. The information contained in the visual representation of the experimenter's face must, therefore, be represented in an amodal f a s h i o n - that is, in a format which is not specifically visual but also accessible to motor output. On the traditional developmental view which follows from the Lockean and Berkeleyan conception of sensation, these findings should have been impossible. The Piagetian (1952) view of development held that perceptual development is, in large part, a matter of the learning of associations among different sensory modalities. Nonetheless, Meltzoff and Moore's findings have been replicated. They have also been found to hold across cultures and subject groups and have been extended to other facial gestures such as head movements (Kellman & Arterberry 1998). Meltzoff & Moore's studies show that newborns have intermodal capabilities, but the phenomena described do not say what cognitive functions underlie them. One of the objections to the original study (Meltzoff & Moore 1977) was that the imitative behaviour was the outcome of a reflex orfixed-action pattern and nothing more (e.g. Bjorklund 1987). Were this the case, facial imitation would not support the claim that infants are capable of imitating in any cognitively substantive way because reflexive imitation does not require an amodal representation and the capacity for voluntary behaviour. In a subsequent study (Meltzoff & Moore 1983),
Phenomenal Qualities and Intermodal Perception
141
however, the subjects were prevented from spontaneously imitating the facial expressions of the experimenter by having a pacifier placed in their mouths. After the experimenter modeled the facial expression, he adopted a passive expression and the pacifier was removed from the baby's mouth. Nonetheless, subjects produced from memory the expression that had been modeled. Further, Meltzoff & Moore (1994) conducted a study with six-week old infants in which one of three gestures - - tongue protrusion at the side of the mouth, tongue protrusion in the middle of the mouth, or mouth opening - - was modeled for subjects. As before, infants imitated the gesture. On the second day of the experiment subjects were presented with a blank face, but they continued to produce the remembered gesture from the day before. Not only did they remember the gesture, but they produced it more frequently in the memory task than in the initial imitation paradigm. Even more remarkably, an analysis of the movements performed by the infants revealed that they seemed to make better approximations of the gesture over the course of the trials. These data supports a number of claims (see Kellman & Arterberry 1998). First, since reflexes are taken to be immediate responses to stimuli, this behaviour could not be reflexive. The subjects were not only able to remember the expressions for a few seconds following modeling by the experimenter, they were able to remember them from one day to the next. Furthermore, reflexes or fixed-action patterns tend to be stereotyped. The fact that infants worked toward an ever better motor behaviour means that they were learning and not merely acting reflexively. In general, these findings provide a picture of the perceptual, motor, and learning capacities of children that is altogether beyond anything that traditional psychology could have countenanced. Indeed, it appears, as Kellman & Arterberry (1998) put it, that this kind of imitation requires that "one must perceive the event as an e v e n t - not as patches of brightness and colour.., as a representation of objects, surfaces, and changes occurring" (p. 268). And, as Meltzoff & Moore (1994) say: "The capacity to organise actions based on a stored representation of perceptually absent events is a starting point in infancy, not its culmination" (p. 95).
Phenomenal Qualities and Perceptual Representation The infant data just reviewed looks to be a decisive refutation of the traditional empiricist view that perceptual learning is required prior to behaviours that presuppose intermodal perception. Empiricism has not of course disappeared, but there is a recognition in psychology that it must be modified. (An illustration of one such modification is Nakayama & Shimojo 1992.)
142
Ian Gold
I characterised phenomenism as involving two important theses: (1) that one's access to representational content supervenes on the phenomenal qualities of the relevant mental state; and (2) that phenomenal qualities are modality-specific. This makes phenomenal qualities in the phenomenist's sense a species of sensations in the traditional empiricist sense. For this reason, I argued that phenomenism has the same consequence as traditional empiricism, namely, that perceptual learning has to take place before intermodal perception is possible. Like empiricism, therefore, phenomenism is at odds with the facts of development. In contrast, a direct theory of perception is not in tension with these results. Direct realism takes the objects of perception to be the objects of the external world. On this view, when Mary perceives a sphere, the content of her perceptual state has, as a component, the spatial properties of the sphere. These properties, as Evans (1985) argues, are amodal. For this reason, Mary's knowledge of the external world is not restricted by the perceptual modalities through which the perceptual information is transmitted. As long as the different modalities are correctly detecting features of the same environment, intermodal transfer is not only possible but, perhaps, to be expected. Any perceptual theory that takes perception to be a transfer of information from the environment to the perceiver (e.g. Dretske 1981; Gibson 1966; see also Heil 1987) will be in a more promising position to explain the infant data than phenomenism and other empiricist views.
Options Where does phenomenism go wrong? I have characterised phenomenism as being committed to two theses, so, in the first instance, there are two possible strategies the phenomenist can attempt. The first is to reject the thesis that phenomenal qualities are modality-specific; the second is to reject the thesis that sensations have representational content. The first has the virtue that it would do less violence to the traditional empiricist picture of perception. Modality-specificity is probably quite widely held and could be given up without losing a good deal of what distinguishes phenomenism from other views about mental representation. The difficulty with this option is that it is unclear what sense could be made of the idea that phenomenal qualities are not modality-specific. As indicated above, one strong motivation for the thesis of modality-specificity is the assumption that there are perceptual modalities to begin with. It is conceivable, however, that some notion of "phenomenal quality" could survive the discovery that there are no distinct perceptual modalities. One recent suggestion (Stoffregen & Bardy 2001) is that the apparently diverse modalities
Phenomenal Qualities and Intermodal Perception
143
are better construed as interdependent parts of a single perceptual function. This suggestion harmonises with the developmental data, and should it turn out to be plausible, the phenomenism-representationism debate could be revisited against the background of new assumptions. The second option involves rejecting the view that phenomenal qualities have representational content. This would be a rejection of a fundamental aspect of traditional empiricism, but it might still provide an appealing alternative to the representationist position. Harman's intentional theory, for example, entirely rejects the existence of any sort of "subjective" quality largely, it seems, in order to avoid any suggestion that there are perceptual intermediaries. But this sometimes stretches intuitions. Harman (1996b) says, for example, that the experience of orgasm "involves a perception of the relevant bodily events and usually includes other things as well, including intense feelings of desire and enjoyment" (p. 76). The claim is that orgasm is exhausted by the representational content of the state m that something is occurring in one's body m and that the feelings of pleasure we associate with orgasm are in fact feelings of desire and enjoyment, both of which can presumably be analysed in non-phenomenal terms. Perhaps Harman is fight about this. What makes phenomenism intuitively plausible, however, is that it can more straightforwardly account for our beliefs that experiences of orgasm or phosphenes do, in fact, have phenomenal, nonrepresentational properties. If phenomenism could retain the view that there are phenomenal qualities this could be seen as a virtue even if those states do not themselves have representational content. Such a view might even be consistent with the intentional theory. Harman denies that our perception of the external world goes by means of the direct perception of intermediaries, but he need not deny that there are any subjective states at all. He does so, presumably, only because philosophers have typically taken subjective states to play a representational function. What might such a view look like? Let us say, following Evans (1985), that when one is looking at a particular scene, one is in an "informational state" and that that state represents the scene. Information is an abstract notion, and it is a commonplace that one can embody the same information in indefinitely many ways. I can represent the spread of my students' grades as a bar graph, a curve, a set of numbers, or by means of any number of other conventions. There is, therefore, a distinction to be made between the informational content of a mental state and its form. In the case of perception, we can distinguish the informational content of a perceptual state and the way in which that state is consciously experienced. Consider the case of colour perception. If colour realists are fight (see, e.g. Hilbert 1987), then colour is identical to surface spectral reflectance. A particular reflectance is usually represented mathematically as a curve of reflected light plotted against incident light. In perception, however, we do not represent the
144
lan Gold
colour of an object as a curve but as a bit of, say, experienced redness. On this suggestion, then, the redness experienced in perception is indeed a property of the external world; it is directly perceived. It is, however, perceived in a particular conventional format the format of red "as we see it" (see Mackie 1976) rather than as a curve of reflectance. What are mistakenly taken to be directly perceived properties of experience are no more than the form taken by the informational state of colour perception when that state is conscious. Part of what makes colors seem like subjective qualities is the fact that nothing like color as we see it exists in the external world. But this is not surprising any more than the fact that curves on cartesian coordinate graphs do not exist as natural properties in the external world. They are rather one way of representing those natural properties. If this were the case, one could make a place for subjective qualities without taking them to be the intermediaries between the perceiver and the environment. On this other conception, they would be aspects of conscious experience that provide the vehicle for the information contained in the perceiver's state. They come after the information about the external world is detected rather than before. Were there to be evidence for such a view, it would satisfy the demands of representationism while leaving room for the feel of experience. And, in the spirit of Block's conception of phenomenism, the nature of phenomenal qualities would indeed be a question about consciousness.
Acknowledgments This research was supported in part by a McDonnell Project in Philosophy and the Neurosciences senior fellowship and a grant from Monash University. I am grateful to John Heil, Daniel Stoljar, and Natalie Stoljar for comments on earlier drafts of this paper.
References Bach-y-Rita, R., Collins, C. C., Saunders, E A., White, B. W., & Scadden, L. (1969). Sensory substitution by tactile image projection. Nature, 221,963-964. Berkeley, G. (1709/1975). An essay towards a new theory of vision. In: Berkeley (Ed.), Philosophical works. Introduction and notes by M. R. Ayers. London: Everyman's Library. Berkeley, G. (1713/1975). Three dialogues between Hylas and Philonous. In: Berkeley (Ed.), Philosophical works. Introduction and notes by M. R. Ayers. London: Everyman's Library.
Phenomenal Qualities and lntermodal Perception
145
Bjorklund, D. E (1987). A note on neonatal imitation. Developmental Review, 7, 86-92. Block, N. (1990). Inverted earth. In: J. Tomberlin (Ed.), Philosophicalperspectives (Vol. 4). Ridgeview: Atascadero. Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences, 18, 227-247. Block, N. (1996). Mental paint and mental latex. In: E. Villanueva (Ed.), Perception. Ridgeview: Atascadero. Cassirer, E. (1951). The philosophy of the enlightenment. Translated by E Koelln & J. Pettegrove. Princeton: Princeton University Press. Clark, A. (2000). A case where access implies phenomenal qualities? Analysis, 60, 30-37. Dancy, J. (1985). Contemporary epistemology. Oxford: Blackwell. Degenaar, M. (1996). Molyneux's problem. Dordrecht: Kluwer Academic. Dretske, E I. (1981). Knowledge and the flow of information. Cambridge: MIT Press. Evans, G. (1985). Molyneux's question. In: A. Phillips (Ed.), Evans, collected papers. Oxford: Clarendon Press. Gibson, J. J. (1966). The senses considered as perceptual systems. Boston: Houghton Mifflin. Girvin, J. E (1988). Current status of artificial vision by electrocortical stimulation. Canadian Journal of Neurological Sciences, 15, 58-62. Grice, H. E (1966). Some remarks about the senses. In: R. J. Butler (Ed.), Analytical philosophy: First series. Oxford: Blackwell. Hamilton, R. H., & Pascual-Leone, A. (1998). Cortical plasticity associated with Braille learning. Trends in Cognitive Science, 2, 168-174. Harman, G. (1990). The intrinsic quality of experience. In: J. Tomberlin (Ed.), Philosophical perspectives (Vol. 4). Ridgeview: Atascadero. Harman, G. (1996a). Explaining objective colour in terms of subjective reactions. In: E. Villanueva (Ed.), Perception. Ridgeview: Atascadero. Harman, G. (1996b). Qualia and color concepts. In: E. Villanueva (Ed.), Perception. Ridgeview: Atascadero. Heil, J. (1987). The Molyneux question. Journal for the Theory of Social Behavior, 17, 227-241. Hilbert, D. (1987). Colour and colour perception. Stanford: Centre for the Study of Language and Information. Jackson, E (1977). Perception. Cambridge: Cambridge University Press. Kay, L. (1974). A sonar device to enhance spatial perception of the blind: Engineering design and evaluation. The Radio and Electronic Engineer, 44, 605-627. Kellman, E J., & Arterberry, M. E. (1998). The cradle of knowledge. Cambridge: MIT Press. Kujala, T., Alho, K., & Naatanen, R. (2000). Cross-modal reorganization of human cortical functions. Trends in Neuroscience, 23, 115-120. Leahey, T. H. (2000). A history of psychology (5th ed.). Upper Saddle River: Prentice-Hall. Locke, J. (1976). The correspondence of John Locke 1632-1704 (Vol. 3). E. S. De Beer (Ed.). Oxford: Clarendon Press. Mackie, J. L. (1976). Problems from Locke. Oxford: Clarendon Press.
146
Ian Gold
Martin, M. (1992). Sight and touch. In: T. Crane (Ed.), The contents of experiences. Cambridge: Cambridge University Press. Meltzoff, A. (1993). Molyneux's babies. In: N. Eilan, R. McCarthy, & B. Brewer (Eds), Spatial representation. Oxford: Blackwell. Meltzoff, A. N., & Borton, R. W. (1979). Intermodal matching by human neonates. Nature, 282, 403-404. Meltzoff, A. N., & Moore, M. K. (1977). Imitation of facial and manual gestures by human neonates. Science, 198, 75-78. Meltzoff, A. N., & Moore, M. K. (1983). Newborn infants imitate adult facial gestures. Child Development, 54, 702-709. Meltzoff, A. N., & Moore, M. K. (1994). Imitation, memory, and the representation of persons. Infant Behavior and Development, 17, 83-89. Morgan, M. (1977). Molyneux's question. Cambridge: Cambridge University Press. Nakayama, K., & Shimojo, S. (1992). Experiencing and perceiving visual surfaces. Science, 257, 1357-1363. Peacocke, C. (1983). Sense and content. Oxford: Oxford University Press. Peacocke, C. (1992). A study of concepts. Cambridge: MIT Press. Piaget, J. (1952). The construction of reality in the child. New York: International Universities Press. Place, U. T. (1956). Is consciousness a brain process? British Journal of Psychology, 47, 44-50. Sacks, O. (1995). To see and not to see. In: Sacks (Ed.), An anthropologist on Mars. New York: Alfred A. Knopf. Stein, B. E., & Meredith, M. A. (1993). The merging of the senses. Cambridge: MIT Press. Stoffregen, T. A., & Bardy, B. G. (2001). On specification and the senses. Behavioral and Brain Sciences.
Chapter 8
On Folk Psychology and Mental Representation Peter Godfrey-Smith
Introduction
In the 1980s the problem of giving a naturalistic theory of mental content beckoned young philosophers like myself; this looked like a philosophical problem that was both fundamental and solvable. The aim was to give a theory of the physical or biological properties of the internal states of organisms that suffice to make these states representations of the world beyond them. Internal states that have these special physical or biological properties have semantic content, and they do so as a matter of objective fact. The folk-psychological concepts of belief and desire were seen as picking out, in a rough and imperfect way, both the kinds of inner states and the kinds of semantic properties that would figure in the more detailed naturalistic theory. A theory of mental representation of this kind would not only become a centerpiece of cognitive science, but would be essential to epistemology and many other areas of philosophy as well. It would be important to all parts of philosophy that must use or assume a theory of thought. The guiding ideas for this project derived largely from the work of Jerry Fodor and Fred Dretske. Daniel Dennett and Stephen Stich looked on as skeptical but constructive critics (see Rey 1997; Sterelny 1990; Stich & Warfield 1994 for reviews). Roughly twenty years on, how has the project fared? With some sadness and much caution, I suggest that things have not gone well for the Dretske-Fodor program. I doubt that we will ever see a satisfactory version of the kind of theory that Representation in Mind: New Approaches to Mental Representation Edited by H. Clapin, P. Staines and P. Slezak Copyright 9 2004 by Elsevier Ltd. All rights of reproduction in any form reserved. ISBN: 0-08-044394-X
148
Peter Godfrey-Smith
Fodor's Psychosemantics (1987) and Dretske's Explaining Behavior (1988) tried to develop. Despite this, I do think we have learned a lot from the development of this literature. Some good partial answers may have been given to important questions m but not the exact questions that Dretske and Fodor were trying to answer. So I think it is time to start looking at different approaches to the network of questions surrounding belief and representation. This rethinking will involve looking again at some of the ideas of the nay-sayers of the 1980s, like Dennett and Stich, but looking further afield as well. In this paper I will begin to sketch one alternative way forward. "Begin" and "sketch" are the right words; this paper will not give anything like a complete theory. It will not even tackle the central problem of saying how the ascription of content actually works. Instead, I will cautiously outline some ideas that might be pieces of a future theory.
Two Sets of Facts We can start in familiar territory by recalling one of the fundamental disagreements that structured much of the discussion in the 1980s. Dretske, Fodor and many others think that some organisms, including people, contain inner states and structures which represent the world, and do so as a matter of objective fact. Both ordinary people and theoreticians of various kinds engage in the interpretation of these organisms; we attribute beliefs, desires, and other less commonsensical intentional states to them. When we do this m both we the folk and we the theoreticians m we are trying to describe real features of how agents are wired and how they are connected to the world. Interpretation is based largely on behavioral evidence, but it is an attempt to describe what is really going on inside. Other philosophers, including Dennett, think that this picture misconstrues the practice of interpretation. (Davidson's views overlap with Dennett's on some of these issues, but I will focus on Dennett here.) For Dennett (1978, 1981), to make an interpretation using folk psychological concepts is not to posit definite causally salient structures inside the head, which have special semantic connections to states of affairs in the environment. For an agent to have a particular belief is merely for the attribution of this belief to be compelling to an interpreter, where an interpreter has a characteristic viewpoint and a special set of goals. Interpretation is holistic, behaviorist and rationalizing. From the point of view of someone like Dretske or Fodor, this "interpretationbased" view seems to have the tail wagging the dog. But for Dennett the dog exists only as a projection made by the tail, and to say anything different is to buy into the old view of the mind as a kind of "ghost inside the machine."
On Folk Psychology and Mental Representation
149
The dispute has proven hard to resolve. But one thing we can do is step back and say this: however the details go, we will in the end have to account for two sets of facts: (1) Facts about the wiring and organization of behaviorally complex organisms, and the connections between their wiring and the world around them. (I will call these "wiring-and-connection facts.") (2) Facts about our actual practices of interpretation and ascription of content. (I will call these "interpretation facts.") Here I described these sets of facts neutrally with respect to the debates between people like Fodor and Dennett. Whether the interpretations made by people are descriptions ofthe wiring-and-connection facts or not, the world does contain these two sets of facts. Both are empirical phenomena, and in principle there could be complete empirical theories of each (see also Stich 1992). So imagine a future state of scientific knowledge in which we have highly detailed empirical theories of people. One thing this body of empirical knowledge will contain is a description of these two sets of facts. But as well as these two bodies of empirical knowledge, we will want a theory of how the two sets of facts are connected to each other. Here we find one of the roles for philosophy m to describe the coordination between the facts about interpretations and the facts about wirings-and-connections. The same sort of thing is true in other parts of philosophy as well. We can think of the goal of philosophical theories of causation, and theories of knowledge, in the same kind of way. Philosophy should aim to describe the connections between facts about the use of difficult and controversial concepts, and facts about the parts of the world that the concepts are in some sense aimed at dealing with. Philosophy should link the empirical facts about human causation-ascriptions with the empirical facts about how the world runs. Philosophy should link the empirical facts about human knowledge-ascriptions with the facts about how beliefs are regulated and how beliefs enable organisms to get around the world. In describing these connections, it is natural and appropriate for philosophy to speculate about how the empirical stories on either side will turn out. We can speculate but also be guided by the empirical information as it comes in. Although this is not the only role for philosophy in these areas, it is one central role. Let us return to the philosophy of mind. Imagine that we have complete descriptions of both the wiring-and-connection facts and the interpretation facts. How will familiar theories of mental representation in the literature look in relation to those descriptions?
150
Peter Godfrey-Smith
The Dretske-Fodor program will look like an assertion of one way the two sets of facts might be tied together. For the Dretske-Fodor program, folk psychological interpretations are controlled by data which contain information about the wiring-and-connection facts. That is, the wiring-and-connection facts generate behavioral data which in turn generate interpretations. These interpretations function as attempts to accurately describe the wiring-and-connection facts. Some interpretations are false, of course, like any claims made about hidden things. We can be data-driven and also be wrong. But for people like Dretske and Fodor, interpretations are often fairly accurate, and this accuracy is a matter of accuracy about wiring-and-connection facts. You might object at this point: how could interpretations given by people in earlier ages with no knowledge of the kinds of "wiring" we have in our heads possibly be trying to describe "wiring-and-connection facts?" This objection focuses too hard on my term "wiring." Any kind of inner structure that figures systematically in the causation of behavior will count for present purposes. The Churchlands' program shows us another way in which the two sets of facts might be related. Folk psychological interpretations are intended to accurately describe wiring-and-connection facts, but these interpretations embody a false theory and fail in systematic ways (P. M. Churchland 1981; P. S. Churchland 1986). Dennett's view shows us yet another option. On this alternative view, it is an error to think that the role of interpretations is to try to describe wiring-andconnection facts. Belief ascriptions are not attempts to pinpoint discrete, causally active, internal states with special semantic properties. Belief-attributions are not like gene-attributions. Instead, we should give a theory of the social role of interpretations that does not treat them as representationally aimed at the wiring-and-connection facts. We should think of belief ascriptions and other interpretations as part of a practice that has to do with various kinds of social coordination. Folk psychology is a "craft" (Dennett 1991). Inner wirings and physical connections between internal and external do exist, but it is a mistake to think that folk psychological interpretations function as attempts to make specific claims about wirings and connections. From the point of view of Dennett's picture, the Dretske-Fodor view is mistaken because it does not take seriously the special properties of human interpretive practices; interpretation is not just an attempt to lay out the hidden structure of a complex machine. From the point of view of the Dretske-Fodor picture, Dennett's view seems to deny that understanding the mind must ultimately be understanding what is going on inside the skin. That is something that ordinary people know, and something psychology knows whenever it is not diverted by bad theory. In the next section I will try to make some progress on these oppositions. Before leaving the methodological discussion I should note some idealizations
On Folk Psychology and Mental Representation
151
I have made. I distinguished two sets of facts, wiring-and-connection facts and interpretation facts, and imagined our gaining good theories of both of these. The relation between the structure of the mind and the structure of our interpretive practices might well be co-evolutionary (see Godfrey-Smith 2002). But beyond that, the relation between both of these sets of facts and our theories of these facts can be co-evolutionary. Highly theoretical ideas about the mind can filter down into everyday practices of interpretation. A recent example is found in Freud's ideas about unconscious desires, which have filtered quite far into everyday interpretation. Older examples might be furnished by religious ideas about the soul, and literary ideas about romantic love. Mental states, interpretations of mental states, and theories of both of these are all interlocked. The extent and importance of the interlocking depends on unresolved questions about modularity, plasticity and cultural change.
On Interpretation After all the years of debate, how can we start to resolve the disagreements outlined in the previous section? The first point that should be clear is that we need more empirical knowledge about folk psychological interpretation. Fortunately, new empirical input is coming in. In particular, there is a growing body of work within psychology on the development of folk psychology in children (Davies & Stone 1995; Stone & Davies 1996). And work within cognitive science on the "wiring and connection" side continues as usual. We can expect the oppositions outlined above to be transformed as this empirical work develops. In the meantime though, I would like to outline some simple ways in which we might find some of the disputes resolved. Consider one very basic feature of the disagreements described earlier: is it the case that folk psychological interpretations make commitments or hypotheses about the internal structure of the person being interpreted? Surely it is remarkable and of some sociological interest that there is still so much disagreement about this simple question. Dennett has argued for many years that folk psychology does not commit its users to anything about the internal causal structure of the people being interpreted. Dennett accepts a kind of "minimal logical behaviorism." Nothing about the insides of a person could affect the truth or falsity of an interpretation of that person in folk psychological terms. So says Dennett. But to others this claim seems nothing less than outrageous. People who differ massively on other matters m Fodor, Stich, Armstrong, and the Churchlands m at least agree that folk psychology is committed to some claims
152
Peter Godfrey-Smith
about what is going on inside people's heads. Surely beliefs and the like are, at least sometimes, internal mental causes of what people do. Once we get past this fundamental opposition, the disagreements among those with a more "realist" interpretation of folk psychology are themselves substantial. Some, like Fodor and Stich, think that folk psychology involves rather strong claims about the insides of our heads, strong enough for eliminativism to at least be a relevant possibility. Others, like Jackson & Pettit (1990), think that folk psychology makes claims so weak that eliminativism is an almost impossible option. Let us focus on the fundamental dispute about whether folk psychology does or does not involve commitments about internal structure. How should we deal with the alarming inability to agree on so basic an issue? Is one side so steeped in ideology that it cannot accept the blindingly obvious? I suggest that some aspects of this position might be resolved with a simple assertion of false dichotomy. Folk psychology need not have a single role, with respect to the question of whether interpretations are directed at describing inner structure. What we might need is some kind of pluralism on the issue. Folk psychological interpretation is a tool that lends itself to several different intellectual and practical tasks. With some kind of pluralist view in hand, we can explain many of the strange battles of the past by noting that different philosophers have focused on different aspects of folk psychological practice. Compare the use of folk psychology in a law court, when the aim is to work out exactly what crime a person committed, and the use of folk psychology in freeway traffic, when the aim is to avoid collision and coordinate one's actions with surrounding drivers. These two uses of folk psychology look different and perhaps they are different. When one focuses on the law court case, it seems transparently clear that the apparatus of interpretation is being used to explore hypotheses about the inner causes of behavior. When one focuses on the freeway traffic case, the postulation of inner structure seems beside the point. All that is relevant in that case is the fast, accurate prediction of behavior. Philosophical theories devised under the influence of each of these two paradigm cases will look very different. So we may need some kind of pluralist option about folk psychology m about the intended relation between wiring-and-connection facts and interpretation facts. But there is a variety of ways in which this pluralist view might be developed. Here I will sketch a couple of options, without taking a firm stand on any side. One thing we have to work out is whether folk psychology contains multiple practices, or a single practice with something like multiple construals. Deciding this issue requires that we work out where folk psychology itself stops, and application of and commentary on it begins. Dennett (1991) has argued that we should distinguish the "craft" from the "ideology" of folk psychology. The "craft"
On Folk Psychology and Mental Representation
153
is the unreflective use of the apparatus; the "ideology" is a set of ideas about the craft of folk psychology. The distinction I am making here is similar, but not the same. The "ideology" of folk psychology, for Dennett, is something that has no role (or very little role) in the ordinary social use of the apparatus. It is purely a piece of theory used to comment on the craft. The distinction I am making here is one between basic features of the craft, and more elaborate comments on and applications of the craft that also have a role within folk interpretation itself. So if we want to develop a pluralist option for folk psychology, one question we have to answer is whether there is a diversity of crafts, or a diversity of construals and applications, or both. I will say something about each possibility. One simple way to resolve the old debates would be to claim that folk psychology is bifurcated all the way down to the most basic features of the practice. It might be argued that people switch between two different modes of interpretation, usually without realizing it. On a freeway, people apply a form of interpretation which is either behaviorist or very close to it. But when working out in a legal setting whether someone is guilty of murder or manslaughter, a different kind of interpretation is used. Then everything hinges on choosing between specific rival hypotheses about inner mental causes. I pause to contrast this strategy with another recent proposal aimed at resolving some of these issues. Jackson & Pettit (1990) argue that folk psychology (as a whole) does make some commitments about internal structure, but these commitments are extremely minimal. So their reply to Dennett is that there is no "instrumentalist" option for folk psychology because the fight "realist" construal is so uncontentious. For Jackson and Pettit the same minimal claims are made when one uses folk psychology in the law court and on the highway. I suggest that their minimal realist construal is unlikely to make much sense of the most causally specific kinds of folk-psychological interpretation m the kind found in courts of law, soap operas, and so on. But more importantly, once we recognize the possibility of a pluralist view about folk psychology, there is no need to search for a single account that simultaneously makes sense of the striking obviousness of interpretations made in some contexts (on the highway), and for the apparent logical strength and specificity of interpretations made in other contexts (law courts). Different practices are at work in each case. How likely is this first possibility to be right? One obvious problem is the fact that it seems strange for there to be two separate interpretive practices with so much in common. The same basic concepts are used in both kinds of interpretation (otherwise it would be much more obvious that there are two practices). It would be interesting to investigate whether some kinds of propositional attitude ascription only appear in one context or the other m whether some folk psychological
154
Peter Godfrey-Smith
concepts are only useful in the service of interpretations aimed at uncovering mental causes. That would be evidence for this first pluralist option. More pressingly though, the psychologist Alison Gopnik tells me that there is no evidence in developmental psychology at all that seems to support this first option especially, no evidence supporting the idea that children learn different practices of interpretation including a near-behaviorist practice for use when prediction is all that matters. So I will now sketch a different, somewhat more complicated, version of the pluralist view. Suppose we think of folk psychology as a model of the mind, rather than (as it is often put) a theory. What is the difference? The difference that I have in mind has to do with the fact that a model, in science, can be interpreted as having many different kinds of relationship to the phenomena it is directed upon. In particular, there is a continuum of possible attitudes with respect to which features of a model are interpreted in a realistic w a y as having real and distinct counterparts in the underlying structure of the world. Two scientists can use the same model while disagreeing about which features of a model are taken to have real counterparts, and a single scientist can use the same model while changing his or her mind about the right interpretation. A model is a conceptual structure which lends itself to a variety of different scientific uses. Sometimes predictive adequacy is central, but sometimes more than this is hoped for or required. Sometimes a model might be favored despite unimpressive predictive power, if it seems to get something about the basic causal structure right. Folk psychology might be considered in something like the same way. The basic folk psychological model of the mind itself is normally acquired at a very young age, as the empirical literature has shown. What a child acquires by the time he or she is five years old might be facility with a model. This can be acquired without a grasp of some distinctions between different construals and applications of the model. Children might pick up different ways of using and construing the model as they grow older m they pick up different ways of highlighting some aspects of the model and downplaying others. A normal adult does not care about the inner structure of the driver alongside on the freeway; only behavioral patterns matter. The folk psychological model is then used as an input-output device. But in other social contexts the model will be used to try to get a grip on the exact nature of mental states which act as intervening variables in the causation of behavior. Behavior now functions only as evidence for hypotheses about inner causes. The aim is to elucidate the fine structure of mental processes ("did she realize the likely effect of those words?", "did he really fear for his life at that point?"). In some social contexts, the details of the underlying mental processes, as far as we can
On Folk Psychology and Mental Representation
155
discern them, do make a difference to how we treat a person or problem, while in other contexts they do not. So on this view there is a unitary, low-level "craft" of folk psychological interpretation. This consists in the ability to apply a model. But a model itself does not determine its proper interpretation; it can be construed either as a set of hypotheses about hidden causes, or as a purely predictive device. It can also function as a mixture of these m some core features of the model might be taken to have real counterparts, while most details of the model are not interpreted this way. My suggestion is that the ability to use the model in a variety of different ways is itself part of the set of skills that a folk-psychological interpreter comes eventually to acquire. The distinction between different construals or applications of the folk-psychological model is not something peculiar to philosophical discussion, but is part of the tool-kit ordinary people use to negotiate different contexts in social life. I am unsure how best to further develop these ideas, and how best to connect them with empirical work on the structure, evolution and ontogeny of the folk psychological model of the mind. If the second pluralist proposal is correct, we should be able to empirically distinguish two different aspects of the acquisition of folk psychology m the acquisition of facility with the model, and the acquisition of alternative construals and applications. In correspondence, Alison Gopnik suggested to me that the ease with which young children handle the concept of pretending might tell against this view. Children do have a grasp of the distinction between literal and non-literal applications of all kinds of frameworks. And their take on mental states is that these states are real. If this is fight, then the acquisition of folk psychology takes place in a way that is accompanied by a simple realist construal of the model. Any pluralism must then result from later revision. Clearly these are subtle empirical issues, and my discussion here has the nature of speculation and sketching. I do think the possibility of a pluralist view of the commitments of folk psychology is important though. It suggests that there is no need to "thread the needle"; no need to give a single account of the commitments made by folk psychology that does simultaneous justice to the role of folk psychology in freeway traffic, soap operas, law courts, historical reconstruction of the causes of World War I, and devising marketing plans for new products. This needle-threading project may well have been, to use an Australian expression, a mug's game. It is possible that the pluralist option might help with some other questions that philosophers battle endlessly about in this area. An example is the problem of animal belief.
156
Peter Godfrey-Smith
On the Representational Concepts Used Within Cognitive Science Suppose the ideas sketched in the previous sections are right. What would this tell us about the role of folk psychological concepts in cognitive science, and the issue of eliminativism? In this section I will discuss that issue, plus another linked set of issues concerning the naturalistic relations between thought and the world that have figured in philosophical attempts to give a reductive analysis of semantic properties. If folk psychology gives us a model that can be applied to mental processes, then folk psychology can be associated with either a strong or a weak set of hypotheses about the mind. We can look for a detailed mapping between the structure of the model and the structure of the mind, or a minimal mapping. Folk psychology itself does not resolve the issue, does not tell us which are the issues that are crucial for assessing whether beliefs (for example) are real or not. If the folk psychological model was both predictively unimpressive and turned out to be structurally very unlike real mental processing, then the answer would be clear. Eliminativism would be vindicated. If folk psychology was predictively impeccable and the mind turned out to work the way Jerry Fodor has claimed it d o e s - with a set of core sentence-like structures representing the world and being manipulated computationally during thought then eliminativism would be clearly false. But the more likely outcomes lie in the middle. Folk psychology is clearly very predictively powerful in many of its domains of normal use. As a picture of mental processing, we do not know what will come of folk psychology yet. But let us look at cases where the two extreme options are false, and some pieces of folk psychology remain useful. In science, a model can be used or rejected as a whole, but it can also be mined, piecemeal, for elements that can be used in new models. This mining is what we often find in the case of folk psychology. When a concept like belief appears in a discussion in cognitive science, what has often happened is that the cognitive scientist is taking the folk psychological picture of mental processing and "stripping it down" to yield a few core structural elements that might have scientific value. For example, a core feature of the folk psychological idea of a belief is the contrast between beliefs on the one hand and wants and desires on the other. A theory about the mind that is in many ways at variance with folk psychology might retain a basic contrast between "how things are" registrations of the world and "what I want" registrations of the world. It might also retain the idea of rational or well-adapted cognitive processes as involving a systematic interaction between the two. Imagine some theory of this k i n d - a theory recognizing the core structural contrast between belief and desire, but embedding this contrast within a detailed
On Folk Psychology and Mental Representation
157
picture that departs in many ways from folk psychology. (This might be a radically connectionist theory, a theory borrowing from dynamical systems theory in the style of Tim Van Gelder, or a model of distributed cognition in the style of Rodney Brooks.) And suppose the theory is successful and developed in detail. Some cognitive scientists will want to retain folk-psychological terms like "belief" for the states that are posited and described in this process; others may want to avoid these terms. To retain the term "belief" is to stress the continuities between the scientific model and the folk psychological model, with respect to the basic interaction between "how things are" and "what I want" states. To drop the term is to stress the discontinuities between the scientific and the folk psychological picture. But there is no fact of the matter about whether the psychological states that appear in such a psychology "really are beliefs" m whether they are the same states posited by folk psychology but more accurately described (as envisaged in Lycan 1988). There is no fact of the matter because folk psychology itself does not commit to a sufficiently definite specification of what beliefs are supposed to be like. However, it will be appropriate to conclude in such a case that there is some non-trivial coordination between the folk-psychological model and the scientific theory. Here is an example to illustrate these general points. Ramsey et al. (1991) claim that propositional attitudes like beliefs are conceived by folk psychology as "functionally discrete, semantically evaluable states" that play a causal role in the production of behavior and other propositional attitudes. The key idea here is "functional discreteness"-- beliefs are seen as individually revisable and deployable. They can be added and lost individually, in virtue of reasonably localized alterations to the cognitive system. Ramsey et al. claim that connectionist models of a certain kind do not treat inner representations as having these features. So if connectionist models of this kind are accurate models of cognition, they support eliminativism about belief. This is a good example of people taking the folk psychological model of the mind, and insisting on a very realistic construal of the structure of the model. The structural match between the folk psychological model and the real nature of cognition has to be very good, or folk psychology has been undermined. But folk psychology itself does not contain any commitment to these "discreteness" properties. When mental processing is modeled by folk psychology, beliefs appear in the model in a way that suggests this kind of discreteness, but only if the model is interpreted in such a way that nearly everything in it is supposed to have a real-world counterpart. And this is not the only, or the most natural, interpretation of the model; there is no one right way to construe the model. So it is a mistake to see folk psychology as definitively committed to the falsity of the class of models Ramsey et al. discuss. But it is true that if these models were accurate, that would significantly reduce the number
158
Peter Godfrey-Smith
of elements of the folk psychological model that can be taken to have real-world counterparts. So far I have mostly discussed folk psychology's handling of inner structure. There is another side to these issues as well, and this has to do with connections between internal structures and the world. To finish the paper I will make some suggestions about this second topic that follow up some of the same themes as the earlier discussion. During the 1980s we saw many attempts to show a coordination between some set of naturalistic relations between internal and external states, and the one hand, and folk psychological semantic relations, on the other. The aim was to use this coordination to reduce semantic properties to more basic, physical properties. I doubt that such a theory can ever work (again, I say this with caution). I will not hazard a guess about which of the existing alternatives to the reductive approach is most close to being fight. Instead I will discuss a related topic. Although the reductive project failed, the philosophers pursuing it did succeed in describing some interesting kinds of natural connection between thought and the world. What should we make of these naturalistic connections? And what should we make of the use of representational concepts in cognitive science that have the same naturalistic orientation? To approach the issue, we can begin by looking again at a passage from a key early discussion of these issues. Here is the passage Dretske used to begin his book
Knowledge and the Flow of lnformation (1981): In the beginning there was information. The word came later. The transition was achieved by the development of organisms with the capacity for selectively exploiting this information in order to survive and perpetuate their kind (p. vii). Properly understood, this is a good summary of some fundamental features of the physical involvement that organisms have with their environments. Information, in the sense analyzed by Dretske, is a resource that can be utilized by organisms, like the carbohydrate molecules in fruits and seeds. The evolution of sensory and cognitive mechanisms is the evolution of ways of making use of this resource. But the relationship between information in Dretske's naturalistic sense and the semantic properties posited in folk psychological interpretation is more complicated than Dretske and others envisaged. Informational properties in the Dretske sense cannot be used to give a "reduction" of folk-theoretic semantic involvement between internal states and external conditions. What Dretske has done instead is isolate a real and important kind of naturalistic relation between organisms and environments, a relation which has a partial
On Folk Psychology and Mental Representation
159
similarity to the kinds of semantic relations that figure in folk psychological interpretation. In that initial statement I focused on Dretske-style informational links, which are in some ways fundamental to a theory of organism/environment relations. But similar points apply to a variety of other naturalistic relations described in the recent literature. In particular, something similar is true of the properties involving biological functions and mapping relations that Millikan has described with such care (1984). These are real relations between internal and external, which probably have a significant role in cognitive science. But they can have this role without being the material for a reductive theory of folk-psychological semantic properties. (I should add that although I would make these claims about both Dretske's informational relations and Millikan's teleo-functional relations, I do not think this is true of Fodor's relations of asymmetric dependence (1987), as these are not fully naturalistic relationships at all.) I suggest that when a cognitive scientist works on mental representation, what we often find is a special kind of meeting between two conceptual frameworks and mindsets. The people doing cognitive science are people, who bring with them ordinary habits of folk psychological interpretation. But they are also scientists, and science brings with it special criteria for what to look for and describe. For example, any connections between brain states and the world which figure in a scientific theory should be describable in physicalistically acceptable terms. What results is a special kind of interpretive practice, born of the meeting of folk interpretive habits and the special features of science. We should think of the "representational" concepts used in cognitive science as amalgams, or hybrids, born of the interaction between the ordinary interpretive habits that cognitive scientists have just in virtue of being people, and the scientific aims of describing precise, naturalistic and empirically studiable relations between organisms and environments. The representational concepts used in cognitive science are products of marriages between folk semantic concepts and a family of naturalistic concepts of physical specificity - - concepts of connection and directedness that are based on causal, nomic and functional concepts. The folk interpretive practices and the scientific concepts of specificity meet.., and what results is a hybrid description tailored to the demands of some part of cognitive science. What results is a description of "what the frog's eye tells the frog's brain," or a description of how the visual system "infers" shape from shading. We see a similar kind of hybrid in some everyday descriptions of computational devices - - when we say "the email program found that it was not connected to a server, so it told the modem to make a connection." (I hope this reminds people of the now-neglected program of "homuncular functionalism" defended by Dennett in one form (1978) and by Lycan in another (1981). The
160
Peter Godfrey-Smith
intermediate levels of description envisaged by homuncular functionalists tend to borrow both from folk-psychological interpretive practices on the one hand and from physical forms of description on the other. They are hybrids too.) Cognitive scientists will sometimes express their ideas as if they are using their scientific concepts to give a reduction or other unitary explanation of folk psychological semantic relations. I think this is a harmless mistake. In support of my claim that this is a mistake, I point to the fact that different cognitive scientists tend to have in mind radically different views about which naturalistic concepts of specificity are most fundamentally connected to representation and meaning. Some think the most important naturalistic relation here is covariation, of a special kind; others think it's resemblance, of a special and abstract kind. Others might think it's a concept of teleo-functional specificity, deriving from the biological concept of function. No one is right. These relations should not be conceived as rival attempts at the reduction of a folk concept of representation or meaning. Cognitive scientists forge different kinds of hybrid semantic concepts in different circumstances in response to different theoretical needs, and different ways in which scientific concepts of specificity and folk habits of interpretation interact with each other. It might turn out that some one of these relations is more scientifically fundamental than the others - - more fundamental to the project of explaining how intelligent systems work. But that, again, does not make this scientific relation into something that yields a reductive explanation of folk-psychological semantic properties.
Conclusion As I said at the outset, this paper is a collection of pieces that might be part of an alternative picture of folk psychology and mental representation. One theme seen in the two main discussions (third and fourth sections) is a kind of pluralism. Old disputes about the role and status of folk psychology might be resolved with the aid of a pluralist view of how folk psychological interpretation operates. There are several different ways in which this idea might be developed. And we might think of much of the work of Dretske, Millikan and others as describing a range of naturalistic relations between internal and external states, relations which may well have importance within cognitive science even though they cannot be used to give a reductive analysis of the content of folk psychological representational states. I do not want to give the impression that "pluralist" alternatives are always the way to resolve problems, however. Pluralism is not always an advance (especially when there is only one thing, as Mark Twain might have said). But one way or another, we do need to explore some new approaches to folk psychology and mental representation.
On Folk Psychology and Mental Representation
161
Acknowledgments Thanks to Hugh Clapin for organizing the Sydney conference, and to all those who participated in discussion of these topics. This paper is partly an attempt to reassess the force of some of Daniel Dennett's and Stephen Stich's arguments (especially those in Dennett 1981). This paper has been also influenced by discussions with many people over the years, but I should give special acknowledgment to the role that a coffee with Huw Price at Badde Manors and a beer with Kim Sterelny, both about nine years ago, had in getting me to think along different lines about these problems.
References Churchland, E M. (1981). Eliminative materialism and the propositional attitudes. Journal of Philosophy, 78, 67-90. Churchland, E S. (1986). Neurophilosophy: Toward a unified science of the mind, rain. Cambridge, MA: MIT Press. Davies, M., & Stone, T. (Eds) (1995). Folk psychology: The theory ofmind debate. Oxford: Blackwell. Dennett, D. C. (1978). Brainstorms. Philosophical essays on mind and psychology. Cambridge, MA: MIT Press. Dennett, D. (1981). Three kinds of intentional psychology. Reprinted in" D. C. Dennett (Ed.), The intentional stance. Cambridge, MA: MIT Press. Dennett, D. C. (1991). Two contrasts: Folk craft vs. folk science, and belief vs. opinion. In: J. Greenwood (Ed.), The future of folk psychology: Intentionality and cognitive science. Cambridge: Cambridge University Press. Dretske, E (1981). Knowledge and the flow of information. Cambridge, MA: MIT Press. Dretske, E (1988). Explaining behavior. Cambridge, MA: MIT Press. Fodor, J. A. (1987). Psychosemantics. Cambridge, MA" MIT Press. Godfrey-Smith, E (2002). On the evolution of representational and interpretive capacities. Monist, 85(1), 50-69. Jackson, E, & Pettit, E (1990). In defence of folk psychology. Philosophical Studies, 59, 31-54. Lycan, W. G. (1981). Form, function and feel. Journal of Philosophy, 78, 24-49. Lycan, W. G. (1988). Judgment and justification. Cambridge: Cambridge University Press. Millikan, R. G. (1984). Language, thought, and other biological categories. Cambridge, MA: MIT Press. Ramsey, W., Stich, S., & Garron, J. (1991). Connectionism, eliminativism and the future of folk psychology. In: J. Greenwood (Ed.), The future offolk psychology: Intentionality and cognitive science. Cambridge: Cambridge University Press.
162
Peter Godfrey-Smith
Rey, G. (1997). Contemporary philosophy of mind: A contentiously classical approach. Oxford: Blackwell. Sterelny, K. (1990). The representational theory of mind: An introduction. Oxford: Blackwell. Stich, S. P. (1992). What is a theory of mental representation? Mind, 101, reprinted in: Stich & Warfield (1994). Stich, S. P., & Warfield, T. A. (Eds) (1994). Mental representation: A reader. Oxford: Blackwell. Stone, T., & Davies, M. (Eds) (1996). Mental simulation: Evaluations and applications. Oxford: Blackwell.
Chapter 9
The World Gone Wrong? Images, Illusions, Mistakes and Misrepresentations Peter Slezak
Introduction" Parallels In seeking to understand the extraordinary persistence and recalcitrance of the problems of intentionality, it is instructive to focus attention on one particular facet of the issue. The question of misrepresentation has been discussed recently as a puzzling aspect of the overall problem of the semantics of mental representation (Dretske 1994; Fodor 1984, 1994) and I propose to explore this issue as a loose thread which may be pulled to unravel the rest of the tangled ball. It has evidently not been remarked that the problem of misrepresentation emerging from m o d e m theories of cognition is itself only the recent guise of an old problem. Indeed, it is a problem familiar to philosophers by other names in other contexts. For example, misrepresentation is, after all, closely akin to illusion in a certain sense of this term and, of course, the so-called "Argument from Illusion" is among the most well known of philosophical issues. 1 The Argument I A degree of confusion has been systematically introduced in these discussions by the failure to distinguish crucially different kinds of "illusion." An illusion in the sense relevant to the argument concerning ideas, sense-data or representations (Locke, Malebranche, Ayer) is, strictly speaking, hallucination or imagery. However, phenomena commonly referred to as "illusions" in this context such as mirages or sticks appearing bent in water are not illusory at all in an important sense. These are veridical perceptions of the light patterns entering the eye unlike cognitive errors such as the Miiller-Lyer illusion. Richard Gregory (1997), for example, has explicitly assimilated these phenomena, but no theory of cognitive processes could explain the "illusion" in the case of mirages and seemingly bent sticks due to refracted light. Representation in Mind: New Approaches to Mental Representation Edited by H. Clapin, P. Staines and P. Slezak 9 2004 Published by Elsevier Ltd. ISBN: 0-08-044394-X
164
Peter Slezak
can be seen employed by Malebranche (1712) and Locke (1690) in favour of a "representative" account of perception which posits "ideas" as an intermediary between the mind and the world. More recently, A. J. Ayer (1940) deployed exactly the same Argument for the reality of sense-data as mental entities that are the direct objects of perception. My suggestion is that the affinity between these problems, inter alia, is not merely a superficial one. Illusions in the relevant sense are actually hallucinations and, thereby, misrepresentations of exactly the sort relevant to the contemporary puzzle for symbolic, computational accounts of cognition. The parallel is hardly surprising when it is recognized that in all cases the fundamental theory in question involves mental representations which intervene between the mind and the world. This is a tri-partite account seen explicitly in Malebranche and Locke and recently endorsed by Bechtel (1998) as fundamental to modern conceptions as well (see Slezak 2002a). Despite its problems, the ubiquity of this view is undoubtedly a consequence of the fact that it seems difficult to imagine any sensible alternative. Nevertheless, the three seemingly indispensable elements of any theory of representation give rise to notorious and seemingly intractable problems. Fodor (1998, 2002) has recently suggested that his own Representational Theory of the Mind (RTM) may be understood on the model of the classical Empiricist conception: "To a first approximation . . . . the idea that there are mental representations is the idea that there are Ideas minus the idea that Ideas are images" (1998: 8). In this light, it is hardly surprising that modern problems might be simply the reinvention of old problems in a new guise. If the correspondence between representation and world fails for some reason we have misrepresentation or illusion. In the modern case, as posed by Fodor (1994) and Dretske (1994), the problem is, given causation between these elements, how to explain the possibility of illusion; in the classical case the problem is, given illusion, how to explain causation? The modern problem of misrepresentation arises because causal or correlational theories don't appear to permit a distinction between true and false representations. If a dog causes a representation of "cat" in mentalese, on the causal account it must ipso facto count as meaning "dog" and is, therefore, not a mistaken representation of cat. Conversely, the classical Argument from Illusion, starts from the other end, as it were. Beginning with the distinction between true and false representations, the Argument recognizes that these cannot both be correlated with an external reality, and concludes that in both veridical and non-veridical cases there must be some other object of direct perception, the "idea" or sense-datum (see S. Reynolds 2000). In view of these analogies, therefore, I suggest it is no coincidence that Fodor's (1980) "methodological solipsism" is strongly evocative of a Berkeleyan idealism. We see this parallel unwittingly brought into relief by Jackendoff (1992) who comments with a mild irony on the unpalatable implications of the modern symbolic, computational view:
The Worm Gone Wrong
165
The position emerging from this approach seems paradoxical. Why, if our understanding has no direct access to the real world, aren't we always bumping into things? (Jackendoff 1992: 161) This comment is reminiscent of Johnson's famous retort to Berkeley's "ingenious sophistry" by kicking a stone and saying "I refute it thus." In both cases, appealing to bumping into things, the responses bring into relief the way in which classical and modern theories seem to entail a disconnection of the mind and the world. As historian of philosophy Nadler notes, Malebranche, too, ... is often portrayed by his critics "as enclosing the mind in a 'palace of ideas', forever cut off from any kind of cognitive or perceptual contact with the material world" (Nadler 1992: 7). Despite its theological trappings, Malebranche held an avowed tri-partite model which was fundamentally the same as current accounts or representation (Bechtel 1998). It is amusing to notice how Malebranche's attempt to articulate his concerns are echoed today by Fodor. Thus, in a famous passage Malebranche wrote: I think everyone agrees that we do not perceive objects external to us by themselves. We see the sun, the stars and an infinity of objects external to us; and it is not likely that the soul should leave the body to stroll about the heavens, as it were, in order to behold all these objects... (Malebranche 1712: 217). Fodor writes in the same vein: It is, to repeat, puzzling how thought could mediate between behavior and the world .... The trouble isn't m anyhow, it isn't solely thinking that thoughts are somehow immaterial. It's rather that thoughts need to be in more places than seems possible if they' re to do the job that they' re assigned to. They have to be, as it were, 'out there' so that things in the world can interact with them, but they also have to be, as it were, 'in here' so that they can proximally cause behaviour.., it's hard to see how anything could be both (Fodor 1994: 83). Drawing attention to these parallels is not only interesting in its own right, but may suggest the source of the long-standing difficulty. The prospect of such
166
Peter Slezak
illumination is enhanced when it is noticed that yet further seemingly independent problems are merely variants of the same one. Thus, it is not a big leap from misrepresentation and illusions to notice that images are species of the same genus. Imagery involves illusory or non-veridical experiences of exactly the sort central to the classical argument for sense-data. The proverbial Pink Elephant of inebriated apprehension is not relevantly different from the subjects' imaginings in the celebrated experiments of Shepard & Metzler (1971) and Kosslyn (1994). Of course, at one time the so-called "Imagery Debate" was described by Block (1981) as "the hottest topic in cognitive science," but the connection with the other problems has not been made. If my conjectured parallel is warranted, it is perhaps no surprise that the imagery debate has been among the most persistently intractable disputes in cognitive science. In particular, the recalcitrance of the problem appears to arise not from its inherent scientific, explanatory difficulty but only from a notoriously seductive mistake (see Pylyshyn 2002, 2003; Slezak 2002b). An independent literature has grown around Putnam's (1975) "Twin Earth" fictional scenario. I will suggest that this, too, may be seen as a variant of a puzzle about meaning and misrepresentation. Less prominent today, the "Gettier Paradox" had generated a significant body of philosophical debate as a puzzle over knowledge conceived as "justified, true belief." Gettier cases involve various ways in which beliefs might be held for good reason, but be irrelevantly true, by mistake, as it were. It is of some interest to see that the current theoretical controversies at the heart of cognitive science merely rehearse this famous paradox. By illuminating one another from different and unusual directions, these parallels among diverse and often unconnected problems suggest the source of the difficulty.
Causation and Mis-causation? The contemporary puzzle of misrepresentation emerges as a paradox for the most plausible account of representation, that is, one which is "naturalistic." Of accounts in this category, Fodor (1984) explains: There are, so far as I know, only two sorts of naturalistic theories of the representation relation that have ever been proposed. And at least one of these is certainly wrong. The two theories are as follows: that C specifies some sort of resemblance relation between R and S: and that C specifies some sort of causal relation between R and S. The one of this pair that is certainly wrong is the resemblance theory (1984: 33).
The World Gone Wrong
167
Fodor notes that the difficulties with resemblance are "old n e w s " - - its symmetry, its being "in the eye of the beholder" etc. He rehearses these matters only to suggest why the idea of a causal theory of representation is attractive. As we will see presently, there is considerable irony in this contrast because, despite its apparent advantages, the causal theory fails for essentially the same reasons as resemblance and is, indeed, at a deeper level merely a variant of the same account. Nevertheless, Fodor observes You might wonder whether resemblance is part of the natural order (or whether it's only, as it were, in the eye of the beholder). But to wonder that about causation is to wonder whether there is a natural order (1984: 33). But this is to misrepresent the issues at stake. We see here that Fodor accords great weight to the causal account in view of its "naturalistic" status, implicitly suggesting that no other account would meet the naturalistic requirement. Fodor appears to suggest that any other account must be occult or "paranormal" in some way: Well, what would it be like to have a serious theory of representation? Here, too, there is some consensus to work from. The worry about representation is above all that the semantic (and/or the intentional) will prove permanently recalcitrant to integration in the natural order; for example, that the semantic/intentional properties of things will fail to supervene upon their physical properties (1984: 32). Thus, Fodor identifies naturalistic accounts of representation with causal accounts, but it is an exaggeration to say that challenging a causal account is ipso facto to question the very existence of a "natural order." The issue here is only a specific account of the representation relation and not a challenge to the naturalistic, causal picture of the world as such. That is, at worst the challenge questions whether representation must be direct causation. Obviously, there might be other kinds of relations which are consistent with naturalism but not causal in the required sense. That is, to doubt whether representation is to be explained by a causal theory is emphatically not to doubt whether representations have causes per se. A lesson from Chomsky is that a causal relation, for example, between stimulus and response, does not exhaust the theoretical, explanatory options consistent with a scientific naturalism. The intentional contents of propositional attitudes are no more likely to be under stimulus control than "verbal behaviour."
168
Peter Slezak
Along the lines we have noted, Fodor elaborates the paradoxical consequences of a causal theory of representation: The point of all this, I emphasize, is not to argue against causal accounts of representation. I think, in fact, that something along the causal line is the best hope we have for saving intentionalist theorizing, both in psychology and in semantics. But I think too that causal theories have some pretty kinky consequences... (1984: 34). This trouble is intrinsic; the conditions that causal theories impose on representation are such that, when they're satisfied, misrepresentation cannot, by that very fact, occur (Fodor 1984: 34). By comparison, the Malebranche-Locke argument for representative ideas takes the case of misrepresentation or illusion and recognizes that this cannot be caused in the usual way - - essentially Fodor's puzzle expressed in reverse. Fodor argues that, if ideas are caused by external objects, we can't have illusions; Malebranche and Locke argue that, if we have illusions, they cannot be caused by external objects. In both cases, the puzzle arises directly from a commitment to the tri-partite conception in which representations intervene between the mind and the world and are somehow correlated with it. In particular, the questions of veridicality for Locke's ideas arose from the impossibility of any comparison between representations and the world, except from the perspective of an independent outside observer. As Berkeley recognized, the very distinction between true and false ideas cannot be made without comparing representations and the world. Of course, this perspective is unavailable to the mind itself. Correspondingly, an explanatory theory cannot make tacit appeal to such a perspective without committing the homunculus error. This means that the veridicality or otherwise of mental representation does not serve an explanatory role and is, therefore, not a legitimate part of a theory of mind. In Berkeley's idealist response to this problem we can see the precursor to Fodor's position. If, as Fodor (1994:112) recommends, "semantics isn't part of psychology," mental representations seem closed off, like Locke-Malebranche ideas, from the external world. Securing the connection between representations and the world through causation simply binds them in such a way as to preclude error and thus causation functions for Fodor in the way that a mysterious correspondence worked for Locke.
Truth Conditions as Psychology? Of course, the problems of explaining truth and error are two sides of the same coin and, therefore, the puzzle of misrepresentation is a symptom of fundamental
The Worm Gone Wrong
169
problems in the conception of mental representations as semantically evaluable. Fodor clearly states the centrality of truth preservation to the computational conception: It is a characteristic of the mental processes.., that they tend to preserve semantic properties like truth. Roughly, if you start out with a true thought, and you proceed to do some thinking, it is very often the case that the thoughts that the thinking leads you to will also be true. This is, in my view, the most important fact we know about minds; no doubt it's why God bothered to give us any. A psychology that can't make sense of such facts as that mental processes are typically truth preserving is ipso facto dead in the water (Fodor 1994: 9). Elsewhere, too, Fodor remarks "we need it [broad content] to make sense of the fact that thoughts have the truth conditions that they do" (1994: 50). However, despite the essential place of truth conditions for a theory of intentionality, dire consequences seem to follow: It is, to put the point starkly, the heart of externalism that semantics isn'tpart of psychology. The content of your thoughts (/utterances), unlike for example, the syntax of your thoughts (/utterances), does not supervene on your mental processes (Fodor 1994: 38). The point is one Fodor had also made earlier in the conclusion to his (1980) paper where he said My point, then, is of course not that solipsism is true; it's just that truth, reference and the rest of the semantic notions aren't psychological categories (1980: 253). Notoriously, this had led to a widespread consensus on Fodor's pessimistic conclusion " . . . of the semanticity of mental representations we have, as things now stand, no adequate account" (Fodor 1985:31). Stalnaker (1991: 229) echoes Fodor saying "There is little agreement about how to do semantics, or even about the questions that define the subject of semantics..." Fodor repeatedly presents the idea that semantics isn't part of psychology as among the paradoxical consequences of the most plausible account of mental representation. On the one hand, solipsism can't be true, but on the other hand, we can't seem to do psychology on any other conception either since the semantic notions of truth and reference can't figure in a psychological theory as such. This
170
Peter Slezak
dilemma is acute for Fodor only because, as we have just seen, Fodor is firmly committed to a conception of mental processes as typically truth preserving. It seems we can't do psychology with semantic notions and we can't do psychology without them.
Hume's Problem Ironically, my formulation of the dilemma here echoes a remark of Dennett's in a quite different context in which he was explaining the difficulties for psychological theorizing that led either to behaviorism or infinite regress. In that context, Dennett remarked, ... psychology without homunculi is impossible. But psychology with homunculi is doomed to circularity or infinite regress, so psychology is impossible (Dennett 1978: 123). While it is perhaps not obvious, I would like to show that these two dilemmas are the same. At root, the puzzle of semantics is a version of the homunculus problem. In the present context this insight entails embracing one horn of Fodor's dilemma, that is acknowledging that truth and reference are not psychological categories and, therefore, have no place in cognitive science as a theory of mental representation. However, abandoning the explanation of truth and reference, and incidentally misrepresentation too, does not entail abandoning an account of the content of mental representations as such. Rather, it is to avoid a certain mistaken conception of these as semantically evaluable, that is, as intelligible to an external observer. We see now the connection with Dennett's paradox, for he was perceptively remarking on what he called "Hume's P r o b l e m " - - namely . . . nothing is intrinsically a representation of anything; something is a representation only for or to someone; any representation or system of representations thus requires at least one user or interpreter of the representation who is external to it (Dennett 1978: 122). Dennett observed that Hume "wisely shunned the notion of an inner self that would intelligently manipulate these items, but this left him with the necessity of getting the ideas and impressions to "think for themselves" (1978: 122). Undoubtedly the obscurity of this last requirement in practice has encouraged a certain neglect of its importance despite the notoriety of the homunculus problem in principle. My
The Worm Gone Wrong
171
suggestion is that the puzzle of semantics for mental representations, as distinct from the real explanatory problem, turns on the failure to heed Dennett's Humean warning. If semantic evaluability for mental representations tacitly involves the theorist understanding the symbol, we have simply fallen into the traditional trap. I wish to argue that truth and related semantic notions such as misrepresentation are implicitly dependent on just such a criterion of user-intelligibility or userinterpretability (see R. Cummins 1996). Rephrasing Dennett, my point will be that nothing is intrinsically a misrepresentation of anything, but only for or to some user or interpreter. That is, the question becomes: Can we capture the notion of intentional content without falling into the error of conceiving mental representations as themselves intelligible (to ourselves)? In particular, is misrepresentation, like resemblance, in the eye of the beholder? The difficulty of actually formulating a substantive explanatory theory which succeeds in getting the ideas to "think for themselves" should not preclude avoiding erroneous theories that demonstrably fail in this regard. In order to forestall misunderstanding, it is important to note that Dretske, for example, clearly recognizes the distinction between "original" and "derived" intentionality, for he begins his discussion of misrepresentation by explaining that "the problem i s . . . one of a system's powers of representation in so far as these powers do not derive from the representational efforts of another source" (p. 297?). Thus, Dretske clearly recognizes that a diagram, for example, might misrepresent only in the sense in which this is "derived from its role as an expressive instrument for agents.., who already have this power" (ibid.). It is important to appreciate that my charge against accounts such as Dretske's or Fodor's does not ascribe to them a lack of awareness of this familiar, important distinction. Nevertheless, it remains that they sin against it unwittingly in the very formulation of the problem. I will suggest that the notion of misrepresentation itself cannot be coherently articulated without tacitly invoking the very semantic concepts to be explained. In a familiar way, the theorist is unknowingly doing the work of the theory which only appears to work through the failure to notice this crucial contribution of a non-trivial nature. Thus, I agree with Fodor (1994) only regarding his prefatorial confession: "It may be that I have gotten myself into a philosophical situation about which all that can helpfully be said is that I ought not to have gotten myself into it" (1994: 1). Fodor's suspicion sums up the strategy of my discussion, for I contend that the problem Fodor has been struggling with for two decades is fundamentally misconceived. Semantic evaluability is not an unproblematic notion in this context since it invites an illegitimate reliance upon the very representational, semantic abilities we seek to explain.
172
Peter Slezak
Gettier Paradox It is instructive to consider a well-known philosophical problem in which interest has declined since its heyday though it bears an unnoticed parallel to the issues we have been considering in cognitive science today. The Gettier Paradox generated a considerable literature as a puzzle about the conception of knowledge as justified true belief. This conception was confronted by certain kinds of counter-example. Though seemingly remote from the concerns of cognitive science, these Gettier cases shed important light on problems of interest here because they are actually cases of misrepresentation. In the typical Gettier case, for example, someone x might see an outline in a field that looks very much like a cow and come to hold the belief"There is a cow in the field." In fact, the cow-looking silhouette is a bush, but unbeknown to x, there is, in fact a cow somewhere else in the field of which he is unaware. In this case, x has a belief which is both justified on the evidence and true, but not deserving to be called knowledge. Or, in a different example, a person might see an object ahead and not notice that it is, in fact, the reflection in a mirror of an object elsewhere. However, unknown to the person, behind the mirror there is an identical object exactly where the person believes one to be. Again, we have a case of justified, true belief that does not deserve to be called knowledge. How are these puzzles relevant to cognitive science? Without entering into the mire of Gettier exegesis, I suggest that the moral of these cases is simply that a conception of knowledge which is tied to truth is inappropriate by invoking a God's eye perspective on ultimate realities that may be unknown or unknowable. Without such omniscience, we can only have justified belief and should dispense with the concept of knowledge as truth. In all Gettier cases, the circumstances that make x's belief true are known only to us from a vantage point which is irrelevant to a characterization of a person's state of mind. These cases might be called veridical or true misrepresentations since the truth of the belief in question is accidental and irrelevant to its warrant, just as a genuine misrepresentation might be false for accidental and irrelevant reasons. Dretske/Fodor examples of misrepresentation and Putnam's Twin Earth cases are of this latter sort and their parallel with Gettier examples has been unnoticed perhaps because they are mirror-images of one-another. The Gettier case of a bush-caused cow belief is exactly the same as Fodor's dog-caused cat belief. The truth of the former and the falsity of the latter are irrelevant to questions of interest to psychology. Indeed, these cases are closer to those of mirages and seemingly bent sticks in water being due to the perversity of the world rather than a lapse of cognition. Certainly when trying to characterize the mind for the purposes of psychology, describing beliefs as "knowledge" is inappropriate if this requires their truth. Any adequate, even complete account, of a person's psychology would have to invoke
The Worm Gone Wrong
173
only the relation of beliefs to available evidence and not their actual, ultimate truth value. The moral of the Gettier cases seems to be that the world can mislead us in various ways, giving us good reasons for things that may be false, bad reasons for things that may be true and good reasons for things that may be true for other reasons. None of this should occasion deep philosophical anxieties, at least for those concerned with psychology. Dretske's (1994) misled magnetic bacterium has a justified belief that the direction he swims is "up" even though it is actually down. Likewise, the frog has a justified belief that there is a fly at 12 o'clock high even though it is actually an experimenter's lead pellet. Such possibilities for a mismatch between beliefs and the world don't appear to raise genuine explanatory questions. Do the odometer and speedometer of a bicycle misrepresent in some sense when the bike is ridden on rollers and not moving? Would someone in a full virtual-reality suit be guilty of murder if they strangled someone who happened to be there in the real world as well as in the virtual world? Or would they simply be acquitted on the mitigating grounds of a Gettier defence? The question may be a good plot for a movie, but does not deserve serious philosophical attention for cognitive science.
The World Gone Wrong? The issues of interest here may be clarified by the simple cases just referred to. Thus, Fodor made an important observation in remarking ... it is not the frog but the world that has gone wrong when a frog snaps at a bee-bee (Fodor, quoted in R. G. Millikan 1991: 161). Drawing out this remark helps us to see how to avoid Fodor's own dilemma noted earlier. In brief, it suggests that the very problem of misrepresentation and, consequently, the entire constellation of semantic notions, are irrelevant to the explanatory enterprise of cognitive psychology. Simply put, the question is: Why should a theory of frog psychology concern itself with explaining the frog's "mistake" at all? Obviously, if the world or experimenter are sufficiently perverse they may contrive to alter things to a degree which will create a mismatch between the frog's representations and the world. Psychology has no obligation to explain why the world has gone wrong. Conversely, however, psychology has no obligation to explain why it usually goes fight either. Hand-waving in the direction of evolution suffices here. Cardiology has no obligation to explain why it is generally blood that flows in our veins. Conceivably, one might contrive things so that Cabernet Sauvignon would replace the usual liquid. However, such a possibility is not likely to cause theoretical anxieties for medical science
174
Peter Slezak
any more than the disorientation of Dretske's (1994) maliciously manipulated monocellular microbes should cause analogous worry for cognitive science. I have been trying to suggest that it is not clear why a scientific theory of the mind should be concerned with the truth conditions of mental states except in an ersatz manner according to which we wish to know the various ways in which minds are more or less adapted to the world. By manipulating their optic flow, Srinivasan's honeybees can be fooled into landing too fast or overshooting their target when foraging (see Zhang 1999, 2000). Such studies tell us a great deal about the mechanisms of internal representation in the bee, but it is difficult to see why Srinivasan should care about the philosophical problems these facts are alleged to give rise to. Knowledge of truth conditions is neither available nor relevant to the bee, but only to us as theorists who interpret the supposed meaning of its symbols as if these were intrinsic relations between symbol and world. The possibility of dislocating the normal arrangements should not occasion deep metaphysical worry. The "pre-established harmony" between representations and the world is not a matter of truth conditions in the philosophers' sense, but only the usual adaptedness of biological systems to their niche. As we will note presently in relation to philosophical conceptions of animal cognition, it is essential to distinguish the truth of beliefs from the knowledge of their truth. The latter involves a reflective judgement and, at best, only the former can be of relevance to the explanatory enterprise of cognitive science. As I have been suggesting, even this must be qualified with scare quotes, since "truth" of beliefs here means only "more or less." The point is closely analogous to H. A. Simon's insight about bounded rationality and the requirement for organisms to make satisficing rather than optimizing decisions. In the same spirit here, and for the same reasons, it is not the truth of our beliefs that is to be explained but only their general reliability.
Symbol-- World Relations By contrast with such "naturalistic" studies of misrepresentation, consider how the philosophical problems are taken to arise. Fodor (1998b: 14) says that "content is constituted, exhaustively, by symbol-world relations" and specifically ... what bestows content on mental representations is something about their causal-cum-nomological relations to the things that fall under them: for example, what bestows upon a mental representation the content dog is something about its tokenings being caused by dogs (ibid.: 12).
The Worm Gone Wrong
175
However, the need to find a way to tie representations to the world is a mistake derived from thinking about them qua understander rather than qua theorist: As understanders we "see" the connection and feel the need to explain its apparent firmness and basis, but as theorists we need only explain how behaviour tends to be more or less successfully coordinated to the world by the mediation of thoughts. The way thoughts mediate in this way to coordinate behaviour is undoubtedly a difficult scientific problem, but does not appear to raise the typical philosophical problems of intentionality. In particular, my thoughts about dogs need not, in general be caused by dogs but only enable me to identify them when present or to find them when not. Likewise, more importantly, for food and a mate. This conception places much less stringent demands on a theory of mental representation and, in particular, removes the seeming paradoxes of misrepresentations or mismatches that may occur. Failures of representation, if systematic, will cause extinction just like any other maladaptation. Causal-cum-nomological relations are too strong to be the connection between representations and things that fall under them precisely because things may so easily fail to fall under them. First, it is hard to see how causal relations could work with anything other than the contents of immediate perception. In the absence of any canine in the vicinity, my current thought of "dog" is not plausibly caused by a dog in any reasonable sense. Second, the mismatches between symbols and world can be generated too easily to be dismissed in ceteris paribus clauses as somehow extraneous or minor perturbations to an otherwise sound theory. How can a Fodorian causal-nomological account of contents possibly explain the generation of ideas and thoughts by fictions, myths, movies, reveries, fantasies and other misrepresentations? A "virtual reality" suit of the science fiction movie Matrix would generate a rich mental life without any causal, nomological relation between contents and the world. Or, more realistically, are we to ascribe Einstein's thoughts about the structure of space-time to actual causal-nomological relations between them? It seems that the characteristic feature of thoughts is, indeed, their very loose connection with the world that they are about. In this sense, misrepresentation seems to be the rule rather than the exception, "wild" tokenings of thoughts being the norm. Except perhaps for perceptual experience, it seems that occurrent thoughts are more often about how things might be or should have been rather than how thing are. Thus, Fodor's contemporary problem of misrepresentation, like Berkeley's problem of veridicality for Locke's theory, appears to arise from tacitly adopting the stance of external interpreter. As Berkeley recognized, the very distinction between veridical and non-veridical itself cannot be coherently formulated except in terms of judgements which cannot be part of the explanatory enterprise. Like the picture on a jig-saw puzzle, the meaning of representations conceived as semantically evaluable is for our own convenience and not intrinsic to the
176
Peter Slezak
arrangements of inter-locking components. This is not quite the same as advocating a "narrow" or "solipsistic" psychology, but only to warn against conceiving the external connections with the world on the basis of our own understanding as theorists for whom the representation may be construed as an external symbol and its systematic relation to the world is then seen as an inexplicable link. The link seems inexplicable because we have been looking in the wrong place: like a Rorschach ink-blot, the pattern is in us, not in the world.
Twin Earth as Misrepresentation A minor philosophical industry had developed conceming Putman's (1975) Twin Earth scenario as yet another puzzle about our concepts of representation and intentionality. However, although treated as an independent puzzle conceming the nature of meaning, the Twin Earth problem, too, seems to be a variant of the problem of misrepresentation. Instead of the familiar scenario comparing myself with my Twin Earth double, we may imagine an equivalent circumstance in which I am unknowingly transported to Twin Earth. There, like my atom-for-atom replica in the original case, I will refer to XYZ as "water" without knowing that i am no longer on earth referring to H20. The usual perplexities conceming "narrow" and "wide" meaning seem more difficult to motivate in this case because we are now dealing with the same individual and so the question of different meanings cannot easily arise. However, this scenario is identical to the original one in every relevant respect by virtue of the assumption of twin identity. The original case is supposed to show that a term may have different "wide" meanings which are not determined by what is in the head. However, since the term in question is now in a single head, my case forces us to give a different analysis. Instead of terms having different meanings on earth and Twin Earth, we now have a case of inadvertent misrepresentation. When transported to Twin Earth, my use of the term "water" is now mistaken. Clearly, however, my mistake may be unknown and unknowable to anyone except to ourselves who tell the story. The case is now obviously identical with that of the hapless frog who has a piece of lead replace his lunch. Thus, the Twin Earth story or my variant of it may be portrayed equivalently by imagining that on this Actual Earth, God secretly switches all H20 to XYZ. As before, my use of the term "water" is now in error when the world is surreptitiously altered. We see that Putnam's Twin Earth story is, indeed, simply another way of telling Dretske's story about the magnetic micro-organisms fooled into thinking up is down. It is the world not the organism that is in error. It is only by illegitimately adopting the external omniscience of the theorist that we can formulate these puzzles.
The Worm Gone Wrong
1'/'/
D a v i d s o n on M i s t a k e s Suggestive evidence that the puzzles concerning belief and misrepresentation arise from a conflation of thinking, on one hand, and reflective thinking about thinking, on the other hand, is seen in Davidson's claim that a creature cannot have a belief if it does not have the concept of belief. This is supposedly because "Someone cannot have a belief unless he understands the possibility of being mistaken, and this requires grasping the contrast between truth and e r r o r - true belief and false belief" (1975: 22). However, Davidson appears to be relying on conflating the mere having of beliefs with having beliefs about beliefs. This assimilation is no mere oversight, for it is explicitly embraced as a doctrine about the nature of belief and thought. Thus, Carruthers says conscious thought is a matter of having second-order thoughts about other thoughts (1996: 157). Carruthers, too, suggests that discrimination between experience requires "that the subject should be capable of thinking about" (and so conceptualising) its own experiences (1996: 157). Nevertheless, if this is not a mere stipulation about how to use the words "belief" and "thinking," it seems exceedingly unsatisfactory from a theoretical point of view to deny that animals might have beliefs because they are unable to know that they have them and reflect on their truth value. It is entirely unclear why a creature could not have both true and false beliefs without also realizing which was the case in a particular instance. A cat can surely be correct in thinking that a mouse is in a certain hole without having the concepts of belief and truth. As Fodor has aptly put the point: Surely what matters to whether it's all right for me to step on the cat's tail is primarily whether it hurts him, not what he thinks about it: still less whether he thinks about what he thinks about it (1999: 12,13). Searle (1994) has also questioned Davidson's demand for second-order notions such as the concept of belief in the attribution of belief, but it is noteworthy that his way of expressing his contrary view subtly encourages the very mistake he attributes to Davidson. Thus, Searle says that animals need not have metalinguistic beliefs or concepts in order to have true or false beliefs, but he expresses his point by saying that an animal is able "to distinguish the state of affairs in which their belief is satisfied from the state of affairs in which it is not satisfied" (1994:212) or an animal "can tell whether its belief is true or false" ( 1994:212). However, such locutions are ambiguous in continuing to suggest that an animal might have beliefs about their beliefs whereas they merely have them. That is, strictly speaking, an animal does not "tell true from false beliefs" as Searle puts it, but only has true or false beliefs.
178
Peter Slezak
We, as theorists, may "tell" true from false beliefs since this is a meta-linguistic judgement, and is not the same as simply having beliefs which may be true or false.
Visual Imagery Perhaps the clearest illustration of the notorious error I have been trying to indicate is seen in the long standing "Imagery Debate." As already noted, the interest of this debate here derives from the fact that images are precisely illusions or misrepresentations of the sort which have been central to the independent issues discussed. In this domain, a "pictorial" or "depictive" theory of visual images has been championed principally by S. Kosslyn (1994) who declares the "resolution" of the controversy, but there remain grounds for skepticism. Although the charge is strenuously rejected by its partisans, the pictorial theory is repeatedly accused of committing the homunculus error as a consequence of postulating representations that resemble their objects. However, resemblance and other pictorial properties only appear to work because we, as observers, can interpret them. Without entering into the details of this notorious debate (see Pylyshyn 2002, 2003; Slezak 1995), it is important only to note here that it is simply one version of a puzzle that arises in different guises at different times and different contexts. Thus, it is significant that the problem of misrepresentation has a certain plausible analysis that corresponds precisely with an independently plausible solution to the imagery d e b a t e - a question-begging dependence on the theorist's intelligence. If misrepresentations are just images by another name, these parallels are just what we would expect.
Verstehen Sie? Intelligibility or Explainability? The problems we have been noting seem to arise from an assumption that we cannot understand a creature's thought unless we can share it and express it as one of our own (Davidson 1995). However, this view involves an egregious equivocation on the notion of understanding which can mean interpreting a meaningful representation (verstehen), or explaining it as in science (erkliiren). Failure to heed this distinction has bedevilled a long dispute in the social sciences (Slezak 1990; Winch 1957) and is at the heart of the vexed debates in cognitive science too (Slezak 1999). The undoubted fact that we are unable to interpret a dog's thoughts and express them in our own language has no bearing on the possibility of understanding them through science. Likewise, the question of semantics or meaning of mental representations in cognitive science is standardly confused between whether representations are intelligible and whether they are
The Worm Gone Wrong
179
explainable as in Davidson' s 'interpretative' approach to meaning. Searle' s (1980) notorious Chinese Room conundrum trades directly on this confusion by asking whether a fully intelligent language understander can interpret computational symbols which are the substrate of thought. Their unintelligibility to such a homunculus intruding into the system has curiously been taken as relevant to the theoretical, explanatory issue of the intentionality of symbols. However, this explanatory issue must be the quite different question of how the symbols relate to one another, the organism's behaviour and the outside world. The puzzle in these related cases is pernicious because we are entangled in a pseudo-explanatory question from the very outset: The very first step identifying the problem assumes that veridicality or otherwise for mental representations is an appropriate scientific problem. However, this assumption is not neutral with regard to certain crucial substantive and problematic claims. Specifically, the explanatory question about misrepresentation, like resemblance of images, can only be formulated from the vantage point of an external intelligence which can understand the representations as meaningful in one sense of this crucially ambiguous term. Undoubtedly the close coordination of internal representations with the external world is required for intelligence as an adaptive feature of organisms, but it is easy to confuse the requirements of a satisficing coordination with truth as this notion has given rise to the recent discussions. The scientific, explanatory question here is distinct from the philosophical pseudo-problem. Asking about truth and its converse invites us to think of mental representations as external objects of apprehension and intelligibility themselves. This is of course the notorious error which has plagued theorizing about the mind throughout its history. Asking about the truth or falsity of a representation is inescapably to invite understanding its meaning. However, this is the crucial slip from the respectable explanatory concept of intentionality to the spurious observer-relative sense of meaningfulness. It should be clear that the role of mental symbols and our scientific concern to explain this must be independent of whether or not we are able to understand the symbols as meaningful to us. Searle's criterion of "meaningfulness," namely, whether or not he, as homunculus, can understand symbols, is surely an irrelevant criterion, though the mistake is not Searle's alone, for his famous scenario correctly captures the assumptions embodied in the Simon-Newell physical symbol system hypothesis. These symbols have been explicitly conceived on the model of an uninterpreted logical calculus with user-ascribed meaning (Newell 1986; Newell & Simon 1976). In this sense, Searle's scenario must be seen, not as a refutation of "strong AI" per se, but as a reductio ad absurdum of the universal assumptions about the nature of the symbols m the same assumption that is at the heart of the problem for pictorial imagery and misrepresentation.
180
Peter Slezak
In one of its metastases throughout cognitive science, the pathology of observerascribed meaning has arisen independently within AI as a debate about the classical logical conception in which an abstract formal system gets its meaning from a model theory. However, Woods (1987), Smith (1987), Rosenschein (1985) and others have been insisting that this logicist view is fundamentally misguided in its conception of the way in which an intelligent system gets to relate to the external world. Confirming its centrality to cognitive science, we see Newell & Simon (1976) give a clear statement of the logical view which Fodor (1980) later called the "formality condition." They trace the roots of their symbol system hypothesis explicitly to the program of Frege & Russell. Explaining the position, Newell says: The idea is that there is a class of systems which manipulate symbols, and the definition of these systems is what's behind the programs in AI. The argument is very simple. We see humans using symbols all the time. They use symbol systems like books, they use fish as a symbol for Christianity, so there is a whole range of symbolic activity, and that clearly appears to be essential to the exercise of mind (1986: 33). This explanation is striking for the explicitness with which is assimilates internal, mental representations with our external communicative symbols such as words and pictograms. Nilsson (1987, 1991), too, has been explicit in citing our symbolic artefacts such as books against the "proceduralist" position that representations must be such as to be utilized by the system rather than understood or ascribed meaning by the designer.
Unifying the Variety? We see the same conflation clearly and self-consciously defended by Dennett (1978): What is needed is nothing less than a completely general theory of representation, with which we can explain how words, thoughts, thinkers, pictures, computers, animals, sentences, mechanisms, states, functions, nerve impulses, and formal models (inter alia) can be said to represent one thing or another (1978: 91). The hoped-for unification in cognitive science was to be achieved precisely by showing that these heterogeneous items are all species of the same genus. Dennett explains further:
The World Gone Wrong
181
It will not do to divide and conquer here m by saying that these various things do not represent in the same sense. Of course that is true, but what is important is that there is something that binds them all together, and we need a theory that can unify the variety (1978: 91). Incidentally, protestations about exactly this conception of unification were at the heart of Arnauld's critique of Malebranche in the seventeenth century. Following Descartes, Arnauld argued ... what has thrown the question of ideas into confusion is the attempt to explain the way in which objects are represented by our ideas by analogy with corporeal things, but there can be no real comparison between bodies and minds on this question (Arnauld 1683/1990: 67). In case the previous quotations are thought to be unrepresentative, it is not difficult to find remarks from leading researchers that reveal precisely this assimilation of internal and external representations. Thus Rumelhart & Norman (1983) wrote: We define a symbol to be an arbitrary entity that stands for or represents something else. By "entity" we mean anything that can be manipulated and examined... Humans also use external devices as symbols, such as the symbols of writing and printing, electronic displays or speech waves (1983: 78). Consider a recent statement by Dan Lloyd which explains the modern notion of representation: Humans are representing animals, and we have built a world crammed with representations of many kinds. Consider, for example, the number and variety of pictorial representations: paintings, photographs, moving pictures, line drawings, caricatures, diagrams, icons, charts, graphs, and maps. Add the variety of linguistic representations, in signs, titles, texts of all kinds, and especially spoken words and sentences . . . . Human life, in short, is largely a cycle of making and interpreting representations (Dan Lloyd 2001). From the earliest remarks of Newell and Simon we see a consistent conflation of our own representational acts and artifacts with mental representations, and it becomes
182
Peter Slezak
clear why truth conditions should be thought relevant to explanation in cognitive science. The paradigmatic cases of semantic evaluability are our representational artifacts which we "use" and whose meaning we understand in a sense necessarily quite different from that of mental representations. Judgements of semantic evaluability are inescapably cognitive judgements requiting the theorist to contribute illegitimately to the explanatory enterprise by tacitly invoking significant abilities which are the very ones to be explained. We see this mistake embodied in the very foundational assumptions of cognitive science, as Birnbaum has perceptively noted, speaking of"the logicist assumption, generally implicit, that a successful organism's beliefs about the world are such that the real world is a model of those beliefs" (1991: 63): The leap of faith in the logicist program is the presumption that, in saying something about what it means for a machine to have beliefs, AI is obliged to reiterate a theory of how logical symbols are to be interpreted, developed over the last century in logic and mathematics for fundamentally different purposes ... (1991: 62).
Glue for Meaning and Reference? Fodor's earlier criticism of proceduralist views has been posed recently as a question for Block's "conceptual role semantics." Fodor & Lepore (1992) note that the arguments owing to Frege and Putnam suggest that "the relation between meaning and reference is deeply problematic" (1992: 166) since Frege shows that sameness of reference does not guarantee sameness of meaning, whereas Putnam's Twin Earth story shows that sameness of meaning does not guarantee sameness of reference. They ask, what is the glue that keeps the two factors, meaning and reference, stuck together (Fodor & Lepore 1992: 170). Fodor and Lepore present the fundamental puzzle as follows: What we need to know is what precludes radical mismatches between intension and extension. Why can't you have a sentence that has an inferential role appropriate to the thought that water is wet, but is true iff 4 is prime? (1992:171). This is, of course, Fodor's question: What makes a computer program play chess rather than simulate the Six Day War (Fodor 1978: 207). Fodor and Lepore say that no adequate semantics could allow an expression whose intension and
The Worm Gone Wrong
183
extension were so radically disconnected as having inferential role appropriate to the thought that water is wet but truth conditions for 4 is prime. However, their following question is most revealing. They ask "What on earth would it mean?" (1992: 170). Presumably, this question asks how we might conceivably understand an expression whose intension and extension diverged in this way. However, unless it is merely a fa9on de parler, the very question is inappropriate and suggests that the problem may arise precisely from conceiving the explanatory problem in terms of how we might understand mental representations as distinct from how we might explain them. It should not be assumed that this assimilation of internal and extemal representations has been merely an unproblematic mode of speech. We can see the ubiquitous sliding from one to another despite a clear understanding of the distinction. Thus, Block (1986) distinguishes between "autonomous" and "inherited" meaning saying "The representations on the page must be read or heard to be understood, but not so for the representation in the brain (1986: 83). So far so good. However, despite this explicit recognition of the distinction, the discussion is conducted in terms which tend to obscure or blur the crucial differences. Thus, apparently talking about the meaning of inner mental representations, Block asks "what it is to grasp or understand meanings?" (1986: 82) and "what is [it] for the brain to grasp meanings...?" (1986: 83). The idea of grasping or understanding here is exactly the wrong one in relation to internal representations which are precisely not grasped or understood. Similarly, Block speaks of a theory of semantics and "using an expression" but again this notion prejudices the discussion in favour of an observer-relative conception of meaning. But whether or not we can understand the system in this sense is irrelevant to the question of its meaning in the sense of intentionality. In particular, when our own comprehension of its intelligibility is not involved, the puzzles of truth conditions and misrepresentation dissolve. These become transformed into questions about the coordination of mental computations through perception and behaviour with the world and how these might break down. The latter problems are hard enough but they are presumably only problems not mysteries.
References Arnauld, A. (1683/1990). On true andfalse ideas. Translated with introductory essay by S. Gaukroger. Manchester: Manchester University Press. Ayer, A. J. (1940). The foundations of empirical knowledge. London: Macmillan. Bechtel, W. (1998). Representations and cognitive explanations: Assessing the dynamicist's challenge in cognitive science. Cognitive Science, 22(3), 295-318. Block, N. (Ed.) (1981). Imagery. Cambridge, MA: MIT Press.
184
Peter Slezak
Block, N. (1986). Advertisement for a semantics for psychology, Midwest studies in philosophy, X; reprinted in: S. Stich, & T. Warfield (Eds), Mental representation: A reader ( 1994, pp. 81-141). Oxford: Blackwell. Carruthers, P. (1996). Language, thought and consciousness: An essay in philosophical psychology. Cambridge: Cambridge University Press. Cummins, R. (1996). Representations, targets and attitudes. Cambridge, MA: MIT Press. Davidson, D. (1975). Thought and talk. In: S. Guttenplan (Ed.), Mind and language. Oxford: Clarendon Press. Davidson, D. (1995). Could there be a science of rationality? International Journal of Philosophical Studies, 3(1), 1-16. Dennett, D. (1978). Artificial intelligence as philosophy and as psychology. In: Brainstorms (pp. 109-126). Vermont: Bradford Books. Dretske, E (1994). Misrepresentation. In: S. Stich, & T. A. Warfield (Eds), Mental representation. Oxford: B lackwell. Fodor, J. (1978). Tom Swift and his procedural grandmother, cognition, 6; reprinted in: Representations (1981, pp. 204-221). Cambridge, MA: MIT Press. Fodor, J. (1980). Methodological solipsism considered as a research strategy in cognitive psychology. Behavioral and Brain Sciences, 3, 63-109. Reprinted in: Representations (1981). Cambridge, MA: MIT Press. Fodor, J. (1984). Semantics, Wisconsin style. Synthese, 59, 231-250; reprinted in: A theory of content and other essays (1990, pp. 31-49). Cambridge, MA: Bradford/MIT Press. Fodor, J. (1985). Fodor's guide to mental representation. Mind, 94. Reprinted in: S. Stich, & T. Warfield (Eds), Mental representation: A reader (1994, pp. 9-33). Oxford: Blackwell. Fodor, J. A. (1994). The elm and the expert. Cambridge, MA: MIT Press. Fodor, J. A. (1998). Concepts: Where cognitive science went wrong. Oxford: Oxford University Press. Fodor, J. A. (1999). Not so clever Hans. London Review of Books (February), 12-13. Fodor, J. A. (2002). Hume variations. Oxford: Oxford University Press. Fodor, J. A., & Lepore, E. (1992). Holism: A shopper's guide. Oxford: Blackwells. Gregory, R. L. (1997). Knowledge in perception and illustration. Philosophical Transactions of the Royal Society London B, 352, 1121-1128. Jackendoff, R. (1992). Languages of the mind. Cambridge, MA: Bradford/MIT Press. Kosslyn, S. (1994). Image and brain: The resolution of the imagery debate. Cambridge, MA: MIT Press. Locke, J. (1690/1975). Essay concerning human understanding. P. H. Nidditch (Ed.). Oxford: Clarendon Press. Lloyd, D. (2001). Representation. Macmillan encyclopedia of cognitive science. London: Macmillan. Malebranche, N. (1712/1997). The search after truth. Edited and translated by T. M. Lennon, & P. J. Olscamp. Cambridge: Cambridge University Press. Millikan, R. G. (1991). Speaking up for Darwin. In: Loewer, & Rey (Eds), Meaning in mind: Fodor and his critics (pp. 151-164). Oxford: Blackwell. Nadler, S. (1992). Malebranche and ideas. Oxford: Oxford University Press.
The World Gone Wrong
185
Newell, A. (1986). The symbol level and the knowledge level. In: Z. Pylyshyn, & W. Demopoulos (Eds), Meaning and cognitive structure. Norwood: Ablex. Newell, A., & Simon, H. A. (1976). Computer science as empirical inquiry: Symbols and search. Communications of the ACM, 19, 113-126. Putnam, H. (1975). The meaning of 'meaning'. In: K. Gunderson (Ed.), Language, mind and knowledge: Minnesota studies in the philosophy of science, Volume 7. Minneapolis: University of Minnesota Press. Reprinted in: H. Putnam, Mind, language and reality: Philosophical papers (Vol. 2, pp. 215-271). Cambridge: Cambridge University Press. Pylyshyn, Z. (2002). Mental imagery: In search of a theory. Behavioral and Brain Sciences, 25(2), 157-237. Pylyshyn, Z. (2003). Seeing and visualizing: Its not what you think. Cambridge, MA: MIT. Reynolds, S. L. (2000). The argument from illusion. Nous, 34(4), 604-621. Rosenschein, S. J. (1985). Formal theories of knowledge in AI and robotics. New Generation Computing, 3, 345-357. Rumelhart, D. E., & Norman, D. A. (1983). Representation in memory, preprint ONR Report 8302, June 1983, for Steven's Handbook of Experimental Psychology, R. C. Atkinson, R. J. Herrnstein, G. Lindzey, & R. D. Luce (Eds). New York: Wiley. Searle, J. (1980). Minds, brains and programs. Behavioral and Brain Sciences, 3, 417-424. Searle, J. (1994). Animal minds. In: E A. French, T. E. Uehling, & H. K. Wettstein (Eds), Philosophical naturalism: Midwest studies in philosophy (Vol. XIX). Indiana: University of Notre Dame Press. Shepard, R. N., & Metzler, J. (1971). Mental rotation of three dimensional objects. Science, 171, 701-703. Slezak, E (1990). Man not a subject for science? Social Epistemology, 4(4), 327-342. Slezak, E (1995). The 'philosophical' case against visual imagery. In: E Slezak, T. Caelli, & R. Clark (Eds), Perspectives on cognitive science: Theories, experiments and foundations. Norwood: Ablex. Slezak, E (1999). Situated cognition: Empirical issue, paradigm shift or conceptual confusion? In: J. Wiles, & T. Dartnall (Eds), Perspectives on cognitive science (Vol. 2). Norwood: Ablex. Slezak, E (2002a). The tri-partite model of representation. Philosophical Psychology, 13(2), 239-270. Slezak, E (2002b, April). The imagery debate: D6j?a vu all over again? Commentary on Zenon Pylyshyn. Behavioural and Brain Sciences, 25(2), 209-210. Smith, B. C. (1987). The correspondence continuum. CSLI Report 87-71. Stalnaker, R. (1991). How to do semantics for the language of thought. In: B. Loewer, & G. Rey (Eds), Meaning in mind: Fodor and his critics (pp. 229-237). Oxford: Blackwell. Winch, E (1957). The idea of a social science. London: Routledge. Woods, W. A. (1987). Don't blame the tools. Computational Intelligence, 3, 228-237. Zhang, S. W. (1999). Honeybee memory: Navigation by associative grouping and recall of visual stimuli. Neurobiology of Learning and Memory, 72, 180-201. Zhang, S. W. (2000). Maze navigation by honeybees: Leaning path regularity. Learning and Memory, 7, 364-374.
This Page Intentionally Left Blank
Chapter 10
Representation, Reduction, and Interdisciplinarity in the Sciences of Memory John Sutton
Introduction" Memory and Interdisciplinary Memory is studied at a bewildering number of levels, in a daunting range of disciplines, and with a vast array of methods. Is there any sense at all in which memory theorists m from neurobiologists to narrative theorists, from the developmental to the postcolonial, from the computational to the cross-cultural m are studying the same phenomena? This exploratory review paper sketches the bare outline of a positive framework for understanding current work on memory, both within the various cognitive sciences and across the gulfs between the cognitive and the social sciences. The project is, obviously, hopelessly ambitious. How could the concepts, models, or practices of such glaringly incompatible activities as clinical neuropsychology and media theory, or developmental psychology and Holocaust studies, ever be imported into neighbouring discursive universes? More to the point, why would anyone bother? Those who work at the "subpersonal" level, trying to understand neural processes or to construct better models of various memory systems, display either respectful neutrality towards social science, or active disdain towards its perceived anti-naturalism. And those who study "social" levels may think of psychology either as important but irrelevant, or as irretrievably marred by individualism. Sensible advice to leave the frameworks distinct can be met, I
Representation in Mind: New Approaches to Mental Representation Edited by H. Ciapin, P. Staines and P. Slezak Copyright 9 2004 by Elsevier Ltd. All rights of reproduction in any form reserved. ISBN: 0-08-044394-X
188
John Sutton
suggest, by arguing not for a unitary view of memory, but for an integrated framework within which different memory-related phenomena might be understood. The skeletal structure of such an integrated framework might already be in place, implicit in overlapping concepts and theories across the levels at which memory is studied. This paper is merely a provocative preliminary vision of this structure. It should elicit either principled defences of the idea that certain disciplines and levels of explanation in the study of memory should remain insulated and autonomous, or help in beginning to forge the elusive connections I seek. My claims are four, claims to be understood more as guiding hypotheses for further work than as results. Firstly, significant problems, as yet neglected by both practitioners and commentators, about reduction and interdisciplinarity arise even within the relatively restricted list of disciplines comprising the cognitive sciences of memory. Secondly, suitably weak approaches to reduction will allow for substantive intertheoretic and interdisciplinary contact across some of those sciences. Thirdly, even within accepted subdisciplines of cognitive psychology, key concepts and frameworks increasingly point to the relevance of factors outside the individual. And finally, if these three claims go through, there is already a growing body of relevant research on "culture-and-cognition" which shows ways to link certain (broadly connectionist/dynamicist) theories of individual memory with the study of external memory systems in technology and society. Few commentators have been equipped to analyse the oddly disconnected parallel courses of the recent explosions of work on memory in the cognitive and the social sciences respectively. Even a listing of the modem disciplines of memory reveals the difficulty of the task. 1 Some set of capacities labelled "memory" is a key focus of research, funding, institutional and pedagogical energy, and writing in (for example) molecular neurobiology, computational neuroscience, cognitive neuropsychology and neuropsychiatry, comparative psychology and ethology, cognitive psychology of many stripes, developmental psychology, personality psychology, social psychology, psychoanalysis, psychiatry, sociology, anthropology, philosophy, history, political theory, media studies, museumology, Holocaust studies, postcolonial studies, and literary theory. Researchers whose attention edges out from their home discipline are thus often driven to express doubts about the unity of the phenomena in question. Susan Engel, for example, notes in the preface to her book Context is Everything that "in I This listing of disciplines, and the claims associated, needs to be defended in secure empirical, bibliographical, and sociologicalstudies. For the moment,the impressionisticsurveymust suffice. The hierarchical order of the list is intendedmerelyto catch an impreciseand general sense of current work, rather than to beg any questions about interdisciplinaryrelations.
Representation, Reduction and Interdisciplinarity
189
recent years the topic of memory has become so popular, it seems both ubiquitous and yet oddly invisible. People glide seamlessly from a discussion of childhood recollections to national memories, as if they were part of the same phenomenon" (1999" viii). 2 Despite Engel's sensible caution, her balanced and elegant book stretches from developmental and personality psychology to include careful discussions of literary autobiography and historical memory. But I also want to persuade psychologists who, unlike Engel, are sceptical about the relevance of social studies of memory to their domain that there are substantive and rarely asked questions here about the scope of the cognitive sciences. Can we use a common commitment to mental representations, or to an information-processing framework within cognitive science, to mark off its boundaries and to exclude the social sciences of memory? This would require confidence that some such commitment can be specified clearly and non-trivially, and that it would indeed encompass all of the desired subdisciplines of cognitive science. But even then such a framework might, I'll suggest, itself be extended to deal with cultural and historical memory phenomena. Most episodes of human remembering have multiple causes, describable in different vocabularies at different levels, and which are not restricted to the past events or experiences remembered. So, in part, questions about interdisciplinarity are questions about which causes matter for specific explanatory purposes. Yet often, when writers on memory do acknowledge legitimate issues about how different levels and disciplines might or might not knit together, they are not sensitive to the difficulty of specifying such purposes, and of carrying out particular programs of translation, reduction, or interdisciplinary theory construction. Hype abounds. The distinguished neuroscientists Larry Squire and Eric Kandel write that "memory promises to be the first mental faculty to be understandable in a language that makes a bridge from molecules to mind, that is, from molecules to cells, to brain systems, and to behavior" (1999: 3). 3 They believe that we're on the verge of a truly multidisciplinary, jointly social and natural, science of memory: "the molecular and cognitive study of memory represents only the most recent attempt, historically, to bridge the sciences, which are traditionally concerned with nature
2 Engel's intention, in one of the best recent syntheses of current psychological perspectives, is to sketch theoretical connections linking "the magnificent but bewildering array of studies and insights that have emerged over the last 10 years," in order to investigate lines along which they might be contrasted with each other. Her attempt "to construct a framework for thinking about memory" touches on more of my wild list of disciplines than any other comparable text, and is particularly strong on developmental studies, for which see the section on Constructive Remembering below. 3 For a more cautious popular assessment of the high-level relevance of the neurobiology of memory see Rose (1992, Chapter 13).
190
John Sutton
and the physical world, and the humanities, which are traditionally concerned with the nature of human experience, and to use this bridge for the improvement of mentally and neurologically ill patients and for the general betterment of humankind" (1999:215). And to take just one example from the other end of the disciplinary spectrum, the media theorist Barbie Zelizer writes in a review paper that "in pace with the constitution of the social sciences themselves . . . . the study of collective memory has virtually erased disciplinary boundaries" (1995:216). Obviously, it's not going to be easy. Introducing a pathbreaking collection of interdisciplinary essays on memory distortion, Fischbach and Coyle demand that "we must crawl into the [black] box and try to understand the brain as it functions in the context of human affairs. This requires that the advances in neuroscience be paralleled by equally creative thinking at the level of psychology and the behavior of societies" (1995: xi). Faced with this requirement, philosophical pessimism may seem appropriate. In a detailed examination of the pitfalls of interdisciplinary theory-construction, Patricia Kitcher uses a historical case study to list "subtle and not so subtle dangers" in moving too fast between disciplines and discourses. Just as did Freud, she argues, so contemporary cognitive scientists often exhibit too much faith in the resources of a neighbouring discipline, disregarding the seriousness of its internal problems; or they take the coherence and harmony of two theories as conclusive evidence for the truth of both (Kitcher 1992: 159-161, 172-174, 180-183). Kitcher's "purpose is not to argue against the interdisciplinary perspective," but she believes that we need to achieve "a more critical understanding of this necessary and exciting project m and so to proceed more cautiously" (1992: 5, 219). Well, perhaps: recent gatherings of memory researchers from different backgrounds may have aided dissemination of specialist work, but don't seem to have left the participants with the desired glowing sense of impending breakthroughs in interdisciplinary theory-construction. 4 But a start must be made somewhere, and occasionally a messy preference for proliferation over prudence in difficult domains may pay off. Kitcher sets up her study of Freud's interdisciplinary science of mind by distinguishing three plausible methods for assessing the progress and success of such grand joint ventures. We could "consider the logical form of interdisciplinary
4 I'm aware of five such conferences and series in the last few years: two in Cambridge, MA, on memory distortion in 1994 and on memory and belief in 1997 (see Schacter 1995a; Schacter & Scarry 2000); the Darwin College lectures in Cambridge, England in 1996 (see Fara & Patterson 1998); a C.N.R.S. "Cross-Disciplinary Encounters" series on memory in Paris in 2000 (see http://www.cnrs.fr/cw/en/pres/compress/memoire/sommaire.htm cited at 7 March 2004) and a University of London series on memory in science in 2000. Each of these series included at least one session on neural bases of memory, and at least one session on memory in sociology or history.
Representation, Reduction and Interdisciplinarity
191
theories and arguments"; try to assess current interdisciplinary work in cognitive science directly; or abstract away from current research to gain a better sense of "the potential strengths and weaknesses of interdisciplinary methodology by examining a historical case" (Kitcher 1993: 4). I agree entirely with her view of the utility of what might be called a historical cognitive science. 5 But my method here is the direct description of interdisciplinary approaches in cognitive science. I don't dispute Kitcher's worry about this method, that "it is perhaps somewhat early to judge the effects of interdisciplinary integration on cognitive science," but I don't see the harm in trying: and, by restricting attention to memory, I hope also to open up the possibility of comparing its potential susceptibility to interdisciplinary analysis with that of other cognitive domains. For example, interesting comparative analyses might be made between the case of memory and the study of colour vision, on the one hand, and of dreaming, on the other. My hunch is that the current sciences of memory look less successfully interdisciplinary than the case of colour vision, but more successfully so than those of dreaming. If this is the case, it could be due either to the peculiar history and developmental course of the relevant sciences, or to something peculiar to the particular phenomena in question. Like the present project, this would be work in the philosophy of cognitive science, seeking a synoptic vision by trying to immerse in some detailed areas of some of the contributing sciences in order to motivate debate, and enlisting the help of experts to point out the gaps in such motivated appropriation of their languages. 6
Notes on Reduction and Interdisciplinarity There already are partial, imperfect, but promising interdisciplinary research programs within the cognitive sciences of memory. What kind of interdisciplinarity
5 In Philosophy andMemory Traces: Descartes to connectionism (Sutton 1998), I tried to do philosophy of cognitive science and historical cognitive science simultaneously,by offering a combined history and defence of the notion of distributed representation in models of memory. The idea was to undermine different patterns of resistance to interdisciplinarity, in the form both of humanists' resistance to cognitive science, and of scientists' resistance to culture and history. But the project was almost entirely underlabouring, and the positive vision of a cultural cognitive science of memory was hardly even sketched there. 6 So, as should be clear by now, this is not a project in the philosophy of memory as traditionally practiced. The abstract epistemologicalquestions about the very existence of memory traces, about representations and realism, and more recently about self-knowledge and externalism which have characterised Anglophone philosophical discussions of memoryhave their own importance, but are of little interest to most cognitive scientists. For introductions see Warnock (1987) and Sutton (2002).
192
John Sutton
do I have in mind? Again, I agree with Kitcher: in difficult and immature domains like this, "it is often extremely difficult to reduce a methodological approach [that of interdisciplinarity] to a set of logical relations" (Kitcher 1992: 4). So I ' m not trying to decide in just what respects the cognitive sciences of memory, despite harnessing the vast apparatus of Kuhnian "normal science," are nevertheless paradoxically pre-paradigmatic. 7 Instead I argue that the apparently universalizing urge to connect quite different languages in the sciences of memory does not commit me to an implausibly strong kind of reductionism, or to any classical, microreductive unity-of-science doctrine. By simply mentioning the availability of a range of alternative, weaker approaches to interdisciplinary theory-construction and to reduction in cognitive science, I hope to ward off the immediate concern that my quest will inevitably impose an artificial unity on complex and highly diverse phenomena. After a general discussion of reduction, I'll look at a particular strand of the sciences of memory which is used by Valerie Gray Hardcastle to exemplify interdisciplinary theory-construction.
Interdependent Phenomena and Patchy Reduction John B ickle takes some pleasure, introducing his persuasive defence of a "newwave" reductionism, in quoting Jaegwon Kim's remark that being a reductionist nowadays "is a bit like being a logical positivist or a member of the Old Left: an aura of doctrinaire naivete hangs over such a person" (Bickle 1998: 1, quoting Kim 1989). Fighting over the word "reduction" is often, and perhaps should only be, a political activity. It's not at all obvious whether the particular interdisciplinary research on memory which I discuss should be seen as "reductionist" or not. My perspective here is intended to be broad enough to encompass both avowed "new-wave" reductionists like Cliff Hooker, Patricia and Paul Churchland, and John Bickle, and those who claim to reject reduction while encouraging detailed engagement in interdisciplinary programs, like Hardcastle and Kitcher. We can see why the retention or rejection of the term may not matter (at least for current purposes) by examining Kitcher's criteria for interdisciplinarity (1992: 6-7). 8
7 On cognitive science in general as a case study in the philosophy of science, see the rigorous notion of "research frameworks" developed by Barbara von Eckardt in her remarkable book What is Cognitive Science? (1993: 13-56, 345-396). Von Eckardt has recently suggested that cognitive science is "still a mere babe in the woods of science" (1999:221). 8 Where Kitcher's three options for assessing interdisciplinary theory-construction, discussed in the first section above, were merely methodological, the three kinds of interdependence she envisages here are metaphysical.
Representation, Reduction and Interdisciplinarity
193
She argues that the idea of mutual influence across apparently separate disciplines requires, minimally, an "assumption that the phenomena dealt with by the different disciplines are in some way interdependent." Such interdependence, for Kitcher, can come in at least three varieties: reduction, which she characterizes classically as involving derivation of reduced laws from reducing laws plus biconditional or bridge laws; supervenience, by which "the properties captured in one science alter if and only if there is some change in the properties characterized by a second science"; or a still weaker kind of genuine interdependence. This last variety of interdependence between phenomena holds "just in case some changes in the properties mentioned by one science alter a significant range of properties mentioned by the second" (1992: 6-7). Yet it is around something like this last kind of weaker, local relation between theories that the "new-wave" reductionists build their more liberal view. 9 Whatever is implied by these preferred accounts of intertheoretic relations, whether characterized as interdependence or reduction, they are definitely not committed to two extreme claims sometimes ascribed to any substantive reductionism. They do not require the complete purging or elimination of all higher-level terms and concepts from lower-level, "reducing" explanations. And they do not imply that ultimately only the vocabulary of fundamental physics will be legitimate, l~ Neurobiological theories retain terms from "higher-level" theories: Kandel's account of associative learning in the sea slug Aplysia, for instance, employs ineliminably psychological terms such as habituation and sensitization. But this doesn't entail that such theories are mere implementations, rather than genuine reductions. 11 "New-wave" reductionists can agree that the causal generalizations of theories like Kandel's are, as Schaffner puts it, "typically not framed in purely biochemical terminology," but still argue that some such theories genuinely describe relations of interdependence between phenomena at different levels. Realistic interesting intertheoretic relations will allow at most for partial reductions. Genuine explanations will produce "many weblike and bushy connections" across levels, with causal sequences described at many different levels of aggregation. In biological
9 The key source for this tradition is Hooker (1981). See also, in addition to the sources cited by Bickle (1998), Churchland & Churchland (1998). 10Both of these positions are ascribed to "radical" reductionists by Gold & Stoljar (1999). For a detailed response see my commentary (Sutton 1999). See also the treatment of long-term potentiation in Stoljar & Gold (1998). II As Gold and Stoljar suggest in a detailed discussion of Kandel: but they don't deal with alternative interpretations of Kandel's work by "new-wave"reductionists, instead assuming an implausibly narrow classical picture of reduction. See Schaffner (1992) and Bickle (1995, 1998: Chapter 5).
194
John Sutton
cases especially, the generalizability of local reductions may be limited, with some perhaps being specific to the particular system under investigation. So if there is reduction, it is "bound to be patchy" (Schaffner 1992: 337). So the serious problem posed to classical microreductive views by the possibility of multirealizability (by which many different lower-level, reducing terms or laws might in different contexts realize the same higher-level terms or laws) does not bite here: interesting reductions are often going to be heterogeneous (Bickle 1998: Chaps 1 and 4). It's just because of this lower-level heterogeneity that many reductions are revisionary, with the explanations of the higher-level theory either merely approximating the reducing explanations, or being fragmented into many different lower-level explanations (Bickle 1998: 200-201). This is how we have to get what Paul Churchland calls "objective knowledge of a highly idiosyncratic reality" (1996: 306). So Bickle sees Kandel's account of associative learning, with its intentionalist vocabulary of surprise, expectation, and predictability, as genuinely reductive just because of the proposed combinatorial dynamics by which "a few fundamental processes of cellular plasticity, internal to individual neurons, are theorized to occur in a variety of sequences and combinations to produce complex forms of behavioral plasticity" (Bickle 1995: 268).
Interdisciplinary Theory in the Sciences of Memory To engage in even this much metaphysics is to go beyond the interests of many cognitive scientists. As Endel Tulving, a leading cognitive psychologist of memory, says, "our research community as a whole does not perceive much value in conceptual analysis; there is no promise of social reinforcement for any single individual who might be attracted to the enterprise" (2000: 34). Tulving of course is, from his position of eminence, recommending new attention to conceptual clarity: but interdisciplinarity may in practice be furthered more by the pursuit of particular empirical cases. In this spirit, I outline a discussion of the psychological distinction between explicit and implicit memory by the philosopher Valerie Gray Hardcastle, who argues that it exemplifies (historically and conceptually) the mutual interdependence between disciplines which she sets at the heart of the cognitive scientific enterprise (Hardcastle 1996" Chapter 6). 12
12The exampleis motivatedby Hardcastle's desire to display particularconnections betweenpsychology and neurosciencewhich do not entail the strict explanatorydependence of the former on the latter. For this reason she thinks "we need to stop arguing over the reductionisrn/antireductionism issue" (1996: 104). Carl Craver is developing an independent account of robust intertheoretic strategies in neuroscience: see Craver & Darden (2001) and Craver (2002).
Representation, Reduction and Interdisciplinarity
195
Hardcastle's case study aims to describe a "strongly interdisciplinary approach," which carves out an explanatory framework not just by relying on evidence from more than one area, but by actively accepting the underlying assumptions of those areas. If successful, the case implicitly answers concerns like Kitcher's about the overhasty importing of alien methods and concepts: "we may use another discipline for collateral support, inspiration, and to help set the parameters of inquiry, but we cannot simply borrow data wholesale from other theories over the same state space" (Hardcastle 1996: 106, 110). The distinction in question, at a first pass, divides explicit or aware remembering from nonconscious influences of past experience on perception, thought, and action. The idea that there are two independent systems in play here is a substantive hypothesis. 13 My concern here is not with its plausibility, but with rehearsing Hardcastle's claim that it is genuinely interdisciplinary. Initial evidence for the explicit/implicit distinction came from the study of amnesia in clinical neuropsychology. The patient H. M. and others proved able to learn new skills (such as solving simple puzzles), and to remember how to engage in certain procedures, even while explicitly denying having experienced them before, and showing no conscious awareness of even recent events (Hardcastle 1996: 112; compare Schacter 1996: 137-142; Squire & Kandel 1999: 11-16; and, on H. M., Hilts 1995). Secondly, various strands of developmental inquiry suggest the existence in some domains of facilitated processing before, and independently of, explicit knowledge: face recognition, for example, seems to operate automatically from birth in a context-bound system which is later either replaced or augmented by the capacity to achieve explicit learning of specific faces (Hardcastle 1996:106-111, with references). In turn, cognitive psychologists have designed tests to tap explicit and implicit memory independently. Whereas performance in explicit recognition of previously-presented words, for example, diminishes severely over a week, the ability successfully to complete fragments of words (such as adding "LE" to "TAB__" after previously being shown "TABLE" in a list) declines much less rapidly. The suggestion is that since this facilitation of performance due to previous exposure varies independently of explicit recognition, it is the work of a distinct implicit priming system (Hardcastle 1996:114). Hardcastle goes on to describe the varying relation of each of these three traditions (in neuropsychology, developmental psychology, and cognitive psychology) to current neurobiological attempts to localize memory systems in animals,
13 I'm not here discussingthe general question of the logic and criteria for postulating and identifying separate memorysystems.For a concise account see Schacter& Tulving(1994).
196
John Sutton
and to recent neuroimaging studies. There are, as she notes, specific puzzles about each strand of evidence here. To what extent, for example, are the implicit memory systems modality-specific? And just how complex and apparently "cognitive" can implicit memory be (compare Schacter 1996: 187-191)? Is it an essential property of implicit systems that their representations can't be manipulated and controlled? Again, rather than evaluating her own preferred synthesis of the evidence, 14 I address Hardcastle's moral: an interdisciplinary theory will include both a general set of principles which apply to phenomena across several different objects of study (different animals, for example), and a series of specific models from which these general principles were derived. The former may be too general to drive specific predictions for any experimental procedures; the latter may be too particular to generalize across many organisms (Hardcastle 1996: 132-133, 138). This is obviously far from a classical reductionist vision: As a true discipline still in the making, cognitive science is an amalgam of different disciplines, with each bringing incompatible research techniques and traditions to bear on a common set of interests. We have no reason to expect that the different domains will ultimately fuse into one; each could retain its autonomy as they address a set of central concerns, linked instead by a fuzzy set of entity attributes across a family of models and certain core framework principles (Hardcastle 1996: 135). The power of Hardcastle's account lies in this unusual combination. On the one hand she accepts the ongoing multiplicity of disciplinary perspectives; but on the other hand she argues that memory scientists can, and often do, use these different perspectives to address just the same event space, in compiling single explanations which are themselves multi-levelled. When the objects and events under investigation are both complex and idiosyncratic, the best explanations may have to be partial or patchy in order to connect to a more general theory which links different but related phenomena. Whether this is seen, with Hardcastle, as a failure of the old reductionist urge to find "a single overarching account"
14Hardcastle argues for a revised terminology for the two systems, suggesting "structural episodic memory" for the implicit, task-specificresponses, and "semantic episodic memory"for explicit, crossmodal memory(1996:131-132). This is controversial: one mightfor examplebe uneasyaboutincluding implicit memoryas a varietyof episodic memory.But this doesn't affectthe metatheoreticalpoint about interdisciplinarity.
Representation, Reduction and Interdisciplinarity
197
(1996: 139), or, with Bickle and the Churchlands, as a triumph of new-wave reductionism, is unimportant: in either case, we have at least the possibility of seeing different (sub)disciplines as dealing with interdependent phenomena.
Constructive Remembering: Source Memory and Development I have suggested that issues about reduction and interdisciplinarity arise, and are already being fruitfully addressed, within the cognitive sciences of memory. In this section I make the further case that key lines of research within those sciences point to the relevance of factors outside the individual, and of explanatory connections between cognitive and social psychology. My examples are from two strands of an emerging consensus about constructive remembering in cognitive and developmental psychology, in work firstly on source monitoring and source amnesia, and secondly on the development of autobiographical memory. In both cases, larger-scale historical or cultural dynamics may be relevant in understanding both distortion and accuracy in individual memory in both adults and children. Cognitive psychology can't neglect the interface between brain and world. In the context of the distinction between explicit and implicit memory discussed in the previous section, the focus here is primarily on certain forms of explicit or declarative memory. I'm looking, specifically, at autobiographical episodic memories, explicit experiential memories of past events and episodes in a personal history. Consensus on a definition of episodic memory, sometimes called "personal memory" by philosophers, is surprisingly hard to find: we can work for now with William Brewer's account of "recollective memory" as "a reliving or reviving of my own past phenomenal experience, with the additional knowledge that I've had that experience before" (1996). 15 It's in episodic memory that we achieve a form of "mental time travel," in which we're oriented to events as occurring at particular past times, events which we sometimes knit into autobiographical narratives (Suddendorf & Corballis 1997; Tulving 1983, 1993, 1999). While Brewer's definition leaves open a number of key questions which I
15One initial respect in which this seems too strong is in the requirement that I know I've had the experience before. Most cognitive psychologists abstract away from the folk use of "remember" as a "success-word," and study the varieties of false memory, thus including cases in which I mistakenly believe I've had the experience before.
198
John Sutton
bypass here, 16 we can focus on clear core cases of episodic remembering, to pick out certain aspects of current cognitive-psychological approaches.
Constructive Remembering "A variety of conditions exist," declares Daniel Schacter, "in which subjectively compelling memories are grossly inaccurate" (1995b: 22). Many cognitive psychologists now accept, at least in principle, that explanations of even relatively simple memory phenomena will invoke multiple causes. This recognition is of course hard to put into practice. But the recovered-vs-false-memory crisis of the 1990s encouraged academic psychologists to shelve disputes between "ecological" and "lab" approaches to memory, and to forge a swift consensus on the multicausal and constructive nature of remembering, in opposition to extreme views on the possibility of total and exact recovery of long-repressed memories. 17 There are two related aspects of this relatively recent consensus in cognitive psychology (for reviews see Roediger 1996; Roediger & McDermott 2000; Schacter 1999). Firstly, attention is focussed increasingly on the context of recollection, rather than solely on encoding or on the nature of encoded traces. Realistic representational theories of memory do not assume that recall is fully determined by the nature of the stored representation. On Tulving's notion of "synergistic ecphory," traces (whatever they may be) are "merely potential contributors to recollection": the engram is not the memory, and instead "the [current] cue and 16 Two issues of considerable interest concern time in memory, and the relations (both developmental and conceptual) between episodic memory and autobiographical memory. John Campbell argues that episodic memory presupposes a particular and objective conception of time as linear, which grounds our ability to locate autobiographical episodes as at particular past times (Campbell 1994: Chapter 2, 1997; Hoerl 1999). Not all autobiographical memories are episodic, since I can (non-experientially) remember facts about my life (such as my date and place of birth), using only semantic memory. But the converse question, whether all episodic memories are autobiographical, remains open. My hunch is that any putative consensus on the nature of episodic memory will show it to be, both conceptually and sociologically, an intrinsically interdisciplinary construct, with, in particular, the methods and results of cognitive ethological research on animal memory needing to be integrated: see for example Campbell (1994: 37-41, 64-71), Tomasello (1999: 124-125) and Dennett (2000). 17 This speculation on the sociology of recent cognitive psychology needs careful historical support. For the earlier battles between "real-life" and cognitivist opponents see for example Middleton & Edwards (1990), and special issues of American Psychologist 46 (1991) and The Psychologist 5 (1992). A later, ecumenical treatment is Koriat & Goldsmith (1996). Ian Hacking's brilliant Rewriting the Soul (1995) was perhaps too early to catch the effects of the recovered memory movement on academic theories of memory.
Representation, Reduction and Interdisciplinarity
199
the engram conspire" in any act of episodic remembering (Schacter 1996: 56-71, 105; 1982: 181-189; Tulving 1983: 12-14). So, as Susan Engel puts it, "one creates the memory at the moment one needs it, rather than merely pulling out an intact item, image, or story," so that specific features of any context of recall may be direct causes of the content and format of the memory-as-retrieved (1999: 6). Secondly, cognitive psychologists have come to accept more flexible and dynamic pictures of long-term "storage." This is partly due to the influence of connectionist or Parallel Distributed Processing (PDP) models, but also to a rich array of studies on misinformation, bias, and the role of schemas in memory, some of which I discuss below. This internal plasticity is one of the most curious and characteristic features of human memory, and one which clearly differentiates our cognitive systems from current digital computers. It's pretty useful for the contents of our files to remain exactly the same from the moment I close them at night to the moment I open them again in the morning. But various kinds of reorganization and realignment often happen to the information retained in my brain over the same period. In us, many memories do not naturally sit still in cold storage. This consensus about the constructive nature of remembering needn't be unrealistically overdescribed. It's not that accuracy and reliability in memory are suddenly shown by science to be impossible or unlikely. Rather, the assumption is that understanding of the mechanisms of distortion will also illuminate the processes operating in veridical remembering (Mitchell & Johnson 2000: 179-180). Neither "accuracy" or "reliability" are transparent notions in this context, and "truth" in memory, though not forever inaccessible, is neither a single nor a simple thing. Verbatim recall and other forms of exact reproduction are rarely necessary for success in remembering (Rubin 1995).
Constructive Memory and Source Monitoring The "source monitoring" framework developed by Marcia Johnson and her colleagues exemplifies a method of "experimental phenomenology" which takes subjective judgements and feelings as core explananda for cognitive psychological theory. "Source monitoring" is the process of recalling when, where, and how some information was acquired (Johnson et al. 1993; Johnson & Raye 2000; Mitchell & Johnson 2000). Johnson argues that there is typically no tag or label on a memory to specify its source. Instead, activated contents must be "evaluated and attributed to particular sources through decision processes performed during remembering" (Johnson et al. 1993: 3). Our subjective experience of autobiographical remembering, it's suggested, depends on our source attributions: when some content doesn't
200
John Sutton
have the right set of qualitative characteristics, it's likely to be experienced not as something I remember through personal experience, but just as something I know (Rajaram & Roediger 1997). 18 These attributions are not determined by properties of the "memory records" alone, but are influenced also by a range of aspects of the current context, including biases, motivations, current agendas, attention, stress, metamemory assumptions, apparent compatibility with other knowledge, and social setting. Johnson suggests, further, that the ease of attribution of a content to a specific source is one "compelling reason for experiencing it as belonging to our personal past," thus offering another promising line of enquiry into the multiple roots of autobiographical remembering (Johnson et al. 1993:21). As a theoretical framework subsuming a variety of experimental results on memory distortion, misinformation, eyewitness testimony and suggestibility, the source monitoring research is driven by evidence from a range of (sub)disciplines. As well as cognitive psychology, it draws on cognitive neuroscience (both clinical studies of frontal lobe lesions, and neuroimaging investigations), and on personality, developmental, and social psychology. In each case, the project is to study conditions under which source monitoring breaks down, in confusions of (for example) reality and fiction, internal and external sources, trustworthy and unreliable sources, or perception and imagination. In each case, contents become unglued from their origins and misattributed to another source or context. There are highly dramatic instances, as when the psychologist Donald Thomson was accused of rape when in fact the victim had been watching him being interviewed on TV prior to the rape; but related processes operate in our mundane tendencies falsely to "remember" seeing a thematically-related word such as "sweet" in a presented list of words like "candy, .... sugar," and "taste" in which "sweet" did not in fact appear (Schacter 1999: 188-190, with references). 19 Source forgetting drives the generation of plausible, but in such cases incorrect, attributions. There is no sharp line between the normal, fallible but functional operation of source monitoring processes, and more pathological and confabulatory contaminations and intrusions. Some forms of source monitoring are harder for children, and develop at different stages; some forms are more difficult for older adults. Some individual difference factors (including imaging ability, hypnotizability, and dissociative tendencies) may be related to particular kinds of
18There are various intermediatepossibilities, in which subjects feel they are neither straightforwardly "remembering," nor just "knowing," yet they somehow have access to the information in question. 19Schacter also includes ordinary "cryptomnesia" and unintentional plagiarism, in which we are remembering but do not subjectively feel that we are, because we misattribute to our own imagination a thought or idea which in fact derives from a specific prior experience.
Representation, Reduction and Interdisciplinarity
201
source confusions. 2~ But misattribution is not always unadaptive: indeed, normal subjects apparently produce false recognitions in certain contexts more often than do amnesics with medial temporal lobe damage (Schacter 1999: 190). If correct, this intriguing evidence suggests that accurate source monitoring is often less important than the kind of generalization on the basis of similarity and theme which (in normal subjects) may "give rise to distortions as an inherent byproduct" (McClelland 1995" 84). 21 Susan Engel may overstate a little in writing that "it is the norm rather than the exception to be unable to distinguish between what happened, what you feel about what happened, and what others may have said about what happened" (1999: 16): but the source monitoring framework at least suggests that understanding the conditions in which such confusions occur will require attention to a large range of affective, motivational, temperamental, and social contextual factors which influence the mapping of information to source. The idea that cognitive psychology studies the individual mind, leaving social processes to be treated by the social sciences, thus looks untenable. The creation of a false memory may often be explained neither by an overconfident therapist alone, nor by a purely internal process. As Janice Haaken argues, the source monitoring framework requires us to discuss the recovered memory crisis without either overconfident individualism, or a focus on "memory deceivers on the other side of the fence": instead, we have to trace the subtle sedimentations of culture in some individual memories. "The source of a remembrance," Haaken concludes, "may not be readily or immediately located in discrete events in an individual past but rather may be found in the complex web of converging group experiences" (1998:111).
Autobiographical Memory Development and Scaffolding Susan Engel's 12-year-old son, working on a class writing assignment, looked seriously at his mother before asking "Mom, what is my most important memory?" (1999: 24). Autobiographical memory develops in a shared environment, in which its content as well as its expression is influenced by that interpersonal and cultural context. Developmental studies are a rich and flourishing area within the sciences of memory. Despite great variety in methods and theoretical assumptions, most 20See the special issue of Applied CognitivePsychology12 (1998), "Individual Differences and Memory Distortion." 21 This point is also relevant to philosophical accounts of the role of memory in maintaining continuity of personal identity over time. Where traditional theories of personal identity focus on current contact in memory with specific episodes in the personal past, it may be, as Marya Schechtman argues, "precisely insofar as our memories smooth over the boundaries between the different moments in our lives.., that we are able to produce a coherent life history" (1994:13).
202
John Sutton
schools of developmental thought are thoroughly interdisciplinary, calling to differing degrees on neuropsychology and social psychology as well as cognitive psychology; and most accept, at least in some explanatory contexts, the significant causal influence over time of the remembering environment. Children learn to remember in company. Children start talking about the past "almost as soon as they begin talking," but the form of their references to past events takes some years to develop (Nelson & Fivush 2000: 286). 22 At early stages, adults provide much of both the structure and the content of young children's references to the past, providing "scaffolding" for the children's memories. Where initially children use genetic event memories implicitly, like scripts, as a basis on which to understand routines and generate expectations, they gradually develop the ability spontaneously to refer to specific past episodes with rich phenomenal content. Although there are a number of different theories of how these changes unfold, with the relative roles of language, temporal representation, theory of mind and metarepresentational capacities, and self-schemata being as yet uncertain, joint reminiscing is a key part of the process (Howe & Courage 1997; Perner & Ruffman 1995; for integrative discussion see Welch-Ross 1995). Children gradually develop perspectival temporal frameworks in which to locate memories of idiosyncratic events. Memory sharing practices, often initiated by adults, encourage the idea of different perspectives on the same once-occupied time (McCormack & Hoerl 1999). In developing this temporal perspective-switching, children start to take memories as objects for negotiation, shared attention, and discussion. Realization of the existence of discrepancies between versions of the past goes along with the development of some kind of self-schema, as children begin to collect stories into some kind of personal history. The ability to view one's life retrospectively is sophisticated, and follows adult guidance in simpler conversations about the past. On the social interactionist picture of the development of memory, defended by Robyn Fivush, Katherine Nelson, and others, parental and cultural models or strategies for the recounting of past events act as initial scaffolding on which children start to hang their own memories. They then internalize the forms and narrative conventions appropriate to their context, so that "early reminiscing begins as an interpersonal process and only becomes intrapersonal over time" (Engel 1999: 27; see for example Fivush 1991, 1994; Nelson 1993). The point is not that children can't remember in solitude, nor that they remember only what they talk about, but that both shared
22There are of course also a range of memoryphenomenabefore language:for surveysof the methods used to study memoryin infants, and current thinking about the results, see Mandler & McDonough (1997) and Rovee-Collier& Hayne (2000).
Representation, Reduction and Interdisciplinarity
203
and inner reminiscing may alter the form and the content of autobiographical memories. Variations in narrative practices may then reappear in the subjective idiosyncrasies of early remembering. Cultural variations in the nature and contexts of talk about the past, and intracultural variation in the motivations for and richness of remembered narratives, have been convincingly demonstrated. 23 It is not clear whether these variations have longer-term effects, nor how narrative practices interact with the simultaneous development of self-schemata and of a theory of mind. But my comment on this work is simpler: I want to insist on the possibility of strong interaction between individual and shared reminiscence without assuming that the developing internal representations are a straightforward projection of the shared narratives. Fivush and her colleagues occasionally write as if the format of autobiographical memory is itself linguistic or language-like, as if children simply incorporate the forms and contents of local external narratives: following Vygotsky, Fivush argues that "the narrative forms that children are learning to organize their recounting of past experiences are also used for organizing their internal representations of past experiences" (1994: 138). This is possible, but the argument slides too quickly: we can accept that what Peggy Miller calls the "distribution of storytelling fights" in a culture or in a family may strongly influence the uses and the contents of individual memories (Miller et al. 1990), without having to assume that either the format or the organization of those individual memories is literally linguistic or narrative. In the final section of the paper, I suggest that this same situation crops up more generally in the sciences of memory. We want to allow a variety of ways in which individual and cultural processes and representations can interact, complementing or conflicting with each other, while retaining asymmetry between those internal and external representations. Social norms and cultural narratives are not simply downloaded into the mind, and yet each mind is more intimately tangled with such norms and narratives than traditional individualistic cognitive science has often allowed.
Cognition and Culture: Collective Memory and External Memory Traces Inside and Outside the Mind The work I've described in cognitive and developmental psychology is as yet unintegrated with the central projects of cognitive science, in that it proceeds
See for example the studies of "elaborative" and "pragmatic" reminiscing styles by Reese et al. (1993) and Welch-Ross (1997); and of psychological effects of cross-cultural linguistic styles by Mullen & Yi (1995) and MacDonald et al. (2000).
23
204
John Sutton
without substantive assumptions about the nature of mental representation and computation. But the experimental evidence gathered in these subdisciplines must be compatible with a theory of mental representation: the constraint works both ways. So if I'm right that such central parts of the cognitive psychology of memory as these already point to the relevance of factors outside the individual, there are two tasks for the wildly optimistic interdisciplinarian to pursue. Firstly, different approaches to mental representation within cognitive science must be related to, and tested against, this data and these cognitive-psychological frameworks. Secondly, we must find vocabularies and methods for connecting the cognitive and the social sciences of memory. In this final section, I address this second task: this requires dropping neutrality between cognitive-scientific theories, and applying to the case of memory some recent work in cognitive anthropology and embodied cognition. The aim is to find a general framework for the sciences of memory in which the concepts of social or "collective" memory, and of "external memory," will be integral parts of cognitive science, rather than social constructionist myths or humanistic curiosities. If memories are not stored independently, permanently, and explicitly within the individual mind (but are, for example, superpositionally retained as dispositions of the connection weights of neural networks), then the relatively unstable individual memory needs support from more stable external scaffolding or props. One theoretical goal might be a broad metaphysics of traces inside and outside the mind (perhaps along the ambitious lines suggested by Leyton 1992). Such a picture would not collapse the distinction between internal and external representation and processing, but would provide a framework for investigating whether and how our interaction with different forms of external information systems might in turn affect the format and processing of individual memories. Culture and technology are products of cognition and action, but in the human case, as Merlin Donald argues, such products in turn "have direct effects upon individual cognition" (1991:10).
Collective Memory Some explanations in the social sciences will refer, among multiple causes, to (appropriately flexible) internal processes of schematization or reconstruction. And some explanations in the cognitive sciences will refer to the transformations of external representations. The resources of situated cognitive science can be put to work in forging fruitful notions of collective and external memory. There's widespread scepticism about the very idea of "collective memory." Naturalists are uneasy about the taint of Jungian archetypes, or morphic resonance. And recently
Representation, Reduction and Interdisciplinarity
205
even sociologists and historians who work on what seem to be social memory phenomena have tried to disclaim the notion. The historian Pieter Lagrou (2000) suggests that the term "memory" amalgamates "memory proper, which is an inalienably individual capacity, and "memory" as a metaphor, in an anthropomorphism that is often not conscious, for the entire set of current representations in a community." James Young, working on memories of the Holocaust, prefers to use the term "collected memory" instead of "collective memory," because "societies cannot remember in any other way than through their constituents' memories" (1993: xi). 24 In discussing the work of the great sociological theorist of collective memory, Maurice Halbwachs, Fentress and Wickham worry that his concept of collective consciousness was "curiously disconnected from the actual thought processes of any particular person," leaving later sociological accounts with the danger of treating the individual as "a sort of automaton, passively obeying the interiorized collective will" (1992: ix-x). This widespread embarrassment is understandable among sociologists and historians seeking explanatory models which are both flexible and naturalistic. But it isn't necessary. Although not my aim here, it's possible to find a more subtle, and more plausible, account of the relations between individual and collective representations in Halbwachs' own work. 25 It's true that he was highly critical of the individualism of psychological theory between the wars, but his positive, anti-individualist views do not rest on mystery. For Halbwachs, briefly, "there is no point in seeking where memories are preserved in my brain or in some nook of my mind to which I alone have access: for they are recalled to me externally" (1925/1992: 38). 26 The people and groups around me normally "give me the means to reconstruct them?' Collective frameworks of memory are not the simple product of isolated individual memories, constructed after the fact by combinations of separate reminiscences, but are rather, in part, their source, the instruments used in the particular acts of recall. There's a sharp contrast, argues Halbwachs in an intriguing chapter on dreams, between remembering and the actual "state of isolation" of the dreamer,
24Compare the odd locution of Jacques Le Goff (1992: 53): "at a metaphorical but important level.... the absence, or voluntary or involuntaryloss, of collectivememoryamongpeoples and nations can cause serious problems of collectiveidentity. The connections between differentforms of memory may also be not metaphoricalbut real in character." 25 Another compatible project here would be the application to memory of recent work in social ontology,which seeksto naturalizenotions like "joint action" and "mutual knowledge."See for example Tuomela (1995). But my argumentabout memory doesn't depend on the details of any particular such project. 26The view that Halbwachs simply neglects psychology is unfortunately and erroneously supported by this translation, which simply omits the bulk of the early chapters of Halbwachs' work, which cover dreams, language, constructivememory, and the localization of memories.
206
John Sutton
who isn't capable directly of reliance on the frameworks of collective memory: "it is not in memory but in the dream that the mind is most removed from society" (1925/1992: 42). 27 So in ordinary remembering, which is either actually or potentially shared, it is the public scaffolding of various forms, in the physical, symbolic, and mnemonic environment, which triggers the specific form and content of individual memory. The contemporary sociologist Paul Connerton puts the point strongly: "it's not because thoughts are similar that we can evoke them; it is rather because the same group is interested in those memories, and is able to evoke them, that they are assembled together in our minds" (1989: 37). Blunt claims like these no doubt strike many readers as not simply antiindividualist, but as anti-cognitivist: how can there be a cognitive science of memory which includes or even allows for such a displacement of explanatory relevance from the individual mind/brain to the natural or cultural world? The assumption to be questioned is that cognitivist, information-processing commitments automatically rule out notions of external or collective memory. 28 I'm going to make this case, and the stronger case that certain cognitive theories specifically require attention to external and collective memory, by direct engagement with the implications for memory science of recent work on the "embedded" and "extended" mind. But first I make two preliminary point about current movements in the social sciences of memory.
Collective Memory and Schema Theory Firstly, there is ample naturalism in history, anthropology, and sociology, in studies of memory which are compatible with, even if not actively constrained by, current work in cognitive science. For example, Michael Schudson has classified forms of collective memory, arguing for distinctions between three kinds: socially mediated individual memories, cultural forms and artifacts which hold and interpret the past for social mediation, and individual memories which are in turn constructed from the cultural forms (1995; compare Zelizer 1995; Assmann 1 9 9 7 : 1 - 2 2 on "mnemohistory"; and the review by Olick & Robbins 1998). The main current concern for a naturalist about this social scientific work should not, I suggest,
27 The point of the contrast is to argue that memories do not endure unchanged in an unconscious state, for later autonomous extraction and manipulation by either the dream or the will: Halbwachs doesn't deny the role of memory in dreams, which are "made of fragments of memory mutilated and mixed up with others," but underlines the absence of interpersonal contact and comparison in dreaming. 28Though not applied to memory, this general point is the core of Ronald McClamrock's book Existential Cognition (1995).
Representation, Reduction and Interdisciplinarity
207
be the danger of seeing collective memory as floating free of individuals, but rather the temptation to overlinguisticize the form of the internal representations which construct and are permeated by collective memories. Halbwachs sometimes wrote like this, arguing for instance that "one cannot think about the events of one's past without discoursing upon them" (1925/1992: 53), and the powerful influence of Russian psychology on Anglophone developmental theory often has the same result (Bakhurst 1990). But this is to project too quickly the format of external, expressed memories back inwards onto internal memory. But this linguistic-constructivist conception of mental representation, which is at odds with the post-connectionist cognitive science I describe below, is not necessary: my memories can be called forth socially, moulded and formed by external influences, without having themselves to be, in their internal aspect, linguaform. Secondly, the history of the concept of "schema" instructively reveals psychologists and cognitive anthropologists struggling to find a vocabulary for relations between internal and external memories which neither collapses the distinction nor sees the internal as simply the reflection of the social. When Bartlett imported the term into the psychology of memory from neurophysiology, he worried about its implications of stasis: I strongly dislike the term 'schema'. It is at once too definite and too sketchy . . . . It suggests some persistent, but fragmentary 'form of arrangement', and it does not indicate what is very essential to the notion, that the organised mass results of past c h a n g e s . . , are actively doing something all the time (Bartlett 1932: 201). 29 So a "schema" is not, in one sense, a definite cognitive structure at all, and Bartlett makes the same point in relation to the notion, closely related in his system, of a memory trace: Though we may still talk of traces, there is no reason in the world for regarding these as made complete, stored up somewhere, and then re-excited at some much later moment. The traces that our evidence allows us to speak of are interest-determined, interest-carried traces. They live with our interests and with them they change (1932: 211-212). 30
29See also Bartlett's Chapter 18 on Halbwachs and collective memory.For Bartlett, a schemahas both a conservative and a creative aspect, tending both to homogenize or conventionalize the new, and to support. See now the essays in Saito (2000), Iran-Nejad & Winsler (2000) and Kashima (2000). 30For an extended defence of such a dynamic conception of traces, see Sutton (1998:277-316).
208
John Sutton
The concept of a schema need not imply a settled structure in order for it to play its required explanatory roles. As an enduring but modifiable set of tendencies or dispositions, a schema may usefully be invoked to explain, for example, the way a story may be normalized in the remembering and retelling, with the schema driving easy inferences to uncertain or untold parts of the story. Cognitive-psychological accounts of the schema were implemented in connectionist models in the 1980s, with the history of past processing being "stored" in the (enduring but modifiable) matrix of connection weights of the neural network, and thus influencing the processing of new and related input (Rumelhart et al. 1986). And more recently, cognitive anthropologists have found this tradition a useful way of modelling both the "centripetal" forces of cultural reproduction and the competing "centrifugal" processes of variation and inconsistency: Claudia Strauss and Naomi Quinn, for example, take the connectionist version of schema theory to show how cultural learning produces responses which are permeated by tradition and yet not rigidly repetitive. Remembering occurs on the spot, in a context, and yet can be guided (without being determined) by cultural norms: because connectionism emphatically rejects a linguistic model of internal memory, it's easier to see that the traces culture leaves on and in individual brains and bodies are not downloaded copies of any specific cultural instructions, but rather flexible and particular action-oriented responses (Strauss & Quinn 1997: Chapter 3). 31 The dynamics of intrapersonal thoughts, feelings, and motives, conclude Strauss and Quinn, may be quite different from those of extrapersonal messages and practices, even if we accept that the boundaries between the two realms are permeable (1997:8 and passim). This is a key point which it's not easy to hold onto in the excitement of the new embedded cognition movements, and to which I'll return in conclusion.
Distributed Cognition and External Memory It's no accident that memory is at the heart of recent work on "the extended mind." On top of the connectionist focus on the plasticity of superpositionally stored memory traces, theorists like Andy Clark, Mark Rowlands, and Edwin Hutchins examine dynamic forms of interplay between such internal representations and the
31 See also Strauss & Quinn (1997: 44--47) on how to reinterpret Bourdieu's notion of the habitus, as a set of embodied dispositions driving regulated improvisation, into a naturalistic cultural cognitive science. Compare D'Andrade (1995); and Barry Smith's account of Hayek's notion of partial and dynamic cognitive maps, which act not as passive memory stores but as active memory-competence, in Smith (1997).
Representation, Reduction and lnterdisciplinarity
209
(natural and social) environment. 32 Linked through what Clark calls "continuous reciprocal causation," brain and world are often engaged in an ongoing interactive dance from which adaptive action results (Clark 1997: 163-166). Tim van Gelder explains the motivation for seeing genuinely cognitive processing as seeping out of the skull: Since the nervous system, body, and environment are all constantly changing and simultaneously influencing each other, the true cognitive system is a single unified system embracing all three. Interaction between the inner and the outer i s . . . a matter of coupling, such that both sets of processes continually influence each other's direction of change (van Gelder 1995: 373). This idea of the "extended mind" has led Clark et al. to argue that memory contents, as well as memory processes, sometimes spread or leak out of the brain and are left in the world (Clark & Chalmers 1998). Just as our unique problemsolving abilities depend most importantly "on our abilities to dissipate reasoning" by building "designer environments" (Clark 1997:180, 191), so our unique abilities to access, manage, and manipulate large bodies of information depend more on the technological and cultural symbolic networks we've constructed to plug ourselves in to (Donald 1991: 269-360; Rowlands 1999). I don't want here either to defend this startling idea against individualist objections, or to suggest that it's compatible only with these connectionist/dynamicist accounts of internal representation. This "active externalism," unlike traditional views about wide content and social externalism, does not violate the general considerations in favour of narrow content mentioned by Frank Jackson in this volume, because we agree with Jackson that knowledge of the world or the past depends on traces left in our region of the world or of time: what the active externalist denies is just that the division between the regions accessible to us, and those not, must coincide with the edges of the individual body. 33 The idea of extended m e m o r y systems leaves
32The best single treatment is Clark (1997). See also Hutchins (1995: Chapter 9) and Rowlands (1999: Chapter 6); for an introduction see now Clark (2001: Chapters 7-8), and for the role of philosophy of biology in understanding embodied cognition see Griffiths & Stotz (2000). 33 See Jackson (2004): the region in which traces are accessible to cognition may extend beyond the individual body, as in Jackson's example of a sealed spaceship, or contract within the body. Jackson's notion of an impermeablebarrier between self and world needs better motivation. For other objections to the very idea of the extended mind, see Adams & Aizawa (2001). The extended mind thesis admits of stronger metaphysical, and weaker methodologicalreadings: although I'm sympathetic to both, as long as the stronger is understood correctly, most of the argument in this section relies only on the weaker point.
210
John Sutton
room for, or positively invites, the correlative extension of the cognitive sciences of memory to include historical, technological, and cross-cultural investigations. The idea that "external memory" or external representation in general is neither merely metaphorical, nor a straightforward expression of more fundamental mental representations, can seem to rest on the claim that the external representations m information in notebooks, for example, which meets certain criteria of accessibility and reliability are functionally isomorphic to mental representations. 34 But whatever the merits of this claim, it sits oddly with connectionism: if internal representations are "stored" only in an implicit and distributed fashion, to be reconstructed only in a context, they don't look very like, and have quite different causal properties to, the fixed and context-independent symbols written in a notebook (the point is forcefully made by O'Brien 1998). But the case for the extended mind doesn't rest on such a case for functional isomorphism between the inner and the outer. Rather, as Clark makes clear in other moods, the core idea is that quite disparate internal and external elements are simultaneously coopted into integrated larger cognitive systems. The external media on which we rely so much as cognitive scaffolding are for Clark "best seen as alien but complementary to the brain's style of storage and computation. The brain need not waste its time replicating such capacities. Rather, it must learn to interface with the external media in ways that maximally exploit their peculiar virtues" (1997: 220). But different external media for the storage, transmission, and transformation of information have their own peculiar virtues. The various kinds of memory scaffolding which humans have used, from knots, rhymes, codes, and sketches to artificial memory techniques, photographs, books, and computers each have different properties. The resources of the historian and social scientist must again be included in cognitive science. 35 Merlin Donald carefully catalogues the differences between (internal) engrams and the (external) "exograms" which humans have produced since the Upper Palaeolithic era: "unlike the constantly-moving and fading contents of biological working memory, the contents of this externally-driven processor can be frozen in time, reviewed, refined, and reformatted" (Donald 1991: 308-319; 2000). While the enduring and expandable nature of many external symbol systems has indeed altered the informational environment in which brains
34This is suggested, notably, by the celebrated case of Inga and Otto described by Clark & Chalmers (1998), where the key argument is that if external information is used in ways which would be seen as cognitive if performed internally, there's then no principled reason not to see it as genuinely cognitive. In consequence, Adams & Aizawa (2001) treat the entire argument for the extended mind as resting on such functional isomorphisms. 35I've tried to put this into practice in one case study in Sutton (2000).
Representation, Reduction and Interdisciplinarity
211
develop, media theorists remind us that not all such systems are designed to hold information permanently in context- or medium-independent fashions, and that not all those systems which are designed to do so actually succeed (see for example Klein 1997; Tofts & McKeich 1997; and in art history, Kwint 1999). 36 Yet the communal, transmissable bodies of knowledge which do pervade human culture are not merely convenient holding-pens for information too unwieldy to store internally. Instead, they are essential tools for thinking which often alter the cognitive processes in which brains, bodies, and external symbol systems are all entangled. As Clark in particular argues, the internalizing of relatively context-free representational formats brings new cognitive possibilities, and burdens: By 'freezing' our own thoughts in the memorable, context-resistant, modality-transcending format of a sentence, we create a special kind of mental object an object that is amenable to scrutiny from multiple cognitive angles, is not doomed to alter or change every time we are exposed to new inputs or information, and fixes the ideas at a high level of abstraction from the idiosyncratic details of their proximal origins in sensory input (Clark 1997:210). By thus assimilating particular kinds of props and pivots over the course of cognitive development, we regulate our own minds, imposing an approximation of rigidity and inflexibility on our own mental representations. In this final section, I've taken a swift tour through some ways of thinking about collective memory and external memory which might be naturalistically acceptable. I've programmatically suggested that the post-connectionist notion of the extended mind, carefully developed, shows us how to treat external memory within a general information-processing framework. Whether this turns into a universalizing cognitive science of society, or a contextualist social cognitive science, depends on the care with which particular projects are pursued. I don't want to appear an ebullient optimist, hyping up the current state of interdisciplinary endeavours. Such schemes in the sciences of memory are as yet remorselessly difficult, and (despite actual local interdisciplinary contacts) there is a general sense that those sciences are disconnected from each other. I hope that this provides sufficient motivation for trying to show how such different memory researchers m from neurobiology to narrative theory, from the developmental to the postcolonial, from the computational to the cross-cultural, might one day be able to talk to each other.
Again, Michael Leyton's provocativetheoryof cognition (1992) as the recoveryof the past from the shape, in particular from the asymmetries,of objects in the present, may offer a unified framework.
36
212
John Sutton
Acknowledgments I gave earlier versions of this paper in 2000 in a seminar series on "Memory in Science" at the London School of Economics, and at the Representation in Mind conference at Sydney University. Thanks to both audiences for helpful comments, especially to Hasok Chang, Wayne Christensen, Ian Gold, Carl Hoefer, Cliff Hooker, Stephen Jacyna, Doris Mcllwain, Daniel Stoljar, Will Sutton, Maria Trochatos, Andrew Warwick, and Elizabeth Wilson.
References Adams, E, & Aizawa, K. (2001). The bounds of cognition. Philosophical Psychology, 14, 43-64. Assmann, J. (1997). Moses the Egyptian. Harvard University Press. Bakhurst, D. (1990). Social memory in Soviet thought. In: Middleton & Edwards (1990) (pp. 203-226). Bartlett, E C. (1932). Remembering: A study in experimental and social psychology. Cambridge University Press. Bickle, J. (1995). Psychoneural reduction of the genuinely cognitive: Some accomplished facts. Philosophical Psychology, 8, 265-285. Bickle, J. (1998). Psychoneural reduction: The new wave. MIT Press. Brewer, W. (1996). What is recollective memory? In: D. Rubin (Ed.), Remembering our past (pp. 19-66). Cambridge University Press. Campbell, J. (1994). Past, space, and self. MIT Press. Campbell, J. (1997). The structure of time in autobiographical memory. European Journal of Philosophy, 5, 105-118. Churchland, P. M. (1996). Flanagan on moral knowledge. In: R. McCauley (Ed.), The churchlands and their critics. Blackwell. Churchland, P. M., & Churchland, P. S. (1998). Intertheoretic reduction: A neuroscientist's field guide. In: On the contrary: Critical essays 1987-1997. MIT Press. Clark, A. (1997). Being there: Putting and brain, body, world together again. MIT Press. Clark, A. (2001). Mindware: An introduction to the philosophy of cognitive science. Oxford University Press. Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58, 7-19. Connerton, P. (1989). How societies remember. Cambridge University Press. Craver, C. E (2002). Interlevel experiments and multilevel mechanisms in the neuroscience of memory. Philosophy of Science, 69 (Supplement) 83-97. Craver, C., & Darden, L. (2001). Discovering mechanisms in neurobiology: The case of spatial memory. In: P. Machamer, R. Grush, & P. McLaughlin (Eds), Theory and method in neuroscience (pp. 112-137). Pittsburgh University Press. D'Andrade, R. (1995). The development of cognitive anthropology. Cambridge University Press.
Representation, Reduction and Interdisciplinarity
213
Dennett, D. (2000). Making tools for thinking. In: D. Sperber (Ed.), Metarepresentations. Oxford University Press. Donald, M. (1991). Origins of the modern mind. Harvard University Press. Donald, M. (2000). The central role of culture in cognitive evolution: A reflection on the myth of the "isolated mind". In: L. Nucci, G. Saxe, & E. Turiel (Eds), Culture, thought, and development (pp. 19-38). Lawrence Erlbaum. Engel, S. (1999). Context is everything: The nature ofmemory. W. H. Freeman & Co. Fara, E, & Patterson, K. (Eds) (1998). Memory. Cambridge University Press. Fentress, J., & Wickham, C. (1992). Social memory. Blackwell. Fischbach, G. D., & Coyle, J. T. (1995). Preface. In: Schacter (1995a) (pp. ix-xi). Fivush, R. (1991). The social construction of personal narratives. Merrill-Palmer Quarterly, 37, 59-81. Fivush, R. (1994). Constructing narrative, emotion, and self in parent-child conversations about the past. In: U. Neisser, & R. Fivush (Eds), The remembering self (pp. 136-157). Cambridge University Press. Gold, I., & Stoljar, D. (1999). A neuron doctrine in the philosophy of neuroscience. Behavioral and Brain Sciences, 22, 809-869. Griffiths, E E., & Stotz, K. (2000). How the mind grows: A developmental perspective on the biology of cognition. Synthese, 122, 29-51. Haaken, J. (1998). Pillar of salt: Gender, memory, and the perils of looking back. Rutgers University Press. Hacking, I. (1995). Rewriting the soul: Multiple personality and the sciences of memory. Princeton University Press. Halbwachs, M. (1925/1992). The social frameworks of memory. In: L. A. Cose (Ed.), Halbwachs, on collective memory. Chicago University Press. Hardcastle, V. G. (1996). How to build a theory in cognitive science. SUNY Press. Hilts, E J. (1995). Memory's ghost. Simon & Schuster. Hoerl, C. (1999). Memory, amnesia, and the past. Mind and Language, 14, 227-251. Hooker, C. (1981). Towards a general theory of reduction. Dialogue, 20, 38-59, 201-236, 496-529. Howe, M. L., & Courage, M. L. (1997). The emergence and development of autobiographical memory. Psychological Review, 104, 499-523. Hutchins, E. (1995). Cognition in the wild. MIT Press. Iran-Nejad, A., & Winsler, A. (2000). Bartlett's schema theory and modem accounts of learning and remembering. Journal of Mind and Behavior, 21, 5-36. Jackson, E (2004). Representation and experience (this volume). Johnson, M. K., Hashtroudi, S., & Lindsay, D. S. (1993). Source monitoring. Psychological Bulletin, 114, 3-28. Johnson, M. K., & Raye, C. L. (2000). Cognitive and brain mechanisms offalse memories and beliefs. In: Schacter & Scarry (2000) (pp. 35-86). Kashima, Y. (2000). Recovering Bartlett's social psychology of cultural dynamics. European Journal of Social Psychology, 30, 383-403. Kim, J. (1989). The myth of nonreductive materialism. Proceedings and Addresses of the American Philosophical Association, 63, 31-47.
214
John Sutton
Kitcher, E (1992). Freud's dream: A complete interdisciplinary science of mind. MIT Press. Klein, N. M. (1997). The history of forgetting. Verso Books. Koriat, A., & Goldsmith, M. (1996). Memory metaphors and the real-life/laboratory controversy. Behavioral and Brain Sciences, 19, 167-228. Kwint, M. (1999). Introduction: The physical past. In: M. Kwint, C. Breward, & J. Aynsley (Eds), Material memories (pp. 1-16). Berg. Lagrou, P. (2000). History and memory: The example of the two world wars. CNRS Cross-Disciplinary Encounters, at http://www.cnrs.fr/cw/en/pres/compress/memoire/ lagrou.htm cited at 22 July 2003. Le Goff, J. (1992). History and memory. Columbia University Press. Leyton, M. (1992). Symmetry, causality, mind. MIT Press. McClamrock, R. (1995). Existential cognition. Chicago University Press. McClelland, J. L. (1995). Constructive memory and memory distortions. In: Schacter (1995a) (pp. 69-90). McCormack, T., & Hoerl, C. (1999). Memory and temporal perspective. Developmental Review, 19, 154-182. MacDonald, S., Uesiliana, K., & Hayne, H. (2000). Cross-cultural and gender differences in childhood amnesia. Memory, 8, 365-376. Mandler, J. M., & McDonough, L. (1997). Nonverbal recall. In: N. L. Stein et al. (Eds), Memory for everyday and emotional events (pp. 141-164). Lawrence Erlbaum. Middleton, D., & Edwards, D. (Eds) (1990). Collective remembering. Sage. Miller, P. J., Potts, R., Fung, H., Hoogstra, L., & Mintz, J. (1990). Narrative practices and the social construction of self in childhood. American Ethnologist, 17, 292-311. Mitchell, K. J., & Johnson, M. K. (2000). Source monitoring: Attributing mental experiences. In: Tulving & Craik (2000) (pp. 179-195). Mullen, M. K., & Yi, S. (1995). The cultural context of talk about the past: Implications for the development of autobiographical memory. Cognitive Development, 10, 407-419. Nelson, K. (1993). The psychological and social origins of autobiographical memory. Psychological Science, 4, 7-14. Nelson, K., & Fivush, R. (2000). Socialization of memory. In: Tulving & Craik (2000) (pp. 283-295). O'Brien, G. (1998). Review of Andy Clark, being there. Metascience, 7, 78-83. Olick, J. K., & Robbins, J. (1998). Social memory studies: From "collective memory" to the historical sociology of mnemonic practices. Annual Review of Sociology, 24, 105-140. Perner, J., & Ruffman, T. (1995). Episodic memory and autonoetic consciousness: Developmental evidence and a theory of childhood amnesia. Journal of Experimental Child Psychology, 59, 516-548. Rajaram, S., & Roediger, H. L. (1997). Remembering and knowing as states of consciousness during retrieval. In: J. Cohen, & J. Schooler (Eds), Scientific approaches to consciousness (pp. 213-240). Lawrence Erlbaum. Reese, E., Haden, C. A., & Fivush, R. (1993). Mother-child conversations about the past: Relationships of style and memory over time. Cognitive Development, 8, 403-430.
Representation, Reduction and lnterdisciplinarity
215
Roediger, H. L. (1996). Memory illusions. Journal of Memory and Language, 35, 76-100. Roediger, H. L., & McDermott, K. B. (2000). Distortions of memory. In: Tulving & Craik (2000) (pp. 149-164). Rose, S. (1992). The making of memory: From molecules to mind. Bantam. Rovee-Collier, C., & Hayne, H. (2000). Memory in infancy and early childhood. In: Tulving & Craik (2000) (pp. 267-282). Rowlands, M. (1999). The body in the mind. Cambridge University Press. Rubin, D. C. (1995). Memory in oral traditions: The cognitive psychology of epic, ballads, and counting-out rhymes. Oxford University Press. Rumelhart, D. E., Smolensky, P., McClelland, J. L., & Hinton, G. E. (1986). Schemata and sequential thought processes in PDP models. In: J. McClelland, & D. Rumelhart (Eds), Parallel distributed processing (Vol. 2, pp. 7-57). MIT Press. Saito, A. (Ed.) (2000). Bartlett, culture, and cognition. Psychology Press. Schacter, D. L. (1982). Stranger behind the Engram: Theories of memory and the psychology of science. Lawrence Erlbaum. Schacter, D. L. (1995a). Memory distortion: How minds, brains, and societies reconstruct the past. Harvard University Press. Schacter, D. L. (1995b). Memory distortion: History and current status. In: Schacter (1995a) (pp. 1-43). Schacter, D. L. (1996). Searching for memory: The brain, the mind, and the past. Basic Books. Schacter, D. L. (1999). The seven sins of memory: Insights from psychology and cognitive neuroscience. American Psychologist, 54, 182-203. Schacter, D. L., & Scarry, E. (Eds) (2000). Memory, brain, and belief Harvard University Press. Schacter, D. L., & Tulving, E. (1994). What are the memory systems of 1994? In: D. Schacter, & E. Tulving (Eds), Memory systems 1994 (pp. 1-38). MIT Press. Schaffner, K. (1992). Philosophy of medicine. In: M. Salmon, J. Earman, C. Glymour, & J. Lennox (Eds), Introduction to the philosophy of science (pp. 323-339). Prentice-Hall. Schechtman, M. (1994). The truth about memory. Philosophical Psychology, 7, 3-18. Schudson, M. (1995). Dynamics of distortion in collective memory. In: Schacter (1995a) (pp. 346-364). Smith, B. (1997). Hayek and connectionism. In: S. E Frowen (Ed.), Hayek: Economist and social philosopher (pp. 9-29). Macmillan. Squire, L. R., & Kandel, E. R. (1999). Memory: From mind to molecules. Scientific American Library. Stoljar, D., & Gold, I. (1998). On biological and cognitive neuroscience. Mind and Language, 13, 110-131. Strauss, C., & Quinn, N. (1997). A cognitive theory of cultural meaning. Cambridge University Press. Suddendorf, T., & Corballis, M. (1997). Mental time travel and the evolution of the human mind. Genetic, Social and General Psychology Monographs, 123, 133-167. Sutton, J. (1998). Philosophy and memory traces: Descartes to connectionism. Cambridge University Press.
216
John Sutton
Sutton, J. (1999). The Churchlands' neuron doctrine: Both cognitive and reductionist. Behavioral and Brain Sciences, 22, 850-851. Sutton, J. (2000). Body, mind, and order: Local memory and the control of mental representations in medieval and renaissance sciences of self. In: G. Freeland, & A. Corones (Eds), 1543 and all that: Word and image in the proto-scientific revolution (pp. 117-150). Kluwer. Sutton, J. (2002). Memory: Philosophical issues. In: L. Nadel (Ed.), Encyclopedia of cognitive science. Macmillan. Tofts, D., & McKeich, M. (1997). Memory trade: A prehistory of cyberculture. Interface Books. Tomasello, M. (1999). The cultural origins of human cognition. Harvard University Press. Tulving, E. (1983). Elements of episodic memory. Oxford University Press. Tulving, E. (1993). What is episodic memory? Current Directions in Psychological Science, 2, 67-70. Tulving, E. (1999). Episodic vs. semantic memory. In: E Keil, & R. Wilson (Eds), The MIT encyclopedia of the cognitive sciences (pp. 278-280). MIT Press. Tulving, E. (2000). Concepts of memory. In: Tulving & Craik (2000) (pp. 33-43). Tuomela, R. (1995). The importance of us. Stanford University Press. van Gelder, T. (1995). What could cognition be, if not computation? Journal of Philosophy, 92, 345-381. von Eckardt, B. (1993). What is cognitive science ? MIT Press. von Eckardt, B. (1999). Critical notice of Hardcastle, How to build a theory in cognitive science. Philosophy and Phenomenological Research, 59, 221-224. Warnock, M. (1987). Memory. Faber. Welch-Ross, M. (1995). An integrative model of the development of autobiographical memory. Developmental Review, 15, 338-365. Welch-Ross, M. (1997). Mother-child participation in conversations about the past: Relations to preschoolers' theory of mind. Developmental Psychology, 33, 618-629. Young, J. (1993). The texture of memory: Holocaust, memorials, and meaning. Yale University Press. Zelizer, B. (1995). Reading the past against the grain: The shape of memory studies. Critical Studies in Mass Communication, 12, 204-239.
Chapter 11
Kant on Constructing Causal Representations Patricia Kitcher
1. Introduction
In his classic discussion of causation, The Cement of the Universe, J. L. Mackie concluded that "neither Kant's explicit arguments nor the general Kantian approach contribute much to our understanding of causation" (1980:112-113). He reached that harsh assessment after critically surveying the standard ways of interpreting and reconstructing the arguments of the Second Analogy. But he also brought his own historical skills to bear on the text and here his judgments were both more flattering and more tentative. He challenged Kant with a fundamental question: what good is a priori knowledge of the principle that all events have causes? He then read Kant as being committed to the following answer: "a priori knowledge of the general principle that every change is covered by some regularity would somehow help us to discover and establish particular causal laws" (1980: 92). He went on to surmise that although these ideas were not elaborated in the text, they might be developed in a way that would permit the deduction of causal laws from limited observations and the general causal principle. In 1997, Patricia Cheng carried out this project, although without (I believe) any direct influence from Mackie. I will lay out Cheng's proposal for using Kantian principles to extract causal information from observed correlations in the next section. Having used her work to redeem some of the promise of Kant's theories, I will display apparent ingratitude by offering Humean and Kantian critiques of her theory. Section three will offer a Humean critique of her definition of "cause"; Representation in Mind: New Approaches to Mental Representation Edited by H. Clapin, P. Staines and P. Slezak Copyright 9 2004 by Elsevier Ltd. All rights of reproduction in any form reserved. ISBN: 0-08-044394-X
218
Patricia Kitcher
section four will offer a Kantian critique of her solution to the puzzle of causal priority. The criticisms are, however, meant to be constructive. In each case, I will try to show how, by adopting some further Humean or Kantian ideas, Cheng's theory can be modified to meet the criticism. In the concluding section, I will offer a more tentative Kantian criticism of Cheng's analysis of the "computational" problem of causal induction.
2. Cheng's Power PC Theory Cheng presents her theory as a combination of the insights of Hume and Kant. She began in the Humean tradition of covariation, with the "contingency" or "probabilistic contrast" model. In the contingency model, a relation of contingency, the difference that the presence of some condition, L makes to some effect, E, is expressed as follows: APrlE = Pr(EII) - Pr(EI ~ I) [where Pr(E[/) is the observed frequency of I in the presence of E]. This formula must be understood as an idealization. It expresses contingency, if we abstract from confounding causes. Since the problem of causal induction takes place in the real world of confounding causes, the probabilistic contrast model will be workable only if it can be enriched to take account of confounding. To deal with confounding by altemative causes, Cheng proposes a "causal power theory of the probabilistic contrast model," or the "Power PC theory" for short. She reads Kant as endorsing the view that people have a priori knowledge that serves as a framework for interpreting data in terms of causes (1997: 368). In particular, she takes the content of that a priori knowledge to be: "there are such things in the world as causes that have the power to produce an effect.., and that only such things influence the occurrence o f an effect" (1997: 372, see also Bullock et al. 1982). She then combines Hume's and Kant's insights as follows: Kantian causal powers are the theoretical constructs for which Humean covariation can provide evidence, if we make the additional Kantian assumption that nothing influences the occurrence of effects except causal powers. (The last clause is meant to capture the familiar Kantian claim that all effects have causes, [1781/1787, A 189/B232].) Before tuming to the details of Cheng's account, I should consider an obvious objection to her attempt to combine Kant's views about causation with the probabilistic contrast model. Because of his views about freedom, Kant is almost always read as a determinist. Conversely, the tradition of probabilistic causality arose in response to the problem of trying to make sense of causality in light of the indeterminism of quantum mechanics. As Paul Guyer has argued, however, nothing in Kant's canonical discussion of causality in the Second Analogy tums on
Kant on Constructing Causal Representations
219
rule-governed succession being understood as universal rather than probabilistic (1987:240-241). Hence Cheng is not trying to mix a little determinist oil with probabilistic water. Since covariation does not, in general, imply causation, there is real work for the Kantian assumption to do. Cheng's basic move is intuitively clear and can be made mathematically precise. The causal power, Po, of a putative cause, L would be the contingency of that cause/effect pair if there were no other causes. 1 But there are. Take alpha (o0 to be the set of alternative causes, A1 through An. We then make the following assumptions. (1) When I occurs, it produces E with probability POl, when ~ occurs, it produces E with probability Po,~; nothing else influences the occurrence of E (1997: 373). Beyond this "Kantian" assumption, Cheng also assumes that (2) I and oL influence the occurrence of E independently (1997: 373). [That is, I and oLoccur independently of each other (see 1997: 374, and note 5).] (3) I and oL influence E with causal powers that are independent of how often I and oL occur. (The probability of E being produced by oL is Pr(o0 • Po,~ [1997, 373].) [Since it claims that if, for example, C1 produces El 80% of the time, then Cl will always produce E1 80% of the time, this principle is a probabilistic generalization of Kant's other key causal principle, "same cause, same effect" (1781/1787, A 189/B232). I will sometimes express this principle more simply as the claim that causal powers are "equipollent."] Assumption two will do some heavy work in the derivation of a formula for extracting causal powers from covariation, so I should make its rationale and implications explicit. Covariation accounts are widely understood to be vulnerable to the problem of "common causes." However frequently a candidate cause and an effect covary, the candidate cannot be considered to be a cause, unless there is some assurance that both it and the effect are not the result of some other cause. Assumption two is a boundary condition for Cheng's model, because it stipulates that there are no common causes. (Suppose there were a common cause of E and L Since that cause is a cause of E, it must belong to a, in which case I would not vary independently of oL.)
1I use a different notation, because I think Cheng's is slightly confusing.
220
Patricia Kitcher
In my slightly different notation, this is Cheng's argument. Intuitively, how often E occurs at all depends on how many times it is produced by I and how many times it is produced by any of its alternative causes. Given the three preceding assumptions, this claim can be made precise. The probability that E will occur is the sum of the probability that I will occur times the causal power o f / , Pol, and the probability that oL will occur times the causal power of oL, P % , minus the number of times that E occurs when I and a are both present. That is: Pr(E) = Pr(1) • POl q- Pr(ot) • Poa - Pr(l) x Pol x Pr(ot) • Poa
(1)
To find the relation between the power of a cause with respect to an effect and the contingency evidence, we need to consider the probability of E occurring, given/, and the probability of E occurring in the absence of I. To calculate the probability of E, given/, conditionalize Equation (1) on the occurrence of I. Since Pr(I/I) is one, the first and third terms of the RHS can be simplified, yielding: Pr(EII) = Pol + Pr(e~ll) • Po~ - POl • Pr(otlI) • Po~
(2)
That is, the probability of E, given I, is the probability that E was produced by I or by e~ in the presence of I. To determine the probability of E given the absence of I, conditionalize Equation (1) on not-/. Since P r ( I / ~ I) is zero, the first and third terms in the RHS disappear, yielding: Pr(EI ~ I) = Pr(otl "~ I) x P %
(3)
Since APrlE = P r ( E I I ) - Pr(EI ~ I), we can use Equations (2) and (3) (plus simplification) to express the contingency in terms of causal powers: APrlE = [1 -- Pr(otll) • Po~] x Pol + [Pr(e~ll) - Pr(cxl ~ I)] • Po~]
(4)
This equation can be rearranged to give an estimate of the causal power of I: POl =
APrlE -- [Pr(otlI) - Pr(otl ~ I)] • P % 1 - Pr(otll) x P %
(5)
As noted in assumption two, however, I and ot are independent. Thus we may assume, Pr(alI) = Pr(otl ~ I) = Pr(a), and the second term in the numerator goes to zero. With this assumption, we may derive: POl = APrIE 1 - Pr(o0 • P o a
(6)
We may seem to have gotten nowhere, since the unknown power of I would still be expressed in terms of the unknown power of oL. Notice, however, that we are not required to calculate P % independently; what we need is the product Pr(e0 x P % , that is, the probability of the effect that is due to oL. But if I and a are independent,
Kant on Constructing Causal Representations
221
then Pr(oLII) = Pr(oLI ~ I) = Pr(o0, and we can substitute Pr(ct) for Pr(oLI "~ I) in Equation (3), yielding an estimate of the product Pr(et) x Po~ solely in terms of the observed frequency of E in the absence of I. In this way, we may derive Equation (7), which gives the powers of I just in terms of observed frequencies relating E and I: Pol = APrtE Pr(EI ~ I)
(7)
Thus, by making the Kantian assumptions that causal powers and only causal powers influence the occurrence of effects (and that causal powers are equipollent), we can use covariation to determine causal p o w e r - in the condition where the potential causes are independent. Cheng's work amply fulfills Mackie's contention that Kant's causal principle could enable us to move from limited observations to particular causal laws (1980: 92). In her own view, the Power PC theory resolves three outstanding problems (1997: 398). First, it explains something that seems mysterious on the covariance view. Since covariation has no intrinsic temporal order, the temporal priority of causes to effects has to be introduced as an unrelated condition. If covariations are the visible expression of causal powers, however, then the conditions belong together: "if reasoners have the intuitive notion that causes produce.., the effect, it follows that the cause must precede the effect (even if only by an infinitesimal amount) because a cause must exist before it can produce any consequences" (1997: 398). Second, it fills in a fatal lacuna in standard power theories. Even if things have causal powers, how does that hypothesis enable us to use available evidence to induce causes? Third, since it appeals only to highly general, though cause-specific knowledge, it avoids positing an implausibly rich innate endowment for causal mechanisms. Cheng also reports on a series of experiments suggesting that reasoners use qualitative versions of the Power PC theory in inducing causes from the evidence. I will present just one class of cases where the Power PC theory is much better than a simple contingency model in providing a d e s c r i p t i v e l y - and normatively correct account of causal induction. I borrow Cheng's own illustration of the class. Cheng was once tested for a food allergy via the usual battery of scratch tests. Every scratch registered positively, because, as it turned out, she was allergic to the scratch procedure. The contingency model would yield a probabilistic contrast of zero, APrlE = Pr(EII) - Pr(EI ~ I) = 0, and hence, no causal relation for any of the particular foods. Her doctor was, however, correctly uncertain about whether or not she was allergic to any of the foods tested (1997: 368). Experimental evidence also shows that, when considering cases of "ceiling effects," where an alternative cause always produces the effect, reasoners are
222
Patricia Kitcher
reliably unsure whether or not a zero contrast indicates an absence of causal relation for the candidate cause (1997: 383-384). Equation (6) of the power PC model predicts this result, since Pr(ot) x Po,~ would be one, making POl undefined. Despite its considerable virtues, Cheng's theory is somewhat limited. In more recent work, she has tried to generalize the case of singular, alternative causes to cover conjunctive causes (2000: 223); Clark Glymour (2000: 223) has suggested that, by using some of the techniques from his own work with Bayes' nets, Cheng's model may be generalized to handle cases of intervening variables and complex situations involving a variety of generative and preventive causes. My concern will not be with possible extensions of the Power PC theory, however, but with its philosophical foundations.
3. H u m e and Kant on Causal "Powers" Although Cheng finds inspiration in Hume and Kant, both philosophers would reject the Power PC theory, and for exactly the same reason. They did not believe that the notion of a "power" was sufficiently clear or precise to be usable in science or philosophy. The broad outlines of Hume's critique are familiar. We cannot perceive "powers" directly, but only the changes that they bring about. Attempts to elucidate "power" in terms that are better understood have created a small circle mutually inter-definable, but equally unclear, concepts, including "cause," "power," "production," "efficacy" and the like (1739: 157). Hume considered the move that Cheng makes: powers might be introduced through an inference from constant conjunctions or correlations. So, for example, f's possession of the causal power to bring about g's would provide a theoretical explanation of the correlation between f's and g's. He rejected such an inference as rationally indefensible (1739: 91). To understand the force of Hume's argument, it is useful to divide the issue into two sorts of cases. First consider the situation where nothing beyond the "brute correlation" between f's and g's is ever discovered. In this case, we might pejoratively, but fairly, describe the power of f's over g's as a "brute" power. In the second class of cases, we discover some previously hidden property of f's, say their molecular structures, that gives some insight into how or why f's regularly produce g's. Such "powers" are more perspicuously labeled "stricter regularity" powers, since other sorts of things might have the same molecular structure and some f's might have only defective versions of the relevant structure. When Hume objected to the invocation of powers, his targets were brute powers; he believed that science progressed by finding stricter regularities. What is gained by theorizing that a particular correlation of f's with g's is the result of the brute power of f's to bring g's about? By hypothesis, there is nothing
Kant on Constructing Causal Representations
223
in the known properties of f's and g's that makes the link between f's and g's any more than a brute correlation. Suppose we then posit that something in the f's, the g-power, brings about g's. As Jonathan Bennett (1971), noted many years ago, Hume believed that under these circumstances, we have simply replaced one brute correlation, of f's and g's, with two brute correlations, f's and g-powers, and g-powers and g's. No advance has been made. Cheng does not regard the "power" aspect of the Power PC theory as having Humean sanction. Rather, she traces "causal powers" back to Kant. Although Kant famously "replied" to Hume's attack on causation, he offered no defense of the notion of "causal power" at all. 2 A perusal of Cheng's other philosophical references yields a more plausible source for the idea of causal powers, namely, Nancy Cartwright. In opposition to the long-standing Humean tradition, Cartwright (1989, e.g. 91ft.) has argued that we apprehend causal powers as directly as we apprehend anything else. In her view, we can just see the power that one action (say a quick shove by a six-year-old) has over the condition of something else (say, a standing toddler quickly becoming supine). 3 The problem is that Cheng tries to combine Cartwright's hypothesis of causal powers, with Hume's thesis that causal power per se cannot be observed. Whether Cartwright is correct or incorrect in her dispute with Hume over the perceptibility of causal powers, her position is consistent. We can legitimately appeal to causal powers to explain phenomena, because, if pressed on what causal powers are, we can reply: they are what you perceive, when, for example, you perceive one child pushing over another. By contrast, Cheng cannot agree with Hume that we cannot perceive causal power (e.g. in 2000: 227) and yet appeal to causal powers in her theory of causality. The problem is not one of theoretical terms in general. We can understand such terms in a variety of indirect ways, most often, perhaps, by making analogies between theoretical and observable entities. The problem is with the notion of"power" itself. Unless it is understood as a stand-in for whatever stricter regularity will be discovered to explain the covariation, the quality-less character of "causal power" brings no clear notion to mind at all.
2 Some commentators (e.g. Falkenstein 1995: 332) suggest that Kant retained an Aristotelian notion of causal efficacy along with a modifiedHumean regularitytheory. However,they see the two types of causality as playingdifferentroles in his system.On this view,Kant neededAristotelianpowercausality not for the purpose of rebutting Hume in the Second Analogy, but to give sense to the possibility of "noumenal causality."Even if Kant believedin unknowablecausalpowers for the unknowablenoumena (which I doubt), this doubly opaque thesis would be of no interest to Cheng, since it would have no relation to the "other" Kantian theses she wants to endorse, "all events have causes" and "same cause, same effect." 3 Cartwright used this example in conversation.
224
Patricia Kitcher
In their positive as well as their negative views, Hume and Kant were in considerable agreement: causal induction is not a matter of finding mysterious powers, but of discovering the rules by which effects covary with causes. Some philosophers may protest that rules come so cheaply that a rule view is vacuous. In particular, it would be possible to construct a rule for the covariance of an effect simply by disjoining the circumstances of each of its occurrences. Although this point is logically impeccable, both Hume and Kant were concerned with the interplay between nature and our faculties. Hence they would presume that the rules at issue be cast in terms of classifications of "causes" and "effects" that are cognizable by us and it is far from clear that gerry-mandered "universal" rules can meet this requirement. It is a relatively straightforward matter to amend Cheng's borrowing from a Hume-inspired Kant. On that view, it is not that cognizers assume the existence of causal powers. Rather, they assume that every occurrence of an event type E can be placed under a rule for the covariance of events of type E with the occurrence of events of type L or A l or A2 or A3 . . . . . Or, in Kant's own language, cognizers presuppose that "everything that happens (begins to be) presupposes something which it follows in accordance with a rule" (A 189). Cheng's three assumptions and her seven step derivation can easily be recast, by substituting Kant's rules (R) for her powers, Po. For example, cognizers would use observed covariations to induce these rules, by assuming that every occurrence can be placed under some rule. Thus assumption one (p. 5 above) would be restated as: when an occurrence of event type I occurs, then an occurrence of an event of type E occurs with probability RI; when an event of type e~occurs, then an occurrence of an event of type E occurs with probability Ra; every occurrence of E can be accounted for by RI or R~. The crucial equation, (7), would be understood as an estimate of the rule by which occurrences of events of type E covary with occurrences of type I: R I = APrlE
1 - Pr(EI ~ I)
(8)
In a telling moment, Cheng observes that the causal power of I "coincides" with Pr(E/I) when there are no other causes (as her Equation (7) shows [1997: 373]). The rule view avoids this unexplained coincidence. When no other causes exist or are present, Pr(E/I) just is the rule for the covariance of occurrences of events of type E upon events of type I. On what is, I believe, Kant's real view, the relation between observed covariation and the actual rules of covariation would be the same as that between observed covariation and causal powers on the Power PC theory, namely, evidence to theory. Further, the "rules" theory of the probabilistic contrast model would preserve the basic structure of Cheng's solution to the problem of causal induction: causal
Kant on Constructing Causal Representations
225
induction requires an interplay of observational data with a priori assumptions about how to interpret evidence. Thus, Cheng could substitute Kant's unmysterious "rules" view for the power view that she and others trace back to Kant and preserve most of the virtues of her account. In particular, although the "rule PC theory" would not fill in a lacuna in the power view (by explaining how the assumption of causal powers helps with the problem of causal induction), as noted, it would fill in the corresponding lacuna (noted by Mackie) in the a priori approach, by showing exactly how various a priori assumptions enable cognizers to move from observable evidence to claims of cause and effect. It also preserves the virtue of using a few high-level principles in the place of an excessively large innate endowment of causal knowledge. The notable exception is the explanation of temporal priority. As Cheng observed, rules of covariation have no temporal implications. As we will see in the next section, however, further Kantian reflections imply that explaining the temporal priority of causes to effects may not be a virtue.
4. Kant on the Simultaneity of Causes and Effects In the Second Analogy, Kant examined several cases that seemed to contradict the standard assumption that effects follow causes. The issue of temporal succession was crucial for Kant, because he believed that causal relations enabled cognizers to determine temporal order. If causes do not invariably precede effects, however, then it is not obvious how understanding causal relations can help with the ordering problem. The worrying cases involved the heating of a room by a stove and a heavy ball denting a pillow. As Mackie argued, subjects' ordinary judgments about causes do not reliably line up with either necessary or sufficient conditions for bringing about an effect (1980, Chapter 2, passim). Hence, his well-known analysis of "cause" as an insufficient, but non-redundant part of an unnecessary but sufficient (INUS) condition (1980: 62). For the purpose of temporal ordering, however, the crucial condition would be the "last" element of a sufficient condition to fall into place. Kant conceptualized events and causes in terms of alterations or changes in the states of enduring things. Let us call the last alteration in one of the causal factors that completes a sufficient condition for the event, the "C-state of a C-substance," and the effect "the alteration in an E-substance from its O[ld]-state to its N[ew]-state." If causes precede effects, then the possession of a C-state by a C-substance should precede m but by "infinitesimal amount" m the possession of the N-state by the E-substance. As Kant realized however, this cannot be correct: "in the instant in which the effect first arises, it is always simultaneous with t h e . . . [C-substance having the
226
Patricia Kitcher
C-state] since if the cause [including the C-state of the C-substance] had ceased to be an instant before, then the effect would never have arisen" (A 203/B 248). Neither can the E-substance's possession of its N-state come after the C-substance having its C-state. Suppose that the C-substance acquired its C-state at tl, then, if that is the last element of a sufficient condition, why did the E-substance not alter from its O-state to its N-state at tl ? Suppose that the E-substance did not acquire its N-state until some te after tl. Since points in time are indistinguishable from each other in themselves and, ex hypothesi, nothing else relevant to the N-state of E has changed between t l and te, why would the E-substance take on its N-state at te ? It was Kant's early recognition that there were no plausible answers to these questions that led him conceptualize causes, not in terms of enduring substances, but in terms of the alteration of a C-substance to its C-state (see Nova Dilucidatio, I: 411). Hence Cheng's claim that causes precede because "a cause must exist before it can produce any consequences" (1997: 398, my emphasis) is wrong. The C-state of the C-substance need not m indeed it cannot m exist before the E-substance having its N-state. Despite his clear appreciation of this point in his early writings, Kant's discussion in the Second Analogy appears somewhat muddled. He analyzes the pillow and ball case as follows: If one lays the ball on the pillow, then the dent follows the previously smooth shape, but if the pillow happened to have a dent, then the ball would not follow. So the laying of the ball is the cause and the dent the effect, and that "dynamical connection" gives the temporal order, even though the laying and the denting are simultaneous. What matters is the order of time, not its lapse; even if no time has elapsed, the temporal order can still be determined (A 203/B 249). Kant seems to be saying that even though there can be no time lapse between the C-state of the C-substance and the N-state of the E-substance, one state precedes the other! Kant's other suggestion is that cognizers get temporal sequence only because most causes do not achieve their effect in an instant. To take his example, a fire is lit in a stove and the effect of the room becoming warm extends past the initial instant when the C-substance acquired it C-state. There are two ways to understand his notion of the effect developing. Suppose a fire is lit at six and the room attains its peak temperature of 68 ~ at 6:10. Cognizers can get the effect to come after the cause by seeing the original lighting as the cause of the maximum temperature, but a more reasonable analysis would view the change from say 67 ~ to 68 ~ as caused by a further change in the logs, for example, their acquiring at 6:10 the C-state of the last inch of the last log burning. With a continuing process such as a fire, there would be a succession of simultaneous causes and effects. Since placing the effect after the cause by identifying the cause as the initial lighting and the effect as the final temperature would involve significant
Kant on Constructing Causal Representations
227
confusion, a second reading of "effects do not develop in an instant" seems more plausible. On this interpretation, Kant's point was not that many causes are ongoing processes, but that the total effect of the room getting warm m the warmth radiating out from the air immediately around the stove to the air in rest of the r o o m - takes a while to develop. Even in that case, however, it is slightly misleading to call the stove, rather than the immediately adjacent air, the "cause" of the outermost air of the room becoming warm. Applying Kant's own test, had the stove ceased to exist just prior to the outermost air receiving heat from the neighboring air, the heat would still be transferred. Kant's discussion of simultaneous causes and effects has several implications for Cheng's view. First and most obviously, it implies that under reasonable ways of conceptualizing "causes" and "effects," causes do not always precede effects. Although Kant presented these examples as somewhat anomalous, his considered position was that the "majority of efficient causes in nature are simultaneous with their effects" (A 203/B 248). From his discussion, it seems to follow that, barring confusion, all causes are simultaneous with the beginning of their effects. His focus on the "last" element of the sufficient condition, the C-state of the C-substance, does not prejudice the issue. In principle, all the (separately) insufficient conditions might occur together, so no piece of the cause would precede the effect. Despite the seemingly universal view that causes precede their effects, Kant's argument that they must be simultaneous receives considerable empirical support. Unless they have some hypothesis about intervening developments, adults (and most children over 4) will reliably select the candidate cause that is as close in time as possible to the effect over its competitors (Bullock et al. 1982: 238-239). These results could be explained by appealing to a latent belief that causes and effects (or the beginnings of effects) are simultaneous. Cheng acknowledges that she is offering her theory of causation without first settling the difficult issue of how to delimit causes and effects (1997: 370). I am not suggesting that Kant's analyses of "cause" in terms of the change of state of a C-substance and "effect" in terms of the change of state of an E-substance provide complete solutions to the demarcation problem. At the least, however, the preceding discussion shows that the thesis of temporal priority cannot be maintained without first settling on some canonical way to understand "causes" and "effects." Beyond casting doubt on the precedence condition, Kant's analyses of various cases also offer some plausible explanations of why subjects believe that causes precede, even while latently believing that causes and effects are simultaneous. Although the change to the N-state of the E-substance is simultaneous with the C-state of the C-substance, the N-state may well persist. This situation is best illustrated in the example of the ball and the pillow. The pillow acquires the state
228
Patricia Kitcher
of being dented exactly when the ball is placed on it, but that state may persevere long after the ball is destroyed. Still, not all N-states persist, so this rationale for the belief in the precedence of causes is pragmatic. If subjects tend to pick out causes as the salient beginning of an ongoing process, then the fact that many causes are continuous (the stove) and many effect states lead to further effect states (the radiating of heat) would provide further cases where subjects place effects after causes for pragmatic reasons. We are now in a position to offer a possible amendment to Cheng's position. Kant's theory does not provide exactly what Cheng thought, namely, a rationale for the precedence condition of the probabilistic contrast model. Instead it makes a correction to the model: N-states of E-substances are simultaneous with C-states of C-substances; subjects commonly believe that all causes precede because many effect states persevere, many causes are continuous, and many effect states lead to further developments in the states of the E-substance. These almost universally encountered circumstances give subjects systematic, but nonetheless pragmatic reasons for placing effects after causes. As Cheng notes, often there are also what might be described as "idiosyncratic" pragmatic factors at work. So smoking is described as the "cause" of lung-cancer, because it is the salient beginning of a long chain of hidden transformations leading to the development of disease (1997:371). Although it is simple and relatively uncontroversial, the proposed amendment is not ideal from either Cheng's or Kant's point of view. Seemingly Cheng wanted a justification for the precedence condition itself, and not an accounting of the pragmatic factors involved in subjects' belief in precedence. Kant would be even less satisfied. Although he acknowledged that causes and effects were simultaneous, he believed that our understanding of the causal order determined our understanding of the temporal order. How can this be if causes and effects are simultaneous? Finally, this amendment to the precedence claim does not fit very well with the earlier substitution of "underlying rules of covariance" for "causal powers." If the correct understanding of the nature of causality is in terms of underlying rules of covariance, then it would seem little short of miraculous that the particular circumstances surrounding covariance invariably lead subjects to place causes before and effects after. 4 But this result is not controversial. Whether subjects think of the cause as coming before the effect, or as an infinitesimal amount before the effect, or as simultaneous with the effect, even young children will not place the cause after the effect (Bullock et al. 1982: 222-223). Since the proposed
4 Sincehe proposes to determinecausal structuresfromnontemporal data, Pearl also raises the question of why dependency and temporal relations always go together (2000: 57-58).
Kant on Constructing Causal Representations
229
emendation gives such an unsatisfactory account of this key explanandum, I will venture a little further into Kant's theories in hopes of doing better. If Kant's considered view was that [at least] the majority of causes and effects were simultaneous, then how could he also maintain that cognizers' appreciation of cause and effect was crucial in their determination of temporal order? Fortunately, Arthur Melnick (1973) cracked this difficult exegetical puzzle some years ago. In his original discussion of alteration and time in the Aesthetic, Kant noted that the only way cognizers can make sense of alteration, an E-substance going from state O to N (or not-S to S), is temporally. Contradiction cannot be avoided unless the states are regarded as existing at different times (A 31/B48). So if a cognizer has perceptions of an existing thing in two different states, O and N, then she knows that one of these states must succeed the other. But which one? Melnick's crucial insight is that Kant conceived of cause and effect as a three-place relation, between (in my terminology) the C-state of the C-substance and the O and N states of the E-substance (1973: 101-102). Hence the rule is not simply that the N-state of the E-substance covaries with the C-state of the C-substance, but that it does so when the E-substance is in state O. Alternatively a complete rule of covariance would be: upon the C-state of the C-substance, the E-substance varies from its O-state to its N-state. On Kant's view, determination of causal relations enables cognizers to place events in temporal order - even though causes do not precede m because the rules indicate how the states of the E-substance succeed each other in the presence of the cause. Melnick's hypothesis that Kant was operating with a three-place relation is given considerable support by looking at the examples. All of them involve three factors: the room goes from cooler to warmer given the fire, the pillow goes from plumped to dented, given the leaden ball, and in the central example of the Second Analogy, the ship moves from its upstream position to its downstream position, given the current. 5 Melnick's reading of the Second Analogy permits a more complex, but more satisfactory, way of amending Cheng's theory. Once again, Kant's view provides a correction to the precedence condition of the probabilistic contrast model. Causes do not precede their effects; one state of an enduring substance succeeds another according to some rule stating that upon a C-substance attaining its C-state, the E-substance alters from its O-state to its N-state with probability Rc. Given the metaphysical principle that no thing can have different states at the same time and the metaphysical principle that all alterations of state fall under some rule of covariance or other, it follows that an effect, the N-state of the E-substance,
5 Paul Guyer also adopts Melnick's three-place interpretation and notes how well it works with Kant's ship example as well as the examples of the stove and the pillow (1987: 257-262).
230
Patricia Kitcher
comes after some earlier state of the E-substance, and that if the attaining of the N-state in this instance falls under rule Rc, then the C-state of the C-substance cannot come after the N-state of the E-substance in this instance. 6 Thus this account provides a clear and comprehensible metaphysical rationale not for the precedence condition, but for something very close to it, the "non-succession" condition m the conviction that no cause can come after its effect.
5. Kant on the Relation Between Causality and Temporality In addition to disagreeing with Cheng about the definition of "cause" in terms of "powers" rather than "rules," Kant would object to her account of the discovery problem. Cheng's goal is to offer a computational level theory of causal induction, in her words, to "specify the abstract function relating the input and the output of the process of causal induction given the constraints that govern the problem of causal induction" (1997: 370). On the probabilistic contrast model, "an initial criterion for identifying potential causes is perceived priority" (1997:371).7 Besides objecting that it is not the cause that precedes, but a different state of the E-substance, Kant would also reject the assumption that cognizers can simply perceive that one state came first. He believed that, although cognizers could recognize that an Esubstance had two different states, its O-state and its N-state, and could know that these could not exist at the same time, the only way that they could determine which state came first and which succeeded was by appeal to the relation of cause and effect (A 189/B 234). If cognizers know a rule stating that an E-substance alters from its O-state to its N-state in the presence of a C-state of a C-substance, then they know that the N-state must be placed after the O-state. Hence the determination of temporal relations itself depends on interpreting the data causally. The venerable objection to Kant's proposal that causal interpretation is required for temporal ordering, which Cheng echoes (1997: 368), is that it is circular. If cognizers did not frequently observe the N-state of the E-substance following its O-state, then how could they ever come to know the above-cited rule? Beyond
6 As William Harper has noted, the repetitive motion of a pendulum offers an interesting case of succession, since, for example, the left apogee of the motionboth precedes and succeeds it central nadir position (1984:120). Harper solves this problem by appealing to a more precise rule that specifies the current position of the pendulum bob. The three-place analysis also resolves this problem,but in a more principled way, since it is always necessary to look to the O-state of the E-substance to state the rule. 7 At one point, Kant observes that temporal sequence is the only "empirical criterion" of the relation of cause and effect (A 203/B 249). But by this he means that we determinetemporal sequenceand relations of cause and effect through each other-- and hence that all effects must involvetemporal sequence not that temporal sequence can be determined in complete independence of causal relations.
Kant on Constructing Causal Representations
231
noting the highly controversial nature of Kant's own position, Cheng could defend against this criticism by pointing to the limited scope of her theory. She need not deny that causal knowledge can be used to determine temporal relations. Having divided the problem of causation into the acquisition of causal knowledge and the use of specific causal knowledge in further causal induction, she explicitly limits her theory to the first task (1997: 370). A better way to put Kant's objection might be that Cheng's division of the problem of causal induction is illegitimate. The causal and temporal interpretations that we place on sensory data are holistic and this is true from the beginning. Indeed, for Kant, it was not just temporality and causality that reciprocally implied each other. Cognizers' judgments about whether some putative event was real or imaginary depended on their being able to fit that event into one all-encompassing spatial-temporal and causal framework. So causal induction could not be a matter of moving from perceived events, or the perceived succession of an N-state of an E-substance upon its O-state, to causes; it had to involve using sensory data to solve simultaneously for temporal relations, causal relations, and real events vs. illusions. The Kantian criticism that Cheng misconceives the discovery problem, because she fails to appreciate the complex interplay between causal and temporal determinations cannot get a foothold without some way of blunting the charge that his rival account is impossible, because circular. To get some sense of how Kant would have avoided the circle, we need to consider the Newtonian context of his causal theorizing. After this brief historical excursus, I will return to the primary task of showing how, once again, a Kantian criticism may lead to a useful amendment to Cheng's account. In presenting Newton's and Kant's approaches to the problem of space and time, I rely on Michael Friedman's (1992) clear account. Newtonian physics referred to real motions, and so to changes in real spatial position through real time. As Newton explained in his "Scholium to Definitions," however, we are not able to determine real change in spatial position or real time directly: "the parts of that immovable space.., do by no means come under the observations of our senses" and although "time flows equably" and we measure time by motions, "it may well be that there is no such thing as a uniform motion by which time may be accurately measured" (Newton 1689). Thus Newton faced exactly the same sort of circle that confronts Kant's theory. To determine real motions and so forces, he needed real time and real space; since neither of the latter could be perceived, they had to be derived from (non-equable) motions. According to Friedman, Newton's solution was to start from apparent motions. From these, we then deduce the force laws. The laws enable us to find the common center of gravity of the solar system, which we can then use as an immovable surrogate for absolute space; this, in turn, enables us to distinguish
232
PatriciaKitcher
relative from true motion (1992: 142). Friedman regards Kant as following a similar procedure in the Metaphysical Foundations of Natural Science, but giving the end result a different (transcendental) gloss. Again we start from apparent motions and then work our way to Newton's laws. From the laws and further observations we try to reach ever more general inertial frames. The end result of this procedure is not, however, an approximation to absolute space. Kant could find no sense to that notion. Rather, he viewed the same Newtonian argument as a constructive procedure for first defining the concept of true motion and real space ( 1992:143). The Second Analogy began with the Newtonian premise that "time cannot be perceived in itself" (A 188/B 233); hence the need for ordering by causal relations. Since the oft-noted circle is obvious, it is a reasonable interpretive hypothesis that Kant believed that something like the procedure that saved Newtonian physics from hopeless circularity could also save his account of cause in general. Cognizers would begin with apparent motions. Although Kant offered no hypothesis about how this was possible, Mackie did. If time cannot be perceived directly, then seemingly cognizers could order events only by being sensitive to certain kinds of patterns in the data. To use Mackie's metaphor, some of the data must appear to be made to fit together as pieces of a jigsaw puzzle (1980: 105). In the parade case for Newtonian physics, different pieces of a motion would fit together as parts of a continuous motion with a beginning and an end. But doesn't this sort of example show that cognizers can get temporal order without causation, but merely by appeal to motion? I think that Kant would agree that subjects can get partial, prima facie orderings from apparent motions. This does not show that human temporal perception is independent of assumptions about causality, however, any more than Newton's initial appeal to apparent motions showed that apparent and real motions could be distinguished without assumptions about laws. Apparent motion is only part of the story. Because of their innate propensities to perceive apparent motions, cognizers could begin to piece together a temporal order. Although they might be fortunate enough to be able to grasp the temporal order of some potential alterations and causes against the background of these motions, it is probably more reasonable to assume that they also have some innate propensities to see certain kinds of things as causes and effects. At this point, cognizers would have very partial causal and temporal orderings. How are they to arrive at what Newton and Kant took to be their mature representation of time, namely, a complete ordering of alterations in the states of substances against in an "equably flowing" time scale? Again, Kant did not develop the account, but the following scenario seems plausible. One way to begin to construct a temporal scale would be to assume, falsely, that any motion is uniform. Then the distance traversed would give
Kant on Constructing Causal Representations
233
an approximation of duration. 8 Further experience will present cognizers with different motions yielding different scales; the assumption of uniform motion will be undermined upon observation of two motions, where one speeds up relative to the other. How can these putative, but manifestly incompatible, clocks be corrected? It is at this point, I believe, that causes enter the Kantian account. In parallel with Newton's procedure, cognizers would use some initial frames of reference and clocks to get a relation between each prima facie cause and its apparent effect. Since causes are equipollent, when the same cause is applied, the same alteration should occur; paradigmatically, the same force operating under the same conditions should set similar bodies in similar motions through space and time. Cognizers would then be in a position to move from their understanding of the relations between the various causes and effects to a universal clock that accurately captured each of those relations. Although the general case involving all causes would be messier than the case of celestial mechanics, the basic movement of thought would be the same, from apparent motions, prima facie causes, and provisional clocks and frames of reference, to causal laws, and from causal laws to ever more universal clocks. Further, the key Kantian insight highlighted by Friedman would be the same: the universal clock is not a clock that approximates some independent time, but that clock that enables the causal laws to come out right. Given the Newtonian procedure for the simultaneous construction of a temporal and causal order of real events, Kant's claim that causal interpretation is itself an ingredient in the recognition of temporal sequence cannot be rejected as incoherent. He need not deny that there are c a s e s - indeed very many cases where cognizers start from correlated sequences and move to causal relations. Rather, the point is that these occur against a background of interdependent causal and temporal determinations, so the direction of inference from time order to causation is not one-way. The importance of Kant's claim for the interdependence of causal and temporal determination for Cheng's project is that it brings out a further, high-level principle that cognizers bring to bear in inducing causes. Besides Cheng's stipulative principle, (2) "I and a influence the occurrence of E independently" (above p. 219), we already have three metaphysical principles in play: (1) When an event of type I occurs, that is, whenever a Ci-type-substance attains its Ci-type-state, then an event of type E occurs, that is, an E-type-substance alters from its O-type-state to its N-type-state, with probability R i; when an event of type e~ occurs, that is whenever a C~-type-substance attains its
8 Among others (Richards 1982) presents data suggesting that children use distance to measure time.
234
Patricia Kitcher
Ca-type-state, then an E-type-substance alters from its O-type-state to its N-type-state with probability Ra; every alteration of an E-type-substance from its O-type-state to its N-type-state can be accounted for by R / o r R~. (2) The causal rules, R/ and Ra are invariant though not necessarily universal. [The rule version of Cheng's third principle.] (3) The C-state of a C-substance is simultaneous with the beginning of the N-state in the E-substance. An effect, the N-state of the E-substance, comes after some earlier state of the E-substance, and if the attaining of the N-state in this instance falls under rule Rc, then the C-state of the C-substance cannot come after the N-state of the E-substance in this instance. [The correction to Cheng's claim that causes precede effects.] The Newtonian procedure reveals a further crucial principle or constraint in Kant's account: (4) The deterministic or probabilistic rules characterize alterations of all real states of substances in one, temporally coherent world. Do cognizers implicitly use such a principle in inducing causes? Cheng does not claim that lay people use the equations presented above to calculate causes; rather her hypothesis is that ordinary subjects use qualitative versions of this principles. Similarly, I am not trying to suggest in Kant's name that children or lay people are constantly calibrating causes against each other in any very precise way to get a uniform time-scale. The idea would be only that they would reject causes that were blatantly out of alignment with the time-scales provided by co-occurring processes and tend to favor causes that were compatible with other salient processes. Seemingly the following experiment might provide evidence one way or another. Suppose subjects were shown a time-lapsed film (so they could not appeal to prior beliefs about scale) in which a person deteriorated at a much greater rate relative to other processes depicted, such as the leafing-out of trees. If asked to identify the cause of the change, then, if Kant is fight, they would try to get time in the film to "flow equably," and hence would fasten on disease rather than aging. If this prediction is borne out, then Cheng would need to invoke something like principle four to explain the induction. It is possible to reply on her behalf that principle four has no beating on her project, the original induction of causes, as opposed to the use of causal knowledge in subsequent inductions. But the reply misses the point at issue. In dividing the problem of causal induction, I believe that she was trying to separate the role of general high-level causal principles from more specific types of causal information. And, although principle four cannot be used without some causal information or other, the principle itself is perfectly
Kant on Constructing Causal Representations
235
general. It would also be possible to reply that principle four concerns the selection of potential causes and not the induction from covariation between potential causes and effects to actual causes. Principle four would not operate along with her three hypotheses to induce cause from covariation; its function would be to screen out some covariational evidence as irrelevant. Again, however, insofar as Cheng wants to separate the task of isolating particular causes from the general principles governing the problem of causal induction, principle four is relevant to her theory. Does Kant's argument for an intimate connection between temporality and causality imply that Cheng must give up the claim that "perceived priority [of an O-state of a E-substance to its N-state and the C-state of a C-substance]" is an input to all causal inductions? This is a delicate question, and not just because "perceived" might mean "apparent." How should we characterize the epistemic states involved in the Newtonian procedure? Do subjects "detect" apparent motion, "recognize" it, "believe" it occurs, or "have a firm conviction" about it? For scientists trying to fathom the forces governing celestial motions, the latter, weightier cognitive attitudes seem fight. But what about children at the beginning of the task of inducing the causal structure of the world? Kant's fundamental antiHumean claim was that whatever cognitive heft our mental attitudes to temporal sequence, cause, and reality have, it is the same heft for each. Since we can grasp time as passing rather than standing still only through grasping alteration, and we can grasp alteration as opposed to illusion only if we can grasp some plausible cause, however inchoate, then our attitudes about the reality of an effect, and so the succession of an N-state upon an O-state, and our attitude towards a possible cause will be on a par. By contrast, Hume believed that questions of temporal precedence could be settled independently of causal hypotheses. In taking temporal priority as an input to causal reasoning, Cheng is thus siding with Hume. It is highly doubtful that Kant was fight that all temporal determinations require the mediation of causal knowledge. Nonetheless, Cheng may not wish to rule out the possibility that, in causal reasoning itself, our attitudes towards temporal and causal relations may sometimes be determined together. To allow for this compromise between the Humean and Kantian camps, Cheng could amend her position as follows: when it can be determined independently, the priority of one state of a substance to another and the simultaneity of the latter state with a potential cause is the input to the process of causal induction; where temporal priority cannot be established independently, subjects use apparent effects to solve simultaneously for a temporal order and a set of causal rules that are mutually consistent. 9
This is a significantly revised version of the paper I read at the conference. I'm grateful to many discussants for criticisms of the earlier version, and especially to Huw Price, Cliff Hooker and Mark Bickhard. I'm also grateful to Philip Kitcher for discussions of several key points.
9
236
Patricia Kitcher
References Bennett, J. (1971). Locke, Berkeley, Hume: Central themes. Oxford: Clarendon Press. Bullock, M., Gelman, R., & Baillargeon, R. (1982). The development of causal reasoning. In: Friedman (1982) (pp. 209-254). Cartwright, N. (1989). Nature's capacities and their measurement. Oxford: Clarendon Press. Cheng, E (1997). From covariation to causation: A causal power theory. Psychological Review, 104, 367-4-5. Cheng, E (2000). Causality in the mind: Estimating contexual and conjunctive causal power. In: E Keil, & R. Wilson (Eds), Cognition and explanation (pp. 227-253). Cambridge: MIT Press. Falkenstein, L. (1995). Kant's intuitionism. Toronto: University of Toronto Press. Friedman, M. (1992). Kant and the exact sciences. Cambridge, MA: Harvard University Press. Glymour, C. (2000). Bays nets as psychological models. In: E Keil, & R. Wilson (Eds), Cognition and explanation (pp. 169-197). Cambridge: MIT Press. Guyer, E (1987). Kant and the claims of knowledge. New York: Cambridge University Press. Harper, W. A. (1984). Subjective and objective sequence. In: W. A. Harper, & R. Meerbote (Eds), Kant on causality, freedom and objectivity. Minneapolis: University of Minnesota Press. Hume, D. (1739/1962). A treatise of human nature. L. A. Selby-Bigge (Ed.). Oxford: Oxford University Press. Kant, I. (1781/1787/1998). Critique of pure reason. In: P. Guyer, & W. Allen (Trans.), The Cambridge edition of the works of Immanuel Kant. New York: Cambridge University Press. Mackie, J. L. (1980). The cement of the universe: A study in causation. Oxford: Clarendon Press. Melnick, A. (1973). Kant's analogies of experience. Chicago: University of Chicago Press. Newton, I. (1689). Philosophiae naturalis principia mathematica, Bk 1, Motte, A (transl.) revised by E Cajori. Berkeley: University of California Press (1934). Pearl, J. (2000). Causality: Models, reasoning, and inference. Cambridge University Press. Richards, D. D. (1982). Children's time concepts: Going the distance. In: W.J. Friedman (Ed.). The developmental psychology of time (pp. 13-45). New York: Academic Press.
Author Index Adams, E, 95, 209, 210 Aitchison, I. J. R., 72 Aizawa, K., 209, 210 Armstrong, D. M., 107, 121 Arnauld, A., 181 Arterberry, M. E., 137-141 Assmann, J., 206 Atick, J., 30 Ayer, A. J., 164 Bach-y-Rita, R., 138 Baker, L. R., 3 Bakhurst, D., 207 Bardy, B. G., 132, 142 Barlow, H. B., 30 Bartlett, E C., 207 Bechtel, W., 164, 165 Beer, R. D., 44, 47 Bennett, J., 223 Berkeley, G., 131, 137 Bickhard, M. H., 56, 57, 65, 72, 75, 77-86 Bickle, J., 192-194 Bjorklund, D. E, 140 Blachowicz, J., 10 Block, N., 3, 13, 107, 113, 125-128, 130, 131,134, 166, 183 Boden, M., 57 Borton, R. W., 138, 139 Braitenberg, V., 43 Brewer, W., 197 Brooks, R. A., xviii, 46, 47, 50-52 Brown, H. R., 72 Bryson, J., 50, 51
Bullock, M., 218, 227, 228 Bunge, M., 10, 11 Campbell, D. T., 77 Campbell, J., 57, 198 Campbell, R. L., 80, 81 Carruthers, E, 177 Cartwright, N., 223 Cassirer, E., 136, 137 Chalmers, D., 209, 210 Chang, C. C., 74 Cheng, E, 217-219, 221,223, 224, 226-228, 230, 231 Chipman, S., 9, 10 Christensen, W. D., 50, 51, 53, 59, 60, 62, 64, 65, 78 Churchland, E M., 15-17, 21, 150, 193, 194 Churchland, E S., 150, 193, 194 Clark, A., 131 Connerton, E, 206 Corballis, M., 197 Courage, M. L., 202 Coyle, J. T., 190 Craver, C. E, 194 Cummins, R., xvi, 2, 3, 5, 8-10, 13, 14, 22-26, 29, 30, 35, 83, 91,104, 171 Dancy, J.,133, 134 D'Andrade, R., 208 Darden, L., 194 Davidson, D., 91, 177, 178 Davies, M., 151 Davies, E C. W., 72
238
Author Index
Dawar, A., 74 Degenaar, M., 136, 137 Dennett, D. C., xiii, 57, 67, 148, 150, 152, 159, 161,170, 180 Donald, M., 204, 209, 210 Dretske, E I., 2, 3, 8, 23, 83, 87, 92, 93, 96, 142, 148, 158, 163, 164, 173, 174 Dreyfus, H., 43 Edwards, D., 198 Elman, J. L., 48 Engel, S., 188, 199, 201,202 Evans, G., 136-138, 142, 143 Falkenstein, L., 223 Fara, E, 190 Fentress, J., 205 Field, D. J., 29-31 Field, H., xii, 2, 13 Files, C., 8 Fischbach, G. D., 190 Fivush, R., 202, 203 Fodor, J. A., 2, 3, 8, 21, 23, 27, 37, 82, 92, 148, 163-169, 171,174, 177, 180, 182 Friedman, M., 231,232 Gardenfors, R, 10 Gibson, J. J., 142 Girvin, J. R, 138 Glymour, C., 222 Godfrey-Smith, E, 92, 151 Gold, I., 193 Goldsmith, M., 198 Gregory, R. L., 163 Grice, H. E, 113, 132 Griffiths, E E., 209 Grush, R., 64 Guyer, E, 218, 229 Haaken, J., 201 Hacking, I., 198
Halbwachs, M., 205, 207 Hamilton, R. H., 138 Hardcastle, V. G., 194-196 Harman, G., 13, 107, 127, 128, 133, 134, 143 Harper, W. A., 230 Harr6, R., 72 Hasemann, J. M., 56 Haugeland, J., 23 Hayne, H., 202 Heil, J., 3, 142 Hella, L., 74 Hendriks-Janson, H., 42, 49, 50 Hexmoor, H., 56 Hilbert, D., 143 Hilts, P. J., 195 Hoerl, C., 198, 202 Hooker, C. A., 51, 53, 59, 60, 62, 64, 65 Hooker, C., 193 Howe, M. L., 202 Hubel, D. H., 22, 30, 32 Hume, D., 222 Hutchins, E., 58, 209 Hylton, P., 86 Iran-Nejad, A., 207 Jackendoff, R., 164, 165 Jackson, E, 108, 111-113, 115, 116, 133, 152, 153, 209 Janlert, L. E., 56 Jenkins, W., 103 Johnson, M. K., 199, 200 Johnson-Laird, E, 10 Kaas, J., 103 Kandel, E. R., 189, 195 Kant, I., 218, 219 Kashima, Y., 207 Kay, L., 138 Keisler, H. J., 74 Kellman, E J., 137-141
Author Index
Kelso, J. A. S., 44, 48 Kim, J., 71,192 Kitcher, E, 190-193 Klein, N. M., 211 Klowden, M. J., 54 Koch, C., 32 Kolaitis, Ph. G., 74 Koriat, A., 198 Kosslyn, S., 166, 178 Kripke, S., 113 Kuffler, S., 30 Kujala, T., 138 Kwint, M., 211 Lagrou, P., 205 Le Goff, J., 205 Leahey, T. H., 133, 135 Legg, C., 66 Leibniz, G., xii LePore, E., 3, 37, 182 Levine, A., 82 Lewicki, M. S., 32 Lewis, D., xi, xiii, 115 Leyton, M., 204, 211 Livingstone, M., 32 Lloyd, D., 2, 181 Loar, B., 13 Locke, J., 136, 164 Loewer, B., 3 Lycan, W. G., xx, 91,107, 157, 159 MacDonald, S., 203 Mackie, J. L., 144, 217, 221,225,232 Malcolm, C., 56 Malebranche, N., 164, 165 Mandler, J. M., 202 Martin, M., 132 Mataric, M. J., 50 McClamrock, R., 206 McClelland, J. L., 201 McCormack, T., 202 McDermott, K. B., 198 McDonough, L., 202
239
McGinn, C., 13 McKeich, M., 211 Mele, A., 3 Melnick, A., 229 Meltzoff, A. N., 138-141 Meredith, M. A., 136 Merleau-Ponty, M., 43, 63 Merzenich, M., 103 Metzler, J., 10, 166 Middleton, D., 198 Miller, P. J., 203 Millikan, R. G., 2, 8, 82, 92, 118, 159, 173 Mitchell, K. J., 199 Montague, P. R., 54 Moore, G. E., 109 Moore, M. K., 139-141 Moreno, A., 77 Morgan, M., 136 Mullen, M. K., 203 Munakata, Y., 31 Nadler, S., 165 Nagel, E., 101 Nakayama, K., 141 Nelson, K., 202 Newell, A., xii, xiii, 29, 179, 180 Newton, I., 231 Norman, D. A., 181 O'Brien, G., 4, 10, 15, 18, 210 Olick, J. K., 206 Olshausen, B. A., 29-32 Opie, J., 4, 15, 18 O'Reilly, R. C., 31 Palmer, S., 10 Papineau, D., 105 Pascual-Leone, A., 138 Patterson, K., 190 Peacocke, C., 111,120, 129 Pearl, J., 228 Perner, J., 202
240
Author Index
Pettit, P., 3, 112, 152, 153 Pfeifer, R., 47, 51 Piaget, J., 85, 137, 140 Place, U. T., 128 Port, R., 44, 45 Putnam, H., 113, 166, 176 Pylyshyn, Z., xii, 166, 178 Quartz, S. R., 59 Quinn, N., 208 Raft, R., 50 Rajaram, S., 200 Ramsey, W., 157 Raphael, B., 56 Raye, C. L., 199 Real, L. A., 54 Recanzone, G., 103 Reese, E., 203 Rey, G., 147 Reynolds, S. L., 164 Richards, D. D., 233 Richie, D. M., 65, 81, 86 Robbins, J., 206 Roediger, H. L., 198, 200 Rose, S., 189 Rosenschein, S. J., 180 Rosenthal, S. B., 80 Rovee-Collier, C., 202 Rowlands, M., 105,209 Rubin, D. C., 199 Ruffman, T., 202 Ruiz-Mirazo, K., 77 Rumelhart, D. E., 181,208 Sacks, O., 137 Saito, A., 207 Saunders, S., 72 Scarry, E., 192 Schacter, D. L., 190, 193-196, 198-201 Schaffner, K., 193, 194 Schechtman, M., 201
Scheier, C., 47, 51 Schiffer, S., 13, 21, 22 Schudson, M., 206 Sciama, D. W., 72 Searle, J., 177, 179 Segal, G., 112 Sejnowski, T. J., 59 Sellars, W., xii Shepard, R. N., 9, 10, 166 Sherman, S. M., 32 Shimojo, S., 141 Simon, H. A., 29, 179, 180 Slezak, P., 164, 166, 178 Smith, B. C., 180, 208 Smith, L. V., 44, 58 Smithers, T., 44, 47, 48 Squire, L. R., 189, 195 Stalnaker, R., 115, 169 Stampe, D., 8, 23 Stein, B. E., 136 Stein, L., 51 Sterelny, K., 91,147 Stich, S. P., 147, 148 Stoffregen, T. A., 132, 142 Stoljar, D., 193 Stone, T., 151 Stotz, K., 209 Strauss, C., 208 Suddendorf, T., 197 Sutton, J., 191, 193, 207, 210 Swoyer, C., 10, 11 Terveen, L., 56, 57, 65, 79, 81, 85 Thelen, E., 44, 48, 58 Tofts, D., 211 Tomasello, M., 198 Tulving, E., 194, 195, 197, 199 Tuomela, R., 205 Tye, M., 107, 109, 120 Van Gelder, T., 44, 45, 48, 209 Von Eckardt, B., xv, 2, 4, 5-7, 14, 192
Author Index
Warfield, T. A., 147 Warnock, M., 191 Weinberg, S., 72 Welch-Ross, M., 203 Wickham, C., 205 Wiesel, T. N., 22 Winch, P., 178 Winsler, A., 207
Wittgenstein, L., 87 Woods, W. A., 180 Yi, S., 203 Young, J., 205 Zelizer, B., 206 Zhang, S. W., 174
241
This Page Intentionally Left Blank
Subject Index accuracy, 36-38 activation space, 16, 17 adaptive interaction, 42, 43, 54, 56, 58, 62-64, 66 asymmetric dependency, 82 attitudes vs mental representation, 1-22 autonomous agent robotics (AAR), 46-49, 55 autonomy, 77 blind sight, 119 causal constraint, 6-8, 14, 18 causal induction, 230-235 causal power theory of the probabilistic contrast model, 218 causal powers, 223-225 causal theories, 8, 14 Chinese Room, 179 cognitivism, 42-45, 48, 49, 55, 56, 58, 61-63, 66 Cog project, 51, 52 connectionism, 15, 191,208, 210 constructive remembering, 197, 198 content, 5-8, 85-87, 110, 114, 120, 121 and causation, 114, 116 egocentric, 115 narrow vs wide, 110-118 non-conceptual, 120-122 convention, 6, 7 craft, 152, 153, 155 cybernetic systems, 94-102
description theory of reference, 114 differentiation, 54, 59, 64, 77, 79-81 disjunction problem, 82, 105 dynamic presupposition, 78 dynamical systems theory (DST), 44-50, 55, 57, 62 dynamically situated approaches, 47-55, 58-66 eliminativism, 50, 62, 152 emergence, 43, 45, 48, 59, 64, 71-73, 75-77, 79 encoding structure, 29 error (misrepresentation), 3, 47, 65, 79, 81-85, 87, 88, 104, 127, 150, 163, 164, 166, 168, 170-179 etiological models, 75, 77, 87 extended mind, 209 externalism, 110-118, 131,169, 209 face recognition network, 15-17 folk psychology, 147-160 frame problem, 56-58 function, 77, 78 functional role theories, 13, 14 Gettier cases, 166, 172, 173 habituation method, 139 homuncular functionalism, 159, 160, 168 homunculus problem, 170, 171 Hume's problem, 170, 171 illusion, 134, 164, 166 indeterminacy, 110
244
Subject Index
indication, 22-24 information, 1, 2, 10, 25, 27, 29, 30, 35, 45, 58-60, 77, 158, 159 intenders, 34, 35 intentionality, 62, 103 derived vs original, 7 interactivism, 74 interdisciplinarity, 187-211 intermodal perception, 136-141 internalism, 118, 130 internal relation, 86 interpretation, 4, 5, 7, 151-155, 193 isomorphism, 9, 13, 14, 84, 87, 210 language of thought (LOT), 27-30 memory, 187-211 collective, 204-208 episodic, 197 explicit vs implicit, 194-196 external, 208-211 mental causation, 3 model, 154 modelling, 61 Molyneux's question, 136-141 motivation principle, 52 multidisciplinary research, 50 naturalism, 4, 6, 7, 14, 18, 71, 72, 75, 167, 187, 206 non-mental representation, 4 non-stationary learning systems, 59 normativity, 71, 75, 77, 78, 83, 84, 87, 93, 96, 98, 100, 101,103, 104 phenomenal qualities, 125-136 phenomenism, 126, 142-144 portability (source independence), 25, 33
process metaphysics, 73 propositions as targets, 35, 37 reduction, 158, 192-194 representationalism, 107-110, 120 representationism, 133, 143, 144 resemblance, 8-14, 167 and functional role theories, 13, 14 first order, 9 second order, 18 structural, 15-18 schema theory, 206, 208 self-directedness, 41, 53-55, 60, 62, 64 self-maintenance (recursive), 76, 77 sense data, 166 sense datum theory, 108, 109 source monitoring, 199-201 structural isomorphism, 84 structuralist theories, 14-18 subsumption architecture, 47, 48, 50-52, 58 swampman, 113 symbol systems, 59 system detectable error, 79, 83 targets, 33-38 teleosemantics, 64, 92, 93 triadic analysis of mental representation, 4-8 Turing machine, 7 twin earth, 112, 113, 115, 166, 176-177 understanding, 179, 183 use theories, 2, 3 visual imagery, 178 visual representation, 30-33