TELEOSEMANTICS
This page intentionally left blank
Teleosemantics New Philosophical Essays Edited by
GRAHAM MACDONA...
29 downloads
599 Views
1MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
TELEOSEMANTICS
This page intentionally left blank
Teleosemantics New Philosophical Essays Edited by
GRAHAM MACDONALD AND DAVID PAPINEAU
CLARENDON PRESS · OXFORD
1
Great Clarendon Street, Oxford Oxford University Press is a department of the University of Oxford. It furthers the University’s objective of excellence in research, scholarship, and education by publishing worldwide in Oxford New York Auckland Cape Town Dar es Salaam Hong Kong Karachi Kuala Lumpur Madrid Melbourne Mexico City Nairobi New Delhi Shanghai Taipei Toronto With offices in Argentina Austria Brazil Chile Czech Republic France Greece Guatemala Hungary Italy Japan Poland Portugal Singapore South Korea Switzerland Thailand Turkey Ukraine Vietnam Oxford is a registered trademark of Oxford University Press in the UK and in certain other countries Published in the United States by Oxford University Press Inc., New York the several contributors 2006 The moral rights of the authors have been asserted Database right Oxford University Press (maker) First published 2006 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, or under terms agreed with the appropriate reprographics rights organization. Enquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above You must not circulate this book in any other binding or cover and you must impose the same condition on any acquirer British Library Cataloguing in Publication Data Data available Library of Congress Cataloging in Publication Data Data available Typeset by Laserwords Private Limited, Chennai, India Printed in Great Britain on acid-free paper by Biddles Ltd., King’s Lynn, Norfolk ISBN 0–19–927026–0 978–0–19–927026–2 ISBN 0–19–927027–9 (Pbk.) 978–0–19–927027–9 (Pbk.) 1 3 5 7 9 10 8 6 4 2
Contents List of Contributors
1. 2. 3.
4.
5. 6. 7. 8. 9.
10.
Introduction: Prospects and Problems for Teleosemantics Graham Macdonald and David Papineau Language, Modularity, and Evolution Kim Sterelny Mental Representation, Naturalism, and Teleosemantics Peter Godfrey-Smith Representation, Teleosemantics, and the Problem of Self-Knowledge Fred Dretske The Epistemological Objection to Opaque Teleological Theories of Content Frank Jackson Useless Content Ruth Millikan On Thinking of Kinds: A Neuroscientific Perspective Dan Ryder Teleosemantics and the Consumer Mohan Matthen Content for Cognitive Science Karen Neander Representation and Unexploited Content Robert Cummins, Jim Blackmon, David Byrd, Alexa Lee, and Martin Roth Fearing Fluffy: The Content of an Emotional Appraisal Carolyn Price
Index
vii 1 23 42
69
85 100 115 146 167 195
208 229
This page intentionally left blank
List of Contributors Jim Blackmon is a doctoral student in the Philosophy Department at the University of California at Davis. David Byrd is a Lecturer in the Philosophy Department at Santa Clara University. Robert Cummins is Professor in the Philosophy Department at the University of Illinois-Urbana/Champaign. Fred Dretske is Emeritus Professor of Philosophy at Stanford University. Peter Godfrey-Smith is Professor in the Philosophy Department at Harvard University. Frank Jackson is Distinguished Professor of Philosophy and Director of the Research School of Social Sciences, the Australian National University. Alexa Lee is a doctoral student in the Philosophy Department at the University of California at Davis. Graham Macdonald is Professor of Philosophy in the School of Philosophy and Religious Studies, University of Canterbury, New Zealand, and Distinguished International Fellow, Institute of Cognition and Culture, Queen’s University, Belfast. Mohan Matthen is Canada Research Chair in Philosophy, Perception, and Communication at the University of Toronto. Ruth Millikan is Board of Trustees Distinguished Professor Emerita in the Philosophy Department at the University of Connecticut. Karen Neander is Professor of Philosophy at the University of California at Davis. David Papineau is Professor of Philosophy of Science at King’s College London. Carolyn Price is Lecturer in Philosophy at the Open University. Martin Roth is an Assistant Professor of Philosophy at Knox College, Illinois. Dan Ryder is an Assistant Professor in the Department of Philosophy at the University of Connecticut. Kim Sterelny is a Professor in the Philosophy Department at the Australian National University and at the Victoria University of Wellington, New Zealand.
This page intentionally left blank
Introduction: Prospects and Problems for Teleosemantics Graham Macdonald and David Papineau 1. NATURALISM The programme that has come to be known as ‘teleosemantics’ aims to offer a naturalistic account of mental representation. That is, it aims to show how the representational powers of mental states fit into the world revealed by the natural sciences. It is distinguished from other naturalistic approaches to mental representation by its reliance on a notion of function that plays a prominent role in biology: thus it explains the truth conditions of belieflike states, say, or the satisfaction conditions of desire-like states, in terms of the biological functions of these states. In the next section we shall give fuller details of the ways in which teleosemanticists have used the notion of function to explain representation. But first it is worth considering exactly how teleosemantics might contribute to the general project of naturalization. The notion of ‘naturalism’ is much contested. The requirement that we should only credit those facts that are recognized by the ‘natural sciences’ has little bite in the absence of some substantial specification of what qualifies a domain of investigation as a ‘natural science’. For many contemporary philosophers, the real significance of naturalism lies in the need to explain what role, if any, mental and other prima facie non-physical phenomena play in the physical world. The rationale for this emphasis is that physics, unlike other natural sciences, seems to be causally complete: physical effects always have physical causes, if they have causes at all. This means that any facts that are capable of producing physical effects cannot be ontologically supplementary to the physical realm; otherwise there would be an absurd proliferation of causal overdetermination. For example, since mental facts often have physical effects in the form of bodily movements and consequent
2
Graham Macdonald and David Papineau
impacts on the wider world, this argues that they cannot themselves lie outside the physical realm, for if they did then the bodily movements and so on would have non-physical causes as well as physical ones. Still, the doctrine that the mental realm cannot be outside the physical realm is less straightforward than it seems. For one thing, it is unclear how ‘physical’ is to be understood in this context. As Carl Hempel pointed out some decades ago (Hempel 1969), our present best physical theory is likely to be overtaken in the future by a better theory, which argues that we would be wrong to constrain our naturalizing ambitions by our present understanding of the physical. On the other hand, the constraint that our ontology must respect some future physical theory seems no constraint at all, given our ignorance of that future theory. So any attempt properly to articulate the idea that the mental cannot be separate from the ‘physical’ threatens to make this doctrine either overly restrictive or quite empty. It seems to us that this challenge can be satisfactorily answered. The recent literature contains a number of alternative suggestions for defining ‘physical’ in a way which leaves physicalism both plausible and contentful (for a brief survey of the options, see Jackson 1998: 6–9). However, even given such a definition of ‘physical’, there is another familiar respect in which physicalism about the mental needs further elucidation: how strictly should we read the requirement that the mental not be ‘ontologically supplementary’ to the physical realm, as we put it above? Few naturalists would want to subscribe to ‘type physicalism’, in the sense of requiring each respectable mental property to be strictly identical with some property describable in the language of physics. Rather, ontological indistinctness from the physical is widely agreed by physicalists to require only that mental properties should supervene on physical facts, in the sense that they are metaphysically determined by the physical facts. But now this threatens to remove the teeth from naturalism once more. Supervenience on the physical is not a strong requirement: for example, moral properties are generally supposed to supervene on physical properties, even by philosophers who would strongly resist the idea that moral facts can somehow be investigated by the methods of the natural sciences. Still, the demands of supervenience are not empty. If you hold that certain properties, while not type-identical to physical properties, are nevertheless metaphysically supervenient on them, then surely you owe some explanation of why this should be so. What is it about mental properties, say—or indeed moral properties—that makes it the case that a mental or moral difference must be due to a physical difference? A satisfactory answer need not type identify mental or moral properties with physical properties, but it will need to give some account of the nature of these properties that
Prospects and Problems for Teleosemantics
3
will explain why their instances should be metaphysically fixed by the physical facts. The view that representational facts are functional facts can be seen as an answer to this challenge. The concept of function that is used in biology is itself a contested notion. In fact, it is likely that there is no ‘one’ notion of function employed in biology. We shall consider some alternative analyses of ‘function’ below. But, on any account, two things are clear. First, functional properties are a paradigm of properties that are not type-reducible to physical properties. There is no strictly physical property that is necessary and sufficient for being a wing, say. All it takes for something to be a wing is that it have the function of enabling flight: beyond this there is no limit to the physical variety of different kinds of wing. And the same goes for other familiar functional categories in biology, like being a stomach, or a heart, or an eye. Second, despite this lack of type reducibility, it is clear that the functional facts are metaphysically determined by the physical facts. Two items could not possibly have all the same physical features (including their physical histories and environments) yet not have the same functional features. Once the (wide) physical properties of something are given, then its functional nature is fixed. (Different analyses of biological function will explain this supervenience on the physical in different ways, but all will agree that the functional facts do supervene on the physical facts.) So an analysis of representational facts as functional facts will imply that representational properties are not type-identical to physical properties, yet at the same time will explain why representational facts must supervene on the physical facts and thus be naturalistically acceptable. Even if teleosemantics does not type reduce representational properties to physical properties, for the reasons just explained, it may reduce them to biological properties. (Thus Ruth Millikan, the most prominent of teleosemanticists, entitled her first book Language, Thought, and Other Biological Categories, 1984.) In defence of this biologically reductionist view, it can be observed that teleosemantics aims to offer an explicit account of representational properties by appealing to a notion of function that is used in biological theorizing. On the other hand, it is unclear whether the facts to which teleosemanticists reduce representational properties should really qualify as biological facts, given that they standardly involve cognitive mechanisms that would normally be counted as in the realm of psychology rather than biology. In the end, we do not think that much hangs on whether we think of teleosemantics as a type reduction to biological properties. The more important point is that teleosemantics offers a naturalistically acceptable explanation of representation, whether or not we also count this as a biological reduction.
4
Graham Macdonald and David Papineau
2. THE TELEOSEMANTIC PROGRAMME In this section we will explain the basic strategy of teleosemantics and give some indication of the different ways this strategy has been developed by different theorists. The simplest way to introduce the teleosemantic programme is to contrast it with an alternative naturalistic approach to mental representation, namely, causal or indicator semantics. On the latter view, the content of a belief-like cognitive state is that condition that typically causes the state, and which the state therefore indicates (Stampe 1977; Dretske 1981). The standard objection to this causal approach is that it has trouble explaining misrepresentation. Misrepresentation by a belieflike state occurs when the state is tokened, but its truth condition is not. However, if the state’s truth condition is simply the range of circumstances that cause the state to be tokened, then it is unclear how the state can be tokened and yet its truth condition not obtain. Take a state that intuitively represents the presence of a snake. Such a state will often be caused, not by real snakes, but also by glimpses of slithery animals, toy snakes, and so on. The problem for the causal theory is that it has no obvious way of excluding these misleading extra causes from this state’s truth condition. So the causal theory seems to end up implying, absurdly, that all tokenings of this belieflike state are true. The teleosemantic approach, by contrast, explains the content of a belieflike state, not in terms of its typical causes, but in terms of how it is biologically designed to function. The snake-registering state will be designed to prompt behaviour that is advantageous specifically in the presence of real snakes, and not in the presence of harmless slithery animals, or toy snakes. So the truth condition of the state is specifically snake, and accordingly it will misrepresent the environment when it is prompted by other causes. Let us spell this out in a little more detail, using some notions introduced by Ruth Millikan (1984, 1993). She distinguishes the mechanisms that produce mental representations from those that consume them. The producing mechanisms will be the sensory and other cerebral mechanisms that give rise to cognitive representations. The consumer mechanisms will be those that use these representations to direct behaviour in pursuit of some biological end. Now, biological functions are in the first instance always a matter of effects: a trait’s function is that effect it is supposed to produce. So the function of a mental representation will lie in the way it contributes to the biological end of the mechanism that consumes it. More specifically, its function will be to enable the consumer mechanism to achieve its end by gearing behaviour to circumstances. Given this, we can think of the
Prospects and Problems for Teleosemantics
5
representation’s truth condition as the circumstance that enables it to fulfil this function—that is, the circumstance in which the behaviour it prompts is designed to produce the consumer mechanism’s end. For example, if we think of the snake representation above as being consumed by a mechanism whose function is to avoid snake bites, then the representation’s truth condition will be snake, rather than harmless look-alikes, since the snakeavoidance behaviour has been designed to have a positive result specifically when a real snake is present. Viewed in this way, teleosemantics has close affinities with the ‘success semantics’ analysis of content within the context of everyday belief–desire psychology (Whyte 1990, 1991). According to success semantics, the truth condition of a given belief is that circumstance which will ensure the satisfaction of whichever desire combines with the belief to prompt action. (More intuitively, what shows that you believe p is that you choose behaviour that will satisfy your desires if p. What shows you believe a snake is present is that you act in way that is sensible if a snake is present.) To see success semantics as a special case of Millikan’s version of teleosemantics, we need only equate the consumer mechanism for a belief with the decision-making processes that use that belief to select behaviour that will satisfy your currently active desires. Given this, the association of a success condition with a belief can be viewed as one example of the way Millikan’s analysis fixes the content of any belief-like representation as that circumstance under which a consumer mechanism guided by that belief will achieve its end. Not everybody views success semantics this way. Most philosophers who have developed success semantics view it, not as an aspect of teleosemantics, but as an alternative to it. And certainly there is no immediate appeal to biological function in the idea of a circumstance that ensures the satisfaction of the desire that combines with some belief to prompt action. On the other hand, the success semantics programme is in one sense obviously incomplete as an explanation of mental representation: it explains truth conditions for beliefs in terms of satisfaction condition for desires. A full account of mental representation needs to set the assumptions of success semantics in a wider context, so as to yield an explanation of desire satisfaction as well. And one obvious option here is to appeal to the notion of biological function, and say that the satisfaction condition of a desire is that result which the desire has the biological function of producing. The resulting combination of success semantics and teleosemantics has been developed in some detail by David Papineau (1984, 1993). This kind of approach is ‘top-down’—it starts with the kind of complex cognitive structure assumed by everyday belief–desire psychology—by contrast with
6
Graham Macdonald and David Papineau
Millikan’s ‘bottom-up’ approach—which begins with primitive biological representations of danger and food in simple non-human animals. In favour of Millikan’s strategy, there is the obvious advantage of more general applicability, and moreover her approach avoids the danger that everyday belief-psychology may offer a misleading picture of actual human cognitive structure. On the other hand, a full account of mental representation will need to cover human cognition too, and Papineau’s approach offers one possible account of this. In the end, perhaps the two approaches are best thought of as complementary rather than competing. Rather than trying to adjudicate between these alternatives, let us focus instead on one feature they have in common, and which differentiates them strikingly from causal or indicator semantics, and indeed from nearly all traditional philosophical analyses of content. Note that, on either account, the processes giving rise to representations (the producing mechanisms) play no particular role in the above analysis of content. On traditional approaches, by contrast, the content of a representation is taken to be some kind of function of the conditions that give rise to the representation, or correctly give rise to it, or verify it, or some such. These approaches thus make it relatively difficult, so to speak, for a representation to be false. (Of course, they aim to avoid the charge, levelled at indicator semantics above, that they make falsity impossible; but, even so, their general tenor is to make falsity unusual.) By contrast, the teleosemantic approach as explained above dissociates the determination of content from input conditions, and correspondingly makes it very easy for representations to be false. On the teleosemantic approach, content depends on how consumer mechanisms interpret representations. It depends on the behavioural output, not the informational input. The content is that condition under which the resulting behaviour would be appropriate, whether or not the actual circumstances that caused the representation are of that type. Take the snake representation again. This is a snake representation because it makes you behave in a way appropriate to snakes, given your biological ends. And this will remain the case even if you are pretty bad at recognizing snakes. The production mechanism for this representation may be triggered by toy snakes, by other slithery animals, indeed by the slightest hint of a slither, yet the representation will still stand for snake, if it is specifically snake-appropriate behaviour that it prompts. True, given that this representation has the content snake, in virtue of this being the condition under which it has functional effects, its producing mechanism will derivatively have the function of producing this representation only when it is true, that is, when snakes are present. Still, the fact that this mechanism has this function by no means implies that it will achieve it particularly often. To cite an oft-repeated example, sperm have the function
Prospects and Problems for Teleosemantics
7
of fertilizing ova, but only one in a zillion actually does this. Provided the pay-off from success sufficiently outweighs the costs of failure, biological mechanisms can have functions that they fail to achieve far more often than not. Indeed there are good reasons why we should expect a snake representation production mechanism to have just this structure: the pay-off from success (avoiding a real danger of snake bite) so far outweighs the costs of representing falsely (needless evasive action) that it makes biological sense to err on the side of caution, and produce the representation in response to even the most fallible signs of snakes. So teleosemantics, as so far outlined, stands diametrically opposed to the kind of input-based causal or verificationist theories that imply that false representations are atypical. Given the frequency with which false representations are in fact found, this would seem to count in favour of the teleosemantic approach. However, not all thinkers within the teleosemantic camp regard its commitment to output-based content as an unalloyed advantage. Consider the following well-known thought-experiment devised by Paul Pietroski (1992), and discussed by a number of contributors to this volume. The kimu are simple creatures, with very limited sensory abilities, whose only enemies are the snorf, who hunt them every day at dawn. A mutation endows one of them with a disposition to sense and approach red things. This disposition is a biological advantage to its possessors, since it leads them to climb a nearby hill every dawn, the better to observe the red sunrise, and means that they thereby avoid the marauding snorf, who do not climb hills. As a result, the disposition spreads through the kimu population. Now, consider the state a kimu gets into when it is stimulated by something red. It seems natural to credit this state with the content red. But an output-based teleosemantics argues differently. Nothing good happens to the kimu just because they approach something red. Most of their redapproaching behaviour is just a waste of time. It is only when this behaviour takes them away from the dangerous snorf that it yields any biological advantage. So an output-based teleosemantics will deem the state in question to represent snorf-free, or predator-free, or some such. This strikes many as strongly counter-intuitive, especially when it is further specified that the kimu cannot tell a snorf from a sausage, and would be perfectly happy to approach any snorf who happened to colour themselves red. Whatever the other virtues of teleosemantics, it seems wrong for it to conclude that the kimu’s state signifies snorf-free, rather than simply red. After all, by hypothesis the kimu’s senses are tracking the presence or absence of redness, not the presence or absence of snorf. There are alternative versions of teleosemantics that promise to analyse cases like these differently. These alternatives place more emphasis on the processes that produce representations than the purely output-based kind
8
Graham Macdonald and David Papineau
of teleosemantics described so far. For example, Fred Dretske builds his version of teleosemantics on a prior notion of indication (1988, 1995). Dretske first specifies that a type of state F indicates a type of state G just in case Fs never occur in relevant environments in the absence of Gs. In this sense, we can expect the products of sensory mechanisms to indicate the stimuli that trigger those outputs, independently of any biological advantages that may then ensue (and thus we can expect the kimu states to indicate redness rather than snorf-freeness). Of course, indication in itself does not amount to semantic representation, for the kind of reasons given earlier: if the indicated condition is always present when the indicator is, as the definition of indication requires, then the possibility of misrepresentation has not yet been explained. So Dretske specifies that true representation occurs only when F in addition has the function of indicating G—for example, the states of the visual system have the function of indicating features of the nearby environment. Since items with biological functions can sometimes fail to perform these functions, so F can on occasion fail to indicate G—this will happen when F is tokened in environments other than those where it has the function of being a sure-fire indicator of G. Dretske’s specific theory assumes a very strong notion of indication—Fs never occur in relevant environments without Gs—and this generates particular problems for his approach (see McLaughlin 1991). Other philosophers have focused on weaker notions of ‘indication’, requiring only that Fs be correlated with Gs, not that they are sure-fire signs of Gs, while still following Dretske’s suggestion that representation should be explained as a matter of states having the function of indicating something, rather the function of guiding behaviour towards some end. (Cf. Neander 1995, Chapter 8 in this volume.) However, whatever notions of indication they use, all such ‘input-based’ versions of teleosemantics face difficulties in explaining which correlations between Fs and Gs count for representational purposes. There will be many different kinds of Gs that any given type F is correlated with (even if we require, with Dretske, that the correlation be perfect over certain environments): a given cortical sensory state, for instance, will co-vary with kinds of surface stimulations of the sensory organs, and with intervening neural activations, as well as with a range of different distal external causes, such as slithery appearances, actual snakes, forked tongues, and so on. True, all versions of teleosemantics, including output-based ones, have some difficulty in explaining how representations can be directed at some specific option from such a range of alternatives (cf. Fodor 1990). However, output-based theories can at least rule out candidate contents that do not ensure that advantageous consequences will follow from resulting behaviour (such as slithery appearances,
Prospects and Problems for Teleosemantics
9
as opposed to actual snakes). But input-based theories cannot appeal to this resource, given that they aim to explain representational relations without attending to behavioural consequences, and so face greater problems in dealing with the threat of representational indeterminacy. What about the counter-intuitive consequences that Pietroski’s kimu thought-experiment seems to foist on output-based versions of teleosemantics? Perhaps there is room for output-based teleosemantic theories to argue that these intuitions depend on reading more into Pietroski’s scenario than is justified by his description. Pietroski says that the kimu evolve some state that is triggered by redness and which has the advantage of keeping them away from the snorf. Given this specification, it is hard to stop ourselves thinking of the kimu as having some generalpurpose visual system which gathers items of visual information which might then be used to inform an open-ended range of behavioural projects directed at different possible ends (such as avoiding blood, or finding postboxes, or indeed wanting to see red things). However, this extra structure in fact takes us significantly beyond what Pietroski’s description actually requires, and it is open to output-based teleosemanticists to argue that their theory is quite able to explain why an organism with all this extra structure would be representing redness rather than snorf-freeness: if the organism’s visual states inform a range of different behaviours directed at different ends, then the content of any such state needs to be fixed as some condition that assists in the achievement of all those ends, and this may well come out as redness. On the other hand, if we do stick to a minimal understanding of the snorf, as having only a special-purpose visual sensitivity that brings no advantage except snorf-avoidance, then it’s not so clear that there is anything wrong with the output-based reading of their states as representing snorf-freeness: after all, if these states never do anything except trigger simple avoidance behaviour, it seems natural enough to read them as representing the danger they are designed to avoid. 3. FUNCTIONS All teleosemanticists seek to analyse content in terms of biological function. But as yet we have said nothing explicit about the notion of biological function itself. In fact, this notion is much disputed. The two main contenders are the ‘historical–etiological’ view, and the ‘systems’ view that is analysed and defended by Robert Cummins. As it happens, the former is favoured by most teleosemanticists. But there is no necessary tie here, and in principle one could combine a teleosemantic approach to representation with a systems view of teleology.
10
Graham Macdonald and David Papineau
The systems account of functions focuses on complexity and aims to understand the complex functioning of biological and other systems in terms of the working of their parts (Cummins 1975, 2002; see also Millikan 2002). Many complex systems can be thought of a goal-directed, and in such cases the systems account of function offers one way of understanding how biologists are able to explain traits by citing their effects: on the systems account, such ‘functional explanations’ show how a part of a system contributes to the system’s production of some goal. Cummins distinguishes explanations of changes, which answer the question ‘What caused system S to acquire property P?’, from explanations of properties, which tell us ‘what it is for system S to instantiate property P’. Property explanations can be given by constructing an analysis which details the properties of S’s components and their organization. For example, the kinetic theory of gases explains what it is for a gas to have temperature via an account of the properties of the molecules contained within the gas. Cummins calls the analysis of such a system ‘compositional analysis’. He distinguishes ‘functional analysis’ as the analysis of dispositional properties, like the ability to see, or digest, or locomote. So for Cummins the ascription of a function to some part of a system specifies how the part has some capacity that helps the system to achieve some result. In biology, organisms are analysed into a number of systems (such as the digestive system, the circulatory system, and so on) with each system having a particular task to perform. How these tasks get performed is in turn explained by functional analysis, such an analysis citing the capacities of parts of these systems, these capacities being the functions of the parts. It will be a constraint on such analyses that the properties of the parts of the system be less sophisticated, less complex, than the property of the system being analysed. By its nature, functional analysis is most fruitfully applied to complex systems whose behaviour is produced by the coordinated action of simpler but well-organized parts. It is unsurprising, then, that the paradigms of systems suitable for functional analysis are those that have been designed to operate via the relevant interdependence of the parts. There is nothing in the systems account of functions, however, that restricts it only to designed systems: in principle any complex system can be subject to a systems-style functional analysis, including such non-designed systems as the terrestrial weather system or the Milky Way galaxy. The historical–etiological account of function, by contrast, explicitly restricts attribution of functions to traits that have in some sense been designed to produce the effect cited as the function. On this account of function, functions are the upshot of prior processes of selection. A trait has a function if it has been designed by some process of selection to produce some effect. In the central cases, where the traits in question are biological
Prospects and Problems for Teleosemantics
11
adaptations, the selection process will be non-intentional natural selection. An effect of a trait counts as its function if the trait has a certain history: in the past possession of that trait produced the relevant effect, which in turn had the consequence facilitating the reproduction of items with that trait. In such cases, it is natural to adopt teleological terminology, and say that, in the normal case, the trait exists because of an effect the trait can produce, or in order to fulfil its function. Take the streamlined shape shared by a number of large underwater predators—dolphins, blue marlin, great white shark. Here we have a trait (‘streamlinedness’) that serves these predators well, and they have the property, it seems, because it serves them so well (cf. Griffiths and Sterelny 1999: 245–6). The fact that a number of different species that don’t share a common ancestor have this trait suggests that the present members of the different species have the streamlined shape because, in the environment they share, that shape was reproductively more beneficial to their ancestors than lack of it was to their ancestors’ competitors. So, in the environment in which the adaptation arose, the shape has the function of facilitating swift movement, which enables more successful predation. Put as briefly and generally as possible, the etiological account says that functionality arises because some individuals in a group acquire novel traits with capacities that are favourable to their ability to reproduce. Such features are transmitted to their descendants, proliferating within the group in the process. Those features will then have as their function the exercise of the favourable capacity. Perhaps the main reason why teleosemantic theorists prefer the etiological to the systems view of functions is that it offers a strong notion of malfunction, which is something teleosemantics needs to account for mispresentation. We saw earlier how non-teleological semantic theories like causal and indicator semantics have trouble explaining mispresentation. (If states ‘represent’ whatever causes them, then how can they ever be tokened falsely?) Teleosemantics offered to deal with this by distinguishing those circumstances where a representation is ‘supposed’ to be present from other circumstances which may happen to cause it. In the latter kind of case, the representation is malfunctioning, that is, not doing what it is supposed to. Clearly, this story needs a robust account of when a trait is doing what it is supposed to do and when it is malfunctioning. It is not clear that the systems approach can offer such an account. On the etiological approach, a trait is supposed to do whatever its predecessors did that gave rise to a reproductive advantage, and the trait is malfunctioning when it doesn’t do this. By contrast, all that the systems account can offer is a statistical criterion: in most systems of a certain kind this kind of trait does F, so here the trait
12
Graham Macdonald and David Papineau
is malfunctioning in not doing F. By contrast with the etiological analysis, this statistical systems account seems to lack any normative content: it doesn’t seem to show that a trait in any sense ought to be doing F; it just says it isn’t doing F, and so is statistically unusual, but nothing more. So the historical–etiological approach to functions has the advantage of making the kind of principled distinction that seems necessary for a substantial account of malfunction in general and misrepresentation in particular. However, in tying themselves to historical–etiological functions for this reason, teleosemanticists might seem to run into a converse difficulty. As we saw earlier, the systems version attributes functionality to any traits that cooperate to produce distinctive behaviour in complex systems. By contrast, the historical theory allows for functions only in systems that have been subject to some process of design. When it comes to biological systems, including our human selves, the only available designer might seem to be the process of natural selection operating intergenerationally on gene frequencies. However, if biological functionality always derives in this way from genetic selection, it is surely unlikely that all representation, including such paradigms as human beliefs and desires, can be explained in terms of functions. After all, most human beliefs and desires are products of ontogeny rather than phylogeny, in the sense that no genes have been selected because they foster those specific beliefs or desires. So there seems no possibility of explaining the contents of these states in terms of etiological design-based functions. Fortunately for the teleosemantic project, the historical account need not restrict functionality to traits that are genetically based in the sense that specific genes have been selected because they give rise to those traits. There are ways in which biological items can be the products of design even though they have no specific genetic basis in this sense. In particular, there are two theoretical resources which often go unnoticed in this context. The first resource, utilized most by Millikan, appeals to a many-layered account of functions. The second resource depends on the idea of non-genetic selection. These resources greatly expand the range of items which possess etiological design-based functions.
Millikan Functions Central to the etiological account is the idea that individuals gain functional traits as a result of being replicated. Millikan (1984, 1993) offers a highly abstract account of replication. A simplified version goes like this: item A is a reproduction of individual B if and only if B has some determinate properties in common with A, and this correlation of properties can be explained by a natural law. These common properties are the reproductively established
Prospects and Problems for Teleosemantics
13
properties of B, and the items sharing these properties form reproductively established families (‘refs’). Genes can form such refs. Other items that are inherited as a result of genetic replication, such as eyes or hearts, can form higher-order refs (horefs). ‘Direct proper functions’ are etiological functions of traits possessed by items that are members of refs or horefs, where such functions are the result of past selection in those families. Millikan’s term for etiological functions is ‘proper’ function. She notes that one kind of proper function is a relational proper function, which is a function to do something only when bearing a certain relation to something else. Many fish, shrimp, and prawns can adjust their colour and pattern to the environment in which they find themselves. One prawn, Hippolyte varians, is called the ‘chameleon prawn’ on account of its ability to adapt its colour to its current environment across a range of yellows, greens, and browns (see Stephenson and Stewart 1955). The prawn’s camouflage mechanism has the relational function of making its colour match that of its environment, whatever that colour may be. Given a specific colour to adapt to, the mechanism then acquires an adapted proper function. So when the prawn is sitting on a particular brown weed, say, the adapted proper function of the mechanism is to make the prawn a matching shade of brown. Crucial here is that this precise colour may have never been produced before, so it is not a member of a ‘ref’, and the production of this particular colour is not a direct proper function of the camouflage mechanism, nor of anything else. One can extend this picture to novel representations within compositional syntactical system. Consider the famous dance of the bees, which acts as a signal to other bees, ‘telling’ them where to go to find nectar. This dance is adapted to the location of the nectar, so it has an adapted function. Again, the dance indicating this specific direction may never have occurred before. Rather, it owes its functionality to the syntax of a system that has been reproduced because possession of such a system has yielded a reproductive advantage in the past. (See Millikan, Chapter 5 in this volume, for further extensions of this approach).
Non-Genetic Selection The first resource enabling extension of the etiological strategy can be called derivative functionality: devices can have the (direct) function of producing effects that themselves have (derived) functions. The second resource available to the teleosemanticist is non-genetic selection. So far we have not paused to analyse the notion of selection. In fact this notion applies much more broadly than to genetically based selection. All it requires is a set of items that that have the characteristics of:
14
Graham Macdonald and David Papineau
(a) variability in the traits possessed, (b) selection of items with certain traits, (c) heritability of traits selected for. Selection cannot take place if there is no initial variation: the same selection forces operating on a homogeneous population will have no discriminating effect. When there is variation, items will then be selected for having some trait if that trait interacts with salient features of the environment in such a way that other items without that trait are seen to ‘suffer’ some loss by contrast. If these favoured traits are then transmitted to descendants of the items initially having those traits, the proportion of items with these traits will increase. Whenever these conditions are satisfied, even if no genes are involved, then it can be said that any selected trait is functional, its function being to produce those types of effects that lead to the differential reproduction of items with the trait in question. There are two different modes of non-genetic selection worth mentioning in this context. The first operates on normal organisms but involves non-genetic intergenerational inheritance. Many traits are passed from parents to children by channels other than the sexual transmission of genetic material: these traits will include the possession of parasites, the products of imprinting mechanisms, and the many cognitive and behavioural traits acquired from parents via social learning. A number of biological theorists are currently interested in the way in which such non-genetically inherited traits can be naturally selected through the normal Darwinian process of differential reproduction of organisms (Jablonka and Lamb 1999; WestEberhard 2003; Mameli 2004). The mere fact that these traits are transmitted non-genetically does not stop their possession satisfying the three conditions above, with advantageous traits consequently becoming prevalent for the standard reason that their possessors have more offspring and those offspring inherit the traits. Non-genetically inherited traits that become prevalent in this way will have functions, namely, the effects that favoured them. It is possible, though this is an area that has yet to be properly explored, that functions of this kind could do much to explain the contents of sophisticated mental representations. After all, it seems a natural enough thought that certain non-genetically inherited ways of thinking are an advantage to their possessors because they make them sensitive to certain features of the environment. On the other hand, it remains an open question how many features of human thought are in fact due to differential reproduction of offspring resulting from such advantages. The second mode of non-genetic selection is more familiar and perhaps more directly relevant to the teleosemantic project. This involves not the
Prospects and Problems for Teleosemantics
15
differential reproduction of organisms over generations, but the differential reproduction of cognitive or behavioural items themselves during the development of a given individual. Such ontogenetic selection takes place, for example, when behaviour is moulded by experience during learning. In such cases we can think of the items selected as having the function of producing those effects in virtue of which they were favoured by the learning mechanism. This kind of ontogenetic selection has been termed ‘vicarious’ or ‘secondary’ selection by Donald Campbell (1974). Campbell’s thought is that the relevant developmental mechanisms have themselves been selected for by genetically based natural selection to be non-genetic selectors. They operate so as to be less severe selectors than death, permitting learning and other adaptational processes to occur. Campbell developed an explicit ‘blind-variation-and-selective-retention’ (BVSR) model of learning. There were three essential aspects to Campbell’s BVSR model: (a) mechanism(s) for introducing variation, (b) consistent selection processes, (c) mechanism(s) for preserving and/or propagating the selected variants. As will be clear, Campbell’s requirements for selective learning correspond precisely to the requirements specified earlier for selection in general. Campbell thought that BVSR learning mechanisms are found throughout nature. At the most basic level, for example, an organism may avoid noxious substances when its chemoreceptors signal that the environment is becoming lethal. Here the chemoreceptor mechanism selects the behavioural responses, this mechanism itself having been selected for precisely this task. More generally, genetic selection spawns BVSR learning processes, which in turn can spawn higher such processes, all the way up the tree of knowledge. Campbell saw BVSR learning processes as yielding economies in the creation of knowledge. At their most sophisticated, such processes underpin the ways we use language to impart knowledge, with language itself functioning as a substitute for the individual organism’s perceptual investigation of its environment. As we observed above, the kind of ontogenetic selection dealt with by Campbell’s model of learning will yield cognitive and behavioural items with specific etiological functions, namely, those effects in virtue of which they were selected. Here too we will have functionality without the selection of genes, and this again expands the range of processes which can be subject to teleosemantic analysis. It should be said that there is as yet not a great deal of detailed work showing how teleosemantics might analyse sophisticated human modes of
16
Graham Macdonald and David Papineau
cognition by appealing to functions other than those deriving directly from the selection of genes. True, Millikan has indicated how her notion of an adapted proper function can be used to account for the representational contents of elements in complex representational systems. And Dretske has focused on selection in learning as one means by which to explain how cognitive states can be teleosemantically targeted on specific contents. Still, much remains to be done in applying teleosemantics to specifically human modes of cognition. Perhaps this is inevitable. Detailed analyses of representational powers in terms of etiological functions must rest on an adequate empirical knowledge of the cognitive mechanisms involved. There is no question of identifying the functions of cognitive items if we don’t know what kinds of mechanisms process these items and how those mechanisms develop in individuals. From this perspective, the teleosemantic project is not so much a theory of content for sophisticated human representation, but a methodology which promises to explain content piecemeal, in the wake of empirical discoveries about human cognitive architecture. Progress in teleosemantic accounts of human representation will come only along with empirical advances in cognitive science. We hope at least to have shown in this section that teleosemantics has plenty of resources to offer this long-term project.
4. SUMMARIES OF CHAPTERS The authors of the chapters in this volume were invited to contribute because of their manifest expertise in areas central to teleosemantics. Their approaches to that programme are diverse, as are their opinions as to whether it possesses the potential to be successful. As editors we hoped for a collection of articles that would enhance the discussion both of foundational matters and of specific problems within the general programme. We have not been disappointed. We are confident that these essays will provide a rich source of material stimulating much further debate, and are grateful to all our authors for their contributions. Close to the heart of teleosemantic theories is the thought that our linguistic capacity has provided our species with a crucial cognitive advantage over rival species. As Kim Sterelny notes (Chapter 1), one view of linguistic competence is that it is ‘massively modular’, encapsulated so as to make the role of experience limited to the ‘evocation’ of one alternative from a pre-existent, fully specified set of alternative languages. Sterelny argues that a modest modularity hypothesis is more plausible, given the assumption that language would have evolved in a Darwinian incremental fashion; his
Prospects and Problems for Teleosemantics
17
suggestion is that only the organizational, syntactic, aspect of language is modular. The proposal that our semantic understanding is analogous to perception, a kind of ‘natural telepathy’ in which beliefs are ‘transported’ from the speaker’s mind to the hearers’ minds, is resisted, but Sterelny does think that ‘natural telepathy’ could work for concepts. Sterelny concludes by warning against seeing the causal process involved in this transmission of concepts as one requiring modular mechanisms. The demands made on linguistic expression by an ever-changing environment necessitate more flexibility than is allowed for on the ‘massively modular’ conception of language. Peter Godfrey-Smith (Chapter 2) locates teleosemantics within the tradition of giving naturalistic explanations of the semantic properties of mental representations. He investigates one basic aspect of representation—the taking of one item, X, to tell us something about another item, Y—seeing this as one model for mental representation. He notes that philosophers commonly assume three features of this model: the user of the representation is distinct from the representation, the representation must have a ‘target’ established in some way, and the representation must be isomorphic, perhaps abstractly so, to what it represents. Godfrey-Smith looks at the work of some cognitive scientists to see how they implement the basic model, noting their lack of concern over the foundational issues troubling philosophers. Teleosemantics can be seen as a philosophical elaboration of the basic model, with the foundational issues, particularly the ‘target’ problem, to the fore. He concludes with a discussion of Ruth Millikan’s particular approach to these topics. Fred Dretske (Chapter 3) confronts an epistemological problem faced by any teleosemanticist, or, more broadly, any externalist. For Dretske, the content of a (mental) representational state is given by what it has the function of indicating about the world. Given the etiological account of what it is to have a function, it looks as though we would need to know the relevant history to know what we are thinking, a strongly counter-intuitive result. Dretske’s response is to separate out carefully two components in what we know when we know what we think. One is that we are thinking, the second is what we are thinking. Introspection can deliver knowledge of the second, but it may need some empirical inquiry (e.g. into history) to deliver knowledge of the first. The epistemology of our knowledge of intentional content is also discussed by Frank Jackson (Chapter 4), who argues that teleological theories appealing to selection cannot explain how it is that the folk have justified opinions about intentional contents. His contention is that such theories make contents consist of properties that the folk need never have heard of, and for that reason cannot have justified beliefs about.
18
Graham Macdonald and David Papineau
This, the ‘folk epistemology objection’, has been aired before, so Jackson replies to various responses to the objection (cf. Braddon-Mitchell and Jackson 1997, 2002; Papineau 2001). One such response claims that correlations between opaque selection states and transparent states could allow for access to the opaque content via the transparent state. But such a correlation, argues Jackson, won’t deliver justified opinion about the opaque content unless the folk have knowledge of the correlation—and in most cases, they won’t. Strengthening the correlation to identity fails the same test—the epistemic properties of (folk-accessible) intentional content differ from the epistemic properties of any selectional role states. Interpreting a teleosemantic theory along the lines of functional rolerealizing state theories, where the teleosemantic state is the realizer, also won’t do; the relevant content property will be the role property, the one to which the folk have epistemic access. Ruth Millikan (Chapter 5) addresses a concern that was raised some time ago by Christopher Peacocke (1992) that the content assigned to any representation by her teleosemantic theory would be essentially anti-realistic. It would be this because content would be sensitive to, and only to, selection processes, and selection can only operate on what is available to those processes. In other words, no selection-transcendent content could be assigned to any representation. But it seems as though we do understand content that has not itself been selected for, content that is ‘useless’ from the point of view of, say, reproductive advantage. Millikan shows how her teleosemantic approach can embrace the idea of ‘useless’ content, content that is derived from selection processes (however these are understood) but is not itself selected for. Her solution makes essential use of extended selection processes: any process involving trial-and-error learning counts as a selection process. Further, any set of systematic mapping rules that are selected for will contain the capacity to generate such ‘useless content’. This theme of non-selected (but still teleosemantic) representational content is pursued by Dan Ryder (Chapter 6), who applies the general form of Millikan’s notion of derived relational proper functions to the brain. Specifically he uses the idea of a modelling machine’s function to be the production of models of those items that are fed into it (its inputs) and applies it to the way in which natural kinds can be modelled in the brain. He shows how a cell can become tuned to a source (a bird, say) that is the location of constantly correlated features (feathers, a beak, etc.). Cellular networks tuned in this way model the environment, which is what they were designed (via natural selection) to do. Ryder shows how such cellular networks can meet a particular challenge, that of showing how the extension of one concept can be determined as different from that of another concept even though the two extensions have superficially resembling members.
Prospects and Problems for Teleosemantics
19
Mohan Matthen (Chapter 7) asks the question ‘What feature is represented by a perceptual experience?’ His answer is Millikanian: a perceptual experience represents an object, say, as having feature F if that is the normal condition for the successful performance of a function of a consumer of that representation. Is the consumer’s function univocal? Matthen claims it is; a perceptual experience, provided by a detector of the system, has the function of eliciting an epistemic response in the organism, where an epistemic response can be as basic as altering the potentialities of connected neurons. And he argues that stimuli evoking the same epistemic response are alike just in case they are treated as similar by the effector (the consumer of the representation). And the meaning (representational function) of a perceptual experience is given by the coordination scheme that emerges in the coevolution of the detector and effector systems. The consumer-oriented version of teleosemantics, favoured by Millikan and Matthen, is examined by Karen Neander (Chapter 8), who thinks it overlooks a plausible constraint on the content assigned to any representational system. The constraint is that the assigned content should play a role in the explanation of the behaviour, perception, or cognition of the organism. The explanatory role is further refined: the content should be suitable for use in explanations employed in mainstream cognitive science. A negative point she makes is that some teleological theories, those that see content being determined by that environmental feature representation of which selectively favoured past users of the representation, fail to respect this explanatory constraint. Neander argues that a theory of content must link the content to the information processing that informs the content. Regarding Pietrowski’s snorf-kimu example (see above), she notes that the information processed is colour information, so any theory assigning content to the representation such as ‘snorf-freeness’ would decouple content from information processing, and this would render it unsuitable for the explanatory purposes of mainstream cognitive science. Neander looks in detail at how toads distinguish their prey, and assesses the suggestion that ‘nutrient’ is the content playing a causal role in prey-catching behaviour. This suggestion is dismissed on the grounds that the toads do not have a capacity to detect nutrients; they do have the capacity to detect various features of the stimulus, these being associated with nutrients often enough for the discriminative capacity to be a fitness-enhancing capacity. Neander then examines several possible objections to her preference for assigning the ‘narrow’ content, finding none of them convincing. The theme of discriminatory capacities is further explored in ‘Representation and Unexploited Content’ by Robert Cummins, Jim Blackmon, David Byrd, Alexa Lee, and Martin Roth (Chapter 9). Their claim is that any theory of content suitable for representationalist theories in cognitive science must
20
Graham Macdonald and David Papineau
allow for the phenomenon of ‘unexploited content’, content that the system containing it is unable to use. One may have a representation at one’s disposal but be unable to exploit all of its representational features, perhaps because we have not been taught how to use those features. Teleosemantics, they argue, requires content to be truly ascribed only after the ability to exploit is acquired—so for the teleosemanticist there can be no ‘unexploited’ content. They trace this failure to take into account unexploited content to a tendency to conflate representation and indication. There are significant differences between the two, but both can have unexploited content, content that is there prior to selection—and, it is argued, the kind of content that is needed for representationalist cognitive science must allow for some part of it to be unexploited. The conclusion is that it is an objection to teleosemantic theories that they cannot accommodate unexploited content. As many of these chapters illustrate, teleosemantics has been used primarily to account for cognitive aspects of mentality. Carolyn Price (Chapter 10) applies the ‘High Church teleosemantic theory’ of Millikan to the determination of the content of emotional appraisals, these being intentional states (e.g. beliefs) triggering the occurrence of our emotions. A particular problem is how to distinguish such appraisals from dispassionate evaluative judgements. Price begins by listing various functions of emotional appraisals, such as providing motivation, focusing attention on relevant information, limiting the set of responses the subject will choose from, triggering expressive behaviour. Given this, she raises the question ‘What kind of content does an emotional appraisal have—descriptive, directive, or mixed?’ In the light of the variety of functions such appraisals are called upon to perform, Price suggests that a mixed, descriptive and directive, content is called for. Do evaluative judgements similarly have mixed content? She suggests that they do, but that the directive content of the two is different, in that the directive content of the emotional appraisal will be more detailed about the response than will the directive content of the evaluative judgement. The descriptive content of the emotional appraisal will also be tied to avoidable threats, this restriction not being applied to the descriptive content of evaluative judgements. Correlatively, the temporal content of an emotional appraisal will be restricted to the present, near past, or near future, whereas the temporal content of an evaluative judgement need not be so restricted.
REFERENCES B-M, D., and J, F. (1997), ‘The Teleological Theory of Content’, Australasian Journal of Philosophy, 75: 474–89.
Prospects and Problems for Teleosemantics
21
(2002), ‘A Pyrrhic Victory for Teleonomy’, Australasian Journal of Philosophy, 80: 372–7. C, D. (1974), ‘Evolutionary Epistemology’, in P. Schilpp (ed.), Essays in Honor of Karl Popper (La Salle, Ill.: Open Court). C, R. (1975), ‘Functional Analysis’, Journal of Philosophy, 72: 741–65. (2002), ‘Neo-Teleology’, in A. Ariew, R. Cummins, and M. Perlman (eds.), Functions: New Essays in Philosophy of Psychology and Biology (Oxford: Oxford University Press). D, F. (1981), Knowledge and the Flow of Information (Cambridge, Mass.: MIT Press). (1988), Explaining Behaviour (Cambridge, Mass.: MIT Press). (1995), Naturalising the Mind (Cambridge, Mass.: MIT Press). F, J. (1990), ‘A Theory of Content’, in Fodor, A Theory of Content and Other Essays (Cambridge, Mass.: MIT Press). G, P., and S, K. (1999), Sex and Death (Chicago: University of Chicago Press). H, C. (1969), ‘Reduction: Ontological and Linguistic Facets’, in S. Morgenbesser, P. Suppes, and M. White (eds.), Essays in Honor of Ernest Nagel (New York: St Martin’s Press). J, E., and L, M. (1999), Epigenetic Inheritance and Evolution (Oxford: Oxford University Press). J, F. (1998), From Metaphysics to Ethics (Oxford: Oxford University Press). ML, B. (ed.) (1991), Dretske and his Critics (Cambridge, Mass.: Blackwell). M, M. (2004), ‘Nongenetic Selection and Nongenetic Inheritance’, British Journal for the Philosophy of Science, 55: 35–71. M, R. (1984), Language, Thought, and Other Biological Categories (Cambridge, Mass.: MIT Press). (1993), White Queen Psychology and Other Essays for Alice (Cambridge, Mass.: Bradford Books, MIT Press). (2002), ‘Biofunctions: Two Paradigms’, in A. Ariew, R. Cummins, and M. Perlman (eds.), Functions: New Essays in Philosophy of Psychology and Biology (Oxford: Oxford University Press). N, K. (1995), ‘Malfunctioning and Misrepresenting’, Philosophical Studies, 79: 109–41. P, D. (1984), ‘Representation and Explanation’, Philosophy of Science, 61: 550–72. (1993), Philosophical Naturalism (Oxford: Blackwell). (2001), ‘The Status of Teleosemantics, or How to Stop Worrying about Swampman’, Australasian Journal of Philosophy, 79: 279–89. P, C. (1992), A Study of Concepts (Cambridge, Mass.: MIT Press). P, P. (1992), ‘Intentional and Teleological Error’, Pacific Philosophical Quarterly, 73: 267–82. S, D. (1977), ‘Toward a Causal Theory of Linguistic Representation’, in P. A. French, T. E. Uehling, Jr., and H. K. Wettstein (eds.), Midwest Studies in
22
Graham Macdonald and David Papineau
Philosophy, ii: Studies in the Philosophy of Language (Minneapolis: University of Minnesota Press). S, E. M., and S, C. (1955), Animal Camouflage (London: A. & C. Black; first pub. Harmondsworth: Penguin, 1946). W-E, M. (2003), Developmental Plasticity and Evolution (Oxford: Oxford University Press). W, J. T. (1990), ‘Success Semantics’, Analysis, 50: 149–57. (1991), ‘The Normal Rewards of Success’, Analysis, 51: 65–73.
1 Language, Modularity, and Evolution Kim Sterelny 1. LANGUAGE AND EVOKED CULTURE Language is at the core of the cognitive revolution that has transformed that discipline over the last forty years or so, and it is also the central paradigm for the most prominent attempt to synthesize psychology and evolutionary theory. A single and distinctively modular view of language has emerged out of both these perspectives, one that encourages a certain idealization. Linguistic competence is uniform, independent of other cognitive capacities, and with a developmental trajectory that is largely independent of environmental input (Pinker 1994, 1997). Thus language is seen as a paradigm of John Tooby and Leda Cosmides’ concept of ‘evoked culture’: linguistic experience serves only to select a specific item from a menu of innately available options (Tooby and Cosmides 1992). In explaining this concept, Tooby and Cosmides appeal to the metaphor of a jukebox. The human genome pre-stores a set of options, and the different experiences provided by different cultures select different elements out of this option set. I shall argue to the contrary that variability between speakers, the sensitivity of linguistic development to environmental input, and the limits of encapsulation are not noise. They are central to the language and its evolution. Evolutionary arguments about language face a problem. Evidentially robust theories of the evolution of language are in short supply. That is no accident. Language is unique; no other living species communicates with even a simple language.¹ Moreover, it leaves no direct trace in the ¹ The claim that there is a qualitative difference between human language and animal signal systems is not controversial, though there are conflicting views on the nature of the crucial differences: see Hauser et al. (2002), Jackendoff and Pinker (2005), and Pinker and Jackendoff (2005) for divergent views on the crucial organizational difference between language and other systems. Hauser and his allies think that recursive structure is the crucial novelty; Jackendoff and Pinker think that this understates the significance of the productive lexicon. Deacon, in contrast to both groups, argues that the crucial
24
Kim Sterelny
archaeological record: communicative abilities are evident, if at all, from their effects on material culture, local ecology, and demography. This problem cannot be completely solved. Nonetheless, a partial solution is possible by making only conservative assumptions about language and its evolution. For that reason I shall make no assumptions about the specific timing of the evolution of language, nor about the specific theories of syntax and semantics that are true of our language or its ancestors. But I shall assume (a) human and pre-human social life was characterized by some unstable mix of cooperation and competition; (b) within this mix, language has played a crucial role in facilitating coordinated, cooperative behaviour; (c) language is a complex adaptation that has evolved in a classically Darwinian way, i.e. from simple to more complex forms by relatively small increments. I shall not assume, however, that this evolutionary process has been entirely biological rather than cultural;² (d ) I shall assume that language expresses a speaker’s thoughts and that words express concepts. In turn, concepts, and hence words, typically stand in some reference-like relationship to items in the speaker’s environment. These assumptions are not wholly uncontroversial but they are not tendentious. Most defenders of a modular conception of language accept them (though some have rejected (c), as Pinker and Bloom (1990) complain). In the next two sections of this chapter I present two arguments, each of which puts the ‘evoked culture’ model of language under pressure. First I show a tension between this model and the idea that language evolved incrementally. In the following section, I discuss Millikan’s ‘natural telepathy’ image of language, arguing that interpreting others is not, contrary to her suspicion, a perception-like and hence modular cognitive capacity. In the final section, I combine these two arguments with other features of human evolution to argue that our environment has been too unstable for selection to hard-wire a specification of language into human minds.
2. LANGUAGE: AN ORIGIN MYTH AND ITS CONSEQUENCES I agree with Pinker and other defenders of the modularity hypothesis that language is very cognitively demanding, and the evolution of language must difference is semantic not organizational: terms in natural language are symbols, not natural signs (Deacon 1997). ² The biology–culture dichotomy is not a good one: (Oyama et al. 2001). But by ‘biological’ I mean that phenotypic similarity is transmitted from parent to offspring exclusively by the flow of genes; not through any form of social learning.
Language, Modularity, and Evolution
25
indeed traverse a cognitive bottleneck. Language cannot evolve without its cognitive preconditions evolving with it. But let me begin my softening-up by noting that many of these cognitive demands are not plausibly construed as requiring the evolution of a specialized module. For one thing, as Merlin Donald emphasizes, language requires a great expansion of voluntary control over physical behaviour. Speech is voluntary. Moreover, it demands very fine motor control (Donald 1991). Yet great-ape vocal communication seems to be at best quasi-voluntary. It is motivationally tied to emotional arousal, and cued by specific stimuli in the ape’s perceptual world. Moreover, language makes intensive demands on memory. The different parties to a conversation must remember who said what to whom. Furthermore, language is not tied to the immediate environment of the language user; we talk about much that is removed in space and time. So an agent must have the ability to recall lexical items, rather than having a lexicon whose entries are accessible only through some particular perceptual stimulus. The speaking ape needs a memory upgrade. Language also makes intensive demands on attention, for speech is a complex multitasking activity. In a conversation, you must do more than recall what has been said: you need to monitor and act on the effects on your utterances and those of others. You need to be alert to signs in the audience of a failure to understand, loss of interest, dissent, and other clues that the conversation is going wrong. As well, you may need to monitor the non-linguistic aspects of the situation. For often the point of talking is to facilitate coordinated action of some kind. Language is integrated into other aspects of social life, and hence talking requires that we divide our attention between different tasks, and between different aspects of our current circumstances. Many of the cues to which we must attend are contingent, varying within and across culture. We cannot be hard-wired with this information. Many of the capacities that make language possible are not encapsulated specializations. For they are interface capacities: their job is to integrate linguistic action within the envelope of the agent’s overall activities and those of the social group of which the agent is a part. As Fodor (2001) notes, these are not plausibly construed as tasks for a module. A defender of the evoked culture model of language might concede all this. The effective social use of language, she might agree, imposes extra demands on cognitive capacities that are not specialized for language. But the evolution of language requires as well the evolution of a language module: an innate specification of the organization of language. In my view it is hard to see how such a module could evolve. For consider what the evolution of language would be like, on a modular conception. Presumably language will have its origins in some simple version of proto-language (which I shall call ‘Habilene’): a system with a relatively small number
26
Kim Sterelny
of word-like elements used in short utterances whose interpretation will be heavily dependent on contextual cues (Jackendoff 1999). Learning and using these tools will not depend on cognitive specializations for language, for these will not yet have evolved. Instead it will depend on the agents’ general learning and problem solving techniques. But (to continue my origin myth) even a simple system of this kind is of great benefit to those who could acquire and use it. As Habilene became increasingly central to hominid social life, those who failed to acquire it, and those who acquired it imperfectly, would be at an increasingly crippling disadvantage. Their prospects for biological success would plummet. These factors would set the stage (a modularity theorist might conjecture) for an evolutionary transformation. A phenotypic trait that was once the result of individual learning—a trait that was a manifestation of phenotypic plasticity—would gradually become developmentally entrenched, with a specific genetic basis. For the establishment of Habilene sets up selection both to accelerate its acquisition and to decouple that acquisition from the accidents of individual experience. If Habilene were to develop through individual experience, and if children of taciturn speakers of Habilene were thereby to develop an impoverished capacity to use it, there would be selection for any mutation making that acquisition process less sensitive to variations in individual experience. A true (proto-)language module will evolve in response to the socio-cultural evolution of protolanguage and its establishment as an essential feature of human social competence. However, once established, such a module would have a profound effect on the further evolution of language. The modularization of language decouples linguistic capacity from linguistic experience. Instead, experience simply plays a triggering or selecting role. On the Tooby–Cosmides view of evoked culture, experience just selects a particular element from the internally specified set of available responses. To take a parallel case, on an evoked culture model of human sexual relations, we have (in that particular jukebox) a pre-specified set of marriage and childcare customs. An agent’s specific social experience would select the local variant from this range of options. Likewise, in the case of language, linguistic experience sets the parameters from a pre-specified set. In contrast, a socio-cultural model of evolution requires that children acquiring local culture are sensitive to variations in experience. For it is through such sensitivity that a successful innovation in parental behaviour is transmitted to their offspring. If children are only crudely responsive to experiences provided by their parents, intergenerational transmission will be of low fidelity, and parental innovation is unlikely to be copied to the next generation (Boyd and Richerson 1985, 1996, 2000; Richerson and Boyd 1998; Tomasello 1999a,b; Sterelny 2006).
Language, Modularity, and Evolution
27
Thus, if complex adaptations are assembled by socio-cultural evolutionary mechanisms, generation N + 1 must be sensitive to the experience provided by generation N, for it is only through such sensitivity that small improvements in the N generation can be transmitted to N + 1. Everyone accepts that some adaptive aspects of language are built by socio-cultural processes that depend on a culturally mediated flow of information across the generations. For that is how specialist vocabularies must be built. At least in this respect, language is a collective product. The vocabulary of any language is coined, refined, and transmitted by many of its speakers. Even in a small-scale society where every speaker ends up with full individual command over each item of the language, no one invents their own vocabulary. Language is a collective product, but one which is transmitted to new members of the group with high fidelity; and it thus meets the conditions for an evolutionary ratchet effect. Useful terminology is transmitted accurately enough for it to be used as a basis for further improvement.³ Once an aspect of language is modularized, though, the language of the N + 1 generation ceases to be sensitive to variations in the language use of the N generation. Thus the modularization of language would change the evolutionary mechanisms of language as well as the developmental mechanisms of language. Once a module evolves that substantially decouples linguistic competence from variations in experience, further evolutionary change in language can take place only through genetically based alterations in the module which are then selected to the extent that they confer on their bearers a selective advantage. How likely are such mechanisms to take us from relatively simple protolanguages to full human language? In my view: not likely at all. A module with its parameters set specifies a language, so the members of a community amongst whom a particular module is fixed will all speak essentially identical languages. Linguistic competence based on a common module is uniform and insensitive to new experience. Recall, too, that a central function of language is co-ordination. How then could speakers of expanded varieties of language invade when they are rare? How could, say, the ability to deploy a modal vocabulary be advantageous in any environment in which hardly anyone else can use such notions? Those who think language depends on an innate language module suppose that that innate system determines the general features of our grammatical, morphological, and phonological systems. If so, and if there are two different ³ Tomasello argues that this is true not just of the vocabulary of a language but also of its structural features: many morphological features—for example, tense–aspect systems—begin as specific items of vocabulary which are then incorporated as markers; for example, ‘going’ and ‘will’ have been converted into future constructions in English (Tomasello 1999; 2003: 42–4).
28
Kim Sterelny
versions of the module within the one population (Chomskysoft TM1.2 and Chomskysoft TM1.3 ), as there must be as it evolves, then some members of the population will speak a language that is partially incommensurable to other members of the same group. Without the right module, even if my conversational partner deploys modal constructions, I will neither understand nor come to understand them.⁴ Remember, too, that structural innovations are not merely additive. If, for example, Chomskysoft TM1.3 differs from its predecessor by having a regularized tense-aspect system modifying the verb phrase, then sentence formation using that module will be pervasively different from those of the predecessor. Incommensurability will not be an occasional glitch in cross-module interactions. It will be a typical and persistent feature of these interactions, for those equipped with the simpler module cannot acquire the new structures (or, at least, cannot acquire native-speaker ease in the use of them). Thus, consider a case where a language module, Ancestral, is fixed in a population. Consider a Variant that would be superior in expressive power if it were universal in a population. It by no means follows that Variant can invade. Even if agents are motivated to cooperate, high error rates make coordination difficult, and the invading form may increase error rates initially, even though it would be a superior system if it were fixed in the population. An ability to embed clauses multiply or to indicate tense and aspect by inflections threatens just to generate extra communication failures. The fitness structure is similar to that of a small number of Tit-for-Tat variants in an otherwise All-Defect population. Those using the Tit-for-Tat strategy find it hard to establish, for they do worse against the majority practice than that practice does against itself. The same is true of Variant, for the pay-off structure is as follows: Variant–Variant > Ancestral–Ancestral > Variant–Ancestral, because Variant–Ancestral interactions will result in more communication and coordination failures than either of the other pairings. But since Variants are rare and Ancestrals are common, unless there is sharp population segregation most Ancestral interactions will be with other Ancestrals, and few Variant interactions will be with other Variants. Thus Variants can invade only if (i) Variant–Variant interactions are much more productive than Ancestral–Ancestral interactions; or (ii) Variant–Ancestral interactions are only slightly less productive than Ancestral–Ancestral interactions; or (iii) Variants rarely interact with Ancestrals. The conditions which would allow Variant to invade may have been satisfied in early hominid populations. But it is not likely that they were. For there are reasons for suspecting that Variant–Ancestral interactions ⁴ Or, at best, I will understand them with more difficulty and with more misunderstandings.
Language, Modularity, and Evolution
29
would have had substantially worse pay-offs than Ancestral–Ancestral interactions. For one thing, speakers of the unimproved variant may both treat linguistic similarity and difference as a marker of group boundaries, and be less apt to act cooperatively to those out of the group. This is quite likely. Robin Dunbar has argued that linguistic difference functions to mark group boundaries in this way. Dunbar suspects that initially accidental variations between adjacent groups evolved independent functional significance by serving as a barrier to a particular free-riding strategy, namely, that of cheating on one’s obligations and then moving on: the drifter’s strategy. The more it costs to shift, the less appealing that strategy becomes. Linguistic differences, especially if they coincide with the boundaries of trust, add to the costs of movement. In the face of increasing linguistic difference and the increasing importance of language, a defect–shift strategy would become increasingly unappealing⁵ (Nettle and Dunbar 1997; Dunbar 1999). So Dunbar thinks that adjacent linguistic communities are under selection to maintain their different identities, not to reduce them. Whether or not the differences between languages are adaptations to mark group differences, they very likely have that effect. Dunbar and Nettle show that language is important to perceptions of group identity, and that perceived group membership is important to our dispositions to cooperate. If, then, Variant were perceived as marking membership of a different community, then the value of Ancestral–Variant interactions would be further degraded. For speakers of Ancestral would be liable to think of them as ‘not one of us’ and be less motivated to act cooperatively towards them. There is a further factor that would add costs to Variant–Ancestral interactions in this scenario. Misunderstandings impose an intrinsic cost, for they lead to coordination failure: hence the lowered value of Ancestral– Variant interactions even in trusting social worlds. But human social life involves competition and manipulation as well as cooperation. Since this fact is hardly a secret, social worlds are not always trusting. That leads to the possibility that misunderstanding may be read as defection, and that will increase the costs of interactions which are prone to generate misunderstanding. Indeed, there is some reason to suspect that linguistic interactions ⁵ Moreover, accent has the further consequence that when individuals do transfer, they carry some of their history with them. So it acts as a membership badge whether you want one or not. Our linguistic identity is both difficult to fake and difficult to conceal. The same is of course true of other social conventions which are both complex and arbitrary: how to use your fish-knife; in which direction the port circulates. These points about the role of language differences in our social life are well taken, but the drifter strategy was probably never a genuine option for our hominid ancestors. Amongst chimps, for example, only female adolescents have an opportunity to migrate.
30
Kim Sterelny
might be treated with particular suspicion. For defection may be especially difficult to detect, especially in cases where other agents have information and say nothing. Who is to know that Two-Aardvarks knew of the bear that ate Uncle Charlie? This problem of reciprocation and policing is made more acute by the fact that in its early stages it is likely that language use was error prone (Deacon 1997). Before the evolution of fine motor control, speech would have been more effortful, and the sounds from which speech was composed would have been less clearly distinguished from one another. In earlier stages of the evolution of language, then, the reception problem was more challenging, and that challenge was met by a less fancy system. Policing defection while maintaining a cooperative relationship with genuine cooperators is much more difficult in a noisy, errorridden environment (Sigmund 1993). The environment, then, may well have been somewhat more suspicious, and communication failures may well have been more likely to erode trust, increasing the social costs of Ancestral–Variant interactions. There is a problem, then, if we suppose that contemporary human language evolved in stages, with earlier populations using a simpler and less expressively powerful version of language, and if we suppose, further, that these earlier populations were equipped with a language module that explained their linguistic capacities in the same way that, allegedly, a language module explains our capacities. However, there is no invasion problem if we suppose that these earlier populations were linguistically heterogeneous and developmentally plastic. Heterogeneity and plasticity both help smooth the way for the invasion of linguistic innovators. To take the simplest kind of example, the ability to use a 500 word vocabulary can invade a community of speakers whose standard limit was 400, if the speakers do not all store the same 400 lexical items. If the community stock vocabulary is of 800 items, then the advantage of the invader would reside in being able to communicate and coordinate on more issues with more of the existing community. It is not (say) that the invader alone would master modal terms; rather, the invader alone would master both modal and functional terms. Furthermore, if the population is sensitive to linguistic experiences, innovations can spread horizontally and obliquely and hence users of the superior variant will more often reap the rewards of interaction with others with the same skills. Overlapping and flexible competences in the original community allow invasion. Thus the invasion problem seems to rule out the following scenario. Protolanguage evolved socio-culturally, but once it became crucial for social life it was modularized; its development became independent of variation in linguistic experience and hence uniform. Subsequent change in fundamental linguistic competence took place by the selective retention of
Language, Modularity, and Evolution
31
favourable variations in the genetic basis of language modules, not by the socio-cultural transmission of individual innovations. This cannot be the right picture if the evolution of new and more expressively powerful variants of language presupposes ongoing real differences in linguistic competence in communities, not the invasion of new variants into uniform communities. This still leaves open the possibility that learning and using language has been and is mediated by language-specific adaptations. But it does seem to undermine an evoked culture model of language—a model in which experience acts only to select from a palette of innately specified alternatives. Even if I am right about the invasion problem, perhaps this does not rule out an evoked culture model of contemporary language. Perhaps it was not modules all the way back. On this suggestion, humans differ from our ancestors not just in the power of our language but also in its developmental basis. Our competence in Sapiensish depends on an innate module that specifies its organization and expressive power. Thus all living humans speak languages which are, specific vocabulary items aside, effectively identical in expressive power. This, however, was an evolutionary achievement; this cognitive uniformity was not a feature of our ancestors’ linguistic lives. The evolution of language is the evolution of a gradually modularizing capacity. Earlier forms of language were cruder, less expressively powerful, and more dependent on experience for their acquisition. While I have no knockdown objection to this suggestion, there are problems with this picture. How could this underlying genetic and cognitive uniformity evolve in the face of linguistic diversity and geographic dispersal? Linguistic heterogeneity even within a single community was not mere noise: it was essential to language dynamics. If humans were dispersed across a number of habitats, that would have generated further diversity, especially in non-functional aspects of language. To put the point in a nutshell: how could the most dispersed hominid, the most culturally varied hominid, and the one for whom the cultural transmission of similarity was the most important, become the most linguistically uniform? What could homogenize heterogeneity, especially in the face of cultural mechanisms that maintain differences once they are established. Moreover, this suggestion seems inconsistent with the standard version of the poverty of the stimulus argument for the modularity of language. For on this suggestion, full human language (or something close to it) evolved by socio-cultural mechanisms, and thence was modularized, through the selection of genetic variants that increased the pace and robustness of acquisition. This picture clearly presupposes that it is possible to acquire human linguistic capacities without a wired-in module. For on this view, we had language before we had a module (though perhaps we had other adaptations for language acquisition and use).
32
Kim Sterelny
Finally, I think there is a more plausible alternative to this Baldwinesque view of the relationship between learned and unlearned components in the evolution of language, Eva Jablonka’s ‘assimilate and stretch’ model (Avital and Jablonka 2000). As with language, it’s a model of the evolution of agents that face cognitive bottlenecks because they are under selection for the elaboration of cognitively demanding skills. In selective regimes of this kind, evolutionary changes that make some aspect of a skill unlearned will be favoured because they free cognitive resources. Those free resources allow the skill to become still more complex. She had in mind such traits as complex courtship displays: complex birdsong, bowerbird bower building. There is a feedback loop in evolutionary systems of this kind, for independently of the average level of complexity, the more complex displays are preferred. Suppose that a species of bowerbird is under selection for bower quality. For an individual male bird, the more elaborate the bower, the better. Suppose further, and plausibly enough, that learning to build a bower is cognitively demanding. If so, there is a cognitive bottleneck: cognitive resources constrain the capacity to build an elaborate bower. If a male must learn everything about bowers, only simple ones can be made. To the extent that elements of bower building are brought under genetic control, the cognitive bottleneck on such birds would be eased. Their cognitive resources would then stretch to a more complex bower. But because there is feedback in the system, bower building as a whole would continue to have both innate and learned elements, even as bowers became ever more elaborate. In this example, the feedback loop of elaboration is driven by female preference. But the model fits language too. Feedback there is driven by the coevolutionary connection between language and the rest of culture: each allows the other to become more elaborate. As human cultural and social worlds become more elaborate, there is selection for languages of greater precision and expressive power. As languages of greater power and accuracy become available, human social worlds become more elaborate. I see no reason to suppose that this coevolutionary feedback loop has shut down. On this picture, bringing elements of language under genetic control has contributed to the increased complexity, power, and precision of language: with more innate, more can be learned. On this picture the expressive power of even contemporary language is not fixed by an innate language module, though innate elements are crucial to that power. In short: this view of language does not involve the wholesale rejection of a nativist, modular conception. But I do take the invasion problem to be grounds for some caution in accepting an evoked culture model of syntax, a model in which the role of experience is merely to select from a menu of pre-specified alternatives by setting parameters to their appropriate settings. Even so, parsing may depend in important ways on an encapsulated,
Language, Modularity, and Evolution
33
developmentally entrenched, and specialized adaptation. It is reasonable to conjecture that many of the organizational features of language have been stable over an evolutionarily significant period. If so, an evolutionarily stable subset of the total information available to an agent may suffice to solve the parsing problem. Hence an encapsulated system can be hardwired into the brain of language-using agents: nature can predict much of the information they will need for language parsing. I do not think this is true of interpretation: of working out what others mean. 3. LANGUAGE AS A SYSTEM OF NATURAL TELEPATHY? Few doubt that some of the cognitive skills involved in language use are specialized and encapsulated: those involved, for example, in the phonological analysis of utterances in your native language. Likewise, few argue that all of those cognitive skills are specialized and encapsulated. Fodor nominated the pragmatics of language use as a paradigm of unencapsulated problem solving (Fodor 1983). For example, there is no telling in advance what information you will need to work out why a speaker says something that is obviously and spectacularly false and who therefore cannot be saying what they mean. So no module can be designed for this task. In this section, I will argue that identifying the literal meaning of an agent’s utterance is more like the case of pragmatics than that of phonological analysis. Grice, Bennett, and others have argued for a metarepresentational picture of language. Understanding the meaning of an utterance involves recognizing and representing the speaker’s communicative intentions (Grice 1957; Bennett 1976). Understanding utterances essentially involves recognizing that speakers are intentional systems. Ruth Millikan does not doubt that we sometimes interpret others via identifying their communicative intentions, especially when participants in an exchange share no language. But she thinks that as a general model of language use this picture is psychologically implausible. Our system of linguistic interpretation is not designed to recognize communicative intentions. Rather, its function is to generate the same belief in the audience that caused the agent’s utterance.⁶ For Millikan, interpretation is much more like perception than it is like inference. The system of meaning conventions form the channel conditions for ‘natural telepathy’. When word tokens fulfil their proper function, a thought in the speaker’s mind appears in the mind of the audience (Millikan 1984, 1998). ⁶ Or to act appropriately, if the utterance is an imperative.
34
Kim Sterelny
Millikan’s picture presupposes an extremely cooperative picture of the use of language. For a system of communication which functioned to insert the same belief in the minds of the audience as that expressed by a speaker would be vulnerable to deceptive manipulation. So if human social environments included a mix of cooperative and competitive interactions, her picture is implausible (Origgi and Sperber 2000; Sterelny 2003). However, Millikan’s metaphor of natural telepathy looks much more apt when expressed as a claim about concepts. The channel conditions of normal language use cause the same concepts to appear in the mind of the audience as those that are expressed in the speaker’s utterance. Two-Aardvarks tells Erectus Jack ‘Tiger by the pond!’ Whatever else is necessary for interpretation, Erectus will not have understood Two-Aardvarks unless he understands what he means by ‘tiger’. Two-Aardvarks and Erectus Jack will have understood one another only if the concept Two-Aardvarks expresses by ‘tiger’ is the concept Erectus Jack deploys in his thoughts in response to Two-Aardvarks’s news. When all goes well, Two-Aardvarks’s utterance of ‘tiger’ causes Erectus Jack to think of tigers. The operation of language as a system of natural telepathy depends on the establishment of regularized, automatized, and coordinated practices of this kind. I have a word for tigers just so long as I have a tiger concept, and I standardly execute my communicative intentions involving tigers using that word. I understand your term ‘tiger’ if I typically recognize your communicative intentions about tigers when you have them, and that recognition is caused by the covariation between your uttering the word ‘tiger’ and your having such a communicative intention concerning tigers. How automatic and perception-like might this causal process be? Is the recognition of meaning mediated by modular mechanisms? Let’s compare recognition of meaning to paradigmatically perceptual processes. Consider such examples as: the use of stereopsis to locate objects in space; estimating the speed of oncoming objects; locating objects in space by attending to the different times signals reach each ear. These all involve perceptual sensitivity to recondite physical features of our environment, and computationally complex exploitation of that sensitivity. But our perceptual response does not require that we conceptualize these facts. Erectus Jack does not have a concept of hue or saturation just in virtue of his colour vision system’s sensitivity to hue and saturation. Colour vision is informationally encapsulated, as the persistence of illusion shows. It is autonomous, operating without cognitive decision or supervision; thus we cannot decide to see in black and white. These factors explain how it is that we are sensitive to hue and saturation without having concepts for hue and saturation. Might language decoding be the same? Erectus Jack is causally responsive to features of Two-Aardvarks’s cognitive and linguistic repertoire, but
Language, Modularity, and Evolution
35
it does not follow that he has the concept of a word, concept, or symbol. It is true that the concept–word correspondences to which Jack is sensitive are less stable than the connections between hue, saturation, and colour judgements. For colour vision depends on constant features of the physical environment, whereas the TIGER–‘tiger’ correspondence depends on linguistic regularities. But those regularities are stable over human generations. Moreover, as with colour vision, the process is involuntary. Once Two-Aardvarks says ‘tiger’, Jack cannot help but hear it as that term. Thus Ruth Millikan argues that once the system of correlations or connections has been established (and it can be established gradually, without deliberate engineering or consciousness), decoding can be causally sensitive to these connections, without agents having to have thoughts about these connections. You can speak and understand without being able to talk (or think) about speaking or understanding. On this picture, understanding the semantics of others’ utterances is assimilated to module-like processes (though Millikan does not use this terminology): it is automatic, unreflective; not subject to central control. I doubt that this is right. ‘Hue’, ‘saturation’, and other terms that pick out the components of visual cognition are not elements of folk vocabulary. The science of human colour vision is not a refined, more subtle, more precise but recognizable version of folk thought on colour experience. Likewise syntax and phonology are not refined versions of folk thought about language. In contrast, ‘word’ is a folk word, as is ‘means’, ‘is a name of ’, ‘is about’, and so forth. Much of semantics is a refined, extended, and more precise version of folk views of language and meaning. I do not think this is an accident. Seeing does not depend on being able to talk about seeing. For the most part, the utility of vision does not depend on our capacity to reflect on vision. In contrast, the utility of language is more closely tied to our capacity to reflect on language. I think there are three reasons why this is so. First, in most circumstances we do not see, or in other ways represent, our retinal image. That image is just a causal intermediary on the road to perceptual belief. What another agent says cannot be a mere causal intermediary in this sense, for that would lead us to be liable to deception, manipulation, and swallowing whole the errors of others. Even if we do typically accept what others tell us, it is still true that what others say is evaluated, and hence must be represented. Furthermore, the fact that we have been told something is often an important datum in itself. It is a clue to what others believe and want. Second, language learning and the division of linguistic labour seems to require that we represent language as well as use it. In learning language, we acquire much of our vocabulary from expressions rather than instances. Ostensive learning plays some role in acquisition, but many terms are acquired from representations of their targets,
36
Kim Sterelny
rather than from those targets themselves. We may not need folk semantic concepts to talk. But how could we acquire a concept from a word without having concepts for ‘word’, ‘about’, and so on? Early language use could not rely on smooth, well-established, uniform, and relatively context-free use of language. Stable word–concept correlations represent coevolutionary achievements, not initial starting points. Many early uses of language would have often been more like language-contact situations using pidgins. And pidgins, as many have pointed out, rely heavily on pragmatics. They rely on speakers thinking about what their conversational partners meant, asking themselves ‘What did he mean by ‘‘pushpush’’. Could he really have meant . . . ?’ Thus, whatever the literal meaning turns out to be, the processes by which one language user identifies that literal meaning of an utterance do not seem to be modular; they are not like perceptual processing. They are conceptualized and they are not encapsulated.
4. LANGUAGE IN A CHANGING WORLD One consequence of an innate language module is that it partially decouples linguistic competence from environmental input. No one suggests, of course, that linguistic competence does now, or ever has, developed in the complete absence of linguistic experience. If we are equipped with an innate language module, development is less sensitive to noise; to idiosyncratic variations in the particular experience a child might have. If a child’s competence depends on an innate module together with a modicum of linguistic experience, the children of taciturn parents will not grow up linguistically disabled. Developmental robustness is good design for traits which are absolutely central to human life. So far so good. But the presumption here is that variation in the linguistic input is noise not signal. We do not want development to be too robust. In a German-speaking community, we want competence to be sensitive to the fact that the input is in German. In general, the idea that good design decouples important adaptations from environment input assumes that the environment is constant. If the environment is variable in some significant respect, the development of the adaptive complex should be variable in response. To some extent, innate systems can generate adaptive response to variation in the environment by wiring in conditional responses. This is the evoked culture model of innately structured yet adaptive and variable response to environmental variation. Thus if human males have, over evolutionarily significant periods, lived both in polygamous societies and in societies which are strictly monogamous, then selection could wire in
Language, Modularity, and Evolution
37
conditional rules: ‘if you are rich and powerful in a polygamous society, marry many wives; if you are rich and powerful in a monogamous environment, be sure to marry a highly nubile wife of high genetic quality.’ However, there are limits on wired-in conditional responses. The range of variation and its significance must itself be constant. So innate modules do presuppose unchanging environments, though perhaps what does not change is the range and significance of environmental variability. Even if humans have been widely distributed through ecospace, innate systems can direct adaptive responses so long as the region of space humans have occupied has been stable over time. However, there is no reason to think that the environment of human language has been stable, even in the extended sense discussed above. For one thing, the physical and biological environment of human evolution has been increasingly unstable, as a result of climate change and of the expansion of hominid species out of their ancestral range (Potts 1996, 1998; Calvin 2002). Moreover, and perhaps more importantly, human language capacities have coevolved with other cognitive capacities and with the processes of culture change. Language is not the only distinctive human cognitive capacity: we differ from our closest living kin in having far greater powers to represent the future (Suddendorf and Corballis 1997), causal interactions (Tomasello 2000), the mental life of other agents (Heyes 1998), moral and other norms (Boyd and Richerson 1992; Richerson and Boyd 1998). If, as is surely likely, these capacities arose with language rather than preceded it, then the coevolution of language with these other cognitive faculties would have greatly altered the expressive demands on human language. Moreover, the coevolution of cognition with culture has built a mechanism that results in the cumulative change of human environments. The generation-by-generation construction of specialist vocabularies discussed earlier is an instance of a more general process. Humans are niche constructors: we rework our own environment; think of shelters and clothes; the domestication of animals; the use of tools (Odling-Smee 1994; Odling-Smee et al. 2003). Moreover, our niche construction is cumulative: generation N + 1 inherits a changed world from generation N and further modifies the world N + 2 will inherit. So Michael Tomasello has argued that, in contrast to the great apes, there are three timescales that matter in understanding human minds and human culture. The great apes have social but not cultural lives, and hence there are only two timescales in their cognitive histories: those of phylogeny and ontogeny. In understanding human cognition, there is a third timescale: that of the history of culture, as complex capacities are assembled. As these are built, they interact with and transform individual ontogeny and biological history (Tomasello 1999a).
38
Kim Sterelny
Crucial targets of language are socio-genetic constructions. What human agents need to discuss has not been a stable target onto which selection could lock, encoding a stable competence that can be built into us all once and for all. The variability of human cultures and the coevolution of language with other cognitive capacities are sources of instability. Over the last few hundred thousand years, there is every reason to suppose that human cultures have varied substantially in their need to encode information linguistically. We need language to describe (and prescribe) norms, other minds, causal interactions, future plans. Let me mention a few other possibilities. Consider first social rank and hierarchy. We do not know the size of human groups 200,000 years ago, but they were likely to be relatively small and egalitarian foraging communities. Contrast these social worlds with the much larger and far more stratified social worlds that have come into existence: stratifications that are in many ways built into language. Second, consider quantitative information. The socio-cultural invention of numerals and mathematical notation enables contemporary humans to express quantitative information far more precisely and extensively than pre-numerate cultures could. Third, consider other representational media. Language now has the resources to represent depictive representational media: pictures, diagrams, drawings. Once more, these public and enduring representational media are important and probably relatively recent (i.e. perhaps 35,000 years old) cultural inventions (Mithen 1998b, 2000). But once invented they are powerful tools in their own right, and important targets of language. Consider finally the non-actual. We have no information, of course, on how long humans have been storytellers. For all we know, this might be a very ancient use of language. But the use of language for mythological and religious purposes may not be ancient, for there is no evidence that religious belief has ancient origins (Mithen 1998a). Of course, a good deal of what I have just been saying is at best fairly plausible speculation. We have no way of assigning relative dates to, say, the emergence of our knowledge of the future, of other minds, and of full language. But if I am right in thinking that the expressive demands on human languages have changed in important ways, and as a consequence human languages have differed significantly in their expressive power—if, for example, the kinds of vocabulary vary—then the structure of the lexicon is not a stable target onto which selection can lock, pre-wiring agents with the information that they need. In fundamental ways, the representational powers of language are sensitive to experience, to local culture. Language is not merely evoked by experience.
Language, Modularity, and Evolution
39
REFERENCES A, E., and J, E. (2000), Animal Traditions: Behavioural Inheritance in Evolution (Cambridge: Cambridge University Press). B, J. (1976), Linguistic Behaviour (Cambridge: Cambridge University Press). B, R., and R, P. (1985), Culture and the Evolutionary Process (Chicago: Chicago University Press). (1992), ‘Punishment Allows the Evolution of Cooperation (or Anything Else) in Sizable Groups’, Ethology and Sociobiology, 13: 171–95. (1996), ‘Why Culture is Common but Cultural Evolution is Rare’, Proceedings of the British Academy, 88: 77–93. (2000), ‘Memes: Universal Acid or a Better Mouse Trap?’, in R. Augner (ed.), Darwinizing Culture: The Status of Memetics as a Science (Oxford: Oxford University Press). C, W. (2002), A Brain for All Seasons: Human Evolution and Abrupt Climate Change (Chicago: University of Chicago Press). D, T. (1997), The Symbolic Species: The Co-evolution of Language and the Brain (New York: Norton). D, M. (1991), Origins of the Modern Mind: Three Stages in the Evolution of Culture and Cognition (Cambridge, Mass.: Harvard University Press). D, R. (1999), ‘Culture, Honesty and the Freerider Problem’, in C. Power, C. Knight, and R. Dunbar (eds.), The Evolution of Culture (Edinburgh: Edinburgh University Press). F, J. (1983), The Modularity of Mind (Cambridge, Mass.: MIT Press). (2001), The Mind Doesn’t Work That Way (Cambridge, Mass.: MIT Press). G, H. P. (1957), ‘Meaning’, Philosophical Review, 66: 377–88. H, M., C, N., and F, T. (2002), ‘The Faculty of Language: What is It, Who has It, and How Does it Evolve?’, Science, 298: 1569–79. H, C. M. (1998), ‘Theory of Mind in Non-Human Primates’, Behavioral and Brain Sciences, 21: 101–48. J, R. (1999), ‘Possible Stages in the Evolution of the Language Capacity’, Trends in Cognitive Science, 3/7: 272–9. and P, S. (2005), ‘The Nature of the Language Faculty and its Implications for Evolution of Language (Reply to Fitch, Hauser and Chomsky)’, Cognition 97: 211–25. M, R. (1984), Language, Thought, and Other Biological Categories (Cambridge, Mass.: MIT Press). (1998), ‘Language Conventions Made Simple’, Journal of Philosophy, 94/4: 161–80. M, S. (1998a), ‘The Supernatural Beings of Prehistory and the External Storage of Religious Ideas’, in C. Renfrew and C. Scarre (eds.), Cognition and Material Culture: The Archaeology of Symbolic Storage (Cambridge: McDonald Institute).
40
Kim Sterelny
M, S. (1998b), ‘A Creative Explosion? Theory of the Mind, Language and the Disembodied Mind of the Upper Palaeolithic’, in Mithen (ed.), Creativity in Human Evolution and Prehistory (New York: Routledge). (2000), ‘Mind, Brain and Material Culture: An Archaeological Perspective’, in P. Carruthers and A. Chamberlain (eds.), Evolution and the Human Mind: Modularity, Language and Metacognition (Cambridge: Cambridge University Press). N, D., and D, R. (1997), ‘Social Markers and the Evolution of Reciprocal Exchange’, Current Anthropology, 38: 93–9. O-S, F. J. (1994), ‘Niche Construction, Evolution and Culture’, in T. Ingold (ed.), Companion Encyclopedia of Anthropology (London: Routledge). L, K., and F, M. (2003), Niche Construction: The Neglected Process in Evolution (Princeton: Princeton University Press). O, G., and S, D. (2000), ‘Evolution, Communication and the Proper Function of Language’, in P. Carruthers and A. Chamberlain (eds.), Evolution and the Human Mind: Modularity, Language and Metacognition (Cambridge: Cambridge University Press). O, S., G, P. E., and G, R. (eds.) (2001), Cycles of Contingency: Developmental Systems and Evolution (Cambridge, Mass.: MIT Press). P, S. (1994), The Language Instinct: How the Mind Creates Language (New York: William Morrow). (1997), How the Mind Works (New York: Norton). and B, P. (1990), ‘Natural Language and Natural Selection’, Behavioral and Brain Sciences, 13/4: 707–84. and J, R. (2005), ‘The Faculty of Language: What’s Special about It’, Cognition, 95: 201–36. P, R. (1996), Humanity’s Descent: The Consequences of Ecological Instability (New York: Avon). (1998), ‘Variability Selection in Hominid Evolution’, Evolutionary Anthropology, 7/3: 81–96. R, P., and B, R. (1998), ‘The Evolution of Human Ultrasociality’, in I. Eibl-Eibisfeldt and F. Salter (eds.), Ideology, Warfare and Indoctrinability (Oxford: Berghahn Books). S, K. (1993), Games of Life: Explorations in Ecology, Evolution and Behaviour (London: Penguin). S, K. (2003), Thought in a Hostile World (New York: Blackwell). (2006), ‘The Evolution and Evaluability of Culture,’ Mind and Language 21: 137–165. S, T., and C, M. (1997), ‘Mental Time Travel and the Evolution of the Human Mind’, Genetic, Social, and General Psychology Monographs, 123: 133–67. T, M. (1999a), The Cultural Origins of Human Cognition (Cambridge, Mass.: Harvard University Press). (1999b), ‘The Human Adaptation for Culture’, Annual Review of Anthropology, 28: 509–29.
Language, Modularity, and Evolution
41
(2000), ‘Two Hypotheses about Primate Cognition’, in C. Heyes and L. Huber (eds.), Evolution of Cognition (Cambridge, Mass.: MIT Press). (2003), Constructing a Language: A Usage Based Theory of Language Acquisition (Cambridge, Mass.: Harvard University Press). T, J., and C, L. (1992), ‘The Psychological Foundations of Culture’, in J. Barkow, L. Cosmides, and J. Tooby (eds.), The Adapted Mind (Oxford: Oxford University Press).
2 Mental Representation, Naturalism, and Teleosemantics Peter Godfrey-Smith 1. INTRODUCTION The ‘teleosemantic’ program is part of the attempt to give a naturalistic explanation of the semantic properties of mental representations. The aim is to show how the internal states of a wholly physical agent could, as a matter of objective fact, represent the world beyond them. The most popular approach to solving this problem has been to use concepts of physical correlation with some kinship to those employed in information theory (Dretske 1981, 1988; Fodor 1987, 1990). Teleosemantics, which tries to solve the problem using a concept of biological function, arrived in the mid-1980s with ground-breaking works by Millikan (1984) and Papineau (1984, 1987).¹ The decade or so from the early 1980s to the early 1990s was the heyday of the program of giving naturalistic theories of mental representation. The work was pervaded by a sense of optimism; here was a philosophical problem that seemed both fundamental and solvable. Its solution would be a major contribution to cognitive science, and would also contribute to many other parts of philosophy, especially epistemology. The work was accompanied by skeptics and naysayers of various kinds (Stich 1983; Dennett 1987), but in many circles optimism prevailed. On some days it seemed that a full solution might be just around the corner. This whole program seems to have lost momentum, at least for now. Fodor, who once had detailed solutions to offer on a regular basis, now ¹ The basic ideas of teleosemantics can also be used to try to explain the semantic properties of public representations (Millikan 1984). But the main focus, both in the initial work and in more recent discussions, has been mental representation. I will say quite a bit in this chapter about how mental and public representation are related, but I will not discuss teleosemantic treatments of public representation itself.
Mental Representation and Naturalism
43
seems to express only a vague hope that some form of informational semantics will succeed (1998: 12). Teleosemantics seems to have a fair number of people still working on it, with various degrees of faith, as can be seen in this volume. Millikan’s enthusiasm about her initial proposals seems undiminished, in contrast to Fodor. But the teleosemantic program is not insulated from the general turn away from optimism. Sometimes an idea loses momentum in philosophy for no good reason—because of a mixture of internal fatigue and a shift in professional fashion, for example. It is possible that this is what happened with naturalistic theories of representation. But I think that many people have been quietly wondering for a few years whether the naysayers might have been right all along. More concretely, I think there is a growing suspicion that we have been looking for the wrong kind of theory, in some big sense. Naturalistic treatments of semantic properties have somehow lost proper contact with the phenomena, both in philosophy of mind and in parts of philosophy of language. But this suspicion is not accompanied by any consensus on how to rectify the problem. In this chapter, my response to this difficult situation is to re-examine some basic issues, put together a sketch of one possible alternative approach, and then work forward again with the aid of this sketch.² So a lot of the chapter is concerned with the idea of mental representation in general, and what philosophy can contribute to our understanding of this phenomenon. These foundational discussions take up the next two sections. Section 4 then looks at some empirical work that makes use of the idea of mental representation—a different empirical literature from the ones that philosophers usually focus on. Then in Section 5 I look at teleosemantics from the perspective established in the preceding sections.
2. REPRESENTATIONALISM REASSESSED According to the main stream of work in naturalistic philosophy of mind in the 1980s, inner states of organisms like us represent the world. ‘Representation’ here is understood as a real, fairly unified natural relation that is picked out and understood in a very vague way by folk theory, and will eventually be described in much more detail by cognitive science and philosophy. One standard form of opposition to this picture is the ‘interpretivist’ ² I have undertaken similar forays in a couple of other papers in edited collections (Godfrey-Smith 2004, 2005). There may be some tensions across these different papers, all of which are expressed in a cautious and exploratory way. The 2004 paper discusses problems with the mainstream 1980s program, and possible concessions to the naysayers, in more detail.
44
Peter Godfrey-Smith
family of positions (Dennett 1987; Davidson 1984), according to which there are no semantic properties over and above those attributed by interpreters, where the role of interpreter is associated with a characteristic set of interests and point of view. This mild caricature of a familiar clash provides a point from which to look for new alternatives. What we want, I suggest, is a view that says something like this: There are indeed various kinds of connectedness and specificity that link inner states with conditions in the external world. But we should not look so directly to the everyday concepts of representation, belief, meaning, and so on, in describing what these connections are. The folk apparatus of everyday interpretation is primarily a social tool. It has genuine descriptive and explanatory uses, but these are mixed in with other features, and it is easy to be misled by socially tuned quirks of the apparatus, when trying to use it to describe real relations between inner states and the world.³ In some ways, this alternative shades into each of the more standard options mentioned above. But it is not supposed to be just a middle road. The idea here is that it is time to consider different possible accounts of what kinds of application semantic descriptions might have, both in principle and in practice, to inner states of physical systems. This chapter explores one possibility of this type. The main idea I will discuss is that we might see the idea of mental representation as the application of a particular model to mental phenomena. More precisely, we might see one kind of application of the idea of mental representation in these terms. The model in question is a schematized version of the pattern seen in one central kind of public representation use. That pattern is extracted and used in an attempt to understand mental processes. I see this attempted model-based understanding of the mind as available to the ‘folk’, and available also to scientists and philosophers who treat the model in more serious and rigorous ways. This model is one ‘route’ to the semantic description of inner states. It is probably not the only one; another, possibly distinct, route is via a concept of computation—via the idea of physical interactions that mirror logical relations among propositions. A third way may be via information theory in Shannon’s (1948) sense. I will leave open whether or not one of the ‘routes’ will turn out to be primary or fundamental. Certainly there come to be connections between them (see the end of Section 4 below). In addition, my aim here is not to offer a theory of how we acquire and use the most basic mentalistic concepts (thought, belief, pain, etc.). My focus is specifically on the idea of representation. ³ Some of Stich’s papers (e.g. 1982) show this rather well.
Mental Representation and Naturalism
45
The emphasis on models in this chapter is influenced by some ideas in philosophy of science, where the distinctive properties of model-based understanding have been much discussed in recent years (see especially Giere 1988). The sense of ‘model’ I use in this chapter is as follows: a model is a hypothetical structure that is supposed to bear some relevant resemblance relation to some real-world system that we are trying to understand. The hypothetical structure may in many cases be derived from another more familiar system, though that is not essential to the strategy. I think that many philosophers, and possibly more scientists, might accept that in some sense the idea of mental representation involves the application to the mind of a model derived from public symbol use. But this fact might usually be seen as not very informative. ‘Yes, sure it’s a model; now let me get back to what I was doing.’ In this chapter I will keep the idea in center stage. Interestingly, Wilfrid Sellars’s famous 1956 discussion of the ‘theoretical’ nature of folk psychological concepts used a very sophisticated conception of theorizing that gave an important role to models, in a sense of ‘model’ fairly close to mine. But subsequent developments of Sellars’s idea have not followed suit. So let us now look at what I will call the ‘basic representationalist model’. This is a structure—a sort of schema or scenario—that furnishes a way of describing agents and their use of symbols to deal with the world. Our starting point here is one familiar everyday sense of the term ‘representation’, as applied to public, external objects. In this sense, a representation is one thing that is taken to stand for another, in a way relevant to the control of behavior or some other decision. More specifically, I take the paradigm case here to be that when a person decides to control their behavior towards one domain, Y, by attending to the state of something else, X. The state of X is ‘consulted’ in working out how to behave in relation to Y. This can take the form of a conscious behavioral strategy, and it is also the topic of a familiar kind of third-person interpretation. You might decide to consult a street map to negotiate your way around a new neighborhood. Someone looking on at you can specify both the map and the mapped domain; they say you are using the map as a guide to a particular territory. This chapter will look at both this very general sense of representation and a more specific subcategory. The general class of cases is those where some X is consulted as a guide to behavior directed on Y. The more specific category is the class of cases where this strategy involves the use of a resemblance relation (perhaps an abstract and limited one) between X and Y. When we consult street maps, we usually do so because we hope to make use of a resemblance relation between map and mapped domain. But the idea of consulting the state of one thing as a guide to another does not always involve a resemblance relation. (Here I mean that we need not
46
Peter Godfrey-Smith
always hope for or rely on a representation relation, not merely that we might sometimes hope for one that is not present.) In the simplest possible case, what is consulted as a guide to behavior could be something as simple as the value of a single binary variable. (One if by land, two if by sea.) So far I have talked about a familiar public phenomenon. But this way of thinking about representation seems to lend itself readily to the case of mental states or brain structures. In this chapter I treat this as a kind of modeling exercise; we take a familiar pattern seen in social phenomena, and apply it to the case of thought. The ‘we’ here includes both ordinary people and cognitive scientists looking for a more scientific handle on mental processes. The view developed here recalls, in several respects, Sellars’s account of the operation of our ordinary mentalist concepts—roughly, what philosophers now call folk psychology (1956/1997). Sellars imagined that folk psychology might first appear as a theory in which inner processes were hypothesized to resemble outward verbal discourse. The present account is similar to Sellars’s in form, but is not the same view or even necessarily linked closely to it. I am supposing that public representation use furnishes a model for inner processes, but speech itself is probably not an especially relevant kind of public representation, in this context. In addition, the special features of propositional attitudes and their ascriptions are not the focus here. The best way to develop the present story in detail might be to tie it to Sellars’s account, but that possibility will not be discussed much below.⁴ One problem with writing about this set of ideas is confusion resulting from the profusion of things that can be called ‘models’. When we consult the state of X in order to determine our behavior towards Y, it can be natural to say that we are using X as a model of Y. This is often a useful way of talking about the phenomenon in question. But what I am concerned with in this chapter is the idea of taking that familiar pattern or situation—where one thing is consulted as a guide to another—and using that as a model for understanding some features of thought. So I will, purely for practical reasons, never use the term ‘model’ in this chapter for an internal or external object (my schematic ‘X’) that is being treated as a representation of something else; I will only use the term ‘model’ when talking about how the public phenomenon of representation use can be used as a source of hypotheses about inner processes. The basic representationalist model is a very natural (in the sense of appealing) way of thinking about some aspects of the mind. I see the model ⁴ My account here is closer to Sellars’s account of sense-impressions, given at the very end of his paper, than it is to his basic account of thoughts.
Mental Representation and Naturalism
47
as something that ordinary folk readily turn to in describing mental processes. Above I used the case of consulting the value of a binary variable as the simplest possible example of the kind of phenomenon seen in the basic representationalist model. Once we say this, it seems obvious that the variable consulted could be either internal or external to the brain, as long as the variable’s value can be read. For those in some intellectual traditions, however, alarm bells are now ringing. The application of representational talk of this kind to internal states is a trap, raising the prospect of regresses, private-language problems, and more. The representationalist adopts an innocent look: ‘Surely you can’t object to the internalization of the value of a binary variable? Would it help if I etched it on my teeth, rather than in my brain?’ And once the basic point about the possible internalization of a simple representation has been accepted, it seems reasonable to conjecture that the complex structures of the brain could contain components that function as more elaborate maps of external things, perhaps exploiting abstract resemblance relations to coordinate behavior with the world. Indeed, some authors are led to formulate very strong hypotheses along these lines. Here is Robert Cummins: ‘what makes sophisticated cognition possible is the fact that the mind can operate on something that has the same structure as the domain it is said to cognize’ (1994: 297–8). I will return to this claim by Cummins below. But for now let us continue discussing the basic model itself, as opposed to versions of the view that include a role for resemblance relations. My aim is to treat the basic representationalist model in a way that avoids philosophical excesses of several kinds. It is a mistake to think that there are no prima facie foundational problems at all with the serious application of the model to inner states (some will be discussed in the next section). But it is also, I suggest, a mistake to think that semantic and representational concepts are so inextricably tied to social interpretive practices that using the model in psychology is just a massive error. Here is another way to think of the situation. Consider the specific case of maps, which various psychologists and philosophers have taken seriously as akin to internal representational states. The familiar phenomena of map use in the public arena have both an ‘empirical skeleton’ and a rich social embedding. Here I do not mean that there must be a way of picking out in bare causal terms all and only the events that would normally be described as the use of maps, and assigning the maps definite semantic properties. I mean to absorb the possibility that this is impossible, because of the openendedness and context-sensitivity of interpretive practices. I just mean that at least many cases of map use have some typical local causal features. Our habits of interpretation of these phenomena are affected by more than the
48
Peter Godfrey-Smith
empirical skeleton, but the empirical skeleton can be used as a source of hypotheses about how the mind works. I see this as applying to both the case of maps, and representations more generally. So the empirical skeleton of public representation or map use might be made the basis for a scientific understanding of the mind—in principle it can do this, but this may or may not be a good idea. Perhaps it is a good idea; perhaps there are special kinds of adaptive or intelligent dealing with the world that are only made possibly by representation use, where this phenomenon is found in public contexts and also in the mind. Ruth Millikan’s theory, for example, treats internal and external signs as merely differently located instances of the same natural kind. Alternatively, this might all be a bad idea. One kind of anti-representationalist holds that the only empiricalskeletal features of representational phenomena that might be found in the mind’s workings are trivial ones. The critic may also argue that using representationalist ideas when formulating structural hypotheses about the mind tends to lead to subtle regresses and pseudo-explanatory traps. So when we use the representationalist model about the mind, we get very little return and we face persistent dangers. It might, alternatively, be a good idea in some sub-fields and at some stages in our understanding, while being misleading elsewhere. In some readers I imagine a feeling of impatience at this point. Do we really need yet another ‘back to square one’ exercise? Surely it is perverse to deny, at the present time, that representationalism has been fruitful in many areas of cognitive science; the problems to work on now in this area are problems of detail. I sympathize with one form of this impatience—a form that accepts that representationalism is something like a model, and insists that the model has done well in recent years. But I would add that it is easy to work within the representationalist model without properly resolving some acute foundational issues. (Indeed, that is one of the things models are good for.) 3. THREE FEATURES AND A CHALLENGE Let us look more closely at the ‘basic representationalist model’, and also at versions that make use of a resemblance relation. In this section I will discuss three characteristics of the model, and will also discuss in more detail the problem of regresses and pseudo-explanations. The first feature of the model I will discuss perhaps looks harmless, but I will say quite a lot about it. When we have a situation that fits the basic
Mental Representation and Naturalism
49
representationalist model, the representation being consulted must be, in some sense, a distinct thing from whatever is consulting it. As the model has it, one thing is used to guide behavior towards another. If we are describing a particular situation as an instance of this phenomenon, there must be a way of recognizing a separation between the representation and whatever is using it. Paradigmatically, there is also some generality or portability to the rule being used to interpret the representation, but I will not treat that as so important here. So if we are applying this representationalist model to the mind, we must have some confidence that representations can in fact be separated from their users or readers. Much of mainstream philosophy has simply accepted this. There is a standard way of talking in philosophy of mind that treats this as no problem. We often posit representational states, or structures, while supposing that in some sense they can be identified as distinct parts of the system. The availability of different ‘levels of description’ is sometimes taken to allay any worries that might arise on this front. This tendency is not exceptionless, but a great deal of representationalist talk simply assumes a separation between a representation and something else that deals with it. This is common in teleosemantics and especially explicit in Millikan. Millikan’s account is focused on things called ‘intentional icons’ (which include beliefs and other mental representations) that are situated ‘midway’ between ‘producer and consumer’ mechanisms. One must make also make a separation assumption in order to say what Cummins said in the quotation I gave in the previous section—that intelligence requires that the mind operate on something with the same structure as the domain it is dealing with. In large parts of cognitive science, standard ways of talking also assume separation, without worrying much about it. On the ‘classical’ computationalist side of cognitive science, there is a good reason for this. One of the distinctive things about ordinary digital computers is the fact that there is a good separation between the data stored in memory and the processing apparatus that makes use of the data. (You can upgrade your memory and your processor separately.) One can talk about a computer in a way that violates the particular location of this distinction that is laid down by the hardware; one can talk of a virtual processor with a different structure from the one in the hardware, for example. But in the machine itself, there is a separation of the right kind from the point of view of the basic representationalist model. So if the mind is being seen as similar to an ordinary digital computer, there is no reason to worry too much about the possibility of data structures being inextricably tied to the processing. From the point of view of mainstream cognitive science, it is presumably
50
Peter Godfrey-Smith
important and non-accidental that we have ended up building computers with this good separation. In less orthodox parts of cognitive science, especially parts associated with connectionism, situated cognition, and dynamical systems, the question of separation is more vivid.⁵ Sometimes a questioning of separation is seen as antithetical to representationalism; sometimes instead it is just described as ‘distributed representation’. Connectionists quite often want to hang onto familiar kinds of representationalist talk. I do not deny that they can do this, but they may have to depart from the basic representationalist model to do so, and this may have consequences. Sometimes it seems that advocates of distributed representation want to talk in two ways at once, both inside and outside the constraints of the basic model. The separation problem also has an interesting role in neuroscientific work. Talk of ‘inner maps’ can be very appealing when talking about various cognitive functions in an abstract way, but it is the neuroscientist who has to deal with the possibility that no straightforward separation may appear between ‘map’ and ‘reader’. I turn to a second feature of the basic model. When we engage in the familiar interpretive practice outlined earlier, saying that X’s state is being used as a guide to Y, we assume an answer to a question about specificity. Why is it Y that is the ‘target’ here? In the everyday cases, a person can say that it is Y that they are using X as a guide to. In the case of maps, for example, they can say that they are treating X as a map of Y. Mapping talk of this kind fits into a larger assumed semantic framework, in which maps, rules of interpretation, and target domains can be picked out and distinguished. Clearly a somewhat different story must be told when using the basic representationalist model to describe internal processing. But I take it that some way of picking out the target domain must be available. This general type of problem has been discussed extensively by Cummins (1996). He sees giving a theory of ‘targets’ and giving a theory of what a representation says about a target as two distinct parts of a theory of mental content. As far as I can tell, Cummins and I do not have exactly the same issue in mind when we talk about the problem of targets. Targets in my sense are bigger and vaguer than they are in his; a typical target for me will not be a particular object but a whole region of the environment. And I do not hold out hope for a unified naturalistic theory of how targets are determined in all real cases. But we are thinking of similar problems, clearly. To make the problem vivid, consider a scientific case. Suppose that there is a structure in a rat’s hippocampus that is said to be a ‘cognitive map’. ⁵ See e.g. Ramsey et al. (1991); van Gelder (1995); Clark (1997).
Mental Representation and Naturalism
51
(This concept will be discussed some more in the next section.) The rat is guiding its behavior, in some specific spatial task, by using this inner structure. It seems we can say that this is a case of the rat using the state of X (the inner structure) as a guide to Y. But, of course, all the rat is doing is receiving input of various kinds, and combining this with various pre-existing inner states to control behavior. It does not single out X, single out Y, and decide to use the former as a guide to the latter. From the point of view of the scientist, there is no problem here. The rat is situated in a particular environment—a maze, for example. If the scientist has reason to posit inner representations, he or she can say that the representations are being used to deal with this particular maze. The scientist applies what I will call a ‘thin behavioral’ specification of the target. This is fine in practice, at least in simple cases. It is also rather philosophically unsatisfying. It is natural from the scientist’s point of view to say that the rat is using X as a guide to Y, but as far as the mechanics of the situation are concerned, the ‘as a guide to Y’ claim seems extraneous. There will also be a lot of vagueness in thin behavioral specifications of targets. We have a different and richer specification of the target when it is picked out explicitly by a separate representational act. Against this, it might be argued that worrying about a richer and sharper specification of the target is worrying about something that is not part of the ‘empirical skeleton’ of representation use, and hence should not detain us. I will return to this issue below. The third issue I will discuss in this section is not an essential part of the basic representationalist model, but is a feature of many applications and developments of it. It is common when talking of mental representation in ways inspired by the kinds of considerations discussed above to posit a resemblance relation, albeit an abstract one, between representation and target. In what I regard as well-developed versions of this idea, the target itself is not specified by the presence of a resemblance relation; the specification of the target is a separate matter. Rather, the idea is that given that some internal structure X is being consulted as a guide to Y, this consultation can only be expected to be successful or adaptive to the extent that there is a suitable resemblance relation between the two. So the goal, in some sense, of consulting a representation is to exploit a resemblance relation between representation and target. At first glance, it surely seems clear that this should be regarded as an optional feature of the representationalist model. Some and only some public representations work via resemblance; why should this not be true also of internal representations? However, it is quite common in this area to
52
Peter Godfrey-Smith
use the notion of resemblance far more broadly, and see the exploitation of resemblance relations as a general or invariable feature of mental representation. Sometimes, it seems to me, these claims are made in a way that uses an extremely weak concept of resemblance or similarity. In other cases, the concept of resemblance being used is not especially diluted, and a genuinely strong claim is being expressed. The underlying line of reasoning might perhaps be something like this. In the public case, the available relations between X and Y that might be exploited are roughly the three distinguished many years ago by C. S. Peirce: resemblance, indication, and conventionally established relations. The last of these is off the table in the case of mental representation. The second can be assimilated to the first, once resemblance or isomorphism is construed in a suitably abstract way. So the only kind of relation that really matters here is resemblance. For this or other reasons, many discussions of mental representation extend the language of resemblance to cover a very broad class of cases. In Randy Gallistel’s entry for ‘Mental Representation’ in the Elsevier Encyclopedia of the Social and Behavioral Sciences (2001) he insists that all representations exhibit an isomorphism with the represented domain. In correspondence, Gallistel confirmed that cases usually discussed by philosophers using concepts of information or indication (thermostats, fuel gauges, etc.) are treated by him as involving abstract isomorphisms. Millikan’s teleosemantic theory uses concepts of mapping and correspondence in similarly broad ways; occasionally she explicitly says that her theory vindicates the idea that inner representations ‘picture’ or ‘mirror’ the world (1984: 233, 314). And earlier I quoted Cummins (1994), who claimed that the exploitation of structural similarity is the key to all sophisticated cognition. I do not want to deny that there are some very subtle but still reasonable notions of resemblance that may be used here, especially those employed in logic and mathematics. My aim is not to restrict the talk of resemblance and mapping to cases where some very obvious notion of picturing is involved. Yet I resist the idea that some suitably abstract resemblance or isomorphism relation is always involved in mental representation. When X is consulted to guide behavior towards Y, this may involve the exploitation of an antecedently specifiable resemblance relation, but it may not. It can be tempting to add here that there must be some natural relation between representation and target that makes the representation worth consulting. And from there, it can seem that resemblance or isomorphism is the only genuine candidate. But this is not so. Once we have an intelligent brain, it can generate and adaptively manipulate representations that do not have any simple, easily exploited relation to their targets. (Strong versions of the ‘language of thought’ hypothesis are expressions of this possibility: Fodor 1975.) For this reason, I see no reason to accept the Cummins hypotheses that was
Mental Representation and Naturalism
53
quoted earlier. That hypothesis arises out of a desire for an overly simple explanation for when and why it is worth consulting X to deal with Y. More precisely, there are (as we often find) strong and weak ways to read the Cummins hypothesis, with the strong way unjustified and the weak way misleading. In strong forms, the hypothesis was criticized in the previous paragraph. In weak forms, the notion of similarity or resemblance is extended too far, and becomes post hoc in its application. (If a representation was successfully and systematically consulted to deal with some target, there must have been a similarity or isomorphism present of some kind . . .) Before leaving this topic, I should note in fairness that the Cummins hypothesis I have focused on here was expressed in a note attached as commentary (1994) to a reprinting of an earlier chapter. The same ideas were followed up in his 1996 book, but I have chosen to focus on a formulation that Cummins presented in a rather ‘unofficial’ way. Secondly, I am aware that the representational role of abstract but not-trivial resemblance relations, especially those with mathematical description, needs a far more detailed treatment than I have given it here. The final topic I will discuss in this section is a general challenge to the usefulness of the representationalist model. I call it a challenge to the ‘usefulness’ of the model, but the challenge is derived from stronger arguments, often directed against the model’s very coherence. My aim here is to modify and moderate an older form of challenge. I argued that the empirical skeleton of public representation use might be used as a model for some kinds of mental processing. But might it be possible to see, in advance, reasons why this will be a bad or misleading model? Famous arguments due to Wittgenstein (1953) and the tradition of work following him are relevant here. One form of argument that is especially relevant holds that if we import the basic structure of representation use into the head, we find that the reader or interpreter part of the mechanism has to be so smart that we have an apparent regress, or pseudoexplanation. A version of this challenge to representationalist explanation in cognitive science is expressed by Warren Goldfarb (1992). He is discussing a hypothesis that people with perfect pitch make use of ‘mental tuning forks’. This concept was introduced in a newspaper discussion of a piece of neuroscientific work on the different neural activity of people with and without perfect pitch. Goldfarb regards the hypothesis of mental tuning forks as pseudo-explanatory in the extreme. Tuning forks! Are they sounding all the time? If so, what a cacophony! How does the subject know which fork’s pitch to pick out of the cacophony when confronted with a tone to identify? If they are not always sounding, how does she know which one to sound when confronted with a tone?
54
Peter Godfrey-Smith
Real tuning forks give us the means to identify pitches, but they do so because we have the practices and abilities to use them. The internal standard is supposed to give us the means to identify items, but without practices and abilities, for the internal standard is also meant to operate by itself, in a self-sufficient manner. (If it were not, it would be otiose: why not settle for practices and abilities themselves? . . .) (Goldfarb 1992: 114–15)
This line of thought might also be used to express a challenge to the Cummins hypothesis that I have discussed several times in this chapter. Cummins wants to explain intelligence by giving the mind access to something with the same structure as its target. Call this structure S. If the mind’s problem is dealing with things that exhibit S, how does it help to put something with S inside the head? The mind still has to detect and respond to S, just as it did when S was outside. When the challenge is expressed in these strong sorts of terms, the right reply to it is to connect the representationalist model to the basic ideas of ‘homuncular functionalism’ (Dennett 1978; Lycan 1981). The internal representation is not supposed to be ‘self-sufficient’, to use Goldfarb’s term. It would need a reader or interpreter; there must be something akin to ‘practices and abilities’. But the mind’s interpreter mechanism need not have the whole set of practices and abilities of a human agent. The interpreter can be much less sophisticated than this (more ‘stupid’, as the homuncular functionalist literature used to say), and might operate in a way that is only somewhat analogous to a human agent using an external representation. The representationalist holds that positing this kind of separation between a representation-like structure with an exploitable relation to a target and a subsystem to make use of that structure is a good hypothesis about the mind. If we put these two components together, some special cognitive capacities become possible. So if the challenge is expressed by saying that we can see in advance that no explanatory progress can be made with the basic representationalist model, then the challenge can be defused. But the fact that we have this in-principle answer does not mean that we will necessarily make progress in the actual world, by using the representationalist model. It may well be that, for reasons akin to those expressed in the traditional challenge, there is little in fact to be gained by employing the model. This will depend on what the mind’s structure is actually like. In order to have some explanatory usefulness, there needs to be the right kind of interaction between a representation and reader in the mind. Putting it in homuncular functionalist terms, the reader needs to be smart enough for its interaction with the representation to be reader-like, but not so smart that the model collapses into homuncularism of the bad kind.
Mental Representation and Naturalism
55
4. INNER MAPS IN THE COGNITIVE SCIENCES This section will look at one family of applications of the basic representational model in psychology and other cognitive sciences. The work discussed in this section makes use of the concept of a mental or cognitive map—a representational structure with some similarity or analogy to familiar external maps, like street maps. This is obviously not the only way to develop and apply the basic representational model in trying to understand mental processes, but it is a very natural way to do so. As I noted in Sections 2 and 3 of this chapter, there is a way of thinking about the representationalist model that leads people to think of resemblance or isomorphism as a crucial relation between internal and external states. Looking for inner map-like structures is a way to develop this idea. The literature on inner maps is also, as I see it, a rather pure and direct way to use the basic representationalist model to think about the mind. The literature on inner maps in the cognitive sciences is partially separate from the tradition that emphasizes computation, logic, and language-like representation. The empirical work on cognitive maps in question is often (unsurprisingly) concerned with spatial skills, usually in non-linguistic animals. So this is a somewhat simpler arena in which the role of the representational model can be investigated. In particular, we do not have to worry about the possible effects of public language capacities on the representational powers of thought.⁶ The notion of inner maps is also interesting because it seems to be a kind of ‘attractor’ concept, one that people come back to over and over again and from different parts of science and philosophy. There is something very appealing about this idea, but of course it also raises in a vivid way the pitfalls discussed at the end of the previous section. I should also emphasize that the discussion in this section is an initial foray into this literature; I hope to discuss it in more detail on another occasion. Here I will also discuss scientific work rather than philosophical work (see Braddon-Mitchell and Jackson 1996 for a relevant philosophical discussion). In psychology, the father of the idea of inner maps is E. C. Tolman (1948). For Tolman, the hypothesis of ‘cognitive maps’ was put forward in response to some particular forms of intelligent behavior, studied primarily in rats and seen especially (though not exclusively) in dealing with space. The crucial contrast that Tolman had in mind when he developed this ⁶ For some speculations on these issues that complement the present discussion, see Godfrey-Smith (forthcoming b).
56
Peter Godfrey-Smith
idea was with strict ‘stimulus–response’ models; the hypothesis of cognitive maps was motivated by the inability of stimulus–response models to account for what his rats could do. In his ‘sunburst maze’ experiment, for example, a rat first learned a highly indirect route to a food source, and was then presented with a large range of new paths, some of which led more or less directly to the food source. Rats chose a nearly direct path much more often than chance would predict. Tolman’s idea was often ignored in its mid-twentieth-century context, but has since become much more influential. There has been a revival of the idea both in comparative psychology and also in neuroscience.⁷ Earlier I distinguished a basic sense of representation in which the state of one thing is used to guide behavior towards another, and a richer notion in which this guidance involves a resemblance relation. Both in philosophy and in the sciences, we find the term ‘map’ used in a range of weaker and stronger senses. In its weakest senses, any internal representation can be described as an internal or cognitive map. In its strongest senses, it involves a notion of resemblance between the map and the target domain. As far as I can tell, Tolman and many of the workers in psychology and neuroscience who have followed him use the term ‘cognitive map’ in a way that is intermediate between the weakest and strongest senses I am distinguishing. That is, a cognitive map is not just any mental representation—it has something extra—but perhaps it need not work via a resemblance relation with what it represents. The term ‘map’ is primarily invoked in connection with spatial cognition, but is sometimes used more generally. For Tolman and others, the hypothesis of a cognitive map is often used to express the hypothesis of some kind of cognitive sophistication, over and above simple associative mechanisms and (especially) stimulus–response processes. But there seems to be a family of relevant kinds of sophistication. In a good deal of the literature that I have looked at so far, the dialectic is something like this. The researcher will have in mind a contrast between two (or more) classes of inner mechanisms that might be used in dealing with a behavioral problem (usually a problem involving space). One might be a family of comparatively simple, associative mechanisms, and the other will be a family of more sophisticated mechanisms that apparently involve a more flexible and intelligent use of information. The term ‘cognitive map’ is associated with the latter.⁸ But the boundary between what counts as a simpler or deflationary explanation and what counts as an explanation in ⁷ For a survey in comparative psychology that uses the concept, see Roberts (1998). For a review of the neuroscientific work, see Jeffrey et al. (forthcoming). ⁸ Tolman himself, after contrasting his mapping idea with the stimulus–response model, then compared more task-specific ‘strip maps’ with more flexible ‘comprehensive maps’. He saw this as a gradient distinction. Once we have this distinction, the contrast
Mental Representation and Naturalism
57
terms of cognitive mapping tends, perhaps, to shift around as background knowledge changes. One of the main contrasts that is salient here is between what is described as an ‘integrated’ representation of spatial structure, and the possession of special-purpose behavioral rules that can only be applied in specific circumstances (Mackintosh 2002). So a good deal of the empirical work tries to find whether animals are able to use experience to come up with solutions to problems that they have not been explicitly trained on. Many of Tolman’s experiments had this character, and so have various later ones. The ability to use novel short cuts in navigation tasks is often taken to suggest the presence of a ‘integrated’ knowledge of spatial structure, for example. However, it turns out that there are associationist mechanisms that do predict the use of some kinds of short cut (Deutsch 1960; Bennett 1996); the explanatory resources of the ‘simpler’ mechanisms are richer than they were in Tolman’s time. In some of Kim Sterelny’s philosophical work (2003) he explores the distinction between ‘decoupled’ representations, that can be put to use in a variety of behavioral tasks, and those that are ‘coupled’ to some specific task or behavior. A lot of the talk of ‘integrated’ knowledge in discussions of cognitive mapping is gesturing towards this same distinction; integration is largely the ability to use learned information in a flexible range of tasks and contexts. The most important landmark after Tolman was the work that introduced the idea of cognitive mapping into neuroscience, O’Keefe and Nadel’s The Hippocampus as a Cognitive Map (1978). O’Keefe and Nadel argued for the existence of a special kind of learning system, the ‘locale’ system, which constructs and uses map-like representations. O’Keefe and Nadel also offered a specific neurological hypothesis: the hippocampus is the part of the brain where this happens, at least in some animals.⁹ The neurological hypothesis was supported with various studies of deficits (especially those associated with failure at harder spatial problems) and also with the discovery of ‘place cells’ in the rat hippocampus. These cells fire when the animal is in a certain place, but are not dependent on the simplest place-related stimuli; they keep firing when the rat’s angle of view is changed or it is plunged into darkness, for example. In much of this literature, I think it would be quite accurate to say that the idea of cognitive mapping is used as a model, in the specific sense between strip maps and associative mechanisms is perhaps a little problematic in Tolman’s paper. ⁹ In humans the hippocampus seems to be associated more with general laying down of memories than with spatial cognition in particular.
58
Peter Godfrey-Smith
developed in the earlier sections of this chapter. A lot of discussion of the distinctive features of maps is discussion at the level of semantic properties; maps are taken to be associated with specific kinds of representational power. The psychologist’s focus is often on something like the question of whether we have reason to think that the organism’s brain contains something with those representational properties. Some of the questions that are most pressing from a philosophical point of view are not much discussed. In particular, there is not a great deal of discussion of how a physical structure in the brain might come to have the distinctive representational properties associated with maps. To illustrate this, let us look at how the target and separation problems appear within this empirical work. The target problem seems to be all but invisible. Suppose the researcher is watching the rat deal with a water maze. (These are pools of colored water in which a platform is hidden just below the surface in a particular location, which the rat must find and later relocate.) The rat solves the problem. The researcher posits a cognitive map, as the water maze is a hard problem that should defeat simple associative mechanisms. What is the target of the hypothesized map? Obviously, it is the water maze—that is what the rat is dealing with. Once one gets inside the mindset of the empirical worker, the idea that there is a foundational problem with the specification of the target seems strange. The focus of the empirical work is a set of contrasting hypotheses about what is going on inside the rat’s head. Whatever is in there is obviously directed, in some sense, on what the researcher can see is the rat’s current behavioral problem—the water maze. The idea that mapping talk might be jeopardized by the need to give some substantive story about why that is the target is quite odd. So the cognitive scientific literature operates with what I called earlier a ‘thin behavioral’ specification of the target. The target of a map is just whatever the map is in fact used to deal with. With respect to the separation problem, we see a similar lack of concern to that found in much philosophical literature. The practice in the empirical literature is often to describe inner maps within the context of an imagined or assumed division between map and reader mechanisms, without treating this as a very substantive hypothesis. This might be seen as a harmless way of talking that may involve an idealization. If so, then from the philosophical point of view this idealization may be bearing quite a lot of weight. So a lot of the time, a representationalist model using the idea of a map is introduced into discussion as a means for describing and investigating some empirically interesting distinctions, without much concern for the foundational problems. But some discussions do take some of the extra steps, and ask how a neural structure could possibly have the properties posited
Mental Representation and Naturalism
59
in the model. One way to do this is to introduce explicitly the concept of computation. I see this concept not as itself part of the basic representationalist model, but as something additional that can be connected to it. John O’Keefe (of O’Keefe and Nadel 1978), for example, has in his later work tried to put the notion of cognitive mapping ‘on a firmer computational basis’ (1997: 280). His strategy is to show how a spatially organized array of neural activity could encode a matrix of numerical values that constitute a map of the environment. This map might be ‘consulted’ by the rat in the control of behavior via mathematical manipulation of the matrix (multiplication, inversion, and so on). So a large part of O’Keefe’s paper is a demonstration that these mathematical manipulations of matrices are all operations that neural circuits could, in principle, perform. It would be interesting (and difficult) to look closely at how O’Keefe’s computational account relates to the issue of separation, and also to the homuncular functionalist treatment of the role of the ‘reader’ mechanisms, as discussed in Section 3 above.
5. WHAT HAVE WE LEARNED FROM TELEOSEMANTICS? Suppose the discussion in the preceding sections is on the right track. What then is the status of the large body of philosophical work on naturalistic theories of mental representation? In particular, what, if anything, have we learned from teleosemantics? I take these two questions in turn. The basic representationalist model is a schematic, vague sort of structure, and also one that is not usually described in rigorously naturalistic terms. So the following question presents itself: supposing that we formulated a version of the model in purely naturalistic terms, exactly what sorts of semantic description could be given a principled basis in the model? When this question is asked about a very simple and stripped-down version of the model, we know from decades of philosophical work that the available semantic descriptions will exhibit a range of indeterminacies and breakdowns. But there is the possibility that richer versions of the model may support more determinate and fine-grained semantic descriptions than stripped-down versions do. So it is possible to take the basic model and embed it in a more elaborate and detailed scenario, where all the extra components of the scenario are described in purely naturalistic terms. We can do this, and then ask which additional kinds of semantic description attach plausibly to the resulting structure. For example, we can try to embed the basic model in a surrounding context that will make a sharp and principled notion of misrepresentation available, or the discrimination of contents that involve coextensive concepts.
60
Peter Godfrey-Smith
What I am describing here is a kind of philosophical model-building, which operates by augmenting and supplementing a basic representationalist model that has its origins outside philosophy. Even those philosophers who are hostile to the general framework presented in this chapter will have to agree that something like this is what a lot of work in the last twenty years has, in fact, involved. The naturalistic philosophical literature has spent a lot of time describing idealized hypothetical scenarios that have reasonable or compelling descriptions in both physicalist and semantic terms. The aim has been to describe as minimal and empirically feasible a structure as possible, while maximizing the number of features of paradigmatic semantic description that attach plausibly to the structure. In particular, this is how I see teleosemantics. The teleosemantic literature takes the basic model and embeds it within a biological setting. A simple causal structure that conforms, roughly, to the basic model is embedded within a context involving an evolutionary history and various forces of natural selection. We take the basic model, embed it in this setting, and see how this affects our responses with respect to the semantic description of the structures in the model. So the basic representationalist model itself has no particular relationship to such things as natural selection and biological function. But it is possible to make use of such biological concepts to show how a rather elaborate semantic description of the basic model can be grounded in naturalistic factors. One of the intuitions that has driven teleosemantics is the idea that rich biological concepts of function pick out a special kind of involvement relation between parts of organisms and their environments. Edging even closer to the semantic domain, there is a kind of specificity or directedness that an evolved structure can have towards an environmental feature that figures in its selective history. I think this idea is basically right; there is an important kind of natural involvement relation that is picked out by selection-based concepts of function. But this relation is found in many cases that do not involve representation or anything close to it. It is found in the case of enzymes, ordinary physical traits, and all the other features of organisms that can have selective histories. There is nothing intrinsically semantic about this involvement relation. But this sort of involvement relation can be added to the basic representationalist model; that is what teleosemantics does. In keeping with the comments at the end of Section 2, I should add that selection-based concepts of function might also be used within other approaches to bridging semantic and physicalist description of inner processes, besides the basic representationalist model. But let us see how selectionist ideas contribute to versions of the model that has been the main topic of this chapter.
Mental Representation and Naturalism
61
I will focus once again on the target problem. I claimed in earlier sections that the basic representationalist model can be employed, and is often employed, with a ‘thin behavioral’ specification of the target. In practice, there is no problem saying that the target of the rat’s inner map (if there is one) is the maze with which it is dealing. This idea was accepted, while noting that from a philosophical point of view the specification of the target here seems somewhat vague and extraneous. In a teleosemantic version of the representationalist model, however, the target becomes far from extraneous. This is because of the role of the target in a feedback process that shapes the representation-using mechanisms. I will use Millikan’s theory to illustrate this. A central concept in her account is that of an ‘indicative intentional icon’. A wide range of semantically evaluable phenomena turn out to involve these structures, including bee dances, indicative natural language sentences, and human beliefs. Millikan says that an indicative intentional icon is a structure that ‘stands midway’ between producer and consumer mechanisms that can both be characterized in terms of biological function. The consumer mechanisms modify their activities in response to the state of the icon in a way that only leads systematically to the performance of the consumers’ biological functions if a particular state of the world obtains. That state is (roughly) the content of the icon. More specifically, though, if we have a set-up of this kind then the icon is ‘supposed to map’ onto the world in a particular way—via the application of a particular rule or (mathematical sense) function. Given the way that the consumers will respond to the state of the icon, if the world is in such-and-such a state then the icon is supposed (in a biological sense) to be in a corresponding state. What this story involves, in abstract terms, is a combination of the basic representationalist model plus a feedback process, in which relations between actions produced and the state of the world can shape the representation-using mechanisms. We suppose that the success of actions controlled by the consultation of an inner representation is determined by the state of some particular part of the world, and these successes and failures have consequences for the modification of the cognitive system. The particular feedback process that Millikan uses is biological natural selection. Withingeneration change is handled by an elaborate story about how selection on learning mechanisms generates teleo-functional characterization of the products of those learning mechanisms. It is the abstract idea of a suitable feedback process (or what Larry Wright once called a ‘consequence etiology’) that is most relevant here, however (see Wright 1976). If we have a feedback process of the right kind, then the representationalist model can be employed in such a way that the specification of the target becomes a natural part of the story—a real part of the mechanism. The aspect of the
62
Peter Godfrey-Smith
world that the organism’s inner representation is directed on is the aspect whose different states control feedback processes shaping relevant parts of the cognitive system, especially the ways in which representations of that kind are produced and consumed. There is also a more famous way in which the biological embedding of the basic representationalist model can be used to motivate richer semantic descriptions. This is the problem of error or misrepresentation. Much of the original appeal of teleosemantics was its ability to employ teleo-functional notions of purpose in order to deal with apparently normative aspects of semantic phenomena. In particular, the biological notion of failure to perform a proper function was used to attack the problem of misrepresentation, which had caused a lot of trouble for information-based theories. Millikan has occasionally said that the sole or primary contribution of the teleo-functional part of her theory (or any theory like it) is to handle this problem. I do not think this is right. As will be clear from the discussion in earlier sections of this chapter, I think the target problem is significant in its own right, and is not (as it is sometimes seen) merely an aspect of, or perspective on, the error problem. But the misrepresentation problem is indeed one place where the teleosemantic machinery has seemed helpful. In the light of the preceding discussion, it is possible to take a different perspective on the use of teleo-functional concepts to deal with error. In general, the ‘normative’ nature of semantic concepts has been overestimated and overemphasized in recent decades. Part of this emphasis came from the remarkable rhetorical power of Kripke’s discussions of semantic phenomena in his book on Wittgenstein (Kripke 1982). Part of it came from other places, including the teleosemantic literature itself. Semantic phenomena do display something a bit like normativity in its more familiar senses. But the relation between the quasi-normative aspect of semantic phenomena and the quasi-normative forms of description made possible via selectionbased or teleonomic concepts of function is really a kind of analogy or mirroring, not a potential reductive relationship. The ways in which misrepresentation can motivate descriptions of what ‘ought’ to have happened and what is ‘supposed’ to be the case derive not from the empirical skeleton of representational phenomena, but from the elaborate network of interpretive social practices that surround it. (This is part of the real message of Kripke’s book.) Above I argued that by embedding the representationalist model in a certain kind of biological context, Millikan could give a good answer to the target question. The fact that this embedding is one way to settle questions about targets does not mean it is the only way. Other ways of augmenting or embedding the basic model may suffice to do the same kind of job. As has repeatedly been noted, the empirical commitments behind any
Mental Representation and Naturalism
63
ascription of content made on Millikan’s basis are strong. A selection process of just the right kind is needed, to ‘aim’ an inner icon at a definite aspect of the world. As I argued in an earlier paper (1994), it is possible to doubt the widespread applicability of this pattern of explanation without bringing in fanciful swampman-like cases, and without casting doubts on Millikan’s interpretation of evolutionary theory and its relation to learning. All that is needed to raise problems is a ‘noisier’ set of processes affecting the evolution of the cognitive mechanisms in question. And in cases where it is available, Millikan’s solution can also diverge in interesting ways from pretheoretic intuitions about the target of a representation. This is the message, as I see it, of Paul Pietroski’s argument against teleosemantics (1992). Pietroski uses an elaborate thought-experiment to make his case, but the argument can also be described more schematically. He describes a case where coordination with a particular environmental variable is responsible for the success of a state of internal wiring in a simple organism, but where the crucial environmental variable is not linked to anything the organism can perceive in a single causal chain, but instead is linked to something the organism can perceive by a common cause pattern. By responding to observable variable Z, the organism coordinates its dealings with Y, where Z and Y are (roughly) products of a common cause. Applying Millikan-style principles to determine the content of an inner state by which the organism guides its behavior, we find that the inner state will represent the state of Y, the variable that matters in the explanation of success and failure. But Y may be a variable whose states no organism of this kind can perceptually discriminate in any circumstances. Pietroski makes a good case for the view that his case shows a real gap between intuition and the application of Millikan’s principles. Millikan’s account looks for the explanatory variable in a feedback process affecting the representational machinery, and this variable can be sufficiently far from the organism’s perceptual capacities (even in optimal circumstances) that this seems an implausible target from the standpoint of everyday interpretation. In retrospect, it is unsurprising that there is no perfect match between the explanatory variable and the pretheoretic assignment of content. What I find surprising is that it takes some real work to come up with a problem case that shows this clearly. I have discussed the relation between Millikan’s work and the target problem in some detail here. I will not say much about the separation problem. It will be obvious that Millikan’s account and (so far as I can tell) other accounts in the teleosemantic literature tend to help themselves to an assumption about separation without worrying much about the issue. As Millikan says, to be an intentional icon something must exist ‘midway between’ producer and consumer mechanisms; at least to some extent, these components all have separable roles. To use one of her standard lists
64
Peter Godfrey-Smith
of examples, bee dances, adrenalin flows, indicative natural language sentences, and human beliefs are all seen as indicative intentional icons. The assumption that beliefs are structures that have the kind of distinct location between producer and consumer mechanisms that we find in the case of bee dances and adrenalin flows is a substantial assumption. The other topic discussed in earlier sections that might be raised here is the role of resemblance relations. I distinguished the basic representationalist model from the more specific version where a resemblance relation between representation and target is exploited. Interestingly, Millikan does present her view as one in which a mapping or resemblance relation is an essential part of the explanatory structure; she sees her account as a vindication of the old idea that thoughts can picture the world, albeit in an abstract way (1984: 233, 314). Other teleosemantic theories (e.g. Papineau, 1987, Neander, in this volume) do not have this feature. I have argued in previous discussions (1994, 1996) that Millikan’s discussions of resemblance and mapping are somewhat misleading. It is only a very attenuated sense in which her account, if true, would vindicate the idea that thoughts picture the world. This point can be made compactly by noting that the general notion of mapping that Millikan uses to describe all indicative intentional icons will apply to natural language sentences, as well as beliefs. True descriptive sentences like ‘snow is white’ map or picture the world, too. I have handled the notion of mapping in this chapter in such a way that map-like representation contrasts with linguistic, conventionally mediated representation. It is possible to describe notions of reference, satisfaction, and truth using the language of abstract picturing, but I do not see much point in it. In Millikan’s account, as I argued in the papers cited above, the talk of mapping boils down to a certain kind of requirement of systematicity. Any representation must be related to other possible representations in systematic ways, and the links between that set of representations and states of the world they represent will have related systematic features. Systematicity of this kind may well be a very important idea. I would distinguish it, however, from what I take to be much stronger uses of the idea of resemblance of the kind seen (for example) in the Cummins hypothesis discussed earlier in this chapter. I am acutely aware that the issues about mapping and resemblance raised here need a more detailed and careful treatment. That will have to wait for another paper. The main thing I have done in this section is give an alternative account of the philosophical role of teleosemantics, focusing on Millikan’s version. But some work in teleosemantics can also be seen as expressing empirical hypotheses. The idea that teleosemantics is a straightforwardly empirical theory has been asserted explicitly in the literature (e.g. Papineau 2001).
Mental Representation and Naturalism
65
But I have in mind a slightly different set of empirical claims, and from the point of view of the present chapter, the more philosophical and more empirical aspects of teleosemantics have often been mixed together in complicated ways. One empirical hypothesis has remained live throughout the chapter: does the basic representationalist model pick out an important natural kind that figures in the causal organization of intelligent organisms? I have described representationalism as a ‘model’, but models can be used to represent accurately the real structure of the world. Models in the present sense should not be associated, in any general way, with instrumentalism. To that hypothesis about the basic model we can add hypotheses about the evolutionarily embedded versions of the model that figure in teleosemantics. One hypothesis is especially vivid. Might it be the case that the only natural systems that instantiate the basic representationalist model have been shaped not just by evolution in general, but by the particular kinds of selective histories that figure in the teleosemantic literature? Might a particular kind of natural selection be the only feasible etiology for the kind of structure seen in the basic representational model? Some advocates of teleosemantics might want to say, at this point, that this was always the whole point of the program—the orientation of the work has always been empirical, though very abstract. To me, however, it seems that if an empirical hypothesis of this kind is supposed to be center stage, it has not been properly separated before now from various other ideas—ideas of the kind discussed in the earlier parts of this section.
6. CONCLUSION This has been a somewhat sprawling chapter, but I will give a brief summary of my main points. We may need some new views about the kind of application that representational talk has to inner states and processes. In this chapter I tried to develop one such view, by treating representationalism as the application of a model. This view is not offered as an account of our most basic mentalistic concepts, although it might be linked with an account like Sellars’s. When taken seriously in a scientific or philosophical context, the basic representationalist model does have interesting foundational problems, but these are not insoluble in principle. Much of the literature on ‘cognitive maps’, especially in comparative psychology, is a rather pure application of the basic representationalist model. Teleosemantics, especially Millikan’s work, can be seen as a philosophical elaboration of the same model. Teleosemantics embeds a version of the basic model in a detailed biological scenario, and
66
Peter Godfrey-Smith
shows how quite elaborate semantic descriptions of elements of this model can be matched up with, or grounded in, carefully described biological factors. More of the baggage of folk semantic talk can be given a precise analogue in a naturalistic, biological scenario than one might expect. The biological treatment may not be the only approach to augmenting the basic model that yields these kinds of results, however. Teleosemantics has not uncovered a set of hidden historical assumptions that must be made when engaging in any psychological or neuroscientific talk of mental representation. In addition, teleosemantic ideas may possibly be applied to content without going via the basic representationalist model (a possibility I did not much discuss). Teleosemantics also contains, in a rather philosophically entangled way, empirical hypotheses about the feasible etiologies by which structures that instantiate the basic representationalist model might arise. I doubt that teleosemantics, or any theory like it, will deliver the direct, reductive, puff-of-papal-smoke solution that the 1980s literature envisaged. We probably need to look for a different kind of philosophical approach to semantic phenomena from what we are used to. But in the meantime, we have learned quite a lot from naturalistic semantics, including teleosemantics, even if we have not learned what we might have originally hoped to learn.¹⁰
REFERENCES B, A. (1996), ‘Do Animals Have Cognitive Maps?’, Journal of Experimental Biology, 199: 219–24. B-M, D., and J, F. (1996), Philosophy of Mind and Cognitive Science (Oxford: Blackwell). C, A. (1997), Being There: Putting Brain, Body and World Together Again (Cambridge, Mass.: MIT Press). C, R. (1994), ‘Interpretational Semantics’, in S. Stich and T. Warfield (eds.), Mental Representation: A Reader (Oxford: Blackwell). (1996), Representations, Targets, and Attitudes (Cambridge, Mass.: MIT Press). D, D. (1984), Essays on Truth and Interpretation (Oxford: Oxford University Press). D, T. (1997), The Symbolic Species: The Co-evolution of Language and the Brain (New York: Norton). ¹⁰ I am indebted to Kim Sterelny for extensive comments on an earlier draft of this chapter. This work has been influenced by discussions with Richard Francis and Alison Gopnik.
Mental Representation and Naturalism
67
D, D. C. (1978), Brainstorms: Philosophical Essays on Mind and Psychology (Cambridge, Mass.: MIT Press). (1987), The Intentional Stance (Cambridge, Mass.: MIT Press). D, J. (1960), The Structural Basis of Behavior (Cambridge: Cambridge University Press). D, F. (1981), Knowledge and the Flow of Information (Cambridge, Mass.: MIT Press). (1988), Explaining Behavior (Cambridge, Mass.: MIT Press). F, J. A. (1975), The Language of Thought (New York: Crowell). (1987), Psychosemantics (Cambridge, Mass.: MIT Press). (1990), A Theory of Content and Other Essays (Cambridge, Mass.: MIT Press). (1998), Concepts: Where Cognitive Science Went Wrong (Oxford: Oxford University Press). G, R. (2001), ‘Mental Representation, Psychology of ’, in P. Baltes and N. Smelser (eds.), The International Encyclopedia of the Social and Behavioral Sciences (New York: Elsevier). G, R. (1988), Explaining Science: A Cognitive Approach (Chicago: Chicago University Press). G-S, P. (1994), ‘A Continuum of Semantic Optimism’, in S. Stich and T. Warfield (eds.), Mental Representation: A Reader (Oxford: Blackwell). (1996), Complexity and the Function of Mind in Nature (Cambridge: Cambridge University Press). (2004), ‘On Folk Psychology and Mental Representation’, in P. Staines (ed.), Representation in Mind: New Approaches to Mental Representation pp. 147–62, (New York: Elsevier). (2005), ‘Untangling the Evolution of Mental Representation’, in A. Zilh˜ao (ed.), Cognition, Evolution, and Rationality: A Cognitive Science for the XXIst Century pp. 85–102, (London: Routledge). G, W. (1992), ‘Wittgenstein on Understanding’, in P. A. French, T. E. Uehling, and H. K. Wettstein (eds.), Midwest Studies in Philosophy, xvii: The Wittgenstein Legacy (Notre Dame, Ind.: University of Notre Dame Press). J, K., A, M., H, R., and C, S. (forthcoming), ‘Studies of the Hippocampal Cognitive Map in Rats and Humans’. K, S. (1982), Wittgenstein on Rules and Private Language (Cambridge, Mass.: Harvard University Press). L, W. (1981), ‘Form, Function, and Feel’, Journal of Philosophy, 78: 24–50. M, N. (2002), ‘Do Not Ask Whether They Have a Cognitive Map, but How They Find Their Way Around’, Psicologica, 23: 165–85. M, R. (1984), Language, Thought, and Other Biological Categories (Cambridge, Mass.: MIT Press). O’K, J. (1997), ‘Computations the Hippocampus Might Perform’, in L. Nadel, L. A. Cooper, P. Culicover, and R. M. Harnish (eds.), Neural Connections, Mental Computation (Cambridge, Mass.: MIT Press). and N, L. (1978), The Hippocampus as a Cognitive Map (Oxford: Clarendon Press).
68
Peter Godfrey-Smith
P, D. (1984), ‘Representation and Explanation’, Philosophy of Science, 51: 550–72. (1987), Reality and Representation (Oxford: Blackwell). (2001), ‘The Status of Teleosemantics, or How to Stop Worrying about Swampman’, Australasian Journal of Philosophy, 79: 279–89. P, P. (1992), ‘Intentionality and Teleological Error’, Pacific Philosophical Quarterly, 73: 267–82. R, W., S, S., and G, J. (1991), ‘Connectionism, Eliminativism and the Future of Folk Psychology’, in J. Greenwood (ed.), The Future of Folk Psychology: Intentionality and Cognitive Science (Cambridge: Cambridge University Press). R, W. A. (1998), Principles of Animal Cognition (Boston: McGraw-Hill). S, W. (1956/1997), Empiricism and the Philosophy of Mind (Cambridge, Mass.: Harvard University Press). S, C. (1948), ‘A Mathematical Theory of Communication’, Bell System Technical Journal, 27: 379–423, 623–56. S, K. (2003), Thought in a Hostile World (Oxford: Blackwell). S, S. P. (1982), ‘On the Ascription of Content’, in A. Woodfield (ed.), Thought and Object: Essays on Intentionality (Oxford: Oxford University Press). (1983), From Folk Psychology to Cognitive Science: The Case Against Belief (Cambridge, Mass.: MIT Press). T, E. C. (1948), ‘Cognitive Maps in Rats and Men’, Psychological Review, 55: 189–208. G, T. (1995), ‘What Might Cognition Be, If Not Computation?’, Journal of Philosophy, 92: 345–81. W, L. (1953), Philosophical Investigations, trans. G. Anscombe (New York: Macmillan). W, L. (1976), Teleological Explanations (Berkeley: University of California Press).
3 Representation, Teleosemantics, and the Problem of Self-Knowledge Fred Dretske A verbal description of Clyde represents him as having three children. A portrait of him as a young man represents him as having had a bushy moustache. A thermometer represents him as now having a temperature of 103 ◦ F. Verbal descriptions, portraits, and thermometers are ordinary physical objects. So thinking of the mind in representational terms is the first step in conceiving of the mind in naturalistically congenial terms—as like other physical objects and events: measuring instruments, words, and pictures. Thoughts and experiences of Clyde, just like thermometers, words, and pictures, are physical objects that say things about Clyde that may or may not be true.¹ Materialists will have little trouble thinking of judgments (propositional attitudes, in general) in this way, but perceptual experiences (and other qualia-laden phenomenal states) are generally regarded as more problematic. A representational account of sense experience seems to me as unavoidable, however, as a representational theory of belief. The properties we use to individuate and classify experiences are, after all, like the properties we use to individuate and classify such obvious instances of representation as stories: they are properties not (necessarily) of the experiences (stories) themselves, but of what these experiences (stories) are experiences (stories) of. The qualities we experience, the ones that make an experience the experience it is (shape, color, movement, pitch, texture, orientation, hardness), needn’t be qualities of anything in the person’s head who is having the experience and, therefore (given that experiences occur in the head) not qualities of the experiences themselves. They needn’t (at least at the time ¹ I focus throughout on mental states—primarily beliefs (the propositional attitudes) and experiences—that can most easily be interpreted in representational terms. This makes my job a lot easier. The task of providing a naturalistic theory of the mind is already hard enough without asking for more trouble.
70
Fred Dretske
of the experience) be qualities of anything. How is this possible? Dualists have a story about how it is possible, but what can materialists say? Materialists should say it is possible in the same way it is possible to have stories—‘Cinderella’, for instance—depicting pumpkins transformed into carriages without there being anything—certainly nothing between the covers of the book (where the story is)—resembling such magical pumpkins. The properties and situations one is consciously aware of in having an experience resemble the properties and situations described in stories: they are intentional entities, properties things are represented as having, and things don’t have to have—nothing needs to have—the properties things are represented as having. Measuring instruments are familiar examples of such representational–intentional systems. Nothing need be going 100 mph for a (malfunctioning) speedometer to represent things as going that fast. Even when the representation is veridical, it (the representation) needn’t have the properties it says or means the automobile has. Ordinarily, of course, a speedometer (located in the car whose speed it represents) has the same speed it represents the car as having, but the police have stationary devices that can represent a car as going 100 mph. According to a representational theory, experiences (of movement, say) are like that. The representational vehicle, the thing in your head that represents the object you see as moving, doesn’t (or needn’t) have the properties (movement) it represents this object as having. That is why looking in a person’s (or bat’s) head won’t reveal the qualities being experienced by the person (bat) in whose head one looks. What I experience (see) when I look in another person’s head are the representations—electrical–chemical events in gray, soggy, brain stuff, that represent there to be a bright orange pumpkin out there. What the person in whose brain I look experiences is a bright orange pumpkin out there. I know of no other theory about the nature of perceptual experience that tells this satisfying a story about the subjective—first-person vs. thirdperson—aspects of experience. If experiences of pumpkins exist inside the person seeing the pumpkins (if they don’t, why would they vanish when the person closes her eyes?), it neatly explains why we can’t tell what other people’s (or animals’) experiences are like just by looking in their head, at the experiences they are having. We can’t tell—at least not just by looking inside—for the same reason I can’t always tell what is happening in a story by looking in the book where the story (the representation) is. If the book is written in Chinese, I can, by looking in the book, see the representation, the Chinese characters, but I fail to ‘see’ (understand, know, appreciate) their meaning or content, and it is the meaning, not the symbols that have it, in which pumpkins and carriages exist.
Representation and Self-Knowledge
71
Since a representational theory accounts for these otherwise puzzling facts about experience, it successfully bridges the explanatory gap. A representational theory of experience doesn’t, I admit, solve the ‘hard’ problem of consciousness. It bridges this explanatory gap only by opening up an equally puzzling gap somewhere else: how do electrical and chemical events in gray soggy brain stuff manage to represent bright orange pumpkins? We all understand how ink marks on paper—namely, the words ‘bright orange pumpkin’—when embedded in an appropriate descriptive context, manage to perform the trick. We—people with certain intentions and purposes—give these symbols this power by making them mean bright, orange, and pumpkin. We could, collectively, make these words mean something else. But no one, presumably, gave the events occurring in our brains their meaning. If, somewhere in our brains, there are representations of orange pumpkins, these representations, unlike words in stories, are not conventional, arbitrary, representations of the objects we see, hear, feel, and think about. Their representational power is entirely independent of our purposes, intentions, desires, and beliefs. So where do brains get this marvelous power to represent (both conceptually in judgment and non-conceptually in experience) the things we think about and perceive? How do they come to mean orange pumpkin out there? Good question. It is a question to which teleosemantics is supposed to provide an answer. Teleosemantics is a theory of meaning—and, therefore, a theory of representation—that explains, or purports to explain, in naturalistically approved terms, how configurations of matter in biological entities can mean (truly or falsely as the case may be) that there is an orange pumpkin nearby. If it is successful, then, teleosemantics provides an explanation of how arrangements of matter inside our heads can qualify as beliefs about and experiences of orange pumpkins. Teleosemantics (I’m here describing only my own version of it—see Dretske 1988, 1995) does this by identifying the meaning (representational content) of a neurobiological state (organ, mechanism, etc.) with what it has the function of indicating about the world. A (veridical) experience of a nearby orange pumpkin is simply a physical state of (presumably) the brain that represents its external cause (in this case a pumpkin) as orange, pumpkin-shaped, and nearby. It represents the pumpkin this way because this state has the function of indicating the occurrence of these properties in its external cause. If the external cause turns out to be a green cabbage, we get misrepresentation—a misperception—of the cabbage as an orange pumpkin. A green cabbage is the object seen, but it looks like an orange pumpkin. If there is no external cause (the representation is triggered by drugs in the bloodstream), the experience is hallucinatory: there is nothing out there
72
Fred Dretske
(e.g. a green cabbage) that the representation represents to be an orange pumpkin. It nonetheless still represents there to be an orange pumpkin out there. The experience is, subjectively, the same. The world is different. This answers the question, it solves the problem, only if there are naturalistically acceptable ways of interpreting the key ingredients of the recipe: function and indication. If we identify the representational content of a neurological structure with its indicator function, we had better be prepared to say where these functions come from and what it is (indication) they have the function of doing, and we should be able to say this, moreover, without taking for granted the ideas—belief and experience (and such related notions as intention and purpose)—we are invoking functions to understand. Otherwise the whole enterprise is circular. I understand where the function of measuring instruments (words, diagrams, etc.) comes from. It comes from us—from the purposes, intentions, and beliefs of designers, builders, and users. That is what makes mercury in a glass tube mean that Clyde’s temperature is 103◦ (whatever his actual temperature happens to be). If, however, our brains have indicator functions, functions that determine what we experience, think, and intend, these functions cannot, like the functions of instruments, words, and diagrams, come from us. I won’t say anything about indication. I tried to do that elsewhere (Dretske 1981). I’m more interested in the idea of function, the idea that binds teleosemanticists together. In my own case, the (informational or indicator) functions in question are either (for sense experience) phylogenetic or (for belief and other conceptual states) ontogenetic (functions a state-type acquires as a result of a certain type of learning). But those are details, and I am not now interested in explaining (or defending) my own theory of representation. Nor am I interested in criticizing other versions of teleosemantics. I want, instead, to circle the wagons and address a common problem, a problem that, as far as I can tell, almost all teleosemanticists face.² For if one traces representational power (the capacity to mean that something is so) to functions, as teleosemanticists do, then one has to deploy a notion of function that carries the kind of normative or intentional punch that psychological affairs so obviously exhibit.³ It won’t do, for ² Teleosemanticists face a lot more problems than the one I will be discussing—see, for example, Perlman (2002), Enc¸ (2002), and Walsh (2002) and, of course, Fodor (1990)—but one thing at a time. ³ I’m not sure the quality we are after is normative (see Dretske 2001 for my reasons). That will depend, I suppose, on what one means by ‘normative’. But whatever it is, it is something capable of grounding the difference between the true and the false, right and wrong, correct and incorrect, valid and invalid. It is, in a word, something capable of putting the mis- into representation and the mal- into function.
Representation and Self-Knowledge
73
instance, to operate with a notion of function that makes a thing’s function whatever it happens to be doing—even if what it happens to be doing helps explain the behavior of the system of which it is a part.⁴ For that doesn’t tell us what the thing is supposed to be doing. It doesn’t tell us, therefore, how misrepresentation (more generally, malfunctions) can occur. It doesn’t tell us how it is possible get things wrong, make a mistake, commit a fallacy, and reason incorrectly. If we are to distinguish meaning from truth—as we must if we are to have a theory in which misrepresentation is possible—we need functions a thing can stop performing without losing, functions that stay fixed (as meaning does) as the world being represented (truth) changes, functions that can survive a change in the way things function. Only by grounding this distinction between how a thing is and how it is supposed to be can functions provide a way of understanding how something can continue to say that p is true when it isn’t.⁵ If one operates (as I do) with indicator (informational) functions, one needs functions (to indicate) that are durable enough to survive a failure to indicate. We need functions that will enable us to distinguish (Grice 1957) non-natural meaning (the sense in which ‘It is raining’ means it is raining) from natural meaning (the sense in which wet streets mean it is raining). I believe that only etiological functions, functions a thing has in virtue of its history, are up to this task. Unless an object’s performance is seen in the light of its history, there is no way it is supposed to be, no way it can get things wrong, no way for it to be incorrect or defective. No way to make things malfunction (Millikan 1989; Neander 1991). To say that a mechanism (or bodily organ) is defective, diseased, or malfunctioning is to say that it is not in the condition it is supposed to be in, and this ‘supposed to be’ only makes sense if there is a standard, a norm, that is independent of what the mechanism is, in fact, doing. If you or I deliberately make something that way for illustrative purposes, as a prototype, or as a work of abstract art, or if it materializes randomly that way out of cosmic dust, it isn’t defective, diseased, or malfunctioning no matter how much it looks and/or behaves like an object (organ, mechanism, part) that (with a different history) is defective, diseased, or malfunctioning. The standards or norms implied in describing a mechanism (organ, etc.) as defective come from its function or purpose, yes, but a function or purpose it has in virtue of its history. In ⁴ Roughly, the kind of functions associated with Cummins (1975) causal-role analysis of function. ⁵ Hardcastle’s (2002) effort to give causal-role functions a normative quality is not much help to materialists. She traces the normativity of a thing’s causal role (what she calls ‘pragmatic’ functions) to the explanatory purposes of scientists (see pp. 152–3). Such normativity is obviously not available to explain, at least not in naturalistic terms, the normativity or intentionality of the mental.
74
Fred Dretske
the case of artifacts, the history that defines the function or purpose of a thing relates to the intentions and purposes of those who design and use it. In the case of biological organs and mechanisms, it comes from—where else?—the evolutionary history or individual development of a system. It is for this reason that defective things have to have a history. If they don’t, they aren’t defective no matter how much they fail to work the way we want them to.⁶ Injured, healthy, strained, stretched, diseased, flawed, ill, sick, damaged, spoiled, ruined, marred, contaminated, defiled, corrupted, infected, malformed —they are all like that. Nothing can be any of these things unless it is the result of some historical process that has defined what that thing’s function is (or the kind of thing it is—see footnote 5) and, therefore, what it is supposed to be doing. If a thing is marred or damaged, for instance, it departs in some degree from a standard that defines how it, or things of that sort, should be. And so it is in psychological affairs. Without a history (intentions of purposeful agents provide their own kind of history) there are no mistakes. Or misrepresentations. There are, therefore, no representations. It is for this reason that teleosemantics is committed to a historical, an etiological, conception of functions. That, at least, is what I believe. I’m not, however, going to spend more time arguing for it.⁷ For present purposes, I will simply assume it. That will suffice because if this assumption is false, if teleosemantics does not need history in its theory of meaning, if, somehow, an object that materializes randomly can, at the first moment of its existence (before it has acquired any relevant history), have an appropriate function and, thereby, misrepresent the world around it, then the ⁶ It is for this reason that I do not think a general goal-contribution analysis of function (according to which an organ’s function is its contribution to the goals of a system (see Boorse’s 2002 excellent discussion) will suffice for teleosemantics. Not unless it has an etiological component to it. It comes closer than a causal-role analysis in capturing the ‘normative’ character of biological (and psychological) affairs—what makes something defective, broken, diseased, mistaken, or wrong—but without an etiological component, it mistakenly assigns the same functions to the parts of physical duplicates even if the duplicates materialized randomly. I don’t think the parts of a system that materializes randomly (however much these systems resemble purposive systems) have functions—things they are supposed to do. Boorse’s suggestion (2002: 76) that health (one of the alleged ‘goals of living systems’) is simply the absence of pathology (which, in turn, is the less than normal functioning of a part) doesn’t seem to me to help. How do we tell whether something that materializes randomly is a healthy human being or a defective chimpanzee, a monstrously deformed chipmunk or a diseased extraterrestrial (one who would quickly die in the habitat to which it ‘belongs’)? Unless we have a way of saying why the creature isn’t one of these other things, and I don’t see any principled way of saying any of it, it makes little or no sense to speak of it as having a broken leg or diseased kidneys, or, indeed, of making a mistake. ⁷ See chapter 5 of Dretske (1995) for more argument.
Representation and Self-Knowledge
75
problem I’m going to discuss (below) is not really a problem for teleosemantics. So the worst-case scenario for teleosemantics is if the assumption is true. So, in defense of teleosemantics, I assume it true in order to show that, even on a worst-case scenario, teleosemantics is a plausible theory of representation. If the function of a thing and, thereby, its representational power, is derived from history, teleosemantics is a particularly strong form of externalism about the mind. The external facts on which mental content (and, therefore, the mental) supervenes are not only facts external to the representation, they are facts that sometimes (in the case of biological functions, for instance) relate to the very remote past. Not only does the mind not supervene on the current physical state of a system, it does not supervene on the current global state of the universe. According to teleosemantics, what we think and experience today—indeed, the fact that we think and experience anything at all today—depends not only on what is going on in us and around us, but on events and conditions that existed long ago and (probably) far away. A physical duplicate of a conscious being, a person (?) who lacked the appropriate history—a history that gave its internal states the requisite functions—would not think and experience anything at all. The internal machinery would function—causally speaking—in the same way, but it wouldn’t have the same (or, indeed, any) function. There would be no representations. The duplicate would be a zombie—devoid of thought and experience.⁸ To describe this result as a problem for teleosemantics may be too generous. Some would describe it as a reductio ad absurdum. This result not only strains the imagination of hard-headed materialists (I have heard respected materialists describe it as preposterous), it seems to many to be at odds with the obvious fact that we know—often enough anyway—what we are thinking and experiencing. And even if we don’t always know exactly what we think and experience, we certainly know that we think and experience. We know that we are conscious beings, and conscious beings (on a representational view of the mind) are beings that have thoughts and experiences. Descartes gives us this much: the first and most indubitable fact is that we think. But if thinking about or experiencing orange pumpkins (or anything else, for that matter) requires us (or the mechanisms and organs in us) to have had a certain history, as teleosemantics (on our assumption) tells us, then we can either know, by introspection, by the fact that we are conscious beings, that we’ve had such a history or we need to study history in order to find out whether we are really conscious. Either way, it is absurd. ⁸ This, of course, is what Davidson’s Swampman (a physical duplicate of D. Davidson that materialized randomly in a swamp) was meant to illustrate.
76
Fred Dretske
This epistemological argument against externalist theories of mental representation has spawned a substantial literature. I do not propose to review this literature (many of the seminal articles are collected in Ludlow and Martin 1998) because, as I see it, this objection to externalism, and to teleosemantics in particular, rests on a false assumption. Everyone (even externalists) assume, mistakenly, that what we know by introspection is not only (in the case of thought)⁹ what we think, the content of our current propositional attitude, but also that we think it, the fact that we occupy a mental state having this proposition as its content. If this assumption is false, if what we know by introspection is that it is pumpkins one is thinking (wondering, worrying, deciding) about without knowing, at least not in the same way, that one is thinking (wondering, worrying, or deciding) about pumpkins, then there is no threat to externalism. What the teleosemanticist says is constituted by external, historical, relations—the fact that one is mentally representing pumpkins—is not the fact that introspection yields: that it is pumpkins one is mentally representing. This sounds paradoxical, I know, so let me begin my defense of it by talking about young children—people who think but do not know, do not even understand, what it means to think. This won’t get me quite where I want to go, but it’s a step in the right direction. Psychologists tell us that a typical 3-year-old does not have a developed concept of belief. These children have beliefs, of course, and it is easy enough to find out, by asking them the right questions, what it is they believe, but that they believe or think these things is beyond their comprehension. They lack the concept of thought; they are, nonetheless, authorities about what they think. They enjoy a special kind of access to the content of their thought. If you want to know whether 3-year-old Suzy thinks Daddy is home or the dog is loose, just ask her. Is Daddy home? Is the dog loose? Her answers will tell you, quite unerringly, what it is she believes. You cannot, to be sure, ask Suzy directly, and in just these words, what it is she believes because (we are assuming) Suzy doesn’t yet understand what it means to believe something. But there are indirect ways of finding out. If Suzy understands what it means for the dog to be loose, you can find out whether she thinks the dog is loose by asking her whether the dog is loose. She is an authority, mind you, not on whether the dog is loose (about this topic you may know better than she), but on whether she thinks the dog is loose. Her authority on this ⁹ Throughout I use (like Descartes) ‘thought’ as a very general category. It (for me) includes all propositional attitudes. In wondering whether P, in hoping, fearing, or regretting that P, and in wanting or desiring that P, one is (in this broad sense) thinking that P. To know that you think (in this broad sense) that P, then, is to know that you occupy a mental state with intentional content. It is not merely to know that you (for instance) think rather than fear that P.
Representation and Self-Knowledge
77
topics is in no way diminished by her ignorance of what it means to think the dog is loose. Children who think (and say) that the dog is loose must, of course, understand what it means for the dog to be loose but they needn’t understand what it means to think it. They would not—indeed, they could not—say (or even think) that they know what they think since saying (or thinking) this requires them to refer or pick out what they think (that the dog is loose) as something they think and, lacking an understanding of what it means to think, they are unable to do this. Nonetheless, we can certainly describe what they know in this way. The child knows what it thinks—namely, that the dog is loose—despite not knowing (not even thinking) that it thinks this. I will be accused of ignoring scope ambiguities. It may be intelligible—even true—to say that Suzy knows something about the proposition—that the dog is loose —she believes, but that is not the same as saying that Suzy knows what she believes. That would be like saying that Suzy knows the answer to a difficult mathematical problem—that it is, say, 24—simply because she knows what number is written on the board (namely, 24) when 24 happens to be the answer to the problem. Suzy may know that the number 24 is on the board, but unless she knows this number under the description ‘the answer to the problem’ she doesn’t know what the answer to the problem is. Likewise, Suzy may think that the dog is loose, and she may know this proposition—that the dog is loose —under some description or other (perhaps as ‘what I told Mommy’), but unless she knows it under the description ‘what I think’ she doesn’t know what she thinks. All Suzy really knows is what she told Mommy. What she told Mommy is what she thinks, of course, but she doesn’t know this.¹⁰ This is a fair objection. At least it is an objection that philosophers are apt to make. So I concede the point. It is why I said at the beginning that consideration of children, people who lack the concept of thought, would not get me quite where I wanted to go. It only gets me to the point of having established that there is some sense in which children can know what it is they think without knowing that they think it. In this sense, knowing that it is X you think does not require you to know you think X. Children ¹⁰ This point could also be expressed by talking about the difference (in the context of knowledge or belief attributions) between attributive vs. referential uses of descriptions. Following Bo¨er and Lycan’s (1986: 18) description of the difference between a referential sense of knowing who the murderer is (in which it is not necessary to know the murderer murdered anyone) and the attributive sense (where this is necessary), we could say that children know what they think in the referential sense (where it is not necessary to know they think it), not the attributive sense (where it is necessary to know this).
78
Fred Dretske
can know what they think without knowing they think it in the same sense you can know what my brother is doing without knowing it is my brother doing it. But this, as my critic has been quick to point out and as I am now willing to concede, only shows that Suzy knows of what she thinks that it is that the dog is loose (the phrase ‘of what she thinks’ kept carefully outside the that-clause that expresses what it is Suzy really knows). It does not show that Suzy knows that what she thinks is that the dog is loose (the phrase ‘what she thinks’ here occurring inside the scope of the knowledge attribution). So it does not, not in any relevant sense, show that Suzy knows what she thinks—much less that she knows what she thinks without knowing she thinks it. So, to make the next step in my argument, let me shift to a person who, unlike a child, possesses the relevant concepts and beliefs. I will describe an analogous situation—knowing what someone said—and suggest that it provides a model for knowing what one thinks. Clyde gets a telephone call from his good friend Harold. Harold tells him that he is going on vacation for two weeks. Clyde hears him say this and, let us suppose, hears him say it under ideal telephonic conditions (no static, clear articulation, etc.), the kind of conditions that would ordinarily prompt us to say that Clyde knows what Harold said. So far, I hope, there is nothing suspicious. Now the twist. There are several people, all practical jokers, who, quite unknown to Clyde, enjoy telephoning Clyde and imitating Harold. They are very good at it. As far as Clyde can tell, the call he received from Harold could have been from any one of these other people. It sometimes is one of these other people. Unaware of the past deceptions, and, therefore, the very real present possibilities they create, Clyde not only believes (correctly as it turns out), without doubt or hesitation, that it is Harold he is talking to, but (incorrectly as it turns out) that he knows it is Harold.¹¹ I have Gettierized¹² Clyde’s belief that it was Harold on the phone while leaving intact his evidence for what it was that Harold said to him. The question I’m interested in is this: does the fact that Clyde does not know it was Harold who said he was going on vacation mean that he doesn’t know what Harold said to him? If asked (‘What did Harold say?’), Clyde will tell you, confidently and truthfully, exactly what Harold said. If asked whether he knows—and, if so, how he knows—that this is what Harold ¹¹ I here assume that if Clyde can’t tell the difference between Harold’s voice on the phone and the voices of several other people, any one of whom might be calling, then, whether or not he realizes it, he doesn’t know it is Harold. He certainly can’t hear that it is Harold. ¹² That is, I have described conditions in which Clyde has a justified true belief (that it is Harold he is talking to) that does not constitute knowledge.
Representation and Self-Knowledge
79
said, Clyde will tell you, once again confidently and (I submit) truthfully, that he knows this because he heard him say it. If anyone ever knows what another person says on the phone, Clyde, given the circumstances, surely, knows what Harold said. Yet, Clyde doesn’t know it was Harold who said it. Clyde thinks he knows. This, indeed, is why he so confidently reports what he knows by referring to the caller as Harold. But the truth of the matter is that Clyde is ignorant about who called him. Unlike the earlier case of the child, we now have an example in which the agent does understand the phrase (‘what Harold said’) being used to pick out the proposition that he heard expressed on the phone. He not only understands it, he confidently (and truly!) believes it refers to what he heard. That is why he describes what he heard the caller say as ‘what Harold said’. Unlike the case of the child who does not believe, does not even understand, that ‘what I think’ (when said or thought by her) is a correct description of the mental state whose content (namely, that the dog is loose) she has special access to, Clyde does understand—indeed, he truly and confidently believes, that ‘what I heard Harold say’ is a correct description of the content he has special (auditory) access to, the proposition he heard expressed on the telephone. Why isn’t this enough to know not (once again) that it was Harold who said he was going on vacation, but that what Harold said was that he was going on vacation? If it is enough, then, it seems, we have an attractive externalist model of introspection. Just as Clyde can know what it was Harold said without knowing, at least not by hearing, that it was Harold who said it, why can’t a person know what it is he thinks (by, say, introspection) without knowing, not by introspection, that he thinks it?¹³ ¹³ Strictly speaking, the analogy with knowing what you think vs. knowing that you think it should contrast knowing what Harold said with knowing that he said it—not, as I have done, with knowing that it was Harold who said it. With a few minor alterations this could be done. All we need to imagine are things—programmed sound synthesizers, for example, or (thanks to Doug MacLean for this suggestion) parrots—that can produce the same sounds as Harold when he says that he is going on vacation without actually saying or asserting anything. I assume here that parrots and machines who make the sounds ‘I am going on vacation’ are not actually saying they are going on vacation. They utter the words (and, therefore, perhaps, in direct discourse say) ‘I am going on vacation’, but they do not, by producing these sounds, say (indirect discourse) that they are going on vacation. When Clyde hears Harold on the phone saying that he is going on vacation, therefore, he can know (by hearing) what Harold said without knowing (at least not by hearing) that anything was said. I have chosen to run the analogy as I have in the text because it is simpler and more intuitive and it makes the point equally well. The important point, once again, is that the way you know the x is y may be, and often is, quite different from the way you know that it is x that is y.
80
Fred Dretske
I will be told (I have been told) by impatient skeptics that I am still ignoring subtle, but nonetheless quite valid, distinctions of scope. Since Clyde doesn’t know it was Harold who said he was going on vacation, it would be wrong to put the phrase ‘what Harold said’ inside the that-clause that expresses what Clyde knows. Clyde doesn’t know that what Harold said is that he is going on vacation. All he really knows is something of or about what Harold said—that it is that he is going on vacation. We can say that Clyde knows what the caller said, and the caller was Harold, but since Clyde doesn’t know the caller was Harold, he doesn’t know what Harold said. Clyde believes the caller was Harold, and he is (we may suppose) fully justified in this belief, but (given the special circumstances) he doesn’t know it. So it would be wrong to describe him as knowing that what Harold said is that he was going on vacation. Clyde doesn’t—not really, not strictly—know what Harold said. I don’t believe in this real, this strict, form of knowledge,¹⁴but I’m willing, once again, to concede the point to those who believe that, strictly speaking, nothing belongs inside the scope of a knowledge attribution that isn’t known to belong there by the agent to whom the knowledge is ascribed. I am, after all, interested in convincing even a stubborn internalist that there is a perfectly workable externalist model of self-knowledge and that the teleosemanticist’s commitment to a historical account of thought and experience is not to be rejected on epistemological grounds. So I’ll work with what I’m given. My third (and final) example—a slight variation on the second—is meant to comply with these more demanding strictures on scope. We have to remember that we needn’t suppose that Clyde can know what Harold said without knowing it was Harold who said it. All I need to show is that the way Clyde knows it was Harold who said it can be different from the way he knows what Harold said. If this is possible, then the way is clear to concede (to the yet unconverted) that although you cannot know what you think without knowing that you think it, the way you know what you think may be entirely different from the way you know that you think it. You can know what you think by introspection, but introspection may not be the way you know you think it. To know that you are thinking about pumpkins may require—or so externalists are free to maintain—a more indirect method, a method (empirical investigation? historical research?) compatible with a teleosemantic theory of what it takes to think about pumpkins. All that introspection tells you is that it is pumpkins you are thinking about. ¹⁴ Readers familiar with my views on contrastive statements (Dretske 1972), closure (Dretske 1970, 1971, 2005) and the incremental character of perceptual claims (Dretske 1969, ch. ) will understand why I don’t accept it.
Representation and Self-Knowledge
81
To illustrate this possibility we need imagine only a small modification of our last example. Clyde (overly suspicious from too much philosophy) finds out who called him by tracing the call. He discovers that the call originated from a phone to which only Harold had access. So he knows it was Harold who called him. He knows both what Harold said (namely, that he was going on vacation) and that it was Harold who said it, but his way of knowing the one is different from the way he knows the other. He heard what Harold said, but he did not—indeed (given the impersonators) could not—have heard that it was Harold who was saying it. Given the conditions, there is nothing distinctive about Harold’s voice to enable one to know, by hearing him talk on the phone, that it is Harold. Clyde knows it was Harold who said that he was going on vacation as a result of an empirical research, but no investigation was required to find out what Harold said. Clyde heard him say he was going on vacation. Using this slightly modified example as a model for introspection, then, the proposal is that the way we find out what it is we think (desire, wonder, fear, expect, etc.) is different from the way we find out that we think (desire, wonder, fear, and expect). The first method we call introspection. Whatever, exactly, introspection comes down to, it does not involve empirical investigation of external circumstances. That is what makes it introspection. But this is quite consistent with an empirical investigation being required to find out that the content revealed by introspection is the content of a mental state, a state whose possession of content is constituted, in part, by a network of external relations some of which are historical. Given our model, this should be as sensible as saying that Clyde heard what Harold said but needed an investigation (tracing the telephone call) to find out it was Harold who said it. Or, to give an example that might appeal to baseball fans, it is like needing a program to find out that it was Lou at bat, but not needing a program to know that Lou hit a home run. You saw him hit a home run. You know what he did by direct perception. But you know who did it—that it was Lou—indirectly, by consulting your program. We have, then, the following picture of self-knowledge: when thinking about or experiencing pumpkins, we can know, with a special kind of first-person Cartesian authority, what it is we are thinking about and experiencing—namely, orange pumpkins. Nothing illicit is smuggled into the scope of the knowledge attribution since we know both that it is pumpkins we are thinking about and that we are thinking about them. So we can, in both word and thought, and with full knowledge, pick out and refer to what we are thinking about—orange pumpkins—as something we are thinking about. Nonetheless, our way of knowing that it is pumpkins we are thinking about is, or may be, quite different from the way we know that we are thinking about them. Although we enjoy first-person
82
Fred Dretske
authority about the first, we enjoy no privileged access to the second fact. It may be, as teleosemanticists have it, that to know you are thinking about or experiencing pumpkins requires information not obtained by looking inward. Introspection doesn’t tell you that you think, only what you think. If this is right, we have an answer to the epistemological objection to teleosemantics. A special authority about, and a privileged access to, one’s own thoughts and experiences is compatible with a historical theory of thought and experience. The only remaining question is whether this answer to the objection gives a plausible account of self-knowledge. Is this really all that introspection yields? Do we, in fact, use a different method to find out that we think from the method (if it is a method) that tells us what we think? My purpose here was only to argue that there was no valid epistemological objection to teleosemantics. I’ve already done this. I should quit now. But I can’t resist a few remarks about the plausibility of this picture of the mind’s knowledge of itself. As my examples show, we often know that x is y by some direct method (hearing, seeing, introspection) without knowing, without being able to know, by that same direct method, that it is x that is y. If we know that it is x that is y, our way of knowing this may be, and often is, quite different from our way of knowing that x is y. I don’t have to see, even be able to see, that it is water that is boiling to see that the water is boiling. There is, after all, nothing about water to distinguish it from gin, vodka, and a variety of other liquids. I needn’t be able to see that it is water in order to knowingly refer to what I see to be boiling as water. So if it is water, and if I reasonably and truly believe it is water, there is nothing to prevent me from saying I can see that the water is boiling. That, I submit, is how I know that the water is boiling. If I know it at all, though, that isn’t how I know it is water that is boiling. If I actually know it is water, I probably know that in some way other than the way I know it is boiling. The fact that I came to know (or believe) it is water by chemical analysis doesn’t mean I can’t see that the water is boiling. Why shouldn’t the same be true of introspection. The fact that I found out I think by having someone (parents? teachers? friends? Descartes?) tell me doesn’t mean I can’t now discover what I think by simple introspection. The analogy with ordinary perception can be pushed a little further. Perception of ordinary dry goods tells us what is in the physical world, not that there is a physical world. I see that there are cookies in the jar, people in the room, and (by the newspapers) continued violence in the Middle East. That is how I know there are cookies, people, and violence in these places. Cookies, people, and violence are physical things that exist independently of my perception of them. Do I, therefore, know, by visual perception, by seeing, that there are things that exist independently of my perception of them? Can I see that there is a material world and that, therefore, solipsism is false?
Representation and Self-Knowledge
83
I don’t think so. It seems more reasonable to say that assuming there is a physical world, or assuming we know (in some other way) that there is a physical world, perception tells us what sorts of things are in it—cookies, people, and violence. Visual perception has the job of telling me what physical objects I see, not that I see physical objects. If my perceptual faculties had the latter job, the job of telling me that I was (in effect) not hallucinating, not aware of some figment of my own imagination, they would be incapable of discharging their responsibilities. For, as we all know, hallucinatory cookie jars can, and sometimes do, look much the same as real cookie jars. You can’t see the difference. If it’s a real object you see, perception will tell you whether it’s an orange or a banana (a difference that is plainly visible), but perception cannot tell you whether it’s a real orange or just a figment of your imagination. That difference isn’t visible. Memory has a similar structure. Memory tells us what happened in the past—the specifics, as it were, of personal history. It does not tell us there is a past. I can remember (hence, know) what I had for breakfast this morning. No trick at all. I distinctly remember that it was granola. Nonetheless, despite this (what I remember) implying that the past is real (if it isn’t real, I didn’t have breakfast this morning; hence, do not remember having granola for breakfast this morning) this doesn’t mean I can remember that the past is real. If I know the past is real, I don’t know this by remembering that it is real. That isn’t a way to answer Russell’s skeptical question about the past.¹⁵ If I know the past is real, I know it in some way other than by memory. Memory is a faculty that tells me what occurred in the past given that there was a past just as perception tells me what is in the material world given that there is a material world. Maybe I have to know the past is real in order to remember what I had for breakfast this morning (I doubt it, but let that pass), and maybe I have to know there is a physical world to see whether there are cookies in the jar (let that pass too), but the point is that I do not have to know these things by memory and vision in order for memory and vision to tell me (give me knowledge of ) what I had for breakfast this morning and what is in the cookie jar. Introspection is like that. Introspection tells me what is in my mind, what it is I am thinking, wanting, hoping, expecting, and the kind of experiences I am having. It doesn’t tell me I really have a mind, mental states with content. If I know that at all, I know it in some way other than by introspection, the faculty that, given that I have thoughts and feelings, tells me what I’m thinking and feeling. ¹⁵ Russell’s question: How do you know the world and all its contents were not created a few moments ago complete with memory traces, fossils, history books, etc.—complete, that is, with all the indicators you rely on to tell you about the past?
84
Fred Dretske REFERENCES
A, A., C, R., and P, M. (eds.) (2002), Functions: New Essays in the Philosophy of Psychology and Biology (Oxford: Oxford University Press). BE¨ , S. E., and L, W. (1986), Knowing Who (Cambridge, Mass.: Bradford Books, MIT Press). B, C. (2002), ‘A Rebuttal on Functions’, in Ariew et al. (2002: 63–112). C, R. (1975), ‘Functional Analysis’, Journal of Philosophy, 72/20: 741–65. D, F. (1969), Seeing and Knowing (Chicago: University of Chicago Press). (1970), ‘Epistemic Operators’, Journal of Philosophy, 68/24: 1007–23. (1971), ‘Conclusive Reasons’, Australasian Journal of Philosophy, 49/1: 1–22. (1972), ‘Contrastive Statements’, Philosophical Review, 81/4: 411–37. (1981), Knowledge and the Flow of Information (Cambridge, Mass.; Bradford Books, MIT Press). (1988), Explaining Behavior (Cambridge, Mass.: Bradford Books, MIT Press). (1995), Naturalizing the Mind (Cambridge, Mass.: Bradford Books, MIT Press). (2001), ‘Norms, History, and the Mental’, in Denis Walsh (ed.), Naturalism, Evolution and Mind (Cambridge: Cambridge University Press); previously pub. in Dretske, Perception, Knowledge, and Belief: Selected Essays (Cambridge: Cambridge University Press, 2000). (2005), ‘The Case against Closure’, in Matthias Steup and Ernest Sosa (eds.), Contemporary Debates in Epistemology (Malden, Mass.; Blackwell). EC¸, B. (2002), ‘Indeterminacy of Function Attributions’, in Ariew et al. (2002: 291–313). F, J. (1990), A Theory of Content and Other Essays (Cambridge, Mass.: Bradford Books, MIT Press). G, P. (1957), ‘Meaning’, Philosophical Review, 66: 377–88. H, V. G. (2002), ‘On the Normativity of Functions’, in Ariew et al. (2002: 144–56). L, P., and M, N. (eds.) (1998), Externalism and Self-Knowledge (Stanford, Calif.; CSLI Publications). M, R. (1989), ‘In Defense of Proper Functions’, Philosophy of Science, 56/2: 288–302. N, K. (1991), ‘The Teleological Notion of Function’, Australasian Journal of Philosophy, 69: 454–68. P, M. (2002), ‘Pagan Teleology: Adaptational Role and the Philosophy of Mind’, in Ariew et al. (2002: 263–90). W, D. M. (2002), ‘Brentano’s Chestnuts’, in Ariew et al. (2002: 314–37).
4 The Epistemological Objection to Opaque Teleological Theories of Content Frank Jackson 1. A COMMONPLACE ABOUT CONTENT After you’ve seen someone walk through a minefield, you have a pretty good idea where they think the mines are located. After you’ve dined regularly with someone, you have a pretty good idea of what they like to eat and drink. And so on and so forth. It is a commonplace that what people do and say tells the folk—that is, you and me, Shakespeare and Aristotle, most of us when we are not drawing on specialist knowledge in cognitive science or whatever—a good deal about the contents of the beliefs and desires of our fellow human beings, and more generally about what they think. The observation of behaviour in circumstances often grounds justified belief about the contents of intentional states, and, what is more, we qua members of the folk know this. This commonplace does not presume behaviourism of course—we all know that the direction of the wind is indicated by the behaviour of a windsock but facts about the wind cannot be analysed in terms of facts about windsocks. The fact that this is a commonplace means that any theory of content should respect it. This chapter argues that certain versions of teleological theories of content are inconsistent with it, and more generally with the fact we folk qua folk have many justified beliefs about the contents of beliefs and desires, and thought more generally. The versions I have in mind are theories offering biconditionals like: (A) x believes that P iff x is in a state that . . . where the ellipsis is filled by a clause that relates to selectional matters that are opaque to the folk, and where these biconditionals are understood
86
Frank Jackson
as part of a theory which identifies content properties with the relevant selectional ones. The last proviso is important. Nothing I will say goes against the views of those who are content to say that biconditionals like (A) are interesting, or reflect something important about content, or are helpful ways of understanding how it is we came to have contentful states as we evolved, or . . . It is theories that identify content with opaque selectional properties (or opaque teleological properties more generally, but we will stick to selectional versions of teleology) that are my target. This is not to make my target boringly small: surely the obvious way to understand someone who offers a theory of content is as offering an identification of content properties, an account of what content is, and so the obvious way to understand someone who offers a teleological theory of content is as offering an identification of content properties with selectional properties. An example of the kind of teleological theory that I am targeting is one that says:¹ (B) x believes that P iff x is in a state that is selected (in the evolutionary sense) to co-vary with P and (C) x desires that P iff x is in a state that is selected (in the evolutionary sense) to bring about P and sees these two clauses as making the case for identifying being a belief that P with being selected to co-vary with P, for identifying believing that P with being in a state that is selected to co-vary with P, for identifying being a desire that P with being selected to bring about P, and for identifying desiring that P with being in a state selected to co-vary with P, or something along these broad lines. In sum, I mean a theory that uses connections between selectional facts and intentional facts to deliver identifications of being and having intentional states that P with being and having certain selectional properties that have the appropriate relation to that P. Now the folk qua folk do not have opinions, let alone justified ones, about the theory of evolution; and even if they did, they do not qua folk have justified views about which structures are or are not selected for—it took serious research to show that the chin is not selected for. Moreover, even if the folk did have opinions including justified ones about which ¹ For a view of this kind, see Papineau (1993: 94). He is not offering it as a finished theory but as a sketch to give the general idea. Our argumentation will be independent of the various qualifications and refinements.
Opaque Teleological Theories of Content
87
structures are selected for, they do not qua folk have opinions, let alone justified ones, about what these structures are selected for —the selectional histories of the ear and the larynx are major research topics. Moreover, no one thinks that observations of subjects interacting with their environments are enough to justify attributing properties like having a state selected to covary with such and such, or having a state selected to bring about so and so, to subjects. To suppose otherwise would make a nonsense of all the work that went into establishing the theory of evolution. At first glance, we seem to have a serious problem for teleological theories of content of the kind in question, henceforth opaque selectional theories of content: they seem to imply that the folk do not have the justified opinions about the contents of the intentional states of their fellow humans that they clearly do have, and that, in particular, even extended observations of interactions with environments are not enough to justify ascribing contents. After all, the folk do not have justified opinions about magnolia metal because they have never heard of it, and, even if they had heard of it, would not normally have justified opinions about which metals are examples of it.² In the same way, it seems that they would not, on opaque selectional theories, have justified opinions about intentional contents because these contents are properties they have never heard of, or need never have heard of. And, even when they have heard of them, the absence or presence of these properties is not something the folk need have a justified view about in order to have justified opinions about what their fellows are thinking, and is not something we get justified belief concerning out of observation of interactions with the environment. My contention in this chapter is that what seems right at first glance seems right after a number of glances. I will argue that opaque selectional theories of content cannot explain how it is that we folk have the justified opinions we do about intentional contents. They cannot explain, for instance, why Shakespeare was so often justified in his views concerning what those around him wanted and believed based on his observations of their behaviour. Our argument will be independent of the detail of the various teleological accounts of content provided only that they are opaque selectional accounts in the sense already explained. Where necessary I will frame matters in terms of the very simple teleological account of the content of belief and desire given above, but the point being made will not depend on the simplification (which all parties can agree is substantial); it will depend on the opacity.
² It is a lead-base alloy used in bearings.
88
Frank Jackson
2. OUR ARGUMENT LAID OUT I know many teleological theorists will insist that their theory cannot be in trouble from such a ‘quick’ objection and that the objection commits some simple mistake or other. Most of the rest of this chapter is a series of replies to the various objections to the folk epistemology objection, or FEO, as I’ll call it, from people who insist that I have made some simple mistake or other. But let me spell out the structure of the objection before we proceed to look at some responses to it. The objection in schematic form runs thus: Premise 1. Having an intentional state with such and such content is a property justifiably ascribed in so and so circumstances. (Premise supported by reflection on Shakespeare and the folk generally.) Premise 2. Having a state, or being in a state, with so and so a selectional history is not justifiably ascribed in so and so circumstances. (Premise supported by reflection on what is needed to justify belief in selectional matters.) Conclusion. Having an intentional state with such and such content is not identical with being in a state with so and so a selectional history. (Leibniz’s Law) This argument is valid and there is no problem about applying Leibniz’s Law to properties of properties. The same style of argument can, of course, be run for the intentional states themselves, for being the belief that P as opposed to having the belief that P, for instance. It is time to look at the possible objections. They are all constructed from objections that I have come across in one form or another but I have not sourced them for fear of misrepresenting. 3. THE OBJECTION THAT THE SELECTIONAL WILL CORRELATE WITH THE FOLK Suppose we have a theory of the form (D) x believes that P iff x is in a state that is K where K is an opaque selectional matter. We might have very good reason to hold that (E) x is in a state that is K iff x is in a state which is J where J is a matter transparent to the folk. It will then be the case, according to the theory, that
Opaque Teleological Theories of Content
89
(F ) x believes that P iff x is in a state that is J. For example, suppose that belief that P is a matter of being in a state selected to co-vary with P but considerations to do, say, with the difficulties of surviving should there be too much dissonance between what co-varies with how things are currently and how things were as we evolved tell us that current functional roles of the kind widely agreed to be transparent to the folk will match up with the relevant selectional roles in ways that mean the two are reliable guides to each other. In that case, runs the objection, it might well be that something like (F ) was true consistently with content being determined by something like (D). In that case, continues the objection, selectional theorists can escape the folk epistemology objection by pointing out that, although content is determined by something opaque to the folk, the something opaque is reliably correlated with something transparent to the folk. Reply. My response to this objection is that it is one thing for there to be a correlation, another for the folk to be aware of it or to have a justified opinion that there is such a correlation. The history of medicine is full of correlations we know about now and wish we had known about much earlier—it would have saved many lives if we had known about them sooner. All the same, we did not know about them sooner and were not in any position to have a justified opinion about them sooner. Perhaps, and anyway I am granting the point here, the selectional roles are correlated with matters transparent to the folk. This will not ground justified opinion on the part of the folk about contents determined by opaque selectional facts if the folk know nothing of the correlation and have no justified belief in its existence. And isn’t this precisely the position of the folk? Shakespeare had no idea, and could have had no idea, that there is a correlation between the facts from which he justifiably inferred contents and the selectional facts of teleological theory. But this did not debar him from having justified opinions about what people were thinking. The key point can be brought out by considering David Papineau’s discussion of an objection of Andrew Woodfield’s. Woodfield asks for a reason to believe that ‘the teleological theory’s ascriptions of content coincide with those made by everyday psychology’.³ Papineau’s answer ‘is simply that it would be a mystery that the desire for some physical result r should do what everyday psychology says it does [deliver what is desired for agents with true beliefs] . . . unless it has been selected to produce r’.⁴ Be this right or not, it is not something available to the folk. It would tell us why the folk get it right ³ Papineau (1993: 94). ⁴ Ibid. 97–8. He makes a similar argument for belief on p. 99.
90
Frank Jackson
given a selectional story but not why they are justified in holding that they get it right.
4. THE OBJECTION FROM SCIENTIFIC IDENTITIES Teleologists often present their view as a kind of scientific identification of content properties and states with selectional properties and states playing selectional roles.⁵ And it might well be thought that these identities tell us that there must be something amiss with the FEO. The distribution of water is not a separate matter from the distribution of H2 O, and yet the folk can have justified opinions about where there is water despite not having justified opinions about where there is H2 O, or even knowing that there is such stuff as H2 O. Shakespeare had many justified opinions about where there is water, despite not knowing any modern chemistry. Teleologists should, runs the objection, hold that as water is to H2 O, so content is to selectional role. Reply. It is important to appreciate what the water–H2 O example tells us, and what it does not tell us. It tells us that someone can consistently offer a claim of the form (D) x believes that P iff x is in a state that is K where K is an opaque selectional matter, and where (D) is advanced as a necessary a posteriori truth and not as an a priori truth or as a conceptual analysis in some traditional sense, without thereby committing themselves to an epistemological claim about justified belief about content requiring justified belief about opaque selectional matters. What it does not tell us is the sense in which offering (D) is offering what might properly be thought of as a teleological theory of content. As it stands, (D) is compatible with theories of content that are antipathetic to teleology. Consider the theory that holds that having a belief with content that P, and being in a state that is K, are distinct properties, and more generally that content is quite distinct from anything teleological: for example, the theory might hold that being a belief that P and being selected to co-vary with P are quite distinct properties, and that believing that P and being in a state selected to co-vary with P are quite distinct properties. ⁵ Papineau is explicit that this is how he understands teleological theories; see his comments on theoretical reductions on (1993: 93), and his comments at the beginning of (2001). Braddon-Mitchell and Jackson (1997) argues from the perspective of the critic rather than the supporter, that this is the best way to read the view.
Opaque Teleological Theories of Content
91
However, the theory also holds that although the relevant content property and the relevant selectional property are distinct, they metaphysically necessitate each other. This is not a teleological theory of content. It is no more a teleological theory of content than are necessitarian dual attribute theories of mind versions of physicalism. Some dual attribute theorists hold that the phenomenal feels of sensory states are properties distinct from any that appear in physicalist theories of mind, but that they are necessarily connected to properties that appear in physicalist theories. The necessary connection does not turn their view into a version of physicalism. Mutatis mutandis for teleology. It follows that we need to add something to (D) in order to get a view that can properly be regarded as a teleological theory of content. The obvious addition, as we in effect noted near the beginning, is an identity claim. In addition to holding that a suitable instance of (D) is a necessary a posteriori biconditional, teleologists should hold that suitable instances of (E ) having such and such content = being a state that plays so and so a selectional role and (E ) being in a state with such and such content = being in a state that plays so and so a selectional role are necessary a posteriori truths. Again, the water–H2 O case might be held to be suggestive. Not only is it a necessary a posteriori truth that x is water iff x is H2 O, it is a necessary a posteriori truth that water is identical with H2 O (modulo worlds where there is no water). But now we have the trouble we noted at the outset. (E ) and (E ) are false: for each it is the case that the property on the left-hand side differs in its epistemic properties from the property on the right-hand side; that’s the nub of the FEO. And a similar consideration applies in the case of scientific identities. When Shakespeare believed that there is water in a glass in front of him, he believed that things are a certain way in the glass, but what he believed about how things are in the glass is not what we believe when we believe that a glass contains H2 O. It follows that how things are believed to be when water is believed to be somewhere is not how things are believed to be when H2 O is believed to be somewhere. And this is what one would expect given the a posteriori nature of the identity between water and H2 O. If what one believes about how things are when one believes that there is water is the same as what one believes about how things are when one believes that there is H2 O, then from the very moment humans believed that there is water, they believed that there is H2 O. But the a posteriori
92
Frank Jackson
nature of the identity between water and H2 O confirms the common-sense view that believing that there is H2 O came later in time. The property we ascribe when we believe that there is water differs from the property we ascribe when we believe that there is H2 O. We believed the first to be instantiated long before we believed the second to be instantiated. In sum, if we call the first property ‘being water’ and the second ‘being H2 O’, there is an epistemic argument that shows that being water = being H2 O, just as there is for having such and such content = being a state that plays so and so a selectional role, and for being in a state with such and such content = being in a state that plays so and so a selectional role. Some will bite the bullet and say that we should think of the water–H2 O case in the way many view the Hesperus–Phosphorus case. Many hold that the proposition that Hesperus = Phosphorus is the same proposition as that Hesperus = Hesperus, and, in consequence, that one believes the first iff one believes the second, and that if one believes that some planet = Hesperus, one believes that that planet = Phosphorus. Surprising but true. In the same way, they might say, what we believe about how things are when we believe that there is water is the same as what we believe about how things are when we believe that there is H2 O, and, in particular, we should say that Shakespeare believed that various glasses contain H2 O precisely because he believed that they contain water. Surprising but true. My claim, therefore, that there is a difference in the epistemic properties of being water and of being H2 O is a mistake, and ditto for selectional and content properties. However, the contention that the proposition that Hesperus = Phosphorus is the same proposition as that Hesperus = Hesperus is part of a package deal that affirms that it is a priori that Hesperus = Phosphorus— surprising but true—because we know that the one thing that those who so contend should not say is that the proposition that Hesperus = Phosphorus is not knowable a priori, whereas the proposition that Hesperus = Hesperus is knowable a priori. For ‘they’ are one and the same proposition on the contention under discussion.⁶ But whatever should be said about the thorny issues surrounding identities expressed using proper names, surely scientific identities like ‘water = H2 O’ and ‘acids = proton donors’ are a posteriori. (And teleologists who draw on scientific identities in explaining their view insist on the point, see Papineau 1993). I am not, of course, saying that water = H2 O. The relation between water and H2 O is like that between the colour of the sky and being blue. The colour of the sky = being blue but the property we ascribe when we ⁶ For some of the issues here, see e.g. Soames 2002: 4 and the discussion that follows.
Opaque Teleological Theories of Content
93
say that something has the colour of the sky is not being blue; it is being same-coloured with the sky. Likewise, the property we believe something to have when we believe it to be the colour of the sky is not being blue (on the obvious reading); it is being same-coloured with the sky. Being water and being H2 O differ just as being blue and being the colour of the sky differ, but water and H2 O are one and the same just as (being) blue and the colour of the sky are one and the same property. 5. THE OBJECTION FROM SECOND-ORDER PROPERTIES Many of the things we say about how things are are to the effect that something has a property that itself has a property. To be fragile is to have some internal nature that causes breaking on dropping: there is the internal nature, and there is what it does or would do. Or take the colour of the sky example just mentioned: to say that something has the colour of the sky is to say that it has a colour that has the property of including the sky in its extension. We might call the properties ascribed in such cases ‘secondorder’, meaning not that they are properties of properties but that they are properties possessed by x in virtue of x’s possession of a property that itself has a property: the property of having whatever property is so and so. An example much discussed in the philosophy of mind is the version of functionalism tailored to be compatible with a type–type mind–brain identity theory.⁷ It draws a sharp distinction between being in pain and pain. To be in pain is to instantiate a kind that fills so and so functional role; pain is the kind (one that may or may not vary from creature to creature, etc.) that does fill the role. This gives two property identities: being in pain = instantiating a property that fills so and so functional role pain = the property that fills so and so functional role (neural state N in such and such creatures, as it might be). In the same way, runs the objection to FEO, selectional theories must distinguish two properties. One is the property we ascribe to someone when we say that they believe that P or (same thing) the property we believe them to have when we believe that they believe that P. This property is not a selectional property—it is, for instance, a property we folk are justified in believing someone to have in circumstances where we are not justified in believing them to have any relevant selectional property. However, this property is the property of having a property which is thus and so, and this ⁷ See e.g. Lewis (1983) and Prior et al. (1982).
94
Frank Jackson
latter property is, as an a posteriori matter, a selectional property. We have, that is to say, two property identities: (F ) believing that P = the property of having a property that is thus and so (G) belief that P = the property that is thus and so. The property referred to in the right-hand side of (F ) is the property we believe that x has when we believe that x believes that P. It is the property that we are entitled to ascribe on the basis of folk-available data including especially interactions with the environment. Therefore, it is not a selectional property. However, this does not preclude the property referred to in the right-hand side of (G) being, as an a posteriori matter, a selectional property. The objection from second-order properties to FEO holds that the sense in which a selectional style of teleological theory of content is correct lies in the truth of (G), where the property referred to in the right-hand side of (G) is indeed a selectional property. Reply. There is a sense in which this is not a selectional theory of content. Content properties are the relevant commonalities among entities with contentful states. The property of having a state with the content that P is the relevant commonality among the creatures that have an intentional state with content P. And the suggestion maintains that the relevant commonality for creatures that believe that P is having a property which is thus and so, where this is not a selectional property. If it were, it would not be a folk-available property. Similar remarks apply to the properties of the intentional states per se. The point here is similar to the one often made against the type–type mind–brain identity theory of mind. Although the theory holds that pain is neural state kind N, the properly psychological property is the relevant commonality among those in, say, pain, and that is a functional property and not a neurological one on the theory. But let’s set the branding issue aside. The substantive problem with the suggestion on behalf of selectional theories that draws on second-order properties arises from the need for the second-order property that figures in the right-hand side of (F ) to be one we may reasonably ascribe on the basis of folk-available data, to be one ‘we are entitled to ascribe on the basis of folk-available data including especially interactions with the environment’. The folk are not cognitive scientists. All the same, they know that when a parcel of matter with clear boundaries displays distinctive, patterned interactions with its environment, there are internal properties of that parcel that play a big role in underpinning the interactions. It is very plausible that when we believe that x believes that P, what we believe about x involves some implicit assumptions about how the interactions are generated. This
Opaque Teleological Theories of Content
95
is confirmed by the fact that when ‘blockhead’ cases and Martian marionette cases are described to us, we immediately designate them as cases where there is no contentful thought.⁸ That which generates ‘their’ interactions is not of the right kind, and we know this qua member of the folk. It is, therefore, plausible that the property we ascribe to x when we believe that x believes that P, or when we use ‘believes that P’ of x, is a second-order property in the requisite sense. But it is important that the ‘thus and so’ in (F ) be given an undemanding reading in the sense of a reading that preserves the point we have been focusing on: the folk qua folk have plenty of justified beliefs about content. For example, if we spelt out ‘thus and so’ in a way that stated that the internal workings be carbon-based, as in (F carbon) believing that P = the property of having a carbon-based property that is thus and so, we would wrongly make it the case that the folk who do not know that we are carbon-based—Shakespeare and Aristotle would be examples—lack justified beliefs about content.⁹ All the same, this leaves quite a bit of room to manoeuvre. There is a lot about internal goings on that is available to the folk from interaction patterns. Consider the phenomenon of imprinting. Baby ducks are disposed to keep company with the first thing of a suitable sort that they see after hatching. Usually it is the mother duck obviously, but sometimes it is a dog, the experimenter, or whatever. They imprint on the first thing that they see, and their behaviour is explained in terms of their having imprinted on the mother, the dog, or . . . Although we ascribe imprinting on the basis of observation of behaviour, what we ascribe, or are in a position to ascribe, goes well beyond behavioural patterns. We know, for instance, that: the duck’s initial sighting lays down a persisting trace inside the chicken, otherwise its following behaviour would fade away quickly; that the nature of the internal trace that is laid down is a function of the nature of the thing first seen, otherwise it would not be able to discriminate between the thing first seen and things seen subsequently; and that the trace laid down is causally connected to the ways the legs and the head operate, otherwise the information being carried by the trace inside the baby duck would be ⁸ ‘Blockhead’ is Block’s (1981) example of a ‘person’ who works by look-up tree. The Martian marionette is Peacocke’s (1983, final chapter) example of a ‘person’ who is effectively a puppet controlled by radio signals from Mars. ⁹ This is not a quick argument for multiple realisability. Rigidifying is epistemically undemanding—knowing that something is F is enough for knowing that it is actually F. A suitable rigid reading of the ‘thus and so’ might, for example, reference-fix on something which turns out to be carbon-based.
96
Frank Jackson
irrelevant to the movements of its head and legs in sustaining the accompanying behaviour. All of this is something available and implicitly known to the folk. The little reasoning sketches given above do not call on specialist knowledge in cognitive science. However, it is also true that what is so available is restricted to the causal, functional, and informational underpinnings of the behavioural interactions. It concerns leaving traces, causal transactions between traces, casual links to legs, the causal role of the eyes, and so on. In order to preserve the folk-availability of content, the ‘thus and so’ in (F ) must pertain to matters of this kind. What is obvious to the folk and something they may justifiably believe and ascribe on the basis inter alia of observations of behaviour can be complex, but the complexity is restricted to complex causal and functional roles, and the like. It does not include selectional roles. If it did Wallace and Darwin’s insights would have been available to the folk qua folk. This blocks the possibility of the identification in (G) being with selectional properties. The situation can be put as follows. We need an undemanding reading of the ‘thus and so’—one that makes the existence of a property that is thus and so folk-available—in order for (F ) to meet the folk-availability constraint. Mark any such reading ‘thus and so*’. But then what we get from (G) is that belief that P = the property that is thus and so* and the property that is thus and so* is not a selectional property. It might be a functional property of the kind we can justifiably believe in given the information available to the folk, or it might be whatever property plays the relevant functional role, because if we can justifiably believe that some functional property is instantiated we can justifiably believe that there is a property playing the functional role. But in neither case is it a selectional property. Selectional properties are not folk-available, and they do not play the relevant functional roles—that is done by neural properties, or maybe internal functional architecture. In sum, the second-order property way of reading selectional cum teleological theories of content faces a dilemma. Any reading of (F ) that allows some instance of (G) to be an identification of a content property with a selectional property is a reading that means that the folk-availability constraint is violated. Any reading of (F ) that meets the folk-availability clause blocks any instance of (G) being an identification of content with a selectional property.
Opaque Teleological Theories of Content
97
6. THE OBJECTION THAT LEIBNIZ’S LAW DOES NOT APPLY TO EPISTEMIC PROPERTIES ‘‘You went wrong near the beginning when you said ‘there is no problem about applying Leibniz’s Law to properties of properties’. There is a problem if the properties of properties are epistemic ones like being justifiably supposed to obtain in so and so circumstances. The opacity of belief contexts tells us that.’’ Reply. This is tantamount to denying that we have beliefs about properties. For consider the right response for those who hold that ‘a = b’, ‘S believes that a is F ’, and ‘It is false that S believes that b is F ’ can be true together. The right response is that when S believes that a is F, S does not have a belief about a, the thing. S has, rather, a belief about the proposition that a is F, and the explanation of how it can be false that S believes that b is F and true that S believes that a is F, when a = b, is that the proposition that b is F is a different proposition from that a is F, and the beliefs in question are about propositions and not objects. Thus there is no violation of Leibniz’s Law. (Of course, some insist that S’s belief that a is F is about a, but they are the same people who insist that if S believes that a is F, then S ipso facto believes that b is F in the case where a = b.) I think we should resist any suggestion that we do not have beliefs about properties. We really do have beliefs about how things are in certain parts of the world and how things are with certain things. Our beliefs really do place things in categories, and that’s to assign them properties. Our assignments may be correct or incorrect but it is properties, not something else or nothing, that gets assigned. It might be objected that an argument like the one just rehearsed for objects can be developed for properties in order to show that we do not strictly have beliefs about properties. Surely the following can be true together (H) S believes that a is blue, (I) It is false that S believes that a has the colour of the sky, ( J) Blue = the colour of the sky. How so if S’s belief is about the single property—blue (= the colour of the sky)? The answer is that we can read (I) in two different ways. We can read the claim that S believes that a has the colour of the sky as being true when S believes that a is blue. Is there some other colour that S believes a to have? On this reading it is not possible for (H), (I), and (J) to be true together. The more obvious way to read (I) is as saying that it is false that
98
Frank Jackson
S believes that a has the property of being same-coloured with the sky. But blue = being same-coloured with the sky. There is no way to read (I) that makes trouble for the intuitive view that when S believes that a is F, S has a belief about a property, namely, the property that S believes a to have. 7. THE OBJECTION FROM THE REVISIONARY NATURE OF OPAQUE SELECTIONAL THEORIES ‘‘To require selectional theories of content to meet the folk-availability constraint is to misunderstand their revisionary nature. Yes, Aristotle and Shakespeare lacked justified belief about what their fellow human beings were thinking, but that is no objection to teleology of the opaque selectional kind once its revisionary nature is taken into account.’’ Reply. There are revisionary theories and there are revisionary theories. In the company of many, I think that there is a revisionary element in many of the most widely discussed analyses in philosophy.¹⁰ Compatibilist analyses of free will and continuity accounts of personal identity often have a normative element; they are not purely descriptive. ‘‘This is what our concept should be when we take into account all we know, identify potential confusions in the folk concept, and the role we give the concept’’ is the tenor of the discussion. However, these analyses preserve our epistemic entitlements. They do not imply across the board that our beliefs about which actions are free and when the person I am talking to today is the person I was talking to yesterday are ones we are not entitled to. For example, continuity theorists who allow that there is a Cartesian element in the folk concept of personal identity allow that Shakespeare was by and large entitled to his opinions about when he was or was not using the same actor to play different parts. Indeed, often the reason for favouring some analysis or other of personal identity or free action is in part that it makes sense of how it is that our judgements about identity and free action are reasonable ones. REFERENCES B, N (1981), ‘Psychologism and Behaviorism’, Philosophical Review, 90: 5–43. B-M, D, and J, F (1997), ‘The Teleological Theory of Content’, Australasian Journal of Philosophy, 75/4: 474–89. ¹⁰ See Jackson (1998: 42–3).
Opaque Teleological Theories of Content
99
J, F (1998), From Metaphysics to Ethics (Oxford: Clarendon Press). L, D (1983), ‘An Argument for the Identity Theory’, repr. in Lewis, Philosophical Papers, i (New York: Oxford University Press), 99–107. P, D (1993), Philosophical Naturalism (Oxford: Blackwell). (2001), ‘The Status of Teleosemantics, or How to Stop Worrying about Swampman’, Australasian Journal of Philosophy, 79/2: 279–89. P, C (1983), Sense and Content (Oxford: Oxford University Press). P, E. W., P, R, and J, F (1982), ‘Functionalism and Type–Type Identity Theories’, Philosophical Studies, 42: 209–25. S, S (2002), Beyond Rigidity (New York: Oxford University Press).
5 Useless Content Ruth Millikan A central claim of teleological theories of content, certainly of my own teleological theory of content, has been that the content of a representation is determined, in a very important part, by the systems that interpret it. I should be more careful: . . . by the interpreting systems with which the representation-producing systems have been designed to cooperate. Content is determined, in large part, by the ways these interpreting systems are designed to react to the representations presented to them. More exactly, content is determined by the rules or functions in accordance with which representations need to correspond to world affairs (to what they represent) if the responses of these interpreters are going to produce the effects they are designed to produce by way of normal mechanisms (LTOBC, VM ).¹ Teleosemantics is in large part ‘consumer semantics’. It cannot be the function of a system to produce representations unless the representations that are produced have some use. But they will not have a use unless they have consumers that know how to use them, consumers such as, perhaps, some further systems in the brain or in the brains of conspecifics. Further, if design is interpreted to mean biological design and use is interpreted to mean biological function, then it looks as though the content of a representation must depend, in the end, on its having biological utility. Various untoward consequences have been claimed to follow from this view. I will discuss some of these alleged consequences. Most of what I will say I have said before one place or another, but in diverse contexts in a variety of different essays. I will try to draw these various thoughts together. The first worry is sometimes raised in a broader context. Nonhuman animals, it is said, mentally represent only those aspects of the world that ¹ I use the following abbreviations: LTOBC for Language, Thought and Other Biological Categories; OCCI for On Clear and Confused Ideas; VM for Varieties of Meaning.
Useless Content
101
they need to represent in order to guide their limited repertoires of behavior. The worlds that they live in are severely restricted; they register only those aspects that are directly relevant to fulfilling their practical needs. Add to this the increasingly popular Gibsonian position that perception is always, at root, perception of affordances. The animal perceives its environment only as a series of possibilities for action, as containing things to approach or to run from, things to climb up on, places to go through, things to eat, things to hide under, things to throw, and so forth. The animal’s world is perceived only as it relates directly to its own interests, needs, and abilities. What is afforded to an animal is necessarily a function of its own peculiar capacities for action. Where animals have different kinds of abilities, then, they must live in different perceived worlds. Nor did Gibson restrict this description of perception to nonhuman animals. Human perception is included. Apples are perceived as for eating, mailboxes as for posting letters in, and so forth. But if this is the character of our basic perception, and if human cognition is built on top of equipment originally designed for this kind of perception, it seems to follow that we too live in a world with subjectively imposed limits. Our world cannot be an objective world, but is a world created in large part by our peculiar human constitutions and abilities. The view that the content of every representation depends in an essential way on its use seems to play directly into the hands of this kind of argument. If teleological theories of content are right, we humans should be intrinsically unable to represent a truly objective reality. The enterprise of empirical science is doomed to presenting only a warped and truncated vision of the world. I agree with Gibson that basic perception is perception for action, indeed, that basic perception is perception of affordances (VM, pt. ). Gibson’s position well deserves to be gaining in popularity. But for reasons I will review below, not all human perception is perception directly for action, nor is all cognition based on perception designed directly for use in action. I will return to this theme a bit later. At the moment there are more elementary points to be made. First, a perception or other representation of the relation of something to oneself or of the utility of something for oneself is not the same as a subjective representation or a representation of something merely subjective. It may be, of course, that when the objects one observes have certain kinds of relations to oneself, utilities or disutilities for oneself, then biases in observation are more likely to occur. But affairs involving relations to oneself or utilities for oneself are, in themselves, perfectly objective affairs. There is no intrinsic reason why an animal that perceives only affairs of practical significance to itself should not perceive these affairs perfectly objectively and perfectly accurately. For example, in order to manipulate any object in my world I will need to perceive
102
Ruth Millikan
its spatial relation to me. But its spatial relation to me is a perfectly objective relation, and I am best off if I perceive it accurately. Second, that a creature represents only a few features of its world does not mean that it represents its world as having only a few features. The limit of one’s representations of the world is not a representation of the limits of one’s world. That two creatures represent different aspects of the world they live in does not imply that they represent different worlds or that they represent the world as being different ways. What I know about the world is very strictly limited. It does not follow that I represent the world as having strict limits. In the vocabulary of OCCI, to suppose that this follows would be to ‘import completeness’ (OCCI, ch. 8). A second worry about how the content of a representation could rest on its biological utility may be eased by understanding clearly the breadth of the notion of biological utility that is intended. The claim is that all human purposes may be seen to be biological purposes when examined in the right light, hence that all human utility is biological utility (VM, ch. 1). Having biological utility is not at all the same as merely serving a purpose for which the organism’s genes were selected. Besides biological purposes that rest on genic selection, such as the purposes of the stomach, heart, liver, kidneys, the purposes of their various activities, and so forth, instrumental conditioning generates a separate level of purposive activity, and explicitly represented goals implemented through practical reasoning yet another. The relations among these various levels of purpose is complicated. I will say something about this. One of the things that has been evolved by natural selection is evolvability itself, for example, in the evolution of homeo box genes and of sexual reproduction. Relevant to our current concerns is the evolution of behavioral systems that learn by trial and error or, as Dennett says, by generate and test. The analogy between genic selection and operant conditioning or instrumental learning is remarkably close (Hull et al. 2001). Processes of practical reasoning also proceed by trial and error, as the thinker mentally experiments, attempting different routes from his current situation to another envisioned situation until a path is found that will connect these two (VM, ch. 17). Different kinds of purposes emerge on each of these levels. These purposes sometimes conflict with one another and also with the original purposes for which the genes were selected, yet all are biological purposes, originally derived from natural selection of genetic materials (VM, chs. 1 and 2). Consider first instrumental learning. Appropriate genes are responsible for the development of the systems that learn by operant conditioning. And
Useless Content
103
appropriate genes are responsible for determining what the original forms of reinforcement will be for a particular animal species. But forms of reinforcement, such as alleviation of hungry or of thirsty feelings, or the presence of a sweet taste in the mouth, are not themselves means toward fulfillment of any purposes of the genes. Sweet tastes, for example, although correlated during the history of the species with the presence of needed nutrition, are not, in themselves, involved in any direct causal process resulting in increased nutrition. Behaviors reinforced by sweet tastes (M&Ms) are selected for producing sweet tastes, so they have as a natural purpose to bring in sweet tastes. Also, true, the disposition to be reinforced by sweet tastes was selected because sweet tastes typically indicated nutritive value. But behaviors can succeed in bringing in sweet tastes without succeeding in increasing nutrition. That is what saccharine is sold for. Although the purpose of procuring sweet tastes on the genetic level is gaining calories, the psychological purpose of behaviors conditioned by sweets is merely to procure sweet tastes. So behaviors can succeed in fulfilling their natural biological purposes of bringing in sweet tastes without fulfilling any more basic functions for which genes were selected. In the broad sense meant here these behaviors have ‘biological utility’, although in a narrower sense—on a shallower level—of course they do not. We generally distinguish these levels of purpose by calling one ‘psychological’ and the other ‘biological’, but looked at carefully, the level of psychological purposes produced by reinforcement is just another layer of biological purpose, though these are not so direct as the purposes, say, of the heart’s beating or the stomach’s juices flowing. It is in the broad sense in which psychological purposes are included among biological purposes that the content of a representation is said to rest on its biological utility. Processes of practical reasoning also proceed by trial and error, generating a third level of natural selection, hence of natural purposes. The natural purpose of behaviors selected through practical reasoning is to reach whatever goals are represented at the start of the reasoning as the ends for which means are to be selected. The final implementing intentions that immediately generate behaviors have been selected because these representations have led, in the inner world of the thinker, to representations of the original goal as fulfilled. If the practical reasoning operates properly, implementing these selected intentions will produce implementation of the original goal. The interesting and challenging question concerns the origin of these represented goals, in particular, their relation to more basic biological purposes resting more directly on genetic selection. What determines which conscious goals we pursue? Perhaps most are derivative from prior goals that we have. We aim for them because we believe they will lead to ends already established as goals. But what mechanism selected the original goals from which these goals were derived?
104
Ruth Millikan
Clearly our original conscious goals are not the same as whatever our genes aim for. Babies do not come into the world fervently desiring to live a long life and produce lots more babies. Nor do they come into the world desiring to obtain just those goodies that will or would serve to reinforce their behaviors through conditioning. This is true in the first instance simply because babies don’t come into the world thinking about anything, hence explicitly desiring anything. They have to develop concepts before they can explicitly desire things. But let us suppose that a child has developed adequate concepts of all the things that do or will happen to reinforce her behaviors, from sweet tastes to sexual pleasure to smiles. She is able to think about each of these things. Merely developing concepts of these things is not the same as conceiving these things AS things that reinforce or AS things wanted, however. Just as there may be a large gap between what reinforces an animal’s behavior and what has deeper biological utility for the animal—between sweet tastes and nutrition or between sexual pleasure and having babies—there may be a large gap between what actually reinforces behavior and what one wants or what one understands to be a cause of reinforcement. Often we may know when we are attracted or repulsed without knowing exactly why we are attracted or repulsed, or we may know that we are happy or sad without knowing why. Knowing exactly what it is that we want is by no means automatic. Likely we only find this out by experience, indeed, by something analogous to hypothesis formation and testing. ‘I don’t know why it is but shopping for Christmas always makes me anxious’, we say, or ‘There is something about electric trains that just fascinates me.’ We may wrongly suppose that having a lot of money is what is required to make us happy, or being allowed to sleep in every day without having morning commitments. Indeed, just what it is that makes people happy has proved to be an extremely challenging question for clinical and social psychology. Moreover, knowing what it is that attracts us or repels us need not lead to a reasoned desire for or against that thing. Often we have other conflicting interests to consider. Still, the mechanisms through which goals are projected by conscious desire and reason are undoubtedly mechanisms that our genes have been selected for engendering.² These mechanisms are still with us because they have sometimes—often enough—produced behaviors that benefited human genes in the past. Aims and goals that are products of these mechanisms are also biological purposes in the broad sense intended, even though it may happen very often that they do not lead to fulfillment of any more direct purposes ² The claim, made by Fodor and others, that our human cognitive mechanisms may actually have arrived of a piece without benefit of natural selection is discussed in VM, ch. 1 n. 2 and ch. 2 n. 5.
Useless Content
105
of the genes, perhaps not even producing behaviors that are psychologically rewarding. Fulfillment of these aims and goals still has biological utility in the broad sense that is meant. Now one of the most important jobs that beliefs are designed to do, surely, is to combine with other beliefs to form new true beliefs. This kind of function contributes to biological utility whenever any of the beliefs formed in this way turns out to have biological utility. Even the most highly theoretical of beliefs are not excluded from having biological utility, then, so long as they participate in chains of reasoning that eventually bear practical fruit. Still, surely many beliefs that we humans have do not ever help to serve any of our goals, so the question does arise how these beliefs can have content on the teleologist’s view. First we should be clear that the teleologist’s position is not that each individual belief must help to serve a biological function. Rather, beliefs must fall within a general system of representation where the semantic rules for the system are determined by reference to the way its consumers, its interpreters, are designed to use these representations. But in order to have been designed to make and use representations in the system in a certain way, the producers and consumers or their ancestors must have been using representations from the same system productively in the past. Otherwise these mechanisms would not be representation producers and consumers. They would not have been selected, through evolutionary history or through prior learning, for their capacity to coordinate through the use of representations. But there is another alternative. There may be ways that the producers have been designed to learn or to be tuned to produce representations that are coordinated with ways the consumers are designed to be tuned to use them, so that the producer’s rules and the consumer’s use dispositions are somehow tailored in advance to match one another. If the former—if the producer and consumer (or their ancestors) have a history of being coordinated through the use of a particular semantic system, we need to know how to generalize from past successful coordinations to determine a unique general semantic rule that determines content in those cases where no actual coordination occurs. If the latter—if producer and consumer have been designed to learn to respond in a coordinated way without actually practicing together, we need to understand how this marvel is accomplished. I have argued that the semantic rules governing representations that humans use in perception and thought are sometimes derived one of these ways and sometimes the other. Let me discuss these alternatives in turn. Suppose, first, that the rules are determined by a series of past successful coordinations between the producers and consumers of the representations.
106
Ruth Millikan
These mechanisms were selected for as a result of these coordinations, selected for either by learning systems or through genetic evolution. The semantic rules for these representations concern mapping relations between them and world affairs they have corresponded to such that two conditions are fulfilled. First, there must be a single kind of causal mechanism or process that operated in these historical cases, reference to which explains how the producers managed on each of these past occasions to produce representations that corresponded to world affairs by these mapping rules. It cannot be the proper function of the producers to make representations that vary in a definite way parallel to affairs obtaining in the world unless there is a systematic way they have been doing this under certain conditions. They must, for example, be sensitive to ‘local natural signs’ of the existence of these affairs (VM, chs. 3–6). Second, there must be uniform explanatory causal mechanisms or processes that have accounted for the fact that the consumers’ responses to the representations have contributed in the past to proper performance of their functions and these explanatory mechanisms must depend on the fact that the representations have varied according to world affairs by some particular semantic rule of correspondence. Representations that are not actually used have content because their semantics is ‘compositional’ in a very broad sense (Millikan 1995, 2004). Complete representations always fall within some system of representations such that variations in the complete representations correspond systematically to variations in what is represented.³ The content of representations that are not used or not used productively is determined by the way other representations in the same system have been successfully used in the past. Still, we might question whether there is only one way to generalize from a series of past successful uses so as to determine a general rule of correspondence between representations and candidates represented. Recall Kripke’s (1982) worries that no matter how many past examples of correct additions one begins with, these examples will never determine the correct way to go on to new examples using the ‘plus rule’. But the two cases are very different. The rule we are seeking is one that not only coincides with past successful cases of coordination but conformity with which, in each case, causally generated the coordination. In each case, had the representation been different the consumer would have reacted differently, and this reaction would not have served its purpose, or would not have served it owing (in part) to the fact or condition that was represented. Kripke’s difficulty was that he could not appeal to dispositions that he has to react to the plus ³ For a discussion of this sense of ‘compositional’ and its application both to intentional representations and to natural signs, see VM, chs. 3–4.
Useless Content
107
sign, because under certain conditions or at certain times his disposition may be to make mistakes. But we can appeal to dispositions. We can appeal to the dispositions that past representation consumers have had at those particular times and in those particular circumstances in which their reactions to representations served their psychological or biological functions. The particular dispositions to which we appeal are relational dispositions. How the properly functioning consumer reacts is determined as a function of its representational input. Its disposition is to react in a way that bears a certain relation to this input. Different inputs would have produced different outputs. Kripke’s deep difficulty was that he could find no way of determining objectively when the attempt at addition gives a correct answer. Indeed, although he doesn’t mention this, he has no way of determining which past additions were correct, hence were suitable examples from which to generalize either. But we do have a way of determining objectively when representation production and consumption have achieved a coordination in the past. The skeptical question can be pressed further, however. Representations that are not in fact used to serve biological utilities are one thing. But could a representation have a semantic value that was so extreme that, supposing it to be true, it would be physically impossible for it’s consumers to make use of it. For example, Caroline Price (2000) asks about a bee dance that indicates nectar in a position that is far too distant for any bee to fly to. The bees do not, presumably, use the information gained from the dances of fellow bees in theoretical inference. The semantics of bee dances derives solely and directly from the use of these dances in directing bees to fly to sources of nectar. Now in point of fact, bees indicate greater distance by slowing down their tail waggles and lengthening the intervals between the bursts of sound they make while dancing. Besides the fact that no bee could return from a distance that was too far to fly to in order to perform a dance indicating this distance, it is also unlikely that bees have the capacity either to dance properly going so slowly or to perceive accurately the speed of waggling and of sound bursts below a certain frequency. For what would they use these capacities? It is unlikely that a dance that, by logical extension of beemese rules, would tell of nectar much too far to fly to could be either danced or, more central, recognized by fellow bees. No ancestor bees have had dispositions to make use of such dances. Such bee dances, then, are meaningless in beemese. After all, bee dances are not designed for human interpreters who carry stopwatches to read them, but for bees. Turning to humans, my claim is that the only human mental representations whose semantic values are determined solely by generalization from past successful uses are those perceptual representations that directly guide action, such as
108
Ruth Millikan
are produced, for example, through the dorsal visual channels (VM, ch. 14). And there are surely limitations on the extremes that humans can perceptually represent or interpret directly for action, just as there are on the dances that bees can dance and interpret. Pressing questions about the representation of affairs with which humans could not possibly interact arise only when we are pretty certain that we do in fact represent these things. For example, Peacocke (1992) has asked about representing propositions that concern things outside our light cone. No interaction with such things is physically possible even in principle. What makes this question legitimate is that we are pretty sure that we can formulate questions about what is outside our light cone, and even though we are not able to answer these questions, it is possible that someone might have an unfounded belief about some such matter, and that the belief would in fact have a truth value. Again, my suggestion on this matter will involve an appeal to compositionality.⁴ But first, I need to locate the problem within the framework of the theory of empirical concepts with which I have been working (LTOBC, OCCI ). Representations produced and consumed by systems that employ empirical concepts are examples of a kind of representation that producers and consumers can learn how or be tuned to use cooperatively without actually practicing together. Their production and use dispositions can be tailored in advance to fit one another. It will take me a few moments to explain how this can be. I will not defend my position on this matter at any length, however. Discussion and defense can be found in VM chapter 19, LTOBC chapters 15–19, and OCCI chapter 7. The implied realist ontology is explained and defended in OCCI chapter 2 and in LTOBC chapters 14–17. Empirical concepts, in basic cases, consist in part of abilities to recognize the objects, kinds, properties, and relations (I will just say entities) of which they are the concepts when these entities are encountered in experience.⁵ Adequate abilities of this sort must be extremely sophisticated abilities. This is because the same entity encountered under different conditions, at different angles and distances from the perceiver or registered through different media of information transmission, will impact the perceiver through a ⁴ So does Caroline Price’s solution (2000). I agree with Price’s analysis as far as it goes. But the question how we acquire concepts of objective kinds, objects, and properties in the first place so as to recombine them compositionally requires explanation. ⁵ In OCCI, ch. 6, and in VM, ch. 9, I argue that gathering information by believing what another human says is in all relevant ways exactly like gathering information through direct perception. The result is that people can have basic concepts of things that they don’t yet know how to recognize ‘in the flesh’.
Useless Content
109
wide variety of different proximal stimulations. The most formidable of all the tasks that the perceptual–cognitive systems must perform is to learn to recognize what is objectively the same as the same when it is encountered again under any of numerous varying circumstances. Lacking this ability, an inner representation-producing mechanism would either fail to recognize the presence of its represented objects nearly all of the time, or be ridiculously redundant, producing many separate and different representations of the same, which the consuming systems would not then recognize as being representations of the same. This would result, first, in an inability to accumulate information about an entity over time in a format allowing recognition that the information all concerned the same entity. Second, it would result in an inability to apply information previously gathered about entities when interacting with them again, for these entities would not be recognized as the same ones the information concerned. Having an adequate ability to recognize when one is encountering information about the same thing again is an essential part of what it is to have an adequate empirical concept of that thing. Assuming that we are not born with conceptual abilities of the required sort, how do our perceptual–cognitive systems learn to reidentify objects, kinds, and properties correctly? What criteria do they use to distinguish successful from unsuccessful attempts at reidentification so as to fashion a nonredundant (and nonequivocal) inner representational system capable of representing these entities and their configurations in a uniform way? A simple reply is that correct reidentification of these entities might be tested during practical activity. What counts as the same is what yields the same results when reacted to or treated in the same way. It is the same, for example, if you can chase and catch it using the same technique, and eat it in the same way, and if it tastes the same way, and nourishes you in the same way. It is the same if in responding to it the same way you get the same proximal results. This answer assumes that you can recognize when your responses or treatment are objectively the same, but although attempting to stabilize two variables at once makes things more difficult, there is no reason in principle why it can’t be done. No doubt this is the way in which natural selection trains the innate perceptual–cognitive mechanisms of a species and how perceptual–cognitive learning takes place both in lower species and, at a certain level, in humans. Psychologists call this ‘generalization and discrimination’, though it is not always recognized that learning what is objectively the same response, with regard to the practical purposes at hand, is part of what is being learned. But this simple reply leads us directly back to the very first problem raised in this chapter. If the only criterion for having successfully learned
110
Ruth Millikan
how to reidentify an object, kind, or property is successful use of that ability in practical activity, then one could only learn to represent those aspects of the world one needed to represent in order to guide one’s behaviors. Elsewhere I have argued that it is likely that the representational abilities of nonhuman animals are limited in exactly that way (VM, chs. 18–19). But there seems reason to believe that human cognitive capacities are not so limited. Is there a way that we humans might learn to identify objective entities with which we have had, as yet, no practical dealings, so as to accumulate information involving them within a non-redundant representational system? If we could develop abilities accurately to reidentify various entities in the world prior to practical uses, we could use these abilities later, much more easily learning to use these entities in a variety of practical ways and to apply prior knowledge of them during practical activities. A general ability to discover where objective distal samenesses lie through the intervening diversity of proximal stimulations would be selected for because it served further biological functions. It would not then be necessary that every conceptual ability developed in this way should actually find practical applications. The belief representations in which a concept that had been developed in this way came to figure would have determinate truth conditions without that. The suggestion on conceptual development that I have made (LTOBC, chs. 18 and 19; OCCI, ch. 7; VM, ch. 19) might be summed up in old-fashioned terminology by saying that coherence in human belief serves as a test of correspondence. It serves as a test of the correspondence, not primarily of individual beliefs to their truth makers but, more fundamentally, as a test of consistent correspondence of the same representational elements to the same elements in objective world affairs. Coherence serves as a test of our capacities correctly to reidentify distal entities through diverse proximal manifestations. Consistent judgment is a natural sign of the representation producer’s capacity to represent the same as the same. Coherence is a psychological goal that reinforces ways that the cognitive systems are producing representations. Like a sweet taste in the mouth, coherence is a sign that has correlated with achievement of a deeper biological purpose, the production of a consistent and accurate representational system, itself only a means to further biological ends. Now I will unpack this a bit. Consistency within a representational system can be used as a test only if inconsistency is possible within that system and if inconsistency shows up on the surface of the representational system. Consider bee dances, for example. Bee dances cannot conflict with one another. If two bees dance different dances when returning to the hive, it is always possible that both
Useless Content
111
dances are correct. There really is nectar both of those places. Nor do the bees have any way of representing where there isn’t any nectar. Bee dances are not sensitive to a negation transformation. A subject–predicate sentence and its negation, on the other hand, are explicitly incompatible, incompatible right on the surface. Similarly, humans can think negative thoughts, and these thoughts contrast explicitly with possible positive thoughts. Whether the way human thoughts are coded resembles the way language is coded in any other way, certainly our thoughts are sensitive to a negation transformation. Whenever we have opportunity to gather the same information in two different ways, through two different natural information channels yielding different proximal stimulations, we have a chance to gain evidence about whether our various methods of attempting to identify the entities represented in the subjects and the predicates of these judgments are each converging to focus on some single objectively same thing. Consistent agreement in judgments is evidence that these various methods of making the same judgment are all converging on the same distal affair, bouncing off the same target, as it were. If the same belief is confirmed by sight, by touch, by hearing, by testimony, by various inductions one has made, and is confirmed also by theoretical considerations (inference is a method of identification too), this is sterling evidence for the univocity of the various methods one has used to identify each of the various facets of the world that the belief concerns. Thus the same object that is square as perceived from here should be square as perceived from there and square by feel and square by checking with a carpenter’s square and square by measuring its diagonals and square by hearing from another person that it is square.⁶ Similarly, if a person is tall and good at mathematics as recognized today, that same person should prove tall and good at mathematics when reidentified tomorrow. Both one’s general methods of reidentifying individuals and one’s methods of recognizing height and mathematical skill are corroborated in this way as methods of reidentifying objective selfsames. That the same chemical substance is found to melt at the same temperature by checking with an alcohol thermometer, a mercury thermometer, a resistance thermometer, a gas thermometer, a bimetal expansion thermometer, and a thermal thermometer is evidence both that one is able to recognize the same chemical substance again and that there is indeed some real quantity (unlike caloric pressure) that is being measured by all of these instruments. If a multitude of different operational definitions are found to correlate with one another exactly, then they can be assumed all to measure the same thing. Moreover, and
⁶ See n. 5 above.
112
Ruth Millikan
much more fundamentally, it can be assumed that there is something real that each one measures.⁷ Notice, however, that agreement in judgments can be a test of correspondence only in so far as it is possible for one’s judgments to disagree with each other. And explicit disagreement in judgments is possible only in so far as negation in judgments is possible.⁸ In its basic form, negation is a semantic operation on the logical predicate of a sentence (Millikan 1984, ch. 14; Horn 1989, ch. 6).⁹ Logicians call this ‘internal negation’. For example, the normal reading, say, of the classic negative sentence ‘The king of France is not bald’ makes it equivalent to ‘The king of France is non-bald’, so that the negative as well as the affirmative presupposes the existence of a king of France. More obviously, ‘John is not tall’ is equivalent to ‘John is non-tall’ and ‘John does not know French’ is equivalent to ‘John is ignorant of French’, and so forth. There are also secondary uses of ‘not’ to reject a sentence on non-truth-conditional grounds, as in ‘The slithy toves did not gyre and gimble in the wabe’ or ‘The square root of two is not blue’ or ‘You didn’t see two mongeese, dear, you saw two mongooses’ or ‘The king of France is not bald, dear; France doesn’t have a king’. But the fundamental use of the negative is not to prohibit assertion of a sentence, but to make a positive, though indefinite, statement to the contrary. The standard negative sentence actually says something positive about its subject, namely, that it is characterized by some contrary or other of the predicate of the sentence. If John is not tall it is because he is short or of medium height. If John does not understand French it is because French sentences either leave his mind blank or produce in it thoughts different than for a Frenchman. This point about negation assumes importance when we turn to epistemology and consider how evidence is gathered for a negative judgment. Begin with the obvious: The absence of a representation of a certain fact is not ⁷ ‘If we see on a road one house nearer to us than another, our other senses will bear out the view that it is nearer; for example, it will be reached sooner if we walk along the road. Other people will agree that the house which looks nearer to us is nearer; the ordinance map will take the same view . . .’ (Russell 1912: 31). This is Russell’s argument that a real spatial relation between the houses corresponds to the nearer than relation of which we are aware between certain sense data. ⁸ The remainder of this chapter follows paragraphs in VM, ch. 19, very closely, with the kind permission of the MIT Press. ⁹ External negation, which operates on the sentence as a whole, is called ‘immunizing’ negation in (Millikan 1984). Horn (1989) gives a parallel analysis calling it ‘metalinguistic’ negation as opposed to ‘descriptive’ negation. The claim is that immunizing or metalinguistic negation is not a semantic operator.
Useless Content
113
equivalent to the presence of a representation showing the negative of that fact. Absence of a belief is not a negative belief. Similarly, absence of perceptual evidence leading one to form or confirm a belief is not perceptual evidence that leads one to form or confirm the negative of that belief. If you look again from another angle at what you took to be a square object but fail this time to see that the object is square, or reach out with your hand but fail to feel that the object is square, this by itself is not evidence against the object’s being square. Perhaps the trouble is that you can no longer see the object at all, or although you see it, you can’t make out its shape against the light. Perhaps the trouble is that the object is not where it appeared to be so that reaching out your hand to feel it you encounter nothing at all. To gather evidence against the object being square, you must first see or feel the object, and then you must see or feel that its shape is some contrary of square, perhaps round or oblong. Gathering evidence for the negative of a proposition is always gathering positive evidence, evidence for some contrary of that proposition. It follows that an ability to recognize contraries of a property through the variety of their diverse manifestations and to recognize them as being contraries, as being incompatible, is required in order to test one’s abilities to identify subjects of judgment, and vice versa. The result is not an epistemological regress or circle. But both of these abilities do have to be in place before stability of judgment over time, over various perspectives, and through diverse media of information transmission can emerge with regard to any particular kind of subject matter. Both these abilities have to be in place before steady evidence can accumulate that successful identifications are being made outside the context of practical activity. True, without doubt the first leg up is still practical. Many of the things recognized as the same again for purposes of practical use turn out to correspond to pretty good subjects or predicates for theoretical judgment as well. (The second leg up is public language—OCCI, ch. 6; VM, chs. 9 and 19.) But the end result is the perfection of concepts by a method that does not rely on practical uses as its criterion of success. It does not rely on practical uses as its criterion of success, and yet this criterion has been selected for because its use has yielded practical results in the past. Beliefs about things outside our light cone, then, are like desires for candies that contain saccharine. Their content is determined biologically, but the biological functions—logically coherent thought, obtaining sweet tastes—that determine their content are not on the same level as any direct purposes of the genes. Nor, in the case of coherent thought, is the content determined by its use in having helped to fulfill past psychological goals, except in so far as avoiding contradiction might be considered a psychological goal.
114
Ruth Millikan REFERENCES
H, L. R. (1989), A Natural History of Negation (Chicago: University of Chicago Press). H, D. L., L, R. E., and G, S. S. (2001), ‘A General Account of Selection: Biology, Immunology and Behavior’, Behavioral and Brain Sciences, 24/2: 511–69. K, S. A. (1982), Wittgenstein on Rules and Private Languages (Cambridge, Mass.: Harvard University Press). M, R. G. (1984), Language, Thought, and Other Biological Categories (Cambridge, Mass.: MIT Press). (1995), ‘A Bet with Peacocke’, in C. Macdonald and G. Macdonald (eds.), Philosophy of Psychology: Debates on Psychological Explanation (Oxford: Blackwell), 285–92. (2000), On Clear and Confused Ideas (Cambridge: Cambridge University Press). (2004), Varieties of Meaning (Cambridge, Mass.: MIT Press). P, C. (1992), A Study of Concepts (Cambridge, Mass.: MIT Press). P, C. (2000), Functions in Mind (Oxford: Oxford University Press). R, B. (1912), The Problems of Philosophy (New York: Holt).
6 On Thinking of Kinds: A Neuroscientific Perspective Dan Ryder 1. KINDS AND PSYCHOSEMANTICS — NOT A MATCH MADE IN HEAVEN Reductive, naturalistic psychosemantic theories do not have a good track record when it comes to accommodating the representation of kinds. In this chapter, I will suggest a particular teleosemantic strategy to solve this problem, grounded in the neurocomputational details of the cerebral cortex. It is a strategy with some parallels to one that Ruth Millikan has suggested, but to which insufficient attention has been paid. This lack of attention is perhaps due to a lack of appreciation for the severity of the problem, so I begin by explaining why the situation is indeed a dire one. One of the main tasks for a naturalistic psychosemantic theory is to describe how the extensions of mental representations are determined. (Such a theory may also attempt to account for other aspects of the ‘meaning’ of mental representations, if there are any.) Some mental representations, e.g. the concept of water, denote kinds (I shall be assuming this is non-negotiable). How is this possible? Unfortunately, I haven’t the space to canvass all the theories out there and show that each one fails to accommodate the representation of kinds, but I will point out the major types of problems that arise for the kinds of theories that, judging by the literature, are considered viable contenders.¹ In general, the theories either ¹ For example, I ignore pure internal conceptual role theories, since they fail to explain how a concept can even have an extension. No psychosemantics has ever been given for ‘long-armed’ role theories (e.g. Harman 1987), and the early informational theories (Stampe 1977; Dretske 1981) have been shown to suffer fatal problems with disjunctivitis and the like (Fodor 1990).
116
Dan Ryder
attempt and fail to account for the representation of kinds, or they fall back on something like an intention to refer to a kind—not exactly the most auspicious move for a reductive theory. There are a number of problems that prevent non-teleosemantic theories from explaining how it is possible to represent kinds. A concept of a kind K must exclude from its extension things that superficially resemble K but whose underlying nature is different, e.g. a concept of water excludes XYZ (Putnam 1975). Any psychosemantic theory that depends exclusively upon the intrinsic properties (including dispositions) of the representer to determine extension will thus fail to provide for the unequivocal representation of kinds, by reason of the familiar twin cases. This problem infects theories based on isomorphism (Cummins 1996), information (Usher 2001), and nomological covariance, including (as Aydede 1997 demonstrates) Fodor’s asymmetric dependence theory in its most recent guise (Fodor 1994). Some information-based and nomological covariance theories avoid the twin problem by adding further conditions on extension determination. Whether or not these moves help with twins, they do not help with kinds. For example, Prinz (2002) requires a representation’s content to be its ‘incipient cause’, the thing that explains the concept’s acquisition. Such modified informational and nomological covariance theories inevitably fall victim to a second problem for such theories, the problem of epistemically ideal conditions. On a nomological covariance theory, surely it is the category that exhibits the best nomological covariance with a representation that is its content. (What else could it be?) Since the covariation of a representation with its content will always be better in ideal epistemic circumstances, a nomological covariance theory will always incorrectly dictate that the content of a representation of kind K is rather K-in-epistemically-ideal-circumstances. Getting the right content cannot be achieved by ruling out those factors that are ‘merely epistemic’, e.g. by adding an ‘ideal conditions’ clause (Stampe 1977; Stalnaker 1984), since which factors are merely epistemic will depend upon which factors are semantic (McLaughlin 1987). If the symbol really means K-in-good-light, being in good light is not a merely epistemic factor. (And if the symbol really means K-in-weak-light, good light will not be ideal.) The only remaining way to rule out the epistemic factors from the content of a representation on a nomological covariance theory would be by ad hoc restriction of the class of candidate contents to those that fit with how we in fact carve up the world. Because we find these categories intuitive, this move can easily go unnoticed. But this is to ignore the fact that a psychosemantics will be part of an explanation for why those categories are the intuitive ones, i.e. why our representations have the contents they do. The restrictions on content must
On Thinking of Kinds
117
be predicted by a complete psychosemantics (perhaps supplemented by an epistemology); they cannot be tacked on in order to fix a flawed theory. An appropriate restriction might appear to emerge naturally from nomological covariance theories, since on such theories, the classes of representable contents are presumably marked by having ‘nomic’ rather than ‘non-nomic’ properties in common. But making that distinction principled while ruling out the problem contents and ruling in the legitimate contents is a tall order (not to mention the collateral problem that arises for representing individuals). For instance, if crumpled, shirtness, and even crumpledshirtness are nomic properties (Fodor 1991), why on earth isn’t gold-ingood-light? (Thus, in Dennett’s Philosophical Lexicon: ‘jerry-mander, v. To tailor one’s metaphysics so as to produce results convenient for the philosophy of mind’; Dennett 1987.) On the other hand, if we were to restrict representable contents to scientific kinds, the content of much of our mental life would go unaccounted for. One might expect teleosemantic theories to do better, but in fact they do not. Consider indicator teleosemantics, for instance (Dretske 1986, 1988; Matthen 1988). If an indicator I is selected for indicating kind K, that means I’s indicating kind K has been causally relevant to the presence of I in this (species of ) organism, or causally relevant to I’s recruitment to perform some biologically relevant task. The problem that arises is a version of Fodor’s (1990) indeterminacy complaint (which is itself a version of the disjunction problem): why suppose that I has been selected to indicate K as opposed to a disjunction of local signs of K? If K has been causally relevant to I’s selection, then the local signs of K that mediate I’s ability to indicate K have also been causally relevant to I’s selection. There is no reason to pick K rather than local signs of K as I’s content, and, again, an ad hoc restriction isn’t kosher. (In correspondence, Dretske has agreed that, for these reasons, he faces difficulties in accounting for the representation of kinds and indeed anything distal.) A similar problem infects the basic version of Millikan’s ‘mapping’ theory, which involves the selectional recruitment of ‘representation producers’ to produce representations that map onto the environment according to a certain rule. This selectional recruitment operates via these representations’ consumers; it is the mapping of the representations onto their contents that has, historically, explained the proper performance of one or more of the consumers’ functions—in Millikan’s terminology, this mapping is a Normal condition on proper performance of the consumers’ functions. She also requires the mapping be something that the producer can bring about or effect, e.g. via a detection mechanism. Now, whenever Millikan is tempted to suppose that it is a kind that a representation is supposed to map onto, we may counter as follows: that
118
Dan Ryder
kind will have a cluster of typical properties, some subset of which will be those that explain its detection (or however the mapping is brought about) and the proper performance of the consumer’s relevant functions. Why not say that the representation denotes either the subset or disjunction of selectionally explanatory properties, rather than the kind itself? A kind is not identical to a subset or disjunction of its properties. This seems to raise the spectre of twin problems for Millikan—the selectionally explanatory properties could be clarity, liquidity, and potability, properties that H2 O and XYZ share. (We shall see later that some additional resources allow Millikan to deal with this problem, if not for the ubiquitous fly-snapping frog, then at least for us.) As far as I know, all other reductive theories face a relative of one of the problems above. We must face the possibility that a reductive naturalistic psychosemantics cannot explain the representation of kinds directly. However, there are also a number of non-reductive strategies for explaining the possibility of representing kinds, strategies that appeal in some way to the representation wielder’s intentions. Such an appeal might be used in a reductionist project if the representations (normally concepts) mobilized in these intentions do not themselves include any representations of kinds. I think this hope is forlorn, but the reductionist has a lesson to learn from the attempt. What sort of intention would do the trick? Well, an intention to pick out a kind, of course. For instance, the representation of specific kinds could be accounted for by mental description. Just as one might say that unicorns are picked out in thought descriptionally (‘a horse with a horn’—otherwise a puzzling case, especially for the naturalist, since unicorns do not exist), perhaps one could pick out a specific kind with ‘a φ that is a kind’, where ‘φ’ denotes some complex property whose representation can be accommodated by your favourite reductive theory. It would remain to give some reductive account of the concept of kindhood; however, if concepts of particular kinds are difficult to account for with current reductive theories of intentionality, then the concept of kindhood seems even more difficult. Further, how would such a concept be acquired without prior concepts of specific kinds as examples? Perhaps a further non-reductive move could be made, filling out the concept of kindhood descriptionally. A well-developed account of kindhood exists that could be put to use in this way. According to the ‘unified property cluster’ account (Boyd 1991; Kornblith 1993; Millikan 1999, 2000), a natural kind is characterized by a set of correlated properties, where some further principle explains why they are correlated, and thus why reliable inductive generalizations can be made over them. For example, water is a substance with multiple correlated properties like liquidity in
On Thinking of Kinds
119
certain conditions, clarity, the ability to dissolve certain other substances, etc., where these ‘surface properties’ are explained by water’s nature or hidden essence, namely its molecular structure. This pattern of regularity organized around a ‘source of correlation’ (as I call it) is not restricted to chemical natural kinds. In the case of biological kinds, these correlations are due, not to an underlying chemical structure, but to the common history shared by their members. Millikan (1998, 1999, 2000; see also Bloom 2000) extends the unified property cluster account beyond natural kinds. Non-natural (but real) kinds also have multiple correlated properties unified by some explanatory reason. Artifacts, for instance, will often have correlated properties because they serve some specific function (e.g. screwdrivers), because they originate from the same plan (e.g. Apple’s iMac), or because they have been copied for sociological reasons (e.g. a coat of arms and its variants). Even kinds of events and processes are sources of correlation, for instance Hallowe’en, biological growth, and atomic fusion. (Millikan also points out that individuals fall into the same pattern, although we will not be concerned with them here.) Now the non-reductionist account of a kind concept becomes ‘a φ that is a member of a class sharing a syndrome of properties with a common underlying explanation’ or something like that, mentioning in φ some specific features of the syndrome (and perhaps restricting it to ‘around here’). This seems to be the idea behind the theory of ‘psychological essentialism’ (Keil 1989; Medin and Ortony 1989), and at least one philosopher has appealed to it in order to explain the possibility of representing kinds (Prinz 2002). Again, though, this seems cold comfort to the reductionist, for at least two reasons. First, the sophistication of this conception is such that the potential for reduction is not getting any clearer (not to mention the fact that it may put kind concepts beyond the reach of a sizeable proportion of the population, and I don’t just mean children). Second, it seems to betray an almost classical empiricist faith that the concepts featured in a kind’s syndrome can ultimately be understood (presumably through many steps of analysis), not as kind concepts themselves, but as some other sort of concept. (Perhaps as representations of properties transduced by perceptual systems? This is the extreme view that Prinz’s theory entails, although he does not assert it.) This seems unlikely, especially for biological kinds. For example, maleness is a biological kind characterized by a biological syndrome including, for instance, produces sperm, which is its own biological kind, etc. Doubtless there is much more to be said about the descriptionist strategy. For instance, the ‘division of linguistic labour’ (Putnam 1975) is available to the descriptionist, whereby people can supposedly think of kinds through
120
Dan Ryder
language, by deference to experts on the referents of public language terms. (But how do the experts think of kinds? And does this deference not involve concepts of kinds, e.g. the kind word ?) Or the description may be ‘naturalized’ by giving it a causal role reading (but how can causal roles determine extensions?) I will not pursue this line of inquiry any further here; for one thing, there are many independent problems that accrue to descriptional accounts of concepts generally (Fodor 1998; Millikan 2000). A direct reductionist approach would clearly be welcome. Teleosemantics is the reductionist approach I advocate, and the teleosemanticist can learn from the non-reductive appeal to further intentions. The standard teleosemantic strategy is to take what might appear to be intentions, an agent’s purposes, and reconstrue them as functions, or biological purposes. This, I suggest, is what we should do with psychological essentialism, eliminating the need for a complex, reduction-resistant conception of kindhood on the part of the agent. What we need is for the mind’s representational system, or a part of it, to have the function of indicating kinds (Dretskean formulation), or that it be supposed to map onto kinds (Millikanian formulation), or something similar. That is, kindhood itself, as characterized by the unified property cluster account, needs to be selectionally relevant to the representational system. This, in fact, is a way of understanding what Millikan has attempted over the course of a number of publications (1984, pt. ; 1998, 2000), except that she claims selectional relevance not for kindhood, but something broader that includes kindhood as a subtype (she calls it substancehood). The account I present below was developed independently, but has much in common with Millikan’s. What is particularly interesting is that my story comes from the neuroscience, whereas hers is derived from abstract psychological considerations. This convergence is surely a sign of truth! The main idea underlying my proposal is that the brain’s predictive network was selected for because of the way it interacts specifically with kinds (actually, the more general class of ‘sources of correlation’). Its special predictive capacities are dependent upon its interacting with kinds qua kinds, so kindhood, as characterized in the unified property cluster account, was selectionally relevant to the design of our representational system. One function of our representational system, therefore, must be characterized with reference to kindhood, and this function ultimately underlies our representation of kinds. But getting to this conclusion will require considerable stage-setting. First I will present a particular teleosemantic framework, and apply it to representation acquisition (for concepts of kinds are acquired ); then I will outline the neuroscientific details that yield an answer to our main question.
On Thinking of Kinds
121
2. THE TELEOSEMANTICS OF MODELS Many representations may be understood in teleosemantic terms. Although a tire gauge carries information about pressure, temperature, and volume, it represents only pressure because that is what it is supposed to carry information about (Dretske 1986, 1988). A map of Oconomowoc, Wisconsin, is two-dimensionally similar to both Oconomowoc and Blow-me-down, Newfoundland, but it only represents Oconomowoc because that is the location to which it is supposed to be two-dimensionally similar. (And a sheet of paper, physically indistinguishable from a map of Oconomowoc, that is not a map but rather a section of wallpaper is not supposed to be twodimensionally similar to anything, and as a result, does not represent anything.) Bar graphs, pictures, and many other representations can be treated analogously. For all of these representations, there is some type of relation that they tend to enter into with the things they represent, and which thing they represent appears to be the thing with which they are supposed to be so related, or to which they have the function of being so related.² Rather than taking indicators (Dretske 1988, 1995), or pictures, or words to be the analogue of mental representation, I believe that neuroscience and psychology recommend that we adopt the representational paradigm of models. Some of the most familiar examples of models are scale models, like a child’s model airplane, or a model of a building that is to be constructed. Models capitalize on isomorphism. Isomorphism is a relation between two structures (e.g. spatial structures), where a mapping of elements from one structure to the other preserves some pattern of relations across the mapping.³ This mirroring of a pattern of relations is what makes a model useful. When our access to the thing a model represents is somehow restricted, we can use the model to reason about that thing (Swoyer 1991; Cummins 1996). For instance, if we do not know what the left wing of the Spirit of St Louis looks like, we can just consult our model to find out. That is an example ² Teleological theories typically put this ‘supposed to’ in terms of function. Here, this stretches the normal use of ‘function’ a little; normally we say that something has the function of doing something, not of being a certain way. However, it is convenient to use the term to cover both sorts of supposed-tos, and this is how I shall use it. What matters is the normativity, not the functionality per se. ³ Consider a structure S1 , where the elements of S1 are interrelated by a single type of two-place relation, R1 , according to some particular pattern. That is, R1 obtains between certain specific pairs of elements of S1 . S1 is isomorphic to another structure, S2 , if there is a relation R2 (also two-place) and a one-to-one function mapping the elements of S1 onto the elements of S2 such that: for all x and y belonging to S1 , xR1 y if and only if f(x)R2 f(y). This definition may be extended to n-place relations in the obvious way (Russell 1927: 249–50; Anderson 1995).
122
Dan Ryder
of using the model to fill in missing information about the world (‘predictive use’, broadly speaking). Another important use of models is in practical reasoning, in figuring out how to act (‘directive use’). For instance, the scale model of a building might be used as a guide for its construction. (Elsewhere, I have argued that the occurrent attitudes are the causal role equivalents of these two uses; 2002.) Just like representation in indicators and maps, representation in models is a functional property—mere isomorphism is insufficient. A rocky outcrop that just happens to be isomorphic to the Spirit of St Louis does not represent the Spirit of St Louis because the isomorphism in question is not a normative one—the rock is not supposed to be isomorphic to the Spirit of St Louis. A model represents because it has the function of mirroring or being isomorphic to some other structure.⁴ Structures are composed of elements that enter into relations. When two structures are isomorphic, an element of one is said to correspond to a particular element in the other, within the context of that isomorphism. These two relations, isomorphism and correspondence, are promoted to being representational properties when they become normative or functional. A model represents a structure S when it has the function of being isomorphic to S, and the model’s elements then represent the elements of S because they have the function of corresponding to them. Thus representation in models comes in two related varieties, one for the model, and the other for its elements. A model of the Spirit of St Louis models the Spirit of St Louis, while the left wingtip of the model stands in for the left wingtip of the Spirit of St Louis.
3. MODEL-BUILDING Any teleological theory of mental representation faces a problem if it relies solely upon natural selection to endow content-determining functions upon brain states. The problem is this: most of our mental representations, including our representations of kinds, are not acquired through evolution, but rather through learning. Suppose that there are models in the brain. The internal model that you have of the rules of chess is not a model whose isomorphism functions were determined by natural selection! Yet in order for it to be a model of the rules of chess, it must have the function of mirroring the rules of chess. Whence this function? ⁴ Actually, normative isomorphism isn’t sufficient for representation—the representing structure must also be actually or normatively put to one or more characteristic uses, e.g. predictive and/or directive use. See Ryder (2002).
On Thinking of Kinds
123
One teleosemantic strategy for answering this question is to treat learning as a selectional process itself, leading to the acquisition of functions (Papineau 1987, 1993; Dretske 1988). There are two types of selectional processes that might be co-opted to perform the design role. At the neurophysiological level, there is ‘neural Darwinism’, according to which neurons die off in a way that is supposed to be analogous to natural selection (Changeux 1985; Edelman 1987). However, even supposing the analogy is good enough to support a notion of function, the empirical evidence suggests that a selectional account of brain development and learning is at best radically incomplete (Quartz and Sejnowski 1997). The other possibility, at the psychological level this time, is some sort of reinforcement learning story—‘design’ through reward and punishment. (Both Dretske and Papineau rely on this.) The problem here is that new representational capacities can be acquired merely through observation, independently of reinforcement (Sagi and Tanne 1994; Bloom 2000). In this section, I describe a framework that yields an account of nonselectional learning in the form of model acquisition. In the next section, I apply this framework to the brain. Briefly, my story is that evolution designs a model-making machine, and the operation of this machine constitutes learning. (The general framework is closely related to and probably an instance of Millikan’s account of ‘derived relational proper functions’. For the purposes of my argument here, though, I prefer to leave it open whether Millikan’s more general story is correct.) I will start with the intentional design of models, like the model airplane. When someone produces a model of the Spirit of St Louis, they typically consult the actual plane in producing the model. (Of course, they may do so indirectly by consulting photographs, for example.) When we consider the model produced, and ask the question ‘What is this a model of ?’, one way of answering our question would be to tell us what object was used as a template for the model. Since the Spirit of St Louis was the template for the model produced, it is a model of the Spirit of St Louis, and not some other plane that it happens to be isomorphic to. Note how this already begins to move us away from a dependence upon intentional design, which is necessary if we eventually want to apply the account to learning. Suppose the model designer has an intention to produce a model of the Wright biplane, but (mistakenly) uses the Spirit of St Louis as his template. Is the model that gets produced a model of the Wright biplane, or the Spirit of St Louis? There are considerations on both sides. We can further reduce the involvement of intentions by moving to a case of automated model production. Consider the following device, ‘the automatic scale modeller’, designed to produce static models. It takes some object as input, and produces a mould from the object. Next it shrinks
124
Dan Ryder
the mould. Then it injects a substance that hardens inside the mould, and finally it breaks the mould and ejects a small-scale model of the original object. Why is it that we can say that the scale model this machine produces is a model of the original object? Suppose the original object is the Spirit of St Louis. (It is a big machine!) There need not be any intention to produce a model of the Spirit of St Louis at work here. Perhaps someone just set this model-making machine loose on the world, letting it wander about, making models of whatever it happens to come across. (Of course, there were intentions operative in the production of the machine; what we have eliminated is any specific intention to produce a model of the Spirit of St Louis.) The scale model produced is a model of the Spirit of St Louis simply because the plane is what served as a template for production of the model. The function of this machine is not to produce isomorphs of particular things; it has the more general function of producing isomorphs of whatever it is given as input. Each individual model inherits its function of mirroring some specific object O from this general function, and the fact that O is the input that figures in its causal history. Consequently, for any particular model the machine produces, we must know that model’s causal history in order to know what it represents. But there is something else we need to know: the machine’s design principles. In our example, the spatial structure of the model represents the spatial structure of the thing modelled. But the model has a number of other structural features besides its spatial structure; for example it has a density structure. However, these other structural features are not representational. Even if it fortuitously turned out that our scale model of the Spirit of St Louis has exactly the same density structure as the Spirit of St Louis, the density structure of the model would not correctly represent the density structure of the plane (just as a black-and-white TV doesn’t correctly represent the colour of a zebra). This is because if the scale model happened to have a density structure that mirrored the density structure of the real plane, it would be entirely by accident, in the sense that it would not be by design. A model-making machine is designed so that certain specific types of relational features of input objects will cause the production of a specific type of isomorphic structure. Those features of the input object that, by design, determine the isomorphism for the automatic scale modeller are spatial relations—and so spatial relations are the only relations the model represents, that it has the function of mirroring. Similarly, the only relational features of the model that are structured by the input object, by design, are spatial relations. Thus the design principles of the automatic scale modeller tell us that only the spatial features of the model it produces do any representing. When supplemented with the production history of a particular
On Thinking of Kinds
125
model, the design principles can tell us exactly what that model and its elements represent, i.e. what the model has the function of being isomorphic to, and what its elements have the function of corresponding to in the context of that isomorphism. Similarly for any other model-making machine: the machine’s design principles plus the causal history of a particular model will tell us what that model represents. Note that the automatic scale modeller is capable of producing inaccurate models. Perhaps a piece of the machine falls off during its operation, and introduces a lump into the model of the plane. This model says something false about the plane’s structure. Alternatively, it may be that the general design principles for the machine fail in certain unforeseen circumstances, e.g. perhaps deep holes in an object cannot be fully penetrated by the modelling clay. In both of these types of inaccuracies, the machine fails to produce what it is supposed to produce, namely a structure spatially isomorphic to its input. In the automatic scale modeller, there are two stages to the production of a genuine model with a specific content. I propose that we can apply these two stages of model production to the brain, in particular to the cerebral cortex (because the thalamocortical system is the most likely brain structure to subserve mentality). The first stage is the design of the model-making machine, either intentional design (the automatic scale modeller) or evolutionary design (the cortex). The second stage is exactly the same in both: template-based production of specific models according to the design principles of the machine, as determined by the first stage. This is what it is to acquire new representations through (non-reinforcement) learning. If we suppose that the seat of the mind, the cerebral cortex, is designed (by natural selection) to build models of the environment, the crucial question that arises is this: what are the design principles of the cortex? In the next section I will describe, from a functional point of view, the essentials of these design principles according to the SINBAD theory. First, though, a little preview of how this foray into neuroscience will help us eventually answer the question we started out with, of how it is possible to represent kinds. The type of models the cortex is designed to build are dynamic models.⁵ The elements of a static model and the isomorphic structure it represents are constants, like the position of the tip of the plane’s wing, and the position of the tip of the model’s wing (relative to other points internal to the plane). By contrast, in a dynamic model the elements in the isomorphic ⁵ The earliest extended physicalist discussion of the dynamic isomorphism idea, and a defence of its relation to the mind, occurs in Kenneth Craik’s The Nature of Explanation (1943). See also Cummins (1989) and McGinn (1989, ch. 3).
126
Dan Ryder
Cell response
structures are variables. Rather than mirroring spatial structure, a dynamic model mirrors covariational structure. For instance, a model used for weather prediction might have elements that correspond to positions in the atmosphere, where these elements can take on different values depending upon whether there is likely to be rain, snow, a hurricane, or clear sky at that position. The values of the elements in the model covary in complex ways, and those covariation relations are meant to mirror covariation relations in the atmosphere. (Weather models are used only predictively, to fill in missing information (about the future); but if we had the ability to manipulate some weather-affecting variables, such models could be given directive use as well.) The SINBAD theory is a theory of cell tuning. A neuron ‘tunes’ to an entity x in the environment when it adjusts its connections from other neurons such that it has a strong response to x and a weak response to other items (see Figure 6.1). The important thing to note is that cell tuning occurs under the influence of the environment. I think that we ought to conceive of multiple cells’ tuning as a process of template-based production of dynamic models. It was important that the automatic scale modeller was designed so that the represented structure influenced the production of the representing structure in the model. What that means for an automatic dynamic modeller is that the regularity or covariational structure of the environment must influence the structuring of the model. A simple example of such dynamic structuring under the influence of the environment would be classical learning by association (Figure 6.2a). The associationist supposes that we begin with internal items that are already ‘tuned’ to particular things in the environment. Taking the neurophysiological point of view, suppose that the internal items are neurons, and that one neuron begins its life tuned to
Stimulus dimension
Figure 6.1. Cell tuning.
On Thinking of Kinds
127
(a)
(b)
External variable Internal variable
Figure 6.2. In (a), mirroring of pairwise correlations (associationism); in (b), mirroring of multiple correlations.
flashes, and another begins its life tuned to booms. Through a process of association, the pairwise correlation between flashes and booms (in thunderstorms) comes to be reflected in a mirroring covariation between the neurons tuned to flashes and booms. There are a number of reasons why the cortical design principles cannot be those of classical associationism. One particularly serious problem with the associationist proposal is that it is too impoverished to explain our capacity to reason (Fodor 1983). In any case, there is neurophysiological evidence that the regularity structure in the environment that guides production of cortical models is not simple pairwise correlational structure, as the associationist supposes. Rather, the template regularity pattern is of multiple correlations, i.e. multiple features that are all mutually correlated (Figure 6.2b) (Favorov and Ryder 2004). This proposal also receives support from psychology. While people tend to be quite poor at learning pairwise correlations, unless the correlated features are highly salient and the correlation is perfect or near-perfect ( Jennings et al. 1982), when multiple mutual correlations are present in a data-set, people suddenly become
128
Dan Ryder
highly sensitive to covariational structure (Billman and Heit 1988; Billman 1996; Billman and Knutson 1996). Already, a special relationship between the cortex and kinds is intimated. Recall the unified property cluster account of kinds; it said that a kind is characterized by a set of correlated properties, i.e. multiple mutual correlations, where some further principle explains why they are correlated. According to the SINBAD theory, the principal cells of the cerebral cortex are built to take advantage of this general pattern of regularity, the pattern due to sources of correlation (including kinds). Interacting with sources of correlation allows SINBAD networks to become dynamically isomorphic to the environment, making them useful for prediction (and direction). My eventual claim will be this: SINBAD networks (and thus the cortical network) are designed to produce isomorphs to regularity structures involving kinds (and other sources of correlation) specifically. The SINBAD design principles designate kinds as part of the proper template for the cortical model-building machine; thus we can say the cortex, and the enclosing agent, genuinely represent kinds.
4. THE SINBAD THEORY OF THE CEREBRAL CORTEX The relevant cortical design principles apply in the first instance to pyramidal cells (see Figure 6.3), the most common neuron type in the cerebral cortex (70 to 80 per cent of the neurons in the cortex fall into this class—see Abeles 1991; Douglas and Martin 1998). Like any other neuron, a pyramidal cell receives inputs on its dendrites, which are the elaborate tree-like structures as depicted on the cell in Figure 6.3. A cortical pyramidal cell typically receives thousands of connections from other neurons, some of which are excitatory, which increase activity, and others of which are inhibitory, which decrease activity. (Activity is a generic term for a signal level.) Each principal dendrite—an entire tree-like structure attached to the cell body—produces an activity determined by all of the excitatory and inhibitory inputs that it receives. This activity is that dendrite’s output, which it passes onto the cell body. The output of the whole cell (which it delivers elsewhere via its axon) is determined in turn by the outputs of its principal dendrites. The input–output profile of a dendrite, and thus its contribution to the whole cell’s output, can be modified by adjusting the strengths of its synaptic connections, and possibly by modifying other properties of the dendrite as well, like its shape (Woolley 1999; McAllister 2000). An important question in neuroscience is: what principles underlie the adjustments a cell makes in order to settle on some input to output causal profile? Why do
On Thinking of Kinds
129
a principal dendrite
cell body (soma)
axon
Figure 6.3. A typical cortical pyramidal cell. The dendrites form the input region of the cell, which transmits its output via the axon. There is a total of five principal dendrites visible on this cell. (‘Dendrite’ can refer either to a principal dendrite or to a sub-branch of a principal dendrite.) Axons from other neurons synapse on one or more of the thousands of tiny spines covering the dendrites; inhibitory synapses may also occur between spines.
certain connections become highly influential, while others get ignored or even dropped? And what determines the nature of the influence they come to exert? For short: What is the pyramidal cell ‘learning rule’? The SINBAD theory provides one plausible answer.⁶ The proposal is this: that each principal dendrite will adjust its connections so that it will tend to contribute the same amount of activity to the cell’s output as the other principal dendrites on the cell. So if there are five principal dendrites, like on the cell in Figure 6.3, they will each tend to adjust their connections over time so that they will consistently contribute one-fifth of the cell’s total output. I’ll put this by saying, ‘They try to match each other’s activities.’ They are not literally trying, of course; it is merely a brute causal tendency that they have. (The acronym ‘SINBAD’ ⁶ For full details of the SINBAD theory, please see Ryder and Favorov (2001), Ryder (2004), and Favorov and Ryder (2004).
130
Dan Ryder
stands for a Set of INteracting Backpropagating Dendrites, which refers to the mechanism by which the dendrites try to match each other’s activities.) For simplicity, consider a SINBAD cell that has only two principal dendrites. They are trying to contribute an equal amount, 50 per cent, to the cell’s output; that is they are trying to match each other’s activities. And they are trying to do that consistently, no matter what inputs they happen to get. Suppose the cell’s two principal dendrites are connected to the same detector, or sensory receptor. In this situation, it will be very easy for them to match. If they both just pass their input on to the cell body without manipulating it in any way, they will always match. However, dendrites do not get the same inputs, as a rule (Favorov and Kelly 1996). Thus in the typical situation, the two dendrites’ matching task will not be trivial. Suppose, for instance, that they receive two completely unrelated inputs. To use a fanciful example, suppose dendrite A receives an input from a green ball detector, while dendrite B receives an input from a whistle detector. Suppose both detectors go off at the same time; i.e. there is a green ball present, and also a whistle sounds. So both dendrites become active at the same level, let’s say 40 units, and they pass that activity on to the cell body, which will become active at 80 units. The dendrites have both passed the same amount of activity on to the cell body, so according to the SINBAD connection adjustment principle, they will not change their connections at all. The next time either one receives its input, it will treat it in the same fashion as it did this time. But remember that it was a coincidence that there was a green ball and a whistle present at the same time. Next time, perhaps there is just a green ball. The output of the cell will then be 40 units, where dendrite A accounts for 100 per cent of this output, while dendrite B accounts for 0 per cent. The dendrites have radically failed to match. The adjustment principle dictates that dendrite A weaken its connection to the green ball detector, and that dendrite B strengthen any active connections (of which there are none, we are supposing). But it is a hopeless case; the two dendrites will never consistently match activities no matter how they adjust their connection strengths, because they are receiving two utterly unrelated inputs. The only way they can match is if their inputs are in some way mutually predictable. The most basic form of mutual predictability is simple pairwise correlation. If green balls and whistles were consistently correlated, then the two dendrites would be able to match their activities consistently. So, for instance, if dendrite A also received a connection from a beak detector, and dendrite B received a connection from a feather detector, the dendrites could learn to match. The beak and feather connections would strengthen, while the green ball and whistle connections would weaken to nothing. The learning rule would make dendrite A come to respond strongly to beaks,
On Thinking of Kinds
131
and dendrite B to feathers. Because beaks and feathers are consistently correlated in the environment, the dendrites will consistently match. Of course, there are more complex forms of mutual predictability than simple correlation. Real dendrites can receive thousands of inputs, and they are capable of integrating these inputs in complex ways. So the dendrites can find not just simple correlations between beaks and feathers, but also what I call ‘complex correlations’ between functions of multiple inputs. Consider another cell. Suppose that amongst the detectors its first dendrite is connected to is a bird detector and a George Washington detector, and for its second dendrite, a roundness detector and a silveriness detector. (Clearly detectors that no well-equipped organism should be without!) There is no consistent simple correlation between any two of these, but there is a consistent complex correlation—bird XOR George Washington is correlated with round AND silvery. So in order to match consistently, the dendrites will have to adjust their input–output profiles to satisfy two truth tables. The first dendrite will learn to contribute 50 per cent when [bird XOR George Washington] is satisfied, and the second one will learn to contribute 50 per cent only when [round AND silvery] is satisfied; otherwise they will both be inactive (output = 0). Since these two functions are correlated in the environment, the two dendrites will now always match their activities, and adjustment in this cell will cease. Consistent environmental correlations are not accidental: there is virtually always a reason behind the correlations. For example, the correlation between beaks and feathers in the first example isn’t accidental—they are correlated because there exists a natural kind, birds, whose historical nature (an evolutionary lineage) explains why they tend to have both beaks and feathers. What will happen to a cell that has one dendrite that comes to respond to beaks, while the other comes to respond to feathers? The cell will respond to birds—the thing that explains the correlations in its inputs. Similarly, the second cell will come to respond to the kind that explains the complex correlations in its inputs, namely American quarters. SINBAD cells thus have a strong tendency to tune to sources of correlation. Different cells will tune to different sources of correlation, depending upon what inputs they receive. Each cell’s tuning is to be explained by a particular source, and the correlations that source is responsible for (Figure 6.4). When this tuning process takes place over an entire network, the network is transformed so that its flows of activation come to mirror regular variation in its containing organism’s environment. Where the environment has some important variable—a source of correlation—the network will have a cell that has tuned to that source of correlation. And where there is a predictive relation among sources of correlation, the network will be disposed to mirror that relation. The activities of the network’s
132
Dan Ryder
External variable Source of correlation Sensory input SINBAD cell tuned to source of correlation
Figure 6.4. A SINBAD cell tunes to a source of correlation by selecting sensory inputs from the mutually correlated external variables that constitute that source of correlation’s syndrome of features.
cells will covary in just the way that their correspondents in the environment do. In short, the network becomes dynamically isomorphic to the environment. The reason that a cortical SINBAD network develops into a dynamic isomorphism is that cells’ inputs are not only sensory, but also (in fact primarily) derived from within the cortical network. A cell’s tuning is guided, in part, by these intracortical connections (Phillips and Singer 1997). Tuning—changing a cell’s dispositions to react to the environment—occurs through the modification of a cell’s dispositions to react to activity in other cells, mediated by intracortical connections. It is these latter dispositions that come to mirror environmental regularities. Remember, on the classical associationist picture, a pairwise correlation in the environment comes to be mirrored in the brain (Figure 6.2a). On the SINBAD picture, it’s not simple pairwise correlation that comes to be mirrored in the brain, but patterns of multiple correlations, plus their sources (Figure 6.4). In Figure 6.5, where one SINBAD cell’s inputs come from other SINBAD cells in the cortical network, it can be seen that, through the process of cell tuning, the regularities obtaining among the
On Thinking of Kinds
133
cell receiving input from other SINBAD cells
Figure 6.5. When a SINBAD cell receives intracortical inputs from other SINBAD cells, the relations among sources of correlation come to be mirrored in those connections. (The mirroring is shown as bidirectional since cortical connections tend to be reciprocal, though the two directions are mediated by distinct ‘cables’).
sources of correlation upon which these cells depend for dendritic matching will come to be reflected in their intracortical connections. Since sources of correlation are interrelated both within and across levels (cats are related to fur and to mice, water is related to taps and salt, and grass is related to greenness and to suburbia), an extensive network develops, as a cell’s dendrites come to use other cells’ outputs in finding a function that allows them to match.⁷ A cell may start with a tenuous correlational seed,⁸ but this subtle sign of those correlations’ source is enough to put the cell on a path towards discovering the multitude of regularities in which that source participates. As the cell achieves more and more robust dendritic matching, ⁷ Naturally, in getting their dendrites to match, cells can take advantage not only of intrinsic features, but also of relational features of sources of correlation. ⁸ If there is no correlation available, the cell’s activity will be low, and it will elaborate its dendrites in search of new inputs until it is able to find a correlation (for a review of dendritic growth, which occurs throughout life, see Quartz and Sejnowski 1997). If it is still unsuccessful, at some point the cell will ‘give up’ and degenerate (Edelman 1987).
134
Dan Ryder
the correlational seed ends up producing a complex dendritic tree, which realizes complex functions relating that source of correlation to many others (or rather cells that have tuned to them). Thus in tuning to a source of correlation, the dendrites of a particular pyramidal cell find mathematical functions that relate that source of correlation, not only to sensory inputs, but also to other sources of correlation via intracortical connections. In this way, the connections among cells gain characteristics that dispose flows of activity to mirror regularities involving these sources.⁹ This is extremely useful. When a SINBAD cell is activated, this amounts to the network ‘inferring’ the presence of a particular source of correlation, both directly from sensory input, and indirectly from other cells that are active owing to the presence of the source of correlation to which they have tuned. A cell that has tuned to a particular variable has a large number of sources from which it can obtain information about that variable, from numerous sensory input channels and also neighbouring cells. (This is due to the inductive richness of sources of correlation, and the capacity of a cell’s dendrites to make use of much of this richness in learning to match.) If one of those sources of information is blocked, e.g. sensory inputs, the others will compensate.¹⁰ In the context of the networks’ dynamic isomorphism to the environment, a cell that corresponds to the kind tiger (because that is what it has tuned to) will light up when all that is seen is a twitching tail, or even a footprint. That is, intracortical connections allow the network to perform the trick of ‘filling in missing information’. Here then, in brief summary, is how SINBAD networks operate. The multiple dendrites on a SINBAD cell must find mathematical functions of their inputs that are correlated. Assuming these correlations are not accidental, the cell will tune to their source. In tuning to a source of correlation, a cell will provide neighbouring cells with a useful input, i.e. an input that helps their dendrites to find correlated functions. Thus these neighbouring cells, in turn, tune to sources of correlation, and the process repeats. The end result of this complex multiple-participant balancing act is that a SINBAD network, richly endowed with internal links, comes to be ⁹ If you are worried that there are far too many sources of correlation our brains need to have some cells tune to, consider the fact that in the densely interconnected human cerebral cortex, there are somewhere between 11 and 25 billion pyramidal cells (Pakkenberg and Gundersen 1997). Compare this to a good adult vocabulary of 50,000 words. (There is also a mechanism to prevent too many cells from tuning to the same source of correlation—see Favorov and Ryder 2004.) ¹⁰ Of course, this will create a mismatch between dendrites; if a previously correlated input is consistently absent, the dendrite will learn to ignore it in order to achieve a match again with the other dendrites on the cell.
On Thinking of Kinds
135
dynamically isomorphic to the environment from which it receives inputs. This dynamic isomorphism mirrors the deep structure of the environment, with elements that correspond not only to sensory features, but also to the kinds, natural and otherwise, and other sources of correlation around which environmental regularities are structured. 5. A TELEONEUROSEMANTICS FOR THE REPRESENTATION OF KINDS Can the SINBAD theory explain how the representation of kinds is possible? It should be uncontroversial that the cortical network, if it is a SINBAD network, is a model-building machine. Clearly the cortical network is supposed to structure itself isomorphically with regularities in the environment; the utility of this isomorphism is undeniable, for filling in missing information about the world, and in practical reasoning. But our main question, of how it is possible to represent kinds, turns upon the nature of the specific design principles of a SINBAD network. The design principles of a model-making machine dictate its general function, and thus what type of structure it represents. We have seen that SINBAD networks have a strong tendency to become dynamic isomorphisms that mirror regularities organized around sources of correlation. The result we now want to get is that this tendency is teleofunctional: that the cortical SINBAD network was designed to develop such isomorphisms, and consequently that SINBAD cells are supposed to correspond to sources of correlation. Given the analysis of model representation from Section 2, it would follow that SINBAD cells represent sources of correlation. Since kinds form one type of source of correlation, we would have shown how it is possible to represent kinds. Note the contrast between this approach and the descriptionist’s. The descriptionist begins with a (complex) representation of a cluster of observable properties that typically characterize some kind. But, observes the descriptionist, no such representation can ever represent a kind—it will never have the right extension to do so (owing to twin problems, for example). The only way to mentally represent a kind, they continue, is to represent it as a kind in a very strong sense: one must have a detailed conception of kindhood, and somehow link this with the representation of the cluster of observable properties. (In Section 1, I tried to show that this was not a very promising approach.) According to my proposal, this detailed conception of kindhood is not necessary. Contra the descriptionist, it is possible for a representation that does not include anything like a conception of kindhood nevertheless to
136
Dan Ryder
‘get the extension right’ in the case of a kind. The automatic scale modeller produces representations of (relative) spatial points on objects in virtue of having the general purpose or function of producing correspondences to spatial points. It need not have anything like a conception of spatial-pointhood in order to do this. Similarly, a representing device may produce representations of kinds in virtue of having the general purpose or function of producing correspondences to kinds, while utterly lacking a conception of kindhood. (In fact, my proposal is slightly different, of course. It says that the cerebral cortex has a general purpose or function of producing correspondences to the broader class of sources of correlation, but that still means it can ‘get the extensions right’ in the case of kinds, since kinds are simply a variety of source of correlation.) Does this count as representing a kind ‘as a kind’? I don’t know. Some would take a detailed conception of kindhood to be necessary for that phrase properly to apply. All I care about at the moment is solving the problem described in Section 1, which was the problem of ‘getting the extensions right’, something that no other extant theory can manage. (Perhaps SINBAD neurosemantics can also help explain how it is possible to acquire the concept of kindhood, as well as the conception that typically accompanies it, but I make no claims about that here.) So all we need to show is that SINBAD cells have the general purpose or function of corresponding to sources of correlation (or, if you prefer, that the cortex has the function of ‘producing’ cells exhibiting such correspondences). Since evolution is the designer here, we need to make it plausible that the SINBAD mechanism was selected for the properties of its interaction specifically with sources of correlation, that its being structured by sources of correlation in particular confers some benefit compared to other types of model-building (e.g. pairwise association). This is eminently plausible. We saw that the clustering of numerous (possibly complex) properties around a source of correlation allows a cell that tunes to that source to have multiple lines of ‘evidence’ for its presence. The result is an extremely powerful predictive network, with multipotent capabilities for filling in.¹¹ Importantly, SINBAD cells must tune to reliable sources of multiple correlations in order for the network to exhibit this sort of power; the particular advantage of the network depends entirely upon the inductive richness of sources of correlation. SINBAD cells are plausibly built (by evolution) to take advantage of this inductive richness—they have a strong tendency to tune to sources of correlation, and this tendency is what ultimately produces a rich isomorphism. ¹¹ Note that several functions may overlap on a single dendrite, and typically cells will operate in population units, with all members of one population corresponding to the same source of correlation. So the capacity for filling in is astronomical (Ryder 2002).
On Thinking of Kinds
137
There are several related aspects to the way in which SINBAD cells take advantage of the inductive richness of sources of correlation.¹² First, this richness permits a SINBAD network, given the nature of its units, to develop a correspondingly rich inductive network, not of scattered pairwise correlations (as in an associative network), but of interrelated regularities grounded in the deep structure of the environment. This richness ensures robust prediction through redundancy —there are many ways to predict the same thing. Second, the inductive richness of sources of correlation facilitates future learning. Because sources of correlation are inductively rich, once a SINBAD cell starts to tune to one by discovering some of the correlated properties it exhibits (the correlational seed), the cell (given its special properties) is in a uniquely advantageous position to discover further correlation.¹³ (This will be the case as long as its dendrites have not come to match their activities perfectly, which they almost never will, owing to the presence of noise in the cortical network.) We saw that in this way, a cell continually adds to its lines of ‘evidence’ for the presence of the source of correlation to which it is tuning, indefinitely enriching the model’s isomorphism. Relatedly, the success of SINBAD networks depends upon a cell, during the course of learning, receiving useful inputs from other cells in the network. These other inputs will be much more useful to aid dendritic matching if these other cells have tuned to real kinds, so that their outputs carry information about real kinds. That is because environmental regularities are fundamentally determined by interactions among real kinds (and other sources of correlation). So, given the nature of the SINBAD mechanism, it’s vital that cells tune to sources of correlation in order to develop a nice isomorphism. A way to think of it at the network level is this: the SINBAD cortical network was selected for mirroring, not just regularities, but grounded regularities (grounded in sources of correlation). But, one might ask, how can mirroring grounded regularities be selected for? In order for a mechanism to be designed to mirror grounded regularities, it must incorporate some way of picking them out. But how could any mechanism do this? ¹² They are relatives of functions that Millikan attributes to empirical concepts (in the previous chapter in this volume, and her 2000, ch. 3): ‘accumulating information’ about substances (including real kinds), and ‘applying information previously gathered’ about substances. ¹³ This has a corollary: learning becomes much more efficient when a SINBAD network creates correspondences to sources of correlation. Compared to the acquisition of a model lacking such correspondences, fewer relations involving fewer variables must be learned in order for the model to be inferentially complete. (See Favorov and Ryder 2004; Kursun and Favorov 2004).
138
Dan Ryder
Well, the SINBAD mechanism does this, as compared to its competitors (i.e. other broadly ‘associative’ mechanisms, including those actually found in other parts of the brain, like the hippocampus, basal ganglia, amygdala, and cerebellum). It does not, of course, incorporate some infallible groundedness detector, but function does not require infallibility. SINBAD networks exhibit a peculiar sensitivity to multiple mutual correlations, and multiple correlations are a sign of groundedness. (Related ideas commonly surface in epistemology and metaphysics. In epistemology: corroborating evidence indicates truth. In metaphysics: multiple causal powers that are only contingently co-instantiated indicate a real object.) SINBAD networks then, as opposed to other correlation-based mechanisms, have a (fallible) way of winnowing out the useful, grounded correlations. Thus they could have been selected for this ability, and plausibly were. So they plausibly have the specific function of mirroring grounded regularities, in particular (given the reasons adduced above) regularities grounded in sources of correlation. So just as the automatic scale modeller has a dispositional ‘fit’ for producing spatial isomorphs (and not density isomorphs), a SINBAD network has a dispositional ‘fit’ for producing isomorphs to regularities centred around sources of correlation (and not to more generic sorts of regularities). It is precisely this fit with sources of correlation that gives SINBAD networks an advantage over other types of mechanism with respect to richness and redundancy of prediction. So it is reasonable to infer that the cortical network, if it is a SINBAD network, was (in part) designed by evolution to come to mirror regularities specifically involving sources of correlation. This is one of the design principles of the cortical model-making machine. In particular, SINBAD cells were designed to tune to and come to correspond to sources of correlation in the context of the SINBAD network’s isomorphism. Since SINBAD cells have the general function of corresponding to sources of correlation, they represent or stand in for sources of correlation in the models the cortex produces. And since kinds are sources of correlation, SINBAD cells can represent kinds, in the sense that they can ‘get the extensions right’. Particular extensions are determined by these general design principles in conjunction with the history of production of a particular SINBAD model.¹⁴ In order to figure out which specific regularity structure is represented by a specific cortical model, we need to be able to identify what regularity structure served as a template for that model. At a finer grain, we can figure out what sources of correlation served as the template for
¹⁴ Here follows a brief account. For more details, see Ryder (2002, 2004).
On Thinking of Kinds
139
production of each of the elements of the model.¹⁵ The elements of a SINBAD model are particular cells that have tuned or are in the process of tuning to a source of correlation. However, the cell’s template is not just any source of correlation that has helped cause it to fire at some point in its past. Something serves as a template for model production only relative to the design principles of the model-building machine. Call the type of structure that a particular model-making machine is supposed to mirror (according to its design principles) ‘the structural type that is proper’ for that machine. Something can both (1) be an element of a structural type that is proper for a particular machine, and (2) causally affect the structure of a model produced by that machine, while nevertheless failing to be a template for that model. Thus when a rock falls in through the output door in the automatic scale modeller causing a lump on a model of the Spirit of St Louis, that lump does not represent the spatial structure of the rock. Rather, the machine has produced an imperfect model of the Spirit of St Louis. The design principles of the machine dictate what object is being modelled; only objects that enter through the input door are modelled. A better example might be a machine that is designed to fly around, find single objects (by detecting coherent outlines, perhaps), circle around while photographing them, and compile the collection of 2D information into a 3D model, discarding information about surrounding objects. Suppose a rock standing in front of the Spirit of St Louis obscures a portion of the fuselage from all the machine’s camera positions, so that it must interpolate part of its 3D model, an interpolation that fails to mirror the actual spatial structure of the plane. Given the design principles of the machine, the correct way to describe this is as an inaccuracy in the model of the Spirit of St Louis. Even though this portion of the model was in part caused by the rock, it does not in any way represent the rock. (The reason that this is a better example is because this machine’s design process is more open to the environment; it is thus more prone to error, just as a SINBAD network is.) SINBAD cells are designed to come to correspond to sources of correlation through their dendrites learning to match, where this learning is dependent upon some particular source of correlation. This is how the process is supposed to proceed: a cell, which starts off with randomly weighted connections to other cells, is exposed to a source of correlation many times. Upon each exposure, it improves its dendritic matching owing to correlations in some ¹⁵ Regularity-structure-based template determination is more holistic and coherencerelated, while element-based template determination is more atomistic and correspondence-related. Both can be relevant to content determination; where conflicts arise, the result may be a specific kind of equivocation (see below). See also Millikan’s contribution to this volume (previous chapter), on the relation between coherence and correspondence.
140
Dan Ryder
of the properties that have helped cause it to fire.¹⁶ These properties are correlated owing to their being properties of this particular source, so the cell finally comes to tune to and correspond to that source of correlation. This is the functionally normal route for a SINBAD model to adopt a particular configuration, the way it was designed to work: some specific source of correlation causes (or explains) each cell’s achievement of dendritic matching. Only in this way can a cell participate in a reliable predictive network. It is equivalent to the automatic scale modeller taking in an object through its input door, producing a nice mould, and spitting out a perfect model. In a SINBAD model, deviations from the way structuring is supposed to proceed by design will be deviations from the way tuning is supposed to proceed by design. These will include causal interactions with things that have inhibited a cell from achieving its current level of matching success, and in most cases, these inhibitors will not be templates for the cell. For example, consider the following history of SINBAD model production. Suppose a cell had been gradually tuning to cats. Perhaps a dog caused the cell to fire at some point, because in certain conditions, dogs look like cats. Let us say that the dog made three of the cell’s dendrites match, while there was a failure to match for two dendrites. The SINBAD learning rule made some of these dendrites modify their connections. But this does not improve their overall matching success. The dendrites will tend to move away from functions that pick out cats (functions they had previously been tending towards), without taking them any closer to functions that pick out dogs, or anything else. Without consistent ‘training’ through exposure to multiple dogs, the dendrites are unlikely to modify their behaviour so as to increase their sensitivity to features characteristic of dogs. (In fact, the only thing they might improve their sensitivity towards, in this case, is this particular dog in this particular circumstance—and that is not even a source of correlation.) Subsequent exposures to cats bring the dendrites back towards the function that picks out cats, and the cell back towards better matching success (and thus predictive utility). That response to a dog inhibited the cell from achieving its current level of matching success; it led it away from finding the correlated functions due to cats, the isomorphism it has now settled on. The dog was something that affected the model, but not ¹⁶ Despite the fact that the causal relation between the activity in a SINBAD cell’s dendrite and a source of correlation is mediated by some intervening physiology, it is still causation in virtue of some determinate property of the stimulus. Which property was causally relevant to the activity in the synapse can be identified by counterfactuals. Suppose an instance of a particular shade of red causes a cell to fire. Had the stimulus been a different shade of red, the cell would have fired anyway. However, had the stimulus been blue, it would not have fired. Then the property that was causally relevant to the synaptic activity was redness, not the particular shade of red, nor colouredness.
On Thinking of Kinds
141
according to design. It features in the history of the cell’s tuning, but it did not cause or explain its matching success, i.e. the aspect of model structuring that occurs by design. That dog was a stray rock in the SINBAD mechanism, while the kind cat was this cell’s template. A cell’s template is a source of correlation that explains its current matching success. This does not mean that the model-building machine was broken when it changed so as to reduce its predictive utility; it was just functioning suboptimally. Also note that our conclusion that this cell represents cats is consistent with an alternative history in which the cell permanently veers off its course, eventually tuning to and representing dogs. In this case, the cell would have different properties with a different explanation for its matching success, and it would be in a different model, with a different history—indeed, at some point in its progress, the kind cat may cease to provide any explanation for its matching success. Its matching success may depend entirely upon its previous exposure to dogs.¹⁷ On the other hand, if the kind cat (as well as dog) continues to explain its matching success, the cell will be disjunctive or ‘equivocal’, where two kinds are confused as being the same (Ryder 2002). (See Millikan 2000 on equivocal concepts, which certainly exist and so ought to be psychosemantically explicable.) This is another way model design can proceed sub-optimally. It is sub-optimal since it will lead to inductive errors. So the design principles of the cortical model-making machine pick out, as a cell’s template, only the things that have helped that cell achieve its current matching success (where that is measured holding its response profile and current broad environment fixed). Anything else does not explain the creation of an internal model according to the cortical design principles. Therefore, a single SINBAD cell¹⁸ has the function of corresponding only to the source of correlation that actually helped it achieve the degree of dendritic matching it has attained thus far. That is the source of correlation ¹⁷ Note that this avoids a problem that arose for Dretske’s (now abandoned) account of representation in Knowledge and the Flow of Information (Dretske 1981). In this book, Dretske identifies the content of an indicator with the information that was instrumental in causing it to develop a particular sensitivity during its ‘learning period’. When it responds to something else after the learning period, it misrepresents. Suppose an indicator has been responding to As, and then it responds to a B. On Dretske’s old theory, if this latter response is part of the learning period, Bs will be part of the indicator’s content, but if it is part of the ‘use’ period, then the indicator misrepresents the B as an A. Which is it? As Loewer (1987) points out, the problem for Dretske is that there is no principled way to distinguish between the ‘learning’ period and the ‘use’ period. In SINBAD neurosemantics, there is no need to distinguish between a learning period and a use period. You just need to ask whether B explains to some non-negligible extent, the cell’s current matching success. ¹⁸ I note again that representations that actually have a cognitive effect will typically involve populations of SINBAD cells.
142
Dan Ryder
the cell represents. Anything else that it responds to, has responded to, or corresponds to in the context of some isomorphism is not part of the cell’s representational content. So not only can SINBAD cells have the function of corresponding to kinds in the context of an isomorphism, the details of the SINBAD mechanism allow us to determine exactly which kind (or other source of correlation) a particular SINBAD cell has the function of corresponding to. An element of a model represents that which it has the function of corresponding to. So if all goes well, that kind will be the unique representational content of the cell. Since SINBAD cells are the basic elements of a SINBAD network, we can also determine which regularity structure a whole network has the function of being isomorphic to, and thus models. Because of their inductive richness and SINBAD’s penchant for such richness, kinds will tend to figure prominently in these internal models. Which, in addition to the evidence linking SINBAD to the cortex, and the cortex to the mind, is an important reason to suppose that mental representation, at least in us, is SINBAD representation. REFERENCES A, M. (1991), Corticonics: Neural Circuits of the Cerebral Cortex (Cambridge: Cambridge University Press). A, C. A. (1995), ‘Isomorphism’, in J. Kim and E. Sosa (eds.), A Companion to Metaphysics (Oxford: Blackwell). A, M. (1997), ‘Has Fodor Really Changed his Mind on Narrow Content?’, Mind and Language, 12/3–4: 422–58. B, D. O. (1996), ‘Structural Biases in Concept Learning: Influences from Multiple Functions’, in D. Medin (ed.), The Psychology of Learning and Motivation (San Diego: Academic Press). and H, E. (1988), ‘Observational Learning from Internal Feedback: A Simulation of an Adaptive Learning Method’, Cognitive Science, 12: 587–625. and K, J. (1996), ‘Unsupervised Concept Learning and Value Systematicity: A Complex Whole Aids Learning the Part’, Journal of Experimental Psychology: Learning, Memory and Cognition, 22: 458–75. B, P. (2000), How Children Learn the Meanings of Words (Cambridge, Mass.: MIT Press). B, R. N. (1991), ‘Realism, Anti-Foundationalism, and the Enthusiasm for Natural Kinds’, Philosophical Studies, 61: 127–48. C, J-P. (1985), Neuronal Man: The Biology of Mind (Oxford: Oxford University Press). C, K. J. (1943), The Nature of Explanation (Cambridge: Cambridge University Press).
On Thinking of Kinds
143
C, R. (1989), Meaning and Mental Representation (Cambridge, Mass.: MIT Press). (1996), Representations, Targets, and Attitudes (Cambridge, Mass.: MIT Press). D, D. (1987), The Philosophical Lexicon, . D, R., and M, K. (1998), ‘Neocortex’, in G. M. Shepherd (ed.), The Synaptic Organization of the Brain (Oxford: Oxford University Press). D, F. (1981), Knowledge and the Flow of Information (Stanford, Calif.: CSLI Publications). (1986), ‘Misrepresentation’, in R. J. Bogdan (ed.), Belief: Form, Content and Function (Oxford: Oxford University Press). (1988), Explaining Behavior (Cambridge, Mass.: MIT Press). (1995), Naturalizing the Mind (Cambridge, Mass.: MIT Press). E, G. M. (1987), Neural Darwinism: The Theory of Neuronal Group Selection (New York: Basic Books). F, O. V., and K, D. G. (1996), ‘Local Receptive Field Diversity within Cortical Neuronal Populations’, in O. Franzen, R. Johansson, and L. Terenius (eds.), Somesthesis and the Neurobiology of the Somatosensory Cortex (Basel: Birkhauser). and R, D. (2004), SINBAD: A Neocortical Mechanism for Discovering Environmental Variables and Regularities Hidden in Sensory Input’, Biological Cybernetics, 90: 191–202. F, J. (1983), The Modularity of Mind (Cambridge, Mass.: MIT Press). (1990), A Theory of Content and Other Essays (Cambridge, Mass.: MIT Press). (1991), ‘Reply to Antony and Levine’, in B. Loewer and G. Rey (eds.), Meaning in Mind: Fodor and his Critics (Oxford: Blackwell). (1994), The Elm and the Expert (Cambridge, Mass.: MIT Press). (1998), Concepts: Where Cognitive Science Went Wrong (Oxford: Oxford University Press). H, G. (1987), ‘(Nonsolipsistic) Conceptual Role Semantics’, in E. Lepore (ed.), Semantics of Natural Language (New York: Academic Press). J, D. L., A, T. M., and R, L. (1982), ‘Informal Covariation Assessment: Data-Based versus Theory-Based Judgments’, in D. Kahneman, P. Slovic, and A. Tversky (eds.), Judgment under Uncertainty: Heuristics and Biases (Cambridge: Cambridge University Press). K, F. (1989), Concepts, Kinds, and Cognitive Development (Cambridge, Mass.: MIT Press). K, H. (1993), Inductive Inference and its Natural Ground (Cambridge, Mass.: MIT Press). K, O., and F, O. (2004), ‘SINBAD Automation of Scientific Discovery: From Factor Analysis to Theory Synthesis’, Natural Computing, 3/2: 207–33. L, B. (1987), ‘From Information to Intentionality, Synthese, 70: 287–317. MA, A. K. (2000), ‘Cellular and Molecular Mechanisms of Dendrite Growth’, Cerebral Cortex, 10: 963–73.
144
Dan Ryder
MG, C. (1989), Mental Content (Oxford: Blackwell). ML, B. (1987), ‘What is Wrong with Correlational Psychosemantics’, Synthese, 70: 271–86. M, M. (1988), ‘Biological Functions and Perceptual Content’, Journal of Philosophy, 85: 5–27. M, D. L., and O, A. (1989), ‘Psychological Essentialism’, in S. Vosinadou and A. Ortony (eds.), Similarity and Analogical Reasoning (Cambridge: Cambridge University Press). M, R. (1984), Language, Thought, and Other Biological Categories (Cambridge, Mass.: MIT Press). (1998), ‘A Common Structure for Concepts of Individuals, Stuffs, and Real Kinds: More Mama, More Milk, and More Mouse’, Behavioral and Brain Sciences, 21/1: 55–65. (1999), ‘Historical Kinds and the ‘‘Special Sciences’’ ’, Philosophical Studies, 95/1–2: 45–65. (2000), On Clear and Confused Ideas (Cambridge: Cambridge University Press). P, B., and G, H. J. G. (1997), ‘Neocortical Neuron Number in Humans: Effect of Sex and Age’, Journal of Comparative Neurology, 384: 312–20. P, D. (1987), Reality and Representation (Oxford: Blackwell). (1993), Philosophical Naturalism (Oxford: Blackwell). P, W. A., and S, W. (1997), ‘In Search of Common Foundations for Cortical Computation’, Behavioral and Brain Sciences, 20: 657–722. P, J. (2002), Furnishing the Mind: Concepts and their Perceptual Basis (Cambridge, Mass.: MIT Press). P, H. (1975), ‘The Meaning of Meaning’, in K. Gunderson (ed.), Language, Mind and Knowledge (Minneapolis: University of Minnesota Press). Q, S. R., and S, T. J. (1997), ‘The Neural Basis of Cognitive Development: A Constructivist Manifesto’, Behavioral and Brain Sciences, 20: 537–96. R, B. (1927), The Analysis of Matter (London: Routledge). R, D. (2002), ‘Neurosemantics: A Theory’, Ph.D. diss., University of North Carolina. (2004), ‘SINBAD Neurosemantics: A Theory of Mental Representation’, Mind and Language, 19/2: 211–40. and F, O. V. (2001), ‘The New Associationism: A Neural Explanation for the Predictive Powers of Cerebral Cortex’, Brain and Mind, 2/2: 161–94. S, D., and T, D. (1994), ‘Perceptual Learning: Learning to See’, Current Opinion in Neurobiology, 4: 195–9. S, R. (1984), Inquiry (Cambridge, Mass.: MIT Press). S, D. (1977), ‘Toward a Causal Theory of Linguistic Representation’, in P. A. French, T. E. Uehling, and H. K. Wettstein (eds.), Midwest Studies in Philosophy, ii: Studies in the Philosophy of Language (Morris: University of Minnesota Press).
On Thinking of Kinds
145
S, C. (1991), ‘Structural Representation and Surrogative Reasoning’, Synthese, 87: 449–508. U, M. (2001), ‘A Statistical Referential Theory of Content: Using Information Theory to Account for Misrepresentation’, Mind and Language, 16/3: 311–34. W, C. S. (1999), ‘Structural Plasticity of Dendrites’, in G. Stuart, N. Spruston, and M. H¨ausser (eds.), Dendrites (Oxford: Oxford University Press).
7 Teleosemantics and the Consumer Mohan Matthen Perceptual systems are automatic sorting machines. (See Matthen 2005, pt. .) By processing the energy or chemical signals emanating from environmental stimuli, they sort these stimuli into classes on some consistent basis. From the perceiver’s point of view, this activity culminates in perceptual appearance. Stimuli that have been assigned to the same class present the same appearance: for example, things that are assigned to the same class by the colour vision system look the same with respect to colour. Conversely, things that have been assigned to different classes present different appearances: things that have been assigned to different classes by colour vision look different with respect to colour. Thus, things that look blue—or the same shade of blue—have been assigned to the same class by colour vision; their blue appearance is an internally generated sign of this. Such classes are nested: thus, two things that have been assigned to blue may at the same time have been assigned to different shades of blue. That they both look blue is a consequence of the first assignment; that they look somewhat different in colour is a consequence of the second. 1. WHERE TELEOSEMANTICS FITS IN The line of thought expressed above epitomizes what J. J. Gibson called an ecological approach to perception. In this way of thinking, a perceptual experience is not merely —as classical empiricists like Locke would have it—an occurrence within the mind of a perceiver from which the state of the external world can be inferred. Rather, it is an engagement with what the sensory system takes to be an environmental object (Matthen 2005, pt. ), and betokens a sub-personal act of classification with regard to this object. (I take ‘sub-personal’ to be roughly equivalent to ‘modular’ in the sense given by Jerry Fodor 1982.)
Teleosemantics and the Consumer
147
Now, if one adopts this ecological line of thought, it is natural to ask: What feature do objects assigned to a single class share? what, for example, is the feature shared by things that appear blue? Is it possible to say in an informative way what distinction a sensory system is making when it classifies an object in a certain way? One may distinguish two ways of answering this question more or less consonant with the ecological approach. The first concentrates on the process leading up to the perceptual experience. Things that appear blue have occasioned a process that culminates in the perceptual experience of them as blue. In view of this, one might propose that blue is the feature of environmental objects that explains this commonality of perceptual process, the feature that stands at the head of a causal process with this result. But this proposal has an obvious flaw. There are things that appear blue but are not really blue. These things also stand at the head of a perceptual process that culminates in the appearance of blue. But they are not blue. Equally, there are things that really are blue, but do not look to be so. These things do not share in the feature in question, but are blue. So it looks as if this process proposal cannot accommodate the evident divergence between appearing to be of a kind and really being of that kind. This flaw can be corrected by concentrating on the range of circumstances, ‘normal’ or ‘standard’ conditions as they are variously called, in which things that appear blue are guaranteed to be blue, and stipulating that blue is the feature of environmental objects that explains the appearance as of blue in these circumstances. But one may also abandon the causal approach altogether, in favour of what one might call a semantic approach. At the heart of this approach is the idea that perceptual experiences are representations of environmental objects: thus, they refer to objects, attribute features to them, and possess in consequence, a truth-value. That something is represented as blue is compatible with its not actually being blue. So we do not have to appeal to non-normal conditions to explain how there can be a mismatch between sensory appearance and reality. This change of perspective comes, however, at the cost of introducing a new and controversial concept—that of representation. The natural home of this concept is in the study of communication between agents who possess intentions and goals. It is not immediately clear how it can be extended to states issued by automatic sub-personal systems. Let us put this problem aside until later. The semantic approach to perceptual experience demands a reformulation of the question asked earlier. Blue is, on this approach, the feature attributed to an object of perception by a perceptual experience of that
148
Mohan Matthen
object: it is the property a thing appears to possess just in virtue of looking blue. Thus: X is blue if and only if X has the feature attributed to something by its looking blue. So the question can now be put in the following way: what feature is attributed to a stimulus by a perceptual experience of that stimulus as blue? Or: what feature does a thing appear to have just in virtue of looking blue? Now, this is not, strictly speaking, a semantic question. All that semantics tells us about simple terms is what they refer to or denote. Further, it tells us how to compute what complex terms refer to or denote, given what their simple components refer to or denote. The question now being considered is about how a simple term gets its semantic value, so to speak, or about the real-world relation that must hold between a term and its semantic value. This goes beyond the scope of semantics itself: it is a meta-semantic question. Summarily: it is a question not about the meaning of something’s looking blue, but about how meaning relations get established in the world. The situation here is similar to that concerning the causal theory of names. David Kaplan (1989: 574) observes that this theory—the ‘historical chain theory’, as he calls it—does not tell us what a name means. The meaning of a name is just its referent, Kaplan insists. Rather, the historical chain theory tells us how—by virtue of what relation—a particular object gets to be the meaning of a name. The theory offers us a genetic explanation of a semantic relation, and is, in this way, meta-semantic. When I use the name ‘Aristotle’, I am simply referring to Aristotle. In order to do this, I do not need to know how this name came to be meaning-related to Aristotle himself. When I use the term ‘Aristotle’ to refer to Aristotle, I am displaying implicit or explicit grasp of the conventions governing the use of established proper names. In order to do this, I do not need to know anything about the social, religious, or legal customs by which such names get attached to individuals, or about the means by which these name–individual connections ground naming relations in the language. This is for a meta-semantic theory, which is of considerable interest to theoreticians of language, but not a part of what a user of language must know. In exactly the same way, the semantic approach to perceptual content should not be taken as aiming to expose how a perceptual experience came to meaning-relate to a feature of things in the world—as above, a perceiver does not need to know this in order to grasp the meaning of an experience of something as blue. The question posed above is meta-semantic. It seeks a specification of the relation that must obtain between a perceptual experience and a sense-feature if such an experience is to denote the feature.
Teleosemantics and the Consumer
149
One simple theory of the meta-semantics of perceptual experience states goes like this: Indicator Meta-Semantic Theory. A perceptual experience E attributes feature F to its object if states of the type E occur only when caused by a thing with feature F. Introducing a technical term to simplify the above cumbersome locution: E attributes F to x if an occurrence of E indicates that x is F. This account imports ideas of causal origin from the process approach sketched above, and fails for exactly the same reason. It does not accommodate the divergence of appearance and reality. It seems clear that it is possible to misperceive something as blue. Such a misperception attributes to its object a feature that the object does not actually possess. Yet, if attribution was a matter of indication, this would be impossible: for an experience that attributes the feature could not have occurred if the feature was not really there. A second meta-semantic theory, not subject to this difficulty, goes like this. Teleo-Meta-Semantic (TMS) Theory. A perceptual experience E attributes feature F to an object x if states of type E have the biological function of getting the perceiver to respond in a functionally appropriate way to x possessing F. Once again, simplifying by the introduction of a technical term meant simply to abbreviate the above: E attributes F to x if an occurrence of E is supposed to initiate the appropriate response to x being F. Teleosemantic theories—for the most part, I drop the ‘meta’ in what follows—treat of perceptual experience as an intermediary between an external situation and what an organism is supposed to do in that situation. The experience tells the organism that the object of perception (i.e. x in the above schema) has the feature (F), and that it should respond appropriately; it is not merely an inert indicator that this situation has occurred. Fred Dretske puts this point in terms of a distinction between ‘natural signs’ which merely indicate, and representations, which are treated as signs for a particular purpose. He says: ‘Putting chilled alcohol in a glass cylinder doesn’t generate a misrepresentation unless somebody calibrates the glass, hangs it on the wall, and calls it a thermometer’ (1988: 67). It is when it is used as a sign of the temperature that the alcohol becomes a representation. This is the central idea of the Teleosemantic Theory. (Later on, we
150
Mohan Matthen
will essay a more comprehensive defence of the idea that perceptual experiences are representations. This paragraph will suffice for now.) Clearly, something can have a certain function but not perform it in a particular instance. Thus, the above theory is compatible with falsity of attribution. A virtue of TMS Theory is that it seems to make the question we are dealing with—‘What is the feature attributed to a thing by an experience of it as blue?’—recognizably scientific. Scientists in many areas of the life sciences—evolutionary biology, physiology, psychology, ethology, etc.— ask about the biological function of animal organs and the various states and conditions of these organs. The methodology of addressing these questions is contested, but constitutes fairly familiar territory nonetheless. Thus, the Teleosemantic Theory serves as a bridge between philosophy and cognitive science.
2. ENTER THE CONSUMER TMS Theory meets an immediate point of resistance when one tries to apply it to cognitively sophisticated organisms. What is the functionally appropriate response to a given perceptually represented situation—for instance, to the experience of something as blue? Is there such a thing? In Matthen (1988) I argued that there is not. It seemed obvious to me then that while sensory states in primitive organisms, or vestiges of these states in sophisticated organisms (e.g. the sneeze, the startle response, etc.), might lead directly to autonomic responses, perceptual states in human and other higher animals might lead to any number of responses depending on other volitional and cognitive states. There is no determinate way that we are supposed to react to the presence of a blue thing much less to one that merely appears blue. Suppose I see a nearly full glass of beer on the table. It is by no means automatic that I will pick it up and take a sip. Is it mine or somebody else’s? Am I drunk or sober? Is it six o’clock on a hot evening when I am looking forward to my first drink of the day? Or is it six in the morning when I have found an unfinished glass that someone left there overnight? These alternatives suggest that the perceptual state feeds into a complicated process of contextualization, cognitive assessment, and decision-making. This is its function, not the initiation of some autonomic response. It seems not just simplistic, then, but flat out wrong-headed to define perceptual content in terms of some single response which perception is supposed to initiate. Call this the Problem of Multiple Responses. Responding to this Problem, I suggested that the states in lower organisms (or in ourselves) that tie into autonomic responses were not full-fledged perceptual states but only ‘quasi-perceptual’. The function of full-fledged
Teleosemantics and the Consumer
151
perceptual states, such as the one illustrated above, was merely to ‘detect’ (or, using the terminology introduced above, ‘indicate’—see Dretske 1988; Matthen 1989) certain situations, leaving it up to the perceiver to decide what ought to be done with the information so provided. Perception is detection, I proposed, not the initiation of an appropriate response. By so arguing, I brought my position close to that of Dretske (1988) as quoted above. In short, I proposed: Weak TMS Theory. E attributes F to x if an occurrence of E is supposed to indicate that x is F. The normative element contained in the words ‘supposed to’ is essential here; it distinguishes TMS Theory from the Indicator Meta-Semantic Theory. In the weaker version of TMS Theory, there is no requirement that a perceptual state should initiate a specific response to the situation represented, as there is in the (stronger) TMS Theory articulated earlier: indeed, the claim was that perceptual states proper (as opposed to quasi-perceptual states) do not initiate specific responses. This proposal immediately raised Ruth Millikan’s ire (1989/1993). Millikan is, of course, a pioneer of teleosemantic theories (though her focus was originally on language and communication, rather than perception), and as such she subsumes perceptual states under the broader category of representations. A representation, she says, is an item made for use by a ‘consumer’. It accords by a certain rule with the situation it represents (‘accords by a certain rule’ is, I think, her version of ‘supposed to indicate’), and the consumer is thereby enabled to respond appropriately to that situation. A beaver splash representing danger is an example. If a beaver splashes its tail in the danger-representing way, and danger in fact exists, then other beavers in the vicinity, the consumers of this representation, will do what they are supposed to do in the presence of danger. If it splashes its tail when there is no danger, then these consumers will thereby be given reason to do what would have been appropriate had there been danger, but not what is appropriate in the actual non-threatening situation. Potentially, this could cause the beaver’s normal activities to be disrupted. Thus, the accuracy of the representation is, as Millikan says, a normal condition of the success of the consumer’s actions. (In response to the Problem of Multiple Responses, this should be amended to read ‘normal condition of the success of the consumer’s choice of actions’.) The Problem of Multiple Responses to a given situation is beside the point, Millikan argued. What is important is that a connection exist between a given cognitive state and the situation or feature that it is supposed to represent. With regard to the glass of beer, the accuracy of my perception is a normal condition for the success both of drinking it (when that is
152
Mohan Matthen
functionally appropriate) and of throwing it away (ditto)—both actions would be guided by the same perception. The correspondence of my perceptual state to the situation that it represents is thus a normal condition for the success of any action I might take with respect to it. Thus, ‘representational content rests not on univocity of consumer function but on sameness of normal conditions for those functions’. To summarize: the same factrepresentation can be a normal condition for a variety of different actions. Thus, according to Millikan, there is no need to revise or weaken TMS Theory. What is needed is to realize that fact-detection underwrites all the different uses that might be made of a representation. Of course, this is not to deny that representations have a function of detecting that which they represent. Organisms have evolved in such a way as to be able to perform certain responsive functions—functions that demand different actions in response to different environmental conditions. This means not only that they will be able to effect actions in its repertoire, but that they will be able to discriminate one situation from another, and perform the right action in the right circumstances. This demands a detector to make sure that the effector is well informed of the circumstances to which it is meant to respond. The function of such a detector will be that of issuing token representations that indicate what circumstances obtain, at least with accuracy sufficient not to disrupt the task at hand. Thus, I take it that Millikan did not mean to disagree with the idea that the function of perceptual states is ‘detecting’ or ‘indicating’—issuing a representation that properly accords with a certain state of affairs (except in so far as it might not have been clear that ‘detecting’ a situation was merely another word for ‘being in accord by a certain rule’ with it). What she meant to challenge was the idea that there is no single function if there is no single response. ‘Matthen is . . . looking pretty squarely at the representation consumers, but at what it is the representation’s job to get these consumers to do, rather than at normal conditions for their proper operation,’ she complains. This focus on initiating a response blinded Matthen (i.e. me) to the single function of representations, she seems to suggest. This is wrong. I did not say that perceptual states had no single function; on the contrary, I said that they had the single function of detecting a certain situation—of registering that the situation has occurred and putting this information at the service of the perceiver’s effector organs. I arrived at the same conclusion as Millikan, but in slightly different terminology by a slightly different route. Millikan too was proposing a Weak TMS Theory in so far as she supposes it to be the function of a representational state simply to be in accord with a particular situation. She takes it that when we are specifying the function of the representation, the specific action-choice for
Teleosemantics and the Consumer
153
which the consumer uses this representational state is irrelevant. All that is relevant is that any action-choice will be disrupted by the representation’s not being in accord with the situation it represents. Millikan, however, went much further than I did with detecting. Whereas I had been content to assume that in very low-level organisms, the representational state that accorded with environmental situations was the very same as the one that initiated the action, Millikan—rightly, I think—insisted on distinguishing between the detector function and the effector-triggering function even in simple ‘pushmi-pullyu’ representations, as she calls them (1995)—states that combine detector and effector functions in a single package. Imagine that a certain cell-body in a unicellular organism enters into a particular chemical state when a particular situation obtains, and that this chemical state initiates a certain behaviour in the organism. (Perhaps the lack of some nutrient causes a molecule to be stripped of a nucleotide, which in turn causes the absent nutrient to be synthesized.) The chemical state of this cell-body both detects the occurrence of the triggering situation and initiates the appropriate response to that situation, and is, as such, a pushmi-pullyu representation. Despite such co-location of detecting and effecting, it is important to distinguish these two capacities, and to link them in precisely the way that Millikan does, i.e. to say that the successful performance of the detector function is a normal condition for the successful performance of the effector function. The detector manufactures representations; the effector is a consumer of the representation. This is an absolutely crucial insight for teleosemantics. However, the point goes deeper than Millikan realized. Effector functions cannot be disregarded or ignored when we ask about detector functions. To be sure, it is a function of detectors to be accurate, and this accuracy forms part of the normal conditions of effector success. But this does not mean that we can state the detector function in abstraction from effector function. For we cannot properly identify which feature the detector is supposed to detect accurately except by reference to effector function. This is what Millikan (and I) failed to appreciate. Both of us assumed (I surmise) that we could find the feature in question by looking ‘upstream’, i.e. by looking at the environment. In fact, we can find it only by looking at the consumer. Or so I shall argue.
3. WHY MULTIPLE RESPONSES ARE NO PROBLEM Let me return to the Problem of Multiple Responses. I now want to argue that both Millikan and I were wrong about this. We both overlooked the
154
Mohan Matthen
fact that even in cognitively sophisticated organisms, there actually is a set of autonomic responses to perceptual experiences, and that these are available for the identification of the feature attributed to objects by perceptual experience. There is hence no need to remove from TMS Theory the reference to a functionally appropriate response (though it is important to recognize that it implies both a detector and an effector function). The reason why this set of autonomic responses was not evident in the first place is that they are not bodily actions which the organism uses to act upon its environment, but self-directed changes to what might be called its representational or epistemic state. The difference between perception in cognitively sophisticated organisms and in detector functions in pushmi-pullyu representations is not that the reaction to the former is ‘free’ while those to the latter are autonomic, but rather that in the more evolved organisms, perception controls bodily action mostly through the mediacy of epistemic operations. However, perception still evokes autonomic responses in cognitively sophisticated organisms: in particular, and crucially for its function in these organisms, it feeds autonomically into a change of epistemic state. At a very simple level, consider the phenomenon known as habituation. If you present a young baby with a blue stimulus repeatedly, its initial interest and attention dies away—on the fifth trial or so, it ignores the presentation. If after such a series of blue presentations, you present the baby with a red stimulus, it regards this new presentation with something approximating the interest and attention it gave the original blue item. The series of blue representations changes the baby’s inner state, which leads it to treat a further repetition with a lower level of interest than it would have met at the start of the series. Notice how sensory classification ties into this pattern. The baby reacts with less interest to the fifth blue presentation because each stimulus in the series was co-classified with the others. It reacts with renewed interest to the red presentation because this one was differentiated from the previous ones. Thus the sorting function of colour perception is essentially involved with the reaction pattern. More abstractly, response depends on sensory similarity or dissimilarity. It is apparently functionally appropriate that organisms should react less strongly to the occurrence of a stimulus similar to something that they have recently encountered, and perceptual sorting is (among other things) a device for detecting whether what they are encountering now is similar to or different from what they encountered earlier. In short, habituation is an autonomic response to perceptual sorting, and perceptual sorting is a detector function for habituation. Notice that regardless of what the baby (or some cognitively sophisticated adult) does with the series of presentations, she cannot avoid the change of internal epistemic state.
Teleosemantics and the Consumer
155
Imagine an actor who is told to react with pretended surprise to the blue presentation. She rehearses this action again and again, and perfects a look of surprise. She has not avoided habituation, but has simply ignored it in her outward reaction. There are multiple responses to the repeated blue presentation, then, but there is also an autonomic response. Habituation illustrates a more general phenomenon. Any perceptual state alters the state of inner epistemic ‘organs’, with the consequence that responses to further presentations of that or other stimuli are altered. In computational terms, these epistemic organs have been represented, ever since Donald Hebb, as networks of connected neurons. Every new perceptual state affects the strengths of the connections between neurons in these networks. In more cognitive terms, one might say that the organism maintains a set of inner expectations, and every incoming perceptual state automatically occasions an update of these expectations. With respect to these expectations, a new stimulus will reinforce an expectation to the degree that is experienced as similar to occurrences encountered before, or weaken an expectation to the degree that it is dissimilar to what has been encountered before. Such mechanisms for creating, maintaining, and extinguishing expectations are, as Quine (1969: 306) argues, innate. For ‘There could be no induction, no habit formation, no conditioning, without prior dispositions on the subject’s part to treat one stimulation as more nearly similar to a second than to a third.’ This implies that there can be no learning without an unlearned, i.e. innate, capacity to measure perceptual similarity. Broadly speaking, these expectations are of two sorts. First, we have expectations concerning individual objects—about where they will be, what colour or shape they will be, etc. Second, we have expectations about associations among features: about what size goes with what weight, what colour goes with what taste, and so on. Though these expectations change measurably only after a series of perceptual presentations, the most plausible way to represent the connection between perception, on the one hand, and learning or memory, on the other, is to posit that each and every perceptual state has some effect on some expectation or ‘epistemic’ net. (The term ‘epistemic’ is perhaps inflated in this context. The point is that even very simple organisms are capable of forming environmental representations that last longer than a single interaction with that environment. In the case of habituation, each presentation lasts only a second or so, but the change to the response disposition lasts much longer. The term ‘epistemic’ is used to mark this longer duration, the change of dispositions over the long run; no suggestion of reasoning or justification is intended.) In view of the occurrence of such autonomic epistemic responses, we may elaborate TMS Theory: the responses initiated by perceptual states
156
Mohan Matthen
are epistemic, especially in cognitively sophisticated animals. Briefly, the similarity of two things depends on how similar the epistemic response to them is. Two exactly similar things will occasion exactly the same response. For example, if R1 and R2 are exactly similar red chips, and a baby has been habituated to blue, then they will evoke the same reaction of attention and surprise. On the other hand, a violet chip which is somewhat similar in colour to both blue and red will evoke a less surprised reaction, or perhaps evoke surprise less often. A sense feature like blue comprises items to which the response is the same, up to some degree of similarity. (We shall discuss these points more fully in Section 4 below.)
4. FITTING PERCEPTION TO THE CONSUMER We need not dwell further on the actual constitution of autonomic epistemic responses. What they are in any given organism is something for psychologists to discover. The point that is important for TMS Theory is merely that there are such autonomic responses. I want now to return to the question posed at the beginning. When objects are assigned to a single sensory class, what feature are they represented as possessing? What, for instance, is the feature attributed to something by the experience of it as blue? I said that when two things look the same (up to some degree of similarity), they will occasion the same epistemic response. Return to the phenomenon of habituation. Whenever a presentation of blue occurs, the ‘surprise value’ of blue is reduced. In other words, a blue-presentation strengthens the expectation that blue occurs in the present environment. After a certain number of repetitions, the strength of this expectation rises above a threshold value, and the perceiver no longer diverts its attention in order to look at new blue-presentations. On the other hand, a presentation of red has a different effect. Having experienced blue repeatedly, the perceiver is surprised by the sudden appearance of red, but would not have been surprised by another appearance of blue. These presentations are sorted by colour, and two are presentations of the same colour if they lead to the same response, and of different colours if they lead to different responses. This gives us a hint as to how we should go about answering the question posed. We ask: what is the principle of sorting that the perceptual system employs? What do two stimuli that evoke the same response share—in what respect are they alike? There is a tendency in the philosophy of perception to search for an answer to this question in the external world. A detector is supposed to
Teleosemantics and the Consumer
157
have states that accord with certain states of the environment. So it seems that the right way to approach the question is to find out what states of the environment detector states are supposed to accord with. The question we are asking demands more than that we establish a correlation between detector states and environmental states. What we are looking for here is a theory of function within which to embed any such correlation. One popular theory about colour perception is that it is supposed to detect surface spectral reflectance (cf. Hilbert 1987, 1992; Matthen 1988). The idea is that the colour vision system sorts things by surface spectral reflectance, and that an organism has colour vision just in case it is able to discriminate such reflectances. This is an organism-independent way of characterizing sensory feature. Having looked at the real-world functional correlate of colour perception states, we arrive at a characterization of colours in purely physical terms. Call this the ‘upstream’ approach: it looks upstream from the perceiver to find the meaning of perceptual representations. The upstream approach neglects the consumer. Every detector function is associated with effector functions within the same organism. As we saw in the previous section, perceptual functions are associated with epistemic functions. These epistemic functions result in new representations, which are associated with various further effector functions, and so on. All these functions together form a system that serve the organism in its interactions with the environment. Consider the detector system from this point of view. It will work well if it sorts things together when they are ‘supposed to’ be treated the same for these epistemic purposes, and sorts them differently, i.e. differentiates them, when they are supposed to be treated differently. To revert to Millikan’s insight about the consumer, perceptual classification is correct if it serves the organism’s effector functions; it is incorrect if it disrupts these functions. For example, the perceptual sorting function that serves habituation is supposed to group things together when it is appropriate to ignore a new presentation as not constituting news. This is the ‘downstream’ approach. It attends to the results of perceptual sorting activities, not to the stimuli that occasion these activities, for the meaning of perceptual experiences. This point of view is dictated by a simple fact. Biological function arises out of evolutionary history. The function of an organ is that feature of it that contributed positively to the selection history of the type of organism in which it occurs. In a system of organ functions like the one sketched above, functions must be coordinated. If a sensory system coclassifies things that need to be differentiated for the normal functioning of effector organs, the organism suffers. Conversely, if the organism develops activities that demand a different classification scheme than its detector organs provide, it will suffer. Early teleosemantic theories concentrated on
158
Mohan Matthen
how perceptual states evolved to correspond with environmental states. It is more to the point to ask why one set of environmental states is singled out rather than another. This question is answered by the following thesis: The Coevolution Thesis. Perceptual systems coevolve with effector systems. Their function is to provide effector systems with information specific to the performance of the behaviours produced by the effector systems. Though it may be true that a particular colour vision system is supposed to detect surface reflectances, why it settled on surface reflectances and how it carves up the domain of surface reflectances depends on the epistemic effector it serves. Ask the downstream question first: how must stimuli be classified if success in epistemic effector activities is to be enabled? The answer could well turn out to be: stimuli must be classified according to their surface spectral reflectance. However, this answer would be subordinate to the downstream question in much the same way as extension is subordinate to intension. The Coevolution Thesis is illustrated by a recent trend in theories of the evolution of colour vision in primates. The central idea is that primate colour vision evolved so that certain kinds of edible vegetation would become more conspicuous in tropical forests. (For a review, see Surridge et al. 2003.) It is not agreed among these scientists exactly what food is involved here: some think it is a certain kind of fruit, others that it is a certain kind of leaf. The important point for our present purposes is that the function of colour vision in these primates is not to maximize our discrimination of surface reflectances, but to maximize the perceived difference between the food and the background vegetation in tropical forests so as to make the search for food more effective. Notice that this way of understanding primate colour vision looks beyond a broad similarity among colour vision systems in different animal phyla, i.e. that they are all directed towards surface reflectance. It treats this broad similarity as merely the background to concentrating on a particular problem special to the ancestral primate environment. Thus, it potentially explains more of primate colour vision than the broader approach can. It explains not only why primates detect reflectance, but also why they sharply distinguish green from red, possibly at the cost of distinguishing among greens and among reds. A salmonid fish would not necessarily make the same discriminations among surface reflectances, and the coevolution thesis—the downstream approach—tells us why. One consequence of this evolutionary history in our present dispositions could be this: we become habituated to series of colour-presentations that differed from one another but which all fell into the general category
Teleosemantics and the Consumer
159
of background vegetation, but be maximally surprised after such habituation by something that fell into the general colour-category of food (and vice versa). In other words, the degree of similarity experienced depends on this historically important ecological problem: whether things are the colour of food or the colour of background vegetation. This not only validates the theory that colour is reflectance, but explains and elaborates it. Why and how are reflectances sorted? By the similarity relation required for the identification of food. (It is a central part of these recent theories that primate food itself evolved so as to take advantage of primate perceptual discriminations: remember that it is an advantage to a plant to have its fruit etc. eaten, since this aids reproduction. But this is not very important here.) Now, given the specificity of primate evolution, it is not surprising that different animals sort things differently, even when they are sorting by ‘colour’, i.e. by wavelength-sensitive discrimination. Birds are specialized to use colour for aerial navigation; honey-bees to find pollen in flowers (not leaves or fruit); fish to discriminate objects in water-filtered sunlight. The emphasis on the consumer—the various systems that use detector functions to serve the interests of the organisms—is necessary not to solve the Problem of Multiple Responses, but in an informative specification of why certain features in the real world are represented by perceptual states. All of these organisms will sort things differently by colour. Some will sort by wavelength-sensitive variables other than reflectance, but even when different organisms are sorting by reflectance, they will use different sorting principles. (See Matthen 1999 for evidence and discussion of this.)
5. HOW WE KNOW SENSE FEATURES The emphasis on the coevolution of the provider and the consumer of representations—i.e. of detector and effector systems within the same organism—clears up a well-known problem for the Teleosemantic Theory. In early versions of teleosemantics, psychological data was used to show that sense features were physically characterized kinds. For example, David Hilbert (1987, 1992) and I (Matthen 1988) argued that colours were surface spectral reflectances. (In light of the argument given in the previous section, the conclusion ought to have been that human colours were surface spectral reflectances grouped by a particular similarity metric.) Recently, however, it has been pointed out (by Boghossian and Velleman 1991 and BraddonMitchell and Jackson 1997) that we do not as naive perceivers know colours under this description. After all, people in ancient times knew colours well enough, just as well as the naive observer today knows them—and surface
160
Mohan Matthen
spectral reflectances had not even been discovered. Yet, it seems that we do know the colours. For sensory features of this sort, to experience them is to know them. So, it appears, the Teleosemantic Theory cannot account for the subjective content of sense features, i.e. for how they present themselves to us. (Actually, it is unclear why a naive perceiver should know a meta-semantic, as opposed to a semantic, fact about the meaning of experience: nevertheless, I will take the difficulty at face value. That is: we do have a naive grasp of colour and TMS Theory should explain this grasp. On the other hand, TMS Theory need not stumble on the fact that the naive colour-perceiver need know nothing of evolutionary theory.) Clearly, the consumer’s perspective helps with this difficulty. For instead of trying to find the definition of sense features upstream at the head of the process that starts with a thing out in the world, the consumer’s perspective looks downstream to the effector system for this definition. Now, one might say that the effector system possesses one kind of ‘knowledge’ of what a detector system’s determinations mean—it knows this in terms of its own response. We claimed at the end of Section 3 that the proper way to define a sense feature was in terms of sameness of response: blue is that feature of things that brings about the same response as some paradigm—the sky, perhaps—up to some degree of similarity. Since the epistemic response to sensory states can be assumed to be known—tacitly, but nonetheless completely—the problem of how we know sense features seems not to arise at all within this framework. The knowledge is instinctive and contained in how we respond to things. (Boghossian and Velleman, who attacked ‘physicalist’ theories of colour, could agree since a response-defined theory is not, by their lights, physicalist. Braddon-Mitchell and Jackson, however, were attacking teleosemantic theories as such, and they should find the present argument liberating.) In the next and final section, we examine the role of perceptual experience in making explicit the kind of tacit knowledge discussed in the present section.
6. JUSTIFYING THE REPRESENTATIONAL FRAMEWORK So far we have not given serious consideration to the central presupposition of the semantic approach, namely that perceptual experiences represent situations outside themselves, that they are semantic entities that refer, describe, and are either true or false. What justifies this assumption? As noted in Section 1, the concept of representation is naturally at home in the arena of human communication. How can we extend it to the activities of sub-personal systems? Here again, the consumer point of view comes to
Teleosemantics and the Consumer
161
our aid. For as we shall now see, it aids us in constructing an important structural analogy between acts that possess conventional meaning and perceptual experiences. We have been supposing that the sorting function of a given perceptual system coordinates with that of an epistemic effector system. The effector system has several actions within its repertoire, A1, A2, . . . , An, and these are appropriate when some target object has feature F1, F2, . . . , Fn. The job of the detector system is to sort objects in its field of operation into classes corresponding to these features. For example, let’s suppose that there are two actions available to the ‘habituation system’ (as one might call the mechanism controlling the baby’s behaviour described in Section 3 above)—A1 is the action of attending closely to a new stimulus and A2 is the action of ignoring it. Let’s suppose that after a series of blue-presentations, another blue object is presented. The system will be behaving correctly if it registers ‘Blue’, and ignores the new stimulus, or registers ‘Red’ and attends to it. Its action will disrupt what the organism is supposed to be doing if it registers ‘Blue’ and pays the new stimulus a great deal of resourcediverting attention, or registers ‘Red’ and ignores it. Notice that this error might well occur because in the odd lighting conditions that prevail, or, because the visual system has become fatigued, the new blue-presentation is not properly classified as blue. In order for the organism to prosper, then, two conditions must obtain: (a) The perceptual system must accurately determine which of Blue or Red obtains, and (b) The effector system must reliably perform Ignore if Blue obtains and Attend if Red obtains. The coevolution thesis stipulates that the action-routines above were shaped in conjunction with the sameness or difference of the new stimulus with respect to those presented before. Thus, the question ‘To what real world feature does blue correspond?’ is answered in the first instance by determining what the Ignore routine is supposed to accomplish. Once this is done, one can infer what relation of similarity will optimally trigger the above routines. This is the relation that defines the sense features in question. One can, of course, try to determine what physical relation of similarity this corresponds to, but, as I argued earlier, this is a subordinate question. An alternative way of putting this point runs as follows. The equivalence relation that defines sense features is: the system is supposed to treat x the same way as y for purposes of initiating effector functions. If two objects are ‘supposed to’ evoke the same response—more ponderously, if the evolutionary function of the effector system will be best served by these two objects being treated the same way—then they belong to the same detector category. If
162
Mohan Matthen
they are supposed to be treated differently, they belong to different detector categories. Thus, similarity is that relation to previous presentations that merits the functionally appropriate reaction Ignore. (One assumes, on the basis of the psychological evidence, that sameness is assessed on the basis of sensory similarity—indeed, habituation is often used as a measure of sensory similarity—but for present purposes, this is unimportant.) Note here that this allows for the possibility of error. The detector system we have been discussing may well classify a newly presented object as similar to the ones presented before, when in fact the appropriate response would have been to treat it as dissimilar. As Millikan urged, this would disrupt the downstream activity. This raises a further question. It is not enough that the detector system should determine which of the above circumstances obtains. It is also necessary to ensure that the effector system does the right thing. Otherwise, we are no further ahead: the effector system would be on its own as far as ‘deciding’ what to do is concerned. Now, in some such cases, the detector system will simply do something that forces the appropriate operation of the effector. It may, for instance, secrete an enzyme, or give off a molecule that gets the effector system going. Let’s call this a coercive signal: it simply causes the requisite action to happen. A coercive signal is what Millikan calls a pushmi-pullyu representation—which is insightful, except that (as I shall now argue) such ‘representations’ are not really semantic in character. The phenomena cited in Section 2 in connection with the Problem of Multiple Responses indicate a situation in which such a coercive signal might be problematic. For, as we saw, sensory output is subjected to an elaborate process of cognitive assessment. This suggests that the sensory determination has to enter into a number of interacting epistemic routines. Coercive signals have to be causally appropriate to each different effector system. A signal that fed into a number of different systems would therefore have to be moulded to have the appropriate effect on all of these. This is difficult to arrange. Thus, it is unlikely that the same coercive signal could be used for a number of different interacting systems. What is required therefore is a non-coercive signal, one that a number of different epistemic systems can take as an indication of what the sensory system has determined to be the case. The detector system has, as it were, to say ‘I have determined that such-and-such is the case’, and then leave the effector system to handle this information as it will. (The effector system’s action may be quite deterministic, but it is not forced by a coercive signal issued by the sensory system.) There is a classic view of consciousness that fits the notion of a non-coercive signal well. Cognitive scientists plausibly
Teleosemantics and the Consumer
163
think that sensory consciousness is a monitoring device that allows the perceiver to know about the states of her own perceptual systems. This is just another way of saying that sensory consciousness, or perceptual experience, makes the determination of sensory systems available to the perceiver. The news of what the sensory system has determined is simply posted in the form of a characteristic form of experience, and the perceiver determines how to use it in epistemic operations and rational decision-making. David Lewis (1969, ch. 4) and Bryan Skyrms (1996, ch. 5) each have seminal accounts of how meaning emerges in non-coercive communicative situations like this. Lewis considers a plan arrived at by a sexton and Paul Revere. Simplifying somewhat, the sexton looks out for one of two situations: (R1) The redcoats set out by land. (R2) The redcoats set out by sea. Once the sexton has ascertained which of the two situations has occurred, he does one of two things: (S1) Hang one lantern in the belfry. (S2) Hang two lanterns in the belfry. Revere for his part looks to see how many lanterns are displayed in the belfry. Depending on how many he sees, he performs one of two actions. (A1) Warn the countryside that R1 has occurred. (A2) Warn the countryside that R2 has occurred. Now, clearly the sexton can adopt one of two action plans. They are: (X1) If R1 then S1, and if R2 then S2. (X2) If R1 then S2, and if R2 then S1. Similarly, Revere can adopt one of two action plans: (V1) If S1 has been executed, then A1, and if S2 has been executed, then A2. (V2) If S1 has been executed, then A2, and if S2 has been executed, then A1. Given that both the sexton and Revere want it to be the case that A1 if and only if R1 and A2 if and only if R2—this is the background coordination problem—they must coordinate action plans. We can call this the Signalling Coordination Problem. Either the sexton should adopt X1, and Revere should adopt V1, or the sexton should adopt X2, and Revere, V2. When one of these combinations of action plans is
164
Mohan Matthen
achieved, Lewis says, S1 and S2 are signals. In Lewis’s case, the requisite combination is achieved by agreement between the parties. But Skyrms (1996) shows that, under natural selection, coordinated action plans have an ‘attractive force’ of their own, and there is no need for extrinsic acts of agreement. As he (1996: 103) says: ‘Signaling system equilibria . . . must emerge in the games of common interest that Lewis originally considered.’ Like Lewis, Skyrms was considering signals between organisms that have an interest in achieving a coordinated signalling system. Here, we are considering a signal coordination problem between subsystems of a single organism. Since the subsystems of a single organism perish or prosper according to whether the organism does, the commonality of ‘interest’ is guaranteed. If coordination is not achieved, the organism will be less fit, and consequently both the detector and the effector will perish, regardless of how effective they may be considered in themselves. Notice that if there is to be a possibility of signalling in such a situation, the sexton has to have at least as many putative signals available to him as there are circumstances that demand different actions on the part of Revere. But neither cares at all which communicative action plan is adopted, as long as the mutually desired result ensues. It follows that there is always a choice as to which signal is associated with which circumstance. Thus, the association between signal and circumstance is a matter of convention: there is always a choice among possible association schemes, and nothing matters other than that both the sexton and Revere agree on which signal is to be used in which set of circumstances. Earlier, I stipulated that there were n features that could be attributed to a stimulus x, and correspondingly n actions, each one appropriate when the corresponding feature is detected. We now see that the sensory system would need n non-coercive signals to inform the effector organs of which of these features is detected. The important point to note is that, as Lewis and Skyrms demonstrate, it does not matter which signal was associated with which feature, so long as the actions taken by the epistemic effector systems coordinated properly with the signalling code chosen by the sensory detector system. A coercive signal has to be chosen for its effects. If I am going to ensure that you do the right thing by directly manipulating your body, then I must choose my actions in such a way as to achieve this end. With a non-coercive signal, all that matters is that my action plan coordinates with yours. Thus, which signal stands for which determination of the sensory system is contingent and historical. What matters in the case of a non-coercive signal is that the ‘speaker’ and the ‘hearer’ should coordinate their attitudes. Where are we now? In Section 4, I proposed that sensory systems coevolve with effector systems. Here I am adding a further wrinkle to that tale.
Teleosemantics and the Consumer
165
The claim is that under certain circumstances, the effector function is such that, instead of direct manipulation, a non-coercive signal of the detector state needs to be sent from detector to effector. Perceptual experience is such a signal. The analysis of these signalling problems by Lewis and Skyrms demonstrates that the meaning of a given experience is determined by an arbitrary coordination scheme which emerges as a part of the coevolution of detector and effector systems. In these circumstances, it is a matter of convention and history which signal is associated with which circumstance for the purposes of communication between systems. It is a convention in the sense that (a) there is a plurality of coordination equilibria, and (b) natural selection does not ‘prefer’ one to another. This is the sense in which we can take perceptual experience to be a representation with conventional meaning. REFERENCES B, P, and V, D (1991), ‘Physicalist Theories of Color’, Philosophical Review, 97: 67–106. B-M, D, and J, F (1997), ‘The Teleological Theory of Content’, Australasian Journal of Philosophy, 75: 474–89. D, F I. (1988), Explaining Behavior: Reasons in a World of Causes (Cambridge, Mass.: Bradford Books, MIT Press). H, D R. (1987), Color and Color Perception, CSLI Lecture Notes 9 (Stanford, Calif.: CSLI Publications). (1992), ‘What is Color Vision?’, Philosophical Studies, 68: 351–70. F, J A. (1982), The Modularity of Mind (Cambridge, Mass.: Bradford Books, MIT Press). K, D (1989), ‘Afterthoughts’, in J. Almog, J. Perry, and H. Wettstein (eds.), Themes from Kaplan (New York: Oxford University Press). L, D (1969), Convention: A Philosophical Study (Cambridge, Mass.: Harvard University Press). M, M (1988), ‘Biological Functions and Perceptual Content’, Journal of Philosophy, 85: 5–27. (1989), ‘Intensionality and Perception: A Reply to Rosenberg’, Journal of Philosophy, 86: 727–33. (1999), ‘The Disunity of Color’, Philosophical Review, 108: 47–84. (2005), Seeing, Doing, and Knowing: A Philosophical Theory of Sense Perception (Oxford: Clarendon Press). M, R G. (1989/1993), ‘Biosemantics’, Journal of Philosophy, 86: 281–97; repr. in Millikan, White Queen Psychology and Other Essays for Alice (Cambridge, Mass.: Bradford Books, MIT Press). (1995), ‘Pushmi-Pullyu Representations’, Philosophical Perspectives, 9: 185–200.
166
Mohan Matthen
Q, W. v. O. (1969), ‘Natural Kinds’, in Quine, Ontological Relativity and Other Essays (New York: Columbia University Press). S, B (1996), Evolution of the Social Contract (Cambridge: Cambridge University Press). S, A K., O, D, and M, N I. (2003), ‘Evolution and Selection of Trichromatic Vision in Primates’, Trends in Ecology and Evolution, 18: 198–205.
8 Content for Cognitive Science Karen Neander 1. INTRODUCTION I see some newspaper blown by the wind as a cat slinking, and thus I represent the newspaper as a cat. Three things are involved: (i) a representation, presumably some neural event, (ii) its target, in this case the newspaper, and (iii) the content of the representation, cat slinking. In this case, the representation misrepresents its target because there is a mismatch between its target and its content. A philosophical theory of mental content is principally concerned with the relation between items of the first and third kind. Such a theory tries to answer the question: in virtue of what does a mental representation have the content it has?¹ An obvious desideratum for such a theory of content is that it gets the contents of representations right. It’s easy to give a theory of mental content that ascribes some content to mental representations. There’s the Today is Tuesday Theory, for example, which says that all of our brain states have the content Today is Tuesday.² This allows for misrepresentation because it entails that all of our brain states are wrong six days of the week. However, it is a terrible theory because it gets the contents of very few mental representations right. What we want is a theory that entails that we are thinking that today is Tuesday only if we are thinking that today is Tuesday, and that entails that we are thinking that the plants need watering if we are thinking that the plants need watering instead. While this desideratum is obvious, it is surprisingly difficult to apply. Consider the notorious case of the frog. A normal frog will snap at anything that’s moving and suitably small and contrastive with its background (for short, at anything small, dark, and moving). At least as philosophers tell the ¹ As I am using the term, a mental representation need not be conscious, or part of a conscious mental state. ² I owe this nice example to Barry Loewer.
168
Karen Neander
tale, in their natural habitat the small, dark, moving things are mostly flies that are nutritious for frogs. There has been debate about whether this or that theory of mental content generates suitable content in this case, and yet—and this is one of the most frustrating things for those fresh to the debate—opinions differ as to what the correct content is. In the case of the frog, philosophers have variously argued that the content of its visual representation is fly; frog food; a parcel of chemicals nutritious for frogs; something small, dark, and moving; small, dark, moving food ; or something indeterminate between these.³ How can we use simple system cases to test our theories if we cannot agree on what the content is? What we need is some independent ground for believing one content ascription rather than another. This chapter tries to provide such a ground. Here I argue that some candidate contents serve the purposes of mainstream cognitive science better than others do, mainstream cognitive science being understood as that science that uses an information-processing approach to provide operational explanations of cognitive capacities. I claim that some candidate contents can and some cannot play a role in such explanations. This is a contentious beginning but I am content to make my conclusion conditional on the assumption that a theory of content should try to meet the needs of mainstream cognitive science. Subject to this condition, if my argument here is along the right lines, we will have a good reason to reject standard teleological theories, such as Ruth Millikan’s (1991), as well as some non-standard ones, such as Kim Sterelny’s (1990) and Nicholas Agar’s (1993). On the positive side, we will also have good reason to take another look at informational theories like those offered by Jerry Fodor (1990b), Fred Dretske (1994), Pierre Jacob (1997), and Neander (1995, forthcoming).⁴ Note that the last three of these are teleological theories of mental content, so this chapter should not be construed as an argument against all teleological theories. In what follows I switch from frogs to toads. The perceptual systems of frogs and toads are very similar and the neuroethological literature on the two overlaps to a great extent, but toads let me make my point a little more vividly. ³ Jerry Fodor (1990a) and Kim Sterelny (1990) say it represents its target as a fly. Ruth Millikan (1991) says it represents it as frog food and Carolyn Price (1998, 2001) says it represents it as a parcel of chemicals nutritious for frogs. Fodor (1990b: 106), who has changed his mind, along with Fred Dretske (1994), who has also changed his, Pierre Jacob (1997), and I (Neander 1995) say it represents it as something small, dark, and moving. Nicholas Agar (1993) says it is small, dark, moving food. And Dretske (1986), David Papineau (1998), and Daniel Dennett (1995) suggest that the content is indeterminate between these things. This is a very incomplete list of those who have participated in this debate, but it is representative. ⁴ This chapter is from my forthcoming book Mental Representation: The Natural and the Normative in a Darwinian World (MIT Press).
Content for Cognitive Science
169
I hope readers will enjoy or at least endure with patience the short excursion into toad neuroethology that takes place in Section 2. I also hope that even those who disagree with the implications I draw from it in Section 3 will find the issues interesting. The smaller goal is to argue for a particular content ascription in a particular case, but the larger goal is obviously more important. It is to illustrate the way in which content ascriptions should cohere with neuroethological analyses of relevant cognitive capacities. As for the relation between neuroethology and cognitive science, the two are continuous. The relevant domain of study for frogs and toads is referred to as ‘‘neuroethology’’ not ‘‘cognitive science’’ since the latter applies mostly to the study of our own species. However, neuroethology does the same sort of thing for other animals as cognitive science does for us; it studies such things as perception, motor control, and decision making in non-human animals, including primates. Its aims and methodological tools are also much the same despite some obvious differences, such as in the ethical constraints that scientists feel themselves to be under and the fact that verbal responses by research subjects have an important place in one and none in the other. In both cases, the aim is to understand the normal flow and transformation of information and its neural substrate. Sometimes explicitly computational models are developed in both cases (e.g. compare Marr 1982 and Cobas and Arbib 1992). Moreover, biologists believe that much of what they have learned about the anuran (frog and toad) nervous system applies—and many of the concepts developed in studying them are applicable—to a wide range of vertebrate species (Ewert et al. 1983: 414). Given this substantial and methodological continuity, I sometimes use ‘‘cognitive science’’ to refer to both cognitive science and neuroethology, generically.
2. TOAD NEUROETHOLOGY Elsewhere (Neander forthcoming) I explain that there is a prima facie reason to think that frogs and toads have mental representations. I argue that if we assume, uncontroversially, that their brain states have the function of carrying information, then they have intensional states in so far as they are not extensional. That is, sentences describing these states guarantee neither existential generalization nor the preservation of truth-value under the substitution of co-referring terms. This creates a presumption that mental representations are involved, which I see no reason to override. So in my view neither the frog nor the toad should be considered merely toy examples. From conversations with philosophers, I have the impression that many think of anuran ‘‘prey’’-recognition as mere transduction, but this is wrong. The it’s-merely-transduction view might have seemed justified in the light
170
Karen Neander
of the seminal paper by Jerome Y. Lettvin and his colleagues (1959), ‘What the Frog’s Eye Tells the Frog’s Brain’, which first sparked philosophical interest in the frog. But this is an oversimplification even of this paper. Lettvin et al.’s claim was that ‘‘prey’’ discrimination occurred in retinal cells, but anuran retinal cells are more complex than mammalian ones and the relevant process was not thought to be mere transduction (even mammalian retinal cells do more than mere transduction). In any case, more recent research has undermined Lettvin et al.’s claim. It turns out that further information processing involving mid-brain structures is required for ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination. Five decades of intensive research further on, biologists are still trying to unravel the complexities. Some philosophers complain that the frog example is ‘numbingly familiar’ but my sympathy with this is tempered by the fact that we have maintained an impressive collective ignorance about the real live case. Besides, both the frog and the toad are genuinely excellent subjects for our purposes. As I have already remarked, the anuran brain is similar in terms of broad principles to those of many other vertebrate species. Also, while it is certainly very complex, it is nonetheless relatively simple compared to other vertebrate brains, and so it is easier to understand. In addition, frogs and toads have been intensively studied—they are the amphibian equivalent of Drosophila. As a result, neuroethologists have a more complete understanding of their nervous systems than they have of most other vertebrate nervous systems (Ewert et al. 1983: 413).
2.1. Sign-Stimuli and Prey-Capture in the Toad So here goes. Some facts. Before we enter the brain, there are some things that can be observed from the outside. The first is that neither frogs nor toads feed only on flies. Lettvin et al.’s (1959) classic paper was on leopard frogs (frogs of the Rana pipiens cluster) and adult leopard frogs are not fussy eaters. In their natural environment they eat a variety of insects, including sowbugs, spiders, damselflies, crickets, leafbugs, spittlebugs, and short-horned grasshoppers. So the content ascription—flies —was never plausible given their real diet. Adult toads are not choosy either. Different toads have different diets but they generally eat a variety of things, such as beetles, bugs, millipedes, slugs, and earthworms. Big toads also hunt larger creatures, such as snakes, small birds, and even frogs. So, for toads, no content ascription that singles out a particular prey species—worms, say—would be suitable either. Those who fancy contents of this sort (e.g. Sterelny 1990) need to think in terms of contents that contain a long
Content for Cognitive Science
171
list of prey-species, or else (better) a more generic content, such as member of one of the toad’s prey-species. Here we will be interested in the toad’s visual representation of its prey and in its perceptual content. One and the same representation might possess motor and/or motivational content as well as perceptual content, but since I only look at the toad’s visual system I only discuss perceptual content here. Note also that prey catching can be triggered by tactile as well as visual stimulation, but this is under the control of a different neural pathway and I do not discuss it further. Normal adult toads can see stationary objects. They don’t splat blindly into walls and tree trunks. But again this is under the control of a different pathway. The toad’s visually induced behavioral responses to moving objects reveals that it can discriminate between at least three kinds of moving objects, which (without prejudging the outcome of this discussion) we can for convenience refer to as ‘‘prey’’, ‘‘predator’’, and ‘‘other’’ (to remind us not to prejudge the outcome, I’ll use scare quotes around these terms in what follows). The typical response to these is in brief to try to catch them, avoid them, and ignore them, respectively. A toad’s prey-catching behavior is affected by its motivational state, which can vary according to season, time of day, and how much it has eaten. Sated toads stop hunting, and it is thought that escaping predators might always have precedence over hunting. There’s also a preferencestructure for selecting a prey to hunt (or, for that matter, a predator to avoid) if more than one prey (or predator) is detected, but this is a complexity I shan’t pursue (Cobas and Arbib 1992 attempt to model it). The adult capacity to distinguish between ‘‘prey’’, ‘‘predator’’, and ‘‘other’’ is innate and differs from what’s found in tadpoles, which are vegetarian. Newly metamorphosed toads that have never before been exposed to prey-like or predator-like stimuli can nonetheless perform ‘‘prey’’/ ‘‘predator’’/‘‘other’’ discrimination. And they can do this even if as tadpoles they were raised in a completely homogeneous environment. However, accuracy (e.g. in judging distance) improves with practice and the full behavioral repertoire of adults develops over several weeks or more. Finally, toads can become habituated to repeated dummy stimuli such as moving dots on a computer screen that reappear at the same location or a looming cardboard square that looms in the same location one too many times. There’s also evidence that individual experience can affect prey selection with respect to surface features of stimuli, like dots and stripes (Ewert and Kehl 1978). It seems that toads can learn to avoid bees and bombardier beetles (Cott 1936; Brower and Brower 1962). Plus
172
Karen Neander
positive conditioning can affect their responses: if, for instance, the odor of mealworms accompanies feeding for a time it can subsequently strengthen their prey-catching response and even override what would otherwise constitute non-prey-like (‘‘other’’) features. What follows concerns their capacity to distinguish ‘‘prey’’, prior to any such conditioning. David Papineau (1998: 5 n. 1) wonders if frogs and the like have a belief–desire psychological structure, and suggests that creatures lack determinate representational content if they lack a belief–desire psychological structure. Whether toads have a belief–desire psychological structure depends on how demanding the notions of belief and desire are, but it’s worth noting in passing that toads have motivational states and therefore states that have a desire-like direction of fit, and they also have informational states and therefore states that have a belief-like direction of fit. In other words, they have states that were designed to tell them what conditions obtain and they have states that were designed to cause certain conditions to obtain. Toad behavior is also somewhat flexible, even aside from its modest learning potential, in the sense that it’s modified in appropriate ways on the basis of different informational and motivational states. Toads respond to large looming predator-like stimuli by a range of behaviors that biologists describe as sidestepping, ducking, puffing up, rising up stiff-legged, excreting toxic oils, and turning, crawling, or leaping away. In response to a prey-like stimulus, in contrast, researchers report that the toad displays a sequence of behavioral elements, which are said to consist typically in orienting (o) toward the stimulus, stalking or approaching it (a), fixating or viewing the prey from front on (f ), and snapping at it (s) by lunging and/or extending its tongue and/or snapping its jaw. There is some flexibility in how these behavioral elements are combined. The toad might simply snap if the prey is in front and within range, and orienting and approaching can be left out or repeated as often as required. So we might have sequences such as f -s, o-f -s, o-a-a-f -s, o-o-a-o-a-o-o-a-o-o-a-a-f -a-f -s, and so on. Usually, a toad’s response to a stimulus deemed prey-like involves an initial orienting toward it, unless the toad is already so oriented. A normal toad does not orient toward predator-like stimuli or (prior to conditioning) other-like stimuli. But if a toad sees a prey-like stimulus move out from behind a barrier it first moves around the barrier, which can involve turning away from the prey. If the toad is placed in a glass dome and a prey-like item is rotated horizontally at a constant distance around the toad the sequence is o-o-o-o-o-o-oo-o . . . and so on, until the toad habituates, which takes about sixty seconds. This rotating procedure has been used in some experiments (that I describe
Content for Cognitive Science
173
below) to gauge the extent to which a stimulus counts as prey-like for the toad. The more turns a motivated toad makes in a thirty-second interval, the more the stimulus is considered prey-like for the toad. Some parts of the entire prey-capture sequence are classified as fixedaction patterns. That’s to say that once they have begun they cannot be modified in the light of further information. For instance, if a dummy prey disappears after a critical point in the fixation phase, the toad still snaps and gulps and, as the neuroethologists report, often licks its mouth in seeming satisfaction. As ethologists use the term, the sign stimuli for an innate releasing mechanism for a fixed-action pattern are those features of the environment that release or trigger the relevant behavior. The sign stimuli for an innate releasing mechanism can be ascertained by purely behavioral studies through the use of dummy stimuli and the careful variation of variables, a practice that goes back to the famous studies of Konrad Lorenz and Niko Tinbergen in the first half of the twentieth century. Behavioral studies show that the toad distinguishes ‘‘prey’’ from ‘‘predator’’ and ‘‘other’’ by some quite specific features of a moving stimulus. The range of the relevant dimensions and the preference curves differ from species to species; there are hundreds of species of toad. However, the story is much the same across the range. In a famous series of studies (summarized in Ewert et al. 1983) Jorg-Peter Ewert and his colleagues used a variety of dummy stimuli including cardboard cutouts with three distinct configurations. These consisted of (i) rectangles of constant width and varying lengths moved in a direction parallel to their longest axis, dubbed ‘‘worms’’, (ii) rectangles of constant width and varying lengths moved in a direction perpendicular to their longest axis, dubbed ‘‘anti-worms’’, and (iii) squares of different sizes, dubbed ‘‘squares’’. In brief, as shown in Figure 8.1, the ‘‘worms’’ provoke prey-capture behavior, although ‘‘anti-worms’’ of the same dimensions are ignored, and the ‘‘squares’’ produce prey-capture behavior if they are small enough and avoidance behavior when they are larger. As you can see, the categories ‘‘worm’’, ‘‘anti-worm’’, and ‘‘square’’ do not quite correspond with the categories ‘‘prey’’, ‘‘other’’, and ‘‘predator’’. However, a toad’s prey tends to be worm-like. That is, it tends to be within certain size parameters and moving in a direction that parallels its longest axis. Again, I use scare quotes to remind us that not all ‘‘worms’’ are real worms and not all real worms are ‘‘worms’’. ‘‘Worms’’ can be crickets or millipedes, or cardboard cutouts, and a real worm can be an ‘‘anti-worm’’ if, for example, it is stunned, hung by its tail, and moved perpendicular to its longest axis. Visual ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination is not affected by features of the stimuli that are not captured by these dummy stimuli. For
174
Karen Neander
Number of turns per minute
Behavior Worm
40
s Square
s s
20 Antiworm 0
Worm
s Antiworm
Square
4 8 16 32 2 Length s (degrees)
Figure 8.1. The toad’s behavioral response to worms, anti-worms, and squares, measured in turns per minute, varies with the length s of each kind of stimulus. After Ewert (1980)
example, changing the velocity of the stimuli makes no difference, nor does changing the style of motion from wriggling to scuttling. A minor qualification is that some changes, such as in the direction of contrast from a black stimulus against a white background to a white stimulus against a black background, can affect acuity and maximum size preference. Neuroethologists stress that this is not a case of mere feature detection, let alone mere transduction. The shape of the stimulus is relevant and so is the direction of motion, but the toad’s response cannot be understood as a response to the mere summation of these two. Two items with the same shape can produce no response or an enthusiastic response, depending on the direction of motion, and two items with the same direction of motion can produce no response or an enthusiastic response, depending on the shape. The response is provoked by what is called a ‘‘configural feature’’ (or sometimes a ‘‘gestalt’’): here, motion relative to shape. Clearly, we can learn the sign stimuli for toad prey-capture behavior before we understand the relevant information processing or the neural substrate.⁵ This is important because unless the sign stimuli are first identified any attempt to develop an adequate information-processing ⁵ This creates problems for Price’s (1998, 2001) solution to the functional indeterminacy problem(s). It conflicts with her claim that her ‘abstraction condition’ rules in favor of the content toad food (she is discussing frog food, but extending the same reasoning to toads, we get toad food). According to Price, when we determine ‘the unique correct description’ of the function of a device and hence, on her view, the description that determines the contents of its representations, we must not consider the internal design of the device or that of the collaborating systems. Where ‘d’ is the relevant device, Price says, ‘we do not need to know how d’s fellow components make their contribution. Nor do we need to know about the design of the device itself ’ (1998: 70). This doesn’t favor toad food over something more like item moving in a direction that is parallel to its longest axis because we don’t need to know how the detection device or its fellow components work (in the sense of knowing anything about their underlying structural
Content for Cognitive Science
175
account of the toad’s capacity or to uncover the relevant neural substrate of that capacity will fail. Ewert et al. (1983: 415) report that some early studies failed for that reason. An adequate model of perceptual processing cannot be developed until we know what information is extracted from the retina. And, as Ewert et al. (1983: 415) comment, when neurobiologists try to uncover the neural substrate of a capacity, such as ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination in the toad, the basic strategy is to look for neurons or clusters of neurons that have the same stimulus preferences as the creature, in this case, the motivated toad. If researchers lack an accurate idea of what the stimulus preferences are, they do not know what to look for. When looking for a needle in a haystack it helps at least to know that it’s a needle you’re looking for! We’ll have cause to return to these points later.
2.2. Information Flow and the Neural Substrate We turn now to the nature of the neural substrate and the information processing it performs. How does the toad distinguish prey-like stimuli? Although toad brains are very simple compared to ours, they are, as I remarked before, still highly complex systems, and as a consequence the answer remains imperfectly understood. However, even an incomplete sketch of what’s known might suffice to show that some candidate contents for the toad’s ‘‘prey’’-representation seem gratuitous from a neuroethological perspective. Of course, the processing of visual information begins with the retina, where receptor cells (rods and cones) transduce light into neural firings. The optic nerve, which mediates between the retinas and mid-brain structures, contains retinal ganglion cells. The receptive field of a retinal ganglion cell is the area of space from which light (or its absence) can affect it. In the toad the retinal ganglion cells have receptive fields composed of two concentric fields. Those that appear to respond differentially to ‘‘worms’’, ‘‘anti-worms’’, and ‘‘squares’’ are the R2, R3, and R4 cells, which have an excitatory inner center and an inhibitory outer surround, meaning that stimulation of the inner region excites the cell whereas stimulation of the outer region inhibits it. (In some retinal ganglion cells this inhibitory and excitatory organization is reversed.) R2, R3, and R4 cells differ with respect to the size of their excitatory receptive fields (ERFs), the strength of their inhibitory receptive fields (IRFs), and the kind of stimulus that excites the center and inhibits the or functional design) in order to know the sign stimuli that trigger the prey-catching response. We can discover them while treating the frog or toad as a black box.
176
Karen Neander
Table 8.1. A comparison of three classes of retinal ganglion cells in a typical toad. Neuronal class R2 R3 R4
Approx. ERF diameter (deg.) 4 8 16
IRF strength
Preferred stimulus
+++ ++ +
Dimming Dimming or brightening Brightening
Source: Adapted from Ewert et al. (1983: 444).
surround. In the common toad, R2s have the smallest receptive fields, with center diameters of 4◦ . They are primarily ‘‘off-center’’, meaning that their centers respond best to a dimming of light. R3s have somewhat larger receptive fields, with center diameters of 8◦ , and they have ‘‘on–offcenters’’, meaning that their centers respond well to either dimming or brightening. R4s have the largest receptive fields, with centers that have diameters of 16◦ and they respond best to an increase of light. Some of these properties are compared in Table 8.1. Since the surround inhibits the cell, each cell responds best when the entire center is stimulated and none of the surround is. Thus, an R2 cell will respond most strongly when a dark circle entirely fills its center’s receptive field. These cells can thus provide information about changes in illumination in the visual field, which can be used to extract information about the size, shape, and motion of moving stimuli. The response patterns of R2, R3, and R4 retinal ganglion cells to the ‘‘worm’’, ‘‘anti-worm’’, and ‘‘square’’ configurations are given in Figure 8.2. If we compare these responses to the behavioral response patterns of the toad, shown in Figure 8.1, we can see that none of these retinal ganglion cells have excitation patterns that mirror the response of a motivated toad. For instance, the behavioral response increases with the length of the wormlike stimuli, but none of the retinal ganglion cells show the same preference. So neurobiologists conclude that ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination requires further processing beyond that performed by the retinal ganglion cells. The retinal ganglion cells primarily extend to the optic tectum (T), a mid-brain structure, and also to the thalamic pretectum (TH), in addition to several other neural structures. Different kinds of retinal ganglion cells go to different layers of the optic tectum. There’s also a crossing over, with retinal ganglion cells from the right eye crossing to the left tectum and those from the left eye going to the right tectum. Neighborhood relations are preserved, which means that nearby retinal ganglion cells that record from nearby retinal receptors and thus from nearby regions of visual space project onto nearby parts of the tectal layers. In this sense, the tectum contains multiple maps of the visual field.
Content for Cognitive Science A. R2 cells
B. R3 cells
40 Impulses per second
177
40 Worm Worm
Antiworm
20 Square
0
2
4
8
Square
20
Antiworm
16
32
0
2
4
8
16
32
Length s (degrees) C. R4 cells
40
Square 20 Antiworm Worm 0 2
4
8
16
32
Figure 8.2. Responses of retinal ganglion cells to three kinds of moving stimuli. After Ewert (1980)
If the optic tectum is removed, visually induced prey-capture and predator-avoidance and indeed all responses to moving stimuli cease. Since the avoidance of stationary barriers remains intact it is understood that the latter is under the control of a different neural pathway. The optic tectum is involved in locating moving stimuli but it is thought to do more than merely locate them. It appears that distinct pathways within the tectum control the frog’s turning away from predators and turning toward prey because these capacities dissociate, which is just to say that they can be disrupted independently given sufficiently small lesions in the area (King and Comer 1996). In so far as the recognition of visual stimuli as prey-like can be localized, neurobiologists consider the activation of a certain class of cells in the optic tectum—the T5(2) cells—to be the best candidate. As some express it, activity in the T5(2) cells ‘reflects a good approximation of the probability that the stimulus fits the prey category’ (Ewert et al. 1983: 450). If you compare the activation pattern of these cells, as shown in Figure 8.3c, with
178
Karen Neander
Number of impulses per second
the response patterns of the motivated toad, as shown in Figure 8.1, you can see that the match is quite close. It is sufficiently close for neuroethologists to think that the T5(2) cells might be the ‘prey-recognition neurons’.⁶ The response pattern of these T5(2) cells is largely explained as a balance of the inputs from two other classes of cells, an excitatory input from another class of tectal cells, the T5(1) cells, and an inhibitory input from some thalamic pretectal cells, the TH3 cells. The activation patterns of TH3 and T5(1) cells are shown in Figures 8.3a and b. Electrode implantation and lesion experiments provide evidence that the TH3 cells have this inhibitory effect. Electrical stimulation of the TH3 cells reduces the response of T5(2) neurons. And if the thalamic pretectum is removed, the toad’s prey-capture response becomes disinhibited, so that the toad acts as if everything that moves is prey: it will then orient toward its own extremities, toward large predator-like cardboard squares, to the hand of the experimenter, and so on. Smaller lesions in the thalamic pretectum produce the same response with respect to smaller parts of the visual field. Neuroethologists conceive of this ensemble of TH3, T5(1), and T5(2) cells as something like an ANDgate (Ewert et al. 1983: 442), or as an AND-gate with a NOT-gate on one of the inputs (it does not have a discrete on–off character, but the analogy with computational components described as AND-gates is apt). A schematic presentation of the proposed operation of this ‘‘gate’’ is given in Figure 8.4. Other areas of the toad’s brain also mediate visual ‘‘prey’’/‘‘predator’’/ ‘‘other’’ discrimination. For example, cells in the telencephalon inhibit the 30
Square
30
20 10 0
Worm 2
4
8
Worm
Anti- 20 worm 10
16 32
(a) TH3 cells
30 Worm
Square
0
Antiworm
2
4
8
16 32
(b) T5(1) cells
20
Square
10 0
Antiworm 2
4
8
16 32
(c) T5(2) cells
Length s (degrees)
Figure 8.3. Responses of thalamic TH3 cells, tectal T5(1) cells, and tectal T5(2) cells. Only (c), the T5(2) cells, has the same pattern as the behavioral response of the toad, shown in Figure 8.1. After Ewert (1980)
⁶ The match is not perfect. One difference is that the maximally stimulating length of the worm-like stimulus is 16◦ for behavior and only 8◦ for the T5(2) cells (for further discussion of this point, see Camhi 1984: 230–7).
Content for Cognitive Science T5(1)
179
+ T5(2)
TH3 A.
T5(1) TH3
− + T5(2) −
B. T5(1) TH3
C. T5(1) TH3 D. T5(1) TH3
+ T5(2) −
+ T5(2) − + T5(2) −
Figure 8.4. The T5(1), TH3, T5(2) ensemble. The response of the T5(2) cells is primarily controlled by the inhibiting influence of the TH3 cells and the excitatory influence of the T5(1) cells. After Ewert (1980), from Carew (2000)
activity of the thalamus. Their removal in the poor toad results in hyperexcited visually induced escape behavior and eliminates visually induced prey-capture behavior altogether. However, as I say, the areas that are primarily responsible for visual ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination are the optic tectum and the thalamus. In sum, it is thought that visually induced prey-capture is primarily mediated by the optic tectum moderated by the thalamus (and that visually induced predator-avoidance is primarily mediated by the thalamus moderated by the optic tectum). 3. IDENTIFYING THE CORRECT CONTENT It is almost time to turn to the implications of the neuroethological findings for the content of the relevant representation. But first, what is the relevant representation? Philosophers have talked of the toad’s or frog’s perceptual representation, but what neurological state, event, or process has been at issue? As we have seen, neither the image on the retina nor
180
Karen Neander
the firing of the optic fibers (the retinal ganglion cells) turns out to be the ‘‘prey’’-representation. Our best bet to date is that ‘‘prey’’-recognition has only occurred once T5(2) excitation has occurred. For this reason, I shall speak of a high frequency of action potential activity in a T5(2) cell—hereafter ‘‘+T5(2)’’ or ‘‘T5(2) excitation’’—as the relevant representation. This is better than the usual vague or worse talk but it could be an oversimplification. There are a number of different features of neuronal events that could carry information. One is the average rate of firing of neurons, and this seems to be involved in this case. However, it might not be a case of local coding (i.e. a single cell for a single feature). Instead, a cluster of cells (e.g. with overlapping tuning curves) might be involved. Luckily, the issue does not substantially impact the main argument in this chapter. Only certain details would need changing. Assuming that the relevant representation is a +T5(2), the question is ‘‘What is the content of a +T5(2)’’? Before this case can be used to test our theories of mental content we need an independent basis for attributing one content to it rather than another. In Sections 3.1 and 3.2, I suggest that we need to observe the coherence constraint: if the contents of mental representations are to play a role in explaining cognitive capacities, they must cohere with the relevant information processing. At the risk of laboring the point, let me stress that this is not offered as a theory of mental content. Coherence is an intentional notion, but that’s fine and dandy here because I am not offering a theory of mental content at this point. I am offering a heuristic or a principle to be used in identifying contents in simple systems so that we can determine pre-theoretically (i.e. pre-philosophical-theory-ofcontent-ly) what the correct content is.
3.1. Localization Content It is easiest to start with the localization content of +T5(2)s. This has not been the subject of controversy in the philosophical literature. The controversy has focused on what a +T5(2) represents, not on where it represents it as being. However, in part for that very reason, the localization content can be a useful illustration of the coherence constraint. Recall that different retinal ganglion cells receive input from different retinal receptor cells, which have different receptive fields, and also that the different retinal ganglion cells extend to different cells within the relevant layer of the optic tectum, so that different T5(2) cells also have different receptive fields. Although the what-component of the information carried is the same for every T5(2) cell, the where-component can differ. In a normal toad, excitation of one T5(2) cell carries the information that there’s something (let’s just call it a whatsit for the moment) within its receptive
Content for Cognitive Science
181
field, whereas excitation of a different T5(2) cell carries the information that there is a whatsit within its receptive field. How precise is this localization information? As yet, I’ve not mentioned that the T5(2) cells are thought to be monocularly driven cells with relatively large receptive fields. If so, excitation of each single cell of the type normally reflects, as the neuroethologists say, the probability that a prey-like stimulus is present in a relatively large region of the visual field. These cells do not normally carry precise information about stimuli location on any dimension: left–right, up–down, near–far. They are precise enough for orienting toward prey but not for accurate snapping, and it is thought that more precise localization is carried out by different sets of tectal and thalamic neurons. One hypothesis is that another class of tectal cells, the T7 cells, has the function of extracting more precise information about location from the overlapping receptive fields of T5 neurons. Also (thalamic) TH6 and (tectal) T3 neurons, though monocularly driven, are apparently sensitive to motion in the near–far dimension, so they might provide information based on the changing size of the image on the retina (things cast a bigger image as they approach) or from disparities in the image when the toad turns (the closer the stimulus the more its retinal image moves). In addition, orienting brings the stimulus into the toad’s restricted binocular range and more precise information about locations could then be derived during fixation. T1(3) neurons are a candidate here since they are binocularly driven. The reasoning outlined in the preceding paragraph is the kind of reasoning that the neuroethologists use. Indeed, it is the reasoning they use, almost verbatim (Ewert et al. 1983: 455). It is also (more or less) the kind of reasoning that we should use when we look for independent (pre-philosophicaltheory-of-content) reasons for attributing contents. In line with this, I suggest that content ascriptions should, in general, be motivated from below by an understanding of the mechanism underlying the information processing and from above by the information-processing explanation of the relevant capacity. I elaborate on this in the remainder of this subsection. Let’s look at the localization content of a particular T5(2) cell. I’ll refer to its +T5(2)s as Rs, and to a particular token (instance) of them as r. Let L be the receptive field of this cell, and let L− be the location of the stimulus (inside L) that causes r to be tokened. Also let L+ be an area larger than and encompassing both L− and L. What is the localization content of r? Does it have the content that there is a whatsit at L, L− , or L+ ? Rs normally reflect with some degree of probability that a prey-like stimulus is in each of these locations, so by itself this does not discriminate between them. I suggest that the content for Rs in general and r in particular is that there is a whatsit in L, even though r’s
182
Karen Neander r
L+ L L−
Figure 8.5. What is the localization content of r, an excitation of a T5(2) cell caused by the worm located at L− ? Is it that the stimulus is at L− , L, or L+ ? The receptive field of the T5(2) cell that produces r is L
stimulus is located at L− and is in L+ as well as in L, and even though r carries the information that this stimulus is located at L− and is in L+ as well as in L. Assuming you agree, the task is to figure out why this is so. Are there identifiable principles involved? First ask, why does it seem overly specific to say that r’s content is that there is a whatsit at L− ? In part it is because we think that the content of r should be the same as the contents of other Rs (i.e. other +T5(2)s produced by the same cell). Although r carries the information that the stimulus is at L− , this is mere happenstance, for on other occasions when this cell fires the stimuli that provoke the Rs could be outside L− , elsewhere in L. The mechanism that normally informs Rs has no special causal or informational relation with L− that it does not also have with the rest of L. It has the function of informing Rs and hence r about stimuli in L, generally, as opposed to only stimuli at L− , more specifically. Next ask, why is it too imprecise to say that the content of r is that there is a whatsit in L+ ? It carries the information that the stimulus is in L+ as well as in L, and this is not mere happenstance. Normally, all Rs carry the information that there is a whatsit in L+ as well as in L. However, Rs are normally provoked by stimuli in L rather than elsewhere in L+ , and this is not mere happenstance because the mechanism that normally informs Rs has a special causal or informational relation with L that it does not have with the rest of L+ . It has the function of informing Rs and hence r about stimuli in L, as opposed to other areas in L+ . R-production is normally insensitive to whatsits that are in L+ if they are not in L. Notice as well that the wrong content ascription can interfere with its explanatory value. If we say that the content of r is a whatsit is in L+ , we undermine the explanation of the toad’s orienting, because it makes it a puzzle how the toad orients as precisely as it does. (Unless we want to admit
Content for Cognitive Science
183
that our content ascription is irrelevant to explaining the orienting.) And if we say that the content of r is a whatsit is in L− , we undermine the explanation of the toad’s orienting again, because it makes it a puzzle why the toad does not orient more precisely than it does. It makes it a puzzle why, after r, the toad’s brain still has to figure out the more precise location of the stimulus before it can send snapping instructions to its motor neurons. The coherence constraint is admittedly a bit vague but I hope that this illustration helps. The general idea is that content ascriptions need to make sense in the context of, and be suitable for playing a role in, an operational explanation of the relevant cognitive capacity.
3.2. What is Represented? Finally, we come to what is represented. Transcribing from frog to toad, popular candidates for what +T5(2)s represent are a member of a prey species, toad food (i.e. a packet of chemicals nutritious for toads), something moving in a direction that parallels its longest axis, and toad food that is moving in a direction that parallels its longest axis. Which should we choose for the purposes of mainstream cognitive science?⁷ I start with what the neuroethologists say. The truth is that they rarely if ever utter sentences beginning with ‘the content of a +T5(2) is . . . ’, or ‘+T5(2)s mean . . . ’, or ‘+T5(2)s represent . . . ’. However, the toad’s ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination is often used in neuroethology texts to illustrate what is referred to as object recognition (see e.g. Camhi 1984, ch. 7; Carew 2000, ch. 4) and in this context neuroethologists do speak in terms of (a) ‘‘prey-recognition’’, (b) ‘‘the recognition of wormlike stimuli’’, (c) ‘‘the recognition of preylike stimuli’’, and (d ) ‘‘the recognition of the configuration of visible features’’ (i.e. those related to the size, shape, and motion relative to shape of the stimulus). Different scientists do not say different things so much as use a similar range of expressions. For example, Jeffrey Camhi says, the toad’s T5(2) neurons are good candidates for prey-recognition neurons. (Camhi 1984: 232; emphasis added)
And then, in the next paragraph, he adds, it is implicit in the definition of recognition neurons that they must not only respond selectively to the particular object being recognized (in this case a moving wormlike stimulus), but they must actually be cells that the brain uses in recognizing this object. (Camhi 1984: 233; emphasis added) ⁷ I hope it’s clear that none of this is meant to solve the so-called functional indeterminacy problem. It’s one thing to argue for a given content, and another to show that one’s theory delivers it.
184
Karen Neander
Thus, in the space of two paragraphs, Camhi speaks of the recognition of prey and the recognition of a moving worm-like stimulus. Because the italicized in (a) through (d ) are used interchangeably in this way, I suggest that (a), (b), and (c) are all shorthand for (d ), which is the most precise but also the most cumbersome. We have just seen (a) and (b) being used interchangeably, and (b) and (c), which refer to worm-like and prey-like stimuli, respectively, presumably refer to stimuli that are worm-like or prey-like with respect to the configuration of visible features, and are thus presumably equivalent to (d ). I have to concede, to those who dislike talk of content in the case of such simple systems, that there is little explicit talk of error concerning what is represented. However, there is no hesitation when it comes to talk of error with respect to the localization content in this or similar cases (e.g. see Carew 2000: 65–70; Grobstein et al. 1983: 334–7) and neuroethologists do talk about the toad trying to catch ‘inappropriate’ stimuli. For example, if after ablation of the thalamus the toad orients toward a large looming square or the experimenter’s hand (remember that a normal toad only orients toward prey-like stimuli) this might be described as a response to an ‘inappropriate’ stimulus (e.g. Carew 2000: 115). It is interesting that the normal toad’s response to a cardboard cutout ‘‘worm’’ is not spoken of as inappropriate (as far as I have seen). On the contrary, it is treated as a paradigm case of the appropriate response. It is also interesting that if a toad orients toward toad food or toward a member of one of its prey species that is moved perpendicular to its longest axis the toad’s response is counted as inappropriate. This kind of case is neglected in the philosophical literature. Philosophers often find it intuitive that the toad misrepresents its target if it snaps at a cardboard cutout with, as I would say, the right configuration of visible features. But what of the case where the toad snaps at something with, as I would say, the wrong configuration of features? Suppose that we kill a millipede and string it up by its tail and move it perpendicular to its longest axis past a hungry toad. If the toad snaps at it, is it representing correctly or incorrectly? Different candidate contents give different answers. If the content is toad food, the relevant +T5(2) is correct. Ditto if the content is a member of one of the toad’s prey species. But if the content is specified in terms of the relevant configuration of visible features, it is incorrect. Which should it be? Here I think we should side with the neuroethologists’ talk of what is inappropriate. Not simply because that is how they talk, but because this is what coheres with our understanding of the information processing and the underlying mechanism.
Content for Cognitive Science
185
Several considerations speak in favor it. One is the fact (at least prior to any conditioning) that a normal toad does not snap at such an ‘‘antiworm’’. Only an abnormal toad, such as a toad with an ablated thalamus, reliably snaps at ‘‘anti-worms’’. It seems unfortunate in a theory of content if it implies that correct representation requires abnormality. Putting neurological impairment aside, the contents toad food or member of a prey species also seem unmotivated from the perspective of the information processing and underlying mechanism. As we’ve seen, the mechanisms that precede +T5(2)s are sensitive to environmental features relevant to the size, shape, and motion relative to shape of the stimulus. They can therefore support a content that concerns the configuration of visible features. They do not support the other contents in so far as they are insensitive to whether the stimulus is nutritious or whether it is a member of a certain species. No detectors of chemicals nutritious for toads are involved; no detectors of individuating characters of species are involved. At most, we can say that the mechanisms that detect the configuration of visible features approximate as detectors of these.⁸ Natural selection has (as biologists say) satisficed. Other things being equal, a capacity to tell nutritious things apart from non-nutritious things would have been good for the toad, as would a capacity to recognize members of species that have historically been caught and eaten by them. But the plain fact is that neither of these capacities evolved. Such capacities might require fancier cognitive equipment, which is expensive to build, operate, and maintain, or the needed mutations might have never arisen. Either way, these are not the capacities that a toad possesses. Instead the toad possesses a different capacity that merely approximates these.⁹ Of course, it is the capacity possessed and not the capacity approximated that neuroethologists aim to explain. In part this is a methodological point. As noted at the end of Section 2.2, neuroethologists need an exact description of the preferences of the motivated toad before looking for their neural basis or else they will be looking for the wrong thing. In part, it is also a logical point. You cannot explain how your car does 30 miles per ⁸ If what’s meant by ‘‘prey’’ is not a member of a species that has historically been hunted by the toad, but rather something that can be caught and eaten, the suggestion is much more promising. Not, in my view, because edibility is a Gibsonian affordance, but because +T5(2)s lead to further information processing that controls catching and eating, and we do not want to preclude the possibility that +T5(2)s have more content than their perceptual content (i.e. motor or motivational content). ⁹ One could argue that the toad has a capacity to recognize that food has a certain probability of being present, but notice that the toad would still not be in error when it tokened at something with the right configuration of visible features.
186
Karen Neander
gallon if it cannot do 30 miles per gallon and can only do 28 instead. You could explain how it could do 30 if it were altered in various ways, but neuroethologists are not primarily interested in explaining how to improve on the design of creatures like toads. What of the content toad food that is moving in a direction that parallels its longest axis? This kind of mixed proposal, intended to be ecumenical, doesn’t disregard the neuroethology as egregiously as others do, but the compromise needs motivating. It is a virtue of it that it lets us count a +T5(2) produced in response to a nutritious ‘‘anti-worm’’ as erroneous. But most of the objections mentioned above in relation to toad food apply equally well to toad food with the right configuration of features. The toad does not have a capacity to recognize whether or not the stimulus is nutritious. It has a capacity to recognize whether the stimulus has the right configuration of features. This latter capacity approximates the former capacity. But a capacity approximated is a capacity approximated, not a capacity possessed.
3.3. Responses to Objections Before closing this chapter, I would like to review a few potential objections. They cannot all be fully answered here, but it will be as well to register them and comment on them briefly. 1. Some worry that eating a whatsit would not be rational on the part of the toad unless the toad thought that whatsits were nutritious (e.g. Price 2001). I do not see this. It is not as if a toad represents a prey-like stimulus as not food or not nutritious. But more importantly, it is controversial whether the point of making content ascriptions is to rationalize behavior. That might be the point of making content ascriptions in the context of folk psychology (Price’s main concern), but rationalizing behavior, as opposed to explaining it, is not the point of content ascriptions in mainstream cognitive science. It does not assume that all doxastic representational systems are rational, let alone that all sub-doxastic representational systems are. Sometimes when this objection is raised it is said that unless the content ascription rationalizes the behavior, no intentional explanation of the behavior can be given. This depends on what counts as an intentional explanation. However, cognitive scientists can give a representational explanation of the toad’s behavior, consistent with the content ascription recommended here. They can explain that the toad is adapted to chase and swallow stimuli that it represents as having the right configuration of visible features. If that does not constitute an intentional explanation, so be it. 2. It might be argued that we are forced to select a content like toad food because an appeal to selection history is the only way to naturalize content
Content for Cognitive Science
187
and this is the content we get when we appeal to selection history. However, even if a theory of content must make reference to selection history, there are different ways to do so, and some theories of content that appeal to selection history generate contents of the sort favored here, as was noted in the introduction. Of course, whether they are satisfactory on other grounds remains to be seen. I am sometimes accused of not being teleological in proposing this argument. However, I see myself as steering us toward a more promising type of teleological theory. 3. Millikan (2000, app. ) argues that arguments like mine lead us to a bad end. She thinks that such arguments lead to the conclusion that content cannot go beyond what can be discriminated. In fact, to my dismay, Millikan interprets an earlier paper of mine (Neander 1995) as asserting this, and as even asserting that there can be no distal contents (and no distal functions)! I did not say this (I certainly did not mean to say this), and nor am I saying it here. I shall not try to unravel the prior confusion but we should be clear that the present argument does not entail that content cannot go beyond what can be discriminated. Millikan maintains that arguments like mine are ‘‘verificationist’’ (Millikan 2000, app. ). She also levels the claim against Jerry Fodor, for his having said that, ‘According to informational semantics, if it’s necessary that a creature can’t distinguish X s from Y s, it follows that a creature can’t have a concept that applies to X s but not to Y s’ (Fodor 1995: 32.) Fodor (1990b: 107–9) makes a similar point, in explaining his asymmetric-dependency theory, which entails that the frog represents (de dicto) small, black dots rather than flies. Fodor defends this outcome because, he says, if we think of the content as fly we would have to allow that some mistakes are nomologically necessary for the frog. He sheds light on what he means by this when he adds that if we think of the content as fly, ‘There is no world compatible with the perceptual mechanisms of frogs in which they can avoid mistaking black dots for flies’ (Fodor 1990b: 107–9). Fodor refers to the implication as ‘an attenuated sort of verificationism’ (1990b: 108), an admission that Millikan turns into a critique. However, Fodor’s principle is not really verificationist. For one thing, it applies to concepts not sentences or sentences in the head. For another, it only applies to conceptual primitives. Admittedly, most concepts are conceptual primitives according to Fodor, or anyway the Fodor from back then, but we need to note that we only approach even an attenuated form of verificationism if we combine Fodor’s principle with this further claim. Even with respect to conceptual primitives, Fodor’s principle only entails that their contents cannot go beyond what can be discriminated in a certain
188
Karen Neander
sense. It is unclear how much of a creature’s psychology Fodor wants us to keep fixed across possible worlds, but a world in which we learn something that we could learn is obviously consistent with our psychology. This means that even the contents of conceptual primitives can, according to Fodor’s principle, go beyond what can presently be discriminated. It allows for a distinction in referential content between ‘‘water’’ in our mouths and ‘‘water’’ in the mouths of our Twin Earth doppelg¨angers for instance (see Putnam 1975). We need to be equally cautious when interpreting my claim that contents have to be supported by relevant information processing. This also permits contents to go beyond what can be discriminated. Consider the water–XYZ case again. In 1600 Oscar refers to water, specifically, despite his inability to tell it apart from XYZ. One popular way to understand this capacity is that it involves what psychologists call ‘‘essentialist thinking’’. This in turn seems to involve an intention to defer, in this case, to nature to delineate the boundaries of the kind term. It seems to involve an intention to refer to a kind that has a certain Lockean nature, which might be hidden or unknown, but which is nonetheless expected to explain the superficial properties by which we recognize instances of the kind as instances of the kind (e.g. instances of water as instances of water). According to Frank Keil (1989), human children do not acquire this capacity until the early years of elementary school, but according to Susan Gelman and Henry Wellman (1991), there is some evidence of its earliest manifestations as early as 4 years old. Its development is uniform enough in children for developmental psychologists to suspect that it is an innate capacity for our species. I do not know if other primates have this capacity, but I am very confident that toads do not. It seems a moral certainty that we have the kind of information processing needed to support essentialist thinking and that toads do not. In fact, any theory of mental content that entails that toads have contents that in us require essentialist thinking is profoundly problematic. It obliterates an important distinction in our capacities versus theirs, and rushes headlong toward sophisticated contents, heedlessly bestowing them where they do not belong. There are many degrees of difference between our brains and those of toads, and a good theory of content needs to observe them all. Thomson’s gazelles surely do not have essentialist thinking either, and if so they must lack a water concept or a lion concept that is exactly like ours. Nonetheless, their perceptual abilities are much more sophisticated than a toad’s. For example, a gazelle can distinguish lion appearances from the appearances of other predators it encounters. When gazelles see predators approach, they monitor them, and whether they flee seems to
Content for Cognitive Science
189
depend on subtle behavioral cues on the predator’s part. How they flee (e.g. whether they engage in stotting, and how close they let the predator get before fleeing) depends on the type of predator (Griffin 2001: 71). Further, they can recognize a lion-like animal from different distances and different angles of view, and on the basis of seeing only small parts of it (Griffin 2001: 128). Compare this to the toad’s simple perceptual capacities. The toad’s detection of prey-like or predator-like stimuli is not mere transduction, but it is nonetheless based on a fairly simple configuration of visible features. A good theory of content should respect these gradations in sophistication. 4. Another way to respond to the argument in this chapter is to argue that an innately programmed inference is involved in the case of the toad. It seems undeniable that the toad is processing information about the configuration of visible features. But it could be argued that the toad’s brain infers the presence of something nutritious (or the presence of a member of a prey-species) from this configuration of visible features. According to most contemporary theories of perception, perceptual processing is often inferential. It often involves innate assumptions that are ‘‘embodied’’ in the processing. The relevant inferences are not deliberate or conscious but are implicit in unconscious processing. So why shouldn’t this be what is happening in the toad’s case? The problem with this response is that there is no motivation from the perspective of neuroethology for thinking that an inference is involved. And we are looking for pre-theory-of-content reasons for preferring one content over another. For one thing, if it seems plausible that the toad infers the presence of nutrients, it should seem equally plausible that it infers the presence of a member of a prey species, since the configuration of features can be considered equally good evidence for both. There is nothing in the neuroethology that distinguishes between these two competing interpretations. For another thing, the claim that there is an inference or something like an inference has empirical implications, if we are realists about representations. Those who press this reply must allow that some representation in the first place represents the configuration of features. There must be one representing the premise (or quasi-premise) and another representing the conclusion (or quasi-conclusion) of the inference (or quasi-inference). To insist that one and the same representation represents both premise and conclusion is an ad hoc move. Let us take a quick look at a case that cognitive scientists count as inferential: our perception of parallel lines converging in the distance. It is thought that we see them as parallel because our perceptual processing embodies the assumption that lines that converge toward the horizon are
190
Karen Neander
parallel and receding. The idea is that we infer, or sort of pseudo-infer, from the presence of converging lines in the 2D sketch that parallel and receding lines are present in the scene, and so our 3D sketch of it represents them as parallel and receding (Marr 1982). Notice that when cognitive scientists make this proposal they posit two stages of representation: the 2D sketch of converging lines and their subsequent 3D interpretation as parallel and receding. In the case of the toad, there is no evidence of a relevant second step. Once the toad tokens a +T5(2), it appears to have done all it does by way of ‘‘prey’’/‘‘predator’’/‘‘other’’ discrimination. From there, it moves on to processes that govern orienting, approaching, more precise localization of the prey, and so on. In any case, the inferential response cannot save certain theories (e.g. Millikan’s, at least on her construal of it), which do not allow that any representation of the configuration of visual features occurs. If there is no representation of the configuration of visual features there can be no inference from such a representation. In relation to this, it is also worth noting that a general principle of information-processing accounts of perception is that visible properties must be represented before invisible ones are (see e.g. Palmer 1999: 85–92,146–50). What is meant by a ‘‘visible property’’ here is not entirely clear. However, the idea is that the perceptual system must begin with the surface features of a perceived object that are presented to the perceptual system. It must begin with the hide of the cow, not its inner nature, and with the side of the cow facing the perceiver, not the parts hidden from view. This principle (often regarded as the downfall of a Gibsonian theory of perception)¹⁰ makes good sense. And any theory that entails that the toad’s perceptual system only represents the stimulus as nutritious violates this principle. 5. Nor does it help to claim that toads need to know about toad food. Millikan says that mice need to know about hawks, not merely about the properties by which they recognize hawks (Millikan 2000, app. ). But this is just plain wrong. The mouse does not need to know about hawks as such. As long as it escapes, the capacity to recognize hawk-like properties suffices (satisfices). Nor do toads need to know about toad food as such. What matters is whether they eat the food, not whether they represent it as such.¹¹ ¹⁰ I ignore Gibson’s theory of perception in this chapter because it’s not mainstream cognitive science. ¹¹ This recalls Fodor’s point: ‘All that’s required for frog snaps to be functional is that they normally succeed in getting the flies into the frogs; and, so long as the little black dots in the frog’s Normal environment are flies, the snaps do this equally well on either account of their intentional objects. The mathematics of survival comes out precisely the
Content for Cognitive Science
191
6. It might also be argued that mainstream cognitive science is just one approach to understanding animal behavior. I grant this. As I said in the introduction to this chapter, my conclusions here are conditional on the assumption that a theory of mental content should meet the needs of mainstream cognitive science. I am aware that there are radical critiques of mainstream cognitive science. I am also aware that there might be other scientific purposes, mainstream but in corners of science not examined here, that pull us in different directions. I leave it to others to argue that these are strong enough to warrant either pluralism about content or a different notion of content. 7. Next to last are concerns about distal content. Items with the right configuration of visual features include worms, and so on, which are distal. So there is nothing in the choice of content (the configuration of visible features) that precludes distal content. But does the principle I am advocating (that the relevant information processing must support the content we ascribe) drive us toward the proximal? My answer is no, not inexorably. Elsewhere (in Neander forthcoming) I propose a solution to the problem of distal content that is sensitive to this principle. (Promises, promises . . . ) 8. Finally, it can seem absurd to think that toads have a concept of something moving in a direction that parallels its longest axis, because this can seem rather a sophisticated concept. Didn’t we learn the concept of parallel lines in school? Is it not more sensible to think that, at most, toads can possess only basic concepts like food or worm or millipede? The problem with this objection is that it reverses the relevant ordering of basic–complex concepts. It would get the order right if what mattered were the order in which children acquire lexical concepts, but what matters is what is elementary from the perspective of information processing. Early visual processing in humans first involves representations of the surface features of objects. Recognition of things as food or worms or millipedes is based on this. The toad, if I am right, does not get that far. Its categories remain closely tied to the visible features of its environment. The present proposal does not attribute overly sophisticated capacities to the toad. On the contrary, compared to its competitors, it attributes less sophisticated capacities to the toad. This reminds me of a time when, after I had given a talk at a colloquium, someone began almost yelling at me that his 2-year-old daughter did not have a concept of an inter-aural disparity. I had been explaining a same either way’ (Fodor 1990b: 106). As Fodor goes on to say, this is the kind of thing that makes philosophers feel that content makes no difference, but we have seen that it makes an explanatory difference.
192
Karen Neander
hypothesis about sound localization. The hypothesis is that, in part, we hear where a sound is coming from by figuring out which ear received the pattern of sound first. Since a sound from the right side enters the right ear a fraction of a second before it enters the left ear, one plausible hypothesis is that the brain can determine the direction of sound (in part) by determining the difference in the time of its arrival at each ear. It’s not a straightforward process, because the brain has to figure out which pattern from one ear matches which pattern from the other. The bird’s trilling has to be matched with the bird’s trilling, and not with your friend’s talking, of the fridge’s humming, which might be heard simultaneously. Of course, it’s not a conscious process. We have no introspective access to the process. We only have introspective access to its results, when we hear sounds as coming from a particular direction. The interjector was confused, despite the impressive volume with which he explained his point. The hypothesis may or may not require his daughter to possess a concept of an inter-aural disparity. That depends on what is meant by ‘‘concept’’. But the proposal does not require her to possess one in any objectionable sense. People mean many different things by ‘‘concept’’, but here are two possibilities. According to one, I possess a concept of X just in case I have introspective access to a semantically structured representation of Xs. According to the other, I possess a concept of X just in case my brain employs a representation of Xs. The first is a much more demanding concept of a concept, and the interjector’s mistake was to think that this account of sound localization required that his daughter, who could hear where sounds were coming from, possessed a concept of an inter-aural disparity in the first, demanding sense. Of course, if it requires her to possess a concept of an inter-aural disparity, it can only require her to possess it in the second, quite undemanding sense. That is to say, it can only require that her brain represent inter-aural disparities. And that is not in the least absurd, for according to plausible psychological theories, she does. The same sort of thing is true for the toad. There is no suggestion in the preceding discussion that the toad has any concepts in the first demanding sense. What is required is merely that its brain has a representation of the relevant configuration of visible features, and thus a ‘‘concept’’ of them (if we choose to call it that) only in the second, undemanding sense.¹² ¹² Early drafts were presented to the Department of Philosophy at Duke University (2004), to the Philosophy of Science Colloquia, University of California at Berkeley ´ (2002), the Jean Nicod Institute and the Ecole Normale Sup´erieure, Paris (2002), the Instituto de Investigaciones Filos´oficas, Mexico City (2002), and the Department of Philosophy, Syracuse University (2001). I benefited a lot from the discussions at these events and would like to thank the participants.
Content for Cognitive Science
193
REFERENCES A, N. (1993), ‘What Do Frogs Really Believe?’, Australasian Journal of Philosophy, 71: 1–12. B, J. V. Z., and B, L. P. (1962), ‘Experimental Studies of Mimicry, 6. The Reaction of Toads (Bufo terrestris) to Honeybees (Apis mellifera) and their Dronefly Mimics (Eristalis vinetorum)’, American Naturalist, 96: 297–308. C, J. (1984), Neuroethology: Nerve Cells and the Natural Behavior of Animals (Sunderland, Mass.: Sinauer Associates). C, T. J. (2000), Behavioral Neurobiology (Sunderland, Mass.: Sinauer Associates). C, C., and A, M. (1992), ‘Prey-Catching and Predator-Avoidance in Frog and Toad: Defining the Schemas’, Journal of Theoretical Biology, 157: 271–304. C, H. B. (1936), ‘The Effectiveness of Protective Adaptation in the Hive Bee, Illustrated by Experiments on the Feeding Reaction, Habit Formation, and Memory of the Common Toad (Bufo bufo bufo)’, Proceedings of the Zoological Society of London (1936), 111–33. D, D. (1995), Darwin’s Dangerous Idea (New York: Simon & Schuster). D, F. (1986), ‘Misrepresentation’, in Radu Bogdan (ed.), Belief: Form, Content and Function (Oxford: Oxford University Press), 17–36. (1994), ‘If You Can’t Make One You Don’t Know How It Works’, in P. A. French, T. E. Uehling, and H. K. Wettstein (eds.), Midwest Studies in Philosophy, xix: Philosophical Naturalism (Notre Dame, Ind.: University of Notre Dame Press), 468–82; repr. with small revisions in D. Chalmers (ed.), Philosophy of Mind (Oxford: Oxford University Press, 2002), 491–9. E, J. P. (1980), Neuroethology: An Introduction to the Neurethological Fundamentals of Behavior, trans. Transemantics Inc. (Berlin: Springer Verlag). and K, W. (1978), ‘Configural Prey-Selection by Individual Experience in Toad BufoBufo’, Journal of Physiology, 126: 105–14. B, H., and S-P, E. (1983), ‘Neuroethological Analysis of the Innate Releasing Mechanism for Prey-Catching in Toads’, in J. P. Ewert, R. Capranica, and D. Ingle (eds.), Advances in Vertebrate Neuroethology (New York: Plenum Press), 413–75. F, J. (1990a), ‘Psychosemantics, or: Where Do Truth Conditions Come From?’, in W. Lycan (ed.), Mind and Cognitio: A Reader (Oxford: Blackwell), 312–37. (1990b), ‘A Theory of Content II’, in Fodor, A Theory of Content and Other Essays (Cambridge, Mass.: Bradford Books, MIT Press). (1995), The Elm and the Expert: Mentalese and its Semantics, The Jean Nicod Lectures (Cambridge, Mass.: MIT Press). G, S., and W, H. (1991), ‘Insides and Essences: Early Understandings of the Non-Obvious’, Cognition, 38: 213–44; repr. in E. Margolis and S. Laurence (eds.), Concepts: Core Readings (Cambridge, Mass.: MIT Press, 1999), 613–37.
194
Karen Neander
G, D. R. (2001), Animal Minds: Beyond Cognition to Consciousness (Chicago: University of Chicago Press). G, P., C, C., and K, S. K. (1983), ‘Frog Prey Capture Behavior: Between Sensory Maps and Directed Motor Output’, in J. P. Ewert, R. Capranica, and D. Ingle (eds.), Advances in Vertebrate Neuroethology (New York: Plenum Press), 331–47. J, P. (1997), What Minds Can Do: Intentionality in a Non-Intentional World (Cambridge: Cambridge University Press). K, F. C. (1989), Concepts, Kinds and Cognitive Development (Cambridge, Mass., MIT Press). K, J. R., and C, C. M. (1996), ‘Visually Elicited Turning Behavior in Rana Pipiens: Comparative Organization and Neural Control of Escape and Prey Capture’, Journal of Comparative Physiology, 178/3: 293–305. L, J. Y., M, H. R., MC, W. S., and P, W. H. (1959), ‘What the Frog’s Eye Tells the Frog’s Brain’, Proceedings of the Institute of Radio Engineers, 1940–51. M, D. (1982), Vision (San Francisco: Freeman). M, R. (1991), ‘Speaking Up for Darwin’, in G. Rey and B. Loewer (eds.), Meaning and Mind: Fodor and his Critics (Oxford: Blackwell), 151–64. (2000), On Clear and Confused Ideas: An Essay about Substance Concepts (Cambridge: Cambridge University Press). N, K. (1995), ‘Misrepresenting and Malfunctioning’, Philosophical Studies, 79: 109–41. (forthcoming), Mental Representation: The Natural and the Normative in a Darwinian World (Cambridge, Mass.: MIT Press). P, S. (1999), Vision Science: Protons to Phenomenology (Cambridge, Mass.: MIT Press). P, D. (1998), ‘Teleosemantics and Indeterminacy’, Australasian Journal of Philosophy, 76: 1–14. P, C. (1998), ‘Determinate Functions’, Nous, 32: 54–75. (2001), Functions in Mind: A Theory of Intentional Content (Oxford: Clarendon Press). P, H. (1975), ‘The Meaning of ‘‘Meaning’’ ’, in K. Gunderson (ed.), Language, Mind and Knowledge (Minneapolis: University of Minnesota Press), 131–93; repr. in Putnam, Philosophical Papers, ii: Mind, Language and Reality (Cambridge: Cambridge University Press). S, K. (1990), The Representational Theory of Mind (Oxford: Blackwell).
9 Representation and Unexploited Content Robert Cummins, Jim Blackmon, David Byrd, Alexa Lee, and Martin Roth 1. INTRODUCTION In this chapter, we introduce a novel difficulty for teleosemantics, namely, its inability to account for what we call unexploited content—content a representation has, but which the system that harbors it is currently unable to exploit. In Section 2, we give a characterization of teleosemantics. Since our critique does not depend on any special details that distinguish the variations in the literature, the characterization is broad, brief, and abstract. In Section 3, we explain what we mean by unexploited content, and argue that any theory of content adequate to ground representationalist theories in cognitive science must allow for it.¹ In Section 4, we show that teleosemantic theories of the sort we identify in Section 2 cannot accommodate unexploited content, and are therefore unacceptable if intended as attempts to ground representationalist cognitive science. Finally, in Section 5, we speculate that the existence and importance of unexploited content has likely been obscured by a failure to distinguish representation from indication, and by a tendency to think of representation as reference. ¹ There are, of course, initiatives in cognitive science that are not representationalist—e.g. the dynamic systems approach advocated by van Gelder and Port (1995) and others. If non-representationalist approaches ultimately carry the day, then disputes about how mental representation should be understood in cognitive theory will have been idle. For the most part, we simply assume in what follows that some form of representationalism is correct. But, now and again, we phrase matters more methodologically, as points about what representationalist explanations of cognition need to assume rather than as points about what cognitive processes actually require.
196
Cummins, Blackmon, Byrd, Lee, and Roth
2. TELEOSEMANTICS Teleological accounts of representational content identify the extension of a representation R with the class of things C such that, historically, it was applications of tokens of R to members of C that led to the selection and replication of the mechanisms that produced or consumed tokens of R. Accounts along these general lines are familiar from the writings of Millikan, Neander, Papineau, and others (Millikan 1984, 1986; Neander 1991; Papineau 1984, 1987; recent anthologies by Allen et al. 1998; Buller 1999; Ariew et al. 2002). For our purposes, the crucial point about all such theories is that a representation R can have the content C for a system only if it had, when selection took place, the ability to apply R to members of C. There cannot be selection for an ability that isn’t there. The scenario underlying teleological accounts of content features a sub-population with a mechanism (what Cummins 1996a calls an intender) in the R-application business. It applies R to a variety of things, including, under certain circumstances, members of C. Those applications—the applications of R to members of C—prove adaptive enough to cause the mechanism in question to spread through the population over time. It is no part of this story that the reliability of R applications (applications of R to Cs vs. non-Cs) or their accuracy (excellence of fit between a token of R and its target C) is ever very good, or improves over time. All that is required is that it is the applications of R to members of C that leads to the spread of the mechanism through the population. The trait—the characteristic pattern of R applications—may go to fixation even though R is applied somewhat inaccurately to members of C, and only under rather special or rare circumstances, and frequently applied to other things. We emphasize this to make it clear that the critique we elaborate in the next section does not depend on the reliability or accuracy of the selected representing or representation-consuming mechanisms. On the other hand, accurate enough applications of R to Cs cannot simply be random accidents: there must be a mechanism in the R-application business to select. 3. UNEXPLOITED CONTENT By unexploited content we mean information or content carried by or present in a representation that its harboring system is, for one reason or another, unable to use or exploit. A common-sense example will help to introduce
Representation and Unexploited Content
197
the basic idea. Imagine someone who learns to use road maps to find a route from point A to point B. A study of the map might lead to the following plan: make a left at the third intersection, then another left at the next cross street, followed by an immediate right. It never occurs to this person to use the map to extract distance information until, one day, someone suggests that the map shows a shorter route than the one generated. Prior to this insight, our imaginary subject uses the map in a way that would be insensitive to various geometrical distortions, such as shrinking the north–south axis relative to the east–west axis. If assignments of representational content are limited by the abilities its user actually has to exploit the map, we will have to say that there is no distance information there to be exploited until after the user has learned to exploit it. And this will evidently make it impossible to explain how the user could learn to effectively compare routes for length: you cannot learn to exploit content that isn’t there. Indeed, it is evident that this story makes no sense at all unless we concede that relative distances are represented before the user learns to exploit that information. Even if the user never exploits relative-distance information, we are forced to allow that it is there to be exploited, since, under the right conditions, the user could have learned to use maps to compare distances. This would not be possible if the map did not represent relative distances. How seriously should we take this sort of example? We think the lesson is far-reaching and fundamental. To begin with, the idea that a brain can learn to exploit previously unexploited structure in its representations is presupposed by all neural network models of learning. Such learning typically consists in adjusting synaptic weights so as to respond properly to input activation patterns. This whole process makes no sense unless it is assumed that input patterns represent proximal stimuli prior to learning, and that the representational content of input patterns remains the same throughout learning. Prior to learning, the network cannot properly exploit input representations: that is precisely what the process of weight adjustment achieves over time.² Having come this far, we can see that the problem of learning to exploit ‘lower-level’ (‘upstream’) representations must be ubiquitous in the brain, if we assume that the brain acquires new knowledge and abilities via synaptic weight adjustment. In perceptual learning, for example, proximal stimuli must be represented before the appropriate cortical structures learn or evolve to exploit those representations in target location and recognition. ² We have heard it said that the network creates the content of its input patterns as learning progresses. But if we say this, we have no reason to say that early responses are errors. And if early responses are not errors, why change the weights in any particular direction? Indeed, why change them at all?
198
Cummins, Blackmon, Byrd, Lee, and Roth
Figure 9.1. Depth form texture gradients
As an example, consider the capacity to exploit texture gradients as visual depth cues. Representations in V1 contain texture gradients but the ability to exploit these as depth cues (as you do when you view Figure 9.1) develops later. Similarly, the ability to exploit retinal disparity in binocular vision as a depth cue develops along with the organization of binocular columns in the visual cortex. This process can be aborted by exotropy, but in such cases, binocular fusion without stereoscopic depth vision can still be achieved as the result of surgical correction and vision training, demonstrating that the retinal disparity information is still present in early visual representations, but unexploited for depth. There is no need to multiply examples. Once our attention is drawn to the phenomenon, it is clear that there must be many features of representations, especially structural features, at nearly all levels of perceptual and cognitive processing, that require learning and/or development for proper exploitation. 4. TELEOSEMANTICS AND UNEXPLOITED CONTENT Situating these facts in an evolutionary context immediately reveals a problem for teleosemantics. It is certainly possible, and probably common, that the abilities required to exploit various features of representations evolved well after those features appeared in the representations themselves. As just remarked, the ability to exploit texture gradients in early visual representation as depth cues might well have evolved well after well-defined gradients were available in those early representations. Now here is the point: the presence of texture gradients in early visual representations could not have been adaptive prior to the evolution of the processes that exploit them. Teleosemantics, however, implies that texture gradients did not represent depth until after it became adaptive for visual representations to include
Representation and Unexploited Content
199
them. In general, content only becomes adaptive, hence a candidate for the kind of content-fixing selection contemplated in teleosemantics, when and if the ability to exploit it is acquired. Evidently, there can be no selection for an ability to exploit content that isn’t there. The ‘opportunity’ to evolve the ability to exploit texture gradients in visual representations as depth cues simply cannot arise unless and until depth-representing texture gradients become available to exploit.³ Reflection on this last point suggests that the same difficulty arises whether the ability to exploit some feature of a representation is learned or evolved. For concreteness, assume that the ability to exploit texture gradients as depth cues is learned. While the current state of neuroscience provides no definitive account of such learning, it is perfectly intelligible to suppose it involves the systematic adjustment of synaptic weights in some substructure of the visual cortex. Evidently, if the ability to learn to exploit texture gradients itself evolved after texture gradients became available in early visual representations, we have a situation exactly like the one just rehearsed: teleosemantics assigns no content to unexploited features of representations, and this undermines the obvious explanation of how the ability to learn to exploit such features might later become adaptive. To sum up: once our attention is directed to the phenomenon of unexploited content, it is natural to ask how the ability to exploit previously unexploited content might be acquired. Learning in the individual, and evolution in the species, are the obvious answers. Equally obvious, however, is that teleosemantics cannot allow for evolving the ability to exploit previously unexploited content: that requires content to pre-date selection, and teleosemantics requires selection to pre-date content. 5. REPRESENTATION AND INDICATION It seems likely that the very possibility of unexploited content has been overlooked in philosophical theories of content because of a failure to distinguish representation from indication. In this section, we digress a bit to explain how we understand this distinction, and conclude by suggesting how exclusive attention to indication tends to make the phenomenon of unexploited content difficult to discern.⁴ ³ The underlying general point here, that selection for a given capacity requires that the capacity already exist in some part of the population, is not new. See e.g. Macdonald (1989). ⁴ This section draws heavily from Cummins and Poirier (2004).
200
Cummins, Blackmon, Byrd, Lee, and Roth
5.1. Terminology 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—by (perhaps computational) mental processes. Such a 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.⁵ An indicator, on the other hand, simply produces structurally arbitrary outputs that signal the presence or magnitude of some property in its ‘receptive field’.
5.2. Indication We begin with some influential examples. • Thermostats typically contain a bimetallic element whose shape indicates the ambient temperature. • Edge detector cells were discovered by David Hubel and Torsten Wiesel (1962). They write: ‘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 of different brightness (edges), and dark rectangular bars against a light background.’ • ‘Idiot lights’ in your car come on when, for example, 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. Since we 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 and Wiesel are indicators, and that the pattern of electrical spikes they emit when they fire are ⁵ It is possible that the brain employs several such schemes. See Cummins (1996b) and Cummins et al. (2001) for further discussion of this possibility.
Representation and Unexploited Content
201
indicator signals. Similarly, the bimetallic element found in most thermostats is an indicator, and its shape is the signal.
5.3. Indication vs. Representation Indication is generally regarded as a species of representation. Indeed, causal and informational theories of representational content assert that representation is, or is inherited from, indicator content.⁶ We think the two should be kept distinct. Indication is transitive, representation is not. If S3 indicates S2, and S2 indicates S1, then S3 indicates S1. Imagine a photo-sensitive cell pointed at an ‘idiot light’ in your car, and attached to a relay activating an audio device that plays a recording: ‘The oil pressure is low.’ If the light indicates low oil pressure, so does the recording. Representation, on the other hand, is not transitive. A representation of the pixel structure of a digitized picture of the Statue of Liberty is not a representation of the statue’s visual appearance, though the later may be recovered from the former.⁷ To anticipate some terminology we will use later, a representation of the pixel structure is an encoding of the statue’s visual appearance.⁸ Indicator signals are arbitrary; representations are not. This is implied by 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. Because indicator signals are arbitrary, disciplined structural transformations of them cannot systematically alter their meanings. Such transformations, however, are precisely what make representations useful. Consider, for example, a software package that takes a digitized image of a face as input and ‘ages’ it, i.e. returns 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—a name, ⁶ The theory is generally credited to Denis Stampe (1977). Its most prominent advocates are Fodor (1987) and Dretske (1981). ⁷ Representation, on the view advocated by Cummins (1996a), is grounded in isomorphism. Since isomorphism is plainly transitive, it might seem that representation must be transitive too. In a sense, this is right: the things that stand in the isomorphism relation are structures—sets of ‘objects’ and relations on them. If S1 is isomorphic to S2, and S2 is isomorphic to S3, then S1 is isomorphic to S3. An actual physical representation, however, is not an abstract object; it has a structure—actually, several—but it isn’t itself a structure. The connected graph structure of a paper road map is isomorphic to the street and intersection structure of a town, but not to the town’s topology. The town’s topology is isomorphic to the topology of a citrus grove. But no structure of the road map need be isomorphic to any structure of the grove. ⁸ It is what Haugeland would call a recording of the picture. See Haugeland (1990).
202
Cummins, Blackmon, Byrd, Lee, and Roth
say—because there is no correlation between the physical characteristics something must have to be a signal that indicates the appearance of a face at age 18 and the physical characteristics of that face at age 18. It follows from the nature of indication that the structural properties of an indicator signal have no significance. 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 coincidentally, detector signals. Indicators are source-dependent in a way that representations 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, while 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 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’.
5.4. Discussion If indication is your paradigm of mental content, as it is bound to be if you hold some form of causal theory, you are going to focus on what fixes the content of an indicator signal.¹⁰ Whatever fixes the content of an indicator signal, it is not its structural properties. In this context, therefore, motivation is lacking for thinking about which aspects of a representation’s 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. ¹⁰ See Cummins (1997) for more on the marriage between causal theories, indication, and the Language of Thought.
Representation and Unexploited Content
203
structure can usefully be processed, and whether the ability to do that processing is learned or evolved or a combination of both. Maps rub your nose in the possibility of unexploited content; idiot lights do not. There can, however, be unexploited indicator signals. Think of the colorcoded idiot lights at intersections: you have to learn that red means stop, green means go. Before learning, this is also unexploited content (though not what we have been calling representational content), and, unsurprisingly, it makes trouble for teleosemantics. Teleosemantics implies that an indicator signal has no content until there has been selection for the indicator that generates it. But the ability to exploit, or to learn to exploit, an indicator signal can only evolve if the indicator is already there signaling its target. Magnetosomes are magnetically polarized structures (typically ferrite surrounded by a membrane) in single-cell ocean-dwelling anaerobic bacteria. The orientation of these structures correlates with the direction of the earth’s magnetic field. By following the magnetic orientation in a particular direction, organisms far from the equator can avoid aerobic water near the surface. For this to work, magnetosomes must be chained and attached at both ends of the cell to form a reasonably straight line in the direction of locomotion (see Figure 9.2). This is because the orientation of the organism is simply a consequence of the orientation of the chain of polarized molecules. The whole body of the bacterium is a floating compass needle. The organism swims, and will move in whatever direction it happens to point. Chaining, of course, is simply a physical consequence of having a lot of little magnets suspended in proximity. They will stick together north to
Figure 9.2. Magnetotactic bacterium from the Chiemsee, Bavaria, Germany (Biomagnetism Group, University of Munich). Dark blobs are sulfur granules
204
Cummins, Blackmon, Byrd, Lee, and Roth
south. What is not so obvious is why the north pole of the string winds up attached at the ‘front’—i.e. direction of locomotion—end of the organism. However this happens, it is certainly possible, indeed probable, that the attachment process evolved after magnetosomes themselves appeared within the cell body of anaerobic bacteria. Selectionist theories imply that magnetosome chains did not indicate the direction of anaerobic water until after it became adaptive to do so, i.e. only after the evolution of attachment. But surely it is in part because they did indicate the direction of anaerobic water that the attachment process was adaptive enough to be selected for.
6. CONCLUSION A very natural response to the foregoing is to say that unexploited content isn’t really content. After all, there is a lot of unexploited information in the environment, information that cognitive systems must acquire the abilities to exploit. We do not call that information content. We are sympathetic with the comparison between learning or evolving an ability to exploit information in a representation or indicator signal and learning or evolving an ability to exploit information in the environment. We think these are, in fact, deeply similar. The importance of this similarity is obscured or lost in theories that essentially take representation to be reference. Theories of content that take representation to be reference perforce focus on devising the conditions that (allegedly) fix the references of semantically primitive terms, relying on the standard truth-conditional combinatorics to fix the references and truth-conditions of complex expressions. Access to something that refers to horses—a primitive term in Mentalese—however, tells you nothing about horses. Actual information about horses, therefore, is to be found only in the (or a) set of Mentalese sentences that contain a |horse| (a Mentalese term referring to horses) and appear in the Belief Box. The only sense in which such an account allows for unexploited content, therefore, is the sense in which a cognitive agent might not exploit all of its beliefs about horses on a particular occasion. While this is undoubtedly a case of unexploited information, it is not a case of the sort we have been discussing. Returning to our analogy, inability to extract relative-distance information from a road map is quite distinct from failing to read or utilize some of the sentences in a book. In the later case, the content of the unread sentences is unproblematically extractable from those sentences; they just are not noticed for one reason or another. The problem is not that one doesn’t know how to read them. In the case of the map, a new skill is required to exploit the needed information. Unexploited
Representation and Unexploited Content
205
information of the sort allowed for in Language of Thought (LOT) theories evidently poses no problem for teleosemantics comparable to the one we have been urging, since the mechanisms responsible for applying the primitive constituents and processing the relevant syntactical machinery may be selected for independently of their occurrence in any particular belief. Cognitive systems need information. LOT accounts attempt to provide for this by giving indication a twofold role. First, indicator signals alert the organism to the instantiation of their target properties in their receptive fields.¹¹ Second, primitive terms of LOT inherit their references from the properties they are used to indicate in the context of detection. Cognition is understood as the application of theories expressed as organized sets of sentences in Mentalese, hence as a species of truth-conditional inference, implemented as computation over symbolic structures with Tarskian logical forms. Perhaps something like this story makes sense for the ‘higher’ cognition of adult humans, especially if, like Plato, one is inclined to suppose that cognition is thinking and that thinking is essentially talking to oneself. But this picture is of little use for understanding phenomena like the capacity to exploit texture gradients in early visual representations as depth cues. When we turn to phenomena such as these, the truth-conditional semantics of propositional attitude contents is of dubious significance to cognitive science. Much more important, we believe, is the conception of representation and indication briefly introduced above. Representations, thus conceived, are of use to cognitive systems as an information source in addition to the environment (i) because they can be stored, and (ii) because they can be structurally transformed in ways that the environment typically cannot be. Sophisticated indicators are of use because they can signal the presence of environmental features—e.g. the presence of a predator or a mate—that are extremely abstract from the point of view of proximal stimulation. Representation and indication thus conceived are what make reference and the propositional attitudes possible. A theory that begins with the truthconditional semantics of propositional attitude contents thus skips over most of the action and begins with a phenomenon that is just the sort of cognitive achievement that mainstream cognitive science seeks to explain in terms of representation. We do not wish to quibble over whether the phenomenon we have called unexploited content is really content. We do contend, however, that what we are calling content is what ultimately does the work in representationalist ¹¹ See Cummins and Poirier (2004) for a discussion of how indicators might become ‘source-free’ and function as terms.
206
Cummins, Blackmon, Byrd, Lee, and Roth
cognitive science.¹² No doubt we need to mark the distinction between exploited and unexploited content. We think ‘‘exploited content’’ and ‘‘unexploited content’’ do the job nicely. Refusing to use the word ‘‘content’’ for as yet unexploited features of structured representations strongly suggests, wrongly, that those features are somehow different from those that are exploited. There is no intrinsic difference between texture gradients that are exploited and texture gradients that are not. To suppose otherwise would be like supposing that road maps cease to represent relative distances in the hands of those who cannot extract that information from them.¹³ In this chapter, we have urged what we think is a novel objection to teleosemantic theories, namely that they cannot accommodate unexploited content or information. Surely, a necessary condition for the plausibility of a theory of mental representation that hopes to ground representationalist cognitive science is that it accommodate unexploited content or information. For it must be possible for a system to be able to learn or evolve the capacity to exploit the information carried by a representation or indicator signal, and this implies that the information is there prior to acquisition of the capacity to exploit it. REFERENCES A, C., B, M., and L, G. (eds.) (1998), Nature’s Purposes (Cambridge, Mass.: MIT Press). A, A., C, R., and P, M. (eds.) (2002), Functions: New Essays in the Philosophy of Psychology and Biology (Oxford: Oxford University Press). B, D. (ed.) (1999), Function, Selection, and Design (New York: SUNY Press). C, P. (1995), The Engine of Reason, the Seat of the Soul (Cambridge, Mass.: MIT Press). ¹² An anonymous reviewer complained that we have not cited actual examples of cognitive scientists appealing to unexploited content. We are saying that they presuppose it whenever they assume that a representational capacity or ability is learned or evolved. Presumably, they were all learned or evolved. ¹³ There is a temptation to think that an unexploited feature of a representation doesn’t represent anything to (or for) the system that harbors it. This is surely right. But to assimilate representation to representation to/for will, like teleosemantics, make it impossible to understand how, for example, the ability to exploit texture gradients as depth cures could be learned or evolved. For more on the representation/representation-to distinction, see Cummins (1996a). Notice, by the way, that what is important about texture gradients is not just that they somehow covary with depth. It is their suitability for structural processing that makes them useful. When covariation is all that matters, an arbitrary indicator signal is all that is required.
Representation and Unexploited Content
207
C, R. (1996a), Representations, Targets, and Attitudes (Cambridge, Mass.: MIT Press). (1996b), ‘Systematicity’, Journal of Philosophy, 93: 591–614. (1997), ‘The LOT of the Causal Theory of Reference’, Journal of Philosophy, 94: 535–42. and P, P. (2004), ‘Representation and Indication’, in H. Clapin, P. Staines, and P. Slezak (eds.), Representation in Mind (New York: Elsevier). B, J., B, D., P, P., R, M., and S, G. (2001), ‘Systematicity and the Cognition of Structured Domains’, Journal of Philosophy, 98: 167–85. D, F. (1981), Knowledge and the Flow of Information (Cambridge, Mass.: MIT Press). F, J. (1975), The Language of Thought (New York: Thomas Y. Crowell). (1987), Psychosemantics: The Problem of Meaning in the Philosophy of Mind (Cambridge, Mass.: MIT Press). H, J. (1990), ‘Representational Genera’, in W. Ramsey, S. Stich, and D. Rumelhart (eds.), Philosophy and Connectionist Theory (Hillsdale, NJ: Lawrence Erlbaum). H, D., and W, T. (1962), ‘Receptive Fields, Binocular Interaction and Functional Architecture in the Cat’s Visual Cortex’, Journal of Physiology of London, 160: 106–54. M, G. (1989), ‘Biology and Representation’, Mind and Language, 4: 186–200. M, R. (1984), Language, Thought, and Other Biological Categories (Cambridge, Mass.: MIT Press). (1986), ‘Thoughts without Laws: Cognitive Science with Content’, Philosophical Review, 95: 47–80. N, K. (1991), ‘The Teleological Notion of Function’, Australasian Journal of Philosophy, 69: 454–68. P, D. (1984), ‘Representation and Explanation’, Philosophy of Science, 51: 550–72. (1987), Reality and Representation (Oxford: Blackwell). S, S. (1987), Remnants of Meaning (Cambridge, Mass.: MIT Press). S, D. (1977), ‘Towards a Causal Theory of Linguistic Representation’, in P. A. French, T. E. Uehling, and H. K. Wettstein (eds.), Midwest Studies in Philosophy, ii: Studies in the Philosophy of Language (Minneapolis: University of Minnesota Press). G, T. J., and P, R. (1995), ‘It’s About Time: An Overview of the Dynamical Approach to Cognition’, in R. Port and T. van Gelder (eds.), Mind as Motion: Explorations in the Dynamics of Cognition (Cambridge, Mass.: MIT Press).
10 Fearing Fluffy: The Content of an Emotional Appraisal Carolyn Price 1. INTRODUCTION Harry was visiting his friend Mel. He had wandered into the kitchen to make a cup of tea, when he heard a sound behind him. He swung round to see Mel’s cat Fluffy hissing at him from the floor. Seeing Fluffy’s aggressive stance, Harry panicked: he kicked out at the cat and bolted from the kitchen. Later, Harry described to Mel what happened: ‘I evaluated the situation’, he said, ‘and judged that Fluffy presented a danger.’ Harry, it seems, was not being entirely honest: there is a difference between the dispassionate evaluation that he described to Mel and the emotional response that he actually produced. But what could the difference be? The problem is particularly puzzling if we accept that emotions are not simply bodily feelings or urges, but involve appraisals of some kind. How can we distinguish emotional appraisals from dispassionate evaluations? There are a number of ways in which the distinctive character of an emotional appraisal might be explained. One strategy is to suggest that emotional appraisals are psychological states of a distinctive type. They are not judgements, thoughts, desires, or perceptual states; nor are they some combination of these. Rather, they are produced by a separate psychological mechanism, and have a special role to play in our psychology (Griffiths 1990). Again, it might be suggested that emotional appraisals have a special kind of content.¹ For example, it is sometimes suggested that the content of emotional appraisals is related in a distinctive way to the subject’s interests or concerns (Averill 1980: 310; Nussbaum 2001: 52 n.). ¹ I use the term ‘content’ to refer to the truth (or correctness) conditions or the satisfaction conditions of intentional states.
The Content of an Emotional Appraisal
209
These strategies might be seen as alternatives; but they are not mutually exclusive. It may be that a complete explanation of the distinctive character of emotional appraisals will refer to their origin, their function, and their content. Indeed, it is reasonable to expect these considerations to be linked. In what follows, I would like to explore these links. In particular, I would like to explore the idea that the content of emotional appraisals is dependent on their function. I shall concentrate on a single case study: Harry’s appraisal of the danger presented by Fluffy. I shall refer to this appraisal as AP . My discussion will assume a certain theoretical background: that is, a teleosemantic theory of intentional content of the kind suggested by Ruth Millikan (1984, 1989). A theory of this kind begins from the claim that the content of an intentional state is determined by its biological function; or, more precisely, by the biological function of the mechanism that produced it, together with the way in which that mechanism normally works. Millikan’s version of the theory is marked by two distinctive features. First, on her approach, the notions of a biological function and of normality are understood historically.² Secondly, on her account, we cannot determine the content of an intentional state by considering only the conditions under which it would normally be produced: we must also consider the way in which a state of that type would normally benefit the subject (Millikan 1984: 100; 1989: 286; Price 2001: 82). I shall follow Karen Neander in referring to a theory of this type as a High Church teleosemantic theory (HCT) (Neander 1995). Elsewhere, I have attempted to apply a version of HCT to the content of relatively simple intentional occurrences, such as sensory signals; and I have attempted to extend that account to increasingly sophisticated intentional devices, including perceptual states with singular content, spatial maps, judgements, and desires (Price 2001). In this chapter, I shall rely on some of the claims and distinctions that I introduced in my earlier discussion. I shall introduce these bits of baggage, as briefly as I can, as I go along. In order to present a teleosemantic account of the content of emotional appraisals, it will be necessary to say something about the function that these states play in our psychology. This is, of course, a highly controversial issue; it is also an empirical issue, on which philosophers are entitled to do no more than speculate. In the next two sections, I shall offer a speculative account of the function of certain emotional appraisals, drawing ² Roughly, an item can be described as having the function to perform a certain task if its presence is explained, in part, by the fact that earlier items of the same type successfully performed that task (Millikan 1984: 25–9; 1989); an item will be carrying out its function in a normal way if its presence is explained, in part, by the fact that earlier items performed that task in the same way (Millikan 1984: 33).
210
Carolyn Price
on suggestions made by a number of philosophers and psychologists. The claims that I shall go on to make about the content of AP must be regarded as conditional on the truth of this account. 2. WHAT IS AN EMOTIONAL APPRAISAL? I shall begin with some brief remarks aimed at clarifying what I mean by an emotional appraisal. I take it that the occurrence of an emotion is a complex event, involving a structured pattern of physiological and psychological changes, triggered by an intentional state of some kind.³ I use the term ‘emotional appraisal’ to refer to the intentional state that triggers these changes.⁴ In this chapter, I shall assume that the changes involved in a particular occurrence of emotion are triggered by a single appraisal. This assumption may well turn out to be false. It has been suggested that an occurrence of emotion involves at least two processes: a speedy, automated process and a slower process that involves higher-level cognitive capacities (Oatley and Johnson-Laird 1987; LeDoux 1998; Toates 2002). If so, it is possible that these processes are initiated by two or more distinct appraisals, each performing a different set of functions. If this turns out to be correct, it will be necessary to adjust my account in order to accommodate this point. 3. GENERAL- AND SPECIAL-PURPOSE SYSTEMS In this section, I shall introduce a distinction that may help us to understand the function of an emotional appraisal. Human beings appear to possess a psychological system that is involved in making reasoned decisions about what to do in any given situation. This reasoning system is capable of drawing on a range of intentional states, including factual judgements, evaluations, and desires. Millikan suggests that there is an important distinction between this system and other intentional systems. Our reasoning system is a general-purpose system: it can trigger behaviour that serves almost any interest that we possess or that we ³ For two different views of what might be included within an occurrence of emotion, see Ekman (1980: 80–95); Goldie (2000: 12–16). ⁴ For present purposes, I shall sidestep questions about how these states are produced, leaving it open whether they are generated by a separate psychological mechanism or whether they are triggered by beliefs or desires. If HCT is correct, a complete account of the content of emotional appraisals will need to take account of the way in which they are produced; but I do not have space to explore this issue properly here.
The Content of an Emotional Appraisal
211
may acquire during the course of our lives (Millikan 1986: 72; Price 2001: 191–211). General-purpose systems benefit from a capacity to deploy information and practical skills in a highly flexible way. There is no telling in advance how any of the intentional states produced by the system will be used in pursuing the subject’s interests. On the other hand, systems of this kind face some significant challenges. First, because a general-purpose system serves a wide range of different interests, it needs to have some means of determining which interest to pursue in any given situation. Secondly, general-purpose systems have access to a broad range of information, only a small subset of which is relevant in any given situation. The system must have some way to ensure that the right information is brought to bear on the problem. A similar problem arises with respect to the behavioural responses that the system is able to deploy: the system must ensure that the responses that it considers are likely to be of some use. These problems are particularly pressing in situations that need to be dealt with quickly. General-purpose systems contrast with special-purpose systems. In particular, they contrast with intentional systems that are interest-dependent: these are systems that serve a specific interest or set of interests—for example, predator-avoidance or nest-building. Interest-dependent systems range from simple recognition–response mechanisms to systems with relatively sophisticated inferential capacities. The information and behavioural responses available to an interest-dependent system are used to promote only the interest or set of interests that the system serves. As a result, organisms that possess only interest-dependent intentional systems will be able to make only limited use of the information and skills that they acquire. Interest-dependent systems avoid some of the problems that confront general-purpose intentional systems. In particular, a system that serves only a single interest will not need to decide which interest to serve in any given situation. Moreover, such a system will not face the same degree of difficulty in focusing on relevant information and effective behavioural responses: a system of this kind will normally have access only to information that is likely to help it to execute the task that it functions to execute; moreover, the behavioural responses that it is able to deploy will normally be limited to responses that have, in the past, proved effective in enabling it to execute this task. The distinction between interest-dependent and general-purpose systems bears some similarities to Fodor’s distinction between modular and non-modular systems. Fodor characterizes modules as systems that are both domain-specific and informationally encapsulated: that is, they have a specific informational or executive task to perform and are able to draw on information from only a limited set of sources (Fodor 1983:
212
Carolyn Price
47–52, 64–86). Like modular systems, interest-dependent systems are domain-specific. However, they need not be informationally encapsulated: as we have seen, they will normally have access only to information that is relevant to the interests that they serve; but this does not, by itself, rule out the possibility that such a system might call on a wide array of informational sources, including intentional states generated by a generalpurpose system. The distinction between general-purpose and interestdependent systems focuses on the way in which the states that they generate are put to use, rather than the way in which they are produced. 4. THE FUNCTION OF AN EMOTIONAL APPRAISAL With this distinction at the ready, we can now turn our attention to the first question that we need to address: what is the biological function of an emotional appraisal? It is not axiomatic that emotional appraisals have a biological function at all. For this to be case, one of two possibilities must be realised. One possibility is that we have inherited a set of emotional capacities that once helped our ancestors to survive and to reproduce, and which thereby help to explain our presence today. If so, our emotional capacities will function to help us in just the way in which they helped our ancestors. Secondly, it is possible that our emotional capacities are not inherited, but develop during the course of our lives, in a way that depends on our experiences as individuals and as members of a certain social group. It might be thought that this scenario is incompatible with the claim that our emotional capacities have a biological function. But the two claims can be made compatible if we assume that the process of learning is controlled by mechanisms that themselves have a biological function. If such mechanisms exist, their function will be to generate emotional capacities suited to our physical and social environment. On this scenario, our emotional capacities could be ascribed functions deriving from the function of these mechanisms, together with the way in which they normally work (Millikan 1984: 46–7; Price 2001: 124–9). In what follows, I shall assume that one of these possibilities is realised. I shall not try to adjudicate between them.⁵ I shall begin from the suggestion that the function of emotions is to enable us to deal with what Paul Ekman calls fundamental life tasks (Ekman 1992: 171). These include hazards, such as an encounter with a large predator; and opportunities, such as an encounter with a potential ⁵ For some different views, see Ekman (1980, 1992); Griffiths (1997); Averill (1980); Goldie (2000: 84–122).
The Content of an Emotional Appraisal
213
mate. These situations are crucial for the subject, in the sense that an inappropriate response can be highly damaging. In addition, they are often emergencies, to which it is necessary to respond very quickly. According to the life task hypothesis, the function of emotional appraisals is to enable us to deal effectively with situations of these kinds.⁶ How do emotional appraisals help us to deal with these situations? Emotional appraisals are produced quickly, enabling the subject to identify the situation without delay. They trigger physiological changes, preparing the subject for action; and they prompt expressive behaviour, signalling to others how the subject is likely to react. In addition, they generate behaviour that is designed to resolve the situation: for example, fleeing from the threat; avenging the insult; celebrating the goal. Unlike some of the expressive behaviours triggered by emotional appraisals, these behaviours are not plausibly regarded as stereotypical responses: they are generated by practical inference. This implies that emotional appraisals are capable of influencing practical decision-making in some way. How do emotional appraisals influence decision-making? First, emotional appraisals are plausibly regarded as sources of motivation. Moreover, emotional motivations are urgent. In other words, they do not compete on an equal footing with the subject’s other goals. Gripped by fear, Harry would not normally balance his desire to avoid being injured by Fluffy against his desire to make himself a cup of tea: his panicky appraisal prompts him to treat the goal of protecting himself from Fluffy as his only current concern. Secondly, emotions seem to influence the way in which we think. We saw earlier that a subject who is capable of general-purpose reasoning faces a fundamental difficulty: they must ensure that, in deciding how to act in a particular situation, they focus primarily on considerations that are relevant to the situation. Ronald de Sousa suggests that emotions provide a solution to this problem: emotions function to frame our reasoning, by focusing our attention on information that is relevant to the problem at hand (de Sousa 1987: 190–6; see also Evans 2002). If this proposal is correct, then one function of AP is to fix Harry’s attention on the threat posed by Fluffy and on any aspect of the situation that might help him to escape or to ward off the threat. De Sousa’s proposal concerns the information that the subject considers before acting. We saw earlier that a similar difficulty arises with respect to ⁶ See also Griffiths (1990); Tooby and Cosmides (1990); Lazarus (1991); JohnsonLaird and Oatley (1992); LeDoux (1998). Theorists who have argued for the life task hypothesis have generally supposed that our emotional capacities are inherited; however, it would be possible to combine the life task hypothesis with the view that our emotional capacities are learned.
214
Carolyn Price
the possible actions that the subject will consider. Oatley and JohnsonLaird suggest that a further function of an emotional appraisal is to focus the subject’s attention on a specific set of responses to the situation (Oatley and Johnson-Laird 1987: 37). This seems plausible. If we are told that Harry is overcome by fear, we would expect him not only to try to avoid being injured by Fluffy, but to do so in certain predictable ways—for example, by running away from her, by trying to fend her off, or by cowering in a corner. We would not expect him to react by talking calmly to her, even if this is known to be an effective way to deal with angry cats. To respond by talking gently to Fluffy would require Harry to master his fear, not to act in accordance with it. This lends support to the suggestion that AP works to narrow down the types of action that Harry is able to consider. De Sousa sums up his proposal by suggesting that the function of emotions is to cause the mechanisms that are responsible for practical reasoning to operate as if they were informationally encapsulated modules. I would like to suggest instead that one function of an emotional appraisal is to cause our general-purpose reasoning system to act as if it were a special-purpose, interest-dependent system. As we saw earlier, such a system will focus exclusively on a single task; it will normally have access only to information that is relevant to that task; and it will be able to generate a limited set of actions, each of which has already proved effective in dealing with that task. All this suggests that an emotional appraisal has a very complex function. We might describe it roughly as follows. In a situation in which the subject is confronted by a certain type of crucial situation, an emotional appraisal functions: 1. to prompt the subject immediately to find a way to resolve the situation, without regard to other considerations; 2. to focus the subject’s attention on a narrow range of possible actions; 3. to focus the subject’s attention on information that will help him or her to choose one of these possible actions and to perform it in an effective way; 4. to trigger physiological changes that prepare the subject to carry out one of those solutions; 5. to trigger expressive behaviour that signals the subject’s situation and likely actions to other organisms. How plausible is this account? Certainly it is not complete: it ignores, for example, the longer-term effects of our emotional experiences on mood, motivation, and memory. Moreover, this account does not fit all types of emotional appraisal equally well. In particular, clause (2) does not apply to
The Content of an Emotional Appraisal
215
all kinds of emotional appraisal. This becomes clear if we contrast Harry’s panicky response to Fluffy with a case of anxiety. Suppose that Harry is anxious because he has been told that he is suffering from a dangerous illness. Knowing that Harry was anxious about the news would not tell us anything about how he might seek to deal with it; instead, we would expect him to keep turning the situation over in his mind, trying to find a solution—any solution—to his problem. This makes sense if we suppose that anxiety is a response to a threat, but not necessarily an imminent one. Anxiety keeps Harry’s attention focused on the threat, but it does not ensure that he deals with it immediately; as a result, it does not need to focus his attention on a limited set of responses. Again, it might be argued that not all types of emotional response function to prompt actions that are designed to resolve the situation. For example, in the case of ‘backward-looking’ emotions, such as sorrow or happiness, there is now nothing that the subject can do to influence what has happened. A similar claim might be made about certain empathic emotional responses—for example, fearing for a tightrope walker who is tottering above a 30 foot drop. Indeed, it is not immediately obvious what we should say about the function of responses of these kinds. These issues require further discussion. For present purposes, what I would like to suggest is that the account works well for certain kinds of emotional appraisal, including paradigmatic cases of anger and panicky fear. It makes sense of the urgency of these responses, their tendency to dominate our attention, and the fact that they are typically associated with particular kinds of action. In what follows, I shall use this account to underwrite some suggestions about the content of AP ; as I go along, I shall suggest some ways in which the content of AP might be contrasted with the content of an evaluative judgement about danger.
5. KINDS OF INTENTIONAL CONTENT Intentional states can be divided into different categories. First, there are states that register that a certain situation obtains or that a certain type of event has occurred. These states have descriptive content, identical with the information that they normally carry. Factual judgements possess purely descriptive content: they function to convey information, which can be used to satisfy a range of different goals. I have argued elsewhere that the states produced by simple signalling systems also have purely descriptive content. The function of a simple signalling system is to trigger some stereotypical response whenever a certain condition arises. The sensory system that triggers the eye blink reflex is a system of this kind: this system
216
Carolyn Price
works by triggering a blink whenever it senses that an object is approaching the eye. To do this, it does not need to represent blinking or protecting the eye as a goal; it just needs to indicate that something is approaching the eye (Price 2001: 138-41).⁷ Secondly, some intentional states represent a goal to be achieved or an activity to be performed. These states have directive content, determined by the goal or activity that they normally prompt. Desires and plans have content of this kind. States with directive content differ from simple signals in that they are normally used in some process of practical inference: in a process of this kind, the system decides how to achieve a certain outcome, given the information at its disposal. This might involve something as sophisticated as practical reasoning, but this need not be the case: for example, a bee might possess a navigational system that enables it to calculate how to get to a nectar source, given its current location. For this kind of process to occur, the system needs to represent ‘reaching the nectar source’ as a goal that the bee’s behaviour is supposed to realize. Desires can be ascribed purely directive content. This is because a desire typically represents a goal that could be worth pursuing in a range of different contexts. As Millikan points out, however, some intentional states represent goals that are supposed to be satisfied only when a certain condition arises: for example, a bee dance will normally prompt the watching bees to fly off to a certain location only if there is nectar at that location. States of this kind possess a combination of descriptive and directive content: a bee dance both conveys the information that there is a source of nectar at a certain location, and tells the watching bees to fly off to that location (Millikan 1984: 100; 1995a:191). In order to ascribe content to AP , we need to begin by deciding what form of content it possesses: does it possess purely descriptive content, purely directive content, or a combination of descriptive and directive content? 6. THE FORM OF AN EMOTIONAL APPRAISAL Many recent discussions treat emotional appraisals as judgements or thoughts, or as in some way analogous to perceptual states, implying that they have descriptive content of some kind. This seems correct: an emotional appraisal is supposed to be produced in a particular type of situation, and it is supposed to ensure that the subject’s response is appropriate to a situation of that type. For example, a panicky appraisal is supposed to be produced in the presence of a threat of some kind. If so, ⁷ Contrast Millikan (1984: 116–18; 1995a).
The Content of an Emotional Appraisal
217
we should treat AP as representing the presence of a threat. For example, we might begin by assigning it something like the following content: that cat is very threatening.⁸ I shall return to this ascription of descriptive content later on. At this point, I would like to turn to a different question: Should AP be ascribed a directive content? The idea that emotional appraisals include a directive element is not new. For example, de Sousa suggests that the content of an emotional appraisal includes an aim or goal (de Sousa 1987: 120–1). Some cognitive theorists have suggested that all or some emotions involve desires (Marks 1982; Lyons 1993: 63–4). Do the considerations set out in the last two sections support the idea that AP represents a goal of some kind? It will be helpful, first, to make clear which aspects of the account do not imply that AP has a directive content. First, the claim that AP functions to trigger certain involuntary physiological and behavioural changes is perfectly compatible with the claim that its content is purely descriptive. Harry does not decide how to produce these changes, and so there is no need to suppose that he represents them as goals. In this respect, AP is operating as a simple signal, triggering a set of stereotypical responses—just like the sensory signal that triggers the eye blink. We might also consider the way in which AP will normally influence Harry’s decision-making. According to the account suggested above, AP functions to ensure that Harry deals with the danger as a matter of urgency; further, it functions to focus his attention on relevant information—for example, the location of the threat. But Harry does not decide how to prioritise the situation, or how to focus his attention on relevant information. And so we do not have to suppose that he needs to represent these acts as goals. It is enough that AP conveys the information that he has encountered a threat, thereby triggering these motivational and cognitive changes. However, I suggested earlier that emotional appraisals function to ensure that the subject produces an effective response to the situation. For example, we might suppose that a function of AP is to ensure that Harry avoids being injured by Fluffy. Avoiding being injured is not a stereotypical response: Harry needs to decide how to achieve this outcome, given the information ⁸ It would be possible to question whether this ascription succeeds in capturing the full descriptive content of AP . In particular, it might be suggested that AP is supposed to reflect, not only how things are in the subject’s environment, but also Harry’s desires (Robinson 1983; Roberts 1988). If this were correct, it would follow that AP will normally carry the information that Harry desires to avoid injury: in other words, the descriptive content of AP should be expressed as: that cat is very threatening; and I desire not to be harmed. In contrast, Michael Stocker argues that our emotions are independent of what we desire (Stocker 1987: 64). If Stocker is right, we do not need to ascribe this more complex descriptive content to AP . I shall not pursue this matter here, because it turns on issues about the normal origin, rather than the function, of AP . But see n. 9.
218
Carolyn Price
that he has about his situation. For example, he might take account of information about the proximity of the threat, the availability of exits or suitable hiding places, and so on. If so, AP will represent avoiding injury as a goal to be achieved. Moreover, I also suggested that some emotional appraisals have a further function—to direct the subject’s attention to a specific set of possible actions. I suggested that this was true of panicky appraisals like the one produced by Harry. The options that occur to Harry might include running away, lashing out at the threat, or hiding from it. Again, these are not stereotypical responses: they are performed in a way that takes account of information about the situation: for example, information about the location of the threat, the position of an exit, and so on. This implies that AP will represent not only the overall goal of avoiding injury, but also a set of sub-goals: running away, lashing out at the threat, hiding, or whatever. Finally, this directive content will include a temporal element: it will represent the time at which the subject is to respond to the situation. Presumably, a panicky appraisal will normally prompt the subject to act immediately, but this may not be true of all types of emotional appraisal: there may be some kinds of emotional appraisal that do not normally prompt the subject to act immediately, but only at some time in the future. We have already seen, for example, that anxious appraisals do not motivate an immediate response. This highlights the need to distinguish between the claim that an emotional appraisal functions to prompt the subject to act immediately and the claim that it functions to ensure that the subject prioritises the problem. An appraisal that prompts the subject to respond at some time in the future may nonetheless function to ensure that the subject begins to search for an effective response straight away. But this ‘straight away’ does not need to be represented, because, as we have seen, prioritising the problem is not something that the subject needs to decide how to do. If all this is correct, it suggests that the content of AP should be expressed in something like the following way: that cat is very threatening; avoid being injured by it—by running away from it now or by lashing out at it now or by hiding from it now. What justifies this ascription of content to AP ? If we accept HCT, this ascription will be justified by the history of this type of appraisal. If the capacity to produce panicky appraisals is inherited, the crucial point will be that this capacity sometimes enabled Harry’s ancestors to avoid being injured by prompting them to run away from the threat, or to lash out at it, or to hide from it. If we assume that Harry’s capacity to produce panicky appraisals is learned, then this ascription will depend on the claim that
The Content of an Emotional Appraisal
219
this capacity has, in the past, enabled Harry (or perhaps others in his community) to avoid being injured by behaving in one of these ways. In either case, the content of AP will reflect the successes of the past. 7. THE FORM OF AN EVALUATIVE JUDGEMENT In the last section, I suggested that AP possesses a combination of directive and descriptive content. In this section, I shall consider what form of content we should ascribe to an evaluative judgement, for example the judgement ‘That cat is dangerous’. In order to determine the content of the evaluative judgement, we need to investigate its function. In other words, we need to specify how it normally contributes to processes of theoretical and practical reasoning. First, an evaluative judgement normally conveys some factual information. For example, the judgement ‘That cat is very dangerous’ will be appropriate only if there is a substantial possibility that the cat will cause serious harm to the subject. However, this evaluation differs from the factual judgement ‘That cat is likely to cause serious harm to me’. There is room for different views about the root of this distinction.⁹ But it is certainly plausible to suppose that one difference between these two kinds of state is that an evaluation normally motivates the subject to respond to the situation in a certain way. For example, if Harry were to judge that Fluffy is very dangerous, we would expect him to be motivated to avoid her, or at least to treat her with caution. If this is correct, evaluative judgements, like emotional appraisals, will possess a combination of descriptive and directive content.¹⁰ We could accept this suggestion without supposing that emotional appraisals and evaluative judgements perform their motivational function in just the same way. I suggested earlier that emotional appraisals function to prioritise certain considerations at the expense of others. There is no suggestion that evaluative judgements have this effect. A general-purpose reasoning system will normally operate by balancing the subject’s evaluations and desires against each other in coming to a decision about what to ⁹ One possibility is that an evaluative judgement depends, not only on how things are in the environment, but also on the subject’s desires or preferences. For example, it might be suggested that an evaluative judgement about danger will normally carry the information that the subject desires to avoid injury. (I find this suggestion attractive, but I am not clear whether it is correct.) If it turned out that evaluations, but not emotional appraisals, normally carry information about the subject’s desires, this would constitute a difference in content between the two states—a contrast explained, not by a difference in function, but by a difference in the way in which these states are normally produced. ¹⁰ Millikan tentatively endorses this suggestion (Millikan 1995a).
220
Carolyn Price
do. As a result, the judgement that Fluffy is dangerous presents the threat posed by Fluffy as one consideration among others. Unlike an emotional appraisal, it might well be overridden in favour of some other goal. It is possible to make a connection between this point and claims made by other writers concerning the distinction between evaluative judgements and emotional appraisals. Both Michael Stocker and Peter Goldie have suggested that the contrast between emotional appraisals and evaluative judgements is not a matter of their content—at least, not in the sense in which I am using the term here—but rather of the way in which their content is entertained by or presented to the subject. Stocker suggests that there is a distinctive sense in which the content of an emotional appraisal is ‘taken seriously’ by the subject (Stocker 1987). Goldie suggests that the difference between an emotional appraisal and an evaluative judgement is analogous to the difference between an indexical and a non-indexical thought: in feeling fear, he suggests, ‘you are emotionally engaged with the world, and typically you are poised for action in a new way’ (Goldie 2000: 61). Similarly, I have suggested that one difference between AP and an evaluative judgement is that the content of AP is presented to Harry as an immediate and overriding priority. As we have seen, however, this is not a difference in content: because AP does not represent prioritising the situation as a goal. However, there are a number of ways which the content of AP might be thought to differ from the content of an evaluative judgement. In what follows, I shall investigate some of these differences. I shall begin by considering the directive content of the two kinds of state.
8. SOUNDING THE RETREAT There is at least one respect in which the directive content of AP appears to differ from the directive content of an evaluative judgement. I have suggested that AP might be ascribed directive content taking roughly the following form: avoid being injured by that cat—by running away from it now or by lashing out at it now or by hiding from it now. In contrast, there is no reason to make such a complex ascription of directive content to an evaluative judgement. This is because an evaluative judgement does not have the function to prompt one of a specific set of possible actions. To judge that a cat is dangerous is not specifically to be motivated to flee from it or to lash out at it, but only to find some way to avoid being injured by it. This evaluation will help to generate an action only by becoming involved in a process of practical reasoning, in which the subject’s other beliefs are brought to bear on the problem. As a result, the
The Content of an Emotional Appraisal
221
way in which a dispassionate subject would respond to the situation will depend on what they believe: if they believe that the best way to deal with an angry cat is to talk gently to it, then this is what they would be likely to do. The evaluative judgement does not embody the lessons of the past in the way that an emotional appraisal does. Moreover, there is no reason to suppose that the evaluative judgement functions to prompt the subject to act immediately, rather than at some suitable time in the future. This is a corollary of the fact that the evaluation does not prompt the subject to produce one of a specific set of actions. A subject who calmly judges that they are in the presence of a dangerous cat might not react immediately: for example, they might decide to wait for something to happen (for the cat to spring, say) before taking action. All this suggests that the directive content of an evaluation will be simpler than the directive content possessed by AP . Consider the evaluative judgement ‘That cat is very dangerous’. The directive content of this judgement might be expressed as: avoid being injured by that cat. This suggests another way in which it would be possible to understand Goldie’s claim that the feeling of fear leaves the subject ‘poised for action in a new way’. The evaluative judgement ‘That cat is very dangerous’ will normally motivate the subject to take steps to avoid being injured by the cat; but a panicky appraisal will motivate the subject to produce one of a specific set of responses; moreover, the appraisal motivates the subject to respond in one of these ways right now. 9. WHAT CONSTITUTES A THREAT? In what remains of the chapter, I shall concentrate on the descriptive content of AP . Earlier, I suggested that the descriptive content of AP might be expressed as: that cat is very threatening. This might prompt us to ask what exactly constitutes a threat: is it the threat of any kind of physical injury to the subject, or is it restricted to only a subset of those? Might it include threats of other kinds—for example, the threat of losing something that the subject values? According to HCT, the answer to this question will depend on how the capacity to produce panicky appraisals has benefited Harry or Harry’s ancestors. It is possible that this explanation is very complicated, and takes in a number of scenarios. If so, the content of AP will relate to some highly complex relational property, or perhaps to a disjunction of such properties. There is no reason to assume that Harry’s evaluative concept ‘dangerous’ picks out exactly the same set of threats.
222
Carolyn Price
Indeed, there is at least one reason to deny that this will be the case. According to HCT, as we have seen, the content of AP will be determined by the nature of situations in which the production of a panicky appraisal has benefited Harry or Harry’s ancestors. This implies that these situations had a feature in common: that is, they all involved a threat that could be avoided or overcome in some way. More precisely, they all involved a threat that could be avoided by running away, or by lashing out, or by hiding. Where a situation involves a threat that cannot (in principle) be avoided in one of these ways, the capacity to produce a panicky appraisal could not normally benefit the subject. This implies that AP will represent the presence, not merely of a threat, but of a threat that can, in principle, be avoided in one of these ways. If Harry is confronted by a threat of some other kind— for example, if he hears that he is suffering from a dangerous illness—it would not be appropriate for him to produce an appraisal of this kind.¹¹ At first glance, it might seem that a similar restriction will apply to Harry’s evaluative judgements. For example, it might be thought that a subject could not normally benefit from a capacity to evaluate a situation as dangerous if that situation cannot be avoided. If this were correct, it would imply that Harry’s evaluative concept ‘dangerous’ will apply only to dangers that he can, in principle, avoid. Of course, even if this were correct, it would imply that the descriptive content of this evaluative judgement is broader than the descriptive content of a panicky appraisal. This is because the directive content of the evaluative judgement ‘That cat is dangerous’ does not concern a specific set of possible actions, but only concerns some more general goal, such as avoiding injury. As a result, there would be no need to suppose that the evaluative concept ‘dangerous’ concerns only dangers that can be escaped by running away, lashing out, or hiding. So the evaluative judgement would apply to a wider range of cases than AP . However, I have argued elsewhere that, in the case of subjects who are capable of a certain rather sophisticated form of reasoning, the content of judgements and desires is not subject to this kind of restriction (Price 2001: 241–50).¹² Subjects who are capable of this kind of reasoning are able to recognise that certain goals are (in principle) beyond their reach, and to understand this as resulting from a combination of their own limited capacities and of some independent feature of the situation—the distance of a place, say, or the complexity of a problem. In what follows, I shall refer to this form of reasoning as R-reasoning. A subject who is capable of R-reasoning might sometimes save themselves time and trouble ¹¹ Anxiety might be an appropriate response, however. ¹² For other discussions of this issue, see Peacocke (1992: 129–32); Millikan (1995b); Papineau (1996).
The Content of an Emotional Appraisal
223
by abandoning the pursuit of an unattainable goal; or they might take steps to prevent a certain outcome from becoming unattainable. In order to obtain these benefits, the subject must be able to represent certain goals as unattainable: they must be able to think of certain places as unreachable, certain problems as insoluble, certain dangers as unavoidable, and so on. If Harry is able to engage in R-reasoning, he might well possess an evaluative concept ‘dangerous’ that applies to all kinds of threat, including threats that he cannot, in principle, avoid. Might similar considerations be made to apply to AP ? They could be made to apply if it could be shown that Harry’s ability to engage in Rreasoning sometimes depends on his capacity to produce panicky appraisals. In particular, it might be suggested that, on some occasions, a panicky appraisal will normally prompt the evaluative judgement that the subject is in a dangerous situation. If this were so, the subject could sometimes benefit by producing a panicky appraisal in response to an unavoidable threat; this is because the appraisal might sometimes prompt a judgement that was later used in a process of R-reasoning. However, it is far from clear that this kind of inferential connection normally exists. Indeed, there are reasons to think that the evaluative judgements and emotional appraisals are normally generated independently of each other (Greenspan 1980; Griffiths 1990).¹³ Unless this view turns out to be mistaken, we should conclude that there is an important distinction between the descriptive content of AP and the descriptive content of an evaluative judgement. The content of AP will concern only threats that the subject is able to defuse by performing one of a limited set of actions. Whether a threat falls into this category will depend, in part, on the practical capacities of the subject. In contrast, the content of the evaluative judgement does not depend on the practical capacities of the subject. As a result, the content of the evaluative judgement will be marked by a kind of objectivity that we cannot attribute to AP .
10. THE TIMING OF THE THREAT Philip Percival has pointed out an intriguing distinction between emotional appraisals and dispassionate evaluative judgements (Percival 1992; see ¹³ The fact that a frightened person is more likely to judge that they are in a dangerous situation is not in itself evidence for the existence of this inferential connection. This phenomenon might arise from the fact that the feeling of fear focuses the subject’s attention on the frightening aspects of the situation, and so on the evidence that supports that evaluation.
224
Carolyn Price
also Maclaurin and Dyke 2002). If someone has suffered a loss, we would expect them to feel sad; but, in most cases, we would expect their sorrow to diminish after a time. Again, if someone has been treated very unjustly, we would expect them to feel angry; but we would expect their anger to fade over time. Something similar applies to fear: we would expect Harry’s panic to fade away as soon as he has escaped from Fluffy. Indeed, in many cases, we would think that there is something inappropriate or irrational about an emotion that does not diminish in intensity over time. In contrast, if someone judges that they have suffered a flagrant injustice or survived a very dangerous encounter, we would not expect their evaluation of what has happened to change as time passes: they should not judge that the injustice was any less flagrant or the encounter any less dangerous because it occurred a long time ago. A consideration of the function of these two kinds of state can help us to make sense of this distinction.¹⁴ I have suggested that the function of a panicky appraisal is to mobilise physiological, cognitive, and behavioural resources to deal with a threat that currently confronts the subject. The ‘currently’ is important: Harry would not normally benefit from a propensity to produce a panicky appraisal in response to the information that some threat had occurred in the past, or that some threat will occur at some time in the future. A panicky appraisal, then, will represent a threat as present or as imminent; but not as past, or in the future. Of course, Harry may remember a past encounter with Fluffy, and feel afraid. But his appraisal will represent the threat as present or imminent, not as past or in the more distant future. The range of temporal content that an emotional appraisal can convey will depend on the nature of the emotion. Panicky actions are likely to be useful only in the midst of the crisis. In contrast, angry actions might help the subject to deal with offences that are just about to happen, that are happening, or that have happened in the recent past. This implies that an angry appraisal might represent an offence as imminent, present, or in the recent past—but not as belonging to the remote past or the distant future. In contrast, there are no such restrictions on the temporal content of an evaluative judgement. For example, if Harry can judge that Fluffy is dangerous as he eyes her across the kitchen, he can also judge that Fluffy was dangerous when he remembers the event the next day. It is easy to see how the capacity to produce this past-tense evaluation might be of use to Harry in drawing further inferences—for example, the inference that he ought to avoid visiting people who live with cats. Again, Harry can also judge that Fluffy will still be dangerous two years from now: this is because he can ¹⁴ Maclaurin and Dyke offer a similar explanation (Maclaurin and Dyke 2002).
The Content of an Emotional Appraisal
225
benefit from a capacity to produce evaluations that concern situations that lie in the distant future when he is making long-term plans. As Percival makes clear, a similar point applies to the representation of hypothetical events. Harry can produce evaluative judgements that concern hypothetical situations, or even situations that are known not to have occurred. In contrast, the account offered here suggests that a panicky appraisal will always represent a situation as actual. This is because panicky appraisals function to prompt an immediate response, which would be redundant or even harmful in the face of a merely possible threat. The situation may be different in the case of other emotions: for example, it is possible to see how a subject might benefit from anxious exploration of a merely possible situation, or even a situation that is known not to have occurred.
11. DEGREES OF THREAT As I have characterized it, AP represents not only the occurrence of a threat, but also its degree: Harry’s wild flight suggests that he represents Fluffy not simply as a threat but as very threatening. Again, he might judge that Fluffy is very dangerous. However, there are some reasons to suppose that there will be some differences in the way in which these kinds of state represent the severity of a threat. We would expect the severity of the threat represented by AP to be linked with the intensity of the emotional episode that it triggers. The intensity of an emotional episode appears to involve a number of elements. The more intense Harry’s panic, the more intently his attention is focused on the situation; the more strongly he is motivated to deal with the threat; and the more vigorous are the physiological changes that he undergoes. Presumably, one of the functions of AP will be to ensure that the intensity of these changes correlates with some variable factor in his situation. The intensity of an episode of fear seems to reflect a number of factors—the severity of the threatened injury, the probability that the injury will actually occur, and its closeness in time. If so, AP will normally carry information about all these factors. There are two things to note about this. First, there is no reason to suppose that these factors are represented by separate elements in the appraisal. We would need to suppose this only if there was some differentiation in the way in which Harry would normally respond to each of these factors. Secondly, it might turn out that the relation between the degree of threat represented by Harry’s appraisal and the extent to which each of these factors is exemplified by the situation is not a linear one. In particular,
226
Carolyn Price
although it seems reasonable to suppose the degree of threat represented by AP will increase with the probability that injury will occur, this will only apply up to a certain point: once injury is certain, as we have seen, Harry could not normally benefit by producing a panicky appraisal. It follows from this that the variation represented by this element of the appraisal might be quite idiosyncratic; there is no reason to assume that it will be captured by some concept ‘degree of danger’ that features in Harry’s evaluative judgements. For example, given that Harry is capable of evaluating situations in the past, present, and future, we would expect an evaluative judgement to distinguish the seriousness and probability of a threat from its closeness in time. But, as we have seen, it is possible that AP does not make this distinction.
12. CONCLUSION I have suggested that one difference between AP and an evaluative judgement has to do with the way in which the content of the two states is entertained by or presented to the subject: in the case of AP , its content is presented as an urgent priority. This is not a difference in content, but a difference in the type of response that these two kinds of state are supposed to provoke. In addition, I have argued that it is possible to identify at least four significant differences between the content of AP and the content of the evaluative judgements that Harry is able to make concerning danger. 1. The directive content of AP is more complex than the directive content of his evaluative judgements. 2. The descriptive content of AP concerns only threats that Harry is able to avoid by running away, lashing out, hiding, and so on. In contrast, evaluative judgements may represent dangers that the subject cannot, in principle, avoid. 3. Panicky appraisals can represent only threats that are present or imminent. In contrast, evaluative judgements may represent dangers in the past, or in the distant future; and they may represent dangers that are merely possible or that have not in fact occurred. 4. AP represents degree of threat in a way that may be quite different from the way in which an evaluative judgement represents degree of danger. My conclusions are limited in two different ways. First, as I stated earlier, they are conditional both on the success of HCT as a theory of intentional content and also on the speculative account of the function of emotional appraisals that I presented earlier. Secondly, in this chapter, I have offered
The Content of an Emotional Appraisal
227
an account only of panicky appraisals. Although I have indicated some of the ways in which the content of other types of emotional appraisal might be thought to differ from the content of panicky appraisals, I have not attempted to offer an account of the content of emotional appraisals in general. Nevertheless, I hope that, by examining one relatively straightforward case study, I have indicated how a teleosemantic approach might be used to generate a credible and informative account of the content of other types of emotional appraisal.¹⁵
REFERENCES A, J. (1980), ‘A Constructivist View of Emotion’, in R. Plutchik and H. Kellerman (eds.), Theories of Emotion (New York: Academic Press), 305–39. S, R. (1987), The Rationality of Emotion (Cambridge, Mass.: MIT Press). E, P. (1980), ‘Biological and Cultural Contributions to Body and Facial Movement’, in A. Rorty (ed.), Explaining Emotions (Berkeley: University of California Press), 73–101. (1992), ‘An Argument for Basic Emotions’, Cognition and Emotion, 6: 169–200. E, D. (2002), ‘The Search Hypothesis of Emotion’, British Journal for the Philosophy of Science, 53: 497–509. F, J. A. (1983), The Modularity of Mind (Cambridge, Mass.: MIT Press). G, P. (2000), The Emotions: A Philosophical Exploration (Oxford: Oxford University Press). G, P. (1980), ‘Ambivalence and the Logic of Emotions’, in A. Rorty (ed.), Explaining Emotions (Berkeley: University of California Press), 223–50. G, P. (1990), ‘Modularity, and the Psychoevolutionary Theory of Emotion’, Biology and Philosophy, 5: 175–96. (1997), What Emotions Really Are (Chicago: University of Chicago Press). J-L, P., and O, K. (1992), ‘Basic Emotions, Rationality and Folk Theory’, Cognition and Emotion, 6: 201–23. L, R. S. (1991), Emotion and Adaptation (Oxford: Oxford University Press). LD, J. E. (1998), The Emotional Brain: The Mysterious Underpinnings of Emotional Life (London: Weidenfield & Nicolson). L, W. (1993), Emotion (Aldershot: Gregg Revivals; first pub. 1980). M, J., and D, H. (2002), ‘Thank Goodness that’s Over: The Evolutionary Story’, Ratio, 15: 276–92. M, J. (1982), ‘A Theory of Emotions’, Philosophical Studies, 42: 227–42. ¹⁵ I have read early drafts of this chapter at seminars at York University and at the Open University. I would like to thank the participants for their thoughtful comments. The chapter has also benefited from conversations with Philip Percival, Kevin Sludds, and Fred Toates. I am grateful for their help.
228
Carolyn Price
M, R. (1984), Language, Thought, and Other Biological Categories (Cambridge, Mass.: MIT Press). (1986), ‘Thoughts without Laws; Cognitive Science with Content’, Philosophical Review, 95: 47–80. (1989), ‘Biosemantics’, Journal of Philosophy, 86: 281–97. (1995a), ‘Pushmi-pullyu Representations’, in J. Tomberlin (ed.), Philosophical Perspectives, ix (Atascadero, Calif.: Ridgeview Press), 185–200. (1995b), ‘A Bet with Peacocke’, in C. Macdonald and G. Macdonald (eds.), Philosophy of Psychology: Debates on Psychological Explanation (Oxford: Blackwell), 285–92. N, K. (1995), ‘Misrepresenting and Malfunctioning’, Philosophical Studies, 79: 109–41. N, M. (2001), Upheavals of Thought: The Intelligence of Emotions (Cambridge: Cambridge University Press). O, K., and J-L, P. N. (1987), ‘Towards a Cognitive Theory of Emotions’, Cognition and Emotion, 1: 29–50. P, D. (1996), ‘Discussion of Christopher Peacocke’s A Study of Concepts’, Philosophy and Phenomenological Research, 56: 425–32. P, C. (1992), A Study of Concepts (Cambridge, Mass.: MIT Press). P, P. (1992), ‘Thanks Goodness that’s Non-Actual’, Philosophical Papers, 21: 191–213. P, C. (2001), Functions in Mind: A Theory of Intentional Content (Oxford: Oxford University Press). R, R. C. (1988), ‘What an Emotion Is: A Sketch’, Philosophical Review, 97: 183–209. R, J. M. (1983), ‘Emotion, Judgement and Desire’, Journal of Philosophy, 80: 731–40. S, M. (1987), ‘Emotional Thoughts’, American Philosophical Quarterly, 24: 59–69. T, F. (2002), ‘Application of a Multilevel Model of Behavioural Control to Understanding Emotion’, Behavioural Processes, 60: 99–114. T, J., and C, L. (1990), ‘The Past Explains the Present: Emotional Adaptations and the Structure of Ancestral Environments’, Ethology and Sociobiology, 11: 375–424.
Index adaptation 11, 24, 27, 29, 31, 33, 36 language-specific 31 affordances 101, 185 fn.8 Agar, N. 168 Aydede, M. 116 bee dance 61, 107, 110, 216 Bennett, J. 33 biological design 100 environment 37 kinds 119 purposes 102–03, 105 utility 100, 102–03, 105 Bloom, P. 24, 119, 123 Boghossian, P. 159–60 Braddon-Mitchell, D. 159–60 Camhi, J. 183–4 Campbell, D. 15 Cartesian 81, 98 coevolution 19, 32, 36–8, 158–9, 161, 164 cognitive science 16, 19, 20, 42–3, 48–50, 55, 96, 150, 167–69, 183, 189–90, 191, 205–6 coherence 110, 139 fn.15, 180 ,183 communication 28, 30, 34, 147, 151, 160, 165 compositional 10, 13, 106–08 computational 34, 59, 155, 169, 178, 200 concept 24, 34–6, 78, 104, 187–88, 191–92 empirical 108–09 evaluative 221–23 mentalistic 44–6, 65, 72 of belief 76 conditioning 104, 155, 172 instrumental 102 operant 102 consumer 19, 49, 100, 105–08, 117–18, 150–53, 157, 159–60, 202 mechanism 4–6, 49, 61, 63–4 semantics 100 contraries 113
correspondence 35, 52, 106, 110, 122, 136–37, 152 Cosmides, L. 23, 26 Craik, K. 125 culture 26, 32, 37–8 material 24 evoked 23, 24–6, 32, 36 Cummins, R. 9–10, 47, 49–50, 51–3, 64, 201 fn.7 defect 28–30 Dennett, D. 117, 168 Descartes, R. 75, 76 fn.9 (see also Cartesian) de Sousa, R. 213–4, 217 dispositions 103, 105–08, 132, 138, 155 relational 107 Donald, M. 25 Dretske, F. 8, 117, 120–21, 123, 141 fn.17, 149, 151, 168, 201 fn.6 Dunbar, R. 29 Dyke, H. 224 Ekman, P. 210, 212 encapsulated 16, 25, 32–4, 36, 211–12, 214 emotions 208–27 Ewert, J.-P. 173, 175 externalism 76 feedback 32, 61–3 fitness 19, 28 Fodor, J. 25, 33, 42–3, 116–17, 146, 168, 187–88, 190 fn.11, 201 fn.6, 211 frog visual system, 168–72, 174–5, 179 functional analysis 10 categories 3 indeterminacy 174 fn.5, 183 fn7. property 94, 96, 122 role 18, 89, 93, 96 teleo- 61–2, 135 traits 12
230 functionalist 54, 59 functions biological 1, 3–5, 9, 12, 42, 61, 75, 100, 105, 107, 110, 144, 149–50, 157, 165, 209, 212 concept of 3, 60, 62 consumer 152 detector 153–54, 157 effector 153–54, 157, 161, 165 etiological 10–3, 15–6, 73–4 Millikan 12–3 normal 74, 157 proper 13, 16, 33, 62, 84, 106 relational 13, 18, 123 systems account of 10–11 Gallistel, R. 52 Gelman, S. 188 genes homeo-box 102 purposes of 103–05, 113 selection of 12–16, 102, 104 Gibson, J. 101, 146, 185, 190 Goldfarb, W. 53–4 Goldie, P. 210, 220–1 Grice, P. 33, 73 habituation, 154–6, 161–2 Hardcastle, V. 73 fn.5 Haugeland, J. 202 fn.6 Hebb, D. 155 Hempel, C. 2 Hilbert, D. 159 hippocampus 51, 57 Hubel, D. 200, 202 indication 8, 20, 52, 72, 117, 149, 199–202, 205, 216 inference 33, 107, 111, 189, 224 practical 213, 216 innate 23, 25, 27, 31–2, 109, 155, 171, 173, 188–89, 212 language module 36 systems 27, 36–7 intentional icon 49, 64 interpretation 26, 33, 44–5, 48, 50, 63, 189, 190 isomorphism 52–3, 201 Jablonka, E. 32 Jackson, F. 159–60 Jacob, P. 168 Johnson-Laird, P. 214
Index Kaplan, D. 148 Keil, F. 188 kimu 7–9 Kripke, S. 62, 106–07 language module 26–8 of thought 200, 205 parsing 33 proto- 26–7 learning 15–18, 26, 31–2, 35, 63, 72, 109, 122–23, 125–27, 130, 137, 139–41, 155, 172, 197–9, 203–04 classical, 126–7, 137 instrumental 102 mechanisms 15, 61, 212 social 14, 24 Fn.2 system 57, 106 trial-and-error 18 Leibniz’s Law 88, 97 Lettvin, J. 170 Lewis, D. 163–5 Locke, J. 188 Loewer, B. 141, 167 Lorentz, K. 173 Macdonald, G. 199 Maclaurin, J. 224 magnetosomes 203–4 malfunction 11, 12, 21, 73, 229, 194 map cognitive 51, 55–9, 65 inner 50, 55, 58, 61 road 197, 201, 204, 206 strip- 56–7 Marr, D. 190 memory 25, 47, 49, 83, 155 Millikan, R. 5, 6, 12–13, 16–20, 24, 33–5, 48–9, 52, 61–5, 115, 117–20, 123, 137 fn.12, 151–53, 157, 162, 168, 187, 190, 196, 209–12, 216, 219 fn.10, 222 fn.12 misrepresentation 4, 12, 59, 62, 73–4, 149, 167–8, 184 models 121–42 module 25–8, 30–3, 35–6, 211, 214 modularity 23–4, 26, 146 Nadel, L. 57, 59 natural kinds 115–42
Index naturalism 1, 2, 43, 48 -istic 1, 3, 4, 17, 42–3, 50, 59, 60, 69, 71–3 Neander, K. 196, 209 negation, 111–12 internal 112 Nettle, D. 29 neuroethology, 169–186 neuroscience 56–7, 199 normative 12, 62, 72–4, 98, 151 -ity 62, 73 fn.5, 121–2 Oatley, K. 214 O’Keefe, J. 57, 59 ontogeny 12, 37 opacity 87, 97 opaque 18, 85–90, 98 Papineau, D. 5–6, 42, 64, 86 fn.1, 89–90, 92, 97, 123, 168 fn.3, 172, 196, 222 fn.12 Peacocke, C. 18, 95 fn.8, 108, 222 fn.12 perceptual experience 19, 69–70, 147–50, 154, 157, 160, 165 Percival, P. 223–5 phenomenal 69, 91 phylogeny 12, 37 Pierce, C.S. 52 Pietrowski, P. 19 Plato 205 Poirier, P. 199, 205 Port, R. 195 Price, C. 107–08, 168 fn.3, 174 fn.5, 186 Prinz, J. 116, 119 propositional attitude 46, 69 fn.1, 76, 205 pyramidal cells 128–42 Quine, W.v.O. 155 reduction 3, 90, 118–20 reference 24, 64, 95 fn.9, 195, 204–05 regress 47–8, 53, 113 reinforcement 103–04, 123 representation (see also misrepresentation) asymmetric dependence theory of, 116, 187 cognitive 4 inner 61–2 internal 56
231 mental 1, 4–6, 14, 17, 42–5, 49, 51–2, 56, 59, 107, 115, 121–22, 142, 167, 195 fn, 200, 206 public 42, 44, 46, 48, 53 pushmi-pullyu 153–4, 162, 165 semantic 8 visual 171 representational content 71–2, 142, 172, 196, 197, 201, 203 model 45–50, 54–5, 59–62, 65–6 properties 3, 58, 122 system 16, 100, 105, 109–10, 120, 186 resemblance relation 45, 47–8, 51–2, 56, 64 Russell, B. 83, 112 fn.7 selection 10–14, 16–18, 20, 29, 31–2, 38, 60–3, 65, 85–94, 96, 98, 102–04, 109, 117–18, 120–03, 157, 185–87, 196, 199, 203–04 non-genetic 13–14 ontogenetic- 15 secondary 15 selectional facts 86, 89 process 123 properties 86, 90–4, 96 recruitment 117 relevance 120 roles 18, 89–92, 96 theories of content 87, 89, 94, 98 Sellars, W. 45–6, 65 semantic conceptual role semantics, 115 consumer semantics 100 description 44, 59, 60, 66 indicator semantics 4, 6, 11, 117, 149, 151 informational semantics 43, 115–16 meta-, 148–9, 151 properties 17, 42–4, 48 rules 105 success semantics 5 SINBAD theory, 125–42 Skyrms, B. 163–5 Stampe, D. 4, 201 fn.6 Sterelny, K. 57, 168, 170 stimulation 8, 155, 171, 175, 178 proximal 109–11, 205
232 stimulus 19, 25, 140 fn.16, 148, 154–55, 161, 164, 172–78, 181–90, 200 poverty of 31 stimuli 8, 57, 146, 156–57, 170–71, 181–84, 189, 197 Stocker, M. 217, 220 substances, 120 supervene 2–3, 75 teleological 94, 100–01, 168, 187, 196 theory of content 85–8, 90–1 telepathy 17, 33–4
Index Tinbergen, N. 173 toad visual system, 169–192 Tolman, E. 55–7 Tooby, J. 23, 26 van Gelder, R. 195 Velleman, D. 159–60 verificationist 7, 187 Wellman, H. 188 Wiesel, T. 200 Wittgenstein, L. 53, 62 Wright, L. 61