Junk Representations Lawrence

Many philosophers and psychologists who approach the issue of representation from a computational or measurement theoretical perspective end up having to deny the possibility of junk representations—representations present in an organism's head but that enter into no psychological processes or produce no behaviour. However, I argue, a more functional perspective makes the possibility of junk representations intuitively quite plausible—so much so that we may wish to question those views of representation that preclude the possibility of junk representations. I explore some of the reasons we should care about the possibility of junk representations and conclude with some speculation about whether junk representations are in fact present in our heads. 1 An introduction to junk representations 2 An example of junk representations 3 Objections to junk representations 3.1 The computational objection 3.2 The measurement theory objection 3.3 Millikan 's objection 4 Conclusion: in search of junk representations 1 An introduction to junk representations Are junk representations possible? My use of the term 'junk' is intended to conjure in the reader thoughts of junk DNA. Biologists accept the possibility that much of the DNA present in an organism's genome may serve no organismic function. This 'junk' DNA takes up room in the nuclei of an organism's cells, but it contributes nothing to the phenotype of the organism. It's like the cobwebbed stationary bicycle you trip over in your basement: it's there, but it might as well not be. In contrast to junk DNA, many philosophers and psychologists fiercely resist the possibility of junk representations. Indeed, theories of representation that agree on little else agree that junk representations are an impossibility. However, I believe that there is a way to understand representation that is consistent with the possibility of junk representations. In the following pages I defend the possibility of junk representations. This project is important for several reasons. First, as I just noted, within both philosophy and psychology the view that representation must be coupled with processing has become an entrenched doctrine. No one sees the need to defend © Oxford University Press 1997 346 Lawrence A. Shapiro or question it. However, if I am right about the possibility of junk representations then any philosophical or psychological account of representation that builds upon the computer metaphor needs to be rethought. Likewise, the possibility of junk representations challenges anyone who thinks representations must produce behaviour. As often has been the case in science and philosophy, it is reflection upon fundamental assumptions that inspires new avenues of investigation. It is my hope that consideration of junk representations will open new paths of exploration within the cognitive sciences. Secondly, the possibility of junk representations may play a pivotal role in settling disputes within psychology that hinge in part upon disagreement over the nature of representation. So, for instance, the possibility of junk representations has significance for the debate between connectionists and classicists in so far as classicists think the notion of representation is incoherent without mention of a CPU that, by using the representation, gives it meaning. As we will see below, the possibility of junk representations suggests a view of representation that is compatible with connectionist architectures. Hence, the classicist who dismisses connectionism on the grounds that connectionist computers cannot support representation must rest her case on other arguments. Furthermore, some philosophers believe that representations should be attributed to an organism only as a last resort—only when explanations of behaviour or cognition are impossible without them. These philosophers tend to be anti-realists about representation. Dennett, for instance, believes representations are merely posits investigators rely upon for successful explanation and prediction: 'all there is to really and truly believing that p (for any proposition p) is being an intentional system for which p occurs as a belief in the best (most predictive) interpretation' (Dennett [1987a], p. 29, his emphasis). However, the possibility of junk representations suggests that representations can be more than mere explanatory posits. This follows, because, by definition, junk representations do not cause behaviour and so cannot be necessary for the explanation or prediction of behaviour. In this respect, the possibility of junk representations suggests the kind of opportunity for psychologists that the discovery of DNA prompted for molecular biologists. Just as genes, which were originally introduced as a tool for predicting the inheritance of phenotypic traits, can now be studied without regard for their phenotypic effects, the possibility of junk representations should encourage us to make representations objects of investigation in their own right—independently of whatever behavioral effects they may or may not produce. Before going any further, I should acknowledge that the view of representation I shall be assuming is teleological in character. Accordingly, those who reject teleological accounts of representation may dismiss my conclusions because of doubts about the foundation on which they rest. In an effort to speak to this audience, I have chosen the following strategy. In the next section I Junk Representations 347 attempt to make the existence of junk representations plausible with an example. It is my hope that the example persuades even teleophobes that there is something to junk representations. If it doesn't, still remaining is the important task of convincing teleophiles that junk representations belong in their ontology. 2 An example of junk representations Stephen Kosslyn's extraordinary work on mental imagery provides a convenient vehicle for illustrating the possibility of junk representations. On Kosslyn's theory, mental images are formed on a visual buffer consisting of numerous representational cells that together function like a circular space ([1980], pp. 84-5). When forming a mental image, according to Kosslyn, these representational cells get 'filled in' with bits of the image. To exploit the same analogy that inspired Kosslyn's theory, imagine a cathode ray tube (e.g. a computer monitor) that is circular in shape. The image depicted on a CRT is composed of thousands of dots, or pixels, each of which is a representational part of the image. These pixels are comparable to the representational cells of the visual buffer. Although to my knowledge Kosslyn never entertained the possibility of junk representations (indeed, for reasons we will discuss in the next section, Kosslyn is committed to their impossibility) the visual buffer can be conceived in a way that makes the possibility of junk representations plausible. Suppose that the visual buffer appears circular only because the psychological processes that 'view' the image on the buffer are limited in such a way that they can view only a circular chunk of the image. Analogously, imagine viewing a rectangular CRT through a tube of the sort you find in the centre of a roll of paper towels. If you are close enough to the screen, the corners will be 'cut off and the image you see will be circular in shape. I claim that the bits of the image in the corners of the CRT, as well as the bits of image in the corner cells of the rectangular visual buffer, are junk representations. They are representational parts of an image despite the fact, let us assume, that no processes are operating upon them and nothing is viewing them. Like the useless DNA crowding the nuclei of your cells, it is conceivable that there are junk representations taking up space in your brain. Of course, the claim that the cells in the corners of the visual buffer represent requires some defence. In particular, I need to defend the claim that these corner cells are representing despite the fact that nothing is viewing them. Towards such a defence, I shall assume that representational states are functional states, where 'functional' is to be understood in a teleological, purposive, sense. Moreover, whatever the function of a representational state may be, I shall assume this function includes carrying information about, or correlating with, 348 Lawrence A. Shapiro that which is represented. Whether all this can be made out in safely nonintentional terms does not concern me here. Many philosophers are of a mind that until we can explicate representation in nonintentional terms we ought to view talk of representation with suspicion. I disagree. Currently there is no commonly received analysis of representation in nonintentional terms and yet appeals to representation continue to bear fruit in empirical investigations of cognitive capacities (e.g. Kosslyn's work in imagery, Marr's [1982] work in vision). If pudding is proved by its eating, representation is proved by its usefulness in theories of mind. So, I propose now to offer two arguments to support my claim that the corner cells of the visual buffer are representations. Both arguments assume that representations are functional kinds. In particular, a representation is a state that has the function to carry information just as a screwdriver is a device that has the function to drive screws and a pen is a device that has the function to write. My strategy in the following arguments will be to argue that just as a screwdriver needn't be used to count as a screwdriver, and a pen needn't be used to count as a pen, a representation needn't be used to count as a representation. The first argument I call the Argument from Genesis. The individuation of functional kinds requires consideration of their origin. More specifically, I shall assume, a token belongs to the functional kind X if it is a reproduction of a token of kind X and part of an explanation for why tokens of X have reproduced (or have been reproduced) is some capacity that things of kind X can effect. Alternatively, something might count as a member of functional kind X if it is the product of something that has reproduced (or has been reproduced) because it has the capacity to produce things of kind X which, in turn, have some capacity that explains why their production is desirable (i.e. will explain why their producers continue to reproduce or be reproduced). Origins of the first sort tend to be relevant to the individuation of artefactual functions. So, for instance, we consider something to be a pen because it is a reproduction of something that has been reproduced because it has the capacity to write. Biological traits tend to be of the second sort: something is a wing because it is the product of certain genes that have reproduced because they are able to produce wings which confer some benefit upon their possessor (which contains the genes that code for wings). 1 A reviewer for this journal noted the following disanalogy between representations and items like pens and screwdrivers: pens are pens whether or not they perform their function of writing, but it doesn't make sense to say representations are representations whether or not they perform their function of carrying information—a representation that does not carry information cannot be a representation. The criticism seems correct. However, the point to draw from the analogy, as we shall see, is that functional kinds, whether pens or representations, are what they are for reasons of aetiology and not because someone or something makes use of their capacities. 2 I am here obviously indebted to Millikan's [1984] account of reproductively established families and proper functions. Junk Representations 349 Schematically, the Argument from Genesis has the following form: Argument from Genesis (1) X has function Y if X has the requisite genesis. (2) X has the requisite genesis. (3) Therefore, X has function Y. According to this argument, if the object that has been lost between your sofa cushions from almost the moment you bought it and so never had the opportunity to mark a piece of paper had the requisite genesis, it would be a pen. Pens don't need to be used to be pens. Pens have the function to write even if they are never used for this purpose because they are reproductions of things that have been reproduced because of their capacity to write. Similarly, although they are not used in any cognitive processes, the states of the cells in the coiners of the visual buffer are representations because they have been produced by a type of cognitive mechanism that continues to be present in organisms because of its capacity to produce representational states that the organism does use. These corner cell representations are like the sperm Millikan mentions so often ([1984], p. 29). Sperm that fail to fertilize an egg—that never do anything useful for the organism—are no less sperm than those that do end up fertilizing an egg. In short, because being used is not an individuating property of functions, there may be states that represent despite their uselessness. These states are junk representations. The second argument suggesting that the comer cells in the visual buffer represent is the Argument from Structure. This argument depends on the observation that structure is often a useful (but not full proof) guide to function. One needn't know the genesis of bat wings to be fairly sure that their function is flight: it's enough to know that they are structurally quite similar to other things that do have flight as their function. Schematically, the Argument from Structure is this: Argument from Structure (1) Things with structure S typically have function Y. (2) X has structure S. (3) Therefore it is likely that X has function Y. To apply the Argument from Structure to the corner cells of the visual buffer we need to suppose that these cells have the same structure as those cells of the buffer that are 'viewed'. In his most recent book, Kosslyn [1994] claims 3 In the case of natural mechanisms, it will be the 'intentions' of natural selection that ultimately ground ascriptions of function. A mechanism might have the function to produce representations because its continued presence in some type of animal is explained by its capacity to produce representations. 350 Lawrence A. Shapiro that the visual buffer is located in the occipital lobe, more specifically in areas VI-V4 of the visual cortex. Suppose our neuroscience were advanced enough to locate those cells of the occipital lobe that implemented the representational cells of the circularly shaped visual buffer. Imagine further that neuroscientists reported the discovery of neurons that behaved exactly like those neurons constituting the visual buffer but were not connected to the 'viewing' mechanism. These neurons, suppose, are physiologically identical to those neurons realizing the visual buffer that admittedly do represent and they are active when and only when the neurons in the visual buffer are active. Furthermore, suppose it were possible to predict the activity of these neurons from evidence about how neurons in the visual buffer behave when visually imaging certain objects (e.g. a red expanse) together with informed speculation about what features of the image would fall into the corners of the visual buffer were it in fact rectangular in shape. Surely the structural similarities between these 'disconnected' neurons and the neurons that are hooked up to the 'viewing' mechanism should lead us to conclude that they very likely have the same function. We should conclude from the investigation of the structure of the visual buffer that connected and disconnected groups of neurons alike represent features of a mental image. The Argument from Genesis and Argument from Structure provide, I think, compelling reasons to grant the possibility of junk representations. Indeed, I hope it seems somewhat puzzling that philosophers and psychologists would deny such a thing. Yet, deny it they do and it is now time to consider and respond to some objections philosophers and psychologists have made to the possibility of junk representations. 3 Objections to junk representations In this section I will consider three objections to junk representations. The first follows from a computational view of representation and the second from a measurement theoretical view. Both objections, I shall argue, miss their mark because they fail to appreciate the sense in which representation is a functional achievement. The third objection comes from Ruth Millikan, a philosopher who has done more than any other to analyse representation in terms of 4 A reviewer for this journal pointed out that the Argument from Genesis and the Argument from Structure might be related in the following way. Quite possibly the mechanism that produces representations produces other things as well. Hence, the claim that some of the unused things the mechanism produces are representations requires additional evidence; Appeals to structure can provide this sort of evidence. While I agree that appeals to structure may be necessary to say which of the effects of a representation producer are representations, it may none the less be true that some of the mechanism's unused effects are representations despite structural dissimilarities with representations that are used. Indeed, perhaps it is because of these structural dissimilarities that the representations cannot be used. Likewise, even a horribly malformed pen is a pen, and this is because it was built for writing even if it cannot. Junk Representations 351 function. In responding to Millikan, I will be trying to convince her that junk representations are in fact consistent with her view. 3.1 The computational objection Most who resist the possibility of junk representations do so, I suspect, because they adhere to a notion of representation that follows from viewing the mind as a classical computer. In a classical computer a token string of Is and Os counts as a symbol only relative to a set of instructions that specify how the string can be manipulated and interact with other strings. It is the CPU that reads the instructions and carries out those operations upon the string that make it a symbol, that permit an assignment of representational content to it. Note that some tokens of a single type of string may be symbols while other tokens may not—a token's status as a symbol depends upon whether the CPU is acting on it according to an appropriate set of instructions, i.e. a set of instructions under which it makes sense to assign to the string an interpretation (see Dennett [1987b]; Fodor [1975]; Haugeland [1985]; Kosslyn and Hatfield [1984]; Pylyshyn [1984, 1989]). Pylyshyn [1984] provides a nice illustration of the representation/processing interdependence that is central in computational theories of representation. The example begins with the supposition that there exists a mapping between physical states of a computer with symbolic expressions. These expressions include tilings like 'o', 'ox', 'xo', 'xx', 'oox', 'oxo', etc. These expressions do not yet mean anything, however, because they do not participate in any operations. But, Let us further suppose that when a certain pattern (designated by the symbol 'ffi') occurs in a portion of the machine called its 'instruction register', the machine's memory registers change states according to a certain, specifiable regularity. For example, when the portion of the machine called register 1 is in the state that maps onto the string xox, and register 2 is in the state that maps onto the string xxo, then register 3 changes its state from whatever it was to the state corresponding to the string xoxx. This regularity might be used to represent addition of numbers, provided we adopt an appropriate semantic function, SF, and that the regularity meets certain requirements. Here, the required semantic function is easily defined; it happens to be the function that maps strings of o's and x's onto numbers, namely, the binary number system (Pylyshyn [1984], pp. 59-60, his emphasis). Pylyshyn's example makes perspicuous the sense in which computational theories of representation bind representation to processing. Until a symbol engages in operations of various sorts and thereby exhibits a regularity that makes it interpretable it has no meaning. Just as (it is said) sharks must move constantly or else drown, symbols in a classical computer must participate in processes or else lack meaning. 352 Lawrence A. Shapiro Obviously, on the view of representation deriving from the classical theory of computation, junk representations are an impossibility. By assumption, the corner cells of the visual buffer are not being operated upon by the biological equivalent of an appropriately programmed CPU. There are, ex hypothesi, no processes in which the corner cells are participants. And, since it is rule-guided processing that, as Kosslyn notes, 'allows stored information to be "interpreted"' ([1983], p. 25), these corner cells cannot be said to represent anything. To turn a popular slogan on its head, it appears there can be no representation without computation. As I understand the computational objection, it is making the following sort of conceptual point. Nothing represents intrinsically. The red light that tells you to stop doesn't intrinsically mean stop. It means stop because of the presence of various conventions that prescribe its use as a stop signal. Thus, being a representation requires that there be some rules of use separate from the representation that gives the representation its meaning, and this is as true for the symbols in a computer as it is for the street lights that regulate traffic. While in the latter case it is individuals who determine the meaning of street lights by their decisions about how to use them, in the former case it is the computational processes that determine whether a symbol will be a representation. But, on the view of representation I sketched above, for r to be a representation it suffices that r is produced by something that has been reproduced because it produces representations. Just as a pen need not have a user to make it a pen, a representation need not have a user to make it a representation. Consequently, it is possible to agree with the (plausible) claim that nothing represents intrinsically and yet deny that use is essential to representation. Facts about genesis are relevant for deciding whether a given state has the function to represent, and these facts are independent of facts about whether the state is participating in some computational (or cognitive) process. If the classical view of computation has it that use is necessary for representation, so much the worse for the classical view. Before turning to the next objection to junk representations, it is worth pausing to consider a response a computationalist might make to the discussion above. I have claimed that the cells composing the corners of the invented rectangular visual buffer are junk representations because they represent even though the viewing operations Kosslyn discusses do not operate upon them— nothing makes use of them. Yet, the response goes, on the computational view of representation it does make sense to consider some strings of 1 s and Os to be representations even when the CPU is not currently processing them in virtue of the fact that the CPU can process them. Strings in RAM, for instance, might count as symbols because the CPU has access to them and may at some point use them. Likewise, although I have portrayed the corner cells of the visual buffer as inaccessible to the viewing processes, what really matters to their Junk Representations 353 status as representations is that they could be used if they were available for viewing. So, the response goes, use is essential to representation, however current use is not. Lloyd [1989] provides a nice illustration of this view. Lloyd begins with a denunciation of junk representations: 'in psychology, representations must be put to work. They must play a role in the explanation of behavior, or there will be no psychological function for them' (ibid., p. 18). Lloyd then offers the caveat sketched above: But the requirement that representations have a cognitive role in the systems which contain them is also a capacity that may go unrealized in specific cases. Particular representations may evaporate without a trace. What is necessary, however, is that the system in which they occurred was one in which they could have made a difference, either to other representations or to behaviour (ibid., p. 18). On this more liberal interpretation of use Lloyd might agree that the corner cells of the visual buffer I described are representations; moreover he'd agree that the corner cells are not at the moment being used. However, Lloyd would say, these cells owe their status as representations to the fact that they could, conceivably, be put to use at some time. Hence, Lloyd might conclude, my description of junk representations must be refined. If by junk representations I mean representations that never get used but could, counterfactually, be used by the cognitive system of which they are a part, then Lloyd would not dispute their possibility. However, if by junk representations I mean representations that could never be used then, Lloyd would insist, such things are impossible— they cannot be representations. As tenable as the response may seem, a bit of scrutiny shows its position on the relation between representation and use to be either vacuous or questionbegging; hence, it does not rescue the computational account from its inability to accommodate junk representations. The vacuity of the response becomes apparent when we ask whether merely counterfactual use is a sufficient condition for making something a representation. Is something a representation simply because it could be used to carry information? The difficulty with this view is that virtually anything could be used to carry information. Information, in the technical sense that I take to be relevant to representation, is every where. The stomach, for instance, may by its increase or decrease in volume carry information about the amount of food it contains. Because it could be used to carry this information, ought we to conclude that the size of the stomach does, at this moment, represent the amount of food contained within it? Likewise, any string of Is and Os could be used to represent, but surely not all strings of Is and Os are representations. Furthermore, depending 5 We may read 'behaviour' generally enough to include things other than bodily movement. Hence, the behaviour of a CPU is, on this understanding, still behaviour. 354 Lawrence A. Shapiro upon the state of the CPU, any given string of Is and Os could represent many things or nothing. Relative to which state of the CPU are we to judge that a string of digits represents? Being the sort of thing that could be used to carry information is simply too profligate a criterion by which to judge representational status and so it is to actual use that the computationalist must appeal. On the other hand, the claim that counterfactual use is a necessary condition for something's being a representation begs the question against junk representations. On the perspective I have taken, representations are functional kinds and have their function in virtue of their aetiology. Like pens, screwdrivers, and sperm, representations owe their function to their special kind of history—to the fact that they were made by something that, in turn, has the function to produce representations. The discussion of the fictitious visual buffer is intended to suggest that the effects of a representation—its potential uses—do not determine its functional nature. Insofar as the discussion of the visual buffer compels the intuition that the corner cells of the visual buffer are representations, the burden rests on the opponent of junk representations to support her claim that use is a necessary condition for representation. Of course, one could simply deny that they find the example of junk representations at all compelling, but this still leaves unanswered the following question: Why should use count as an individuating property of the function of representation? If, as I suspect, use is rarely an individuating property of a functional kind, it falls to the opponent of junk representations to say why the functional nature of representational states is so terribly unique. 3.2 The measurement theory objection Like Kosslyn, the psychologist Stephen Palmer believes '[i]t is axiomatic within an information-processing framework that one cannot discuss representation without considering processes . . . Without those processes, the representation is meaningless' ([1978], p. 265). However, Palmer believes junk representations are impossible not because of any commitment to the computational theory of mind, but because he subscribes to a measurement theory analysis of representation. According to Palmer's measurement theory analysis, '[a] world, X, is a representation of another world, Y, if at least some of the relations for objects of X are preserved by relations for corresponding objects of Y' (ibid., p. 267). Representation, on this view, involves a mapping of relations between elements in one domain upon relations between elements in another. Processing is essential, Palmer believes, because there must be something that establishes the mapping of one set of relations onto the other. So, for instance, Palmer discusses an example in which line lengths (in the representing world) represent the heights of rectangles (in the represented world). However, there is nothing intrinsic about relationships between line Junk Representations 355 lengths that make them suitable candidates for representing block heights. "The height of rectangles, for example, might be represented by line length where there is no intuitively obvious relationship (such as our usual concepts of "longer than" or "shorter than") to model "taller than'" (ibid., p. 265). So, what justifies the claim that lines of particular lengths represent rectangles of particular heights? The answer to this question introduces the importance of processing in Palmer's account of representation: 'if there is a process that interprets the length of lines—whatever they may be—such that corresponding lengths are functionally ordered just as the rectangle heights are ordered, then there is an operational relation defined by this process that corresponds to the "taller than" relation in the represented world' (ibid., p. 265). In short, processing establishes an isomorphism between relations of objects in the representing world and relations of objects in the represented world. It is in virtue of this isomorphism that objects in the former domain represent objects in the latter. In spirit, the measurement theory argument for the representation-processing pair does not differ from the computational theorist's argument. Palmer, like the computational theorist, is concerned to deny that things represent simpliciter. Just as the operations a CPU performs upon a string of 1 s and Os make the string a symbol by making the string interpretable as a symbol, some sort of operational processes are necessary to determine which relations between worlds define representations. 'Longer than' means 'taller than' only relative to a process that uses 'longer than' to mean 'taller than'. Accordingly, junk representations are impossible because they are, by definition, not processed by anything. In so far as the measurement theory analysis of representation subscribes to the same conceptual point grounding the computational theorist's demand for a representation-processing pair, my response to the computational theorist works here as well. One needn't look to processing to determine representation. Something can be a representation simply because its function is to carry information. Whether processes actually make use of a representation is an issue quite separate from whether something is a representation, just as whether you ever find the pen you dropped between your sofa cushions is an issue quite separate from whether that object between the sofa cushions is a pen. 3.3 Millikan's objection Despite Millikan's staunchly teleological analysis of representation, she seems adverse to junk representations: If it is actually one of a system's functions to produce representations, as 6 Indeed, Palmer sees his operational relations as close kin to Pylyshyn's semantic interpretation functions ([1978], p. 265). 356 Lawrence A. Shapiro we have said, these representations must Junction as representations for the system itself. Let us view the system, then, as divided into two parts or two aspects, one of which produces representations for the other to consume. What we need to look at is the consumer part, at what it is to use a thing as a representation. Indeed, a good look at the consumer part of the system ought to be all that is needed to determine not only representational status but representational content ([1989], pp. 285-6, my emphasis). In her discussions of representation Millikan places heavy emphasis upon the need for a representation consumer—a device that will actually use a representation. Representation consumers are necessary for making something a representation, Millikan believes, because, first, 'not every device whose job description includes producing items that vary with the world is a representation producer. The devices in me that produce calluses are supposed to vary their placement according to where the friction is, but calluses are not representations' (ibid., p. 283). In order to make a state that varies with states of the world a representation, Millikan thinks, there must be something that makes use of the state because it varies with the world. This is the first job of the representation consumer. Secondly, because everything, representation or not, varies with the world in numerous ways, a representation consumer is necessary to define which variations are relevant for fixing content. The relevant variations are simply the ones that must be maintained in order for the representation consumer to perform its function properly (ibid., p. 287). Given the double role Millikan assigns to representation consumers in her analysis of representation, it appears that she must deny that the corner cells of the visual buffer are representations. With nothing to consume them, why suppose their variations with things in the world constitute representations of these things? Which variations are the ones that define their content? Though I think Millikan's concerns are well motivated, I also believe it is possible to define representations independently of relations to current representation consumers. Consider, for example, a photograph of the Parthenon. Suppose this photograph is, for some reason, never viewed (after being developed it's lost in the mail for ever). I claim that this photograph none the less represents the Parthenon. How can it without the presence of a representation consumer? Recall that Millikan thinks a representation consumer is necessary to (i) justify the claim that a varying state is a representation and (ii) define the content of the varying state. Let's now consider whether we can satisfy (i) and (ii) without requiring the presence of a representation consumer. Is there a motivation for saying that the never-viewed photograph of the Parthenon is, first, a representation, and, second, a representation of the Parthenon? The justification for the claim that the photograph is a representation comes from consideration of the fact that the photograph was produced by a Junk Representations 357 device—a camera—that has the function of producing representations. Of course, it is possible that we might fail to recognize that something is a representation because nothing uses it as a representation and because we are ignorant of its origins. But, presumably, Millikan requires the presence of representation consumers for more than epistemological reasons. It won't do to say that representation consumers are necessary to make something a representation if in fact they are necessary only to make apparent that something is a representation. If representations are defined by their functions, and their functions do not include that they be used in a particular way, then representation consumers are no more relevant to the individuation of something as a representation than carpenters are relevant to the individuation of something as a screwdriver. Turning now to the question of the photograph's content, we see there is a need to invoke a representation consumer; however, I submit, the representation consumer can determine the content of a given representation without having to consume (i.e. use) it. Assume that people want pictorial representations of objects in their environments. In pursuit of these desires, the camera is invented. Now, as Millikan observes, everything varies in many ways with the world, and this is no less true of photographs. How are we to say which of the photograph's variations with the world are the relevant ones for determining its representational content? Relevance is always relative, and Millikan determines which variations are relevant by appeal to the needs of the consumer. So, the explanation for why a photograph of the Parthenon is a representation of the Parthenon (rather than, say, a representation of some pattern of light on the lens) is, roughly, because it is the variation of the photograph with the Parthenon that satisfies the representation consumer's desire for a pictorial representation of that object. In Millikan's terms, photographs represent the object that reflects the light which has exposed the film because it is by varying with that object that photographs 'give' consumers what they want. However, these considerations do not imply that a token representation has no content prior to its consumption. If Millikan's story about what fixes the content of photographs is correct, then there is a fact of the matter about which relations between a photograph and the world define the photograph's representational content, whether or not a consumer happens to be present. Photographs represent the objects they depict because it is by performing this job that 7 Cameras produce flashes as well, but cameras produce these only because they must do so in order to produce the representations that they must produce if they are going to continue to be reproduced. 8 I put it in this awkward way rather than appealing to something like resemblance because I wish to avoid the consequence that a photograph of a Pepsi can represents any Pepsi can. I think photographs represent those objects that cause the exposure of the film. 358 Lawrence A. Shapiro they would satisfy a consumer's needs were a consumer present. Likewise, if the viewer of the visual buffer can satisfy its needs by exploiting the fact that the cells of the visual buffer are correlated with that which causes in them certain sorts of activity, then corner cells of the visual buffer that are exhibiting these sorts of activity represent particular things despite never being viewed. So, Millikan might be right that a representation consumer is necessary to select those relations that establish the content of representational types (e.g. photographs or cells of the visual buffer), but she is wrong if she takes this to imply that a given representational token has no content until it is used by a representation consumer. Once the relevant relations that determine representational content are fixed, token representations (e.g. a photograph of the Parthenon or particular states of cells in the corners of the visual buffer) may have content in the absence of consumers. 4 Conclusion: in search of junk representations In the introductory section I provided several reasons to care about the possibility of junk representations. In accepting their possibility, one is calling into question views of representation that have been inspired by the computer metaphor (or measurement theory) or that require that representations produce some kind of behaviour. I also noted that in so far as we accept the possibility of junk representations, we are taking sides on the debate between classicists and connectionists.BBecause the classical environment is hostile to junk representations, ifVswe think that junk representations are possible we should prefer cognitive architectures that allow for their existence. To be sure, there is no consensus among philosophers or psychologists about how we are to understand representation in connectionist nets, but those accounts of representation that have been offered are consistent with the possibility of junk representations (e.g. Hatfield [1991]). Finally, I noted that the possibility of junk representations is important because such a possibility suggests that we can investigate representation outside contexts of psychological explanation. Philosophers have often defended attributions of representational states with the claim that they are indispensable in psychological explanations. Some, however, have extended this claim, arguing that representations are to be valued purely for their instrumental utility (Dennett [1978, 1987a]). Yet, in accepting the possibility of junk representations, one is granting that representations need not partake in explanations of psychological capacities: there is more to representation than explanatory .utility. Hence, recognition of junk 9 Earlier in response to the computational objection I argued that counterfactual use could not suffice to make something a representation,'lest everything becomes a representation. In the present context, I am appealing to counterfactual use to fix representational content, which differs from representational status—a point Millikan appreciates. Junk Representations 359 representations justifies questions about the nature of representational states themselves apart from their explanatory utility. So far, of course, I have considered only the: possibility of junk representations. But, do these things really exist? Before answering this question, I must urge that the mere possibility of junk representations is significant. This follows, for junk representations are possible only under a certain conception of what representation is and what it is not. Of real importance is not whether junk representations do in fact exist, but whether the approach to representation that affords their possibility is the right one. If we agree they are possible, we agree that there is something wrong or incomplete about standard computational/measurement theory analyses of representation. Moreover, we agree that representations should not be maligned as mere explanatory posits. But what of the dispute between classicists and connectionists? Here matters become a bit more difficult. For our purposes, the debate between classicists and connectionists is over the kind of architecture that underlies cognition. Accordingly, an open-minded classicist might agree that junk representations are possible under some analyses of representation, but unless there really are junk representations in human heads, the classicist might observe, there is no reason to prefer the connectionist architecture to the classical one. For this reason it would indeed be valuable to know whether junk representations do in fact exist. So, are there junk representations? Quite naturally, there is no explicit discussion of such things in neuroscientific texts. Why should the neuroscientist care about something that doesn't do anything? Moreover, direct enquiries of neuroscientists have borne more frustration than fruit. Nevertheless, my quest for junk representations has uncovered the following. Many aphasias and vision deficiencies occur because representations that need to be used are not. For instance, there are varieties of colour blindness that are due to cortical damage (Plant [1991]). In such cases, the cells in the retina that represent colour continue to do so, but their, representational content is ignored in the later stages of colour processing. More telling, junk representations appear to be present in normal adults. Individuals exhibit vast differences in peripheral vision. However, most of these differences are a result of cortical, rather than retinal, processes. At the retinal level, people with narrower visual fields have the same representational capacities as people with wider visual fields, but because their higher visual processes ignore some of the information their retinas provide, their peripheral vision is inferior. Indeed, we might suppose, whenever a representational capacity, involves the hierarchical processing of information there must be a chance that junk representations will be cast off at various stages along the way. 10 The neuroscientist Deric Bownds provided me with this example. 360 Lawrence A. Shapiro These examples of junk representations are of course not as neat as the one I invented in Section 2. In particular, they do not illustrate a species-wide organ that routinely spins off a predictable quantity of junk representations. However, the peripheral vision example comes close. If it is true for no individual that all retinal information makes it to the cortex, then we have a case where every individual will have in their visual systems some junk representations. Whether these examples suffice to cast further doubt upon the computationalist conception of representation is, I suppose, for the reader to decide. At the very least I hope these examples inspire further thought about junk representations and their consequences for philosophy and psychology.