This paper was published in Proceedings of the Conference on Cognition and Representation, Technical Report of the SUNY at Buffalo Center for Cognitive Science, 1992.
Many of you may have had a neighbour or a grandfather who was a fanatic about tools. Grandpa owned everything Sears ever made, and took it as a sign of moral turpitude if you used, say, a flathead screwdriver to take out a Phillips-head screw. Now I confess that I still do not own a Phillips screwdriver, and also use my flathead screwdriver to open paint cans. But Grandpa did rub off on me a little bit: I eventually became a philosopher, and take it that an important part of what philosophers do consists in the examination of conceptual tools.
What I am going to do here is look at one of the main conceptual tools used in cognitive science: namely, the notion of representation. Due to limitations of time, the discussion will be be very programmatic in nature, and much of the detail that might be supplied in a longer discussion is left out. I hope to persuade you of three main points:
First, that there is real unclarity and even ambiguity in the usage of the word 'representation'
Second, that different notions of representation are suited to different tasks&emdash;in particular, that some philosophical projects require a stronger sense of the word than do most empirical theories.
Third, that it is possible to adopt a weak construal of the word 'representation' that lets us have a good scientific psychology and lets us have intentional realism too, but without providing a representational explanation of intentionality.
Now my interest in clarifying the very notion of representation may come as a bit of a surprise. There seems to be a widespread assumption in cognitive science circles that we know perfectly well what we mean when we say something "is a representation," and hence the only interesting questions are about what sorts of representations are used, how they are realized in the brain, and how they get this peculiar property of representing something. And the assumption that we know what it is to be a representation is quite understandable, since the notion of representation looks like an old friend. After all, we have known about representation as long as we have known about symbols and pictures and maps and leitmotifs. Anybody who knows English knows what we mean when we say that a picture or a symbol is a representation.
While this attitude is perfectly natural, it also involves some implicit (and arguably none-too-innocent) assumptions about the word 'representation.' I shall call this set of assumptions the "Received View."
1) There is a single, univocal and familiar notion called 'representation.'
2) There is a corresponding property called "being a representation."
3) There are a number of paradigm examples of things that possess this property, e.g.:
- schematic drawings
- leitmotifs, etc.
4) Philosophers and psychologists have long known that mind/world relations are mediated by mental representations.
5) Until recently, it was impossible to hook up a representational theory of content with a causal account of mental processes.
6) The paradigm of machine computation has shown us how to make this connection.
1. Problems with the Received View
Once we spell things out this way, some of the assumptions begin to look a bit dubious. First, it is undoubtedly true that there is an abundance of paradigm examples to which the word 'representation' is applied. But it is not at all clear that the same notion is expressed by that word in different cases, nor that there is a property called "being a representation" that is shared in common by pictures, symbols, and the rest. It seems just as likely that the word 'representation' is homonymous, or that it signifies a family resemblance rather than a common property. So here is the first problem: it is not clear that there is one thing that we are saying even of all the familiar cases when we call them "representations." And hence it is not clear what we are saying of the new cases postulated by cognitive science when we call them "representations."
Second, some people have tried to articulate an analysis of 'representation' broad enough to cover the paradigm cases. I am thinking of a diverse group of writers that includes Thomas Reid , Edmund Husserl , A.J.Ayer , Daniel Dennett , Kieth Lehrer  and Brenda Judge . According to these people, what it is to be a representation is to be used and interpreted in a particular way:
To say that A is a representation of B is (implicitly) to say that there is some interpreter I who uses A to stand for B.
As these writers have been quick to point out, there are disastrous consequences if you try to apply this interpretation-dependent notion of representation as an explanation of meaningful mental states, as it leads to a regress of homuncular interpreters, and requires you to assume the very things you are trying to explain&emdash;in particular, the mind's access to extra-mental reality. I shall call this line of criticism the Reid/Husserl objection. One might see the complaints about the interpretation-dependence and "derived intentionality" of mental symbols raised by John Searle [1980, 1984] and Kenneth Sayre  as falling essentially along these same lines.
It is important to emphasize that the point of the Reid/Husserl objection is not that pictures and symbols and the like have two separate features: one of being representations and another of being convention- and interpretation-dependent. The point, rather, is that this convention- and interpretation-dependence is built right into the notion of 'representation.' On this view, someone who speaks of "non-interpretation-dependent representations" is merely showing that she does not know how the word 'representation' is used. It would be a mistake on the order of saying "tool without a purpose." What it is to be a tool is precisely to be usable for some purpose, and what it is to be a representation is precisely to be used to stand for something else.
So now we have two reasons to be interested in an analysis of the notion&emdash;or, better, of various alternative notions&emdash;of representation:
1) We may already have several usages of the word, and hence it is not clear which, if any, is being used when we speak of "mental representations."
2) There is an argument of respectable lineage to the effect that a correct understanding of the notion of representation reveals a kind of dependence on interpretation that would render that notion useless for explanation in cognitive science.
To these two reasons I shall add a third: namely, that there is some danger that we may argue for the need for one notion of representation, then invalidly draw conclusions based on a subtle elision to an alternative meaning of the word. We might, for example, argue that we need something like mental pictures, then argue as though we had proven that our "representations" had to have syntactic features, or argue that thought involves "representing the world as being thus" and then proceed upon the stronger assumption that thought involves entities called "representations" having semantic properties of the same sort that are attributed to symbols. In these sorts of cases, there would be an ambiguity of the term 'representation' in the argument that would render the argument invalid and paralogistic.
2. The Search for a Rule for the Use of 'Representation'
The task, I take it, is to supply a rule for the usage of the word 'representation' such that:
a) it is adequate to the explanatory tasks to which the word is put in theories in cognitive science, and
b) it manages to avoid the Reid/Husserl objection
Now I am aware of four basic strategies for supplying a rule for the use of a word:
- Continuity with ordinary usage
- Stipulative definition
- Theoretical definition
Some of these strategies seem to be present in the philosophical literature on representation, and others are not. I take it, for example, that Fodor's language of thought hypothesis involves an extension of the familiar notion of symbolic representation to a new domain, and that Fodor [1975, 1981, 1987] intends us to apply what we already understand by the word 'representation' when it is applied to symbols when we conceive of "mental representations"&emdash;in particular, we are to think of them as entities having both syntactic and semantic properties. By contrast, Robert Cummins  develops a technical notion of "S-representation" which could well serve as a stipulative definition of the word 'representation' as Cummins uses the term. It seems rather doubtful that the third method, the method of ostensive definition, would be used for mental representations, since these are generally regarded as theoretical entities, and theoretical entities are, in principle, difficult to point to. The very fact that mental representations are conceived as theoretical entities, however, suggests the fourth strategy: namely, that we regard the expression 'mental representation' as being an expression that is theoretically defined.
It is perhaps important to be clear about what I mean by a "theoretical definition" of a term, and how this might work with respect to the word 'representation'. What I mean by "theoretical definition" is that sometimes the best way to determine what a technical expression e means in a theory T is to forget entirely about what it means in other contexts, and look exclusively at the explanatory work it does in T. When a term's meaning is entirely determined by the explanatory work it does in a theory, I shall say that it is "theoretically defined." A basic form of such a definition for the word 'representation' might look something like this:
'representation' in theory T =df "whatever it is that does x"
where 'x' names some subset of the things explained by T. For example, 'x' might mean one or more of the following:
- accounts for the intentionality of mental states
- stands in appropriate informational relationships with environmental factors
- tracks enviornmental factors in a fashion that allows for adaptation, etc.
Of course, one should not assume from the outset that the word 'representation' does the same work in every theory in which it is used.
3. Robust and Theoretical Construals Constrasted
Granted that there is a dry philosophical question to be asked about different senses of the word 'representation,' we might well ask who ought to care and why. The first part of my answer is that I think the Reid/Husserl objection needs to be taken seriously. If theoretical work in cognitive science really depends on the notion of 'representation' that Reid and Husserl articulate, it is built on an untenable conceptual foundation. (I think some people have really steered clear of cognitive science because of just this kind of worry.) And the best way to show that it does not depend on that notion is to articulate an alternative sense for the word 'representation' that fits the theories and escapes the problems of conventionality.
But there is also another issue here. There are important differences between what you can do with a theory that takes 'representation' as a theoretical term and what you can do if you build some semantic presuppositions into your definition of 'representation,' either by continuity with familiar paradigms or by stipulative definition. There are six main points about these differences:
1) The structure of the explanations is different for the two cases.
2) The theoretical construal involves no commitments to dubious semantic properties.
3) The theoretical construal avoids the Reid/Husserl objection entirely.
4) The theoretical construal does not explain the intentionality of mental states.
5) The theoretical construal is very attractive as an account of the use of the word 'representation' in empirical theories in cognitive science.
6) The theoretical construal does not do all that is desired by some philosophers.
To see how the structure of the explanations is affected, consider as an example how theoretical definition differs from making wholesale use of the familiar notion of symbolic representation. The interpretation of cognitive science most familiar to philosophers treats intentional states as relations to symbolic representations that may literally be said to have both syntactic and semantic properties. Now part of what makes the notion of symbolic representation so interesting is that, in addition to having the syntactic overtones needed for computation, it also has semantic overtones built into it. To call something a "symbolic representation" is to impute to it semantic properties. And semantic properties seem like the right kinds of things to explain semantic properties: hence one can plausibly view the semantic properties of mental states as "inherited" from those of the symbolic representations they contain.
It is thus tempting to see the computer paradigm as leading to a three-tier account of intentionality:
Mentalistic properties, lying at the highest level, are explained as relations to symbolic representations, lying at a second level of description. Here, what it is to be a symbolic representation is not defined in terms of the contribution to mental states, but has some independent meaning&emdash;presumably the same meaning that the expression "symbolic representation" usually has when applied to symbols. So there is one explanation that links the upper two levels: the intentionality of mental states is explained in terms of representation. There is then a further question of how to link up the representational level with more straightforwardly naturalistic properties. This would be a separate explanation linking the lower two levels of the hierarchy. These two explanations are easily regarded as separable if there is some property of "being a representation." In the Introduction to RePresentations, for example, Fodor seems to regard the one sort of explanation as having already been carried out, while the lower sort&emdash;a "theory of how representations represent"&emdash;might not realistically be available.
Constrast with this what happens if you take "representation" as a truly theoretical term.
Two-Tiered (Theoretical) Account
In this case, the term 'representation' cannot do any truly independent explanatory work. If you say, for example, that "representational properties are those properties of the objects of inner computation, whatever they turn out to be, that accout for intentionality," then you don't have some property called "being a representation" that is offered as an explanation of the intentionality of the mental. (If you do try to do it that way, "representational properties" end up lacking explanatory power for the same reasons that dormative virtues lack explanatory power.) Instead, the word 'representation' functions as a kind of place-holder for whatever naturalistic properties might turn out to account for intentionality. There is no middle level of explanation, and indeed as yet there is no account of intentionality: one is merely pointing to the possibility of a hoped-for explanation somewhere down the road.
4. The Theoretical Construal and the Divergence of Interests
This kind of theoretical construal of 'representation' has a significant virtue and a nontrivial price. The virtue is that it completely avoids the Reid/Husserl objection. That objection arises only when we try to transport some convention- or interpretation-dependent notions into our explanation of intentionality, and the theoretical construal does not do this. The price is that we lose the ability&emdash;or perhaps it was merely the appearance&emdash;of having a notion of representation that is itself robust enough to do some real explanatory work in the absence of a full-blown naturalized psychology. In particular, we do not have a notion of representation that can do any work in explaining the meaningfulness of mental states.
Here, I think, there could be a certain parting of the ways between empirical scientists and at least some philosophers. For the desire for a semantically robust notion of representation is driven in large measure, if not in full, by a distinctively philosophical wish list: in particular, by a desire to provide an analysis of necessary and sufficient conditions for meaningfulness for the mental, and to do so in accordance with particular ontological predilections. (That is, materialist ones.) But it is not at all clear that these are issues that the empirical psychologist needs to worry about. It is certainly my experience that scientists do not, by and large, get very excited about these issues, much to the frustration of some philosophers. And so, if the distinctively scientific requirements for a notion of representation do not themselves require a semantically robust construal, scientists might well have reason to prefer the unproblemmatic theoretical construal. And this, I think, is really the right line to take.
In my view, what makes cognitive science so attractive is that it seems to hold some promise of supplying psychological explanations that bear two crucial marks of scientific maturity. First, they might describe the systemmatic relations of the domain of inquiry in a fashion that is rigorously mathematizable. Second, they may describe connections between different levels of explanation&emdash;e.g., upwards to economics and downwards to neuroscience. And this project, I think, is quite neutral with respect to ontological preference and to whether the relations between the levels is one of reduction or supervenience or mere correlation. From the standpoint of science, it is good enough to specify the mechanisms through which a psychological process is, to use an intentionally neutral term, realized. The question of whether realization involves supervenience or contingent identity or correlation or perhaps some relation that is sui generis in the case of psycho-physical relations is really quite beside the point so long as we are just doing empirical science. And that, I suspect, makes up a good healthy part of the reason that it is so hard to get psychologists interested in Cartesian demons, brains in vats and counter-earthers.
We might make an analogy with Newtonian mechanics. Newton's laws describe the how&emdash;the "in what fashion"&emdash;of gravitationally-induced motion without saying a word about why gravitational bodies attract. And while this may leave us, as it apparantly sometimes left Newton, a bit dissatisfied, it is perfectly good science nonetheless. My suggestion is that psychology can be a perfectly respectable science by specifying the how of cognition&emdash;the functional relations and the mechanisms through which they are realized&emdash;even if it has nothing to say about the why of meaningfulness.
Moreover, I strongly suspect that most if not all cognitivist explanation in psychology could be accommodated by a theoretical construal of 'representation' along the following lines, which I like to describe as a Bowdlerized version of computational theory of mind.
Bowdlerized Computational Theory of Mind (BCTM):
(B1) The mind's cognitive aspects are functionally describable in the form of something like a machine table.
(B2) This functional description is such that(a) attitudes are described by functions, and
(b) contents are associated with local machine states. Call these cognitive counters.
(B3) These cognitive counters are physically instantiable.
(B4) Intentional states are realized through relationships between the cognizer and cognitive counters. In particular, for every every attitude A and every content C of an organism O, there is a functional realtion R and a cognitive counter type T such that O takes attitude A[C] just in case O is in relation R to a tokening of T.
Here "representations" = cognitive counters = "the things that (a) are the objects of computation and (b) are the things that covary with content in the realization of intentional states."
I think that BCTM captures the heart of what cognitivist researchers are after with respect to both the computational and the representational sides of the computer paradigm. And it does so without setting off any Reid/Husserl concerns by using words like 'symbol' or 'meaning' to describe cognitive counters. So if empirical researchers were to agree that all they mean by 'representation' is something like "cognitive counter", they would be safe from attacks on the flank occupied by writers like Reid and Ayer and Searle and Sayre. This would, of course, leave plenty of room for discussion of what kinds of formal and causal properties cognitive counters would have to have to render them suitable for the realization of content. All it would prohibit would be invoking some additional property called "representing x" that was supposed to do some additional work beyond what is done by BCTM supplemented by some further specification of cognitive counters.
5. Locating the Issues
The same issues can be approached from the perspective of locating different positions on the notion of representation in relation to cognitive science. We might see a project in philosophical psychology like Fodor's as standing at midstream. Fodor seems to want us to adopt a robust, semantically-pregnant notion of representation that is closely tied to, if not indeed identical with, the familiar notion of symbolic representation. There are at least three key reasons Fodor thinks this is the right way to look at representation:
1) He thinks it is forced upon us by the success of cognitivist theories in psychology.
2) He thinks it provides an explanation of the intentionality of mental states.
3) He thinks it provides a way of "vindicating" intentional psychology against charges of methodological and ontological impropriety.
To the left of this position are two groups of people who do not draw the same moral that Fodor draws from empirical research. The first group is made up of philosophers such as Stephen Stich, who hold to a "syntactic theory of mind." Stich believes that recent cognitivist research in psychology may require the view that psychological processes are functionally describable and syntactically-driven, but not that they are manipulations of meaningful representations. Stich is also inclined to draw the stronger moral that we are entitled to dismiss all intentional states on the grounds that they are unnecessary theoretical posits. The second group to the left of Fodor's position is made up of empirical researchers such my old teacher Stephen Grossberg, who do not insist on anything like symbolic representation and find philosophical issues somewhat alien.
On the right hand side stands a different sort of philosopher, who is concerned with issues like the Reid/Husserl objection, or Searle's "derived intentionality," or homuncular regress arguments. This is really a variation on the Reid/Husserl concerns: the only semantically-rich notions of representation we are familiar with get us in a regressive tangle if we try to apply them to things in the mind. So if you want to account for intentionality by way of a rich notion of representation, you have to articulate a usage of that word that (a) avoids the regressive tangle, and (b) provides enough semantic "umph" to explain meaningfulness. I for one do not think that this has been done. Sometimes writers on the right, however, also tend to draw a stronger conclusion, to the effect that cognitivist/computational theories are somehow fundamentally wrong-headed, and this precisely because they depend upon the notions of representation and computation. But this conclusion only follows if you buy into the claim that theoretical research in cognitive science is committed to a semantically robust notion of representation.
Now I should very much like to get a more reliable canvass of how many researchers do, upon reflection, think they need something stronger than BCTM to provide theoretical underpinnings for their research. If I am right in thinking that, by and large, they will not, I think there is a way of playing both ends against the middle here and still being intentional realists. The essentials of my view are as follows:
1) Psychological questions about the form and the physical underpinnings of the mechanisms through which cognitive processes are realized can effectively be separated from philosophical questions about what kind of relation "realization" might involve (e.g., reduction, supervenience, contingent identity, correlation, or some relationship that is sui generis).
2) Two marks of maturity in a science are (a) the mathematization of its explanations and (b) the ability to form connections between higher and/or lower levels of explanation.
3) Intentional states are among the data to be explained by psychology.
4) Intentional states are also among the theoretical supposita of psychology.
5) Either 3 or 4 individually offsets any skepticism about intentional states based upon ontological predilections&emdash;hence, intentional psychology is not currently in need of vindication.
6) Cognitive/computational psychology, if successful, can provide for mathematical and connective maturity without a semantically robust notion of representation.
7) If (6) is true, concerns about the viability of more robust notions of representation in psychology, such as the Reid/Husserl objection, do not impugn cognitivist research.
8) A theory like BCTM is in principle compatible with a naturalisitic theory of intentionality tied closely to properties of cognitive counters, but does not supply such a theory, and the notion of "representation" would not really figure in such a theory.
9) A good deal of what is meant by a "theory of intentionality" is really just a specification of how it is, and how it must be, realized&emdash;i.e., a specification of the formal and causal constraints upon the realizing structures&emdash;and a theory consonant with BCTM might provide this without thereby providing an explanation of meaningfulness.
10) What is left unexplained about the meaningfulness of the mental may well be inscrutable and fundamental. That is, we may be left with a theory of the "how" (the "in what fashion") of cognition without an account of what makes it meaningful, much as Newtonian mechanics supplies an account of the "how" of gravity without saying anything about the "why" thereof.
11) This would in no way impugn the scientific status of cognitivist theories consonant with BCTM.
12) Nor would it weaken cognitivism on philosophical grounds, because it is no shame to be unable to explain what is incapable of explanation.
Of course this is just a very brief articulation of a position, not a defense of it. Let me conclude by pointing to what I think needs to be argued against alternative positions.
I think I have two chief differences with Fodor. First, he seems to believe that there is a semantically-rich notion of representation to be had that (a) can explain the intentionality of the mental yet (b) is not fraught with conventionality or interpretation. I say that if there is such a notion, no one has spelled out to me what it might be. Here, I think, the burden of proof&emdash;and indeed of exposition&emdash;is largely on the other side. Second, he believes that psychological theories require us literally to ascribe semantic properties (of this sort that I am thus far unaware of) to the functionally-individuated cognitive counters required for the formal side of the theories. I, by contrast, think that it does not matter a whit whether the cognitive counters may be said to have semantic properties, so long as they are properly matched up with the semantic properties of mental states in whose realization they are supposed to play a role. Likewise, I think that it does not matter what kind of architecture they are implemented in, so long as it has the right formal properties. Here, I think, the next move is mine&emdash;to show why the literal attribution of such properties does not matter. On the issue of vindicating intentional psychology, I think Fodor and I really agree deep down that as far as we personally are concerned it has never convincingly been called into question. There may be a need to "vindicate" it in the sense of convincing others; but one's inability to supply, say, a proof of the consistency of materialism with mentalism need not compel one to dismiss either the one or the other. (After all, the computational theory could equally well be regarded as a "vindication of materialism" for the convinced intentional realist.)
My differences with Stich are slightly different. I agree with Stich that it does not matter for psychological theory whether cognitive counters are "representations" in the semantically rich sense. Where we differ is on whether all cognitive states ought to be jettisoned in the process. I say no, in large measure because I believe that cognitive states are not only the theoretical supposita of psychology, but also make up a good part of its data, and data die harder than theoretical entities.
With writers like Searle and Sayre I have only minor differences&emdash;perhaps none at all with Husserl. Searle and Sayre seem to think, not only that Fodor's theory does not explain intentionality, but that such an explanation ought to be given. Hence Searle makes vague appeals to some "causal powers of the brain" which a computer, as such, lacks, and Sayre tries to build up intentionality out of information in the technical sense. As a woefully brief line on these views, I should say that what Searle offers is as of yet far from being a theory of how the properties of brain explains mental states, and that what Sayre offers is not sufficient conditions for intentionality, but specifications of some of the formal and causal properties of intentionality of specific kinds (e.g., in perception) in a vocabulary that is perhaps well-suited to quantifying the veridicality of perceptual states. And none of that conflicts with what I have to say.
Finally, let me make it clear that BCTM is meant to be the strongest interpretation of the "semantics" of representations needed for traditional "symbol-processing" theories. There are of course researchers in cognitive science&emdash;e.g., connectionists&emdash;who use the word 'representation' in a way that probably cannot be forced into the confines of BCTM, as they do not associate the realization of content with discrete storage locations. To accommodate such usages of the word 'representation' would probably require an even broader rule for the use of the term&emdash;perhaps something on the order of: "things in the organism that systemmatically covary with states of the environment in such a fashion as to allow the organism to respond adaptively to its surroundings."
Regardless of the significant heresies that I have expressed here, my conclusion is meant to be largely ecumenical and conciliatory: there really are different notions of representation in the air, and serious problems can arise if you fail to pay close attention to the differences. But there is a way of looking at representation as a fairly bland theoretical term. And this, I argue, is strong enough to serve for the empirical scientist while weak enough to avoid the Reid/Husserl objection. It is thoroughly consistent with intentional realism, but does not provide an account of intentionality for the mind. (Though it is in principle compatible with such an account as a supplement.) It is mainly certain philosophers who are concerned with the latter; and if they insist on having it out of a representational theory, it is here that the interests of philosophy and those of research diverge, as this project requires something out of a notion of representation that empirical researchers can do without.
Ayer, A.J. 1968. The Origins of Pragmatism. London: Macmillan.
Cummins, Robert. 1989. Meaning and Mental Representation. Cambridge, Massachussetts: Bradford Books.
Dennett, Daniel C. 1969. Content and Consciousness. London: Routledge & Kegan Paul.
Fodor, Jerrold. 1975. The Language of Thought. New York: Thomas Crowell.
Fodor, Jerry. 1981. Representations. Cambridge, Massachussetts: Bradford Books/MIT Press.
Fodor, Jerry. 1987. Psychosemantics. Cambridge, Massachussetts: Bradford Books.
Horst, Steven. 1990. Symbols, Computation and Intentionality: A Critique of the Computational Theory of Mind. Doctoral Dissertation.
Husserl, Edmund. 1913. Ideen au einer reinen Phänomenologie und phänomenologischen Philosohie. The Hague: Nijhoff. English edition, Ideas: General introduction to pure phenomenolgy. Trans. W.R. Boyce Gibson. Collier Books.
Judge, Brenda. 1985. Thinking About Things. Edingburgh: Scottish Academic Press.
Lehrer, Kieth. 1989. "Conception without Representation&emdash;Justification without Inference: Reid's Theory," Noûs XXIII, number 2 (April, 1989), pages 145&emdash;154.
Reid, Thomas. 1983. Thomas Reid's Inquiry and Essays. Edited by Ronald E. Beanblossom and Keith Lehrer. Indianapolis: Bobbs-Merrill.
Sayre, Kenneth. 1986. "Intentionality and Information Processing: An Alternative Model for Cognitive Science" Behavioral and Brain Sciences, Volume 9, no. 1 (March, 1986):121&endash;138.
Searle, John. 1980. "Minds, Brains and Programs." Behavioral and Brain Sciences 3:417-424.
Searle, John. 1984. Minds, Brains and Science. Cambridge, Massachussetts: Harvard University Press.