Saku Mantere (SM): Perhaps we should start with Your new book, The Mind Doesn't Work That Way . Could you summarize its main message?
Jerry Fodor (JF): There's a lot of stuff around that tries to synthesize a sort of a fusion of computational psychology with social Darwinism and a kind of nativism that we have become familiar with in Chomsky's work. What I was interested in was: what parts of that might be salvageable and with what generality? There is also another member of this collection of doctrines, which is the idea that a lot of psychological processes are modular, encapsulated, in some way or the other. So, the book is attempting to set out a view of cognitive architecture, which accepts some of these views and not others. It's not very enthusiastic about social Darwinism, it opts for a kind of limited modularity theory and it wants to argue that there are serious, possibly principled constraints on the generality of the computational theory of mind. It is an attempt to provide a short but reasonably systematic alternative to what I call the standard view, which has become popular as a result of the work of Steve Pinker and other psychologists recently.
SM: As I was reading In Critical Condition a few years ago, I was surprised
to notice that You 
criticized
authors who were defending the modularity of mind.
JF: I think there are two questions about modularity that are both interesting, and both open. The first is whether any mental processes can be profitably viewed as modular or encapsulated. The other is, if the answer to the first one is yes, whether it is true of all mental processes. I suspect that there are modular systems, that they are involved largely with perception and the articulation of action, and that the design of most of the cognitive mind is not modularized. There are principled differences between the kinds of cognitive problems, which can be solved by modularized systems and the ones that can't. I wanted to talk about the latter in the book.
SM: In the book You make a distinction between global and local mental processes. If I understood You correctly, the local processes can be perceived as modular.
JF: I think that's probably right. I think modularity goes with locality and also with nativism. I think that's the natural pattern of doctrines to hold it all together.
SM: And the global processes, the ones that are not modular, are they conscious by definition?
JF: Not necessarily. I think the distinction between conscious and unconscious is probably just about orthogonal to the distinction between global and modular. I don't know of any good examples of modular processes that are conscious, but for all I know there might be some.
SM: So what is the paradigmatic example of a global process?
JF: Thinking, reasoning, inference to the best explanation, problem solving, theory construction - all the stuff that we don't understand (laughs)...
SM: All the stuff that computers can't do.
JF: And all the stuff that computers can't do. I think that's not an accident. I mean, computational locality in a certain sense is built-in to the notion of computation on which computers run and I suspect it doesn't exhaust the kinds of rational processes that there can be.
SM: Why isn't it a good idea to combine Darwinism and modularity?
JF: I do think it is a good idea; in fact, I think there's a persuasive argument that whatever parts of the mind are modular are very probably adaptations. But I don't think that the mind is 'massively' modular; and I know of no serious adaptationist story about the nonmodular parts.
SM: So is the situation hopeless for cognitive science?
JF: Hopeless is putting it a bit strongly. We have been running for the last fifty years on a notion of what it is for something to compute, that we got from Turing. It looks to me like a lot of what goes on in at least modularized, localized or encapsulated cognition does fit very nicely with that notion of computation. The question is whether it's general. My guess is that it's not, my guess is that there are relatively principled reasons why some kinds of mentation cannot be modeled with the Turing notion of computation. If that's so, we need some alternative notion of computation to work with. My guess is the kind of progress we'll be making until somebody does something as fundamental as Turing did, is the kind of progress we have been making this far. Namely: progress on systems that are local and modular, largely innate. Like for example visual perception, language production and perception, that sort of stuff. There isn't anything like a theory of reasoning in cognitive science or psychology or anywhere else, and I suspect that's because the notion of computation that we have is simply not appropriate for those kinds of processes. So my guess is we won't make serious progress in that direction until somebody comes up with a notion of computation richer that the one we've got. Whether it's hopeless - who knows? It looked hopeless to suppose that there could be any notion of mental processes except association until Turing figured out how to build a much more general notion.
I suppose that at some point someone will figure out a reasonable notion of computation for a theory of global reasoning. It's just that we don't have one now and we don't have even a rough idea on what one would look like. There's a big part of the theory of mind on which we've made no progress, and don't know how to as things now stand - over and above the standard problems about consciousness and things of that kind, which we also don't have any theory of.
SM: About consciousness: I recall that in the Elm and the Expert , You limited the discussion about consciousness to a footnote in which You quoted the end of Wittgenstein's Tractatus. What did You mean by that?
JF: It seems to me that problems about consciousness are a bit similar to the problems about the kind of reasoning that's not computational. One doesn't know how to approach either problem, one doesn't know how to make what Tom Kuhn used to call normal science out of them. The discussion in both cases, notably in cases with consciousness consists of people writing articles that say: "Gee, it would really be wonderful if somebody had an idea about this. Somebody really ought to have an idea about it." The trouble is nobody does and what you can't do, you can't do.
SM: So, consciousness is a tricky topic for a cognitive scientist to work on. Damasio's work on consciousness has received attention in Finland lately, so I would like to ask your opinion on his work.
JF: I don't know much about Damasio's work. I've only read bits and pieces, so I'm certainly no expert. One reason for this is, as I said earlier, the problem of consciousness seems to me pretty hopeless. Another reason is that I deeply don't believe in neuropsychology as a method. It seems to me that the inferences made from single cases are so ambivalent and so hard to constrain, that I find it very hard to take them very seriously as a source of data for psychological theories.
I spent about twenty years in a psychology department, in which that was the principle activity: to go from descriptions of pathologies of one kind or another to psychological theory or at least to constraints on psychological theory. It seems to me a very poor source of data. The symptoms are typically highly confounded cases and they are also very often reported by people who don't ask the right psychological questions because they don't know much psychology, they know a lot about the brain's physiology instead.
I'm very dubious about that whole line of work, but I have no right to be dubious about Damasio in particular, because I know his work only slightly, and he may for all I know have the neurological key to consciousness. I find that hard to believe, though. The interesting question anyway is not where in the brain it is, but how a neurological system or anything else can produce consciousness. I don't see how exploring brain damage cases is likely to answer that question. In fact I don't know what's likely to answer it. It seems to me completely enigmatic.
SM: What empirical evidence do You view as most fruitful for a theory of mind?
JF: Of course there is a relevant class of empirical information. For example, if as I would assume, mental representations are like sentences, does that mean they have the same sort of constituent structure that sentences have? If they do, how do you do negation, how do you do quantifiers? Those are empirical questions. If somebody says that the way you do quantifiers in thought is that you line them all up on the left hand side the way you do in quantificational logic, that's an empirical claim and you ought to be able to run an experiment to see if it is true. It might be constrained by neurology, but I don't see how. It is more likely to be constrained by the sort of data you get by doing psychological experiments. If you think that quantifiers are handled in a certain way in thought, then you will be pushed into making certain predictions about what kinds of thoughts are easier to think, what kinds of arguments are easier to follow, and so on.
The trouble with the neurological stuff is that we don't know of any kind of representational system that you couldn't, in principle, build with a brain! So neurology doesn't seem to constrain psychology in any very exiting way. Maybe it will when we know more about psychology and about the brain, but at the moment it is very hard to make inferences from neurological data to answer questions like: what is the logic of mental processes, what is the nature of mental representation, and so forth.
Is the mind still a computer?
SM: We have discussed on the problems of writing a computational story on global processes. Do You still think that thinking is computation in general? Is the mind still a computer?
It depends a lot on how strictly you are insisting on the notion of computation. If what you mean is the kinds of symbolic operations that computers do, then I doubt that that picture is general. If you are taking a more relaxed view on what counts as computation, so that computation is, roughly speaking, a causal process that operates on say, symbols or mental representations or something, and does so in virtue of their formal or non-semantic properties, that's likely to turn out to be right. At least I don't know of any clearer, current reason for doubting what it is. The problem is that that loose way of putting the thing collapses two kinds of operations: those which are local and those which aren't. And the formalism that we have only works for the local ones.
There's a general picture of the mind that goes back to... I don't know - probably Aristotle, certainly Descartes. Is says that the mind, or mental processes anyway, are causal operations and they are defined on mental representations. The traditional view was that they are associative operations defined on mental representations. That sort of picture, according to which mental processes should be understood as causal chains which interrelate mental representations seems to me to be very plausible, a natural way of characterizing both global and local mental processes. It seems in fact that the only real difference between what we are doing now and what classical associationism more or less back to Aristotle has done is that we have replaced the notion of an association with the notion of a computation in the Turing sense. The problem is that it looks to me that some mental processing - though it is instantiated in causal relations among mental representations - doesn't work the way Turing architectures do. There is a general sense in which the computational theory of the mind continues to hold even if you believe what I say in my book about global processes. But, there is a more specific sense that relies on Turing's way of realizing a computational system which is not likely not to work for global mental processes if my analysis is correct.
SM: In Your book You state that there are two ways of trying to solve the problems of the computational mind: the classical, Turing-based model and on the other hand connectionist network models. You claim that the connectionist approach has the "worst of both worlds". Would You like to elaborate on this?
JF: As far as I can see, connectionism is just a way of resuscitating associationism and putting it on a machine. The problem about associationism is that it failed to explain how ideas can be connected in terms of their logical relations, as opposed to relations in frequency or contiguity or whatever the classical principles of associationism were. The Turing revolution provided a mechanism, which would allow representations to interact causally in virtue of their logical structure. That's basically the breakthrough. You could now explain, roughly speaking, how mechanisms could instantiate rational processes, where rational processes are relations that hold among ideas in virtue of their logical type. Associationism wasn't able to do this because it didn't have any notion of the logical, or syntactic, structure of a mental representation, nor is there any such notion in connectionism. So, from the point of view of the classical computational tradition, connectionism loses what was in a way most important about the Turing insight, namely that you can get causal processes to model logical and rational processes by exploiting the logical syntax of mental representations. Connectionist architectures lose this great advantage of the Turing model of computation.
On the other hand, if you
are worried about the globality problem, connectionism doesn't help. Connections
have problems of the same kind, largely because connectionism implies a
kind of radical holism, i.e. the content of an idea is determined by the
entirety of its connections to other ideas in the network. The consequence
is that the global cognitive processes which manifest themselves in processes
like theory change, confirmation and disconfirmation are extremely hard
to model in a connectionist system, because if we change any part, we change
all of it. A part of the problem about globality is that if you want to
change parts of your belief system when your beliefs get disconfirmed, you
want to change parts of them but not all of them. And you want to do this
consonant with global principles such as simplicity and so on. It's very
hard to do this in a connectionist system, because if you have a radical
holist theory of beliefs if you change any belief you change all of them.
So, on the one hand, there is a globality problem for connectionism quite parallel to the globality problem for classical theories. And, on the other, you can't do with connectionism what classical theories do very nicely, which is as I said, represent thought processes as determined by the logical form of thoughts if the only mental representations you have are unstructured thoughts and the only causal relation among thoughts is association. And that's basically what a connectionist system is.
Here is another way to put this point: Association is a paradigmatically local relation, in that respect it's just like Turing's notion of computation. According to the standard associationist story, if you have idea A, which is followed by idea B and this happens very often, you get a mental path from A to B. It doesn't matter what other ideas you have; that's a paradigm of a local operation. The global structure of your belief system doesn't impinge upon the association formation mechanisms at all.
But look, you have these globality problems! Where do you infer B from A? This depends on all sorts of other things you believe and it is very hard to say which sorts of things they are. I take Quine's epistemological holism very seriously in a certain sense. It looks like whether you are prepared to infer from A to B can be determined from very general properties of your belief system: conservatism, simplicity, all sorts of stuff.
So, Hume had a locality problem. Any associationist does. And though the Turing picture allows us to understand how the syntax, or logical structure, of thoughts can determine the causal processes of inference, it doesn't solve the globality problem any more than associationism did. Turing said that if you have the thought A & B, then syntactically that thought contains B. So, if you have a machine, which given the sentence, the mental representation A & B, can perform the act of detaching the B, then that machine will do something that can be construed as conferring from A & B to B. Now that is wonderful, because now we are saying something that Hume didn't say, namely how causal relations among ideas could be reflections, not just of their statistics or something of that sort, but also of their logical structure.
But it does doesn't answer the question of how your willingness to infer from A to B in a case when it is not a purely logical inference, but depends on other thoughts, beliefs and commitments. Turing and we are no better that the associationists in answering that. Not terribly surprisingly, when you get the whole theory fixed up to the best form that we can, we still have a representational theory which has, it seems to me - and quite aside from consciousness and the really hard stuff - problems that we don't know how to solve.
SM: You said earlier that if you don't have a very strict idea of computation, you can still say the mind is still a computer in a sense...
JF: In some sense, that is, it processes symbols. If you mean something that loose by a computer, on any version of the current view I know the mind is some sort of a computer. The big question is: which sort?
SM: So, You still believe in the language of thought?
JF: There is a distinction that I like to make between representational theories of mind and computational theories of mind. Representational theories of mind postulate something like mental representations, internal symbols, languages of thought, however you like to look at this. They used to call these things ideas, now they call them mental representations. The representational theory of mind is common ground to all of them. It's very hard to find theories of the mind in the western tradition, which aren't some version or other of the representational theory of mind. The examples that come to mind are "direct realists," people like Gibson or McDowell... Putnam has recently been making these kinds sounds too. And Skinner since with behaviorism there obviously is no mental representation at all. What these kinds of theories want to do is they want to say that somehow thinking is some kind of direct, unmediated mind-world connection...
SM: Externalism?
JF: Not just externalism, but non-representationalism. Take the case of perception. It's one thing to say that what perceptual state we are in depends on what the object in the world is that we are perceiving. So if we are in the state of perceiving a chair, that's externalistically determined, namely if there weren't a chair, you wouldn't be perceiving one. So in that sense externalism isn't in dispute. What the standard treatments have, as I said from Aristotle on, is the assumption that whatever the relation between you and the chair is in virtue of what you are able to perceive it or think about it or whatever, it is mediated by some internal representation, some sort of a symbol. Thinking of a chair or perceiving one is something like having such a symbol in mind, where the symbol denotes the chair. And then, mental processes are understood as causal relations among such symbols. That's the language of thought picture, at least it's the representational theory of mind picture.
The language of thought story doesn't really add a lot to that; it's just a special case of a representational theory of mind. It says: let's suppose that there are mental representations that play the kind of role in thought and in mental processes that people have traditionally thought they do, but let's add the following wrinkle: mental representations have traditionally been traditionally thought of in terms of pictures. That has led to all sorts of problems that Wittgenstein and others made familiar. So let's think of mental representations, instead of being something like pictures, as being something like sentences. When you think about the world, you have a sentence in your head, which expresses the content of your thought, the semantic interpretation of the mental sentence is the content of the thought.
What is the advantage of doing this? Well, there are two kinds of advantages. One is you get away from the limitations on expressive power that you inherit if you identify mental representations with pictures. It's very hard to say not-p in a picture. It is easy to say not-p in a sentence. So, you get a more reasonable account of the expressive power of thought. The other advantage connects with the points about Turing computation that we were talking about before. Turing computation is set up to take sentence-like representations onto other sentence-like representations in virtue of their internal structure. So if you substitute a sort of a sentence-like view of mental representation for a picture-view, you can, first, get closer to a notion of representation which does express what we can think about. Se if you substitute a sentence-like view of mental representation for a picture-live view you can co-opt the Turing notion of what mental processing might be like and substitute it for the associative notion. And, as I said before, the Turing notion allows you to relate sentences in virtue of their structure, something that association is unable to do.
So, the language of thought is really a sort of minor variation on the classical representational picture. But though minor, it does two things you need to do. Firstly, it allows you to get away from the picture theory of ideas and secondly, it allows you to get away from the associative theory of mental processes. And it does so just the way you want to, namely, it allows you to think of mental processes as causal interactions that mental representations enter into in virtue of their syntactic and logical structure. So now we have got a theory of rational mental processes, where before we had only a theory of statistical mental processes, ones that depended upon statistical relations among ideas.
SM: So is the language of thought hypothesis about better explanatory power than for example pictures, or is it an ontological hypothesis? Is there really, I mean, REALLY, just a language in the head and no pictures?
JF: There is a long and complicated controversy over the state of mental images. I don't have any great investment on one side or the other. I'm perfectly happy if what happens in say, certain kinds of memory or whatever, involves pictures. What I want is that the resources of the linguistic notion of representation are also available to the system, because for all sorts of processes, like for example reasoning, as opposed to imagining, access to sentence-like representation gives you vastly more expressive power than access to images. Roughly, the object of belief and the object of thought in some sense is typically propositional. Images don't have any propositional content whereas sentences typically do.
Whether or not there are other modes of representation available to the mind aside from linguistic, syntactic representation, my guess is anything you can think of it probably exploits (laughs)! I have strong opinions about what the lower bound must be. The lower bound must give you a rich enough notion of syntax to reconstruct notions of logical form and so on. But what the upper bound is, what other mechanisms of representations there are, God only knows.
As for whether it is ontological or merely explanatory, I take it that the way you find about ontology is to find out which explanations work. I'm a fervent believer in arguments to the best explanation. I don't know how else you do science. So I assume that if we can show that given the available data the mind seems to work the way it would work if it employed a sentential means of representation, then that's a pretty good reason to think that it employs a sentential means of representation.
SM: This brings us nicely toYour ontology. Did I just detect a pragmatist flair?
JF: No. Pragmatism says "whatever works is ipso facto true. I was saying only that if a theory works, then, all else equal, the best hypothesis is that that's the true theory. So I was endorsing a kind of realism, not a kind of pragmatism.
SM: You have spoken of "hairy realism" in some of Your work ...
JF: I suppose that I believe in some kind of materialism. I don't know why it matters, really (laughs). But I suppose that there aren't any states and processes except for physical states and processes - if only one knew exactly what that is supposed to mean... But I'm not expecting a reduction of mental apparatus to neurological apparatus. In that sense I'm a materialist without being a reductionist.
SM: In Your In Critical Condition You criticize Peacocke of using a transcendental argument.
JF: There is a species of the transcendental argument in theories of meaning and theories of mind, which takes it a priori that there must be certain a priori constraints between intentional states of any kind. A typical such doctrine is that you couldn't have a language that couldn't be learned from the epistemic stance of the radical translator. Now that's supposed to be an a priori principle. It's supposed to say that facts about a language, including the facts about the meanings of expressions in the language must all be accessible from a certain epistemic position, from the position of an observer who has a certain kind of data.
Another version of this, very widely held, is that languages are necessarily public. So, there couldn't be a fact about meaning, which isn't publicly accessible for a language learner. Or a Dummet-Wittgenstein kind of picture, which says that every axiom of a semantic theory has to have its counterpart in some or other capacity of the guy who speaks the language. They are all supposed to be a priori constraints on the relation between intentional facts and, loosely speaking, observable facts. These constraints are supposed to be guaranteed by a transcendental argument, of which the form is roughly: if the constraints weren't satisfied, the language couldn't be learnable, or there wouldn't be any difference between using the language correctly and using it incorrectly. I take that as a form of transcendental argument.
People run this in all sorts of ways, the goal of which is in every case to try to establish exactly what I don't think you can establish, namely a priori constraints on the ontology of the mental. A priori constraints on what kinds of constructs can be used in mentalistic theories or in theories of language.
Now, I don't know of any arguments of those kinds that work. Consider, for example, the argument that runs: language wouldn't be learnable except if it is learnable from the point of view of say, a radical translator. That argument is supposed to be something like: children are actually in that situation. So any constraints for a radical translator apply to a child who is learning the language as well. We know the language can be learnt, and since the data for learning must satisfy the condition on data for radical translation the information that his linguistic environment provides the child must be sufficient to determine the correct theory of his language. Well, that's fine if assume the crucial premise, which says that the child's epistemic position is that of somebody who is faced with a corpus of utterances and behavior and he has no prior information about the language the utterances are drawn from. But, you have to assume a highly non-nativist view on what children are like to accept that kind of an argument. In general I think the way these transcendental arguments work is that they assume an extremely empiricist theory either of what learning is like or what communication is like. That assumption cannot be abandoned without giving up the argument.
SM: This brings me to nativism. The question that has been haunting me is: which concepts are innate? What's Your final word on the subject?
JF: Let's back up a bit. There is a general theory about concepts that has been around since the empiricists, and as far as I know there are no exceptions to it. The theory is that most concepts are some sorts of constructions out of a relatively small number of primitive concepts. One way of running this kind of theory is to say that most concepts are basically definitions. The concept BACHELOR is a construction out of the concepts UNMARRIED and MAN, and so forth. For a long while that was supposed to be the model. Putting aside questions about how many definitions there actually are, it's clear that that model can't itself be general, because it doesn't tell you how the primitive concepts are learned. Characteristically, though usually implicitly, the assumption was that the primitive concepts are innate. That didn't bother anybody for centuries because it was thought that all of the primitive concepts are sensory. And then you could construct complex concepts like TABLE and CHAIR and so on out of the concepts of sensations. You end up thinking that a table is really a bundle of sensations.
However, if you find it hard to believe that tables and chairs are bundles of sensations, then presumably you are going to need a richer primitive conceptual basis than what the sensory concepts make available. The embarrassment is quite straightforward and quite persuasive, I think. If you think that CHAIR is a construction out of sensory concepts, then all you need to assume that you bring to learning that concept, are those sensory concepts. But if you assume (what I take to be plausible) that it's not a construction at all, and that you cannot define CHAIR then you are taking it to be a basic concept. And the only model that we have for basic concepts is that they are innate.
By the way, the empiricists were very clear on the sensory concepts being innate, people just don't read them that way. You know, Hume says that it is self evident, not worth arguing about. So does practically every other empiricist. They all held that what is really innate is the sensorium. The sensorium is a device which, if stimulated in a certain way produces a certain sensory concept. The structure of the sensorium is innate and the mapping it preserves from sensory events on the sensory concepts is just a part of human nature.
SM: What is your critique to Hume's conception of complex concepts being built out of basic, sensory concepts (i.e. the output of the sensorium)?
JF: This is a very long story; but here are some of the relevant bits: Firstly, Hume's account commits one to there being lots of analytic definitions since complex concepts entail their constituents and vice versa. But, in fact, there are very few. One more or less runs out after "bachelor = unmarried male." Also, I find Quine's arguments about revisability are persuasive; and (to put it roughly) if all statements are revisable, then none of them are analytic.
Secondly, I know of no case of a term that can be defined in PURELY SENSORY terms (unless it is itself purely sensory). The total failure of attempts in philosophy of science to 'reduce theory terms to observation terms' is testimony to this. As is the total failure in semantics and epistemology to articulate a defensible version of the 'empiricist principle'.
Thirdly, there is no empirical evidence that most of our concepts are, in fact, complex. Mill noticed that, although WINDOW is supposed to be a constituent of HOUSE, it is intuitively quite possible to think about a house without thinking about its windows. Modern experiments in cognitive psychology and psycholinguistics generally accord with Mill's intuition.
Finally, the same goes to concept learning. If, for example, the concept ANIMAL is a constituent of the concept DOG, then you'd expect the two concepts to be learned in that order. The failure of the purported constituency relations correctly to predict the order in which concepts are acquired is just about exceptionless.
We are in the following situation: The lack of definability of most concepts makes them look like they must be basic, we can't treat them as constructions out of other concepts. But on the other hand, it looks like the basic concepts have to be innate, since there hasn't ever been another proposal. So you put these two together and it looks like chairs and carburetors and all that stuff has to be innate. Which, as I said, is embarrassing. Can you do any better? Well, I think you can do a little better, but not a lot. And that's what Concepts was supposed to be about.
One thing that's wrong with this radically nativist picture is, aside from its seeming so crazy, is that you typically acquire concepts from their instances. If concepts are innate, if experience only makes them accessible - is simply some triggering of an innate repertoire, then it's hard to see why concepts should be learned from their instances. They could be learned from any experience or no experience. The trick is to get enough induction into concept acquisition to account for the fact I just mentioned, that the typical way of learning the concept CHAIR is from chairs. One would like to get that in without supposing that the way you do it is that somehow you construct CHAIR from the defining properties of being a chair. This is because there doesn't seem to be any defining properties! That's the problem of there not being definitions.
So there is a suggestion in the books , which I'm actually inclined to think is really not too bad. It goes like this. It is widely recognized these days that for many concepts there are prototypes, prototypes being spectacularly good instances of the concept, typically also the high frequency instances. The prototype of a dog is a sort of a middle-size dog, the prototype of a chair is really a chair, not a stool, etc. Prototypes are constructions, too. So roughly the same arguments that suggest that concepts can't be definitional constructions also suggest that they can't be identified with their prototypes.
Think of concept acquisition: here is how the mind is constructed. It has got an inductive, essentially statistical, device for building prototypes, that being statistical and so on. The innate part comes in as linking the prototype to its concept. The prototype as I said can't be typically identified with the concept for a variety of fairly deep reasons. But you can imagine a kind of a mind, which, given a learned prototype then formulates a concept by exploiting some sort of innate mechanism. The condition on the concept that it formulates is what determines the right extension. So, the concept CHAIR, unlike the prototypical chair, has to apply to all chairs. So you can now think of the mind as a mechanism that learns prototypes and, in some way or the other, pairs them with concepts.
This is partly a nativist view and partly a constructionist view and it's probably not very happily labeled as either. The major work is being done by what is assumed to be an innate component, because you have to have something about the mind that produces the appropriate concept given the appropriate prototype. The burden of nativism is still very strong, but in this picture there is at least some way of getting in an inductive part of concept acquisition, which is something that people have wanted. This kind of picture at least allows you to be not surprised by the fact that people learn the concept CHAIR from chairs, even though you think that the mechanism by which they do so is essentially innate and perhaps chair-specific.
I think that the issues about concepts are almost entirely empirical, and that the weight of the evidence is very largely on the atomist side.
SM: What do You mean by the approach You call informational atomism?
JF: Suppose concepts were constructions of some sort: statistical constructions like prototypes or logical constructions like definitions. We would then have the following consequence: you couldn't have a concept, unless you also had the concepts that also were involved in the construction. If it's a point of definition that horses are animals, then presumably you can't have the concept HORSE without having the concept ANIMAL. If some concepts are constructions out of out of other concepts, then there are some concepts that you can't have unless you have others.
If on the other hand you think what I'm inclined to, in a nativistic way, that concepts typically are not identifiable with any structured object, then that leaves open the possibility that the conditions for having a concept don't include having any other concepts. In that sense the view is atomistic. It says that, the relation between the concepts HORSE and ANIMAL is not such that you can't have one without having the other.
By the way this is connected with Hume's view that all basic, sensory concepts are independent of one another. What I am saying is really a generalization out of what he's saying. You don't have to have any other concept to have the concept RED, if you take the Humean view that all you have to have is experience of red things. Hume is a radical atomist about basic concepts. I want to be a radical atomist about basic concepts, too, except that I think that most concepts are basic, i.e. are not constructions.
So, you end up being a radical atomist about lots of concepts. Which concepts I'm not a radical atomist about? Well, the ones that show their structure on the surface: The concept BROWN COW is clearly a construction out of the concepts BROWN and COW. Very plausibly you cannot have the concept BROWN COW without having the concepts BROWN and COW. I'm inclined to doubt that you can say the same about HORSE and ANIMAL or CHAIR and FURNITURE. I believe that the only structured object that is connected with having a concept is its prototype, and that doesn't play the role the concept does; it just plays the role of a stage in concept acquisition.
It's conceivable, according to the kind of semantics I'm running, that the only condition on having the concept CHAIR is that your mind be in the right causal connection with chairs. There are no further conditions on other concepts that it has to have. That couldn't be right if concepts were definitions, because definitions are structures out of concepts. And it couldn't be right if concepts were prototypes, because prototypes are constructions out of concepts too - they are statistical rather than logical constructions. If you think most concepts aren't constructions, then you open the possibility that concepts are radically atomistic. Their content is determined by mind-world relations, not by their relations to one another.
This is a really serious issue. Suppose you take the opposite position, the one almost everybody holds, that there are some interconceptual constraints on concept possession, so you can't have the concept HORSE unless you have the concept ANIMAL, or maybe you can't have the concept RED unless you have the concept COLOR. Can you give a characterization of the concepts that you have to have, given that you have a certain concept? The standard view is that you have to have the concept ANIMAL to have the concept HORSE, but you don't have to have the concept CIRCUS to have the concept HORSE - that would be bizarre! What makes the difference? Well, the classical way of making the difference is to say that the relation between HORSE and ANIMAL is analytic, whereas the relation between HORSE and CIRCUS isn't. So if you want interconceptual relations to be constitutive of concepts, then you got to draw the analytic/synthetic distinction. I don't think that's possible.
That's why I want to get rid of all that stuff. I have no principled way of saying which connections among concepts are constitutive of the connected concepts, so I say none of them are. Of course there are infinitely many analytic sentences, for example "brown cows are brown". It's also analytic that "brown cows are cows". That goes along with the fact that BROWN COW is a construction out of those concepts. But, if HORSE isn't a construction out of ANIMAL, then I don't know the grounds for distinguishing "horses are animals", which is supposed to be analytic, from "horses run in circuses", which isn't supposed to be analytic. I take it very seriously that Quine was probably right in saying that you can't run a principled analytic/synthetic distinction. I think there is then a very short series of intermediate steps that end you up with saying: "Right, you have to be a conceptual atomist."
A way of illuminating this is looking at the history of analyticity. It originated from the notion of concept containment. What is it for "horses are animals" to be analytic? Well, it is for the concept ANIMAL to be part of the concept HORSE. The notion that concepts are constructions, have internal structure, and that there are analyticity relations among concepts run together. If you give up the analyticity, then you have to give up the internal structure. If you give up the internal structure, then it's not obvious that any connections between concepts are constitutive. The only thing that's obvious is that if one concept is a part of another, then the inference from one concept to another has to be constitutive. So, if there aren't any analyticities that a concept is involved in, then the concept can't be structured.
SM: So, looking at the philosophical influences behind your work, I detect Quine in this sense...
JF: In just this respect. It turns out that the notion of definition, which is as Quine says interdefinable with analyticity, really doesn't do any work in semantics. Very few terms are definable. So in this respect Quine is one influence.
On the other hand, Quine, at least officially, is a behaviorist. I've heard Quine say that he has learned all his psychology from his colleague, Professor Skinner, I don't know if that's literally true. I assume, anyhow, that you can't get rich explanations of behavior without assuming that it's a product of all sorts of elaborate cognitive processes. I think the reasons for believing that are just overwhelming by now. So that means behaviorism is not true. OK, so it's not. So unlike Quine I'm a robust realist about mental operations and mental causation. I guess that's not been the mainstream tradition, certainly not in Wittgenstein, certainly not in Ryle, certainly not in Quine. But the location of the mainstream changes radically with Chomsky. I think most people now think that there are no a priori constraints on what kind of apparatus you might have to postulate in behavioral explanation. Not everybody thinks this, though - I guess for example Dennett disagrees.
I think that, just as there aren't any a priori on what kinds of particles physics is allowed to postulate, so psychology just makes the normal assumption that any science does; that ontology is not an a priori issue. You just have to find out how the world works and then postulate whatever mechanisms it seems to employ in order to work that way.
SM: And then there's Chomsky?
JF: Yeah, then there's Chomsky. You take Quine's being serious about there not being an analytic/synthetic distinction and combine it with Chomsky's being serious about mentalism, i.e. with the notion that behavior has typically got mental causes that are ontologically distinct from behavior. You put those together and shake them up for a while, I think you end up pretty much with the position I'm endorsing.
SM: How about Putnam? I understand that he was Your teacher?
JF: He was my thesis supervisor. When I knew Hilary well, he used to be what he called a scientific realist. He's moved off to a position, which I find increasingly unintelligible, but I got stuck at the scientific realism stage. That's what I got from Hilary. If you are going to be a serious mentalist, you are going to say: "Look, there are mental states and they causally determine behavior." You can't have causal relations between things that don't really exist and things that really do! So, you have to say mental states are among things that really, honest to God, exist. That's what it's like to be a scientific realist about psychology.
SM: How about brains in vats then? At least after the movie Matrix, this question has been the major philosophical question in any high school student's mind these days.
JF: I'll tell you a principle I learned from Chomsky, which I think is a very deep methodological insight. Don't try to solve in your psychology philosophical problems that you would have, even if you didn't have a psychology. If you are worrying about mental states, whether mental states are real, and you are worrying about this for the usual reasons, because of brains in vats and so on, then you ought also to worry about whether mountains are real. But within geology, while doing the science of mountains and streams and stuff like that, one doesn't say: "Well, the first thing I have to do is convince myself that there are mountains and streams by some philosophical argument, and then I'll go ahead and do the science."
Whatever philosophical interest there is in questions about skepticism and whether or not these problems are solvable, or dissolvable or whatever, they are not problems one expects the science of the subject matter to solve. They are external to that. One just starts off by taking the ontology for observable, middle-sized objects for granted and proceeds from there. You can do all the geology you like and still have the question open whether there really are mountains, or whether the world was created ten seconds ago... that was one that used to worry Darwinians. You just take the ontology that there are mountains and streams and so on at face value and go on to build a theory of those. It seems to me that's what one has to do in psychology, too. One asks in psychology questions like: how is perception possible? How is it possible that if I'm situated in the right position and the lights are on, blah, blah, blah, I can see a chair. It's not a possible answer that there are philosophical reasons for there not being any chairs, so you can't!
In fact, I think something is really going wrong when you start out with one of these "how is it possible" -questions and end up on philosophical grounds that well, as a matter of fact it's not. Like the question: how can there be mind-independent objects? I think science should start out with a common sense ontology and work from there. In any case I don't expect a psychological theory to answer the question: how do you know you are not a brain in a vat? But it might very well answer the question: given that you are not a brain in a vat, how is it that you can perceptually identify chairs? I don't know if there is an answer to the brain-in-a-vat question, what I do know is that the question is not in the domain of the theory of mind, any more than the question whether mountains are just ideas of mountains is in the domain of geology.
SM: Do you think there is any sense in writing philosophy about such questions?
JF: Very often skepticism is pointing to something real. Skepticism about other minds, having now grown through behaviorism, was pointing to a perfectly solid truth namely, that there are no conceptual connections between facts about behavior and facts about minds. Whether there's more to be found by exploring skeptical worries, I don't know.
The way it usually works, though, and I think probably properly so, is that you start out by assuming that the possibility that the skeptic is raising isn't real. And now the question is: What does the world have to be like so that the skeptic's possibility isn't real in that world? For example, given the fact that you obviously do know about other people's mental states in all sorts of ways, ask how is that possible?
SM: So skepticism can produce healthy influences for a larger discussion?
JF: A lot of good philosophy has come from thinking seriously about skeptical arguments. But, the way one thinks seriously is starting out by assuming that there is something wrong and trying to find out what it is. The force of the brains-in-vats case isn't to get you worried about whether you are really a brain in a vat, because you know perfectly well that you are not, but to get you thinking how it is possible to know you are not. So one is assuming that there is something wrong about the skeptical challenge and the interest of exploring the challenge is finding out what it is.
SM: One example of a philosopher who tried to find a positive solution to a famous skeptical argument was Saul Kripke in his reading of Wittgenstein, his skeptical solution to Wittgenstein's problem of following a rule.
JF: Wittgenstein thought, and I think correctly, that you can't show that you following a certain rule of behavior just by saying that you take the rule to have been satisfied by what you did it to be such. But a point that a lot of people have made is, I again think correctly, is that you also can't show that you are following a rule that just saying that everybody takes the rule to be satisfied by what you did. But, I guess this is what Kripke claims to be the skeptical solution.
Kripke thinks that the dispositional notion of following a rule can't be sustained, but I don't think he has made a case for that. Kripke in effect, that the behavior we produce could be the first part of the actualization of lots of different dispositions. I'm not deeply moved by this, unless it is seen as general skepticism about induction. I guess the way to put this is the following. If there are real psychological states, then being in them de facto makes some hypotheticals true and other hypotheticals not true. That seems no different from states like solidity. If something is solid, it is disposed to react in certain ways if you push it. I don't see why that picture doesn't apply to psychological dispositions just as it does to physical ones.
Of course, you can't get more than a finite number of instances of hypotheticals that are satisfied; people die after a while. So why is it that the rule covers infinitely many cases and not just of the finite number that actually occur? Well, I think that's a perfectly reasonable question, but I don't understand why it doesn't apply to all contingent hypotheticals. And if it applies to all contingent hypotheticals, then this isn't a problem about the mind, it is a problem about induction. It's a problem about saying what inductive generalizations have priority over what others, it's a problem of saying how counterfactual support works. Kripke's problem looks to me an awful lot like that. If you see a guy saying 2+2=4 +2=6, +2=8, +2=10, and then he stops doing that because he is hit by a lightning bolt, then how do you know that his notion of plus two isn't such that he would have given fifteen as the next number had you asked him? That sounds to me awfully like "how do you know the ruby after t is going to be red and not green?" And the answer is: I don't know how you know. That's a really deep problem about induction. If we understood counterfactuals, we would understand a lot more about inductive inference.
So I would really like to have solved that problem - I would be rich and famous if I did, but that doesn't have anything to do with the mind, any more than it does with rubies. If you are going to worry about that problem, why do it in psychology. Somebody ought to say what it is that warrants your saying "if I have got one red ruby and another and so on till t, then the right expectation at t is that the next ruby will be red." Somebody ought to say what the logic of that kind of an inference is, but it's not a problem in pscyhology, or even in the philosophy of mind.
The move that gets made at this point is one for which I have no sympathy at all. People the rule case is different from the induction case because it's normative. And that it's normative means that it's not just he does say six at t, but that's what he ought to say if he is following the rule. I honestly don't understand what the claim is. He will certainly say six if he is in the dispositional state that we take everybody else who is adding two to be in. That there's more to the normative force of following a rule - that I don't understand. I'm afraid normativity leaves me kind of cold.
SM: You talk about the mind being computational, at least in a loose sense. Does this also mean that there is one, single rationality for all human beings? Isn't this just one more Grand Story or something?
JF: I'm not quite sure what the issue is. I take it that p & q entails p for just about everybody. I don't think that's relativized, you know. What would it be like to find somebody, for whom it's correct, not just that they are inclined to say p & q but not-p, but they are right in saying that? I mean, that it is true when they say it? I give up, I don't know what that would be like.
There are, I suppose, all sorts of non-logical - in some strict sense of logic - constraints on action and inference that we pick up as we go along. Logic doesn't by any means exhaust our intuitions about rationality. In fact it is a rather narrow subset of those intuitions. The intuitions we have are presumably codified in theories and traditions and practices that we think are approximately correct, so that these theories and practices are approximately warranted. Of course we could be wrong about that. All sorts of inferences that are in a certain sense straightforwardly empirical, like "if it's water then it's H2O", are no better than the theories behind them. So you could say that these inferences are relativized to theories. But so what? The inferences are sound if the theories are true. This is all sort of banal. What I'm not seeing is a deep sense in which rationality would be relativized to something interesting.
SM: What are You actually: a philosopher, or a psychologist, a linguist, or what?
JF: I thought I was a philosopher for a long time. Then I discovered that philosophy is supposed to be an a priori discipline like logic, and I think logic. So I guess I'm not a philosopher.
I'm interested in how the mind works. In particular I'm interested in representational theories of mind. What you can do with them and can't do with them, which phenomena seem to be treatable in those terms, what ontological commitments they make, what ontological commitments they don't make - those kinds of questions. One reason such questions are interesting is that they are connected with a lot of other questions that are interesting. If you want to have a theory of mind, you have to have a theory about lots of other things, for example word meaning and which things are learnt and how.
As I said, I don't think that philosophy is an a priori science. I think doing this kind of philosophy is a great deal like theory construction. You try something to see if it works, if it doesn't then you take it back and try something else. The result is, if you are lucky, you can verge on, if not the whole truth, then at least more things you can explain about what is connected to what.
All in all, I'd really like to do Hume right (laughs). Getting rid of associations, getting rid of the picture theory, getting rid of all the aspects of traditional representational theories that we now do without because we have got a rich logic for sentences and we have got a notion of computation to replace the notion of association, and so on. So starting with Hume and not worrying about innateness as he does, one might end up with a modernized representational theory of mind, and the question I am really interested in is: What could you do with such a theory, what could you explain and what couldn't you?
SM: You have mentioned Hume quite a few times during the interview. Could you sketch out the historical and philosophical path that You feel most at home with?
JF: There is a whole tradition of thinking about the mind as a representation-manipulation device. That goes back to Plato. He speaks of memories as like birds in cages that you reach in and grab. That tradition is supplemented by associationism What determines the causal relations among ideas is whatever it is that determines association: frequency, similarity, etc.
Then that tradition runs all the way through classical rationalist theories of mind. It runs through the empiricist theories as well. I consider my work to be just an extension of the same view.
My kind of philosophy of mind probably started in a cave somewhere! The representational tradition is one of the oldest in philosophy. It's very hard to think of anybody who has ever thought seriously about the mind - except for the behaviorists of course - who didn't end up with some theory of generally this same general shape. I find that very gratifying, because if I am making a mistake, then practically everybody else is, too!
