A DEDUCTIVE ARGUMENT FOR 

THE REPRESENTATIONAL THEORY OF THINKING

William G. Lycan

Wilfrid Sellars, Jerry Fodor and others[1] have argued that to think is to harbor and deploy a physically realized mental representation.To put the thesis not much less crudely: for S to think or "occurrently believe" that P is for there to be a state of S's central nervous system that (a) plays a characteristic information-storing role in S's behavioral economy and (b) bears the semantic or propositional content that P.The state bears that content in much the same sense and in much the same way that a sentence of English or another natural language means that P??at least compositionally, in the sense of having proper parts and its meaning being a function of the meanings of those parts[2] (though the state may be diffuse and highly distributed rather than modular or physically salient as public linguistic tokens are, and it has its content naturally rather than (in part) conventionally as public linguistic tokens do).The representational theory, as I am calling it, is thoroughly empirical rather than purely philosophical, since the existence of any central nervous system whatever is a thoroughly empirical fact.

The arguments for the representational theory take a number of impressively different forms.The theory has also come in for a good deal of criticism in recent years,[3] and there too the arguments take a number of impressively different forms.I have vigorously if not conclusively defended representationalism before; in this paper I shall add one more argument to the "pro" side of the profuse debate.

1. The Deductive Argument

I begin with an interesting rhetorical fact that has been underemphasized:We representationalists tend to think not only that our view is true, but that it is fairly obvious.Sometimes in philosophical debate one needs to know not only what a party is claiming but what epistemic status that party attaches to the claim.[4]Opponents of the representational theory should be aware that this is one of those philosophical occasions.Representationalism is not supposed to be a hunch, a bold speculation or a wild throw of the dice, but rather a truth that is pretty plain once a few not very controversial assumptions have been accepted.Of course some philosophers have rejected representationalism nevertheless, perhaps disagreeing with the assumptions I have in mind.But some philosophers who have not visibly repudiated the assumptions have also rejected representationalism, sometimes suggesting that the representational theory is gratuitously daring and/or outlandish.(Thus, "Of course I'm a physicalist and a realist about beliefs and desires, but I don't believe in little sentences written in brain matter or anything crazy like that."My main opponents are philosophers who are physicalists and intentional realists but who do not believe in little sentences written in brain matter.)

Though I shall be arguing that the representational theory is fairly obvious, I emphasize that I am not about to indulge in brute a priorism.Representationalism is an unmistakeably empirical and fallible claim.But as Fodor said in The Language of Thought, representationalism is also "the only game in town."No credible alternative picture of thinking has taken the field??no matter how strongly one may suspect that thinking is something other than physical mental representation.[5]My argument is intended to bring out this empirical obviousness as well as the truth of representationalism.[6]

Let us start the argument by considering what is at the same time one of the most humdrum facts of human life and one of the most amazing facts of the physical universe: natural-language speakers' staggering ability to understand indefinitely long, totally novel sentences on the spot.Since the 1950's, linguists have rightly marvelled at that fact, modestly called by Chomsky "the creative aspect of language use," and at the uncountable number of possible sentences that a normal speaker could and would understand on the spot but for performance limitations.Linguists have been driven by this infinitary or at least unbounded ability to suppose that the sentences that are understood must be composed of atomic, semantically primitive elements drawn from a finite and finitely learnable stock of morphemes, and that the meanings of sentences are compositionally determined by productive, recursive rules for arranging morphemes into certain combinations and orders.(That, surely, is what syntax or grammars of natural languages are for.)Speakers must somehow have mastered those recursive rules even if they have never explicitly represented the rules to themselves.

Notice that in saying all this the linguists have not just rallied to an a priori guidon.The morphemes-plus-compositional-rules theory is empirical and fallible; it might be false.But it is the only game in town.One cannot now imagine any other explanation for the striking facts of speakers' near-universal understanding (call those facts collectively "the unboundedness of understanding").[7]For well-known reasons,[8] Behaviorism will never do.Divine inspiration is always a possibility, as Socrates says at the end of Meno, but not a very likely one. 

The Chomskyan argument from unbounded competence out of finite resourcesto recursive structure is familiar and overwhelmingly persuasive.How could the thing be otherwise, barring either magic or divine intervention?[9]

But now forget public natural languages and consider thinking.As Dennett has noted,[10] we have "essentially limitless powers" to think about different things and states of affairs.(If the universe contains things we simply cannot and could not in principle think about, those things are remote to the nth degree.)We may distinguish at least four claims here, in decreasing order of strength: 

(i) that any possible fact could in principle be an object of human thought; 

(ii) that any actual fact or nomologically possible fact could in principle be an object of human thought; 

(iii) that within reason any actual or actually possible fact could in principle be thought??that is, that although human capacities are ultimately limited, the limitations have no measurable day-to-day effect and for practical purposes, die Gedanken sind frei

(iv) that in real life, we can think nondenumerably many things, and one is hard put to imagine any particular source of things we intractably could not think or imagine.

In one or more such sense, thinking is a universal capacity.Thought can be about anything, where `can' and `anything' have one or another nondenumerable range of values; possible propositional contents of thought are practically unlimited.For short, I shall refer to this unobjectionably vague feature as "the unboundedness of thinking."[11]

Take this unboundedness of thinking as a premise, (I).Add physicalism regarding the mind: the thesis that every psychological fact about any human being is token-identical with some physical fact or other.Alternatively, so as not to deter those who argue that physicalism is false of sensations, add just (II) physicalism about thinking in particular.I ask anyone who doubts physicalism about thinking to grant it just now for the sake of discussion. 

Now, if the Chomskyan morphemes-plus-compositional-rules theory is the only game in towm as regards the creativity/novelty/unboundedness of public-linguistic understanding, then presumably a parallel theory is the only game in town as regards the unboundedness of thinking.If we cannot come up with a credible alternative in the former case, then we cannot come up with one in the latter, for structurally the two are exactly alike.If physicalism (II) is true, then to explain the unboundedness of thinking, we need a finite stock of semantically primitive physical elements and a compositional system of recursive rules to project the atomic semantic values of the elements into the composite semantic values of whole complex thoughts.

One might suppose that if physicalism were not true, the unboundedness of thinking would be easier to explain.Indeed, for the nonphysicalist the two explanatory problems are not structurally alike, since it is open to the nonphysicalist to insist that die Gedanken are indeed frei in a way that physical public speech is not: the mind has the power to think any uncountable number of things, each one simply, primitively, and all of a piece,without any sort of compositional analysis or recursive strategies.Quite so.But the nonphysicalist does not thus offer an explanation, but mystery and an appeal to magic.How does the nonphysical mind accomplish its feat of unbounded thinking?It just does; it is wonderful.My point is that if one does seek an explanation of the unboundedness of thinking, that extraordinary and very striking human capacity, one will not find it in nonphysicalism, and so one is advised to accept (II) after all.[12]

Thus we should proceed to (III): If the unboundedness of thinking holds, then so far as one can see, we need a finite stock of primitive elements and a compositional system of recursive rules to project the atomic semantic values of the elements into the composite semantic values of whole complex thoughts.And again given (II), the elements must be physical and the rules must be physically realized (of course not by being themselves represented).Thus (IV): Human beings have systems of physical states, and/or host physical events, that serve as semantically primitive elements of a lexicon, and human beings (somehow) physically realize rules that combine strings of those elements into configurations having arbitrarily complex propositional contents in virtue of the semantic values of the primitives' having been projected by the principles of composition.

(IV) is the doctrine of representationalism??QED.And granted (I)-(III), (IV) should be obvious.

2. Obviousness and the Empirical

How could a thesis about people's central nervous systems be both entirely empirical and obvious to a philosopher who has done no relevant scientific or other experimental research whatever?

Happily for philosophers everywhere, there is a range of entirely empirical fact that is accessible from one's armchair.Viz.:One can often attend to some indisputable everyday pre- and non-philosophical facts of life and draw some indisputably reasonable inferences from them, without injecting any tendentious philosophical ideas.I shall illustrate this possibility by elaborating a philosophically harmless example.

Imagine that we are aliens, perhaps visiting Jovians, inspecting Los Angeles and scrutinizing in particular its dominant life form, the automobile.We see only the automobile's outside, as it variously crawls and races along the asphalt, but we note that its behavior is systematic: people who visibly intend to go to certain places get into such a contrivance and then (usually) end up in the places they wanted to go.An automobile, we discern, is a device that takes travellers' travel-intentions as inputand outputs the destinations intended.How might such a thing do such a thing?

Without doing more than looking at a functioning automobile from a distance, we Jovians could reasonably infer the following:There being no sail, poles or jetstream, the auto moves by dint of its high-friction contact with the road, specifically because its wheels (for such they are) turn without slipping.Thus some inner mechanism applies torque to the wheels to make them turn.A power source is required to drive that mechanism, and there must also be some system of cogs, gears, belts, rods, or their functional equivalents to convert whatever sort of energy is produced by the power source into kinetic energy and transmit kinetic energy to the wheels.

Meanwhile, the driver must have constant and minute control over the transmission process, since both the speed and the direction of the vehicle are a function of the driver's intentions.There must be a device (which might naturally be called an "accelerator"), directly accessible to the driver inside the car, that affects either the amount of energy generated by the power source or the flow of energy from the power source to the wheels??presumably the former, since otherwise there would be superfluous energy to be got rid of and wasted.Moreover the accelerator's effect various continuously, like that of a rheostat, rather than like that of a switch.Further, there must be a similarly accessible steering mechanism that controls direction, by turning the front wheels continuously left or right.Still further, there must be a device to overcome the heavy vehicle's inertia when a sudden slowing or stop is required??call it a "brake."The brake may be a special feature of the acceleration system (as when jet pilots slow down by reversing their jets), or it may be a mechanically separate component that acts more directly on the wheels; but it is not a sail, a parachute or an anchor.

Thus we already know a good deal about auto mechanics without even looking through the car window, much less raising the hood.If we allow ourselves a closer look and listen, we can infer still more.The power source very probably is a combustion engine of some sort, since neither an electrical nor a nuclear engine would make noise??the noise is in particular one of muffled explosions??and the auto emits gases typical of combustion byproducts.Periodically the auto is seen to stop for infusion of a specially marketed fluid, presumably fuel of some sort.

Assuming the power source is indeed a combustion engine, at least five more components must be posited.There must be a fuel reservoir, a fuel line, a carbureting device to mix oxygen into the fuel, and a combustion chamber with at least one compressible wall.Moreover there must be a standing auxiliary power source at all times available to ignite the fuel when the auto has not been used for some time; a battery is the obvious though by no means only choice.

I have said that "we already know" a good deal of auto mechanics without looking inside even a single auto.Skeptically inclined philosophers who are stingy in the matter of knowledge ascriptions would be reluctant to give in on this one.We speculating Jovians do not really know the things we have speculated.There are alternative possibilities that we have not specifically ruled out.Perhaps cars are shot out of their garages by powerful slingshots, maintained in motion by internal gravity wheels, and steered by pre-programmed home computers.Or perhaps they are pulled along their actual routes by powerful magnets hidden underground.Or perhaps they obey divine commands, and their innards are irrelevant.Or perhaps they work by telepathy, or simply by magic. 

The existence of such alternatives is real, but not very impressive.The slingshot and magnet hypotheses could be ruled out by only slightly closer external observation of drivers and autos.The remaining competitors cannot be ruled out in that way, but are implausible given our general knowledge of the world and the methodological assumptions of science.None of the alternatives is a live option.The auto must instead work in something like the way we have hypothesized.But this "must" is still empirical in character, and expresses nothing even so strong as nomic necessity.

So too with representationalism.The representationalist asks, "What else could it be?," meaning not that there is no conceivable alternative but only that at least for an intentional realist, the conceivable alternatives are not live options.Representationalism may still turn out to be wrong, but it is hard feasibly even to imagine a real-world alternative.

3. Objections and Replies

I claim that the argument of section 1 constitutes a powerful defense of representationalism as I have characterized that doctrine.Now, there are further claims that have sometimes obligingly been tacked onto representationalism by its critics, such as that the brain-realized representations are physically salient, or that they actually have syntactic forms characteristic of natural languages, or that they have syntactic forms characteristic of the predicate calculus, or whatever.I doubt that any of those generously supplied further claims is true.But no representationalist qua representationalist has ever been committed to any of them.

There is a gap, however, between my conclusion (IV) and a fully-fledged representational theory of thinking:Though I claim that physical states of people's nervous systems have semantic values, which (off the record) I understand in terms of reference and truth-conditions, I have said nothing about what it is that endows a bit of brain matter with its semantic value.If the semantic value of a mental morpheme is a referent or extension in particular, in virtue of what does the referential relation hold between the morpheme and the external referent?This is Wittgenstein's question, "What makes my image of him into an image of him?"[13]An answer to it would be what Fodor calls a "psychosemantics."A fully-fledged representational theory of thinking would incorporate a psychosemantics.

I emphasize that I offer no psychosemantics here.[14]If my conclusion (IV) is correct, there must exist a correct psychosemantics; but its details must be left for another occasion.

Let me now address two quick objections to the deductive argument itself; then I shall go on to anticipate three more substantive criticisms.The first objection[15] is that premise (III) simply begs the entire question; at any rate representationalism appears to me as the "only game in town" only because (III) gives all other games away.If one accepts (III), then of course representationalism seems painfully obvious.But why should we accept (III)?

I reply that, as before, we should accept (III) because its consequent is the only known or imagined viable explanation of its antecedent.Indeed, in one very crude sense of "recursive," to say that a device produces infinitary output from a finite stock of elements just is practically to call the device recursive.Again, we must grant??as I do??that (III) might be false, but since we cannot see how (III) could be false, we are entitled to assume (III) to be true, until someone should come up with an alternative, nonrecursive account of the unboundedness of thinking.I shall consider one attempt at one, below.

The second quick objection is that even if my deductive argument is in fact sound, I have oversold it.I have billed it as a deductive proof, but since premise (III) rests on admittedly explanatory considerations as we have just seen, we should consider the argument instead as a piece of abductive reasoning, as an inference to the best available explanation.Thus more modestly understood, the argument is more likely to convince, and in particular the now only tentative assumption of (III) will seem less arrogant and be less rash.Representationalism is the best going explanation of the unboundedness of thinking in a physical system, even if it is not the only possible explanation.

If forced to rhetorical retreat of this sort, I would of course accept the weaker explanationist position.And I grant that my deductive argument is at bottom a special case of inference to the best explanation.[16]But what I am emphasizing in this paper is that it is a very special case.The point of my automobile example was that from time to time the world presents us with a striking sort of datum, viz., a datum that, practically speaking, admits of only one explanation.In such a case, it is tempting to describe one's reasoning as "inference to the only explanation," an inference perhaps intermediate in strength between ordinary explanatory inference and deductive demonstration.[17]Admittedly, the difference between this inference and ordinary abduction is one of degree, not of kind, since it is fallible and relies on one's overall view of the world; "only" in "only explanation" means "only feasible," not "only conceivable."But I think the rough distinction is worth preserving, and that is why I resist the proposed deflation of Fodor's claim and the forensic status of my argument.

I now turn to tougher opposition: an eliminativist response, an instrumentalist and/or "environmentalist" response, and a Connectionist response.

1. An eliminativist denies the argument's first premise (I), on the grounds that no one ever thinks anything at all.[18]Eliminativism is a position I respect, but reject (on Moorean grounds combined with the inadequacy of arguments for the position[19]).This paper offers no answer to a staunch eliminativist, but if one really is an eliminativist one already has trouble enough.

2. There is a type of philosopher, descended from Heidegger, Dewey, Wittgenstein or Ryle, who abhors inner-process, causal-mechanical views of thinking and who sees thinking as rather a matter of outward aspect and/or environmental surround.This "instrumentalist/ environmentalist" (i/e) view has many importantly distinct versions, but here I can address it only generically.[20]The i/e-ist claims that although brain processes are causally implicated in thinking, they do not constitute the truth-makers of thought-ascriptions; they are not part of what thinking itself is.What a person thinks is determined by that person's behavior in response to stimuli and/or that person's social context and/or the explanatory interests of that person's friends and/or "interpretation" on the friends' part.If this is right, then (I)-(III) do not after all entail (IV):There is an alternative physicalist explanation of the unboundedness of thinking, one that does not invoke semantically charged brain processes.Physically realized mental-morphemes-plus-compositional-rules might exist without being localized in individual humans' central nervous systems.

True, they might, and the objection is correct: (IV) is not a logical consequence of (I)-(III).But this is yet another case of the only game in town.For to pursue the i/e alternative we have to ask what outer, behavioral-cum-social-cum-predictive physical states of affairs could physically realize mental-morphemes-plus-compositional-rules.Behaviorist resources are notoriously not up to the task.Conventions and social practices are rich and various, but they all presupposes thinking on the part of a social group's members.This does not necessarily lead to circularity, since someone might devise a recursive social theory of thinking that specified a base set of thoughts without reference to other thoughts or thinkers and then defined a larger class of thoughts by reference to the base thoughts.But what would actually be the base thoughts in such an account?It is hard even to speculate.Moreover, wherein would the base thoughts have their semantic values, without reference to further thoughts?I am not suggesting such a theory is impossible.I am saying only that we cannot now begin to see how it would go, much less provide even a sketch of it.

If my argument from unboundedness is to be valid, we need a bridge premise to get (IV) from (I)-(III).I propose: 

(B)If (I), (II) and (III), then (IV).

(B) neatly restores validity to the argument.(B) is not logically true, but I have a hard time doubting that (B) is true.[21]

3. It may be complained that as a matter of plain fact the representational theory is not the "only game in town," but has real-world competitors.The leading alleged competitor would be contemporary Connectionism.[22](In capitalizing the term I am following recent neologism and referring to a philosophical contention again reminiscent of Wittgenstein and Ryle, roughly that intelligent human capacities, thinking and rationality are somehow holistically emergent from connectionist architecture in the brain rather than being a matter of the manipulation of internal representations according to rules.[23]I shall use lower-case `c' when meaning just the type of engineering or programming architecture.)

But, logically speaking, either connectionism or Connectionism is entirely compatible with representationalism, and arguably it is an instance of representationalism.[24]In any case, any Connectionist who wanted to deny the representational theory would have to find something wrong with my deductive argument, and nothing about connectionist architecture itself immediately suggests any criticism of the argument.That connectionist architecture differs topologically from von Neumann architecture seems to affect the argument not at all, though the difference is of much importance to the practice of AI itself.In particular, that representation is distributed rather than modular is no embarrassment to the representational theory; in fact, it entails it.

But the matter cannot be dismissed quite so handily.As Stanley Munsat and Stephen Stich have pointed out to me, a Connectionist might deny (III), the Chomskyan premise that if thinking is unbounded in my sense then there must be a recursion on a finite stock of semantic primitives; connectionist architecture might be put forward as a genuinely alternative, competing way of getting unbounded competence from finite resources.(This would be a very radical and important suggestion, given the enormous power and persuasiveness of Chomsky's original argument.)

But it is still hard to see how connectionist architecture alone could do that.Consider again the case of a subject's hearing a long, novel sentence and immediately producing an equally complex and contextually appropriate utterance in response.Our Connectionist would have to claim that the response was mediated by the activation pattern on the system's hidden units but also that no mental morphemes or semantic primitives could be abstracted from the activation pattern itself.This impossibility of abstraction seems to me very unlikely.Take the example of NETtalk, a connectionist program that (after training) audibly pronounces English words given written text.[25]NETtalk's accomplishment is very impressive, for each of three paradigmatically connectionist and Connectionist reasons.First, the task is very tricky, since the actual function from English spelling to oral pronunciation is highly irregular.Second, the function is also many-many rather than one-one or many-one, since the pronunciation of a given letter or letter-group depends heavily on intralexical context.Third and most notably for philosophy of mind, the machine is initially given neither rules of phonology nor representations that look like rules of any other kind, but only a general and topic-neutral learning algorithm that adjusts various activation levels between word-presentation trials.At first the feat seems like magic.And philosophically, to some, it seems like a Rylean/Wittgensteinian dream come true??highly intelligent behavior unmediated by representations and rules.

However:As is well known, mature NETtalk activation patterns at the hidden layer are found to be partitioned into disjoint classes.There are 79 of those classes, and as many people have pointed out, this is no accident; they correspond to the 79 distinct letter-to-phoneme moves that are mastered by a normal competent speaker of English.Now it also turns out, under a "cluster analysis" carried out by Rosenberg and Sejnowski (op. cit.) that involved grouping by similarity along certain parameters and averaging of values, that the 79 classes are also grouped into two main superclasses and various hierarchically organized subdivisions of those.The two main superclasses correspond to the vowels and consonants of English, and the subdivisions (again, made on the basis of a general similarity metric) correspond to familiar vowel and consonant subtypes.Thus, out fall all the phonemes of English phonology.Though NETtalk is handed no phonological rules by its creators, it acquires phonological categories in the process of learning its job, and, we may say counterfactually, it would not have achieved its extraordinary accomplishment had it not acquired those categories.It is still extraordinary, as well, that the categories cannot in any simple way be read off the raw graphological input; nonetheless NETtalk acquires them, and accordingly, it seems to me natural to ascribe phonological concepts to NETtalk (though one may quibble over what rich computational properties a feature of a system might be required to have in order to count as a bonafide concept rather than merely a discriminative ability).[26]

In similar wise, one would expect, the familiar morphemes would fall out of a connectionist syntax and semantics for a general natural-language-understanding machine.(This expectation is reinforced by the known psychological robustness of morphemes in humans,[27] though I do not like to sully the a priori nature of my argument by officially appealing to psychological experiment.)Indeed, one can argue from the presumed supervenience of actual and counterfactual linguistic differences on the connectionist architecture, to the claim that there would have to be systematic counterfactual differences in the activation pattern, however holistically they would need to be characterized, and those systematic differences would presumably constitute differences of semantic elements.[28]Notice also that in the case of human language (as Jay Rosenberg has pointed out in conversation), the back-propagation of error used to "train up" connectionist networks is conducted in semantic terms, e.g., by misunderstanding or by correction. 

But all this is beside the point, for our concern here is not with the explanation of public linguistic behavior, but with the unboundedness of thinking.Notice that even if the Connectionist objection did work against Chomsky for the case of natural language understanding, it does not obviously apply to thinking at all.Thinking itself is not a matter of turning an input into an output, even though it is functionally specified ultimately in terms of inputs and outputs and other psychological states.How would a connectionist machine itself have the capacity to think an indefinite number of indefinitely complex and novel thoughts?Either it could not, or (surely) it would do so by having a finite repertoire of physical states it could be in that would implement those thoughts.Any Connectionist is free to seize the former disjunct and add that a great virtue of connectionist machines is that they do not need to have individually specifiable "thoughts" in order to transform their inputs into intelligent outputs; this claim would be very much to the liking of some Connectionist converts I know.But notice (carefully!) that to say this is to suggest the quining or elimination of thinking, not an alternative to the Chomskyan explanation of the unboundedness of thinking.So I do not see that (III) is in danger.[29]

Still, the NETtalk example strongly suggests a possible revision in the representationalist's conception of mental morphemes.

4. An Alternative Picture

It is not unnatural, though it is inaccurate, to characterize representationalists as advocates of "brain writing."We representationalists do often write or speak as if it is bits of brain matter that have semantical properties, just as bits of ink do when suitably arranged on pieces of paper.More properly, since propositional attitudes are not physical individuals but states of or events occurring within physical systems, what we have meant is that brain states and events have the semantical properties.But either way, admittedly, we have conceived the relevant attitude-tokens as admitting of forthrightly neurophysiological description, whether or not those tokens are morphologically salient in the brain.

Now, consider NETtalk again.Even after NETtalk has been trained up (cf. fn 26), no state of its units described in hardware terms corresponds to its concept or "concept" of /ey/, nor any such state to its sophisticated "belief that" the input word "weigh" is pronounced /wey/.NETtalk's concepts were discovered only by a very abstract statistical analysis, and, more to the point, they can noncircularly be characterized only in the same abstract statistical terms.They are none the less real for this, but they are the less meaty (so far as the "units" of a "connectionist network" being simulated on a digital computer can themselves be called meaty).NETtalk's having a concept or a "belief" is (a large mass of) NETtalk's units' collectively having a certain statistical, hence essentially mathematical, property.

So too, perhaps, with human concepts and beliefs.Dennett notes that if our brains' architecture is connectionist, the "brain-thingamabob [that] refers to Chicago" would per se have to be described statistically and in terms of the whole connectionist system or a very large sector of it.[30]And although our brains' architecture is very unlikely to be literally connectionist (there being no internal "supervisor" or "teacher" to provide instant correction of error), our brains are more connectionist than they are von Neumann.It is a strong possibility that the NETtalk model extrapolates to brains??that human belief-states are not hardware-described states, but only statistical patterns manifested by largeish regions of the brain.For a person to have the belief that P is for a functionally cognitive/representational[31] region of that person's brain to exhibit such-and-such a statistical feature, and??in light of Twin-Earth examples??for that feature to be causally grounded in whatever objects and properties figure in the proposition that P.As with NETtalk, this grounded-statistical-abstraction picture makes intentionality and belief none the less real or objective.(For reasons of euphony, let me revise that label and make it "the grounded abstract-statistical picture," acronym "GASP.")An i/e-ist who thinks in particular that intentional content depends on interpretation could not accept GASP, for the statistical properties of brain regions and their causal groundings are entirely independent of interpretation.

The last two sentences sound alarmingly like Dennett's protestations in "True Believers"[32] that his own interpretion-based view does not rob his concept of belief of its objectivity.Alarmingly, not because Dennett's view does not preserve objectivity (it does so, by making attitude content a matter of whether someone's taking a certain interpretive stance would in fact achieve good predictive results).Rather, because (a) for Dennett, attitude content is still conceptually constructed out of purpose-relative interpretation, in my view an unhealthy feature for a theory of content to have; (b) Dennett's objectivity does not amount to determinacy, and he explicitly endorses Quine's indeterminacy thesis for content; (c) for Dennett, a thing's innards (including the matter of whether it even has any) are conceptually irrelevant to its having intentional states; (d) Dennett himself has always used mathematical abstractions, such as centers of gravity, kinematic vectors, and the Equator, to illustrate the sort of view he favors, and in an important recent paper, "Real Patterns,"[33] he emphasizes the reality of mathematical abstractions and de-emphasizes interpretation.

I am all right as regards (a), (b), and (c).GASP has none of the three implications I find objectionable.In particular, abstract statistical properties of NETtalk's sort owe nothing to interpretation, even hypothetical interpretation; GASP-content is still determinate (so far as has been shown); and a believer's innards must have highly articulate structure in order to support the type of statistical abstraction to which I am appealing.But what of (d)?What distinguishes GASP from Dennett's view on that score? 

Actually I think we need to note what distinguishes Dennett's mathematical examples themselves from his view.I have always found his appeal to mathematical abstractions incongruous with his own underlying interpretivism.Centers of gravity and the Equator are real and as they are, independently of (even hypothetical) interpretation; they are determinate features of physical systems; and the systems whose features they are have, and must have, the physical structure needed to support the abstraction that reveals them.Thus, they do not well illustrate the basic position he actually defends.For that reason, I am not uncomfortable with (d) once (a)-(c) have been set aside.[34]

But to return to this paper's central theme:Arguments for representationalism often double or triple as arguments for physicalism about thinking and as defenses of the propositional-attitude part of folk psychology, and the three issues tend to get mixed up and run together.My main minimal point here is that if one simply assumes both physicalism about thinking and the probity of folk psychology, representationalism is, in light of the unboundedness of thinking, almost inevitable.Even better, if one defends both physicalism about thinking and the probity of folk psychology, representationalism is almost inevitable.If on the other hand one chooses to question either physicalism about thinking or the probity of folk psychology, one is fighting on another front entirely, and against very well entrenched and powerful opponents.[35]

Department of Philosophy

University of North Carolina

Chapel Hill, NC27599-3125

USA

FOOTNOTES



[1]Sellars, Science, Perception and Reality (London: Routledge and Kegan Paul, 1963), Science and Metaphysics (London: Routledge and Kegan Paul, 1968), and, to say the least, elsewhere; Fodor, The Language of Thought (New York: Crowell, 1975); RePresentations (Cambridge, MA: Bradford Books / MIT Press, 1981); Psychosemantics (Cambridge, MA: Bradford Books / MIT Press, 1987), and, t.s.t.l., e.See also Hartry Field, "Mental Representation," Erkenntnis, Vol. 13 (1978), and the papers collected in Part I of my Judgement and Justification (Cambridge: Cambridge University Press, 1988).

[2]This compositionality requirement is roughly what Brian Loar demands of any "language of thought" theory worth the name, in "Must Beliefs Be Sentences?" (in P. Asquith and T. Nickles (eds.), Proceedings of the PSA, 1982 (East Lansing, Michigan, 1983)).

[3]The most sustained criticism has been waged by D.C. Dennett, in Brainstorms (Montgomery, VT: Bradford Books, 1978) and The Intentional Stance (Cambridge, MA: Bradford Books / MIT Press, 1987).Paul Churchland and Patricia Churchland have battled also, especially in "Stalking the Wild Epistemic Engine" (Noûs 17 (1983)), but (as I read them) they argue more narrowly against a more specific form of representationalism than I am concerned to defend here.

[4]My favorite instance of this point is Quine's so-called "criterion of ontological commitment" (see particularly "On What There Is" and "Logic and the Reification of Universals," collected in From a Logical Point of View (New York: Harper Torchbooks, 1953))Some commentators have taken Quine to be defending an actually or potentially controversial philosophical doctrine.But such commentators misunderstand him entirely; if one is to grasp what he is saying??even if one ultimately disagrees??one must see that the "criterion" is supposed to be obvious insofar as it is true.(On this, see pp. 280-281 of my "The Trouble with Possible Worlds," in M. Loux (ed.), The Possible and the Actual (Ithaca, NY: Cornell University Press, 1979). 

[5]I have in mind here the ostensibly "new paradigm" afforded by contemporary Connectionism in AI.I shall turn to it in section 3 below.

[6]I originally entitled this paper "A New Argument...."But since first drafting it, I found something very close to my deductive argument on pp. 147-148 of Fodor's Psychosemantics, loc. cit., in Fodor's appeal to the "productivity" of intentional states.It had also been hinted at in Gilbert Harman's formal comment on Loar (op. cit.), and encapsulated in a sentence on p. 93 of John Haugeland's Artificial Intelligence: The Very Idea (Cambridge, MA: Bradford Books / MIT Press, 1985).And, more recently, see J.A. Fodor and Z.W. Pylyshyn, "Connectionism and Cognitive Architecture: A Critical Analysis," Cognition 28 (1988).

[7]Chomsky himself does not believe that understanding is literally, utterly unbounded.Recently he has argued that because the human language faculty has a rich innate structure, that structure imposes significant limits on what can be understood:

To the extent that we can discover the properties of the language faculty, we can construct `unlearnable languages,' languages that cannot be acquired by the language faculty because at every point it will make the wrong choices, the wrong guesses as to the nature of the language.(Language and Problems of Knowledge (Cambridge, MA:MIT Press, 1988), p. 150)


Indeed, Chomsky points out, any imaginable creature's language faculty would have to have some structure or other, and so could be crashed in the way described.(See also J.A. Fodor, The Modularity of Mind (Cambridge, MA: MIT Press / Bradford Books, 1983), pp. 120-126).But structural constraints of this sort do not affect Chomsky's argument for recursion, for they place no practical limits on what human beings can understand in the course of everyday life, and in particular they do not reduce the infinitude of competence.


[8]See Chomsky's review of Skinner's Verbal Behavior, Language 35 (1959).


[9]A few philosophers have bravely resisted Chomsky's argument:Elliott Sober, "Computability and Cognition," Synthese 39 (1978); Robert Matthews, "Learnability of Semantic Theory," in E. LePore (ed.), Truth and Interpretation (Oxford: Basil Blackwell, 1986); Stephen Schiffer, Remnants of Meaning (Cambridge, MA: Bradford Books / MIT Press, 1987).Schiffer actually constructs (secs. 7.4 and 7.5) an alternative mechanical description of how one might have an infinitary capacity for language understanding without embodying any Chomskyan or Davidsonian recursion.Personally I do not think Schiffer's method succeeds as advertised, for it works by a trick of translation that I believe conceals just such a recursion.But that matter need not be adjudicated here, for even if Schiffer's alternative does work for the interpretation of a public natural language, it will not carry over to the case of thinking.Thinking is not a matter of interpretation by translation of one idiom into another.

[10]The Intentional Stance, loc. cit., p. 149.

[11]It may be that there are some structural bounds on thinking, much as there may be such bounds on language acquisition and understanding.(Colin McGinn has gestured toward one possible type of bound, in The Problem.of Consciousness (Oxford: Blackwell, 1991.)But for the same reasons I offered in fn 7 above, limits of this sort do not affect the Chomskyan argument I am about to offer; thinking is unbounded for all practical purposes and for worlds upon worlds of impractical ones.

[12]This is not to deny that a nonphysicalist might adopt the Chomskyan model, so long as the nonphysicalist's idea of mental contents were sufficiently articulated.I thank Bill Alston for pointing out that the present claim offers no embarrassment to nonphysicalism per se.

[13]Philosophical Investigations (New York: Macmillan), p. 177.

[14]I have been content to leave the task largely to others, particularly Ruth Garrett Millikan in Language, Thought, and Other Biological Categories (Cambridge, MA: Bradford Books / MIT Press, 1984), and Jerry Fodor in "Psychosemantics," in Lycan (ed.), Mind and Cognition, loc. cit.See also Robert Stalnaker, Inquiry (Cambridge, MA: Bradford Books / MIT Press, 1984); Devitt and Sterelny, Language and Reality (Cambridge, MA: Bradford Books / MIT Press, 1987); Fodor, Psychosemantics (loc. cit., which criticizes "Psychosemantics" aforementioned); Dretske, Explaining Behavior (Cambridge, MA: Bradford Books / MIT Press, 1988); Fodor, A Theory of Content and Other Essays (Cambridge, MA: Bradford Books / MIT Press, 199x) and authors cited therein.For a prolegomenon to my own view, see my "Ideas of Representation," in D. Weissbord (ed.), Mind, Value, and Culture: Essays in Honor of E.M. Adams (Atascadero, CA: Ridgeview Publishing, 1989).

[15]Put to me by Stephen Schiffer.The second objection was raised by John Barker.

[16]For the record, I think all deductive arguments are special cases of inference to the best explanation; see Chs. 6 and 8 of my Judgement and Justification, loc. cit.

[17]But unhappily not intermediate in etymology: no Latin prefix stands comfortably between "ab-" and "de-."

[18]E.g., P.M. Churchland, "Eliminative Materialism and the Propositional Attitudes," Journal of Philosophy 78 (1981).

[19]For brief criticism of the main arguments, see my "Ideas of Representation," loc. cit.Briefly:The (in one way) scientific ill-behavedness of intentional phenomena per seis no reason for doubting the existence of intentional items; nor does the alleged poverty of folk psychology considered as science convince me that folk psychology has not succeeded brilliantly for a folk science, or that that success should not be taken at face value as licensing the standard abductive inference to the truth of the view, or that the success is not quite amazing considered on its own.And I am more certain that I myself and other people have thoughts than I am of the premises of any purely philosophical argument designed to convince me to the contrary.

[20]The genus includes at least Behaviorists, Wittgensteinians, hermeneuticists, and Dennett.

[21]In a formal comment on an earlier version of this paper, Robert Stalnaker offered a parody of my argument designed to show that (B) simply begs the question.If (B) were understood as a material conditional, it certainly would beg, being then a bare denial of the conjunction(I) & (II) & (III) & not-(IV).But I intend (B) as a considerably stronger, intensional conditional, backed by my "only game in town" argument preceding.

[22]See particularly J.L. McClelland, D.E. Rumelhart and the PDP Research Group, Parallel Distributed Processing: Explorations in the Microstructures of Cognition, Vols. 1 and 2(Cambridge, MA: Bradford Books / MIT Press, 1986).

[23]There is the intermediate psychological thesis that such-and-such behavioral capacities are explained by connectionist architecture actually realized in organisms' brains.

[24]"Localist" connectionist networks, in which individual units have specific propositional contents, are uncontroversially representors.But the argument can be made even for genuinely distributive "globalist" networks.See J.A. Fodor and Z.W. Pylyshyn, "Connectionism and Cognitive Architecture: A Critical Analysis," loc. cit.; Paul Smolensky, "On the Proper Treatment of Connectionism," Behavioral and Brain Sciences 11 (1988); my "Symbols, Subsymbols, Neurons," ibid., and other commentaries on Smolensky's paper; P.W. Bechtel's "Connectionism and the Philosophy of Mind: An Overview," and other essays in the Southern Journal of Philosophy, Supplementary Vol. XXVI (1988); and my "Homuncular Functionalism Meets PDP," in W. Ramsey, S. Stich and D. Rumelhart (eds.), Philosophy and Connectionist Theory (Hillsdale, NJ: Lawrence Erlbaum, 1991). 

[25]T.J. Sejnowski and C.R. Rosenberg, "NETtalk: A Parallel Network that Learns to Read Aloud," Johns Hopkins University EE and CS Technical Report, January, 1986, reprinted inJ.A. Anderson and E. Rosenfeld (eds.), Neurocomputing (Cambridge, MA: Bradfor