Chapter Ten—
An Alternative Approach to Computational Psychology

The preceding chapters have presented an extended argument to the effect that CTM does not make good on either of its principal philosophical claims—that is, it provides neither an explanation of intentionality nor a vindication of intentional psychology. At best, we are left with a weakened, "bowdlerized" version of CTM (really a strong version of machine functionalism) that purports to describe the form of mental processes while treating the nature and legitimacy of intentional categories as a background assumption. Deprived of its impressive philosophical claims, CTM will seem to some to have lost its appeal entirely. And perhaps even more importantly, some writers seem drawn to infer that the failure of CTM as a philosophical project entails the bankruptcy of computational psychology as an empirical research strategy as well. Computationalism's critics and defenders alike often seem to assume that a successful scientific psychology ought to solve philosophical problems (like the mind-body problem) as well. On this view, the strong naturalization of the mental becomes a criterion for a successful scientific psychology, and even for the legitimacy of intentional categories. As a result, the failure of CTM to meet the philosopher's difficult metaphysical and explanatory criteria is thought to impugn the scientific enterprise of computational psychology out of which CTM arose as well.

I think this view is distinctly wrongheaded. (And, incidentally, I have yet to find a practicing scientist who shares it.) I make no claims about whether computational psychology will turn out to provide the foundations for a mature science of cognition, but it seems clear that what is

needed for a successful scientific research programme is substantially weaker than what is needed for a solution to the mind-body problem or a strong naturalization of the mental. I said in an earlier chapter that in the end we would need to distinguish between the claims of computational psychology as an empirical research programme and CTM's distinctly philosophical claims. I intend to make good on that claim in this final section of the book by presenting the outline of an alternative interpretation of the importance of computational psychology as an empirical research programme. On this view, what the computer metaphor tries to provide for psychology are two features that have widely been considered cardinal virtues of mature sciences: namely, (1) mathematically exact descriptions and explanations, and (2) strategies for connecting our discourse about mental events with other kinds of discourse—particularly "lower-level" descriptions such as those provided by neuroscience. It is presently unclear whether computational psychology will succeed in either of these goals, or whether it will perform better than other competitor theories arising from neuroscience or neural network approaches. But were it to succeed in these goals, even without strongly naturalizing the mental, this would count as substantive scientific progress, and in ways that have important precedents in the history of science. And all that is needed for this kind of progress is the kind of weak naturalization called for by BCTM. Strong naturalization, I shall argue, is not a requirement internal to the practice of science, but rather an externally imposed criterion deriving from a particular philosophical ideology. From the standpoint of the practicing scientist, it is not necessary to produce an instantiation analysis of mental states—a realization account is enough. And likewise for the scientist it is not necessary to vindicate things like perceptual gestalts—such things are assumed as data to be explained, and their status is never called into question.

10.1—
A Story about the Maturation of Sciences

I should like to begin by pointing to two features that seem to be common to the sciences that have traditionally been regarded as "mature" and "hard." The first characteristic of such sciences is that they have achieved a certain degree of rigor in their explanations. In particular, they have discovered mathematical expressions that capture, with greater or lesser degrees of exactitude, the relationships among the objects form-

ing the domain of a given science insofar as they take part in the phenomena that that science seeks to explain. The second characteristic of such sciences is that they involve what are sometimes called "structural explanations" or "microexplanations"—or, more generally, connections between domains of discourse. These explanations relate phenomena that occur at one level of description L1 to the objects, relationships, and processes at a more basic level of description L2 that are ultimately responsible for phenomena at L1 . Microexplanation can occur wholly within the bounds of a single science, and it can also occur across the boundaries of sciences, as in the case of the explanation of the combinatorial properties of the elements (chemistry) by reference to the behaviors of charged particles (physics). The issue of whether features such as these are necessary for scientific maturity is an important locus of contention in philosophy of science. Notably, there have been heated disputes about the status of biology in this regard. I wish to sidestep such issues here: I embrace Newton-Smith's (1981) idea that scientific theories can enjoy a plurality of "good-making" qualities. Mathematization and connectivity are two such qualities, and happen to be ones that have been emphasized in the "modern" view of science. My claim here is not that they are essential to a discipline's scientific status, nor that they are the only virtues relevant to scientific maturation, but merely that computational psychology may fruitfully be seen as an attempt to endow psychology with these virtues; and that if it succeeded in doing so, this would be a significant achievement.

A third feature that is often closely connected to the maturation of a science is the occurrence of a conceptual revolution that involves seeing the phenomena a science sets itself to describing and explaining in a fundamentally new way. The use of metaphor often plays a crucial role in such conceptual revolutions, though as often as not the metaphor is abandoned once rigorous mathematical description of the domain in its own right has been achieved. I am inclined to regard conceptual revolution and the use of guiding metaphor more as a feature of crucial stages in the process of maturation rather than a feature of mature sciences as such.^[1] (After all, conceptual change and the use of metaphor are just as much a part of attempts at science that never get off the ground as they are of successful science.) Thus the conclusion of previous chapters that computing machines provide only metaphorical inspiration for computational psychology is in keeping with the role of metaphor in other sciences as well.

A few examples of mathematization and microexplanation may be of

use in setting the stage for a discussion of psychology and computer science.

10.1.1—
Copernicus, Galileo, Newton

The emergence of the "new science" of the sixteenth and seventeenth centuries is sometimes referred to under the heading of the "Copernican revolution" in physics. And it is true that Copernicus played a crucial role in starting a conceptual revolution in astronomy that paved the way for the development of what was to become Newtonian mechanics. It is important to see, however, that Copernican astronomy in its own right is only the first step towards a mature physics.^[2] Copernicus's own concerns were still largely those of an astronomer . He was concerned with finding a description of planetary motion. His own model of that motion, however, was highly influenced by his Ptolemaic predecessors: a system of circular orbits and epicycles around a point close to the sun.^[3] (The sun was not at the center of Copernicus's system; it, like the planets, orbited another point in space. It is also worth noting that Copernicus's model contained more epicycles than Ptolemy's. The virtue of this model lay neither in its elegance nor in its predictive accuracy [see Kuhn 1957: 169-171].) Kepler, by contrast, was engaged in a project of finding a kind of mathematical description of planetary motion that would at once be elegant and exact. One important breakthrough—the one we probably imprinted upon when we learned about the progress of modern physics—was Kepler's discovery of the fact that planetary orbits are elliptical and that orbital speed can be determined on the basis of the area of the ellipse subtended by a portion of the orbit.^[4] But when we learned this, we probably overlooked what was really important about this discovery. The fact that orbits take the form of an ellipse rather than some other conic form is really irrelevant to the progress of physics. What is crucially important is that the motions of the planets can be described exactly by mathematical expressions, regardless of which ones, and that they can all be described by the same kinds of expressions.^[5] Physics would have done just as well if planetary orbits had been of a different, yet precisely describable shape. Celestial mechanics would have gotten nowhere so long as the only descriptions of planetary motion were in terms of a motley batch of epicycles having no discernible overarching pattern.

The further progress of modern physics was facilitated by the emergence of two other mathematical innovations: the development of algebraic geometry allowed for the possibility of performing algebraic cal-

culations upon the motions of the planets through time, and the calculus provided for the possibility of making calculations about acceleration. The culmination of these advances was Newtonian mechanics, which summarized the interactions of gravitational bodies in a set of extremely elegant mathematical equations that came to be known as "Newton's laws." Newtonian mechanics unified the fields of astronomy, celestial mechanics, and sublunary physics under one set of mathematical descriptions, and stood as the standard for scientific theories until displaced by relativity theory (which was, itself, dependent upon the development of differential geometry for its descriptions of space and time).

It is worth emphasizing that Newton's achievement lies in his having left us a rigorous and general description of the effects of gravitational bodies upon one another. His ambivalent attempts to address the why of gravity (his much-discussed flirtation with "forces") add nothing to the picture, and his failure to solve the "why" of gravity detracts not in the least from the power and the utility of Newtonian mechanics.^[6] One might well suspect that gravity amounts to something more than the empirical regularities of how bodies move in relation to one another, and it is appealing to seek some insight into this "something more," but such insight is not needed in order to make Newtonian mechanics "good science."

In brief, Newtonian mechanics provides for the mathematical maturity of a large portion of physics without providing any microexplanation for gravitational attraction in terms of some subgravitational level of explanation. Gravitation is treated as fundamental . And the lack of such an explanatory connection is not generally viewed as a fatal objection to Newtonian mechanics as good science, even though it is in some ways dissatisfying. Moreover, it displaced a Cartesian physics which did offer a microexplanation of gravitation in terms of mechanical interactions of particles.^[7]

10.1.2—
Chemistry

A second example can be supplied by chemistry, which experienced distinctly separate stages of progress towards mathematical and connective maturity. There was a time when chemistry was largely independent of physics. In fact, chemistry attained a remarkable degree of mathematical maturity with very little help from physics, and it is possible to learn large portions of chemistry with little or no knowledge of physics. (Indeed, I believe it is still the practice in teaching chemistry in the schools

first to present a chemistry that involves little or no physics before moving on to those parts where physics becomes crucial.) For there was substantial progress in understanding basic laws governing combinations of the elements before there was any underlying physical theory about what sorts of microstructure might account for these laws. The periodic table (Mendeleev around 1869), the notion of valence (Frankland in 1852), and a remarkable set of laws governing combinations of elements (as early as Lavoisier's work published in 1787) were developed long before these notions were further grounded in a theory of subatomic particles in the twentieth century. With the development of the periodic table, the notion of valence, and laws of combination, chemistry achieved a significant degree of mathematical maturity. In this respect, the age of Lavoisier made a significant step beyond the procedures and wisdom of previous chemists and alchemists, however great their technical skill, because there was, for the first time, a rigorous and systematic description of how the elements reacted in combination.

The major progresses in theoretical chemistry since the time of Mendeleev have been in the connections, the border marches, between chemistry and other disciplines such as physics and biology. In order to understand reactions between large molecules, for example, it was necessary to understand something of their physical structure. The propensities of molecules to combine in certain ways (some of which seemed anomalous) called for an explanation in terms of underlying structure, an explanation supplied by such notions as electron orbitals, ionic and covalent bonding, and the postulation of charged and uncharged subatomic particles. In the process, chemistry became increasingly connected to physics. At the same time, it became evident that many biological phenomena could be accounted for by chemical explanations: the bonding between hemoglobin and oxygen accounts for the transport of oxygen through the circulatory system to cells throughout the body (and hemoglobin's preference for bonding with carbon monoxide explains the ease of carbon monoxide suffocation); important parts of processes such as the Krebs cycle are chemical in nature; and of course the basic element underlying genetics, the DNA molecule, is typified by a particular molecular structure. With the advent of discoveries such as these, chemistry—which already had a high degree of mathematical maturity—acquired a large amount of what we might call connective maturity as well. It is worth noting, however, that most of these connections were not made until the latter half of the twentieth century, more than a century after

chemistry had gained significant mathematical maturity. Some things take time.

It seems safe to say that sciences tend to become both more powerful and more firmly established as they become more intimately connected with one another. Chemistry raises questions that physics has to answer, and provides answers to questions raised by biology. Astronomy provides a lab for physics to study things that cannot be reproduced here and now, and physics provides a lab for testing hypotheses about things that are too far away to investigate firsthand.

10.2—
The Appeal of a Mature Psychology

It should be abundantly clear that any developments that could bring all or part of psychology towards either or both of these kinds of maturity would be of major importance. Indeed, in the case of psychology, the absence of these kinds of maturity has been a large factor contributing to the widespread sentiment that psychology is not and perhaps cannot be a mature science. Consider first the matter of mathematization. Psychological explanation has traditionally been among the least systematic bodies of explanation among those disciplines that aspire to the name of science. Even higher-level disciplines such as economics have a stock of mathematical laws that describe their subject matter, even if only under "ideal" conditions. But while psychology has made inroads in terms of measurement of abilities (particularly in perceptual psychophysics), and seems susceptible to statistical generalizations over populations, the kind of explanation that takes place about and in terms of cognitive states has been notoriously resistant even to rough generalization, much less mathematization.^[8]

The situation is little better with connectivity. While it is the case that some higher-level disciplines such as economics proceed on assumptions about cognition (e.g., rational decision making), the connections between psychology and lower -level disciplines such as neurology and biology (not to mention physics) have been at once contentious and unedifying. On the one hand, there has long been almost universal agreement that there are systematic and "special" connections between mind and brain. Even Descartes, notorious to many as the arch-dualist, attributed a wide array of psychological processes to the brain and nervous system, reserving only language, reasoning, and the will for the immaterial soul.^[9] Descartes also viewed the connection between soul and body

as extremely intimate—far more so than that between a pilot and the ship he steers—and probably sui generis (Meditation VI [AT VII. 81]). On the other hand, the nature of such a "special connection" has been elusive both philosophically and empirically. The philosophers cannot seem to agree on what the precise nature of the "special connection" might (or must) be, and the empirical scientists have been hard pressed to discern what the elements on the neurological (physiological, physical) side of the relationship might be. If one of the marks of a mature psychology would be having discoveries of the form "Mental phenomenon M bears special relation R to neurological phenomenon N ," there seem to be two problems: the scientists cannot discover what N is, and the philosophers cannot decide what R has to be. To put it very mildly, it would be great progress if one could find a way beyond this old and frustrating impasse.

Now modern psychology has, in fact, made some progress on some fronts. There has been some significant quantification of perceptual psychophysics, and quantification of at least some of the observations in cognitive psychology. At the same time, neuroscience has emerged as a distinct offshoot of physiology that can draw upon other formal and empirical disciplines. Problems that were known to Helmholtz but unsolvable in his day are now solvable due to advances in mathematics (see Grossberg 1980). And the localization of mental functions in the brain has been greatly aided by more exacting and less intrusive observational techniques, such as those supplied by magnetic resonance imaging. But until recently the domain of cognition seemed largely untouched by these advances.

10.3—
Computation, Mathematization, and Connectivity

It is here that the computer paradigm may prove to be of considerable worth. What computer science provides is a rigorous set of terms and methods for talking about certain kinds of systems: systems whose distinctive characteristic is their functional organization. What can be characterized in functional terms can be described rigorously by computer science. Now in order for this to be of use to psychology, several things must be the case. First, psychological phenomena must be functionally describable. And here the sense of "function" is the technical mathematical sense. To put it differently, psychological phenomena must be such as to he describable by an algorithm or effective procedure expressible in the form of a machine table. Here computer science

supplies two things: first, a language (or set of languages) for the rigorous specification of algorithms; and second, an assurance that a very large class of algorithms (the finite ones) have a structure that can be instantiated by a physical system. So the computer paradigm might do two things here for psychology: it might provide a rigorous language for characterizing the system of causal interrelations between psychological phenomena, and at the same time provide assurance that this characterization can be realized in a physical mechanism that does not simply flout every law of nature. In short, the computer paradigm might provide the right tools for the mathematization of at least some part of psychology.

If computer science might directly provide the right tools for psychology to progress towards mathematical maturity, it might thereby indirectly provide an important contribution towards connective maturity as well. Of course, computation is not the right sort of notion to provide everything needed for connective maturity. Computation is an abstract or formal notion, and is therefore neutral, in important ways, about what sorts of things it describes. This is not to say that it does not itself specify functionally delimited kinds, but rather that in so doing it remains absolutely agnostic about (a ) what the nature of these kinds may be, apart from their formal interrelations, and (b ) how these functions are realized. A single computational description could apply equally well to a set of silicon chips, a network of cells, a structure of gears and levers, a set of galaxies, or the changes in affections of a Cartesian immaterial substance. Hence, even the best imaginable computational description of cognition would, in and of itself, do nothing about connecting psychology with other disciplines. For all that computational description buys us, it might still turn out that the things so described are not brain processes after all, but processes in an immaterial soul without even any analogous processes taking place in a brain. To be sure, the fact that computational structures can be physically instantiated is "bracing stuff" to someone who feels committed both to cognitivism and to materialism. It shows that the evidence for intentional realism may not be evidence against materialism, and vice versa. But the claim that cognitive processes are functionally describable has no consequences for the debate over whether materialism is correct in its ontological claims.

Nor does any computational description of cognitive processes have any direct consequences for how they are realized in the nervous system. This is, of course, a famous benefit of the computational approach: it allows for the possibility of the realization of equivalent functions in vastly

different architectures—human, angelic, Martian, or Macintosh. The strongest constraint the computational description might place upon the realizing system is that it share the functional structure characteristic of the realized cognitive process. That is, if cognitive phenomenon C is realized through a realizing system R , and C is characterized by functional structure F , it must be the case that R is also characterized by F .

But while this does not directly connect psychology with, say, neuroscience, it may provide just the sort of link that is needed to forge a connection between the two. The brain, after all, is a complex and bewildering set of interrelated units, and those who wander in its tractless wastes are constantly groping to discern what are the significant units and relations. The availability of careful characterizations of cognitive processes is the sort of thing that might serve, if not as a Rosetta stone for the brain, at least as a hastily scribbled map. Indeed, the grand appeal of the functionalist strategy in empirical psychology lies largely in the fact that starting "top-down" and unlocking black boxes one stage at a time has often seemed to be the only way one can proceed if one is interested in phenomena lying at a higher level than, say, on-center off-surround structures. As a somewhat idealized characterization, sometimes the only way to proceed is to get as clear as possible on the form of the process you wish to describe and then look for some candidate realizing system that has the right "shape" to match it.

It thus appears that progress towards mathematical maturity is one of the more likely roads towards connective maturity as well. The link between the two is not hard and fast: one might get a good descriptive functional psychology without making much progress in seeing how the functional structures are realized in the brain, much as we have no microexplanations for gravity or magnetism. But then again, progress in mathematization might bring connective progress in its wake, as combinatorial chemistry was eventually supplemented by a structurally oriented chemistry that is strongly linked to physics. One simply does not know in advance how the cards will fall.

10.4—
The Implicit Form of Cognitive Psychology

These considerations suggest an outline of how cognitive science proceeds and how it is related to intentional psychology. It is perhaps worth a brief digression to emphasize a few basic points. First, cognitive science is not

so much in the business of justifying psychological phenomena as trying to bring them to some clarity. In particular, it is concerned with developing models of mental states and processes that get their formal properties right, and in a fashion that is precisely statable. Second, to employ this sort of strategy, it is necessary to proceed from some precomputational understanding of mental states and processes. To be sure, the process of modeling often alters our precritical understanding of our subject matter (be it psychology or fluid dynamics); but we have to start from something like commonsense belief-desire psychology (or one of the precomputational attempts to make it more rigorous). The application of the computer paradigm is an attempt to clarify a mode of description and explanation we already use, and intentional states are involved in what one would normally take to be both the explanatory posits of psychology and the data to be explained.

Third, research tends to proceed top-down, from behavior and consciously entertained intentional states and processes, to hypotheses about underlying intentional states that could explain them, to mechanisms that could support such states and processes. Fourth, the initial specification of underlying mechanisms emphasizes their formal properties rather than their physical nature. Eventually one would wish to reach a stage where the formal properties necessary to explain some higher-order process are precisely those of simple physical mechanisms like neurons or even complex mechanisms like fields of interconnected neurons. But it is not clear how many intermediate formal "information-processing" levels are needed to mediate between intentional description and neurological description.

Finally, there are clearly distinct projects of mathematization and connection, and it might be possible to make progress in one without progress in the other. Notably, it might be possible to achieve considerable insight into the formal structure of cognition through computer modeling without thereby achieving much progress towards knowing how that structure is realized through brain tissue. It is therefore conceivable that computational research in psychology could produce mathematical maturity without connective maturity. On the other hand, it is similarly possible that neuroscience and connectionist research would produce models that would, to the great surprise of many, exhibit emergent formal properties that are much like those independently desirable for description of intentional states and processes, in which case progress in mathematization and connection might come together.

10.5—
Intentionality

Let us now make our discussion somewhat more concrete by looking at how one might apply the resources of computational psychology to the description—and, so far as possible, the explanation—of intentionality. In so doing, we shall take careful note of what kinds of connections are forged between domains and what kinds of explanation are actually likely to arise. It will be helpful to distinguish several different kinds of "accounts" that might be given, or perhaps several different kinds of description that might enter into a general account of intentionality. I wish to suggest that we may distinguish three separate components that an account of intentionality might have: (1) a "pure logical analysis" of intentionality, which describes the necessary structures of intentional states, (2) an abstract description of the formal properties of what is given in the logical analysis, and (3) an account of how the properties described in the vocabulary of the logical analysis are related to the realm of nature. These initial descriptions are necessarily a bit unclear, but will be expanded upon in the following pages.

10.5.1—
The Pure Logical Analysis of Intentionality

The topic of intentionality has received a great deal of attention in the century or so since Brentano (1874) reintroduced it into the European philosophical milieu. Much of this attention (e.g., by writers such as Brentano, Husserl, and most of the continental tradition, and writers like Chisholm and Searle in the English-speaking world) has been devoted to the examination of what one might call the "logical structure" of intentionality—that is, of properties that intentional states have just by virtue of being intentional states, or by virtue of being intentional states of a particular sort (e.g., judgments, conjectures, perceptual gestalts). A number of such properties stand immediately to the fore.

All intentional states involve an attitude-content structure.

Every intentional state is "directed towards" something—its "intentional object"—whether anything actually exists corresponding to that object or not.

Every intentional state is the intentional state of some intending subject.

Intentional states can have other intentional states as their intentional objects.

Every intentional state presents its object in some fashion or under some description (as being thus ) and under some intentional modality (judging, hoping, desiring, etc.).

Every intentional state has properties that determine its "conditions of satisfaction"—that is, that determine what would have to be true of the world in order for that state to be felicitous. And so on.

Features such as these are features of intentionality per se, and not features of intentionality that accrue to it specifically as it occurs in some particular kind of being. So whereas, for example, the claim that all desires are realized in brains is at best a contingent truth (it seems logically possible that things with different bodies—or perhaps even no bodies—could have desires), it is a necessary (and indeed analytic) truth about desire that every desire is a desire for something. The process of clarifying such features seems to be more a kind of analysis than empirical inquiry, and seems to be in large measure concerned with what might be called the "logical form" of intentional states—that is, the fact that they have an attitude-content structure, the fact that they posit an object or state of affairs under some description, and so on.

These "logical" properties of intentionality were given some attention by Brentano, and have been more carefully developed by writers like Roderick Chisholm (1957, 1968, 1984b), John Searle (1983), and particularly Edmund Husserl (1900, 1913), who devotes several volumes to the explication of intentionality.^[10] It might be appropriate to call this kind of description of intentionality "pure logical analysis" of intentionality. Husserl's expression "pure phenomenology" is also appropriate, though it may prove misleading to readers who associate the word 'phenomenology' with things having to do with qualitative feels and not with intentional states. What is properly suggested by the term involves the claims that (1) (occurrent) intentional states are things we experience, (2) they can also become the objects of our inquiry and analysis, (3) such intentional states "have a phenomenology" in the sense that features such as the attitude-content structure of intentional states are part of the "what-it's-like" (see Nagel 1974) of intentional states, and (4) these features can be discovered by phenomenological reflection. The "what-it'slike," of course, is not a qualitative "what-it's-like" (a "what-it-feels -like") but a logical "what-it's-like" (a "what-form-it-has"). Chisholm's linguistically based approach to intentionality is an attempt to attain greater clarity about mental states by attending to the logical forms of sentences used to report them. (Popular myths to the contrary notwith-

standing, the focus of Chisholm's interest is the intentionality of mental states, and he approaches them through an analysis of the sentences used to report them because of the difficulty of addressing the topic of mental states directly. Chisholm, like Husserl, thinks that phenomenology is difficult and elusive.)

In addition to the analysis of features common to all intentional states, this kind of pure analysis could reveal features peculiar to particular kinds of intentional states.

It is part of the very nature of PERCEPTUAL experiences that they set conditions of fulfillment involving a state of affairs in which something corresponding to the intentional object actually caused the state.

It is part of the essence of states having the modality of RECOLLECTION (things that present themselves as memories) that they be founded upon previously experienced immediate experiences of PERCEPTUAL PRESENTATION or some other intentional modality, and so on.

This kind of analysis of intentionality stands in some ways prior to the kind of investigation of the mind undertaken by computational psychology and BCTM. Computer modeling and artificial intelligence might, of course, provide very useful tools in pursuing such an analysis, as computer science provides ways of talking about inference and data structures that can greatly enrich one's ability to talk about logical form and conceptual relationships. It may also be that certain ideas that have emerged out of computer science (procedural representation, to name one notable example) may provide tools for the logical analysis of intentionality that would not otherwise have been available. But by and large, it is our intuitions about our mental states that constrain our computational descriptions, and not vice versa. If computational description is useful, it is useful in furthering a project to which we are already committed when we undertake the analysis of intentionality.

10.5.2—
The Formal Description of Intentionality

There is, however, a second level of description at which computational description might really add something new to an account of intentionality. For while traditional logical analyses yield numerous essential insights into intentionality, they tend to do very little to give an overarching model of how these insights fit together, and in particular they do not give the kind of model that would seem to be of much use in the project of building an empirical theory. Here the resources of computer sci-

ence may be of some use precisely in their ability to supply descriptions of the formal properties of certain kinds of systems. And the insights gained through logical and phenomenological analysis might be interpretable as formal constraints placed on a mathematical description of the "form" of intentional states and processes. This line of thought has been pursued by writers like Dreyfus and Hall (1984) and Haugeland (1978, 1981, 1985), who have seen a certain continuity between the Husserlian approach to intentionality and computer modeling. I shall not go into detail about where I agree and disagree with the analysis presented by these writers but shall supply a few examples of how I think this sort of intuition might be fleshed out.

(1) One insight to be gained from the logical analysis of intentionality is that intentional states can be about other intentional states. I can, for example, wish I could believe that my neighbor was trustworthy (WISH [BELIEF [my neighbor is trustworthy]]), or remember once having believed in the lost continent of Atlantis (RECOLLECTION [BELIEF [Atlantis exists]]). And such an insight is all very well and good, not to mention true. This same insight, however, can also be cashed out as a more interesting claim about the possible structures of intentional states: namely, that the structure permits of recursion. Or, to put it differently, if we were to give a formal description of the form of intentional states, it would have to involve a rule that allowed for recursion by embedding reference to one intentional state within the content of another. And since we have formal ways of talking about recursion, we have now taken a small step towards being able to say something about the abstract formal properties of intentionality. Such an insight might also provide the basis for other hypotheses—such as that the distinction between competence and performance can be applied to this embedding of intentional states, and that there might be general rules governing what intentional states can take particular other intentional states as arguments.

(2) Some insights gained from logical analysis take the form of either normative or productive rules concerning intentional states. For example, an analysis of the intentional modality of recollection reveals that it presents its object as having been previously experienced in some other intentional mode (e.g., perception). This sets normative constraints on the satisfaction of such a state: you cannot felicitously remember seeing Y unless you have at some previous time had a perceptual gestalt of Y . You can, however, experience a state whose intentional modality is RECOLLECTION and whose content is that of oneself having seen Y without actually having had a perceptual gestalt of Y in the past. (There are false

memories, after all.) So we would wish to describe our intentional processes in such a fashion that

(1) it is possible to experience RECOLLECTION [self having seen Y ] without having previously experienced PERCEPTUAL PRESENTATION [Y ], but

(2) the satisfaction conditions for RECOLLECTION [self having seen Y ] cannot be fulfilled unless PERCEPTUAL PRESENTAION [Y ] has previously been experienced.^[11]

Such rules can, of course, be characterized in terms of purely formal relationships expressed in the form of normative licensing rules (which set constraints on satisfaction conditions) and productive rules which describe what combinations of intentional states actually result in the generation of particular new intentional states.

(3) To take a somewhat different example, the analysis of intentionality may show us how to separate the issue of "being about something" in the sense conveyed by the opaque construal of intentional verbs from the issue of the fulfillment of such states in veridical intentional states. There is, I think, a good case to be made to the effect that, once this is done, we already have a mathematical format for talking about the fidelity of at least some intentional states (e.g., the perceptual ones): namely, the Mathematical Theory of Communication (MTC). Even if one is wary of the claims made by Sayre (1986) that one can build semantic content out of the technical notion of information employed in MTC, it nonetheless seems that MTC might be telling a perspicuous story about the difference between veridical perception and perceptual gestalts that result from illusions, hallucinations, and the like.

Now it is important to see how this story differs from some other stories about computers and the mind. The point here is not that intentional states are just functional relationships to symbols and hence precisely analogous to computing machines. The point, rather, is that there is a system of abstract properties to be found in the system of intentional states and processes, and these might very well be the same abstract properties that are being explored in computer science, in much the same sense that the calculus provided an appropriate set of mathematical forms for problems in classical mechanics. The question is that of finding the right description for the formal features of intentionality, and not that of whether anything sharing those formal features would thereby have intentionality as well. The answer to that latter question is surely no: there

will always be purely abstract objects having any given formal structure, and these do not have intentionality. And in general we should not expect any two isomorphic systems to be identical in all properties: for example, thermodynamics and Mathematical Theory of Communication share a formalism, but have different subject matters. The "intentionality" of symbols in computers may seem to track the intentionality of mental states, but only because symbols in computers are representations that have semiotic-meanings and hence are designed to express the mental-meanings of mental states.

To put matters somewhat differently, if we start with an analysis of intentionality and add the resources of computer science, we might end up with a useful set of formal constraints upon the shape intentional systems can take. On the other hand, if we merely start with formal properties, we will never develop notions such as mental-meaning out of those, and hence will never get intentionality as opposed to getting the formal shape shared by intentionality and perhaps any number of other things. Moreover, we need to start with our intuitions about intentionality to know which formal properties are relevant. There are many possible formal descriptions which might be interesting but are not viable as descriptions of cognition. The only way to get a formal description of intentionality is to start top-down from our intuitions about the intentional states we already know about—namely, our own—and study their formal "shape" by a process of abstraction.

10.5.3—
Intentionality and the Realm of Nature

We have thus far discussed two possible components of an account of intentionality: a pure logical analysis and a more mathematically perspicuous description of the formal relations revealed in the logical analysis. The remaining portion of such an account—and the portion that has seemed to be of greatest interest to writers in the philosophy of cognitive science—is an account of how intentional phenomena relate to the natural world. Of course, what many people really want is a way to see intentionality as itself being a natural phenomenon; but as we do not at this point know whether that is possible, it seems a bit strong a desideratum to set for an account of intentionality. It seems a more sober approach to begin by asking what can be done to relate intentionality to natural categories and then assess the relation between our conclusions and our previous metaphysical and methodological commitments.

It is, I think, agreed by almost everyone who believes in mental states

at all that there is some sort of special and intimate relationship between mental states and particular kinds of bodily states. What is open to investigation and dispute are the following three questions: (1) What is the right inventory of mental states? (To what extent is our common-sense inventory accurate? Are there "states" called "beliefs"? Was common-sense explanation ever intended to imply that there were, or is this an error of philosophical analysis, as implied by Wittgenstein and some in the continental camp ?) (2) What bodily states are thus "specially related" to particular mental states? And (3) what, precisely, does this "special relationship" consist in? I shall say very little about the first question here. Let the reader simply consider the remaining questions with regard to those mental states she does feel committed to.

Now the question of what bodily states particular mental states are "specially related" to seems to present a reasonable agenda for empirical psychology without shackling the psychologist to a burden of metaphysical proof. Empirical psychology can show such things as that there is a special relationship between C-fiber firings and the experience of pain. It cannot derive the qualitative state from a description of the physiology of C-fibers, nor from a description of how they interact with the rest of the body. And the result that C-fiber firings are "the physiological side of pain" is agreeable to philosophers who fall into very different metaphysical camps. Where they differ is on what to say about the precise nature of the relationship between C-fiber firings and the experience of pain: whether they are contingently identical, or that one supervenes on the other, or one causes the other, or that they are causally unrelated but perfectly coordinated by some preestablished harmony, and so on. And it seems quite clear (a ) that scientists do not, by and large, care about these further issues, and (b ) that, qua scientists, they are right not to care. (Consider what a burden upon science it would be if scientists waited until all the metaphysical disputes could be resolved!) And while the opacity of qualia to scientific analysis (i.e., the fact that you cannot "derive" qualitative states from neuroscience by way of something like an instantiation analysis) may seem a distressing anomaly to some, it is an anomaly that philosophers are, by and large, deciding that we have to live with.

I think the situation is very much the same with respect to intentionality. To spell matters out more explicitly: (1) Intentional states have a phenomenology, a "what-it's-like," though it is not a qualitative "whatit's-like" but a logical and semantical "what-it's-like." (2) Psychology might, in principle, be able to identify bodily states that are "specially

related" to intentional states such as occurrent judgments or perceptual gestalts. (3) If it can do this at all, it can do so in a scientifically respectable way without settling questions about ontology. (4) As argued in the previous chapter, the phenomenological element of intentional states is not subject to the kind of "strong naturalization" involving an instantiation analysis. And, hence, (5) intentionality is not subject to strong naturalization.

There are, of course, important differences between qualia like pain and intentional states. First, intentionality has a rich logical structure that pain lacks. And it is for this reason that a simple quality such as pain can be realized by a physiological mechanism with so few dimensions of freedom as the firings of certain kinds of nerve cells. A phenomenon such as judgment could not, even in principle, be realized through the firings of particular cells, because the physiological phenomenon involved does not have the right logical structure to support the logical structure of judgment. And here we have an important link between the formal analysis of intentionality and any account we might give of its realization: namely, that the analysis of intentionality places formal and in some cases causal constraints upon the kinds of mechanisms through which intentional states can be realized . This is the sort of issue that was being explored through the work done in "knowledge representation" by artificial intelligence researchers during the 1970s. It is also of fundamental importance to psychology, for the mathematical description of the mechanisms both specifies the functional properties that constitute it as a mechanism (and not an accidental by-product) and gives an important clue to identifying the tissue in which it is realized and the kinds of activity in that tissue that are of interest. (There can, or course, be other clues, such as evidence of activation through magnetic resonance scanning and the topology of neural connections.) And here too computer modeling (both conventional and connectionist) can be of crucial importance in determining whether a given architecture can support the formal features necessary to a particular kind of state or process.

10.5.4—
BCTM and Accounting for Intentionality

Given the foregoing analysis, what can BCTM and computational psychology do by way of providing an "account" of intentionality? The first thing they might be able to do is to supply a way of taking a pure logical analysis of intentionality of the sort offered by Brentano, Chisholm, Husserl, or Searle and teasing out a more rigorous description of the

formal properties of intentionality. This would be the kind of project that would move (intentional) psychology towards mathematical maturity. This, however, holds a further possibility: the analysis of intentionality places constraints upon the formal and causal features that a physical system must have in order to realize intentional states, and this might be of use in the project of providing a realization account for intentional states, thereby providing a measure of connective maturity for psychology as well.

Of course, whether this connective maturity could actually accrue to psychology is an empirical question. For a formal specification of intentionality would open the doors to a number of alternative possibilities. It seems to me that any of the following could turn out to be the case:

(1) Intentionality has formal properties that can be physically realized, and we can find mechanisms in the body that share those properties and whose activation is correlated with the experience of the corresponding intentional states.

(2) Intentionality has formal properties that cannot be realized by any physical system.

(3) Intentionality has formal properties that can be physically realized, but not by a digital machine (hence we need a noncomputational psychology if we are to provide a realization account for intentionality).

(4) Intentionality has formal properties that are not in fact shared by any mechanisms in the body, and hence at least some intentional states are not realized through bodily states.

(5) Intentional states are individually matched with physiological states sharing their formal properties, but this typing of states is not relevant to causal regularities.

I suspect that there are a number of other possibilities as well. But this selection should be sufficient to show that empirical study of intentionality could have some ramifications for metaphysics, albeit not definitive ones. If intentional states and physiological states are nicely correlated in a way that preserves causal regularities, a great number of ontological possibilities remain open. If intentionality has formal properties that cannot be realized by any physical system, intentionality and materialism are incompatible, and most dualists are likely to be surprised as well. (Perhaps Platonists or Kantians would find this possibility less jarring; I

am not sure.) It might be the case that something like the frame problem could be made to pose a case for something like (3), in which case we need some noncomputational approach to psychology. Possibility (4), again, opposes intentional realism and materialism, though again it might surprise dualists as well. And (5) might well be very welcome to both interactionists (who might want individual thoughts to have physical correlates through which the body is influenced while reserving the causal regularities for the nonmaterial soul) and epiphenomenalists. The kind of analysis I suggest thus does some limited metaphysical work, but not in a way that is question-begging and ideological: it is only by getting the best analysis of intentionality we can get, and seeing how it might match up with natural phenomena, that we really know what is at stake metaphysically in an account of intentionality.

The research programme associated with BCTM thus might do something very significant by way of providing an "account" of intentionality: it might render the logical analysis of intentionality formally perspicuous, and it might provide the key to a realization account of intentionality as well. What it does not do, of course, is produce an account of the nature of intentionality—of what it is to be an intentional state—in terms of some other kinds of categories (for instance, naturalistic ones). The key notions of "aboutness" and "(mental-)meaning" are left unexplained even if there should turn out to be some particular naturalistic relationships through which they are realized.

10.6—
A Reorientation in the Philosophy of Cognitive Science

It is my belief that the thumbnail sketch of computational psychology presented above presents an interesting scientific research programme that might turn out to produce theories with the two important scientific "good-making" qualities that I have described. The preceding sections should, I think, make it clear that it might endow psychology with these virtues even without providing a strong naturalization of the mental. Mathematization distills the form of a process by abstracting from the nature of the things that are related, and hence the mathematically reduced description does not provide sufficient conditions for the properties and objects it relates. Moreover, mathematization is methodologically dependent upon the prior assumption of the phenomena to be related by the mathematical descriptions.

It is likewise possible to form linkages between domains of discourse

that fall short of metaphysical sufficiency and conceptual adequacy. We can accept the wave-particle duality of matter without being able to derive one from the other, and we can accept that the detection of oriented lines is realized through columns in the V1 region of the visual cortex without being able to derive phenomenological properties from neural states, or vice versa. Such a situation leaves questions open, of course, and they are nagging questions. But (1) it is not clear at the outset whether they are really scientific questions or philosophical questions that ultimately turn upon the way we have misconstrued the problem, and (2) in the meantime, the existence of these questions does not impugn the progress that has been made along the way.

The big point here is that the needs of the philosopher and those of the empirical scientist diverge in important ways (see Horst 1992). And this is reflected in the ways we tend to regard one another's projects. In my experience, scientists tend to be utterly mystified at what philosophers could desire beyond (a ) a good model of mental processes, and (b ) localization of the functional units of the model in the nervous system. To them, the issue of whether the connections found are reductions, or supervenience relations, or merely empirically adequate generalizations is virtually unintelligible, and surely of no interest for the practice of science. I think they are right to think this way so long as the issue is one of empirical science . Empirical science aims at finding the regularities and connections that are there to be found, and seeks as strong an explanatory relation as it can uncover. Science is blind to distinctions beyond empirical adequacy, as such distinctions cannot be decided by experiment. And from the empirical perspective, it counts as progress to assert what you have found, even if you think you should have found something more. But from the empirical standpoint, it is also true that theoretical ideologies (say, that we must have a certain kind of explanation, or a certain view of the world) stand in the dock against the evidence of data and successful theories, and not just the other way around. The Cartesian view of the essence of matter as extension implied that there must be a mechanical explanation of light and gravitation, and indeed of all material phenomena. But better experimentation and better theories forced us to abandon Cartesian physics. Likewise, field theories now stand alongside theories of contact interactions. And teleological categories in biology are being accepted against an older mechanistic ideology. A psychology with sufficient internal good-making qualities that was not a strong naturalization would itself call the need for strong naturalization into question.

I therefore think that philosophers have been wrong to yoke the sci -

entific importance of computational psychology to its potential philosophical benefits. (And hence the failure of CTM to produce those philosophical results need not impugn computational psychology as an empirical research programme.) My alternative suggestion is that we look at what computational psychology might provide in the way of good-making qualities internal to scientific practice, and that mathematical and connective maturity stand out in this regard. This alternative has implications for how we ought to go about the philosophical study of cognitive science. For example, if this approach is the right approach, the best way to study cognitive science as science would be through careful case studies and comparisons between different theories (much the way one studies the history of any other science). Such an endeavor lies outside the scope of this book. But it is possible briefly to assess some of the comparative merits of my alternative construal of the importance of computational psychology with approaches that bind science and metaphysics in a tighter yoke. Here are several advantages that I think my alternative approach enjoys.

10.6.1—
A Better Description of Scientific Practice

First, the alternative approach is in closer accordance with the facts of scientific practice. The kinds of metaphysical and epistemic constraints required for strong naturalization are often missing in paradigm examples of good theories in other sciences. Scientific progress has often come, for example, in the form of laws relating several variables in the absence of any metaphysical necessity or conceptual adequacy relating those variables, and also in the absence of any microexplanation of why the law should hold. Most laws, when they were discovered at least, expressed relationships that were metaphysically contingent and epistemically opaque. If this is good enough for, say, Newton's laws, why should we hold psychology to a more stringent standard?

It is also plainly the case that, say, experimental psychologists and psychophysicists do not seem to feel a need to vindicate the phenomena they study (at any rate, no more than do practitioners of any other sciences), and that phenomenological and intentional description often play important roles in the initial description of problems that it is the job of theoretical psychology to solve. And the kind of "explanation" of, say, "seeing a red square" that is sought by a vision theorist does not require anything like metaphysically sufficient conditions.

Finally, one can point to anecdotal evidence from joint meetings in which psychologists are frustrated and baffled by the problems that divide

philosophers. My experience thus far has been that the characterization I have presented of where psychological concerns end and purely philosophical ones begin has been almost universally well received by psychologists, modelers, and neuroscientists, though often more controversial among philosophers.

10.6.2—
Ideology and Theory—Bad Precedents

Second, one can hardly be optimistic about the precedents for holding scientific theories to the litmus of a particular metaphysical or methodological ideology. Cartesian mechanism, Humean views on induction and causation, logical positivism, behaviorism—these stand as just a few prominent examples of research programmes that were based on metaphysical or methodological views with heavy theoretical implications. All of them seemed very compelling in their time, as they canonized metaphysical views or forms of explanation that held a firm grip on the imaginations of their day. But each was eventually eroded by successful scientific work that belied their metatheoretic assumptions. It is one thing to study the mutual influence of scientific theory and metaphysics or theories of scientific method. It is quite another to take a view like strong naturalism and use it as a test for scientific legitimacy. This kind of move has a poor track record. Better to look at the kinds of explanations that psychological theory does provide and draw one's conclusions from there.

10.6.3—
Scientific Progress without Naturalization

Third, as argued above, it is possible to have important kinds of scientific progress without also achieving interesting metaphysical results or solving traditional philosophical problems. A philosophy of cognitive science that is tied down to particular metaphysical goals is not free to assess the kind of good-making qualities that psychological theories may enjoy (and which perhaps some already do) that fall short of these goals. My approach, in contrast, emphasizes such achievements, while remaining open to the investigation of results with more metaphysical bite should they arise.

10.6.4—
Comparisons with other Research Programmes

Finally, my alternative approach yields a strategy for comparing "traditional" computational approaches to psychology that center on notions

of "rules" and "representations" with rival approaches arising from other sources such as neuroscience and neural network theories. For most of the time that the computer metaphor has been exploited in psychology, other approaches have also been explored, even if they have only recently been brought to the awareness of a broad audience of philosophers. For example, the Mathematical Theory of Communication of Shannon and Weaver (1949) has been explored by Sayre (1969, 1976, 1986) as an alternative basis for characterizing mental processes from the early 1960s to the present, and "neural network" approaches based on attempts to provide a mathematical characterization of the interactions of large numbers of cells in the brain were pioneered by McCulloch and Pitts (1943) and have been developed over the space of more than three decades by researchers such as Grossberg (1982) and Anderson (1973), as well as more recent researchers better known to philosophers such as Rumelhart and McClelland (1986) and Smolensky (1988).

Each of these approaches has its own preferred mathematical tools (sometimes including new mathematical machinery developed to solve particular problems) and its characteristic formally exact models of processes underlying cognition. Much of the debate between proponents of different models centers upon the features gained and lost by different kinds of mathematical apparatus: for example, the use of differential versus difference equations, or additive versus multiplicative shunting. (A good survey of mathematical differences in neural modeling is found in Levine [1991].) A second area of difference, both within the neural network camp and between its members and traditional artificial intelligence comes in the relationships assumed between the project of mathematical modeling and the project of connecting the model upwards to the data supplied by psychophysics and downward to that supplied by neuroscience. Some models are designed only to fit particular data curves, while others are intended additionally to be neurologically plausible.

In short, there is a great deal to be understood about the major research programmes in this area by looking at their mathematics, looking at their commitments to forming ties to other domains, and looking at their strategies for doing so. This kind of approach has some hope of shining light on individual theories, and also of clarifying the real differences between them. By contrast, most of what has come out of the philosophy of psychology with respect to neural networks so far has been centered on one of two issues: (1) whether connectionist theories are really the same as (or compatible with, or reducible to, or implementable

through) computational theories (and vice versa), and (2) how the availability of connectionist models vitiates Fodor's "only game in town" argument to the effect that we need intentional states because the only models we have to explain behavior require them as theoretical posits. Very little has been said of a philosophical nature about neural network theories in their own right, as opposed to how they compare to the kind of view espoused in CTM. Since I have argued that CTM does not in fact bear any philosophical fruit after all, I find comparisons with neural networks along that axis to be pretty much beside the point. Looking in detail at their mathematical repertoires and their commitments to kinds of interdomain connections in various directions, by contrast, gives us a concrete project in philosophy of science that we can sink our teeth into.

10.7—
Computation and Its Competition

This mention of competing research programmes provides a natural transition to a final point to be made in this chapter. What I have tried to provide here is an alternative approach to assessing the importance of the computational paradigm in psychology—a way of looking at the question, "If computational psychology is a successful research programme, what is it that it will have contributed to psychology?" My answer has been that computational psychology tries to endow psychology with two good-making qualities that have often been viewed as highly (even crucially) important to the maturation of sciences: namely, mathematical and connective maturity. But we should note that both the question and the answer are highly conditional: they concern what computational psychology would do if carried out successfully. Of course it is an open question whether it can or will be carried out successfully, so none of the preceding is meant as an endorsement of the computational approach to psychology as the right approach. It is an approach that is on the table, and as philosophers of science we are obliged to assess its promise.

At the same time, it is important to acknowledge that there are serious issues concerning the viability and prospects of the research programme. First, like any research programme, it may simply not succeed even on its own terms by failing to achieve any fundamental explanatory successes. Second, there have been serious arguments raised by writers like Dreyfus (1972) and Winograd and Flores (1986) to the effect that there are properties of the mind that symbol-manipulating systems cannot duplicate, and even more fundamental objections by writers like Ryle

(1949), Wittgenstein (1958), and the Verstehen tradition to the interpretation of psychological ascriptions now called the "theory-theory." I think these are all serious issues, and any one of them could turn out to have serious implications for the possible applications of rival models of the mind.

Finally, any successes of the computational approach to the mind in accordance with BCTM would also have to be assessed by comparison with the successes of rival research projects such as those arising out of neural network approaches or information theory. A brief list of issues might include but not be limited to:

The availability of exact mathematical descriptions for a wide variety of psychological phenomena.

The "naturalness" of these descriptions to their subject matter. (For example, the classical computational approach seems to have a natural way of approaching the attitude-content structure of intentional states. Do other approaches have an equally intuitive way of reflecting this feature in their models? Likewise, connectionist models seem naturally suited to modeling the behavior of fields of neurons, and information theory seems to have a natural way of talking about fidelity of intentional states.)

The comparative elegance of the models. (Can one approach supply straightforward descriptions and explanations where another requires a mass of ugly kludges?)

The tradeoffs between having a general framework (such as rule-conforming counter transformations or the technical notion of information) and having the freedom to employ an eclectic batch of mathematical tools.

The ways in which alternative research programmes are really competitors, and the extent to which they are ultimately compatible—because their formalisms turn out to be equivalent, for example, or because they are really engaged with different aspects of cognition or different questions about the mind.

The investigation of these questions will constitute a serious philosophical research programme in its own right, and will not be undertaken here. However, one important result of the discussion that has preceded in this book is the following: one might have thought that the approach to the mind found in CTM should enjoy pride of place over some of its

competitors because it solves certain philosophical problems (explaining intentionality and vindicating intentional psychology) that its competitors have no strategies for solving. But I have tried to argue that it fails to solve these problems, and that its true benefits lie in how it might provide virtues wholly internal to the science of psychology. But without the philosophical claims to confer pride of place upon CTM, there is a level playing field. We may now assess the comparative merits of CTM, connectionism, and neuroscience on wholly scientific grounds as scientific research programmes. This, I believe, effectively separates a set of questions about the philosophy of mind (such as the mind-body problem and the question of the precise metaphysical relationship between mental states and the bodily states through which they are realized) from questions about the science of the mind (such as what the important goodmaking qualities are for such a science). And this, I believe, is progress.