Intelligence, information processing and explanation S. SILVERS Tilburg University, Hogeschoollaan 225, Tilburg, the Netherlands (Received 25-I-1981)
Abstract. The concepts referred to in the title of this paper have been subjected to radical reappraisal in recent years as a result of seemingly disparate developments in psychology, the mathematical theory of computation, and the philosophy of science. The discussion is an attempt to draw the disparities together in order to show the conceptual connectedness of the notions of intelligence, information processing, and explanation as these occur in psychological theory. It is argued that behaviorism (the received theory in psychology during the first half of this century) has been shown to be inadequate as an explanatory thesis as regards our best prescientific intuitions with respect to cognition and intelligent action, the range of entities (i.e., the kinds of beings) to whose behavior the concepts of cognition can be appropriately applied, and distorts the criteria of scientific explanation in order to adhere to the canons of the Deductive-Nomological Model of explanation. The information processing (or computational) model of intelligence (or cognitive processes) is introduced historically by way of a brief discussion of the concept of a Turing machine and its antecedants in proof theory or metarnathematics. The I-P model is displayed in terms of three overlapping interpretations: computer similation studies of cognitive processes in humans, artificial intelligence (AI), and robotics. A brief discussion is devoted to some 'far-out' speculations concerning some of the implications of a general theory of intelligence. In the penultimate section the virtues of the I-P model are explored and praised vis-/~-vis post-Positivist criteria of scientific explanation. The final section suggests where the vices of the I-P model are located, viz., in the necessity but unlikelihood of a computational model of language translation since this would pressuppose a formalized yet global theory of interpretation.
During the first half of the 20th century, psychology, at least the kind of psychological theory favored by Anglo- and American psychologists played a dirty trick on us. This dirty trick consisted in trying to convince us, by all manner and means of experiment, that some of the concepts dearest not only to our hearts, but also and more pertinently to our minds, were nothing but ways of speaking. This behaviorist psychology heralded both the metaphysical and methodological presuppositions of the physical sciences by adopting them as their own, thus Acta Biotheoretica 30, 177-198 (1981) 0001-5342/81/0303-0177 $03.30 @ 1981 Martinus Niihoff Publishers, The Hague. Printed in the Netherlands.
178 acknowledging the verificationist epistemology which the Logical Empiricists were busily refining (eventually refining them out of existence). In so doing, behaviorist psychology taught us that the concept of the mental was scientifically irresponsible, a vestige of a naive and fantasy-laden philosophy of science. One consequence, among many, of this radical behaviorist theorizing was that features normally associated with mentality, such as intelligence, became looked upon as a specifically personal property(ies). Moreover, instead of providing an explanatory analysis of intelligence, the behaviorist approach characterized intelligence in terms of how they chose to measure it. Of course their choice of parameters for the measurement reflected the underlying assumption that the only acceptable kind of evidence for the determination of intelligence was the results of strange tests such as learning lists of nonsense syllables and the like. The reason nonsense syllables and the like were chosen to be the elements of behaviorist learning theory was their acknowledged lack of meaning. And, since cognition and intelligence were just the kinds of phenomena the behaviorists wished to explain without appealing to meaning (meaning, of course, is one of those concepts dearest to our minds which the behaviorists refused to acknowledge), learning, which presumably requires something like intelligence, had to be accounted for in terms of the conditions under which humans learned lists of nonsense and rats ran through mazes.It is peculiar to us now that such enormous effort was expended in testing for intelligence by having subjects do things that nobody would reasonably expect of an intelligent being. (Why would anybody think that you could tell how clever a being is by seeing how good he is at learning nonsense?) This is only to say that if the behavior theorists had given any thought to the concept of intelligent action, i.e. analyzed it in terms of its function in an individual, then perhaps the rote memorization of lists of nonsense would not have been conceived as learning and that results of such experiments would then not have been explained away in terms of schedules of stimulus and response reinforcement and the law of effect. So, rather than telling us what we could reasonably be expected to want to know about intelligence, behaviorism told rather how to measure it. And so it came to pass that cynicism replaced curiosity to the extent that intelligence became whatever it was that intelligence tests measured ... heardly enlightening. A crucial consequence of the behaviorist attitude was that the notion of intelligence was restricted (in the sense that it was looked upon as) to an individual rather than generalizable characteristic. Such an attitude, moreover, makes it exceedingly difficult to describe infra-human
179 behavior such that of infants, animals, and certain machines in terms of intelligence despite the fact that such behavior is readily, if intuitively, perceived as intelligent. What seems clear now is that the concept of an intelligent action is not one which includes a specification of the kind of being, which might satisfy the conditions (or criteria) for its application. Thus in addition to its many other failings 1, the behaviorist theory of learning simply could not explain what it was intended to, viz. actions which by virtue of their non-arbitrary character qualified as intelligent. Behaviorist explanation, for all its positivist virtue could not 'save the phenomena' it sought to explain, and sought to fill the lacunae in its explanations by the addition of ever more complicated ceteris paribus clauses. The explanatory scheme which came to supplant behaviorism as an account of intelligence and cognition is the information processing model, also known as the computational model of cognition.
I! To begin with a simple characterization of the information processing model of cognition, it may be said that the various kinds of typically mental or psychological phenomena associated with cognitive processes such as thinking, believing, deciding, and other such propositional attitudes (described by verbs of intentionality) are specified in terms of a system consisting of receptors and effectors for the receipt and discharge of signals from the environment to the system and then from the system back out into the environment, a processing unit of some kind, and a memory. The'processor performs operations on the outputs from the receptors and stores them in the memory; the processor also retrieves such stored data and processes it for discharge. It is obvious that the concept of a processor is made to do much work for it must together with the other elements of the system be capable of providing a convincing account of cognition and hence intelligent action. And it is also obvious that the I-P model must employ (systems of) representations i.e. symbolically coded descriptions or other means of representing the environmental situation insofar the receptors or sensors are selectively sensitive to that environment. Hence it can be seen that the I-P model seeks to explain cognitive processes in terms of the kinds of properties said to be characteristic of systems of representations. To see how this notion of the mind as a symbol manipulating system developed it is instructive to briefly examine its surprising history. The computational or information processing model of cognition has a most intriguing conceptual background, deriving from the much and J
r
180 justly heralded' work in metamathematics (or proof theory). The roots of the fundamental idea of calculability or computability as an analogue of (and hence theory of) cognition goes back to Hilbert's questions concerning the computability, completeness, and consistency with regard to proofs within axiomatic systems. The results of theorems established by GOdel (the incompleteness of any consistent axiom system rich enough to express the theory of numbers), Tarski (the impossibility of such a system's containing its own truth theorem), and Church (the impossibility of formally distinguishing theoremhood within a system) led to an examination of the issues of computability and completeness by A.M. Turing. Turing's idea, in effect was to approach the issue of the formal impossibility of devising an effective decision procedure for the solution of solvable formal problems by considering the thesis that any effective procedure which can be carried out by a human problem solver can be accomplished by a machine with a bare minimum of features. What has become known as a Turing machine is, in fact, an abstract set of operations to be carried out in a simple step-by-step procedure in accordance with a precise and unambiguous set of rules (of transformation). Any Turing machine can be fully described in terms of a quintuple of properties as follows: the machine is an abstract construct involving a finite number of internal configurations that are its machine states and these are determined in accordance with a set of transformation rules which define the states the machine is said to be in. Such a machine is usually portrayed as consisting in a length of tape divided in a series of squares, a reader or scanning device which reads the content o f the squares and a writer or mechanism for putting down or erasing symbols in the squares while the tape moves under the reader. The quintuple which defines the machine is then the current state, symbol read, next state, symbol written, direction of tape motion. The tape can be shuttled in both left and right directions, one section at a time. The machine operates in accordance with procedural rules such that if the machine is in a given state, i.e., a certain symbol (or no symbol) is printed in the square being read, then depending upon what symbol occurs in that square, the machine is unambiguously instructed to either change or not change the symbol being scanned and to shuttle the tape either to the left or right and to repeat the procedures. This brief excursion into the notion of a Turing machine is to point out that the set of transformation rules which determines the machine's operations is actually an abstract description of such a machine. And in general, any such well formulated set of transformation rules is a machine description regardless of the character of the realization of
181 those rules. This is the sense in which the notion of a Turing machine is actually a set of operating instructions or program. Moreover, when what appears on the tape (of a Turing machine), is an encoded description of another machine, in other words, an appropriately encoded set of transformation rules and data, which is itself a machine description, it follows that a Turing machine can, by virtue of its input, be made to imitate the operations of the machine whose description is given as input. Such a Turing machine T, which is capable of being transformed into any other machine M by following in a step-by-step procedure the operations that are carried out by M, is said to be a Universal Turing machine. Weizenbaum restates the results of Turing's 1936 paper, 'there exists a Turing machine U (actually a whole class of machines) whose alphabet consists of the two symbols '0' and '1' such that, given any procedure written in any precise and unambiguous language and a Turing machine L embodying the transformation rules of that language, Turing machine U can imitate the Turing machine L in L's execution of that procedure2. ' It follows immediately from this that a Turing machine can describe and imitate itself. With the advent of the idea of a universal Turing machine it eventually became clear that computational operations were not limited to the manipulation of numbers nor restricted to the domain of logic and mathematics. Nevertheless, there does still seem to be a rather popular notion that computers are number crunchers. But given the capacity to encode rules concerned with the manipulation of symbols of all kinds, computers can take as input symbolically represented information of any kind. It is this generality of purpose which allows for the limitless number of tasks to which computational machinery can be put in providing answers to difficult and important questions by simulating the behavior required for the solution to problems. Anything which can be symbolically represented and for which successive states are specifiable can be simulated by a computer. The connection between the developments in computer soence and the psychological analysis of cognition becomes clear in light of the concept of the modern computer as a general purpose symbol manipulating system. Thus both minds or cognitive processes in humans and general purpose symbol manipulators are characterized as instances or distinct realizations of the same underlying concept of an information processing system. The implementation of the idea that both machines and minds shared in the conceptual foundations of the theory of information processing was due largely to the well-known efforts of Newell, Simon and Shaw. The general goal of the I-P model,
182 however, was to specify the character of the flow of information through the system. The I-P model of cognition may be conveniently, if somewhat arbitrarily divided into three not terribly distinct categories: cognitive simulation studies, artificial intelligence, and robotics. Cognitive simulation (or CS) was the primary concern of a group of workers headed by Newell and Simon who where interested in accommodating hard-won experimental data on human problem solving performance to the I-P model. Ironically, perhaps, it was the attempt to simulate the cognitive processes involved in human problem solving performance that led to the at least partial demise of the project to explain intelligent action by simulating human performance. For it was soon found that whatever the crucial mark of human intelligence is, it seems to be dependent upon the kinds of subtle psychological phenomena associated with being sentient as well as sapient, and these features proved resistent to even the most sophisticated heuristic programming 3. Thus the somewhat exaggerated predictions that a computer program would soon be world chess champion fell considerably short of the mark; we still have Karpov. To be sure, the successes of the Logic Theorist (LT) and the General Problem Solver (GPS) programs are to be hailed, if only for confirming 'that heuristics, or rules of thumb, form an integral core of problem solving processes '4. A less empirically constrained version of the I-P model assumes a more abstract interpretation of the concept of intelligence, it has come to be known as AI for artificial intelligence. The distinction is that AI studies are not constrained by trying to model intelligence in terms of empirically observed results of human performance. AI is a 'top-down' analysis of the concept of intelligence which means that the analysis begins with a different set of priorities. 'In artificial intelligence (a discipline generally viewed as a top-down analysis of intelligent action), the primary goal is to discover the methods which are sufficient for some particular problem domain, and to postpone concern over a number of important questions, such as its detailed similarity to human mechanismsS. ' Here the aim is the explanation of intelligent action regardless of the stuff of the agent and its comparability with humans. As Pylyshyn notes, AI is interested in and methodologically constrained by the consideration of completeness of explanation and exchanges (or seeks to exchange) the incompleteness of the explanatory accounts of intelligence yielded by typically 'bottom-up' theories like behaviorism and what Dennet calls 'neuron signal physiological psychology' for a different kind of incompleteness of explanation, namely the 'incompleteness on specifiability of detailed correspondence with
183 experiments in favor of accounting for the possibility of certain performance skills.' (ibid. p. 429) The point is that where the adherence to experimental data in explaining intelligence sacrifices an explanatory account and hence tolerates a non-explanation for intelligent action (because the data do not explain the performance) the top-down h i strategy is to opt for an explanation for the intelligent performance even if that explanation is irrelevant to the experimental data. In this salient sense then AI lays claim to being a kind of non empirical analysis usually attributed to philosophy, and in particular to epistemologists who also pose topdown queries like 'how is knowledge possible?'. The AI version of such questions is 'how is it possible to develop a computer program which functions in such a way as to satisfy criteria which if satisfied by a human agent would be called intelligent.' The kinds of answers AI workers have provided have taken the form of constructing ever more sophisticated programs and hence AI may be fairly seen as epistemic engineering. Perhaps the clearest insight into the just noted engineering aspect of AI is to be found in robotics. Even the most sophisticated of programs like Winngrad's SHRDLU, a program for understanding natural language, which carries out instructions involving the spatial rearrangement of blocks of varying geometrical form, executes the task symbolically, i.e., by rearranging projected representations of the blocks, it does not physically move three-dimensional blocks. Robotics is concerned with an even more pronounced sense of intelligent action as performance; it is concerned with '... the problems associated with the building of machines that sense aspects of their environments, e.g. with the aid of television eyes, and that are capable of acting on it, e.g., by means of computer controlled mechanical hands and arms. (This) work has, as might be expected, generated a host of subproblems in such areas as vision, computer understanding of natural language, and pattern recognition6. ' An obvious consideration regarding the desirability or need of a discipline like robotics is that the concept of intelligent action is one which must be analyzed in terms of an agent's dealing with a continually changing environment, an environment, moreover, that changes as a result of the agent's interaction with it. This requires that sensory information of all kinds become processed appropriate to changing environmental needs. The most celebrated robot project to date (and perhaps for a long time to come because robotics research has become dominated by the demands of the industry) is the Stanford Research Institute's Shakey, a self-propelled robot on wheels. (Shakey has
184 become immobile, however, and is now an interesting vestige of past research interests. Hollywood and the toy industry keep our interest up now with R2D2 and the like) 7. Robotics is the one area of AI that really makes our eyes pop out, perhaps because of the obvious toy-like character of such devices which revive or continue our childhood fantasies, perhaps because in making robots in something like our own image we place ourselves in the ranks of the gods, or correspondingly perhaps because we are fascinated by the unknown limits of our own creativity. In any case we have a long history and rich literature on the relation of man and robot, recall Norbert Wiener's book God and Golem, Inc. The term 'Golem' comes from the name Joseph Golem which was given to a creature fashioned from the cl~iy from the banks of the river Moldau in Praag by Rabbi Judah Ben Loew in 1548 for the dual purpose of spying on the Gentiles to learn if there were to be pogroms and for doing janitorial work in the synagogue. Joseph Golem, however, got out of hand and had to be destroyed by his creator. Other Golems also appear in historical records. To bring these matters up to date it seems that a number of workers in AI claim 'that they grew up with a family tradition that they are descendants of Rabbi Loew, though they doubt that this belief has had much influence. Among them are Marvin Minsky and Joel Moses of M.I.T. Further, Moses tells (me) that a number of other American scientists have considered themselves to be descendants of Rabbi Loew, including John von Neumann... and Norbert Wiener .... '.'
III We have now seen 3 ways in which the information processing model of cognition has proliferated itself into so many dimensions of our ideas concerning the nature of intelligent action. This, of course, is an implicit requirement of any adequate theory; it must have broad powers of systematization. Before proceeding to the examination of the explanatory force of computational psychology, I shall, in this brief section pursue some of the consequences of the AI perspective as regards the notion of intelligence. The preceeding discussion began with a simple rehearsal of the metamathematical foundations of what was to eventually become the I-P model in psychology. Douglas Hofstadter has happily provided us with a thoroughgoing, thoroughly ingenious, and certainly thoroughly delightful excursion into the formal world of AI in his book G~del,
185
Escher, Bach: An Eternal Golden Braid 8. In it he offers several versions of the fundamental thesis of AI. Two are of particular interest: 'The Church-Turing Thesis, AI version: Mental processes of any sort can be simulated by a computer program whose underlying language is a power equal to that ... in which all partial recursive functions can be programmedL' It is this version of the thesis from which is derived the well-known functionalist thesis that if psychological predicates mark out natural psychological kinds then there is an infinite number if systems which will satisfy these predicates 1°. This identification of mental states with computational states is what distinguishes functionalist psychology from reductionist materialism; for a computational state may be realized by anything whatever given a system such that '... there is a correspondence between the states of one and the states of the other that preserves functional relations'~. ' A program then is a theory describing the autonomous psychological states of a system and a functional isomorphism between two systems will in general be established if the theory T is correct and there is a mapping of every 'property and relation defined in system 2 in such a way that T comes out true when all references to system 1 are replaced by references to system 2, and all property and relational symbols in T are reinterpreted according to the mappingS2. ' It follows from this somewhat libertarian point of view that intelligent action by whatever else we seek to constrain it, is not to be constrained by considerations about the kind of stuff constituent of the agent. 'Strange as it may seem to common sense and sophisticated intuition alike, the question of the autonomy of our mental life does not hinge on and has nothing to do with that all to popular, al to old question about matter or soul-stuff. We could be made of Swiss cheese and it wouldn't matter~3. ' A somewhat more pragmatic version of the AI thesis, in practice at least the version of AI that constitutes a motive force, is 'As the intelligence of machines evolves, its underlying mechanism will gradually converge to the mechanisms underlying human intelligence~'. ' This view has not only the libertarian consequences of the ChurchTuring thesis of intelligence, it goes considerably further in its implications. It follows, for example, that human intelligence is merely one interesting variation on a general theme of intelligence. Moreover, it goes without saying (although it is said more and more frequently these days in extra-theological circles) that human intelligence might very well lie beyond the scope of human understanding. That is to say, there may very well be perfectly good non-GOdelian reasons why we cannot ever come to fully understand human intelligence.
186 If we make the reasonable assumption that the seat of our peculair kind of intelligence is in the brain, then it just might turn out that its structure may not yield to an analysis in terms of even the strangest and most complicated kinds of recursive operations. In other words, the fundamental processes in the central nervous system may be such that the resultant processes which derive from them and which identify with being intelligent are simply not capable of 'understanding' their own underlying structures. The reasons need not be GOdelian since it might be that it is a brute fact about our brains or that stultifying boredom ensues long before any positive seeming analysis can be completed, or more generously, it may be that there simply isn't enought time in the life of the physical universe to permit us to complete the job ~s. Two views support this contention. The first is that the more we learn about the brain, the more we learn how complex it is and, correspondingly, how much time is needed, even given the most optimistic estimates, to obtain even partial results. Second, that the more advanced computational machinery becomes and the more flexible the programming languages are, the further away we move, both mechanically and conceptually, from the underlying fundamental computational operations in terms of which the instructions are executed. It is the rare programmer or writer of program languages who knows the program gets carried out in terms of machine languages and operations. But the story becomes really fascinating when we realize that if our own (and other) kind(s) of intelligence are not comprehensible to us for non-mysterious reasons, it would seem to follow that beings more intelligent than us may have a clearer picture than we do. The (only) trouble is, if this supposition holds, that we should very likely have considerable difficulty in recognizing such beings at all; or if we should be so perspicuously perceptive, the difficulty would then be to recognize their behavior as intelligent (could we 'understand' them?) and accommodate ourselves to it and them. These considerations conjure the consistently thinkable prospect that intelligence, assuming it to be the markedly general, abstract, and explanatory concept the various sciences of behavior suggest it is, may be embodied or realized in systems which are, quite literally, beyond our conceptualizability. But I hasten to remind you that in these conjuries we must be guided by what are by our lights the best indicators of what is to be counted as intelligent behavior. And this, I submit, is the essentially healthy and perhaps common sense view that the vague concept of science, that itself continues to defy satisfactory explication, is in Peirce's sense, our best bet; for it presupposes, as no other
187 alternative, the demand to know and hence to enquire, without presuming truth. There is, therefore, here no consolation for theology in the I-P model, except perhaps that possible instantiations of intelligence are not rejected a priori and hence prejudiciously.
IV The I-P model of cognition and intelligence attemps to explain the facts of cognition by assuming that intelligent action is the product of some sort of systematic processing of data such that those results which the system delivers will satisfy the criteria presumed to hold for intelligence. The analogy which is claimed to hold and hence provide an explanatory account of cognition is between the central nervous system and a computing device that the following kinds of analogues, for example, are said to obtain: neuron system (in the brain)--subsystem (in the computer), nervous conduction-information flow, sensor-terminal, sensory input--software input, memory-trace-memory (storage), mental function-information processing. It is readily apparent how the I-P model differs from its behavior predecessor as the paradigm (to use a perfectly good word which has suffered the slings and arrows of outrageous fortunes) of explanation in psychology. In the first and most obvious case, the I-P model explicitly assumes what the S-R model explicitly denies, viz., the existence of the facts of mental phenomena, i.e., of consciousness and all that is acknowledged to follow from the concept of mental life. I'll spare one and all the belaboring of the ridicule to which such a radical behaviorist view has been justly treated and suffice here with the marvelous comment of Ogden and Richards in their The Meaning of Meaning where they accuse behaviorism of obligating us to feign anaesthesia 1~. The I-P model then saves the observed phenomena to be explained, and in a sense that is certainly important as regards its claim to being a better explanation than behaviorism can offer. This is the dirty trick that I alluded to at the outset: behaviorism tried to convince us that we are all stupid. Secondly, like all good explanatory models, the I-P model is an account of how the phenomena to be explained come about, rather than why they occur. In the heyday of Logical Empiricism explanations were concerned with answers to causal questions interpreted as asking in a Humean way, why did event E occur. The answers were to be constrained by the Deductive-Nomological model which was then taken as explicating the Humean causal analysis, relieving science of the
188 burden of the search for causative forces and the like. But here, as elsewhere in philosophy, the resolution of a given problem is more likely to substitute one set of problems for another rather than actually dissolve the issue. Goodman's nettlesome question of distinguishing true, law-like but accidental generalizations from laws was but one outstanding consequence of the D-N model's presuppositions. Let us not suppose, however, that simply transferring from the 'why' to the 'how' mode of asking questions in science clears up the problem of causality 17. The point, however, is that in acknowledging psychological states, the I-P model focuses attention on what seems to be the obvious consideration, that given the wealth of neurophysiological activity which constitutes the human central nervous system, the question is, 'what is the nature of the connection between the non-intentional and subpersonal activities on the one hand and the intentional and personlevel activities we experience as consciousness and personal identity on the other'. In other words, how does all that non-intentional neural activity get transformed or 'translated' into the kinds of events for which intentional explanation seems irresistably appropriate. As I have formulated the I-P model's explanatory concern it is clear that the model shares with its behaviorist precursor the presumption of a physicalist-based theory of science; put somewhat more perspicuously, the I-P model is at least compatible with a physicalist metaphysics but it is distinct from behaviorism in being non-reductive. That is to say, it is not a feature of the computer model as an explanation of intelligent action that events to be explained are accounted for exclusively in terms of physical properties. As we have seen the model provides for an abstract characterization of cognition infinitely realizable. This leads directly to the third reason that can be adduced for the explanatory superiority of the I-P model, viz. that is provides for a far more general account of intelligence than behaviorism. It might also be noted here that with the advent of the I-P model both genuine intellectual challenge and fun find each other within the domain of tough-minded thinking about psychology. And if you think about it, while there is some genuine intellectual challenge involved with the attempt to establish that what for eons has suckered us all into believing was genius, clear-minded thought, creative imagination, and common sense is really nothing other than the literally stupid association of positive and negative reinforcings, it heardly seems a happy thing to do. What genuine excitement must have been derived from the attempt to reduce rationality to a tropism. (Whatever turns you on!). One wonders if the behaviorists viewed their own explanatory hypothesis as wellthought out and at the same time saw those explanations as just so
189 much meaningless verbal behavior. But the fun I speak of here and which I see as a worthy feature of scientific theories (which must satisfy us as well as our criteria) is the opening of the concept of cognition and in particular to the idea of intelligent action and seeing in that new access possibilities for the scientific exploration of commonalities among what were previously held to be unbridgable divergencies. For with such commonalities there develops a real potential for understanding the mechanisms underlying the production of intelligent behavior, whether such behavior is exhibited by human or infra-human entities. This increase in the generalizability of hypotheses regarding intelligence speaks strongly in favor of the I-P model as the best explanation. For it is a requirement of a successful explanatory scheme that it accounts for phenomena which its predecessor failed to do. The abstract and non-reductionist character of the I-P model allows for its realization in terms of any thinkable set of parameters, which as intellectual fun demands, presumes some deep thinking about the concept of intelligence and the kinds of systems and structures that exhibit it. Still another important consideration in the analysis of scientific explanatory force concerns the issue of the kind of language within which our theories are formulated and our explanations are couched. Regard, if you will, the sort of effort it would require, if instead of describing behavior as we are prone to do, in terms of the intelligence, rationality, and insight which lots of interesting behavior seems to exhibit, we were to adopt the rigid habits of the S-R theorist. The critics of Skinner (Chomsky, Dennett et al.), Ryle (Fodor), and even of Carnap (Chisholm) have given both the methodological and metaphysical arguments in favor of a radical behaviorist language a thorough thrashing and I will not rehearse the reading of that riot-act here. The point I wish to make is the broader one, akin to Putnam's argument against the adoption of a sense-datum language as the language of science. Putnam argues that a thing-language gains preference over sense-data languages in science because the very descriptions of the phenomena we seek to explain are already described in thing-language and we should have a difficult time indeed recognizing the observed phenomena that are to be accounted for if they were to be described in sensation language. In other words, the descriptive utility of languages presupposing material things is part of the argument for acknowledging a thing as scientifically acceptable. The question he poses is one which Feyerabend at an earlier moment in his career discussed, viz., how can we prefer the thing-language concept to that of the sense-data languages when there really is no suitable alternative
190 against which a thing-language can be tested and compared. As Popper too, has repeatedly reminded us, science is not marked by a method of empirically testing all testable hypothesis for there are simply too many, nor is it the search for the simplest testable theories since there are various and incompatible measures of simplicity of hypothesis. Hence the search for explanatory power presupposes 'some kind of a priori ordening of hypotheses '18 that are plausible from the standpoint of worthy of being tested, given what we already know. Of course, as with all philosophic arguments, the foregoing might very well obtain and it still be the case that we are completely mistaken about the existence of material things, scientific simplicity, and plausibility; we might still be just 'brains in a vat '19. But in the absence of devastating evidence and knock-out argument to the contrary, our most cherished and wellthought out intuitions clearly favor the acknowledgement of material things, not because material things are so wonderful but because they are so unavoidable.., we just keep on philosophically bumping into things that go bump in the night. What has been called a robust realism about physical things is not only a healthy common sense attitude toward the phenomena we observe, it is also the stuff of which the conceptual scheme in which we make observations consists. Applying this familiar post-Positivist reasoning to the question of the explanatory strength of the I-P model in psychology, it seems clear that the phenomena to be explained by theories of learning and perception are describable in terms of the cognitive processes of the agents displaying the described behavior. Harking back to the second point concerning the facts of cognition, the intimate relationship between acknowledgment of such facts and the kind of conceptual scheme within which such facts become irresistible, itself becomes transparent. Certainly Chisholm and Dennett have made good cases for the view that the non-intentional characterization of behavior patterns in learning, i.e., the non-intentional predictions of animal behavior are themselves based on fully intentional presuppositions concerning the normal behavior of animals; e.g., when they are hungry they want food and will do what they believe will bring it about, even if it is doing something stupid like pressing a bar which they have been artificially trained to associate with food. The point in all this is quite simply that, scientifically speaking, the language of intelligence, rationality, perception and the like form a background for the formulation and testing of theories of cognition in much the same manner as the language of things does for the testing of (plausible) hypotheses about other phenomena we observe. The question then is not always 'are the various kinds of behavior which we
191 observe to be intelligent really intelligent?' but rather 'how can we explain the phenomena which we observe to be intelligent?'. In other words, we seem irresistably 'conditioned' or inclined or disposed to characterize all kinds of actions and behaviors as intelligent on the basis of an admittedly a priori consideration of plausibility. What I am arguing then is that just as the only alternative to thing theory is a competent 'no-thing' theory so the only alternative to a theory of cognition is a theory of 'no-cognition' and that is what we all got in a big dose of behaviorism. But if behaviorists were trying to make psychology scientifically competent and responsible then it would have served their purpose far better to have had a clearer view of science. By that I mean that just as the totality of the scientific enterprise seems to converge asymptotically toward the factual existence of things in the sense that although there is no final confirmation of the existence of material things, the lack of a viable alternative lends reasonable scientific credence to the fact of material things, so too the convergence of separate disciplines toward the existence of intelligent action, in the absence of a viable rival suggests that would we do well to conclude to the view there is something to be called intelligent action. My claim thus far is that the I-P model is the best explanation of intelligence and cognition. It shares with realism the postulation of the kinds of ententies which fall outside the scope of observationality at least as this latter notion is defined by classicial empiricism, and moreover, it develops a theory which is intended to account for the kinds of phenomena we do observe (with our senses) in terms of the properties of the unobservables. Put so crudely, all this sounds incredible tendentious except for the fact that the history of science continues to evidence precisely this kind of account. It is no longer big news that the stuff of which empiricism is made is not the instrumentalist-idealist fluff that Carnap and Goodman, among others, unsuccessfully tried to reconstrue into physical objects. In other words, sensory experience is not the measure of the existence of material things. An important analogue to the points I have been laboring can be found in the theory of biological evolution. Explanatory hypotheses like the theory of special creation and spontaneous generation have been held to account for the totality of phenomena in the biosphere yet it seems clear that the evolutionary hypothesis is more than an explanatory account of some interesting but unrelated phenomena of biological development. Indeed, the theory of evolution attempts to almost literally provide the connecting links in the chain which it is believed ties the living world together. In short, the theory of evolution constitutes a framework of biosystematics. As the Medawars state,
192 Pedagogic 'proofs' of the past occurrence of evolution are of the same kind and unfortunately the same intellectual stature as those 'proofs' of the roundness of the earth which we learnt in our earliest schooldays. It is not upon these so-called 'proofs', however, that the acceptance of such a hypothesisdepends. It is rather that the hypothesis of evolution pervades, underlies and makes sense of the whole of biological science in much the same way as the idea of the roundness of the earth permeates the whole of geodesy, chronology, navigation and cosmology.The evolutionaryhypothesis is part of the very fabric of the way we think in biology. Only the hypothesis of evolution makes sense of the obvious inter-relationships between organisms, the phenomena of heridity and the pattern of development. For a biologist the alternative to thinking in evolutionary terms is not to think at all2°. The moral I wish to draw here is that for a behavioral scientist whether psychologist, anthropologist, ethologist or computer scientist, the alternative to thinking in terms of the concept of cognition and intelligence is not to think at all. Without rehearsing the points thus far let me at least collect the main considerations. I began by outlining a case against the behaviorist view of the psychology of intelligent action which does presuppose a reasonable familiarity with the underlying principles of behaviorist reasoning. Next the computational model was introduced in general terms and attention was focused on the differences and different virtues of cognitive simulation studies, AI, and robotics. Then I pleaded for the I-P model on the grounds of its being both methodologically and metaphysically a superior candidate as an explanatory theory than its predecessor. There are still other such meta-theoretical points to be made in favor of the computational approach to the theory of cognition concerning the nature of the relationship between explanation and understanding. These relate to the reasons for preferring explanation as redescription to explanation as deductive-nomological inference and preferring understanding as increased systematic coherence to understanding as deductive acumen. But these issues must be foresworn here as they carry us too far afield. With these many virtues in mind let me turn finally to some of the disabilities of the model and hence to an examination of its explanatory weakness.
V The I-P model of explanation is an intentional model insofar as it attemps to account for psychological states by reference to the goals of the system as well as the means available for subserving the achievement of those goals. It is thus a formal analogue of explanation by belief and desire. Nevertheless, it underwrites a mechanistic way of explaining as is
193 obvious from the discussion of Turing machines. This mechanistic explanation of mental operations leads to the charge that it is the point of the explanation of intelligent action to account for the purposefulness of such action without presuming the truth of mechanism. For this would beg every non-mechanist account of intelligence. But this charge overlooks the relevance of Church's theorem concerning the impossibility of a formal decision procedure, which amounts to the claim that anything computable is Turingmachine computable. For if intentional states are to be accounted for by reference to the cognitive processes underlying them then the truth of Church's theorem establishes that if that relationship can be computed it can be mechanistically explained. 'The constraint of mechanism is no more severe than the constraint of begging the question in psychology and who would wish to do that TM. Thus the I-P model is not obviously question begging. It is the point of the computational model to be an explanation of human intelligence in the sense of associating within the model analogues of the kinds of processes typically identified with mental functions, e.g., thinking, deciding, believing, understanding, etc. in short, the model provides analogues for human reasoning. Fodor has put it snappily, 'according to the model, deciding is a computational process; the act the agent performs is a consequence of computations defined over representations of possible actions. No representations, no computations, no computations, no model22.' He goes on immediately to say that he might have well said that 'the model presupposes a language' which makes it clear that the computational systems are to be taken as analogues to what has become popularly termed the languages of the brain. Given these considerations it is no wonder that the primary point of interest in computational psychology is with the character of the representational system, for the overriding concern is and must be with the properties of the various systems of representation which we know must be such that what is happening externally to the system can be read off from the internal representations. This overriding concern had lead some theoreticians (e.g.Z.W. Pylyshyn) to link the entire fate of cognitive psychology to the success of determining what is meant for two different mental representations to have the same content. The current view is that the computational system and the representational system are the same, i.e., that the structure of a formalized language is at once the system within which the external world is represented. Putnam refers to this union of the 'medium of computation and the medium of representation' as 'the working hypothesis of cognitive psychology today'23/24.
194
It is with this point that the explanatory quality of the computational model becomes manifest. Within the model there are said to be analogues of the brain languages and the explanation of the operations of the brain language is then in terms of the computational processes delineated in the model. Witness Hofstadter's view: 'In principle, I have no doubt that a totally reductionistic but incomprehensible explanation of the brain exists; the problem is how to translate (my italics) it into a language we ourselves can fathom. Surely we don't want a description in terms of the position and momenta of particles; we want a description which relates neural activity to 'signals' (intermediate level phenomena) and wich relates 'signals', in turn, to 'symbols' and subsystems, including the presumed-to-exist 'self-symbol'. This act of translation from low-level physical hardware to high-level psychological software is analogous to the translation of number-theoretical statements into metamathematical statements~S. '
Here then is a genuine case of question begging. For if the key to the explanation (and hence understanding) of intelligent action is to be found in the ever so elusive code which will translate the various brain languages into languages 'we ourselves can fathom' then that explanatory key will not be found. The computational model presupposes that translations are algorithms for the mapping of formal properties of classes of symbols onto one another. But there is serious reason to doubt that this kind of algorithmic translation is an appropriate analogue of the way that humans go about the business of translating natural languages. One part of the reason is empirical, viz., the acknowledged failures of the long-pursued project of machine translation o f natural language. The other part of the reason is conceptual, viz., the relationship between translation and understanding is not as is presumed by the computational model. If we can speak of brain language and translation then cognitive psychology must be careful in specifying the explanatory relationship between the concept of translation and the concept of understanding. For among others, the argument for the indeterminacy of translation makes a compelling case against the view that (a) translation is a piecemeal and context-free affair; that is it possible to develop a manual of translation by 'discovering' in a step-by-step way expressions which 'in fact' correspond, and (b) that understanding a language consists in being able to translate it into one we already know. For unless it is realized that what we behold is a language then what is beheld will just be some much noise (or if written, just so many marks). Even if we seem to understand something to be a language because we detect patterns which corrspond to what we recognize, it is still possible to be grossly misled as we certainly would be if we stumbled across a tree in the forest on which the
195 nooks and crannies in the bark (from a certain point of view) spelled out the Bill of Rights to the U.S. Constitution. In other words, we understand something to be a language not in vacuo and hence to the exclusion of other features which contribute to the recognition of language. It is precisely that our recognition of language and hence our attempt to translate it is something that occurs in a somewhat loosely differentiated context in which the elements, even presuming we could define them into necessary and sufficient conditions, meld or cohere and which give us reason to believe that it is a language (e.g. like that of Von Frisch's dancing bees) and makes the attempt at translation a plausible undertaking. If this is right, then the explanatory mechanisms proposed in the computer model, translation schemas, cannot be used to account for understanding and intelligent action because translation, if we are to make sense of the notion, must be assumed to occur within contexts in which a host of additional factors must be present which together contribute to the judgment that translation is a reasonable course to pursue. To understand something as a language is to perceive or interpret a range of behaviors as a whole as being of the kind in which language and hence translation appropriately fit, and this is clearly a 'holistic' endeavor. Such a holistic account is one in which explanation (and hence understanding) is achieved by the determination of what is to be expected to follow from the making of assumptions (or interpretations) regarding 'overall' behavior. Here we seem to be guided by what has been called the principle of charity (or humanity) to the effect that if the attemps at translation produce logically peculiar statements like 'my brother is an only child' then instead of ascribing such unusual beliefs to the speakers of such a language we assume that our translational hypotheses are badly construed. For instance, if it is a language we confront then we can expect to find analogies of the kinds of properties our languages have, e.g., communicative, descriptive, expressive, etc. and that its stock of logical truths is not radically different from ours. Another way of making this general point is to say that the difference between the kinds of explanation of intelligence envisaged by computational (functionalist) psychology and our hopefully healthy intuitions about the explanation of intelligence is this: the former accounts for it by restricting that concept to those areas of investigation in which the current and increasingly large body of information processing principles have shown themselves to be capable of yielding important and interesting insights, albeit partial ones, into crucial aspects of human cognition which had heretofore proven intransigent to
196 analysis. The latter which we might call interpretational explanation (Erklarendes Verstehen) imposes no such restrictions. In assuming the mind to be a computer-like device, a meat machine, as Minsky calls it, the best that can be expected is a description of how such computing systems and their inherent representations operate. This is, to be sure, no mean feat. But such descriptions, important as they are, by themselves are not explanations in any enlightening sense, unless accompanied by some kind of interpretation of the descriptive elements which enable us to accept those descriptions as significant and explanatory. But the rub is that the required interpretational skills are themselves not computable. The computational rules, as we have seen from the discussion of Turing machines, are instructions which the machine follows in an exceedingly local, step-by-step way. Interpretational explanation is not fixed in this way to a finite set of rules and data but is guided rather by varying interests and changing purposes and hence subject to continual modification. This point has also been made in this way: information processing psychology keeps on looking for the concept of intelligence under a streetlamp because that is the only place that is illuminated. The underlying consideration is that our best lights tell us that intelligent action cannot be separated off from the complex of opaque features of behavior that are associated with being the kind of entity for which it is appropriate to claim intelligence. In this it is important that the computational model takes care in postulating intelligent subsystems in order to explain intelligent systems for then very little has been explained. The suggestion that we might continue to search under the streetlamp for a means of developing a computational theory of interpretation meets the same frustrations, but then in spades, that were experienced by those who attempted to mechanize translation only to come up with results like 'invisible idiot' as a translation of 'out of sight, out of mind' and the well-known machine translation of 'the spirit is willing but the flesh is weak' into Russian as 'the vodka is drinkable but the meat is spoiled'. In sum then it would seem that there is a need to critically consider the tendency toward extravagant claims regarding the explanatory power of the computational model. I believe that such consideration involves argumentation about the interpretational and hence normative character of such concepts as intelligence, rationality, humanity, and personhood; what Locke referred to as the forensic nature of these notions, and the way these and other derivative ideas are connected is indeed the business of the philosopher 26.
197
This objection, however, is not meant to undermine the project of computational psychology but instead to put it in an Minsky-like frame or Shank-like script. To see this point as the death-knell of functionalist psychology because the I-P approach cannot be expected to explain all of what counts as cognition would be, as Hubert Dreyfus has done, to throw out the baby with the bath water. If there is a distinct discipline which we could call interpretation theory, then it would of necessity involve things like the principle of charity and hence of interest-relative explanation. What I should like to be clear about is that the analysis of a theory of interpretation is really not at cross-purposes with the computational analysis of mental states. Rather than implying one another and hence explaining one another the computational and interpretational theory of intelligence are guided by different principles and distinct goals. Whatever the character of the tensions between them, these approaches will probably still be around at some time in the far distant future (if such there be) when we know much more about the fine-structure of both.
Notes 1. D.C. Dennett, 'Skinner Skinned' in his Brainstorms, Bradford Books and Harvester Press, N.Y. and London, 1978. 2. Joseph Weizenbaum, Computer Power and Human Reason, W.H. Freeman, San Francisco, 1976, p. 62. 3. See K. Gunderson, Mentality and Machines, Doubleday Anchor Books, Garden City, N.Y., 1971, and H. Dreyfus, What computers can't do, Harper Colophon Books, N.Y., 1972, revised ed., 1979. 4. A. Newell and H. Simon 'Computer Simulation of Human Thinking', Science, vol. 134, Dec. 12, 1961, p. 12; quoted in H. Dreyfus, What computers can't do, rev. ed. op. tit. p. 94. 5. Z.W. Pylyshyn, Metaphorical Imprecision and the 'Top-Down' Research Strategy, in A. Ortony, ed., Metaphor and Thought, Cambridge University Press, N.Y., 1979, p. 428. 6. Joseph Weizenbaum, op. cit. p. 167. 7. Pamela McCorduck, Machines Who Think, W.H. Freeman, San Francisco, 1979, p. 13. 8. Basic Books, N.Y., 1979. 9. Ibid. p. 578-9. 10. See, H. Putnam, 'Minds and Machines', in his Mind, Language and Reality, Cambridge University Press, London, 1975 and the series of related articles therein; J.A. Fodor, Psychological Explanation, Random House, N.Y., 1968, The Language of Thought, Thomas Crowell, N.Y., 1975; N.R. Block, 'Troubles with Functionalism', in C.W. Savage, ed., Perception and Cognition, Minnesota Studies in the Philosophy of Science, Vol. IX, Univ., of Minnesota Press, Minneapolis, 1978. 11. H. Putnam, 'Philosophy and Our Mental Life' in Mind, Language and Reality, p. 291. 12. Ibid. p. 292. 13. Ibid. p. 291.
198 14. D. Hofstadter, op. cit. p. 579. 15. Cf. H. Putnam, Meaning and the Moral Sciences, Routledge & Kegan Paul, 1978 and D.C. Dennett, 'Mechanism and Responsibility' in Brainstorms, op. cit. 16. Cited in A.J. Ayer, The Central Questions of Philosophy, Weidenfeld and Nicolson, London, 1973, p. 127. 17. As Sylvan Bromberger's often cited article 'Why-Questions' in R. Colodny, ed. Mind and Cosmos, vol. III, in the University of Pittsburgh Series in the Philosophy of Science, University of Pittsburgh Press, 1966, illustrates. 18. H. Putnam, 'Language and Philosophy' in Mind, Language and Reality, p. 27. 19. H. Putnam, 'Realism and Reason' in his Meaning and the Moral Sciences, op. cit. 20. Peter and Jean Medawar, The Life Science, Harper Colophon Books, N.Y. 1977, p. 23 -24. 21. D.C. Dennett, 'Artificial Intelligence as Philososophy and Psychology' in Brainstorms op. cir. p. 112. 22. J.A. Fodor, The Language of Thought, op. cit. p. 31. 23. H. Putnam, 'Computational Psychology and Interpretation Theory', forthcoming, 24. It is in this sense that cognitive psychology has resuscitated something akin to the private language concept which an entire generation of analytic philosophers believed had been pronounced effectively dead by Wittgenstein. 25. G6del, Escher, Bach, op. cit. p. 709. 26. Cf. H. Putnam, 'Computational Psychology and Interpretation Theory', forthcoming, in which these issues are discussed.